Profiler Code Overview¶
This is an overview of the code that implements the Profiler inside Firefox with dome details around tricky subjects, or pointers to more detailed documentation and/or source code.
It assumes familiarity with Firefox development, including Mercurial (hg), mach, moz.build files, Try, Phabricator, etc.
It also assumes knowledge of the user-visible part of the Firefox Profiler, that is: How to use the Firefox Profiler, and what profiles contain that is shown when capturing a profile. See the main website https://profiler.firefox.com, and its documentation.
For just an “overview”, it may look like a huge amount of information, but the Profiler code is indeed quite expansive, so it takes a lot of words to explain even just a high-level view of it! For on-the-spot needs, it should be possible to search for some terms here and follow the clues. But for long-term maintainers, it would be worth skimming this whole document to get a grasp of the domain, and return to get some more detailed information before diving into the code.
WIP note: This document should be correct at the time it is written, but the profiler code constantly evolves to respond to bugs or to provide new exciting features, so this document could become obsolete in parts! It should still be useful as an overview, but its correctness should be verified by looking at the actual code. If you notice any significant discrepancy or broken links, please help by filing a bug.
This is the common usage for some frequently-used terms, as understood by the Dev Tools team. But incorrect usage can sometimes happen, context is key!
profiler (a): Generic name for software that enables the profiling of code. (“Profiling” on Wikipedia)
Profiler (the): All parts of the profiler code inside Firefox.
Base Profiler (the): Parts of the Profiler that live in mozglue/baseprofiler, and can be used from anywhere, but has limited functionality.
Gecko Profiler (the): Parts of the Profiler that live in tools/profiler, and can only be used from other code in the XUL library.
Profilers (the): Both the Base Profiler and the Gecko Profiler.
profiling session: This is the time during which the profiler is running and collecting data.
profile (a): The output from a profiling session, either as a file, or a shared viewable profile on https://profiler.firefox.com
Profiler back-end (the): Other name for the Profiler code inside Firefox, to distinguish it from…
Profiler front-end (the): The website https://profiler.firefox.com that displays profiles captured by the back-end.
Firefox Profiler (the): The whole suite comprised of the back-end and front-end.
When working on the profiler, here are some guiding principles to keep in mind:
Low profiling overhead in cpu and memory. For the Profiler to provide the best value, it should stay out of the way and consume as few resources (in time and memory) as possible, so as not to skew the actual Firefox code too much.
Common data structures and code should be in the Base Profiler when possible.
WIP note: Deduplication is slowly happening, see meta bug 1557566. This document focuses on the Profiler back-end, and mainly the Gecko Profiler (because this is where most of the code lives, the Base Profiler is mostly a subset, originally just a cut-down version of the Gecko Profiler); so unless specified, descriptions below are about the Gecko Profiler, but know that there may be some equivalent code in the Base Profiler as well.
Use appropriate programming-language features where possible to reduce coding errors in both our code, and our users’ usage of it. In C++, this can be done by using a specific class/struct types for a given usage, to avoid misuse (e.g., an generic integer representing a process could be incorrectly given to a function expecting a thread; we have specific types for these instead, more below.)
Follow the Coding Style.
Whenever possible, write tests (if not present already) for code you add or modify – but this may be too difficult in some case, use good judgement and at least test manually instead.
Here is a high-level view of the Base or Gecko Profiler lifecycle, as part of a Firefox run. The following sections will go into much more details.
Profiler initialization, preparing some common data.
Threads de/register themselves as they start and stop.
During each User/test-controlled profiling session:
Profiler start, preparing data structures that will store the profiling data.
Periodic sampling from a separate thread, happening at a user-selected frequency (usually once every 1-2 ms), and recording snapshots of what Firefox is doing:
CPU sampling, measuring how much time each thread has spent actually running on the CPU.
Stack sampling, capturing a stack of functions calls from whichever leaf function the program is in at this point in time, up to the top-most caller (i.e., at least the
main()function, or its callers if any). Note that unlike most external profilers, the Firefox Profiler back-end is capable or getting more useful information than just native functions calls (compiled from C++ or Rust):
Labels added by Firefox developers along the stack, usually to identify regions of code that perform “interesting” operations (like layout, file I/Os, etc.).
Java function calls.
At any time, Markers may record more specific details of what is happening, e.g.: User operations, page rendering steps, garbage collection, etc.
Optional profiler pause, which stops most recording, usually near the end of a session so that no data gets recorded past this point.
Profile JSON output, generated from all the recorded profiling data.
Profiler stop, tearing down profiling session objects.
Note that the Base Profiler can start earlier, and then the data collected so far, as well as the responsibility for periodic sampling, is handed over to the Gecko Profiler:
Base Profiler init
Base Profiler start
(Firefox loads the libxul library and initializes XPCOM)
Gecko Profiler init
Gecko Profiler start
Handover from Base to Gecko
Base Profiler stop
(Bulk of the profiling session)
Gecko Profiler stop
Gecko Profiler shutdown
(Firefox ends XPCOM)
Base Profiler shutdown
Base Profiler functions that add data (mostly markers and labels) may be called from anywhere, and will be recorded by either Profiler. The corresponding functions in Gecko Profiler can only be called from other libxul code, and can only be recorded by the Gecko Profiler.
