#include "dm.h"
#include "dm-bio-list.h"
-#include "dm-io.h"
-#include "dm-log.h"
-#include "kcopyd.h"
+#include "dm-bio-record.h"
#include <linux/ctype.h>
#include <linux/init.h>
#include <linux/time.h>
#include <linux/vmalloc.h>
#include <linux/workqueue.h>
+#include <linux/log2.h>
+#include <linux/hardirq.h>
+#include <linux/dm-io.h>
+#include <linux/dm-dirty-log.h>
+#include <linux/dm-kcopyd.h>
#define DM_MSG_PREFIX "raid1"
+#define DM_IO_PAGES 64
-static struct workqueue_struct *_kmirrord_wq;
-static struct work_struct _kmirrord_work;
+#define DM_RAID1_HANDLE_ERRORS 0x01
+#define errors_handled(p) ((p)->features & DM_RAID1_HANDLE_ERRORS)
-static inline void wake(void)
-{
- queue_work(_kmirrord_wq, &_kmirrord_work);
-}
+static DECLARE_WAIT_QUEUE_HEAD(_kmirrord_recovery_stopped);
/*-----------------------------------------------------------------
* Region hash
unsigned region_shift;
/* holds persistent region state */
- struct dirty_log *log;
+ struct dm_dirty_log *log;
/* hash table */
rwlock_t hash_lock;
struct list_head *buckets;
spinlock_t region_lock;
+ atomic_t recovery_in_flight;
struct semaphore recovery_count;
struct list_head clean_regions;
struct list_head quiesced_regions;
struct list_head recovered_regions;
+ struct list_head failed_recovered_regions;
};
enum {
/*-----------------------------------------------------------------
* Mirror set structures.
*---------------------------------------------------------------*/
+enum dm_raid1_error {
+ DM_RAID1_WRITE_ERROR,
+ DM_RAID1_SYNC_ERROR,
+ DM_RAID1_READ_ERROR
+};
+
struct mirror {
+ struct mirror_set *ms;
atomic_t error_count;
+ unsigned long error_type;
struct dm_dev *dev;
sector_t offset;
};
struct dm_target *ti;
struct list_head list;
struct region_hash rh;
- struct kcopyd_client *kcopyd_client;
+ struct dm_kcopyd_client *kcopyd_client;
+ uint64_t features;
- spinlock_t lock; /* protects the next two lists */
+ spinlock_t lock; /* protects the lists */
struct bio_list reads;
struct bio_list writes;
+ struct bio_list failures;
+
+ struct dm_io_client *io_client;
+ mempool_t *read_record_pool;
/* recovery */
region_t nr_regions;
int in_sync;
+ int log_failure;
+ atomic_t suspend;
+
+ atomic_t default_mirror; /* Default mirror */
- struct mirror *default_mirror; /* Default mirror */
+ struct workqueue_struct *kmirrord_wq;
+ struct work_struct kmirrord_work;
+ struct timer_list timer;
+ unsigned long timer_pending;
+
+ struct work_struct trigger_event;
unsigned int nr_mirrors;
struct mirror mirror[0];
return region << rh->region_shift;
}
+static void wake(struct mirror_set *ms)
+{
+ queue_work(ms->kmirrord_wq, &ms->kmirrord_work);
+}
+
+static void delayed_wake_fn(unsigned long data)
+{
+ struct mirror_set *ms = (struct mirror_set *) data;
+
+ clear_bit(0, &ms->timer_pending);
+ wake(ms);
+}
+
+static void delayed_wake(struct mirror_set *ms)
+{
+ if (test_and_set_bit(0, &ms->timer_pending))
+ return;
+
+ ms->timer.expires = jiffies + HZ / 5;
+ ms->timer.data = (unsigned long) ms;
+ ms->timer.function = delayed_wake_fn;
+ add_timer(&ms->timer);
+}
+
/* 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,
+ struct dm_dirty_log *log, uint32_t region_size,
region_t nr_regions)
{
unsigned int nr_buckets, max_buckets;
spin_lock_init(&rh->region_lock);
sema_init(&rh->recovery_count, 0);
+ atomic_set(&rh->recovery_in_flight, 0);
INIT_LIST_HEAD(&rh->clean_regions);
INIT_LIST_HEAD(&rh->quiesced_regions);
INIT_LIST_HEAD(&rh->recovered_regions);
+ INIT_LIST_HEAD(&rh->failed_recovered_regions);
rh->region_pool = mempool_create_kmalloc_pool(MIN_REGIONS,
sizeof(struct region));
}
if (rh->log)
- dm_destroy_dirty_log(rh->log);
+ dm_dirty_log_destroy(rh->log);
if (rh->region_pool)
mempool_destroy(rh->region_pool);
vfree(rh->buckets);
}
}
+static void complete_resync_work(struct region *reg, int success)
+{
+ struct region_hash *rh = reg->rh;
+
+ rh->log->type->set_region_sync(rh->log, reg->key, success);
+
+ /*
+ * Dispatch the bios before we call 'wake_up_all'.
+ * This is important because if we are suspending,
+ * we want to know that recovery is complete and
+ * the work queue is flushed. If we wake_up_all
+ * before we dispatch_bios (queue bios and call wake()),
+ * then we risk suspending before the work queue
+ * has been properly flushed.
