Skip to contentSkip to conversation

Instrumentation and TLB stats

PATCH In review Pranith Kumar 4 participants 6 messages Jun 28, 2017 – Jul 10, 2017

lore
Share & tools
lore
#0 lore raw

cover Instrumentation and TLB stats

The following two patches are what I use to instrument guest code and collect TLB hit/miss information. These patches are for informational and discussion purposes only.

Pranith Kumar (2):
[TEST] aarch64: Use pmuserenr_el0 register for instrumentation [TEST] Collect TLB stats along with victim TLB

 accel/tcg/cputlb.c        | 12 ++++++++++++
 cpus.c                    | 26 ++++++++++++++++++++++++++
 include/exec/cpu-defs.h   |  4 ++++
 include/sysemu/cpus.h     |  2 ++
 target/arm/helper.c       | 23 +++++++++++++++++++++--
 tcg/i386/tcg-target.inc.c | 16 ++++++++++++++--
 vl.c                      |  3 +++
 7 files changed, 82 insertions(+), 4 deletions(-)
Signature · 1 line
2.13.0
#1 lore raw

PATCH 1/2 aarch64: Use pmuserenr_el0 register for instrumentation

We need a way for the benchmark running in the guest to indicate us to start/stop our instrumentation. On x86, we could use the 'cpuid' instruction along with an appropriately populated 'eax' register. However, no such dummy instruction exists for aarch64. So we modify the permission bits for 'pmuserenr_el0' register and tap that to instrument the guest code.

You can use the following annotations on your region-of-interest to instrument the code.

#define magic_enable() \
asm volatile ("msr pmuserenr_el0, %0" :: "r" (0xaaaaaaaa));
#define magic_disable() \
asm volatile ("msr pmuserenr_el0, %0" :: "r" (0xfa11dead));

---
 target/arm/helper.c | 19 +++++++++++++++++--
 1 file changed, 17 insertions(+), 2 deletions(-)
target/arm/helper.c +17 -2
index 2594faa9b8..dfbf03676c 100644
@@ -1124,9 +1124,24 @@ static uint64_t pmxevtyper_read(CPUARMState *env, const ARMCPRegInfo *ri)
     }
 }
 
+bool enable_instrumentation;
+
 static void pmuserenr_write(CPUARMState *env, const ARMCPRegInfo *ri,
                             uint64_t value)
 {
+    ARMCPU *cpu = arm_env_get_cpu(env);
+    CPUState *cs = CPU(cpu);
+
+    if (value == 0xaaaaaaaa) {
+        printf("Enabling instrumentation\n");
+        enable_instrumentation = true;
+        tb_flush(cs);
+    } else if (value == 0xfa11dead) {
+        printf("Disabling instrumentation\n");
+        enable_instrumentation = false;
+        tb_flush(cs);
+    }
+
     if (arm_feature(env, ARM_FEATURE_V8)) {
         env->cp15.c9_pmuserenr = value & 0xf;
     } else {
@@ -1316,13 +1331,13 @@ static const ARMCPRegInfo v7_cp_reginfo[] = {
       .access = PL0_RW, .type = ARM_CP_CONST, .resetvalue = 0,
       .accessfn = pmreg_access_xevcntr },
     { .name = "PMUSERENR", .cp = 15, .crn = 9, .crm = 14, .opc1 = 0, .opc2 = 0,
-      .access = PL0_R | PL1_RW, .accessfn = access_tpm,
+      .access = PL0_RW | PL1_RW, .accessfn = access_tpm,
       .fieldoffset = offsetof(CPUARMState, cp15.c9_pmuserenr),
       .resetvalue = 0,
       .writefn = pmuserenr_write, .raw_writefn = raw_write },
     { .name = "PMUSERENR_EL0", .state = ARM_CP_STATE_AA64,
       .opc0 = 3, .opc1 = 3, .crn = 9, .crm = 14, .opc2 = 0,
-      .access = PL0_R | PL1_RW, .accessfn = access_tpm, .type = ARM_CP_ALIAS,
+      .access = PL0_RW | PL1_RW, .accessfn = access_tpm, .type = ARM_CP_ALIAS,
       .fieldoffset = offsetof(CPUARMState, cp15.c9_pmuserenr),
       .resetvalue = 0,
       .writefn = pmuserenr_write, .raw_writefn = raw_write },
Signature · 1 line
2.13.0
#2 lore raw

Re: [Qemu-devel] [PATCH 1/2] [TEST] aarch64: Use pmuserenr_el0 register for instrumentation

Quoted 6 lines · Pranith Kumar

We need a way for the benchmark running in the guest to indicate us to start/stop our instrumentation. On x86, we could use the 'cpuid' instruction along with an appropriately populated 'eax' register. However, no such dummy instruction exists for aarch64. So we modify the permission bits for 'pmuserenr_el0' register and tap that to instrument the guest code.

