-
Alex Kodat authored
While the sampler checked if the sampled thread had the Isolate locked (if locks are being used) under Linux, the check was not done under Windows (or Fuchsia) which meant that in a multi-threading application under Windows, thread locking was not checked making it prone to seg faults and the like as the profiler would be using isolate->js_entry_sp to determine the stack to walk but isolate->js_entry_sp is the stack pointer for the thread that currently has the Isolate lock so, if the sampled thread does not have the lock, the sampler woud be iterating over the wrong stack, one that might actually be actively changing on another thread. The fix was to move the lock check into CpuSampler and Ticker (--prof) so all OSes would do the correct check. The basic concept is that on all operating systems a CpuProfiler, and so its corresponding CpuCampler, the profiler is tied to a thread. This is not based on first principles or anything, it's simply the way it works in V8, though it is a useful conceit as it makes visualization and interpretation of profile data much easier. To collect a sample on a thread associated with a profiler the thread must be stopped for obvious reasons -- walking the stack of a running thread is a formula for disaster. The mechanism for stopping a thread is OS-specific and is done in sample.cc. There are currently three basic approaches, one for Linux/Unix variants, one for Windows and one for Fuchsia. The approaches vary as to which thread actually collects the sample -- under Linux the sample is actually collected on the (interrupted) sampled thread whereas under Fuchsia/Windows it's on a separate thread. However, in a multi-threaded environment (where Locker is used), it's not sufficient for the sampled thread to be stopped. Because the stack walk involves looking in the Isolate heap, no other thread can be messing with the heap while the sample is collected. The only ways to ensure this would be to either stop all threads whenever collecting a sample, or to ensure that the thread being sampled holds the Isolate lock so prevents other threads from messing with the heap. While there might be something to be said for the "stop all threads" approach, the current approach in V8 is to only stop the sampled thread so, if in a multi-threaded environment, the profiler must check if the thread being sampled holds the Isolate lock. Since this check must be done, independent of which thread the sample is being collected on (since it varies from OS to OS), the approach is to save the thread id of the thread to be profiled/sampled when the CpuSampler is instantiated (on all OSes it is instantiated on the sampled thread) and then check that thread id against the Isolate lock holder thread id before collecting a sample. If it matches, we know sample.cc has stop the sampled thread, one way or another, and we know that no other thread can mess with the heap (since the stopped thread holds the Isolate lock) so it's safe to walk the stack and collect data from the heap so the sample can be taken. It it doesn't match, we can't safely collect the sample so we don't. Bug: v8:10850 Change-Id: Iba6cabcd3e11a19c261c004103e37e806934dc6f Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2411343Reviewed-by: Peter Marshall <petermarshall@chromium.org> Commit-Queue: Peter Marshall <petermarshall@chromium.org> Cr-Commit-Position: refs/heads/master@{#69952}
76217f57
Name |
Last commit
|
Last update |
---|---|---|
.. | ||
benchmarks | ||
cctest | ||
common | ||
debugger | ||
debugging | ||
fuzzer | ||
fuzzilli | ||
inspector | ||
intl | ||
js-perf-test | ||
memory | ||
message | ||
mjsunit | ||
mkgrokdump | ||
mozilla | ||
test262 | ||
torque | ||
unittests | ||
wasm-api-tests | ||
wasm-js | ||
wasm-spec-tests | ||
webkit | ||
BUILD.gn | ||
OWNERS |