block: make barrier completion more robust
authorTejun Heo <tj@kernel.org>
Fri, 28 Nov 2008 04:32:05 +0000 (13:32 +0900)
committerJens Axboe <jens.axboe@oracle.com>
Mon, 29 Dec 2008 07:28:45 +0000 (08:28 +0100)
Barrier completion had the following assumptions.

* start_ordered() couldn't finish the whole sequence properly.  If all
  actions are to be skipped, q->ordseq is set correctly but the actual
  completion was never triggered thus hanging the barrier request.

* Drain completion in elv_complete_request() assumed that there's
  always at least one request in the queue when drain completes.

Both assumptions are true but these assumptions need to be removed to
improve empty barrier implementation.  This patch makes the following
changes.

* Make start_ordered() use blk_ordered_complete_seq() to mark skipped
  steps complete and notify __elv_next_request() that it should fetch
  the next request if the whole barrier has completed inside
  start_ordered().

* Make drain completion path in elv_complete_request() check whether
  the queue is empty.  Empty queue also indicates drain completion.

* While at it, convert 0/1 return from blk_do_ordered() to false/true.

Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
block/blk-barrier.c
block/elevator.c
include/linux/blkdev.h

index 1efabf829c538ae5af7613ef0b0d19d06c111b76..b03d88013e1ef6a2cacf0b9e8317813fcfcbbb27 100644 (file)
@@ -88,7 +88,7 @@ unsigned blk_ordered_req_seq(struct request *rq)
                return QUEUE_ORDSEQ_DONE;
 }
 
-void blk_ordered_complete_seq(struct request_queue *q, unsigned seq, int error)
+bool blk_ordered_complete_seq(struct request_queue *q, unsigned seq, int error)
 {
        struct request *rq;
 
@@ -99,7 +99,7 @@ void blk_ordered_complete_seq(struct request_queue *q, unsigned seq, int error)
        q->ordseq |= seq;
 
        if (blk_ordered_cur_seq(q) != QUEUE_ORDSEQ_DONE)
-               return;
+               return false;
 
        /*
         * Okay, sequence complete.
@@ -109,6 +109,8 @@ void blk_ordered_complete_seq(struct request_queue *q, unsigned seq, int error)
 
        if (__blk_end_request(rq, q->orderr, blk_rq_bytes(rq)))
                BUG();
+
+       return true;
 }
 
 static void pre_flush_end_io(struct request *rq, int error)
@@ -151,9 +153,11 @@ static void queue_flush(struct request_queue *q, unsigned which)
        elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
 }
 
-static inline struct request *start_ordered(struct request_queue *q,
-                                           struct request *rq)
+static inline bool start_ordered(struct request_queue *q, struct request **rqp)
 {
+       struct request *rq = *rqp;
+       unsigned skip = 0;
+
        q->orderr = 0;
        q->ordered = q->next_ordered;
        q->ordseq |= QUEUE_ORDSEQ_STARTED;
@@ -177,7 +181,7 @@ static inline struct request *start_ordered(struct request_queue *q,
                queue_flush(q, QUEUE_ORDERED_DO_POSTFLUSH);
                rq = &q->post_flush_rq;
        } else
-               q->ordseq |= QUEUE_ORDSEQ_POSTFLUSH;
+               skip |= QUEUE_ORDSEQ_POSTFLUSH;
 
        if (q->ordered & QUEUE_ORDERED_DO_BAR) {
                rq = &q->bar_rq;
@@ -193,35 +197,40 @@ static inline struct request *start_ordered(struct request_queue *q,
 
                elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
        } else
-               q->ordseq |= QUEUE_ORDSEQ_BAR;
+               skip |= QUEUE_ORDSEQ_BAR;
 
        if (q->ordered & QUEUE_ORDERED_DO_PREFLUSH) {
                queue_flush(q, QUEUE_ORDERED_DO_PREFLUSH);
                rq = &q->pre_flush_rq;
        } else
-               q->ordseq |= QUEUE_ORDSEQ_PREFLUSH;
+               skip |= QUEUE_ORDSEQ_PREFLUSH;
 
        if ((q->ordered & QUEUE_ORDERED_BY_DRAIN) && q->in_flight)
                rq = NULL;
        else
-               q->ordseq |= QUEUE_ORDSEQ_DRAIN;
+               skip |= QUEUE_ORDSEQ_DRAIN;
+
+       *rqp = rq;
 
