Merge tag 'dm-3.12-changes' of git://git.kernel.org/pub/scm/linux/kernel/git/device-mapper/linux-dm

Pull device-mapper updates from Mike Snitzer:
 "Add the ability to collect I/O statistics on user-defined regions of a
  device-mapper device.  This dm-stats code required the reintroduction
  of a div64_u64_rem() helper, but as a separate method that doesn't
  slow down div64_u64() -- especially on 32-bit systems.

  Allow the error target to replace request-based DM devices (e.g.
  multipath) in addition to bio-based DM devices.

  Various other small code fixes and improvements to thin-provisioning,
  DM cache and the DM ioctl interface"

* tag 'dm-3.12-changes' of git://git.kernel.org/pub/scm/linux/kernel/git/device-mapper/linux-dm:
  dm stripe: silence a couple sparse warnings
  dm: add statistics support
  dm thin: always return -ENOSPC if no_free_space is set
  dm ioctl: cleanup error handling in table_load
  dm ioctl: increase granularity of type_lock when loading table
  dm ioctl: prevent rename to empty name or uuid
  dm thin: set pool read-only if breaking_sharing fails block allocation
  dm thin: prefix pool error messages with pool device name
  dm: allow error target to replace bio-based and request-based targets
  math64: New separate div64_u64_rem helper
  dm space map: optimise sm_ll_dec and sm_ll_inc
  dm btree: prefetch child nodes when walking tree for a dm_btree_del
  dm btree: use pop_frame in dm_btree_del to cleanup code
  dm cache: eliminate holes in cache structure
  dm cache: fix stacking of geometry limits
  dm thin: fix stacking of geometry limits
  dm thin: add data block size limits to Documentation
  dm cache: add data block size limits to code and Documentation
  dm cache: document metadata device is exclussive to a cache
  dm: stop using WQ_NON_REENTRANT

18 files changed, 1359 insertions, 151 deletions

diff --git a/drivers/md/Makefile b/drivers/md/Makefile
index 5ef78efc27f2..2acc43fe0229 100644
--- a/drivers/md/Makefile
+++ b/drivers/md/Makefile
@@ -3,7 +3,7 @@
 #
 
 dm-mod-y	+= dm.o dm-table.o dm-target.o dm-linear.o dm-stripe.o \
-		   dm-ioctl.o dm-io.o dm-kcopyd.o dm-sysfs.o
+		   dm-ioctl.o dm-io.o dm-kcopyd.o dm-sysfs.o dm-stats.o
 dm-multipath-y	+= dm-path-selector.o dm-mpath.o
 dm-snapshot-y	+= dm-snap.o dm-exception-store.o dm-snap-transient.o \
 		    dm-snap-persistent.o
diff --git a/drivers/md/dm-cache-target.c b/drivers/md/dm-cache-target.c
index 0df3ec085ebb..29569768ffbf 100644
--- a/drivers/md/dm-cache-target.c
+++ b/drivers/md/dm-cache-target.c
@@ -67,9 +67,11 @@ static void free_bitset(unsigned long *bits)
 #define MIGRATION_COUNT_WINDOW 10
 
 /*
- * The block size of the device holding cache data must be >= 32KB
+ * The block size of the device holding cache data must be
+ * between 32KB and 1GB.
  */
 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
+#define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
 
 /*
  * FIXME: the cache is read/write for the time being.
@@ -101,6 +103,8 @@ struct cache {
 	struct dm_target *ti;
 	struct dm_target_callbacks callbacks;
 
+	struct dm_cache_metadata *cmd;
+
 	/*
 	 * Metadata is written to this device.
 	 */
@@ -117,11 +121,6 @@ struct cache {
 	struct dm_dev *cache_dev;
 
 	/*
-	 * Cache features such as write-through.
-	 */
-	struct cache_features features;
-
-	/*
 	 * Size of the origin device in _complete_ blocks and native sectors.
 	 */
 	dm_oblock_t origin_blocks;
@@ -138,8 +137,6 @@ struct cache {
 	uint32_t sectors_per_block;
 	int sectors_per_block_shift;
 
-	struct dm_cache_metadata *cmd;
-
 	spinlock_t lock;
 	struct bio_list deferred_bios;
 	struct bio_list deferred_flush_bios;
@@ -148,8 +145,8 @@ struct cache {
 	struct list_head completed_migrations;
 	struct list_head need_commit_migrations;
 	sector_t migration_threshold;
-	atomic_t nr_migrations;
 	wait_queue_head_t migration_wait;
+	atomic_t nr_migrations;
 
