2012-04-09 20:50:52 +04:00
|
|
|
/*
|
|
|
|
* Common CPU TLB handling
|
|
|
|
*
|
|
|
|
* Copyright (c) 2003 Fabrice Bellard
|
|
|
|
*
|
|
|
|
* This library is free software; you can redistribute it and/or
|
|
|
|
* modify it under the terms of the GNU Lesser General Public
|
|
|
|
* License as published by the Free Software Foundation; either
|
|
|
|
* version 2 of the License, or (at your option) any later version.
|
|
|
|
*
|
|
|
|
* This library is distributed in the hope that it will be useful,
|
|
|
|
* but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
|
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
|
|
* Lesser General Public License for more details.
|
|
|
|
*
|
|
|
|
* You should have received a copy of the GNU Lesser General Public
|
|
|
|
* License along with this library; if not, see <http://www.gnu.org/licenses/>.
|
|
|
|
*/
|
|
|
|
|
|
|
|
#include "config.h"
|
|
|
|
#include "cpu.h"
|
2012-12-17 21:19:49 +04:00
|
|
|
#include "exec/exec-all.h"
|
|
|
|
#include "exec/memory.h"
|
|
|
|
#include "exec/address-spaces.h"
|
2014-03-28 22:42:10 +04:00
|
|
|
#include "exec/cpu_ldst.h"
|
2012-04-09 20:50:52 +04:00
|
|
|
|
2012-12-17 21:19:49 +04:00
|
|
|
#include "exec/cputlb.h"
|
2012-04-09 20:50:52 +04:00
|
|
|
|
2012-12-17 21:19:49 +04:00
|
|
|
#include "exec/memory-internal.h"
|
2013-10-14 19:13:59 +04:00
|
|
|
#include "exec/ram_addr.h"
|
2014-03-28 20:55:24 +04:00
|
|
|
#include "tcg/tcg.h"
|
2012-04-09 20:50:52 +04:00
|
|
|
|
|
|
|
//#define DEBUG_TLB
|
|
|
|
//#define DEBUG_TLB_CHECK
|
|
|
|
|
|
|
|
/* statistics */
|
|
|
|
int tlb_flush_count;
|
|
|
|
|
|
|
|
/* NOTE:
|
|
|
|
* If flush_global is true (the usual case), flush all tlb entries.
|
|
|
|
* If flush_global is false, flush (at least) all tlb entries not
|
|
|
|
* marked global.
|
|
|
|
*
|
|
|
|
* Since QEMU doesn't currently implement a global/not-global flag
|
|
|
|
* for tlb entries, at the moment tlb_flush() will also flush all
|
|
|
|
* tlb entries in the flush_global == false case. This is OK because
|
|
|
|
* CPU architectures generally permit an implementation to drop
|
|
|
|
* entries from the TLB at any time, so flushing more entries than
|
|
|
|
* required is only an efficiency issue, not a correctness issue.
|
|
|
|
*/
|
2013-09-04 04:19:44 +04:00
|
|
|
void tlb_flush(CPUState *cpu, int flush_global)
|
2012-04-09 20:50:52 +04:00
|
|
|
{
|
2013-09-04 04:19:44 +04:00
|
|
|
CPUArchState *env = cpu->env_ptr;
|
2012-04-09 20:50:52 +04:00
|
|
|
|
|
|
|
#if defined(DEBUG_TLB)
|
|
|
|
printf("tlb_flush:\n");
|
|
|
|
#endif
|
|
|
|
/* must reset current TB so that interrupts cannot modify the
|
|
|
|
links while we are modifying them */
|
2013-01-16 22:29:31 +04:00
|
|
|
cpu->current_tb = NULL;
|
2012-04-09 20:50:52 +04:00
|
|
|
|
2013-12-07 01:44:51 +04:00
|
|
|
memset(env->tlb_table, -1, sizeof(env->tlb_table));
|
implementing victim TLB for QEMU system emulated TLB
QEMU system mode page table walks are expensive. Taken by running QEMU
qemu-system-x86_64 system mode on Intel PIN , a TLB miss and walking a
4-level page tables in guest Linux OS takes ~450 X86 instructions on
average.
QEMU system mode TLB is implemented using a directly-mapped hashtable.
This structure suffers from conflict misses. Increasing the
associativity of the TLB may not be the solution to conflict misses as
all the ways may have to be walked in serial.
A victim TLB is a TLB used to hold translations evicted from the
primary TLB upon replacement. The victim TLB lies between the main TLB
and its refill path. Victim TLB is of greater associativity (fully
associative in this patch). It takes longer to lookup the victim TLB,
but its likely better than a full page table walk. The memory
translation path is changed as follows :
Before Victim TLB:
1. Inline TLB lookup
2. Exit code cache on TLB miss.
3. Check for unaligned, IO accesses
4. TLB refill.
5. Do the memory access.
6. Return to code cache.
After Victim TLB:
1. Inline TLB lookup
2. Exit code cache on TLB miss.
3. Check for unaligned, IO accesses
4. Victim TLB lookup.
5. If victim TLB misses, TLB refill
6. Do the memory access.
7. Return to code cache
The advantage is that victim TLB can offer more associativity to a
directly mapped TLB and thus potentially fewer page table walks while
still keeping the time taken to flush within reasonable limits.
