Use AUTO_PROFILER_STATS in both profilers, in:
- SamplerThread::Run() calling DoPeriodicSample()
- racy_profiler_add_marker
- ProfileBuffer::DeleteExpiredStoredMarkers()
This should cover all areas affected by the upcoming changes to the
ProfileBuffer storage, and how markers are stored.
Differential Revision: https://phabricator.services.mozilla.com/D42826
--HG--
extra : moz-landing-system : lando
AUTO_PROFILER_STATS(name) can be used to time a {block}.
Statistics are gathered in a function-static variable, and printf'd when the
program ends.
Differential Revision: https://phabricator.services.mozilla.com/D42825
--HG--
extra : moz-landing-system : lando
Some users will want to lock the buffer but not do any specific operation with
it.
Differential Revision: https://phabricator.services.mozilla.com/D42824
--HG--
extra : moz-landing-system : lando
Backtraces (that are kept in some marker payloads) are stored in a small
ProfileBuffer, we will need to store that data, which will happen to be inside
a BlockRingBuffer, so BlockRingBuffer needs to be able to (de)serialize itself!
This is done by storing the contents in the active buffer range, and some extra
data, to later reconstruct a BlocksRingBuffer that looks like the original.
Depends on D42496
Differential Revision: https://phabricator.services.mozilla.com/D42634
--HG--
extra : moz-landing-system : lando
Markers and their payloads contain all kinds of objects that we'll need to
serialize into a BlocksRingBuffer (new ProfileBuffer storage).
This patch will add functions to:
- Compute the size needed to store objects,
- Write multiple objects into a BlockRingBuffer entry,
- Read objects back from an entry.
And it will provide a number of useful de/serialization helpers for:
- Trivially-copyable objects,
- Strings of different types,
- Raw pointers (with some safety guards to avoid surprises),
- Tuples (to store multiple sub-objects),
- Spans,
- Maybe (for optional objects),
- Variant.
This should be enough to store most kinds of data. Further specializations
can&will be written as necessary for more complex or obscure types.
Differential Revision: https://phabricator.services.mozilla.com/D42496
--HG--
extra : moz-landing-system : lando
This allows its use in std algorithms and types that require a real iterator
(like `template <typename InputIt> std::string::string(InputIt, InputIt)`).
Differential Revision: https://phabricator.services.mozilla.com/D42452
--HG--
extra : moz-landing-system : lando
`Reader::At()` can be used to get a `BlockIterator` at a given `BlockIndex`,
clamped between `begin()` and `end()`.
This will be useful when we want to iterate starting at a given index, e.g.,
when duplicating stacks.
Differential Revision: https://phabricator.services.mozilla.com/D42449
--HG--
extra : moz-landing-system : lando
The main goal of these bugs is to move markers to a new storage, so I'm adding
lots of markers to TestBaseProfiler.
Also adding labels, easier to read unsymbolicated profiles, and gives a bit more
coverage too.
And adding a separate "fibonacci canceller" thread, which is needed on some
slower platforms (e.g., Linux 64 ASAN times out otherwise); as a bonus, this
tests AUTO_BASE_PROFILER_REGISTER_THREAD.
Differential Revision: https://phabricator.services.mozilla.com/D42448
--HG--
extra : moz-landing-system : lando
(Also MOZ_BASE_PROFILER_STARTUP.)
This makes it easier to keep that variable setup in the environment, and change
its value to switch between enabling and disabling the profiler.
Differential Revision: https://phabricator.services.mozilla.com/D42447
--HG--
extra : moz-landing-system : lando
In practice the Reader doesn't need to be copied/moved/reassign.
BlocksRingBuffer::Read() can just instantiate one on the stack, and pass it by
reference to callbacks.
Differential Revision: https://phabricator.services.mozilla.com/D42118
--HG--
extra : moz-landing-system : lando
The point of the EntryReserver was mainly to have an object that represented a
writing lock on BlocksRingBuffer, so potentially perform multiple consecutive
writes.
After some experience implementing bug 1562604, there's actually no need for it.
So instead of having `Put()` create an `EntryReserver`, we now have
`ReserveAndPut()` that does the whole work in one function.
Differential Revision: https://phabricator.services.mozilla.com/D42116
--HG--
extra : moz-landing-system : lando
EntryWriter doesn't even need to be moveable, as BlocksRingBuffer can just
create one on the stack, and pass it by reference to callbacks.
