mon: do crushtool test with fork and timeout, but w/o exec of crushtool #16025
Conversation
liewegas
force-pushed
the
wip-19964
branch
2 times, most recently
from
2b169fe
to
093415d
Compare
src/common/fork_function.h
Outdated
| return 128 + WTERMSIG(status); | ||
| } | ||
| if (WIFEXITED(status)) { | ||
| int r = WEXITSTATUS(status); |
There was a problem hiding this comment.
@liewegas
WEXITSTATUS(status) evaluates to the low-order 8 bits of the argument passed to exit() by the child.
We could make the std::function int8_t, and change this to:
int8_t r = WEXITSTATUS(status);
I suppose this would allow to avoid "unsigned awkwardness".
src/common/fork_function.h
Outdated
| errstr << ": timed out\n"; | ||
| return -ETIMEDOUT; | ||
| } | ||
| errstr << ": exit status: " << r << "\n"; |
There was a problem hiding this comment.
Probably it is better to output r -1 value?
|
After |
|
Updated, mind looking? Thanks! |
src/common/fork_function.h
Outdated
|
|
||
| // Run a function post-fork, with a timeout. The exit code encoding | ||
| // weirdness does seem to want me to allow negative values, so we make | ||
| // the std::function unsigned. fork_function() itself is signed |
There was a problem hiding this comment.
@liewegas need to update the comment and the commit log message?
Signed-off-by: Sage Weil <sage@redhat.com>
Not strictly necessary, but a tidier. Signed-off-by: Sage Weil <sage@redhat.com>
We see timeouts here, but I very much suspect they are due to the overhead of launching the crushtool process and not the test itself. We have perfectly code already in our process, though; we just want to isolate failure and time out reliably. So, fork and timeout, without executing a new binary. Hopefully-fixes: http://tracker.ceph.com/issues/19964 Signed-off-by: Sage Weil <sage@redhat.com>
test_with_fork is superior in all ways :) Signed-off-by: Sage Weil <sage@redhat.com>
Signed-off-by: Sage Weil <sage@redhat.com>
|
made a few changes, will test again. |
This basically copies the Subprocess hackery for the timeout of the forked
subprocess, but operates on a lamba (in this case, CrushTester::test()) instead
of running crushtool via exec(2).