Add 'resume' to exception store API.

This is the first step to replace 'read_metadata' with 'resume'.
'resume' will populate the exception store implementation's
exception cache - rather than loading the snapshot's exception
cache.

Relocating the 'resume' is more in line with the way device-mapper
targets (and device-mapper mirror logs) work. Additionally, with
future shared exception store types, we don't duplicate the cache.

RFC-by: Jonathan Brassow <jbrassow@redhat.com>

Index: linux-2.6/drivers/md/dm-exception-store.h
================================================== =================
--- linux-2.6.orig/drivers/md/dm-exception-store.h
+++ linux-2.6/drivers/md/dm-exception-store.h
@@ -31,6 +31,8 @@ struct dm_exception_store_type {
*/
void (*dtr) (struct dm_exception_store *store);

+ int (*resume) (struct dm_exception_store *store);
+
/*
* The target shouldn't read the COW device until this is
* called. As exceptions are read from the COW, they are
Index: linux-2.6/drivers/md/dm-snap-persistent.c
================================================== =================
--- linux-2.6.orig/drivers/md/dm-snap-persistent.c
+++ linux-2.6/drivers/md/dm-snap-persistent.c
@@ -89,6 +89,7 @@ struct commit_callback {
*/
struct pstore {
struct dm_exception_store *store;
+ struct dm_exception_table *table;
int version;
int valid;
uint32_t exceptions_per_area;
@@ -130,6 +131,24 @@ struct pstore {
struct workqueue_struct *metadata_wq;
};

+static struct kmem_cache *exception_cache;
+
+static struct dm_exception *alloc_exception(void *unused)
+{
+ struct dm_exception *e;
+
+ e = kmem_cache_alloc(exception_cache, GFP_NOIO);
+ if (!e)
+ e = kmem_cache_alloc(exception_cache, GFP_ATOMIC);
+
+ return e;
+}
+
+static void free_exception(struct dm_exception *e, void *unused)
+{
+ kmem_cache_free(exception_cache, e);
+}
+
static unsigned sectors_to_pages(unsigned sectors)
{
return DIV_ROUND_UP(sectors, PAGE_SIZE >> 9);
@@ -279,7 +298,7 @@ static int read_header(struct pstore *ps
*/
if (!ps->store->chunk_size) {
ps->store->chunk_size = max(DM_CHUNK_SIZE_DEFAULT_SECTORS,
- bdev_hardsect_size(ps->store->cow->bdev) >> 9);
+ bdev_hardsect_size(ps->store->cow->bdev) >> 9);
ps->store->chunk_mask = ps->store->chunk_size - 1;
ps->store->chunk_shift = ffs(ps->store->chunk_size) - 1;
chunk_size_supplied = 0;
@@ -485,6 +504,7 @@ static void persistent_dtr(struct dm_exc
dm_io_client_destroy(ps->io_client);
vfree(ps->callbacks);
free_area(ps);
+ dm_exception_table_destroy(ps->table);
kfree(ps);
}

@@ -496,6 +516,13 @@ static int persistent_read_metadata(stru
int r, uninitialized_var(new_snapshot);
struct pstore *ps = get_info(store);

+ if (ps->callbacks)
+ /*
+ * Temporary work around for having two different functions
+ * that get us going... 'read_metadata' and 'resume'.
+ */
+ goto read_metadata;
+
/*
* Read the snapshot header.
*/
@@ -507,7 +534,7 @@ static int persistent_read_metadata(stru
* Now we know correct chunk_size, complete the initialisation.
*/
ps->exceptions_per_area = (ps->store->chunk_size << SECTOR_SHIFT) /
- sizeof(struct disk_exception);
+ sizeof(struct disk_exception);
ps->callbacks = dm_vcalloc(ps->exceptions_per_area,
sizeof(*ps->callbacks));
if (!ps->callbacks)
@@ -548,11 +575,44 @@ static int persistent_read_metadata(stru
/*
* Read the metadata.
*/
+read_metadata:
r = read_exceptions(ps, callback, callback_context);

