Occasional segfaults when running with @threads 3 #45196
Comments
|
For the record, with this code I cannot reproduce the crash nor the valgrind errors NUM_THREADS = 30
c=Channel()
function worker()
while true
take!(c)
a = Float32[]
for i in 1:rand(1:2000)
push!(a, 0.3f0)
end
end
end
tasks = [schedule(Task(worker)) for i in 1:NUM_THREADS]
for i in 1:100_000_0
i % 1000 == 0 && @info "i" i
for t in 1:NUM_THREADS
put!(c, nothing)
end
end
EDIT: The code above does not reproduce the crash because all workers run on the same thread. With the code below I can also reproduce the issues. So, the problem is not when recreating the threads, but when allocating memory on a spawned thread. |
Have you tried it under rr? If you can catch the crash there and it still appears to be a bug in the julia runtime, you may have to send me the trace privately. |
|
@JeffBezanson The problem is that:
thanks |
|
Hi, from julia import Main
import pygame
Main.eval("""
function worker()
while true
a = Float64[]
push!(a, 0.42)
end
end
t= Threads.@spawn worker()
wait(t)
""") |
|
Wow, interesting. Thanks for digging in to that. |
|
Doss pygame do anything with signals? |
|
It is likely a PythonCall issue, since I think that package currently disables Julia's signal handling (signal handling is mandatory to use threads) |
This is a followup from #44460 (comment)
My real application often crashes when executed normally. Once I got this error (without valgrind):
The MWE below (which just created arrays from threads) never crashes when run normally or with rr.
However, it always quickly crashes with valgrind. I could not run it with julia 1.9 due to this, so I used julia 1.8.3. I use a sysimg with cpu_target=generic.
I get different error reports depending on the execution, but it usually ends with
Once I got these more informative errors about mutexes & GC.
The text was updated successfully, but these errors were encountered: