linux-sg2042/mm/shuffle.c

208 lines
5.7 KiB
C
Raw Normal View History

mm: shuffle initial free memory to improve memory-side-cache utilization Patch series "mm: Randomize free memory", v10. This patch (of 3): Randomization of the page allocator improves the average utilization of a direct-mapped memory-side-cache. Memory side caching is a platform capability that Linux has been previously exposed to in HPC (high-performance computing) environments on specialty platforms. In that instance it was a smaller pool of high-bandwidth-memory relative to higher-capacity / lower-bandwidth DRAM. Now, this capability is going to be found on general purpose server platforms where DRAM is a cache in front of higher latency persistent memory [1]. Robert offered an explanation of the state of the art of Linux interactions with memory-side-caches [2], and I copy it here: It's been a problem in the HPC space: http://www.nersc.gov/research-and-development/knl-cache-mode-performance-coe/ A kernel module called zonesort is available to try to help: https://software.intel.com/en-us/articles/xeon-phi-software and this abandoned patch series proposed that for the kernel: https://lkml.kernel.org/r/20170823100205.17311-1-lukasz.daniluk@intel.com Dan's patch series doesn't attempt to ensure buffers won't conflict, but also reduces the chance that the buffers will. This will make performance more consistent, albeit slower than "optimal" (which is near impossible to attain in a general-purpose kernel). That's better than forcing users to deploy remedies like: "To eliminate this gradual degradation, we have added a Stream measurement to the Node Health Check that follows each job; nodes are rebooted whenever their measured memory bandwidth falls below 300 GB/s." A replacement for zonesort was merged upstream in commit cc9aec03e58f ("x86/numa_emulation: Introduce uniform split capability"). With this numa_emulation capability, memory can be split into cache sized ("near-memory" sized) numa nodes. A bind operation to such a node, and disabling workloads on other nodes, enables full cache performance. However, once the workload exceeds the cache size then cache conflicts are unavoidable. While HPC environments might be able to tolerate time-scheduling of cache sized workloads, for general purpose server platforms, the oversubscribed cache case will be the common case. The worst case scenario is that a server system owner benchmarks a workload at boot with an un-contended cache only to see that performance degrade over time, even below the average cache performance due to excessive conflicts. Randomization clips the peaks and fills in the valleys of cache utilization to yield steady average performance. Here are some performance impact details of the patches: 1/ An Intel internal synthetic memory bandwidth measurement tool, saw a 3X speedup in a contrived case that tries to force cache conflicts. The contrived cased used the numa_emulation capability to force an instance of the benchmark to be run in two of the near-memory sized numa nodes. If both instances were placed on the same emulated they would fit and cause zero conflicts. While on separate emulated nodes without randomization they underutilized the cache and conflicted unnecessarily due to the in-order allocation per node. 2/ A well known Java server application benchmark was run with a heap size that exceeded cache size by 3X. The cache conflict rate was 8% for the first run and degraded to 21% after page allocator aging. With randomization enabled the rate levelled out at 11%. 3/ A MongoDB workload did not observe measurable difference in cache-conflict rates, but the overall throughput dropped by 7% with randomization in one case. 4/ Mel Gorman ran his suite of performance workloads with randomization enabled on platforms without a memory-side-cache and saw a mix of some improvements and some losses [3]. While there is potentially significant improvement for applications that depend on low latency access across a wide working-set, the performance may be negligible to negative for other workloads. For this reason the shuffle capability defaults to off unless a direct-mapped memory-side-cache is detected. Even then, the page_alloc.shuffle=0 parameter can be specified to disable the randomization on those systems. Outside of memory-side-cache utilization concerns there is potentially security benefit from randomization. Some