CFQ: add think time check for service tree
[linux-2.6.git] / block / blk-flush.c
index 62b7df9..bb21e4c 100644 (file)
@@ -1,6 +1,69 @@
 /*
  * Functions to sequence FLUSH and FUA writes.
+ *
+ * Copyright (C) 2011          Max Planck Institute for Gravitational Physics
+ * Copyright (C) 2011          Tejun Heo <tj@kernel.org>
+ *
+ * This file is released under the GPLv2.
+ *
+ * REQ_{FLUSH|FUA} requests are decomposed to sequences consisted of three
+ * optional steps - PREFLUSH, DATA and POSTFLUSH - according to the request
+ * properties and hardware capability.
+ *
+ * If a request doesn't have data, only REQ_FLUSH makes sense, which
+ * indicates a simple flush request.  If there is data, REQ_FLUSH indicates
+ * that the device cache should be flushed before the data is executed, and
+ * REQ_FUA means that the data must be on non-volatile media on request
+ * completion.
+ *
+ * If the device doesn't have writeback cache, FLUSH and FUA don't make any
+ * difference.  The requests are either completed immediately if there's no
+ * data or executed as normal requests otherwise.
+ *
+ * If the device has writeback cache and supports FUA, REQ_FLUSH is
+ * translated to PREFLUSH but REQ_FUA is passed down directly with DATA.
+ *
+ * If the device has writeback cache and doesn't support FUA, REQ_FLUSH is
+ * translated to PREFLUSH and REQ_FUA to POSTFLUSH.
+ *
+ * The actual execution of flush is double buffered.  Whenever a request
+ * needs to execute PRE or POSTFLUSH, it queues at
+ * q->flush_queue[q->flush_pending_idx].  Once certain criteria are met, a
+ * flush is issued and the pending_idx is toggled.  When the flush
+ * completes, all the requests which were pending are proceeded to the next
+ * step.  This allows arbitrary merging of different types of FLUSH/FUA
+ * requests.
+ *
+ * Currently, the following conditions are used to determine when to issue
+ * flush.
+ *
+ * C1. At any given time, only one flush shall be in progress.  This makes
+ *     double buffering sufficient.
+ *
+ * C2. Flush is deferred if any request is executing DATA of its sequence.
+ *     This avoids issuing separate POSTFLUSHes for requests which shared
+ *     PREFLUSH.
+ *
+ * C3. The second condition is ignored if there is a request which has
+ *     waited longer than FLUSH_PENDING_TIMEOUT.  This is to avoid
+ *     starvation in the unlikely case where there are continuous stream of
+ *     FUA (without FLUSH) requests.
+ *
+ * For devices which support FUA, it isn't clear whether C2 (and thus C3)
+ * is beneficial.
+ *
+ * Note that a sequenced FLUSH/FUA request with DATA is completed twice.
+ * Once while executing DATA and again after the whole sequence is
+ * complete.  The first completion updates the contained bio but doesn't
+ * finish it so that the bio submitter is notified only after the whole
+ * sequence is complete.  This is implemented by testing REQ_FLUSH_SEQ in
+ * req_bio_endio().
+ *
+ * The above peculiarity requires that each FLUSH/FUA request has only one
+ * bio attached to it, which is guaranteed as they aren't allowed to be
+ * merged in the usual way.
  */
+
 #include <linux/kernel.h>
 #include <linux/module.h>
 #include <linux/bio.h>
 
 /* FLUSH/FUA sequences */
 enum {
-       QUEUE_FSEQ_STARTED      = (1 << 0), /* flushing in progress */
-       QUEUE_FSEQ_PREFLUSH     = (1 << 1), /* pre-flushing in progress */
-       QUEUE_FSEQ_DATA         = (1 << 2), /* data write in progress */
-       QUEUE_FSEQ_POSTFLUSH    = (1 << 3), /* post-flushing in progress */
-       QUEUE_FSEQ_DONE         = (1 << 4),
+       REQ_FSEQ_PREFLUSH       = (1 << 0), /* pre-flushing in progress */
+       REQ_FSEQ_DATA           = (1 << 1), /* data write in progress */
+       REQ_FSEQ_POSTFLUSH      = (1 << 2), /* post-flushing in progress */
+       REQ_FSEQ_DONE           = (1 << 3),
+
+       REQ_FSEQ_ACTIONS        = REQ_FSEQ_PREFLUSH | REQ_FSEQ_DATA |
+                                 REQ_FSEQ_POSTFLUSH,
+
+       /*
+        * If flush has been pending longer than the following timeout,
+        * it's issued even if flush_data requests are still in flight.
+        */
+       FLUSH_PENDING_TIMEOUT   = 5 * HZ,
 };
 
