[PATCH] 01/05 Implement generic dispatch queue
authorTejun Heo <htejun@gmail.com>
Thu, 20 Oct 2005 14:23:44 +0000 (16:23 +0200)
committerJens Axboe <axboe@nelson.home.kernel.dk>
Fri, 28 Oct 2005 06:44:24 +0000 (08:44 +0200)
Implements generic dispatch queue which can replace all
dispatch queues implemented by each iosched.  This reduces
code duplication, eases enforcing semantics over dispatch
queue, and simplifies specific ioscheds.

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

index 4144f30d82a97aeb6de21d1a87ba463e20e18c1f..a27555908d359654a619209ff34dad174d018247 100644 (file)
 static DEFINE_SPINLOCK(elv_list_lock);
 static LIST_HEAD(elv_list);
 
+static inline sector_t rq_last_sector(struct request *rq)
+{
+       return rq->sector + rq->nr_sectors;
+}
+
 /*
  * can we safely merge with this request?
  */
@@ -143,6 +148,9 @@ static int elevator_attach(request_queue_t *q, struct elevator_type *e,
        INIT_LIST_HEAD(&q->queue_head);
        q->last_merge = NULL;
        q->elevator = eq;
+       q->last_sector = 0;
+       q->boundary_rq = NULL;
+       q->max_back_kb = 0;
 
        if (eq->ops->elevator_init_fn)
                ret = eq->ops->elevator_init_fn(q, eq);
@@ -225,6 +233,48 @@ void elevator_exit(elevator_t *e)
        kfree(e);
 }
 
+/*
+ * Insert rq into dispatch queue of q.  Queue lock must be held on
+ * entry.  If sort != 0, rq is sort-inserted; otherwise, rq will be
+ * appended to the dispatch queue.  To be used by specific elevators.
+ */
+void elv_dispatch_insert(request_queue_t *q, struct request *rq, int sort)
+{
+       sector_t boundary;
+       unsigned max_back;
+       struct list_head *entry;
+
+       if (!sort) {
+               /* Specific elevator is performing sort.  Step away. */
+               q->last_sector = rq_last_sector(rq);
+               q->boundary_rq = rq;
+               list_add_tail(&rq->queuelist, &q->queue_head);
+               return;
+       }
+
+       boundary = q->last_sector;
+       max_back = q->max_back_kb * 2;
+       boundary = boundary > max_back ? boundary - max_back : 0;
+
+       list_for_each_prev(entry, &q->queue_head) {
+               struct request *pos = list_entry_rq(entry);
+
+               if (pos->flags & (REQ_SOFTBARRIER|REQ_HARDBARRIER|REQ_STARTED))
+                       break;
+               if (rq->sector >= boundary) {
+                       if (pos->sector < boundary)
+                               continue;
+               } else {
+                       if (pos->sector >= boundary)
+                               break;
+               }
+               if (rq->sector >= pos->sector)
+                       break;
+       }
+
+       list_add(&rq->queuelist, entry);
+}
+
 int elv_merge(request_queue_t *q, struct request **req, struct bio *bio)
 {
        elevator_t *e = q->elevator;
@@ -255,13 +305,7 @@ void elv_merge_requests(request_queue_t *q, struct request *rq,
                e->ops->elevator_merge_req_fn(q, rq, next);
 }
 
-/*
- * For careful internal use by the block layer. Essentially the same as
- * a requeue in that it tells the io scheduler that this request is not
- * active in the driver or hardware anymore, but we don't want the request
- * added back to the scheduler. Function is not exported.
- */
-void elv_deactivate_request(request_queue_t *q, struct request *rq)
+void elv_requeue_request(request_queue_t *q, struct request *rq)
 {
        elevator_t *e = q->elevator;
 
@@ -269,19 +313,14 @@ void elv_deactivate_request(request_queue_t *q, struct request *rq)
         * it already went through dequeue, we need to decrement the
         * in_flight count again
         */
-       if (blk_account_rq(rq))
+       if (blk_account_rq(rq)) {
                q->in_flight--;
+               if (blk_sorted_rq(rq) && e->ops->elevator_deactivate_req_fn)
+                       e->ops->elevator_deactivate_req_fn(q, rq);
+       }
 
        rq->flags &= ~REQ_STARTED;
 
