OpenCloudOS-Kernel/arch/s390/mm/gmap.c

2204 lines
63 KiB
C
Raw Normal View History

/*
* KVM guest address space mapping code
*
* Copyright IBM Corp. 2007, 2016
* Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
*/
#include <linux/kernel.h>
#include <linux/mm.h>
#include <linux/swap.h>
#include <linux/smp.h>
#include <linux/spinlock.h>
#include <linux/slab.h>
#include <linux/swapops.h>
#include <linux/ksm.h>
#include <linux/mman.h>
#include <asm/pgtable.h>
#include <asm/pgalloc.h>
#include <asm/gmap.h>
#include <asm/tlb.h>
s390/mm: support EDAT1 for gmap shadows If the guest is enabled for EDAT1, we can easily create shadows for guest2 -> guest3 provided tables that make use of EDAT1. If guest2 references a 1MB page, this memory looks consecutive for guest2, but it might not be so for us. Therefore we have to create fake page tables. We can easily add that to our existing infrastructure. The invalidation mechanism will make sure that fake page tables are removed when the parent table (sgt table entry) is changed. As EDAT1 also introduced protection on all page table levels, we have to also shadow these correctly. We don't have to care about: - ACCF-Validity Control in STE - Access-Control Bits in STE - Fetch-Protection Bit in STE - Common-Segment Bit in STE As all bits might be dropped and there is no guaranteed that they are active ("unpredictable whether the CPU uses these bits", "may be used"). Without using EDAT1 in the shadow ourselfes (STE-format control == 0), simply shadowing these bits would not be enough. They would be ignored. Please note that we are using the "fake" flag to make this look consistent with further changes (EDAT2, real-space designation support) and don't let the shadow functions handle fc=1 stes. In the future, with huge pages in the host, gmap_shadow_pgt() could simply try to map a huge host page if "fake" is set to one and indicate via return value that no lower fake tables / shadow ptes are required. Acked-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-04-18 19:24:52 +08:00
#define GMAP_SHADOW_FAKE_TABLE 1ULL
/**
* gmap_alloc - allocate and initialize a guest address space
* @mm: pointer to the parent mm_struct
* @limit: maximum address of the gmap address space
*
* Returns a guest address space structure.
*/
static struct gmap *gmap_alloc(unsigned long limit)
{
struct gmap *gmap;
struct page *page;
unsigned long *table;
unsigned long etype, atype;
if (limit < (1UL << 31)) {
limit = (1UL << 31) - 1;
atype = _ASCE_TYPE_SEGMENT;
etype = _SEGMENT_ENTRY_EMPTY;
} else if (limit < (1UL << 42)) {
limit = (1UL << 42) - 1;
atype = _ASCE_TYPE_REGION3;
etype = _REGION3_ENTRY_EMPTY;
} else if (limit < (1UL << 53)) {
limit = (1UL << 53) - 1;
atype = _ASCE_TYPE_REGION2;
etype = _REGION2_ENTRY_EMPTY;
} else {
limit = -1UL;
atype = _ASCE_TYPE_REGION1;
etype = _REGION1_ENTRY_EMPTY;
}
gmap = kzalloc(sizeof(struct gmap), GFP_KERNEL);
if (!gmap)
goto out;
INIT_LIST_HEAD(&gmap->crst_list);
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
INIT_LIST_HEAD(&gmap->children);
INIT_LIST_HEAD(&gmap->pt_list);
INIT_RADIX_TREE(&gmap->guest_to_host, GFP_KERNEL);
INIT_RADIX_TREE(&gmap->host_to_guest, GFP_ATOMIC);
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
INIT_RADIX_TREE(&gmap->host_to_rmap, GFP_ATOMIC);
spin_lock_init(&gmap->guest_table_lock);
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
spin_lock_init(&gmap->shadow_lock);
atomic_set(&gmap->ref_count, 1);
page = alloc_pages(GFP_KERNEL, 2);
if (!page)
goto out_free;
page->index = 0;
list_add(&page->lru, &gmap->crst_list);
table = (unsigned long *) page_to_phys(page);
crst_table_init(table, etype);
gmap->table = table;
gmap->asce = atype | _ASCE_TABLE_LENGTH |
_ASCE_USER_BITS | __pa(table);
gmap->asce_end = limit;
return gmap;
out_free:
kfree(gmap);
out:
return NULL;
}
/**
* gmap_create - create a guest address space
* @mm: pointer to the parent mm_struct
* @limit: maximum size of the gmap address space
*
* Returns a guest address space structure.
*/
struct gmap *gmap_create(struct mm_struct *mm, unsigned long limit)
{
struct gmap *gmap;
gmap = gmap_alloc(limit);
if (!gmap)
return NULL;
gmap->mm = mm;
spin_lock(&mm->context.gmap_lock);
list_add_rcu(&gmap->list, &mm->context.gmap_list);
spin_unlock(&mm->context.gmap_lock);
return gmap;
}
EXPORT_SYMBOL_GPL(gmap_create);
static void gmap_flush_tlb(struct gmap *gmap)
{
if (MACHINE_HAS_IDTE)
__tlb_flush_idte(gmap->asce);
else
__tlb_flush_global();
}
static void gmap_radix_tree_free(struct radix_tree_root *root)
{
struct radix_tree_iter iter;
unsigned long indices[16];
unsigned long index;
void **slot;
int i, nr;
/* A radix tree is freed by deleting all of its entries */
index = 0;
do {
nr = 0;
radix_tree_for_each_slot(slot, root, &iter, index) {
indices[nr] = iter.index;
if (++nr == 16)
break;
}
for (i = 0; i < nr; i++) {
index = indices[i];
radix_tree_delete(root, index);
}
} while (nr > 0);
}
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
static void gmap_rmap_radix_tree_free(struct radix_tree_root *root)
{
struct gmap_rmap *rmap, *rnext, *head;
struct radix_tree_iter iter;
unsigned long indices[16];
unsigned long index;
void **slot;
int i, nr;
/* A radix tree is freed by deleting all of its entries */
index = 0;
do {
nr = 0;
radix_tree_for_each_slot(slot, root, &iter, index) {
indices[nr] = iter.index;
if (++nr == 16)
break;
}
for (i = 0; i < nr; i++) {
index = indices[i];
head = radix_tree_delete(root, index);
gmap_for_each_rmap_safe(rmap, rnext, head)
kfree(rmap);
}
} while (nr > 0);
}
/**
* gmap_free - free a guest address space
* @gmap: pointer to the guest address space structure
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
*
* No locks required. There are no references to this gmap anymore.
*/
static void gmap_free(struct gmap *gmap)
{
struct page *page, *next;
/* Flush tlb of all gmaps (if not already done for shadows) */
if (!(gmap_is_shadow(gmap) && gmap->removed))
gmap_flush_tlb(gmap);
/* Free all segment & region tables. */
list_for_each_entry_safe(page, next, &gmap->crst_list, lru)
__free_pages(page, 2);
gmap_radix_tree_free(&gmap->guest_to_host);
gmap_radix_tree_free(&gmap->host_to_guest);
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
/* Free additional data for a shadow gmap */
if (gmap_is_shadow(gmap)) {
/* Free all page tables. */
list_for_each_entry_safe(page, next, &gmap->pt_list, lru)
page_table_free_pgste(page);
gmap_rmap_radix_tree_free(&gmap->host_to_rmap);
/* Release reference to the parent */
gmap_put(gmap->parent);
}
kfree(gmap);
}
/**
* gmap_get - increase reference counter for guest address space
* @gmap: pointer to the guest address space structure
*
* Returns the gmap pointer
*/
struct gmap *gmap_get(struct gmap *gmap)
{
atomic_inc(&gmap->ref_count);
return gmap;
}
EXPORT_SYMBOL_GPL(gmap_get);
/**
* gmap_put - decrease reference counter for guest address space
* @gmap: pointer to the guest address space structure
*
* If the reference counter reaches zero the guest address space is freed.
*/
void gmap_put(struct gmap *gmap)
{
if (atomic_dec_return(&gmap->ref_count) == 0)
gmap_free(gmap);
}
EXPORT_SYMBOL_GPL(gmap_put);
/**
* gmap_remove - remove a guest address space but do not free it yet
* @gmap: pointer to the guest address space structure
*/
void gmap_remove(struct gmap *gmap)
{
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
struct gmap *sg, *next;
/* Remove all shadow gmaps linked to this gmap */
if (!list_empty(&gmap->children)) {
spin_lock(&gmap->shadow_lock);
list_for_each_entry_safe(sg, next, &gmap->children, list) {
list_del(&sg->list);
gmap_put(sg);
}
spin_unlock(&gmap->shadow_lock);
}
/* Remove gmap from the pre-mm list */
spin_lock(&gmap->mm->context.gmap_lock);
list_del_rcu(&gmap->list);
spin_unlock(&gmap->mm->context.gmap_lock);
synchronize_rcu();
/* Put reference */
gmap_put(gmap);
}
EXPORT_SYMBOL_GPL(gmap_remove);
/**
* gmap_enable - switch primary space to the guest address space
* @gmap: pointer to the guest address space structure
*/
void gmap_enable(struct gmap *gmap)
{
S390_lowcore.gmap = (unsigned long) gmap;
}
EXPORT_SYMBOL_GPL(gmap_enable);
/**
* gmap_disable - switch back to the standard primary address space
* @gmap: pointer to the guest address space structure
*/
void gmap_disable(struct gmap *gmap)
{
S390_lowcore.gmap = 0UL;
}
EXPORT_SYMBOL_GPL(gmap_disable);
/**
* gmap_get_enabled - get a pointer to the currently enabled gmap
*
* Returns a pointer to the currently enabled gmap. 0 if none is enabled.
*/
struct gmap *gmap_get_enabled(void)
{
return (struct gmap *) S390_lowcore.gmap;
}
EXPORT_SYMBOL_GPL(gmap_get_enabled);
/*
* gmap_alloc_table is assumed to be called with mmap_sem held
*/
static int gmap_alloc_table(struct gmap *gmap, unsigned long *table,
unsigned long init, unsigned long gaddr)
{
struct page *page;
unsigned long *new;
/* since we dont free the gmap table until gmap_free we can unlock */
page = alloc_pages(GFP_KERNEL, 2);
if (!page)
return -ENOMEM;
new = (unsigned long *) page_to_phys(page);
crst_table_init(new, init);
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
spin_lock(&gmap->guest_table_lock);
if (*table & _REGION_ENTRY_INVALID) {
list_add(&page->lru, &gmap->crst_list);
*table = (unsigned long) new | _REGION_ENTRY_LENGTH |
(*table & _REGION_ENTRY_TYPE_MASK);
page->index = gaddr;
page = NULL;
}
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
spin_unlock(&gmap->guest_table_lock);
if (page)
__free_pages(page, 2);
return 0;
}
/**
* __gmap_segment_gaddr - find virtual address from segment pointer
* @entry: pointer to a segment table entry in the guest address space
*
* Returns the virtual address in the guest address space for the segment
*/
static unsigned long __gmap_segment_gaddr(unsigned long *entry)
{
struct page *page;
unsigned long offset, mask;
offset = (unsigned long) entry / sizeof(unsigned long);
offset = (offset & (PTRS_PER_PMD - 1)) * PMD_SIZE;
mask = ~(PTRS_PER_PMD * sizeof(pmd_t) - 1);
page = virt_to_page((void *)((unsigned long) entry & mask));
return page->index + offset;
}
/**
* __gmap_unlink_by_vmaddr - unlink a single segment via a host address
* @gmap: pointer to the guest address space structure
* @vmaddr: address in the host process address space
*
* Returns 1 if a TLB flush is required
*/
static int __gmap_unlink_by_vmaddr(struct gmap *gmap, unsigned long vmaddr)
{
unsigned long *entry;
int flush = 0;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
BUG_ON(gmap_is_shadow(gmap));
spin_lock(&gmap->guest_table_lock);
entry = radix_tree_delete(&gmap->host_to_guest, vmaddr >> PMD_SHIFT);
if (entry) {
flush = (*entry != _SEGMENT_ENTRY_INVALID);
*entry = _SEGMENT_ENTRY_INVALID;
}
spin_unlock(&gmap->guest_table_lock);
return flush;
}
/**
* __gmap_unmap_by_gaddr - unmap a single segment via a guest address
* @gmap: pointer to the guest address space structure
* @gaddr: address in the guest address space
*
* Returns 1 if a TLB flush is required
*/
static int __gmap_unmap_by_gaddr(struct gmap *gmap, unsigned long gaddr)
{
unsigned long vmaddr;
vmaddr = (unsigned long) radix_tree_delete(&gmap->guest_to_host,
gaddr >> PMD_SHIFT);
return vmaddr ? __gmap_unlink_by_vmaddr(gmap, vmaddr) : 0;
}
/**
* gmap_unmap_segment - unmap segment from the guest address space
* @gmap: pointer to the guest address space structure
* @to: address in the guest address space
* @len: length of the memory area to unmap
*
* Returns 0 if the unmap succeeded, -EINVAL if not.
