xemu/cputlb.c

732 lines
23 KiB
C
Raw Normal View History

/*
* 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 "qemu/osdep.h"
#include "cpu.h"
#include "exec/exec-all.h"
#include "exec/memory.h"
#include "exec/address-spaces.h"
#include "exec/cpu_ldst.h"
#include "exec/cputlb.h"
#include "exec/memory-internal.h"
#include "exec/ram_addr.h"
#include "tcg/tcg.h"
#include "qemu/error-report.h"
#include "exec/log.h"
#include "exec/helper-proto.h"
#include "qemu/atomic.h"
/* DEBUG defines, enable DEBUG_TLB_LOG to log to the CPU_LOG_MMU target */
/* #define DEBUG_TLB */
/* #define DEBUG_TLB_LOG */
#ifdef DEBUG_TLB
# define DEBUG_TLB_GATE 1
# ifdef DEBUG_TLB_LOG
# define DEBUG_TLB_LOG_GATE 1
# else
# define DEBUG_TLB_LOG_GATE 0
# endif
#else
# define DEBUG_TLB_GATE 0
# define DEBUG_TLB_LOG_GATE 0
#endif
#define tlb_debug(fmt, ...) do { \
if (DEBUG_TLB_LOG_GATE) { \
qemu_log_mask(CPU_LOG_MMU, "%s: " fmt, __func__, \
## __VA_ARGS__); \
} else if (DEBUG_TLB_GATE) { \
fprintf(stderr, "%s: " fmt, __func__, ## __VA_ARGS__); \
} \
} while (0)
/* statistics */
int tlb_flush_count;
/* 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.
*/
void tlb_flush(CPUState *cpu)
{
CPUArchState *env = cpu->env_ptr;
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 01:35:23 +00:00
memset(env->tlb_v_table, -1, sizeof(env->tlb_v_table));
memset(cpu->tb_jmp_cache, 0, sizeof(cpu->tb_jmp_cache));
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 01:35:23 +00:00
env->vtlb_index = 0;
env->tlb_flush_addr = -1;
env->tlb_flush_mask = 0;
tlb_flush_count++;
}
static inline void v_tlb_flush_by_mmuidx(CPUState *cpu, va_list argp)
{
CPUArchState *env = cpu->env_ptr;
tlb_debug("start\n");
for (;;) {
int mmu_idx = va_arg(argp, int);
if (mmu_idx < 0) {
break;
}
tlb_debug("%d\n", mmu_idx);
memset(env->tlb_table[mmu_idx], -1, sizeof(env->tlb_table[0]));
memset(env->tlb_v_table[mmu_idx], -1, sizeof(env->tlb_v_table[0]));
}
memset(cpu->tb_jmp_cache, 0, sizeof(cpu->tb_jmp_cache));
}
void tlb_flush_by_mmuidx(CPUState *cpu, ...)
{
va_list argp;
va_start(argp, cpu);
v_tlb_flush_by_mmuidx(cpu, argp);
va_end(argp);
}
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))) {
memset(tlb_entry, -1, sizeof(*tlb_entry));
}
}
void tlb_flush_page(CPUState *cpu, target_ulong addr)
{
CPUArchState *env = cpu->env_ptr;
int i;
int mmu_idx;
tlb_debug("page :" TARGET_FMT_lx "\n", addr);
/* Check if we need to flush due to large pages. */
if ((addr & env->tlb_flush_mask) == env->tlb_flush_addr) {
tlb_debug("forcing full flush ("
TARGET_FMT_lx "/" TARGET_FMT_lx ")\n",
env->tlb_flush_addr, env->tlb_flush_mask);
tlb_flush(cpu);
return;
}
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 01:35:23 +00:00
/* 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);
}
}
tb_flush_jmp_cache(cpu, addr);
}
void tlb_flush_page_by_mmuidx(CPUState *cpu, target_ulong addr, ...)
