dm cache: add mq policy

A cache policy that uses a multiqueue ordered by recent hit
count to select which blocks should be promoted and demoted.
This is meant to be a general purpose policy.  It prioritises
reads over writes.

Signed-off-by: Joe Thornber <ejt@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>

4 files changed, 1279 insertions, 0 deletions

diff --git a/Documentation/device-mapper/cache-policies.txt b/Documentation/device-mapper/cache-policies.txt
new file mode 100644
index 0000000..731879f9
--- /dev/null
+++ b/Documentation/device-mapper/cache-policies.txt
@@ -0,0 +1,72 @@
+Guidance for writing policies
+=============================
+
+Try to keep transactionality out of it.  The core is careful to
+avoid asking about anything that is migrating.  This is a pain, but
+makes it easier to write the policies.
+
+Mappings are loaded into the policy at construction time.
+
+Every bio that is mapped by the target is referred to the policy.
+The policy can return a simple HIT or MISS or issue a migration.
+
+Currently there's no way for the policy to issue background work,
+e.g. to start writing back dirty blocks that are going to be evicte
+soon.
+
+Because we map bios, rather than requests it's easy for the policy
+to get fooled by many small bios.  For this reason the core target
+issues periodic ticks to the policy.  It's suggested that the policy
+doesn't update states (eg, hit counts) for a block more than once
+for each tick.  The core ticks by watching bios complete, and so
+trying to see when the io scheduler has let the ios run.
+
+
+Overview of supplied cache replacement policies
+===============================================
+
+multiqueue
+----------
+
+This policy is the default.
+
+The multiqueue policy has two sets of 16 queues: one set for entries
+waiting for the cache and another one for those in the cache.
+Cache entries in the queues are aged based on logical time. Entry into
+the cache is based on variable thresholds and queue selection is based
+on hit count on entry. The policy aims to take different cache miss
+costs into account and to adjust to varying load patterns automatically.
+
+Message and constructor argument pairs are:
+	'sequential_threshold <#nr_sequential_ios>' and
+	'random_threshold <#nr_random_ios>'.
+
+The sequential threshold indicates the number of contiguous I/Os
+required before a stream is treated as sequential.  The random threshold
+is the number of intervening non-contiguous I/Os that must be seen
+before the stream is treated as random again.
+
+The sequential and random thresholds default to 512 and 4 respectively.
+
+Large, sequential ios are probably better left on the origin device
+since spindles tend to have good bandwidth. The io_tracker counts
+contiguous I/Os to try to spot when the io is in one of these sequential
+modes.
+
+Examples
+========
+
+The syntax for a table is:
+	cache <metadata dev> <cache dev> <origin dev> <block size>
+	<#feature_args> [<feature arg>]*
+	<policy> <#policy_args> [<policy arg>]*
+
+The syntax to send a message using the dmsetup command is:
+	dmsetup message <mapped device> 0 sequential_threshold 1024
+	dmsetup message <mapped device> 0 random_threshold 8
+
+Using dmsetup:
+	dmsetup create blah --table "0 268435456 cache /dev/sdb /dev/sdc \
+	    /dev/sdd 512 0 mq 4 sequential_threshold 1024 random_threshold 8"
+	creates a 128GB large mapped device named 'blah' with the
+	sequential threshold set to 1024 and the random_threshold set to 8.
diff --git a/drivers/md/Kconfig b/drivers/md/Kconfig
index 1a4fbcd..1a96cbc 100644
--- a/drivers/md/Kconfig
+++ b/drivers/md/Kconfig
@@ -281,6 +281,16 @@ config DM_CACHE
          algorithms used to select which blocks are promoted, demoted,
          cleaned etc.  It supports writeback and writethrough modes.
 
+config DM_CACHE_MQ
+       tristate "MQ Cache Policy (EXPERIMENTAL)"
+       depends on DM_CACHE
+       default y
+       ---help---
+         A cache policy that uses a multiqueue ordered by recent hit
+         count to select which blocks should be promoted and demoted.
+         This is meant to be a general purpose policy.  It prioritises
+         reads over writes.
+
 config DM_MIRROR
        tristate "Mirror target"
        depends on BLK_DEV_DM
diff --git a/drivers/md/Makefile b/drivers/md/Makefile
index 24b5256..adc8710 100644
--- a/drivers/md/Makefile
+++ b/drivers/md/Makefile
@@ -12,6 +12,7 @@ dm-log-userspace-y \
 		+= dm-log-userspace-base.o dm-log-userspace-transfer.o
 dm-thin-pool-y	+= dm-thin.o dm-thin-metadata.o
 dm-cache-y	+= dm-cache-target.o dm-cache-metadata.o dm-cache-policy.o
+dm-cache-mq-y   += dm-cache-policy-mq.o
 md-mod-y	+= md.o bitmap.o
 raid456-y	+= raid5.o
 
@@ -46,6 +47,7 @@ obj-$(CONFIG_DM_RAID)	+= dm-raid.o
 obj-$(CONFIG_DM_THIN_PROVISIONING)	+= dm-thin-pool.o
 obj-$(CONFIG_DM_VERITY)		+= dm-verity.o
 obj-$(CONFIG_DM_CACHE)		+= dm-cache.o
+obj-$(CONFIG_DM_CACHE_MQ)	+= dm-cache-mq.o
 
 ifeq ($(CONFIG_DM_UEVENT),y)
 dm-mod-objs			+= dm-uevent.o
diff --git a/drivers/md/dm-cache-policy-mq.c b/drivers/md/dm-cache-policy-mq.c
new file mode 100644
index 0000000..9641532
--- /dev/null
+++ b/drivers/md/dm-cache-policy-mq.c
@@ -0,0 +1,1195 @@
+/*
+ * Copyright (C) 2012 Red Hat. All rights reserved.
