[BLOCK] reimplement handling of barrier request
authorTejun Heo <htejun@gmail.com>
Fri, 6 Jan 2006 08:51:03 +0000 (09:51 +0100)
committerJens Axboe <axboe@suse.de>
Fri, 6 Jan 2006 08:51:03 +0000 (09:51 +0100)
Reimplement handling of barrier requests.

* Flexible handling to deal with various capabilities of
  target devices.
* Retry support for falling back.
* Tagged queues which don't support ordered tag can do ordered.

Signed-off-by: Tejun Heo <htejun@gmail.com>
Signed-off-by: Jens Axboe <axboe@suse.de>
block/elevator.c
block/ll_rw_blk.c
include/linux/blkdev.h
include/linux/elevator.h

index 85a11ce..39dcccc 100644 (file)
@@ -304,15 +304,7 @@ void elv_requeue_request(request_queue_t *q, struct request *rq)
 
        rq->flags &= ~REQ_STARTED;
 
-       /*
-        * if this is the flush, requeue the original instead and drop the flush
-        */
-       if (rq->flags & REQ_BAR_FLUSH) {
-               clear_bit(QUEUE_FLAG_FLUSH, &q->queue_flags);
-               rq = rq->end_io_data;
-       }
-
-       __elv_add_request(q, rq, ELEVATOR_INSERT_FRONT, 0);
+       __elv_add_request(q, rq, ELEVATOR_INSERT_REQUEUE, 0);
 }
 
 static void elv_drain_elevator(request_queue_t *q)
@@ -332,8 +324,19 @@ static void elv_drain_elevator(request_queue_t *q)
 void __elv_add_request(request_queue_t *q, struct request *rq, int where,
                       int plug)
 {
+       struct list_head *pos;
+       unsigned ordseq;
+
+       if (q->ordcolor)
+               rq->flags |= REQ_ORDERED_COLOR;
+
        if (rq->flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
                /*
+                * toggle ordered color
+                */
+               q->ordcolor ^= 1;
+
+               /*
                 * barriers implicitly indicate back insertion
                 */
                if (where == ELEVATOR_INSERT_SORT)
@@ -393,6 +396,30 @@ void __elv_add_request(request_queue_t *q, struct request *rq, int where,
                q->elevator->ops->elevator_add_req_fn(q, rq);
                break;
 
+       case ELEVATOR_INSERT_REQUEUE:
+               /*
+                * If ordered flush isn't in progress, we do front
+                * insertion; otherwise, requests should be requeued
+                * in ordseq order.
+                */
+               rq->flags |= REQ_SOFTBARRIER;
+
+               if (q->ordseq == 0) {
+                       list_add(&rq->queuelist, &q->queue_head);
+                       break;
+               }
+
+               ordseq = blk_ordered_req_seq(rq);
+
+               list_for_each(pos, &q->queue_head) {
+                       struct request *pos_rq = list_entry_rq(pos);
+                       if (ordseq <= blk_ordered_req_seq(pos_rq))
+                               break;
+               }
+
+               list_add_tail(&rq->queuelist, pos);
+               break;
+
        default:
                printk(KERN_ERR "%s: bad insertion point %d\n",
                       __FUNCTION__, where);
@@ -422,25 +449,16 @@ static inline struct request *__elv_next_request(request_queue_t *q)
 {
        struct request *rq;
 
-       if (unlikely(list_empty(&q->queue_head) &&
-                    !q->elevator->ops->elevator_dispatch_fn(q, 0)))
-               return NULL;
-
-       rq = list_entry_rq(q->queue_head.next);
-
-       /*
-        * if this is a barrier write and the device has to issue a
-        * flush sequence to support it, check how far we are
-        */
-       if (blk_fs_request(rq) && blk_barrier_rq(rq)) {
-               BUG_ON(q->ordered == QUEUE_ORDERED_NONE);
+       while (1) {
+               while (!list_empty(&q->queue_head)) {
+                       rq = list_entry_rq(q->queue_head.next);
+                       if (blk_do_ordered(q, &rq))
+                               return rq;
+               }
 