{
CPUArchState *env = cpu->env_ptr;
int i, k;
va_list argp;
va_start(argp, addr);
tlb_debug("addr "TARGET_FMT_lx"\n", addr);
/* Check if we need to flush due to large pages. */
if ((addr & env->tlb_flush_mask) == env->tlb_flush_addr) {
tlb_debug("forced full flush ("
TARGET_FMT_lx "/" TARGET_FMT_lx ")\n",
env->tlb_flush_addr, env->tlb_flush_mask);
v_tlb_flush_by_mmuidx(cpu, argp);
va_end(argp);
return;
}
addr &= TARGET_PAGE_MASK;
i = (addr >> TARGET_PAGE_BITS) & (CPU_TLB_SIZE - 1);
for (;;) {
int mmu_idx = va_arg(argp, int);
if (mmu_idx < 0) {
break;
}
tlb_debug("idx %d\n", mmu_idx);
tlb_flush_entry(&env->tlb_table[mmu_idx][i], addr);
/* check whether there are vltb entries that need to be flushed */
for (k = 0; k < CPU_VTLB_SIZE; k++) {
tlb_flush_entry(&env->tlb_v_table[mmu_idx][k], addr);
}
}
va_end(argp);
tb_flush_jmp_cache(cpu, addr);
}
/* 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)
{
cpu_physical_memory_test_and_clear_dirty(ram_addr, TARGET_PAGE_SIZE,
DIRTY_MEMORY_CODE);
}
/* update the TLB so that writes in physical page 'phys_addr' are no longer
tested for self modifying code */
void tlb_unprotect_code(ram_addr_t ram_addr)
{
cpu_physical_memory_set_dirty_flag(ram_addr, DIRTY_MEMORY_CODE);
}
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;
}
}
}
static inline ram_addr_t qemu_ram_addr_from_host_nofail(void *ptr)
{
ram_addr_t ram_addr;
ram_addr = qemu_ram_addr_from_host(ptr);
if (ram_addr == RAM_ADDR_INVALID) {
fprintf(stderr, "Bad ram pointer %p\n", ptr);
abort();
}
return ram_addr;
}
void tlb_reset_dirty(CPUState *cpu, ram_addr_t start1, ram_addr_t length)
{
CPUArchState *env;
int mmu_idx;
env = cpu->env_ptr;
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 01:35:23 +00:00
for (i = 0; i < CPU_VTLB_SIZE; i++) {
tlb_reset_dirty_range(&env->tlb_v_table[mmu_idx][i],
start1, length);
}
}
}
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(CPUState *cpu, target_ulong vaddr)
{
CPUArchState *env = cpu->env_ptr;
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 01:35:23 +00:00
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);
}
}
}
/* 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.
*
* Called from TCG-generated code, which is under an RCU read-side
* critical section.
*/
void tlb_set_page_with_attrs(CPUState *cpu, target_ulong vaddr,
hwaddr paddr, MemTxAttrs attrs, int prot,
int mmu_idx, target_ulong size)
{
CPUArchState *env = cpu->env_ptr;
MemoryRegionSection *section;
unsigned int index;
target_ulong address;
target_ulong code_address;
uintptr_t addend;
CPUTLBEntry *te;
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 01:35:23 +00:00
unsigned vidx = env->vtlb_index++ % CPU_VTLB_SIZE;
int asidx = cpu_asidx_from_attrs(cpu, attrs);
assert(size >= TARGET_PAGE_SIZE);
if (size != TARGET_PAGE_SIZE) {
tlb_add_large_page(env, vaddr, size);
}
sz = size;
section = address_space_translate_for_iotlb(cpu, asidx, paddr, &xlat, &sz);
assert(sz >= TARGET_PAGE_SIZE);
tlb_debug("vaddr=" TARGET_FMT_lx " paddr=0x" TARGET_FMT_plx
" prot=%x idx=%d\n",
vaddr, paddr, prot, mmu_idx);
address = vaddr;
if (!memory_region_is_ram(section->mr) && !memory_region_is_romd(section->mr)) {
/* IO memory case */
address |= TLB_MMIO;
addend = 0;
} else {
/* TLB_MMIO for rom/romd handled below */
addend = (uintptr_t)memory_region_get_ram_ptr(section->mr) + xlat;
}
code_address = address;
iotlb = memory_region_section_get_iotlb(cpu, section, vaddr, paddr, xlat,
prot, &address);
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 01:35:23 +00:00
/* 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].addr = iotlb - vaddr;
env->iotlb[mmu_idx][index].attrs = attrs;
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)
|| memory_region_is_romd(section->mr)) {
/* Write access calls the I/O callback. */
te->addr_write = address | TLB_MMIO;
} else if (memory_region_is_ram(section->mr)
&& cpu_physical_memory_is_clean(
memory_region_get_ram_addr(section->mr) + xlat)) {
te->addr_write = address | TLB_NOTDIRTY;
} else {
te->addr_write = address;
}
} else {
te->addr_write = -1;
}
}
/* Add a new TLB entry, but without specifying the memory
* transaction attributes to be used.