+ *
+ * This file is released under the GPL.
+ */
+
+#include "dm-cache-policy.h"
+#include "dm.h"
+
+#include <linux/hash.h>
+#include <linux/module.h>
+#include <linux/mutex.h>
+#include <linux/slab.h>
+#include <linux/vmalloc.h>
+
+#define DM_MSG_PREFIX "cache-policy-mq"
+#define MQ_VERSION	"1.0.0"
+
+static struct kmem_cache *mq_entry_cache;
+
+/*----------------------------------------------------------------*/
+
+static unsigned next_power(unsigned n, unsigned min)
+{
+	return roundup_pow_of_two(max(n, min));
+}
+
+/*----------------------------------------------------------------*/
+
+static unsigned long *alloc_bitset(unsigned nr_entries)
+{
+	size_t s = sizeof(unsigned long) * dm_div_up(nr_entries, BITS_PER_LONG);
+	return vzalloc(s);
+}
+
+static void free_bitset(unsigned long *bits)
+{
+	vfree(bits);
+}
+
+/*----------------------------------------------------------------*/
+
+/*
+ * Large, sequential ios are probably better left on the origin device since
+ * spindles tend to have good bandwidth.
+ *
+ * The io_tracker tries to spot when the io is in one of these sequential
+ * modes.
+ *
+ * Two thresholds to switch between random and sequential io mode are defaulting
+ * as follows and can be adjusted via the constructor and message interfaces.
+ */
+#define RANDOM_THRESHOLD_DEFAULT 4
+#define SEQUENTIAL_THRESHOLD_DEFAULT 512
+
+enum io_pattern {
+	PATTERN_SEQUENTIAL,
+	PATTERN_RANDOM
+};
+
+struct io_tracker {
+	enum io_pattern pattern;
+
+	unsigned nr_seq_samples;
+	unsigned nr_rand_samples;
+	unsigned thresholds[2];
+
+	dm_oblock_t last_end_oblock;
+};
+
+static void iot_init(struct io_tracker *t,
+		     int sequential_threshold, int random_threshold)
+{
+	t->pattern = PATTERN_RANDOM;
+	t->nr_seq_samples = 0;
+	t->nr_rand_samples = 0;
+	t->last_end_oblock = 0;
+	t->thresholds[PATTERN_RANDOM] = random_threshold;
+	t->thresholds[PATTERN_SEQUENTIAL] = sequential_threshold;
+}
+
+static enum io_pattern iot_pattern(struct io_tracker *t)
+{
+	return t->pattern;
+}
+
+static void iot_update_stats(struct io_tracker *t, struct bio *bio)
+{
+	if (bio->bi_sector == from_oblock(t->last_end_oblock) + 1)
+		t->nr_seq_samples++;
+	else {
+		/*
+		 * Just one non-sequential IO is enough to reset the
+		 * counters.
+		 */
+		if (t->nr_seq_samples) {
+			t->nr_seq_samples = 0;
+			t->nr_rand_samples = 0;
+		}
+
+		t->nr_rand_samples++;
+	}
+
+	t->last_end_oblock = to_oblock(bio->bi_sector + bio_sectors(bio) - 1);
+}
+
+static void iot_check_for_pattern_switch(struct io_tracker *t)
+{
+	switch (t->pattern) {
+	case PATTERN_SEQUENTIAL:
+		if (t->nr_rand_samples >= t->thresholds[PATTERN_RANDOM]) {
+			t->pattern = PATTERN_RANDOM;
+			t->nr_seq_samples = t->nr_rand_samples = 0;
+		}
+		break;
+
+	case PATTERN_RANDOM:
+		if (t->nr_seq_samples >= t->thresholds[PATTERN_SEQUENTIAL]) {
+			t->pattern = PATTERN_SEQUENTIAL;
+			t->nr_seq_samples = t->nr_rand_samples = 0;
+		}
+		break;
+	}
+}
+
+static void iot_examine_bio(struct io_tracker *t, struct bio *bio)
+{
+	iot_update_stats(t, bio);
+	iot_check_for_pattern_switch(t);
+}
+
+/*----------------------------------------------------------------*/
+
+
+/*
+ * This queue is divided up into different levels.  Allowing us to push
+ * entries to the back of any of the levels.  Think of it as a partially
+ * sorted queue.
+ */
+#define NR_QUEUE_LEVELS 16u
+
+struct queue {
+	struct list_head qs[NR_QUEUE_LEVELS];
+};
+
+static void queue_init(struct queue *q)
+{
+	unsigned i;
+
+	for (i = 0; i < NR_QUEUE_LEVELS; i++)
+		INIT_LIST_HEAD(q->qs + i);
+}
+
+/*
+ * Insert an entry to the back of the given level.
+ */
+static void queue_push(struct queue *q, unsigned level, struct list_head *elt)
+{
+	list_add_tail(elt, q->qs + level);
+}
+
+static void queue_remove(struct list_head *elt)
+{
+	list_del(elt);
+}
+
+/*
+ * Shifts all regions down one level.  This has no effect on the order of
+ * the queue.
