Merge branch 'x86/hyperv' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip

Topic branch for stable KVM clockource under Hyper-V.

Thanks to Christoffer Dall for resolving the ARM conflict.

14 files changed, 347 insertions, 141 deletions

diff --git a/drivers/md/bcache/alloc.c b/drivers/md/bcache/alloc.c
index a0cc1bc6d884..6cc6c0f9c3a9 100644
--- a/drivers/md/bcache/alloc.c
+++ b/drivers/md/bcache/alloc.c
@@ -525,15 +525,21 @@ struct open_bucket {
 
 /*
  * We keep multiple buckets open for writes, and try to segregate different
- * write streams for better cache utilization: first we look for a bucket where
- * the last write to it was sequential with the current write, and failing that
- * we look for a bucket that was last used by the same task.
+ * write streams for better cache utilization: first we try to segregate flash
+ * only volume write streams from cached devices, secondly we look for a bucket
+ * where the last write to it was sequential with the current write, and
+ * failing that we look for a bucket that was last used by the same task.
  *
  * The ideas is if you've got multiple tasks pulling data into the cache at the
  * same time, you'll get better cache utilization if you try to segregate their
  * data and preserve locality.
  *
- * For example, say you've starting Firefox at the same time you're copying a
+ * For example, dirty sectors of flash only volume is not reclaimable, if their
+ * dirty sectors mixed with dirty sectors of cached device, such buckets will
+ * be marked as dirty and won't be reclaimed, though the dirty data of cached
+ * device have been written back to backend device.
+ *
+ * And say you've starting Firefox at the same time you're copying a
  * bunch of files. Firefox will likely end up being fairly hot and stay in the
  * cache awhile, but the data you copied might not be; if you wrote all that
  * data to the same buckets it'd get invalidated at the same time.
@@ -550,7 +556,10 @@ static struct open_bucket *pick_data_bucket(struct cache_set *c,
 	struct open_bucket *ret, *ret_task = NULL;
 
 	list_for_each_entry_reverse(ret, &c->data_buckets, list)
-		if (!bkey_cmp(&ret->key, search))
+		if (UUID_FLASH_ONLY(&c->uuids[KEY_INODE(&ret->key)]) !=
+		    UUID_FLASH_ONLY(&c->uuids[KEY_INODE(search)]))
+			continue;
+		else if (!bkey_cmp(&ret->key, search))
 			goto found;
 		else if (ret->last_write_point == write_point)
 			ret_task = ret;
diff --git a/drivers/md/bcache/bcache.h b/drivers/md/bcache/bcache.h
index 843877e017e1..5e2d4e80198e 100644
--- a/drivers/md/bcache/bcache.h
+++ b/drivers/md/bcache/bcache.h
@@ -320,14 +320,15 @@ struct cached_dev {
 	 */
 	atomic_t		has_dirty;
 
-	struct bch_ratelimit	writeback_rate;
-	struct delayed_work	writeback_rate_update;
-
 	/*
-	 * Internal to the writeback code, so read_dirty() can keep track of
-	 * where it's at.
+	 * Set to zero by things that touch the backing volume-- except
+	 * writeback.  Incremented by writeback.  Used to determine when to
+	 * accelerate idle writeback.
 	 */
-	sector_t		last_read;
+	atomic_t		backing_idle;
+
+	struct bch_ratelimit	writeback_rate;
+	struct delayed_work	writeback_rate_update;
 
 	/* Limit number of writeback bios in flight */
 	struct semaphore	in_flight;
@@ -336,6 +337,14 @@ struct cached_dev {
 
 	struct keybuf		writeback_keys;
 
+	/*
+	 * Order the write-half of writeback operations strongly in dispatch
+	 * order.  (Maintain LBA order; don't allow reads completing out of
+	 * order to re-order the writes...)
+	 */
+	struct closure_waitlist writeback_ordering_wait;
+	atomic_t		writeback_sequence_next;
+
 	/* For tracking sequential IO */
 #define RECENT_IO_BITS	7
 #define RECENT_IO	(1 << RECENT_IO_BITS)
@@ -488,6 +497,7 @@ struct cache_set {
 	int			caches_loaded;
 
 	struct bcache_device	**devices;
+	unsigned		devices_max_used;
 	struct list_head	cached_devs;
 	uint64_t		cached_dev_sectors;
 	struct closure		caching;
@@ -852,7 +862,7 @@ static inline void wake_up_allocators(struct cache_set *c)
 
 /* Forward declarations */
 
-void bch_count_io_errors(struct cache *, blk_status_t, const char *);
+void bch_count_io_errors(struct cache *, blk_status_t, int, const char *);
 void bch_bbio_count_io_errors(struct cache_set *, struct bio *,
 			      blk_status_t, const char *);
 void bch_bbio_endio(struct cache_set *, struct bio *, blk_status_t,
diff --git a/drivers/md/bcache/btree.c b/drivers/md/bcache/btree.c
index 81e8dc3dbe5e..bf3a48aa9a9a 100644
--- a/drivers/md/bcache/btree.c
+++ b/drivers/md/bcache/btree.c
@@ -419,7 +419,7 @@ static void do_btree_node_write(struct btree *b)
 	SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) +
 		       bset_sector_offset(&b->keys, i));
 
