Skip to content

Commit

Permalink
refine per api function
Browse files Browse the repository at this point in the history 
1. Replace the reading syscall with getting event counter registers
   directly.
2. Support arm32 architecture event counter regiters.
3. Only collect the perf data for current threads.
4. Modify the test8 to enable our new per api function.
5. Forbid per api function on ANDROID platform.
  • Loading branch information
@zhochi01
zhochi01 committed Jan 14, 2025
1 parent a1a6691 commit 2eb694b
Show file tree
Hide file tree
Showing 6 changed files with 259 additions and 281 deletions.
2 changes: 0 additions & 2 deletions collectors/collector_utility.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -2,7 +2,6 @@

// This module includes utilities for writing collectors

#include <vector>
#include <sys/stat.h>

#include "interface.hpp"
Expand All @@ -16,7 +15,6 @@ std::string getMidgardInstrOutputPath();

// Hack to workaround strange missing support for std::to_string in Android
#ifdef __ANDROID__
#include <string>
#include <sstream>

template <typename T>
Expand Down
189 changes: 85 additions & 104 deletions collectors/perf.cpp
View file
Original file line number Diff line number Diff line change
@@ -1,23 +1,18 @@
#include "perf.hpp"

#include <errno.h>
#include <stdlib.h>
#include <stdio.h>
#include <unistd.h>
#include <string.h>
#include <asm/unistd.h>
#include <sys/types.h>
#include <pthread.h>
#include <sstream>
#include <sys/ioctl.h>
#include <sys/syscall.h>
#if !defined(ANDROID)
#include <linux/perf_event.h>
#else
#include "perf_event.h"
#endif
#include <asm/unistd.h>
#include <sys/types.h>
#include <dirent.h>
#include <pthread.h>
#include <sstream>
#include <fstream>

static std::map<int, std::vector<struct event>> EVENTS = {
{0, { {"CPUInstructionRetired", PERF_TYPE_HARDWARE, PERF_COUNT_HW_INSTRUCTIONS, false, false, hw_cnt_length::b32, false},
Expand Down Expand Up @@ -79,15 +74,30 @@ static inline uint64_t makeup_booker_ci_config(int nodetype, int eventid, int by
return config;
}

PerfCollector::PerfCollector(const Json::Value& config, const std::string& name) : Collector(config, name)
PerfCollector::PerfCollector(const Json::Value& config, const std::string& name, bool enablePerapiPerf) : Collector(config, name)
{
// libcollector doesn't support any per api function on ANDROID platforms.
#if defined(ANDROID) || defined(__ANDROID__)
mEnablePerapiPerf = false;
#else
mEnablePerapiPerf = enablePerapiPerf;
if (mEnablePerapiPerf)
{
volatile uint64_t pmcr_el0;
asm volatile("mrs %0, PMCR_EL0" : "=r"(pmcr_el0));
pmu_counter_bits = ((pmcr_el0 & 0x80) == 0x80 ? 64 : 32);
DBG_LOG("pmu counter bits are: %u\n", pmu_counter_bits);
DBG_LOG("pmcr_el0 is: %lu\n", pmcr_el0);
}
#endif
struct event leader = {"CPUCycleCount", PERF_TYPE_HARDWARE, PERF_COUNT_HW_CPU_CYCLES, false, false, hw_cnt_length::b32};
bool leaderOnce = true;

mSet = mConfig.get("set", -1).asInt();
mInherit = mConfig.get("inherit", 1).asInt();

leader.inherited = mInherit;
leader.cspmu = false;
leader.device = "single";

if ((0 <= mSet) && (mSet <= 3))
Expand Down Expand Up @@ -206,7 +216,7 @@ PerfCollector::PerfCollector(const Json::Value& config, const std::string& name)
}
}

mAllThread = mConfig.get("allthread", true).asBool();
mAllThread = mConfig.get("allthread", !mEnablePerapiPerf).asBool();
}

static inline long perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
Expand Down Expand Up @@ -249,9 +259,12 @@ static int add_event(const struct event &e, int tid, int cpu, int group = -1)
bool PerfCollector::init()
{
create_perf_thread();

for (perf_thread& t : mReplayThreads)
{
if (mEnablePerapiPerf)
{
t.eventCtx.setEnablePerApi();
}
t.eventCtx.init(mEvents[t.device_name], t.tid, -1);
}