+ */
+static void queue_shift_down(struct queue *q)
+{
+	unsigned level;
+
+	for (level = 1; level < NR_QUEUE_LEVELS; level++)
+		list_splice_init(q->qs + level, q->qs + level - 1);
+}
+
+/*
+ * Gives us the oldest entry of the lowest popoulated level.  If the first
+ * level is emptied then we shift down one level.
+ */
+static struct list_head *queue_pop(struct queue *q)
+{
+	unsigned level;
+	struct list_head *r;
+
+	for (level = 0; level < NR_QUEUE_LEVELS; level++)
+		if (!list_empty(q->qs + level)) {
+			r = q->qs[level].next;
+			list_del(r);
+
+			/* have we just emptied the bottom level? */
+			if (level == 0 && list_empty(q->qs))
+				queue_shift_down(q);
+
+			return r;
+		}
+
+	return NULL;
+}
+
+static struct list_head *list_pop(struct list_head *lh)
+{
+	struct list_head *r = lh->next;
+
+	BUG_ON(!r);
+	list_del_init(r);
+
+	return r;
+}
+
+/*----------------------------------------------------------------*/
+
+/*
+ * Describes a cache entry.  Used in both the cache and the pre_cache.
+ */
+struct entry {
+	struct hlist_node hlist;
+	struct list_head list;
+	dm_oblock_t oblock;
+	dm_cblock_t cblock;	/* valid iff in_cache */
+
+	/*
+	 * FIXME: pack these better
+	 */
+	bool in_cache:1;
+	unsigned hit_count;
+	unsigned generation;
+	unsigned tick;
+};
+
+struct mq_policy {
+	struct dm_cache_policy policy;
+
+	/* protects everything */
+	struct mutex lock;
+	dm_cblock_t cache_size;
+	struct io_tracker tracker;
+
+	/*
+	 * We maintain two queues of entries.  The cache proper contains
+	 * the currently active mappings.  Whereas the pre_cache tracks
+	 * blocks that are being hit frequently and potential candidates
+	 * for promotion to the cache.
+	 */
+	struct queue pre_cache;
+	struct queue cache;
+
+	/*
+	 * Keeps track of time, incremented by the core.  We use this to
+	 * avoid attributing multiple hits within the same tick.
+	 *
+	 * Access to tick_protected should be done with the spin lock held.
+	 * It's copied to tick at the start of the map function (within the
+	 * mutex).
+	 */
+	spinlock_t tick_lock;
+	unsigned tick_protected;
+	unsigned tick;
+
+	/*
+	 * A count of the number of times the map function has been called
+	 * and found an entry in the pre_cache or cache.  Currently used to
+	 * calculate the generation.
+	 */
+	unsigned hit_count;
+
+	/*
+	 * A generation is a longish period that is used to trigger some
+	 * book keeping effects.  eg, decrementing hit counts on entries.
+	 * This is needed to allow the cache to evolve as io patterns
+	 * change.
+	 */
+	unsigned generation;
+	unsigned generation_period; /* in lookups (will probably change) */
+
+	/*
+	 * Entries in the pre_cache whose hit count passes the promotion
+	 * threshold move to the cache proper.  Working out the correct
+	 * value for the promotion_threshold is crucial to this policy.
+	 */
+	unsigned promote_threshold;
+
+	/*
+	 * We need cache_size entries for the cache, and choose to have
+	 * cache_size entries for the pre_cache too.  One motivation for
+	 * using the same size is to make the hit counts directly
+	 * comparable between pre_cache and cache.
+	 */
+	unsigned nr_entries;
+	unsigned nr_entries_allocated;
+	struct list_head free;
+
+	/*
+	 * Cache blocks may be unallocated.  We store this info in a
+	 * bitset.
+	 */
+	unsigned long *allocation_bitset;
+	unsigned nr_cblocks_allocated;
+	unsigned find_free_nr_words;
+	unsigned find_free_last_word;
+
+	/*
+	 * The hash table allows us to quickly find an entry by origin
+	 * block.  Both pre_cache and cache entries are in here.
+	 */
+	unsigned nr_buckets;
+	dm_block_t hash_bits;
+	struct hlist_head *table;
+};
+
+/*----------------------------------------------------------------*/
+/* Free/alloc mq cache entry structures. */
+static void takeout_queue(struct list_head *lh, struct queue *q)
+{
+	unsigned level;
+
+	for (level = 0; level < NR_QUEUE_LEVELS; level++)
+		list_splice(q->qs + level, lh);
+}
+
+static void free_entries(struct mq_policy *mq)
+{
+	struct entry *e, *tmp;
+
+	takeout_queue(&mq->free, &mq->pre_cache);
+	takeout_queue(&mq->free, &mq->cache);
+
+	list_for_each_entry_safe(e, tmp, &mq->free, list)
+		kmem_cache_free(mq_entry_cache, e);
+}
+
+static int alloc_entries(struct mq_policy *mq, unsigned elts)
+{
+	unsigned u = mq->nr_entries;
+
+	INIT_LIST_HEAD(&mq->free);
+	mq->nr_entries_allocated = 0;
+
+	while (u--) {
+		struct entry *e = kmem_cache_zalloc(mq_entry_cache, GFP_KERNEL);
+
+		if (!e) {
+			free_entries(mq);
+			return -ENOMEM;
+		}
+
+
+		list_add(&e->list, &mq->free);
+	}
+
+	return 0;
+}
+
+/*----------------------------------------------------------------*/
+
+/*
+ * Simple hash table implementation.  Should replace with the standard hash
+ * table that's making its way upstream.