 	/*
 	 * cache_size entries, dirty if set
@@ -160,9 +157,16 @@ struct cache {
 	/*
 	 * origin_blocks entries, discarded if set.
 	 */
-	uint32_t discard_block_size; /* a power of 2 times sectors per block */
 	dm_dblock_t discard_nr_blocks;
 	unsigned long *discard_bitset;
+	uint32_t discard_block_size; /* a power of 2 times sectors per block */
+
+	/*
+	 * Rather than reconstructing the table line for the status we just
+	 * save it and regurgitate.
+	 */
+	unsigned nr_ctr_args;
+	const char **ctr_args;
 
 	struct dm_kcopyd_client *copier;
 	struct workqueue_struct *wq;
@@ -187,14 +191,12 @@ struct cache {
 	bool loaded_mappings:1;
 	bool loaded_discards:1;
 
-	struct cache_stats stats;
-
 	/*
-	 * Rather than reconstructing the table line for the status we just
-	 * save it and regurgitate.
+	 * Cache features such as write-through.
 	 */
-	unsigned nr_ctr_args;
-	const char **ctr_args;
+	struct cache_features features;
+
+	struct cache_stats stats;
 };
 
 struct per_bio_data {
@@ -1687,24 +1689,25 @@ static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
 static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
 			    char **error)
 {
-	unsigned long tmp;
+	unsigned long block_size;
 
 	if (!at_least_one_arg(as, error))
 		return -EINVAL;
 
-	if (kstrtoul(dm_shift_arg(as), 10, &tmp) || !tmp ||
-	    tmp < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
-	    tmp & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
+	if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
+	    block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
+	    block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
+	    block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
 		*error = "Invalid data block size";
 		return -EINVAL;
 	}
 
-	if (tmp > ca->cache_sectors) {
+	if (block_size > ca->cache_sectors) {
 		*error = "Data block size is larger than the cache device";
 		return -EINVAL;
 	}
 
-	ca->block_size = tmp;
+	ca->block_size = block_size;
 
 	return 0;
 }
@@ -2609,9 +2612,17 @@ static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
 static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
 {
 	struct cache *cache = ti->private;
+	uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
 
-	blk_limits_io_min(limits, 0);
-	blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
+	/*
+	 * If the system-determined stacked limits are compatible with the
+	 * cache's blocksize (io_opt is a factor) do not override them.
+	 */
+	if (io_opt_sectors < cache->sectors_per_block ||
+	    do_div(io_opt_sectors, cache->sectors_per_block)) {
+		blk_limits_io_min(limits, 0);
+		blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
+	}
 	set_discard_limits(cache, limits);
 }
 
diff --git a/drivers/md/dm-crypt.c b/drivers/md/dm-crypt.c
index 6d2d41ae9e32..0fce0bc1a957 100644
--- a/drivers/md/dm-crypt.c
+++ b/drivers/md/dm-crypt.c
@@ -1645,20 +1645,14 @@ static int crypt_ctr(struct dm_target *ti, unsigned int argc, char **argv)
 	}
 
 	ret = -ENOMEM;
-	cc->io_queue = alloc_workqueue("kcryptd_io",
-				       WQ_NON_REENTRANT|
-				       WQ_MEM_RECLAIM,
-				       1);
+	cc->io_queue = alloc_workqueue("kcryptd_io", WQ_MEM_RECLAIM, 1);
 	if (!cc->io_queue) {
 		ti->error = "Couldn't create kcryptd io queue";
 		goto bad;
 	}
 
 	cc->crypt_queue = alloc_workqueue("kcryptd",
-					  WQ_NON_REENTRANT|
-					  WQ_CPU_INTENSIVE|
-					  WQ_MEM_RECLAIM,
-					  1);
+					  WQ_CPU_INTENSIVE | WQ_MEM_RECLAIM, 1);
 	if (!cc->crypt_queue) {
 		ti->error = "Couldn't create kcryptd queue";
 		goto bad;
diff --git a/drivers/md/dm-ioctl.c b/drivers/md/dm-ioctl.c
index f1b758675ec7..afe08146f73e 100644
--- a/drivers/md/dm-ioctl.c
+++ b/drivers/md/dm-ioctl.c
@@ -877,7 +877,7 @@ static int dev_rename(struct dm_ioctl *param, size_t param_size)
 	unsigned change_uuid = (param->flags & DM_UUID_FLAG) ? 1 : 0;
 
 	if (new_data < param->data ||
-	    invalid_str(new_data, (void *) param + param_size) ||
+	    invalid_str(new_data, (void *) param + param_size) || !*new_data ||
 	    strlen(new_data) > (change_uuid ? DM_UUID_LEN - 1 : DM_NAME_LEN - 1)) {
 		DMWARN("Invalid new mapped device name or uuid string supplied.");
 		return -EINVAL;
@@ -1262,44 +1262,37 @@ static int table_load(struct dm_ioctl *param, size_t param_size)
 
 	r = dm_table_create(&t, get_mode(param), param->target_count, md);
 	if (r)
-		goto out;
+		goto err;
 
