Files
sousa-gecko/tools/profiler/gecko/ProfilerChild.cpp
T
Damien Carver 3cf8d2d46c Bug 2033816 - Part 2: Remove the CPUUtilization profiler feature and always gather thread CPU. r=canova,profiler-reviewers,jdescottes
Thread CPU utilization is now always recorded, even when periodic stack
sampling is disabled. Leaving it off only produced misleading profiles, where
idle threads appeared busy because we couldn't tell they were idle.

Remove the "cpu" (CPUUtilization) feature entirely from ProfilerState.h and
BaseProfilerState.h (renumbering the following features), the WebExtensions
geckoProfiler schema, and the default feature sets, and drop the conditionals
that only recorded CPU when the feature was enabled. Per-thread CPU recording
is still driven by the selected threads and the CPUAllThreads feature.

Since the sampler now always completes a sampling pass, the
SamplingState::NoStackSamplingCompleted state is unreachable; remove it (its
consumers already treated it identically to SamplingCompleted).

Update the gtests accordingly. Also update browser_test_feature_nostacksampling.js:
because CPU is now always gathered, the sampler still records samples while the
No Stack Sampling feature is on, but those samples carry no captured stack (only
the synthetic root frame). Assert on captured stack depth rather than sample
count.

Drop the now-removed "cpu" feature from the remaining test harnesses that still
requested it, which caused the previous backouts: the xpcshell harness'
default MOZ_PROFILER_STARTUP_FEATURES set (it made every launched binary print
an "Unrecognized feature" warning, breaking test_install_hash.js) and the wpt
webdriver bidi profiler start/stop tests (stop asserted every requested feature
is reflected in the saved profile).

Differential Revision: https://phabricator.services.mozilla.com/D306736
2026-07-03 10:13:57 +00:00

603 lines
23 KiB
C++

/* This Source Code Form is subject to the terms of the Mozilla Public
* License, v. 2.0. If a copy of the MPL was not distributed with this
* file, You can obtain one at http://mozilla.org/MPL/2.0/. */
#include "ProfilerChild.h"
#include "GeckoProfiler.h"
#include "platform.h"
#include "ProfilerCodeAddressService.h"
#include "ProfilerControl.h"
#include "ProfilerParent.h"
#include "chrome/common/ipc_channel.h"
#include "nsPrintfCString.h"
#include "nsThreadUtils.h"
#include <memory>
namespace mozilla {
MOZ_RUNINIT /* static */ DataMutexBase<ProfilerChild::ProfilerChildAndUpdate,
baseprofiler::detail::BaseProfilerMutex>
ProfilerChild::sPendingChunkManagerUpdate{
"ProfilerChild::sPendingChunkManagerUpdate"};
ProfilerChild::ProfilerChild()
: mThread(NS_GetCurrentThread()), mDestroyed(false) {
MOZ_COUNT_CTOR(ProfilerChild);
}
ProfilerChild::~ProfilerChild() { MOZ_COUNT_DTOR(ProfilerChild); }
void ProfilerChild::ResolveChunkUpdate(
PProfilerChild::AwaitNextChunkManagerUpdateResolver& aResolve) {
MOZ_ASSERT(!!aResolve,
"ResolveChunkUpdate should only be called when there's a pending "
"resolver");
MOZ_ASSERT(
!mChunkManagerUpdate.IsNotUpdate(),
"ResolveChunkUpdate should only be called with a real or final update");
MOZ_ASSERT(
!mDestroyed,
"ResolveChunkUpdate should not be called if the actor was destroyed");
if (mChunkManagerUpdate.IsFinal()) {
// Final update, send a special "unreleased value", but don't clear the
// local copy so we know we got the final update.
std::move(aResolve)(ProfilerParent::MakeFinalUpdate());
} else {
// Optimization note: The ProfileBufferChunkManagerUpdate constructor takes
// the newly-released chunks nsTArray by reference-to-const, therefore
// constructing and then moving the array here would make a copy. So instead
// we first give it an empty array, and then we can write the data directly
// into the update's array.
