bcache: Fix a writeback performance regression

Background writeback works by scanning the btree for dirty data and
adding those keys into a fixed size buffer, then for each dirty key in
the keybuf writing it to the backing device.

When read_dirty() finishes and it's time to scan for more dirty data, we
need to wait for the outstanding writeback IO to finish - they still
take up slots in the keybuf (so that foreground writes can check for
them to avoid races) - without that wait, we'll continually rescan when
we'll be able to add at most a key or two to the keybuf, and that takes
locks that starves foreground IO.  Doh.

Signed-off-by: Kent Overstreet <kmo@daterainc.com>
Cc: linux-stable <stable@vger.kernel.org> # >= v3.10
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

4 files changed, 43 insertions, 30 deletions

diff --git a/drivers/md/bcache/bcache.h b/drivers/md/bcache/bcache.h
index b39f6f0b45f2..0f12382aa35d 100644
--- a/drivers/md/bcache/bcache.h
+++ b/drivers/md/bcache/bcache.h
@@ -498,7 +498,7 @@ struct cached_dev {
          */
         atomic_t                has_dirty;
 
-        struct ratelimit        writeback_rate;
+        struct bch_ratelimit    writeback_rate;
         struct delayed_work     writeback_rate_update;
 
         /*
@@ -507,10 +507,9 @@ struct cached_dev {
          */
         sector_t                last_read;
 
-        /* Number of writeback bios in flight */
-        atomic_t                in_flight;
+        /* Limit number of writeback bios in flight */
+        struct semaphore        in_flight;
         struct closure_with_timer writeback;
-        struct closure_waitlist writeback_wait;
 
         struct keybuf           writeback_keys;
 
diff --git a/drivers/md/bcache/util.c b/drivers/md/bcache/util.c
index 98eb81159a22..420dad545c7d 100644
--- a/drivers/md/bcache/util.c
+++ b/drivers/md/bcache/util.c
@@ -190,7 +190,16 @@ void bch_time_stats_update(struct time_stats *stats, uint64_t start_time)
         stats->last = now ?: 1;
 }
 
-unsigned bch_next_delay(struct ratelimit *d, uint64_t done)
+/**
+ * bch_next_delay() - increment @d by the amount of work done, and return how
+ * long to delay until the next time to do some work.
+ *
+ * @d - the struct bch_ratelimit to update
+ * @done - the amount of work done, in arbitrary units
+ *
+ * Returns the amount of time to delay by, in jiffies
+ */
+uint64_t bch_next_delay(struct bch_ratelimit *d, uint64_t done)
 {
         uint64_t now = local_clock();
 
diff --git a/drivers/md/bcache/util.h b/drivers/md/bcache/util.h
index 1ae2a73ad85f..ea345c6896f4 100644
--- a/drivers/md/bcache/util.h
+++ b/drivers/md/bcache/util.h
@@ -450,17 +450,23 @@ read_attribute(name ## _last_ ## frequency_units)
         (ewma) >> factor;                                               \
 })
 
-struct ratelimit {
+struct bch_ratelimit {
+        /* Next time we want to do some work, in nanoseconds */
         uint64_t                next;
+
+        /*
+         * Rate at which we want to do work, in units per nanosecond
+         * The units here correspond to the units passed to bch_next_delay()
+         */
         unsigned                rate;
 };
 
-static inline void ratelimit_reset(struct ratelimit *d)
+static inline void bch_ratelimit_reset(struct bch_ratelimit *d)
 {
         d->next = local_clock();
 }
 
-unsigned bch_next_delay(struct ratelimit *d, uint64_t done);
+uint64_t bch_next_delay(struct bch_ratelimit *d, uint64_t done);
 
 #define __DIV_SAFE(n, d, zero)                                          \
 ({                                                                      \
diff --git a/drivers/md/bcache/writeback.c b/drivers/md/bcache/writeback.c
index 22cbff551628..27ac51934822 100644
--- a/drivers/md/bcache/writeback.c
+++ b/drivers/md/bcache/writeback.c
@@ -94,11 +94,15 @@ static void update_writeback_rate(struct work_struct *work)
 
 static unsigned writeback_delay(struct cached_dev *dc, unsigned sectors)
 {
+        uint64_t ret;
+
         if (atomic_read(&dc->disk.detaching) ||
             !dc->writeback_percent)
                 return 0;
 