+	/* Protect md->type and md->queue against concurrent table loads. */
+	dm_lock_md_type(md);
 	r = populate_table(t, param, param_size);
-	if (r) {
-		dm_table_destroy(t);
-		goto out;
-	}
+	if (r)
+		goto err_unlock_md_type;
 
 	immutable_target_type = dm_get_immutable_target_type(md);
 	if (immutable_target_type &&
 	    (immutable_target_type != dm_table_get_immutable_target_type(t))) {
 		DMWARN("can't replace immutable target type %s",
 		       immutable_target_type->name);
-		dm_table_destroy(t);
 		r = -EINVAL;
-		goto out;
+		goto err_unlock_md_type;
 	}
 
-	/* Protect md->type and md->queue against concurrent table loads. */
-	dm_lock_md_type(md);
 	if (dm_get_md_type(md) == DM_TYPE_NONE)
 		/* Initial table load: acquire type of table. */
 		dm_set_md_type(md, dm_table_get_type(t));
 	else if (dm_get_md_type(md) != dm_table_get_type(t)) {
 		DMWARN("can't change device type after initial table load.");
-		dm_table_destroy(t);
-		dm_unlock_md_type(md);
 		r = -EINVAL;
-		goto out;
+		goto err_unlock_md_type;
 	}
 
 	/* setup md->queue to reflect md's type (may block) */
 	r = dm_setup_md_queue(md);
 	if (r) {
 		DMWARN("unable to set up device queue for new table.");
-		dm_table_destroy(t);
-		dm_unlock_md_type(md);
-		goto out;
+		goto err_unlock_md_type;
 	}
 	dm_unlock_md_type(md);
 
@@ -1309,9 +1302,8 @@ static int table_load(struct dm_ioctl *param, size_t param_size)
 	if (!hc || hc->md != md) {
 		DMWARN("device has been removed from the dev hash table.");
 		up_write(&_hash_lock);
-		dm_table_destroy(t);
 		r = -ENXIO;
-		goto out;
+		goto err_destroy_table;
 	}
 
 	if (hc->new_map)
@@ -1322,7 +1314,6 @@ static int table_load(struct dm_ioctl *param, size_t param_size)
 	param->flags |= DM_INACTIVE_PRESENT_FLAG;
 	__dev_status(md, param);
 
-out:
 	if (old_map) {
 		dm_sync_table(md);
 		dm_table_destroy(old_map);
@@ -1330,6 +1321,15 @@ out:
 
 	dm_put(md);
 
+	return 0;
+
+err_unlock_md_type:
+	dm_unlock_md_type(md);
+err_destroy_table:
+	dm_table_destroy(t);
+err:
+	dm_put(md);
+
 	return r;
 }
 
@@ -1455,20 +1455,26 @@ static int table_status(struct dm_ioctl *param, size_t param_size)
 	return 0;
 }
 
-static bool buffer_test_overflow(char *result, unsigned maxlen)
-{
-	return !maxlen || strlen(result) + 1 >= maxlen;
-}
-
 /*
- * Process device-mapper dependent messages.
+ * Process device-mapper dependent messages.  Messages prefixed with '@'
+ * are processed by the DM core.  All others are delivered to the target.
  * Returns a number <= 1 if message was processed by device mapper.
  * Returns 2 if message should be delivered to the target.
  */
 static int message_for_md(struct mapped_device *md, unsigned argc, char **argv,
 			  char *result, unsigned maxlen)
 {
-	return 2;
+	int r;
+
+	if (**argv != '@')
+		return 2; /* no '@' prefix, deliver to target */
+
+	r = dm_stats_message(md, argc, argv, result, maxlen);
+	if (r < 2)
+		return r;
+
+	DMERR("Unsupported message sent to DM core: %s", argv[0]);
+	return -EINVAL;
 }
 
