hollalinux
/
linux-sg2042
mirror of https://github.com/sophgo/linux-riscv.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
linux-sg2042/arch/powerpc/mm/8xx_mmu.c

206 lines
5.1 KiB
C
Raw Normal View History

powerpc/8xx: Map linear kernel RAM with 8M pages On a live running system (VoIP gateway for Air Trafic Control), over a 10 minutes period (with 277s idle), we get 87 millions DTLB misses and approximatly 35 secondes are spent in DTLB handler. This represents 5.8% of the overall time and even 10.8% of the non-idle time. Among those 87 millions DTLB misses, 15% are on user addresses and 85% are on kernel addresses. And within the kernel addresses, 93% are on addresses from the linear address space and only 7% are on addresses from the virtual address space. MPC8xx has no BATs but it has 8Mb page size. This patch implements mapping of kernel RAM using 8Mb pages, on the same model as what is done on the 40x. In 4k pages mode, each PGD entry maps a 4Mb area: we map every two entries to the same 8Mb physical page. In each second entry, we add 4Mb to the page physical address to ease life of the FixupDAR routine. This is just ignored by HW. In 16k pages mode, each PGD entry maps a 64Mb area: each PGD entry will point to the first page of the area. The DTLB handler adds the 3 bits from EPN to map the correct page. With this patch applied, we now get only 13 millions TLB misses during the 10 minutes period. The idle time has increased to 313s and the overall time spent in DTLB miss handler is 6.3s, which represents 1% of the overall time and 2.2% of non-idle time. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-02-10 00:07:50 +08:00
/*
* This file contains the routines for initializing the MMU
* on the 8xx series of chips.
* -- christophe
*
* Derived from arch/powerpc/mm/40x_mmu.c:
*
* This program is free software; you can redistribute it and/or
* modify it under the terms of the GNU General Public License
* as published by the Free Software Foundation; either version
* 2 of the License, or (at your option) any later version.
*
*/
#include <linux/memblock.h>
powerpc/8xx: Map IMMR area with 512k page at a fixed address Once the linear memory space has been mapped with 8Mb pages, as seen in the related commit, we get 11 millions DTLB missed during the reference 600s period. 77% of the misses are on user addresses and 23% are on kernel addresses (1 fourth for linear address space and 3 fourth for virtual address space) Traditionaly, each driver manages one computer board which has its own components with its own memory maps. But on embedded chips like the MPC8xx, the SOC has all registers located in the same IO area. When looking at ioremaps done during startup, we see that many drivers are re-mapping small parts of the IMMR for their own use and all those small pieces gets their own 4k page, amplifying the number of TLB misses: in our system we get 0xff000000 mapped 31 times and 0xff003000 mapped 9 times. Even if each part of IMMR was mapped only once with 4k pages, it would still be several small mappings towards linear area. This patch maps the IMMR with a single 512k page. With this patch applied, the number of DTLB misses during the 10 min period is reduced to 11.8 millions for a duration of 5.8s, which represents 2% of the non-idle time hence yet another 10% reduction. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:45 +08:00
#include <asm/fixmap.h>
#include <asm/code-patching.h>
powerpc/8xx: Map linear kernel RAM with 8M pages On a live running system (VoIP gateway for Air Trafic Control), over a 10 minutes period (with 277s idle), we get 87 millions DTLB misses and approximatly 35 secondes are spent in DTLB handler. This represents 5.8% of the overall time and even 10.8% of the non-idle time. Among those 87 millions DTLB misses, 15% are on user addresses and 85% are on kernel addresses. And within the kernel addresses, 93% are on addresses from the linear address space and only 7% are on addresses from the virtual address space. MPC8xx has no BATs but it has 8Mb page size. This patch implements mapping of kernel RAM using 8Mb pages, on the same model as what is done on the 40x. In 4k pages mode, each PGD entry maps a 4Mb area: we map every two entries to the same 8Mb physical page. In each second entry, we add 4Mb to the page physical address to ease life of the FixupDAR routine. This is just ignored by HW. In 16k pages mode, each PGD entry maps a 64Mb area: each PGD entry will point to the first page of the area. The DTLB handler adds the 3 bits from EPN to map the correct page. With this patch applied, we now get only 13 millions TLB misses during the 10 minutes period. The idle time has increased to 313s and the overall time spent in DTLB miss handler is 6.3s, which represents 1% of the overall time and 2.2% of non-idle time. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-02-10 00:07:50 +08:00
#include "mmu_decl.h"
powerpc/8xx: Map IMMR area with 512k page at a fixed address Once the linear memory space has been mapped with 8Mb pages, as seen in the related commit, we get 11 millions DTLB missed during the reference 600s period. 77% of the misses are on user addresses and 23% are on kernel addresses (1 fourth for linear address space and 3 fourth for virtual address space) Traditionaly, each driver manages one computer board which has its own components with its own memory maps. But on embedded chips like the MPC8xx, the SOC has all registers located in the same IO area. When looking at ioremaps done during startup, we see that many drivers are re-mapping small parts of the IMMR for their own use and all those small pieces gets their own 4k page, amplifying the number of TLB misses: in our system we get 0xff000000 mapped 31 times and 0xff003000 mapped 9 times. Even if each part of IMMR was mapped only once with 4k pages, it would still be several small mappings towards linear area. This patch maps the IMMR with a single 512k page. With this patch applied, the number of DTLB misses during the 10 min period is reduced to 11.8 millions for a duration of 5.8s, which represents 2% of the non-idle time hence yet another 10% reduction. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:45 +08:00
#define IMMR_SIZE (FIX_IMMR_SIZE << PAGE_SHIFT)
powerpc/8xx: Map linear kernel RAM with 8M pages On a live running system (VoIP gateway for Air Trafic Control), over a 10 minutes period (with 277s idle), we get 87 millions DTLB misses and approximatly 35 secondes are spent in DTLB handler. This represents 5.8% of the overall time and even 10.8% of the non-idle time. Among those 87 millions DTLB misses, 15% are on user addresses and 85% are on kernel addresses. And within the kernel addresses, 93% are on addresses from the linear address space and only 7% are on addresses from the virtual address space. MPC8xx has no BATs but it has 8Mb page size. This patch implements mapping of kernel RAM using 8Mb pages, on the same model as what is done on the 40x. In 4k pages mode, each PGD entry maps a 4Mb area: we map every two entries to the same 8Mb physical page. In each second entry, we add 4Mb to the page physical address to ease life of the FixupDAR routine. This is just ignored by HW. In 16k pages mode, each PGD entry maps a 64Mb area: each PGD entry will point to the first page of the area. The DTLB handler adds the 3 bits from EPN to map the correct page. With this patch applied, we now get only 13 millions TLB misses during the 10 minutes period. The idle time has increased to 313s and the overall time spent in DTLB miss handler is 6.3s, which represents 1% of the overall time and 2.2% of non-idle time. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-02-10 00:07:50 +08:00
extern int __map_without_ltlbs;
powerpc/8xx: Map IMMR area with 512k page at a fixed address Once the linear memory space has been mapped with 8Mb pages, as seen in the related commit, we get 11 millions DTLB missed during the reference 600s period. 77% of the misses are on user addresses and 23% are on kernel addresses (1 fourth for linear address space and 3 fourth for virtual address space) Traditionaly, each driver manages one computer board which has its own components with its own memory maps. But on embedded chips like the MPC8xx, the SOC has all registers located in the same IO area. When looking at ioremaps done during startup, we see that many drivers are re-mapping small parts of the IMMR for their own use and all those small pieces gets their own 4k page, amplifying the number of TLB misses: in our system we get 0xff000000 mapped 31 times and 0xff003000 mapped 9 times. Even if each part of IMMR was mapped only once with 4k pages, it would still be several small mappings towards linear area. This patch maps the IMMR with a single 512k page. With this patch applied, the number of DTLB misses during the 10 min period is reduced to 11.8 millions for a duration of 5.8s, which represents 2% of the non-idle time hence yet another 10% reduction. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:45 +08:00
powerpc/8xx: Ensures RAM mapped with LTLB is seen as block mapped on 8xx. On the 8xx, the RAM mapped with LTLBs must be seen as block mapped, just like areas mapped with BATs on standard PPC32. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2017-07-12 18:08:45 +08:00
static unsigned long block_mapped_ram;
powerpc/8xx: Map IMMR area with 512k page at a fixed address Once the linear memory space has been mapped with 8Mb pages, as seen in the related commit, we get 11 millions DTLB missed during the reference 600s period. 77% of the misses are on user addresses and 23% are on kernel addresses (1 fourth for linear address space and 3 fourth for virtual address space) Traditionaly, each driver manages one computer board which has its own components with its own memory maps. But on embedded chips like the MPC8xx, the SOC has all registers located in the same IO area. When looking at ioremaps done during startup, we see that many drivers are re-mapping small parts of the IMMR for their own use and all those small pieces gets their own 4k page, amplifying the number of TLB misses: in our system we get 0xff000000 mapped 31 times and 0xff003000 mapped 9 times. Even if each part of IMMR was mapped only once with 4k pages, it would still be several small mappings towards linear area. This patch maps the IMMR with a single 512k page. With this patch applied, the number of DTLB misses during the 10 min period is reduced to 11.8 millions for a duration of 5.8s, which represents 2% of the non-idle time hence yet another 10% reduction. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:45 +08:00
/*
powerpc/8xx: Ensures RAM mapped with LTLB is seen as block mapped on 8xx. On the 8xx, the RAM mapped with LTLBs must be seen as block mapped, just like areas mapped with BATs on standard PPC32. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2017-07-12 18:08:45 +08:00
* Return PA for this VA if it is in an area mapped with LTLBs.