-	if (!bio_alloc_pages(b->bio, __GFP_NOWARN|GFP_NOWAIT)) {
+	if (!bch_bio_alloc_pages(b->bio, __GFP_NOWARN|GFP_NOWAIT)) {
 		int j;
 		struct bio_vec *bv;
 		void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
@@ -432,6 +432,7 @@ static void do_btree_node_write(struct btree *b)
 
 		continue_at(cl, btree_node_write_done, NULL);
 	} else {
+		/* No problem for multipage bvec since the bio is just allocated */
 		b->bio->bi_vcnt = 0;
 		bch_bio_map(b->bio, i);
 
@@ -1678,7 +1679,7 @@ static void bch_btree_gc_finish(struct cache_set *c)
 
 	/* don't reclaim buckets to which writeback keys point */
 	rcu_read_lock();
-	for (i = 0; i < c->nr_uuids; i++) {
+	for (i = 0; i < c->devices_max_used; i++) {
 		struct bcache_device *d = c->devices[i];
 		struct cached_dev *dc;
 		struct keybuf_key *w, *n;
@@ -1803,10 +1804,7 @@ static int bch_gc_thread(void *arg)
 int bch_gc_thread_start(struct cache_set *c)
 {
 	c->gc_thread = kthread_run(bch_gc_thread, c, "bcache_gc");
-	if (IS_ERR(c->gc_thread))
-		return PTR_ERR(c->gc_thread);
-
-	return 0;
+	return PTR_ERR_OR_ZERO(c->gc_thread);
 }
 
 /* Initial partial gc */
diff --git a/drivers/md/bcache/closure.c b/drivers/md/bcache/closure.c
index 1841d0359bac..7f12920c14f7 100644
--- a/drivers/md/bcache/closure.c
+++ b/drivers/md/bcache/closure.c
@@ -8,6 +8,7 @@
 #include <linux/debugfs.h>
 #include <linux/module.h>
 #include <linux/seq_file.h>
+#include <linux/sched/debug.h>
 
 #include "closure.h"
 
@@ -18,10 +19,6 @@ static inline void closure_put_after_sub(struct closure *cl, int flags)
 	BUG_ON(flags & CLOSURE_GUARD_MASK);
 	BUG_ON(!r && (flags & ~CLOSURE_DESTRUCTOR));
 
-	/* Must deliver precisely one wakeup */
-	if (r == 1 && (flags & CLOSURE_SLEEPING))
-		wake_up_process(cl->task);
-
 	if (!r) {
 		if (cl->fn && !(flags & CLOSURE_DESTRUCTOR)) {
 			atomic_set(&cl->remaining,
@@ -100,28 +97,34 @@ bool closure_wait(struct closure_waitlist *waitlist, struct closure *cl)
 }
 EXPORT_SYMBOL(closure_wait);
 
-/**
- * closure_sync - sleep until a closure has nothing left to wait on
- *
- * Sleeps until the refcount hits 1 - the thread that's running the closure owns
- * the last refcount.
- */
-void closure_sync(struct closure *cl)
+struct closure_syncer {
+	struct task_struct	*task;
+	int			done;
+};
+
+static void closure_sync_fn(struct closure *cl)
 {
-	while (1) {
-		__closure_start_sleep(cl);
-		closure_set_ret_ip(cl);
+	cl->s->done = 1;
+	wake_up_process(cl->s->task);
+}
 
-		if ((atomic_read(&cl->remaining) &
-		     CLOSURE_REMAINING_MASK) == 1)
-			break;
+void __sched __closure_sync(struct closure *cl)
+{
+	struct closure_syncer s = { .task = current };
 
+	cl->s = &s;
+	continue_at(cl, closure_sync_fn, NULL);
+
+	while (1) {
+		set_current_state(TASK_UNINTERRUPTIBLE);
+		if (s.done)
+			break;
 		schedule();
 	}
 
-	__closure_end_sleep(cl);
+	__set_current_state(TASK_RUNNING);
 }
-EXPORT_SYMBOL(closure_sync);
+EXPORT_SYMBOL(__closure_sync);
 
 #ifdef CONFIG_BCACHE_CLOSURES_DEBUG
 
@@ -168,12 +171,10 @@ static int debug_seq_show(struct seq_file *f, void *data)
 			   cl, (void *) cl->ip, cl->fn, cl->parent,
 			   r & CLOSURE_REMAINING_MASK);
 
-		seq_printf(f, "%s%s%s%s\n",
+		seq_printf(f, "%s%s\n",
 			   test_bit(WORK_STRUCT_PENDING_BIT,
 				    work_data_bits(&cl->work)) ? "Q" : "",
-			   r & CLOSURE_RUNNING	? "R" : "",
-			   r & CLOSURE_STACK	? "S" : "",
-			   r & CLOSURE_SLEEPING	? "Sl" : "");
+			   r & CLOSURE_RUNNING	? "R" : "");
 
 		if (r & CLOSURE_WAITING)
 			seq_printf(f, " W %pF\n",
diff --git a/drivers/md/bcache/closure.h b/drivers/md/bcache/closure.h
index ccfbea6f9f6b..3b9dfc9962ad 100644
--- a/drivers/md/bcache/closure.h
+++ b/drivers/md/bcache/closure.h
@@ -103,6 +103,7 @@
  */
 