-               if (q->ordered == QUEUE_ORDERED_FLUSH &&
-                   !blk_barrier_preflush(rq))
-                       rq = blk_start_pre_flush(q, rq);
+               if (!q->elevator->ops->elevator_dispatch_fn(q, 0))
+                       return NULL;
        }
-
-       return rq;
 }
 
 struct request *elv_next_request(request_queue_t *q)
@@ -593,7 +611,21 @@ void elv_completed_request(request_queue_t *q, struct request *rq)
         * request is released from the driver, io must be done
         */
        if (blk_account_rq(rq)) {
+               struct request *first_rq = list_entry_rq(q->queue_head.next);
+
                q->in_flight--;
+
+               /*
+                * Check if the queue is waiting for fs requests to be
+                * drained for flush sequence.
+                */
+               if (q->ordseq && q->in_flight == 0 &&
+                   blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
+                   blk_ordered_req_seq(first_rq) > QUEUE_ORDSEQ_DRAIN) {
+                       blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
+                       q->request_fn(q);
+               }
+
                if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
                        e->ops->elevator_completed_req_fn(q, rq);
        }
index 65c4efc..91d3b48 100644 (file)
@@ -290,8 +290,8 @@ static inline void rq_init(request_queue_t *q, struct request *rq)
 
 /**
  * blk_queue_ordered - does this queue support ordered writes
- * @q:     the request queue
- * @flag:  see below
+ * @q:        the request queue
+ * @ordered:  one of QUEUE_ORDERED_*
  *
  * Description:
  *   For journalled file systems, doing ordered writes on a commit
@@ -300,28 +300,30 @@ static inline void rq_init(request_queue_t *q, struct request *rq)
  *   feature should call this function and indicate so.
  *
  **/
-void blk_queue_ordered(request_queue_t *q, int flag)
-{
-       switch (flag) {
-               case QUEUE_ORDERED_NONE:
-                       if (q->flush_rq)
-                               kmem_cache_free(request_cachep, q->flush_rq);
-                       q->flush_rq = NULL;
-                       q->ordered = flag;
-                       break;
-               case QUEUE_ORDERED_TAG:
-                       q->ordered = flag;
-                       break;
-               case QUEUE_ORDERED_FLUSH:
-                       q->ordered = flag;
-                       if (!q->flush_rq)
-                               q->flush_rq = kmem_cache_alloc(request_cachep,
-                                                               GFP_KERNEL);
-                       break;
-               default:
-                       printk("blk_queue_ordered: bad value %d\n", flag);
-                       break;
+int blk_queue_ordered(request_queue_t *q, unsigned ordered,
+                     prepare_flush_fn *prepare_flush_fn)
+{
+       if (ordered & (QUEUE_ORDERED_PREFLUSH | QUEUE_ORDERED_POSTFLUSH) &&
+           prepare_flush_fn == NULL) {
+               printk(KERN_ERR "blk_queue_ordered: prepare_flush_fn required\n");
+               return -EINVAL;
+       }
+
+       if (ordered != QUEUE_ORDERED_NONE &&
+           ordered != QUEUE_ORDERED_DRAIN &&
+           ordered != QUEUE_ORDERED_DRAIN_FLUSH &&
+           ordered != QUEUE_ORDERED_DRAIN_FUA &&
+           ordered != QUEUE_ORDERED_TAG &&
+           ordered != QUEUE_ORDERED_TAG_FLUSH &&
+           ordered != QUEUE_ORDERED_TAG_FUA) {
+               printk(KERN_ERR "blk_queue_ordered: bad value %d\n", ordered);
+               return -EINVAL;
        }
+
+       q->next_ordered = ordered;
+       q->prepare_flush_fn = prepare_flush_fn;
+
+       return 0;
 }
 
 EXPORT_SYMBOL(blk_queue_ordered);
@@ -346,167 +348,265 @@ EXPORT_SYMBOL(blk_queue_issue_flush_fn);
 /*
  * Cache flushing for ordered writes handling
  */
-static void blk_pre_flush_end_io(struct request *flush_rq, int error)
+inline unsigned blk_ordered_cur_seq(request_queue_t *q)
 {
-       struct request *rq = flush_rq->end_io_data;
-       request_queue_t *q = rq->q;
-
-       elv_completed_request(q, flush_rq);
-
-       rq->flags |= REQ_BAR_PREFLUSH;
-
-       if (!flush_rq->errors)
-               elv_requeue_request(q, rq);
-       else {
-               q->end_flush_fn(q, flush_rq);
-               clear_bit(QUEUE_FLAG_FLUSH, &q->queue_flags);
-               q->request_fn(q);
-       }
+       if (!q->ordseq)
+               return 0;
+       return 1 << ffz(q->ordseq);
 }
 