+ */
+static void hash_insert(struct mq_policy *mq, struct entry *e)
+{
+	unsigned h = hash_64(from_oblock(e->oblock), mq->hash_bits);
+
+	hlist_add_head(&e->hlist, mq->table + h);
+}
+
+static struct entry *hash_lookup(struct mq_policy *mq, dm_oblock_t oblock)
+{
+	unsigned h = hash_64(from_oblock(oblock), mq->hash_bits);
+	struct hlist_head *bucket = mq->table + h;
+	struct entry *e;
+
+	hlist_for_each_entry(e, bucket, hlist)
+		if (e->oblock == oblock) {
+			hlist_del(&e->hlist);
+			hlist_add_head(&e->hlist, bucket);
+			return e;
+		}
+
+	return NULL;
+}
+
+static void hash_remove(struct entry *e)
+{
+	hlist_del(&e->hlist);
+}
+
+/*----------------------------------------------------------------*/
+
+/*
+ * Allocates a new entry structure.  The memory is allocated in one lump,
+ * so we just handing it out here.  Returns NULL if all entries have
+ * already been allocated.  Cannot fail otherwise.
+ */
+static struct entry *alloc_entry(struct mq_policy *mq)
+{
+	struct entry *e;
+
+	if (mq->nr_entries_allocated >= mq->nr_entries) {
+		BUG_ON(!list_empty(&mq->free));
+		return NULL;
+	}
+
+	e = list_entry(list_pop(&mq->free), struct entry, list);
+	INIT_LIST_HEAD(&e->list);
+	INIT_HLIST_NODE(&e->hlist);
+
+	mq->nr_entries_allocated++;
+	return e;
+}
+
+/*----------------------------------------------------------------*/
+
+/*
+ * Mark cache blocks allocated or not in the bitset.
+ */
+static void alloc_cblock(struct mq_policy *mq, dm_cblock_t cblock)
+{
+	BUG_ON(from_cblock(cblock) > from_cblock(mq->cache_size));
+	BUG_ON(test_bit(from_cblock(cblock), mq->allocation_bitset));
+
+	set_bit(from_cblock(cblock), mq->allocation_bitset);
+	mq->nr_cblocks_allocated++;
+}
+
+static void free_cblock(struct mq_policy *mq, dm_cblock_t cblock)
+{
+	BUG_ON(from_cblock(cblock) > from_cblock(mq->cache_size));
+	BUG_ON(!test_bit(from_cblock(cblock), mq->allocation_bitset));
+
+	clear_bit(from_cblock(cblock), mq->allocation_bitset);
+	mq->nr_cblocks_allocated--;
+}
+
+static bool any_free_cblocks(struct mq_policy *mq)
+{
+	return mq->nr_cblocks_allocated < from_cblock(mq->cache_size);
+}
+
+/*
+ * Fills result out with a cache block that isn't in use, or return
+ * -ENOSPC.  This does _not_ mark the cblock as allocated, the caller is
+ * reponsible for that.
+ */
+static int __find_free_cblock(struct mq_policy *mq, unsigned begin, unsigned end,
+			      dm_cblock_t *result, unsigned *last_word)
+{
+	int r = -ENOSPC;
+	unsigned w;
+
+	for (w = begin; w < end; w++) {
+		/*
+		 * ffz is undefined if no zero exists
+		 */
+		if (mq->allocation_bitset[w] != ~0UL) {
+			*last_word = w;
+			*result = to_cblock((w * BITS_PER_LONG) + ffz(mq->allocation_bitset[w]));
+			if (from_cblock(*result) < from_cblock(mq->cache_size))
+				r = 0;
+
+			break;
+		}
+	}
+
+	return r;
+}
+
+static int find_free_cblock(struct mq_policy *mq, dm_cblock_t *result)
+{
+	int r;
+
+	if (!any_free_cblocks(mq))
+		return -ENOSPC;
+
+	r = __find_free_cblock(mq, mq->find_free_last_word, mq->find_free_nr_words, result, &mq->find_free_last_word);
+	if (r == -ENOSPC && mq->find_free_last_word)
+		r = __find_free_cblock(mq, 0, mq->find_free_last_word, result, &mq->find_free_last_word);
+
+	return r;
+}
+
+/*----------------------------------------------------------------*/
+
+/*
+ * Now we get to the meat of the policy.  This section deals with deciding
+ * when to to add entries to the pre_cache and cache, and move between
+ * them.
+ */
+
+/*
+ * The queue level is based on the log2 of the hit count.
+ */
+static unsigned queue_level(struct entry *e)
+{
+	return min((unsigned) ilog2(e->hit_count), NR_QUEUE_LEVELS - 1u);
+}
+
+/*
+ * Inserts the entry into the pre_cache or the cache.  Ensures the cache
+ * block is marked as allocated if necc.  Inserts into the hash table.  Sets the
+ * tick which records when the entry was last moved about.
+ */
+static void push(struct mq_policy *mq, struct entry *e)
+{
+	e->tick = mq->tick;
+	hash_insert(mq, e);
+
+	if (e->in_cache) {
+		alloc_cblock(mq, e->cblock);
+		queue_push(&mq->cache, queue_level(e), &e->list);
+	} else
+		queue_push(&mq->pre_cache, queue_level(e), &e->list);
+}
+
+/*
+ * Removes an entry from pre_cache or cache.  Removes from the hash table.