However, placing a victim TLB before the refill path increase TLB
refill path as the victim TLB is consulted before the TLB refill. The
performance results demonstrate that the pros outweigh the cons.
some performance results taken on SPECINT2006 train
datasets and kernel boot and qemu configure script on an
Intel(R) Xeon(R) CPU E5620 @ 2.40GHz Linux machine are shown in the
Google Doc link below.
https://docs.google.com/spreadsheets/d/1eiItzekZwNQOal_h-5iJmC4tMDi051m9qidi5_nwvH4/edit?usp=sharing
In summary, victim TLB improves the performance of qemu-system-x86_64 by
11% on average on SPECINT2006, kernelboot and qemu configscript and with
highest improvement of in 26% in 456.hmmer. And victim TLB does not result
in any performance degradation in any of the measured benchmarks. Furthermore,
the implemented victim TLB is architecture independent and is expected to
benefit other architectures in QEMU as well.
Although there are measurement fluctuations, the performance
improvement is very significant and by no means in the range of
noises.
Signed-off-by: Xin Tong <trent.tong@gmail.com>
Message-id: 1407202523-23553-1-git-send-email-trent.tong@gmail.com
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
2014-08-05 05:35:23 +04:00
|
|
|
memset(env->tlb_v_table, -1, sizeof(env->tlb_v_table));
|
2013-08-26 08:03:38 +04:00
|
|
|
memset(cpu->tb_jmp_cache, 0, sizeof(cpu->tb_jmp_cache));
|
2012-04-09 20:50:52 +04:00
|
|
|
|
implementing victim TLB for QEMU system emulated TLB
QEMU system mode page table walks are expensive. Taken by running QEMU
qemu-system-x86_64 system mode on Intel PIN , a TLB miss and walking a
4-level page tables in guest Linux OS takes ~450 X86 instructions on
average.
QEMU system mode TLB is implemented using a directly-mapped hashtable.
This structure suffers from conflict misses. Increasing the
associativity of the TLB may not be the solution to conflict misses as
all the ways may have to be walked in serial.
A victim TLB is a TLB used to hold translations evicted from the
primary TLB upon replacement. The victim TLB lies between the main TLB
and its refill path. Victim TLB is of greater associativity (fully
associative in this patch). It takes longer to lookup the victim TLB,
but its likely better than a full page table walk. The memory
translation path is changed as follows :
Before Victim TLB:
1. Inline TLB lookup
2. Exit code cache on TLB miss.
3. Check for unaligned, IO accesses
4. TLB refill.
5. Do the memory access.
6. Return to code cache.
After Victim TLB:
1. Inline TLB lookup
2. Exit code cache on TLB miss.
3. Check for unaligned, IO accesses
4. Victim TLB lookup.
5. If victim TLB misses, TLB refill
6. Do the memory access.
7. Return to code cache
The advantage is that victim TLB can offer more associativity to a
directly mapped TLB and thus potentially fewer page table walks while
still keeping the time taken to flush within reasonable limits.
However, placing a victim TLB before the refill path increase TLB
refill path as the victim TLB is consulted before the TLB refill. The
performance results demonstrate that the pros outweigh the cons.
some performance results taken on SPECINT2006 train
datasets and kernel boot and qemu configure script on an
Intel(R) Xeon(R) CPU E5620 @ 2.40GHz Linux machine are shown in the
Google Doc link below.
https://docs.google.com/spreadsheets/d/1eiItzekZwNQOal_h-5iJmC4tMDi051m9qidi5_nwvH4/edit?usp=sharing
In summary, victim TLB improves the performance of qemu-system-x86_64 by
11% on average on SPECINT2006, kernelboot and qemu configscript and with
highest improvement of in 26% in 456.hmmer. And victim TLB does not result
in any performance degradation in any of the measured benchmarks. Furthermore,
the implemented victim TLB is architecture independent and is expected to
benefit other architectures in QEMU as well.
Although there are measurement fluctuations, the performance
improvement is very significant and by no means in the range of
noises.
Signed-off-by: Xin Tong <trent.tong@gmail.com>
Message-id: 1407202523-23553-1-git-send-email-trent.tong@gmail.com
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
2014-08-05 05:35:23 +04:00
|
|
|
env->vtlb_index = 0;
|
2012-04-09 20:50:52 +04:00
|
|
|
env->tlb_flush_addr = -1;
|
|
|
|
env->tlb_flush_mask = 0;
|
|
|
|
