repair: fix an ABBA deadlock in inode prefetching

The inode prefetching code has a fixed limit of inodes that might are
submitted at a time.  Unfortunately the buffers for them get locked
once the prefetching starts.  That way the threads processing the inode
might get stuck on buffer locked, but not submitted for reading yet.

Fix this by kicking the queue as soon as we would have to wait on the
ra_count semaphore.

Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reported-by: Arkadiusz Miśkiewicz <arekm@maven.pl>
Tested-by: Arkadiusz Miśkiewicz <arekm@maven.pl>
Signed-off-by: Christoph Hellwig <hch@lst.de>

diff --git a/repair/prefetch.c b/repair/prefetch.c
index 0e1167ec63dbeec9dbd03e62e4007223507a41f9..c0b10022566b3cb3584e115caa32643dd830d635 100644 (file)
--- a/repair/prefetch.c
+++ b/repair/prefetch.c
@@ -641,7 +641,18 @@ pf_queuing_worker(
                pftrace("queuing irec %p in AG %d, sem count = %d",
                        irec, args->agno, i);
 #endif
-               sem_wait(&args->ra_count);
+               err = sem_trywait(&args->ra_count);
+               if (err == EAGAIN) {
+                       /*
+                        * Kick the queue once we have reached the limit;
+                        * without this the threads processing the inodes
+                        * might get stuck on a buffer that has been locked
+                        * and added to the I/O queue but is waiting for
+                        * the thread to be woken.
+                        */
+                       pf_start_io_workers(args);
+                       sem_wait(&args->ra_count);
+               }
 
                num_inos = 0;
                bno = XFS_AGINO_TO_AGBNO(mp, cur_irec->ino_startnum);