return r;
}

+/* This function is temporary for patch cleanliness */
+static int add_exception(void *context, chunk_t old, chunk_t new)
+{
+ struct dm_exception_store *store = context;
+ struct pstore *ps = get_info(store);
+ struct dm_exception *e;
+
+ e = dm_alloc_exception(ps->table);
+ if (!e)
+ return -ENOMEM;
+
+ e->old_chunk = old;
+ e->new_chunk = new;
+
+ dm_insert_exception(ps->table, e);
+
+ return 0;
+}
+
+/*
+ * persistent_resume
+ * @store
+ *
+ * Read metadata of the disk and store in our exception table cache.
+ *
+ * Returns: 0 on success, -Exxx on error
+ */
+static int persistent_resume(struct dm_exception_store *store)
+{
+ return persistent_read_metadata(store, add_exception, store);
+}
+
static int persistent_prepare_exception(struct dm_exception_store *store,
struct dm_exception *e)
{
@@ -655,6 +715,7 @@ static int persistent_ctr(struct dm_exce
unsigned argc, char **argv)
{
struct pstore *ps;
+ sector_t hash_size, cow_dev_size, max_buckets;

/* allocate the pstore */
ps = kmalloc(sizeof(*ps), GFP_KERNEL);
@@ -672,8 +733,26 @@ static int persistent_ctr(struct dm_exce
atomic_set(&ps->pending_count, 0);
ps->callbacks = NULL;

+ cow_dev_size = get_dev_size(store->cow->bdev);
+ max_buckets = (2 * 1024 * 1024)/sizeof(struct list_head);
+
+ hash_size = cow_dev_size >> store->chunk_shift;
+ hash_size = min(hash_size, max_buckets);
+
+ hash_size = rounddown_pow_of_two(hash_size);
+
+ ps->table = dm_exception_table_create(hash_size,
+ DM_CHUNK_CONSECUTIVE_BITS,
+ alloc_exception, NULL,
+ free_exception, NULL);
+ if (!ps->table) {
+ kfree(ps);
+ return -ENOMEM;
+ }
+
ps->metadata_wq = create_singlethread_workqueue("ksnaphd");
if (!ps->metadata_wq) {
+ dm_exception_table_destroy(ps->table);
kfree(ps);
DMERR("couldn't start header metadata update thread");
return -ENOMEM;
@@ -711,6 +790,7 @@ static struct dm_exception_store_type _p
.module = THIS_MODULE,
.ctr = persistent_ctr,
.dtr = persistent_dtr,
+ .resume = persistent_resume,
.read_metadata = persistent_read_metadata,
.prepare_exception = persistent_prepare_exception,
.commit_exception = persistent_commit_exception,
@@ -724,6 +804,7 @@ static struct dm_exception_store_type _p
.module = THIS_MODULE,
.ctr = persistent_ctr,
.dtr = persistent_dtr,
+ .resume = persistent_resume,
.read_metadata = persistent_read_metadata,
.prepare_exception = persistent_prepare_exception,
.commit_exception = persistent_commit_exception,
@@ -736,9 +817,16 @@ int dm_persistent_snapshot_init(void)
{
int r;

+ exception_cache = KMEM_CACHE(dm_exception, 0);
+ if (!exception_cache) {
+ DMERR("Couldn't create persistent exception cache.");
+ return -ENOMEM;
+ }
+
r = dm_exception_store_type_register(&_persistent_type );
if (r) {
DMERR("Unable to register persistent exception store type");
+ kmem_cache_destroy(exception_cache);
return r;
}

@@ -747,6 +835,7 @@ int dm_persistent_snapshot_init(void)
DMERR("Unable to register old-style persistent exception "
"store type");
dm_exception_store_type_unregister(&_persistent_ty pe);
+ kmem_cache_destroy(exception_cache);
return r;
}

@@ -757,4 +846,5 @@ void dm_persistent_snapshot_exit(void)
{
dm_exception_store_type_unregister(&_persistent_ty pe);
dm_exception_store_type_unregister(&_persistent_co mpat_type);
+ kmem_cache_destroy(exception_cache);
}
Index: linux-2.6/drivers/md/dm-snap-transient.c
================================================== =================
--- linux-2.6.orig/drivers/md/dm-snap-transient.c
+++ linux-2.6/drivers/md/dm-snap-transient.c
@@ -19,12 +19,29 @@
* Implementation of the store for non-persistent snapshots.
*---------------------------------------------------------------*/
struct transient_c {
+ struct dm_exception_table *table;
+
sector_t next_free;
};

