On Sun, Jul 31, 2011 at 1:47 PM, Dave Chinner <david@xxxxxxxxxxxxx> wrote: > > Yes, I already have, a couple of hours before you sent this: > > http://www.spinics.net/lists/linux-fsdevel/msg47357.html > > We haven't found the root cause of the problem, and writeback cannot > hold off grab_super_passive() because writeback only holds read > locks on s_umount, just like grab_super_passive. With read-write semaphores, even read-vs-read recursion is a deadlock possibility. Why? Because if a writer comes in on another thread, while the read lock is initially held, then the writer will now block. And due to fairness, now a subsequent reader will *also* block. So no, nesting readers is *not* allowed for rw_semaphores even if naively you'd think it should work. So if xfstests 073 does mount/umount testing, then it is entirely possible that a reader blocks another reader due to a pending writer. NOTE! The rwlock *spinlocks* are designed to be unfair to writers, and by design allow recursive readers. That's important and very much by design: it is ok to take a rwlock for reading without disabling interrupts even if there may be *interrupts* that also need it for reading. With the spinning rwlock, there is also much less chance of starvation due to this unfairness. In contrast, the rw_semaphores really can be starved pretty easily if you are excessively unfair to writers. Linus -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