 /*
@@ -1542,7 +1548,7 @@ static int target_message(struct dm_ioctl *param, size_t param_size)
 
 	if (r == 1) {
 		param->flags |= DM_DATA_OUT_FLAG;
-		if (buffer_test_overflow(result, maxlen))
+		if (dm_message_test_buffer_overflow(result, maxlen))
 			param->flags |= DM_BUFFER_FULL_FLAG;
 		else
 			param->data_size = param->data_start + strlen(result) + 1;
diff --git a/drivers/md/dm-kcopyd.c b/drivers/md/dm-kcopyd.c
index d581fe5d2faf..3a7cade5e27d 100644
--- a/drivers/md/dm-kcopyd.c
+++ b/drivers/md/dm-kcopyd.c
@@ -833,8 +833,7 @@ struct dm_kcopyd_client *dm_kcopyd_client_create(struct dm_kcopyd_throttle *thro
 		goto bad_slab;
 
 	INIT_WORK(&kc->kcopyd_work, do_work);
-	kc->kcopyd_wq = alloc_workqueue("kcopyd",
-					WQ_NON_REENTRANT | WQ_MEM_RECLAIM, 0);
+	kc->kcopyd_wq = alloc_workqueue("kcopyd", WQ_MEM_RECLAIM, 0);
 	if (!kc->kcopyd_wq)
 		goto bad_workqueue;
 
diff --git a/drivers/md/dm-raid1.c b/drivers/md/dm-raid1.c
index 699b5be68d31..9584443c5614 100644
--- a/drivers/md/dm-raid1.c
+++ b/drivers/md/dm-raid1.c
@@ -1080,8 +1080,7 @@ static int mirror_ctr(struct dm_target *ti, unsigned int argc, char **argv)
 	ti->per_bio_data_size = sizeof(struct dm_raid1_bio_record);
 	ti->discard_zeroes_data_unsupported = true;
 
