]> git.kernelconcepts.de Git - karo-tx-linux.git/commitdiff
dm cache: defer whole cells
authorJoe Thornber <ejt@redhat.com>
Fri, 15 May 2015 14:26:08 +0000 (15:26 +0100)
committerMike Snitzer <snitzer@redhat.com>
Fri, 29 May 2015 18:19:06 +0000 (14:19 -0400)
Currently individual bios are deferred to the worker thread if they
cannot be processed immediately (eg, a block is in the process of
being moved to the fast device).

This patch passes whole cells across to the worker.  This saves
reaquiring the cell, and also collects bios destined for the same block
together, which allows them to be mapped with a single look up to the
policy.  This reduces the overhead of using dm-cache.

Signed-off-by: Joe Thornber <ejt@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
drivers/md/dm-cache-target.c

index 5d3b20b91ba366834b1897541445324efa864e9d..d2d91c164420a5badc384581a7dbb0e941a3d390 100644 (file)
@@ -257,6 +257,7 @@ struct cache {
        int sectors_per_block_shift;
 
        spinlock_t lock;
+       struct list_head deferred_cells;
        struct bio_list deferred_bios;
        struct bio_list deferred_flush_bios;
        struct bio_list deferred_writethrough_bios;
@@ -969,26 +970,63 @@ static void dec_io_migrations(struct cache *cache)
        atomic_dec(&cache->nr_io_migrations);
 }
 
-static void __cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell,
-                        bool holder)
+static void __cell_release(struct cache *cache, struct dm_bio_prison_cell *cell,
+                          bool holder, struct bio_list *bios)
 {
        (holder ? dm_cell_release : dm_cell_release_no_holder)
-               (cache->prison, cell, &cache->deferred_bios);
+               (cache->prison, cell, bios);
        free_prison_cell(cache, cell);
 }
 
-static void cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell,
-                      bool holder)
+static bool discard_or_flush(struct bio *bio)
+{
+       return bio->bi_rw & (REQ_FLUSH | REQ_FUA | REQ_DISCARD);
+}
+
+static void __cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell)
+{
+       if (discard_or_flush(cell->holder))
+               /*
+                * We have to handle these bios
+                * individually.
+                */
+               __cell_release(cache, cell, true, &cache->deferred_bios);
+
+       else
+               list_add_tail(&cell->user_list, &cache->deferred_cells);
+}
+
+static void cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell, bool holder)
 {
        unsigned long flags;
 
+       if (!holder && dm_cell_promote_or_release(cache->prison, cell)) {
+               /*
+                * There was no prisoner to promote to holder, the
+                * cell has been released.
+                */
+               free_prison_cell(cache, cell);
+               return;
+       }
+
        spin_lock_irqsave(&cache->lock, flags);
-       __cell_defer(cache, cell, holder);
+       __cell_defer(cache, cell);
        spin_unlock_irqrestore(&cache->lock, flags);
 
        wake_worker(cache);
 }
 
+static void cell_error_with_code(struct cache *cache, struct dm_bio_prison_cell *cell, int err)
+{
+       dm_cell_error(cache->prison, cell, err);
+       dm_bio_prison_free_cell(cache->prison, cell);
+}
+
+static void cell_requeue(struct cache *cache, struct dm_bio_prison_cell *cell)
+{
+       cell_error_with_code(cache, cell, DM_ENDIO_REQUEUE);
+}
+
 static void free_io_migration(struct dm_cache_migration *mg)
 {
        dec_io_migrations(mg->cache);
@@ -1525,6 +1563,107 @@ static void inc_miss_counter(struct cache *cache, struct bio *bio)
 
