path: root/drivers/md/dm-raid1.c

/*
 * Copyright (C) 2003 Sistina Software Limited.
 *
 * This file is released under the GPL.
 */

#include "dm.h"
#include "dm-bio-list.h"
#include "dm-io.h"
#include "dm-log.h"
#include "kcopyd.h"

#include <linux/ctype.h>
#include <linux/init.h>
#include <linux/mempool.h>
#include <linux/module.h>
#include <linux/pagemap.h>
#include <linux/slab.h>
#include <linux/time.h>
#include <linux/vmalloc.h>
#include <linux/workqueue.h>

static struct workqueue_struct *_kmirrord_wq;
static struct work_struct _kmirrord_work;

static inline void wake(void)
{
	queue_work(_kmirrord_wq, &_kmirrord_work);
}

/*-----------------------------------------------------------------
 * Region hash
 *
 * The mirror splits itself up into discrete regions.  Each
 * region can be in one of three states: clean, dirty,
 * nosync.  There is no need to put clean regions in the hash.
 *
 * In addition to being present in the hash table a region _may_
 * be present on one of three lists.
 *
 *   clean_regions: Regions on this list have no io pending to
 *   them, they are in sync, we are no longer interested in them,
 *   they are dull.  rh_update_states() will remove them from the
 *   hash table.
 *
 *   quiesced_regions: These regions have been spun down, ready
 *   for recovery.  rh_recovery_start() will remove regions from
 *   this list and hand them to kmirrord, which will schedule the
 *   recovery io with kcopyd.
 *
 *   recovered_regions: Regions that kcopyd has successfully
 *   recovered.  rh_update_states() will now schedule any delayed
 *   io, up the recovery_count, and remove the region from the
 *   hash.
 *
 * There are 2 locks:
 *   A rw spin lock 'hash_lock' protects just the hash table,
 *   this is never held in write mode from interrupt context,
 *   which I believe means that we only have to disable irqs when
 *   doing a write lock.
 *
 *   An ordinary spin lock 'region_lock' that protects the three
 *   lists in the region_hash, with the 'state', 'list' and
 *   'bhs_delayed' fields of the regions.  This is used from irq
 *   context, so all other uses will have to suspend local irqs.
 *---------------------------------------------------------------*/
struct mirror_set;
struct region_hash {
	struct mirror_set *ms;
	uint32_t region_size;
	unsigned region_shift;

	/* holds persistent region state */
	struct dirty_log *log;

	/* hash table */
	rwlock_t hash_lock;
	mempool_t *region_pool;
	unsigned int mask;
	unsigned int nr_buckets;
	struct list_head *buckets;

	spinlock_t region_lock;
	struct semaphore recovery_count;
	struct list_head clean_regions;
	struct list_head quiesced_regions;
	struct list_head recovered_regions;
};

enum {
	RH_CLEAN,
	RH_DIRTY,
	RH_NOSYNC,
	RH_RECOVERING
};

struct region {
	struct region_hash *rh;	/* FIXME: can we get rid of this ? */
	region_t key;
	int state;

	struct list_head hash_list;
	struct list_head list;

	atomic_t pending;
	struct bio_list delayed_bios;
};

/*
 * Conversion fns
 */
static inline region_t bio_to_region(struct region_hash *rh, struct bio *bio)
{
	return bio->bi_sector >> rh->region_shift;
}

static inline sector_t region_to_sector(struct region_hash *rh, region_t region)
{
	return region << rh->region_shift;
}

/* FIXME move this */
static void queue_bio(struct mirror_set *ms, struct bio *bio, int rw);

#define MIN_REGIONS 64
#define MAX_RECOVERY 1
static int rh_init(struct region_hash *rh, struct mirror_set *ms,
		   struct dirty_log *log, uint32_t region_size,
		   region_t nr_regions)
{
	unsigned int nr_buckets, max_buckets;
	size_t i;

	/*
	 * Calculate a suitable number of buckets for our hash
	 * table.
	 */
	max_buckets = nr_regions >> 6;
	for (nr_buckets = 128u; nr_buckets < max_buckets; nr_buckets <<= 1)
		;
	nr_buckets >>= 1;

	rh->ms = ms;
	rh->log = log;
	rh->region_size = region_size;
	rh->region_shift = ffs(region_size) - 1;
	rwlock_init(&rh->hash_lock);
	rh->mask = nr_buckets - 1;
	rh->nr_buckets = nr_buckets;

	rh->buckets = vmalloc(nr_buckets * sizeof(*rh->buckets));
	if (!rh->buckets) {
		DMERR("unable to allocate region hash memory");
		return -ENOMEM;
	}

	for (i = 0; i < nr_buckets; i++)
		INIT_LIST_HEAD(rh->buckets + i);

	spin_lock_init(&rh->region_lock);
	sema_init(&rh->recovery_count, 0);
	INIT_LIST_HEAD(&rh->clean_regions);
	INIT_LIST_HEAD(&rh->quiesced_regions);
	INIT_LIST_HEAD(&rh->recovered_regions);

	rh->region_pool = mempool_create_kmalloc_pool(MIN_REGIONS,
						      sizeof(struct region));
	if (!rh->region_pool) {
		vfree(rh->buckets);
		rh->buckets = NULL;
		return -ENOMEM;
	}

	return 0;
}

static void rh_exit(struct region_hash *rh)
{
	unsigned int h;
	struct region *reg, *nreg;

	BUG_ON(!list_empty(&rh->quiesced_regions));
	for (h = 0; h < rh->nr_buckets; h++) {
		list_for_each_entry_safe(reg, nreg, rh->buckets + h, hash_list) {
			BUG_ON(atomic_read(&reg->pending));
			mempool_free(reg, rh->region_pool);
		}
	}

	if (rh->log)
		dm_destroy_dirty_log(rh->log);
	if (rh->region_pool)
		mempool_destroy(rh->region_pool);
	vfree(rh->buckets);
}

#define RH_HASH_MULT 2654435387U

static inline unsigned int rh_hash(struct region_hash *rh, region_t region)
{
	return (<