tlb_flush_count++;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline void tlb_flush_entry(CPUTLBEntry *tlb_entry, target_ulong addr)
|
|
|
|
{
|
|
|
|
if (addr == (tlb_entry->addr_read &
|
|
|
|
(TARGET_PAGE_MASK | TLB_INVALID_MASK)) ||
|
|
|
|
addr == (tlb_entry->addr_write &
|
|
|
|
(TARGET_PAGE_MASK | TLB_INVALID_MASK)) ||
|
|
|
|
addr == (tlb_entry->addr_code &
|
|
|
|
(TARGET_PAGE_MASK | TLB_INVALID_MASK))) {
|
2013-12-07 01:44:51 +04:00
|
|
|
memset(tlb_entry, -1, sizeof(*tlb_entry));
|
2012-04-09 20:50:52 +04:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-09-04 03:29:02 +04:00
|
|
|
void tlb_flush_page(CPUState *cpu, target_ulong addr)
|
2012-04-09 20:50:52 +04:00
|
|
|
{
|
2013-09-04 03:29:02 +04:00
|
|
|
CPUArchState *env = cpu->env_ptr;
|
2012-04-09 20:50:52 +04:00
|
|
|
int i;
|
|
|
|
int mmu_idx;
|
|
|
|
|
|
|
|
#if defined(DEBUG_TLB)
|
|
|
|
printf("tlb_flush_page: " TARGET_FMT_lx "\n", addr);
|
|
|
|
#endif
|
|
|
|
/* Check if we need to flush due to large pages. */
|
|
|
|
if ((addr & env->tlb_flush_mask) == env->tlb_flush_addr) {
|
|
|
|
#if defined(DEBUG_TLB)
|
|
|
|
printf("tlb_flush_page: forced full flush ("
|
|
|
|
TARGET_FMT_lx "/" TARGET_FMT_lx ")\n",
|
|
|
|
env->tlb_flush_addr, env->tlb_flush_mask);
|
|
|
|
#endif
|
2013-09-04 04:19:44 +04:00
|
|
|
tlb_flush(cpu, 1);
|
2012-04-09 20:50:52 +04:00
|
|
|
return;
|
|
|
|
}
|
|
|
|
/* must reset current TB so that interrupts cannot modify the
|
|
|
|
links while we are modifying them */
|
2013-01-16 22:29:31 +04:00
|
|
|
cpu->current_tb = NULL;
|
2012-04-09 20:50:52 +04:00
|
|
|
|
|
|
|
addr &= TARGET_PAGE_MASK;
|
|
|
|
i = (addr >> TARGET_PAGE_BITS) & (CPU_TLB_SIZE - 1);
|
|
|
|
for (mmu_idx = 0; mmu_idx < NB_MMU_MODES; mmu_idx++) {
|
|
|
|
tlb_flush_entry(&env->tlb_table[mmu_idx][i], addr);
|
|
|
|
}
|
|
|
|
|
implementing victim TLB for QEMU system emulated TLB
QEMU system mode page table walks are expensive. Taken by running QEMU
qemu-system-x86_64 system mode on Intel PIN , a TLB miss and walking a
4-level page tables in guest Linux OS takes ~450 X86 instructions on
average.
QEMU system mode TLB is implemented using a directly-mapped hashtable.
This structure suffers from conflict misses. Increasing the
associativity of the TLB may not be the solution to conflict misses as
all the ways may have to be walked in serial.
A victim TLB is a TLB used to hold translations evicted from the
primary TLB upon replacement. The victim TLB lies between the main TLB
and its refill path. Victim TLB is of greater associativity (fully
associative in this patch). It takes longer to lookup the victim TLB,
but its likely better than a full page table walk. The memory
translation path is changed as follows :
Before Victim TLB:
1. Inline TLB lookup
2. Exit code cache on TLB miss.
3. Check for unaligned, IO accesses
4. TLB refill.
5. Do the memory access.
6. Return to code cache.
After Victim TLB:
1. Inline TLB lookup
2. Exit code cache on TLB miss.
3. Check for unaligned, IO accesses
4. Victim TLB lookup.
5. If victim TLB misses, TLB refill
6. Do the memory access.
7. Return to code cache
The advantage is that victim TLB can offer more associativity to a
directly mapped TLB and thus potentially fewer page table walks while
still keeping the time taken to flush within reasonable limits.
However, placing a victim TLB before the refill path increase TLB
refill path as the victim TLB is consulted before the TLB refill. The
performance results demonstrate that the pros outweigh the cons.
some performance results taken on SPECINT2006 train
datasets and kernel boot and qemu configure script on an
Intel(R) Xeon(R) CPU E5620 @ 2.40GHz Linux machine are shown in the
Google Doc link below.
https://docs.google.com/spreadsheets/d/1eiItzekZwNQOal_h-5iJmC4tMDi051m9qidi5_nwvH4/edit?usp=sharing
In summary, victim TLB improves the performance of qemu-system-x86_64 by
11% on average on SPECINT2006, kernelboot and qemu configscript and with
highest improvement of in 26% in 456.hmmer. And victim TLB does not result
in any performance degradation in any of the measured benchmarks. Furthermore,
the implemented victim TLB is architecture independent and is expected to
benefit other architectures in QEMU as well.
Although there are measurement fluctuations, the performance
improvement is very significant and by no means in the range of
noises.
