-
Notifications
You must be signed in to change notification settings - Fork 3.5k
Fix a deadlock issue with emscripten_lock_async_acquire() #25670
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
base: main
Are you sure you want to change the base?
Changes from all commits
File filter
Filter by extension
Conversations
Jump to
Diff view
Diff view
There are no files selected for viewing
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -188,6 +188,10 @@ void emscripten_lock_busyspin_waitinf_acquire(emscripten_lock_t *lock __attribut | |
| // NOTE: This function can be called in both main thread and in Workers. If you | ||
| // use this API in Worker, you cannot utilise an infinite loop programming | ||
| // model. | ||
| // NOTE 2: This function will always acquire the lock asynchronously. That is, | ||
| // the lock will only be attempted to acquire after current control flow | ||
| // yields back to the browser, so that the Wasm call stack is empty. | ||
| // This is to guarantee an uniform control flow. | ||
|
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. This extra note seems redundant with the second sentence above? Maybe you can tune the second sentence instead of re-stating here if you think there is a difference? |
||
| void emscripten_lock_async_acquire(emscripten_lock_t *lock __attribute__((nonnull)), | ||
| emscripten_async_wait_volatile_callback_t asyncWaitFinished __attribute__((nonnull)), | ||
| void *userData, | ||
|
|
||
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -5329,6 +5329,12 @@ def test_wasm_worker_lock_wait2(self): | |
| def test_wasm_worker_lock_async_acquire(self): | ||
| self.btest_exit('wasm_worker/lock_async_acquire.c', cflags=['--closure=1', '-sWASM_WORKERS']) | ||
|
|
||
| # Tests emscripten_lock_async_acquire() function when lock is acquired both synchronously and asynchronously. | ||
| @also_with_minimal_runtime | ||
| @flaky('https://github.com/emscripten-core/emscripten/issues/25270') | ||
|
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. Because this test doesn't actually create any workers this flaky decorator isn't applicable is it? |
||
| def test_wasm_worker_lock_async_and_sync_acquire(self): | ||
| self.btest('wasm_worker/lock_async_and_sync_acquire.c', expected='1', cflags=['--closure=1', '-sWASM_WORKERS']) | ||
|
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. Can this use There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. (Also does it need |
||
|
|
||
| # Tests emscripten_lock_busyspin_wait_acquire() in Worker and main thread. | ||
| @also_with_minimal_runtime | ||
| def test_wasm_worker_lock_busyspin_wait(self): | ||
|
|
||
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,29 @@ | ||
| #include <emscripten/wasm_worker.h> | ||
| #include <emscripten/threading.h> | ||
| #include <stdio.h> | ||
|
|
||
| emscripten_lock_t lock = EMSCRIPTEN_LOCK_T_STATIC_INITIALIZER; | ||
|
|
||
| int result = 0; | ||
|
|
||
| void on_acquire(volatile void* address, uint32_t value, | ||
| ATOMICS_WAIT_RESULT_T waitResult, void* userData) { | ||
| printf("on_acquire: releasing lock.\n"); | ||
| emscripten_lock_release(&lock); | ||
| printf("on_acquire: released lock.\n"); | ||
| #ifdef REPORT_RESULT | ||
| REPORT_RESULT(result); | ||
|
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. Can't you replace this with |
||
| #endif | ||
| } | ||
|
|
||
| int main() { | ||
| printf("main: async acquiring lock.\n"); | ||
| emscripten_lock_async_acquire(&lock, on_acquire, 0, 100); | ||
| printf("main: busy-spin acquiring lock.\n"); | ||
| emscripten_lock_busyspin_waitinf_acquire(&lock); | ||
| printf("main: lock acquired.\n"); | ||
| emscripten_lock_release(&lock); | ||
| printf("main: lock released.\n"); | ||
| result += 1; | ||
| emscripten_exit_with_live_runtime(); | ||
| } | ||
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
As discussed in b25abd5#r168975511
I prefer this to run as fast as possible if possible. I don't think it is unreasonable to expect the
emscripten_lock_async_acquireto run directly with the things that it entails if the lock is free.But if we are to do this, would
queueMicrotaskwork here instead? That would avoid things like renders to run before we try to acquire the lock.There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
It might be possible, although the consensus here is that a microtask should be a short-lived task. We have no possibility to state the semantics on behalf of the user. If this was a new API, we could freely say that should be the model - but since this is an already shipped API, we cannot change/impose semantics on existing users.
I think it would be best to compose using the existing functions. For example,
or
Would give a convenient way to get the fast access path synchronously.
Performance here should be optimal whenever there is no long-lived contention. And in the case there is > 0.5msec contention, latency will be on the slow path in any case since
emscripten_lock_async_acquire()will yield to the event loop (the fact that it performs a single extra CAS will not be observable.. e.g.emscripten_lock_busyspin_wait_acquire()will have performed millions of CASes already, one more won't matter)There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Do you have any links to these discussions?
The Task vs Microtask has no semantic difference in "size", just when they are expected to run.
The thing we need to ask ourself is if we want the
Atomic.awaitSyncto be able to happen before or after a page render if it is queued (or other similar tasks). Eg. should we always yield to the eventloop when we want to aquire the lock async. I would argue to only yield if the lock is blocking.But in the end, I don't think this matters a lot compared to the big difference of removing the yielding in the critical section which this PR fixes 👍
Uh oh!
There was an error while loading. Please reload this page.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This was from
https://developer.mozilla.org/en-US/docs/Web/API/Window/queueMicrotask and https://developer.mozilla.org/en-US/docs/Web/API/HTML_DOM_API/Microtask_guide : "The microtask is a short function which will run after ... "
I understand you're in the mindset of designing what would be the best API. And I agree with that, though given this is an already shipped API, that would change the semantics. For example, if user writes
Should the async timeout callback trigger first? Or the async lock acquire trigger first?
Currently the computation model is consistent to always trigger the timeout first. Using a microtask would change the async lock callback to trigger either before or after the timeout depending on whether there was contention from other threads or not.
Maybe it would be a better design for one to say "well the above should be unspecified, don't rely on it as a end-user." But given this is an already shipped API, I am very cautious to change that behavior.
It is possible to manually control this behavior with the above two constructs to get that sync functionality, so one can already get the necessary thing with a couple of extra lines without a performance penalty.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I missed the current description.
"The calling thread will asynchronously try to obtain the given lock after the calling thread yields back to the event loop"
With it being explicitly defined, I agree we should not change its behavior, compared to it being undefined.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Personally I think using the microtask queue should be fine here, and we could update the documentation to say "yields back to the microtask queue".
If there is code out there that is dependant on the ordering the the two callback about that seems way to fragile. Also, such code would have been broken by the existence of the current bug (i.e. the current bug this this PR fixes basically ensures that no such code exists in the wild yet, so now would be good time to switch to the microtask queue.. although that should be a followup PR I think).