*/
int gmap_unmap_segment(struct gmap *gmap, unsigned long to, unsigned long len)
{
unsigned long off;
int flush;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
BUG_ON(gmap_is_shadow(gmap));
if ((to | len) & (PMD_SIZE - 1))
return -EINVAL;
if (len == 0 || to + len < to)
return -EINVAL;
flush = 0;
down_write(&gmap->mm->mmap_sem);
for (off = 0; off < len; off += PMD_SIZE)
flush |= __gmap_unmap_by_gaddr(gmap, to + off);
up_write(&gmap->mm->mmap_sem);
if (flush)
gmap_flush_tlb(gmap);
return 0;
}
EXPORT_SYMBOL_GPL(gmap_unmap_segment);
/**
* gmap_map_segment - map a segment to the guest address space
* @gmap: pointer to the guest address space structure
* @from: source address in the parent address space
* @to: target address in the guest address space
* @len: length of the memory area to map
*
* Returns 0 if the mmap succeeded, -EINVAL or -ENOMEM if not.
*/
int gmap_map_segment(struct gmap *gmap, unsigned long from,
unsigned long to, unsigned long len)
{
unsigned long off;
int flush;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
BUG_ON(gmap_is_shadow(gmap));
if ((from | to | len) & (PMD_SIZE - 1))
return -EINVAL;
if (len == 0 || from + len < from || to + len < to ||
from + len - 1 > TASK_MAX_SIZE || to + len - 1 > gmap->asce_end)
return -EINVAL;
flush = 0;
down_write(&gmap->mm->mmap_sem);
for (off = 0; off < len; off += PMD_SIZE) {
/* Remove old translation */
flush |= __gmap_unmap_by_gaddr(gmap, to + off);
/* Store new translation */
if (radix_tree_insert(&gmap->guest_to_host,
(to + off) >> PMD_SHIFT,
(void *) from + off))
break;
}
up_write(&gmap->mm->mmap_sem);
if (flush)
gmap_flush_tlb(gmap);
if (off >= len)
return 0;
gmap_unmap_segment(gmap, to, len);
return -ENOMEM;
}
EXPORT_SYMBOL_GPL(gmap_map_segment);
/**
* __gmap_translate - translate a guest address to a user space address
* @gmap: pointer to guest mapping meta data structure
* @gaddr: guest address
*
* Returns user space address which corresponds to the guest address or
* -EFAULT if no such mapping exists.
* This function does not establish potentially missing page table entries.
* The mmap_sem of the mm that belongs to the address space must be held
* when this function gets called.
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
*
* Note: Can also be called for shadow gmaps.
*/
unsigned long __gmap_translate(struct gmap *gmap, unsigned long gaddr)
{
unsigned long vmaddr;
vmaddr = (unsigned long)
radix_tree_lookup(&gmap->guest_to_host, gaddr >> PMD_SHIFT);
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
/* Note: guest_to_host is empty for a shadow gmap */
return vmaddr ? (vmaddr | (gaddr & ~PMD_MASK)) : -EFAULT;
}
EXPORT_SYMBOL_GPL(__gmap_translate);
/**
* gmap_translate - translate a guest address to a user space address
* @gmap: pointer to guest mapping meta data structure
* @gaddr: guest address
*
* Returns user space address which corresponds to the guest address or
* -EFAULT if no such mapping exists.
* This function does not establish potentially missing page table entries.
*/
unsigned long gmap_translate(struct gmap *gmap, unsigned long gaddr)
{
unsigned long rc;
down_read(&gmap->mm->mmap_sem);
rc = __gmap_translate(gmap, gaddr);
up_read(&gmap->mm->mmap_sem);
return rc;
}
EXPORT_SYMBOL_GPL(gmap_translate);
/**
* gmap_unlink - disconnect a page table from the gmap shadow tables
* @gmap: pointer to guest mapping meta data structure
* @table: pointer to the host page table
* @vmaddr: vm address associated with the host page table
*/
void gmap_unlink(struct mm_struct *mm, unsigned long *table,
unsigned long vmaddr)
{
struct gmap *gmap;
int flush;
rcu_read_lock();
list_for_each_entry_rcu(gmap, &mm->context.gmap_list, list) {
flush = __gmap_unlink_by_vmaddr(gmap, vmaddr);
if (flush)
gmap_flush_tlb(gmap);
}
rcu_read_unlock();
}
/**
* gmap_link - set up shadow page tables to connect a host to a guest address
* @gmap: pointer to guest mapping meta data structure
* @gaddr: guest address
* @vmaddr: vm address
*
* Returns 0 on success, -ENOMEM for out of memory conditions, and -EFAULT
* if the vm address is already mapped to a different guest segment.
* The mmap_sem of the mm that belongs to the address space must be held
* when this function gets called.
*/
int __gmap_link(struct gmap *gmap, unsigned long gaddr, unsigned long vmaddr)
{
struct mm_struct *mm;
unsigned long *table;
spinlock_t *ptl;
pgd_t *pgd;
pud_t *pud;
pmd_t *pmd;
int rc;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
BUG_ON(gmap_is_shadow(gmap));
/* Create higher level tables in the gmap page table */
table = gmap->table;
if ((gmap->asce & _ASCE_TYPE_MASK) >= _ASCE_TYPE_REGION1) {
table += (gaddr >> 53) & 0x7ff;
if ((*table & _REGION_ENTRY_INVALID) &&
gmap_alloc_table(gmap, table, _REGION2_ENTRY_EMPTY,
gaddr & 0xffe0000000000000UL))
return -ENOMEM;
table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
}
if ((gmap->asce & _ASCE_TYPE_MASK) >= _ASCE_TYPE_REGION2) {
table += (gaddr >> 42) & 0x7ff;
if ((*table & _REGION_ENTRY_INVALID) &&
gmap_alloc_table(gmap, table, _REGION3_ENTRY_EMPTY,
gaddr & 0xfffffc0000000000UL))
return -ENOMEM;
table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
}
if ((gmap->asce & _ASCE_TYPE_MASK) >= _ASCE_TYPE_REGION3) {
table += (gaddr >> 31) & 0x7ff;
if ((*table & _REGION_ENTRY_INVALID) &&
gmap_alloc_table(gmap, table, _SEGMENT_ENTRY_EMPTY,
gaddr & 0xffffffff80000000UL))
return -ENOMEM;
table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
}
table += (gaddr >> 20) & 0x7ff;
/* Walk the parent mm page table */
mm = gmap->mm;
pgd = pgd_offset(mm, vmaddr);
VM_BUG_ON(pgd_none(*pgd));
pud = pud_offset(pgd, vmaddr);
VM_BUG_ON(pud_none(*pud));
/* large puds cannot yet be handled */
if (pud_large(*pud))
return -EFAULT;
pmd = pmd_offset(pud, vmaddr);
VM_BUG_ON(pmd_none(*pmd));
/* large pmds cannot yet be handled */
if (pmd_large(*pmd))
return -EFAULT;
/* Link gmap segment table entry location to page table. */
rc = radix_tree_preload(GFP_KERNEL);
if (rc)
return rc;
ptl = pmd_lock(mm, pmd);
spin_lock(&gmap->guest_table_lock);
if (*table == _SEGMENT_ENTRY_INVALID) {
rc = radix_tree_insert(&gmap->host_to_guest,
vmaddr >> PMD_SHIFT, table);
if (!rc)
*table = pmd_val(*pmd);
} else
rc = 0;
spin_unlock(&gmap->guest_table_lock);
spin_unlock(ptl);
radix_tree_preload_end();
return rc;
}
/**
* gmap_fault - resolve a fault on a guest address
* @gmap: pointer to guest mapping meta data structure
* @gaddr: guest address
* @fault_flags: flags to pass down to handle_mm_fault()
*
* Returns 0 on success, -ENOMEM for out of memory conditions, and -EFAULT
* if the vm address is already mapped to a different guest segment.
*/
int gmap_fault(struct gmap *gmap, unsigned long gaddr,
unsigned int fault_flags)
{
unsigned long vmaddr;
int rc;
bool unlocked;
down_read(&gmap->mm->mmap_sem);
retry:
unlocked = false;
vmaddr = __gmap_translate(gmap, gaddr);
if (IS_ERR_VALUE(vmaddr)) {
rc = vmaddr;
goto out_up;
}
if (fixup_user_fault(current, gmap->mm, vmaddr, fault_flags,
&unlocked)) {
rc = -EFAULT;
goto out_up;
}
/*
* In the case that fixup_user_fault unlocked the mmap_sem during
* faultin redo __gmap_translate to not race with a map/unmap_segment.
*/
if (unlocked)
goto retry;
rc = __gmap_link(gmap, gaddr, vmaddr);
out_up:
up_read(&gmap->mm->mmap_sem);
return rc;
}
EXPORT_SYMBOL_GPL(gmap_fault);
/*
* this function is assumed to be called with mmap_sem held
*/
void __gmap_zap(struct gmap *gmap, unsigned long gaddr)
{
unsigned long vmaddr;
spinlock_t *ptl;
pte_t *ptep;
/* Find the vm address for the guest address */
vmaddr = (unsigned long) radix_tree_lookup(&gmap->guest_to_host,
gaddr >> PMD_SHIFT);
if (vmaddr) {
vmaddr |= gaddr & ~PMD_MASK;
/* Get pointer to the page table entry */
ptep = get_locked_pte(gmap->mm, vmaddr, &ptl);
if (likely(ptep))
ptep_zap_unused(gmap->mm, vmaddr, ptep, 0);
pte_unmap_unlock(ptep, ptl);
}
}
EXPORT_SYMBOL_GPL(__gmap_zap);
void gmap_discard(struct gmap *gmap, unsigned long from, unsigned long to)
{
unsigned long gaddr, vmaddr, size;
struct vm_area_struct *vma;
down_read(&gmap->mm->mmap_sem);
for (gaddr = from; gaddr < to;
gaddr = (gaddr + PMD_SIZE) & PMD_MASK) {
/* Find the vm address for the guest address */
vmaddr = (unsigned long)
radix_tree_lookup(&gmap->guest_to_host,
gaddr >> PMD_SHIFT);
if (!vmaddr)
continue;
vmaddr |= gaddr & ~PMD_MASK;
/* Find vma in the parent mm */
vma = find_vma(gmap->mm, vmaddr);
size = min(to - gaddr, PMD_SIZE - (gaddr & ~PMD_MASK));
zap_page_range(vma, vmaddr, size, NULL);
}
up_read(&gmap->mm->mmap_sem);
}
EXPORT_SYMBOL_GPL(gmap_discard);
static LIST_HEAD(gmap_notifier_list);
static DEFINE_SPINLOCK(gmap_notifier_lock);
/**
* gmap_register_pte_notifier - register a pte invalidation callback
* @nb: pointer to the gmap notifier block
*/
void gmap_register_pte_notifier(struct gmap_notifier *nb)
{
spin_lock(&gmap_notifier_lock);
list_add_rcu(&nb->list, &gmap_notifier_list);
spin_unlock(&gmap_notifier_lock);
}
EXPORT_SYMBOL_GPL(gmap_register_pte_notifier);
/**
* gmap_unregister_pte_notifier - remove a pte invalidation callback
* @nb: pointer to the gmap notifier block
*/
void gmap_unregister_pte_notifier(struct gmap_notifier *nb)
{
spin_lock(&gmap_notifier_lock);
list_del_rcu(&nb->list);
spin_unlock(&gmap_notifier_lock);
synchronize_rcu();
}
EXPORT_SYMBOL_GPL(gmap_unregister_pte_notifier);
/**
* gmap_call_notifier - call all registered invalidation callbacks
* @gmap: pointer to guest mapping meta data structure
* @start: start virtual address in the guest address space
* @end: end virtual address in the guest address space
*/
static void gmap_call_notifier(struct gmap *gmap, unsigned long start,
unsigned long end)
{
struct gmap_notifier *nb;
list_for_each_entry(nb, &gmap_notifier_list, list)
nb->notifier_call(gmap, start, end);
}
/**
* gmap_table_walk - walk the gmap page tables
* @gmap: pointer to guest mapping meta data structure
* @gaddr: virtual address in the guest address space
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
* @level: page table level to stop at
*
* Returns a table entry pointer for the given guest address and @level
* @level=0 : returns a pointer to a page table table entry (or NULL)
* @level=1 : returns a pointer to a segment table entry (or NULL)
* @level=2 : returns a pointer to a region-3 table entry (or NULL)
* @level=3 : returns a pointer to a region-2 table entry (or NULL)
* @level=4 : returns a pointer to a region-1 table entry (or NULL)
*
* Returns NULL if the gmap page tables could not be walked to the
* requested level.