Signed-off-by: Xin Tong <trent.tong@gmail.com>
Message-id: 1407202523-23553-1-git-send-email-trent.tong@gmail.com
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
2014-08-05 05:35:23 +04:00
|
|
|
/* check whether there are entries that need to be flushed in the vtlb */
|
|
|
|
for (mmu_idx = 0; mmu_idx < NB_MMU_MODES; mmu_idx++) {
|
|
|
|
int k;
|
|
|
|
for (k = 0; k < CPU_VTLB_SIZE; k++) {
|
|
|
|
tlb_flush_entry(&env->tlb_v_table[mmu_idx][k], addr);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-09-01 19:52:07 +04:00
|
|
|
tb_flush_jmp_cache(cpu, addr);
|
2012-04-09 20:50:52 +04:00
|
|
|
}
|
|
|
|
|
|
|
|
/* update the TLBs so that writes to code in the virtual page 'addr'
|
|
|
|
can be detected */
|
|
|
|
void tlb_protect_code(ram_addr_t ram_addr)
|
|
|
|
{
|
2013-10-10 13:49:53 +04:00
|
|
|
cpu_physical_memory_reset_dirty(ram_addr, TARGET_PAGE_SIZE,
|
2013-10-08 14:44:04 +04:00
|
|
|
DIRTY_MEMORY_CODE);
|
2012-04-09 20:50:52 +04:00
|
|
|
}
|
|
|
|
|
|
|
|
/* update the TLB so that writes in physical page 'phys_addr' are no longer
|
|
|
|
tested for self modifying code */
|
2013-09-03 12:51:26 +04:00
|
|
|
void tlb_unprotect_code_phys(CPUState *cpu, ram_addr_t ram_addr,
|
2012-04-09 20:50:52 +04:00
|
|
|
target_ulong vaddr)
|
|
|
|
{
|
2013-10-08 14:44:04 +04:00
|
|
|
cpu_physical_memory_set_dirty_flag(ram_addr, DIRTY_MEMORY_CODE);
|
2012-04-09 20:50:52 +04:00
|
|
|
}
|
|
|
|
|
|
|
|
static bool tlb_is_dirty_ram(CPUTLBEntry *tlbe)
|
|
|
|
{
|
|
|
|
return (tlbe->addr_write & (TLB_INVALID_MASK|TLB_MMIO|TLB_NOTDIRTY)) == 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
void tlb_reset_dirty_range(CPUTLBEntry *tlb_entry, uintptr_t start,
|
|
|
|
uintptr_t length)
|
|
|
|
{
|
|
|
|
uintptr_t addr;
|
|
|
|
|
|
|
|
if (tlb_is_dirty_ram(tlb_entry)) {
|
|
|
|
addr = (tlb_entry->addr_write & TARGET_PAGE_MASK) + tlb_entry->addend;
|
|
|
|
if ((addr - start) < length) {
|
|
|
|
tlb_entry->addr_write |= TLB_NOTDIRTY;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-06-03 14:44:02 +04:00
|
|
|
static inline ram_addr_t qemu_ram_addr_from_host_nofail(void *ptr)
|
|
|
|
{
|
|
|
|
ram_addr_t ram_addr;
|
|
|
|
|
2013-05-06 16:36:15 +04:00
|
|
|
if (qemu_ram_addr_from_host(ptr, &ram_addr) == NULL) {
|
2013-06-03 14:44:02 +04:00
|
|
|
fprintf(stderr, "Bad ram pointer %p\n", ptr);
|
|
|
|
abort();
|
|
|
|
}
|
|
|
|
return ram_addr;
|
|
|
|
}
|
|
|
|
|
2012-04-09 20:50:52 +04:00
|
|
|
void cpu_tlb_reset_dirty_all(ram_addr_t start1, ram_addr_t length)
|
|
|
|
{
|
2013-05-30 00:29:20 +04:00
|
|
|
CPUState *cpu;
|
2012-04-09 20:50:52 +04:00
|
|
|
CPUArchState *env;
|
|
|
|
|
2013-06-25 01:50:24 +04:00
|
|
|
CPU_FOREACH(cpu) {
|
2012-04-09 20:50:52 +04:00
|
|
|
int mmu_idx;
|
|
|
|
|
2013-05-30 00:29:20 +04:00
|
|
|
env = cpu->env_ptr;
|
2012-04-09 20:50:52 +04:00
|
|
|
for (mmu_idx = 0; mmu_idx < NB_MMU_MODES; mmu_idx++) {
|
|
|
|
unsigned int i;
|
|
|
|
|
|
|
|
for (i = 0; i < CPU_TLB_SIZE; i++) {
|
|
|
|
tlb_reset_dirty_range(&env->tlb_table[mmu_idx][i],
|
|
|
|
start1, length);
|
|
|
|
}
|
implementing victim TLB for QEMU system emulated TLB
QEMU system mode page table walks are expensive. Taken by running QEMU
qemu-system-x86_64 system mode on Intel PIN , a TLB miss and walking a
4-level page tables in guest Linux OS takes ~450 X86 instructions on
average.
QEMU system mode TLB is implemented using a directly-mapped hashtable.
This structure suffers from conflict misses. Increasing the
associativity of the TLB may not be the solution to conflict misses as
all the ways may have to be walked in serial.
A victim TLB is a TLB used to hold translations evicted from the
primary TLB upon replacement. The victim TLB lies between the main TLB
and its refill path. Victim TLB is of greater associativity (fully
associative in this patch). It takes longer to lookup the victim TLB,
but its likely better than a full page table walk. The memory
translation path is changed as follows :
Before Victim TLB:
1. Inline TLB lookup
2. Exit code cache on TLB miss.
3. Check for unaligned, IO accesses
4. TLB refill.
5. Do the memory access.
6. Return to code cache.
After Victim TLB:
1. Inline TLB lookup
2. Exit code cache on TLB miss.
3. Check for unaligned, IO accesses
4. Victim TLB lookup.
5. If victim TLB misses, TLB refill
6. Do the memory access.
7. Return to code cache
The advantage is that victim TLB can offer more associativity to a
directly mapped TLB and thus potentially fewer page table walks while
still keeping the time taken to flush within reasonable limits.
However, placing a victim TLB before the refill path increase TLB
refill path as the victim TLB is consulted before the TLB refill. The
performance results demonstrate that the pros outweigh the cons.
some performance results taken on SPECINT2006 train
datasets and kernel boot and qemu configure script on an
Intel(R) Xeon(R) CPU E5620 @ 2.40GHz Linux machine are shown in the
Google Doc link below.
https://docs.google.com/spreadsheets/d/1eiItzekZwNQOal_h-5iJmC4tMDi051m9qidi5_nwvH4/edit?usp=sharing
In summary, victim TLB improves the performance of qemu-system-x86_64 by
11% on average on SPECINT2006, kernelboot and qemu configscript and with
highest improvement of in 26% in 456.hmmer. And victim TLB does not result
in any performance degradation in any of the measured benchmarks. Furthermore,
the implemented victim TLB is architecture independent and is expected to
benefit other architectures in QEMU as well.