Expand Down Expand Up @@ -415,81 +428,31 @@ bool PerfCollector::collect(int64_t now)
return true;
}

bool PerfCollector::perf_counter_pause() {
#if defined(__aarch64__)
asm volatile("mrs %0, PMCNTENSET_EL0" : "=r" (PMCNTENSET_EL0_safe));
// stop counters for arm64
asm volatile("mrs %0, PMCR_EL0" : "=r" (PMCR_EL0_safe));
asm volatile("msr PMCR_EL0, %0" : : "r" (PMCR_EL0_safe & 0xFFFFFFFFFFFFFFFE));
#elif defined(__arm__)
asm volatile("mrc p15, 0, %0, c9, c12, 1" : "=r"(PMCNTENSET_EL0_safe));
// stop counters for arm32
asm volatile("mrc p15, 0, %0, c9, c12, 0" : "=r"(PMCR_EL0_safe));
asm volatile("mcr p15, 0, %0, c9, c12, 0" : : "r"(PMCR_EL0_safe & 0xFFFFFFFE));
#endif
return true;
}

bool PerfCollector::perf_counter_resume() {
#if defined(__aarch64__)
// start counters for arm64
asm volatile("msr PMCNTENSET_EL0, %0" : : "r" (PMCNTENSET_EL0_safe));
asm volatile("msr PMCR_EL0, %0" : : "r" (PMCR_EL0_safe));
#elif defined(__arm__)
// start counters for arm32
asm volatile("mcr p15, 0, %0, c9, c12, 1" : : "r"(PMCNTENSET_EL0_safe));
asm volatile("mcr p15, 0, %0, c9, c12, 0" : : "r"(PMCR_EL0_safe));
#endif
return true;
}


bool PerfCollector::collect_scope_start(int64_t now, uint16_t func_id, int32_t flags) {
bool PerfCollector::collect_scope_start(uint16_t func_id, int32_t flags, int tid) {
#if defined(__x86_64__)
if (!attempt_collect_scope_x64) {
attempt_collect_scope_x64 = true;
DBG_LOG("WARNING: Frequent invocation of collect_scope on x64 devices may introduce "
"significant overhead to the kernel perf counter data.\n");
}
#endif
if (!perf_counter_pause()) return false;
if (!mCollecting) return false;
struct snapshot snap;
if (flags & COLLECT_REPLAY_THREADS || flags & COLLECT_ALL_THREADS)
{
for (perf_thread &t : mReplayThreads)
{
t.eventCtx.collect_scope(now, func_id, false);
}
}
if (flags & COLLECT_BG_THREADS || flags & COLLECT_ALL_THREADS)
{
for (perf_thread &t : mBgThreads)
{
t.eventCtx.collect_scope(now, func_id, false);
}
}
if (flags & COLLECT_BOOKER_THREADS || flags & COLLECT_ALL_THREADS)
{
for (perf_thread &t : mBookerThread)
{
t.eventCtx.collect_scope(now, func_id, false);
}
}
if (flags & COLLECT_CSPMU_THREADS || flags & COLLECT_ALL_THREADS)
{
for (perf_thread &t : mCSPMUThreads)
for (auto &thread: mReplayThreads)
{
t.eventCtx.collect_scope(now, func_id, false);
}
if (thread.tid == tid)
{
thread.eventCtx.collect_scope(func_id, false, get_pmu_bits());
}
}
}
last_collect_scope_flags = flags;
if (!perf_counter_resume()) return false;
return true;
}

bool PerfCollector::collect_scope_stop(int64_t now, uint16_t func_id, int32_t flags) {
if (!perf_counter_pause()) return false;
bool PerfCollector::collect_scope_stop(uint16_t func_id, int32_t flags, int tid) {
if (!mCollecting) return false;
if (last_collect_scope_flags != flags) {
DBG_LOG("Error: Could not find the corresponding collect_scope_start call for func_id %ud.\n", func_id);
Expand All @@ -498,42 +461,17 @@ bool PerfCollector::collect_scope_stop(int64_t now, uint16_t func_id, int32_t fl
struct snapshot snap_start, snap_stop;
if (flags & COLLECT_REPLAY_THREADS || flags & COLLECT_ALL_THREADS)
{
for (perf_thread &t : mReplayThreads)
{
snap_start = t.eventCtx.last_snap;
snap_stop = t.eventCtx.collect_scope(now, func_id, true);
t.update_data_scope(func_id, snap_start, snap_stop);
}
}
if (flags & COLLECT_BG_THREADS || flags & COLLECT_ALL_THREADS)
{
for (perf_thread &t : mBgThreads)
{
snap_start = t.eventCtx.last_snap;
snap_stop = t.eventCtx.collect_scope(now, func_id, true);
t.update_data_scope(func_id, snap_start, snap_stop);
}
}
if (flags & COLLECT_BOOKER_THREADS || flags & COLLECT_ALL_THREADS)
{
for (perf_thread &t : mBookerThread)
{
snap_start = t.eventCtx.last_snap;
snap_stop = t.eventCtx.collect_scope(now, func_id, true);
t.update_data_scope(func_id, snap_start, snap_stop);
}
}
if (flags & COLLECT_CSPMU_THREADS || flags & COLLECT_ALL_THREADS)
{
for (perf_thread &t : mCSPMUThreads)
for (auto &thread: mReplayThreads)
{
snap_start = t.eventCtx.last_snap;
snap_stop = t.eventCtx.collect_scope(now, func_id, true);
t.update_data_scope(func_id, snap_start, snap_stop);
if (thread.tid == tid)
{
snap_start = thread.eventCtx.last_snap;
snap_stop = thread.eventCtx.collect_scope(func_id, true, get_pmu_bits());
thread.update_data_scope(func_id, snap_start, snap_stop);
}
}
}
if (!perf_counter_resume()) return false;
return false;
return true;
}

