Source code

Revision control

Copy as Markdown

Other Tools

/* -*- Mode: C++; tab-width: 2; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
/* vim: set ts=8 sts=2 et sw=2 tw=80: */
/* 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/. */
#ifndef ProfileChunkedBuffer_h
#define ProfileChunkedBuffer_h
#include "mozilla/Attributes.h"
#include "mozilla/BaseProfilerDetail.h"
#include "mozilla/NotNull.h"
#include "mozilla/ProfileBufferChunkManager.h"
#include "mozilla/ProfileBufferChunkManagerSingle.h"
#include "mozilla/ProfileBufferEntrySerialization.h"
#include "mozilla/ProfileChunkedBufferDetail.h"
#include "mozilla/RefPtr.h"
#include "mozilla/ScopeExit.h"
#include "mozilla/Unused.h"
#include <utility>
#ifdef DEBUG
# include <cstdio>
#endif
namespace mozilla {
// Thread-safe buffer that can store blocks of different sizes during defined
// sessions, using Chunks (from a ChunkManager) as storage.
//
// Each *block* contains an *entry* and the entry size:
// [ entry_size | entry ] [ entry_size | entry ] ...
//
// *In-session* is a period of time during which `ProfileChunkedBuffer` allows
// reading and writing.
// *Out-of-session*, the `ProfileChunkedBuffer` object is still valid, but
// contains no data, and gracefully denies accesses.
//
// To write an entry, the buffer reserves a block of sufficient size (to contain
// user data of predetermined size), writes the entry size, and lets the caller
// fill the entry contents using a ProfileBufferEntryWriter. E.g.:
// ```
// ProfileChunkedBuffer cb(...);
// cb.ReserveAndPut([]() { return sizeof(123); },
// [&](Maybe<ProfileBufferEntryWriter>& aEW) {
// if (aEW) { aEW->WriteObject(123); }
// });
// ```
// Other `Put...` functions may be used as shortcuts for simple entries.
// The objects given to the caller's callbacks should only be used inside the
// callbacks and not stored elsewhere, because they keep their own references to
// chunk memory and therefore should not live longer.
// Different type of objects may be serialized into an entry, see
// `ProfileBufferEntryWriter::Serializer` for more information.
//
// When reading data, the buffer iterates over blocks (it knows how to read the
// entry size, and therefore move to the next block), and lets the caller read
// the entry inside of each block. E.g.:
// ```
// cb.ReadEach([](ProfileBufferEntryReader& aER) {
// /* Use ProfileBufferEntryReader functions to read serialized objects. */
// int n = aER.ReadObject<int>();
// });
// ```
// Different type of objects may be deserialized from an entry, see
// `ProfileBufferEntryReader::Deserializer` for more information.
//
// Writers may retrieve the block index corresponding to an entry
// (`ProfileBufferBlockIndex` is an opaque type preventing the user from easily
// modifying it). That index may later be used with `ReadAt` to get back to the
// entry in that particular block -- if it still exists.
class ProfileChunkedBuffer {
public:
using Byte = ProfileBufferChunk::Byte;
using Length = ProfileBufferChunk::Length;
enum class ThreadSafety { WithoutMutex, WithMutex };
// Default constructor starts out-of-session (nothing to read or write).
explicit ProfileChunkedBuffer(ThreadSafety aThreadSafety)
: mMutex(aThreadSafety != ThreadSafety::WithoutMutex) {}
// Start in-session with external chunk manager.
ProfileChunkedBuffer(ThreadSafety aThreadSafety,
ProfileBufferChunkManager& aChunkManager)
: mMutex(aThreadSafety != ThreadSafety::WithoutMutex) {
SetChunkManager(aChunkManager);
}
// Start in-session with owned chunk manager.
ProfileChunkedBuffer(ThreadSafety aThreadSafety,
UniquePtr<ProfileBufferChunkManager>&& aChunkManager)
: mMutex(aThreadSafety != ThreadSafety::WithoutMutex) {
SetChunkManager(std::move(aChunkManager));
}
~ProfileChunkedBuffer() {
// Do proper clean-up by resetting the chunk manager.
ResetChunkManager();
}
// This cannot change during the lifetime of this buffer, so there's no need
// to lock.
[[nodiscard]] bool IsThreadSafe() const { return mMutex.IsActivated(); }
[[nodiscard]] bool IsInSession() const {
baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
return !!mChunkManager;
}
// Stop using the current chunk manager.
// If we own the current chunk manager, it will be destroyed.
// This will always clear currently-held chunks, if any.
void ResetChunkManager() {
baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
Unused << ResetChunkManager(lock);
}
// Set the current chunk manager.
// The caller is responsible for keeping the chunk manager alive as along as
// it's used here (until the next (Re)SetChunkManager, or
// ~ProfileChunkedBuffer).
void SetChunkManager(ProfileBufferChunkManager& aChunkManager) {
baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
Unused << ResetChunkManager(lock);
SetChunkManager(aChunkManager, lock);
}
// Set the current chunk manager, and keep ownership of it.
void SetChunkManager(UniquePtr<ProfileBufferChunkManager>&& aChunkManager) {
baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
Unused << ResetChunkManager(lock);
mOwnedChunkManager = std::move(aChunkManager);
if (mOwnedChunkManager) {
SetChunkManager(*mOwnedChunkManager, lock);
}
}
// Set the current chunk manager, except if it's already the one provided.
// The caller is responsible for keeping the chunk manager alive as along as
// it's used here (until the next (Re)SetChunkManager, or
// ~ProfileChunkedBuffer).
void SetChunkManagerIfDifferent(ProfileBufferChunkManager& aChunkManager) {
baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
if (!mChunkManager || mChunkManager != &aChunkManager) {
Unused << ResetChunkManager(lock);
SetChunkManager(aChunkManager, lock);
}
}
// Clear the contents of this buffer, ready to receive new chunks.
// Note that memory is not freed: No chunks are destroyed, they are all
// receycled.
// Also the range doesn't reset, instead it continues at some point after the
// previous range. This may be useful if the caller may be keeping indexes
// into old chunks that have now been cleared, using these indexes will fail
// gracefully (instead of potentially pointing into new data).
void Clear() {
baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
if (MOZ_UNLIKELY(!mChunkManager)) {
// Out-of-session.
return;
}
mRangeStart = mRangeEnd = mNextChunkRangeStart;
mPushedBlockCount = 0;
mClearedBlockCount = 0;
mFailedPutBytes = 0;
// Recycle all released chunks as "next" chunks. This will reduce the number
// of future allocations. Also, when using ProfileBufferChunkManagerSingle,
// this retrieves the one chunk if it was released.
UniquePtr<ProfileBufferChunk> releasedChunks =
mChunkManager->GetExtantReleasedChunks();
if (releasedChunks) {
// Released chunks should be in the "Done" state, they need to be marked
// "recycled" before they can be reused.
for (ProfileBufferChunk* chunk = releasedChunks.get(); chunk;
chunk = chunk->GetNext()) {
chunk->MarkRecycled();
}
mNextChunks = ProfileBufferChunk::Join(std::move(mNextChunks),
std::move(releasedChunks));
}
if (mCurrentChunk) {
// We already have a current chunk (empty or in-use), mark it "done" and
// then "recycled", ready to be reused.
mCurrentChunk->MarkDone();
mCurrentChunk->MarkRecycled();
} else {
if (!mNextChunks) {
// No current chunk, and no next chunks to recycle, nothing more to do.
// The next "Put" operation will try to allocate a chunk as needed.
return;
}
// No current chunk, take a next chunk.
mCurrentChunk = std::exchange(mNextChunks, mNextChunks->ReleaseNext());
}
// Here, there was already a current chunk, or one has just been taken.