*
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
* Note: Can also be called for shadow gmaps.
*/
static inline unsigned long *gmap_table_walk(struct gmap *gmap,
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
unsigned long gaddr, int level)
{
unsigned long *table;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
if ((gmap->asce & _ASCE_TYPE_MASK) + 4 < (level * 4))
return NULL;
if (gmap_is_shadow(gmap) && gmap->removed)
return NULL;
if (gaddr & (-1UL << (31 + ((gmap->asce & _ASCE_TYPE_MASK) >> 2)*11)))
return NULL;
table = gmap->table;
switch (gmap->asce & _ASCE_TYPE_MASK) {
case _ASCE_TYPE_REGION1:
table += (gaddr >> 53) & 0x7ff;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
if (level == 4)
break;
if (*table & _REGION_ENTRY_INVALID)
return NULL;
table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
/* Fallthrough */
case _ASCE_TYPE_REGION2:
table += (gaddr >> 42) & 0x7ff;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
if (level == 3)
break;
if (*table & _REGION_ENTRY_INVALID)
return NULL;
table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
/* Fallthrough */
case _ASCE_TYPE_REGION3:
table += (gaddr >> 31) & 0x7ff;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
if (level == 2)
break;
if (*table & _REGION_ENTRY_INVALID)
return NULL;
table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
/* Fallthrough */
case _ASCE_TYPE_SEGMENT:
table += (gaddr >> 20) & 0x7ff;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
if (level == 1)
break;
if (*table & _REGION_ENTRY_INVALID)
return NULL;
table = (unsigned long *)(*table & _SEGMENT_ENTRY_ORIGIN);
table += (gaddr >> 12) & 0xff;
}
return table;
}
/**
* gmap_pte_op_walk - walk the gmap page table, get the page table lock
* and return the pte pointer
* @gmap: pointer to guest mapping meta data structure
* @gaddr: virtual address in the guest address space
* @ptl: pointer to the spinlock pointer
*
* Returns a pointer to the locked pte for a guest address, or NULL
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
*
* Note: Can also be called for shadow gmaps.
*/
static pte_t *gmap_pte_op_walk(struct gmap *gmap, unsigned long gaddr,
spinlock_t **ptl)
{
unsigned long *table;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
if (gmap_is_shadow(gmap))
spin_lock(&gmap->guest_table_lock);
/* Walk the gmap page table, lock and get pte pointer */
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
table = gmap_table_walk(gmap, gaddr, 1); /* get segment pointer */
if (!table || *table & _SEGMENT_ENTRY_INVALID) {
if (gmap_is_shadow(gmap))
spin_unlock(&gmap->guest_table_lock);
return NULL;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
}
if (gmap_is_shadow(gmap)) {
*ptl = &gmap->guest_table_lock;
return pte_offset_map((pmd_t *) table, gaddr);
}
return pte_alloc_map_lock(gmap->mm, (pmd_t *) table, gaddr, ptl);
}
/**
* gmap_pte_op_fixup - force a page in and connect the gmap page table
* @gmap: pointer to guest mapping meta data structure
* @gaddr: virtual address in the guest address space
* @vmaddr: address in the host process address space
* @prot: indicates access rights: PROT_NONE, PROT_READ or PROT_WRITE
*
* Returns 0 if the caller can retry __gmap_translate (might fail again),
* -ENOMEM if out of memory and -EFAULT if anything goes wrong while fixing
* up or connecting the gmap page table.
*/
static int gmap_pte_op_fixup(struct gmap *gmap, unsigned long gaddr,
unsigned long vmaddr, int prot)
{
struct mm_struct *mm = gmap->mm;
unsigned int fault_flags;
bool unlocked = false;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
BUG_ON(gmap_is_shadow(gmap));
fault_flags = (prot == PROT_WRITE) ? FAULT_FLAG_WRITE : 0;
if (fixup_user_fault(current, mm, vmaddr, fault_flags, &unlocked))
return -EFAULT;
if (unlocked)
/* lost mmap_sem, caller has to retry __gmap_translate */
return 0;
/* Connect the page tables */
return __gmap_link(gmap, gaddr, vmaddr);
}
/**
* gmap_pte_op_end - release the page table lock
* @ptl: pointer to the spinlock pointer
*/
static void gmap_pte_op_end(spinlock_t *ptl)
{
spin_unlock(ptl);
}
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
/*
* gmap_protect_range - remove access rights to memory and set pgste bits
* @gmap: pointer to guest mapping meta data structure
* @gaddr: virtual address in the guest address space
* @len: size of area
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
* @prot: indicates access rights: PROT_NONE, PROT_READ or PROT_WRITE
* @bits: pgste notification bits to set
*
* Returns 0 if successfully protected, -ENOMEM if out of memory and
* -EFAULT if gaddr is invalid (or mapping for shadows is missing).
*
* Called with sg->mm->mmap_sem in read.
*
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
* Note: Can also be called for shadow gmaps.
*/
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
static int gmap_protect_range(struct gmap *gmap, unsigned long gaddr,
unsigned long len, int prot, unsigned long bits)
{
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
unsigned long vmaddr;
spinlock_t *ptl;
pte_t *ptep;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
int rc;
while (len) {
rc = -EAGAIN;
ptep = gmap_pte_op_walk(gmap, gaddr, &ptl);
if (ptep) {
rc = ptep_force_prot(gmap->mm, gaddr, ptep, prot, bits);
gmap_pte_op_end(ptl);
}
if (rc) {
vmaddr = __gmap_translate(gmap, gaddr);
if (IS_ERR_VALUE(vmaddr))
return vmaddr;
rc = gmap_pte_op_fixup(gmap, gaddr, vmaddr, prot);
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
if (rc)
return rc;
continue;
}
gaddr += PAGE_SIZE;
len -= PAGE_SIZE;
}
return 0;
}
/**
* gmap_mprotect_notify - change access rights for a range of ptes and
* call the notifier if any pte changes again
* @gmap: pointer to guest mapping meta data structure
* @gaddr: virtual address in the guest address space
* @len: size of area
* @prot: indicates access rights: PROT_NONE, PROT_READ or PROT_WRITE
*
* Returns 0 if for each page in the given range a gmap mapping exists,
* the new access rights could be set and the notifier could be armed.
* If the gmap mapping is missing for one or more pages -EFAULT is
* returned. If no memory could be allocated -ENOMEM is returned.
* This function establishes missing page table entries.
*/
int gmap_mprotect_notify(struct gmap *gmap, unsigned long gaddr,
unsigned long len, int prot)
{
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
int rc;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
if ((gaddr & ~PAGE_MASK) || (len & ~PAGE_MASK) || gmap_is_shadow(gmap))
return -EINVAL;
if (!MACHINE_HAS_ESOP && prot == PROT_READ)
return -EINVAL;
down_read(&gmap->mm->mmap_sem);
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
rc = gmap_protect_range(gmap, gaddr, len, prot, PGSTE_IN_BIT);
up_read(&gmap->mm->mmap_sem);
return rc;
}
EXPORT_SYMBOL_GPL(gmap_mprotect_notify);
/**
* gmap_read_table - get an unsigned long value from a guest page table using
* absolute addressing, without marking the page referenced.
* @gmap: pointer to guest mapping meta data structure
* @gaddr: virtual address in the guest address space
* @val: pointer to the unsigned long value to return
*
* Returns 0 if the value was read, -ENOMEM if out of memory and -EFAULT
* if reading using the virtual address failed.
*
* Called with gmap->mm->mmap_sem in read.
*/
int gmap_read_table(struct gmap *gmap, unsigned long gaddr, unsigned long *val)
{
unsigned long address, vmaddr;
spinlock_t *ptl;
pte_t *ptep, pte;
int rc;
while (1) {
rc = -EAGAIN;
ptep = gmap_pte_op_walk(gmap, gaddr, &ptl);
if (ptep) {
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
pte = *ptep;
if (pte_present(pte) && (pte_val(pte) & _PAGE_READ)) {
address = pte_val(pte) & PAGE_MASK;
address += gaddr & ~PAGE_MASK;
*val = *(unsigned long *) address;
pte_val(*ptep) |= _PAGE_YOUNG;
/* Do *NOT* clear the _PAGE_INVALID bit! */
rc = 0;
}
gmap_pte_op_end(ptl);
}
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
if (!rc)
break;
vmaddr = __gmap_translate(gmap, gaddr);
if (IS_ERR_VALUE(vmaddr)) {
rc = vmaddr;
break;
}
rc = gmap_pte_op_fixup(gmap, gaddr, vmaddr, PROT_READ);
if (rc)
break;
}
return rc;
}
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
EXPORT_SYMBOL_GPL(gmap_read_table);
/**
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
* gmap_insert_rmap - add a rmap to the host_to_rmap radix tree
* @sg: pointer to the shadow guest address space structure
* @vmaddr: vm address associated with the rmap
* @rmap: pointer to the rmap structure
*
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
* Called with the sg->guest_table_lock
*/
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
static inline void gmap_insert_rmap(struct gmap *sg, unsigned long vmaddr,
struct gmap_rmap *rmap)
{
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
void **slot;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
BUG_ON(!gmap_is_shadow(sg));
slot = radix_tree_lookup_slot(&sg->host_to_rmap, vmaddr >> PAGE_SHIFT);
if (slot) {
rmap->next = radix_tree_deref_slot_protected(slot,
&sg->guest_table_lock);
radix_tree_replace_slot(slot, rmap);
} else {
rmap->next = NULL;
radix_tree_insert(&sg->host_to_rmap, vmaddr >> PAGE_SHIFT,
rmap);
}
}
/**
* gmap_protect_rmap - modify access rights to memory and create an rmap
* @sg: pointer to the shadow guest address space structure
* @raddr: rmap address in the shadow gmap
* @paddr: address in the parent guest address space
* @len: length of the memory area to protect
* @prot: indicates access rights: none, read-only or read-write
*
* Returns 0 if successfully protected and the rmap was created, -ENOMEM
* if out of memory and -EFAULT if paddr is invalid.