Although there are measurement fluctuations, the performance
improvement is very significant and by no means in the range of
noises.
Signed-off-by: Xin Tong <trent.tong@gmail.com>
Message-id: 1407202523-23553-1-git-send-email-trent.tong@gmail.com
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
2014-08-05 05:35:23 +04:00
|
|
|
|
|
|
|
for (i = 0; i < CPU_VTLB_SIZE; i++) {
|
|
|
|
tlb_reset_dirty_range(&env->tlb_v_table[mmu_idx][i],
|
|
|
|
start1, length);
|
|
|
|
}
|
2012-04-09 20:50:52 +04:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline void tlb_set_dirty1(CPUTLBEntry *tlb_entry, target_ulong vaddr)
|
|
|
|
{
|
|
|
|
if (tlb_entry->addr_write == (vaddr | TLB_NOTDIRTY)) {
|
|
|
|
tlb_entry->addr_write = vaddr;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* update the TLB corresponding to virtual page vaddr
|
|
|
|
so that it is no longer dirty */
|
|
|
|
void tlb_set_dirty(CPUArchState *env, target_ulong vaddr)
|
|
|
|
{
|
|
|
|
int i;
|
|
|
|
int mmu_idx;
|
|
|
|
|
|
|
|
vaddr &= TARGET_PAGE_MASK;
|
|
|
|
i = (vaddr >> TARGET_PAGE_BITS) & (CPU_TLB_SIZE - 1);
|
|
|
|
for (mmu_idx = 0; mmu_idx < NB_MMU_MODES; mmu_idx++) {
|
|
|
|
tlb_set_dirty1(&env->tlb_table[mmu_idx][i], vaddr);
|
|
|
|
}
|
implementing victim TLB for QEMU system emulated TLB
QEMU system mode page table walks are expensive. Taken by running QEMU
qemu-system-x86_64 system mode on Intel PIN , a TLB miss and walking a
4-level page tables in guest Linux OS takes ~450 X86 instructions on
average.
QEMU system mode TLB is implemented using a directly-mapped hashtable.
This structure suffers from conflict misses. Increasing the
associativity of the TLB may not be the solution to conflict misses as
all the ways may have to be walked in serial.
A victim TLB is a TLB used to hold translations evicted from the
primary TLB upon replacement. The victim TLB lies between the main TLB
and its refill path. Victim TLB is of greater associativity (fully
associative in this patch). It takes longer to lookup the victim TLB,
but its likely better than a full page table walk. The memory
translation path is changed as follows :
Before Victim TLB:
1. Inline TLB lookup
2. Exit code cache on TLB miss.
3. Check for unaligned, IO accesses
4. TLB refill.
5. Do the memory access.
6. Return to code cache.
After Victim TLB:
1. Inline TLB lookup
2. Exit code cache on TLB miss.
3. Check for unaligned, IO accesses
4. Victim TLB lookup.
5. If victim TLB misses, TLB refill
6. Do the memory access.
7. Return to code cache
The advantage is that victim TLB can offer more associativity to a
directly mapped TLB and thus potentially fewer page table walks while
still keeping the time taken to flush within reasonable limits.
However, placing a victim TLB before the refill path increase TLB
refill path as the victim TLB is consulted before the TLB refill. The
performance results demonstrate that the pros outweigh the cons.
some performance results taken on SPECINT2006 train
datasets and kernel boot and qemu configure script on an
Intel(R) Xeon(R) CPU E5620 @ 2.40GHz Linux machine are shown in the
Google Doc link below.
https://docs.google.com/spreadsheets/d/1eiItzekZwNQOal_h-5iJmC4tMDi051m9qidi5_nwvH4/edit?usp=sharing
In summary, victim TLB improves the performance of qemu-system-x86_64 by
11% on average on SPECINT2006, kernelboot and qemu configscript and with
highest improvement of in 26% in 456.hmmer. And victim TLB does not result
in any performance degradation in any of the measured benchmarks. Furthermore,
the implemented victim TLB is architecture independent and is expected to
benefit other architectures in QEMU as well.
Although there are measurement fluctuations, the performance
improvement is very significant and by no means in the range of
noises.
Signed-off-by: Xin Tong <trent.tong@gmail.com>
Message-id: 1407202523-23553-1-git-send-email-trent.tong@gmail.com
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
2014-08-05 05:35:23 +04:00
|
|
|
|
|
|
|
for (mmu_idx = 0; mmu_idx < NB_MMU_MODES; mmu_idx++) {
|
|
|
|
int k;
|
|
|
|
for (k = 0; k < CPU_VTLB_SIZE; k++) {
|
|
|
|
tlb_set_dirty1(&env->tlb_v_table[mmu_idx][k], vaddr);
|
|
|
|
}
|
|
|
|
}
|
2012-04-09 20:50:52 +04:00
|
|
|
}
|
|
|
|
|
|
|
|
/* Our TLB does not support large pages, so remember the area covered by
|
|
|
|
large pages and trigger a full TLB flush if these are invalidated. */
|
|
|
|
static void tlb_add_large_page(CPUArchState *env, target_ulong vaddr,
|
|
|
|
target_ulong size)
|
|
|
|
{
|
|
|
|
target_ulong mask = ~(size - 1);
|
|
|
|
|
|
|
|
if (env->tlb_flush_addr == (target_ulong)-1) {
|
|
|
|
env->tlb_flush_addr = vaddr & mask;
|
|
|
|
env->tlb_flush_mask = mask;
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
/* Extend the existing region to include the new page.