// Make sure it's ready to receive new entries.
InitializeCurrentChunk(lock);
}
// Buffer maximum length in bytes.
Maybe<size_t> BufferLength() const {
baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
if (!mChunkManager) {
return Nothing{};
}
return Some(mChunkManager->MaxTotalSize());
}
[[nodiscard]] size_t SizeOfExcludingThis(MallocSizeOf aMallocSizeOf) const {
baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
return SizeOfExcludingThis(aMallocSizeOf, lock);
}
[[nodiscard]] size_t SizeOfIncludingThis(MallocSizeOf aMallocSizeOf) const {
baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
return aMallocSizeOf(this) + SizeOfExcludingThis(aMallocSizeOf, lock);
}
// Snapshot of the buffer state.
struct State {
// Index to/before the first block.
ProfileBufferIndex mRangeStart = 1;
// Index past the last block. Equals mRangeStart if empty.
ProfileBufferIndex mRangeEnd = 1;
// Number of blocks that have been pushed into this buffer.
uint64_t mPushedBlockCount = 0;
// Number of blocks that have been removed from this buffer.
// Note: Live entries = pushed - cleared.
uint64_t mClearedBlockCount = 0;
// Number of bytes that could not be put into this buffer.
uint64_t mFailedPutBytes = 0;
};
// Get a snapshot of the current state.
// When out-of-session, mFirstReadIndex==mNextWriteIndex, and
// mPushedBlockCount==mClearedBlockCount==0.
// Note that these may change right after this thread-safe call, so they
// should only be used for statistical purposes.
[[nodiscard]] State GetState() const {
baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
return {mRangeStart, mRangeEnd, mPushedBlockCount, mClearedBlockCount,
mFailedPutBytes};
}
// In in-session, return the start TimeStamp of the earliest chunk.
// If out-of-session, return a null TimeStamp.
[[nodiscard]] TimeStamp GetEarliestChunkStartTimeStamp() const {
baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
if (MOZ_UNLIKELY(!mChunkManager)) {
// Out-of-session.
return {};
}
return mChunkManager->PeekExtantReleasedChunks(
[&](const ProfileBufferChunk* aOldestChunk) -> TimeStamp {
if (aOldestChunk) {
return aOldestChunk->ChunkHeader().mStartTimeStamp;
}
if (mCurrentChunk) {
return mCurrentChunk->ChunkHeader().mStartTimeStamp;
}
return {};
});
}
[[nodiscard]] bool IsEmpty() const {
baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
return mRangeStart == mRangeEnd;
}
// True if this buffer is already locked on this thread.
// This should be used if some functions may call an already-locked buffer,
// e.g.: Put -> memory hook -> profiler_add_native_allocation_marker -> Put.
[[nodiscard]] bool IsThreadSafeAndLockedOnCurrentThread() const {
return mMutex.IsActivatedAndLockedOnCurrentThread();
}
// Lock the buffer mutex and run the provided callback.
// This can be useful when the caller needs to explicitly lock down this
// buffer, but not do anything else with it.
template <typename Callback>
auto LockAndRun(Callback&& aCallback) const {
baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
return std::forward<Callback>(aCallback)();
}
// Reserve a block that can hold an entry of the given `aCallbackEntryBytes()`
// size, write the entry size (ULEB128-encoded), and invoke and return
// `aCallback(Maybe<ProfileBufferEntryWriter>&)`.
// Note: `aCallbackEntryBytes` is a callback instead of a simple value, to
// delay this potentially-expensive computation until after we're checked that
// we're in-session; use `Put(Length, Callback)` below if you know the size
// already.
template <typename CallbackEntryBytes, typename Callback>
auto ReserveAndPut(CallbackEntryBytes&& aCallbackEntryBytes,
Callback&& aCallback)
-> decltype(std::forward<Callback>(aCallback)(
std::declval<Maybe<ProfileBufferEntryWriter>&>())) {
baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
// This can only be read in the 2nd lambda below after it has been written
// by the first lambda.
Length entryBytes;
return ReserveAndPutRaw(
[&]() {
entryBytes = std::forward<CallbackEntryBytes>(aCallbackEntryBytes)();
MOZ_ASSERT(entryBytes != 0, "Empty entries are not allowed");
return ULEB128Size(entryBytes) + entryBytes;
},
[&](Maybe<ProfileBufferEntryWriter>& aMaybeEntryWriter) {
if (aMaybeEntryWriter.isSome()) {
aMaybeEntryWriter->WriteULEB128(entryBytes);
MOZ_ASSERT(aMaybeEntryWriter->RemainingBytes() == entryBytes);
}
return std::forward<Callback>(aCallback)(aMaybeEntryWriter);
},
lock);
}
template <typename Callback>
auto Put(Length aEntryBytes, Callback&& aCallback) {
return ReserveAndPut([aEntryBytes]() { return aEntryBytes; },
std::forward<Callback>(aCallback));
}
// Add a new entry copied from the given buffer, return block index.
ProfileBufferBlockIndex PutFrom(const void* aSrc, Length aBytes) {
return ReserveAndPut(
[aBytes]() { return aBytes; },
[aSrc, aBytes](Maybe<ProfileBufferEntryWriter>& aMaybeEntryWriter) {
if (aMaybeEntryWriter.isNothing()) {
return ProfileBufferBlockIndex{};
}
aMaybeEntryWriter->WriteBytes(aSrc, aBytes);
return aMaybeEntryWriter->CurrentBlockIndex();
});
}
// Add a new single entry with *all* given object (using a Serializer for
// each), return block index.
template <typename... Ts>
ProfileBufferBlockIndex PutObjects(const Ts&... aTs) {
static_assert(sizeof...(Ts) > 0,
"PutObjects must be given at least one object.");
return ReserveAndPut(
[&]() { return ProfileBufferEntryWriter::SumBytes(aTs...); },
[&](Maybe<ProfileBufferEntryWriter>& aMaybeEntryWriter) {
if (aMaybeEntryWriter.isNothing()) {
return ProfileBufferBlockIndex{};
}
aMaybeEntryWriter->WriteObjects(aTs...);
return aMaybeEntryWriter->CurrentBlockIndex();
});
}
// Add a new entry copied from the given object, return block index.
template <typename T>
ProfileBufferBlockIndex PutObject(const T& aOb) {
return PutObjects(aOb);
}
// Get *all* chunks related to this buffer, including extant chunks in its
// ChunkManager, and yet-unused new/recycled chunks.
// We don't expect this buffer to be used again, though it's still possible
// and will allocate the first buffer when needed.
[[nodiscard]] UniquePtr<ProfileBufferChunk> GetAllChunks() {
baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
if (MOZ_UNLIKELY(!mChunkManager)) {
// Out-of-session.
return nullptr;
}
UniquePtr<ProfileBufferChunk> chunks =
mChunkManager->GetExtantReleasedChunks();
Unused << HandleRequestedChunk_IsPending(lock);
if (MOZ_LIKELY(!!mCurrentChunk)) {
mCurrentChunk->MarkDone();
chunks =
ProfileBufferChunk::Join(std::move(chunks), std::move(mCurrentChunk));
}
chunks =
ProfileBufferChunk::Join(std::move(chunks), std::move(mNextChunks));
mChunkManager->ForgetUnreleasedChunks();
mRangeStart = mRangeEnd = mNextChunkRangeStart;
return chunks;
}
// True if the given index points inside the current chunk (up to the last
// written byte).
// This could be used to check if an index written now would have a good
// chance of referring to a previous block that has not been destroyed yet.
// But use with extreme care: This information may become incorrect right
// after this function returns, because new writes could start a new chunk.