* Otherwise, returns 0
powerpc/8xx: Map IMMR area with 512k page at a fixed address Once the linear memory space has been mapped with 8Mb pages, as seen in the related commit, we get 11 millions DTLB missed during the reference 600s period. 77% of the misses are on user addresses and 23% are on kernel addresses (1 fourth for linear address space and 3 fourth for virtual address space) Traditionaly, each driver manages one computer board which has its own components with its own memory maps. But on embedded chips like the MPC8xx, the SOC has all registers located in the same IO area. When looking at ioremaps done during startup, we see that many drivers are re-mapping small parts of the IMMR for their own use and all those small pieces gets their own 4k page, amplifying the number of TLB misses: in our system we get 0xff000000 mapped 31 times and 0xff003000 mapped 9 times. Even if each part of IMMR was mapped only once with 4k pages, it would still be several small mappings towards linear area. This patch maps the IMMR with a single 512k page. With this patch applied, the number of DTLB misses during the 10 min period is reduced to 11.8 millions for a duration of 5.8s, which represents 2% of the non-idle time hence yet another 10% reduction. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:45 +08:00
*/
phys_addr_t v_block_mapped(unsigned long va)
{
unsigned long p = PHYS_IMMR_BASE;
if (__map_without_ltlbs)
return 0;
if (va >= VIRT_IMMR_BASE && va < VIRT_IMMR_BASE + IMMR_SIZE)
return p + va - VIRT_IMMR_BASE;
powerpc/8xx: Ensures RAM mapped with LTLB is seen as block mapped on 8xx. On the 8xx, the RAM mapped with LTLBs must be seen as block mapped, just like areas mapped with BATs on standard PPC32. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2017-07-12 18:08:45 +08:00
if (va >= PAGE_OFFSET && va < PAGE_OFFSET + block_mapped_ram)
return __pa(va);
powerpc/8xx: Map IMMR area with 512k page at a fixed address Once the linear memory space has been mapped with 8Mb pages, as seen in the related commit, we get 11 millions DTLB missed during the reference 600s period. 77% of the misses are on user addresses and 23% are on kernel addresses (1 fourth for linear address space and 3 fourth for virtual address space) Traditionaly, each driver manages one computer board which has its own components with its own memory maps. But on embedded chips like the MPC8xx, the SOC has all registers located in the same IO area. When looking at ioremaps done during startup, we see that many drivers are re-mapping small parts of the IMMR for their own use and all those small pieces gets their own 4k page, amplifying the number of TLB misses: in our system we get 0xff000000 mapped 31 times and 0xff003000 mapped 9 times. Even if each part of IMMR was mapped only once with 4k pages, it would still be several small mappings towards linear area. This patch maps the IMMR with a single 512k page. With this patch applied, the number of DTLB misses during the 10 min period is reduced to 11.8 millions for a duration of 5.8s, which represents 2% of the non-idle time hence yet another 10% reduction. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:45 +08:00
return 0;
}
/*
powerpc/8xx: Ensures RAM mapped with LTLB is seen as block mapped on 8xx. On the 8xx, the RAM mapped with LTLBs must be seen as block mapped, just like areas mapped with BATs on standard PPC32. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2017-07-12 18:08:45 +08:00
* Return VA for a given PA mapped with LTLBs or 0 if not mapped
powerpc/8xx: Map IMMR area with 512k page at a fixed address Once the linear memory space has been mapped with 8Mb pages, as seen in the related commit, we get 11 millions DTLB missed during the reference 600s period. 77% of the misses are on user addresses and 23% are on kernel addresses (1 fourth for linear address space and 3 fourth for virtual address space) Traditionaly, each driver manages one computer board which has its own components with its own memory maps. But on embedded chips like the MPC8xx, the SOC has all registers located in the same IO area. When looking at ioremaps done during startup, we see that many drivers are re-mapping small parts of the IMMR for their own use and all those small pieces gets their own 4k page, amplifying the number of TLB misses: in our system we get 0xff000000 mapped 31 times and 0xff003000 mapped 9 times. Even if each part of IMMR was mapped only once with 4k pages, it would still be several small mappings towards linear area. This patch maps the IMMR with a single 512k page. With this patch applied, the number of DTLB misses during the 10 min period is reduced to 11.8 millions for a duration of 5.8s, which represents 2% of the non-idle time hence yet another 10% reduction. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:45 +08:00
*/
unsigned long p_block_mapped(phys_addr_t pa)
{
unsigned long p = PHYS_IMMR_BASE;
if (__map_without_ltlbs)
return 0;
if (pa >= p && pa < p + IMMR_SIZE)
return VIRT_IMMR_BASE + pa - p;
powerpc/8xx: Ensures RAM mapped with LTLB is seen as block mapped on 8xx. On the 8xx, the RAM mapped with LTLBs must be seen as block mapped, just like areas mapped with BATs on standard PPC32. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2017-07-12 18:08:45 +08:00
if (pa < block_mapped_ram)
return (unsigned long)__va(pa);
powerpc/8xx: Map IMMR area with 512k page at a fixed address Once the linear memory space has been mapped with 8Mb pages, as seen in the related commit, we get 11 millions DTLB missed during the reference 600s period. 77% of the misses are on user addresses and 23% are on kernel addresses (1 fourth for linear address space and 3 fourth for virtual address space) Traditionaly, each driver manages one computer board which has its own components with its own memory maps. But on embedded chips like the MPC8xx, the SOC has all registers located in the same IO area. When looking at ioremaps done during startup, we see that many drivers are re-mapping small parts of the IMMR for their own use and all those small pieces gets their own 4k page, amplifying the number of TLB misses: in our system we get 0xff000000 mapped 31 times and 0xff003000 mapped 9 times. Even if each part of IMMR was mapped only once with 4k pages, it would still be several small mappings towards linear area. This patch maps the IMMR with a single 512k page. With this patch applied, the number of DTLB misses during the 10 min period is reduced to 11.8 millions for a duration of 5.8s, which represents 2% of the non-idle time hence yet another 10% reduction. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:45 +08:00
return 0;
}
powerpc/8xx: Rework CONFIG_PIN_TLB handling On recent kernels, with some debug options like for instance CONFIG_LOCKDEP, the BSS requires more than 8M memory, allthough the kernel code fits in the first 8M. Today, it is necessary to activate CONFIG_PIN_TLB to get more than 8M at startup, allthough pinning TLB is not necessary for that. We could have inconditionaly mapped 16 or 24M bytes at startup but some old hardware only have 8M and mapping non-existing RAM would be an issue due to speculative accesses. With the preceding patch however, the TLB entries are populated on demand. By setting up the TLB miss handler to handle up to 24M until the handler is patched for the entire memory space, it is possible to allow access up to more memory without mapping non-existing RAM. It is therefore not needed anymore to map memory data at all at startup. It will be handled by the TLB miss handler. One might still want to PIN the IMMR and the first 24M of RAM. It is now possible to do it in the C memory initialisation functions. In addition, we now know how much memory we have when we do it, so we are able to adapt the pining to the real amount of memory available. So boards with less than 24M can now also benefit from PIN_TLB. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:54 +08:00
#define LARGE_PAGE_SIZE_8M (1<<23)
