github.com/moarvm/moarvm | IRC logs at colabti.org/irclogger/irclogger_logs/moarvm
Set by AlexDaniel on 12 June 2018.
00:58 leont joined 01:49 Altai-man_ joined 01:52 sena_kun left 02:15 sivoais left 02:20 leont left 03:50 sena_kun joined 03:52 Altai-man_ left 05:50 Altai-man_ joined 05:52 sena_kun left 07:47 gugod1 is now known as gugod 07:51 sena_kun joined 07:52 Altai-man_ left
nine dogbert17: oh, I would love to know that! 08:17
dogbert17: can you reproduce it somehow at least some times?
08:38 patrickb joined 09:50 Altai-man_ joined 09:53 sena_kun left 10:39 leont joined
dogbert17 nine: I can reproduce the problem if I run the scripts in a loop 11:27
it looks like this: gist.github.com/dogbert17/892d5892...580cf59e39 11:30
11:51 sena_kun joined 11:52 Altai-man_ left 13:14 lucasb joined 13:17 regreg left 13:50 Altai-man_ joined 13:52 sena_kun left 14:36 patrickb left 15:42 patrickb joined 15:51 sena_kun joined 15:52 Altai-man_ left 16:43 lucasb left
nine dogbert17: oh, so it's failing to acquire some lock 16:45
17:32 domidumont joined
nine Of course the issue refuses to get reproduced on my machine 17:35
17:49 Altai-man_ joined 17:52 sena_kun left
nine dogbert17: it would be interesting to know what return value causes the locking failure. We have that in r in 3rdparty/libuv/src/unix/signal.c:121 but change all return values to -1 in line 124. That seems unnecessary as all non-0 return values are interpreted as failure anyway. 18:10
dogbert17: you could change that to just return r and print the error value to stderr in line 150 before calling abort() 18:11
dogbert17 nine: I'll give it a shot 18:19
I usually run a spectest in another window in order to stress the system a bit 18:24
nine tried that already with TEST_JOBS=80 18:25
dogbert17 Now that's a high number :) 18:27
Now the question is, will it crash again ... 18:32
patrickb Can someone have a look at github.com/MoarVM/MoarVM/pull/1248 It's a pretty basic PR. 18:34
nine patrickb: why is that even using qx in the first place? 18:51
patrickb nine: What would be the alternative? 18:52
nine system of course 18:53
Or open if you need to read the command's output 18:56
Both take lists of arguments which is much safer
open my $fh, '-|', $cmd, $arg1, $arg2, ...; 18:57
18:58 domidumont left
dogbert17 Heh, the bug seems to have gone on strike 18:59
Noticed another strange thing though. I'm running while ./nqp t/nqp/111-spawnprocasync.t; do :; done 19:00
and suddenly some odd text was printed between tests. It turned out to be this: 19:01
Out-of-sync package detected in package_declarator_class at oder') {}
my $dec := nqp::crea
(value in braid: GLOBALish, value in $*PACKAGE: GLOBALish)
patrickb nine: I'll look
dogbert17 gets something to eat 19:09
19:16 zakharyas joined 19:50 sena_kun joined 19:52 Altai-man_ left
dogbert17 nine: it seems as if the 'real' return value is also -1 19:57
nine dogbert17: oh, damn....should have read the docs first: "On error, -1 is returned, and errno is set appropriately." 20:07
according to read(2)
dogbert17 fortunately I printed that value as well :-) 20:08
it's YOU 20:09
i.e. errno = 9 :)
patrickb nine: How would I go about using system or open when I want to suppress and capture stdout and stderr?
moritz patrickb: is that a perl 5 or a raku question? 20:12
nine moritz: perl 5 20:13
moritz then IPC::Run is an often cited answer to that question 20:14
or IPC::Open2 / IPC::Open3 if you are looking for something in perl core
patrickb moritz: Perl5. I was just tasked with removing qx from github.com/MoarVM/MoarVM/blob/mast...modules.pl 20:16
nine patrickb: why do you even capture STDERR? Frankly, I don't think it's worth the hassle 20:17
patrickb It's used to provide an error message containing the original output. 20:18
nine But if you don't capture the original output it will still be there 20:19
The order may be reversed, but that's still better than not working at all if your directories are named a bit unusual 20:20
patrickb True. I'm more concerned about 2&>1 not working on Windows than I am about the dir quoting though.