-static struct request *queue_next_fseq(struct request_queue *q);
+static bool blk_kick_flush(struct request_queue *q);
 
-unsigned blk_flush_cur_seq(struct request_queue *q)
+static unsigned int blk_flush_policy(unsigned int fflags, struct request *rq)
 {
-       if (!q->flush_seq)
-               return 0;
-       return 1 << ffz(q->flush_seq);
+       unsigned int policy = 0;
+
+       if (fflags & REQ_FLUSH) {
+               if (rq->cmd_flags & REQ_FLUSH)
+                       policy |= REQ_FSEQ_PREFLUSH;
+               if (blk_rq_sectors(rq))
+                       policy |= REQ_FSEQ_DATA;
+               if (!(fflags & REQ_FUA) && (rq->cmd_flags & REQ_FUA))
+                       policy |= REQ_FSEQ_POSTFLUSH;
+       }
+       return policy;
 }
 
-static struct request *blk_flush_complete_seq(struct request_queue *q,
-                                             unsigned seq, int error)
+static unsigned int blk_flush_cur_seq(struct request *rq)
 {
-       struct request *next_rq = NULL;
-
-       if (error && !q->flush_err)
-               q->flush_err = error;
-
-       BUG_ON(q->flush_seq & seq);
-       q->flush_seq |= seq;
-
-       if (blk_flush_cur_seq(q) != QUEUE_FSEQ_DONE) {
-               /* not complete yet, queue the next flush sequence */
-               next_rq = queue_next_fseq(q);
-       } else {
-               /* complete this flush request */
-               __blk_end_request_all(q->orig_flush_rq, q->flush_err);
-               q->orig_flush_rq = NULL;
-               q->flush_seq = 0;
-
-               /* dispatch the next flush if there's one */
-               if (!list_empty(&q->pending_flushes)) {
-                       next_rq = list_entry_rq(q->pending_flushes.next);
-                       list_move(&next_rq->queuelist, &q->queue_head);
-               }
-       }
-       return next_rq;
+       return 1 << ffz(rq->flush.seq);
 }
 
-static void blk_flush_complete_seq_end_io(struct request_queue *q,
-                                         unsigned seq, int error)
+static void blk_flush_restore_request(struct request *rq)
 {
-       bool was_empty = elv_queue_empty(q);
-       struct request *next_rq;
-
-       next_rq = blk_flush_complete_seq(q, seq, error);
-
        /*
-        * Moving a request silently to empty queue_head may stall the
-        * queue.  Kick the queue in those cases.
+        * After flush data completion, @rq->bio is %NULL but we need to
+        * complete the bio again.  @rq->biotail is guaranteed to equal the
+        * original @rq->bio.  Restore it.
         */
-       if (was_empty && next_rq)
-               __blk_run_queue(q);
+       rq->bio = rq->biotail;
+
+       /* make @rq a normal request */
+       rq->cmd_flags &= ~REQ_FLUSH_SEQ;
+       rq->end_io = NULL;
 }
 
-static void pre_flush_end_io(struct request *rq, int error)
+/**
+ * blk_flush_complete_seq - complete flush sequence
+ * @rq: FLUSH/FUA request being sequenced
+ * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero)
+ * @error: whether an error occurred
+ *
+ * @rq just completed @seq part of its flush sequence, record the
+ * completion and trigger the next step.
+ *
+ * CONTEXT:
+ * spin_lock_irq(q->queue_lock)
+ *
+ * RETURNS:
+ * %true if requests were added to the dispatch queue, %false otherwise.
+ */
+static bool blk_flush_complete_seq(struct request *rq, unsigned int seq,
+                                  int error)
 {
-       elv_completed_request(rq->q, rq);
-       blk_flush_complete_seq_end_io(rq->q, QUEUE_FSEQ_PREFLUSH, error);
+       struct request_queue *q = rq->q;
+       struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
+       bool queued = false;
+
+       BUG_ON(rq->flush.seq & seq);
+       rq->flush.seq |= seq;
+
+       if (likely(!error))
+               seq = blk_flush_cur_seq(rq);
+       else
+               seq = REQ_FSEQ_DONE;
+
+       switch (seq) {
+       case REQ_FSEQ_PREFLUSH:
+       case REQ_FSEQ_POSTFLUSH:
+               /* queue for flush */
+               if (list_empty(pending))
+                       q->flush_pending_since = jiffies;
+               list_move_tail(&rq->flush.list, pending);
+               break;
+
+       case REQ_FSEQ_DATA:
+               list_move_tail(&rq->flush.list, &q->flush_data_in_flight);
+               list_add(&rq->queuelist, &q->queue_head);
+               queued = true;
+               break;
+
+       case REQ_FSEQ_DONE:
+               /*
+                * @rq was previously adjusted by blk_flush_issue() for
+                * flush sequencing and may already have gone through the
+                * flush data request completion path.  Restore @rq for
+                * normal completion and end it.
+                */
+               BUG_ON(!list_empty(&rq->queuelist));
+               list_del_init(&rq->flush.list);
+               blk_flush_restore_request(rq);
+               __blk_end_request_all(rq, error);
+               break;
+
+       default:
+               BUG();
+       }
+
+       return blk_kick_flush(q) | queued;
 }
 