ProfileBufferChunkManagerUpdate update{
mChunkManagerUpdate.UnreleasedBytes(),
mChunkManagerUpdate.ReleasedBytes(),
mChunkManagerUpdate.OldestDoneTimeStamp(),
{}};
update.newlyReleasedChunks().SetCapacity(
mChunkManagerUpdate.NewlyReleasedChunksRef().size());
for (const ProfileBufferControlledChunkManager::ChunkMetadata& chunk :
mChunkManagerUpdate.NewlyReleasedChunksRef()) {
update.newlyReleasedChunks().EmplaceBack(chunk.mDoneTimeStamp,
chunk.mBufferBytes);
}
std::move(aResolve)(update);
// Clear the update we just sent, so it's ready for later updates to be
// folded into it.
mChunkManagerUpdate.Clear();
}
// Discard the resolver, so it's empty next time there's a new request.
aResolve = nullptr;
}
void ProfilerChild::ProcessChunkManagerUpdate(
ProfileBufferControlledChunkManager::Update&& aUpdate) {
if (mDestroyed) {
return;
}
// Always store the data, it could be the final update.
mChunkManagerUpdate.Fold(std::move(aUpdate));
if (mAwaitNextChunkManagerUpdateResolver) {
// There is already a pending resolver, give it the info now.
ResolveChunkUpdate(mAwaitNextChunkManagerUpdateResolver);
}
}
/* static */ void ProfilerChild::ProcessPendingUpdate() {
// sPendingChunkManagerUpdate must not be held while dispatching or while
// processing the update: Dispatch() takes the target thread's event queue
// mutex, and ProcessChunkManagerUpdate() can resolve an await-next-update
// request, which sends an IPC reply and takes the IPC channel lock. Both of
// those locks can be held while a profiler marker is recorded (which takes
// the profiler buffer lock), and that buffer lock is held when the chunk
// manager update callback takes sPendingChunkManagerUpdate. Holding it across
// either operation would close that cycle and deadlock.
nsCOMPtr<nsIThread> thread;
{
auto lockedUpdate = sPendingChunkManagerUpdate.Lock();
if (!lockedUpdate->mProfilerChild || lockedUpdate->mUpdate.IsNotUpdate()) {
return;
}
thread = lockedUpdate->mProfilerChild->mThread;
}
thread->Dispatch(NS_NewRunnableFunction(
"ProfilerChild::ProcessPendingUpdate", []() mutable {
RefPtr<ProfilerChild> profilerChild;
ProfileBufferControlledChunkManager::Update update;
{
auto lockedUpdate = sPendingChunkManagerUpdate.Lock();
if (!lockedUpdate->mProfilerChild ||
lockedUpdate->mUpdate.IsNotUpdate()) {
return;
}
profilerChild = lockedUpdate->mProfilerChild;
update = std::move(lockedUpdate->mUpdate);
lockedUpdate->mUpdate.Clear();
}
profilerChild->ProcessChunkManagerUpdate(std::move(update));
}));
}
/* static */ bool ProfilerChild::IsLockedOnCurrentThread() {
return sPendingChunkManagerUpdate.Mutex().IsLockedOnCurrentThread();
}
void ProfilerChild::SetupChunkManager() {
mChunkManager = profiler_get_controlled_chunk_manager();
if (NS_WARN_IF(!mChunkManager)) {
return;
}
// Make sure there are no updates (from a previous run).
mChunkManagerUpdate.Clear();
{
auto lockedUpdate = sPendingChunkManagerUpdate.Lock();
lockedUpdate->mProfilerChild = this;
lockedUpdate->mUpdate.Clear();
}
mChunkManager->SetUpdateCallback(
[](ProfileBufferControlledChunkManager::Update&& aUpdate) {
// Updates from the chunk manager are stored for later processing.
// We avoid dispatching a task, as this could deadlock (if the queueing
// mutex is held elsewhere).
auto lockedUpdate = sPendingChunkManagerUpdate.Lock();
if (!lockedUpdate->mProfilerChild) {
return;
}
lockedUpdate->mUpdate.Fold(std::move(aUpdate));
});
}
/* static */ void ProfilerChild::ClearPendingUpdate() {
auto lockedUpdate = sPendingChunkManagerUpdate.Lock();
lockedUpdate->mProfilerChild = nullptr;
lockedUpdate->mUpdate.Clear();
}
void ProfilerChild::ResetChunkManager() {
if (!mChunkManager) {
return;
}
// We have a chunk manager, reset the callback, which will add a final
// pending update.
mChunkManager->SetUpdateCallback({});
// Clear the pending update.