-	ms->kmirrord_wq = alloc_workqueue("kmirrord",
-					  WQ_NON_REENTRANT | WQ_MEM_RECLAIM, 0);
+	ms->kmirrord_wq = alloc_workqueue("kmirrord", WQ_MEM_RECLAIM, 0);
 	if (!ms->kmirrord_wq) {
 		DMERR("couldn't start kmirrord");
 		r = -ENOMEM;
diff --git a/drivers/md/dm-stats.c b/drivers/md/dm-stats.c
new file mode 100644
index 000000000000..8ae31e8d3d64
--- /dev/null
+++ b/drivers/md/dm-stats.c
@@ -0,0 +1,969 @@
+#include <linux/errno.h>
+#include <linux/numa.h>
+#include <linux/slab.h>
+#include <linux/rculist.h>
+#include <linux/threads.h>
+#include <linux/preempt.h>
+#include <linux/irqflags.h>
+#include <linux/vmalloc.h>
+#include <linux/mm.h>
+#include <linux/module.h>
+#include <linux/device-mapper.h>
+
+#include "dm.h"
+#include "dm-stats.h"
+
+#define DM_MSG_PREFIX "stats"
+
+static int dm_stat_need_rcu_barrier;
+
+/*
+ * Using 64-bit values to avoid overflow (which is a
+ * problem that block/genhd.c's IO accounting has).
+ */
+struct dm_stat_percpu {
+	unsigned long long sectors[2];
+	unsigned long long ios[2];
+	unsigned long long merges[2];
+	unsigned long long ticks[2];
+	unsigned long long io_ticks[2];
+	unsigned long long io_ticks_total;
+	unsigned long long time_in_queue;
+};
+
+struct dm_stat_shared {
+	atomic_t in_flight[2];
+	unsigned long stamp;
+	struct dm_stat_percpu tmp;
+};
+
+struct dm_stat {
+	struct list_head list_entry;
+	int id;
+	size_t n_entries;
+	sector_t start;
+	sector_t end;
+	sector_t step;
+	const char *program_id;
+	const char *aux_data;
+	struct rcu_head rcu_head;
+	size_t shared_alloc_size;
+	size_t percpu_alloc_size;
+	struct dm_stat_percpu *stat_percpu[NR_CPUS];
+	struct dm_stat_shared stat_shared[0];
+};
+
+struct dm_stats_last_position {
+	sector_t last_sector;
+	unsigned last_rw;
+};
+
+/*
+ * A typo on the command line could possibly make the kernel run out of memory
+ * and crash. To prevent the crash we account all used memory. We fail if we
+ * exhaust 1/4 of all memory or 1/2 of vmalloc space.
+ */
+#define DM_STATS_MEMORY_FACTOR		4
+#define DM_STATS_VMALLOC_FACTOR		2
+
+static DEFINE_SPINLOCK(shared_memory_lock);
+
+static unsigned long shared_memory_amount;
+
+static bool __check_shared_memory(size_t alloc_size)
+{
+	size_t a;
+
+	a = shared_memory_amount + alloc_size;
+	if (a < shared_memory_amount)
+		return false;
+	if (a >> PAGE_SHIFT > totalram_pages / DM_STATS_MEMORY_FACTOR)
+		return false;
+#ifdef CONFIG_MMU
+	if (a > (VMALLOC_END - VMALLOC_START) / DM_STATS_VMALLOC_FACTOR)
+		return false;
+#endif
+	return true;
+}
+
+static bool check_shared_memory(size_t alloc_size)
+{
+	bool ret;
+
+	spin_lock_irq(&shared_memory_lock);
+
+	ret = __check_shared_memory(alloc_size);
+
+	spin_unlock_irq(&shared_memory_lock);
+
+	return ret;
+}
+
+static bool claim_shared_memory(size_t alloc_size)
+{
+	spin_lock_irq(&shared_memory_lock);
+
+	if (!__check_shared_memory(alloc_size)) {
+		spin_unlock_irq(&shared_memory_lock);
+		return false;
+	}
+
+	shared_memory_amount += alloc_size;
+
+	spin_unlock_irq(&shared_memory_lock);
+
+	return true;
+}
+
+static void free_shared_memory(size_t alloc_size)
+{
+	unsigned long flags;
+
+	spin_lock_irqsave(&shared_memory_lock, flags);
+
+	if (WARN_ON_ONCE(shared_memory_amount < alloc_size)) {
+		spin_unlock_irqrestore(&shared_memory_lock, flags);
+		DMCRIT("Memory usage accounting bug.");
+		return;
+	}
+
+	shared_memory_amount -= alloc_size;
+
+	spin_unlock_irqrestore(&shared_memory_lock, flags);
+}
+
+static void *dm_kvzalloc(size_t alloc_size, int node)
+{
+	void *p;
+
+	if (!claim_shared_memory(alloc_size))
+		return NULL;
+
+	if (alloc_size <= KMALLOC_MAX_SIZE) {
+		p = kzalloc_node(alloc_size, GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN, node);
+		if (p)
+			return p;
+	}
+	p = vzalloc_node(alloc_size, node);
+	if (p)
+		return p;
+
+	free_shared_memory(alloc_size);
+
+	return NULL;
+}
+
+static void dm_kvfree(void *ptr, size_t alloc_size)