-static void flush_data_end_io(struct request *rq, int error)
+static void flush_end_io(struct request *flush_rq, int error)
 {
-       elv_completed_request(rq->q, rq);
-       blk_flush_complete_seq_end_io(rq->q, QUEUE_FSEQ_DATA, error);
+       struct request_queue *q = flush_rq->q;
+       struct list_head *running = &q->flush_queue[q->flush_running_idx];
+       bool queued = false;
+       struct request *rq, *n;
+
+       BUG_ON(q->flush_pending_idx == q->flush_running_idx);
+
+       /* account completion of the flush request */
+       q->flush_running_idx ^= 1;
+       elv_completed_request(q, flush_rq);
+
+       /* and push the waiting requests to the next stage */
+       list_for_each_entry_safe(rq, n, running, flush.list) {
+               unsigned int seq = blk_flush_cur_seq(rq);
+
+               BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
+               queued |= blk_flush_complete_seq(rq, seq, error);
+       }
+
+       /*
+        * Kick the queue to avoid stall for two cases:
+        * 1. Moving a request silently to empty queue_head may stall the
+        * queue.
+        * 2. When flush request is running in non-queueable queue, the
+        * queue is hold. Restart the queue after flush request is finished
+        * to avoid stall.
+        * This function is called from request completion path and calling
+        * directly into request_fn may confuse the driver.  Always use
+        * kblockd.
+        */
+       if (queued || q->flush_queue_delayed)
+               blk_run_queue_async(q);
+       q->flush_queue_delayed = 0;
 }
 
-static void post_flush_end_io(struct request *rq, int error)
+/**
+ * blk_kick_flush - consider issuing flush request
+ * @q: request_queue being kicked
+ *
+ * Flush related states of @q have changed, consider issuing flush request.
+ * Please read the comment at the top of this file for more info.
+ *
+ * CONTEXT:
+ * spin_lock_irq(q->queue_lock)
+ *
+ * RETURNS:
+ * %true if flush was issued, %false otherwise.
+ */
+static bool blk_kick_flush(struct request_queue *q)
 {
-       elv_completed_request(rq->q, rq);
-       blk_flush_complete_seq_end_io(rq->q, QUEUE_FSEQ_POSTFLUSH, error);
+       struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
+       struct request *first_rq =
+               list_first_entry(pending, struct request, flush.list);
+
+       /* C1 described at the top of this file */
+       if (q->flush_pending_idx != q->flush_running_idx || list_empty(pending))
+               return false;
+
+       /* C2 and C3 */
+       if (!list_empty(&q->flush_data_in_flight) &&
+           time_before(jiffies,
+                       q->flush_pending_since + FLUSH_PENDING_TIMEOUT))
+               return false;
+
+       /*
+        * Issue flush and toggle pending_idx.  This makes pending_idx
+        * different from running_idx, which means flush is in flight.
+        */
+       blk_rq_init(q, &q->flush_rq);
+       q->flush_rq.cmd_type = REQ_TYPE_FS;
+       q->flush_rq.cmd_flags = WRITE_FLUSH | REQ_FLUSH_SEQ;
+       q->flush_rq.rq_disk = first_rq->rq_disk;
+       q->flush_rq.end_io = flush_end_io;
+
+       q->flush_pending_idx ^= 1;
+       list_add_tail(&q->flush_rq.queuelist, &q->queue_head);
+       return true;
 }
 