Whenever possible, Gecko Profiler functions should be preferred if accessible, as they may provide extended functionality (e.g., better stacks with JS in markers). Otherwise fallback on Base Profiler functions.
Non-Profiler supporting code
mozglue/baseprofiler - Base Profiler code, usable from anywhere in Firefox. Because it lives in mozglue, it’s loaded right at the beginning, so it’s possible to start the profiler very early, even before Firefox loads its big&heavy “xul” library.
tools/profiler - Gecko Profiler code, only usable from the xul library. That library is loaded a short time after Firefox starts, so the Gecko Profiler is not able to profile the early phase of the application, Base Profiler handles that, and can pass its collected data to the Gecko Profiler when the latter starts.
toolkit/components/extensions/schemas/geckoProfiler.json - File that needs to be updated when Profiler features change.
Out of scope for this document, but its code and bug repository can be found at: https://github.com/firefox-devtools/profiler . Sometimes work needs to be done on both the back-end of the front-end, especially when modifying the back-end’s JSON output format.
The most central public header is GeckoProfiler.h, from which almost everything else can be found, it can be a good starting point for exploration. It includes other headers, which together contain important top-level macros and functions.
WIP note: GeckoProfiler.h used to be the header that contained everything! To better separate areas of functionality, and to hopefully reduce compilation times, parts of it have been split into smaller headers, and this work will continue, see bug 1681416.
MOZ_GECKO_PROFILER and Macros¶
Mozilla officially supports the Profiler on tier-1 platforms: Windows, macos, Linux and Android. There is also some code running on tier 2-3 platforms (e.g., for FreeBSD), but the team at Mozilla is not obligated to maintain it; we do try to keep it running, and some external contributors are keeping an eye on it and provide patches when things do break.
To reduce the burden on unsupported platforms, a lot of the Profilers code is
only compiled when
MOZ_GECKO_PROFILER is #defined. This means that some
public functions may not always be declared or implemented, and should be
surrounded by guards like
Some commonly-used functions offer an empty definition in the
MOZ_GECKO_PROFILER case, so these functions may be called from anywhere
Other functions have associated macros that can always be used, and resolve to
nothing on unsupported platforms. E.g.,
profiler_register_thread where supported,
otherwise does nothing.
WIP note: There is an effort to eventually get rid of
its associated macros, see
RAII “Auto” macros and classes¶
A number of functions are intended to be called in pairs, usually to start and then end some operation. To ease their use, and ensure that both functions are always called together, they usually have an associated class and/or macro that may be called only once. This pattern of using an object’s destructor to ensure that some action always eventually happens, is called RAII in C++, with the common prefix “auto”.
This section describes some platform abstractions that are used throughout the Profilers. (Other platform abstractions will be described where they are used.)
Process and Thread IDs¶
The Profiler back-end often uses process and thread IDs (aka “pid” and “tid”),
which are commonly just a number.
For better code correctness, and to hide specific platform details, they are
encapsulated in opaque types
These types should be used wherever possible.
When interfacing with other code, they may be converted using the member
The main thread ID is available through profiler_main_thread_id (assuming profiler_init_main_thread_id was called when the application started – especially important in stand-alone test programs.) And profiler_is_main_thread is a quick way to find out if the current thread is the main thread.
The locking primitives in PlatformMutex.h are not supposed to be used as-is, but through a user-accessible implementation. For the Profilers, this is in BaseProfilerDetail.h.
In addition to the usual
objects have a name (which may be helpful when debugging),
they record the thread on which they are locked (making it possible to know if
the mutex is locked on the current thread), and in
DEBUG builds there are
assertions verifying that the mutex is not incorrectly used recursively, to
verify the correct ordering of different Profiler mutexes, and that it is
unlocked before destruction.
Mutexes should preferably be locked within C++ block scopes, or as class members, by using BaseProfilerAutoLock.
Some classes give the option to use a mutex or not (so that single-threaded code can more efficiently bypass locking operations), for these we have BaseProfilerMaybeMutex and BaseProfilerMaybeAutoLock.
There is also a special type of shared lock (aka RWLock, see
RWLock on wikipedia),
which may be locked in multiple threads (through
LockShared or preferably
or locked exclusively, preventing any other locking (through
LockExclusive or preferably
Main Profiler Classes¶
Diagram showing the most important Profiler classes, see details in the following sections:
(As noted, the “RegisteredThread” classes are now obsolete in the Gecko Profiler, see the “Thread Registration” section below for an updated diagram and description.)
Making sure the main thread ID is recorded.
MOZ_PROFILER_HELP=1 ./mach runto display the command-line help.
CorePSinstance – more details below.
Registering the main thread.
Initializing some platform-specific code.
Handling other environment variables that are used to immediately start the profiler, with optional settings provided in other env-vars.
The CorePS class has a single instance that should live for the duration of the Firefox application, and contains important information that could be needed even when the Profiler is not running.
A static pointer to its single instance.
The process start time.
A list of registered PageInformations, used to keep track of the tabs that this process handles.