+/* Could use better allocation policies - like in dm-snap-persistent.c */
+static struct dm_exception *alloc_exception(void *unused)
+{
+ return kmalloc(sizeof(struct dm_exception), GFP_KERNEL);
+}
+
+static void free_exception(struct dm_exception *e, void *unused)
+{
+ kfree(e);
+}
+
static void transient_dtr(struct dm_exception_store *store)
{
- kfree(store->context);
+ struct transient_c *tc = store->context;
+
+ dm_exception_table_destroy(tc->table);
+
+ kfree(tc);
}

static int transient_read_metadata(struct dm_exception_store *store,
@@ -35,6 +52,11 @@ static int transient_read_metadata(struc
return 0;
}

+static int transient_resume(struct dm_exception_store *store)
+{
+ return 0;
+}
+
static int transient_prepare_exception(struct dm_exception_store *store,
struct dm_exception *e)
{
@@ -70,6 +92,7 @@ static int transient_ctr(struct dm_excep
unsigned argc, char **argv)
{
struct transient_c *tc;
+ sector_t hash_size, cow_dev_size, max_buckets;

tc = kmalloc(sizeof(struct transient_c), GFP_KERNEL);
if (!tc)
@@ -78,6 +101,23 @@ static int transient_ctr(struct dm_excep
tc->next_free = 0;
store->context = tc;

+ cow_dev_size = get_dev_size(store->cow->bdev);
+ max_buckets = (2 * 1024 * 1024)/sizeof(struct list_head);
+
+ hash_size = cow_dev_size >> store->chunk_shift;
+ hash_size = min(hash_size, max_buckets);
+
+ hash_size = rounddown_pow_of_two(hash_size);
+
+ tc->table = dm_exception_table_create(hash_size,
+ DM_CHUNK_CONSECUTIVE_BITS,
+ alloc_exception, NULL,
+ free_exception, NULL);
+ if (!tc->table) {
+ kfree(tc);
+ return -ENOMEM;
+ }
+
return 0;
}

@@ -108,6 +148,7 @@ static struct dm_exception_store_type _t
.module = THIS_MODULE,
.ctr = transient_ctr,
.dtr = transient_dtr,
+ .resume = transient_resume,
.read_metadata = transient_read_metadata,
.prepare_exception = transient_prepare_exception,
.commit_exception = transient_commit_exception,
@@ -120,6 +161,7 @@ static struct dm_exception_store_type _t
.module = THIS_MODULE,
.ctr = transient_ctr,
.dtr = transient_dtr,
+ .resume = transient_resume,
.read_metadata = transient_read_metadata,
.prepare_exception = transient_prepare_exception,
.commit_exception = transient_commit_exception,
Index: linux-2.6/drivers/md/dm-snap.c
================================================== =================
--- linux-2.6.orig/drivers/md/dm-snap.c
+++ linux-2.6/drivers/md/dm-snap.c
@@ -1071,10 +1071,22 @@ static int snapshot_end_io(struct dm_tar

static void snapshot_resume(struct dm_target *ti)
{
+ int r;
struct dm_snapshot *s = ti->private;

+ /*
+ * Target resumes cannot fail, which leaves us in a tight spot.
+ * We read the exception store, the snapshot may be invalid
+ * or we may have failed to resume for a different reason (EIO?).
+ * If invalid, mark the snapshot as such. However, if other,
+ * what can we do? Mark 'not active'?
+ */
+ r = s->store->type->resume(s->store);
+ if (r == -EINVAL)
+ r = s->valid = 0;
+
down_write(&s->lock);
- s->active = 1;
+ s->active = (r) ? 0 : 1;
up_write(&s->lock);
}



--
dm-devel mailing list
dm-devel@redhat.com
https://www.redhat.com/mailman/listinfo/dm-devel

Add 'resume' to exception store API.