+ * Frees off the cache block if necc.
+ */
+static void del(struct mq_policy *mq, struct entry *e)
+{
+	queue_remove(&e->list);
+	hash_remove(e);
+	if (e->in_cache)
+		free_cblock(mq, e->cblock);
+}
+
+/*
+ * Like del, except it removes the first entry in the queue (ie. the least
+ * recently used).
+ */
+static struct entry *pop(struct mq_policy *mq, struct queue *q)
+{
+	struct entry *e = container_of(queue_pop(q), struct entry, list);
+
+	if (e) {
+		hash_remove(e);
+
+		if (e->in_cache)
+			free_cblock(mq, e->cblock);
+	}
+
+	return e;
+}
+
+/*
+ * Has this entry already been updated?
+ */
+static bool updated_this_tick(struct mq_policy *mq, struct entry *e)
+{
+	return mq->tick == e->tick;
+}
+
+/*
+ * The promotion threshold is adjusted every generation.  As are the counts
+ * of the entries.
+ *
+ * At the moment the threshold is taken by averaging the hit counts of some
+ * of the entries in the cache (the first 20 entries of the first level).
+ *
+ * We can be much cleverer than this though.  For example, each promotion
+ * could bump up the threshold helping to prevent churn.  Much more to do
+ * here.
+ */
+
+#define MAX_TO_AVERAGE 20
+
+static void check_generation(struct mq_policy *mq)
+{
+	unsigned total = 0, nr = 0, count = 0, level;
+	struct list_head *head;
+	struct entry *e;
+
+	if ((mq->hit_count >= mq->generation_period) &&
+	    (mq->nr_cblocks_allocated == from_cblock(mq->cache_size))) {
+
+		mq->hit_count = 0;
+		mq->generation++;
+
+		for (level = 0; level < NR_QUEUE_LEVELS && count < MAX_TO_AVERAGE; level++) {
+			head = mq->cache.qs + level;
+			list_for_each_entry(e, head, list) {
+				nr++;
+				total += e->hit_count;
+
+				if (++count >= MAX_TO_AVERAGE)
+					break;
+			}
+		}
+
+		mq->promote_threshold = nr ? total / nr : 1;
+		if (mq->promote_threshold * nr < total)
+			mq->promote_threshold++;
+	}
+}
+
+/*
+ * Whenever we use an entry we bump up it's hit counter, and push it to the
+ * back to it's current level.
+ */
+static void requeue_and_update_tick(struct mq_policy *mq, struct entry *e)
+{
+	if (updated_this_tick(mq, e))
+		return;
+
+	e->hit_count++;
+	mq->hit_count++;
+	check_generation(mq);
+
+	/* generation adjustment, to stop the counts increasing forever. */
+	/* FIXME: divide? */
+	/* e->hit_count -= min(e->hit_count - 1, mq->generation - e->generation); */
+	e->generation = mq->generation;
+
+	del(mq, e);
+	push(mq, e);
+}
+
+/*
+ * Demote the least recently used entry from the cache to the pre_cache.
+ * Returns the new cache entry to use, and the old origin block it was
+ * mapped to.
+ *
+ * We drop the hit count on the demoted entry back to 1 to stop it bouncing
+ * straight back into the cache if it's subsequently hit.  There are
+ * various options here, and more experimentation would be good:
+ *
+ * - just forget about the demoted entry completely (ie. don't insert it
+     into the pre_cache).
+ * - divide the hit count rather that setting to some hard coded value.
+ * - set the hit count to a hard coded value other than 1, eg, is it better
+ *   if it goes in at level 2?
+ */
+static dm_cblock_t demote_cblock(struct mq_policy *mq, dm_oblock_t *oblock)
+{
+	dm_cblock_t result;
+	struct entry *demoted = pop(mq, &mq->cache);
+
+	BUG_ON(!demoted);
+	result = demoted->cblock;
+	*oblock = demoted->oblock;
+	demoted->in_cache = false;
+	demoted->hit_count = 1;
+	push(mq, demoted);
+
+	return result;
+}
+
+/*
+ * We modify the basic promotion_threshold depending on the specific io.
+ *
+ * If the origin block has been discarded then there's no cost to copy it
+ * to the cache.
+ *
+ * We bias towards reads, since they can be demoted at no cost if they
+ * haven't been dirtied.
+ */
+#define DISCARDED_PROMOTE_THRESHOLD 1
+#define READ_PROMOTE_THRESHOLD 4
+#define WRITE_PROMOTE_THRESHOLD 8
+
+static unsigned adjusted_promote_threshold(struct mq_policy *mq,
+					   bool discarded_oblock, int data_dir)
+{
+	if (discarded_oblock && any_free_cblocks(mq) && data_dir == WRITE)
+		/*
+		 * We don't need to do any copying at all, so give this a
+		 * very low threshold.  In practice this only triggers
+		 * during initial population after a format.
+		 */
+		return DISCARDED_PROMOTE_THRESHOLD;
+
+	return data_dir == READ ?