-static void init_flush_request(struct request *rq, struct gendisk *disk)
+static void flush_data_end_io(struct request *rq, int error)
 {
-       rq->cmd_type = REQ_TYPE_FS;
-       rq->cmd_flags = WRITE_FLUSH;
-       rq->rq_disk = disk;
+       struct request_queue *q = rq->q;
+
+       /*
+        * After populating an empty queue, kick it to avoid stall.  Read
+        * the comment in flush_end_io().
+        */
+       if (blk_flush_complete_seq(rq, REQ_FSEQ_DATA, error))
+               blk_run_queue_async(q);
 }
 
-static struct request *queue_next_fseq(struct request_queue *q)
+/**
+ * blk_insert_flush - insert a new FLUSH/FUA request
+ * @rq: request to insert
+ *
+ * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
+ * @rq is being submitted.  Analyze what needs to be done and put it on the
+ * right queue.
+ *
+ * CONTEXT:
+ * spin_lock_irq(q->queue_lock)
+ */
+void blk_insert_flush(struct request *rq)
 {
-       struct request *orig_rq = q->orig_flush_rq;
-       struct request *rq = &q->flush_rq;
+       struct request_queue *q = rq->q;
+       unsigned int fflags = q->flush_flags;   /* may change, cache */
+       unsigned int policy = blk_flush_policy(fflags, rq);
 
-       blk_rq_init(q, rq);
+       BUG_ON(rq->end_io);
+       BUG_ON(!rq->bio || rq->bio != rq->biotail);
 
-       switch (blk_flush_cur_seq(q)) {
-       case QUEUE_FSEQ_PREFLUSH:
-               init_flush_request(rq, orig_rq->rq_disk);
-               rq->end_io = pre_flush_end_io;
-               break;
-       case QUEUE_FSEQ_DATA:
-               init_request_from_bio(rq, orig_rq->bio);
-               /*
-                * orig_rq->rq_disk may be different from
-                * bio->bi_bdev->bd_disk if orig_rq got here through
-                * remapping drivers.  Make sure rq->rq_disk points
-                * to the same one as orig_rq.
-                */
-               rq->rq_disk = orig_rq->rq_disk;
-               rq->cmd_flags &= ~(REQ_FLUSH | REQ_FUA);
-               rq->cmd_flags |= orig_rq->cmd_flags & (REQ_FLUSH | REQ_FUA);
-               rq->end_io = flush_data_end_io;
-               break;
-       case QUEUE_FSEQ_POSTFLUSH:
-               init_flush_request(rq, orig_rq->rq_disk);
-               rq->end_io = post_flush_end_io;
-               break;
-       default:
-               BUG();
+       /*
+        * @policy now records what operations need to be done.  Adjust
+        * REQ_FLUSH and FUA for the driver.
+        */
+       rq->cmd_flags &= ~REQ_FLUSH;
+       if (!(fflags & REQ_FUA))
+               rq->cmd_flags &= ~REQ_FUA;
+
+       /*
+        * If there's data but flush is not necessary, the request can be
+        * processed directly without going through flush machinery.  Queue
+        * for normal execution.
+        */
+       if ((policy & REQ_FSEQ_DATA) &&
+           !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
+               list_add_tail(&rq->queuelist, &q->queue_head);
+               return;
        }
 
-       elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
-       return rq;
+       /*
+        * @rq should go through flush machinery.  Mark it part of flush
+        * sequence and submit for further processing.
+        */
+       memset(&rq->flush, 0, sizeof(rq->flush));
+       INIT_LIST_HEAD(&rq->flush.list);
+       rq->cmd_flags |= REQ_FLUSH_SEQ;
+       rq->end_io = flush_data_end_io;
+
+       blk_flush_complete_seq(rq, REQ_FSEQ_ACTIONS & ~policy, 0);
 }
 
-struct request *blk_do_flush(struct request_queue *q, struct request *rq)
+/**
+ * blk_abort_flushes - @q is being aborted, abort flush requests
+ * @q: request_queue being aborted
+ *
+ * To be called from elv_abort_queue().  @q is being aborted.  Prepare all
+ * FLUSH/FUA requests for abortion.
+ *
+ * CONTEXT:
+ * spin_lock_irq(q->queue_lock)
+ */
+void blk_abort_flushes(struct request_queue *q)
 {
-       unsigned int fflags = q->flush_flags; /* may change, cache it */
-       bool has_flush = fflags & REQ_FLUSH, has_fua = fflags & REQ_FUA;
-       bool do_preflush = has_flush && (rq->cmd_flags & REQ_FLUSH);
-       bool do_postflush = has_flush && !has_fua && (rq->cmd_flags & REQ_FUA);
-       unsigned skip = 0;
+       struct request *rq, *n;
+       int i;
 