*/
static int gmap_protect_rmap(struct gmap *sg, unsigned long raddr,
unsigned long paddr, unsigned long len, int prot)
{
struct gmap *parent;
struct gmap_rmap *rmap;
unsigned long vmaddr;
spinlock_t *ptl;
pte_t *ptep;
int rc;
BUG_ON(!gmap_is_shadow(sg));
parent = sg->parent;
while (len) {
vmaddr = __gmap_translate(parent, paddr);
if (IS_ERR_VALUE(vmaddr))
return vmaddr;
rmap = kzalloc(sizeof(*rmap), GFP_KERNEL);
if (!rmap)
return -ENOMEM;
rmap->raddr = raddr;
rc = radix_tree_preload(GFP_KERNEL);
if (rc) {
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
kfree(rmap);
return rc;
}
rc = -EAGAIN;
ptep = gmap_pte_op_walk(parent, paddr, &ptl);
if (ptep) {
spin_lock(&sg->guest_table_lock);
rc = ptep_force_prot(parent->mm, paddr, ptep, prot,
PGSTE_VSIE_BIT);
if (!rc)
gmap_insert_rmap(sg, vmaddr, rmap);
spin_unlock(&sg->guest_table_lock);
gmap_pte_op_end(ptl);
}
radix_tree_preload_end();
if (rc) {
kfree(rmap);
rc = gmap_pte_op_fixup(parent, paddr, vmaddr, prot);
if (rc)
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
return rc;
continue;
}
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
paddr += PAGE_SIZE;
len -= PAGE_SIZE;
}
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
return 0;
}
#define _SHADOW_RMAP_MASK 0x7
#define _SHADOW_RMAP_REGION1 0x5
#define _SHADOW_RMAP_REGION2 0x4
#define _SHADOW_RMAP_REGION3 0x3
#define _SHADOW_RMAP_SEGMENT 0x2
#define _SHADOW_RMAP_PGTABLE 0x1
/**
* gmap_idte_one - invalidate a single region or segment table entry
* @asce: region or segment table *origin* + table-type bits
* @vaddr: virtual address to identify the table entry to flush
*
* The invalid bit of a single region or segment table entry is set
* and the associated TLB entries depending on the entry are flushed.
* The table-type of the @asce identifies the portion of the @vaddr
* that is used as the invalidation index.
*/
static inline void gmap_idte_one(unsigned long asce, unsigned long vaddr)
{
asm volatile(
" .insn rrf,0xb98e0000,%0,%1,0,0"
: : "a" (asce), "a" (vaddr) : "cc", "memory");
}
/**
* gmap_unshadow_page - remove a page from a shadow page table
* @sg: pointer to the shadow guest address space structure
* @raddr: rmap address in the shadow guest address space
*
* Called with the sg->guest_table_lock
*/
static void gmap_unshadow_page(struct gmap *sg, unsigned long raddr)
{
unsigned long *table;
BUG_ON(!gmap_is_shadow(sg));
table = gmap_table_walk(sg, raddr, 0); /* get page table pointer */
if (!table || *table & _PAGE_INVALID)
return;
gmap_call_notifier(sg, raddr, raddr + (1UL << 12) - 1);
ptep_unshadow_pte(sg->mm, raddr, (pte_t *) table);
}
/**
* __gmap_unshadow_pgt - remove all entries from a shadow page table
* @sg: pointer to the shadow guest address space structure
* @raddr: rmap address in the shadow guest address space
* @pgt: pointer to the start of a shadow page table
*
* Called with the sg->guest_table_lock
*/
static void __gmap_unshadow_pgt(struct gmap *sg, unsigned long raddr,
unsigned long *pgt)
{
int i;
BUG_ON(!gmap_is_shadow(sg));
for (i = 0; i < 256; i++, raddr += 1UL << 12)
pgt[i] = _PAGE_INVALID;
}
/**
* gmap_unshadow_pgt - remove a shadow page table from a segment entry
* @sg: pointer to the shadow guest address space structure
* @raddr: address in the shadow guest address space
*
* Called with the sg->guest_table_lock
*/
static void gmap_unshadow_pgt(struct gmap *sg, unsigned long raddr)
{
unsigned long sto, *ste, *pgt;
struct page *page;
BUG_ON(!gmap_is_shadow(sg));
ste = gmap_table_walk(sg, raddr, 1); /* get segment pointer */
if (!ste || !(*ste & _SEGMENT_ENTRY_ORIGIN))
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
return;
gmap_call_notifier(sg, raddr, raddr + (1UL << 20) - 1);
sto = (unsigned long) (ste - ((raddr >> 20) & 0x7ff));
gmap_idte_one(sto | _ASCE_TYPE_SEGMENT, raddr);
pgt = (unsigned long *)(*ste & _SEGMENT_ENTRY_ORIGIN);
*ste = _SEGMENT_ENTRY_EMPTY;
__gmap_unshadow_pgt(sg, raddr, pgt);
/* Free page table */
page = pfn_to_page(__pa(pgt) >> PAGE_SHIFT);
list_del(&page->lru);
page_table_free_pgste(page);
}
/**
* __gmap_unshadow_sgt - remove all entries from a shadow segment table
* @sg: pointer to the shadow guest address space structure
* @raddr: rmap address in the shadow guest address space
* @sgt: pointer to the start of a shadow segment table
*
* Called with the sg->guest_table_lock
*/
static void __gmap_unshadow_sgt(struct gmap *sg, unsigned long raddr,
unsigned long *sgt)
{
unsigned long asce, *pgt;
struct page *page;
int i;
BUG_ON(!gmap_is_shadow(sg));
asce = (unsigned long) sgt | _ASCE_TYPE_SEGMENT;
for (i = 0; i < 2048; i++, raddr += 1UL << 20) {
if (!(sgt[i] & _SEGMENT_ENTRY_ORIGIN))
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
continue;
pgt = (unsigned long *)(sgt[i] & _REGION_ENTRY_ORIGIN);
sgt[i] = _SEGMENT_ENTRY_EMPTY;
__gmap_unshadow_pgt(sg, raddr, pgt);
/* Free page table */
page = pfn_to_page(__pa(pgt) >> PAGE_SHIFT);
list_del(&page->lru);
page_table_free_pgste(page);
}
}
/**
* gmap_unshadow_sgt - remove a shadow segment table from a region-3 entry
* @sg: pointer to the shadow guest address space structure
* @raddr: rmap address in the shadow guest address space
*
* Called with the shadow->guest_table_lock
*/
static void gmap_unshadow_sgt(struct gmap *sg, unsigned long raddr)
{
unsigned long r3o, *r3e, *sgt;
struct page *page;
BUG_ON(!gmap_is_shadow(sg));
r3e = gmap_table_walk(sg, raddr, 2); /* get region-3 pointer */
if (!r3e || !(*r3e & _REGION_ENTRY_ORIGIN))
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
return;
gmap_call_notifier(sg, raddr, raddr + (1UL << 31) - 1);
r3o = (unsigned long) (r3e - ((raddr >> 31) & 0x7ff));
gmap_idte_one(r3o | _ASCE_TYPE_REGION3, raddr);
sgt = (unsigned long *)(*r3e & _REGION_ENTRY_ORIGIN);
*r3e = _REGION3_ENTRY_EMPTY;
__gmap_unshadow_sgt(sg, raddr, sgt);
/* Free segment table */
page = pfn_to_page(__pa(sgt) >> PAGE_SHIFT);
list_del(&page->lru);
__free_pages(page, 2);
}
/**
* __gmap_unshadow_r3t - remove all entries from a shadow region-3 table
* @sg: pointer to the shadow guest address space structure
* @raddr: address in the shadow guest address space
* @r3t: pointer to the start of a shadow region-3 table
*
* Called with the sg->guest_table_lock
*/
static void __gmap_unshadow_r3t(struct gmap *sg, unsigned long raddr,
unsigned long *r3t)
{
unsigned long asce, *sgt;
struct page *page;
int i;
BUG_ON(!gmap_is_shadow(sg));
asce = (unsigned long) r3t | _ASCE_TYPE_REGION3;
for (i = 0; i < 2048; i++, raddr += 1UL << 31) {
if (!(r3t[i] & _REGION_ENTRY_ORIGIN))
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
continue;
sgt = (unsigned long *)(r3t[i] & _REGION_ENTRY_ORIGIN);
r3t[i] = _REGION3_ENTRY_EMPTY;
__gmap_unshadow_sgt(sg, raddr, sgt);
/* Free segment table */
page = pfn_to_page(__pa(sgt) >> PAGE_SHIFT);
list_del(&page->lru);
__free_pages(page, 2);
}
}
/**
* gmap_unshadow_r3t - remove a shadow region-3 table from a region-2 entry
* @sg: pointer to the shadow guest address space structure
* @raddr: rmap address in the shadow guest address space
*
* Called with the sg->guest_table_lock
*/
static void gmap_unshadow_r3t(struct gmap *sg, unsigned long raddr)
{
unsigned long r2o, *r2e, *r3t;
struct page *page;
BUG_ON(!gmap_is_shadow(sg));
r2e = gmap_table_walk(sg, raddr, 3); /* get region-2 pointer */
if (!r2e || !(*r2e & _REGION_ENTRY_ORIGIN))
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
return;
gmap_call_notifier(sg, raddr, raddr + (1UL << 42) - 1);
r2o = (unsigned long) (r2e - ((raddr >> 42) & 0x7ff));
gmap_idte_one(r2o | _ASCE_TYPE_REGION2, raddr);
r3t = (unsigned long *)(*r2e & _REGION_ENTRY_ORIGIN);
*r2e = _REGION2_ENTRY_EMPTY;
__gmap_unshadow_r3t(sg, raddr, r3t);
/* Free region 3 table */
page = pfn_to_page(__pa(r3t) >> PAGE_SHIFT);
list_del(&page->lru);
__free_pages(page, 2);
}
/**
* __gmap_unshadow_r2t - remove all entries from a shadow region-2 table
* @sg: pointer to the shadow guest address space structure
* @raddr: rmap address in the shadow guest address space
* @r2t: pointer to the start of a shadow region-2 table
*
* Called with the sg->guest_table_lock
*/
static void __gmap_unshadow_r2t(struct gmap *sg, unsigned long raddr,
unsigned long *r2t)
{
unsigned long asce, *r3t;
struct page *page;
int i;
BUG_ON(!gmap_is_shadow(sg));
asce = (unsigned long) r2t | _ASCE_TYPE_REGION2;
for (i = 0; i < 2048; i++, raddr += 1UL << 42) {
if (!(r2t[i] & _REGION_ENTRY_ORIGIN))
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
continue;
r3t = (unsigned long *)(r2t[i] & _REGION_ENTRY_ORIGIN);
r2t[i] = _REGION2_ENTRY_EMPTY;
__gmap_unshadow_r3t(sg, raddr, r3t);
/* Free region 3 table */
page = pfn_to_page(__pa(r3t) >> PAGE_SHIFT);
list_del(&page->lru);
__free_pages(page, 2);
}
}
/**
* gmap_unshadow_r2t - remove a shadow region-2 table from a region-1 entry
* @sg: pointer to the shadow guest address space structure
* @raddr: rmap address in the shadow guest address space
*
* Called with the sg->guest_table_lock
*/
static void gmap_unshadow_r2t(struct gmap *sg, unsigned long raddr)
{
unsigned long r1o, *r1e, *r2t;
struct page *page;
BUG_ON(!gmap_is_shadow(sg));
r1e = gmap_table_walk(sg, raddr, 4); /* get region-1 pointer */
if (!r1e || !(*r1e & _REGION_ENTRY_ORIGIN))
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
return;
gmap_call_notifier(sg, raddr, raddr + (1UL << 53) - 1);
r1o = (unsigned long) (r1e - ((raddr >> 53) & 0x7ff));
gmap_idte_one(r1o | _ASCE_TYPE_REGION1, raddr);
r2t = (unsigned long *)(*r1e & _REGION_ENTRY_ORIGIN);
*r1e = _REGION1_ENTRY_EMPTY;
__gmap_unshadow_r2t(sg, raddr, r2t);
/* Free region 2 table */
page = pfn_to_page(__pa(r2t) >> PAGE_SHIFT);
list_del(&page->lru);
__free_pages(page, 2);
}
/**
* __gmap_unshadow_r1t - remove all entries from a shadow region-1 table
* @sg: pointer to the shadow guest address space structure
* @raddr: rmap address in the shadow guest address space
* @r1t: pointer to the start of a shadow region-1 table
*
* Called with the shadow->guest_table_lock
*/
static void __gmap_unshadow_r1t(struct gmap *sg, unsigned long raddr,
unsigned long *r1t)
{
unsigned long asce, *r2t;
struct page *page;
int i;
BUG_ON(!gmap_is_shadow(sg));
asce = (unsigned long) r1t | _ASCE_TYPE_REGION1;
for (i = 0; i < 2048; i++, raddr += 1UL << 53) {
if (!(r1t[i] & _REGION_ENTRY_ORIGIN))
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
continue;
r2t = (unsigned long *)(r1t[i] & _REGION_ENTRY_ORIGIN);
__gmap_unshadow_r2t(sg, raddr, r2t);
/* Clear entry and flush translation r1t -> r2t */
gmap_idte_one(asce, raddr);
r1t[i] = _REGION1_ENTRY_EMPTY;
/* Free region 2 table */
page = pfn_to_page(__pa(r2t) >> PAGE_SHIFT);
list_del(&page->lru);
__free_pages(page, 2);
}
}
/**
* gmap_unshadow - remove a shadow page table completely
* @sg: pointer to the shadow guest address space structure
*
* Called with sg->guest_table_lock
*/
static void gmap_unshadow(struct gmap *sg)
{
unsigned long *table;
BUG_ON(!gmap_is_shadow(sg));
if (sg->removed)
return;
sg->removed = 1;
gmap_call_notifier(sg, 0, -1UL);
gmap_flush_tlb(sg);
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
table = (unsigned long *)(sg->asce & _ASCE_ORIGIN);
switch (sg->asce & _ASCE_TYPE_MASK) {
case _ASCE_TYPE_REGION1:
__gmap_unshadow_r1t(sg, 0, table);
break;
case _ASCE_TYPE_REGION2:
__gmap_unshadow_r2t(sg, 0, table);
break;
case _ASCE_TYPE_REGION3:
__gmap_unshadow_r3t(sg, 0, table);
break;
case _ASCE_TYPE_SEGMENT:
__gmap_unshadow_sgt(sg, 0, table);
break;
}
}
/**
* gmap_find_shadow - find a specific asce in the list of shadow tables
* @parent: pointer to the parent gmap
* @asce: ASCE for which the shadow table is created
* @edat_level: edat level to be used for the shadow translation
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
*
* Returns the pointer to a gmap if a shadow table with the given asce is
* already available, ERR_PTR(-EAGAIN) if another one is just being created,
* otherwise NULL
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
*/
static struct gmap *gmap_find_shadow(struct gmap *parent, unsigned long asce,
int edat_level)
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
{
struct gmap *sg;
list_for_each_entry(sg, &parent->children, list) {
if (sg->orig_asce != asce || sg->edat_level != edat_level ||
sg->removed)
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
continue;
if (!sg->initialized)
return ERR_PTR(-EAGAIN);
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
atomic_inc(&sg->ref_count);
return sg;
}
return NULL;
}
/**
* gmap_shadow_valid - check if a shadow guest address space matches the
* given properties and is still valid
* @sg: pointer to the shadow guest address space structure
* @asce: ASCE for which the shadow table is requested
* @edat_level: edat level to be used for the shadow translation
*
* Returns 1 if the gmap shadow is still valid and matches the given
* properties, the caller can continue using it. Returns 0 otherwise, the
* caller has to request a new shadow gmap in this case.