powerpc/8xx: Map linear kernel RAM with 8M pages On a live running system (VoIP gateway for Air Trafic Control), over a 10 minutes period (with 277s idle), we get 87 millions DTLB misses and approximatly 35 secondes are spent in DTLB handler. This represents 5.8% of the overall time and even 10.8% of the non-idle time. Among those 87 millions DTLB misses, 15% are on user addresses and 85% are on kernel addresses. And within the kernel addresses, 93% are on addresses from the linear address space and only 7% are on addresses from the virtual address space. MPC8xx has no BATs but it has 8Mb page size. This patch implements mapping of kernel RAM using 8Mb pages, on the same model as what is done on the 40x. In 4k pages mode, each PGD entry maps a 4Mb area: we map every two entries to the same 8Mb physical page. In each second entry, we add 4Mb to the page physical address to ease life of the FixupDAR routine. This is just ignored by HW. In 16k pages mode, each PGD entry maps a 64Mb area: each PGD entry will point to the first page of the area. The DTLB handler adds the 3 bits from EPN to map the correct page. With this patch applied, we now get only 13 millions TLB misses during the 10 minutes period. The idle time has increased to 313s and the overall time spent in DTLB miss handler is 6.3s, which represents 1% of the overall time and 2.2% of non-idle time. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-02-10 00:07:50 +08:00
/*
* MMU_init_hw does the chip-specific initialization of the MMU hardware.
*/
void __init MMU_init_hw(void)
{
powerpc/8xx: Rework CONFIG_PIN_TLB handling On recent kernels, with some debug options like for instance CONFIG_LOCKDEP, the BSS requires more than 8M memory, allthough the kernel code fits in the first 8M. Today, it is necessary to activate CONFIG_PIN_TLB to get more than 8M at startup, allthough pinning TLB is not necessary for that. We could have inconditionaly mapped 16 or 24M bytes at startup but some old hardware only have 8M and mapping non-existing RAM would be an issue due to speculative accesses. With the preceding patch however, the TLB entries are populated on demand. By setting up the TLB miss handler to handle up to 24M until the handler is patched for the entire memory space, it is possible to allow access up to more memory without mapping non-existing RAM. It is therefore not needed anymore to map memory data at all at startup. It will be handled by the TLB miss handler. One might still want to PIN the IMMR and the first 24M of RAM. It is now possible to do it in the C memory initialisation functions. In addition, we now know how much memory we have when we do it, so we are able to adapt the pining to the real amount of memory available. So boards with less than 24M can now also benefit from PIN_TLB. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:54 +08:00
/* PIN up to the 3 first 8Mb after IMMR in DTLB table */
powerpc/8xx: Make pinning of ITLBs optional As stated in a comment in head_8xx.S, today we "Always pin the first 8 MB ITLB to prevent ITLB misses while mucking around with SRR0/SRR1 in asm". This issue has just been cleared by the preceding patch, therefore we can make this pinning optional (on by default) and independent of DATA pinning. This patch also makes pinning of IMMR independent of pinning of DATA. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2017-07-12 18:08:51 +08:00
#ifdef CONFIG_PIN_TLB_DATA
powerpc/8xx: Rework CONFIG_PIN_TLB handling On recent kernels, with some debug options like for instance CONFIG_LOCKDEP, the BSS requires more than 8M memory, allthough the kernel code fits in the first 8M. Today, it is necessary to activate CONFIG_PIN_TLB to get more than 8M at startup, allthough pinning TLB is not necessary for that. We could have inconditionaly mapped 16 or 24M bytes at startup but some old hardware only have 8M and mapping non-existing RAM would be an issue due to speculative accesses. With the preceding patch however, the TLB entries are populated on demand. By setting up the TLB miss handler to handle up to 24M until the handler is patched for the entire memory space, it is possible to allow access up to more memory without mapping non-existing RAM. It is therefore not needed anymore to map memory data at all at startup. It will be handled by the TLB miss handler. One might still want to PIN the IMMR and the first 24M of RAM. It is now possible to do it in the C memory initialisation functions. In addition, we now know how much memory we have when we do it, so we are able to adapt the pining to the real amount of memory available. So boards with less than 24M can now also benefit from PIN_TLB. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:54 +08:00
unsigned long ctr = mfspr(SPRN_MD_CTR) & 0xfe000000;
powerpc/mm: extend _PAGE_PRIVILEGED to all CPUs commit ac29c64089b74 ("powerpc/mm: Replace _PAGE_USER with _PAGE_PRIVILEGED") introduced _PAGE_PRIVILEGED for BOOK3S/64 This patch generalises _PAGE_PRIVILEGED for all CPUs, allowing to have either _PAGE_PRIVILEGED or _PAGE_USER or both. PPC_8xx has a _PAGE_SHARED flag which is set for and only for all non user pages. Lets rename it _PAGE_PRIVILEGED to remove confusion as it has nothing to do with Linux shared pages. On BookE, there's a _PAGE_BAP_SR which has to be set for kernel pages: defining _PAGE_PRIVILEGED as _PAGE_BAP_SR will make this generic Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-01-12 20:45:27 +08:00
unsigned long flags = 0xf0 | MD_SPS16K | _PAGE_PRIVILEGED | _PAGE_DIRTY;
powerpc/8xx: add CONFIG_PIN_TLB_IMMR CONFIG_PIN_TLB maps IMMR area and the first 24 Mbytes of memory. In some circunstances it might be more interesting to not map IMMR but map 32 Mbytes of memory instead. Therefore we add config option CONFIG_PIN_TLB_IMMR to select if IMMR shall be pinned or not, hence whether we pin 24 or 32 Mbytes of RAM Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:56 +08:00
#ifdef CONFIG_PIN_TLB_IMMR
int i = 29;
#else
int i = 28;
#endif
powerpc/8xx: Rework CONFIG_PIN_TLB handling On recent kernels, with some debug options like for instance CONFIG_LOCKDEP, the BSS requires more than 8M memory, allthough the kernel code fits in the first 8M. Today, it is necessary to activate CONFIG_PIN_TLB to get more than 8M at startup, allthough pinning TLB is not necessary for that. We could have inconditionaly mapped 16 or 24M bytes at startup but some old hardware only have 8M and mapping non-existing RAM would be an issue due to speculative accesses. With the preceding patch however, the TLB entries are populated on demand. By setting up the TLB miss handler to handle up to 24M until the handler is patched for the entire memory space, it is possible to allow access up to more memory without mapping non-existing RAM. It is therefore not needed anymore to map memory data at all at startup. It will be handled by the TLB miss handler. One might still want to PIN the IMMR and the first 24M of RAM. It is now possible to do it in the C memory initialisation functions. In addition, we now know how much memory we have when we do it, so we are able to adapt the pining to the real amount of memory available. So boards with less than 24M can now also benefit from PIN_TLB. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:54 +08:00
unsigned long addr = 0;
unsigned long mem = total_lowmem;
powerpc/8xx: add CONFIG_PIN_TLB_IMMR CONFIG_PIN_TLB maps IMMR area and the first 24 Mbytes of memory. In some circunstances it might be more interesting to not map IMMR but map 32 Mbytes of memory instead. Therefore we add config option CONFIG_PIN_TLB_IMMR to select if IMMR shall be pinned or not, hence whether we pin 24 or 32 Mbytes of RAM Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:56 +08:00
for (; i < 32 && mem >= LARGE_PAGE_SIZE_8M; i++) {
powerpc/8xx: Rework CONFIG_PIN_TLB handling On recent kernels, with some debug options like for instance CONFIG_LOCKDEP, the BSS requires more than 8M memory, allthough the kernel code fits in the first 8M. Today, it is necessary to activate CONFIG_PIN_TLB to get more than 8M at startup, allthough pinning TLB is not necessary for that. We could have inconditionaly mapped 16 or 24M bytes at startup but some old hardware only have 8M and mapping non-existing RAM would be an issue due to speculative accesses. With the preceding patch however, the TLB entries are populated on demand. By setting up the TLB miss handler to handle up to 24M until the handler is patched for the entire memory space, it is possible to allow access up to more memory without mapping non-existing RAM. It is therefore not needed anymore to map memory data at all at startup. It will be handled by the TLB miss handler. One might still want to PIN the IMMR and the first 24M of RAM. It is now possible to do it in the C memory initialisation functions. In addition, we now know how much memory we have when we do it, so we are able to adapt the pining to the real amount of memory available. So boards with less than 24M can now also benefit from PIN_TLB. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:54 +08:00
mtspr(SPRN_MD_CTR, ctr | (i << 8));
mtspr(SPRN_MD_EPN, (unsigned long)__va(addr) | MD_EVALID);