*/
void tlb_set_page(CPUState *cpu, target_ulong vaddr,
hwaddr paddr, int prot,
int mmu_idx, target_ulong size)
{
tlb_set_page_with_attrs(cpu, vaddr, paddr, MEMTXATTRS_UNSPECIFIED,
prot, mmu_idx, size);
}
static void report_bad_exec(CPUState *cpu, target_ulong addr)
{
/* Accidentally executing outside RAM or ROM is quite common for
* several user-error situations, so report it in a way that
* makes it clear that this isn't a QEMU bug and provide suggestions
* about what a user could do to fix things.
*/
error_report("Trying to execute code outside RAM or ROM at 0x"
TARGET_FMT_lx, addr);
error_printf("This usually means one of the following happened:\n\n"
"(1) You told QEMU to execute a kernel for the wrong machine "
"type, and it crashed on startup (eg trying to run a "
"raspberry pi kernel on a versatilepb QEMU machine)\n"
"(2) You didn't give QEMU a kernel or BIOS filename at all, "
"and QEMU executed a ROM full of no-op instructions until "
"it fell off the end\n"
"(3) Your guest kernel has a bug and crashed by jumping "
"off into nowhere\n\n"
"This is almost always one of the first two, so check your "
"command line and that you are using the right type of kernel "
"for this machine.\n"
"If you think option (3) is likely then you can try debugging "
"your guest with the -d debug options; in particular "
"-d guest_errors will cause the log to include a dump of the "
"guest register state at this point.\n\n"
"Execution cannot continue; stopping here.\n\n");
/* Report also to the logs, with more detail including register dump */
qemu_log_mask(LOG_GUEST_ERROR, "qemu: fatal: Trying to execute code "
"outside RAM or ROM at 0x" TARGET_FMT_lx "\n", addr);
log_cpu_state_mask(LOG_GUEST_ERROR, cpu, CPU_DUMP_FPU | CPU_DUMP_CCOP);
}
/* NOTE: this function can trigger an exception */
/* NOTE2: the returned address is not exactly the physical address: it
* is actually a ram_addr_t (in system mode; the user mode emulation
* version of this function returns a guest virtual address).
*/
tb_page_addr_t get_page_addr_code(CPUArchState *env1, target_ulong addr)
{
int mmu_idx, page_index, pd;
void *p;
MemoryRegion *mr;
CPUState *cpu = ENV_GET_CPU(env1);
CPUIOTLBEntry *iotlbentry;
page_index = (addr >> TARGET_PAGE_BITS) & (CPU_TLB_SIZE - 1);
mmu_idx = cpu_mmu_index(env1, true);
if (unlikely(env1->tlb_table[mmu_idx][page_index].addr_code !=
(addr & TARGET_PAGE_MASK))) {
cpu_ldub_code(env1, addr);
}
iotlbentry = &env1->iotlb[mmu_idx][page_index];
pd = iotlbentry->addr & ~TARGET_PAGE_MASK;
mr = iotlb_to_region(cpu, pd, iotlbentry->attrs);
if (memory_region_is_unassigned(mr)) {
CPUClass *cc = CPU_GET_CLASS(cpu);
if (cc->do_unassigned_access) {
cc->do_unassigned_access(cpu, addr, false, true, 0, 4);
} else {
report_bad_exec(cpu, addr);
exit(1);
}
}
p = (void *)((uintptr_t)addr + env1->tlb_table[mmu_idx][page_index].addend);
return qemu_ram_addr_from_host_nofail(p);
}
static uint64_t io_readx(CPUArchState *env, CPUIOTLBEntry *iotlbentry,
target_ulong addr, uintptr_t retaddr, int size)
{
CPUState *cpu = ENV_GET_CPU(env);
hwaddr physaddr = iotlbentry->addr;
MemoryRegion *mr = iotlb_to_region(cpu, physaddr, iotlbentry->attrs);
uint64_t val;
physaddr = (physaddr & TARGET_PAGE_MASK) + addr;
cpu->mem_io_pc = retaddr;
if (mr != &io_mem_rom && mr != &io_mem_notdirty && !cpu->can_do_io) {
cpu_io_recompile(cpu, retaddr);
}
cpu->mem_io_vaddr = addr;