A list of BaseProfilerCounts, used to record things like the process memory usage.
The process name, and optionally the “eTLD+1” (roughly sub-domain) that this process handles.
In the Base Profiler only, a list of RegisteredThreads. WIP note: This storage has been reworked in the Gecko Profiler (more below), and in practice the Base Profiler only registers the main thread. This should eventually disappear as part of the de-duplication work (bug 1557566).
Threads need to register themselves in order to get fully profiled. This section describes the main data structures that record the list of registered threads and their data.
WIP note: There is some work happening to add limited profiling of unregistered threads, with the hope that more and more functionality could be added to eventually use the same registration data structures.
Diagram showing the relevant classes, see details in the following sub-sections:
static ProfilerThreadRegistry object
contains a list of
OffThreadRef points to a
ProfilerThreadRegistration, and restricts
access to a safe subset of the thread data, and forces a mutex lock if necessary
(more information under ProfilerThreadRegistrationData below).
A ProfilerThreadRegistration object contains a lot of information relevant to its thread, to help with profiling it.
This data is accessible from the thread itself through an
object, which points to the
ThreadRegistration, and restricts access to a
safe subset of thread data, and forces a mutex lock if necessary (more
information under ProfilerThreadRegistrationData below).
ThreadRegistrationData and accessors¶
The ProfilerThreadRegistrationData.h header contains a hierarchy of classes that encapsulate all the thread-related data.
ThreadRegistrationData contains all the actual data members, including:
Some long-lived ThreadRegistrationInfo, containing the thread name, its registration time, the thread ID, and whether it’s the main thread.
ProfilingStackthat gathers developer-provided pseudo-frames, and JS frames.
PlatformData(usually required to actually record profiling measurements for that thread).
A pointer to the top of the stack.
A shared pointer to the thread’s
A pointer to the
An optional pre-allocated
JsFramebuffer used during stack-sampling.
Some JS flags.
Sleep-related data (to avoid costly sampling while the thread is known to not be doing anything).
ThreadProfilingFeatures, to know what kind of data to record.
When profiling, a pointer to a
ProfiledThreadData, which contains some more data needed during and just after profiling.
As described in their respective code comments, each data member is supposed to
be accessed in certain ways, e.g., the
JSContext should only be “written
from thread, read from thread and suspended thread”. To enforce these rules,
data members can only be accessed through certain classes, which themselves can
only be instantiated in the correct conditions.
The accessor classes are, from base to most-derived:
ThreadRegistrationData, not an accessor itself, but it’s the base class with all the
ThreadRegistrationUnlockedConstReader, giving unlocked
PlatformData, and stack top.
ThreadRegistrationUnlockedConstReaderAndAtomicRW, giving unlocked access to the atomic data members:
ProfilingStack, sleep-related data,
ThreadRegistrationUnlockedRWForLockedProfiler, giving access that’s protected by the Profiler’s main lock, but doesn’t require a
ThreadRegistrationlock, to the
ThreadRegistrationUnlockedReaderAndAtomicRWOnThread, giving unlocked mutable access, but only on the thread itself, to the
ThreadRegistrationLockedRWFromAnyThread, giving locked access from any thread to mutex-protected data:
nsIThread, and the JS flags.
ThreadRegistrationLockedRWOnThread, giving locked access, but only from the thread itself, to the
JSContextand a JS flag-related operation.
ThreadRegistration::EmbeddedData, containing all of the above, and stored as a data member in each
To recapitulate, if some code needs some data on the thread, it can use
ThreadRegistration functions to request access (with the required rights,
like a mutex lock).
To access data about another thread, use similar functions from
You may find some examples in the implementations of the functions in
ProfilerThreadState.h (see the following section).
The ProfilerThreadState.h header provides a few helpful functions related to threads, including:
profiler_thread_is_being_profiled(for the current thread or another thread, and for a given set of features)
There are multiple ways to start the profiler, through command line env-vars, and programmatically in C++ and JS.
The main public C++ function is
It takes all the features specifications, and returns a promise that gets
resolved when the Profiler has fully started in all processes (multi-process
profiling is described later in this document, for now the focus will be on each
process running its instance of the Profiler). It first calls
if needed, and also
profiler_stop if the profiler was already running.
The main implementation, which can be called from multiple sources, is locked_profiler_start. It performs a number of operations to start the profiling session, including:
Record the session start time.
Pre-allocate some work buffer to capture stacks for markers on the main thread.
In the Gecko Profiler only: If the Base Profiler was running, take ownership of the data collected so far, and stop the Base Profiler (we don’t want both trying to collect the same data at the same time!)
Create the ActivePS, which keeps track of most of the profiling session information, more about it below.
For each registered thread found in the
ThreadRegistry, check if it’s one of the threads to profile, and if yes set the appropriate data into the corresponding
ThreadRegistrationData(including informing the JS engine to start recording profiling data).
On Android, start the Java sampler.
If native allocations are to be profiled, setup the appropriate hooks.
Start the audio callback tracing if requested.
Set the public shared “active” state, used by many functions to quickly assess whether to actually record profiling data.
