On Windows, the profiler changes the timer resolution when the sampling interval is less than 10ms.
However this change affects all of Firefox, which can change other behaviors, sometimes causing confusion when refresh-rate issues magically disappear when the profiler is running.
So now by default the profiler will not change the timer resolution. This should rarely affect the profiler's effective sampling rate, unless all threads in a process are in a waiting state, in which case there is nothing new to sample anyway! The front-end is investigating ways to make sampling gaps more obvious, see https://github.com/firefox-devtools/profiler/issues/2962
If some power users really need the added sampling precision in some profiling sessions, the timer resolution change may be requested before running Firefox, by setting the environment variable "PROFILER_ADJUST_TIMER_RESOLUTION" to any value.
Differential Revision: https://phabricator.services.mozilla.com/D115451
`SharedLibraryInfo::GetInfoForSelf()` can use `PEHeaders::GetBounds` instead of
`GetModuleInformation` to get the start/end address of a module's mapped region
in the local process. It's roughly 100x faster because `GetModuleInformation`
invokes two system calls `NtQueryInformationProcess` and `NtReadVirtualMemory`
while `nt::PEHeaders` does not.
Depends on D115254
Differential Revision: https://phabricator.services.mozilla.com/D115255
This patch replaces two versions of `GetVersion` in Gecko profiler and
baseprofiler with `PEHeaders::GetVersionInfo`.
Depends on D115253
Differential Revision: https://phabricator.services.mozilla.com/D115254
This patch is the actual fix for Bug 1702086. The problem of Bug 1702086 is that
`LoadLibraryExW` loaded the module onto an address different from the original
mapped addresss because it was unloaded after we started enumeration. Calling
`GetPdbInfo` with the original address `module.lpBaseOfDll` caused a crash.
The proposed fix consists of three parts.
The first part is to get PDB information from `handleLock`, which is always valid
even if the original address was unloaded. With this, we don't need a check
of `VirtualQuery`.
The second part is to add `LOAD_LIBRARY_AS_IMAGE_RESOURCE` along with
`LOAD_LIBRARY_AS_DATAFILE` to the call to `LoadLibraryEx`. This is needed
to read information from the sections outside the PE headers because
RVA (= relative virtual address) is an address after relocation.
Without `LOAD_LIBRARY_AS_IMAGE_RESOURCE`, a module is mapped without relocation,
so `GetPdbInfo()` accesses wrong memory resulting in a crash.
The third part is to introduce `PEHeaders::GetPdbInfo`, replacing two versions
of `GetPdbInfo` in Gecko profiler and baseprofiler.
Depends on D115252
Differential Revision: https://phabricator.services.mozilla.com/D115253
This patch introduces `EnumerateProcessModules` to enumerate all loaded modules
in the local process so that Gecko profiler and baseprofiler can use it.
Differential Revision: https://phabricator.services.mozilla.com/D115252
SprintfLiteral doesn't support "%#llx" anymore.
So this patch switches to an explicit "0x%llx" to output addresses as the symbolicator expects.
Differential Revision: https://phabricator.services.mozilla.com/D114914
Before `SharedLibraryInfo::GetInfoForSelf` calls `GetPdbInfo` to parse a module's
PE header, it calls `LoadLibraryEx` to prevent the module from being unloaded
during `GetPdbInfo`. If the module was already unloaded before `LoadLibraryEx`,
however, `LoadLibraryEx` maps the module onto an address different from the original
mapped address, so that `module.lpBaseOfDll` becomes invalid.
This patch is to call `LoadLibraryEx` before `GetModuleInformation` to make sure
`LoadLibraryEx` increments the module's refcount and does not map the module onto
a new address. With this, `module.lpBaseOfDll` is always valid, thus we don't have
to call `VirtualQuery`.
Differential Revision: https://phabricator.services.mozilla.com/D110421
Lastly, we are changing the parts that requires a version bump and bumping the
profiler version in the end. This will require a PR in the front-end for a
version upgrader and changes related to the renaming.
Differential Revision: https://phabricator.services.mozilla.com/D109282
This patch is only about renaming the internals of the profiler codebase and
it doesn't touch any parts that requires a backwards compatibility or version
bump.
Differential Revision: https://phabricator.services.mozilla.com/D109280
It requires including <windows.h>, preventing the inclusion of StackWalk.h
from some places (and upcoming changes will make StackWalk.h included in
more places).
Differential Revision: https://phabricator.services.mozilla.com/D108910
In the case of FramePointerStackwalk, the caller gives a pointer to the
top-most frame to walk from. There isn't really a reason to give a
number of frames to skip, as the right frame pointer could be given in
the first place if that was really necessary. And in practice, it's
hasn't been used so far.
In the case of MozStackWalkThread, the caller presumably doesn't know
what the thread the stack is being walked for is doing, and it would be
a guesswork to pass a valid number of frames to skip. In practice, it's
also not used.
The aSkipFrames is already a footgun on MozStackWalk (and we're going to
change that in bug 1515229), we don't need to keep a footgun on these
other stack walking methods.
Differential Revision: https://phabricator.services.mozilla.com/D108563
No code changes.