[[nodiscard]] bool IsIndexInCurrentChunk(ProfileBufferIndex aIndex) const {
baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
if (MOZ_UNLIKELY(!mChunkManager || !mCurrentChunk)) {
// Out-of-session, or no current chunk.
return false;
}
return (mCurrentChunk->RangeStart() <= aIndex) &&
(aIndex < (mCurrentChunk->RangeStart() +
mCurrentChunk->OffsetPastLastBlock()));
}
class Reader;
// Class that can iterate through blocks and provide
// `ProfileBufferEntryReader`s.
// Created through `Reader`, lives within a lock guard lifetime.
class BlockIterator {
public:
#ifdef DEBUG
~BlockIterator() {
// No BlockIterator should live outside of a mutexed call.
mBuffer->mMutex.AssertCurrentThreadOwns();
}
#endif // DEBUG
// Comparison with other iterator, mostly used in range-for loops.
[[nodiscard]] bool operator==(const BlockIterator& aRhs) const {
MOZ_ASSERT(mBuffer == aRhs.mBuffer);
return mCurrentBlockIndex == aRhs.mCurrentBlockIndex;
}
[[nodiscard]] bool operator!=(const BlockIterator& aRhs) const {
MOZ_ASSERT(mBuffer == aRhs.mBuffer);
return mCurrentBlockIndex != aRhs.mCurrentBlockIndex;
}
// Advance to next BlockIterator.
BlockIterator& operator++() {
mBuffer->mMutex.AssertCurrentThreadOwns();
mCurrentBlockIndex =
ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
mNextBlockPointer.GlobalRangePosition());
mCurrentEntry =
mNextBlockPointer.EntryReader(mNextBlockPointer.ReadEntrySize());
return *this;
}
// Dereferencing creates a `ProfileBufferEntryReader` object for the entry
// inside this block.
// (Note: It would be possible to return a `const
// ProfileBufferEntryReader&`, but not useful in practice, because in most
// case the user will want to read, which is non-const.)
[[nodiscard]] ProfileBufferEntryReader operator*() const {
return mCurrentEntry;
}
// True if this iterator is just past the last entry.
[[nodiscard]] bool IsAtEnd() const {
return mCurrentEntry.RemainingBytes() == 0;
}
// Can be used as reference to come back to this entry with `GetEntryAt()`.
[[nodiscard]] ProfileBufferBlockIndex CurrentBlockIndex() const {
return mCurrentBlockIndex;
}
// Index past the end of this block, which is the start of the next block.
[[nodiscard]] ProfileBufferBlockIndex NextBlockIndex() const {
MOZ_ASSERT(!IsAtEnd());
return ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
mNextBlockPointer.GlobalRangePosition());
}
// Index of the first block in the whole buffer.
[[nodiscard]] ProfileBufferBlockIndex BufferRangeStart() const {
mBuffer->mMutex.AssertCurrentThreadOwns();
return ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
mBuffer->mRangeStart);
}
// Index past the last block in the whole buffer.
[[nodiscard]] ProfileBufferBlockIndex BufferRangeEnd() const {
mBuffer->mMutex.AssertCurrentThreadOwns();
return ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
mBuffer->mRangeEnd);
}
private:
// Only a Reader can instantiate a BlockIterator.
friend class Reader;
BlockIterator(const ProfileChunkedBuffer& aBuffer,
const ProfileBufferChunk* aChunks0,
const ProfileBufferChunk* aChunks1,
ProfileBufferBlockIndex aBlockIndex)
: mNextBlockPointer(aChunks0, aChunks1, aBlockIndex),
mCurrentBlockIndex(
ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
mNextBlockPointer.GlobalRangePosition())),
mCurrentEntry(
mNextBlockPointer.EntryReader(mNextBlockPointer.ReadEntrySize())),
mBuffer(WrapNotNull(&aBuffer)) {
// No BlockIterator should live outside of a mutexed call.
mBuffer->mMutex.AssertCurrentThreadOwns();
}
profiler::detail::InChunkPointer mNextBlockPointer;
ProfileBufferBlockIndex mCurrentBlockIndex;
ProfileBufferEntryReader mCurrentEntry;
// Using a non-null pointer instead of a reference, to allow copying.
// This BlockIterator should only live inside one of the thread-safe
// ProfileChunkedBuffer functions, for this reference to stay valid.
NotNull<const ProfileChunkedBuffer*> mBuffer;
};
// Class that can create `BlockIterator`s (e.g., for range-for), or just
// iterate through entries; lives within a lock guard lifetime.
class MOZ_RAII Reader {
public:
Reader(const Reader&) = delete;
Reader& operator=(const Reader&) = delete;
Reader(Reader&&) = delete;
Reader& operator=(Reader&&) = delete;
#ifdef DEBUG
~Reader() {
// No Reader should live outside of a mutexed call.
mBuffer.mMutex.AssertCurrentThreadOwns();
}
#endif // DEBUG
// Index of the first block in the whole buffer.
[[nodiscard]] ProfileBufferBlockIndex BufferRangeStart() const {
mBuffer.mMutex.AssertCurrentThreadOwns();
return ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
mBuffer.mRangeStart);
}
// Index past the last block in the whole buffer.
[[nodiscard]] ProfileBufferBlockIndex BufferRangeEnd() const {
mBuffer.mMutex.AssertCurrentThreadOwns();
return ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
mBuffer.mRangeEnd);
}
// Iterators to the first and past-the-last blocks.
// Compatible with range-for (see `ForEach` below as example).
[[nodiscard]] BlockIterator begin() const {
return BlockIterator(mBuffer, mChunks0, mChunks1, nullptr);
}
// Note that a `BlockIterator` at the `end()` should not be dereferenced, as
// there is no actual block there!
[[nodiscard]] BlockIterator end() const {
return BlockIterator(mBuffer, nullptr, nullptr, nullptr);
}
// Get a `BlockIterator` at the given `ProfileBufferBlockIndex`, clamped to
// the stored range. Note that a `BlockIterator` at the `end()` should not
// be dereferenced, as there is no actual block there!
[[nodiscard]] BlockIterator At(ProfileBufferBlockIndex aBlockIndex) const {
if (aBlockIndex < BufferRangeStart()) {
// Anything before the range (including null ProfileBufferBlockIndex) is
// clamped at the beginning.
return begin();
}
// Otherwise we at least expect the index to be valid (pointing exactly at
// a live block, or just past the end.)
return BlockIterator(mBuffer, mChunks0, mChunks1, aBlockIndex);
}
// Run `aCallback(ProfileBufferEntryReader&)` on each entry from first to
// last. Callback should not store `ProfileBufferEntryReader`, as it may
// become invalid after this thread-safe call.
template <typename Callback>
void ForEach(Callback&& aCallback) const {
for (ProfileBufferEntryReader reader : *this) {
aCallback(reader);
}
}
// If this reader only points at one chunk with some data, this data will be
// exposed as a single entry.
[[nodiscard]] ProfileBufferEntryReader SingleChunkDataAsEntry() {
const ProfileBufferChunk* onlyNonEmptyChunk = nullptr;
for (const ProfileBufferChunk* chunkList : {mChunks0, mChunks1}) {
for (const ProfileBufferChunk* chunk = chunkList; chunk;
chunk = chunk->GetNext()) {
if (chunk->OffsetFirstBlock() != chunk->OffsetPastLastBlock()) {
if (onlyNonEmptyChunk) {
// More than one non-empty chunk.
return ProfileBufferEntryReader();
}
onlyNonEmptyChunk = chunk;
}
}
}
if (!onlyNonEmptyChunk) {
// No non-empty chunks.
return ProfileBufferEntryReader();
}
// Here, we have found one chunk that had some data.
return ProfileBufferEntryReader(
onlyNonEmptyChunk->BufferSpan().FromTo(
onlyNonEmptyChunk->OffsetFirstBlock(),
onlyNonEmptyChunk->OffsetPastLastBlock()),
ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
onlyNonEmptyChunk->RangeStart()),
ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
onlyNonEmptyChunk->RangeStart() +
(onlyNonEmptyChunk->OffsetPastLastBlock() -
onlyNonEmptyChunk->OffsetFirstBlock())));
}
private:
friend class ProfileChunkedBuffer;
explicit Reader(const ProfileChunkedBuffer& aBuffer,
const ProfileBufferChunk* aChunks0,
const ProfileBufferChunk* aChunks1)
: mBuffer(aBuffer), mChunks0(aChunks0), mChunks1(aChunks1) {
// No Reader should live outside of a mutexed call.
mBuffer.mMutex.AssertCurrentThreadOwns();
}
// This Reader should only live inside one of the thread-safe
// ProfileChunkedBuffer functions, for this reference to stay valid.
const ProfileChunkedBuffer& mBuffer;
const ProfileBufferChunk* mChunks0;
const ProfileBufferChunk* mChunks1;
};
// In in-session, call `aCallback(ProfileChunkedBuffer::Reader&)` and return
// true. Callback should not store `Reader`, because it may become invalid
// after this call.
// If out-of-session, return false (callback is not invoked).
template <typename Callback>
[[nodiscard]] auto Read(Callback&& aCallback) const {
baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
if (MOZ_UNLIKELY(!mChunkManager)) {
// Out-of-session.
return std::forward<Callback>(aCallback)(static_cast<Reader*>(nullptr));
}
return mChunkManager->PeekExtantReleasedChunks(
[&](const ProfileBufferChunk* aOldestChunk) {
Reader reader(*this, aOldestChunk, mCurrentChunk.get());
return std::forward<Callback>(aCallback)(&reader);
});
}
// Invoke `aCallback(ProfileBufferEntryReader& [, ProfileBufferBlockIndex])`
// on each entry, it must read or at least skip everything. Either/both chunk
// pointers may be null.
template <typename Callback>
static void ReadEach(const ProfileBufferChunk* aChunks0,
const ProfileBufferChunk* aChunks1,
Callback&& aCallback) {
static_assert(std::is_invocable_v<Callback, ProfileBufferEntryReader&> ||
std::is_invocable_v<Callback, ProfileBufferEntryReader&,
ProfileBufferBlockIndex>,
"ReadEach callback must take ProfileBufferEntryReader& and "
"optionally a ProfileBufferBlockIndex");
profiler::detail::InChunkPointer p{aChunks0, aChunks1};
while (!p.IsNull()) {
// The position right before an entry size *is* a block index.
const ProfileBufferBlockIndex blockIndex =
ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
p.GlobalRangePosition());
Length entrySize = p.ReadEntrySize();
if (entrySize == 0) {
return;
}
ProfileBufferEntryReader entryReader = p.EntryReader(entrySize);
if (entryReader.RemainingBytes() == 0) {
return;
}
MOZ_ASSERT(entryReader.RemainingBytes() == entrySize);
if constexpr (std::is_invocable_v<Callback, ProfileBufferEntryReader&,
ProfileBufferBlockIndex>) {
aCallback(entryReader, blockIndex);
} else {
Unused << blockIndex;
aCallback(entryReader);
}
MOZ_ASSERT(entryReader.RemainingBytes() == 0);
}
}
// Invoke `aCallback(ProfileBufferEntryReader& [, ProfileBufferBlockIndex])`
// on each entry, it must read or at least skip everything.
template <typename Callback>
void ReadEach(Callback&& aCallback) const {
baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
if (MOZ_UNLIKELY(!mChunkManager)) {
// Out-of-session.
return;
}
mChunkManager->PeekExtantReleasedChunks(
[&](const ProfileBufferChunk* aOldestChunk) {
ReadEach(aOldestChunk, mCurrentChunk.get(),
std::forward<Callback>(aCallback));
});
}
// Call `aCallback(Maybe<ProfileBufferEntryReader>&&)` on the entry at
// the given ProfileBufferBlockIndex; The `Maybe` will be `Nothing` if
// out-of-session, or if that entry doesn't exist anymore, or if we've reached
// just past the last entry. Return whatever `aCallback` returns. Callback
// should not store `ProfileBufferEntryReader`, because it may become invalid
// after this call.
// Either/both chunk pointers may be null.
template <typename Callback>
[[nodiscard]] static auto ReadAt(ProfileBufferBlockIndex aMinimumBlockIndex,
const ProfileBufferChunk* aChunks0,
const ProfileBufferChunk* aChunks1,
Callback&& aCallback) {
static_assert(
std::is_invocable_v<Callback, Maybe<ProfileBufferEntryReader>&&>,
"ReadAt callback must take a Maybe<ProfileBufferEntryReader>&&");
Maybe<ProfileBufferEntryReader> maybeEntryReader;
if (profiler::detail::InChunkPointer p{aChunks0, aChunks1}; !p.IsNull()) {
// If the pointer position is before the given position, try to advance.
if (p.GlobalRangePosition() >=
aMinimumBlockIndex.ConvertToProfileBufferIndex() ||
p.AdvanceToGlobalRangePosition(
aMinimumBlockIndex.ConvertToProfileBufferIndex())) {
MOZ_ASSERT(p.GlobalRangePosition() >=
aMinimumBlockIndex.ConvertToProfileBufferIndex());
// Here we're pointing at the start of a block, try to read the entry
// size. (Entries cannot be empty, so 0 means failure.)
if (Length entrySize = p.ReadEntrySize(); entrySize != 0) {
maybeEntryReader.emplace(p.EntryReader(entrySize));
if (maybeEntryReader->RemainingBytes() == 0) {
// An empty entry reader means there was no complete block at the
// given index.
maybeEntryReader.reset();
} else {
MOZ_ASSERT(maybeEntryReader->RemainingBytes() == entrySize);
}
}
}
}
#ifdef DEBUG
auto assertAllRead = MakeScopeExit([&]() {
MOZ_ASSERT(!maybeEntryReader || maybeEntryReader->RemainingBytes() == 0);
});
#endif // DEBUG
return std::forward<Callback>(aCallback)(std::move(maybeEntryReader));
}
// Call `aCallback(Maybe<ProfileBufferEntryReader>&&)` on the entry at
// the given ProfileBufferBlockIndex; The `Maybe` will be `Nothing` if
// out-of-session, or if that entry doesn't exist anymore, or if we've reached
// just past the last entry. Return whatever `aCallback` returns. Callback
// should not store `ProfileBufferEntryReader`, because it may become invalid
// after this call.
template <typename Callback>
[[nodiscard]] auto ReadAt(ProfileBufferBlockIndex aBlockIndex,
Callback&& aCallback) const {
baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
if (MOZ_UNLIKELY(!mChunkManager)) {
// Out-of-session.
return std::forward<Callback>(aCallback)(Nothing{});
}
return mChunkManager->PeekExtantReleasedChunks(
[&](const ProfileBufferChunk* aOldestChunk) {
return ReadAt(aBlockIndex, aOldestChunk, mCurrentChunk.get(),
std::forward<Callback>(aCallback));
});
}
// Append the contents of another ProfileChunkedBuffer to this one.