The ActivePS class has a single instance at a time, that should live for the length of the profiling session.
The session start time.
A way to track “generations” (in case an old ActivePS still lives when the next one starts, so that in-flight data goes to the correct place.)
Requested features: Buffer capacity, periodic sampling interval, feature set, list of threads to profile, optional: specific tab to profile.
The profile data storage buffer and its chunk manager (see “Storage” section below for details.)
More data about live and dead profiled threads.
Optional counters for per-process CPU usage, and power usage.
A pointer to the
SamplerThreadobject (see “Periodic Sampling” section below for details.)
During a session, the profiling data is serialized into a buffer, which is made of “chunks”, each of which contains “blocks”, which have a size and the “entry” data.
During a profiling session, there is one main profile buffer, which may be started by the Base Profiler, and then handed over to the Gecko Profiler when the latter starts.
The buffer is divided in chunks of equal size, which are allocated before they are needed. When the data reaches a user-set limit, the oldest chunk is recycled. This means that for long-enough profiling sessions, only the most recent data (that could fit under the limit) is kept.
Each chunk stores a sequence of blocks of variable length. The chunk itself only knows where the first full block starts, and where the last block ends, which is where the next block will be reserved.
To add an entry to the buffer, a block is reserved, the size is written first (so that readers can find the start of the next block), and then the entry bytes are written.
The following sessions give more technical details.
They are an efficient way to serialize numbers that are usually small, e.g., numbers up to 127 only take one byte, two bytes up to 16,383, etc.
A ProfileBufferBlockIndex object encapsulates a block index that is known to be the valid start of a block. It is created when a block is reserved, or when trusted code computes the start of a block in a chunk.
The more generic ProfileBufferIndex type is used when working inside blocks.
A ProfileBufferChunk is a variable-sized object. It contains:
A public copyable header, itself containing:
The local offset to the first full block (a chunk may start with the end of a block that was started at the end of the previous chunk). That offset in the very first chunk is the natural start to read all the data in the buffer.
The local offset past the last reserved block. This is where the next block should be reserved, unless it points past the end of this chunk size.
The timestamp when the chunk was first used.
The timestamp when the chunk became full.
The number of bytes that may be stored in this chunk.
The number of reserved blocks.
The global index where this chunk starts.
The process ID writing into this chunk.
An owning unique pointer to the next chunk. It may be null for the last chunk in a chain.
DEBUGbuilds, a state variable, which is used to ensure that the chunk goes through a known sequence of states (e.g., Created, then InUse, then Done, etc.) See the sequence diagram where the member variable is defined.
The actual buffer data.
Because a ProfileBufferChunk is variable-size, it must be created through its
Create function, which takes care of allocating the correct amount
of bytes, at the correct alignment.
The ProfileBufferChunkManager abstract class defines the interface of classes that manage chunks.
Concrete implementations are responsible for: * Creating chunks for their user, with a mechanism to pre-allocate chunks before they are actually needed. * Taking back and owning chunks when they are “released” (usually when full). * Automatically destroying or recycling the oldest released chunks. * Giving temporary access to extant released chunks.
A ProfileBufferChunkManagerSingle object manages a single chunk.
That chunk is always the same, it is never destroyed. The user may use it and optionally release it. The manager can then be reset, and that one chunk will be available again for use.
A request for a second chunk would always fail.
This manager is short-lived and not thread-safe. It is useful when there is some limited data that needs to be captured without blocking the global profiling buffer, usually one stack sample. This data may then be extracted and quickly added to the global buffer.
A ProfileBufferChunkManagerWithLocalLimit object
ProfileBufferChunkManager interface fully, managing a number
of chunks, and making sure their total combined size stays under a given limit.
This is the main chunk manager user during a profiling session.
Note: It also implements the
this is explained in the later section “Multi-Process Profiling”.
It is thread-safe, and one instance is shared by both Profilers.
A ProfileChunkedBuffer object
ProfilerBufferChunkManager to store data, and handles the different
C++ types of data that the Profilers want to read/write as entries in buffer
Its main function is
It takes an invocable object (like a lambda) that should return the size of the entry to store, this is to potentially avoid costly operations just to compute a size, when the profiler may not be running.
It attempts to reserve the space in its chunks, requesting a new chunk if necessary.
It then calls a provided invocable object with a ProfileBufferEntryWriter, which offers a range of functions to help serialize C++ objects. The de/serialization functions are found in specializations of ProfileBufferEntryWriter::Serializer and ProfileBufferEntryReader::Deserializer.
More “put” functions use
ReserveAndPut to more easily serialize blocks of
memory, or C++ objects.
ProfileChunkedBuffer is optionally thread-safe, using a
WIP note: Using a mutex makes this storage too noisy for profiling some real-time (like audio processing). Bug 1697953 will look at switching to using atomic variables instead. An alternative would be to use a totally separate non-thread-safe buffers for each real-time thread that requires it (see bug 1754889).
A ProfileBuffer object
ProfileChunkedBuffer to store data, and handles the different kinds
of entries that the Profilers want to read/write.
Each entry starts with a tag identifying a kind. These kinds can be found in ProfileBufferEntryKinds.h.