+{
+	if (!ptr)
+		return;
+
+	free_shared_memory(alloc_size);
+
+	if (is_vmalloc_addr(ptr))
+		vfree(ptr);
+	else
+		kfree(ptr);
+}
+
+static void dm_stat_free(struct rcu_head *head)
+{
+	int cpu;
+	struct dm_stat *s = container_of(head, struct dm_stat, rcu_head);
+
+	kfree(s->program_id);
+	kfree(s->aux_data);
+	for_each_possible_cpu(cpu)
+		dm_kvfree(s->stat_percpu[cpu], s->percpu_alloc_size);
+	dm_kvfree(s, s->shared_alloc_size);
+}
+
+static int dm_stat_in_flight(struct dm_stat_shared *shared)
+{
+	return atomic_read(&shared->in_flight[READ]) +
+	       atomic_read(&shared->in_flight[WRITE]);
+}
+
+void dm_stats_init(struct dm_stats *stats)
+{
+	int cpu;
+	struct dm_stats_last_position *last;
+
+	mutex_init(&stats->mutex);
+	INIT_LIST_HEAD(&stats->list);
+	stats->last = alloc_percpu(struct dm_stats_last_position);
+	for_each_possible_cpu(cpu) {
+		last = per_cpu_ptr(stats->last, cpu);
+		last->last_sector = (sector_t)ULLONG_MAX;
+		last->last_rw = UINT_MAX;
+	}
+}
+
+void dm_stats_cleanup(struct dm_stats *stats)
+{
+	size_t ni;
+	struct dm_stat *s;
+	struct dm_stat_shared *shared;
+
+	while (!list_empty(&stats->list)) {
+		s = container_of(stats->list.next, struct dm_stat, list_entry);
+		list_del(&s->list_entry);
+		for (ni = 0; ni < s->n_entries; ni++) {
+			shared = &s->stat_shared[ni];
+			if (WARN_ON(dm_stat_in_flight(shared))) {
+				DMCRIT("leaked in-flight counter at index %lu "
+				       "(start %llu, end %llu, step %llu): reads %d, writes %d",
+				       (unsigned long)ni,
+				       (unsigned long long)s->start,
+				       (unsigned long long)s->end,
+				       (unsigned long long)s->step,
+				       atomic_read(&shared->in_flight[READ]),
+				       atomic_read(&shared->in_flight[WRITE]));
+			}
+		}
+		dm_stat_free(&s->rcu_head);
+	}
+	free_percpu(stats->last);
+}
+
+static int dm_stats_create(struct dm_stats *stats, sector_t start, sector_t end,
+			   sector_t step, const char *program_id, const char *aux_data,
+			   void (*suspend_callback)(struct mapped_device *),
+			   void (*resume_callback)(struct mapped_device *),
+			   struct mapped_device *md)
+{
+	struct list_head *l;
+	struct dm_stat *s, *tmp_s;
+	sector_t n_entries;
+	size_t ni;
+	size_t shared_alloc_size;
+	size_t percpu_alloc_size;
+	struct dm_stat_percpu *p;
+	int cpu;
+	int ret_id;
+	int r;
+
+	if (end < start || !step)
+		return -EINVAL;
+
+	n_entries = end - start;
+	if (dm_sector_div64(n_entries, step))
+		n_entries++;
+
+	if (n_entries != (size_t)n_entries || !(size_t)(n_entries + 1))
+		return -EOVERFLOW;
+
+	shared_alloc_size = sizeof(struct dm_stat) + (size_t)n_entries * sizeof(struct dm_stat_shared);
+	if ((shared_alloc_size - sizeof(struct dm_stat)) / sizeof(struct dm_stat_shared) != n_entries)
+		return -EOVERFLOW;
+
+	percpu_alloc_size = (size_t)n_entries * sizeof(struct dm_stat_percpu);
+	if (percpu_alloc_size / sizeof(struct dm_stat_percpu) != n_entries)
+		return -EOVERFLOW;
+
+	if (!check_shared_memory(shared_alloc_size + num_possible_cpus() * percpu_alloc_size))
+		return -ENOMEM;
+
+	s = dm_kvzalloc(shared_alloc_size, NUMA_NO_NODE);
+	if (!s)
+		return -ENOMEM;
+
+	s->n_entries = n_entries;
+	s->start = start;
+	s->end = end;
+	s->step = step;
+	s->shared_alloc_size = shared_alloc_size;
+	s->percpu_alloc_size = percpu_alloc_size;
+
+	s->program_id = kstrdup(program_id, GFP_KERNEL);
+	if (!s->program_id) {
+		r = -ENOMEM;
+		goto out;
+	}
+	s->aux_data = kstrdup(aux_data, GFP_KERNEL);
+	if (!s->aux_data) {
+		r = -ENOMEM;
+		goto out;
+	}
+
+	for (ni = 0; ni < n_entries; ni++) {
+		atomic_set(&s->stat_shared[ni].in_flight[READ], 0);
+		atomic_set(&s->stat_shared[ni].in_flight[WRITE], 0);
+	}
+
+	for_each_possible_cpu(cpu) {
+		p = dm_kvzalloc(percpu_alloc_size, cpu_to_node(cpu));
+		if (!p) {
+			r = -ENOMEM;
+			goto out;
+		}
+		s->stat_percpu[cpu] = p;
+	}
+
+	/*