-       if (e->ops->elevator_deactivate_req_fn)
-               e->ops->elevator_deactivate_req_fn(q, rq);
-}
-
-void elv_requeue_request(request_queue_t *q, struct request *rq)
-{
-       elv_deactivate_request(q, rq);
-
        /*
         * if this is the flush, requeue the original instead and drop the flush
         */
@@ -290,55 +329,89 @@ void elv_requeue_request(request_queue_t *q, struct request *rq)
                rq = rq->end_io_data;
        }
 
-       /*
-        * the request is prepped and may have some resources allocated.
-        * allowing unprepped requests to pass this one may cause resource
-        * deadlock.  turn on softbarrier.
-        */
-       rq->flags |= REQ_SOFTBARRIER;
-
-       /*
-        * if iosched has an explicit requeue hook, then use that. otherwise
-        * just put the request at the front of the queue
-        */
-       if (q->elevator->ops->elevator_requeue_req_fn)
-               q->elevator->ops->elevator_requeue_req_fn(q, rq);
-       else
-               __elv_add_request(q, rq, ELEVATOR_INSERT_FRONT, 0);
+       __elv_add_request(q, rq, ELEVATOR_INSERT_FRONT, 0);
 }
 
 void __elv_add_request(request_queue_t *q, struct request *rq, int where,
                       int plug)
 {
-       /*
-        * barriers implicitly indicate back insertion
-        */
-       if (rq->flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER) &&
-           where == ELEVATOR_INSERT_SORT)
-               where = ELEVATOR_INSERT_BACK;
+       if (rq->flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
+               /*
+                * barriers implicitly indicate back insertion
+                */
+               if (where == ELEVATOR_INSERT_SORT)
+                       where = ELEVATOR_INSERT_BACK;
+
+               /*
+                * this request is scheduling boundary, update last_sector
+                */
+               if (blk_fs_request(rq)) {
+                       q->last_sector = rq_last_sector(rq);
+                       q->boundary_rq = rq;
+               }
+       }
 
        if (plug)
                blk_plug_device(q);
 
        rq->q = q;
 
-       if (!test_bit(QUEUE_FLAG_DRAIN, &q->queue_flags)) {
-               q->elevator->ops->elevator_add_req_fn(q, rq, where);
-
-               if (blk_queue_plugged(q)) {
-                       int nrq = q->rq.count[READ] + q->rq.count[WRITE]
-                                 - q->in_flight;
-
-                       if (nrq >= q->unplug_thresh)
-                               __generic_unplug_device(q);
-               }
-       } else
+       if (unlikely(test_bit(QUEUE_FLAG_DRAIN, &q->queue_flags))) {
                /*
                 * if drain is set, store the request "locally". when the drain
                 * is finished, the requests will be handed ordered to the io
                 * scheduler
                 */
                list_add_tail(&rq->queuelist, &q->drain_list);
+               return;
+       }
+
+       switch (where) {
+       case ELEVATOR_INSERT_FRONT:
+               rq->flags |= REQ_SOFTBARRIER;
+
+               list_add(&rq->queuelist, &q->queue_head);
+               break;
+
+       case ELEVATOR_INSERT_BACK:
+               rq->flags |= REQ_SOFTBARRIER;
+
+               while (q->elevator->ops->elevator_dispatch_fn(q, 1))
+                       ;
+               list_add_tail(&rq->queuelist, &q->queue_head);
+               /*
+                * We kick the queue here for the following reasons.
+                * - The elevator might have returned NULL previously
+                *   to delay requests and returned them now.  As the
+                *   queue wasn't empty before this request, ll_rw_blk
+                *   won't run the queue on return, resulting in hang.
+                * - Usually, back inserted requests won't be merged
+                *   with anything.  There's no point in delaying queue
+                *   processing.
+                */
+               blk_remove_plug(q);
+               q->request_fn(q);
+               break;
+
+       case ELEVATOR_INSERT_SORT:
+               BUG_ON(!blk_fs_request(rq));
+               rq->flags |= REQ_SORTED;
+               q->elevator->ops->elevator_add_req_fn(q, rq);
+               break;
+
+       default:
+               printk(KERN_ERR "%s: bad insertion point %d\n",
+                      __FUNCTION__, where);
+               BUG();
+       }
+
+       if (blk_queue_plugged(q)) {
+               int nrq = q->rq.count[READ] + q->rq.count[WRITE]
+                       - q->in_flight;
+
+               if (nrq >= q->unplug_thresh)
+                       __generic_unplug_device(q);
+       }
 }
 