There are “legacy” kinds, which are small fixed-length entries, such as: Categories, labels, frame information, counters, etc. These can be stored in ProfileBufferEntry objects
And there are “modern” kinds, which have variable sizes, such as: Markers, CPU
running times, full stacks, etc. These are more directly handled by code that
can access the underlying
The other major responsibility of a
ProfileChunkedBuffer is to read back all
this data, sometimes during profiling (e.g., to duplicate a stack), but mainly
at the end of a session when generating the output JSON profile.
Probably the most important job of the Profiler is to sample stacks of a number of running threads, to help developers know which functions get used a lot when performing some operation on Firefox.
This is accomplished from a special thread, which regularly springs into action and captures all this data.
The SamplerThread object
manages the information needed during sampling. It is created when the profiler
starts, and is stored inside the
ActivePS, see above for details.
Samplerobject that contains platform-specific details, which are implemented in separate files like platform-win32.cpp, etc.
The same generation index as its owning
The requested interval between samples.
A handle to the thread where the sampling happens, its main function is Run() function.
A list of callbacks to invoke after the next sampling. These may be used by tests to wait for sampling to actually happen.
The unregistered-thread-spy data, and an optional handle on another thread that takes care of “spying” on unregistered thread (on platforms where that operation is too expensive to run directly on the sampling thread).
Run() function takes care of performing the periodic sampling work:
(more details in the following sections)
Retrieve the sampling parameters.
ProfileBufferon the stack, to capture samples from other threads.
Loop until a
Lock the main profiler mutex, and do:
Check if sampling should stop, and break from the loop.
Clean-up exit profiles (these are profiles sent from dying sub-processes, and are kept for as long as they overlap with this process’ own buffer range).
Record the CPU utilization of the whole process.
Record the power consumption.
Sample each registered counter, including the memory counter.
For each registered thread to be profiled:
Record the CPU utilization.
If the thread is marked as “still sleeping”, record a “same as before” sample, otherwise suspend the thread and take a full stack sample.
On some threads, record the event delay to compute the (un)responsiveness. WIP note: This implementation may change.
Record profiling overhead durations.
Unlock the main profiler mutex.
Invoke registered post-sampling callbacks.
Spy on unregistered threads.
Based on the requested sampling interval, and how much time this loop took, compute when the next sampling loop should start, and make the thread sleep for the appropriate amount of time. The goal is to be as regular as possible, but if some/all loops take too much time, don’t try too hard to catch up, because the system is probably under stress already.
Go back to the top of the loop.
If we’re here, we hit a loop
Invoke registered post-sampling callbacks, to let them know that sampling stopped.
CPU Utilization is stored as a number of milliseconds that a thread or process has spent running on the CPU since the previous sampling.
Energy probes added in 2022.
Stacks are the sequence of calls going from the entry point in the program
main() and some OS-specific functions above), down to the
function where code is currently being executed.
Compiled code, from C++ and Rust source.
Pseudo-frames with arbitrary text, added from any language, mostly C++.
JS, Wasm Frames¶
Recorded by the JavaSampler.
The above types of frames are all captured in different ways, and when finally taking an actual stack sample (apart from Java), they get merged into one stack.
All frames have an associated address in the call stack, and can therefore be merged mostly by ordering them by this stack address. See MergeStacks for the implementation details.
Counters are a special kind of probe, which can be continuously updated during
profiling, and the
SamplerThread will sample their value at every loop.
This is the main counter. During a profiling session, hooks into the memory manager keep track of each de/allocation, so at each sampling we know how many operations were performed, and what is the current memory usage compared to the previous sampling.
SamplerThread records timestamps between parts of its sampling loop, and
records this as the sampling overhead. This may be useful to determine if the
profiler itself may have used too much of the computer resources, which could
skew the profile and give wrong impressions.
Unregistered Thread Profiling¶
At some intervals (not necessarily every sampling loop, depending on the OS), the profiler may attempt to find unregistered threads, and record some information about them.
WIP note: This feature is experimental, and data is captured in markers on the main thread. More work is needed to put this data in tracks like regular registered threads, and capture more data like stack samples and markers.
Markers are events with a precise timestamp or time range, they have a name, a category, options (out of a few choices), and optional marker-type-specific payload data.
Before describing the implementation, it is useful to be familiar with how markers are natively added from C++, because this drives how the implementation takes all this information and eventually outputs it in the final JSON profile.
Adding Markers from C++¶
The main function that records markers is
It’s a variadic templated function that takes the different the expected
arguments, first checks if the marker should actually be recorded (the profiler
should be running, and the target thread should be profiled), and then calls
into the deeper implementation function
AddMarkerToBuffer with a reference
to the main profiler buffer.
AddMarkerToBuffer takes the marker type as an object, removes it from the function parameter list, and calls the next function with the marker type as an explicit template parameter, and also a pointer to the function that can capture the stack (because it is different between Base and Gecko Profilers, in particular the latter one knows about JS).