*
*/
int gmap_shadow_valid(struct gmap *sg, unsigned long asce, int edat_level)
{
if (sg->removed)
return 0;
return sg->orig_asce == asce && sg->edat_level == edat_level;
}
EXPORT_SYMBOL_GPL(gmap_shadow_valid);
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
/**
* gmap_shadow - create/find a shadow guest address space
* @parent: pointer to the parent gmap
* @asce: ASCE for which the shadow table is created
* @edat_level: edat level to be used for the shadow translation
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
*
* The pages of the top level page table referred by the asce parameter
* will be set to read-only and marked in the PGSTEs of the kvm process.
* The shadow table will be removed automatically on any change to the
* PTE mapping for the source table.
*
* Returns a guest address space structure, ERR_PTR(-ENOMEM) if out of memory,
* ERR_PTR(-EAGAIN) if the caller has to retry and ERR_PTR(-EFAULT) if the
* parent gmap table could not be protected.
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
*/
struct gmap *gmap_shadow(struct gmap *parent, unsigned long asce,
int edat_level)
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
{
struct gmap *sg, *new;
unsigned long limit;
int rc;
BUG_ON(gmap_is_shadow(parent));
spin_lock(&parent->shadow_lock);
sg = gmap_find_shadow(parent, asce, edat_level);
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
spin_unlock(&parent->shadow_lock);
if (sg)
return sg;
/* Create a new shadow gmap */
limit = -1UL >> (33 - (((asce & _ASCE_TYPE_MASK) >> 2) * 11));
if (asce & _ASCE_REAL_SPACE)
limit = -1UL;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
new = gmap_alloc(limit);
if (!new)
return ERR_PTR(-ENOMEM);
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
new->mm = parent->mm;
new->parent = gmap_get(parent);
new->orig_asce = asce;
new->edat_level = edat_level;
new->initialized = false;
spin_lock(&parent->shadow_lock);
/* Recheck if another CPU created the same shadow */
sg = gmap_find_shadow(parent, asce, edat_level);
if (sg) {
spin_unlock(&parent->shadow_lock);
gmap_free(new);
return sg;
}
if (asce & _ASCE_REAL_SPACE) {
/* only allow one real-space gmap shadow */
list_for_each_entry(sg, &parent->children, list) {
if (sg->orig_asce & _ASCE_REAL_SPACE) {
spin_lock(&sg->guest_table_lock);
gmap_unshadow(sg);
spin_unlock(&sg->guest_table_lock);
list_del(&sg->list);
gmap_put(sg);
break;
}
}
}
atomic_set(&new->ref_count, 2);
list_add(&new->list, &parent->children);
if (asce & _ASCE_REAL_SPACE) {
/* nothing to protect, return right away */
new->initialized = true;
spin_unlock(&parent->shadow_lock);
return new;
}
spin_unlock(&parent->shadow_lock);
/* protect after insertion, so it will get properly invalidated */
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
down_read(&parent->mm->mmap_sem);
rc = gmap_protect_range(parent, asce & _ASCE_ORIGIN,
((asce & _ASCE_TABLE_LENGTH) + 1) * 4096,
PROT_READ, PGSTE_VSIE_BIT);
up_read(&parent->mm->mmap_sem);
spin_lock(&parent->shadow_lock);
new->initialized = true;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
if (rc) {
list_del(&new->list);
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
gmap_free(new);
new = ERR_PTR(rc);
}
spin_unlock(&parent->shadow_lock);
return new;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
}
EXPORT_SYMBOL_GPL(gmap_shadow);
/**
* gmap_shadow_r2t - create an empty shadow region 2 table
* @sg: pointer to the shadow guest address space structure
* @saddr: faulting address in the shadow gmap
* @r2t: parent gmap address of the region 2 table to get shadowed
* @fake: r2t references contiguous guest memory block, not a r2t
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
*
* The r2t parameter specifies the address of the source table. The
* four pages of the source table are made read-only in the parent gmap
* address space. A write to the source table area @r2t will automatically
* remove the shadow r2 table and all of its decendents.
*
* Returns 0 if successfully shadowed or already shadowed, -EAGAIN if the
* shadow table structure is incomplete, -ENOMEM if out of memory and
* -EFAULT if an address in the parent gmap could not be resolved.
*
* Called with sg->mm->mmap_sem in read.
*/
int gmap_shadow_r2t(struct gmap *sg, unsigned long saddr, unsigned long r2t,
int fake)
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
{
unsigned long raddr, origin, offset, len;
unsigned long *s_r2t, *table;
struct page *page;
int rc;
BUG_ON(!gmap_is_shadow(sg));
/* Allocate a shadow region second table */
page = alloc_pages(GFP_KERNEL, 2);
if (!page)
return -ENOMEM;
page->index = r2t & _REGION_ENTRY_ORIGIN;
if (fake)
page->index |= GMAP_SHADOW_FAKE_TABLE;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
s_r2t = (unsigned long *) page_to_phys(page);
/* Install shadow region second table */
spin_lock(&sg->guest_table_lock);
table = gmap_table_walk(sg, saddr, 4); /* get region-1 pointer */
if (!table) {
rc = -EAGAIN; /* Race with unshadow */
goto out_free;
}
if (!(*table & _REGION_ENTRY_INVALID)) {
rc = 0; /* Already established */
goto out_free;
} else if (*table & _REGION_ENTRY_ORIGIN) {
rc = -EAGAIN; /* Race with shadow */
goto out_free;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
}
crst_table_init(s_r2t, _REGION2_ENTRY_EMPTY);
/* mark as invalid as long as the parent table is not protected */
*table = (unsigned long) s_r2t | _REGION_ENTRY_LENGTH |
_REGION_ENTRY_TYPE_R1 | _REGION_ENTRY_INVALID;
s390/mm: support EDAT1 for gmap shadows If the guest is enabled for EDAT1, we can easily create shadows for guest2 -> guest3 provided tables that make use of EDAT1. If guest2 references a 1MB page, this memory looks consecutive for guest2, but it might not be so for us. Therefore we have to create fake page tables. We can easily add that to our existing infrastructure. The invalidation mechanism will make sure that fake page tables are removed when the parent table (sgt table entry) is changed. As EDAT1 also introduced protection on all page table levels, we have to also shadow these correctly. We don't have to care about: - ACCF-Validity Control in STE - Access-Control Bits in STE - Fetch-Protection Bit in STE - Common-Segment Bit in STE As all bits might be dropped and there is no guaranteed that they are active ("unpredictable whether the CPU uses these bits", "may be used"). Without using EDAT1 in the shadow ourselfes (STE-format control == 0), simply shadowing these bits would not be enough. They would be ignored. Please note that we are using the "fake" flag to make this look consistent with further changes (EDAT2, real-space designation support) and don't let the shadow functions handle fc=1 stes. In the future, with huge pages in the host, gmap_shadow_pgt() could simply try to map a huge host page if "fake" is set to one and indicate via return value that no lower fake tables / shadow ptes are required. Acked-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-04-18 19:24:52 +08:00
if (sg->edat_level >= 1)
*table |= (r2t & _REGION_ENTRY_PROTECT);
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
list_add(&page->lru, &sg->crst_list);
if (fake) {
/* nothing to protect for fake tables */
*table &= ~_REGION_ENTRY_INVALID;
spin_unlock(&sg->guest_table_lock);
return 0;
}
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
spin_unlock(&sg->guest_table_lock);
/* Make r2t read-only in parent gmap page table */
raddr = (saddr & 0xffe0000000000000UL) | _SHADOW_RMAP_REGION1;
origin = r2t & _REGION_ENTRY_ORIGIN;
offset = ((r2t & _REGION_ENTRY_OFFSET) >> 6) * 4096;
len = ((r2t & _REGION_ENTRY_LENGTH) + 1) * 4096 - offset;
rc = gmap_protect_rmap(sg, raddr, origin + offset, len, PROT_READ);
spin_lock(&sg->guest_table_lock);
if (!rc) {
table = gmap_table_walk(sg, saddr, 4);
if (!table || (*table & _REGION_ENTRY_ORIGIN) !=
(unsigned long) s_r2t)
rc = -EAGAIN; /* Race with unshadow */
else
*table &= ~_REGION_ENTRY_INVALID;
} else {
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
gmap_unshadow_r2t(sg, raddr);
}
spin_unlock(&sg->guest_table_lock);
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
return rc;
out_free:
spin_unlock(&sg->guest_table_lock);
__free_pages(page, 2);
return rc;
}
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
EXPORT_SYMBOL_GPL(gmap_shadow_r2t);
/**
* gmap_shadow_r3t - create a shadow region 3 table
* @sg: pointer to the shadow guest address space structure
* @saddr: faulting address in the shadow gmap
* @r3t: parent gmap address of the region 3 table to get shadowed
* @fake: r3t references contiguous guest memory block, not a r3t
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
*
* Returns 0 if successfully shadowed or already shadowed, -EAGAIN if the
* shadow table structure is incomplete, -ENOMEM if out of memory and
* -EFAULT if an address in the parent gmap could not be resolved.
*
* Called with sg->mm->mmap_sem in read.