 void elv_add_request(request_queue_t *q, struct request *rq, int where,
@@ -353,13 +426,19 @@ void elv_add_request(request_queue_t *q, struct request *rq, int where,
 
 static inline struct request *__elv_next_request(request_queue_t *q)
 {
-       struct request *rq = q->elevator->ops->elevator_next_req_fn(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 (rq && blk_fs_request(rq) && blk_barrier_rq(rq)) {
+       if (blk_fs_request(rq) && blk_barrier_rq(rq)) {
                BUG_ON(q->ordered == QUEUE_ORDERED_NONE);
 
                if (q->ordered == QUEUE_ORDERED_FLUSH &&
@@ -376,16 +455,34 @@ struct request *elv_next_request(request_queue_t *q)
        int ret;
 
        while ((rq = __elv_next_request(q)) != NULL) {
-               /*
-                * just mark as started even if we don't start it, a request
-                * that has been delayed should not be passed by new incoming
-                * requests
-                */
-               rq->flags |= REQ_STARTED;
+               if (!(rq->flags & REQ_STARTED)) {
+                       elevator_t *e = q->elevator;
+
+                       /*
+                        * This is the first time the device driver
+                        * sees this request (possibly after
+                        * requeueing).  Notify IO scheduler.
+                        */
+                       if (blk_sorted_rq(rq) &&
+                           e->ops->elevator_activate_req_fn)
+                               e->ops->elevator_activate_req_fn(q, rq);
+
+                       /*
+                        * just mark as started even if we don't start
+                        * it, a request that has been delayed should
+                        * not be passed by new incoming requests
+                        */
+                       rq->flags |= REQ_STARTED;
+               }
 
                if (rq == q->last_merge)
                        q->last_merge = NULL;
 
+               if (!q->boundary_rq || q->boundary_rq == rq) {
+                       q->last_sector = rq_last_sector(rq);
+                       q->boundary_rq = NULL;
+               }
+
                if ((rq->flags & REQ_DONTPREP) || !q->prep_rq_fn)
                        break;
 
@@ -396,9 +493,9 @@ struct request *elv_next_request(request_queue_t *q)
                        /*
                         * the request may have been (partially) prepped.
                         * we need to keep this request in the front to
-                        * avoid resource deadlock.  turn on softbarrier.
+                        * avoid resource deadlock.  REQ_STARTED will
+                        * prevent other fs requests from passing this one.
                         */
-                       rq->flags |= REQ_SOFTBARRIER;
                        rq = NULL;
                        break;
                } else if (ret == BLKPREP_KILL) {
@@ -421,16 +518,16 @@ struct request *elv_next_request(request_queue_t *q)
        return rq;
 }
 
-void elv_remove_request(request_queue_t *q, struct request *rq)
+void elv_dequeue_request(request_queue_t *q, struct request *rq)
 {
-       elevator_t *e = q->elevator;
+       BUG_ON(list_empty(&rq->queuelist));
+
+       list_del_init(&rq->queuelist);
 
        /*
         * the time frame between a request being removed from the lists
         * and to it is freed is accounted as io that is in progress at
-        * the driver side. note that we only account requests that the
-        * driver has seen (REQ_STARTED set), to avoid false accounting
-        * for request-request merges
+        * the driver side.
         */
        if (blk_account_rq(rq))
                q->in_flight++;
@@ -444,19 +541,19 @@ void elv_remove_request(request_queue_t *q, struct request *rq)
         */
        if (rq == q->last_merge)
                q->last_merge = NULL;
-
-       if (e->ops->elevator_remove_req_fn)
-               e->ops->elevator_remove_req_fn(q, rq);
 }
 
 int elv_queue_empty(request_queue_t *q)
 {
        elevator_t *e = q->elevator;
 
+       if (!list_empty(&q->queue_head))
+               return 0;
+
        if (e->ops->elevator_queue_empty_fn)
                return e->ops->elevator_queue_empty_fn(q);
 
-       return list_empty(&q->queue_head);
+       return 1;
 }
 
 struct request *elv_latter_request(request_queue_t *q, struct request *rq)
@@ -528,11 +625,11 @@ 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))
+       if (blk_account_rq(rq)) {
                q->in_flight--;
-
-       if (e->ops->elevator_completed_req_fn)
-               e->ops->elevator_completed_req_fn(q, rq);
+               if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
+                       e->ops->elevator_completed_req_fn(q, rq);
+       }
 }
 
 int elv_register_queue(struct request_queue *q)
@@ -705,11 +802,12 @@ ssize_t elv_iosched_show(request_queue_t *q, char *name)
        return len;
 }
 
