On 05/03/13 22:24, Benjamin Marzinski wrote:
On Fri, May 03, 2013 at 08:36:19AM +0200, Bart Van Assche wrote:
On 05/02/13 23:46, Benjamin Marzinski wrote:
The patch cleans up some signal issues.
First, when the vecs locking around reconfigure got shuffled
around earlier, it was removed from sighup. This patch restores
that.
Second, a new sigusr1 handler was created. However the existing
one was never removed. Since signal handlers are per-process, and
not per-thread, the original handler will get overwritten by the
new one, so this patch deletes the original handler.
Third, sighup locks the vecs lock and sigusr1 locks logq_lock.
However, these signals weren't being blocked before threads locked
those locks. This patch blocks those signals while those locks
are being taken to avoid locking deadlocks.
Are you aware that POSIX does not allow any locking function to be invoked
from inside a signal handler ? See e.g.
http://pubs.opengroup.org/onlinepubs/9699919799/functions/V2_chap02.html#tag_15_04
and the text that starts with "The following table defines a set of
functions that shall be async-signal-safe. Therefore, applications can
invoke them, without restriction, from signal-catching functions" for a
list of C library functions that may be invoked from inside a signal
handler.
Um. no. I clearly wasn't aware of that. Nuts. So is the risk that a
signal will come into some thread that's not blocking it, and try to
acquire the pthread mutex at the same time as another thread that is
blocking the signal is trying to aquire the same mutex and that this
will corrupt the lock?
The Ubuntu man page for pthread_mutex_lock says
"The mutex functions are not async-signal safe. What this means is that
they should not be called from a signal handler. In particular, calling
pthread_mutex_lock or pthread_mutex_unlock from a signal handler may
deadlock the calling thread."
If that's the only risk, then blocking the signal before we lock the
mutexes should avoid the issue. Since the point of mutexes is do deal
with multiple threads that are trying to acquire them at the same time,
it seems like they should be able to handle this when one of the threads
happens to be in a signal handler. Also, the locks in sighup were there
for a long time before they were removed, and while I have definitely
seen deadlocks when we don't remember to mask the signal before locking
on that thread, I've never encountered a bug that seems to be related to
a locking corruption like I speculated in the beginning of my reply.
Now, since my patch was to fix some potential deadlocks from using
pthread_mutex_locks in signal handlers, I do realize that the way
we are doing things isn't the safest way around. Also, it seems pretty
obvious from looking at the URL you posted that we are in undefined
behavior territory, where if we are safe, it's solely based on the
some non-defined parts of the implementation. So I agree, we aren't
following the specs, and I'll work on redesigning things to avoid this.
But the locking I added in this patch fixes corruption that definitely
happens, and is very much able to crash multipathd. I still think this
patch should go in, since those locks were previously there for a long
time, and I can easily crash multipathd by repeadly sending it sighups
without this patch.
Does that sound reasonable?
Sorry but in my opinion the patch at the start of this thread makes the
approach for signal handling in multipathd more complex and harder to
maintain than strictly necessary. Will e.g. the next person who modifies
this code be aware of all these subtleties ? The approach I follow in
multithreaded code that I maintain myself for handling signals is to let
each signal handler write some data into a pipe, wait on that pipe in
the main thread and let the main thread take the appropriate action.
This approach is easy to maintain, does not require to block and unblock
signals at runtime, is portable between Unix systems and POSIX
compliant. See e.g.
http://github.com/bvanassche/srptools/commit/b6589892206ca628dbbb7fd8a1d613bf4f442ee6
for an example. Note: a possible alternative is to use signalfd()
instead of creating a pipe.
Bart.
--
dm-devel mailing list
dm-devel@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/dm-devel