blkio: Implementation of IOPS limit logic
Vivek Goyal [Wed, 15 Sep 2010 21:06:37 +0000 (17:06 -0400)]
o core logic of implementing IOPS throttling.

Signed-off-by: Vivek Goyal <vgoyal@redhat.com>
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>

block/blk-throttle.c

index 4b49201..af53f37 100644 (file)
@@ -59,8 +59,13 @@ struct throtl_grp {
        /* bytes per second rate limits */
        uint64_t bps[2];
 
+       /* IOPS limits */
+       unsigned int iops[2];
+
        /* Number of bytes disptached in current slice */
        uint64_t bytes_disp[2];
+       /* Number of bio's dispatched in current slice */
+       unsigned int io_disp[2];
 
        /* When did we start a new slice */
        unsigned long slice_start[2];
@@ -194,6 +199,8 @@ static struct throtl_grp * throtl_find_alloc_tg(struct throtl_data *td,
 
        tg->bps[READ] = blkcg_get_read_bps(blkcg, tg->blkg.dev);
        tg->bps[WRITE] = blkcg_get_write_bps(blkcg, tg->blkg.dev);
+       tg->iops[READ] = blkcg_get_read_iops(blkcg, tg->blkg.dev);
+       tg->iops[WRITE] = blkcg_get_write_iops(blkcg, tg->blkg.dev);
 
        hlist_add_head(&tg->tg_node, &td->tg_list);
        td->nr_undestroyed_grps++;
@@ -335,6 +342,7 @@ static inline void
 throtl_start_new_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
 {
        tg->bytes_disp[rw] = 0;
+       tg->io_disp[rw] = 0;
        tg->slice_start[rw] = jiffies;
        tg->slice_end[rw] = jiffies + throtl_slice;
        throtl_log_tg(td, tg, "[%c] new slice start=%lu end=%lu jiffies=%lu",
@@ -365,7 +373,7 @@ throtl_slice_used(struct throtl_data *td, struct throtl_grp *tg, bool rw)
 static inline void
 throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
 {
-       unsigned long nr_slices, bytes_trim, time_elapsed;
+       unsigned long nr_slices, bytes_trim, time_elapsed, io_trim;
 
        BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
 
@@ -385,8 +393,9 @@ throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
                return;
 
        bytes_trim = (tg->bps[rw] * throtl_slice * nr_slices)/HZ;
+       io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
 
-       if (!bytes_trim)
+       if (!bytes_trim && !io_trim)
                return;
 
        if (tg->bytes_disp[rw] >= bytes_trim)
@@ -394,51 +403,62 @@ throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
        else
                tg->bytes_disp[rw] = 0;
 
+       if (tg->io_disp[rw] >= io_trim)
+               tg->io_disp[rw] -= io_trim;
+       else
+               tg->io_disp[rw] = 0;
+
        tg->slice_start[rw] += nr_slices * throtl_slice;
 
-       throtl_log_tg(td, tg, "[%c] trim slice nr=%lu bytes=%lu"
+       throtl_log_tg(td, tg, "[%c] trim slice nr=%lu bytes=%lu io=%lu"
                        " start=%lu end=%lu jiffies=%lu",
-                       rw == READ ? 'R' : 'W', nr_slices, bytes_trim,
+                       rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
                        tg->slice_start[rw], tg->slice_end[rw], jiffies);
 }
 
-/*
- * Returns whether one can dispatch a bio or not. Also returns approx number
- * of jiffies to wait before this bio is with-in IO rate and can be dispatched
- */
-static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg,
-                               struct bio *bio, unsigned long *wait)
+static bool tg_with_in_iops_limit(struct throtl_data *td, struct throtl_grp *tg,
+               struct bio *bio, unsigned long *wait)
 {
        bool rw = bio_data_dir(bio);
-       u64 bytes_allowed, extra_bytes;
+       unsigned int io_allowed;
        unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 
-       /*
-        * Currently whole state machine of group depends on first bio
-        * queued in the group bio list. So one should not be calling
-        * this function with a different bio if there are other bios
-        * queued.
-        */
-       BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw]));
+       jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
 
-       /* If tg->bps = -1, then BW is unlimited */
-       if (tg->bps[rw] == -1) {
+       /* Slice has just started. Consider one slice interval */
+       if (!jiffy_elapsed)
+               jiffy_elapsed_rnd = throtl_slice;
+
+       jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
+
+       io_allowed = (tg->iops[rw] * jiffies_to_msecs(jiffy_elapsed_rnd))
+                               / MSEC_PER_SEC;
+
+       if (tg->io_disp[rw] + 1 <= io_allowed) {
                if (wait)
                        *wait = 0;
                return 1;
        }
 
-       /*
-        * If previous slice expired, start a new one otherwise renew/extend
-        * existing slice to make sure it is at least throtl_slice interval
-        * long since now.
-        */
-       if (throtl_slice_used(td, tg, rw))
-               throtl_start_new_slice(td, tg, rw);
-       else {
-               if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
-                       throtl_extend_slice(td, tg, rw, jiffies + throtl_slice);
-       }
+       /* Calc approx time to dispatch */
+       jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
+
+       if (jiffy_wait > jiffy_elapsed)
+               jiffy_wait = jiffy_wait - jiffy_elapsed;
+       else
+               jiffy_wait = 1;
+
+       if (wait)
+               *wait = jiffy_wait;
+       return 0;
+}
+
+static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg,
+               struct bio *bio, unsigned long *wait)
+{
+       bool rw = bio_data_dir(bio);
+       u64 bytes_allowed, extra_bytes;
+       unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 
        jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
 