 /*----------------------------------------------------------------*/
 
+struct inc_detail {
+       struct cache *cache;
+       struct bio_list bios_for_issue;
+       struct bio_list unhandled_bios;
+       bool any_writes;
+};
+
+static void inc_fn(void *context, struct dm_bio_prison_cell *cell)
+{
+       struct bio *bio;
+       struct inc_detail *detail = context;
+       struct cache *cache = detail->cache;
+
+       inc_ds(cache, cell->holder, cell);
+       if (bio_data_dir(cell->holder) == WRITE)
+               detail->any_writes = true;
+
+       while ((bio = bio_list_pop(&cell->bios))) {
+               if (discard_or_flush(bio)) {
+                       bio_list_add(&detail->unhandled_bios, bio);
+                       continue;
+               }
+
+               if (bio_data_dir(bio) == WRITE)
+                       detail->any_writes = true;
+
+               bio_list_add(&detail->bios_for_issue, bio);
+               inc_ds(cache, bio, cell);
+       }
+}
+
+// FIXME: refactor these two
+static void remap_cell_to_origin_clear_discard(struct cache *cache,
+                                              struct dm_bio_prison_cell *cell,
+                                              dm_oblock_t oblock, bool issue_holder)
+{
+       struct bio *bio;
+       unsigned long flags;
+       struct inc_detail detail;
+
+       detail.cache = cache;
+       bio_list_init(&detail.bios_for_issue);
+       bio_list_init(&detail.unhandled_bios);
+       detail.any_writes = false;
+
+       spin_lock_irqsave(&cache->lock, flags);
+       dm_cell_visit_release(cache->prison, inc_fn, &detail, cell);
+       bio_list_merge(&cache->deferred_bios, &detail.unhandled_bios);
+       spin_unlock_irqrestore(&cache->lock, flags);
+
+       remap_to_origin(cache, cell->holder);
+       if (issue_holder)
+               issue(cache, cell->holder);
+       else
+               accounted_begin(cache, cell->holder);
+
+       if (detail.any_writes)
+               clear_discard(cache, oblock_to_dblock(cache, oblock));
+
+       while ((bio = bio_list_pop(&detail.bios_for_issue))) {
+               remap_to_origin(cache, bio);
+               issue(cache, bio);
+       }
+}
+
+static void remap_cell_to_cache_dirty(struct cache *cache, struct dm_bio_prison_cell *cell,
+                                     dm_oblock_t oblock, dm_cblock_t cblock, bool issue_holder)
+{
+       struct bio *bio;
+       unsigned long flags;
+       struct inc_detail detail;
+
+       detail.cache = cache;
+       bio_list_init(&detail.bios_for_issue);
+       bio_list_init(&detail.unhandled_bios);
+       detail.any_writes = false;
+
+       spin_lock_irqsave(&cache->lock, flags);
+       dm_cell_visit_release(cache->prison, inc_fn, &detail, cell);
+       bio_list_merge(&cache->deferred_bios, &detail.unhandled_bios);
+       spin_unlock_irqrestore(&cache->lock, flags);
+
+       remap_to_cache(cache, cell->holder, cblock);
+       if (issue_holder)
+               issue(cache, cell->holder);
+       else
+               accounted_begin(cache, cell->holder);
+
+       if (detail.any_writes) {
+               set_dirty(cache, oblock, cblock);
+               clear_discard(cache, oblock_to_dblock(cache, oblock));
+       }
+
+       while ((bio = bio_list_pop(&detail.bios_for_issue))) {
+               remap_to_cache(cache, bio, cblock);
+               issue(cache, bio);
+       }
+}
+
+/*----------------------------------------------------------------*/
+
 struct old_oblock_lock {
        struct policy_locker locker;
        struct cache *cache;
@@ -1549,28 +1688,18 @@ static int cell_locker(struct policy_locker *locker, dm_oblock_t b)
                          l->structs, &l->cell);
 }
 
-static void process_bio(struct cache *cache, struct prealloc *structs,
-                       struct bio *bio)
+static void process_cell(struct cache *cache, struct prealloc *structs,
+                        struct dm_bio_prison_cell *new_ocell)
 {
        int r;
        bool release_cell = true;
+       struct bio *bio = new_ocell->holder;
        dm_oblock_t block = get_bio_block(cache, bio);
-       struct dm_bio_prison_cell *cell_prealloc, *new_ocell;
        struct policy_result lookup_result;
        bool passthrough = passthrough_mode(&cache->features);
        bool discarded_block, can_migrate;
        struct old_oblock_lock ool;
 