 struct closure;
+struct closure_syncer;
 typedef void (closure_fn) (struct closure *);
 
 struct closure_waitlist {
@@ -115,10 +116,6 @@ enum closure_state {
 	 * the thread that owns the closure, and cleared by the thread that's
 	 * waking up the closure.
 	 *
-	 * CLOSURE_SLEEPING: Must be set before a thread uses a closure to sleep
-	 * - indicates that cl->task is valid and closure_put() may wake it up.
-	 * Only set or cleared by the thread that owns the closure.
-	 *
 	 * The rest are for debugging and don't affect behaviour:
 	 *
 	 * CLOSURE_RUNNING: Set when a closure is running (i.e. by
@@ -128,22 +125,16 @@ enum closure_state {
 	 * continue_at() and closure_return() clear it for you, if you're doing
 	 * something unusual you can use closure_set_dead() which also helps
 	 * annotate where references are being transferred.
-	 *
-	 * CLOSURE_STACK: Sanity check - remaining should never hit 0 on a
-	 * closure with this flag set
 	 */
 
-	CLOSURE_BITS_START	= (1 << 23),
-	CLOSURE_DESTRUCTOR	= (1 << 23),
-	CLOSURE_WAITING		= (1 << 25),
-	CLOSURE_SLEEPING	= (1 << 27),
-	CLOSURE_RUNNING		= (1 << 29),
-	CLOSURE_STACK		= (1 << 31),
+	CLOSURE_BITS_START	= (1U << 26),
+	CLOSURE_DESTRUCTOR	= (1U << 26),
+	CLOSURE_WAITING		= (1U << 28),
+	CLOSURE_RUNNING		= (1U << 30),
 };
 
 #define CLOSURE_GUARD_MASK					\
-	((CLOSURE_DESTRUCTOR|CLOSURE_WAITING|CLOSURE_SLEEPING|	\
-	  CLOSURE_RUNNING|CLOSURE_STACK) << 1)
+	((CLOSURE_DESTRUCTOR|CLOSURE_WAITING|CLOSURE_RUNNING) << 1)
 
 #define CLOSURE_REMAINING_MASK		(CLOSURE_BITS_START - 1)
 #define CLOSURE_REMAINING_INITIALIZER	(1|CLOSURE_RUNNING)
@@ -152,7 +143,7 @@ struct closure {
 	union {
 		struct {
 			struct workqueue_struct *wq;
-			struct task_struct	*task;
+			struct closure_syncer	*s;
 			struct llist_node	list;
 			closure_fn		*fn;
 		};
@@ -178,7 +169,19 @@ void closure_sub(struct closure *cl, int v);
 void closure_put(struct closure *cl);
 void __closure_wake_up(struct closure_waitlist *list);
 bool closure_wait(struct closure_waitlist *list, struct closure *cl);
-void closure_sync(struct closure *cl);
+void __closure_sync(struct closure *cl);
+
+/**
+ * closure_sync - sleep until a closure a closure has nothing left to wait on
+ *
+ * Sleeps until the refcount hits 1 - the thread that's running the closure owns
+ * the last refcount.
+ */
+static inline void closure_sync(struct closure *cl)
+{
+	if ((atomic_read(&cl->remaining) & CLOSURE_REMAINING_MASK) != 1)
+		__closure_sync(cl);
+}
 
 #ifdef CONFIG_BCACHE_CLOSURES_DEBUG
 
@@ -215,24 +218,6 @@ static inline void closure_set_waiting(struct closure *cl, unsigned long f)
 #endif
 }
 
-static inline void __closure_end_sleep(struct closure *cl)
-{
-	__set_current_state(TASK_RUNNING);
-
-	if (atomic_read(&cl->remaining) & CLOSURE_SLEEPING)
-		atomic_sub(CLOSURE_SLEEPING, &cl->remaining);
-}
-
-static inline void __closure_start_sleep(struct closure *cl)
-{
-	closure_set_ip(cl);
-	cl->task = current;
-	set_current_state(TASK_UNINTERRUPTIBLE);
-
-	if (!(atomic_read(&cl->remaining) & CLOSURE_SLEEPING))
-		atomic_add(CLOSURE_SLEEPING, &cl->remaining);
-}
-
 static inline void closure_set_stopped(struct closure *cl)
 {
 	atomic_sub(CLOSURE_RUNNING, &cl->remaining);
@@ -241,7 +226,6 @@ static inline void closure_set_stopped(struct closure *cl)
 static inline void set_closure_fn(struct closure *cl, closure_fn *fn,
 				  struct workqueue_struct *wq)
 {
-	BUG_ON(object_is_on_stack(cl));
 	closure_set_ip(cl);
 	cl->fn = fn;
 	cl->wq = wq;
@@ -300,7 +284,7 @@ static inline void closure_init(struct closure *cl, struct closure *parent)
 static inline void closure_init_stack(struct closure *cl)
 {
 	memset(cl, 0, sizeof(struct closure));
-	atomic_set(&cl->remaining, CLOSURE_REMAINING_INITIALIZER|CLOSURE_STACK);
+	atomic_set(&cl->remaining, CLOSURE_REMAINING_INITIALIZER);
 }
 
