CVS commit: src/sys/netipsec

["Ryota Ozaki" <ozaki-r%netbsd.org@localhost>]

Module Name:    src
Committed By:   ozaki-r
Date:           Wed Sep 27 07:27:29 UTC 2017

Modified Files:
        src/sys/netipsec: key.c

Log Message:
Fix deadlock between pserialize_perform and localcount_drain

A typical ussage of localcount_drain looks like this:

  mutex_enter(&mtx);
  item = remove_from_list();
  pserialize_perform(psz);
  localcount_drain(&item->localcount, &cv, &mtx);
  mutex_exit(&mtx);

This sequence can cause a deadlock which happens for example on the following
situation:

- Thread A calls localcount_drain which calls xc_broadcast after releasing
  a specified mutex
- Thread B enters the sequence and calls pserialize_perform with holding
  the mutex while pserialize_perform also calls xc_broadcast
- Thread C (xc_thread) that calls an xcall callback of localcount_drain tries
  to hold the mutex

xc_broadcast of thread B doesn't start until xc_broadcast of thread A
finishes, which is a feature of xcall(9). This means that pserialize_perform
never complete until xc_broadcast of thread A finishes. On the other hand,
thread C that is a callee of xc_broadcast of thread A sticks on the mutex.
Finally the threads block each other (A blocks B, B blocks C and C blocks A).

A possible fix is to serialize executions of the above sequence by another
mutex, but adding another mutex makes the code complex, so fix the deadlock
by another way; the fix is to release the mutex before pserialize_perform
and instead use a condvar to prevent pserialize_perform from being called
simultaneously.

Note that the deadlock has happened only if NET_MPSAFE is enabled.


To generate a diff of this commit:
cvs rdiff -u -r1.225 -r1.226 src/sys/netipsec/key.c

Please note that diffs are not public domain; they are subject to the
copyright notices on the relevant files.