+		(mq->promote_threshold + READ_PROMOTE_THRESHOLD) :
+		(mq->promote_threshold + WRITE_PROMOTE_THRESHOLD);
+}
+
+static bool should_promote(struct mq_policy *mq, struct entry *e,
+			   bool discarded_oblock, int data_dir)
+{
+	return e->hit_count >=
+		adjusted_promote_threshold(mq, discarded_oblock, data_dir);
+}
+
+static int cache_entry_found(struct mq_policy *mq,
+			     struct entry *e,
+			     struct policy_result *result)
+{
+	requeue_and_update_tick(mq, e);
+
+	if (e->in_cache) {
+		result->op = POLICY_HIT;
+		result->cblock = e->cblock;
+	}
+
+	return 0;
+}
+
+/*
+ * Moves and entry from the pre_cache to the cache.  The main work is
+ * finding which cache block to use.
+ */
+static int pre_cache_to_cache(struct mq_policy *mq, struct entry *e,
+			      struct policy_result *result)
+{
+	dm_cblock_t cblock;
+
+	if (find_free_cblock(mq, &cblock) == -ENOSPC) {
+		result->op = POLICY_REPLACE;
+		cblock = demote_cblock(mq, &result->old_oblock);
+	} else
+		result->op = POLICY_NEW;
+
+	result->cblock = e->cblock = cblock;
+
+	del(mq, e);
+	e->in_cache = true;
+	push(mq, e);
+
+	return 0;
+}
+
+static int pre_cache_entry_found(struct mq_policy *mq, struct entry *e,
+				 bool can_migrate, bool discarded_oblock,
+				 int data_dir, struct policy_result *result)
+{
+	int r = 0;
+	bool updated = updated_this_tick(mq, e);
+
+	requeue_and_update_tick(mq, e);
+
+	if ((!discarded_oblock && updated) ||
+	    !should_promote(mq, e, discarded_oblock, data_dir))
+		result->op = POLICY_MISS;
+	else if (!can_migrate)
+		r = -EWOULDBLOCK;
+	else
+		r = pre_cache_to_cache(mq, e, result);
+
+	return r;
+}
+
+static void insert_in_pre_cache(struct mq_policy *mq,
+				dm_oblock_t oblock)
+{
+	struct entry *e = alloc_entry(mq);
+
+	if (!e)
+		/*
+		 * There's no spare entry structure, so we grab the least
+		 * used one from the pre_cache.
+		 */
+		e = pop(mq, &mq->pre_cache);
+
+	if (unlikely(!e)) {
+		DMWARN("couldn't pop from pre cache");
+		return;
+	}
+
+	e->in_cache = false;
+	e->oblock = oblock;
+	e->hit_count = 1;
+	e->generation = mq->generation;
+	push(mq, e);
+}
+
+static void insert_in_cache(struct mq_policy *mq, dm_oblock_t oblock,
+			    struct policy_result *result)
+{
+	struct entry *e;
+	dm_cblock_t cblock;
+
+	if (find_free_cblock(mq, &cblock) == -ENOSPC) {
+		result->op = POLICY_MISS;
+		insert_in_pre_cache(mq, oblock);
+		return;
+	}
+
+	e = alloc_entry(mq);
+	if (unlikely(!e)) {
+		result->op = POLICY_MISS;
+		return;
+	}
+
+	e->oblock = oblock;
+	e->cblock = cblock;
+	e->in_cache = true;
+	e->hit_count = 1;
+	e->generation = mq->generation;
+	push(mq, e);
+
+	result->op = POLICY_NEW;
+	result->cblock = e->cblock;
+}
+
+static int no_entry_found(struct mq_policy *mq, dm_oblock_t oblock,
+			  bool can_migrate, bool discarded_oblock,
+			  int data_dir, struct policy_result *result)
+{
+	if (adjusted_promote_threshold(mq, discarded_oblock, data_dir) == 1) {
+		if (can_migrate)
+			insert_in_cache(mq, oblock, result);
+		else
+			return -EWOULDBLOCK;
+	} else {
+		insert_in_pre_cache(mq, oblock);
+		result->op = POLICY_MISS;
+	}
+
+	return 0;
+}
+
+/*
+ * Looks the oblock up in the hash table, then decides whether to put in
+ * pre_cache, or cache etc.
+ */
+static int map(struct mq_policy *mq, dm_oblock_t oblock,
+	       bool can_migrate, bool discarded_oblock,
+	       int data_dir, struct policy_result *result)
+{
+	int r = 0;
+	struct entry *e = hash_lookup(mq, oblock);
+
+	if (e && e->in_cache)
+		r = cache_entry_found(mq, e, result);
+	else if (iot_pattern(&mq->tracker) == PATTERN_SEQUENTIAL)
+		result->op = POLICY_MISS;
+	else if (e)
+		r = pre_cache_entry_found(mq, e, can_migrate, discarded_oblock,
+					  data_dir, result);
+	else
+		r = no_entry_found(mq, oblock, can_migrate, discarded_oblock,
+				   data_dir, result);
+
+	if (r == -EWOULDBLOCK)
+		result->op = POLICY_MISS;
+
+	return r;
+}
+
+/*----------------------------------------------------------------*/
+
+/*
+ * Public interface, via the policy struct.  See dm-cache-policy.h for a
+ * description of these.