Why not just define a QEMU-specific system register in the reserved-for-implementation-defined-registers encoding range, rather than modifying the behaviour of an architecturally defined register?

(Though overall I feel we should be able to do better as a mechanism.)

thanks
-- PMM

#3 lore raw

PATCH 2/2 Collect TLB and victim TLB hit/miss stats

In reply to Pranith Kumar

I used the following patch to collect hit/miss TLB ratios for a few benchmarks. The results can be found here: http://imgur.com/a/gee1o

Please note that these results also include boot/shutdown as the per-region instrumentation patch came later.

---
 accel/tcg/cputlb.c        | 12 ++++++++++++
 cpus.c                    | 26 ++++++++++++++++++++++++++
 include/exec/cpu-defs.h   |  4 ++++
 include/sysemu/cpus.h     |  2 ++
 target/arm/helper.c       |  6 +++++-
 tcg/i386/tcg-target.inc.c | 16 ++++++++++++++--
 vl.c                      |  3 +++
 7 files changed, 66 insertions(+), 3 deletions(-)
accel/tcg/cputlb.c +12 -0
index ef52a7e5e0..2ac2397431 100644
@@ -864,12 +864,19 @@ static void io_writex(CPUArchState *env, CPUIOTLBEntry *iotlbentry,
     }
 }
 
+extern bool enable_instrumentation;
+
 /* 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;
+
+    if (enable_instrumentation) {
+        env->tlb_access_victim++;
+    }
+
     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);
@@ -885,6 +892,11 @@ static bool victim_tlb_hit(CPUArchState *env, size_t mmu_idx, size_t index,
             CPUIOTLBEntry tmpio, *io = &env->iotlb[mmu_idx][index];
             CPUIOTLBEntry *vio = &env->iotlb_v[mmu_idx][vidx];
             tmpio = *io; *io = *vio; *vio = tmpio;
+
+            if (enable_instrumentation) {
+                env->tlb_access_victim_hit++;
+            }
+
             return true;
         }
     }
cpus.c +26 -0
index 14bb8d552e..14669b3469 100644
@@ -1602,6 +1602,32 @@ static bool all_vcpus_paused(void)
     return true;
 }
 
+void print_tlb_stats(void)
+{
+    CPUState *cpu;
+    CPU_FOREACH(cpu) {
+        CPUArchState *cs = cpu->env_ptr;
+
+        fprintf(stderr, "TLB accesses %lu, hits %lu, victim accesses %lu, hits %lu\n",
+                cs->tlb_access_total, cs->tlb_access_hit, cs->tlb_access_victim,
+                cs->tlb_access_victim_hit);
+    }
+}
+
+void clear_tlb_stats(void)
+{
+    CPUState *cpu;
+    CPU_FOREACH(cpu) {
+        CPUArchState *cs = cpu->env_ptr;
+
+        cs->tlb_access_total        = 0;
+        cs->tlb_access_hit          = 0;
+        cs->tlb_access_victim       = 0;
+        cs->tlb_access_victim       = 0;
+        cs->tlb_access_victim_hit   = 0;
+    }
+}
+
 void pause_all_vcpus(void)
 {
     CPUState *cpu;
include/exec/cpu-defs.h +4 -0
index 5f4e303635..29b3c2ada8 100644
@@ -138,6 +138,10 @@ typedef struct CPUIOTLBEntry {
     target_ulong tlb_flush_addr;                                        \
     target_ulong tlb_flush_mask;                                        \
     target_ulong vtlb_index;                                            \
+    target_ulong tlb_access_hit;                                        \
+    target_ulong tlb_access_total;                                      \
+    target_ulong tlb_access_victim;                                     \
+    target_ulong tlb_access_victim_hit;                                 \
 
 #else
 
include/sysemu/cpus.h +2 -0
index 731756d948..7d8d92646c 100644
@@ -10,6 +10,8 @@ void resume_all_vcpus(void);
 void pause_all_vcpus(void);
 void cpu_stop_current(void);
 void cpu_ticks_init(void);
+void print_tlb_stats(void);
+void clear_tlb_stats(void);
 
 void configure_icount(QemuOpts *opts, Error **errp);
 extern int use_icount;
target/arm/helper.c +5 -1
index dfbf03676c..d2e75b0f20 100644
@@ -1124,7 +1124,9 @@ static uint64_t pmxevtyper_read(CPUARMState *env, const ARMCPRegInfo *ri)
     }
 }
 