powerpc/8xx: Use L1 entry APG to handle _PAGE_ACCESSED for CONFIG_SWAP When CONFIG_SWAP is set, the TLB miss handlers have to also take into account _PAGE_ACCESSED flag. At the moment it is done by anding _PAGE_ACCESSED into _PAGE_PRESENT using 3 instructions. This patch uses APG for handling _PAGE_ACCESSED, allowing to just copy _PAGE_ACCESSED bit into APG field, hence reducing the action to a single instruction. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-01-12 20:45:33 +08:00
mtspr(SPRN_MD_TWC, MD_PS8MEG | MD_SVALID | M_APG2);
powerpc/8xx: Rework CONFIG_PIN_TLB handling On recent kernels, with some debug options like for instance CONFIG_LOCKDEP, the BSS requires more than 8M memory, allthough the kernel code fits in the first 8M. Today, it is necessary to activate CONFIG_PIN_TLB to get more than 8M at startup, allthough pinning TLB is not necessary for that. We could have inconditionaly mapped 16 or 24M bytes at startup but some old hardware only have 8M and mapping non-existing RAM would be an issue due to speculative accesses. With the preceding patch however, the TLB entries are populated on demand. By setting up the TLB miss handler to handle up to 24M until the handler is patched for the entire memory space, it is possible to allow access up to more memory without mapping non-existing RAM. It is therefore not needed anymore to map memory data at all at startup. It will be handled by the TLB miss handler. One might still want to PIN the IMMR and the first 24M of RAM. It is now possible to do it in the C memory initialisation functions. In addition, we now know how much memory we have when we do it, so we are able to adapt the pining to the real amount of memory available. So boards with less than 24M can now also benefit from PIN_TLB. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:54 +08:00
mtspr(SPRN_MD_RPN, addr | flags | _PAGE_PRESENT);
addr += LARGE_PAGE_SIZE_8M;
mem -= LARGE_PAGE_SIZE_8M;
}
#endif
powerpc/8xx: Map linear kernel RAM with 8M pages On a live running system (VoIP gateway for Air Trafic Control), over a 10 minutes period (with 277s idle), we get 87 millions DTLB misses and approximatly 35 secondes are spent in DTLB handler. This represents 5.8% of the overall time and even 10.8% of the non-idle time. Among those 87 millions DTLB misses, 15% are on user addresses and 85% are on kernel addresses. And within the kernel addresses, 93% are on addresses from the linear address space and only 7% are on addresses from the virtual address space. MPC8xx has no BATs but it has 8Mb page size. This patch implements mapping of kernel RAM using 8Mb pages, on the same model as what is done on the 40x. In 4k pages mode, each PGD entry maps a 4Mb area: we map every two entries to the same 8Mb physical page. In each second entry, we add 4Mb to the page physical address to ease life of the FixupDAR routine. This is just ignored by HW. In 16k pages mode, each PGD entry maps a 64Mb area: each PGD entry will point to the first page of the area. The DTLB handler adds the 3 bits from EPN to map the correct page. With this patch applied, we now get only 13 millions TLB misses during the 10 minutes period. The idle time has increased to 313s and the overall time spent in DTLB miss handler is 6.3s, which represents 1% of the overall time and 2.2% of non-idle time. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-02-10 00:07:50 +08:00
}
powerpc/8xx: mark init functions with __init setup_initial_memory_limit() is only called during init. mmu_patch_cmp_limit() is only called from 8xx_mmu.c Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2017-07-12 18:08:55 +08:00
static void __init mmu_mapin_immr(void)
powerpc/8xx: Map IMMR area with 512k page at a fixed address Once the linear memory space has been mapped with 8Mb pages, as seen in the related commit, we get 11 millions DTLB missed during the reference 600s period. 77% of the misses are on user addresses and 23% are on kernel addresses (1 fourth for linear address space and 3 fourth for virtual address space) Traditionaly, each driver manages one computer board which has its own components with its own memory maps. But on embedded chips like the MPC8xx, the SOC has all registers located in the same IO area. When looking at ioremaps done during startup, we see that many drivers are re-mapping small parts of the IMMR for their own use and all those small pieces gets their own 4k page, amplifying the number of TLB misses: in our system we get 0xff000000 mapped 31 times and 0xff003000 mapped 9 times. Even if each part of IMMR was mapped only once with 4k pages, it would still be several small mappings towards linear area. This patch maps the IMMR with a single 512k page. With this patch applied, the number of DTLB misses during the 10 min period is reduced to 11.8 millions for a duration of 5.8s, which represents 2% of the non-idle time hence yet another 10% reduction. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:45 +08:00
{
unsigned long p = PHYS_IMMR_BASE;
unsigned long v = VIRT_IMMR_BASE;
unsigned long f = pgprot_val(PAGE_KERNEL_NCG);
int offset;
for (offset = 0; offset < IMMR_SIZE; offset += PAGE_SIZE)
powerpc/mm: Rename map_page() to map_kernel_page() on 32-bit These two functions implement the same semantics, so unify their naming so we can share code that calls them. The longer name is more descriptive so use it. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Acked-by: Balbir Singh <bsingharora@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2017-05-29 23:31:56 +08:00
map_kernel_page(v + offset, p + offset, f);
powerpc/8xx: Map IMMR area with 512k page at a fixed address Once the linear memory space has been mapped with 8Mb pages, as seen in the related commit, we get 11 millions DTLB missed during the reference 600s period. 77% of the misses are on user addresses and 23% are on kernel addresses (1 fourth for linear address space and 3 fourth for virtual address space) Traditionaly, each driver manages one computer board which has its own components with its own memory maps. But on embedded chips like the MPC8xx, the SOC has all registers located in the same IO area. When looking at ioremaps done during startup, we see that many drivers are re-mapping small parts of the IMMR for their own use and all those small pieces gets their own 4k page, amplifying the number of TLB misses: in our system we get 0xff000000 mapped 31 times and 0xff003000 mapped 9 times. Even if each part of IMMR was mapped only once with 4k pages, it would still be several small mappings towards linear area. This patch maps the IMMR with a single 512k page. With this patch applied, the number of DTLB misses during the 10 min period is reduced to 11.8 millions for a duration of 5.8s, which represents 2% of the non-idle time hence yet another 10% reduction. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:45 +08:00
}
/* Address of instructions to patch */
powerpc/8xx: add CONFIG_PIN_TLB_IMMR CONFIG_PIN_TLB maps IMMR area and the first 24 Mbytes of memory. In some circunstances it might be more interesting to not map IMMR but map 32 Mbytes of memory instead. Therefore we add config option CONFIG_PIN_TLB_IMMR to select if IMMR shall be pinned or not, hence whether we pin 24 or 32 Mbytes of RAM Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:56 +08:00
#ifndef CONFIG_PIN_TLB_IMMR
powerpc/8xx: Map IMMR area with 512k page at a fixed address Once the linear memory space has been mapped with 8Mb pages, as seen in the related commit, we get 11 millions DTLB missed during the reference 600s period. 77% of the misses are on user addresses and 23% are on kernel addresses (1 fourth for linear address space and 3 fourth for virtual address space) Traditionaly, each driver manages one computer board which has its own components with its own memory maps. But on embedded chips like the MPC8xx, the SOC has all registers located in the same IO area. When looking at ioremaps done during startup, we see that many drivers are re-mapping small parts of the IMMR for their own use and all those small pieces gets their own 4k page, amplifying the number of TLB misses: in our system we get 0xff000000 mapped 31 times and 0xff003000 mapped 9 times. Even if each part of IMMR was mapped only once with 4k pages, it would still be several small mappings towards linear area. This patch maps the IMMR with a single 512k page. With this patch applied, the number of DTLB misses during the 10 min period is reduced to 11.8 millions for a duration of 5.8s, which represents 2% of the non-idle time hence yet another 10% reduction. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:45 +08:00
extern unsigned int DTLBMiss_jmp;
#endif