data exfiltration and return-oriented-programming attacks rely on the ability to infer the location of sensitive data objects. The kernel page allocator, especially early in system boot, has predictable first-in-first out behavior for physical pages. Pages are freed in physical address order when first onlined. Quoting Kees: "While we already have a base-address randomization (CONFIG_RANDOMIZE_MEMORY), attacks against the same hardware and memory layouts would certainly be using the predictability of allocation ordering (i.e. for attacks where the base address isn't important: only the relative positions between allocated memory). This is common in lots of heap-style attacks. They try to gain control over ordering by spraying allocations, etc. I'd really like to see this because it gives us something similar to CONFIG_SLAB_FREELIST_RANDOM but for the page allocator." While SLAB_FREELIST_RANDOM reduces the predictability of some local slab caches it leaves vast bulk of memory to be predictably in order allocated. However, it should be noted, the concrete security benefits are hard to quantify, and no known CVE is mitigated by this randomization. Introduce shuffle_free_memory(), and its helper shuffle_zone(), to perform a Fisher-Yates shuffle of the page allocator 'free_area' lists when they are initially populated with free memory at boot and at hotplug time. Do this based on either the presence of a page_alloc.shuffle=Y command line parameter, or autodetection of a memory-side-cache (to be added in a follow-on patch). The shuffling is done in terms of CONFIG_SHUFFLE_PAGE_ORDER sized free pages where the default CONFIG_SHUFFLE_PAGE_ORDER is MAX_ORDER-1 i.e. 10, 4MB this trades off randomization granularity for time spent shuffling. MAX_ORDER-1 was chosen to be minimally invasive to the page allocator while still showing memory-side cache behavior improvements, and the expectation that the security implications of finer granularity randomization is mitigated by CONFIG_SLAB_FREELIST_RANDOM. The performance impact of the shuffling appears to be in the noise compared to other memory initialization work. This initial randomization can be undone over time so a follow-on patch is introduced to inject entropy on page free decisions. It is reasonable to ask if the page free entropy is sufficient, but it is not enough due to the in-order initial freeing of pages. At the start of that process putting page1 in front or behind page0 still keeps them close together, page2 is still near page1 and has a high chance of being adjacent. As more pages are added ordering diversity improves, but there is still high page locality for the low address pages and this leads to no significant impact to the cache conflict rate. [1]: https://itpeernetwork.intel.com/intel-optane-dc-persistent-memory-operating-modes/ [2]: https://lkml.kernel.org/r/AT5PR8401MB1169D656C8B5E121752FC0F8AB120@AT5PR8401MB1169.NAMPRD84.PROD.OUTLOOK.COM [3]: https://lkml.org/lkml/2018/10/12/309 [dan.j.williams@intel.com: fix shuffle enable] Link: http://lkml.kernel.org/r/154943713038.3858443.4125180191382062871.stgit@dwillia2-desk3.amr.corp.intel.com [cai@lca.pw: fix SHUFFLE_PAGE_ALLOCATOR help texts] Link: http://lkml.kernel.org/r/20190425201300.75650-1-cai@lca.pw Link: http://lkml.kernel.org/r/154899811738.3165233.12325692939590944259.stgit@dwillia2-desk3.amr.corp.intel.com Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Qian Cai <cai@lca.pw> Reviewed-by: Kees Cook <keescook@chromium.org> Acked-by: Michal Hocko <mhocko@suse.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Keith Busch <keith.busch@intel.com> Cc: Robert Elliott <elliott@hpe.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-15 06:41:28 +08:00
// SPDX-License-Identifier: GPL-2.0
// Copyright(c) 2018 Intel Corporation. All rights reserved.
#include <linux/mm.h>
#include <linux/init.h>
#include <linux/mmzone.h>
#include <linux/random.h>
#include <linux/moduleparam.h>
#include "internal.h"
#include "shuffle.h"
DEFINE_STATIC_KEY_FALSE(page_alloc_shuffle_key);
static unsigned long shuffle_state __ro_after_init;
/*
* Depending on the architecture, module parameter parsing may run
* before, or after the cache detection. SHUFFLE_FORCE_DISABLE prevents,
* or reverts the enabling of the shuffle implementation. SHUFFLE_ENABLE
* attempts to turn on the implementation, but aborts if it finds
* SHUFFLE_FORCE_DISABLE already set.