+	 * Suspend/resume to make sure there is no i/o in flight,
+	 * so that newly created statistics will be exact.
+	 *
+	 * (note: we couldn't suspend earlier because we must not
+	 * allocate memory while suspended)
+	 */
+	suspend_callback(md);
+
+	mutex_lock(&stats->mutex);
+	s->id = 0;
+	list_for_each(l, &stats->list) {
+		tmp_s = container_of(l, struct dm_stat, list_entry);
+		if (WARN_ON(tmp_s->id < s->id)) {
+			r = -EINVAL;
+			goto out_unlock_resume;
+		}
+		if (tmp_s->id > s->id)
+			break;
+		if (unlikely(s->id == INT_MAX)) {
+			r = -ENFILE;
+			goto out_unlock_resume;
+		}
+		s->id++;
+	}
+	ret_id = s->id;
+	list_add_tail_rcu(&s->list_entry, l);
+	mutex_unlock(&stats->mutex);
+
+	resume_callback(md);
+
+	return ret_id;
+
+out_unlock_resume:
+	mutex_unlock(&stats->mutex);
+	resume_callback(md);
+out:
+	dm_stat_free(&s->rcu_head);
+	return r;
+}
+
+static struct dm_stat *__dm_stats_find(struct dm_stats *stats, int id)
+{
+	struct dm_stat *s;
+
+	list_for_each_entry(s, &stats->list, list_entry) {
+		if (s->id > id)
+			break;
+		if (s->id == id)
+			return s;
+	}
+
+	return NULL;
+}
+
+static int dm_stats_delete(struct dm_stats *stats, int id)
+{
+	struct dm_stat *s;
+	int cpu;
+
+	mutex_lock(&stats->mutex);
+
+	s = __dm_stats_find(stats, id);
+	if (!s) {
+		mutex_unlock(&stats->mutex);
+		return -ENOENT;
+	}
+
+	list_del_rcu(&s->list_entry);
+	mutex_unlock(&stats->mutex);
+
+	/*
+	 * vfree can't be called from RCU callback
+	 */
+	for_each_possible_cpu(cpu)
+		if (is_vmalloc_addr(s->stat_percpu))
+			goto do_sync_free;
+	if (is_vmalloc_addr(s)) {
+do_sync_free:
+		synchronize_rcu_expedited();
+		dm_stat_free(&s->rcu_head);
+	} else {
+		ACCESS_ONCE(dm_stat_need_rcu_barrier) = 1;
+		call_rcu(&s->rcu_head, dm_stat_free);
+	}
+	return 0;
+}
+
+static int dm_stats_list(struct dm_stats *stats, const char *program,
+			 char *result, unsigned maxlen)
+{
+	struct dm_stat *s;
+	sector_t len;
+	unsigned sz = 0;
+
+	/*
+	 * Output format:
+	 *   <region_id>: <start_sector>+<length> <step> <program_id> <aux_data>
+	 */
+
+	mutex_lock(&stats->mutex);
+	list_for_each_entry(s, &stats->list, list_entry) {
+		if (!program || !strcmp(program, s->program_id)) {
+			len = s->end - s->start;
+			DMEMIT("%d: %llu+%llu %llu %s %s\n", s->id,
+				(unsigned long long)s->start,
+				(unsigned long long)len,
+				(unsigned long long)s->step,
+				s->program_id,
+				s->aux_data);
+		}
+	}
+	mutex_unlock(&stats->mutex);
+
+	return 1;
+}
+
+static void dm_stat_round(struct dm_stat_shared *shared, struct dm_stat_percpu *p)
+{
+	/*
+	 * This is racy, but so is part_round_stats_single.
+	 */
+	unsigned long now = jiffies;
+	unsigned in_flight_read;
+	unsigned in_flight_write;
+	unsigned long difference = now - shared->stamp;
+
+	if (!difference)
+		return;
+	in_flight_read = (unsigned)atomic_read(&shared->in_flight[READ]);
+	in_flight_write = (unsigned)atomic_read(&shared->in_flight[WRITE]);
+	if (in_flight_read)
+		p->io_ticks[READ] += difference;
+	if (in_flight_write)
+		p->io_ticks[WRITE] += difference;
+	if (in_flight_read + in_flight_write) {
+		p->io_ticks_total += difference;
+		p->time_in_queue += (in_flight_read + in_flight_write) * difference;
+	}
+	shared->stamp = now;
+}
+
+static void dm_stat_for_entry(struct dm_stat *s, size_t entry,
+			      unsigned long bi_rw, sector_t len, bool merged,
+			      bool end, unsigned long duration)
+{
+	unsigned long idx = bi_rw & REQ_WRITE;
+	struct dm_stat_shared *shared = &s->stat_shared[entry];
+	struct dm_stat_percpu *p;
+
+	/*
+	 * For strict correctness we should use local_irq_disable/enable
+	 * instead of preempt_disable/enable.
+	 *
+	 * This is racy if the driver finishes bios from non-interrupt
+	 * context as well as from interrupt context or from more different
+	 * interrupts.
+	 *
+	 * However, the race only results in not counting some events,
+	 * so it is acceptable.
+	 *