+ */
+ dispatch_bios(rh->ms, ®->delayed_bios);
+ if (atomic_dec_and_test(&rh->recovery_in_flight))
+ wake_up_all(&_kmirrord_recovery_stopped);
+ up(&rh->recovery_count);
+}
+
static void rh_update_states(struct region_hash *rh)
{
struct region *reg, *next;
LIST_HEAD(clean);
LIST_HEAD(recovered);
+ LIST_HEAD(failed_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_splice_init(&rh->clean_regions, &clean);
- list_for_each_entry (reg, &clean, list) {
- rh->log->type->clear_region(rh->log, reg->key);
+ list_for_each_entry(reg, &clean, list)
list_del(®->hash_list);
- }
}
if (!list_empty(&rh->recovered_regions)) {
- list_splice(&rh->recovered_regions, &recovered);
- INIT_LIST_HEAD(&rh->recovered_regions);
+ list_splice_init(&rh->recovered_regions, &recovered);
list_for_each_entry (reg, &recovered, list)
list_del(®->hash_list);
}
+
+ if (!list_empty(&rh->failed_recovered_regions)) {
+ list_splice_init(&rh->failed_recovered_regions,
+ &failed_recovered);
+
+ list_for_each_entry(reg, &failed_recovered, list)
+ list_del(®->hash_list);
+ }
+
spin_unlock(&rh->region_lock);
write_unlock_irq(&rh->hash_lock);
*/
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, ®->delayed_bios);
- up(&rh->recovery_count);
+ complete_resync_work(reg, 1);
mempool_free(reg, rh->region_pool);
}
- if (!list_empty(&recovered))
- rh->log->type->flush(rh->log);
+ list_for_each_entry_safe(reg, next, &failed_recovered, list) {
+ complete_resync_work(reg, errors_handled(rh->ms) ? 0 : 1);
+ mempool_free(reg, rh->region_pool);
+ }
- list_for_each_entry_safe (reg, next, &clean, list)
+ list_for_each_entry_safe(reg, next, &clean, list) {
+ rh->log->type->clear_region(rh->log, reg->key);
mempool_free(reg, rh->region_pool);
+ }
+
+ rh->log->type->flush(rh->log);
}
static void rh_inc(struct region_hash *rh, region_t region)
spin_unlock_irqrestore(&rh->region_lock, flags);
if (should_wake)
- wake();
+ wake(rh->ms);
}
/*
static void rh_recovery_prepare(struct region_hash *rh)
{
- while (!down_trylock(&rh->recovery_count))
+ /* Extra reference to avoid race with rh_stop_recovery */
+ atomic_inc(&rh->recovery_in_flight);
+
+ while (!down_trylock(&rh->recovery_count)) {
+ atomic_inc(&rh->recovery_in_flight);
if (__rh_recovery_prepare(rh) <= 0) {
+ atomic_dec(&rh->recovery_in_flight);
up(&rh->recovery_count);
break;
}
+ }
+
+ /* Drop the extra reference */
+ if (atomic_dec_and_test(&rh->recovery_in_flight))
+ wake_up_all(&_kmirrord_recovery_stopped);
}
/*
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(®->list, ®->rh->recovered_regions);
+ if (success)
+ list_add(®->list, ®->rh->recovered_regions);
+ else {
+ reg->state = RH_NOSYNC;
+ list_add(®->list, ®->rh->failed_recovered_regions);
+ }
spin_unlock_irq(&rh->region_lock);
- wake();
+ wake(rh->ms);
}
-static void rh_flush(struct region_hash *rh)
+static int rh_flush(struct region_hash *rh)
{
- rh->log->type->flush(rh->log);
+ return rh->log->type->flush(rh->log);
}
static void rh_delay(struct region_hash *rh, struct bio *bio)
for (i = 0; i < MAX_RECOVERY; i++)
up(&rh->recovery_count);
- wake();
+ wake(rh->ms);
}
+#define MIN_READ_RECORDS 20
+struct dm_raid1_read_record {
+ struct mirror *m;
+ struct dm_bio_details details;
+};
+
/*
* 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
+ * This is yucky. We squirrel the mirror struct away inside
+ * bi_next for read/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)
+static struct mirror *bio_get_m(struct bio *bio)
+{
+ return (struct mirror *) bio->bi_next;
+}
+
+static void bio_set_m(struct bio *bio, struct mirror *m)
+{
+ bio->bi_next = (struct bio *) m;
+}
+
+static struct mirror *get_default_mirror(struct mirror_set *ms)
+{
+ return &ms->mirror[atomic_read(&ms->default_mirror)];
+}
+
+static void set_default_mirror(struct mirror *m)
{
- return (struct mirror_set *) bio->bi_next;
+ struct mirror_set *ms = m->ms;
+ struct mirror *m0 = &(ms->mirror[0]);
+
+ atomic_set(&ms->default_mirror, m - m0);
}
-static void bio_set_ms(struct bio *bio, struct mirror_set *ms)
+/* fail_mirror
+ * @m: mirror device to fail
+ * @error_type: one of the enum's, DM_RAID1_*_ERROR
+ *
+ * If errors are being handled, record the type of
+ * error encountered for this device. If this type
+ * of error has already been recorded, we can return;
+ * otherwise, we must signal userspace by triggering
+ * an event. Additionally, if the device is the
+ * primary device, we must choose a new primary, but
+ * only if the mirror is in-sync.
+ *
+ * This function must not block.
+ */
+static void fail_mirror(struct mirror *m, enum dm_raid1_error error_type)
{
- bio->bi_next = (struct bio *) ms;
+ struct mirror_set *ms = m->ms;
+ struct mirror *new;
+
+ if (!errors_handled(ms))
+ return;
+
+ /*
+ * error_count is used for nothing more than a
+ * simple way to tell if a device has encountered
+ * errors.
+ */
+ atomic_inc(&m->error_count);
+
+ if (test_and_set_bit(error_type, &m->error_type))
+ return;
+
+ if (m != get_default_mirror(ms))
+ goto out;
+
+ if (!ms->in_sync) {
+ /*
+ * Better to issue requests to same failing device
+ * than to risk returning corrupt data.
+ */
+ DMERR("Primary mirror (%s) failed while out-of-sync: "
+ "Reads may fail.", m->dev->name);
+ goto out;
+ }
+
+ for (new = ms->mirror; new < ms->mirror + ms->nr_mirrors; new++)
+ if (!atomic_read(&new->error_count)) {
+ set_default_mirror(new);
+ break;
+ }
+
+ if (unlikely(new == ms->mirror + ms->nr_mirrors))
+ DMWARN("All sides of mirror have failed.");
+
+out:
+ schedule_work(&ms->trigger_event);
}
/*-----------------------------------------------------------------
* 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,
+static void recovery_complete(int read_err, unsigned long write_err,
void *context)
{
- struct region *reg = (struct region *) context;
+ struct region *reg = (struct region *)context;
+ struct mirror_set *ms = reg->rh->ms;
+ int m, bit = 0;
+
+ if (read_err) {
+ /* Read error means the failure of default mirror. */
+ DMERR_LIMIT("Unable to read primary mirror during recovery");
+ fail_mirror(get_default_mirror(ms), DM_RAID1_SYNC_ERROR);
+ }
+
+ if (write_err) {
+ DMERR_LIMIT("Write error during recovery (error = 0x%lx)",
+ write_err);
+ /*
+ * Bits correspond to devices (excluding default mirror).