@@ -469,12 +489,62 @@ static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg,
         * up we did. Add that time also.
         */
        jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
-
        if (wait)
                *wait = jiffy_wait;
+       return 0;
+}
+
+/*
+ * Returns whether one can dispatch a bio or not. Also returns approx number
+ * of jiffies to wait before this bio is with-in IO rate and can be dispatched
+ */
+static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg,
+                               struct bio *bio, unsigned long *wait)
+{
+       bool rw = bio_data_dir(bio);
+       unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
+
+       /*
+        * Currently whole state machine of group depends on first bio
+        * queued in the group bio list. So one should not be calling
+        * this function with a different bio if there are other bios
+        * queued.
+        */
+       BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw]));
 
-       if (time_before(tg->slice_end[rw], jiffies + jiffy_wait))
-               throtl_extend_slice(td, tg, rw, jiffies + jiffy_wait);
+       /* If tg->bps = -1, then BW is unlimited */
+       if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
+               if (wait)
+                       *wait = 0;
+               return 1;
+       }
+
+       /*
+        * If previous slice expired, start a new one otherwise renew/extend
+        * existing slice to make sure it is at least throtl_slice interval
+        * long since now.
+        */
+       if (throtl_slice_used(td, tg, rw))
+               throtl_start_new_slice(td, tg, rw);
+       else {
+               if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
+                       throtl_extend_slice(td, tg, rw, jiffies + throtl_slice);
+       }
+
+       if (tg_with_in_bps_limit(td, tg, bio, &bps_wait)
+           && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) {
+               if (wait)
+                       *wait = 0;
+               return 1;
+       }
+
+       max_wait = max(bps_wait, iops_wait);
+
+       if (wait)
+               *wait = max_wait;
+
+       if (time_before(tg->slice_end[rw], jiffies + max_wait))
+               throtl_extend_slice(td, tg, rw, jiffies + max_wait);
 
        return 0;
 }
@@ -486,13 +556,13 @@ static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
 
        /* Charge the bio to the group */
        tg->bytes_disp[rw] += bio->bi_size;
+       tg->io_disp[rw]++;
 
        /*
         * TODO: This will take blkg->stats_lock. Figure out a way
         * to avoid this cost.
         */
        blkiocg_update_dispatch_stats(&tg->blkg, bio->bi_size, rw, sync);
-
 }
 
 static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg,
@@ -763,6 +833,18 @@ static void throtl_update_blkio_group_write_bps (struct blkio_group *blkg,
        tg_of_blkg(blkg)->bps[WRITE] = write_bps;
 }
 
+static void throtl_update_blkio_group_read_iops (struct blkio_group *blkg,
+                       unsigned int read_iops)
+{
+       tg_of_blkg(blkg)->iops[READ] = read_iops;
+}
+
+static void throtl_update_blkio_group_write_iops (struct blkio_group *blkg,
+                       unsigned int write_iops)
+{
+       tg_of_blkg(blkg)->iops[WRITE] = write_iops;
+}
+
 void throtl_shutdown_timer_wq(struct request_queue *q)
 {
        struct throtl_data *td = q->td;
@@ -777,7 +859,12 @@ static struct blkio_policy_type blkio_policy_throtl = {
                                        throtl_update_blkio_group_read_bps,
                .blkio_update_group_write_bps_fn =
                                        throtl_update_blkio_group_write_bps,
+               .blkio_update_group_read_iops_fn =
+                                       throtl_update_blkio_group_read_iops,
+               .blkio_update_group_write_iops_fn =
+                                       throtl_update_blkio_group_write_iops,
        },
+       .plid = BLKIO_POLICY_THROTL,
 };
 
 int blk_throtl_bio(struct request_queue *q, struct bio **biop)
@@ -811,9 +898,11 @@ int blk_throtl_bio(struct request_queue *q, struct bio **biop)
        }
 
 queue_bio:
-       throtl_log_tg(td, tg, "[%c] bio. disp=%u sz=%u bps=%llu"
-                       " queued=%d/%d", rw == READ ? 'R' : 'W',
+       throtl_log_tg(td, tg, "[%c] bio. bdisp=%u sz=%u bps=%llu"
+                       " iodisp=%u iops=%u queued=%d/%d",
+                       rw == READ ? 'R' : 'W',
                        tg->bytes_disp[rw], bio->bi_size, tg->bps[rw],
+                       tg->io_disp[rw], tg->iops[rw],
                        tg->nr_queued[READ], tg->nr_queued[WRITE]);
 
        throtl_add_bio_tg(q->td, tg, bio);
@@ -850,6 +939,7 @@ int blk_throtl_init(struct request_queue *q)
 
        /* Practically unlimited BW */
        tg->bps[0] = tg->bps[1] = -1;
+       tg->iops[0] = tg->iops[1] = -1;
        atomic_set(&tg->ref, 1);
 
        INIT_DELAYED_WORK(&td->throtl_work, blk_throtl_work);