From here, we enter the land of BaseProfilerMarkersDetail.h, which employs some heavy template techniques, in order to most efficiently serialize the given marker payload arguments, in order to make them deserializable when outputting the final JSON. In previous implementations, for each new marker type, a new C++ class derived from a payload abstract class was required, that had to implement all the constructors and virtual functions to:
Create the payload object.
Serialize the payload into the profile buffer.
Deserialize from the profile buffer to a new payload object.
Convert the payload into the final output JSON.
Now, the templated functions automatically take care of serializing all given function call arguments directly (instead of storing them somewhere first), and preparing a deserialization function that will recreate them on the stack and directly call the user-provided JSONification function with these arguments.
Continuing from the public
sets some defaults if not specified by the caller: Target the current thread,
use the current time.
Then if a stack capture was requested, attempt to do it in the most efficient way, using a pre-allocated buffer if possible.
WIP note: This potential allocation should be avoided in time-critical thread. There is already a buffer for the main thread (because it’s the busiest thread), but there could be more pre-allocated threads, for specific real-time thread that need it, or picked from a pool of pre-allocated buffers. See bug 1578792.
From there, AddMarkerWithOptionalStackToBuffer
NoPayload markers (usually added with
in a special way, mostly to avoid the extra work associated with handling
payloads. Otherwise it continues with the following function.
MarkerTypeSerialization<MarkerType>::Serialize retrieves the deserialization tag associated with the marker type. If it’s the first time this marker type is used, Streaming::TagForMarkerTypeFunctions adds it to the global list (which stores some function pointers used during deserialization).
Then the main serialization happens in StreamFunctionTypeHelper<decltype(MarkerType::StreamJSONMarkerData)>::Serialize. Deconstructing this mouthful of an template:
MarkerType::StreamJSONMarkerDatais the user-provided function that will eventually produce the final JSON, but here it’s only used to know the parameter types that it expects.
StreamFunctionTypeHelpertakes that function prototype, and can extract its argument by specializing on
`R(SpliceableJSONWriter&, As...), now
As...is a parameter pack matching the function parameters.
Serializealso takes a parameter pack, which contains all the referenced arguments given to the top
AddBufferToMarkercall. These two packs are supposed to match, at least the given arguments should be convertible to the target pack parameter types.
Serializefunction calls the buffer’s
PutObjectsvariadic function to write all the marker data, that is:
The entry kind that must be at the beginning of every buffer entry, in this case ProfileBufferEntryKind::Marker.
The common marker data (options first, name, category, deserialization tag).
Then all the marker-type-specific arguments. Note that the C++ types are those extracted from the deserialization function, so we know that whatever is serialized here can be later deserialized using those same types.
The deserialization side is described in the later section “JSON output of Markers”.
Adding Markers from Rust¶
Adding Markers from JS¶
During a profiling session, some profiler-related events may be recorded using ProfilingLog::Access.
The resulting JSON object is added near the end of the process’ JSON generation, in a top-level property named “profilingLog”. This object is free-form, and is not intended to be displayed, or even read by most people. But it may include interesting information for advanced users, or could be an early temporary prototyping ground for new features.
See “profileGatheringLog” for another log related to late events.
WIP note: This was introduced shortly before this documentation, so at this time it doesn’t do much at all.
Usually at the end of a profiling session, a profile is “captured”, and either saved to disk, or sent to the front-end https://profiler.firefox.com for analysis. This section describes how the captured data is converted to the Gecko Profiler JSON format.
The FailureLatch interface is used during the JSON generation, in order to catch any unrecoverable error (such as running Out Of Memory), to exit the process early, and to forward the error to callers.
There are two main implementations, suffixed “source” as they are the one source
of failure-handling, which is passed as
FailureLatch& throughout the code:
FailureLatchInfallibleSource is an “infallible” latch, meaning that it doesn’t expect any failure. So if a failure actually happened, the program would immediately terminate! (This was the default behavior prior to introducing these latches.)
FailureLatchSource is a “fallible” latch, it will record the first failure that happens, and “latch” into the failure state. The code should regularly examine this state, and return early when possible. Eventually this failure state may be exposed to end users.
A ProgressLogger object is used to track the progress of a long operation, in this case the JSON generation process.
To match how the JSON generation code works (as a tree of C++ functions calls),
ProgressLogger in a function usually records progress from 0 to 100%
locally inside that function. If that function calls a sub-function, it gives it
a sub-logger, which in the caller function is set to represent a local sub-range
(like 20% to 40%), but to the called function it will look like its own local
ProgressLogger that goes from 0 to 100%. The very top
converts the deepest local progress value to the corresponding global progress.
Progress values are recorded in ProportionValue objects, which effectively record fractional value with no loss of precision.
This progress is most useful when the parent process is waiting for child processes to do their work, to make sure progress does happen, otherwise to stop waiting for frozen processes. More about that in the “Multi-Process Profiling” section below.
While these classes live outside of the Profiler directories, it may sometimes be worth maintaining and/or modifying them to better serve the Profiler’s needs. But there are other users, so be careful not to break other things!