+ * The default mirror cannot change during recovery.
+ */
+ for (m = 0; m < ms->nr_mirrors; m++) {
+ if (&ms->mirror[m] == get_default_mirror(ms))
+ continue;
+ if (test_bit(bit, &write_err))
+ fail_mirror(ms->mirror + m,
+ DM_RAID1_SYNC_ERROR);
+ bit++;
+ }
+ }
- /* FIXME: better error handling */
rh_recovery_end(reg, !(read_err || write_err));
}
{
int r;
unsigned int i;
- struct io_region from, to[KCOPYD_MAX_REGIONS], *dest;
+ struct dm_io_region from, to[DM_KCOPYD_MAX_REGIONS], *dest;
struct mirror *m;
unsigned long flags = 0;
/* fill in the source */
- m = ms->default_mirror;
+ m = get_default_mirror(ms);
from.bdev = m->dev->bdev;
from.sector = m->offset + region_to_sector(reg->rh, reg->key);
if (reg->key == (ms->nr_regions - 1)) {
/* fill in the destinations */
for (i = 0, dest = to; i < ms->nr_mirrors; i++) {
- if (&ms->mirror[i] == ms->default_mirror)
+ if (&ms->mirror[i] == get_default_mirror(ms))
continue;
m = ms->mirror + i;
}
/* hand to kcopyd */
- set_bit(KCOPYD_IGNORE_ERROR, &flags);
- r = kcopyd_copy(ms->kcopyd_client, &from, ms->nr_mirrors - 1, to, flags,
- recovery_complete, reg);
+ set_bit(DM_KCOPYD_IGNORE_ERROR, &flags);
+ r = dm_kcopyd_copy(ms->kcopyd_client, &from, ms->nr_mirrors - 1, to,
+ flags, recovery_complete, reg);
return r;
}
{
int r;
struct region *reg;
- struct dirty_log *log = ms->rh.log;
+ struct dm_dirty_log *log = ms->rh.log;
/*
* Start quiescing some regions.
*---------------------------------------------------------------*/
static struct mirror *choose_mirror(struct mirror_set *ms, sector_t sector)
{
- /* FIXME: add read balancing */
- return ms->default_mirror;
+ struct mirror *m = get_default_mirror(ms);
+
+ do {
+ if (likely(!atomic_read(&m->error_count)))
+ return m;
+
+ if (m-- == ms->mirror)
+ m += ms->nr_mirrors;
+ } while (m != get_default_mirror(ms));
+
+ return NULL;
+}
+
+static int default_ok(struct mirror *m)
+{
+ struct mirror *default_mirror = get_default_mirror(m->ms);
+
+ return !atomic_read(&default_mirror->error_count);
+}
+
+static int mirror_available(struct mirror_set *ms, struct bio *bio)
+{
+ region_t region = bio_to_region(&ms->rh, bio);
+
+ if (ms->rh.log->type->in_sync(ms->rh.log, region, 0))
+ return choose_mirror(ms, bio->bi_sector) ? 1 : 0;
+
+ return 0;
}
/*
* remap a buffer to a particular mirror.
*/
-static void map_bio(struct mirror_set *ms, struct mirror *m, struct bio *bio)
+static sector_t map_sector(struct mirror *m, struct bio *bio)
+{
+ return m->offset + (bio->bi_sector - m->ms->ti->begin);
+}
+
+static void map_bio(struct mirror *m, struct bio *bio)
{
bio->bi_bdev = m->dev->bdev;
- bio->bi_sector = m->offset + (bio->bi_sector - ms->ti->begin);
+ bio->bi_sector = map_sector(m, bio);
+}
+
+static void map_region(struct dm_io_region *io, struct mirror *m,
+ struct bio *bio)
+{
+ io->bdev = m->dev->bdev;
+ io->sector = map_sector(m, bio);
+ io->count = bio->bi_size >> 9;
+}
+
+/*-----------------------------------------------------------------
+ * Reads
+ *---------------------------------------------------------------*/
+static void read_callback(unsigned long error, void *context)
+{
+ struct bio *bio = context;
+ struct mirror *m;
+
+ m = bio_get_m(bio);
+ bio_set_m(bio, NULL);
+
+ if (likely(!error)) {
+ bio_endio(bio, 0);
+ return;
+ }
+
+ fail_mirror(m, DM_RAID1_READ_ERROR);
+
+ if (likely(default_ok(m)) || mirror_available(m->ms, bio)) {
+ DMWARN_LIMIT("Read failure on mirror device %s. "
+ "Trying alternative device.",
+ m->dev->name);
+ queue_bio(m->ms, bio, bio_rw(bio));
+ return;
+ }
+
+ DMERR_LIMIT("Read failure on mirror device %s. Failing I/O.",
+ m->dev->name);
+ bio_endio(bio, -EIO);
+}
+
+/* Asynchronous read. */
+static void read_async_bio(struct mirror *m, struct bio *bio)
+{
+ struct dm_io_region io;
+ struct dm_io_request io_req = {
+ .bi_rw = READ,
+ .mem.type = DM_IO_BVEC,
+ .mem.ptr.bvec = bio->bi_io_vec + bio->bi_idx,
+ .notify.fn = read_callback,
+ .notify.context = bio,
+ .client = m->ms->io_client,
+ };
+
+ map_region(&io, m, bio);
+ bio_set_m(bio, m);
+ (void) dm_io(&io_req, 1, &io, NULL);
}
static void do_reads(struct mirror_set *ms, struct bio_list *reads)
while ((bio = bio_list_pop(reads))) {
region = bio_to_region(&ms->rh, bio);
+ m = get_default_mirror(ms);
/*
* We can only read balance if the region is in sync.