*/
__meminit void page_alloc_shuffle(enum mm_shuffle_ctl ctl)
{
if (ctl == SHUFFLE_FORCE_DISABLE)
set_bit(SHUFFLE_FORCE_DISABLE, &shuffle_state);
if (test_bit(SHUFFLE_FORCE_DISABLE, &shuffle_state)) {
if (test_and_clear_bit(SHUFFLE_ENABLE, &shuffle_state))
static_branch_disable(&page_alloc_shuffle_key);
} else if (ctl == SHUFFLE_ENABLE
&& !test_and_set_bit(SHUFFLE_ENABLE, &shuffle_state))
static_branch_enable(&page_alloc_shuffle_key);
}
static bool shuffle_param;
static int shuffle_show(char *buffer, const struct kernel_param *kp)
mm: shuffle initial free memory to improve memory-side-cache utilization Patch series "mm: Randomize free memory", v10. This patch (of 3): Randomization of the page allocator improves the average utilization of a direct-mapped memory-side-cache. Memory side caching is a platform capability that Linux has been previously exposed to in HPC (high-performance computing) environments on specialty platforms. In that instance it was a smaller pool of high-bandwidth-memory relative to higher-capacity / lower-bandwidth DRAM. Now, this capability is going to be found on general purpose server platforms where DRAM is a cache in front of higher latency persistent memory [1]. Robert offered an explanation of the state of the art of Linux interactions with memory-side-caches [2], and I copy it here: It's been a problem in the HPC space: http://www.nersc.gov/research-and-development/knl-cache-mode-performance-coe/ A kernel module called zonesort is available to try to help: https://software.intel.com/en-us/articles/xeon-phi-software and this abandoned patch series proposed that for the kernel: https://lkml.kernel.org/r/20170823100205.17311-1-lukasz.daniluk@intel.com Dan's patch series doesn't attempt to ensure buffers won't conflict, but also reduces the chance that the buffers will. This will make performance more consistent, albeit slower than "optimal" (which is near impossible to attain in a general-purpose kernel). That's better than forcing users to deploy remedies like: "To eliminate this gradual degradation, we have added a Stream measurement to the Node Health Check that follows each job; nodes are rebooted whenever their measured memory bandwidth falls below 300 GB/s." A replacement for zonesort was merged upstream in commit cc9aec03e58f ("x86/numa_emulation: Introduce uniform split capability"). With this numa_emulation capability, memory can be split into cache sized ("near-memory" sized) numa nodes. A bind operation to such a node, and disabling workloads on other nodes, enables full cache performance. However, once the workload exceeds the cache size then cache conflicts are unavoidable. While HPC environments might be able to tolerate time-scheduling of cache sized workloads, for general purpose server platforms, the oversubscribed cache case will be the common case. The worst case scenario is that a server system owner benchmarks a workload at boot with an un-contended cache only to see that performance degrade over time, even below the average cache performance due to excessive conflicts. Randomization clips the peaks and fills in the valleys of cache utilization to yield steady average performance. Here are some performance impact details of the patches: 1/ An Intel internal synthetic memory bandwidth measurement tool, saw a 3X speedup in a contrived case that tries to force cache conflicts. The contrived cased used the numa_emulation capability to force an instance of the benchmark to be run in two of the near-memory sized numa nodes. If both instances were placed on the same emulated they would fit and cause zero conflicts. While on separate emulated nodes without randomization they underutilized the cache and conflicted unnecessarily due to the in-order allocation per node. 2/ A well known Java server application benchmark was run with a heap size that exceeded cache size by 3X. The cache conflict rate was 8% for the first run and degraded to 21% after page allocator aging. With randomization enabled the rate levelled out at 11%. 