dm-raid1.c

Linux Kernel 2.6.9 for OMAP1710

/*
 * 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;
	sector_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);

static void *region_alloc(int gfp_mask, void *pool_data)
{
	return kmalloc(sizeof(struct region), gfp_mask);
}

static void region_free(void *element, void *pool_data)
{
	kfree(element);
}

#define MIN_REGIONS 64
#define MAX_RECOVERY 1
static int rh_init(struct region_hash *rh, struct mirror_set *ms,
		   struct dirty_log *log, sector_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(MIN_REGIONS, region_alloc,
					 region_free, NULL);
	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 (unsigned int) ((region * RH_HASH_MULT) >> 12) & rh->mask;
}

static struct region *__rh_lookup(struct region_hash *rh, region_t region)
{
	struct region *reg;

	list_for_each_entry (reg, rh->buckets + rh_hash(rh, region), hash_list)
		if (reg->key == region)
			return reg;

	return NULL;
}

static void __rh_insert(struct region_hash *rh, struct region *reg)
{
	unsigned int h = rh_hash(rh, reg->key);
	list_add(&reg->hash_list, rh->buckets + h);
}

static struct region *__rh_alloc(struct region_hash *rh, region_t region)
{
	struct region *reg, *nreg;

	read_unlock(&rh->hash_lock);
	nreg = mempool_alloc(rh->region_pool, GFP_NOIO);
	nreg->state = rh->log->type->in_sync(rh->log, region, 1) ?
		RH_CLEAN : RH_NOSYNC;
	nreg->rh = rh;
	nreg->key = region;

	INIT_LIST_HEAD(&nreg->list);

	atomic_set(&nreg->pending, 0);
	bio_list_init(&nreg->delayed_bios);
	write_lock_irq(&rh->hash_lock);

	reg = __rh_lookup(rh, region);
	if (reg)
		/* we lost the race */
		mempool_free(nreg, rh->region_pool);

	else {
		__rh_insert(rh, nreg);
		if (nreg->state == RH_CLEAN) {
			spin_lock_irq(&rh->region_lock);
			list_add(&nreg->list, &rh->clean_regions);
			spin_unlock_irq(&rh->region_lock);
		}
		reg = nreg;
	}
	write_unlock_irq(&rh->hash_lock);
	read_lock(&rh->hash_lock);

	return reg;
}

static inline struct region *__rh_find(struct region_hash *rh, region_t region)
{
	struct region *reg;

	reg = __rh_lookup(rh, region);
	if (!reg)
		reg = __rh_alloc(rh, region);

	return reg;
}

static int rh_state(struct region_hash *rh, region_t region, int may_block)
{
	int r;
	struct region *reg;

	read_lock(&rh->hash_lock);
	reg = __rh_lookup(rh, region);
	read_unlock(&rh->hash_lock);

	if (reg)
		return reg->state;

	/*
	 * The region wasn't in the hash, so we fall back to the
	 * dirty log.
	 */
	r = rh->log->type->in_sync(rh->log, region, may_block);

	/*
	 * Any error from the dirty log (eg. -EWOULDBLOCK) gets
	 * taken as a RH_NOSYNC
	 */
	return r == 1 ? RH_CLEAN : RH_NOSYNC;
}

static inline int rh_in_sync(struct region_hash *rh,
			     region_t region, int may_block)
{
	int state = rh_state(rh, region, may_block);
	return state == RH_CLEAN || state == RH_DIRTY;
}

static void dispatch_bios(struct mirror_set *ms, struct bio_list *bio_list)
{
	struct bio *bio;

	while ((bio = bio_list_pop(bio_list))) {
		queue_bio(ms, bio, WRITE);
	}
}

static void rh_update_states(struct region_hash *rh)
{
	struct region *reg, *next;

	LIST_HEAD(clean);
	LIST_HEAD(recovered);