-bool enable_instrumentation;
+extern bool enable_instrumentation;
+extern void print_tlb_stats(void);
+extern void clear_tlb_stats(void);
 
 static void pmuserenr_write(CPUARMState *env, const ARMCPRegInfo *ri,
                             uint64_t value)
@@ -1139,6 +1141,8 @@ static void pmuserenr_write(CPUARMState *env, const ARMCPRegInfo *ri,
     } else if (value == 0xfa11dead) {
         printf("Disabling instrumentation\n");
         enable_instrumentation = false;
+        print_tlb_stats();
+        clear_tlb_stats();
         tb_flush(cs);
     }
 
tcg/i386/tcg-target.inc.c +14 -2
index 9d7d25c017..b75bd54c35 100644
@@ -1250,6 +1250,8 @@ static void * const qemu_st_helpers[16] = {
     [MO_BEQ]  = helper_be_stq_mmu,
 };
 
+extern bool enable_instrumentation;
+
 /* Perform the TLB load and compare.
 
    Inputs:
@@ -1300,6 +1302,12 @@ static inline void tcg_out_tlb_load(TCGContext *s, TCGReg addrlo, TCGReg addrhi,
         }
     }
 
+    if (enable_instrumentation) {
+        tcg_out_ld(s, TCG_TYPE_I64, r0, TCG_AREG0, offsetof(CPUArchState, tlb_access_total));
+        tcg_out_addi(s, r0, 1);
+        tcg_out_st(s, TCG_TYPE_I64, r0, TCG_AREG0, offsetof(CPUArchState, tlb_access_total));
+    }
+
     tcg_out_mov(s, tlbtype, r0, addrlo);
     tlb_mask = (target_ulong)TARGET_PAGE_MASK | a_mask;
 
@@ -1348,11 +1356,15 @@ static inline void tcg_out_tlb_load(TCGContext *s, TCGReg addrlo, TCGReg addrhi,
         s->code_ptr += 4;
     }
 
-    /* TLB Hit.  */
-
     /* add addend(r0), r1 */
     tcg_out_modrm_offset(s, OPC_ADD_GvEv + hrexw, r1, r0,
                          offsetof(CPUTLBEntry, addend) - which);
+
+    if (enable_instrumentation) {
+        tcg_out_ld(s, TCG_TYPE_I64, r0, TCG_AREG0, offsetof(CPUArchState, tlb_access_hit));
+        tcg_out_addi(s, r0, 1);
+        tcg_out_st(s, TCG_TYPE_I64, r0, TCG_AREG0, offsetof(CPUArchState, tlb_access_hit));
+    }
 }
 
 /*
vl.c +3 -0
index 59fea15488..7fa392c79e 100644
@@ -192,6 +192,8 @@ int only_migratable; /* turn it off unless user states otherwise */
 
 int icount_align_option;
 
+bool enable_instrumentation;
+
 /* The bytes in qemu_uuid are in the order specified by RFC4122, _not_ in the
  * little-endian "wire format" described in the SMBIOS 2.6 specification.
  */
@@ -4761,5 +4763,6 @@ int main(int argc, char **argv, char **envp)
     qemu_chr_cleanup();
     /* TODO: unref root container, check all devices are ok */
 
+    print_tlb_stats();
     return 0;
 }
Signature · 1 line
2.13.0
#4 lore raw

Re: [Qemu-devel] [PATCH 2/2] [TEST] Collect TLB and victim TLB hit/miss stats

Quoted 22 lines · Pranith Kumar

I used the following patch to collect hit/miss TLB ratios for a few benchmarks. The results can be found here: http://imgur.com/a/gee1o

Please note that these results also include boot/shutdown as the per-region instrumentation patch came later.

---
 accel/tcg/cputlb.c        | 12 ++++++++++++
 cpus.c                    | 26 ++++++++++++++++++++++++++
 include/exec/cpu-defs.h   |  4 ++++
 include/sysemu/cpus.h     |  2 ++
 target/arm/helper.c       |  6 +++++-
 tcg/i386/tcg-target.inc.c | 16 ++++++++++++++--
 vl.c                      |  3 +++
 7 files changed, 66 insertions(+), 3 deletions(-)
accel/tcg/cputlb.c +0 -0
index ef52a7e5e0..2ac2397431 100644

accel/tcg/cputlb.c static void io_writex(CPUArchState *env, CPUIOTLBEntry *iotlbentry, · line 868

Quoted 5 lines
accel/tcg/cputlb.c +2 -0
@@ -864,12 +864,19 @@ static void io_writex(CPUArchState *env, CPUIOTLBEntry *iotlbentry,
     }
 }
 
+extern bool enable_instrumentation;
+

Is there a better place for this than a static global? I was pondering tcg_ctx but that's not really visible to the runtime. Making it part of the TB flags might be useful for only instrumenting certain segments of the code but I suspect I'm bike-shedding at this point.