powerpc/8xx: Don't use page table for linear memory space Instead of using the first level page table to define mappings for the linear memory space, we can use direct mapping from the TLB handling routines. This has several advantages: * No need to read the tables at each TLB miss * No issue in 16k pages mode where the 1st level table maps 64 Mbytes The size of the available linear space is known at system startup. In order to avoid data access at each TLB miss to know the memory size, the TLB routine is patched at startup with the proper size This patch provides a 10%-15% improvment of TLB miss handling for kernel addresses Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:51 +08:00
extern unsigned int DTLBMiss_cmp, FixupDAR_cmp;
powerpc/8xx: Make pinning of ITLBs optional As stated in a comment in head_8xx.S, today we "Always pin the first 8 MB ITLB to prevent ITLB misses while mucking around with SRR0/SRR1 in asm". This issue has just been cleared by the preceding patch, therefore we can make this pinning optional (on by default) and independent of DATA pinning. This patch also makes pinning of IMMR independent of pinning of DATA. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2017-07-12 18:08:51 +08:00
#ifndef CONFIG_PIN_TLB_TEXT
extern unsigned int ITLBMiss_cmp;
#endif
powerpc/8xx: Map IMMR area with 512k page at a fixed address Once the linear memory space has been mapped with 8Mb pages, as seen in the related commit, we get 11 millions DTLB missed during the reference 600s period. 77% of the misses are on user addresses and 23% are on kernel addresses (1 fourth for linear address space and 3 fourth for virtual address space) Traditionaly, each driver manages one computer board which has its own components with its own memory maps. But on embedded chips like the MPC8xx, the SOC has all registers located in the same IO area. When looking at ioremaps done during startup, we see that many drivers are re-mapping small parts of the IMMR for their own use and all those small pieces gets their own 4k page, amplifying the number of TLB misses: in our system we get 0xff000000 mapped 31 times and 0xff003000 mapped 9 times. Even if each part of IMMR was mapped only once with 4k pages, it would still be several small mappings towards linear area. This patch maps the IMMR with a single 512k page. With this patch applied, the number of DTLB misses during the 10 min period is reduced to 11.8 millions for a duration of 5.8s, which represents 2% of the non-idle time hence yet another 10% reduction. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:45 +08:00
powerpc/8xx: mark init functions with __init setup_initial_memory_limit() is only called during init. mmu_patch_cmp_limit() is only called from 8xx_mmu.c Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2017-07-12 18:08:55 +08:00
static void __init mmu_patch_cmp_limit(unsigned int *addr, unsigned long mapped)
powerpc/8xx: Map linear kernel RAM with 8M pages On a live running system (VoIP gateway for Air Trafic Control), over a 10 minutes period (with 277s idle), we get 87 millions DTLB misses and approximatly 35 secondes are spent in DTLB handler. This represents 5.8% of the overall time and even 10.8% of the non-idle time. Among those 87 millions DTLB misses, 15% are on user addresses and 85% are on kernel addresses. And within the kernel addresses, 93% are on addresses from the linear address space and only 7% are on addresses from the virtual address space. MPC8xx has no BATs but it has 8Mb page size. This patch implements mapping of kernel RAM using 8Mb pages, on the same model as what is done on the 40x. In 4k pages mode, each PGD entry maps a 4Mb area: we map every two entries to the same 8Mb physical page. In each second entry, we add 4Mb to the page physical address to ease life of the FixupDAR routine. This is just ignored by HW. In 16k pages mode, each PGD entry maps a 64Mb area: each PGD entry will point to the first page of the area. The DTLB handler adds the 3 bits from EPN to map the correct page. With this patch applied, we now get only 13 millions TLB misses during the 10 minutes period. The idle time has increased to 313s and the overall time spent in DTLB miss handler is 6.3s, which represents 1% of the overall time and 2.2% of non-idle time. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-02-10 00:07:50 +08:00
{
powerpc/8xx: Don't use page table for linear memory space Instead of using the first level page table to define mappings for the linear memory space, we can use direct mapping from the TLB handling routines. This has several advantages: * No need to read the tables at each TLB miss * No issue in 16k pages mode where the 1st level table maps 64 Mbytes The size of the available linear space is known at system startup. In order to avoid data access at each TLB miss to know the memory size, the TLB routine is patched at startup with the proper size This patch provides a 10%-15% improvment of TLB miss handling for kernel addresses Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:51 +08:00
unsigned int instr = *addr;
powerpc/8xx: Map linear kernel RAM with 8M pages On a live running system (VoIP gateway for Air Trafic Control), over a 10 minutes period (with 277s idle), we get 87 millions DTLB misses and approximatly 35 secondes are spent in DTLB handler. This represents 5.8% of the overall time and even 10.8% of the non-idle time. Among those 87 millions DTLB misses, 15% are on user addresses and 85% are on kernel addresses. And within the kernel addresses, 93% are on addresses from the linear address space and only 7% are on addresses from the virtual address space. MPC8xx has no BATs but it has 8Mb page size. This patch implements mapping of kernel RAM using 8Mb pages, on the same model as what is done on the 40x. In 4k pages mode, each PGD entry maps a 4Mb area: we map every two entries to the same 8Mb physical page. In each second entry, we add 4Mb to the page physical address to ease life of the FixupDAR routine. This is just ignored by HW. In 16k pages mode, each PGD entry maps a 64Mb area: each PGD entry will point to the first page of the area. The DTLB handler adds the 3 bits from EPN to map the correct page. With this patch applied, we now get only 13 millions TLB misses during the 10 minutes period. The idle time has increased to 313s and the overall time spent in DTLB miss handler is 6.3s, which represents 1% of the overall time and 2.2% of non-idle time. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-02-10 00:07:50 +08:00
powerpc/8xx: Don't use page table for linear memory space Instead of using the first level page table to define mappings for the linear memory space, we can use direct mapping from the TLB handling routines. This has several advantages: * No need to read the tables at each TLB miss * No issue in 16k pages mode where the 1st level table maps 64 Mbytes The size of the available linear space is known at system startup. In order to avoid data access at each TLB miss to know the memory size, the TLB routine is patched at startup with the proper size This patch provides a 10%-15% improvment of TLB miss handling for kernel addresses Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:51 +08:00
instr &= 0xffff0000;
instr |= (unsigned long)__va(mapped) >> 16;
patch_instruction(addr, instr);
}
unsigned long __init mmu_mapin_ram(unsigned long top)
{
unsigned long mapped;
powerpc/8xx: Map linear kernel RAM with 8M pages On a live running system (VoIP gateway for Air Trafic Control), over a 10 minutes period (with 277s idle), we get 87 millions DTLB misses and approximatly 35 secondes are spent in DTLB handler. This represents 5.8% of the overall time and even 10.8% of the non-idle time. Among those 87 millions DTLB misses, 15% are on user addresses and 85% are on kernel addresses. And within the kernel addresses, 93% are on addresses from the linear address space and only 7% are on addresses from the virtual address space. MPC8xx has no BATs but it has 8Mb page size. This patch implements mapping of kernel RAM using 8Mb pages, on the same model as what is done on the 40x. In 4k pages mode, each PGD entry maps a 4Mb area: we map every two entries to the same 8Mb physical page. In each second entry, we add 4Mb to the page physical address to ease life of the FixupDAR routine. This is just ignored by HW. In 16k pages mode, each PGD entry maps a 64Mb area: each PGD entry will point to the first page of the area. The DTLB handler adds the 3 bits from EPN to map the correct page. With this patch applied, we now get only 13 millions TLB misses during the 10 minutes period. The idle time has increased to 313s and the overall time spent in DTLB miss handler is 6.3s, which represents 1% of the overall time and 2.2% of non-idle time. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-02-10 00:07:50 +08:00
