Use an exit flag as well as pthread_cancel to tell worker threads to terminate #8
Conversation
|
pedantically, you want "its" wrt pthread_cancel changing, imo the commit message needs more details/references. you should also leave comments in the code as to why you're switching to a manual polling method. by not using canceling, you possibly allow threads to stay hung up in loops/syscalls when a cancel request would kick them out automatically (look at all the "cancellation endpoints"). |
chutz
force-pushed
the
use-flag-rather-than-pthread-cancel
branch
from
ba49a49 to
290d1ee
Compare
|
That's a good point. I have updated the patch to take both approaches, so we will get all the advantages of pthread_cancel, and the backup of an exit flag for OSX , so if the cancel doesn't propagate for some reason we will still exit. |
chutz
force-pushed
the
use-flag-rather-than-pthread-cancel
branch
from
290d1ee to
a1851d4
Compare
chutz
changed the title
chutz
force-pushed
the
use-flag-rather-than-pthread-cancel
branch
from
a1851d4 to
ac462a9
Compare
Since this variable is protected by locks everywhere it is used, volatile is useless, and just makes things slower.
chutz
force-pushed
the
use-flag-rather-than-pthread-cancel
branch
2 times, most recently
from
13b407d to
21c72dd
Compare
|
i think their preferred term now is "macOS" :) i'm still confused by what you mean by "misbehaves". does it randomly crash ? randomly hang ? randomly make chicken soup ? or do the other threads just randomly fail to notice they've been canceled ? |
|
Sorry about that, I hadn't realized that I didn't properly articulate that problem. It seems that the threads occasionally fail to notice that they have been cancelled, and never terminate. When the main thread joins them, it ends up hanging forever. |
|
OK -- throw that in the comment. guess there isn't a whole lot we could do if calling pthread_cancel() goes unnoticed but the threads keep checking. can you do some measurements on perf impact ? maybe we should ifdef the damage away to |
chutz
force-pushed
the
use-flag-rather-than-pthread-cancel
branch
from
21c72dd to
8b1a78d
Compare
|
OK, I just pushed over some changes that refactor so all the thread stopping/cancellation code is in one place, then gated the damage away to |
chutz
force-pushed
the
use-flag-rather-than-pthread-cancel
branch
2 times, most recently
from
216c62b to
72a53a2
Compare
| #ifdef __APPLE__ | ||
| /* Check if we should exit, we are doing both cancel and exit condition | ||
| * since on OSX threads seem to occasionally fail to notice when they have | ||
| * been cancelled. We want to havea backup to make sure that we won't hang |
|
overall seems fine to me now |
chutz
force-pushed
the
use-flag-rather-than-pthread-cancel
branch
2 times, most recently
from
020749e to
d648cef
Compare
|
OK, final changes made. |
…tion This code is pretty repetitive, so let's split it out to its own function so we won't have so much repeated code.
chutz
force-pushed
the
use-flag-rather-than-pthread-cancel
branch
from
d648cef to
efaeffa
Compare
|
thanks! let's see if @rapier1 has anything to add. |
It turns out that apparently on OSX, pthread_cancel is not completely reliable. It seems that occasionally a thread will fail to notice that it has been cancelled. Repeatedly running the openssh integrity test suite will eventually deadlock on exit on OSX (I think it took about an hour for the person I had testing this). This changes the code to use an exit flag as well as pthread_cancel to tell the worker threads to exit. This should give us the best of both worlds.
chutz
force-pushed
the
use-flag-rather-than-pthread-cancel
branch
from
efaeffa to
526b442
Compare
|
One final minor fix I though of this morning. After stopping the threads in |
|
merged! |
It turns out that OSX's pthread_cancel implementation is not very reliable, so this switches to a flag variable rather than using pthread_cancel to tell threads that they should exit.
This also removes the volatile keyword from the qstate variable. It's not necessary since the variable is always protected by a mutex, and it causes the compiler to not be able to optimize properly.