This removes risks, and potential data copies.
Differential Revision: https://phabricator.services.mozilla.com/D42115
--HG--
extra : moz-landing-system : lando
Actually, we use clang for all Linux and Android platform, so it is no reason to use frame pointer based stack walker for LUL on Android/x86 and Android/x86-64 if no DWARF rule.
Differential Revision: https://phabricator.services.mozilla.com/D41320
--HG--
extra : moz-landing-system : lando
While mozglue continues to be the correct location for calling the affected
code in this patch, the calls requiring profiler labels will soon be
originating from firefox.exe via the launcher process.
mozglue will be supplying the launcher process with an interface that consists
of what are effectively "OnBeginDllLoad" and "OnEndDllLoad" callback
notifications; obviously an RAII class is not going to be useful for that case.
We still want to keep the RAII stuff around, however, since we still need it
for cases where we need to fall back to using the legacy DLL blocklist.
Differential Revision: https://phabricator.services.mozilla.com/D41807
--HG--
extra : moz-landing-system : lando
While mozglue continues to be the correct location for calling the affected
code in this patch, the calls requiring stackwalk suppression will soon be
originating from firefox.exe via the launcher process.
mozglue will be supplying the launcher process with an interface that consists
of what are effectively "OnBeginDllLoad" and "OnEndDllLoad" callback
notifications; obviously an RAII class is not going to be useful for that case.
We still want to keep the RAII stuff around, however, since we still need it
for cases where we need to fall back to using the legacy DLL blocklist.
Differential Revision: https://phabricator.services.mozilla.com/D41808
--HG--
extra : moz-landing-system : lando
`BlocksRingBuffer` will be used both inside and outside `ProfileBuffer`:
- Inside to serve as `ProfileBuffer`'s main storage for stack traces,
- Outside to allow marker storage even when `ProfileBuffer` is locked during
stack sampling.
`ProfileBuffer` only exists while `ActivePS` is alive, but because of the
potential outside accesses above (due to small races between ProfileBuffer
shutdown, and thread-local IsBeingProfiled() flags), we cannot just do the same
for BlocksRingBuffer, and it must remain alive to gracefully deny these accesses
around the profiler startup and shutdown times.
To accomplish this, `BlocksRingBuffer` may be in different states:
- "In-session", we have a real buffer to write to and read from,
- "Out-of-session", without buffer so reads&writes do nothing.
This is implemented by enclosing the underlying `ModuloBuffer` and the entry
deleter in a `Maybe`, which may be `Nothing` when the profiler is not running
and the `ProfileBuffer`'s `BlocksRingBuffer` is out-of-session.
Differential Revision: https://phabricator.services.mozilla.com/D41519
--HG--
extra : moz-landing-system : lando
`BlocksRingBuffer` will be used both inside and outside `ProfileBuffer`:
- Inside to serve as `ProfileBuffer`'s main storage for stack traces,
- Outside to allow marker storage even when `ProfileBuffer` is locked during
stack sampling.
`ProfileBuffer` only exists while `ActivePS` is alive, but because of the
potential outside accesses above (due to small races between ProfileBuffer
shutdown, and thread-local IsBeingProfiled() flags), we cannot just do the same
for BlocksRingBuffer, and it must remain alive to gracefully deny these accesses
around the profiler startup and shutdown times.
To accomplish this, `BlocksRingBuffer` may be in different states:
- "In-session", we have a real buffer to write to and read from,
- "Out-of-session", without buffer so reads&writes do nothing.
This is implemented by enclosing the underlying `ModuloBuffer` and the entry
deleter in a `Maybe`, which may be `Nothing` when the profiler is not running
and the `ProfileBuffer`'s `BlocksRingBuffer` is out-of-session.
Differential Revision: https://phabricator.services.mozilla.com/D41519
--HG--
extra : moz-landing-system : lando
After some bad experiences, I think EntryReader should be move-only:
- It needs to be moveable so it can be created from a function, and move-
constructed into a Maybe<> if needed.
- It can be passed around as a reference.
Previously, it could be passed by value, but it was too easy to create bugs,
e.g.: A function delegates to a sub-function to read something at the beginning,
then the first function wants to read more past that, but if the reader was
passed by value the first function would not see past what the sub-function did
read.