Build issues were found by renaming `MOZ_GECKO_PROFILER` to something else in toolkit/moz.configure, in both unified and non-unified builds, on all supported platforms.
Also updated some profiler-related comments.
Differential Revision: https://phabricator.services.mozilla.com/D105375
New headers BaseProfilerLabels.h and ProfilerLabels.h now contain all label-related APIs.
These files were hg-copied from the main headers, to preserve history, and then non-label content was removed from the main headers.
The "RAII" macros were moved to these *ProfilerLabels.h headers, because that's the most-common header in which they're needed. Meta-bug 1681416 will probably move these again as needed.
Differential Revision: https://phabricator.services.mozilla.com/D104587
{Base,}ProfilerMarkers.h can now rely on {Base,}ProfilerState.h (and just one function forward declaration), so they don't need to include {Base,Gecko}Profiler.h anymore.
Thanks to that, {Base,Gecko}Profiler.h can now include {Base,}ProfilerMarkers.h at the top.
Also, BaseProfilingStack.h should not include BaseProfiler.h (include loop, can cause failures) not algorithm (not used).
***
roll up into 2nd patch
Differential Revision: https://phabricator.services.mozilla.com/D104969
New headers BaseProfilerState.h and ProfilerState.h now contain most state-reading APIs.
These files were hg-copied from the main headers, to preserve history, and then chosen declarations were kept only in the relevant header.
This is needed in a following patch, where new headers *ProfilerLabels.h use `profiler_is_active()`.
Differential Revision: https://phabricator.services.mozilla.com/D104968
ProfileBufferCollector::SamplePositionInBuffer() and BufferRangeStart() need to provide indices in the same main buffer, because they will be used to discard old data (at some previous `SamplePositionInBuffer`) once the `BufferRangeStart` indicates that it is not referenced by the profiler anymore.
Because the periodic sampler uses a local buffer (to avoid allocations and locks), we need to record the special location from the main profiler buffer in ProfileBufferCollector.
Differential Revision: https://phabricator.services.mozilla.com/D104497
The `RunningTimes` class stores CPU measurements. It may seem overkill for only one value, but in the future more measurements will be added.
During sampling, CPU measurements are collected by platform-specific code. This patch doesn't produce anything yet, see later patches.
These are stored with the samples.
Note that for duplicated samples (when a thread is known to be "asleep"), we still need to collect new measurements, because there could potentially be some activity happening, e.g. in system calls.
Finally the measurements are output as extra "samples" values.
Units for these values may platform-specific, so they are stored in the top-level JSON "meta" object.
We don't collect running times in the Base Profiler (yet), but we still need to add the appropriate field names in the samples' "schema", as expected by profiler.firefox.com.
Differential Revision: https://phabricator.services.mozilla.com/D99413
This patch adds "CPU Utilization" ("cpu" for short) as a new feature that will control the upcoming still-experimental CPU measurements.
Differential Revision: https://phabricator.services.mozilla.com/D99054
Instead of only capturing one feature (NoStackSampling), the sampler thread now stores all features so that any feature can be quickly looked at during sampling.
Currently this is still limited to NoStackSampling, a later patch will start using another feature.
Differential Revision: https://phabricator.services.mozilla.com/D99053
Some sites do have stacks that require more than 64KB to store in the profiler buffer.
Note that this only affects one semi-permanent buffer per process during profiling, and short-lived buffers when capturing stacks in markers.
Differential Revision: https://phabricator.services.mozilla.com/D99981
The "expected maximum stack size" (currently 64KiB) value was present in multiple places.
Now it's accessible from everywhere as ProfileBufferChunkManager::scExpectedMaximumStackSize, so it's easier to modify as needed.
Differential Revision: https://phabricator.services.mozilla.com/D100222
Instead of discarding released chunks, we keep them as "next" chunks, and we make sure there's a valid "current" chunk when possible.
Recycling released chunks means that when using the `ProfileBufferChunkManagerSingle`, the one chunk, in whichever state it may be, will be kept alive and reused.
Differential Revision: https://phabricator.services.mozilla.com/D99979
When a "Put" operation fails (most probably because no chunk was available to store the data), we remember the number of bytes that couldn't be stored.
This can be useful to give an indication of how much more memory would have been needed for successful puts.
Differential Revision: https://phabricator.services.mozilla.com/D99977
In DEBUG builds, instead of only testing for the expected chunk state(s), there are now separate assertions for each *un*expected state, which makes it much easier to track the source of failures.
Differential Revision: https://phabricator.services.mozilla.com/D99974
TimeStamps in markers must now be streamed through `SpliceableJSONWriter::TimeProperty(name, timestamp)`.
This is consistent with all other JSON-writing functions being in `SpliceableJSONWriter` (and base class `JSONWriter`).
Depends on D97556
Differential Revision: https://phabricator.services.mozilla.com/D97557
Gerald Squelart
Toshihito Kikuchi
Andi-Bogdan Postelnicu
Florian Quèze
Alexandru Michis
Nazım Can Altınova
Jeff Muizelaar
Mike Hommey
Markus Stange
Greg Tatum
Masatoshi Kimura
Steve Fink
Simon Giesecke
Bogdan Tara