 /**
@@ -322,6 +306,8 @@ static inline void closure_wake_up(struct closure_waitlist *list)
  * This is because after calling continue_at() you no longer have a ref on @cl,
  * and whatever @cl owns may be freed out from under you - a running closure fn
  * has a ref on its own closure which continue_at() drops.
+ *
+ * Note you are expected to immediately return after using this macro.
  */
 #define continue_at(_cl, _fn, _wq)					\
 do {									\
diff --git a/drivers/md/bcache/debug.c b/drivers/md/bcache/debug.c
index c7a02c4900da..af89408befe8 100644
--- a/drivers/md/bcache/debug.c
+++ b/drivers/md/bcache/debug.c
@@ -116,7 +116,7 @@ void bch_data_verify(struct cached_dev *dc, struct bio *bio)
 		return;
 	check->bi_opf = REQ_OP_READ;
 
-	if (bio_alloc_pages(check, GFP_NOIO))
+	if (bch_bio_alloc_pages(check, GFP_NOIO))
 		goto out_put;
 
 	submit_bio_wait(check);
@@ -251,8 +251,7 @@ void bch_debug_exit(void)
 
 int __init bch_debug_init(struct kobject *kobj)
 {
-	int ret = 0;
-
 	debug = debugfs_create_dir("bcache", NULL);
-	return ret;
+
+	return IS_ERR_OR_NULL(debug);
 }
diff --git a/drivers/md/bcache/io.c b/drivers/md/bcache/io.c
index fac97ec2d0e2..a783c5a41ff1 100644
--- a/drivers/md/bcache/io.c
+++ b/drivers/md/bcache/io.c
@@ -51,7 +51,10 @@ void bch_submit_bbio(struct bio *bio, struct cache_set *c,
 
 /* IO errors */
 
-void bch_count_io_errors(struct cache *ca, blk_status_t error, const char *m)
+void bch_count_io_errors(struct cache *ca,
+			 blk_status_t error,
+			 int is_read,
+			 const char *m)
 {
 	/*
 	 * The halflife of an error is:
@@ -94,8 +97,9 @@ void bch_count_io_errors(struct cache *ca, blk_status_t error, const char *m)
 		errors >>= IO_ERROR_SHIFT;
 
 		if (errors < ca->set->error_limit)
-			pr_err("%s: IO error on %s, recovering",
-			       bdevname(ca->bdev, buf), m);
+			pr_err("%s: IO error on %s%s",
+			       bdevname(ca->bdev, buf), m,
+			       is_read ? ", recovering." : ".");
 		else
 			bch_cache_set_error(ca->set,
 					    "%s: too many IO errors %s",
@@ -108,6 +112,7 @@ void bch_bbio_count_io_errors(struct cache_set *c, struct bio *bio,
 {
 	struct bbio *b = container_of(bio, struct bbio, bio);
 	struct cache *ca = PTR_CACHE(c, &b->key, 0);
+	int is_read = (bio_data_dir(bio) == READ ? 1 : 0);
 
 	unsigned threshold = op_is_write(bio_op(bio))
 		? c->congested_write_threshold_us
@@ -129,7 +134,7 @@ void bch_bbio_count_io_errors(struct cache_set *c, struct bio *bio,
 			atomic_inc(&c->congested);
 	}
 
-	bch_count_io_errors(ca, error, m);
+	bch_count_io_errors(ca, error, is_read, m);
 }
 
 void bch_bbio_endio(struct cache_set *c, struct bio *bio,
diff --git a/drivers/md/bcache/movinggc.c b/drivers/md/bcache/movinggc.c
index d50c1c97da68..a24c3a95b2c0 100644
--- a/drivers/md/bcache/movinggc.c
+++ b/drivers/md/bcache/movinggc.c
@@ -162,7 +162,7 @@ static void read_moving(struct cache_set *c)
 		bio_set_op_attrs(bio, REQ_OP_READ, 0);
 		bio->bi_end_io	= read_moving_endio;
 
-		if (bio_alloc_pages(bio, GFP_KERNEL))
+		if (bch_bio_alloc_pages(bio, GFP_KERNEL))
 			goto err;
 
 		trace_bcache_gc_copy(&w->key);
diff --git a/drivers/md/bcache/request.c b/drivers/md/bcache/request.c
index 643c3021624f..1a46b41dac70 100644
--- a/drivers/md/bcache/request.c
+++ b/drivers/md/bcache/request.c
@@ -576,6 +576,7 @@ static void cache_lookup(struct closure *cl)
 {
 	struct search *s = container_of(cl, struct search, iop.cl);
 	struct bio *bio = &s->bio.bio;
+	struct cached_dev *dc;
 	int ret;
 
 	bch_btree_op_init(&s->op, -1);
@@ -588,6 +589,27 @@ static void cache_lookup(struct closure *cl)
 		return;
 	}
 
+	/*
+	 * We might meet err when searching the btree, If that happens, we will
+	 * get negative ret, in this scenario we should not recover data from
+	 * backing device (when cache device is dirty) because we don't know
+	 * whether bkeys the read request covered are all clean.
+	 *
+	 * And after that happened, s->iop.status is still its initial value
+	 * before we submit s->bio.bio
+	 */
+	if (ret < 0) {
+		BUG_ON(ret == -EINTR);
+		if (s->d && s->d->c &&
+				!UUID_FLASH_ONLY(&s->d->c->uuids[s->d->id])) {
+			dc = container_of(s->d, struct cached_dev, disk);
+			if (dc && atomic_read(&dc->has_dirty))
+				s->recoverable = false;
+		}
+		if (!s->iop.status)
+			s->iop.status = BLK_STS_IOERR;
+	}
+
 	closure_return(cl);
 }
 