-static void blk_post_flush_end_io(struct request *flush_rq, int error)
+unsigned blk_ordered_req_seq(struct request *rq)
 {
-       struct request *rq = flush_rq->end_io_data;
        request_queue_t *q = rq->q;
 
-       elv_completed_request(q, flush_rq);
+       BUG_ON(q->ordseq == 0);
 
-       rq->flags |= REQ_BAR_POSTFLUSH;
+       if (rq == &q->pre_flush_rq)
+               return QUEUE_ORDSEQ_PREFLUSH;
+       if (rq == &q->bar_rq)
+               return QUEUE_ORDSEQ_BAR;
+       if (rq == &q->post_flush_rq)
+               return QUEUE_ORDSEQ_POSTFLUSH;
 
-       q->end_flush_fn(q, flush_rq);
-       clear_bit(QUEUE_FLAG_FLUSH, &q->queue_flags);
-       q->request_fn(q);
+       if ((rq->flags & REQ_ORDERED_COLOR) ==
+           (q->orig_bar_rq->flags & REQ_ORDERED_COLOR))
+               return QUEUE_ORDSEQ_DRAIN;
+       else
+               return QUEUE_ORDSEQ_DONE;
 }
 
-struct request *blk_start_pre_flush(request_queue_t *q, struct request *rq)
+void blk_ordered_complete_seq(request_queue_t *q, unsigned seq, int error)
 {
-       struct request *flush_rq = q->flush_rq;
-
-       BUG_ON(!blk_barrier_rq(rq));
+       struct request *rq;
+       int uptodate;
 
-       if (test_and_set_bit(QUEUE_FLAG_FLUSH, &q->queue_flags))
-               return NULL;
+       if (error && !q->orderr)
+               q->orderr = error;
 
-       rq_init(q, flush_rq);
-       flush_rq->elevator_private = NULL;
-       flush_rq->flags = REQ_BAR_FLUSH;
-       flush_rq->rq_disk = rq->rq_disk;
-       flush_rq->rl = NULL;
+       BUG_ON(q->ordseq & seq);
+       q->ordseq |= seq;
 
-       /*
-        * prepare_flush returns 0 if no flush is needed, just mark both
-        * pre and post flush as done in that case
-        */
-       if (!q->prepare_flush_fn(q, flush_rq)) {
-               rq->flags |= REQ_BAR_PREFLUSH | REQ_BAR_POSTFLUSH;
-               clear_bit(QUEUE_FLAG_FLUSH, &q->queue_flags);
-               return rq;
-       }
+       if (blk_ordered_cur_seq(q) != QUEUE_ORDSEQ_DONE)
+               return;
 
        /*
-        * some drivers dequeue requests right away, some only after io
-        * completion. make sure the request is dequeued.
+        * Okay, sequence complete.
         */
-       if (!list_empty(&rq->queuelist))
-               blkdev_dequeue_request(rq);
+       rq = q->orig_bar_rq;
+       uptodate = q->orderr ? q->orderr : 1;
 
-       flush_rq->end_io_data = rq;
-       flush_rq->end_io = blk_pre_flush_end_io;
+       q->ordseq = 0;
 
-       __elv_add_request(q, flush_rq, ELEVATOR_INSERT_FRONT, 0);
-       return flush_rq;
+       end_that_request_first(rq, uptodate, rq->hard_nr_sectors);
+       end_that_request_last(rq, uptodate);
 }
 
-static void blk_start_post_flush(request_queue_t *q, struct request *rq)
+static void pre_flush_end_io(struct request *rq, int error)
 {
-       struct request *flush_rq = q->flush_rq;
+       elv_completed_request(rq->q, rq);
+       blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_PREFLUSH, error);
+}
 
-       BUG_ON(!blk_barrier_rq(rq));
+static void bar_end_io(struct request *rq, int error)
+{
+       elv_completed_request(rq->q, rq);
+       blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_BAR, error);
+}
 
-       rq_init(q, flush_rq);
-       flush_rq->elevator_private = NULL;
-       flush_rq->flags = REQ_BAR_FLUSH;
-       flush_rq->rq_disk = rq->rq_disk;
-       flush_rq->rl = NULL;
+static void post_flush_end_io(struct request *rq, int error)
+{
+       elv_completed_request(rq->q, rq);
+       blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_POSTFLUSH, error);
+}
 