As a bonus, `mRing` can now be a reference instead of a pointer, and other
members can be const.
Differential Revision: https://phabricator.services.mozilla.com/D40958
--HG--
extra : moz-landing-system : lando
This patch does two things:
1. We refactor the resolution of function pointer and return type so that we
may support additional calling conventions besides just __stdcall;
2. We refactor DynamicallyLinkedFunctionPtr into a base class, and create
StaticDynamicallyLinkedFunctionPtr to specifically handle the static local
use case.
Differential Revision: https://phabricator.services.mozilla.com/D40885
--HG--
extra : moz-landing-system : lando
It makes little sense to copy a writer (also an output iterator).
We're keeping it move-constructible (so it can be passed around for construction
purposes), but not move-assignable to help make more members `const`.
Differential Revision: https://phabricator.services.mozilla.com/D40622
--HG--
extra : moz-landing-system : lando
This makes it easier to grab all BlocksRingBuffer state variables:
- Range start and end.
- Number of pushed blocks/entries, number of cleared blocks/entries.
The function is thread-safe, and the returned values are consistent with each
other, but they may become stale straight after the function returns (and the
lock is released).
They are still valuable to statistics, and to know how far the range has at
least reached (but may go further soon).
Differential Revision: https://phabricator.services.mozilla.com/D40621
--HG--
extra : moz-landing-system : lando
Renamed `BPAutoLock` to `BaseProfilerAutoLock`.
DEBUG-build `~BaseProfilerMutex()` checks that it is unlocked.
Prevent `BaseProfilerMutex` and `BaseProfilerAutoLock` copies&moves.
DEBUG-build check that `Lock()` sees `mOwningThreadId`==0 (because that is the
initial value, and the value after a previous `Unlock()`).
Don't preserve atomic `mOwningThreadId` in JS recording.
Differential Revision: https://phabricator.services.mozilla.com/D40620
--HG--
extra : moz-landing-system : lando
`ProfilerMarkerPayload::Set...()` functions are only used by derived classes in
the same files, and these values could just be set during construction.
Differential Revision: https://phabricator.services.mozilla.com/D40619
--HG--
extra : moz-landing-system : lando
cppunittest TestBaseProfiler and gtest GeckoProfiler.Markers now show overhead
stats.
(Separate patch, because we may want to remove them after a while.)
Differential Revision: https://phabricator.services.mozilla.com/D39642
--HG--
extra : moz-landing-system : lando
`ProfileBuffer` is now responsible for collecting overhead stats, and adding
them to the struct returned by `profiler_get_buffer_info()`.
Differential Revision: https://phabricator.services.mozilla.com/D39641
--HG--
extra : moz-landing-system : lando
`SamplerThread` inheriting from `Sampler` was a bit confusing, and scary with no
virtual destructor&functions.
`SamplerThread` only uses `Sampler`'s `Disable()` and
`SuspendAndSampleAndResumeThread()` functions, and `SamplerThread` is never
accessed through a `Sampler` reference/pointer.
So `SamplerThread` can just own a `Sampler` to make that relationship clearer.
Differential Revision: https://phabricator.services.mozilla.com/D39640
--HG--
extra : moz-landing-system : lando
This of course checks that the mutex is locked as expected in non-public APIs.
It also checks that user callbacks will not keep readers/writers longer than
they should.
Differential Revision: https://phabricator.services.mozilla.com/D39625
--HG--
extra : moz-landing-system : lando
`BaeProfilerMutex` is a concrete mutex based on MutexImpl, which was previously
implemented twice in both platform.h and BlocksRingBuffer.h.
This combined mutex has some DEBUG code (when MOZ_BASE_PROFILER is #defined) to
catch recursive locking, and to assert that the mutex is held (for code that
cannot easily use the "proof of lock" pattern; e.g., going through user-provided
callbacks).
This class needs to be public (because it is used in public headers), but is an
implementation detail, so it is located in a new header
"mozilla/BaseProfilerDetail.h" that will collect `mozilla::baseprofiler::detail`
code that may be useful to a few files in Base Profiler.
Differential Revision: https://phabricator.services.mozilla.com/D39624
--HG--
extra : moz-landing-system : lando
`ClearBefore()` with a past-the-end `BlockIndex` was calling `Clear()`, which
tried to take the lock again! Also we didn't return after that.
Fixed, and added corresponding test.