@@ -611,8 +633,8 @@ static void request_endio(struct bio *bio)
 static void bio_complete(struct search *s)
 {
 	if (s->orig_bio) {
-		struct request_queue *q = s->orig_bio->bi_disk->queue;
-		generic_end_io_acct(q, bio_data_dir(s->orig_bio),
+		generic_end_io_acct(s->d->disk->queue,
+				    bio_data_dir(s->orig_bio),
 				    &s->d->disk->part0, s->start_time);
 
 		trace_bcache_request_end(s->d, s->orig_bio);
@@ -841,7 +863,7 @@ static int cached_dev_cache_miss(struct btree *b, struct search *s,
 	cache_bio->bi_private	= &s->cl;
 
 	bch_bio_map(cache_bio, NULL);
-	if (bio_alloc_pages(cache_bio, __GFP_NOWARN|GFP_NOIO))
+	if (bch_bio_alloc_pages(cache_bio, __GFP_NOWARN|GFP_NOIO))
 		goto out_put;
 
 	if (reada)
@@ -974,6 +996,7 @@ static blk_qc_t cached_dev_make_request(struct request_queue *q,
 	struct cached_dev *dc = container_of(d, struct cached_dev, disk);
 	int rw = bio_data_dir(bio);
 
+	atomic_set(&dc->backing_idle, 0);
 	generic_start_io_acct(q, rw, bio_sectors(bio), &d->disk->part0);
 
 	bio_set_dev(bio, dc->bdev);
diff --git a/drivers/md/bcache/super.c b/drivers/md/bcache/super.c
index b4d28928dec5..133b81225ea9 100644
--- a/drivers/md/bcache/super.c
+++ b/drivers/md/bcache/super.c
@@ -211,7 +211,7 @@ static void write_bdev_super_endio(struct bio *bio)
 
 static void __write_super(struct cache_sb *sb, struct bio *bio)
 {
-	struct cache_sb *out = page_address(bio->bi_io_vec[0].bv_page);
+	struct cache_sb *out = page_address(bio_first_page_all(bio));
 	unsigned i;
 
 	bio->bi_iter.bi_sector	= SB_SECTOR;
@@ -274,7 +274,9 @@ static void write_super_endio(struct bio *bio)
 {
 	struct cache *ca = bio->bi_private;
 
-	bch_count_io_errors(ca, bio->bi_status, "writing superblock");
+	/* is_read = 0 */
+	bch_count_io_errors(ca, bio->bi_status, 0,
+			    "writing superblock");
 	closure_put(&ca->set->sb_write);
 }
 
@@ -721,6 +723,9 @@ static void bcache_device_attach(struct bcache_device *d, struct cache_set *c,
 	d->c = c;
 	c->devices[id] = d;
 
+	if (id >= c->devices_max_used)
+		c->devices_max_used = id + 1;
+
 	closure_get(&c->caching);
 }
 
@@ -906,6 +911,12 @@ static void cached_dev_detach_finish(struct work_struct *w)
 
 	mutex_lock(&bch_register_lock);
 
+	cancel_delayed_work_sync(&dc->writeback_rate_update);
+	if (!IS_ERR_OR_NULL(dc->writeback_thread)) {
+		kthread_stop(dc->writeback_thread);
+		dc->writeback_thread = NULL;
+	}
+
 	memset(&dc->sb.set_uuid, 0, 16);
 	SET_BDEV_STATE(&dc->sb, BDEV_STATE_NONE);
 
@@ -1166,7 +1177,7 @@ static void register_bdev(struct cache_sb *sb, struct page *sb_page,
 	dc->bdev->bd_holder = dc;
 
 	bio_init(&dc->sb_bio, dc->sb_bio.bi_inline_vecs, 1);
-	dc->sb_bio.bi_io_vec[0].bv_page = sb_page;
+	bio_first_bvec_all(&dc->sb_bio)->bv_page = sb_page;
 	get_page(sb_page);
 
 	if (cached_dev_init(dc, sb->block_size << 9))
@@ -1261,7 +1272,7 @@ static int flash_devs_run(struct cache_set *c)
 	struct uuid_entry *u;
 
 	for (u = c->uuids;
-	     u < c->uuids + c->nr_uuids && !ret;
+	     u < c->uuids + c->devices_max_used && !ret;
 	     u++)
 		if (UUID_FLASH_ONLY(u))
 			ret = flash_dev_run(c, u);
@@ -1427,7 +1438,7 @@ static void __cache_set_unregister(struct closure *cl)
 
 	mutex_lock(&bch_register_lock);
 