	/*
	 * Quickly grab the lists.
	 */
	write_lock_irq(&rh->hash_lock);
	spin_lock(&rh->region_lock);
	if (!list_empty(&rh->clean_regions)) {
		list_splice(&rh->clean_regions, &clean);
		INIT_LIST_HEAD(&rh->clean_regions);

		list_for_each_entry (reg, &clean, list) {
			rh->log->type->clear_region(rh->log, reg->key);
			list_del(&reg->hash_list);
		}
	}

	if (!list_empty(&rh->recovered_regions)) {
		list_splice(&rh->recovered_regions, &recovered);
		INIT_LIST_HEAD(&rh->recovered_regions);

		list_for_each_entry (reg, &recovered, list)
			list_del(&reg->hash_list);
	}
	spin_unlock(&rh->region_lock);
	write_unlock_irq(&rh->hash_lock);

	/*
	 * All the regions on the recovered and clean lists have
	 * now been pulled out of the system, so no need to do
	 * any more locking.
	 */
	list_for_each_entry_safe (reg, next, &recovered, list) {
		rh->log->type->clear_region(rh->log, reg->key);
		rh->log->type->complete_resync_work(rh->log, reg->key, 1);
		dispatch_bios(rh->ms, &reg->delayed_bios);
		up(&rh->recovery_count);
		mempool_free(reg, rh->region_pool);
	}

	if (!list_empty(&recovered))
		rh->log->type->flush(rh->log);

	list_for_each_entry_safe (reg, next, &clean, list)
		mempool_free(reg, rh->region_pool);
}

static void rh_inc(struct region_hash *rh, region_t region)
{
	struct region *reg;

	read_lock(&rh->hash_lock);
	reg = __rh_find(rh, region);
	if (reg->state == RH_CLEAN) {
		rh->log->type->mark_region(rh->log, reg->key);

		spin_lock_irq(&rh->region_lock);
		reg->state = RH_DIRTY;
		list_del_init(&reg->list);	/* take off the clean list */
		spin_unlock_irq(&rh->region_lock);
	}

	atomic_inc(&reg->pending);
	read_unlock(&rh->hash_lock);
}

static void rh_inc_pending(struct region_hash *rh, struct bio_list *bios)
{
	struct bio *bio;

	for (bio = bios->head; bio; bio = bio->bi_next)
		rh_inc(rh, bio_to_region(rh, bio));
}

static void rh_dec(struct region_hash *rh, region_t region)
{
	unsigned long flags;
	struct region *reg;
	int should_wake = 0;

	read_lock(&rh->hash_lock);
	reg = __rh_lookup(rh, region);
	read_unlock(&rh->hash_lock);

	if (atomic_dec_and_test(&reg->pending)) {
		spin_lock_irqsave(&rh->region_lock, flags);
		if (reg->state == RH_RECOVERING) {
			list_add_tail(&reg->list, &rh->quiesced_regions);
		} else {
			reg->state = RH_CLEAN;
			list_add(&reg->list, &rh->clean_regions);
		}
		spin_unlock_irqrestore(&rh->region_lock, flags);
		should_wake = 1;
	}

	if (should_wake)
		wake();
}

/*
 * Starts quiescing a region in preparation for recovery.
 */
static int __rh_recovery_prepare(struct region_hash *rh)
{
	int r;
	struct region *reg;
	region_t region;

	/*
	 * Ask the dirty log what's next.
	 */
	r = rh->log->type->get_resync_work(rh->log, &region);
	if (r <= 0)
		return r;

	/*
	 * Get this region, and start it quiescing by setting the
	 * recovering flag.
	 */
	read_lock(&rh->hash_lock);
	reg = __rh_find(rh, region);
	read_unlock(&rh->hash_lock);

	spin_lock_irq(&rh->region_lock);
	reg->state = RH_RECOVERING;

	/* Already quiesced ? */
	if (atomic_read(&reg->pending))
		list_del_init(&reg->list);

	else {
		list_del_init(&reg->list);
		list_add(&reg->list, &rh->quiesced_regions);
	}
	spin_unlock_irq(&rh->region_lock);

	return 1;
}

static void rh_recovery_prepare(struct region_hash *rh)
{
	while (!down_trylock(&rh->recovery_count))
		if (__rh_recovery_prepare(rh) <= 0) {
			up(&rh->recovery_count);
			break;
		}
}