-       if (q->prepare_flush_fn(q, flush_rq)) {
-               flush_rq->end_io_data = rq;
-               flush_rq->end_io = blk_post_flush_end_io;
+static void queue_flush(request_queue_t *q, unsigned which)
+{
+       struct request *rq;
+       rq_end_io_fn *end_io;
 
-               __elv_add_request(q, flush_rq, ELEVATOR_INSERT_FRONT, 0);
-               q->request_fn(q);
+       if (which == QUEUE_ORDERED_PREFLUSH) {
+               rq = &q->pre_flush_rq;
+               end_io = pre_flush_end_io;
+       } else {
+               rq = &q->post_flush_rq;
+               end_io = post_flush_end_io;
        }
+
+       rq_init(q, rq);
+       rq->flags = REQ_HARDBARRIER;
+       rq->elevator_private = NULL;
+       rq->rq_disk = q->bar_rq.rq_disk;
+       rq->rl = NULL;
+       rq->end_io = end_io;
+       q->prepare_flush_fn(q, rq);
+
+       __elv_add_request(q, rq, ELEVATOR_INSERT_FRONT, 0);
 }
 
-static inline int blk_check_end_barrier(request_queue_t *q, struct request *rq,
-                                       int sectors)
+static inline struct request *start_ordered(request_queue_t *q,
+                                           struct request *rq)
 {
-       if (sectors > rq->nr_sectors)
-               sectors = rq->nr_sectors;
+       q->bi_size = 0;
+       q->orderr = 0;
+       q->ordered = q->next_ordered;
+       q->ordseq |= QUEUE_ORDSEQ_STARTED;
+
+       /*
+        * Prep proxy barrier request.
+        */
+       blkdev_dequeue_request(rq);
+       q->orig_bar_rq = rq;
+       rq = &q->bar_rq;
+       rq_init(q, rq);
+       rq->flags = bio_data_dir(q->orig_bar_rq->bio);
+       rq->flags |= q->ordered & QUEUE_ORDERED_FUA ? REQ_FUA : 0;
+       rq->elevator_private = NULL;
+       rq->rl = NULL;
+       init_request_from_bio(rq, q->orig_bar_rq->bio);
+       rq->end_io = bar_end_io;
+
+       /*
+        * Queue ordered sequence.  As we stack them at the head, we
+        * need to queue in reverse order.  Note that we rely on that
+        * no fs request uses ELEVATOR_INSERT_FRONT and thus no fs
+        * request gets inbetween ordered sequence.
+        */
+       if (q->ordered & QUEUE_ORDERED_POSTFLUSH)
+               queue_flush(q, QUEUE_ORDERED_POSTFLUSH);
+       else
+               q->ordseq |= QUEUE_ORDSEQ_POSTFLUSH;
+
+       __elv_add_request(q, rq, ELEVATOR_INSERT_FRONT, 0);
+
+       if (q->ordered & QUEUE_ORDERED_PREFLUSH) {
+               queue_flush(q, QUEUE_ORDERED_PREFLUSH);
+               rq = &q->pre_flush_rq;
+       } else
+               q->ordseq |= QUEUE_ORDSEQ_PREFLUSH;
 
-       rq->nr_sectors -= sectors;
-       return rq->nr_sectors;
+       if ((q->ordered & QUEUE_ORDERED_TAG) || q->in_flight == 0)
+               q->ordseq |= QUEUE_ORDSEQ_DRAIN;
+       else
+               rq = NULL;
+
+       return rq;
 }
 
-static int __blk_complete_barrier_rq(request_queue_t *q, struct request *rq,
-                                    int sectors, int queue_locked)
+int blk_do_ordered(request_queue_t *q, struct request **rqp)
 {
-       if (q->ordered != QUEUE_ORDERED_FLUSH)
-               return 0;
-       if (!blk_fs_request(rq) || !blk_barrier_rq(rq))
-               return 0;
-       if (blk_barrier_postflush(rq))
-               return 0;
+       struct request *rq = *rqp, *allowed_rq;
+       int is_barrier = blk_fs_request(rq) && blk_barrier_rq(rq);
 
-       if (!blk_check_end_barrier(q, rq, sectors)) {
-               unsigned long flags = 0;
+       if (!q->ordseq) {
+               if (!is_barrier)
+                       return 1;
 
