Hi Thomas,
Can you tell me what is the version of the lustre servers and the lustre
clients ?
Thanks,
Maxime Boissonneault
HPC specialist @ Calcul Québec.
Le 2013-01-21 11:11, Thomas Rast a écrit :
Erik Faye-Lund kusmab...@gmail.com writes:
On Fri, Jan 18, 2013 at 6:50 PM, Eric Chamberland
eric.chamberl...@giref.ulaval.ca wrote:
Good idea!
I did a strace and here is the output with the error:
http://www.giref.ulaval.ca/~ericc/strace_git_error.txt
Hope it will be insightful!
This trace doesn't seem to contain child-processes, but instead having
their stderr inlined into the log. Try using strace -f instead...
I happen to have access to a lustre FS on the brutus cluster of ETH
Zurich, so I figured I could give it a shot.
What's odd is that while I cannot reproduce the original problem, there
seems to be another issue/bug with utime():
$ strace -f -o ~/gc.trace git gc
warning: failed utime() on .git/objects/68/tmp_obj_sCAEVc: Interrupted
system call
warning: failed utime() on .git/objects/a6/tmp_obj_3cdB2c: Interrupted
system call
warning: failed utime() on .git/objects/69/tmp_obj_lbU3Xc: Interrupted
system call
warning: failed utime() on .git/objects/c3/tmp_obj_EU97Wc: Interrupted
system call
warning: failed utime() on .git/objects/3e/tmp_obj_tb2j3c: Interrupted
system call
warning: failed utime() on .git/objects/15/tmp_obj_e6zMXc: Interrupted
system call
warning: failed utime() on .git/objects/54/tmp_obj_ExOJVc: Interrupted
system call
warning: failed utime() on .git/objects/e3/tmp_obj_GtPw4c: Interrupted
system call
warning: failed utime() on .git/objects/21/tmp_obj_Xex32c: Interrupted
system call
warning: failed utime() on .git/objects/1a/tmp_obj_CzwsZc: Interrupted
system call
warning: failed utime() on .git/objects/18/tmp_obj_o6fp3c: Interrupted
system call
warning: failed utime() on .git/objects/32/tmp_obj_Ih0G4c: Interrupted
system call
warning: failed utime() on .git/objects/41/tmp_obj_1RXV1c: Interrupted
system call
Counting objects: 137744, done.
Delta compression using up to 48 threads.
Compressing objects: 100% (36510/36510), done.
Writing objects: 100% (137744/137744), done.
Total 137744 (delta 101591), reused 135512 (delta 99472)
The trace is here (2.1MB compressed):
http://thomasrast.ch/download/gc.trace.bz2
For the test I used a clone of another git.git I had around. I think
the error is from sha1_file.c:2564. While that doesn't look too
important (see ca11b212 for context), it does raise the question: what
other system calls that we never expect to EINTR can do this on
sufficiently arcane system/FS combinations?
Peff's test ran without any apparent issue for a few minutes. I also
ran an extended version (at the end) that sets alarms, so as to actually
get interrupted. That proved more interesting. I had to fix verify()
and write_in_full() to account for EINTR in read()/write(), as those
seem likely to fail. I also got link() to fail:
$ ~/lustre-peff-reproducer
unable to create hard link: Interrupted system call
unable to open index file: No such file or directory
but it took a long time. Unfortunately, when running it with strace I
managed to panic the host I ran it on:
$ strace -o ~/peff-reproducer.trace ~/lustre-peff-reproducer
Message from syslogd@brutus1 at Jan 21 17:09:43 ...
kernel:LustreError: 37417:0:(osc_lock.c:1182:osc_lock_enqueue()) ASSERTION(
ols-ols_state == OLS_NEW ) failed: Impossible state: 4
Message from syslogd@brutus1 at Jan 21 17:09:43 ...
kernel:LustreError: 37417:0:(osc_lock.c:1182:osc_lock_enqueue()) LBUG
Message from syslogd@brutus1 at Jan 21 17:09:43 ...
kernel:Kernel panic - not syncing: LBUG
Yay for now having to explain this to the cluster team.
I tried finding a standard that limits the syscalls to which EINTR
applies, without too much success. I'm not sure how far I should trust
my manpages, but while some of them explicitly list EINTR as a possible
error (read, write, etc.) link() does not. (And the linux manpages
agree with the POSOIX ones for once.)
If somebody finds such a standard, we could of course use it to blame
lustre instead :-)
In the absence of it, wouldn't we in theory have to write a simple
loop-on-EINTR wrapper for *all* syscalls?
Of course there's the added problem that when open(O_CREAT|O_EXCL) fails
with EINTR, it's hard to tell whether a file that may now exist is
indeed yours or some other process's.
--- 8
#include fcntl.h
#include unistd.h
#include stdlib.h
#include stdio.h
#include string.h
#include sys/time.h
#include signal.h
#include errno.h
struct itimerval itv;
static int randomize(unsigned char *buf, int len)
{
int i;
len = rand() % len;
for (i = 0; i len; i++)
buf[i] = rand() 0xff;
return len;
}
static int check_eof(int fd)
{
int ch;
int r = read(fd, ch, 1);
if (r 0) {
perror(read error after expected EOF);
return -1;
}
if (r 0