*/
int gmap_shadow_r3t(struct gmap *sg, unsigned long saddr, unsigned long r3t,
int fake)
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
{
unsigned long raddr, origin, offset, len;
unsigned long *s_r3t, *table;
struct page *page;
int rc;
BUG_ON(!gmap_is_shadow(sg));
/* Allocate a shadow region second table */
page = alloc_pages(GFP_KERNEL, 2);
if (!page)
return -ENOMEM;
page->index = r3t & _REGION_ENTRY_ORIGIN;
if (fake)
page->index |= GMAP_SHADOW_FAKE_TABLE;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
s_r3t = (unsigned long *) page_to_phys(page);
/* Install shadow region second table */
spin_lock(&sg->guest_table_lock);
table = gmap_table_walk(sg, saddr, 3); /* get region-2 pointer */
if (!table) {
rc = -EAGAIN; /* Race with unshadow */
goto out_free;
}
if (!(*table & _REGION_ENTRY_INVALID)) {
rc = 0; /* Already established */
goto out_free;
} else if (*table & _REGION_ENTRY_ORIGIN) {
rc = -EAGAIN; /* Race with shadow */
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
}
crst_table_init(s_r3t, _REGION3_ENTRY_EMPTY);
/* mark as invalid as long as the parent table is not protected */
*table = (unsigned long) s_r3t | _REGION_ENTRY_LENGTH |
_REGION_ENTRY_TYPE_R2 | _REGION_ENTRY_INVALID;
s390/mm: support EDAT1 for gmap shadows If the guest is enabled for EDAT1, we can easily create shadows for guest2 -> guest3 provided tables that make use of EDAT1. If guest2 references a 1MB page, this memory looks consecutive for guest2, but it might not be so for us. Therefore we have to create fake page tables. We can easily add that to our existing infrastructure. The invalidation mechanism will make sure that fake page tables are removed when the parent table (sgt table entry) is changed. As EDAT1 also introduced protection on all page table levels, we have to also shadow these correctly. We don't have to care about: - ACCF-Validity Control in STE - Access-Control Bits in STE - Fetch-Protection Bit in STE - Common-Segment Bit in STE As all bits might be dropped and there is no guaranteed that they are active ("unpredictable whether the CPU uses these bits", "may be used"). Without using EDAT1 in the shadow ourselfes (STE-format control == 0), simply shadowing these bits would not be enough. They would be ignored. Please note that we are using the "fake" flag to make this look consistent with further changes (EDAT2, real-space designation support) and don't let the shadow functions handle fc=1 stes. In the future, with huge pages in the host, gmap_shadow_pgt() could simply try to map a huge host page if "fake" is set to one and indicate via return value that no lower fake tables / shadow ptes are required. Acked-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-04-18 19:24:52 +08:00
if (sg->edat_level >= 1)
*table |= (r3t & _REGION_ENTRY_PROTECT);
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
list_add(&page->lru, &sg->crst_list);
if (fake) {
/* nothing to protect for fake tables */
*table &= ~_REGION_ENTRY_INVALID;
spin_unlock(&sg->guest_table_lock);
return 0;
}
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
spin_unlock(&sg->guest_table_lock);
/* Make r3t read-only in parent gmap page table */
raddr = (saddr & 0xfffffc0000000000UL) | _SHADOW_RMAP_REGION2;
origin = r3t & _REGION_ENTRY_ORIGIN;
offset = ((r3t & _REGION_ENTRY_OFFSET) >> 6) * 4096;
len = ((r3t & _REGION_ENTRY_LENGTH) + 1) * 4096 - offset;
rc = gmap_protect_rmap(sg, raddr, origin + offset, len, PROT_READ);
spin_lock(&sg->guest_table_lock);
if (!rc) {
table = gmap_table_walk(sg, saddr, 3);
if (!table || (*table & _REGION_ENTRY_ORIGIN) !=
(unsigned long) s_r3t)
rc = -EAGAIN; /* Race with unshadow */
else
*table &= ~_REGION_ENTRY_INVALID;
} else {
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
gmap_unshadow_r3t(sg, raddr);
}
spin_unlock(&sg->guest_table_lock);
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
return rc;
out_free:
spin_unlock(&sg->guest_table_lock);
__free_pages(page, 2);
return rc;
}
EXPORT_SYMBOL_GPL(gmap_shadow_r3t);
/**
* gmap_shadow_sgt - create a shadow segment table
* @sg: pointer to the shadow guest address space structure
* @saddr: faulting address in the shadow gmap
* @sgt: parent gmap address of the segment table to get shadowed
* @fake: sgt references contiguous guest memory block, not a sgt
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
*
* Returns: 0 if successfully shadowed or already shadowed, -EAGAIN if the
* shadow table structure is incomplete, -ENOMEM if out of memory and
* -EFAULT if an address in the parent gmap could not be resolved.
*
* Called with sg->mm->mmap_sem in read.
*/
int gmap_shadow_sgt(struct gmap *sg, unsigned long saddr, unsigned long sgt,
int fake)
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
{
unsigned long raddr, origin, offset, len;
unsigned long *s_sgt, *table;
struct page *page;
int rc;
BUG_ON(!gmap_is_shadow(sg) || (sgt & _REGION3_ENTRY_LARGE));
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
/* Allocate a shadow segment table */
page = alloc_pages(GFP_KERNEL, 2);
if (!page)
return -ENOMEM;
page->index = sgt & _REGION_ENTRY_ORIGIN;
if (fake)
page->index |= GMAP_SHADOW_FAKE_TABLE;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
s_sgt = (unsigned long *) page_to_phys(page);
/* Install shadow region second table */
spin_lock(&sg->guest_table_lock);
table = gmap_table_walk(sg, saddr, 2); /* get region-3 pointer */
if (!table) {
rc = -EAGAIN; /* Race with unshadow */
goto out_free;
}
if (!(*table & _REGION_ENTRY_INVALID)) {
rc = 0; /* Already established */
goto out_free;
} else if (*table & _REGION_ENTRY_ORIGIN) {
rc = -EAGAIN; /* Race with shadow */
goto out_free;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
}
crst_table_init(s_sgt, _SEGMENT_ENTRY_EMPTY);
/* mark as invalid as long as the parent table is not protected */
*table = (unsigned long) s_sgt | _REGION_ENTRY_LENGTH |
_REGION_ENTRY_TYPE_R3 | _REGION_ENTRY_INVALID;
s390/mm: support EDAT1 for gmap shadows If the guest is enabled for EDAT1, we can easily create shadows for guest2 -> guest3 provided tables that make use of EDAT1. If guest2 references a 1MB page, this memory looks consecutive for guest2, but it might not be so for us. Therefore we have to create fake page tables. We can easily add that to our existing infrastructure. The invalidation mechanism will make sure that fake page tables are removed when the parent table (sgt table entry) is changed. As EDAT1 also introduced protection on all page table levels, we have to also shadow these correctly. We don't have to care about: - ACCF-Validity Control in STE - Access-Control Bits in STE - Fetch-Protection Bit in STE - Common-Segment Bit in STE As all bits might be dropped and there is no guaranteed that they are active ("unpredictable whether the CPU uses these bits", "may be used"). Without using EDAT1 in the shadow ourselfes (STE-format control == 0), simply shadowing these bits would not be enough. They would be ignored. Please note that we are using the "fake" flag to make this look consistent with further changes (EDAT2, real-space designation support) and don't let the shadow functions handle fc=1 stes. In the future, with huge pages in the host, gmap_shadow_pgt() could simply try to map a huge host page if "fake" is set to one and indicate via return value that no lower fake tables / shadow ptes are required. Acked-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-04-18 19:24:52 +08:00
if (sg->edat_level >= 1)
*table |= sgt & _REGION_ENTRY_PROTECT;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
list_add(&page->lru, &sg->crst_list);
if (fake) {
/* nothing to protect for fake tables */
*table &= ~_REGION_ENTRY_INVALID;
spin_unlock(&sg->guest_table_lock);
return 0;
}
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
spin_unlock(&sg->guest_table_lock);
/* Make sgt read-only in parent gmap page table */
raddr = (saddr & 0xffffffff80000000UL) | _SHADOW_RMAP_REGION3;
origin = sgt & _REGION_ENTRY_ORIGIN;
offset = ((sgt & _REGION_ENTRY_OFFSET) >> 6) * 4096;
len = ((sgt & _REGION_ENTRY_LENGTH) + 1) * 4096 - offset;
rc = gmap_protect_rmap(sg, raddr, origin + offset, len, PROT_READ);
spin_lock(&sg->guest_table_lock);
if (!rc) {
table = gmap_table_walk(sg, saddr, 2);
if (!table || (*table & _REGION_ENTRY_ORIGIN) !=
(unsigned long) s_sgt)
rc = -EAGAIN; /* Race with unshadow */
else
*table &= ~_REGION_ENTRY_INVALID;
} else {
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
gmap_unshadow_sgt(sg, raddr);
}
spin_unlock(&sg->guest_table_lock);
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
return rc;
out_free:
spin_unlock(&sg->guest_table_lock);
__free_pages(page, 2);
return rc;
}
EXPORT_SYMBOL_GPL(gmap_shadow_sgt);
/**
* gmap_shadow_lookup_pgtable - find a shadow page table
* @sg: pointer to the shadow guest address space structure
* @saddr: the address in the shadow aguest address space
* @pgt: parent gmap address of the page table to get shadowed
* @dat_protection: if the pgtable is marked as protected by dat
s390/mm: support EDAT1 for gmap shadows If the guest is enabled for EDAT1, we can easily create shadows for guest2 -> guest3 provided tables that make use of EDAT1. If guest2 references a 1MB page, this memory looks consecutive for guest2, but it might not be so for us. Therefore we have to create fake page tables. We can easily add that to our existing infrastructure. The invalidation mechanism will make sure that fake page tables are removed when the parent table (sgt table entry) is changed. As EDAT1 also introduced protection on all page table levels, we have to also shadow these correctly. We don't have to care about: - ACCF-Validity Control in STE - Access-Control Bits in STE - Fetch-Protection Bit in STE - Common-Segment Bit in STE As all bits might be dropped and there is no guaranteed that they are active ("unpredictable whether the CPU uses these bits", "may be used"). Without using EDAT1 in the shadow ourselfes (STE-format control == 0), simply shadowing these bits would not be enough. They would be ignored. Please note that we are using the "fake" flag to make this look consistent with further changes (EDAT2, real-space designation support) and don't let the shadow functions handle fc=1 stes. In the future, with huge pages in the host, gmap_shadow_pgt() could simply try to map a huge host page if "fake" is set to one and indicate via return value that no lower fake tables / shadow ptes are required. Acked-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-04-18 19:24:52 +08:00
* @fake: pgt references contiguous guest memory block, not a pgtable
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
*
* Returns 0 if the shadow page table was found and -EAGAIN if the page
* table was not found.
*
* Called with sg->mm->mmap_sem in read.
