[openib-general] [PATCH]deadlock problem in ipoib

[Roland Dreier]

    Shirley> This patch has fixed a deadlock problem: the caller calls
    Shirley> ipoib_put_ah() while holding priv->lock. (In
    Shirley> ipoib_free_ah() the same lock is reacquired.) This also
    Shirley> fixes the ipoib_flush_paths() calls __patch_free()
    Shirley> without holding priv->lock.

You definitely found some bugs but I think the real fix is to make
sure that ipoib_put_ah() is called -without- any locks held.  That's
because for future devices, ib_destroy_ah() may actually sleep
(although it won't for mthca).

The __path_free in unicast_arp_send() is actually safe to call inside
of the lock because we know that we just created the path struct, and
therefore it won't have an attached ah, so ipoib_put_ah() can never be
called.  (Actually it looks like there's a bug in path_rec_create()
also -- it doesn't set path->ah to NULL)

I think the proper fix for those places calling ipoib_put_ah() inside
a lock is to move the ipoib_ah structs to a temporary list and then
put them outside of the lock.

Thanks,
  Roland