-               if (!queue_locked)
-                       spin_lock_irqsave(q->queue_lock, flags);
+               if (q->next_ordered != QUEUE_ORDERED_NONE) {
+                       *rqp = start_ordered(q, rq);
+                       return 1;
+               } else {
+                       /*
+                        * This can happen when the queue switches to
+                        * ORDERED_NONE while this request is on it.
+                        */
+                       blkdev_dequeue_request(rq);
+                       end_that_request_first(rq, -EOPNOTSUPP,
+                                              rq->hard_nr_sectors);
+                       end_that_request_last(rq, -EOPNOTSUPP);
+                       *rqp = NULL;
+                       return 0;
+               }
+       }
 
-               blk_start_post_flush(q, rq);
+       if (q->ordered & QUEUE_ORDERED_TAG) {
+               if (is_barrier && rq != &q->bar_rq)
+                       *rqp = NULL;
+               return 1;
+       }
 
-               if (!queue_locked)
-                       spin_unlock_irqrestore(q->queue_lock, flags);
+       switch (blk_ordered_cur_seq(q)) {
+       case QUEUE_ORDSEQ_PREFLUSH:
+               allowed_rq = &q->pre_flush_rq;
+               break;
+       case QUEUE_ORDSEQ_BAR:
+               allowed_rq = &q->bar_rq;
+               break;
+       case QUEUE_ORDSEQ_POSTFLUSH:
+               allowed_rq = &q->post_flush_rq;
+               break;
+       default:
+               allowed_rq = NULL;
+               break;
        }
 
+       if (rq != allowed_rq &&
+           (blk_fs_request(rq) || rq == &q->pre_flush_rq ||
+            rq == &q->post_flush_rq))
+               *rqp = NULL;
+
        return 1;
 }
 
-/**
- * blk_complete_barrier_rq - complete possible barrier request
- * @q:  the request queue for the device
- * @rq:  the request
- * @sectors:  number of sectors to complete
- *
- * Description:
- *   Used in driver end_io handling to determine whether to postpone
- *   completion of a barrier request until a post flush has been done. This
- *   is the unlocked variant, used if the caller doesn't already hold the
- *   queue lock.
- **/
-int blk_complete_barrier_rq(request_queue_t *q, struct request *rq, int sectors)
+static int flush_dry_bio_endio(struct bio *bio, unsigned int bytes, int error)
 {
-       return __blk_complete_barrier_rq(q, rq, sectors, 0);
+       request_queue_t *q = bio->bi_private;
+       struct bio_vec *bvec;
+       int i;
+
+       /*
+        * This is dry run, restore bio_sector and size.  We'll finish
+        * this request again with the original bi_end_io after an
+        * error occurs or post flush is complete.
+        */
+       q->bi_size += bytes;
+
+       if (bio->bi_size)
+               return 1;
+
+       /* Rewind bvec's */
+       bio->bi_idx = 0;
+       bio_for_each_segment(bvec, bio, i) {
+               bvec->bv_len += bvec->bv_offset;
+               bvec->bv_offset = 0;
+       }
+
+       /* Reset bio */
+       set_bit(BIO_UPTODATE, &bio->bi_flags);
+       bio->bi_size = q->bi_size;
+       bio->bi_sector -= (q->bi_size >> 9);
+       q->bi_size = 0;
+
+       return 0;
 }
-EXPORT_SYMBOL(blk_complete_barrier_rq);
 
-/**
- * blk_complete_barrier_rq_locked - complete possible barrier request
- * @q:  the request queue for the device
- * @rq:  the request
- * @sectors:  number of sectors to complete
- *
- * Description:
- *   See blk_complete_barrier_rq(). This variant must be used if the caller
- *   holds the queue lock.
- **/
-int blk_complete_barrier_rq_locked(request_queue_t *q, struct request *rq,
-                                  int sectors)
+static inline int ordered_bio_endio(struct request *rq, struct bio *bio,
+                                   unsigned int nbytes, int error)
 {
-       return __blk_complete_barrier_rq(q, rq, sectors, 1);
+       request_queue_t *q = rq->q;
+       bio_end_io_t *endio;
+       void *private;
+
+       if (&q->bar_rq != rq)
+               return 0;
+
+       /*
+        * Okay, this is the barrier request in progress, dry finish it.
+        */
+       if (error && !q->orderr)
+               q->orderr = error;
+
+       endio = bio->bi_end_io;
+       private = bio->bi_private;
+       bio->bi_end_io = flush_dry_bio_endio;
+       bio->bi_private = q;
+
+       bio_endio(bio, nbytes, error);
+
+       bio->bi_end_io = endio;
+       bio->bi_private = private;
+
+       return 1;
 }
-EXPORT_SYMBOL(blk_complete_barrier_rq_locked);
 