ClearPendingUpdate();
// And process a final update right now.
ProcessChunkManagerUpdate(
ProfileBufferControlledChunkManager::Update(nullptr));
mChunkManager = nullptr;
mAwaitNextChunkManagerUpdateResolver = nullptr;
}
mozilla::ipc::IPCResult ProfilerChild::RecvStart(
const ProfilerInitParams& params, StartResolver&& aResolve) {
nsTArray<const char*> filterArray;
for (size_t i = 0; i < params.filters().Length(); ++i) {
filterArray.AppendElement(params.filters()[i].get());
}
profiler_start(PowerOfTwo32(params.entries()), params.interval(),
params.features(), filterArray.Elements(),
filterArray.Length(), params.activeTabID(), params.duration());
SetupChunkManager();
aResolve(/* unused */ true);
return IPC_OK();
}
mozilla::ipc::IPCResult ProfilerChild::RecvEnsureStarted(
const ProfilerInitParams& params, EnsureStartedResolver&& aResolve) {
nsTArray<const char*> filterArray;
for (size_t i = 0; i < params.filters().Length(); ++i) {
filterArray.AppendElement(params.filters()[i].get());
}
profiler_ensure_started(PowerOfTwo32(params.entries()), params.interval(),
params.features(), filterArray.Elements(),
filterArray.Length(), params.activeTabID(),
params.duration());
SetupChunkManager();
aResolve(/* unused */ true);
return IPC_OK();
}
mozilla::ipc::IPCResult ProfilerChild::RecvStop(StopResolver&& aResolve) {
ResetChunkManager();
profiler_stop();
aResolve(/* unused */ true);
return IPC_OK();
}
mozilla::ipc::IPCResult ProfilerChild::RecvPause(PauseResolver&& aResolve) {
profiler_pause();
aResolve(/* unused */ true);
return IPC_OK();
}
mozilla::ipc::IPCResult ProfilerChild::RecvResume(ResumeResolver&& aResolve) {
profiler_resume();
aResolve(/* unused */ true);
return IPC_OK();
}
mozilla::ipc::IPCResult ProfilerChild::RecvPauseSampling(
PauseSamplingResolver&& aResolve) {
profiler_pause_sampling();
aResolve(/* unused */ true);
return IPC_OK();
}
mozilla::ipc::IPCResult ProfilerChild::RecvResumeSampling(
ResumeSamplingResolver&& aResolve) {
profiler_resume_sampling();
aResolve(/* unused */ true);
return IPC_OK();
}
mozilla::ipc::IPCResult ProfilerChild::RecvWaitOnePeriodicSampling(
WaitOnePeriodicSamplingResolver&& aResolve) {
std::shared_ptr<WaitOnePeriodicSamplingResolver> resolve =
std::make_shared<WaitOnePeriodicSamplingResolver>(std::move(aResolve));
if (!profiler_callback_after_sampling(
[self = RefPtr(this), resolve](SamplingState aSamplingState) mutable {
if (self->mDestroyed) {
return;
}
MOZ_RELEASE_ASSERT(self->mThread);
self->mThread->Dispatch(NS_NewRunnableFunction(
"nsProfiler::WaitOnePeriodicSampling result on main thread",
[resolve = std::move(resolve), aSamplingState]() {
(*resolve)(aSamplingState ==
SamplingState::SamplingCompleted);
}));
})) {
// Callback was not added (e.g., profiler is not running) and will never be
// invoked, so we need to resolve the promise here.
(*resolve)(false);
}
return IPC_OK();
}
mozilla::ipc::IPCResult ProfilerChild::RecvClearAllPages() {
profiler_clear_all_pages();
return IPC_OK();
}
mozilla::ipc::IPCResult ProfilerChild::RecvAwaitNextChunkManagerUpdate(
AwaitNextChunkManagerUpdateResolver&& aResolve) {
MOZ_ASSERT(!mDestroyed,
"Recv... should not be called if the actor was destroyed");
// Pick up pending updates if any.