memory_region_dispatch_read(mr, physaddr, &val, size, iotlbentry->attrs);
return val;
}
static void io_writex(CPUArchState *env, CPUIOTLBEntry *iotlbentry,
uint64_t val, target_ulong addr,
uintptr_t retaddr, int size)
{
CPUState *cpu = ENV_GET_CPU(env);
hwaddr physaddr = iotlbentry->addr;
MemoryRegion *mr = iotlb_to_region(cpu, physaddr, iotlbentry->attrs);
physaddr = (physaddr & TARGET_PAGE_MASK) + addr;
if (mr != &io_mem_rom && mr != &io_mem_notdirty && !cpu->can_do_io) {
cpu_io_recompile(cpu, retaddr);
}
cpu->mem_io_vaddr = addr;
cpu->mem_io_pc = retaddr;
memory_region_dispatch_write(mr, physaddr, val, size, iotlbentry->attrs);
}
/* Return true if ADDR is present in the victim tlb, and has been copied
back to the main tlb. */
static bool victim_tlb_hit(CPUArchState *env, size_t mmu_idx, size_t index,
size_t elt_ofs, target_ulong page)
{
size_t vidx;
for (vidx = 0; vidx < CPU_VTLB_SIZE; ++vidx) {
CPUTLBEntry *vtlb = &env->tlb_v_table[mmu_idx][vidx];
target_ulong cmp = *(target_ulong *)((uintptr_t)vtlb + elt_ofs);
if (cmp == page) {
/* Found entry in victim tlb, swap tlb and iotlb. */
CPUTLBEntry tmptlb, *tlb = &env->tlb_table[mmu_idx][index];
CPUIOTLBEntry tmpio, *io = &env->iotlb[mmu_idx][index];
CPUIOTLBEntry *vio = &env->iotlb_v[mmu_idx][vidx];
tmptlb = *tlb; *tlb = *vtlb; *vtlb = tmptlb;
tmpio = *io; *io = *vio; *vio = tmpio;
return true;
}
}
return false;
}
/* Macro to call the above, with local variables from the use context. */
#define VICTIM_TLB_HIT(TY, ADDR) \
victim_tlb_hit(env, mmu_idx, index, offsetof(CPUTLBEntry, TY), \
(ADDR) & TARGET_PAGE_MASK)
/* Probe for whether the specified guest write access is permitted.
* If it is not permitted then an exception will be taken in the same
* way as if this were a real write access (and we will not return).
* Otherwise the function will return, and there will be a valid
* entry in the TLB for this access.
*/
void probe_write(CPUArchState *env, target_ulong addr, int mmu_idx,
uintptr_t retaddr)
{
int index = (addr >> TARGET_PAGE_BITS) & (CPU_TLB_SIZE - 1);
target_ulong tlb_addr = env->tlb_table[mmu_idx][index].addr_write;
if ((addr & TARGET_PAGE_MASK)
!= (tlb_addr & (TARGET_PAGE_MASK | TLB_INVALID_MASK))) {
/* TLB entry is for a different page */
if (!VICTIM_TLB_HIT(addr_write, addr)) {
tlb_fill(ENV_GET_CPU(env), addr, MMU_DATA_STORE, mmu_idx, retaddr);
}
}
}
/* Probe for a read-modify-write atomic operation. Do not allow unaligned
* operations, or io operations to proceed. Return the host address. */
static void *atomic_mmu_lookup(CPUArchState *env, target_ulong addr,
TCGMemOpIdx oi, uintptr_t retaddr)
{
size_t mmu_idx = get_mmuidx(oi);
size_t index = (addr >> TARGET_PAGE_BITS) & (CPU_TLB_SIZE - 1);
CPUTLBEntry *tlbe = &env->tlb_table[mmu_idx][index];
target_ulong tlb_addr = tlbe->addr_write;
TCGMemOp mop = get_memop(oi);
int a_bits = get_alignment_bits(mop);
int s_bits = mop & MO_SIZE;
/* Adjust the given return address. */
retaddr -= GETPC_ADJ;
/* Enforce guest required alignment. */
if (unlikely(a_bits > 0 && (addr & ((1 << a_bits) - 1)))) {
/* ??? Maybe indicate atomic op to cpu_unaligned_access */
cpu_unaligned_access(ENV_GET_CPU(env), addr, MMU_DATA_STORE,
mmu_idx, retaddr);
}
/* Enforce qemu required alignment. */
if (unlikely(addr & ((1 << s_bits) - 1))) {
/* We get here if guest alignment was not requested,
or was not enforced by cpu_unaligned_access above.