 /**
  * blk_queue_bounce_limit - set bounce buffer limit for queue
@@ -1047,6 +1147,7 @@ static const char * const rq_flags[] = {
        "REQ_SORTED",
        "REQ_SOFTBARRIER",
        "REQ_HARDBARRIER",
+       "REQ_FUA",
        "REQ_CMD",
        "REQ_NOMERGE",
        "REQ_STARTED",
@@ -1066,6 +1167,7 @@ static const char * const rq_flags[] = {
        "REQ_PM_SUSPEND",
        "REQ_PM_RESUME",
        "REQ_PM_SHUTDOWN",
+       "REQ_ORDERED_COLOR",
 };
 
 void blk_dump_rq_flags(struct request *rq, char *msg)
@@ -1643,8 +1745,6 @@ void blk_cleanup_queue(request_queue_t * q)
        if (q->queue_tags)
                __blk_queue_free_tags(q);
 
-       blk_queue_ordered(q, QUEUE_ORDERED_NONE);
-
        kmem_cache_free(requestq_cachep, q);
 }
 
@@ -2714,7 +2814,7 @@ static int __make_request(request_queue_t *q, struct bio *bio)
        spin_lock_prefetch(q->queue_lock);
 
        barrier = bio_barrier(bio);
-       if (unlikely(barrier) && (q->ordered == QUEUE_ORDERED_NONE)) {
+       if (unlikely(barrier) && (q->next_ordered == QUEUE_ORDERED_NONE)) {
                err = -EOPNOTSUPP;
                goto end_io;
        }
@@ -3075,7 +3175,8 @@ static int __end_that_request_first(struct request *req, int uptodate,
                if (nr_bytes >= bio->bi_size) {
                        req->bio = bio->bi_next;
                        nbytes = bio->bi_size;
-                       bio_endio(bio, nbytes, error);
+                       if (!ordered_bio_endio(req, bio, nbytes, error))
+                               bio_endio(bio, nbytes, error);
                        next_idx = 0;
                        bio_nbytes = 0;
                } else {
@@ -3130,7 +3231,8 @@ static int __end_that_request_first(struct request *req, int uptodate,
         * if the request wasn't completed, update state
         */
        if (bio_nbytes) {
-               bio_endio(bio, bio_nbytes, error);
+               if (!ordered_bio_endio(req, bio, bio_nbytes, error))
+                       bio_endio(bio, bio_nbytes, error);
                bio->bi_idx += next_idx;
                bio_iovec(bio)->bv_offset += nr_bytes;
                bio_iovec(bio)->bv_len -= nr_bytes;
index a0ce8c5..15db0f1 100644 (file)
@@ -207,6 +207,7 @@ enum rq_flag_bits {
        __REQ_SORTED,           /* elevator knows about this request */
        __REQ_SOFTBARRIER,      /* may not be passed by ioscheduler */
        __REQ_HARDBARRIER,      /* may not be passed by drive either */
+       __REQ_FUA,              /* forced unit access */
        __REQ_CMD,              /* is a regular fs rw request */
        __REQ_NOMERGE,          /* don't touch this for merging */
        __REQ_STARTED,          /* drive already may have started this one */
@@ -230,9 +231,7 @@ enum rq_flag_bits {
        __REQ_PM_SUSPEND,       /* suspend request */
        __REQ_PM_RESUME,        /* resume request */
        __REQ_PM_SHUTDOWN,      /* shutdown request */
-       __REQ_BAR_PREFLUSH,     /* barrier pre-flush done */
-       __REQ_BAR_POSTFLUSH,    /* barrier post-flush */
-       __REQ_BAR_FLUSH,        /* rq is the flush request */
+       __REQ_ORDERED_COLOR,    /* is before or after barrier */
        __REQ_NR_BITS,          /* stops here */
 };
 