Quoted 46 lines
accel/tcg/cputlb.c +10 -0
 /* 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;
+
+    if (enable_instrumentation) {
+        env->tlb_access_victim++;
+    }
+
     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);
@@ -885,6 +892,11 @@ static bool victim_tlb_hit(CPUArchState *env, size_t mmu_idx, size_t index,
             CPUIOTLBEntry tmpio, *io = &env->iotlb[mmu_idx][index];
             CPUIOTLBEntry *vio = &env->iotlb_v[mmu_idx][vidx];
             tmpio = *io; *io = *vio; *vio = tmpio;
+
+            if (enable_instrumentation) {
+                env->tlb_access_victim_hit++;
+            }
+
             return true;
         }
     }
cpus.c +12 -0
index 14bb8d552e..14669b3469 100644
@@ -1602,6 +1602,32 @@ static bool all_vcpus_paused(void)
     return true;
 }
 
+void print_tlb_stats(void)
+{
+    CPUState *cpu;
+    CPU_FOREACH(cpu) {
+        CPUArchState *cs = cpu->env_ptr;
+
+        fprintf(stderr, "TLB accesses %lu, hits %lu, victim accesses %lu, hits %lu\n",
+                cs->tlb_access_total, cs->tlb_access_hit, cs->tlb_access_victim,
+                cs->tlb_access_victim_hit);
+    }
+}
+

cpus.c static bool all_vcpus_paused(void) · line 1626

Quoted 10 lines
cpus.c +10 -0
+void clear_tlb_stats(void)
+{
+    CPUState *cpu;
+    CPU_FOREACH(cpu) {
+        CPUArchState *cs = cpu->env_ptr;
+
+        cs->tlb_access_total        = 0;
+        cs->tlb_access_hit          = 0;
+        cs->tlb_access_victim       = 0;
+        cs->tlb_access_victim       = 0;

Duplicate line here.

Quoted 51 lines
cpus.c +4 -0
+        cs->tlb_access_victim_hit   = 0;
+    }
+}
+
 void pause_all_vcpus(void)
 {
     CPUState *cpu;
include/exec/cpu-defs.h +4 -0
index 5f4e303635..29b3c2ada8 100644
@@ -138,6 +138,10 @@ typedef struct CPUIOTLBEntry {
     target_ulong tlb_flush_addr;                                        \
     target_ulong tlb_flush_mask;                                        \
     target_ulong vtlb_index;                                            \
+    target_ulong tlb_access_hit;                                        \
+    target_ulong tlb_access_total;                                      \
+    target_ulong tlb_access_victim;                                     \
+    target_ulong tlb_access_victim_hit;                                 \
 
 #else
 
include/sysemu/cpus.h +2 -0
index 731756d948..7d8d92646c 100644
@@ -10,6 +10,8 @@ void resume_all_vcpus(void);
 void pause_all_vcpus(void);
 void cpu_stop_current(void);
 void cpu_ticks_init(void);
+void print_tlb_stats(void);
+void clear_tlb_stats(void);
 
 void configure_icount(QemuOpts *opts, Error **errp);
 extern int use_icount;
target/arm/helper.c +3 -1
index dfbf03676c..d2e75b0f20 100644
@@ -1124,7 +1124,9 @@ static uint64_t pmxevtyper_read(CPUARMState *env, const ARMCPRegInfo *ri)
     }
 }
 
-bool enable_instrumentation;
+extern bool enable_instrumentation;
+extern void print_tlb_stats(void);
+extern void clear_tlb_stats(void);
 
 static void pmuserenr_write(CPUARMState *env, const ARMCPRegInfo *ri,
                             uint64_t value)

target/arm/helper.c static void pmuserenr_write(CPUARMState *env, const ARMCPRegInfo *ri, · line 1147

Quoted 7 lines
target/arm/helper.c +2 -0
@@ -1139,6 +1141,8 @@ static void pmuserenr_write(CPUARMState *env, const ARMCPRegInfo *ri,
     } else if (value == 0xfa11dead) {
         printf("Disabling instrumentation\n");
         enable_instrumentation = false;
+        print_tlb_stats();
+        clear_tlb_stats();
         tb_flush(cs);
     }

This needs to be part of the cputlb API so only one call is made from the architecture helpers. I would expect this patch to be the first and the pmuserenr_el0 (or whatever else) to be a per-arch enhancement patch on top.