SpliceableJSONWriter and SpliceableChunkedJSONWriter¶
Because the Profiler deals with large amounts of data (big profiles can take tens to hundreds of megabytes!), some specialized wrappers add better handling of these large JSON streams.
is a subclass of
JSONWriter, and allows the “splicing” of JSON strings,
i.e., being able to take a whole well-formed JSON string, and directly inserting
it as a JSON object in the target JSON being streamed.
It also offers some functions that are often useful for the Profiler, such as: * Converting a timestamp into a JSON object in the stream, taking care of keeping a nanosecond precision, without unwanted zeroes or nines at the end. * Adding a number of null elements. * Adding a unique string index, and add that string to a provided unique-string list if necessary. (More about UniqueStrings below.)
is a subclass of
SpliceableJSONWriter. Its main attribute is that it provides its own writer
which stores the stream as a sequence of “chunks” (heap-allocated buffers).
It starts with a chunk of a default size, and writes incoming data into it,
later allocating more chunks as needed. This avoids having massive buffers being
resized all the time.
It also offers the same splicing abilities as its parent class, but in case an
incoming JSON string comes from another
able to just steal the chunks and add them to its list, thereby avoiding
expensive allocations and copies and destructions.
Because a lot of strings would be repeated in profiles (e.g., frequent marker names), such strings are stored in a separate JSON array of strings, and an index into this list is used instead of that full string object.
Note that these unique-string indices are currently only located in specific spots in the JSON tree, they cannot be used just anywhere strings are accepted.
The UniqueJSONStrings class
stores this list of unique strings in a
Given a string, it takes care of storing it if encountered for the first time,
and inserts the index into a target
The “Gecko Profile Format” can be found at https://github.com/firefox-devtools/profiler/blob/main/docs-developer/gecko-profile-format.md .
The implementation in the back-end is locked_profiler_stream_json_for_this_process. It outputs each JSON top-level JSON object, mostly in sequence. See the code for how each object is output. Note that there is special handling for samples and markers, as explained in the following section.
ProcessStreamingContext and ThreadStreamingContext¶
In JSON profiles, samples and markers are separated by thread and by samples/markers. Because there are potentially tens to a hundred threads, it would be very costly to read the full profile buffer once for each of these groups. So instead the buffer is read once, and all samples and markers are handled as they are read, and their JSON output is sent to separate JSON writers.
A ProcessStreamingContext object
contains all the information to facilitate this output, including a list of
which each contain one
SpliceableChunkedJSONWriter for the samples, and one
for the markers in this thread.
When reading entries from the profile buffer, samples and markers are found by
ProfileBufferEntryKind, and as part of deserializing either kind (more
about each below), the thread ID is read, and determines which
ThreadStreamingContext will receive the JSON output.
At the end of this process, all
SpliceableChunkedJSONWriters are efficiently
spliced (mainly a pointer move) into the final JSON output.
JSON output of Samples¶
This work is done in ProfileBuffer::DoStreamSamplesAndMarkersToJSON.
From the main
ProfileChunkedBuffer, each entry is visited, its
ProfileBufferEntryKind is read first, and for samples all frames from
captured stack are converted to the appropriate JSON.
A UniqueStacks object is used to de-duplicate frames and even sub-stacks:
Each unique frame string is written into a JSON array inside a
SpliceableChunkedJSONWriter, and its index is the frame identifier.
Each stack level is also de-duplicated, and identifies the associated frame string, and points at the calling stack level (i.e., closer to the root).
Finally, the identifier for the top of the stack is stored, along with a timestamp (and potentially some more information) as the sample.
For example, if we have collected the following samples:
A -> B -> C
A -> B
A -> B -> D
The frame table would contain each frame name, something like:
["A", "B", "C", "D"]. So the frame containing “A” has index 0, “B” is at 1,
The stack table would contain each stack level, something like:
[[0, null], [1, 0], [2, 1], [3, 1]].
[0, null] means the frame is 0
(“A”), and it has no caller, it’s the root frame.
[1, 0] means the frame is
1 (“B”), and its caller is stack 0, which is just the previous one in this
And the three samples stored in the thread data would be therefore be: 2, 1, 3 (E.g.: “2” points in the stack table at the frame [2,1] with “C”, and from them down to “B”, then “A”).
All this contains all the information needed to reconstruct all full stack samples.
JSON output of Markers¶
This also happens inside ProfileBuffer::DoStreamSamplesAndMarkersToJSON.
ProfileBufferEntryKind::Marker is encountered,
the DeserializeAfterKindAndStream function
MarkerOptions (stored as explained above), which include the
thread ID, identifying which
SpliceableChunkedJSONWriter to use.
After that, the common marker data (timing, category, etc.) is output.
Streaming::DeserializerTag identifies which type of marker this is.
The special case of
0 (no payload) means nothing more is output.
Otherwise some more common data is output as part of the payload if present, in particular the “inner window id” (used to match markers with specific html frames), and stack.
In case of a C++-written payload, the
DeserializerTag identifies the
MarkerDataDeserializer function to use. This is part of the heavy templated
code in BaseProfilerMarkersDetail.h, the function is defined as
which outputs the marker type name, and then each marker payload argument. The
latter is done by using the user-defined
parameter list, and recursively deserializing each parameter from the profile
buffer into an on-stack variable of a corresponding type, at the end of which
MarkerType::StreamJSONMarkerData can be called with all of these arguments
at it expects, and that function does the actual JSON streaming as the user
See “Profiler Start” and do the reverse!