+EXPORT_SYMBOL(elv_dispatch_insert);
 EXPORT_SYMBOL(elv_add_request);
 EXPORT_SYMBOL(__elv_add_request);
 EXPORT_SYMBOL(elv_requeue_request);
 EXPORT_SYMBOL(elv_next_request);
-EXPORT_SYMBOL(elv_remove_request);
+EXPORT_SYMBOL(elv_dequeue_request);
 EXPORT_SYMBOL(elv_queue_empty);
 EXPORT_SYMBOL(elv_completed_request);
 EXPORT_SYMBOL(elevator_exit);
index 4e2b1b06b411d12accfad7ae28ef00cc28dbe95b..d2a66fd309c3df6c323dbb00156450679e623227 100644 (file)
@@ -353,6 +353,8 @@ static void blk_pre_flush_end_io(struct request *flush_rq)
        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)
@@ -369,6 +371,8 @@ static void blk_post_flush_end_io(struct request *flush_rq)
        struct request *rq = flush_rq->end_io_data;
        request_queue_t *q = rq->q;
 
+       elv_completed_request(q, flush_rq);
+
        rq->flags |= REQ_BAR_POSTFLUSH;
 
        q->end_flush_fn(q, flush_rq);
@@ -408,8 +412,6 @@ struct request *blk_start_pre_flush(request_queue_t *q, struct request *rq)
        if (!list_empty(&rq->queuelist))
                blkdev_dequeue_request(rq);
 
-       elv_deactivate_request(q, rq);
-
        flush_rq->end_io_data = rq;
        flush_rq->end_io = blk_pre_flush_end_io;
 
@@ -1040,6 +1042,7 @@ EXPORT_SYMBOL(blk_queue_invalidate_tags);
 static char *rq_flags[] = {
        "REQ_RW",
        "REQ_FAILFAST",
+       "REQ_SORTED",
        "REQ_SOFTBARRIER",
        "REQ_HARDBARRIER",
        "REQ_CMD",
@@ -2456,6 +2459,8 @@ static void __blk_put_request(request_queue_t *q, struct request *req)
        if (unlikely(--req->ref_count))
                return;
 
+       elv_completed_request(q, req);
+
        req->rq_status = RQ_INACTIVE;
        req->rl = NULL;
 
@@ -2466,8 +2471,6 @@ static void __blk_put_request(request_queue_t *q, struct request *req)
        if (rl) {
                int rw = rq_data_dir(req);
 
-               elv_completed_request(q, req);
-
                BUG_ON(!list_empty(&req->queuelist));
 
                blk_free_request(q, req);
@@ -2477,14 +2480,14 @@ static void __blk_put_request(request_queue_t *q, struct request *req)
 
 void blk_put_request(struct request *req)
 {
+       unsigned long flags;
+       request_queue_t *q = req->q;
+
        /*
-        * if req->rl isn't set, this request didnt originate from the
-        * block layer, so it's safe to just disregard it
+        * Gee, IDE calls in w/ NULL q.  Fix IDE and remove the
+        * following if (q) test.
         */
-       if (req->rl) {
-               unsigned long flags;
-               request_queue_t *q = req->q;
-
+       if (q) {
                spin_lock_irqsave(q->queue_lock, flags);
                __blk_put_request(q, req);
                spin_unlock_irqrestore(q->queue_lock, flags);
index efdc9b5bc05c8687380200f8a871975c91709860..2c7b9154927ae849803f62587b5cd12b95b5153c 100644 (file)
@@ -203,6 +203,7 @@ struct request {
 enum rq_flag_bits {
        __REQ_RW,               /* not set, read. set, write */
        __REQ_FAILFAST,         /* no low level driver retries */
+       __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_CMD,              /* is a regular fs rw request */
@@ -235,6 +236,7 @@ enum rq_flag_bits {
 
 #define REQ_RW         (1 << __REQ_RW)
 #define REQ_FAILFAST   (1 << __REQ_FAILFAST)
+#define REQ_SORTED     (1 << __REQ_SORTED)
 #define REQ_SOFTBARRIER        (1 << __REQ_SOFTBARRIER)
 #define REQ_HARDBARRIER        (1 << __REQ_HARDBARRIER)
 #define REQ_CMD                (1 << __REQ_CMD)
@@ -332,6 +334,13 @@ struct request_queue
        prepare_flush_fn        *prepare_flush_fn;
        end_flush_fn            *end_flush_fn;
 