+ */
+
+static struct mq_policy *to_mq_policy(struct dm_cache_policy *p)
+{
+	return container_of(p, struct mq_policy, policy);
+}
+
+static void mq_destroy(struct dm_cache_policy *p)
+{
+	struct mq_policy *mq = to_mq_policy(p);
+
+	free_bitset(mq->allocation_bitset);
+	kfree(mq->table);
+	free_entries(mq);
+	kfree(mq);
+}
+
+static void copy_tick(struct mq_policy *mq)
+{
+	unsigned long flags;
+
+	spin_lock_irqsave(&mq->tick_lock, flags);
+	mq->tick = mq->tick_protected;
+	spin_unlock_irqrestore(&mq->tick_lock, flags);
+}
+
+static int mq_map(struct dm_cache_policy *p, dm_oblock_t oblock,
+		  bool can_block, bool can_migrate, bool discarded_oblock,
+		  struct bio *bio, struct policy_result *result)
+{
+	int r;
+	struct mq_policy *mq = to_mq_policy(p);
+
+	result->op = POLICY_MISS;
+
+	if (can_block)
+		mutex_lock(&mq->lock);
+	else if (!mutex_trylock(&mq->lock))
+		return -EWOULDBLOCK;
+
+	copy_tick(mq);
+
+	iot_examine_bio(&mq->tracker, bio);
+	r = map(mq, oblock, can_migrate, discarded_oblock,
+		bio_data_dir(bio), result);
+
+	mutex_unlock(&mq->lock);
+
+	return r;
+}
+
+static int mq_lookup(struct dm_cache_policy *p, dm_oblock_t oblock, dm_cblock_t *cblock)
+{
+	int r;
+	struct mq_policy *mq = to_mq_policy(p);
+	struct entry *e;
+
+	if (!mutex_trylock(&mq->lock))
+		return -EWOULDBLOCK;
+
+	e = hash_lookup(mq, oblock);
+	if (e && e->in_cache) {
+		*cblock = e->cblock;
+		r = 0;
+	} else
+		r = -ENOENT;
+
+	mutex_unlock(&mq->lock);
+
+	return r;
+}
+
+static int mq_load_mapping(struct dm_cache_policy *p,
+			   dm_oblock_t oblock, dm_cblock_t cblock,
+			   uint32_t hint, bool hint_valid)
+{
+	struct mq_policy *mq = to_mq_policy(p);
+	struct entry *e;
+
+	e = alloc_entry(mq);
+	if (!e)
+		return -ENOMEM;
+
+	e->cblock = cblock;
+	e->oblock = oblock;
+	e->in_cache = true;
+	e->hit_count = hint_valid ? hint : 1;
+	e->generation = mq->generation;
+	push(mq, e);
+
+	return 0;
+}
+
+static int mq_walk_mappings(struct dm_cache_policy *p, policy_walk_fn fn,
+			    void *context)
+{
+	struct mq_policy *mq = to_mq_policy(p);
+	int r = 0;
+	struct entry *e;
+	unsigned level;
+
+	mutex_lock(&mq->lock);
+
+	for (level = 0; level < NR_QUEUE_LEVELS; level++)
+		list_for_each_entry(e, &mq->cache.qs[level], list) {
+			r = fn(context, e->cblock, e->oblock, e->hit_count);
+			if (r)
+				goto out;
+		}
+
+out:
+	mutex_unlock(&mq->lock);
+
+	return r;
+}
+
+static void remove_mapping(struct mq_policy *mq, dm_oblock_t oblock)
+{
+	struct entry *e = hash_lookup(mq, oblock);
+
+	BUG_ON(!e || !e->in_cache);
+
+	del(mq, e);
+	e->in_cache = false;
+	push(mq, e);
+}
+
+static void mq_remove_mapping(struct dm_cache_policy *p, dm_oblock_t oblock)
+{
+	struct mq_policy *mq = to_mq_policy(p);
+
+	mutex_lock(&mq->lock);
+	remove_mapping(mq, oblock);
+	mutex_unlock(&mq->lock);
+}
+
+static void force_mapping(struct mq_policy *mq,
+			  dm_oblock_t current_oblock, dm_oblock_t new_oblock)
+{
+	struct entry *e = hash_lookup(mq, current_oblock);
+
+	BUG_ON(!e || !e->in_cache);
+
+	del(mq, e);
+	e->oblock = new_oblock;
+	push(mq, e);
+}
+
+static void mq_force_mapping(struct dm_cache_policy *p,
+			     dm_oblock_t current_oblock, dm_oblock_t new_oblock)
+{
+	struct mq_policy *mq = to_mq_policy(p);
+
+	mutex_lock(&mq->lock);
+	force_mapping(mq, current_oblock, new_oblock);
+	mutex_unlock(&mq->lock);
+}
+
+static dm_cblock_t mq_residency(struct dm_cache_policy *p)
+{
+	struct mq_policy *mq = to_mq_policy(p);
+
+	/* FIXME: lock mutex, not sure we can block here */
+	return to_cblock(mq->nr_cblocks_allocated);
+}
+
+static void mq_tick(struct dm_cache_policy *p)
+{
+	struct mq_policy *mq = to_mq_policy(p);
+	unsigned long flags;
+
+	spin_lock_irqsave(&mq->tick_lock, flags);
+	mq->tick_protected++;
+	spin_unlock_irqrestore(&mq->tick_lock, flags);
+}
+
+static int mq_set_config_value(struct dm_cache_policy *p,
+			       const char *key, const char *value)
+{
+	struct mq_policy *mq = to_mq_policy(p);
+	enum io_pattern pattern;
+	unsigned long tmp;
+
+	if (!strcasecmp(key, "random_threshold"))
+		pattern = PATTERN_RANDOM;