-       /*
-        * Check to see if that block is currently migrating.
-        */
-       cell_prealloc = prealloc_get_cell(structs);
-       r = bio_detain(cache, block, bio, cell_prealloc,
-                      (cell_free_fn) prealloc_put_cell,
-                      structs, &new_ocell);
-       if (r > 0)
-               return;
-
        discarded_block = is_discarded_oblock(cache, block);
        can_migrate = !passthrough && (discarded_block || spare_migration_bandwidth(cache));
 
@@ -1615,9 +1744,9 @@ static void process_bio(struct cache *cache, struct prealloc *structs,
                                remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
                                inc_and_issue(cache, bio, new_ocell);
 
-                       } else  {
-                               remap_to_cache_dirty(cache, bio, block, lookup_result.cblock);
-                               inc_and_issue(cache, bio, new_ocell);
+                       } else {
+                               remap_cell_to_cache_dirty(cache, new_ocell, block, lookup_result.cblock, true);
+                               release_cell = false;
                        }
                }
 
@@ -1625,8 +1754,8 @@ static void process_bio(struct cache *cache, struct prealloc *structs,
 
        case POLICY_MISS:
                inc_miss_counter(cache, bio);
-               remap_to_origin_clear_discard(cache, bio, block);
-               inc_and_issue(cache, bio, new_ocell);
+               remap_cell_to_origin_clear_discard(cache, new_ocell, block, true);
+               release_cell = false;
                break;
 
        case POLICY_NEW:
@@ -1654,10 +1783,30 @@ static void process_bio(struct cache *cache, struct prealloc *structs,
                cell_defer(cache, new_ocell, false);
 }
 
+static void process_bio(struct cache *cache, struct prealloc *structs,
+                       struct bio *bio)
+{
+       int r;
+       dm_oblock_t block = get_bio_block(cache, bio);
+       struct dm_bio_prison_cell *cell_prealloc, *new_ocell;
+
+       /*
+        * Check to see if that block is currently migrating.
+        */
+       cell_prealloc = prealloc_get_cell(structs);
+       r = bio_detain(cache, block, bio, cell_prealloc,
+                      (cell_free_fn) prealloc_put_cell,
+                      structs, &new_ocell);
+       if (r > 0)
+               return;
+
+       process_cell(cache, structs, new_ocell);
+}
+
 static int need_commit_due_to_time(struct cache *cache)
 {
-       return !time_in_range(jiffies, cache->last_commit_jiffies,
-                             cache->last_commit_jiffies + COMMIT_PERIOD);
+       return jiffies < cache->last_commit_jiffies ||
+              jiffies > cache->last_commit_jiffies + COMMIT_PERIOD;
 }
 
 static int commit_if_needed(struct cache *cache)
@@ -1716,6 +1865,40 @@ static void process_deferred_bios(struct cache *cache)
        prealloc_free_structs(cache, &structs);
 }
 
+static void process_deferred_cells(struct cache *cache)
+{
+       unsigned long flags;
+       struct dm_bio_prison_cell *cell, *tmp;
+       struct list_head cells;
+       struct prealloc structs;
+
+       memset(&structs, 0, sizeof(structs));
+
+       INIT_LIST_HEAD(&cells);
+
+       spin_lock_irqsave(&cache->lock, flags);
+       list_splice_init(&cache->deferred_cells, &cells);
+       spin_unlock_irqrestore(&cache->lock, flags);
+
+       list_for_each_entry_safe(cell, tmp, &cells, user_list) {
+               /*
+                * If we've got no free migration structs, and processing
+                * this bio might require one, we pause until there are some
+                * prepared mappings to process.
+                */
+               if (prealloc_data_structs(cache, &structs)) {
+                       spin_lock_irqsave(&cache->lock, flags);
+                       list_splice(&cells, &cache->deferred_cells);
+                       spin_unlock_irqrestore(&cache->lock, flags);
+                       break;
+               }
+
+               process_cell(cache, &structs, cell);
+       }
+
+       prealloc_free_structs(cache, &structs);
+}
+
 static void process_deferred_flush_bios(struct cache *cache, bool submit_bios)
 {
        unsigned long flags;
@@ -1883,7 +2066,22 @@ static void stop_worker(struct cache *cache)
        flush_workqueue(cache->wq);
 }
 