nine dogbert17: errno 9 is EBADF /* Bad file number */ 20:21
dogbert17: so my guess would be that it's -1 in this case 20:27
dogbert17 can file descriptors be negative? 20:28
nine No. -1 is used in libuv to signify that they are not yet opened: static int uv__signal_lock_pipefd[2] = { -1, -1 }; 20:29
20:30 sivoais joined
dogbert17 nine: I think I managed to catch the problem in strace 20:44
nine: gist.github.com/dogbert17/c2b2687e...6ebedc72ea 20:46
20:47 zakharyas left
nine dogbert17: but that looks like a normal successful run? 20:49
dogbert17 but what about the line: [pid 9135] tgkill(9133, 9135, SIGABRT <unfinished ...>
line 122 in the gist
nine oh....got distracted by all the exited with 0 20:50
dogbert17 I have seen the r = -1 and errno = 9 sometimes but my while loop running the script didn't quit. This is just like it I believe. 20:51
do you want a normal run to compare with? 20:54
nine I've done that locally 20:55
Also of interest: [pid 9135] <... write resumed> ) = -1 EPIPE (Broken pipe) 20:57
The writes and reads of "*" are uv__signal_lock and uv__signal_unlock 20:58
So the write to the broken pipe is uv__signal_unlock
vrurg How hard would it be to dump stack trace in MVM_exception_throwpayload prior to a panic? I'm getting "MoarVM panic: Internal error: Unwound entire stack and missed handler" and can't locate the problem. 20:59
nine vrurg: I'd just set a break point there and call MVM_dump_backtrace(tc) 21:00
vrurg nine: Sorry, missed the key point: raku/nqp stack. 21:01
nine vrurg: MVM_dump_backtrace will give you exactly that 21:02
vrurg Oh, thanks!
nine dogbert17: I'm pretty sure "[pid 9133] close(4 <unfinished ...>" is closing uv__signal_lock_pipefd[0], and that's why the "[pid 9135] write(5, "*", 1 <unfinished ...>" immediately after it fails 21:03
vrurg Though I'm afraid there could be another problem. The panic is produced by frame unwinding. And it lefts NULL in tc->cur_frame. Can I just temporarily store the original pointer without taking any measures agains gc collection? 21:04
nine dogbert17: can you try it with libuv commit 1de151700d30629090d6a06e47cb92d538601577 reverted? 21:07
dogbert17 ah, you suspect a libuv bug? 21:12
nine Well, I suspect it's a regression and we did have a libuv bump recently. And that bump brings a commit that touches the signal code and is related to closing. Worth a try I'd say 21:14
dogbert17 gah, my revert gets reverted when building moarvm 21:22
nine dogbert17: looks like you need to commit that submodule change in the MoarVM repo, too 21:25
dogbert17 yeah 21:26
do I have to rebuild rakudo as well? 21:29
nine no
dogbert17 unless I did something wrong it soes not help, I still see 21:34
[pid 13713] exit_group(0) = ?
[pid 13739] tgkill(13713, 13739, SIGABRT <unfinished ...
nine Well, would have been nice... 21:45
dogbert17 for how long have you noticed this problem? 21:46
21:50 Altai-man_ joined
nine First confirmed time seems to be February 5th 21:50
no, the 7th
dogbert17 That's the same day that we updated to libuv 1.34.2 21:51
nine Can you try downgrading again? 21:52
Maybe I just suspected the wrong commit
21:52 sena_kun left
patrickb nine: Now I switched over to open(). Mind to have another look? 21:54
dogbert17 the previous version we used was 1.33
which was merged on Oct. 18 2019 21:55
Geth MoarVM: 9a706ba082 | (Patrick Böker)++ | tools/update-submodules.pl
Change update-submodules.pl to not use the shell at all

This should make this more platform agnostic and a bit safer. As a side effect of this the detailed error reporting that checked the STDERR and reported based on that is now gone, because it's not easily possible in perl to capture STDERR without using backticks.
22:02
MoarVM: bb69a97426 | niner++ (committed using GitHub Web editor) | tools/update-submodules.pl
Merge pull request #1248 from patrickbkr/git-cache-dir-with-spaces

Fix buiding when git reference dir has spaces in its path
dogbert17 I get tgkill(4179, 4190, SIGABRT <unfinished ...> with version 1.33 of libuv as well 22:03
something must have changes so that the exitcode is not always zero
*changed
nine So you get that SIGABRT in strace but no core dump? 22:05
dogbert17 correct
nine so it _is_ the change in libuv
is the same true for the revert you tried previously? 22:06
dogbert17 yes
the coredumps quite seldom but the tgkill line can be reproed quite easily by having a spectest running in the background 22:07
nine Oooh...could it just be that it previously swallowed that SIGABRT?
dogbert17 quite possible
nine * Calls to raise() or abort() to programmatically raise a signal are not detected by libuv; these will not trigger a signal watcher. 22:08
according to the docs
22:10 zakharyas joined
leont Catching SIGABRT is generally a terrible idea, only acceptable use-case would be attaching a debugger or some such 22:12
nine is off to bed now 22:21
patrickb o/ 22:22
22:33 zakharyas left
Geth MoarVM: patrickbkr++ created pull request #1256:
Fix submodule updating
22:40
patrickb The above should be merged soon. Build from fresh checkouts broke with the previous commit. The above is the corresponding fix. 22:41
23:19 harrow left 23:45 patrickb left 23:47 harrow joined 23:50 sena_kun joined 23:52 Altai-man_ left