|
|
|
|
This is a compromise between unnecessary flushes and the cost
|
|
|
|
of maintaining a full variable size TLB. */
|
|
|
|
mask &= env->tlb_flush_mask;
|
|
|
|
while (((env->tlb_flush_addr ^ vaddr) & mask) != 0) {
|
|
|
|
mask <<= 1;
|
|
|
|
}
|
|
|
|
env->tlb_flush_addr &= mask;
|
|
|
|
env->tlb_flush_mask = mask;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Add a new TLB entry. At most one entry for a given virtual address
|
|
|
|
is permitted. Only a single TARGET_PAGE_SIZE region is mapped, the
|
|
|
|
supplied size is only used by tlb_flush_page. */
|
2013-09-03 15:59:37 +04:00
|
|
|
void tlb_set_page(CPUState *cpu, target_ulong vaddr,
|
2012-10-23 14:30:10 +04:00
|
|
|
hwaddr paddr, int prot,
|
2012-04-09 20:50:52 +04:00
|
|
|
int mmu_idx, target_ulong size)
|
|
|
|
{
|
2013-09-03 15:59:37 +04:00
|
|
|
CPUArchState *env = cpu->env_ptr;
|
2012-04-09 20:50:52 +04:00
|
|
|
MemoryRegionSection *section;
|
|
|
|
unsigned int index;
|
|
|
|
target_ulong address;
|
|
|
|
target_ulong code_address;
|
|
|
|
uintptr_t addend;
|
|
|
|
CPUTLBEntry *te;
|
2013-05-24 14:59:37 +04:00
|
|
|
hwaddr iotlb, xlat, sz;
|
implementing victim TLB for QEMU system emulated TLB
QEMU system mode page table walks are expensive. Taken by running QEMU
qemu-system-x86_64 system mode on Intel PIN , a TLB miss and walking a
4-level page tables in guest Linux OS takes ~450 X86 instructions on
average.
QEMU system mode TLB is implemented using a directly-mapped hashtable.
This structure suffers from conflict misses. Increasing the
associativity of the TLB may not be the solution to conflict misses as
all the ways may have to be walked in serial.
A victim TLB is a TLB used to hold translations evicted from the
primary TLB upon replacement. The victim TLB lies between the main TLB
and its refill path. Victim TLB is of greater associativity (fully
associative in this patch). It takes longer to lookup the victim TLB,
but its likely better than a full page table walk. The memory
translation path is changed as follows :
Before Victim TLB:
1. Inline TLB lookup
2. Exit code cache on TLB miss.
3. Check for unaligned, IO accesses
4. TLB refill.
5. Do the memory access.
6. Return to code cache.
After Victim TLB:
1. Inline TLB lookup
2. Exit code cache on TLB miss.
3. Check for unaligned, IO accesses
4. Victim TLB lookup.
5. If victim TLB misses, TLB refill
6. Do the memory access.
7. Return to code cache
The advantage is that victim TLB can offer more associativity to a
directly mapped TLB and thus potentially fewer page table walks while
still keeping the time taken to flush within reasonable limits.
However, placing a victim TLB before the refill path increase TLB
refill path as the victim TLB is consulted before the TLB refill. The
performance results demonstrate that the pros outweigh the cons.
some performance results taken on SPECINT2006 train
datasets and kernel boot and qemu configure script on an
Intel(R) Xeon(R) CPU E5620 @ 2.40GHz Linux machine are shown in the
Google Doc link below.
https://docs.google.com/spreadsheets/d/1eiItzekZwNQOal_h-5iJmC4tMDi051m9qidi5_nwvH4/edit?usp=sharing
In summary, victim TLB improves the performance of qemu-system-x86_64 by
11% on average on SPECINT2006, kernelboot and qemu configscript and with
highest improvement of in 26% in 456.hmmer. And victim TLB does not result
in any performance degradation in any of the measured benchmarks. Furthermore,
the implemented victim TLB is architecture independent and is expected to
benefit other architectures in QEMU as well.
Although there are measurement fluctuations, the performance
improvement is very significant and by no means in the range of
noises.