3/ A MongoDB workload did not observe measurable difference in cache-conflict rates, but the overall throughput dropped by 7% with randomization in one case. 4/ Mel Gorman ran his suite of performance workloads with randomization enabled on platforms without a memory-side-cache and saw a mix of some improvements and some losses [3]. While there is potentially significant improvement for applications that depend on low latency access across a wide working-set, the performance may be negligible to negative for other workloads. For this reason the shuffle capability defaults to off unless a direct-mapped memory-side-cache is detected. Even then, the page_alloc.shuffle=0 parameter can be specified to disable the randomization on those systems. Outside of memory-side-cache utilization concerns there is potentially security benefit from randomization. Some data exfiltration and return-oriented-programming attacks rely on the ability to infer the location of sensitive data objects. The kernel page allocator, especially early in system boot, has predictable first-in-first out behavior for physical pages. Pages are freed in physical address order when first onlined. Quoting Kees: "While we already have a base-address randomization (CONFIG_RANDOMIZE_MEMORY), attacks against the same hardware and memory layouts would certainly be using the predictability of allocation ordering (i.e. for attacks where the base address isn't important: only the relative positions between allocated memory). This is common in lots of heap-style attacks. They try to gain control over ordering by spraying allocations, etc. I'd really like to see this because it gives us something similar to CONFIG_SLAB_FREELIST_RANDOM but for the page allocator." While SLAB_FREELIST_RANDOM reduces the predictability of some local slab caches it leaves vast bulk of memory to be predictably in order allocated. However, it should be noted, the concrete security benefits are hard to quantify, and no known CVE is mitigated by this randomization. Introduce shuffle_free_memory(), and its helper shuffle_zone(), to perform a Fisher-Yates shuffle of the page allocator 'free_area' lists when they are initially populated with free memory at boot and at hotplug time. Do this based on either the presence of a page_alloc.shuffle=Y command line parameter, or autodetection of a memory-side-cache (to be added in a follow-on patch). The shuffling is done in terms of CONFIG_SHUFFLE_PAGE_ORDER sized free pages where the default CONFIG_SHUFFLE_PAGE_ORDER is MAX_ORDER-1 i.e. 10, 4MB this trades off randomization granularity for time spent shuffling. MAX_ORDER-1 was chosen to be minimally invasive to the page allocator while still showing memory-side cache behavior improvements, and the expectation that the security implications of finer granularity randomization is mitigated by CONFIG_SLAB_FREELIST_RANDOM. The performance impact of the shuffling appears to be in the noise compared to other memory initialization work. This initial randomization can be undone over time so a follow-on patch is introduced to inject entropy on page free decisions. It is reasonable to ask if the page free entropy is sufficient, but it is not enough due to the in-order initial freeing of pages. At the start of that process putting page1 in front or behind page0 still keeps them close together, page2 is still near page1 and has a high chance of being adjacent. As more pages are added ordering diversity improves, but there is still high page locality for the low address pages and this leads to no significant impact to the cache conflict rate. [1]: https://itpeernetwork.intel.com/intel-optane-dc-persistent-memory-operating-modes/ [2]: https://lkml.kernel.org/r/AT5PR8401MB1169D656C8B5E121752FC0F8AB120@AT5PR8401MB1169.NAMPRD84.PROD.OUTLOOK.COM [3]: https://lkml.org/lkml/2018/10/12/309 [dan.j.williams@intel.com: fix shuffle enable] Link: http://lkml.kernel.org/r/154943713038.3858443.4125180191382062871.stgit@dwillia2-desk3.amr.corp.intel.com [cai@lca.pw: fix SHUFFLE_PAGE_ALLOCATOR help texts] Link: http://lkml.kernel.org/r/20190425201300.75650-1-cai@lca.pw Link: http://lkml.kernel.org/r/154899811738.3165233.12325692939590944259.stgit@dwillia2-desk3.amr.corp.intel.com Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Qian Cai <cai@lca.pw> Reviewed-by: Kees Cook <keescook@chromium.org> Acked-by: Michal Hocko <mhocko@suse.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Keith Busch <keith.busch@intel.com> Cc: Robert Elliott <elliott@hpe.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-15 06:41:28 +08:00
{
return sprintf(buffer, "%c\n", test_bit(SHUFFLE_ENABLE, &shuffle_state)
? 'Y' : 'N');
}
static __meminit int shuffle_store(const char *val,
const struct kernel_param *kp)
{
int rc = param_set_bool(val, kp);
if (rc < 0)
return rc;
if (shuffle_param)
page_alloc_shuffle(SHUFFLE_ENABLE);
else
page_alloc_shuffle(SHUFFLE_FORCE_DISABLE);
return 0;
}
module_param_call(shuffle, shuffle_store, shuffle_show, &shuffle_param, 0400);
/*
* For two pages to be swapped in the shuffle, they must be free (on a
* 'free_area' lru), have the same order, and have the same migratetype.
*/
static struct page * __meminit shuffle_valid_page(unsigned long pfn, int order)
{
struct page *page;
/*
* Given we're dealing with randomly selected pfns in a zone we
* need to ask questions like...
