md/r5cache: fix io_unit handling in r5l_log_endio()
authorSong Liu <songliubraving@fb.com>
Thu, 3 Aug 2017 17:03:17 +0000 (10:03 -0700)
committerShaohua Li <shli@fb.com>
Tue, 8 Aug 2017 14:42:37 +0000 (07:42 -0700)
In r5l_log_endio(), once log->io_list_lock is released, the io unit
may be accessed (or even freed) by other threads. Current code
doesn't handle the io_unit properly, which leads to potential race
conditions.

This patch solves this race condition by:

1. Add a pending_stripe count flush_payload. Multiple flush_payloads
   are counted as only one pending_stripe. Flag has_flush_payload is
   added to show whether the io unit has flush_payload;
2. In r5l_log_endio(), check flags has_null_flush and
   has_flush_payload with log->io_list_lock held. After the lock
   is released, this IO unit is only accessed when we know the
   pending_stripe counter cannot be zeroed by other threads.

Signed-off-by: Song Liu <songliubraving@fb.com>
Signed-off-by: Shaohua Li <shli@fb.com>
drivers/md/raid5-cache.c

index ce98414a6e34e62d049dd84b34fb1b74fcb02003..2dcbafa8e66ca7f5418fce8434fe9674977f0832 100644 (file)
@@ -236,9 +236,10 @@ struct r5l_io_unit {
        bool need_split_bio;
        struct bio *split_bio;
 
-       unsigned int has_flush:1;      /* include flush request */
-       unsigned int has_fua:1;        /* include fua request */
-       unsigned int has_null_flush:1; /* include empty flush request */
+       unsigned int has_flush:1;               /* include flush request */
+       unsigned int has_fua:1;                 /* include fua request */
+       unsigned int has_null_flush:1;          /* include null flush request */
+       unsigned int has_flush_payload:1;       /* include flush payload  */
        /*
         * io isn't sent yet, flush/fua request can only be submitted till it's
         * the first IO in running_ios list
@@ -571,6 +572,8 @@ static void r5l_log_endio(struct bio *bio)
        struct r5l_io_unit *io_deferred;
        struct r5l_log *log = io->log;
        unsigned long flags;
+       bool has_null_flush;
+       bool has_flush_payload;
 
        if (bio->bi_status)
                md_error(log->rdev->mddev, log->rdev);
@@ -580,6 +583,16 @@ static void r5l_log_endio(struct bio *bio)
 
        spin_lock_irqsave(&log->io_list_lock, flags);
        __r5l_set_io_unit_state(io, IO_UNIT_IO_END);
+
+       /*
+        * if the io doesn't not have null_flush or flush payload,
+        * it is not safe to access it after releasing io_list_lock.
+        * Therefore, it is necessary to check the condition with
+        * the lock held.
+        */
+       has_null_flush = io->has_null_flush;
+       has_flush_payload = io->has_flush_payload;
+
        if (log->need_cache_flush && !list_empty(&io->stripe_list))
                r5l_move_to_end_ios(log);
        else
@@ -600,19 +613,23 @@ static void r5l_log_endio(struct bio *bio)
        if (log->need_cache_flush)
                md_wakeup_thread(log->rdev->mddev->thread);
 
-       if (io->has_null_flush) {
+       /* finish flush only io_unit and PAYLOAD_FLUSH only io_unit */
+       if (has_null_flush) {
                struct bio *bi;
 
                WARN_ON(bio_list_empty(&io->flush_barriers));
                while ((bi = bio_list_pop(&io->flush_barriers)) != NULL) {
                        bio_endio(bi);
-                       atomic_dec(&io->pending_stripe);
+                       if (atomic_dec_and_test(&io->pending_stripe)) {
+                               __r5l_stripe_write_finished(io);
+                               return;
+                       }
                }
        }
-
-       /* finish flush only io_unit and PAYLOAD_FLUSH only io_unit */
-       if (atomic_read(&io->pending_stripe) == 0)
-               __r5l_stripe_write_finished(io);
+       /* decrease pending_stripe for flush payload */
+       if (has_flush_payload)
+               if (atomic_dec_and_test(&io->pending_stripe))
+                       __r5l_stripe_write_finished(io);
 }
 
 static void r5l_do_submit_io(struct r5l_log *log, struct r5l_io_unit *io)
@@ -881,6 +898,11 @@ static void r5l_append_flush_payload(struct r5l_log *log, sector_t sect)
        payload->size = cpu_to_le32(sizeof(__le64));
        payload->flush_stripes[0] = cpu_to_le64(sect);
        io->meta_offset += meta_size;
+       /* multiple flush payloads count as one pending_stripe */
+       if (!io->has_flush_payload) {
+               io->has_flush_payload = 1;
+               atomic_inc(&io->pending_stripe);
+       }
        mutex_unlock(&log->io_mutex);
 }