*/
int gmap_shadow_pgt_lookup(struct gmap *sg, unsigned long saddr,
s390/mm: support EDAT1 for gmap shadows If the guest is enabled for EDAT1, we can easily create shadows for guest2 -> guest3 provided tables that make use of EDAT1. If guest2 references a 1MB page, this memory looks consecutive for guest2, but it might not be so for us. Therefore we have to create fake page tables. We can easily add that to our existing infrastructure. The invalidation mechanism will make sure that fake page tables are removed when the parent table (sgt table entry) is changed. As EDAT1 also introduced protection on all page table levels, we have to also shadow these correctly. We don't have to care about: - ACCF-Validity Control in STE - Access-Control Bits in STE - Fetch-Protection Bit in STE - Common-Segment Bit in STE As all bits might be dropped and there is no guaranteed that they are active ("unpredictable whether the CPU uses these bits", "may be used"). Without using EDAT1 in the shadow ourselfes (STE-format control == 0), simply shadowing these bits would not be enough. They would be ignored. Please note that we are using the "fake" flag to make this look consistent with further changes (EDAT2, real-space designation support) and don't let the shadow functions handle fc=1 stes. In the future, with huge pages in the host, gmap_shadow_pgt() could simply try to map a huge host page if "fake" is set to one and indicate via return value that no lower fake tables / shadow ptes are required. Acked-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-04-18 19:24:52 +08:00
unsigned long *pgt, int *dat_protection,
int *fake)
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
{
unsigned long *table;
struct page *page;
int rc;
BUG_ON(!gmap_is_shadow(sg));
spin_lock(&sg->guest_table_lock);
table = gmap_table_walk(sg, saddr, 1); /* get segment pointer */
if (table && !(*table & _SEGMENT_ENTRY_INVALID)) {
/* Shadow page tables are full pages (pte+pgste) */
page = pfn_to_page(*table >> PAGE_SHIFT);
s390/mm: support EDAT1 for gmap shadows If the guest is enabled for EDAT1, we can easily create shadows for guest2 -> guest3 provided tables that make use of EDAT1. If guest2 references a 1MB page, this memory looks consecutive for guest2, but it might not be so for us. Therefore we have to create fake page tables. We can easily add that to our existing infrastructure. The invalidation mechanism will make sure that fake page tables are removed when the parent table (sgt table entry) is changed. As EDAT1 also introduced protection on all page table levels, we have to also shadow these correctly. We don't have to care about: - ACCF-Validity Control in STE - Access-Control Bits in STE - Fetch-Protection Bit in STE - Common-Segment Bit in STE As all bits might be dropped and there is no guaranteed that they are active ("unpredictable whether the CPU uses these bits", "may be used"). Without using EDAT1 in the shadow ourselfes (STE-format control == 0), simply shadowing these bits would not be enough. They would be ignored. Please note that we are using the "fake" flag to make this look consistent with further changes (EDAT2, real-space designation support) and don't let the shadow functions handle fc=1 stes. In the future, with huge pages in the host, gmap_shadow_pgt() could simply try to map a huge host page if "fake" is set to one and indicate via return value that no lower fake tables / shadow ptes are required. Acked-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-04-18 19:24:52 +08:00
*pgt = page->index & ~GMAP_SHADOW_FAKE_TABLE;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
*dat_protection = !!(*table & _SEGMENT_ENTRY_PROTECT);
s390/mm: support EDAT1 for gmap shadows If the guest is enabled for EDAT1, we can easily create shadows for guest2 -> guest3 provided tables that make use of EDAT1. If guest2 references a 1MB page, this memory looks consecutive for guest2, but it might not be so for us. Therefore we have to create fake page tables. We can easily add that to our existing infrastructure. The invalidation mechanism will make sure that fake page tables are removed when the parent table (sgt table entry) is changed. As EDAT1 also introduced protection on all page table levels, we have to also shadow these correctly. We don't have to care about: - ACCF-Validity Control in STE - Access-Control Bits in STE - Fetch-Protection Bit in STE - Common-Segment Bit in STE As all bits might be dropped and there is no guaranteed that they are active ("unpredictable whether the CPU uses these bits", "may be used"). Without using EDAT1 in the shadow ourselfes (STE-format control == 0), simply shadowing these bits would not be enough. They would be ignored. Please note that we are using the "fake" flag to make this look consistent with further changes (EDAT2, real-space designation support) and don't let the shadow functions handle fc=1 stes. In the future, with huge pages in the host, gmap_shadow_pgt() could simply try to map a huge host page if "fake" is set to one and indicate via return value that no lower fake tables / shadow ptes are required. Acked-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-04-18 19:24:52 +08:00
*fake = !!(page->index & GMAP_SHADOW_FAKE_TABLE);
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
rc = 0;
} else {
rc = -EAGAIN;
}
spin_unlock(&sg->guest_table_lock);
return rc;
}
EXPORT_SYMBOL_GPL(gmap_shadow_pgt_lookup);
/**
* gmap_shadow_pgt - instantiate a shadow page table
* @sg: pointer to the shadow guest address space structure
* @saddr: faulting address in the shadow gmap
* @pgt: parent gmap address of the page table to get shadowed
s390/mm: support EDAT1 for gmap shadows If the guest is enabled for EDAT1, we can easily create shadows for guest2 -> guest3 provided tables that make use of EDAT1. If guest2 references a 1MB page, this memory looks consecutive for guest2, but it might not be so for us. Therefore we have to create fake page tables. We can easily add that to our existing infrastructure. The invalidation mechanism will make sure that fake page tables are removed when the parent table (sgt table entry) is changed. As EDAT1 also introduced protection on all page table levels, we have to also shadow these correctly. We don't have to care about: - ACCF-Validity Control in STE - Access-Control Bits in STE - Fetch-Protection Bit in STE - Common-Segment Bit in STE As all bits might be dropped and there is no guaranteed that they are active ("unpredictable whether the CPU uses these bits", "may be used"). Without using EDAT1 in the shadow ourselfes (STE-format control == 0), simply shadowing these bits would not be enough. They would be ignored. Please note that we are using the "fake" flag to make this look consistent with further changes (EDAT2, real-space designation support) and don't let the shadow functions handle fc=1 stes. In the future, with huge pages in the host, gmap_shadow_pgt() could simply try to map a huge host page if "fake" is set to one and indicate via return value that no lower fake tables / shadow ptes are required. Acked-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-04-18 19:24:52 +08:00
* @fake: pgt references contiguous guest memory block, not a pgtable
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
*
* Returns 0 if successfully shadowed or already shadowed, -EAGAIN if the
* shadow table structure is incomplete, -ENOMEM if out of memory,
* -EFAULT if an address in the parent gmap could not be resolved and
*
* Called with gmap->mm->mmap_sem in read
*/
s390/mm: support EDAT1 for gmap shadows If the guest is enabled for EDAT1, we can easily create shadows for guest2 -> guest3 provided tables that make use of EDAT1. If guest2 references a 1MB page, this memory looks consecutive for guest2, but it might not be so for us. Therefore we have to create fake page tables. We can easily add that to our existing infrastructure. The invalidation mechanism will make sure that fake page tables are removed when the parent table (sgt table entry) is changed. As EDAT1 also introduced protection on all page table levels, we have to also shadow these correctly. We don't have to care about: - ACCF-Validity Control in STE - Access-Control Bits in STE - Fetch-Protection Bit in STE - Common-Segment Bit in STE As all bits might be dropped and there is no guaranteed that they are active ("unpredictable whether the CPU uses these bits", "may be used"). Without using EDAT1 in the shadow ourselfes (STE-format control == 0), simply shadowing these bits would not be enough. They would be ignored. Please note that we are using the "fake" flag to make this look consistent with further changes (EDAT2, real-space designation support) and don't let the shadow functions handle fc=1 stes. In the future, with huge pages in the host, gmap_shadow_pgt() could simply try to map a huge host page if "fake" is set to one and indicate via return value that no lower fake tables / shadow ptes are required. Acked-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-04-18 19:24:52 +08:00
int gmap_shadow_pgt(struct gmap *sg, unsigned long saddr, unsigned long pgt,
int fake)
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
{
unsigned long raddr, origin;
unsigned long *s_pgt, *table;
struct page *page;
int rc;
s390/mm: support EDAT1 for gmap shadows If the guest is enabled for EDAT1, we can easily create shadows for guest2 -> guest3 provided tables that make use of EDAT1. If guest2 references a 1MB page, this memory looks consecutive for guest2, but it might not be so for us. Therefore we have to create fake page tables. We can easily add that to our existing infrastructure. The invalidation mechanism will make sure that fake page tables are removed when the parent table (sgt table entry) is changed. As EDAT1 also introduced protection on all page table levels, we have to also shadow these correctly. We don't have to care about: - ACCF-Validity Control in STE - Access-Control Bits in STE - Fetch-Protection Bit in STE - Common-Segment Bit in STE As all bits might be dropped and there is no guaranteed that they are active ("unpredictable whether the CPU uses these bits", "may be used"). Without using EDAT1 in the shadow ourselfes (STE-format control == 0), simply shadowing these bits would not be enough. They would be ignored. Please note that we are using the "fake" flag to make this look consistent with further changes (EDAT2, real-space designation support) and don't let the shadow functions handle fc=1 stes. In the future, with huge pages in the host, gmap_shadow_pgt() could simply try to map a huge host page if "fake" is set to one and indicate via return value that no lower fake tables / shadow ptes are required. Acked-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-04-18 19:24:52 +08:00
BUG_ON(!gmap_is_shadow(sg) || (pgt & _SEGMENT_ENTRY_LARGE));
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
/* Allocate a shadow page table */
page = page_table_alloc_pgste(sg->mm);
if (!page)
return -ENOMEM;
page->index = pgt & _SEGMENT_ENTRY_ORIGIN;
s390/mm: support EDAT1 for gmap shadows If the guest is enabled for EDAT1, we can easily create shadows for guest2 -> guest3 provided tables that make use of EDAT1. If guest2 references a 1MB page, this memory looks consecutive for guest2, but it might not be so for us. Therefore we have to create fake page tables. We can easily add that to our existing infrastructure. The invalidation mechanism will make sure that fake page tables are removed when the parent table (sgt table entry) is changed. As EDAT1 also introduced protection on all page table levels, we have to also shadow these correctly. We don't have to care about: - ACCF-Validity Control in STE - Access-Control Bits in STE - Fetch-Protection Bit in STE - Common-Segment Bit in STE As all bits might be dropped and there is no guaranteed that they are active ("unpredictable whether the CPU uses these bits", "may be used"). Without using EDAT1 in the shadow ourselfes (STE-format control == 0), simply shadowing these bits would not be enough. They would be ignored. Please note that we are using the "fake" flag to make this look consistent with further changes (EDAT2, real-space designation support) and don't let the shadow functions handle fc=1 stes. In the future, with huge pages in the host, gmap_shadow_pgt() could simply try to map a huge host page if "fake" is set to one and indicate via return value that no lower fake tables / shadow ptes are required. Acked-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-04-18 19:24:52 +08:00
if (fake)
page->index |= GMAP_SHADOW_FAKE_TABLE;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
s_pgt = (unsigned long *) page_to_phys(page);
/* Install shadow page table */
spin_lock(&sg->guest_table_lock);
table = gmap_table_walk(sg, saddr, 1); /* get segment pointer */
if (!table) {
rc = -EAGAIN; /* Race with unshadow */
goto out_free;
}
if (!(*table & _SEGMENT_ENTRY_INVALID)) {
rc = 0; /* Already established */
goto out_free;
} else if (*table & _SEGMENT_ENTRY_ORIGIN) {
rc = -EAGAIN; /* Race with shadow */
goto out_free;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
}
/* mark as invalid as long as the parent table is not protected */
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
*table = (unsigned long) s_pgt | _SEGMENT_ENTRY |
(pgt & _SEGMENT_ENTRY_PROTECT) | _SEGMENT_ENTRY_INVALID;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
list_add(&page->lru, &sg->pt_list);
s390/mm: support EDAT1 for gmap shadows If the guest is enabled for EDAT1, we can easily create shadows for guest2 -> guest3 provided tables that make use of EDAT1. If guest2 references a 1MB page, this memory looks consecutive for guest2, but it might not be so for us. Therefore we have to create fake page tables. We can easily add that to our existing infrastructure. The invalidation mechanism will make sure that fake page tables are removed when the parent table (sgt table entry) is changed. As EDAT1 also introduced protection on all page table levels, we have to also shadow these correctly. We don't have to care about: - ACCF-Validity Control in STE - Access-Control Bits in STE - Fetch-Protection Bit in STE - Common-Segment Bit in STE As all bits might be dropped and there is no guaranteed that they are active ("unpredictable whether the CPU uses these bits", "may be used"). Without using EDAT1 in the shadow ourselfes (STE-format control == 0), simply shadowing these bits would not be enough. They would be ignored. Please note that we are using the "fake" flag to make this look consistent with further changes (EDAT2, real-space designation support) and don't let the shadow functions handle fc=1 stes. In the future, with huge pages in the host, gmap_shadow_pgt() could simply try to map a huge host page if "fake" is set to one and indicate via return value that no lower fake tables / shadow ptes are required. Acked-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-04-18 19:24:52 +08:00
if (fake) {
/* nothing to protect for fake tables */
*table &= ~_SEGMENT_ENTRY_INVALID;
spin_unlock(&sg->guest_table_lock);
return 0;
}
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
spin_unlock(&sg->guest_table_lock);
/* Make pgt read-only in parent gmap page table (not the pgste) */
raddr = (saddr & 0xfffffffffff00000UL) | _SHADOW_RMAP_SEGMENT;
origin = pgt & _SEGMENT_ENTRY_ORIGIN & PAGE_MASK;
rc = gmap_protect_rmap(sg, raddr, origin, PAGE_SIZE, PROT_READ);
spin_lock(&sg->guest_table_lock);
if (!rc) {
table = gmap_table_walk(sg, saddr, 1);
if (!table || (*table & _SEGMENT_ENTRY_ORIGIN) !=
(unsigned long) s_pgt)
rc = -EAGAIN; /* Race with unshadow */
else
*table &= ~_SEGMENT_ENTRY_INVALID;
} else {
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
gmap_unshadow_pgt(sg, raddr);
}
spin_unlock(&sg->guest_table_lock);
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
return rc;
out_free:
spin_unlock(&sg->guest_table_lock);
page_table_free_pgste(page);
return rc;
}
EXPORT_SYMBOL_GPL(gmap_shadow_pgt);
/**
* gmap_shadow_page - create a shadow page mapping
* @sg: pointer to the shadow guest address space structure
* @saddr: faulting address in the shadow gmap
* @pte: pte in parent gmap address space to get shadowed
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
*
* Returns 0 if successfully shadowed or already shadowed, -EAGAIN if the
* shadow table structure is incomplete, -ENOMEM if out of memory and
* -EFAULT if an address in the parent gmap could not be resolved.