@@ -241,6 +240,7 @@ enum rq_flag_bits {
 #define REQ_SORTED     (1 << __REQ_SORTED)
 #define REQ_SOFTBARRIER        (1 << __REQ_SOFTBARRIER)
 #define REQ_HARDBARRIER        (1 << __REQ_HARDBARRIER)
+#define REQ_FUA                (1 << __REQ_FUA)
 #define REQ_CMD                (1 << __REQ_CMD)
 #define REQ_NOMERGE    (1 << __REQ_NOMERGE)
 #define REQ_STARTED    (1 << __REQ_STARTED)
@@ -260,9 +260,7 @@ enum rq_flag_bits {
 #define REQ_PM_SUSPEND (1 << __REQ_PM_SUSPEND)
 #define REQ_PM_RESUME  (1 << __REQ_PM_RESUME)
 #define REQ_PM_SHUTDOWN        (1 << __REQ_PM_SHUTDOWN)
-#define REQ_BAR_PREFLUSH       (1 << __REQ_BAR_PREFLUSH)
-#define REQ_BAR_POSTFLUSH      (1 << __REQ_BAR_POSTFLUSH)
-#define REQ_BAR_FLUSH  (1 << __REQ_BAR_FLUSH)
+#define REQ_ORDERED_COLOR      (1 << __REQ_ORDERED_COLOR)
 
 /*
  * State information carried for REQ_PM_SUSPEND and REQ_PM_RESUME
@@ -292,8 +290,7 @@ struct bio_vec;
 typedef int (merge_bvec_fn) (request_queue_t *, struct bio *, struct bio_vec *);
 typedef void (activity_fn) (void *data, int rw);
 typedef int (issue_flush_fn) (request_queue_t *, struct gendisk *, sector_t *);
-typedef int (prepare_flush_fn) (request_queue_t *, struct request *);
-typedef void (end_flush_fn) (request_queue_t *, struct request *);
+typedef void (prepare_flush_fn) (request_queue_t *, struct request *);
 
 enum blk_queue_state {
        Queue_down,
@@ -335,7 +332,6 @@ struct request_queue
        activity_fn             *activity_fn;
        issue_flush_fn          *issue_flush_fn;
        prepare_flush_fn        *prepare_flush_fn;
-       end_flush_fn            *end_flush_fn;
 
        /*
         * Dispatch queue sorting
@@ -420,14 +416,11 @@ struct request_queue
        /*
         * reserved for flush operations
         */
-       struct request          *flush_rq;
-       unsigned char           ordered;
-};
-
-enum {
-       QUEUE_ORDERED_NONE,
-       QUEUE_ORDERED_TAG,
-       QUEUE_ORDERED_FLUSH,
+       unsigned int            ordered, next_ordered, ordseq;
+       int                     orderr, ordcolor;
+       struct request          pre_flush_rq, bar_rq, post_flush_rq;
+       struct request          *orig_bar_rq;
+       unsigned int            bi_size;
 };
 
 #define RQ_INACTIVE            (-1)
@@ -445,12 +438,51 @@ enum {
 #define QUEUE_FLAG_REENTER     6       /* Re-entrancy avoidance */
 #define QUEUE_FLAG_PLUGGED     7       /* queue is plugged */
 #define QUEUE_FLAG_ELVSWITCH   8       /* don't use elevator, just do FIFO */
-#define QUEUE_FLAG_FLUSH       9       /* doing barrier flush sequence */
+
+enum {
+       /*
+        * Hardbarrier is supported with one of the following methods.
+        *
+        * NONE         : hardbarrier unsupported
+        * DRAIN        : ordering by draining is enough
+        * DRAIN_FLUSH  : ordering by draining w/ pre and post flushes
+        * DRAIN_FUA    : ordering by draining w/ pre flush and FUA write
+        * TAG          : ordering by tag is enough
+        * TAG_FLUSH    : ordering by tag w/ pre and post flushes
+        * TAG_FUA      : ordering by tag w/ pre flush and FUA write
+        */
+       QUEUE_ORDERED_NONE      = 0x00,
+       QUEUE_ORDERED_DRAIN     = 0x01,
+       QUEUE_ORDERED_TAG       = 0x02,
+
+       QUEUE_ORDERED_PREFLUSH  = 0x10,
+       QUEUE_ORDERED_POSTFLUSH = 0x20,
+       QUEUE_ORDERED_FUA       = 0x40,
+
+       QUEUE_ORDERED_DRAIN_FLUSH = QUEUE_ORDERED_DRAIN |
+                       QUEUE_ORDERED_PREFLUSH | QUEUE_ORDERED_POSTFLUSH,
+       QUEUE_ORDERED_DRAIN_FUA = QUEUE_ORDERED_DRAIN |
+                       QUEUE_ORDERED_PREFLUSH | QUEUE_ORDERED_FUA,
+       QUEUE_ORDERED_TAG_FLUSH = QUEUE_ORDERED_TAG |
+                       QUEUE_ORDERED_PREFLUSH | QUEUE_ORDERED_POSTFLUSH,
+       QUEUE_ORDERED_TAG_FUA   = QUEUE_ORDERED_TAG |
+                       QUEUE_ORDERED_PREFLUSH | QUEUE_ORDERED_FUA,
+
+       /*
+        * Ordered operation sequence
+        */
+       QUEUE_ORDSEQ_STARTED    = 0x01, /* flushing in progress */
+       QUEUE_ORDSEQ_DRAIN      = 0x02, /* waiting for the queue to be drained */
+       QUEUE_ORDSEQ_PREFLUSH   = 0x04, /* pre-flushing in progress */
+       QUEUE_ORDSEQ_BAR        = 0x08, /* original barrier req in progress */
+       QUEUE_ORDSEQ_POSTFLUSH  = 0x10, /* post-flushing in progress */
+       QUEUE_ORDSEQ_DONE       = 0x20,
+};
 