powerpc/8xx: Map IMMR area with 512k page at a fixed address Once the linear memory space has been mapped with 8Mb pages, as seen in the related commit, we get 11 millions DTLB missed during the reference 600s period. 77% of the misses are on user addresses and 23% are on kernel addresses (1 fourth for linear address space and 3 fourth for virtual address space) Traditionaly, each driver manages one computer board which has its own components with its own memory maps. But on embedded chips like the MPC8xx, the SOC has all registers located in the same IO area. When looking at ioremaps done during startup, we see that many drivers are re-mapping small parts of the IMMR for their own use and all those small pieces gets their own 4k page, amplifying the number of TLB misses: in our system we get 0xff000000 mapped 31 times and 0xff003000 mapped 9 times. Even if each part of IMMR was mapped only once with 4k pages, it would still be several small mappings towards linear area. This patch maps the IMMR with a single 512k page. With this patch applied, the number of DTLB misses during the 10 min period is reduced to 11.8 millions for a duration of 5.8s, which represents 2% of the non-idle time hence yet another 10% reduction. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:45 +08:00
if (__map_without_ltlbs) {
powerpc/8xx: Don't use page table for linear memory space Instead of using the first level page table to define mappings for the linear memory space, we can use direct mapping from the TLB handling routines. This has several advantages: * No need to read the tables at each TLB miss * No issue in 16k pages mode where the 1st level table maps 64 Mbytes The size of the available linear space is known at system startup. In order to avoid data access at each TLB miss to know the memory size, the TLB routine is patched at startup with the proper size This patch provides a 10%-15% improvment of TLB miss handling for kernel addresses Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:51 +08:00
mapped = 0;
powerpc/8xx: Map IMMR area with 512k page at a fixed address Once the linear memory space has been mapped with 8Mb pages, as seen in the related commit, we get 11 millions DTLB missed during the reference 600s period. 77% of the misses are on user addresses and 23% are on kernel addresses (1 fourth for linear address space and 3 fourth for virtual address space) Traditionaly, each driver manages one computer board which has its own components with its own memory maps. But on embedded chips like the MPC8xx, the SOC has all registers located in the same IO area. When looking at ioremaps done during startup, we see that many drivers are re-mapping small parts of the IMMR for their own use and all those small pieces gets their own 4k page, amplifying the number of TLB misses: in our system we get 0xff000000 mapped 31 times and 0xff003000 mapped 9 times. Even if each part of IMMR was mapped only once with 4k pages, it would still be several small mappings towards linear area. This patch maps the IMMR with a single 512k page. With this patch applied, the number of DTLB misses during the 10 min period is reduced to 11.8 millions for a duration of 5.8s, which represents 2% of the non-idle time hence yet another 10% reduction. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:45 +08:00
mmu_mapin_immr();
powerpc/8xx: add CONFIG_PIN_TLB_IMMR CONFIG_PIN_TLB maps IMMR area and the first 24 Mbytes of memory. In some circunstances it might be more interesting to not map IMMR but map 32 Mbytes of memory instead. Therefore we add config option CONFIG_PIN_TLB_IMMR to select if IMMR shall be pinned or not, hence whether we pin 24 or 32 Mbytes of RAM Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:56 +08:00
#ifndef CONFIG_PIN_TLB_IMMR
powerpc/8xx: Map IMMR area with 512k page at a fixed address Once the linear memory space has been mapped with 8Mb pages, as seen in the related commit, we get 11 millions DTLB missed during the reference 600s period. 77% of the misses are on user addresses and 23% are on kernel addresses (1 fourth for linear address space and 3 fourth for virtual address space) Traditionaly, each driver manages one computer board which has its own components with its own memory maps. But on embedded chips like the MPC8xx, the SOC has all registers located in the same IO area. When looking at ioremaps done during startup, we see that many drivers are re-mapping small parts of the IMMR for their own use and all those small pieces gets their own 4k page, amplifying the number of TLB misses: in our system we get 0xff000000 mapped 31 times and 0xff003000 mapped 9 times. Even if each part of IMMR was mapped only once with 4k pages, it would still be several small mappings towards linear area. This patch maps the IMMR with a single 512k page. With this patch applied, the number of DTLB misses during the 10 min period is reduced to 11.8 millions for a duration of 5.8s, which represents 2% of the non-idle time hence yet another 10% reduction. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:45 +08:00
patch_instruction(&DTLBMiss_jmp, PPC_INST_NOP);
powerpc/8xx: Make pinning of ITLBs optional As stated in a comment in head_8xx.S, today we "Always pin the first 8 MB ITLB to prevent ITLB misses while mucking around with SRR0/SRR1 in asm". This issue has just been cleared by the preceding patch, therefore we can make this pinning optional (on by default) and independent of DATA pinning. This patch also makes pinning of IMMR independent of pinning of DATA. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2017-07-12 18:08:51 +08:00
#endif
#ifndef CONFIG_PIN_TLB_TEXT
mmu_patch_cmp_limit(&ITLBMiss_cmp, 0);
powerpc/8xx: Map IMMR area with 512k page at a fixed address Once the linear memory space has been mapped with 8Mb pages, as seen in the related commit, we get 11 millions DTLB missed during the reference 600s period. 77% of the misses are on user addresses and 23% are on kernel addresses (1 fourth for linear address space and 3 fourth for virtual address space) Traditionaly, each driver manages one computer board which has its own components with its own memory maps. But on embedded chips like the MPC8xx, the SOC has all registers located in the same IO area. When looking at ioremaps done during startup, we see that many drivers are re-mapping small parts of the IMMR for their own use and all those small pieces gets their own 4k page, amplifying the number of TLB misses: in our system we get 0xff000000 mapped 31 times and 0xff003000 mapped 9 times. Even if each part of IMMR was mapped only once with 4k pages, it would still be several small mappings towards linear area. This patch maps the IMMR with a single 512k page. With this patch applied, the number of DTLB misses during the 10 min period is reduced to 11.8 millions for a duration of 5.8s, which represents 2% of the non-idle time hence yet another 10% reduction. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:45 +08:00
#endif
powerpc/8xx: Don't use page table for linear memory space Instead of using the first level page table to define mappings for the linear memory space, we can use direct mapping from the TLB handling routines. This has several advantages: * No need to read the tables at each TLB miss * No issue in 16k pages mode where the 1st level table maps 64 Mbytes The size of the available linear space is known at system startup. In order to avoid data access at each TLB miss to know the memory size, the TLB routine is patched at startup with the proper size This patch provides a 10%-15% improvment of TLB miss handling for kernel addresses Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:51 +08:00
} else {
mapped = top & ~(LARGE_PAGE_SIZE_8M - 1);
powerpc/8xx: Map linear kernel RAM with 8M pages On a live running system (VoIP gateway for Air Trafic Control), over a 10 minutes period (with 277s idle), we get 87 millions DTLB misses and approximatly 35 secondes are spent in DTLB handler. This represents 5.8% of the overall time and even 10.8% of the non-idle time. Among those 87 millions DTLB misses, 15% are on user addresses and 85% are on kernel addresses. And within the kernel addresses, 93% are on addresses from the linear address space and only 7% are on addresses from the virtual address space. MPC8xx has no BATs but it has 8Mb page size. This patch implements mapping of kernel RAM using 8Mb pages, on the same model as what is done on the 40x. In 4k pages mode, each PGD entry maps a 4Mb area: we map every two entries to the same 8Mb physical page. In each second entry, we add 4Mb to the page physical address to ease life of the FixupDAR routine. This is just ignored by HW. In 16k pages mode, each PGD entry maps a 64Mb area: each PGD entry will point to the first page of the area. The DTLB handler adds the 3 bits from EPN to map the correct page. With this patch applied, we now get only 13 millions TLB misses during the 10 minutes period. The idle time has increased to 313s and the overall time spent in DTLB miss handler is 6.3s, which represents 1% of the overall time and 2.2% of non-idle time. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-02-10 00:07:50 +08:00