-	for (i = 0; i < c->nr_uuids; i++)
+	for (i = 0; i < c->devices_max_used; i++)
 		if (c->devices[i]) {
 			if (!UUID_FLASH_ONLY(&c->uuids[i]) &&
 			    test_bit(CACHE_SET_UNREGISTERING, &c->flags)) {
@@ -1490,7 +1501,7 @@ struct cache_set *bch_cache_set_alloc(struct cache_sb *sb)
 	c->bucket_bits		= ilog2(sb->bucket_size);
 	c->block_bits		= ilog2(sb->block_size);
 	c->nr_uuids		= bucket_bytes(c) / sizeof(struct uuid_entry);
-
+	c->devices_max_used	= 0;
 	c->btree_pages		= bucket_pages(c);
 	if (c->btree_pages > BTREE_MAX_PAGES)
 		c->btree_pages = max_t(int, c->btree_pages / 4,
@@ -1810,7 +1821,7 @@ void bch_cache_release(struct kobject *kobj)
 		free_fifo(&ca->free[i]);
 
 	if (ca->sb_bio.bi_inline_vecs[0].bv_page)
-		put_page(ca->sb_bio.bi_io_vec[0].bv_page);
+		put_page(bio_first_page_all(&ca->sb_bio));
 
 	if (!IS_ERR_OR_NULL(ca->bdev))
 		blkdev_put(ca->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
@@ -1864,7 +1875,7 @@ static int register_cache(struct cache_sb *sb, struct page *sb_page,
 	ca->bdev->bd_holder = ca;
 
 	bio_init(&ca->sb_bio, ca->sb_bio.bi_inline_vecs, 1);
-	ca->sb_bio.bi_io_vec[0].bv_page = sb_page;
+	bio_first_bvec_all(&ca->sb_bio)->bv_page = sb_page;
 	get_page(sb_page);
 
 	if (blk_queue_discard(bdev_get_queue(ca->bdev)))
diff --git a/drivers/md/bcache/util.c b/drivers/md/bcache/util.c
index e548b8b51322..a23cd6a14b74 100644
--- a/drivers/md/bcache/util.c
+++ b/drivers/md/bcache/util.c
@@ -249,6 +249,13 @@ uint64_t bch_next_delay(struct bch_ratelimit *d, uint64_t done)
 		: 0;
 }
 
+/*
+ * Generally it isn't good to access .bi_io_vec and .bi_vcnt directly,
+ * the preferred way is bio_add_page, but in this case, bch_bio_map()
+ * supposes that the bvec table is empty, so it is safe to access
+ * .bi_vcnt & .bi_io_vec in this way even after multipage bvec is
+ * supported.
+ */
 void bch_bio_map(struct bio *bio, void *base)
 {
 	size_t size = bio->bi_iter.bi_size;
@@ -276,6 +283,33 @@ start:		bv->bv_len	= min_t(size_t, PAGE_SIZE - bv->bv_offset,
 	}
 }
 
+/**
+ * bch_bio_alloc_pages - allocates a single page for each bvec in a bio
+ * @bio: bio to allocate pages for
+ * @gfp_mask: flags for allocation
+ *
+ * Allocates pages up to @bio->bi_vcnt.
+ *
+ * Returns 0 on success, -ENOMEM on failure. On failure, any allocated pages are
+ * freed.
+ */
+int bch_bio_alloc_pages(struct bio *bio, gfp_t gfp_mask)
+{
+	int i;
+	struct bio_vec *bv;
+
+	bio_for_each_segment_all(bv, bio, i) {
+		bv->bv_page = alloc_page(gfp_mask);
+		if (!bv->bv_page) {
+			while (--bv >= bio->bi_io_vec)
+				__free_page(bv->bv_page);
+			return -ENOMEM;
+		}
+	}
+
+	return 0;
+}
+
 /*
  * Portions Copyright (c) 1996-2001, PostgreSQL Global Development Group (Any
  * use permitted, subject to terms of PostgreSQL license; see.)
diff --git a/drivers/md/bcache/util.h b/drivers/md/bcache/util.h
index ed5e8a412eb8..4df4c5c1cab2 100644
--- a/drivers/md/bcache/util.h
+++ b/drivers/md/bcache/util.h
@@ -558,6 +558,7 @@ static inline unsigned fract_exp_two(unsigned x, unsigned fract_bits)
 }
 
 void bch_bio_map(struct bio *bio, void *base);
+int bch_bio_alloc_pages(struct bio *bio, gfp_t gfp_mask);
 
 static inline sector_t bdev_sectors(struct block_device *bdev)
 {
diff --git a/drivers/md/bcache/writeback.c b/drivers/md/bcache/writeback.c
index 56a37884ca8b..51306a19ab03 100644
--- a/drivers/md/bcache/writeback.c
+++ b/drivers/md/bcache/writeback.c
@@ -18,17 +18,39 @@
 #include <trace/events/bcache.h>
 
 /* Rate limiting */
-
-static void __update_writeback_rate(struct cached_dev *dc)
+static uint64_t __calc_target_rate(struct cached_dev *dc)
 {
 	struct cache_set *c = dc->disk.c;
+
+	/*
+	 * This is the size of the cache, minus the amount used for
+	 * flash-only devices
+	 */
 	uint64_t cache_sectors = c->nbuckets * c->sb.bucket_size -
 				bcache_flash_devs_sectors_dirty(c);
+
+	/*
+	 * Unfortunately there is no control of global dirty data.  If the
+	 * user states that they want 10% dirty data in the cache, and has,
+	 * e.g., 5 backing volumes of equal size, we try and ensure each
+	 * backing volume uses about 2% of the cache for dirty data.
+	 */
+	uint32_t bdev_share =
+		div64_u64(bdev_sectors(dc->bdev) << WRITEBACK_SHARE_SHIFT,
+				c->cached_dev_sectors);
+
 	uint64_t cache_dirty_target =
 		div_u64(cache_sectors * dc->writeback_percent, 100);
-	int64_t target = div64_u64(cache_dirty_target * bdev_sectors(dc->bdev),
-				   c->cached_dev_sectors);
 
