LU-14713 llite: mend the trunc_sem_up_write()
The original lli_trunc_sem replace change (commit
e5914a61ac) fixed a
lock scenario:
t1 (page fault) t2 (dio read) t3 (truncate)
|- vm_mmap_pgoff() |- vvp_io_read_start() |- vvp_io_setattr
|- down_write(mmap_sem) |- down_read(trunc_sem) _start()
|- do_map() |- ll_direct_IO_impl()
|- vvp_io_fault_start |- ll_get_user_pages()
|- down_write(
|- down_read(mmap_sem) trunc_sem)
|- down_read(trunc_sem)
t1 waits for read semaphore of trunc_sem which is hindered by t3,
since t3 is waiting for the write semaphore while t2 take its read
semaphore,and t2 is waiting for mmap_sem which has been taken by t1,
and a deadlock ensues.
commit
e5914a61ac changes the down_read(trunc_sem) to
trunc_sem_down_read_nowait() in page fault path, to make it ignore
that there is a down_write(trunc_sem) waiting, just takes the read
semaphore if no writer has taken the semaphore, and breaks the
deadlock.
But there is a delicacy in using wake_up_var(), wake_up_var()->
__wake_up_bit()->waitqueue_active() locklessly test for waiters on the
queue, and if it's called without explicit smp_mb() it's possible for
the waitqueue_active() to ge hoisted before the condition store such
that we'll observe an empty wait list and the waiter might not
observe the condition, and the waiter won't get woke up whereafter.
Fixes:
e5914a61ac ("LU-12460 llite: replace lli_trunc_sem")
Change-Id: Ifdda2c1c8a4171466be1723923c136e84de8ce0e
Signed-off-by: Bobi Jam <bobijam@whamcloud.com>
Reviewed-on: https://review.whamcloud.com/43844
Reviewed-by: Andreas Dilger <adilger@whamcloud.com>
Reviewed-by: Neil Brown <neilb@suse.de>
Reviewed-by: Patrick Farrell <pfarrell@whamcloud.com>
Tested-by: Maloo <maloo@whamcloud.com>
Tested-by: Andreas Dilger <adilger@whamcloud.com>
Reviewed-by: Oleg Drokin <green@whamcloud.com>