ProfileBufferBlockIndex AppendContents(const ProfileChunkedBuffer& aSrc) {
ProfileBufferBlockIndex firstBlockIndex;
// If we start failing, we'll stop writing.
bool failed = false;
aSrc.ReadEach([&](ProfileBufferEntryReader& aER) {
if (failed) {
return;
}
failed =
!Put(aER.RemainingBytes(), [&](Maybe<ProfileBufferEntryWriter>& aEW) {
if (aEW.isNothing()) {
return false;
}
if (!firstBlockIndex) {
firstBlockIndex = aEW->CurrentBlockIndex();
}
aEW->WriteFromReader(aER, aER.RemainingBytes());
return true;
});
});
return failed ? nullptr : firstBlockIndex;
}
#ifdef DEBUG
void Dump(std::FILE* aFile = stdout) const {
baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
fprintf(aFile,
"ProfileChunkedBuffer[%p] State: range %u-%u pushed=%u cleared=%u "
"(live=%u) failed-puts=%u bytes",
this, unsigned(mRangeStart), unsigned(mRangeEnd),
unsigned(mPushedBlockCount), unsigned(mClearedBlockCount),
unsigned(mPushedBlockCount) - unsigned(mClearedBlockCount),
unsigned(mFailedPutBytes));
if (MOZ_UNLIKELY(!mChunkManager)) {
fprintf(aFile, " - Out-of-session\n");
return;
}
fprintf(aFile, " - chunks:\n");
bool hasChunks = false;
mChunkManager->PeekExtantReleasedChunks(
[&](const ProfileBufferChunk* aOldestChunk) {
for (const ProfileBufferChunk* chunk = aOldestChunk; chunk;
chunk = chunk->GetNext()) {
fprintf(aFile, "R ");
chunk->Dump(aFile);
hasChunks = true;
}
});
if (mCurrentChunk) {
fprintf(aFile, "C ");
mCurrentChunk->Dump(aFile);
hasChunks = true;
}
for (const ProfileBufferChunk* chunk = mNextChunks.get(); chunk;
chunk = chunk->GetNext()) {
fprintf(aFile, "N ");
chunk->Dump(aFile);
hasChunks = true;
}
switch (mRequestedChunkHolder->GetState()) {
case RequestedChunkRefCountedHolder::State::Unused:
fprintf(aFile, " - No request pending.\n");
break;
case RequestedChunkRefCountedHolder::State::Requested:
fprintf(aFile, " - Request pending.\n");
break;
case RequestedChunkRefCountedHolder::State::Fulfilled:
fprintf(aFile, " - Request fulfilled.\n");
break;
}
if (!hasChunks) {
fprintf(aFile, " No chunks.\n");
}
}
#endif // DEBUG
private:
// Used to de/serialize a ProfileChunkedBuffer (e.g., containing a backtrace).
friend ProfileBufferEntryWriter::Serializer<ProfileChunkedBuffer>;
friend ProfileBufferEntryReader::Deserializer<ProfileChunkedBuffer>;
friend ProfileBufferEntryWriter::Serializer<UniquePtr<ProfileChunkedBuffer>>;
friend ProfileBufferEntryReader::Deserializer<
UniquePtr<ProfileChunkedBuffer>>;
[[nodiscard]] UniquePtr<ProfileBufferChunkManager> ResetChunkManager(
const baseprofiler::detail::BaseProfilerMaybeAutoLock&) {
UniquePtr<ProfileBufferChunkManager> chunkManager;
if (mChunkManager) {
mRequestedChunkHolder = nullptr;
mChunkManager->ForgetUnreleasedChunks();
#ifdef DEBUG
mChunkManager->DeregisteredFrom(this);
#endif
mChunkManager = nullptr;
chunkManager = std::move(mOwnedChunkManager);
if (mCurrentChunk) {
mCurrentChunk->MarkDone();
mCurrentChunk = nullptr;
}
mNextChunks = nullptr;
mNextChunkRangeStart = mRangeEnd;
mRangeStart = mRangeEnd;
mPushedBlockCount = 0;
mClearedBlockCount = 0;
mFailedPutBytes = 0;
}
return chunkManager;
}
void SetChunkManager(
ProfileBufferChunkManager& aChunkManager,
const baseprofiler::detail::BaseProfilerMaybeAutoLock& aLock) {
MOZ_ASSERT(!mChunkManager);
mChunkManager = &aChunkManager;
#ifdef DEBUG
mChunkManager->RegisteredWith(this);
#endif
mChunkManager->SetChunkDestroyedCallback(
[this](const ProfileBufferChunk& aChunk) {
for (;;) {
ProfileBufferIndex rangeStart = mRangeStart;
if (MOZ_LIKELY(rangeStart <= aChunk.RangeStart())) {
if (MOZ_LIKELY(mRangeStart.compareExchange(
rangeStart,
aChunk.RangeStart() + aChunk.BufferBytes()))) {
break;
}
}
}
mClearedBlockCount += aChunk.BlockCount();
});
// We start with one chunk right away, and request a following one now
// so it should be available before the current chunk is full.
SetAndInitializeCurrentChunk(mChunkManager->GetChunk(), aLock);
mRequestedChunkHolder = MakeRefPtr<RequestedChunkRefCountedHolder>();
RequestChunk(aLock);
}
[[nodiscard]] size_t SizeOfExcludingThis(
MallocSizeOf aMallocSizeOf,
const baseprofiler::detail::BaseProfilerMaybeAutoLock&) const {
if (MOZ_UNLIKELY(!mChunkManager)) {
// Out-of-session.
return 0;
}
size_t size = mChunkManager->SizeOfIncludingThis(aMallocSizeOf);
if (mCurrentChunk) {
size += mCurrentChunk->SizeOfIncludingThis(aMallocSizeOf);
}
if (mNextChunks) {
size += mNextChunks->SizeOfIncludingThis(aMallocSizeOf);
}
return size;
}
void InitializeCurrentChunk(
const baseprofiler::detail::BaseProfilerMaybeAutoLock&) {
MOZ_ASSERT(!!mCurrentChunk);
mCurrentChunk->SetRangeStart(mNextChunkRangeStart);
mNextChunkRangeStart += mCurrentChunk->BufferBytes();
Unused << mCurrentChunk->ReserveInitialBlockAsTail(0);
}
void SetAndInitializeCurrentChunk(
UniquePtr<ProfileBufferChunk>&& aChunk,
const baseprofiler::detail::BaseProfilerMaybeAutoLock& aLock) {
mCurrentChunk = std::move(aChunk);
if (mCurrentChunk) {
InitializeCurrentChunk(aLock);
}
}
void RequestChunk(
const baseprofiler::detail::BaseProfilerMaybeAutoLock& aLock) {
if (HandleRequestedChunk_IsPending(aLock)) {
// There is already a pending request, don't start a new one.
return;
}
// Ensure the `RequestedChunkHolder` knows we're starting a request.
mRequestedChunkHolder->StartRequest();
// Request a chunk, the callback carries a `RefPtr` of the
// `RequestedChunkHolder`, so it's guaranteed to live until it's invoked,
// even if this `ProfileChunkedBuffer` changes its `ChunkManager` or is
// destroyed.
mChunkManager->RequestChunk(
[requestedChunkHolder = RefPtr<RequestedChunkRefCountedHolder>(
mRequestedChunkHolder)](UniquePtr<ProfileBufferChunk> aChunk) {
requestedChunkHolder->AddRequestedChunk(std::move(aChunk));
});
}
[[nodiscard]] bool HandleRequestedChunk_IsPending(
const baseprofiler::detail::BaseProfilerMaybeAutoLock& aLock) {
MOZ_ASSERT(!!mChunkManager);
MOZ_ASSERT(!!mRequestedChunkHolder);
if (mRequestedChunkHolder->GetState() ==
RequestedChunkRefCountedHolder::State::Unused) {
return false;
}
// A request is either in-flight or fulfilled.
Maybe<UniquePtr<ProfileBufferChunk>> maybeChunk =
mRequestedChunkHolder->GetChunkIfFulfilled();
if (maybeChunk.isNothing()) {
// Request is still pending.
return true;
}
// Since we extracted the provided chunk, the holder should now be unused.