-static void requeue_deferred_io(struct cache *cache)
+static void requeue_deferred_cells(struct cache *cache)
+{
+       unsigned long flags;
+       struct list_head cells;
+       struct dm_bio_prison_cell *cell, *tmp;
+
+       INIT_LIST_HEAD(&cells);
+       spin_lock_irqsave(&cache->lock, flags);
+       list_splice_init(&cache->deferred_cells, &cells);
+       spin_unlock_irqrestore(&cache->lock, flags);
+
+       list_for_each_entry_safe(cell, tmp, &cells, user_list)
+               cell_requeue(cache, cell);
+}
+
+static void requeue_deferred_bios(struct cache *cache)
 {
        struct bio *bio;
        struct bio_list bios;
@@ -1904,6 +2102,7 @@ static int more_work(struct cache *cache)
                        !list_empty(&cache->need_commit_migrations);
        else
                return !bio_list_empty(&cache->deferred_bios) ||
+                       !list_empty(&cache->deferred_cells) ||
                        !bio_list_empty(&cache->deferred_flush_bios) ||
                        !bio_list_empty(&cache->deferred_writethrough_bios) ||
                        !list_empty(&cache->quiesced_migrations) ||
@@ -1921,6 +2120,7 @@ static void do_worker(struct work_struct *ws)
                        writeback_some_dirty_blocks(cache);
                        process_deferred_writethrough_bios(cache);
                        process_deferred_bios(cache);
+                       process_deferred_cells(cache);
                        process_invalidation_requests(cache);
                }
 
@@ -1935,6 +2135,7 @@ static void do_worker(struct work_struct *ws)
                         * FIXME: rollback metadata or just go into a
                         * failure mode and error everything
                         */
+
                } else {
                        process_deferred_flush_bios(cache, true);
                        process_migrations(cache, &cache->need_commit_migrations,
@@ -2525,6 +2726,7 @@ static int cache_create(struct cache_args *ca, struct cache **result)
        }
 
        spin_lock_init(&cache->lock);
+       INIT_LIST_HEAD(&cache->deferred_cells);
        bio_list_init(&cache->deferred_bios);
        bio_list_init(&cache->deferred_flush_bios);
        bio_list_init(&cache->deferred_writethrough_bios);
@@ -2682,9 +2884,14 @@ out:
        return r;
 }
 
-static int __cache_map(struct cache *cache, struct bio *bio, struct dm_bio_prison_cell **cell)
+/*----------------------------------------------------------------*/
+
+static int cache_map(struct dm_target *ti, struct bio *bio)
 {
+       struct cache *cache = ti->private;
+
        int r;
+       struct dm_bio_prison_cell *cell = NULL;
        dm_oblock_t block = get_bio_block(cache, bio);
        size_t pb_data_size = get_per_bio_data_size(cache);
        bool can_migrate = false;
@@ -2702,10 +2909,11 @@ static int __cache_map(struct cache *cache, struct bio *bio, struct dm_bio_priso
                 * Just remap to the origin and carry on.
                 */
                remap_to_origin(cache, bio);
+               accounted_begin(cache, bio);
                return DM_MAPIO_REMAPPED;
        }
 
-       if (bio->bi_rw & (REQ_FLUSH | REQ_FUA | REQ_DISCARD)) {
+       if (discard_or_flush(bio)) {
                defer_bio(cache, bio);
                return DM_MAPIO_SUBMITTED;
        }
@@ -2713,15 +2921,15 @@ static int __cache_map(struct cache *cache, struct bio *bio, struct dm_bio_priso
        /*
         * Check to see if that block is currently migrating.
         */
-       *cell = alloc_prison_cell(cache);
-       if (!*cell) {
+       cell = alloc_prison_cell(cache);
+       if (!cell) {
                defer_bio(cache, bio);
                return DM_MAPIO_SUBMITTED;
        }
 
-       r = bio_detain(cache, block, bio, *cell,
+       r = bio_detain(cache, block, bio, cell,
                       (cell_free_fn) free_prison_cell,
-                      cache, cell);
+                      cache, &cell);
        if (r) {
                if (r < 0)
                        defer_bio(cache, bio);
@@ -2734,12 +2942,12 @@ static int __cache_map(struct cache *cache, struct bio *bio, struct dm_bio_priso
        r = policy_map(cache->policy, block, false, can_migrate, discarded_block,
                       bio, &ool.locker, &lookup_result);
        if (r == -EWOULDBLOCK) {
-               cell_defer(cache, *cell, true);
+               cell_defer(cache, cell, true);
                return DM_MAPIO_SUBMITTED;
 