*/
- if (rh_in_sync(&ms->rh, region, 0))
+ if (likely(rh_in_sync(&ms->rh, region, 1)))
m = choose_mirror(ms, bio->bi_sector);
- else
- m = ms->default_mirror;
+ else if (m && atomic_read(&m->error_count))
+ m = NULL;
- map_bio(ms, m, bio);
- generic_make_request(bio);
+ if (likely(m))
+ read_async_bio(m, bio);
+ else
+ bio_endio(bio, -EIO);
}
}
* RECOVERING: delay the io until recovery completes
* NOSYNC: increment pending, just write to the default mirror
*---------------------------------------------------------------*/
+
+/* __bio_mark_nosync
+ * @ms
+ * @bio
+ * @done
+ * @error
+ *
+ * The bio was written on some mirror(s) but failed on other mirror(s).
+ * We can successfully endio the bio but should avoid the region being
+ * marked clean by setting the state RH_NOSYNC.
+ *
+ * This function is _not_ safe in interrupt context!
+ */
+static void __bio_mark_nosync(struct mirror_set *ms,
+ struct bio *bio, unsigned done, int error)
+{
+ unsigned long flags;
+ struct region_hash *rh = &ms->rh;
+ struct dm_dirty_log *log = ms->rh.log;
+ struct region *reg;
+ region_t region = bio_to_region(rh, bio);
+ int recovering = 0;
+
+ /* We must inform the log that the sync count has changed. */
+ log->type->set_region_sync(log, region, 0);
+ ms->in_sync = 0;
+
+ read_lock(&rh->hash_lock);
+ reg = __rh_find(rh, region);
+ read_unlock(&rh->hash_lock);
+
+ /* region hash entry should exist because write was in-flight */
+ BUG_ON(!reg);
+ BUG_ON(!list_empty(®->list));
+
+ spin_lock_irqsave(&rh->region_lock, flags);
+ /*
+ * Possible cases:
+ * 1) RH_DIRTY
+ * 2) RH_NOSYNC: was dirty, other preceeding writes failed
+ * 3) RH_RECOVERING: flushing pending writes
+ * Either case, the region should have not been connected to list.
+ */
+ recovering = (reg->state == RH_RECOVERING);
+ reg->state = RH_NOSYNC;
+ BUG_ON(!list_empty(®->list));
+ spin_unlock_irqrestore(&rh->region_lock, flags);
+
+ bio_endio(bio, error);
+ if (recovering)
+ complete_resync_work(reg, 0);
+}
+
static void write_callback(unsigned long error, void *context)
{
- unsigned int i;
- int uptodate = 1;
+ unsigned i, ret = 0;
struct bio *bio = (struct bio *) context;
struct mirror_set *ms;
+ int uptodate = 0;
+ int should_wake = 0;
+ unsigned long flags;
- ms = bio_get_ms(bio);
- bio_set_ms(bio, NULL);
+ ms = bio_get_m(bio)->ms;
+ bio_set_m(bio, NULL);
/*
* NOTE: We don't decrement the pending count here,
* This way we handle both writes to SYNC and NOSYNC
* regions with the same code.
*/
+ if (likely(!error))
+ goto out;
+
+ for (i = 0; i < ms->nr_mirrors; i++)
+ if (test_bit(i, &error))
+ fail_mirror(ms->mirror + i, DM_RAID1_WRITE_ERROR);
+ else
+ uptodate = 1;
- if (error) {
+ if (unlikely(!uptodate)) {
+ DMERR("All replicated volumes dead, failing I/O");
+ /* None of the writes succeeded, fail the I/O. */
+ ret = -EIO;
+ } else if (errors_handled(ms)) {
/*
- * only error the io if all mirrors failed.
- * FIXME: bogus
+ * Need to raise event. Since raising
+ * events can block, we need to do it in
+ * the main thread.
*/
- uptodate = 0;
- for (i = 0; i < ms->nr_mirrors; i++)
- if (!test_bit(i, &error)) {
- uptodate = 1;
- break;
- }
+ spin_lock_irqsave(&ms->lock, flags);
+ if (!ms->failures.head)
+ should_wake = 1;
+ bio_list_add(&ms->failures, bio);
+ spin_unlock_irqrestore(&ms->lock, flags);
+ if (should_wake)
+ wake(ms);
+ return;
}
- bio_endio(bio, bio->bi_size, 0);
+out:
+ bio_endio(bio, ret);
}
static void do_write(struct mirror_set *ms, struct bio *bio)
{
unsigned int i;
- struct io_region io[KCOPYD_MAX_REGIONS+1];
+ struct dm_io_region io[ms->nr_mirrors], *dest = io;
struct mirror *m;
+ struct dm_io_request io_req = {
+ .bi_rw = WRITE,
+ .mem.type = DM_IO_BVEC,
+ .mem.ptr.bvec = bio->bi_io_vec + bio->bi_idx,
+ .notify.fn = write_callback,
+ .notify.context = bio,
+ .client = ms->io_client,
+ };
+
+ for (i = 0, m = ms->mirror; i < ms->nr_mirrors; i++, m++)
+ map_region(dest++, m, bio);
- for (i = 0; i < ms->nr_mirrors; i++) {
- m = ms->mirror + i;
-
- io[i].bdev = m->dev->bdev;
- io[i].sector = m->offset + (bio->bi_sector - ms->ti->begin);
- io[i].count = bio->bi_size >> 9;
- }
+ /*
+ * Use default mirror because we only need it to retrieve the reference
+ * to the mirror set in write_callback().
+ */
+ bio_set_m(bio, get_default_mirror(ms));
- bio_set_ms(bio, ms);
- dm_io_async_bvec(ms->nr_mirrors, io, WRITE,
- bio->bi_io_vec + bio->bi_idx,
- write_callback, bio);
+ (void) dm_io(&io_req, ms->nr_mirrors, io, NULL);
}
static void do_writes(struct mirror_set *ms, struct bio_list *writes)
*/
rh_inc_pending(&ms->rh, &sync);
rh_inc_pending(&ms->rh, &nosync);
- rh_flush(&ms->rh);
+ ms->log_failure = rh_flush(&ms->rh) ? 1 : 0;
/*
* Dispatch io.