MOZ_ASSERT(mRequestedChunkHolder->GetState() ==
RequestedChunkRefCountedHolder::State::Unused);
// Request has been fulfilled.
UniquePtr<ProfileBufferChunk>& chunk = *maybeChunk;
if (chunk) {
// Try to use as current chunk if needed.
if (!mCurrentChunk) {
SetAndInitializeCurrentChunk(std::move(chunk), aLock);
// We've just received a chunk and made it current, request a next chunk
// for later.
MOZ_ASSERT(!mNextChunks);
RequestChunk(aLock);
return true;
}
if (!mNextChunks) {
mNextChunks = std::move(chunk);
} else {
mNextChunks->InsertNext(std::move(chunk));
}
}
return false;
}
// Get a pointer to the next chunk available
[[nodiscard]] ProfileBufferChunk* GetOrCreateCurrentChunk(
const baseprofiler::detail::BaseProfilerMaybeAutoLock& aLock) {
ProfileBufferChunk* current = mCurrentChunk.get();
if (MOZ_UNLIKELY(!current)) {
// No current chunk ready.
MOZ_ASSERT(!mNextChunks,
"There shouldn't be next chunks when there is no current one");
// See if a request has recently been fulfilled, ignore pending status.
Unused << HandleRequestedChunk_IsPending(aLock);
current = mCurrentChunk.get();
if (MOZ_UNLIKELY(!current)) {
// There was no pending chunk, try to get one right now.
// This may still fail, but we can't do anything else about it, the
// caller must handle the nullptr case.
// Attempt a request for later.
SetAndInitializeCurrentChunk(mChunkManager->GetChunk(), aLock);
current = mCurrentChunk.get();
}
}
return current;
}
// Get a pointer to the next chunk available
[[nodiscard]] ProfileBufferChunk* GetOrCreateNextChunk(
const baseprofiler::detail::BaseProfilerMaybeAutoLock& aLock) {
MOZ_ASSERT(!!mCurrentChunk,
"Why ask for a next chunk when there isn't even a current one?");
ProfileBufferChunk* next = mNextChunks.get();
if (MOZ_UNLIKELY(!next)) {
// No next chunk ready, see if a request has recently been fulfilled,
// ignore pending status.
Unused << HandleRequestedChunk_IsPending(aLock);
next = mNextChunks.get();
if (MOZ_UNLIKELY(!next)) {
// There was no pending chunk, try to get one right now.
mNextChunks = mChunkManager->GetChunk();
next = mNextChunks.get();
// This may still fail, but we can't do anything else about it, the
// caller must handle the nullptr case.
if (MOZ_UNLIKELY(!next)) {
// Attempt a request for later.
RequestChunk(aLock);
}
}
}
return next;
}
// Reserve a block of `aCallbackBlockBytes()` size, and invoke and return
// `aCallback(Maybe<ProfileBufferEntryWriter>&)`. Note that this is the "raw"
// version that doesn't write the entry size at the beginning of the block.
// Note: `aCallbackBlockBytes` is a callback instead of a simple value, to
// delay this potentially-expensive computation until after we're checked that
// we're in-session; use `Put(Length, Callback)` below if you know the size
// already.
template <typename CallbackBlockBytes, typename Callback>
auto ReserveAndPutRaw(CallbackBlockBytes&& aCallbackBlockBytes,
Callback&& aCallback,
baseprofiler::detail::BaseProfilerMaybeAutoLock& aLock,
uint64_t aBlockCount = 1) {
// The entry writer that will point into one or two chunks to write
// into, empty by default (failure).
Maybe<ProfileBufferEntryWriter> maybeEntryWriter;
// The current chunk will be filled if we need to write more than its
// remaining space.
bool currentChunkFilled = false;
// If the current chunk gets filled, we may or may not initialize the next
// chunk!
bool nextChunkInitialized = false;
if (MOZ_LIKELY(mChunkManager)) {
// In-session.
const Length blockBytes =
std::forward<CallbackBlockBytes>(aCallbackBlockBytes)();
if (ProfileBufferChunk* current = GetOrCreateCurrentChunk(aLock);
MOZ_LIKELY(current)) {
if (blockBytes <= current->RemainingBytes()) {
// Block fits in current chunk with only one span.
currentChunkFilled = blockBytes == current->RemainingBytes();
const auto [mem0, blockIndex] = current->ReserveBlock(blockBytes);
MOZ_ASSERT(mem0.LengthBytes() == blockBytes);
maybeEntryWriter.emplace(
mem0, blockIndex,
ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
blockIndex.ConvertToProfileBufferIndex() + blockBytes));
MOZ_ASSERT(maybeEntryWriter->RemainingBytes() == blockBytes);
mRangeEnd += blockBytes;
mPushedBlockCount += aBlockCount;
} else if (blockBytes >= current->BufferBytes()) {
// Currently only two buffer chunks are held at a time and it is not
// possible to write an object that takes up more space than this. In
// this scenario, silently discard this block of data if it is unable
// to fit into the two reserved profiler chunks.
mFailedPutBytes += blockBytes;
} else {
// Block doesn't fit fully in current chunk, it needs to overflow into
// the next one.
// Whether or not we can write this entry, the current chunk is now
// considered full, so it will be released. (Otherwise we could refuse
// this entry, but later accept a smaller entry into this chunk, which
// would be somewhat inconsistent.)
currentChunkFilled = true;
// Make sure the next chunk is available (from a previous request),
// otherwise create one on the spot.
if (ProfileBufferChunk* next = GetOrCreateNextChunk(aLock);
MOZ_LIKELY(next)) {
// Here, we know we have a current and a next chunk.
// Reserve head of block at the end of the current chunk.
const auto [mem0, blockIndex] =
current->ReserveBlock(current->RemainingBytes());
MOZ_ASSERT(mem0.LengthBytes() < blockBytes);
MOZ_ASSERT(current->RemainingBytes() == 0);
// Set the next chunk range, and reserve the needed space for the
// tail of the block.
next->SetRangeStart(mNextChunkRangeStart);
mNextChunkRangeStart += next->BufferBytes();
const auto mem1 = next->ReserveInitialBlockAsTail(
blockBytes - mem0.LengthBytes());
MOZ_ASSERT(next->RemainingBytes() != 0);
nextChunkInitialized = true;
// Block is split in two spans.
maybeEntryWriter.emplace(
mem0, mem1, blockIndex,
ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
blockIndex.ConvertToProfileBufferIndex() + blockBytes));
MOZ_ASSERT(maybeEntryWriter->RemainingBytes() == blockBytes);
mRangeEnd += blockBytes;
mPushedBlockCount += aBlockCount;
} else {
// Cannot get a new chunk. Record put failure.
mFailedPutBytes += blockBytes;
}
}
} else {
// Cannot get a current chunk. Record put failure.
mFailedPutBytes += blockBytes;
}
} // end of `if (MOZ_LIKELY(mChunkManager))`
// Here, we either have a `Nothing` (failure), or a non-empty entry writer
// pointing at the start of the block.
// After we invoke the callback and return, we may need to handle the
// current chunk being filled.
auto handleFilledChunk = MakeScopeExit([&]() {
// If the entry writer was not already empty, the callback *must* have
// filled the full entry.
MOZ_ASSERT(!maybeEntryWriter || maybeEntryWriter->RemainingBytes() == 0);
if (currentChunkFilled) {
// Extract current (now filled) chunk.
UniquePtr<ProfileBufferChunk> filled = std::move(mCurrentChunk);
if (mNextChunks) {
// Cycle to the next chunk.
mCurrentChunk =
std::exchange(mNextChunks, mNextChunks->ReleaseNext());
// Make sure it is initialized (it is now the current chunk).
if (!nextChunkInitialized) {
InitializeCurrentChunk(aLock);
}
}
// And finally mark filled chunk done and release it.
filled->MarkDone();
mChunkManager->ReleaseChunk(std::move(filled));
// Request another chunk if needed.