*
* Called with sg->mm->mmap_sem in read.
*/
int gmap_shadow_page(struct gmap *sg, unsigned long saddr, pte_t pte)
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
{
struct gmap *parent;
struct gmap_rmap *rmap;
unsigned long vmaddr, paddr;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
spinlock_t *ptl;
pte_t *sptep, *tptep;
int prot;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
int rc;
BUG_ON(!gmap_is_shadow(sg));
parent = sg->parent;
prot = (pte_val(pte) & _PAGE_PROTECT) ? PROT_READ : PROT_WRITE;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
rmap = kzalloc(sizeof(*rmap), GFP_KERNEL);
if (!rmap)
return -ENOMEM;
rmap->raddr = (saddr & PAGE_MASK) | _SHADOW_RMAP_PGTABLE;
while (1) {
paddr = pte_val(pte) & PAGE_MASK;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
vmaddr = __gmap_translate(parent, paddr);
if (IS_ERR_VALUE(vmaddr)) {
rc = vmaddr;
break;
}
rc = radix_tree_preload(GFP_KERNEL);
if (rc)
break;
rc = -EAGAIN;
sptep = gmap_pte_op_walk(parent, paddr, &ptl);
if (sptep) {
spin_lock(&sg->guest_table_lock);
/* Get page table pointer */
tptep = (pte_t *) gmap_table_walk(sg, saddr, 0);
if (!tptep) {
spin_unlock(&sg->guest_table_lock);
gmap_pte_op_end(ptl);
radix_tree_preload_end();
break;
}
rc = ptep_shadow_pte(sg->mm, saddr, sptep, tptep, pte);
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
if (rc > 0) {
/* Success and a new mapping */
gmap_insert_rmap(sg, vmaddr, rmap);
rmap = NULL;
rc = 0;
}
gmap_pte_op_end(ptl);
spin_unlock(&sg->guest_table_lock);
}
radix_tree_preload_end();
if (!rc)
break;
rc = gmap_pte_op_fixup(parent, paddr, vmaddr, prot);
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
if (rc)
break;
}
kfree(rmap);
return rc;
}
EXPORT_SYMBOL_GPL(gmap_shadow_page);
/**
* gmap_shadow_notify - handle notifications for shadow gmap
*
* Called with sg->parent->shadow_lock.
*/
static void gmap_shadow_notify(struct gmap *sg, unsigned long vmaddr,
unsigned long offset, pte_t *pte)
{
struct gmap_rmap *rmap, *rnext, *head;
unsigned long gaddr, start, end, bits, raddr;
unsigned long *table;
BUG_ON(!gmap_is_shadow(sg));
spin_lock(&sg->parent->guest_table_lock);
table = radix_tree_lookup(&sg->parent->host_to_guest,
vmaddr >> PMD_SHIFT);
gaddr = table ? __gmap_segment_gaddr(table) + offset : 0;
spin_unlock(&sg->parent->guest_table_lock);
if (!table)
return;
spin_lock(&sg->guest_table_lock);
if (sg->removed) {
spin_unlock(&sg->guest_table_lock);
return;
}
/* Check for top level table */
start = sg->orig_asce & _ASCE_ORIGIN;
end = start + ((sg->orig_asce & _ASCE_TABLE_LENGTH) + 1) * 4096;
if (!(sg->orig_asce & _ASCE_REAL_SPACE) && gaddr >= start &&
gaddr < end) {
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
/* The complete shadow table has to go */
gmap_unshadow(sg);
spin_unlock(&sg->guest_table_lock);
list_del(&sg->list);
gmap_put(sg);
return;
}
/* Remove the page table tree from on specific entry */
head = radix_tree_delete(&sg->host_to_rmap, vmaddr >> 12);
gmap_for_each_rmap_safe(rmap, rnext, head) {
bits = rmap->raddr & _SHADOW_RMAP_MASK;
raddr = rmap->raddr ^ bits;
switch (bits) {
case _SHADOW_RMAP_REGION1:
gmap_unshadow_r2t(sg, raddr);
break;
case _SHADOW_RMAP_REGION2:
gmap_unshadow_r3t(sg, raddr);
break;
case _SHADOW_RMAP_REGION3:
gmap_unshadow_sgt(sg, raddr);
break;
case _SHADOW_RMAP_SEGMENT:
gmap_unshadow_pgt(sg, raddr);
break;
case _SHADOW_RMAP_PGTABLE:
gmap_unshadow_page(sg, raddr);
break;
}
kfree(rmap);
}
spin_unlock(&sg->guest_table_lock);
}
/**
* ptep_notify - call all invalidation callbacks for a specific pte.
* @mm: pointer to the process mm_struct
* @addr: virtual address in the process address space
* @pte: pointer to the page table entry
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
* @bits: bits from the pgste that caused the notify call
*
* This function is assumed to be called with the page table lock held
* for the pte to notify.
*/
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
void ptep_notify(struct mm_struct *mm, unsigned long vmaddr,
pte_t *pte, unsigned long bits)
{
unsigned long offset, gaddr;
unsigned long *table;
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
struct gmap *gmap, *sg, *next;
offset = ((unsigned long) pte) & (255 * sizeof(pte_t));
offset = offset * (4096 / sizeof(pte_t));
rcu_read_lock();
list_for_each_entry_rcu(gmap, &mm->context.gmap_list, list) {
s390/mm: add shadow gmap support For a nested KVM guest the outer KVM host needs to create shadow page tables for the nested guest. This patch adds the basic support to the guest address space (gmap) code. For each guest address space the inner KVM host creates, the first outer KVM host needs to create shadow page tables. The address space is identified by the ASCE loaded into the control register 1 at the time the inner SIE instruction for the second nested KVM guest is executed. The outer KVM host creates the shadow tables starting with the table identified by the ASCE on a on-demand basis. The outer KVM host will get repeated faults for all the shadow tables needed to run the second KVM guest. While a shadow page table for the second KVM guest is active the access to the origin region, segment and page tables needs to be restricted for the first KVM guest. For region and segment and page tables the first KVM guest may read the memory, but write attempt has to lead to an unshadow. This is done using the page invalid and read-only bits in the page table of the first KVM guest. If the first guest re-accesses one of the origin pages of a shadow, it gets a fault and the affected parts of the shadow page table hierarchy needs to be removed again. PGSTE tables don't have to be shadowed, as all interpretation assist can't deal with the invalid bits in the shadow pte being set differently than the original ones provided by the first KVM guest. Many bug fixes and improvements by David Hildenbrand. Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
2016-03-08 19:12:18 +08:00
if (!list_empty(&gmap->children) && (bits & PGSTE_VSIE_BIT)) {
spin_lock(&gmap->shadow_lock);
list_for_each_entry_safe(sg, next,
&gmap->children, list)
gmap_shadow_notify(sg, vmaddr, offset, pte);
spin_unlock(&gmap->shadow_lock);
}
if (!(bits & PGSTE_IN_BIT))
continue;
spin_lock(&gmap->guest_table_lock);
table = radix_tree_lookup(&gmap->host_to_guest,
vmaddr >> PMD_SHIFT);
if (table)
gaddr = __gmap_segment_gaddr(table) + offset;
spin_unlock(&gmap->guest_table_lock);
if (table)
gmap_call_notifier(gmap, gaddr, gaddr + PAGE_SIZE - 1);
}
rcu_read_unlock();
}
EXPORT_SYMBOL_GPL(ptep_notify);
static inline void thp_split_mm(struct mm_struct *mm)
{
#ifdef CONFIG_TRANSPARENT_HUGEPAGE
struct vm_area_struct *vma;
unsigned long addr;
for (vma = mm->mmap; vma != NULL; vma = vma->vm_next) {
for (addr = vma->vm_start;
addr < vma->vm_end;
addr += PAGE_SIZE)
follow_page(vma, addr, FOLL_SPLIT);
vma->vm_flags &= ~VM_HUGEPAGE;
vma->vm_flags |= VM_NOHUGEPAGE;
}
mm->def_flags |= VM_NOHUGEPAGE;
#endif
}
/*
* switch on pgstes for its userspace process (for kvm)
*/
int s390_enable_sie(void)
{
struct mm_struct *mm = current->mm;
/* Do we have pgstes? if yes, we are done */
if (mm_has_pgste(mm))
return 0;
/* Fail if the page tables are 2K */
if (!mm_alloc_pgste(mm))
return -EINVAL;
down_write(&mm->mmap_sem);
mm->context.has_pgste = 1;
/* split thp mappings and disable thp for future mappings */
thp_split_mm(mm);
up_write(&mm->mmap_sem);
return 0;
}
EXPORT_SYMBOL_GPL(s390_enable_sie);
/*
* Enable storage key handling from now on and initialize the storage
* keys with the default key.
*/
static int __s390_enable_skey(pte_t *pte, unsigned long addr,
unsigned long next, struct mm_walk *walk)
{
/*
* Remove all zero page mappings,
* after establishing a policy to forbid zero page mappings
* following faults for that page will get fresh anonymous pages
*/
if (is_zero_pfn(pte_pfn(*pte)))
ptep_xchg_direct(walk->mm, addr, pte, __pte(_PAGE_INVALID));
/* Clear storage key */
ptep_zap_key(walk->mm, addr, pte);
return 0;
}
int s390_enable_skey(void)
{
struct mm_walk walk = { .pte_entry = __s390_enable_skey };
struct mm_struct *mm = current->mm;
struct vm_area_struct *vma;
int rc = 0;
down_write(&mm->mmap_sem);
if (mm_use_skey(mm))
goto out_up;
mm->context.use_skey = 1;
for (vma = mm->mmap; vma; vma = vma->vm_next) {
if (ksm_madvise(vma, vma->vm_start, vma->vm_end,
MADV_UNMERGEABLE, &vma->vm_flags)) {
mm->context.use_skey = 0;
rc = -ENOMEM;
goto out_up;
}
}
mm->def_flags &= ~VM_MERGEABLE;
walk.mm = mm;
walk_page_range(0, TASK_SIZE, &walk);
out_up:
up_write(&mm->mmap_sem);
return rc;
}
EXPORT_SYMBOL_GPL(s390_enable_skey);
/*
* Reset CMMA state, make all pages stable again.
*/
static int __s390_reset_cmma(pte_t *pte, unsigned long addr,
unsigned long next, struct mm_walk *walk)
{
ptep_zap_unused(walk->mm, addr, pte, 1);
return 0;
}
void s390_reset_cmma(struct mm_struct *mm)
{
struct mm_walk walk = { .pte_entry = __s390_reset_cmma };
down_write(&mm->mmap_sem);
walk.mm = mm;
walk_page_range(0, TASK_SIZE, &walk);
up_write(&mm->mmap_sem);
}
EXPORT_SYMBOL_GPL(s390_reset_cmma);