{
auto lockedUpdate = sPendingChunkManagerUpdate.Lock();
if (lockedUpdate->mProfilerChild && !lockedUpdate->mUpdate.IsNotUpdate()) {
mChunkManagerUpdate.Fold(std::move(lockedUpdate->mUpdate));
lockedUpdate->mUpdate.Clear();
}
}
if (mChunkManagerUpdate.IsNotUpdate()) {
// No data yet, store the resolver for later.
mAwaitNextChunkManagerUpdateResolver = std::move(aResolve);
} else {
// We have data, send it now.
ResolveChunkUpdate(aResolve);
}
return IPC_OK();
}
mozilla::ipc::IPCResult ProfilerChild::RecvDestroyReleasedChunksAtOrBefore(
const TimeStamp& aTimeStamp) {
if (mChunkManager) {
mChunkManager->DestroyChunksAtOrBefore(aTimeStamp);
}
return IPC_OK();
}
struct GatherProfileThreadParameters
: public external::AtomicRefCounted<GatherProfileThreadParameters> {
MOZ_DECLARE_REFCOUNTED_TYPENAME(GatherProfileThreadParameters)
GatherProfileThreadParameters(
RefPtr<ProfilerChild> aProfilerChild,
RefPtr<ProgressLogger::SharedProgress> aProgress,
ProfilerChild::GatherProfileResolver&& aResolver)
: profilerChild(std::move(aProfilerChild)),
progress(std::move(aProgress)),
resolver(std::move(aResolver)) {}
RefPtr<ProfilerChild> profilerChild;
FailureLatchSource failureLatchSource;
// Separate RefPtr used when working on separate thread. This way, if the
// "ProfilerChild" thread decides to overwrite its mGatherProfileProgress with
// a new one, the work done here will still only use the old one.
RefPtr<ProgressLogger::SharedProgress> progress;
// Resolver for the GatherProfile promise. Must only be called on the
// "ProfilerChild" thread.
ProfilerChild::GatherProfileResolver resolver;
};
/* static */
void ProfilerChild::GatherProfileThreadFunction(
void* already_AddRefedParameters) {
PR_SetCurrentThreadName("GatherProfileThread");
RefPtr<GatherProfileThreadParameters> parameters =
already_AddRefed<GatherProfileThreadParameters>{
static_cast<GatherProfileThreadParameters*>(
already_AddRefedParameters)};
ProgressLogger progressLogger(
parameters->progress, "Gather-profile thread started", "Profile sent");
using namespace mozilla::literals::ProportionValue_literals; // For `1_pc`.
auto writer =
MakeUnique<SpliceableChunkedJSONWriter>(parameters->failureLatchSource);
auto rv =
profiler_get_profile_json(*writer,
/* aSinceTime */ 0,
/* aIsShuttingDown */ false,
progressLogger.CreateSubLoggerFromTo(
1_pc, "profiler_get_profile_json started",
99_pc, "profiler_get_profile_json done"));
if (rv.isErr()) {
// Failed to get a profile, reset the writer pointer, so that we'll send a
// failure message.
writer.reset();
}
if (NS_WARN_IF(NS_FAILED(
parameters->profilerChild->mThread->Dispatch(NS_NewRunnableFunction(
"ProfilerChild::ProcessPendingUpdate",
[parameters,
// Forward progress logger to on-ProfilerChild-thread task, so
// that it doesn't get marked as 100% done when this off-thread
// function ends.
progressLogger = std::move(progressLogger),
writer = std::move(writer), rv = std::move(rv)]() mutable {
// We are now on the ProfilerChild thread, about to send the
// completed profile. Any incoming progress request will now be
// handled after this task ends, so updating the progress is now
// useless and we can just get rid of the progress storage.
if (parameters->profilerChild->mGatherProfileProgress ==
parameters->progress) {
// The ProfilerChild progress is still the one we know.
parameters->profilerChild->mGatherProfileProgress = nullptr;
}
// Shmem allocation and promise resolution must be made on the
// ProfilerChild thread, that's why this task was needed here.