// In most cases, here we should have one current chunk and no next
// chunk, so we want to do a request so there hopefully will be a next
// chunk available when the current one gets filled.
// But we also for a request if we don't even have a current chunk (if
// it's too late, it's ok because the next `ReserveAndPutRaw` wil just
// allocate one on the spot.)
// And if we already have a next chunk, there's no need for more now.
if (!mCurrentChunk || !mNextChunks) {
RequestChunk(aLock);
}
}
});
return std::forward<Callback>(aCallback)(maybeEntryWriter);
}
// Reserve a block of `aBlockBytes` size, and invoke and return
// `aCallback(Maybe<ProfileBufferEntryWriter>&)`. Note that this is the "raw"
// version that doesn't write the entry size at the beginning of the block.
template <typename Callback>
auto ReserveAndPutRaw(Length aBlockBytes, Callback&& aCallback,
uint64_t aBlockCount) {
baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
return ReserveAndPutRaw([aBlockBytes]() { return aBlockBytes; },
std::forward<Callback>(aCallback), lock,
aBlockCount);
}
// Mutex guarding the following members.
mutable baseprofiler::detail::BaseProfilerMaybeMutex mMutex;
// Pointer to the current Chunk Manager (or null when out-of-session.)
// It may be owned locally (see below) or externally.
ProfileBufferChunkManager* mChunkManager = nullptr;
// Only non-null when we own the current Chunk Manager.
UniquePtr<ProfileBufferChunkManager> mOwnedChunkManager;
UniquePtr<ProfileBufferChunk> mCurrentChunk;
UniquePtr<ProfileBufferChunk> mNextChunks;
// Class used to transfer requested chunks from a `ChunkManager` to a
// `ProfileChunkedBuffer`.
// It needs to be ref-counted because the request may be fulfilled
// asynchronously, and either side may be destroyed during the request.
// It cannot use the `ProfileChunkedBuffer` mutex, because that buffer and its
// mutex could be destroyed during the request.
class RequestedChunkRefCountedHolder {
public:
enum class State { Unused, Requested, Fulfilled };
// Get the current state. Note that it may change after the function
// returns, so it should be used carefully, e.g., `ProfileChunkedBuffer` can
// see if a request is pending or fulfilled, to avoid starting another
// request.
[[nodiscard]] State GetState() const {
baseprofiler::detail::BaseProfilerAutoLock lock(mRequestMutex);
return mState;
}
// Must be called by `ProfileChunkedBuffer` when it requests a chunk.
// There cannot be more than one request in-flight.
void StartRequest() {
baseprofiler::detail::BaseProfilerAutoLock lock(mRequestMutex);
MOZ_ASSERT(mState == State::Unused, "Already requested or fulfilled");
mState = State::Requested;
}
// Must be called by the `ChunkManager` with a chunk.
// If the `ChunkManager` cannot provide a chunk (because of memory limits,
// or it gets destroyed), it must call this anyway with a nullptr.
void AddRequestedChunk(UniquePtr<ProfileBufferChunk>&& aChunk) {
baseprofiler::detail::BaseProfilerAutoLock lock(mRequestMutex);
MOZ_ASSERT(mState == State::Requested);
mState = State::Fulfilled;
mRequestedChunk = std::move(aChunk);
}
// The `ProfileChunkedBuffer` can try to extract the provided chunk after a
// request:
// - Nothing -> Request is not fulfilled yet.
// - Some(nullptr) -> The `ChunkManager` was not able to provide a chunk.
// - Some(chunk) -> Requested chunk.
[[nodiscard]] Maybe<UniquePtr<ProfileBufferChunk>> GetChunkIfFulfilled() {
Maybe<UniquePtr<ProfileBufferChunk>> maybeChunk;
baseprofiler::detail::BaseProfilerAutoLock lock(mRequestMutex);
MOZ_ASSERT(mState == State::Requested || mState == State::Fulfilled);
if (mState == State::Fulfilled) {
mState = State::Unused;
maybeChunk.emplace(std::move(mRequestedChunk));
}
return maybeChunk;
}
// Ref-counting implementation. Hand-rolled, because mozilla::RefCounted
// logs AddRefs and Releases in xpcom, but this object could be AddRef'd
// by the Base Profiler before xpcom starts, then Release'd by the Gecko
// Profiler in xpcom, leading to apparent negative leaks.
void AddRef() {
baseprofiler::detail::BaseProfilerAutoLock lock(mRequestMutex);
++mRefCount;
}
void Release() {
{
baseprofiler::detail::BaseProfilerAutoLock lock(mRequestMutex);
if (--mRefCount > 0) {
return;
}
}
delete this;
}
private:
~RequestedChunkRefCountedHolder() = default;
// Mutex guarding the following members.
mutable baseprofiler::detail::BaseProfilerMutex mRequestMutex;
int mRefCount = 0;
State mState = State::Unused;
UniquePtr<ProfileBufferChunk> mRequestedChunk;
};
// Requested-chunk holder, kept alive when in-session, but may also live
// longer if a request is in-flight.
RefPtr<RequestedChunkRefCountedHolder> mRequestedChunkHolder;
// Range start of the next chunk to become current. Starting at 1 because
// 0 is a reserved index similar to nullptr.
ProfileBufferIndex mNextChunkRangeStart = 1;
// Index to the first block.
// Atomic because it may be increased when a Chunk is destroyed, and the
// callback may be invoked from anywhere, including from inside one of our
// locked section, so we cannot protect it with a mutex.
Atomic<ProfileBufferIndex, MemoryOrdering::ReleaseAcquire> mRangeStart{1};
// Index past the last block. Equals mRangeStart if empty.
ProfileBufferIndex mRangeEnd = 1;
// Number of blocks that have been pushed into this buffer.
uint64_t mPushedBlockCount = 0;
// Number of blocks that have been removed from this buffer.
// Note: Live entries = pushed - cleared.
// Atomic because it may be updated when a Chunk is destroyed, and the
// callback may be invoked from anywhere, including from inside one of our
// locked section, so we cannot protect it with a mutex.
Atomic<uint64_t, MemoryOrdering::ReleaseAcquire> mClearedBlockCount{0};
// Number of bytes that could not be put into this buffer.
uint64_t mFailedPutBytes = 0;
};
// ----------------------------------------------------------------------------
// ProfileChunkedBuffer serialization
// A ProfileChunkedBuffer can hide another one!
// This will be used to store marker backtraces; They can be read back into a
// UniquePtr<ProfileChunkedBuffer>.
// Format: len (ULEB128) | start | end | buffer (len bytes) | pushed | cleared
// len==0 marks an out-of-session buffer, or empty buffer.
template <>
struct ProfileBufferEntryWriter::Serializer<ProfileChunkedBuffer> {
static Length Bytes(const ProfileChunkedBuffer& aBuffer) {
return aBuffer.Read([&](ProfileChunkedBuffer::Reader* aReader) {
if (!aReader) {
// Out-of-session, we only need 1 byte to store a length of 0.
return ULEB128Size<Length>(0);
}
ProfileBufferEntryReader reader = aReader->SingleChunkDataAsEntry();
const ProfileBufferIndex start =
reader.CurrentBlockIndex().ConvertToProfileBufferIndex();
const ProfileBufferIndex end =
reader.NextBlockIndex().ConvertToProfileBufferIndex();
MOZ_ASSERT(end - start <= std::numeric_limits<Length>::max());
const Length len = static_cast<Length>(end - start);
if (len == 0) {
// In-session but empty, also store a length of 0.
return ULEB128Size<Length>(0);
}
// In-session.
return static_cast<Length>(ULEB128Size(len) + sizeof(start) + len +
sizeof(aBuffer.mPushedBlockCount) +
sizeof(aBuffer.mClearedBlockCount));
});
}
static void Write(ProfileBufferEntryWriter& aEW,
const ProfileChunkedBuffer& aBuffer) {
aBuffer.Read([&](ProfileChunkedBuffer::Reader* aReader) {
if (!aReader) {
// Out-of-session, only store a length of 0.
aEW.WriteULEB128<Length>(0);
return;
}
ProfileBufferEntryReader reader = aReader->SingleChunkDataAsEntry();
const ProfileBufferIndex start =
reader.CurrentBlockIndex().ConvertToProfileBufferIndex();
const ProfileBufferIndex end =
reader.NextBlockIndex().ConvertToProfileBufferIndex();
MOZ_ASSERT(end - start <= std::numeric_limits<Length>::max());
const Length len = static_cast<Length>(end - start);
MOZ_ASSERT(len <= aEW.RemainingBytes());
if (len == 0) {
// In-session but empty, only store a length of 0.
aEW.WriteULEB128<Length>(0);
return;
}
// In-session.