*/
- while ((bio = bio_list_pop(&sync)))
- do_write(ms, bio);
+ if (unlikely(ms->log_failure)) {
+ spin_lock_irq(&ms->lock);
+ bio_list_merge(&ms->failures, &sync);
+ spin_unlock_irq(&ms->lock);
+ wake(ms);
+ } else
+ while ((bio = bio_list_pop(&sync)))
+ do_write(ms, bio);
while ((bio = bio_list_pop(&recover)))
rh_delay(&ms->rh, bio);
while ((bio = bio_list_pop(&nosync))) {
- map_bio(ms, ms->default_mirror, bio);
+ map_bio(get_default_mirror(ms), bio);
generic_make_request(bio);
}
}
+static void do_failures(struct mirror_set *ms, struct bio_list *failures)
+{
+ struct bio *bio;
+
+ if (!failures->head)
+ return;
+
+ if (!ms->log_failure) {
+ while ((bio = bio_list_pop(failures)))
+ __bio_mark_nosync(ms, bio, bio->bi_size, 0);
+ return;
+ }
+
+ /*
+ * If the log has failed, unattempted writes are being
+ * put on the failures list. We can't issue those writes
+ * until a log has been marked, so we must store them.
+ *
+ * If a 'noflush' suspend is in progress, we can requeue
+ * the I/O's to the core. This give userspace a chance
+ * to reconfigure the mirror, at which point the core
+ * will reissue the writes. If the 'noflush' flag is
+ * not set, we have no choice but to return errors.
+ *
+ * Some writes on the failures list may have been
+ * submitted before the log failure and represent a
+ * failure to write to one of the devices. It is ok
+ * for us to treat them the same and requeue them
+ * as well.
+ */
+ if (dm_noflush_suspending(ms->ti)) {
+ while ((bio = bio_list_pop(failures)))
+ bio_endio(bio, DM_ENDIO_REQUEUE);
+ return;
+ }
+
+ if (atomic_read(&ms->suspend)) {
+ while ((bio = bio_list_pop(failures)))
+ bio_endio(bio, -EIO);
+ return;
+ }
+
+ spin_lock_irq(&ms->lock);
+ bio_list_merge(&ms->failures, failures);
+ spin_unlock_irq(&ms->lock);
+
+ delayed_wake(ms);
+}
+
+static void trigger_event(struct work_struct *work)
+{
+ struct mirror_set *ms =
+ container_of(work, struct mirror_set, trigger_event);
+
+ dm_table_event(ms->ti->table);
+}
+
/*-----------------------------------------------------------------
* kmirrord
*---------------------------------------------------------------*/
-static LIST_HEAD(_mirror_sets);
-static DECLARE_RWSEM(_mirror_sets_lock);
-
-static void do_mirror(struct mirror_set *ms)
+static void do_mirror(struct work_struct *work)
{
- struct bio_list reads, writes;
+ struct mirror_set *ms =container_of(work, struct mirror_set,
+ kmirrord_work);
+ struct bio_list reads, writes, failures;
+ unsigned long flags;
- spin_lock(&ms->lock);
+ spin_lock_irqsave(&ms->lock, flags);
reads = ms->reads;
writes = ms->writes;
+ failures = ms->failures;
bio_list_init(&ms->reads);
bio_list_init(&ms->writes);
- spin_unlock(&ms->lock);
+ bio_list_init(&ms->failures);
+ spin_unlock_irqrestore(&ms->lock, flags);
rh_update_states(&ms->rh);
do_recovery(ms);
do_reads(ms, &reads);
do_writes(ms, &writes);
-}
-
-static void do_work(void *ignored)
-{
- struct mirror_set *ms;
+ do_failures(ms, &failures);
- down_read(&_mirror_sets_lock);
- list_for_each_entry (ms, &_mirror_sets, list)
- do_mirror(ms);
- up_read(&_mirror_sets_lock);
+ dm_table_unplug_all(ms->ti->table);
}
+
/*-----------------------------------------------------------------
* Target functions
*---------------------------------------------------------------*/
static struct mirror_set *alloc_context(unsigned int nr_mirrors,
uint32_t region_size,
struct dm_target *ti,
- struct dirty_log *dl)
+ struct dm_dirty_log *dl)
{
size_t len;
struct mirror_set *ms = NULL;
len = sizeof(*ms) + (sizeof(ms->mirror[0]) * nr_mirrors);
- ms = kmalloc(len, GFP_KERNEL);
+ ms = kzalloc(len, GFP_KERNEL);
if (!ms) {
ti->error = "Cannot allocate mirror context";
return NULL;
}
- memset(ms, 0, len);
spin_lock_init(&ms->lock);
ms->ti = ti;
ms->nr_mirrors = nr_mirrors;
ms->nr_regions = dm_sector_div_up(ti->len, region_size);
ms->in_sync = 0;
- ms->default_mirror = &ms->mirror[DEFAULT_MIRROR];
+ ms->log_failure = 0;
+ atomic_set(&ms->suspend, 0);
+ atomic_set(&ms->default_mirror, DEFAULT_MIRROR);
+
+ len = sizeof(struct dm_raid1_read_record);
+ ms->read_record_pool = mempool_create_kmalloc_pool(MIN_READ_RECORDS,
+ len);
+ if (!ms->read_record_pool) {
+ ti->error = "Error creating mirror read_record_pool";
+ kfree(ms);
+ return NULL;
+ }
+
+ ms->io_client = dm_io_client_create(DM_IO_PAGES);
+ if (IS_ERR(ms->io_client)) {
+ ti->error = "Error creating dm_io client";
+ mempool_destroy(ms->read_record_pool);
+ kfree(ms);
+ return NULL;
+ }
if (rh_init(&ms->rh, ms, dl, region_size, ms->nr_regions)) {