There is some special handling of the
SampleThread object, just to ensure
that it gets deleted outside of the main profiler mutex being locked, otherwise
this could result in a deadlock (because it needs to take the lock before being
able to check the state variable indicating that the sampling loop and thread
See “Profiler Initialization” and do the reverse!
One additional action is handling the optional
environment variable, to output a profile if the profiler was running.
All of the above explanations focused on what the profiler is doing is each process: Starting, running and collecting samples, markers, and more data, outputting JSON profiles, and stopping.
But Firefox is a multi-process program, since Electrolysis aka e10s introduce child processes to handle web content and extensions, and especially since Fission forced even parts of the same webpage to run in separate processes, mainly for added security. Since then Firefox can spawn many processes, sometimes 10 to 20 when visiting busy sites.
The following sections explains how profiling Firefox as a whole works.
IPC (Inter-Process Communication)¶
As a quick summary, some message-passing function-like declarations live in
RecvX C++ functions are respectively
and virtually declared (for user implementation) in
It’s called from
just before a child process ends, and if the profiler was running, to ensure
that the profile data is collected and sent to the parent, for storage in its
for details. Note that the current “buffer position at gathering time” (which is
effectively the largest
ProfileBufferBlockIndex that is present in the
global profile buffer) is recorded. Later,
looks at the smallest
ProfileBufferBlockIndex still present in the
buffer (because early chunks may have been discarded to limit memory usage), and
discards exit profiles that were recorded before, because their data is now
older than anything stored in the parent.
Profile Buffer Global Memory Control¶
Each process runs its own profiler, with each its own profile chunked buffer. To keep the overall memory usage of all these buffers under the user-picked limit, processes work together, with the parent process overseeing things.
Diagram showing the relevant classes, see details in the following sub-sections:
The ProfileBufferControlledChunkManager interface allows a controller to get notified about all chunk updates, and to force the destruction/recycling of old chunks. The ProfileBufferChunkManagerWithLocalLimit class implements it.
An Update object
contains all information related to chunk changes: How much memory is currently
used by the local chunk manager, how much has been “released” (and therefore
could be destroyed/recycled), and a list of all chunks that were released since
the previous update; it also has a special state meaning that the child is
shutting down so there won’t be updates anymore. An
Update may be “folded”
into a previous one, to create a combined update equivalent to the two separate
ones one after the other.
Update Handling in the ProfilerChild¶
When the profiler starts in a child process, the
starts to listen for updates.
These updates are stored and folded into previous ones (if any). At some point, an AwaitNextChunkManagerUpdate message will be received, and any update can be forwarded to the parent. The local update is cleared, ready to store future updates.
Update Handling in the ProfilerParent¶
When the profiler starts AND when there are child processes, the ProfilerParent’s ProfilerParentTracker creates a ProfileBufferGlobalController, which starts to listen for updates from the local chunk manager.
ProfilerParentTracker is also responsible for keeping track of child
processes, and to regularly
send them AwaitNextChunkManagerUpdate messages,
that the child’s
ProfilerChild answers to with updates. The update may
indicate that the child is shutting down, in which case the tracker will stop
All these updates (from the local chunk manager, and from child processes’ own chunk managers) are processed in ProfileBufferGlobalController::HandleChunkManagerNonFinalUpdate. Based on this stream of updates, it is possible to calculate the total memory used by all profile buffers in all processes, and to keep track of all chunks that have been “released” (i.e., are full, and can be destroyed). When the total memory usage reaches the user-selected limit, the controller can lookup the oldest chunk, and get it destroyed (either a local call for parent chunks, or by sending a DestroyReleasedChunksAtOrBefore message to the owning child).
Historical note: Prior to Fission, the Profiler used to keep one fixed-size circular buffer in each process, but as Fission made the possible number of processes unlimited, the memory consumption grew too fast, and required the implementation of the above system. But there may still be mentions of “circular buffers” in the code or documents; these have effectively been replaced by chunked buffers, with centralized chunk control.
Gathering Child Profiles¶
When it’s time to capture a full profile, the parent process performs its own JSON generation (as described above), and sends a GatherProfile message to all child processes, which will make them generate their JSON profile and send it back to the parent.
All child profiles, including the exit profiles collected during profiling, are stored as elements of a top-level array with property name “processes”.
During the gathering phase, while the parent is waiting for child responses, it regularly sends GetGatherProfileProgress messages to all child processes that have not sent their profile yet, and the parent expects responses within a short timeframe. The response carries a progress value. If at some point two messages went with no progress was made anywhere (either there was no response, or the progress value didn’t change), the parent assumes that remaining child processes may be frozen indefinitely, stops the gathering and considers the JSON generation complete.
During all of the above work, events are logged (especially issues with child processes), and are added at the end of the JSON profile, in a top-level object with property name “profileGatheringLog”. This object is free-form, and is not intended to be displayed, or even read by most people. But it may include interesting information for advanced users regarding the profile-gathering phase.