+	/* Ensure each backing dev gets at least one dirty share */
+	if (bdev_share < 1)
+		bdev_share = 1;
+
+	return (cache_dirty_target * bdev_share) >> WRITEBACK_SHARE_SHIFT;
+}
+
+static void __update_writeback_rate(struct cached_dev *dc)
+{
 	/*
 	 * PI controller:
 	 * Figures out the amount that should be written per second.
@@ -49,6 +71,7 @@ static void __update_writeback_rate(struct cached_dev *dc)
 	 * This acts as a slow, long-term average that is not subject to
 	 * variations in usage like the p term.
 	 */
+	int64_t target = __calc_target_rate(dc);
 	int64_t dirty = bcache_dev_sectors_dirty(&dc->disk);
 	int64_t error = dirty - target;
 	int64_t proportional_scaled =
@@ -116,6 +139,7 @@ static unsigned writeback_delay(struct cached_dev *dc, unsigned sectors)
 struct dirty_io {
 	struct closure		cl;
 	struct cached_dev	*dc;
+	uint16_t		sequence;
 	struct bio		bio;
 };
 
@@ -194,6 +218,27 @@ static void write_dirty(struct closure *cl)
 {
 	struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 	struct keybuf_key *w = io->bio.bi_private;
+	struct cached_dev *dc = io->dc;
+
+	uint16_t next_sequence;
+
+	if (atomic_read(&dc->writeback_sequence_next) != io->sequence) {
+		/* Not our turn to write; wait for a write to complete */
+		closure_wait(&dc->writeback_ordering_wait, cl);
+
+		if (atomic_read(&dc->writeback_sequence_next) == io->sequence) {
+			/*
+			 * Edge case-- it happened in indeterminate order
+			 * relative to when we were added to wait list..
+			 */
+			closure_wake_up(&dc->writeback_ordering_wait);
+		}
+
+		continue_at(cl, write_dirty, io->dc->writeback_write_wq);
+		return;
+	}
+
+	next_sequence = io->sequence + 1;
 
 	/*
 	 * IO errors are signalled using the dirty bit on the key.
@@ -211,6 +256,9 @@ static void write_dirty(struct closure *cl)
 		closure_bio_submit(&io->bio, cl);
 	}
 
+	atomic_set(&dc->writeback_sequence_next, next_sequence);
+	closure_wake_up(&dc->writeback_ordering_wait);
+
 	continue_at(cl, write_dirty_finish, io->dc->writeback_write_wq);
 }
 
@@ -219,8 +267,10 @@ static void read_dirty_endio(struct bio *bio)
 	struct keybuf_key *w = bio->bi_private;
 	struct dirty_io *io = w->private;
 
+	/* is_read = 1 */
 	bch_count_io_errors(PTR_CACHE(io->dc->disk.c, &w->key, 0),
-			    bio->bi_status, "reading dirty data from cache");
+			    bio->bi_status, 1,
+			    "reading dirty data from cache");
 
 	dirty_endio(bio);
 }
@@ -237,10 +287,15 @@ static void read_dirty_submit(struct closure *cl)
 static void read_dirty(struct cached_dev *dc)
 {
 	unsigned delay = 0;
-	struct keybuf_key *w;
+	struct keybuf_key *next, *keys[MAX_WRITEBACKS_IN_PASS], *w;
+	size_t size;
+	int nk, i;
 	struct dirty_io *io;
 	struct closure cl;
+	uint16_t sequence = 0;
 
+	BUG_ON(!llist_empty(&dc->writeback_ordering_wait.list));
+	atomic_set(&dc->writeback_sequence_next, sequence);
 	closure_init_stack(&cl);
 
 	/*
@@ -248,45 +303,109 @@ static void read_dirty(struct cached_dev *dc)
 	 * mempools.
 	 */
 