-       return rq;
+       /*
+        * Complete skipped sequences.  If whole sequence is complete,
+        * return false to tell elevator that this request is gone.
+        */
+       return !blk_ordered_complete_seq(q, skip, 0);
 }
 
-int blk_do_ordered(struct request_queue *q, struct request **rqp)
+bool blk_do_ordered(struct request_queue *q, struct request **rqp)
 {
        struct request *rq = *rqp;
        const int is_barrier = blk_fs_request(rq) && blk_barrier_rq(rq);
 
        if (!q->ordseq) {
                if (!is_barrier)
-                       return 1;
+                       return true;
 
-               if (q->next_ordered != QUEUE_ORDERED_NONE) {
-                       *rqp = start_ordered(q, rq);
-                       return 1;
-               } else {
+               if (q->next_ordered != QUEUE_ORDERED_NONE)
+                       return start_ordered(q, rqp);
+               else {
                        /*
                         * Queue ordering not supported.  Terminate
                         * with prejudice.
@@ -231,7 +240,7 @@ int blk_do_ordered(struct request_queue *q, struct request **rqp)
                                              blk_rq_bytes(rq)))
                                BUG();
                        *rqp = NULL;
-                       return 0;
+                       return false;
                }
        }
 
@@ -242,7 +251,7 @@ int blk_do_ordered(struct request_queue *q, struct request **rqp)
        /* Special requests are not subject to ordering rules. */
        if (!blk_fs_request(rq) &&
            rq != &q->pre_flush_rq && rq != &q->post_flush_rq)
-               return 1;
+               return true;
 
        if (q->ordered & QUEUE_ORDERED_BY_TAG) {
                /* Ordered by tag.  Blocking the next barrier is enough. */
@@ -255,7 +264,7 @@ int blk_do_ordered(struct request_queue *q, struct request **rqp)
                        *rqp = NULL;
        }
 
-       return 1;
+       return true;
 }
 
 static void bio_end_empty_barrier(struct bio *bio, int err)
index 86836dd179c05476078719869a0481777e999bc1..261ffaaf47bdbd125e3c91e44a4e9f7b3ed59fc0 100644 (file)
@@ -944,10 +944,14 @@ void elv_completed_request(struct request_queue *q, struct request *rq)
         * drained for flush sequence.
         */
        if (unlikely(q->ordseq)) {
-               struct request *first_rq = list_entry_rq(q->queue_head.next);
-               if (q->in_flight == 0 &&
+               struct request *next = NULL;
+
+               if (!list_empty(&q->queue_head))
+                       next = list_entry_rq(q->queue_head.next);
+
+               if (!q->in_flight &&
                    blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
-                   blk_ordered_req_seq(first_rq) > QUEUE_ORDSEQ_DRAIN) {
+                   (!next || blk_ordered_req_seq(next) > QUEUE_ORDSEQ_DRAIN)) {
                        blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
                        blk_start_queueing(q);
                }
index b044267009ed6d83eab91fe5a7e53fdcf1dcea56..3c7078e0129d2bab84198627de19164126e08b5d 100644 (file)
@@ -866,10 +866,10 @@ extern void blk_queue_rq_timed_out(struct request_queue *, rq_timed_out_fn *);
 extern void blk_queue_rq_timeout(struct request_queue *, unsigned int);
 extern struct backing_dev_info *blk_get_backing_dev_info(struct block_device *bdev);
 extern int blk_queue_ordered(struct request_queue *, unsigned, prepare_flush_fn *);
-extern int blk_do_ordered(struct request_queue *, struct request **);
+extern bool blk_do_ordered(struct request_queue *, struct request **);
 extern unsigned blk_ordered_cur_seq(struct request_queue *);
 extern unsigned blk_ordered_req_seq(struct request *);
-extern void blk_ordered_complete_seq(struct request_queue *, unsigned, int);
+extern bool blk_ordered_complete_seq(struct request_queue *, unsigned, int);
 
 extern int blk_rq_map_sg(struct request_queue *, struct request *, struct scatterlist *);
 extern void blk_dump_rq_flags(struct request *, char *);