+	 * part_stat_lock()/part_stat_unlock() have this race too.
+	 */
+	preempt_disable();
+	p = &s->stat_percpu[smp_processor_id()][entry];
+
+	if (!end) {
+		dm_stat_round(shared, p);
+		atomic_inc(&shared->in_flight[idx]);
+	} else {
+		dm_stat_round(shared, p);
+		atomic_dec(&shared->in_flight[idx]);
+		p->sectors[idx] += len;
+		p->ios[idx] += 1;
+		p->merges[idx] += merged;
+		p->ticks[idx] += duration;
+	}
+
+	preempt_enable();
+}
+
+static void __dm_stat_bio(struct dm_stat *s, unsigned long bi_rw,
+			  sector_t bi_sector, sector_t end_sector,
+			  bool end, unsigned long duration,
+			  struct dm_stats_aux *stats_aux)
+{
+	sector_t rel_sector, offset, todo, fragment_len;
+	size_t entry;
+
+	if (end_sector <= s->start || bi_sector >= s->end)
+		return;
+	if (unlikely(bi_sector < s->start)) {
+		rel_sector = 0;
+		todo = end_sector - s->start;
+	} else {
+		rel_sector = bi_sector - s->start;
+		todo = end_sector - bi_sector;
+	}
+	if (unlikely(end_sector > s->end))
+		todo -= (end_sector - s->end);
+
+	offset = dm_sector_div64(rel_sector, s->step);
+	entry = rel_sector;
+	do {
+		if (WARN_ON_ONCE(entry >= s->n_entries)) {
+			DMCRIT("Invalid area access in region id %d", s->id);
+			return;
+		}
+		fragment_len = todo;
+		if (fragment_len > s->step - offset)
+			fragment_len = s->step - offset;
+		dm_stat_for_entry(s, entry, bi_rw, fragment_len,
+				  stats_aux->merged, end, duration);
+		todo -= fragment_len;
+		entry++;
+		offset = 0;
+	} while (unlikely(todo != 0));
+}
+
+void dm_stats_account_io(struct dm_stats *stats, unsigned long bi_rw,
+			 sector_t bi_sector, unsigned bi_sectors, bool end,
+			 unsigned long duration, struct dm_stats_aux *stats_aux)
+{
+	struct dm_stat *s;
+	sector_t end_sector;
+	struct dm_stats_last_position *last;
+
+	if (unlikely(!bi_sectors))
+		return;
+
+	end_sector = bi_sector + bi_sectors;
+
+	if (!end) {
+		/*
+		 * A race condition can at worst result in the merged flag being
+		 * misrepresented, so we don't have to disable preemption here.
+		 */
+		last = __this_cpu_ptr(stats->last);
+		stats_aux->merged =
+			(bi_sector == (ACCESS_ONCE(last->last_sector) &&
+				       ((bi_rw & (REQ_WRITE | REQ_DISCARD)) ==
+					(ACCESS_ONCE(last->last_rw) & (REQ_WRITE | REQ_DISCARD)))
+				       ));
+		ACCESS_ONCE(last->last_sector) = end_sector;
+		ACCESS_ONCE(last->last_rw) = bi_rw;
+	}
+
+	rcu_read_lock();
+
+	list_for_each_entry_rcu(s, &stats->list, list_entry)
+		__dm_stat_bio(s, bi_rw, bi_sector, end_sector, end, duration, stats_aux);
+
+	rcu_read_unlock();
+}
+
+static void __dm_stat_init_temporary_percpu_totals(struct dm_stat_shared *shared,
+						   struct dm_stat *s, size_t x)
+{
+	int cpu;
+	struct dm_stat_percpu *p;
+
+	local_irq_disable();
+	p = &s->stat_percpu[smp_processor_id()][x];
+	dm_stat_round(shared, p);
+	local_irq_enable();
+
+	memset(&shared->tmp, 0, sizeof(shared->tmp));
+	for_each_possible_cpu(cpu) {
+		p = &s->stat_percpu[cpu][x];
+		shared->tmp.sectors[READ] += ACCESS_ONCE(p->sectors[READ]);
+		shared->tmp.sectors[WRITE] += ACCESS_ONCE(p->sectors[WRITE]);
+		shared->tmp.ios[READ] += ACCESS_ONCE(p->ios[READ]);
+		shared->tmp.ios[WRITE] += ACCESS_ONCE(p->ios[WRITE]);
+		shared->tmp.merges[READ] += ACCESS_ONCE(p->merges[READ]);
+		shared->tmp.merges[WRITE] += ACCESS_ONCE(p->merges[WRITE]);
+		shared->tmp.ticks[READ] += ACCESS_ONCE(p->ticks[READ]);
+		shared->tmp.ticks[WRITE] += ACCESS_ONCE(p->ticks[WRITE]);
+		shared->tmp.io_ticks[READ] += ACCESS_ONCE(p->io_ticks[READ]);
+		shared->tmp.io_ticks[WRITE] += ACCESS_ONCE(p->io_ticks[WRITE]);
+		shared->tmp.io_ticks_total += ACCESS_ONCE(p->io_ticks_total);
+		shared->tmp.time_in_queue += ACCESS_ONCE(p->time_in_queue);
+	}
+}
+
+static void __dm_stat_clear(struct dm_stat *s, size_t idx_start, size_t idx_end,
+			    bool init_tmp_percpu_totals)
+{
+	size_t x;
+	struct dm_stat_shared *shared;
+	struct dm_stat_percpu *p;