+       /*
+        * Dispatch queue sorting
+        */
+       sector_t                last_sector;
+       struct request          *boundary_rq;
+       unsigned int            max_back_kb;
+
        /*
         * Auto-unplugging state
         */
@@ -454,6 +463,7 @@ enum {
 #define blk_pm_request(rq)     \
        ((rq)->flags & (REQ_PM_SUSPEND | REQ_PM_RESUME))
 
+#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)
@@ -611,12 +621,7 @@ extern void end_request(struct request *req, int uptodate);
 
 static inline void blkdev_dequeue_request(struct request *req)
 {
-       BUG_ON(list_empty(&req->queuelist));
-
-       list_del_init(&req->queuelist);
-
-       if (req->rl)
-               elv_remove_request(req->q, req);
+       elv_dequeue_request(req->q, req);
 }
 
 /*
index ea6bbc2d7407e9fb48de7c67756e4463bba82970..76f4f6920744496ded49763acac725c88822a857 100644 (file)
@@ -8,18 +8,17 @@ typedef void (elevator_merge_req_fn) (request_queue_t *, struct request *, struc
 
 typedef void (elevator_merged_fn) (request_queue_t *, struct request *);
 
-typedef struct request *(elevator_next_req_fn) (request_queue_t *);
+typedef int (elevator_dispatch_fn) (request_queue_t *, int);
 
-typedef void (elevator_add_req_fn) (request_queue_t *, struct request *, int);
+typedef void (elevator_add_req_fn) (request_queue_t *, struct request *);
 typedef int (elevator_queue_empty_fn) (request_queue_t *);
-typedef void (elevator_remove_req_fn) (request_queue_t *, struct request *);
-typedef void (elevator_requeue_req_fn) (request_queue_t *, struct request *);
 typedef struct request *(elevator_request_list_fn) (request_queue_t *, struct request *);
 typedef void (elevator_completed_req_fn) (request_queue_t *, struct request *);
 typedef int (elevator_may_queue_fn) (request_queue_t *, int, struct bio *);
 
 typedef int (elevator_set_req_fn) (request_queue_t *, struct request *, struct bio *, int);
 typedef void (elevator_put_req_fn) (request_queue_t *, struct request *);
+typedef void (elevator_activate_req_fn) (request_queue_t *, struct request *);
 typedef void (elevator_deactivate_req_fn) (request_queue_t *, struct request *);
 
 typedef int (elevator_init_fn) (request_queue_t *, elevator_t *);
@@ -31,10 +30,9 @@ struct elevator_ops
        elevator_merged_fn *elevator_merged_fn;
        elevator_merge_req_fn *elevator_merge_req_fn;
 
-       elevator_next_req_fn *elevator_next_req_fn;
+       elevator_dispatch_fn *elevator_dispatch_fn;
        elevator_add_req_fn *elevator_add_req_fn;
-       elevator_remove_req_fn *elevator_remove_req_fn;
-       elevator_requeue_req_fn *elevator_requeue_req_fn;
+       elevator_activate_req_fn *elevator_activate_req_fn;
        elevator_deactivate_req_fn *elevator_deactivate_req_fn;
 
        elevator_queue_empty_fn *elevator_queue_empty_fn;
@@ -81,15 +79,15 @@ struct elevator_queue
 /*
  * block elevator interface
  */
+extern void elv_dispatch_insert(request_queue_t *, struct request *, int);
 extern void elv_add_request(request_queue_t *, struct request *, int, int);
 extern void __elv_add_request(request_queue_t *, struct request *, int, int);
 extern int elv_merge(request_queue_t *, struct request **, struct bio *);
 extern void elv_merge_requests(request_queue_t *, struct request *,
                               struct request *);
 extern void elv_merged_request(request_queue_t *, struct request *);
-extern void elv_remove_request(request_queue_t *, struct request *);
+extern void elv_dequeue_request(request_queue_t *, struct request *);
 extern void elv_requeue_request(request_queue_t *, struct request *);
-extern void elv_deactivate_request(request_queue_t *, struct request *);
 extern int elv_queue_empty(request_queue_t *);
 extern struct request *elv_next_request(struct request_queue *q);
 extern struct request *elv_former_request(request_queue_t *, struct request *);