This is the first step to replace 'read_metadata' with 'resume'.
'resume' will populate the exception store implementation's
exception cache - rather than loading the snapshot's exception
cache.

Relocating the 'resume' is more in line with the way device-mapper
targets (and device-mapper mirror logs) work. Additionally, with
future shared exception store types, we don't duplicate the cache.

RFC-by: Jonathan Brassow <jbrassow@redhat.com>

Index: linux-2.6/drivers/md/dm-exception-store.h
================================================== =================
--- linux-2.6.orig/drivers/md/dm-exception-store.h
+++ linux-2.6/drivers/md/dm-exception-store.h
@@ -31,6 +31,8 @@ struct dm_exception_store_type {
*/
void (*dtr) (struct dm_exception_store *store);

+ int (*resume) (struct dm_exception_store *store);
+
/*
* The target shouldn't read the COW device until this is
* called. As exceptions are read from the COW, they are
Index: linux-2.6/drivers/md/dm-snap-persistent.c
================================================== =================
--- linux-2.6.orig/drivers/md/dm-snap-persistent.c
+++ linux-2.6/drivers/md/dm-snap-persistent.c
@@ -89,6 +89,7 @@ struct commit_callback {
*/
struct pstore {
struct dm_exception_store *store;
+ struct dm_exception_table *table;
int version;
int valid;
uint32_t exceptions_per_area;
@@ -130,6 +131,24 @@ struct pstore {
struct workqueue_struct *metadata_wq;
};

+static struct kmem_cache *exception_cache;
+
+static struct dm_exception *alloc_exception(void *unused)
+{
+ struct dm_exception *e;
+
+ e = kmem_cache_alloc(exception_cache, GFP_NOIO);
+ if (!e)
+ e = kmem_cache_alloc(exception_cache, GFP_ATOMIC);
+
+ return e;
+}
+
+static void free_exception(struct dm_exception *e, void *unused)
+{
+ kmem_cache_free(exception_cache, e);
+}
+
static unsigned sectors_to_pages(unsigned sectors)
{
return DIV_ROUND_UP(sectors, PAGE_SIZE >> 9);
@@ -282,7 +301,7 @@ static int read_header(struct pstore *ps
*/
if (!ps->store->chunk_size) {
ps->store->chunk_size = max(DM_CHUNK_SIZE_DEFAULT_SECTORS,
- bdev_hardsect_size(ps->store->cow->bdev) >> 9);
+ bdev_hardsect_size(ps->store->cow->bdev) >> 9);
ps->store->chunk_mask = ps->store->chunk_size - 1;
ps->store->chunk_shift = ffs(ps->store->chunk_size) - 1;
chunk_size_supplied = 0;
@@ -495,6 +514,7 @@ static void persistent_dtr(struct dm_exc

/* Don't need to check these, because they are done in ctr */
destroy_workqueue(ps->metadata_wq);
+ dm_exception_table_destroy(ps->table);
kfree(ps);
}

@@ -506,6 +526,13 @@ static int persistent_read_metadata(stru
int r, uninitialized_var(new_snapshot);
struct pstore *ps = get_info(store);

+ if (ps->callbacks)
+ /*
+ * Temporary work around for having two different functions
+ * that get us going... 'read_metadata' and 'resume'.
+ */
+ goto read_metadata;
+
/*
* Read the snapshot header.
*/
@@ -517,7 +544,7 @@ static int persistent_read_metadata(stru
* Now we know correct chunk_size, complete the initialisation.
*/
ps->exceptions_per_area = (ps->store->chunk_size << SECTOR_SHIFT) /
- sizeof(struct disk_exception);
+ sizeof(struct disk_exception);
ps->callbacks = dm_vcalloc(ps->exceptions_per_area,
sizeof(*ps->callbacks));
if (!ps->callbacks)
@@ -558,11 +585,44 @@ static int persistent_read_metadata(stru
/*
* Read the metadata.
*/
+read_metadata:
r = read_exceptions(ps, callback, callback_context);

