Make sleep(n) and Timer creation more MT-safe #32174
Conversation
|
|
src/jl_uv.c
Outdated
| return err; | ||
| } | ||
|
|
||
| jl_uv_associate_julia_struct((uv_handle_t *)handle, (jl_value_t *)handle); |
There was a problem hiding this comment.
one of these things is not like the other—you can't cast to both libuv and julia
There was a problem hiding this comment.
Sorry, I'm getting slightly confused by what's what, since in the following two locations it seems that we're already passing objects which (to my uneducated eyes) seem to end up with casts that are essentially the same as what you pointed out at:
Line 90 in 4b0c8e7
Lines 102 to 103 in 4b0c8e7
Those two libuv calls don't expect a jl_value_t*, even though that's what we're doing there, right? I assume it works because this.handle is the first field in the Timer struct so the pointers are the same anyway?
src/jl_uv.c
Outdated
| int err = uv_timer_init(loop, handle); | ||
| if (err) { | ||
| // TODO: this codepath is currently not tested | ||
| free(handle); |
There was a problem hiding this comment.
missing JL_UV_UNLOCK on this codepath, also probably better to keep the free call in Julia
There was a problem hiding this comment.
Hopefully my latest commit addresses both of these properly 😄
src/jl_uv.c
Outdated
| { | ||
| JL_UV_LOCK(); | ||
| int err = uv_timer_init(loop, uvtimer); | ||
| if (err) { |
There was a problem hiding this comment.
this actually can't fail, so we can just delete the error checking code (or make it an abort)
|
|
||
| jl_uv_associate_julia_struct((uv_handle_t*)uvtimer, jltimer); | ||
| uv_update_time(loop); | ||
| err = uv_timer_start(uvtimer, cb, timeout, repeat); |
There was a problem hiding this comment.
similarly, this can't actually fail either, we can just abort if it returns non-zero
JeffBezanson
changed the title
|
I believe I encountered one of the hanging conditions mentioned at the top of this PR, and wanted to note for others that |
This returns some of the locking removed by 93e3d28 (which @JeffBezanson said was removed because it was probably incorrect), but this time the calls to
uv_update_timeanduv_timer_startare done within a single lock/unlock of the global libuv lock (instead of two separate lock/unlock pairs).This still hangs every so often on my system with
JULIA_NUM_THREADS=8when running a slightly longer version of the test laid out in #32152, which I'd like to have this PR fix once I figure out the root cause. I also think this PR should have at least one test to ensure highly-contended calls tosleeporTimerdon't cause segfaults or hangs; I'll add a simple test to this PR soon to get something on the table.(Eventually) fixes #32152