*/
/* ...is the pfn even in the memmap? */
if (!pfn_valid_within(pfn))
return NULL;
/* ...is the pfn in a present section or a hole? */
if (!pfn_present(pfn))
return NULL;
/* ...is the page free and currently on a free_area list? */
page = pfn_to_page(pfn);
if (!PageBuddy(page))
return NULL;
/*
* ...is the page on the same list as the page we will
* shuffle it with?
*/
if (page_order(page) != order)
return NULL;
return page;
}
/*
* Fisher-Yates shuffle the freelist which prescribes iterating through an
* array, pfns in this case, and randomly swapping each entry with another in
* the span, end_pfn - start_pfn.
*
* To keep the implementation simple it does not attempt to correct for sources
* of bias in the distribution, like modulo bias or pseudo-random number
* generator bias. I.e. the expectation is that this shuffling raises the bar
* for attacks that exploit the predictability of page allocations, but need not
* be a perfect shuffle.
*/
#define SHUFFLE_RETRY 10
void __meminit __shuffle_zone(struct zone *z)
{
unsigned long i, flags;
unsigned long start_pfn = z->zone_start_pfn;
unsigned long end_pfn = zone_end_pfn(z);
const int order = SHUFFLE_ORDER;
const int order_pages = 1 << order;
spin_lock_irqsave(&z->lock, flags);
start_pfn = ALIGN(start_pfn, order_pages);
for (i = start_pfn; i < end_pfn; i += order_pages) {
unsigned long j;
int migratetype, retry;
struct page *page_i, *page_j;
/*
* We expect page_i, in the sub-range of a zone being added
* (@start_pfn to @end_pfn), to more likely be valid compared to
* page_j randomly selected in the span @zone_start_pfn to
* @spanned_pages.
*/
page_i = shuffle_valid_page(i, order);
if (!page_i)
continue;
for (retry = 0; retry < SHUFFLE_RETRY; retry++) {
/*
* Pick a random order aligned page in the zone span as
* a swap target. If the selected pfn is a hole, retry
* up to SHUFFLE_RETRY attempts find a random valid pfn
* in the zone.
*/
j = z->zone_start_pfn +
ALIGN_DOWN(get_random_long() % z->spanned_pages,
order_pages);
page_j = shuffle_valid_page(j, order);
if (page_j && page_j != page_i)
break;
}
if (retry >= SHUFFLE_RETRY) {
pr_debug("%s: failed to swap %#lx\n", __func__, i);
continue;
}
/*
* Each migratetype corresponds to its own list, make sure the
* types match otherwise we're moving pages to lists where they
* do not belong.
*/
migratetype = get_pageblock_migratetype(page_i);
if (get_pageblock_migratetype(page_j) != migratetype) {
pr_debug("%s: migratetype mismatch %#lx\n", __func__, i);
continue;
}
list_swap(&page_i->lru, &page_j->lru);
pr_debug("%s: swap: %#lx -> %#lx\n", __func__, i, j);
/* take it easy on the zone lock */
if ((i % (100 * order_pages)) == 0) {
spin_unlock_irqrestore(&z->lock, flags);
cond_resched();
spin_lock_irqsave(&z->lock, flags);
}
}
spin_unlock_irqrestore(&z->lock, flags);
}
/**
* shuffle_free_memory - reduce the predictability of the page allocator
* @pgdat: node page data
*/
void __meminit __shuffle_free_memory(pg_data_t *pgdat)
{
struct zone *z;
for (z = pgdat->node_zones; z < pgdat->node_zones + MAX_NR_ZONES; z++)
shuffle_zone(z);
}
void add_to_free_area_random(struct page *page, struct free_area *area,
int migratetype)
{
static u64 rand;
static u8 rand_bits;
/*
* The lack of locking is deliberate. If 2 threads race to
* update the rand state it just adds to the entropy.
*/
if (rand_bits == 0) {
rand_bits = 64;
rand = get_random_u64();
}
if (rand & 1)
add_to_free_area(page, area, migratetype);
else
add_to_free_area_tail(page, area, migratetype);
rand_bits--;
rand >>= 1;
}