return r;
}

+/* This function is temporary for patch cleanliness */
+static int add_exception(void *context, chunk_t old, chunk_t new)
+{
+ struct dm_exception_store *store = context;
+ struct pstore *ps = get_info(store);
+ struct dm_exception *e;
+
+ e = dm_alloc_exception(ps->table);
+ if (!e)
+ return -ENOMEM;
+
+ e->old_chunk = old;
+ e->new_chunk = new;
+
+ dm_insert_exception(ps->table, e);
+
+ return 0;
+}
+
+/*
+ * persistent_resume
+ * @store
+ *
+ * Read metadata of the disk and store in our exception table cache.
+ *
+ * Returns: 0 on success, -Exxx on error
+ */
+static int persistent_resume(struct dm_exception_store *store)
+{
+ return persistent_read_metadata(store, add_exception, store);
+}
+
static int persistent_prepare_exception(struct dm_exception_store *store,
struct dm_exception *e)
{
@@ -665,6 +725,7 @@ static int persistent_ctr(struct dm_exce
unsigned argc, char **argv)
{
struct pstore *ps;
+ sector_t hash_size, cow_dev_size, max_buckets;

/* allocate the pstore */
ps = kzalloc(sizeof(*ps), GFP_KERNEL);
@@ -682,8 +743,26 @@ static int persistent_ctr(struct dm_exce
atomic_set(&ps->pending_count, 0);
ps->callbacks = NULL;

+ cow_dev_size = get_dev_size(store->cow->bdev);
+ max_buckets = (2 * 1024 * 1024)/sizeof(struct list_head);
+
+ hash_size = cow_dev_size >> store->chunk_shift;
+ hash_size = min(hash_size, max_buckets);
+
+ hash_size = rounddown_pow_of_two(hash_size);
+
+ ps->table = dm_exception_table_create(hash_size,
+ DM_CHUNK_CONSECUTIVE_BITS,
+ alloc_exception, NULL,
+ free_exception, NULL);
+ if (!ps->table) {
+ kfree(ps);
+ return -ENOMEM;
+ }
+
ps->metadata_wq = create_singlethread_workqueue("ksnaphd");
if (!ps->metadata_wq) {
+ dm_exception_table_destroy(ps->table);
kfree(ps);
DMERR("couldn't start header metadata update thread");
return -ENOMEM;
@@ -721,6 +800,7 @@ static struct dm_exception_store_type _p
.module = THIS_MODULE,
.ctr = persistent_ctr,
.dtr = persistent_dtr,
+ .resume = persistent_resume,
.read_metadata = persistent_read_metadata,
.prepare_exception = persistent_prepare_exception,
.commit_exception = persistent_commit_exception,
@@ -734,6 +814,7 @@ static struct dm_exception_store_type _p
.module = THIS_MODULE,
.ctr = persistent_ctr,
.dtr = persistent_dtr,
+ .resume = persistent_resume,
.read_metadata = persistent_read_metadata,
.prepare_exception = persistent_prepare_exception,
.commit_exception = persistent_commit_exception,
@@ -746,9 +827,16 @@ int dm_persistent_snapshot_init(void)
{
int r;

+ exception_cache = KMEM_CACHE(dm_exception, 0);
+ if (!exception_cache) {
+ DMERR("Couldn't create persistent exception cache.");
+ return -ENOMEM;
+ }
+
r = dm_exception_store_type_register(&_persistent_type );
if (r) {
DMERR("Unable to register persistent exception store type");
+ kmem_cache_destroy(exception_cache);
return r;
}

@@ -757,6 +845,7 @@ int dm_persistent_snapshot_init(void)
DMERR("Unable to register old-style persistent exception "
"store type");
dm_exception_store_type_unregister(&_persistent_ty pe);
+ kmem_cache_destroy(exception_cache);
return r;
}

@@ -767,4 +856,5 @@ void dm_persistent_snapshot_exit(void)
{
dm_exception_store_type_unregister(&_persistent_ty pe);
dm_exception_store_type_unregister(&_persistent_co mpat_type);
+ kmem_cache_destroy(exception_cache);
}
Index: linux-2.6/drivers/md/dm-snap-transient.c
================================================== =================
--- linux-2.6.orig/drivers/md/dm-snap-transient.c
+++ linux-2.6/drivers/md/dm-snap-transient.c
@@ -19,12 +19,29 @@
* Implementation of the store for non-persistent snapshots.
*---------------------------------------------------------------*/
struct transient_c {
+ struct dm_exception_table *table;
+
sector_t next_free;
};