bool PerfCollector::postprocess(const std::vector<int64_t>& timing)
Expand Down Expand Up @@ -672,6 +610,15 @@ bool event_context::deinit()
return true;
}


#define F_BIT_0 ((uint32_t)0x00000001)
#define F_BIT_2 ((uint32_t)0x00000004)
#define F_BIT_3 ((uint32_t)0x00000008)
#define CINSTRP_ARMV8_PMCR_E ((unsigned long long)F_BIT_0) /* Enable all counters */
#define CINSTRP_ARMV8_PMCR_C ((unsigned long long)F_BIT_2) /* Cycle counter reset */
#define CINSTRP_ARMV8_PMCR_R ((unsigned long long)F_BIT_3) /* Cycle counter reset */


bool event_context::start()
{
if (ioctl(group, PERF_EVENT_IOC_RESET, PERF_IOC_FLAG_GROUP) == -1)
Expand All @@ -684,6 +631,24 @@ bool event_context::start()
perror("ioctl PERF_EVENT_IOC_ENABLE");
return false;
}

#if !defined(ANDROID) && !defined(__ANDROID__)
if (getEnablePerApi())
{
volatile uint64_t el0_access = 0;
#if defined(__aarch64__)
asm volatile("mrs %0, PMUSERENR_EL0" : "=r"(el0_access));
#elif defined(__arm__)
asm volatile("mrc p15, 0, %0, c9, c14, 0" : "=r"(el0_access));
#endif
if ((el0_access & (CINSTRP_ARMV8_PMCR_E | CINSTRP_ARMV8_PMCR_C | CINSTRP_ARMV8_PMCR_R)) != (CINSTRP_ARMV8_PMCR_E | CINSTRP_ARMV8_PMCR_C | CINSTRP_ARMV8_PMCR_R))
{
DBG_LOG("EL0 access to PMU is required! Please set the appropriate bits in PMUSERENR_EL0. Current settings: %08x\n", (uint32_t)el0_access);
exit(EXIT_FAILURE);
}
}
#endif

return true;
}

Expand Down Expand Up @@ -732,13 +697,29 @@ struct snapshot event_context::collect(int64_t now)
return snap;
}

struct snapshot event_context::collect_scope(int64_t now, uint16_t func_id, bool stopping)
struct snapshot event_context::collect_scope(uint16_t func_id, bool stopping, uint8_t pmu_bits)
{
if (stopping && last_snap_func_id != func_id) {
DBG_LOG("Error: Could not find the corresponding collect_scope_start call for func_id %ud.\n", func_id);
exit(EXIT_FAILURE);
}
struct snapshot snap;
#if defined(__aarch64__)
if (pmu_bits == 32)
{
asm volatile("mrs %0, PMCCNTR_EL0" : "=r"(snap.values[0]));
}
else
{
asm volatile("mrs %0, PMEVCNTR2_EL0" : "=r"(snap.values[0]));
}
#elif defined(__arm__)
volatile uint32_t PMCCNTR_EL0_lo, PMCCNTR_EL0_hi;
asm volatile("mrrc p15, 0, %0, %1, c9" : "=r"(PMCCNTR_EL0_lo), "=r"(PMCCNTR_EL0_hi));
snap.values[0] = (((uint64_t)PMCCNTR_EL0_hi) << 32) | ((uint64_t)PMCCNTR_EL0_lo);
#else
if (read(group, &snap, sizeof(snap)) == -1) perror("read");
#endif
if (stopping) {
last_snap_func_id = -1;
} else {
Expand Down
Loading

0 comments on commit 2eb694b

Please sign in to comment.