+	else if (!strcasecmp(key, "sequential_threshold"))
+		pattern = PATTERN_SEQUENTIAL;
+	else
+		return -EINVAL;
+
+	if (kstrtoul(value, 10, &tmp))
+		return -EINVAL;
+
+	mq->tracker.thresholds[pattern] = tmp;
+
+	return 0;
+}
+
+static int mq_emit_config_values(struct dm_cache_policy *p, char *result, unsigned maxlen)
+{
+	ssize_t sz = 0;
+	struct mq_policy *mq = to_mq_policy(p);
+
+	DMEMIT("4 random_threshold %u sequential_threshold %u",
+	       mq->tracker.thresholds[PATTERN_RANDOM],
+	       mq->tracker.thresholds[PATTERN_SEQUENTIAL]);
+
+	return 0;
+}
+
+/* Init the policy plugin interface function pointers. */
+static void init_policy_functions(struct mq_policy *mq)
+{
+	mq->policy.destroy = mq_destroy;
+	mq->policy.map = mq_map;
+	mq->policy.lookup = mq_lookup;
+	mq->policy.load_mapping = mq_load_mapping;
+	mq->policy.walk_mappings = mq_walk_mappings;
+	mq->policy.remove_mapping = mq_remove_mapping;
+	mq->policy.writeback_work = NULL;
+	mq->policy.force_mapping = mq_force_mapping;
+	mq->policy.residency = mq_residency;
+	mq->policy.tick = mq_tick;
+	mq->policy.emit_config_values = mq_emit_config_values;
+	mq->policy.set_config_value = mq_set_config_value;
+}
+
+static struct dm_cache_policy *mq_create(dm_cblock_t cache_size,
+					 sector_t origin_size,
+					 sector_t cache_block_size)
+{
+	int r;
+	struct mq_policy *mq = kzalloc(sizeof(*mq), GFP_KERNEL);
+
+	if (!mq)
+		return NULL;
+
+	init_policy_functions(mq);
+	iot_init(&mq->tracker, SEQUENTIAL_THRESHOLD_DEFAULT, RANDOM_THRESHOLD_DEFAULT);
+
+	mq->cache_size = cache_size;
+	mq->tick_protected = 0;
+	mq->tick = 0;
+	mq->hit_count = 0;
+	mq->generation = 0;
+	mq->promote_threshold = 0;
+	mutex_init(&mq->lock);
+	spin_lock_init(&mq->tick_lock);
+	mq->find_free_nr_words = dm_div_up(from_cblock(mq->cache_size), BITS_PER_LONG);
+	mq->find_free_last_word = 0;
+
+	queue_init(&mq->pre_cache);
+	queue_init(&mq->cache);
+	mq->generation_period = max((unsigned) from_cblock(cache_size), 1024U);
+
+	mq->nr_entries = 2 * from_cblock(cache_size);
+	r = alloc_entries(mq, mq->nr_entries);
+	if (r)
+		goto bad_cache_alloc;
+
+	mq->nr_entries_allocated = 0;
+	mq->nr_cblocks_allocated = 0;
+
+	mq->nr_buckets = next_power(from_cblock(cache_size) / 2, 16);
+	mq->hash_bits = ffs(mq->nr_buckets) - 1;
+	mq->table = kzalloc(sizeof(*mq->table) * mq->nr_buckets, GFP_KERNEL);
+	if (!mq->table)
+		goto bad_alloc_table;
+
+	mq->allocation_bitset = alloc_bitset(from_cblock(cache_size));
+	if (!mq->allocation_bitset)
+		goto bad_alloc_bitset;
+
+	return &mq->policy;
+
+bad_alloc_bitset:
+	kfree(mq->table);
+bad_alloc_table:
+	free_entries(mq);
+bad_cache_alloc:
+	kfree(mq);
+
+	return NULL;
+}
+
+/*----------------------------------------------------------------*/
+
+static struct dm_cache_policy_type mq_policy_type = {
+	.name = "mq",
+	.hint_size = 4,
+	.owner = THIS_MODULE,
+	.create = mq_create
+};
+
+static struct dm_cache_policy_type default_policy_type = {
+	.name = "default",
+	.hint_size = 4,
+	.owner = THIS_MODULE,
+	.create = mq_create
+};
+
+static int __init mq_init(void)
+{
+	int r;
+
+	mq_entry_cache = kmem_cache_create("dm_mq_policy_cache_entry",
+					   sizeof(struct entry),
+					   __alignof__(struct entry),
+					   0, NULL);
+	if (!mq_entry_cache)
+		goto bad;
+
+	r = dm_cache_policy_register(&mq_policy_type);
+	if (r) {
+		DMERR("register failed %d", r);
+		goto bad_register_mq;
+	}
+
+	r = dm_cache_policy_register(&default_policy_type);
+	if (!r) {
+		DMINFO("version " MQ_VERSION " loaded");
+		return 0;
+	}
+
+	DMERR("register failed (as default) %d", r);
+
+	dm_cache_policy_unregister(&mq_policy_type);
+bad_register_mq:
+	kmem_cache_destroy(mq_entry_cache);
+bad:
+	return -ENOMEM;
+}
+
+static void __exit mq_exit(void)
+{
+	dm_cache_policy_unregister(&mq_policy_type);
+	dm_cache_policy_unregister(&default_policy_type);
+
+	kmem_cache_destroy(mq_entry_cache);
+}
+
+module_init(mq_init);
+module_exit(mq_exit);
+
+MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
+MODULE_LICENSE("GPL");
+MODULE_DESCRIPTION("mq cache policy");
+
+MODULE_ALIAS("dm-cache-default");