+/* Could use better allocation policies - like in dm-snap-persistent.c */
+static struct dm_exception *alloc_exception(void *unused)
+{
+ return kmalloc(sizeof(struct dm_exception), GFP_KERNEL);
+}
+
+static void free_exception(struct dm_exception *e, void *unused)
+{
+ kfree(e);
+}
+
static void transient_dtr(struct dm_exception_store *store)
{
- kfree(store->context);
+ struct transient_c *tc = store->context;
+
+ dm_exception_table_destroy(tc->table);
+
+ kfree(tc);
}

static int transient_read_metadata(struct dm_exception_store *store,
@@ -35,6 +52,11 @@ static int transient_read_metadata(struc
return 0;
}

+static int transient_resume(struct dm_exception_store *store)
+{
+ return 0;
+}
+
static int transient_prepare_exception(struct dm_exception_store *store,
struct dm_exception *e)
{
@@ -70,6 +92,7 @@ static int transient_ctr(struct dm_excep
unsigned argc, char **argv)
{
struct transient_c *tc;
+ sector_t hash_size, cow_dev_size, max_buckets;

tc = kmalloc(sizeof(struct transient_c), GFP_KERNEL);
if (!tc)
@@ -78,6 +101,23 @@ static int transient_ctr(struct dm_excep
tc->next_free = 0;
store->context = tc;

+ cow_dev_size = get_dev_size(store->cow->bdev);
+ max_buckets = (2 * 1024 * 1024)/sizeof(struct list_head);
+
+ hash_size = cow_dev_size >> store->chunk_shift;
+ hash_size = min(hash_size, max_buckets);
+
+ hash_size = rounddown_pow_of_two(hash_size);
+
+ tc->table = dm_exception_table_create(hash_size,
+ DM_CHUNK_CONSECUTIVE_BITS,
+ alloc_exception, NULL,
+ free_exception, NULL);
+ if (!tc->table) {
+ kfree(tc);
+ return -ENOMEM;
+ }
+
return 0;
}

@@ -108,6 +148,7 @@ static struct dm_exception_store_type _t
.module = THIS_MODULE,
.ctr = transient_ctr,
.dtr = transient_dtr,
+ .resume = transient_resume,
.read_metadata = transient_read_metadata,
.prepare_exception = transient_prepare_exception,
.commit_exception = transient_commit_exception,
@@ -120,6 +161,7 @@ static struct dm_exception_store_type _t
.module = THIS_MODULE,
.ctr = transient_ctr,
.dtr = transient_dtr,
+ .resume = transient_resume,
.read_metadata = transient_read_metadata,
.prepare_exception = transient_prepare_exception,
.commit_exception = transient_commit_exception,
Index: linux-2.6/drivers/md/dm-snap.c
================================================== =================
--- linux-2.6.orig/drivers/md/dm-snap.c
+++ linux-2.6/drivers/md/dm-snap.c
@@ -1072,10 +1072,22 @@ static int snapshot_end_io(struct dm_tar

static void snapshot_resume(struct dm_target *ti)
{
+ int r;
struct dm_snapshot *s = ti->private;

+ /*
+ * Target resumes cannot fail, which leaves us in a tight spot.
+ * We read the exception store, the snapshot may be invalid
+ * or we may have failed to resume for a different reason (EIO?).
+ * If invalid, mark the snapshot as such. However, if other,
+ * what can we do? Mark 'not active'?
+ */
+ r = s->store->type->resume(s->store);
+ if (r == -EINVAL)
+ r = s->valid = 0;
+
down_write(&s->lock);
- s->active = 1;
+ s->active = (r) ? 0 : 1;
up_write(&s->lock);
}


--
dm-devel mailing list
dm-devel@redhat.com
https://www.redhat.com/mailman/listinfo/dm-devel