Aiden Grossman 980d66caae
[llvm-exegesis] Error Out If Perf Counter is Not Fully Enabled (#132892)
Perf counters can be multiplexed if there are too many that need to be
scheduled on a core at the same time (and they exceed the available
PMUs). Other processes (especially system ones in certain environments,
not commonly on Desktop Linux from what I've seen) can also interfere.
This will impact the measurement fidelity as the counter is not actually
counting cycles/uops the entire time. This patch makes it so that we
error out in these cases so the user gets a visible indication things
have gone wrong rather than things failing silently.
2025-03-31 10:58:24 -07:00

256 lines
7.5 KiB
C++

//===-- PerfHelper.cpp ------------------------------------------*- C++ -*-===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
#include "PerfHelper.h"
#include "Error.h"
#include "llvm/Config/config.h"
#include "llvm/Support/Errc.h"
#include "llvm/Support/Error.h"
#include "llvm/Support/raw_ostream.h"
#ifdef HAVE_LIBPFM
#include <perfmon/perf_event.h>
#include <perfmon/pfmlib.h>
#include <perfmon/pfmlib_perf_event.h>
#endif
#include <cassert>
#include <cstddef>
#include <errno.h> // for erno
#include <string.h> // for strerror()
namespace llvm {
namespace exegesis {
namespace pfm {
#ifdef HAVE_LIBPFM
static bool isPfmError(int Code) { return Code != PFM_SUCCESS; }
#endif
bool pfmInitialize() {
#ifdef HAVE_LIBPFM
return isPfmError(pfm_initialize());
#else
return true;
#endif
}
void pfmTerminate() {
#ifdef HAVE_LIBPFM
pfm_terminate();
#endif
}
// Performance counters may be unavailable for a number of reasons (such as
// kernel.perf_event_paranoid restriction or CPU being unknown to libpfm).
//
// Dummy event can be specified to skip interaction with real performance
// counters while still passing control to the generated code snippet.
const char *const PerfEvent::DummyEventString = "not-really-an-event";
PerfEvent::~PerfEvent() {
#ifdef HAVE_LIBPFM
delete Attr;
;
#endif
}
PerfEvent::PerfEvent(PerfEvent &&Other)
: EventString(std::move(Other.EventString)),
FullQualifiedEventString(std::move(Other.FullQualifiedEventString)),
Attr(Other.Attr) {
Other.Attr = nullptr;
}
PerfEvent::PerfEvent(StringRef PfmEventString)
: EventString(PfmEventString.str()), Attr(nullptr) {
if (PfmEventString != DummyEventString)
initRealEvent(PfmEventString);
else
FullQualifiedEventString = PfmEventString;
}
void PerfEvent::initRealEvent(StringRef PfmEventString) {
#ifdef HAVE_LIBPFM
char *Fstr = nullptr;
pfm_perf_encode_arg_t Arg = {};
Attr = new perf_event_attr();
Arg.attr = Attr;
Arg.fstr = &Fstr;
Arg.size = sizeof(pfm_perf_encode_arg_t);
const int Result = pfm_get_os_event_encoding(EventString.c_str(), PFM_PLM3,
PFM_OS_PERF_EVENT, &Arg);
if (isPfmError(Result)) {
// We don't know beforehand which counters are available (e.g. 6 uops ports
// on Sandybridge but 8 on Haswell) so we report the missing counter without
// crashing.