-	while (!kthread_should_stop()) {
-
-		w = bch_keybuf_next(&dc->writeback_keys);
-		if (!w)
-			break;
-
-		BUG_ON(ptr_stale(dc->disk.c, &w->key, 0));
-
-		if (KEY_START(&w->key) != dc->last_read ||
-		    jiffies_to_msecs(delay) > 50)
-			while (!kthread_should_stop() && delay)
-				delay = schedule_timeout_interruptible(delay);
-
-		dc->last_read	= KEY_OFFSET(&w->key);
-
-		io = kzalloc(sizeof(struct dirty_io) + sizeof(struct bio_vec)
-			     * DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS),
-			     GFP_KERNEL);
-		if (!io)
-			goto err;
-
-		w->private	= io;
-		io->dc		= dc;
-
-		dirty_init(w);
-		bio_set_op_attrs(&io->bio, REQ_OP_READ, 0);
-		io->bio.bi_iter.bi_sector = PTR_OFFSET(&w->key, 0);
-		bio_set_dev(&io->bio, PTR_CACHE(dc->disk.c, &w->key, 0)->bdev);
-		io->bio.bi_end_io	= read_dirty_endio;
-
-		if (bio_alloc_pages(&io->bio, GFP_KERNEL))
-			goto err_free;
+	next = bch_keybuf_next(&dc->writeback_keys);
+
+	while (!kthread_should_stop() && next) {
+		size = 0;
+		nk = 0;
+
+		do {
+			BUG_ON(ptr_stale(dc->disk.c, &next->key, 0));
+
+			/*
+			 * Don't combine too many operations, even if they
+			 * are all small.
+			 */
+			if (nk >= MAX_WRITEBACKS_IN_PASS)
+				break;
+
+			/*
+			 * If the current operation is very large, don't
+			 * further combine operations.
+			 */
+			if (size >= MAX_WRITESIZE_IN_PASS)
+				break;
+
+			/*
+			 * Operations are only eligible to be combined
+			 * if they are contiguous.
+			 *
+			 * TODO: add a heuristic willing to fire a
+			 * certain amount of non-contiguous IO per pass,
+			 * so that we can benefit from backing device
+			 * command queueing.
+			 */
+			if ((nk != 0) && bkey_cmp(&keys[nk-1]->key,
+						&START_KEY(&next->key)))
+				break;
+
+			size += KEY_SIZE(&next->key);
+			keys[nk++] = next;
+		} while ((next = bch_keybuf_next(&dc->writeback_keys)));
+
+		/* Now we have gathered a set of 1..5 keys to write back. */
+		for (i = 0; i < nk; i++) {
+			w = keys[i];
+
+			io = kzalloc(sizeof(struct dirty_io) +
+				     sizeof(struct bio_vec) *
+				     DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS),
+				     GFP_KERNEL);
+			if (!io)
+				goto err;
+
+			w->private	= io;
+			io->dc		= dc;
+			io->sequence    = sequence++;
+
+			dirty_init(w);
+			bio_set_op_attrs(&io->bio, REQ_OP_READ, 0);
+			io->bio.bi_iter.bi_sector = PTR_OFFSET(&w->key, 0);
+			bio_set_dev(&io->bio,
+				    PTR_CACHE(dc->disk.c, &w->key, 0)->bdev);
+			io->bio.bi_end_io	= read_dirty_endio;
+
+			if (bch_bio_alloc_pages(&io->bio, GFP_KERNEL))
+				goto err_free;
+
+			trace_bcache_writeback(&w->key);
+
+			down(&dc->in_flight);
+
+			/* We've acquired a semaphore for the maximum
+			 * simultaneous number of writebacks; from here
+			 * everything happens asynchronously.
+			 */
+			closure_call(&io->cl, read_dirty_submit, NULL, &cl);
+		}
 
-		trace_bcache_writeback(&w->key);
+		delay = writeback_delay(dc, size);
 
-		down(&dc->in_flight);
-		closure_call(&io->cl, read_dirty_submit, NULL, &cl);
+		/* If the control system would wait for at least half a
+		 * second, and there's been no reqs hitting the backing disk
+		 * for awhile: use an alternate mode where we have at most
+		 * one contiguous set of writebacks in flight at a time.  If
+		 * someone wants to do IO it will be quick, as it will only
+		 * have to contend with one operation in flight, and we'll
+		 * be round-tripping data to the backing disk as quickly as
+		 * it can accept it.
+		 */
+		if (delay >= HZ / 2) {
+			/* 3 means at least 1.5 seconds, up to 7.5 if we
+			 * have slowed way down.
+			 */
+			if (atomic_inc_return(&dc->backing_idle) >= 3) {
+				/* Wait for current I/Os to finish */
+				closure_sync(&cl);
+				/* And immediately launch a new set. */
+				delay = 0;
+			}
+		}
 
-		delay = writeback_delay(dc, KEY_SIZE(&w->key));
+		while (!kthread_should_stop() && delay) {
+			schedule_timeout_interruptible(delay);
+			delay = writeback_delay(dc, 0);
+		}
 	}
 
 	if (0) {
diff --git a/drivers/md/bcache/writeback.h b/drivers/md/bcache/writeback.h
index a9e3ffb4b03c..66f1c527fa24 100644
--- a/drivers/md/bcache/writeback.h
+++ b/drivers/md/bcache/writeback.h
@@ -5,6 +5,16 @@
 #define CUTOFF_WRITEBACK	40
 #define CUTOFF_WRITEBACK_SYNC	70
 
+#define MAX_WRITEBACKS_IN_PASS  5
+#define MAX_WRITESIZE_IN_PASS   5000	/* *512b */
+
+/*
+ * 14 (16384ths) is chosen here as something that each backing device
+ * should be a reasonable fraction of the share, and not to blow up
+ * until individual backing devices are a petabyte.
+ */
+#define WRITEBACK_SHARE_SHIFT   14
+
 static inline uint64_t bcache_dev_sectors_dirty(struct bcache_device *d)
 {
 	uint64_t i, ret = 0;
@@ -21,7 +31,7 @@ static inline uint64_t  bcache_flash_devs_sectors_dirty(struct cache_set *c)
 
 	mutex_lock(&bch_register_lock);
 
-	for (i = 0; i < c->nr_uuids; i++) {
+	for (i = 0; i < c->devices_max_used; i++) {
 		struct bcache_device *d = c->devices[i];
 
 		if (!d || !UUID_FLASH_ONLY(&c->uuids[i]))