We might widen the access and emulate, but for now
mark an exception and exit the cpu loop. */
goto stop_the_world;
}
/* Check TLB entry and enforce page permissions. */
if ((addr & TARGET_PAGE_MASK)
!= (tlb_addr & (TARGET_PAGE_MASK | TLB_INVALID_MASK))) {
if (!VICTIM_TLB_HIT(addr_write, addr)) {
tlb_fill(ENV_GET_CPU(env), addr, MMU_DATA_STORE, mmu_idx, retaddr);
}
tlb_addr = tlbe->addr_write;
}
/* Notice an IO access, or a notdirty page. */
if (unlikely(tlb_addr & ~TARGET_PAGE_MASK)) {
/* There's really nothing that can be done to
support this apart from stop-the-world. */
goto stop_the_world;
}
/* Let the guest notice RMW on a write-only page. */
if (unlikely(tlbe->addr_read != tlb_addr)) {
tlb_fill(ENV_GET_CPU(env), addr, MMU_DATA_LOAD, mmu_idx, retaddr);
/* Since we don't support reads and writes to different addresses,
and we do have the proper page loaded for write, this shouldn't
ever return. But just in case, handle via stop-the-world. */
goto stop_the_world;
}
return (void *)((uintptr_t)addr + tlbe->addend);
stop_the_world:
cpu_loop_exit_atomic(ENV_GET_CPU(env), retaddr);
}
#ifdef TARGET_WORDS_BIGENDIAN
# define TGT_BE(X) (X)
# define TGT_LE(X) BSWAP(X)
#else
# define TGT_BE(X) BSWAP(X)
# define TGT_LE(X) (X)
#endif
#define MMUSUFFIX _mmu
#define DATA_SIZE 1
#include "softmmu_template.h"
#define DATA_SIZE 2
#include "softmmu_template.h"
#define DATA_SIZE 4
#include "softmmu_template.h"
#define DATA_SIZE 8
#include "softmmu_template.h"
/* First set of helpers allows passing in of OI and RETADDR. This makes
them callable from other helpers. */
#define EXTRA_ARGS , TCGMemOpIdx oi, uintptr_t retaddr
#define ATOMIC_NAME(X) \
HELPER(glue(glue(glue(atomic_ ## X, SUFFIX), END), _mmu))
#define ATOMIC_MMU_LOOKUP atomic_mmu_lookup(env, addr, oi, retaddr)
#define DATA_SIZE 1
#include "atomic_template.h"
#define DATA_SIZE 2
#include "atomic_template.h"
#define DATA_SIZE 4
#include "atomic_template.h"
#ifdef CONFIG_ATOMIC64
#define DATA_SIZE 8
#include "atomic_template.h"
#endif
#ifdef CONFIG_ATOMIC128
#define DATA_SIZE 16
#include "atomic_template.h"
#endif
/* Second set of helpers are directly callable from TCG as helpers. */
#undef EXTRA_ARGS
#undef ATOMIC_NAME
#undef ATOMIC_MMU_LOOKUP
#define EXTRA_ARGS , TCGMemOpIdx oi
#define ATOMIC_NAME(X) HELPER(glue(glue(atomic_ ## X, SUFFIX), END))
#define ATOMIC_MMU_LOOKUP atomic_mmu_lookup(env, addr, oi, GETPC())
#define DATA_SIZE 1
#include "atomic_template.h"
#define DATA_SIZE 2
#include "atomic_template.h"
#define DATA_SIZE 4
#include "atomic_template.h"
#ifdef CONFIG_ATOMIC64
#define DATA_SIZE 8
#include "atomic_template.h"
#endif
/* Code access functions. */
#undef MMUSUFFIX
#define MMUSUFFIX _cmmu
#undef GETPC
#define GETPC() ((uintptr_t)0)
#define SOFTMMU_CODE_ACCESS
#define DATA_SIZE 1
#include "softmmu_template.h"
#define DATA_SIZE 2
#include "softmmu_template.h"
#define DATA_SIZE 4
#include "softmmu_template.h"
#define DATA_SIZE 8
#include "softmmu_template.h"