shrink buf changed behavior

[ofirvin]

Hi,
We have been going over a few recent commits in libntirpc.
and we saw 1352b71
and it kind of looks like chunk_unref_locked can do a double unlock. we are not very familiar with that layer, but do you think that can happen? and if so is that ok?

[dang]

I'm confused, there appear to be no unlocks in chunk_unref_locked(). Did you mean a different function?

[ofirvin]

chunk_unref_locked calls get_lru_chunk_with_lock. hich sometimes leaves ioq unlocked.

and then do_shrink can unlock ioq again if !is_same_buflist(io_buf, new_io_buf, rdma_xprt)

[dang]

I think it's safe. get_lru_chunk_with_lock() only unlocks ioqh->qmutex if pthread_mutex_trylock() succeeded, which means that ioq was not already locked, and we succeeded in taking the lock. The correct thing to do then is unlock it. If trylock() fails (we didn't get the lock, it was already locked, potentially by us), then we skip to the next buffer.