        } else if (r) {
                DMERR_LIMIT("Unexpected return from cache replacement policy: %d", r);
-               cell_defer(cache, *cell, false);
+               cell_defer(cache, cell, false);
                bio_io_error(bio);
                return DM_MAPIO_SUBMITTED;
        }
@@ -2753,21 +2961,30 @@ static int __cache_map(struct cache *cache, struct bio *bio, struct dm_bio_priso
                                 * We need to invalidate this block, so
                                 * defer for the worker thread.
                                 */
-                               cell_defer(cache, *cell, true);
+                               cell_defer(cache, cell, true);
                                r = DM_MAPIO_SUBMITTED;
 
                        } else {
                                inc_miss_counter(cache, bio);
                                remap_to_origin_clear_discard(cache, bio, block);
+                               accounted_begin(cache, bio);
+                               inc_ds(cache, bio, cell);
+                               // FIXME: we want to remap hits or misses straight
+                               // away rather than passing over to the worker.
+                               cell_defer(cache, cell, false);
                        }
 
                } else {
                        inc_hit_counter(cache, bio);
                        if (bio_data_dir(bio) == WRITE && writethrough_mode(&cache->features) &&
-                           !is_dirty(cache, lookup_result.cblock))
+                           !is_dirty(cache, lookup_result.cblock)) {
                                remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
-                       else
-                               remap_to_cache_dirty(cache, bio, block, lookup_result.cblock);
+                               accounted_begin(cache, bio);
+                               inc_ds(cache, bio, cell);
+                               cell_defer(cache, cell, false);
+
+                       } else
+                               remap_cell_to_cache_dirty(cache, cell, block, lookup_result.cblock, false);
                }
                break;
 
@@ -2779,18 +2996,18 @@ static int __cache_map(struct cache *cache, struct bio *bio, struct dm_bio_priso
                         * longer needed because the block has been demoted.
                         */
                        bio_endio(bio, 0);
-                       cell_defer(cache, *cell, false);
+                       // FIXME: remap everything as a miss
+                       cell_defer(cache, cell, false);
                        r = DM_MAPIO_SUBMITTED;
 
                } else
-                       remap_to_origin_clear_discard(cache, bio, block);
-
+                       remap_cell_to_origin_clear_discard(cache, cell, block, false);
                break;
 
        default:
                DMERR_LIMIT("%s: erroring bio: unknown policy op: %u", __func__,
                            (unsigned) lookup_result.op);
-               cell_defer(cache, *cell, false);
+               cell_defer(cache, cell, false);
                bio_io_error(bio);
                r = DM_MAPIO_SUBMITTED;
        }
@@ -2798,25 +3015,6 @@ static int __cache_map(struct cache *cache, struct bio *bio, struct dm_bio_priso
        return r;
 }
 
-static int cache_map(struct dm_target *ti, struct bio *bio)
-{
-       int r;
-       struct dm_bio_prison_cell *cell = NULL;
-       struct cache *cache = ti->private;
-
-       r = __cache_map(cache, bio, &cell);
-       if (r == DM_MAPIO_REMAPPED) {
-               accounted_begin(cache, bio);
-
-               if (cell) {
-                       inc_ds(cache, bio, cell);
-                       cell_defer(cache, cell, false);
-               }
-       }
-
-       return r;
-}
-
 static int cache_end_io(struct dm_target *ti, struct bio *bio, int error)
 {
        struct cache *cache = ti->private;
@@ -2913,7 +3111,8 @@ static void cache_postsuspend(struct dm_target *ti)
        start_quiescing(cache);
        wait_for_migrations(cache);
        stop_worker(cache);
-       requeue_deferred_io(cache);
+       requeue_deferred_bios(cache);
+       requeue_deferred_cells(cache);
        stop_quiescing(cache);
 
        (void) sync_metadata(cache);