// Store buffer length, and start index.
aEW.WriteULEB128(len);
aEW.WriteObject(start);
// Write all the bytes.
aEW.WriteFromReader(reader, reader.RemainingBytes());
// And write stats.
aEW.WriteObject(static_cast<uint64_t>(aBuffer.mPushedBlockCount));
aEW.WriteObject(static_cast<uint64_t>(aBuffer.mClearedBlockCount));
// Note: Failed pushes are not important to serialize.
});
}
};
// A serialized ProfileChunkedBuffer can be read into an empty buffer (either
// out-of-session, or in-session with enough room).
template <>
struct ProfileBufferEntryReader::Deserializer<ProfileChunkedBuffer> {
static void ReadInto(ProfileBufferEntryReader& aER,
ProfileChunkedBuffer& aBuffer) {
// Expect an empty buffer, as we're going to overwrite it.
MOZ_ASSERT(aBuffer.GetState().mRangeStart == aBuffer.GetState().mRangeEnd);
// Read the stored buffer length.
const auto len = aER.ReadULEB128<ProfileChunkedBuffer::Length>();
if (len == 0) {
// 0-length means an "uninteresting" buffer, just return now.
return;
}
// We have a non-empty buffer to read.
// Read start and end indices.
const auto start = aER.ReadObject<ProfileBufferIndex>();
aBuffer.mRangeStart = start;
// For now, set the end to be the start (the buffer is still empty). It will
// be updated in `ReserveAndPutRaw()` below.
aBuffer.mRangeEnd = start;
if (aBuffer.IsInSession()) {
// Output buffer is in-session (i.e., it already has a memory buffer
// attached). Make sure the caller allocated enough space.
MOZ_RELEASE_ASSERT(aBuffer.BufferLength().value() >= len);
} else {
// Output buffer is out-of-session, set a new chunk manager that will
// provide a single chunk of just the right size.
aBuffer.SetChunkManager(MakeUnique<ProfileBufferChunkManagerSingle>(len));
MOZ_ASSERT(aBuffer.BufferLength().value() >= len);
}
// Copy bytes into the buffer.
aBuffer.ReserveAndPutRaw(
len,
[&](Maybe<ProfileBufferEntryWriter>& aEW) {
MOZ_RELEASE_ASSERT(aEW.isSome());
aEW->WriteFromReader(aER, len);
},
0);
// Finally copy stats.
aBuffer.mPushedBlockCount = aER.ReadObject<uint64_t>();
aBuffer.mClearedBlockCount = aER.ReadObject<uint64_t>();
// Failed puts are not important to keep.
aBuffer.mFailedPutBytes = 0;
}
// We cannot output a ProfileChunkedBuffer object (not copyable), use
// `ReadInto()` or `aER.ReadObject<UniquePtr<BlocksRinbBuffer>>()` instead.
static ProfileChunkedBuffer Read(ProfileBufferEntryReader& aER) = delete;
};
// A ProfileChunkedBuffer is usually refererenced through a UniquePtr, for
// convenience we support (de)serializing that UniquePtr directly.
// This is compatible with the non-UniquePtr serialization above, with a null
// pointer being treated like an out-of-session or empty buffer; and any of
// these would be deserialized into a null pointer.
template <>
struct ProfileBufferEntryWriter::Serializer<UniquePtr<ProfileChunkedBuffer>> {
static Length Bytes(const UniquePtr<ProfileChunkedBuffer>& aBufferUPtr) {
if (!aBufferUPtr) {
// Null pointer, treat it like an empty buffer, i.e., write length of 0.
return ULEB128Size<Length>(0);
}
// Otherwise write the pointed-at ProfileChunkedBuffer (which could be
// out-of-session or empty.)
return SumBytes(*aBufferUPtr);
}
static void Write(ProfileBufferEntryWriter& aEW,
const UniquePtr<ProfileChunkedBuffer>& aBufferUPtr) {
if (!aBufferUPtr) {
// Null pointer, treat it like an empty buffer, i.e., write length of 0.
aEW.WriteULEB128<Length>(0);
return;
}
// Otherwise write the pointed-at ProfileChunkedBuffer (which could be
// out-of-session or empty.)
aEW.WriteObject(*aBufferUPtr);
}
};
// Serialization of a raw pointer to ProfileChunkedBuffer.
// Use Deserializer<UniquePtr<ProfileChunkedBuffer>> to read it back.
template <>
struct ProfileBufferEntryWriter::Serializer<ProfileChunkedBuffer*> {
static Length Bytes(ProfileChunkedBuffer* aBufferUPtr) {
if (!aBufferUPtr) {
// Null pointer, treat it like an empty buffer, i.e., write length of 0.
return ULEB128Size<Length>(0);
}
// Otherwise write the pointed-at ProfileChunkedBuffer (which could be
// out-of-session or empty.)
return SumBytes(*aBufferUPtr);
}
static void Write(ProfileBufferEntryWriter& aEW,
ProfileChunkedBuffer* aBufferUPtr) {
if (!aBufferUPtr) {
// Null pointer, treat it like an empty buffer, i.e., write length of 0.
aEW.WriteULEB128<Length>(0);
return;
}
// Otherwise write the pointed-at ProfileChunkedBuffer (which could be
// out-of-session or empty.)
aEW.WriteObject(*aBufferUPtr);
}
};
template <>
struct ProfileBufferEntryReader::Deserializer<UniquePtr<ProfileChunkedBuffer>> {
static void ReadInto(ProfileBufferEntryReader& aER,
UniquePtr<ProfileChunkedBuffer>& aBuffer) {
aBuffer = Read(aER);
}
static UniquePtr<ProfileChunkedBuffer> Read(ProfileBufferEntryReader& aER) {
UniquePtr<ProfileChunkedBuffer> bufferUPtr;
// Keep a copy of the reader before reading the length, so we can restart
// from here below.
ProfileBufferEntryReader readerBeforeLen = aER;
// Read the stored buffer length.
const auto len = aER.ReadULEB128<ProfileChunkedBuffer::Length>();
if (len == 0) {
// 0-length means an "uninteresting" buffer, just return nullptr.
return bufferUPtr;
}
// We have a non-empty buffer.
// allocate an empty ProfileChunkedBuffer without mutex.
bufferUPtr = MakeUnique<ProfileChunkedBuffer>(
ProfileChunkedBuffer::ThreadSafety::WithoutMutex);
// Rewind the reader before the length and deserialize the contents, using
// the non-UniquePtr Deserializer.
aER = readerBeforeLen;
aER.ReadIntoObject(*bufferUPtr);
return bufferUPtr;
}
};
} // namespace mozilla
#endif // ProfileChunkedBuffer_h