Quoted 15 lines
tcg/i386/tcg-target.inc.c +2 -0
index 9d7d25c017..b75bd54c35 100644
@@ -1250,6 +1250,8 @@ static void * const qemu_st_helpers[16] = {
     [MO_BEQ]  = helper_be_stq_mmu,
 };
 
+extern bool enable_instrumentation;
+
 /* Perform the TLB load and compare.
 
    Inputs:

tcg/i386/tcg-target.inc.c static inline void tcg_out_tlb_load(TCGContext *s, TCGReg addrlo, TCGReg addrhi, · line 1305

Quoted 4 lines
tcg/i386/tcg-target.inc.c +1 -0
@@ -1300,6 +1302,12 @@ static inline void tcg_out_tlb_load(TCGContext *s, TCGReg addrlo, TCGReg addrhi,
         }
     }
 
+    if (enable_instrumentation) {

Certainly inside the code generation I'd see this being controlled by TCGContext, e.g. s->tlb_instruction

Quoted 9 lines
tcg/i386/tcg-target.inc.c +5 -0
+        tcg_out_ld(s, TCG_TYPE_I64, r0, TCG_AREG0, offsetof(CPUArchState, tlb_access_total));
+        tcg_out_addi(s, r0, 1);
+        tcg_out_st(s, TCG_TYPE_I64, r0, TCG_AREG0, offsetof(CPUArchState, tlb_access_total));
+    }
+
     tcg_out_mov(s, tlbtype, r0, addrlo);
     tlb_mask = (target_ulong)TARGET_PAGE_MASK | a_mask;
 

tcg/i386/tcg-target.inc.c static inline void tcg_out_tlb_load(TCGContext *s, TCGReg addrlo, TCGReg addrhi, · line 1352

Quoted 5 lines
tcg/i386/tcg-target.inc.c +0 -2
@@ -1348,11 +1356,15 @@ static inline void tcg_out_tlb_load(TCGContext *s, TCGReg addrlo, TCGReg addrhi,
         s->code_ptr += 4;
     }
 
-    /* TLB Hit.  */
-

why drop this comment?

Quoted 26 lines
tcg/i386/tcg-target.inc.c +6 -0
     /* add addend(r0), r1 */
     tcg_out_modrm_offset(s, OPC_ADD_GvEv + hrexw, r1, r0,
                          offsetof(CPUTLBEntry, addend) - which);
+
+    if (enable_instrumentation) {
+        tcg_out_ld(s, TCG_TYPE_I64, r0, TCG_AREG0, offsetof(CPUArchState, tlb_access_hit));
+        tcg_out_addi(s, r0, 1);
+        tcg_out_st(s, TCG_TYPE_I64, r0, TCG_AREG0, offsetof(CPUArchState, tlb_access_hit));
+    }
 }
 
 /*
vl.c +2 -0
index 59fea15488..7fa392c79e 100644
@@ -192,6 +192,8 @@ int only_migratable; /* turn it off unless user states otherwise */
 
 int icount_align_option;
 
+bool enable_instrumentation;
+
 /* The bytes in qemu_uuid are in the order specified by RFC4122, _not_ in the
  * little-endian "wire format" described in the SMBIOS 2.6 specification.
  */

vl.c int main(int argc, char **argv, char **envp) · line 4768

Quoted 6 lines
vl.c +1 -0
@@ -4761,5 +4763,6 @@ int main(int argc, char **argv, char **envp)
     qemu_chr_cleanup();
     /* TODO: unref root container, check all devices are ok */
 
+    print_tlb_stats();
     return 0;
 }

I appreciate this is currently test code for gathering numbers but it would be nice to see if there is a nice way to integrate it upstream (maybe for --enable-debug-tcg builds).

--
Alex Bennée

#5 lore raw

Re: [Qemu-devel] [PATCH 2/2] [TEST] Collect TLB and victim TLB hit/miss stats

Quoted 6 lines · Alex Bennée

Review

Quoted 2 lines
(quoted diff) +2 -0
+extern bool enable_instrumentation;
+

Is there a better place for this than a static global? I was pondering tcg_ctx but that's not really visible to the runtime. Making it part of the TB flags might be useful for only instrumenting certain segments of the code but I suspect I'm bike-shedding at this point.

Why not use tracepoints? This seems to be a natural candidate for systemtap-like tracing scripts.

Paolo