Signed-off-by: Xin Tong <trent.tong@gmail.com>
Message-id: 1407202523-23553-1-git-send-email-trent.tong@gmail.com
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
2014-08-05 05:35:23 +04:00
|
|
|
unsigned vidx = env->vtlb_index++ % CPU_VTLB_SIZE;
|
2012-04-09 20:50:52 +04:00
|
|
|
|
|
|
|
assert(size >= TARGET_PAGE_SIZE);
|
|
|
|
if (size != TARGET_PAGE_SIZE) {
|
|
|
|
tlb_add_large_page(env, vaddr, size);
|
|
|
|
}
|
2013-05-24 14:59:37 +04:00
|
|
|
|
|
|
|
sz = size;
|
2013-12-17 07:06:51 +04:00
|
|
|
section = address_space_translate_for_iotlb(cpu->as, paddr,
|
2013-05-26 23:46:51 +04:00
|
|
|
&xlat, &sz);
|
2013-05-24 14:59:37 +04:00
|
|
|
assert(sz >= TARGET_PAGE_SIZE);
|
|
|
|
|
2012-04-09 20:50:52 +04:00
|
|
|
#if defined(DEBUG_TLB)
|
|
|
|
printf("tlb_set_page: vaddr=" TARGET_FMT_lx " paddr=0x" TARGET_FMT_plx
|
2013-06-05 16:16:42 +04:00
|
|
|
" prot=%x idx=%d\n",
|
|
|
|
vaddr, paddr, prot, mmu_idx);
|
2012-04-09 20:50:52 +04:00
|
|
|
#endif
|
|
|
|
|
|
|
|
address = vaddr;
|
2013-05-24 18:45:30 +04:00
|
|
|
if (!memory_region_is_ram(section->mr) && !memory_region_is_romd(section->mr)) {
|
|
|
|
/* IO memory case */
|
2012-04-09 20:50:52 +04:00
|
|
|
address |= TLB_MMIO;
|
2013-05-24 18:45:30 +04:00
|
|
|
addend = 0;
|
|
|
|
} else {
|
|
|
|
/* TLB_MMIO for rom/romd handled below */
|
2013-05-24 14:59:37 +04:00
|
|
|
addend = (uintptr_t)memory_region_get_ram_ptr(section->mr) + xlat;
|
2012-04-09 20:50:52 +04:00
|
|
|
}
|
|
|
|
|
|
|
|
code_address = address;
|
2013-09-03 15:32:01 +04:00
|
|
|
iotlb = memory_region_section_get_iotlb(cpu, section, vaddr, paddr, xlat,
|
2013-05-24 14:59:37 +04:00
|
|
|
prot, &address);
|
2012-04-09 20:50:52 +04:00
|
|
|
|
|
|
|
index = (vaddr >> TARGET_PAGE_BITS) & (CPU_TLB_SIZE - 1);
|
|
|
|
te = &env->tlb_table[mmu_idx][index];
|
implementing victim TLB for QEMU system emulated TLB
QEMU system mode page table walks are expensive. Taken by running QEMU
qemu-system-x86_64 system mode on Intel PIN , a TLB miss and walking a
4-level page tables in guest Linux OS takes ~450 X86 instructions on
average.
QEMU system mode TLB is implemented using a directly-mapped hashtable.
This structure suffers from conflict misses. Increasing the
associativity of the TLB may not be the solution to conflict misses as
all the ways may have to be walked in serial.
A victim TLB is a TLB used to hold translations evicted from the
primary TLB upon replacement. The victim TLB lies between the main TLB
and its refill path. Victim TLB is of greater associativity (fully
associative in this patch). It takes longer to lookup the victim TLB,
but its likely better than a full page table walk. The memory
translation path is changed as follows :
Before Victim TLB:
1. Inline TLB lookup
2. Exit code cache on TLB miss.
3. Check for unaligned, IO accesses
4. TLB refill.
5. Do the memory access.
6. Return to code cache.
After Victim TLB:
1. Inline TLB lookup
2. Exit code cache on TLB miss.
3. Check for unaligned, IO accesses
4. Victim TLB lookup.
5. If victim TLB misses, TLB refill
6. Do the memory access.
7. Return to code cache
The advantage is that victim TLB can offer more associativity to a
directly mapped TLB and thus potentially fewer page table walks while
still keeping the time taken to flush within reasonable limits.
However, placing a victim TLB before the refill path increase TLB
refill path as the victim TLB is consulted before the TLB refill. The
performance results demonstrate that the pros outweigh the cons.
some performance results taken on SPECINT2006 train
datasets and kernel boot and qemu configure script on an
Intel(R) Xeon(R) CPU E5620 @ 2.40GHz Linux machine are shown in the
Google Doc link below.
https://docs.google.com/spreadsheets/d/1eiItzekZwNQOal_h-5iJmC4tMDi051m9qidi5_nwvH4/edit?usp=sharing
In summary, victim TLB improves the performance of qemu-system-x86_64 by
11% on average on SPECINT2006, kernelboot and qemu configscript and with
highest improvement of in 26% in 456.hmmer. And victim TLB does not result
in any performance degradation in any of the measured benchmarks. Furthermore,
the implemented victim TLB is architecture independent and is expected to
benefit other architectures in QEMU as well.
Although there are measurement fluctuations, the performance
improvement is very significant and by no means in the range of
noises.
Signed-off-by: Xin Tong <trent.tong@gmail.com>
Message-id: 1407202523-23553-1-git-send-email-trent.tong@gmail.com
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
2014-08-05 05:35:23 +04:00
|
|
|
|
|
|
|
/* do not discard the translation in te, evict it into a victim tlb */
|
|
|
|
env->tlb_v_table[mmu_idx][vidx] = *te;
|
|
|
|
env->iotlb_v[mmu_idx][vidx] = env->iotlb[mmu_idx][index];
|
|
|
|
|
|
|
|
/* refill the tlb */
|
|
|
|
env->iotlb[mmu_idx][index] = iotlb - vaddr;
|
2012-04-09 20:50:52 +04:00
|
|
|
te->addend = addend - vaddr;
|
|
|
|
if (prot & PAGE_READ) {
|
|
|
|
te->addr_read = address;
|
|
|
|
} else {
|
|
|
|
te->addr_read = -1;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (prot & PAGE_EXEC) {
|
|
|
|
te->addr_code = code_address;
|
|
|
|
} else {
|
|
|
|
te->addr_code = -1;
|
|
|
|
}
|
|
|
|
if (prot & PAGE_WRITE) {
|
|
|
|
if ((memory_region_is_ram(section->mr) && section->readonly)
|
2012-04-14 18:56:48 +04:00
|
|
|
|| memory_region_is_romd(section->mr)) {
|
2012-04-09 20:50:52 +04:00
|
|
|
/* Write access calls the I/O callback. */
|
|
|
|
te->addr_write = address | TLB_MMIO;
|
|
|
|
} else if (memory_region_is_ram(section->mr)
|
2013-10-10 13:20:22 +04:00
|
|
|
&& cpu_physical_memory_is_clean(section->mr->ram_addr
|
|
|
|
+ xlat)) {
|
2012-04-09 20:50:52 +04:00
|
|
|
te->addr_write = address | TLB_NOTDIRTY;
|
|
|
|
} else {
|
|
|
|
te->addr_write = address;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
te->addr_write = -1;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* NOTE: this function can trigger an exception */
|
|
|
|
/* NOTE2: the returned address is not exactly the physical address: it
|
2012-08-10 20:14:05 +04:00
|
|
|
* is actually a ram_addr_t (in system mode; the user mode emulation
|
|
|
|
* version of this function returns a guest virtual address).