ti->error = "Error creating dirty region hash";
+ dm_io_client_destroy(ms->io_client);
+ mempool_destroy(ms->read_record_pool);
kfree(ms);
return NULL;
}
while (m--)
dm_put_device(ti, ms->mirror[m].dev);
+ dm_io_client_destroy(ms->io_client);
rh_exit(&ms->rh);
+ mempool_destroy(ms->read_record_pool);
kfree(ms);
}
static inline int _check_region_size(struct dm_target *ti, uint32_t size)
{
- return !(size % (PAGE_SIZE >> 9) || (size & (size - 1)) ||
+ return !(size % (PAGE_SIZE >> 9) || !is_power_of_2(size) ||
size > ti->len);
}
return -ENXIO;
}
+ ms->mirror[mirror].ms = ms;
+ atomic_set(&(ms->mirror[mirror].error_count), 0);
+ ms->mirror[mirror].error_type = 0;
ms->mirror[mirror].offset = offset;
return 0;
}
-static int add_mirror_set(struct mirror_set *ms)
-{
- down_write(&_mirror_sets_lock);
- list_add_tail(&ms->list, &_mirror_sets);
- up_write(&_mirror_sets_lock);
- wake();
-
- return 0;
-}
-
-static void del_mirror_set(struct mirror_set *ms)
-{
- down_write(&_mirror_sets_lock);
- list_del(&ms->list);
- up_write(&_mirror_sets_lock);
-}
-
/*
* Create dirty log: log_type #log_params <log_params>
*/
-static struct dirty_log *create_dirty_log(struct dm_target *ti,
+static struct dm_dirty_log *create_dirty_log(struct dm_target *ti,
unsigned int argc, char **argv,
unsigned int *args_used)
{
unsigned int param_count;
- struct dirty_log *dl;
+ struct dm_dirty_log *dl;
if (argc < 2) {
ti->error = "Insufficient mirror log arguments";
return NULL;
}
- dl = dm_create_dirty_log(argv[0], ti, param_count, argv + 2);
+ dl = dm_dirty_log_create(argv[0], ti, param_count, argv + 2);
if (!dl) {
ti->error = "Error creating mirror dirty log";
return NULL;
if (!_check_region_size(ti, dl->type->get_region_size(dl))) {
ti->error = "Invalid region size";
- dm_destroy_dirty_log(dl);
+ dm_dirty_log_destroy(dl);
return NULL;
}
return dl;
}
+static int parse_features(struct mirror_set *ms, unsigned argc, char **argv,
+ unsigned *args_used)
+{
+ unsigned num_features;
+ struct dm_target *ti = ms->ti;
+
+ *args_used = 0;
+
+ if (!argc)
+ return 0;
+
+ if (sscanf(argv[0], "%u", &num_features) != 1) {
+ ti->error = "Invalid number of features";
+ return -EINVAL;
+ }
+
+ argc--;
+ argv++;
+ (*args_used)++;
+
+ if (num_features > argc) {
+ ti->error = "Not enough arguments to support feature count";
+ return -EINVAL;
+ }
+
+ if (!strcmp("handle_errors", argv[0]))
+ ms->features |= DM_RAID1_HANDLE_ERRORS;
+ else {
+ ti->error = "Unrecognised feature requested";
+ return -EINVAL;
+ }
+
+ (*args_used)++;
+
+ return 0;
+}
+
/*
* Construct a mirror mapping:
*
* log_type #log_params <log_params>
* #mirrors [mirror_path offset]{2,}
+ * [#features <features>]
*
* log_type is "core" or "disk"
* #log_params is between 1 and 3
+ *
+ * If present, features must be "handle_errors".
*/
-#define DM_IO_PAGES 64
static int mirror_ctr(struct dm_target *ti, unsigned int argc, char **argv)
{
int r;
unsigned int nr_mirrors, m, args_used;
struct mirror_set *ms;
- struct dirty_log *dl;
+ struct dm_dirty_log *dl;
dl = create_dirty_log(ti, argc, argv, &args_used);
if (!dl)
argc -= args_used;
if (!argc || sscanf(argv[0], "%u", &nr_mirrors) != 1 ||
- nr_mirrors < 2 || nr_mirrors > KCOPYD_MAX_REGIONS + 1) {
+ nr_mirrors < 2 || nr_mirrors > DM_KCOPYD_MAX_REGIONS + 1) {
ti->error = "Invalid number of mirrors";
- dm_destroy_dirty_log(dl);
+ dm_dirty_log_destroy(dl);
return -EINVAL;
}
argv++, argc--;
- if (argc != nr_mirrors * 2) {
- ti->error = "Wrong number of mirror arguments";
- dm_destroy_dirty_log(dl);
+ if (argc < nr_mirrors * 2) {
+ ti->error = "Too few mirror arguments";
+ dm_dirty_log_destroy(dl);
return -EINVAL;
}
ms = alloc_context(nr_mirrors, dl->type->get_region_size(dl), ti, dl);
if (!ms) {
- dm_destroy_dirty_log(dl);
+ dm_dirty_log_destroy(dl);
return -ENOMEM;
}
ti->private = ms;
ti->split_io = ms->rh.region_size;
- r = kcopyd_client_create(DM_IO_PAGES, &ms->kcopyd_client);
- if (r) {
- free_context(ms, ti, ms->nr_mirrors);
- return r;
+ ms->kmirrord_wq = create_singlethread_workqueue("kmirrord");
+ if (!ms->kmirrord_wq) {
+ DMERR("couldn't start kmirrord");
+ r = -ENOMEM;
+ goto err_free_context;
}
+ INIT_WORK(&ms->kmirrord_work, do_mirror);
+ init_timer(&ms->timer);
+ ms->timer_pending = 0;
+ INIT_WORK(&ms->trigger_event, trigger_event);
+
+ r = parse_features(ms, argc, argv, &args_used);
+ if (r)
+ goto err_destroy_wq;
+
+ argv += args_used;
+ argc -= args_used;
- add_mirror_set(ms);
+ /*
+ * Any read-balancing addition depends on the
+ * DM_RAID1_HANDLE_ERRORS flag being present.