 #define blk_queue_plugged(q)   test_bit(QUEUE_FLAG_PLUGGED, &(q)->queue_flags)
 #define blk_queue_tagged(q)    test_bit(QUEUE_FLAG_QUEUED, &(q)->queue_flags)
 #define blk_queue_stopped(q)   test_bit(QUEUE_FLAG_STOPPED, &(q)->queue_flags)
-#define blk_queue_flushing(q)  test_bit(QUEUE_FLAG_FLUSH, &(q)->queue_flags)
+#define blk_queue_flushing(q)  ((q)->ordseq)
 
 #define blk_fs_request(rq)     ((rq)->flags & REQ_CMD)
 #define blk_pc_request(rq)     ((rq)->flags & REQ_BLOCK_PC)
@@ -466,8 +498,7 @@ enum {
 
 #define blk_sorted_rq(rq)      ((rq)->flags & REQ_SORTED)
 #define blk_barrier_rq(rq)     ((rq)->flags & REQ_HARDBARRIER)
-#define blk_barrier_preflush(rq)       ((rq)->flags & REQ_BAR_PREFLUSH)
-#define blk_barrier_postflush(rq)      ((rq)->flags & REQ_BAR_POSTFLUSH)
+#define blk_fua_rq(rq)         ((rq)->flags & REQ_FUA)
 
 #define list_entry_rq(ptr)     list_entry((ptr), struct request, queuelist)
 
@@ -665,11 +696,12 @@ extern void blk_queue_prep_rq(request_queue_t *, prep_rq_fn *pfn);
 extern void blk_queue_merge_bvec(request_queue_t *, merge_bvec_fn *);
 extern void blk_queue_dma_alignment(request_queue_t *, int);
 extern struct backing_dev_info *blk_get_backing_dev_info(struct block_device *bdev);
-extern void blk_queue_ordered(request_queue_t *, int);
+extern int blk_queue_ordered(request_queue_t *, unsigned, prepare_flush_fn *);
 extern void blk_queue_issue_flush_fn(request_queue_t *, issue_flush_fn *);
-extern struct request *blk_start_pre_flush(request_queue_t *,struct request *);
-extern int blk_complete_barrier_rq(request_queue_t *, struct request *, int);
-extern int blk_complete_barrier_rq_locked(request_queue_t *, struct request *, int);
+extern int blk_do_ordered(request_queue_t *, struct request **);
+extern unsigned blk_ordered_cur_seq(request_queue_t *);
+extern unsigned blk_ordered_req_seq(struct request *);
+extern void blk_ordered_complete_seq(request_queue_t *, unsigned, int);
 
 extern int blk_rq_map_sg(request_queue_t *, struct request *, struct scatterlist *);
 extern void blk_dump_rq_flags(struct request *, char *);
index a74c27e..fb80fa4 100644 (file)
@@ -130,6 +130,7 @@ extern int elv_try_last_merge(request_queue_t *, struct bio *);
 #define ELEVATOR_INSERT_FRONT  1
 #define ELEVATOR_INSERT_BACK   2
 #define ELEVATOR_INSERT_SORT   3
+#define ELEVATOR_INSERT_REQUEUE        4
 
 /*
  * return values from elevator_may_queue_fn