mozilla::ipc::Shmem shmem;
if (writer) {
if (const size_t len = writer->ChunkedWriteFunc().Length();
len < UINT32_MAX) {
bool shmemSuccess = true;
const bool copySuccess =
writer->ChunkedWriteFunc()
.CopyDataIntoLazilyAllocatedBuffer(
[&](size_t allocationSize) -> char* {
MOZ_ASSERT(allocationSize == len + 1);
if (parameters->profilerChild->AllocShmem(
allocationSize, &shmem)) {
return shmem.get<char>();
}
shmemSuccess = false;
return nullptr;
});
if (!shmemSuccess || !copySuccess) {
const nsPrintfCString message(
(!shmemSuccess)
? "*Could not create shmem for profile from pid "
"%u (%zu B)"
: "*Could not write profile from pid %u (%zu B)",
unsigned(profiler_current_process_id().ToNumber()),
len);
if (parameters->profilerChild->AllocShmem(
message.Length() + 1, &shmem)) {
strcpy(shmem.get<char>(), message.get());
}
}
} else {
const nsPrintfCString message(
"*Profile from pid %u bigger (%zu) than shmem max "
"(%zu)",
unsigned(profiler_current_process_id().ToNumber()), len,
size_t(UINT32_MAX));
if (parameters->profilerChild->AllocShmem(
message.Length() + 1, &shmem)) {
strcpy(shmem.get<char>(), message.get());
}
}
writer = nullptr;
} else {
// No profile.
const char* failure =
parameters->failureLatchSource.GetFailure();
const nsPrintfCString message(
"*Could not generate profile from pid %u%s%s",
unsigned(profiler_current_process_id().ToNumber()),
failure ? ", failure: " : "", failure ? failure : "");
if (parameters->profilerChild->AllocShmem(
message.Length() + 1, &shmem)) {
strcpy(shmem.get<char>(), message.get());
}
}
Maybe<ProfileGenerationAdditionalInformation> additionalInfo =
rv.isOk() ? Some(rv.unwrap()) : Nothing();
parameters->resolver(IPCProfileAndAdditionalInformation{
std::move(shmem), std::move(additionalInfo)});
// Let's join the gather profile thread now since it's done.
// Note that this gets called inside the ProfilerChild thread
// and not inside the gather profile thread itself.
parameters->profilerChild->JoinGatherProfileThread();
}))))) {
// Failed to dispatch the task to the ProfilerChild thread. The IPC cannot
// be resolved on this thread, so it will never be resolved!
// And it would be unsafe to modify mGatherProfileProgress; But the parent
// should notice that's it's not advancing anymore.
// If this fails, it usually means that the process is shutting down. This
// thread is not joined yet and we can't join it here since we are still in
// the thread, but it will happen inside ProfilerChild::ActorDestroy.
}
}
mozilla::ipc::IPCResult ProfilerChild::RecvGatherProfile(
GatherProfileResolver&& aResolve) {
mGatherProfileProgress = MakeRefPtr<ProgressLogger::SharedProgress>();
mGatherProfileProgress->SetProgress(ProportionValue{0.0},
"Received gather-profile request");
auto parameters = MakeRefPtr<GatherProfileThreadParameters>(
this, mGatherProfileProgress, std::move(aResolve));
// The GatherProfileThreadFunction thread function will cast its void*
// argument to already_AddRefed<GatherProfileThreadParameters>.
parameters.get()->AddRef();
mGatherProfileThread = PR_CreateThread(
PR_SYSTEM_THREAD, GatherProfileThreadFunction, parameters.get(),
PR_PRIORITY_NORMAL, PR_GLOBAL_THREAD, PR_JOINABLE_THREAD, 0);
if (!mGatherProfileThread) {
// Failed to create and start worker thread, resolve with an empty profile.
mozilla::ipc::Shmem shmem;
if (AllocShmem(1, &shmem)) {
shmem.get<char>()[0] = '\0';
}
parameters->resolver(IPCProfileAndAdditionalInformation{shmem, Nothing()});
// And clean up.