+ * This is because the decision to balance depends
+ * on the sync state of a region. If the above
+ * flag is not present, we ignore errors; and
+ * the sync state may be inaccurate.
+ */
+
+ if (argc) {
+ ti->error = "Too many mirror arguments";
+ r = -EINVAL;
+ goto err_destroy_wq;
+ }
+
+ r = dm_kcopyd_client_create(DM_IO_PAGES, &ms->kcopyd_client);
+ if (r)
+ goto err_destroy_wq;
+
+ wake(ms);
return 0;
+
+err_destroy_wq:
+ destroy_workqueue(ms->kmirrord_wq);
+err_free_context:
+ free_context(ms, ti, ms->nr_mirrors);
+ return r;
}
static void mirror_dtr(struct dm_target *ti)
{
struct mirror_set *ms = (struct mirror_set *) ti->private;
- del_mirror_set(ms);
- kcopyd_client_destroy(ms->kcopyd_client);
+ del_timer_sync(&ms->timer);
+ flush_workqueue(ms->kmirrord_wq);
+ dm_kcopyd_client_destroy(ms->kcopyd_client);
+ destroy_workqueue(ms->kmirrord_wq);
free_context(ms, ti, ms->nr_mirrors);
}
static void queue_bio(struct mirror_set *ms, struct bio *bio, int rw)
{
+ unsigned long flags;
int should_wake = 0;
struct bio_list *bl;
bl = (rw == WRITE) ? &ms->writes : &ms->reads;
- spin_lock(&ms->lock);
+ spin_lock_irqsave(&ms->lock, flags);
should_wake = !(bl->head);
bio_list_add(bl, bio);
- spin_unlock(&ms->lock);
+ spin_unlock_irqrestore(&ms->lock, flags);
if (should_wake)
- wake();
+ wake(ms);
}
/*
int r, rw = bio_rw(bio);
struct mirror *m;
struct mirror_set *ms = ti->private;
-
- map_context->ll = bio_to_region(&ms->rh, bio);
+ struct dm_raid1_read_record *read_record = NULL;
if (rw == WRITE) {
+ /* Save region for mirror_end_io() handler */
+ map_context->ll = bio_to_region(&ms->rh, bio);
queue_bio(ms, bio, rw);
- return 0;
+ return DM_MAPIO_SUBMITTED;
}
r = ms->rh.log->type->in_sync(ms->rh.log,
if (r < 0 && r != -EWOULDBLOCK)
return r;
- if (r == -EWOULDBLOCK) /* FIXME: ugly */
- r = 0;
-
/*
- * We don't want to fast track a recovery just for a read
- * ahead. So we just let it silently fail.
- * FIXME: get rid of this.
+ * If region is not in-sync queue the bio.
*/
- if (!r && rw == READA)
- return -EIO;
+ if (!r || (r == -EWOULDBLOCK)) {
+ if (rw == READA)
+ return -EWOULDBLOCK;
- if (!r) {
- /* Pass this io over to the daemon */
queue_bio(ms, bio, rw);
- return 0;
+ return DM_MAPIO_SUBMITTED;
}
+ /*
+ * The region is in-sync and we can perform reads directly.
+ * Store enough information so we can retry if it fails.
+ */
m = choose_mirror(ms, bio->bi_sector);
- if (!m)
+ if (unlikely(!m))
return -EIO;
- map_bio(ms, m, bio);
- return 1;
+ read_record = mempool_alloc(ms->read_record_pool, GFP_NOIO);
+ if (likely(read_record)) {
+ dm_bio_record(&read_record->details, bio);
+ map_context->ptr = read_record;
+ read_record->m = m;
+ }
+
+ map_bio(m, bio);
+
+ return DM_MAPIO_REMAPPED;
}
static int mirror_end_io(struct dm_target *ti, struct bio *bio,
{
int rw = bio_rw(bio);
struct mirror_set *ms = (struct mirror_set *) ti->private;
- region_t region = map_context->ll;
+ struct mirror *m = NULL;
+ struct dm_bio_details *bd = NULL;
+ struct dm_raid1_read_record *read_record = map_context->ptr;
/*
* We need to dec pending if this was a write.
*/
- if (rw == WRITE)
- rh_dec(&ms->rh, region);
+ if (rw == WRITE) {
+ rh_dec(&ms->rh, map_context->ll);
+ return error;
+ }
- return 0;
+ if (error == -EOPNOTSUPP)
+ goto out;
+
+ if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
+ goto out;
+
+ if (unlikely(error)) {
+ if (!read_record) {
+ /*
+ * There wasn't enough memory to record necessary
+ * information for a retry or there was no other
+ * mirror in-sync.
+ */
+ DMERR_LIMIT("Mirror read failed.");
+ return -EIO;
+ }
+
+ m = read_record->m;
+
+ DMERR("Mirror read failed from %s. Trying alternative device.",
+ m->dev->name);
+
+ fail_mirror(m, DM_RAID1_READ_ERROR);
+
+ /*
+ * A failed read is requeued for another attempt using an intact
+ * mirror.
+ */
+ if (default_ok(m) || mirror_available(ms, bio)) {
+ bd = &read_record->details;
+
+ dm_bio_restore(bd, bio);
+ mempool_free(read_record, ms->read_record_pool);
+ map_context->ptr = NULL;
+ queue_bio(ms, bio, rw);
+ return 1;
+ }
+ DMERR("All replicated volumes dead, failing I/O");
+ }
+
+out:
+ if (read_record) {
+ mempool_free(read_record, ms->read_record_pool);
+ map_context->ptr = NULL;
+ }
+
+ return error;
}
-static void mirror_postsuspend(struct dm_target *ti)
+static void mirror_presuspend(struct dm_target *ti)
{
struct mirror_set *ms = (struct mirror_set *) ti->private;
- struct dirty_log *log = ms->rh.log;
+ struct dm_dirty_log *log = ms->rh.log;
+
+ atomic_set(&ms->suspend, 1);
+ /*
+ * We must finish up all the work that we've
+ * generated (i.e. recovery work).