errs() << pfm_strerror(Result) << " - cannot create event " << EventString
<< "\n";
}
if (Fstr) {
FullQualifiedEventString = Fstr;
free(Fstr);
}
#endif
}
StringRef PerfEvent::name() const { return EventString; }
bool PerfEvent::valid() const { return !FullQualifiedEventString.empty(); }
const perf_event_attr *PerfEvent::attribute() const { return Attr; }
StringRef PerfEvent::getPfmEventString() const {
return FullQualifiedEventString;
}
ConfiguredEvent::ConfiguredEvent(PerfEvent &&EventToConfigure)
: Event(std::move(EventToConfigure)) {
assert(Event.valid());
}
#ifdef HAVE_LIBPFM
void ConfiguredEvent::initRealEvent(const pid_t ProcessID, const int GroupFD) {
const int CPU = -1;
const uint32_t Flags = 0;
perf_event_attr AttrCopy = *Event.attribute();
AttrCopy.read_format =
PERF_FORMAT_TOTAL_TIME_ENABLED | PERF_FORMAT_TOTAL_TIME_RUNNING;
FileDescriptor = perf_event_open(&AttrCopy, ProcessID, CPU, GroupFD, Flags);
if (FileDescriptor == -1) {
errs() << "Unable to open event. ERRNO: " << strerror(errno)
<< ". Make sure your kernel allows user "
"space perf monitoring.\nYou may want to try:\n$ sudo sh "
"-c 'echo -1 > /proc/sys/kernel/perf_event_paranoid'.\n"
<< "If you are debugging and just want to execute the snippet "
"without actually reading performance counters, "
"pass --use-dummy-perf-counters command line option.\n";
}
assert(FileDescriptor != -1 && "Unable to open event");
}
Expected<SmallVector<int64_t>>
ConfiguredEvent::readOrError(StringRef /*unused*/) const {
int64_t EventInfo[3] = {0, 0, 0};
ssize_t ReadSize = ::read(FileDescriptor, &EventInfo, sizeof(EventInfo));
if (ReadSize != sizeof(EventInfo))
return make_error<StringError>("Failed to read event counter",
errc::io_error);
int64_t EventTimeEnabled = EventInfo[1];
int64_t EventTimeRunning = EventInfo[2];
if (EventTimeEnabled != EventTimeRunning)
return make_error<PerfCounterNotFullyEnabled>();
SmallVector<int64_t, 1> Result;
Result.push_back(EventInfo[0]);
return Result;
}
ConfiguredEvent::~ConfiguredEvent() { close(FileDescriptor); }
#else
void ConfiguredEvent::initRealEvent(pid_t ProcessID, const int GroupFD) {}
Expected<SmallVector<int64_t>>
ConfiguredEvent::readOrError(StringRef /*unused*/) const {
return make_error<StringError>("Not implemented",
errc::function_not_supported);
}
ConfiguredEvent::~ConfiguredEvent() = default;
#endif // HAVE_LIBPFM
CounterGroup::CounterGroup(PerfEvent &&E, std::vector<PerfEvent> &&ValEvents,
pid_t ProcessID)
: EventCounter(std::move(E)) {
IsDummyEvent = EventCounter.isDummyEvent();
for (auto &&ValEvent : ValEvents)
ValidationEventCounters.emplace_back(std::move(ValEvent));
if (!IsDummyEvent)
initRealEvent(ProcessID);
}
#ifdef HAVE_LIBPFM
void CounterGroup::initRealEvent(pid_t ProcessID) {
EventCounter.initRealEvent(ProcessID);
for (auto &ValCounter : ValidationEventCounters)
ValCounter.initRealEvent(ProcessID, getFileDescriptor());
}
void CounterGroup::start() {
if (!IsDummyEvent)
ioctl(getFileDescriptor(), PERF_EVENT_IOC_RESET, PERF_IOC_FLAG_GROUP);
}
void CounterGroup::stop() {
if (!IsDummyEvent)
ioctl(getFileDescriptor(), PERF_EVENT_IOC_DISABLE, PERF_IOC_FLAG_GROUP);
}
Expected<SmallVector<int64_t, 4>>
CounterGroup::readOrError(StringRef FunctionBytes) const {
if (!IsDummyEvent)
return EventCounter.readOrError(FunctionBytes);
else
return SmallVector<int64_t, 1>(1, 42);
}
Expected<SmallVector<int64_t>>
CounterGroup::readValidationCountersOrError() const {
SmallVector<int64_t, 4> Result;
for (const auto &ValCounter : ValidationEventCounters) {
Expected<SmallVector<int64_t>> ValueOrError =
ValCounter.readOrError(StringRef());
if (!ValueOrError)
return ValueOrError.takeError();
// Reading a validation counter will only return a single value, so it is
// safe to only append the first value here. Also assert that this is true.
assert(ValueOrError->size() == 1 &&
"Validation counters should only return a single value");
Result.push_back((*ValueOrError)[0]);
}
return Result;
}
int CounterGroup::numValues() const { return 1; }
#else
void CounterGroup::initRealEvent(pid_t ProcessID) {}
void CounterGroup::start() {}
void CounterGroup::stop() {}
Expected<SmallVector<int64_t, 4>>
CounterGroup::readOrError(StringRef /*unused*/) const {
if (IsDummyEvent) {
SmallVector<int64_t, 4> Result;
Result.push_back(42);
return Result;
}
return make_error<StringError>("Not implemented", errc::io_error);
}
Expected<SmallVector<int64_t>>
CounterGroup::readValidationCountersOrError() const {
return SmallVector<int64_t>(0);
}
int CounterGroup::numValues() const { return 1; }
#endif
} // namespace pfm
} // namespace exegesis
} // namespace llvm