/*
 * Returns any quiesced regions.
 */
static struct region *rh_recovery_start(struct region_hash *rh)
{
	struct region *reg = NULL;

	spin_lock_irq(&rh->region_lock);
	if (!list_empty(&rh->quiesced_regions)) {
		reg = list_entry(rh->quiesced_regions.next,
				 struct region, list);
		list_del_init(&reg->list);	/* remove from the quiesced list */
	}
	spin_unlock_irq(&rh->region_lock);

	return reg;
}

/* FIXME: success ignored for now */
static void rh_recovery_end(struct region *reg, int success)
{
	struct region_hash *rh = reg->rh;

	spin_lock_irq(&rh->region_lock);
	list_add(&reg->list, &reg->rh->recovered_regions);
	spin_unlock_irq(&rh->region_lock);

	wake();
}

static void rh_flush(struct region_hash *rh)
{
	rh->log->type->flush(rh->log);
}

static void rh_delay(struct region_hash *rh, struct bio *bio)
{
	struct region *reg;

	read_lock(&rh->hash_lock);
	reg = __rh_find(rh, bio_to_region(rh, bio));
	bio_list_add(&reg->delayed_bios, bio);
	read_unlock(&rh->hash_lock);
}

static void rh_stop_recovery(struct region_hash *rh)
{
	int i;

	/* wait for any recovering regions */
	for (i = 0; i < MAX_RECOVERY; i++)
		down(&rh->recovery_count);
}

static void rh_start_recovery(struct region_hash *rh)
{
	int i;

	for (i = 0; i < MAX_RECOVERY; i++)
		up(&rh->recovery_count);

	wake();
}

/*-----------------------------------------------------------------
 * Mirror set structures.
 *---------------------------------------------------------------*/
struct mirror {
	atomic_t error_count;
	struct dm_dev *dev;
	sector_t offset;
};

struct mirror_set {
	struct dm_target *ti;
	struct list_head list;
	struct region_hash rh;
	struct kcopyd_client *kcopyd_client;

	spinlock_t lock;	/* protects the next two lists */
	struct bio_list reads;
	struct bio_list writes;

	/* recovery */
	region_t nr_regions;
	int in_sync;

	unsigned int nr_mirrors;
	struct mirror mirror[0];
};

/*
 * Every mirror should look like this one.
 */
#define DEFAULT_MIRROR 0

/*
 * This is yucky.  We squirrel the mirror_set struct away inside
 * bi_next for write buffers.  This is safe since the bh
 * doesn't get submitted to the lower levels of block layer.
 */
static struct mirror_set *bio_get_ms(struct bio *bio)
{
	return (struct mirror_set *) bio->bi_next;
}

static void bio_set_ms(struct bio *bio, struct mirror_set *ms)
{
	bio->bi_next = (struct bio *) ms;
}

/*-----------------------------------------------------------------
 * Recovery.
 *
 * When a mirror is first activated we may find that some regions
 * are in the no-sync state.  We have to recover these by
 * recopying from the default mirror to all the others.
 *---------------------------------------------------------------*/
static void recovery_complete(int read_err, unsigned int write_err,
			      void *context)
{
	struct region *reg = (struct region *) context;

	/* FIXME: better error handling */
	rh_recovery_end(reg, read_err || write_err);
}

static int recover(struct mirror_set *ms, struct region *reg)
{
	int r;
	unsigned int i;
	struct io_region from, to[KCOPYD_MAX_REGIONS], *dest;
	struct mirror *m;
	unsigned long flags = 0;

	/* fill in the source */
	m = ms->mirror + DEFAULT_MIRROR;
	from.bdev = m->dev->bdev;
	from.sector = m->offset + region_to_sector(reg->rh, reg->key);
	if (reg->key == (ms->nr_regions - 1)) {
		/*
		 * The final region may be smaller than
		 * region_size.
		 */
		from.count = ms->ti->len & (reg->rh->region_size - 1);
		if (!from.count)
			from.count = reg->rh->region_size;
	} else
		from.count = reg->rh->region_size;

	/* fill in the destinations */
	for (i = 0, dest = to; i < ms->nr_mirrors; i++) {
		if (i == DEFAULT_MIRROR)
			continue;

		m = ms->mirror + i;
		dest->bdev = m->dev->bdev;
		dest->sector = m->offset + region_to_sector(reg->rh, reg->key);
		dest->count = from.count;
		dest++;
	}

	/* hand to kcopyd */
	set_bit(KCOPYD_IGNORE_ERROR, &flags);