+ */
rh_stop_recovery(&ms->rh);
- if (log->type->suspend && log->type->suspend(log))
+
+ wait_event(_kmirrord_recovery_stopped,
+ !atomic_read(&ms->rh.recovery_in_flight));
+
+ if (log->type->presuspend && log->type->presuspend(log))
/* FIXME: need better error handling */
- DMWARN("log suspend failed");
+ DMWARN("log presuspend failed");
+
+ /*
+ * Now that recovery is complete/stopped and the
+ * delayed bios are queued, we need to wait for
+ * the worker thread to complete. This way,
+ * we know that all of our I/O has been pushed.
+ */
+ flush_workqueue(ms->kmirrord_wq);
+}
+
+static void mirror_postsuspend(struct dm_target *ti)
+{
+ struct mirror_set *ms = ti->private;
+ struct dm_dirty_log *log = ms->rh.log;
+
+ if (log->type->postsuspend && log->type->postsuspend(log))
+ /* FIXME: need better error handling */
+ DMWARN("log postsuspend failed");
}
static void mirror_resume(struct dm_target *ti)
{
- struct mirror_set *ms = (struct mirror_set *) ti->private;
- struct dirty_log *log = ms->rh.log;
+ struct mirror_set *ms = ti->private;
+ struct dm_dirty_log *log = ms->rh.log;
+
+ atomic_set(&ms->suspend, 0);
if (log->type->resume && log->type->resume(log))
/* FIXME: need better error handling */
DMWARN("log resume failed");
rh_start_recovery(&ms->rh);
}
+/*
+ * device_status_char
+ * @m: mirror device/leg we want the status of
+ *
+ * We return one character representing the most severe error
+ * we have encountered.
+ * A => Alive - No failures
+ * D => Dead - A write failure occurred leaving mirror out-of-sync
+ * S => Sync - A sychronization failure occurred, mirror out-of-sync
+ * R => Read - A read failure occurred, mirror data unaffected
+ *
+ * Returns: <char>
+ */
+static char device_status_char(struct mirror *m)
+{
+ if (!atomic_read(&(m->error_count)))
+ return 'A';
+
+ return (test_bit(DM_RAID1_WRITE_ERROR, &(m->error_type))) ? 'D' :
+ (test_bit(DM_RAID1_SYNC_ERROR, &(m->error_type))) ? 'S' :
+ (test_bit(DM_RAID1_READ_ERROR, &(m->error_type))) ? 'R' : 'U';
+}
+
+
static int mirror_status(struct dm_target *ti, status_type_t type,
char *result, unsigned int maxlen)
{
- unsigned int m, sz;
+ unsigned int m, sz = 0;
struct mirror_set *ms = (struct mirror_set *) ti->private;
-
- sz = ms->rh.log->type->status(ms->rh.log, type, result, maxlen);
+ struct dm_dirty_log *log = ms->rh.log;
+ char buffer[ms->nr_mirrors + 1];
switch (type) {
case STATUSTYPE_INFO:
DMEMIT("%d ", ms->nr_mirrors);
- for (m = 0; m < ms->nr_mirrors; m++)
+ for (m = 0; m < ms->nr_mirrors; m++) {
DMEMIT("%s ", ms->mirror[m].dev->name);
+ buffer[m] = device_status_char(&(ms->mirror[m]));
+ }
+ buffer[m] = '\0';
+
+ DMEMIT("%llu/%llu 1 %s ",
+ (unsigned long long)log->type->get_sync_count(ms->rh.log),
+ (unsigned long long)ms->nr_regions, buffer);
+
+ sz += log->type->status(ms->rh.log, type, result+sz, maxlen-sz);
- DMEMIT("%llu/%llu",
- (unsigned long long)ms->rh.log->type->
- get_sync_count(ms->rh.log),
- (unsigned long long)ms->nr_regions);
break;
case STATUSTYPE_TABLE:
- DMEMIT("%d ", ms->nr_mirrors);
+ sz = log->type->status(ms->rh.log, type, result, maxlen);
+
+ DMEMIT("%d", ms->nr_mirrors);
for (m = 0; m < ms->nr_mirrors; m++)
- DMEMIT("%s %llu ", ms->mirror[m].dev->name,
- (unsigned long long)ms->mirror[m].offset);
+ DMEMIT(" %s %llu", ms->mirror[m].dev->name,
+ (unsigned long long)ms->mirror[m].offset);
+
+ if (ms->features & DM_RAID1_HANDLE_ERRORS)
+ DMEMIT(" 1 handle_errors");
}
return 0;
static struct target_type mirror_target = {
.name = "mirror",
- .version = {1, 0, 2},
+ .version = {1, 0, 20},
.module = THIS_MODULE,
.ctr = mirror_ctr,
.dtr = mirror_dtr,
.map = mirror_map,
.end_io = mirror_end_io,
+ .presuspend = mirror_presuspend,
.postsuspend = mirror_postsuspend,
.resume = mirror_resume,
.status = mirror_status,
{
int r;
- r = dm_dirty_log_init();
- if (r)
- return r;
-
- _kmirrord_wq = create_singlethread_workqueue("kmirrord");
- if (!_kmirrord_wq) {
- DMERR("couldn't start kmirrord");
- dm_dirty_log_exit();
- return r;
- }
- INIT_WORK(&_kmirrord_work, do_work, NULL);
-
r = dm_register_target(&mirror_target);
- if (r < 0) {
- DMERR("%s: Failed to register mirror target",
- mirror_target.name);
- dm_dirty_log_exit();
- destroy_workqueue(_kmirrord_wq);
- }
+ if (r < 0)
+ DMERR("Failed to register mirror target");
return r;
}
r = dm_unregister_target(&mirror_target);
if (r < 0)
- DMERR("%s: unregister failed %d", mirror_target.name, r);
-
- destroy_workqueue(_kmirrord_wq);
- dm_dirty_log_exit();
+ DMERR("unregister failed %d", r);
}
/* Module hooks */
Code Review / linux-2.6.git / blobdiff