}
powerpc/8xx: Don't use page table for linear memory space Instead of using the first level page table to define mappings for the linear memory space, we can use direct mapping from the TLB handling routines. This has several advantages: * No need to read the tables at each TLB miss * No issue in 16k pages mode where the 1st level table maps 64 Mbytes The size of the available linear space is known at system startup. In order to avoid data access at each TLB miss to know the memory size, the TLB routine is patched at startup with the proper size This patch provides a 10%-15% improvment of TLB miss handling for kernel addresses Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:51 +08:00
mmu_patch_cmp_limit(&DTLBMiss_cmp, mapped);
mmu_patch_cmp_limit(&FixupDAR_cmp, mapped);
powerpc/8xx: Map linear kernel RAM with 8M pages On a live running system (VoIP gateway for Air Trafic Control), over a 10 minutes period (with 277s idle), we get 87 millions DTLB misses and approximatly 35 secondes are spent in DTLB handler. This represents 5.8% of the overall time and even 10.8% of the non-idle time. Among those 87 millions DTLB misses, 15% are on user addresses and 85% are on kernel addresses. And within the kernel addresses, 93% are on addresses from the linear address space and only 7% are on addresses from the virtual address space. MPC8xx has no BATs but it has 8Mb page size. This patch implements mapping of kernel RAM using 8Mb pages, on the same model as what is done on the 40x. In 4k pages mode, each PGD entry maps a 4Mb area: we map every two entries to the same 8Mb physical page. In each second entry, we add 4Mb to the page physical address to ease life of the FixupDAR routine. This is just ignored by HW. In 16k pages mode, each PGD entry maps a 64Mb area: each PGD entry will point to the first page of the area. The DTLB handler adds the 3 bits from EPN to map the correct page. With this patch applied, we now get only 13 millions TLB misses during the 10 minutes period. The idle time has increased to 313s and the overall time spent in DTLB miss handler is 6.3s, which represents 1% of the overall time and 2.2% of non-idle time. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-02-10 00:07:50 +08:00
/* If the size of RAM is not an exact power of two, we may not
* have covered RAM in its entirety with 8 MiB
* pages. Consequently, restrict the top end of RAM currently
* allocable so that calls to the MEMBLOCK to allocate PTEs for "tail"
* coverage with normal-sized pages (or other reasons) do not
* attempt to allocate outside the allowed range.
*/
powerpc/8xx: Don't use page table for linear memory space Instead of using the first level page table to define mappings for the linear memory space, we can use direct mapping from the TLB handling routines. This has several advantages: * No need to read the tables at each TLB miss * No issue in 16k pages mode where the 1st level table maps 64 Mbytes The size of the available linear space is known at system startup. In order to avoid data access at each TLB miss to know the memory size, the TLB routine is patched at startup with the proper size This patch provides a 10%-15% improvment of TLB miss handling for kernel addresses Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-05-17 15:02:51 +08:00
if (mapped)
memblock_set_current_limit(mapped);
powerpc/8xx: Map linear kernel RAM with 8M pages On a live running system (VoIP gateway for Air Trafic Control), over a 10 minutes period (with 277s idle), we get 87 millions DTLB misses and approximatly 35 secondes are spent in DTLB handler. This represents 5.8% of the overall time and even 10.8% of the non-idle time. Among those 87 millions DTLB misses, 15% are on user addresses and 85% are on kernel addresses. And within the kernel addresses, 93% are on addresses from the linear address space and only 7% are on addresses from the virtual address space. MPC8xx has no BATs but it has 8Mb page size. This patch implements mapping of kernel RAM using 8Mb pages, on the same model as what is done on the 40x. In 4k pages mode, each PGD entry maps a 4Mb area: we map every two entries to the same 8Mb physical page. In each second entry, we add 4Mb to the page physical address to ease life of the FixupDAR routine. This is just ignored by HW. In 16k pages mode, each PGD entry maps a 64Mb area: each PGD entry will point to the first page of the area. The DTLB handler adds the 3 bits from EPN to map the correct page. With this patch applied, we now get only 13 millions TLB misses during the 10 minutes period. The idle time has increased to 313s and the overall time spent in DTLB miss handler is 6.3s, which represents 1% of the overall time and 2.2% of non-idle time. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-02-10 00:07:50 +08:00
powerpc/8xx: Ensures RAM mapped with LTLB is seen as block mapped on 8xx. On the 8xx, the RAM mapped with LTLBs must be seen as block mapped, just like areas mapped with BATs on standard PPC32. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2017-07-12 18:08:45 +08:00
block_mapped_ram = mapped;
powerpc/8xx: Map linear kernel RAM with 8M pages On a live running system (VoIP gateway for Air Trafic Control), over a 10 minutes period (with 277s idle), we get 87 millions DTLB misses and approximatly 35 secondes are spent in DTLB handler. This represents 5.8% of the overall time and even 10.8% of the non-idle time. Among those 87 millions DTLB misses, 15% are on user addresses and 85% are on kernel addresses. And within the kernel addresses, 93% are on addresses from the linear address space and only 7% are on addresses from the virtual address space. MPC8xx has no BATs but it has 8Mb page size. This patch implements mapping of kernel RAM using 8Mb pages, on the same model as what is done on the 40x. In 4k pages mode, each PGD entry maps a 4Mb area: we map every two entries to the same 8Mb physical page. In each second entry, we add 4Mb to the page physical address to ease life of the FixupDAR routine. This is just ignored by HW. In 16k pages mode, each PGD entry maps a 64Mb area: each PGD entry will point to the first page of the area. The DTLB handler adds the 3 bits from EPN to map the correct page. With this patch applied, we now get only 13 millions TLB misses during the 10 minutes period. The idle time has increased to 313s and the overall time spent in DTLB miss handler is 6.3s, which represents 1% of the overall time and 2.2% of non-idle time. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-02-10 00:07:50 +08:00
return mapped;
}
powerpc/8xx: move setup_initial_memory_limit() into 8xx_mmu.c Now we have a 8xx specific .c file for that so put it in there as other powerpc variants do Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-02-10 00:07:54 +08:00
powerpc/8xx: mark init functions with __init setup_initial_memory_limit() is only called during init. mmu_patch_cmp_limit() is only called from 8xx_mmu.c Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2017-07-12 18:08:55 +08:00
void __init setup_initial_memory_limit(phys_addr_t first_memblock_base,
phys_addr_t first_memblock_size)
powerpc/8xx: move setup_initial_memory_limit() into 8xx_mmu.c Now we have a 8xx specific .c file for that so put it in there as other powerpc variants do Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-02-10 00:07:54 +08:00
{
/* We don't currently support the first MEMBLOCK not mapping 0
* physical on those processors
*/
BUG_ON(first_memblock_base != 0);
/* 8xx can only access 24MB at the moment */
memblock_set_current_limit(min_t(u64, first_memblock_size, 0x01800000));
}
powerpc/8xx: rewrite set_context() in C There is no real need to have set_context() in assembly. Now that we have mtspr() handling CPU6 ERRATA directly, we can rewrite set_context() in C language for easier maintenance. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-02-10 00:08:18 +08:00
/*
* Set up to use a given MMU context.