-        return bch_next_delay(&dc->writeback_rate, sectors * 10000000ULL);
+        ret = bch_next_delay(&dc->writeback_rate, sectors * 10000000ULL);
+
+        return min_t(uint64_t, ret, HZ);
 }
 
 /* Background writeback */
@@ -208,7 +212,7 @@ normal_refill:
 
         up_write(&dc->writeback_lock);
 
-        ratelimit_reset(&dc->writeback_rate);
+        bch_ratelimit_reset(&dc->writeback_rate);
 
         /* Punt to workqueue only so we don't recurse and blow the stack */
         continue_at(cl, read_dirty, dirty_wq);
@@ -318,9 +322,7 @@ static void write_dirty_finish(struct closure *cl)
         }
 
         bch_keybuf_del(&dc->writeback_keys, w);
-        atomic_dec_bug(&dc->in_flight);
-
-        closure_wake_up(&dc->writeback_wait);
+        up(&dc->in_flight);
 
         closure_return_with_destructor(cl, dirty_io_destructor);
 }
@@ -349,7 +351,7 @@ static void write_dirty(struct closure *cl)
 
         closure_bio_submit(&io->bio, cl, &io->dc->disk);
 
-        continue_at(cl, write_dirty_finish, dirty_wq);
+        continue_at(cl, write_dirty_finish, system_wq);
 }
 
 static void read_dirty_endio(struct bio *bio, int error)
@@ -369,7 +371,7 @@ static void read_dirty_submit(struct closure *cl)
 
         closure_bio_submit(&io->bio, cl, &io->dc->disk);
 
-        continue_at(cl, write_dirty, dirty_wq);
+        continue_at(cl, write_dirty, system_wq);
 }
 
 static void read_dirty(struct closure *cl)
@@ -394,12 +396,9 @@ static void read_dirty(struct closure *cl)
 
                 if (delay > 0 &&
                     (KEY_START(&w->key) != dc->last_read ||
-                     jiffies_to_msecs(delay) > 50)) {
-                        w->private = NULL;
-
-                        closure_delay(&dc->writeback, delay);
-                        continue_at(cl, read_dirty, dirty_wq);
-                }
+                     jiffies_to_msecs(delay) > 50))
+                        while (delay)
+                                delay = schedule_timeout(delay);
 
                 dc->last_read   = KEY_OFFSET(&w->key);
 
@@ -424,15 +423,10 @@ static void read_dirty(struct closure *cl)
 
                 trace_bcache_writeback(&w->key);
 
-                closure_call(&io->cl, read_dirty_submit, NULL, &dc->disk.cl);
+                down(&dc->in_flight);
+                closure_call(&io->cl, read_dirty_submit, NULL, cl);
 
                 delay = writeback_delay(dc, KEY_SIZE(&w->key));
-
-                atomic_inc(&dc->in_flight);
-
-                if (!closure_wait_event(&dc->writeback_wait, cl,
-                                        atomic_read(&dc->in_flight) < 64))
-                        continue_at(cl, read_dirty, dirty_wq);
         }
 
         if (0) {
@@ -442,7 +436,11 @@ err:
                 bch_keybuf_del(&dc->writeback_keys, w);
         }
 
-        refill_dirty(cl);
+        /*
+         * Wait for outstanding writeback IOs to finish (and keybuf slots to be
+         * freed) before refilling again
+         */
+        continue_at(cl, refill_dirty, dirty_wq);
 }
 
 /* Init */
@@ -484,6 +482,7 @@ void bch_sectors_dirty_init(struct cached_dev *dc)
 
 void bch_cached_dev_writeback_init(struct cached_dev *dc)
 {
+        sema_init(&dc->in_flight, 64);
         closure_init_unlocked(&dc->writeback);
         init_rwsem(&dc->writeback_lock);
 
@@ -513,7 +512,7 @@ void bch_writeback_exit(void)
 
 int __init bch_writeback_init(void)
 {
-        dirty_wq = create_singlethread_workqueue("bcache_writeback");
+        dirty_wq = create_workqueue("bcache_writeback");
         if (!dirty_wq)
                 return -ENOMEM;