        /*
-        * Special case.  If there's data but flush is not necessary,
-        * the request can be issued directly.
-        *
-        * Flush w/o data should be able to be issued directly too but
-        * currently some drivers assume that rq->bio contains
-        * non-zero data if it isn't NULL and empty FLUSH requests
-        * getting here usually have bio's without data.
+        * Requests in flight for data are already owned by the dispatch
+        * queue or the device driver.  Just restore for normal completion.
         */
-       if (blk_rq_sectors(rq) && !do_preflush && !do_postflush) {
-               rq->cmd_flags &= ~REQ_FLUSH;
-               if (!has_fua)
-                       rq->cmd_flags &= ~REQ_FUA;
-               return rq;
+       list_for_each_entry_safe(rq, n, &q->flush_data_in_flight, flush.list) {
+               list_del_init(&rq->flush.list);
+               blk_flush_restore_request(rq);
        }
 
        /*
-        * Sequenced flushes can't be processed in parallel.  If
-        * another one is already in progress, queue for later
-        * processing.
+        * We need to give away requests on flush queues.  Restore for
+        * normal completion and put them on the dispatch queue.
         */
-       if (q->flush_seq) {
-               list_move_tail(&rq->queuelist, &q->pending_flushes);
-               return NULL;
+       for (i = 0; i < ARRAY_SIZE(q->flush_queue); i++) {
+               list_for_each_entry_safe(rq, n, &q->flush_queue[i],
+                                        flush.list) {
+                       list_del_init(&rq->flush.list);
+                       blk_flush_restore_request(rq);
+                       list_add_tail(&rq->queuelist, &q->queue_head);
+               }
        }
-
-       /*
-        * Start a new flush sequence
-        */
-       q->flush_err = 0;
-       q->flush_seq |= QUEUE_FSEQ_STARTED;
-
-       /* adjust FLUSH/FUA of the original request and stash it away */
-       rq->cmd_flags &= ~REQ_FLUSH;
-       if (!has_fua)
-               rq->cmd_flags &= ~REQ_FUA;
-       blk_dequeue_request(rq);
-       q->orig_flush_rq = rq;
-
-       /* skip unneded sequences and return the first one */
-       if (!do_preflush)
-               skip |= QUEUE_FSEQ_PREFLUSH;
-       if (!blk_rq_sectors(rq))
-               skip |= QUEUE_FSEQ_DATA;
-       if (!do_postflush)
-               skip |= QUEUE_FSEQ_POSTFLUSH;
-       return blk_flush_complete_seq(q, skip, 0);
 }
 
 static void bio_end_flush(struct bio *bio, int err)
@@ -205,7 +386,6 @@ static void bio_end_flush(struct bio *bio, int err)
  * @bdev:      blockdev to issue flush for
  * @gfp_mask:  memory allocation flags (for bio_alloc)
  * @error_sector:      error sector
- * @flags:     BLKDEV_IFL_* flags to control behaviour
  *
  * Description:
  *    Issue a flush for the block device in question. Caller can supply
@@ -214,7 +394,7 @@ static void bio_end_flush(struct bio *bio, int err)
  *    request was pushed in some internal queue for later handling.
  */
 int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
-               sector_t *error_sector, unsigned long flags)
+               sector_t *error_sector)
 {
        DECLARE_COMPLETION_ONSTACK(wait);
        struct request_queue *q;
@@ -240,21 +420,19 @@ int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
        bio = bio_alloc(gfp_mask, 0);
        bio->bi_end_io = bio_end_flush;
        bio->bi_bdev = bdev;
-       if (test_bit(BLKDEV_WAIT, &flags))
-               bio->bi_private = &wait;
+       bio->bi_private = &wait;
 
        bio_get(bio);
        submit_bio(WRITE_FLUSH, bio);
-       if (test_bit(BLKDEV_WAIT, &flags)) {
-               wait_for_completion(&wait);
-               /*
-                * The driver must store the error location in ->bi_sector, if
-                * it supports it. For non-stacked drivers, this should be
-                * copied from blk_rq_pos(rq).
-                */
-               if (error_sector)
-                       *error_sector = bio->bi_sector;
-       }
+       wait_for_completion(&wait);
+
+       /*
+        * The driver must store the error location in ->bi_sector, if
+        * it supports it. For non-stacked drivers, this should be
+        * copied from blk_rq_pos(rq).
+        */
+       if (error_sector)
+               *error_sector = bio->bi_sector;
 
        if (!bio_flagged(bio, BIO_UPTODATE))
                ret = -EIO;