Also: Removed ambiguous "delete" word, now using more precise "destroy" or
"entry destructor".
Differential Revision: https://phabricator.services.mozilla.com/D38846
--HG--
extra : moz-landing-system : lando
This is a similar concept as `nullptr` is to a pointer.
`BlocksRingBuffer` now skips the first byte in the buffer, so that no entries
start at 0 (the internal default `BlockIndex` value).
All `BlocksRingBuffer` public APIs handle this default value, and do nothing
and/or return Nothing (as if it pointed at an already-deleted entry).
Added tests for this, and for all BlockIndex operations.
Differential Revision: https://phabricator.services.mozilla.com/D38667
--HG--
extra : moz-landing-system : lando
Without declaring them, ModuloBuffer had its copy&move constructor&assignments
defaulted. This means it could have been copied, and then both objects would now
own the same resource and attempt to free it on destruction!
So now:
- Copy construction&assignment are now explicitly disallowed.
- Move assignment is disallowed, to keep some members `const`.
- Move construction is allowed (so a function can return a ModuloBuffer), and
ensures that the moved-from object won't free the resource anymore.
Bonus: `mBuffer` is now `const`, to ensure that it cannot point at something
else, but note the pointed-at bytes are *not* const.
So ModuloBuffer is like an unchanging resource, but it allows to be moved-from
as an xvalue that should not be used after the move.
Differential Revision: https://phabricator.services.mozilla.com/D38665
--HG--
extra : moz-landing-system : lando
By default `ModuloBuffer` allocates its own buffer on the heap.
Now `ModuloBuffer` adds two alternatives:
- Take ownership of a pre-allocated `UniquePtr<uint8_t>` buffer.
- Work over an unowned `uint8_t*` array. The caller is responsible for
ownership, and ensuring that the array lives at least as long as the
`ModuloBuffer`/`BlocksRingBuffer`.
`BlocksRingBuffer` can pass along these new options to its underlying
`ModuloBuffer`.
The main use will be for small on-stack `BlocksRingBuffer` that can store a
stack trace, or to more easily collect data (without allocating anything on the
heap) that can then go into the upcoming `ProfileBuffer`'s `BlocksRingBuffer`.
Differential Revision: https://phabricator.services.mozilla.com/D38285
--HG--
extra : moz-landing-system : lando
This adds to the byte-oriented ModuloBuffer from bug 1563425:
- Thread-safety: All APIs may be called at any time from any thread.
- Structure: The buffer will be divided in "blocks" of different size, with some
block meta-data and space for the user "entry".
- Capable of handling user resources: The user may provide a "deleter" that will
be informed about soon-to-be-destroyed entries; so if some entries reference
outside resources, these references may be properly released.
Note: This first implementation still only allows the user to manipulate bytes
and trivially-copyable objects (same as with the ModuloBuffer iterators). A
follow-up bug will introduce better serialization capabilities, with the aim to
eventually store everything that current Profiler Markers and their payloads
contain.
Differential Revision: https://phabricator.services.mozilla.com/D37702
--HG--
extra : moz-landing-system : lando
As we are increasingly moving toward enabling new types of DLL blocking across
our various process types, we need to be able to generate various headers in
various distinct formats.
This script enables us to use a unified DLL blocklist input that generates
these distinct headers. From WindowsDllBlocklistDefs.in, we generate:
WindowsDllBlocklistA11yDefs.h - definitions for a11y
WindowsDllBlocklistLauncherDefs.h - definitions for the launcher process
WindowsDllBlocklistLegacyDefs.h - definitions for the legacy mozglue blocklist
WindowsDllBlocklistTestDefs.h - test-only definitions
These headers are then exported to mozilla.
Note that not all headers use the same format, as not all consumers of these
headers have identical workings. There will be additional header types added
in the future which diverge even more from the standard blocklist format. While
this work may seem a bit pointless at the moment, it will become more necessary
in the future. In particular, this work is a prerequisite for bug 1238735.
Differential Revision: https://phabricator.services.mozilla.com/D36993
--HG--
extra : moz-landing-system : lando
Gerald Squelart
Cameron Kaiser
Nazım Can Altınova
Makoto Kato
Aaron Klotz
Emilio Cobos Álvarez
Sylvestre Ledru
Brindusan Cristian
Toshihito Kikuchi
Greg Tatum