parameters.get()->Release();
mGatherProfileProgress = nullptr;
}
return IPC_OK();
}
mozilla::ipc::IPCResult ProfilerChild::RecvGetGatherProfileProgress(
GetGatherProfileProgressResolver&& aResolve) {
if (mGatherProfileProgress) {
aResolve(GatherProfileProgress{
mGatherProfileProgress->Progress().ToUnderlyingType(),
nsCString(mGatherProfileProgress->LastLocation())});
} else {
aResolve(
GatherProfileProgress{ProportionValue::MakeInvalid().ToUnderlyingType(),
nsCString("No gather-profile in progress")});
}
return IPC_OK();
}
void ProfilerChild::JoinGatherProfileThread() {
if (!mGatherProfileThread) {
// There is no thread running to join.
return;
}
PR_JoinThread(mGatherProfileThread);
mGatherProfileThread = nullptr;
}
void ProfilerChild::ActorDestroy(ActorDestroyReason aActorDestroyReason) {
mDestroyed = true;
// Join the profile gathering thread in case it's not exited yet to prevent
// any leaks during shutdown.
JoinGatherProfileThread();
}
void ProfilerChild::Destroy() {
ClearPendingUpdate();
if (!mDestroyed) {
Close();
}
}
ProfileAndAdditionalInformation ProfilerChild::GrabShutdownProfile() {
LOG("GrabShutdownProfile");
UniquePtr<ProfilerCodeAddressService> service =
profiler_code_address_service_for_presymbolication();
FailureLatchSource failureLatch;
SpliceableChunkedJSONWriter writer{failureLatch};
writer.Start();
auto rv = profiler_stream_json_for_this_process(
writer, /* aSinceTime */ 0,
/* aIsShuttingDown */ true, service.get(), ProgressLogger{});
if (rv.isErr()) {
const char* failure = writer.GetFailure();
return ProfileAndAdditionalInformation(
nsPrintfCString("*Profile unavailable for pid %u%s%s",
unsigned(profiler_current_process_id().ToNumber()),
failure ? ", failure: " : "", failure ? failure : ""));
}
auto additionalInfo = rv.unwrap();
writer.StartArrayProperty("processes");
writer.EndArray();
writer.End();
const size_t len = writer.ChunkedWriteFunc().Length();
// This string and information are destined to be sent as a shutdown profile,
// which is limited by the maximum IPC message size.
// TODO: IPC to change to shmem (bug 1780330), raising this limit to
// JS::MaxStringLength.
if (len + additionalInfo.SizeOf() >=
size_t(IPC::Channel::kMaximumMessageSize)) {
return ProfileAndAdditionalInformation(
nsPrintfCString("*Profile from pid %u bigger (%zu) than IPC max (%zu)",
unsigned(profiler_current_process_id().ToNumber()), len,
size_t(IPC::Channel::kMaximumMessageSize)));
}
nsCString profileCString;
if (!profileCString.SetLength(len, fallible)) {
return ProfileAndAdditionalInformation(nsPrintfCString(
"*Could not allocate %zu bytes for profile from pid %u", len,
unsigned(profiler_current_process_id().ToNumber())));
}
MOZ_ASSERT(*(profileCString.Data() + len) == '\0',
"We expected a null at the end of the string buffer, to be "
"rewritten by CopyDataIntoLazilyAllocatedBuffer");
char* const profileBeginWriting = profileCString.BeginWriting();
if (!profileBeginWriting) {
return ProfileAndAdditionalInformation(
nsPrintfCString("*Could not write profile from pid %u",
unsigned(profiler_current_process_id().ToNumber())));
}
// Here, we have enough space reserved in `profileCString`, starting at
// `profileBeginWriting`, copy the JSON profile there.
if (!writer.ChunkedWriteFunc().CopyDataIntoLazilyAllocatedBuffer(
[&](size_t aBufferLen) -> char* {
MOZ_RELEASE_ASSERT(aBufferLen == len + 1);
return profileBeginWriting;
})) {
return ProfileAndAdditionalInformation(
nsPrintfCString("*Could not copy profile from pid %u",
unsigned(profiler_current_process_id().ToNumber())));
}
MOZ_ASSERT(*(profileCString.Data() + len) == '\0',
"We still expected a null at the end of the string buffer");
return ProfileAndAdditionalInformation{std::move(profileCString),
std::move(additionalInfo)};
}
} // namespace mozilla