|
|
|
|
*/
|
2012-04-09 20:50:52 +04:00
|
|
|
tb_page_addr_t get_page_addr_code(CPUArchState *env1, target_ulong addr)
|
|
|
|
{
|
|
|
|
int mmu_idx, page_index, pd;
|
|
|
|
void *p;
|
|
|
|
MemoryRegion *mr;
|
2013-12-17 07:06:51 +04:00
|
|
|
CPUState *cpu = ENV_GET_CPU(env1);
|
2012-04-09 20:50:52 +04:00
|
|
|
|
|
|
|
page_index = (addr >> TARGET_PAGE_BITS) & (CPU_TLB_SIZE - 1);
|
|
|
|
mmu_idx = cpu_mmu_index(env1);
|
|
|
|
if (unlikely(env1->tlb_table[mmu_idx][page_index].addr_code !=
|
|
|
|
(addr & TARGET_PAGE_MASK))) {
|
|
|
|
cpu_ldub_code(env1, addr);
|
|
|
|
}
|
|
|
|
pd = env1->iotlb[mmu_idx][page_index] & ~TARGET_PAGE_MASK;
|
2013-12-17 07:06:51 +04:00
|
|
|
mr = iotlb_to_region(cpu->as, pd);
|
2012-04-09 20:50:52 +04:00
|
|
|
if (memory_region_is_unassigned(mr)) {
|
2013-05-27 08:49:53 +04:00
|
|
|
CPUClass *cc = CPU_GET_CLASS(cpu);
|
|
|
|
|
|
|
|
if (cc->do_unassigned_access) {
|
|
|
|
cc->do_unassigned_access(cpu, addr, false, true, 0, 4);
|
|
|
|
} else {
|
2013-09-03 19:38:47 +04:00
|
|
|
cpu_abort(cpu, "Trying to execute code outside RAM or ROM at 0x"
|
2013-05-27 08:49:53 +04:00
|
|
|
TARGET_FMT_lx "\n", addr);
|
|
|
|
}
|
2012-04-09 20:50:52 +04:00
|
|
|
}
|
|
|
|
p = (void *)((uintptr_t)addr + env1->tlb_table[mmu_idx][page_index].addend);
|
|
|
|
return qemu_ram_addr_from_host_nofail(p);
|
|
|
|
}
|
|
|
|
|
2014-03-28 20:55:24 +04:00
|
|
|
#define MMUSUFFIX _mmu
|
|
|
|
|
|
|
|
#define SHIFT 0
|
2014-03-28 21:00:25 +04:00
|
|
|
#include "softmmu_template.h"
|
2014-03-28 20:55:24 +04:00
|
|
|
|
|
|
|
#define SHIFT 1
|
2014-03-28 21:00:25 +04:00
|
|
|
#include "softmmu_template.h"
|
2014-03-28 20:55:24 +04:00
|
|
|
|
|
|
|
#define SHIFT 2
|
2014-03-28 21:00:25 +04:00
|
|
|
#include "softmmu_template.h"
|
2014-03-28 20:55:24 +04:00
|
|
|
|
|
|
|
#define SHIFT 3
|
2014-03-28 21:00:25 +04:00
|
|
|
#include "softmmu_template.h"
|
2014-03-28 20:55:24 +04:00
|
|
|
#undef MMUSUFFIX
|
|
|
|
|
2012-04-09 20:50:52 +04:00
|
|
|
#define MMUSUFFIX _cmmu
|
2014-04-28 21:20:00 +04:00
|
|
|
#undef GETPC_ADJ
|
|
|
|
#define GETPC_ADJ 0
|
|
|
|
#undef GETRA
|
|
|
|
#define GETRA() ((uintptr_t)0)
|
2012-04-09 20:50:52 +04:00
|
|
|
#define SOFTMMU_CODE_ACCESS
|
|
|
|
|
|
|
|
#define SHIFT 0
|
2014-03-28 21:00:25 +04:00
|
|
|
#include "softmmu_template.h"
|
2012-04-09 20:50:52 +04:00
|
|
|
|
|
|
|
#define SHIFT 1
|
2014-03-28 21:00:25 +04:00
|
|
|
#include "softmmu_template.h"
|
2012-04-09 20:50:52 +04:00
|
|
|
|
|
|
|
#define SHIFT 2
|
2014-03-28 21:00:25 +04:00
|
|
|
#include "softmmu_template.h"
|
2012-04-09 20:50:52 +04:00
|
|
|
|
|
|
|
#define SHIFT 3
|
2014-03-28 21:00:25 +04:00
|
|
|
#include "softmmu_template.h"
|