* id is context number, pgd is PGD pointer.
*
* We place the physical address of the new task page directory loaded
* into the MMU base register, and set the ASID compare register with
* the new "context."
*/
void set_context(unsigned long id, pgd_t *pgd)
{
s16 offset = (s16)(__pa(swapper_pg_dir));
#ifdef CONFIG_BDI_SWITCH
pgd_t **ptr = *(pgd_t ***)(KERNELBASE + 0xf0);
/* Context switch the PTE pointer for the Abatron BDI2000.
* The PGDIR is passed as second argument.
*/
*(ptr + 1) = pgd;
#endif
/* Register M_TW will contain base address of level 1 table minus the
* lower part of the kernel PGDIR base address, so that all accesses to
* level 1 table are done relative to lower part of kernel PGDIR base
* address.
*/
mtspr(SPRN_M_TW, __pa(pgd) - offset);
/* Update context */
powerpc/mm/slice: Fix hugepage allocation at hint address on 8xx On the 8xx, the page size is set in the PMD entry and applies to all pages of the page table pointed by the said PMD entry. When an app has some regular pages allocated (e.g. see below) and tries to mmap() a huge page at a hint address covered by the same PMD entry, the kernel accepts the hint allthough the 8xx cannot handle different page sizes in the same PMD entry. 10000000-10001000 r-xp 00000000 00:0f 2597 /root/malloc 10010000-10011000 rwxp 00000000 00:0f 2597 /root/malloc mmap(0x10080000, 524288, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS|0x40000, -1, 0) = 0x10080000 This results the app remaining forever in do_page_fault()/hugetlb_fault() and when interrupting that app, we get the following warning: [162980.035629] WARNING: CPU: 0 PID: 2777 at arch/powerpc/mm/hugetlbpage.c:354 hugetlb_free_pgd_range+0xc8/0x1e4 [162980.035699] CPU: 0 PID: 2777 Comm: malloc Tainted: G W 4.14.6 #85 [162980.035744] task: c67e2c00 task.stack: c668e000 [162980.035783] NIP: c000fe18 LR: c00e1eec CTR: c00f90c0 [162980.035830] REGS: c668fc20 TRAP: 0700 Tainted: G W (4.14.6) [162980.035854] MSR: 00029032 <EE,ME,IR,DR,RI> CR: 24044224 XER: 20000000 [162980.036003] [162980.036003] GPR00: c00e1eec c668fcd0 c67e2c00 00000010 c6869410 10080000 00000000 77fb4000 [162980.036003] GPR08: ffff0001 0683c001 00000000 ffffff80 44028228 10018a34 00004008 418004fc [162980.036003] GPR16: c668e000 00040100 c668e000 c06c0000 c668fe78 c668e000 c6835ba0 c668fd48 [162980.036003] GPR24: 00000000 73ffffff 74000000 00000001 77fb4000 100fffff 10100000 10100000 [162980.036743] NIP [c000fe18] hugetlb_free_pgd_range+0xc8/0x1e4 [162980.036839] LR [c00e1eec] free_pgtables+0x12c/0x150 [162980.036861] Call Trace: [162980.036939] [c668fcd0] [c00f0774] unlink_anon_vmas+0x1c4/0x214 (unreliable) [162980.037040] [c668fd10] [c00e1eec] free_pgtables+0x12c/0x150 [162980.037118] [c668fd40] [c00eabac] exit_mmap+0xe8/0x1b4 [162980.037210] [c668fda0] [c0019710] mmput.part.9+0x20/0xd8 [162980.037301] [c668fdb0] [c001ecb0] do_exit+0x1f0/0x93c [162980.037386] [c668fe00] [c001f478] do_group_exit+0x40/0xcc [162980.037479] [c668fe10] [c002a76c] get_signal+0x47c/0x614 [162980.037570] [c668fe70] [c0007840] do_signal+0x54/0x244 [162980.037654] [c668ff30] [c0007ae8] do_notify_resume+0x34/0x88 [162980.037744] [c668ff40] [c000dae8] do_user_signal+0x74/0xc4 [162980.037781] Instruction dump: [162980.037821] 7fdff378 81370000 54a3463a 80890020 7d24182e 7c841a14 712a0004 4082ff94 [162980.038014] 2f890000 419e0010 712a0ff0 408200e0 <0fe00000> 54a9000a 7f984840 419d0094 [162980.038216] ---[ end trace c0ceeca8e7a5800a ]--- [162980.038754] BUG: non-zero nr_ptes on freeing mm: 1 [162985.363322] BUG: non-zero nr_ptes on freeing mm: -1 In order to fix this, this patch uses the address space "slices" implemented for BOOK3S/64 and enhanced to support PPC32 by the preceding patch. This patch modifies the context.id on the 8xx to be in the range [1:16] instead of [0:15] in order to identify context.id == 0 as not initialised contexts as done on BOOK3S This patch activates CONFIG_PPC_MM_SLICES when CONFIG_HUGETLB_PAGE is selected for the 8xx Alltough we could in theory have as many slices as PMD entries, the current slices implementation limits the number of low slices to 16. This limitation is not preventing us to fix the initial issue allthough it is suboptimal. It will be cured in a subsequent patch. Fixes: 4b91428699477 ("powerpc/8xx: Implement support of hugepages") Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Reviewed-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-02-22 22:27:26 +08:00
mtspr(SPRN_M_CASID, id - 1);
powerpc/8xx: rewrite set_context() in C There is no real need to have set_context() in assembly. Now that we have mtspr() handling CPU6 ERRATA directly, we can rewrite set_context() in C language for easier maintenance. Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-02-10 00:08:18 +08:00
/* sync */
mb();
}
powerpc/8xx: rewrite flush_instruction_cache() in C On PPC8xx, flushing instruction cache is performed by writing in register SPRN_IC_CST. This registers suffers CPU6 ERRATA. The patch rewrites the fonction in C so that CPU6 ERRATA will be handled transparently Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr> Signed-off-by: Scott Wood <oss@buserror.net>
2016-02-10 00:08:21 +08:00
void flush_instruction_cache(void)
{
isync();
mtspr(SPRN_IC_CST, IDC_INVALL);
isync();
}