Btrfs: use bytes_may_use for all ENOSPC reservations
authorJosef Bacik <josef@redhat.com>
Tue, 26 Jul 2011 21:00:46 +0000 (17:00 -0400)
committerJosef Bacik <josef@redhat.com>
Wed, 19 Oct 2011 19:12:30 +0000 (15:12 -0400)
We have been using bytes_reserved for metadata reservations, which is wrong
since we use that to keep track of outstanding reservations from the allocator.
This resulted in us doing a lot of silly things to make sure we don't allocate a
bunch of metadata chunks since we never had a real view of how much space was
actually in use by metadata.

This passes Arne's enospc test and xfstests as well as my own enospc tests.
Hopefully this will get us moving in the right direction.  Thanks,

Signed-off-by: Josef Bacik <josef@redhat.com>
fs/btrfs/ctree.h
fs/btrfs/extent-tree.c
fs/btrfs/free-space-cache.c

index 03912c5..332cbdc 100644 (file)
@@ -2196,8 +2196,6 @@ int btrfs_free_extent(struct btrfs_trans_handle *trans,
                      u64 root_objectid, u64 owner, u64 offset);
 
 int btrfs_free_reserved_extent(struct btrfs_root *root, u64 start, u64 len);
-int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
-                               u64 num_bytes, int reserve, int sinfo);
 int btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
                                struct btrfs_root *root);
 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
index 03edac4..fbe6278 100644 (file)
@@ -52,6 +52,21 @@ enum {
        CHUNK_ALLOC_LIMITED = 2,
 };
 
+/*
+ * Control how reservations are dealt with.
+ *
+ * RESERVE_FREE - freeing a reservation.
+ * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
+ *   ENOSPC accounting
+ * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
+ *   bytes_may_use as the ENOSPC accounting is done elsewhere
+ */
+enum {
+       RESERVE_FREE = 0,
+       RESERVE_ALLOC = 1,
+       RESERVE_ALLOC_NO_ACCOUNT = 2,
+};
+
 static int update_block_group(struct btrfs_trans_handle *trans,
                              struct btrfs_root *root,
                              u64 bytenr, u64 num_bytes, int alloc);
@@ -81,6 +96,8 @@ static int find_next_key(struct btrfs_path *path, int level,
                         struct btrfs_key *key);
 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
                            int dump_block_groups);
+static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
+                                      u64 num_bytes, int reserve);
 
 static noinline int
 block_group_cache_done(struct btrfs_block_group_cache *cache)
@@ -3128,9 +3145,7 @@ commit_trans:
 }
 
 /*
- * called when we are clearing an delalloc extent from the
- * inode's io_tree or there was an error for whatever reason
- * after calling btrfs_check_data_free_space
+ * Called if we need to clear a data reservation for this inode.
  */
 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
 {
@@ -3163,6 +3178,7 @@ static int should_alloc_chunk(struct btrfs_root *root,
                              struct btrfs_space_info *sinfo, u64 alloc_bytes,
                              int force)
 {
+       struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
        u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
        u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
        u64 thresh;
@@ -3170,6 +3186,13 @@ static int should_alloc_chunk(struct btrfs_root *root,
        if (force == CHUNK_ALLOC_FORCE)
                return 1;
 
+       /*
+        * We need to take into account the global rsv because for all intents
+        * and purposes it's used space.  Don't worry about locking the
+        * global_rsv, it doesn't change except when the transaction commits.
+        */
+       num_allocated += global_rsv->size;
+
        /*
         * in limited mode, we want to have some free space up to
         * about 1% of the FS size.
@@ -3317,7 +3340,7 @@ static int shrink_delalloc(struct btrfs_trans_handle *trans,
        space_info = block_rsv->space_info;
 
        smp_mb();
-       reserved = space_info->bytes_reserved;
+       reserved = space_info->bytes_may_use;
        progress = space_info->reservation_progress;
 
        if (reserved == 0)
@@ -3341,9 +3364,9 @@ static int shrink_delalloc(struct btrfs_trans_handle *trans,
                writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages);
 
                spin_lock(&space_info->lock);
-               if (reserved > space_info->bytes_reserved)
-                       reclaimed += reserved - space_info->bytes_reserved;
-               reserved = space_info->bytes_reserved;
+               if (reserved > space_info->bytes_may_use)
+                       reclaimed += reserved - space_info->bytes_may_use;
+               reserved = space_info->bytes_may_use;
                spin_unlock(&space_info->lock);
 
                loops++;
@@ -3401,7 +3424,6 @@ static int reserve_metadata_bytes(struct btrfs_trans_handle *trans,
        int ret = 0;
        bool committed = false;
        bool flushing = false;
-
 again:
        ret = 0;
        spin_lock(&space_info->lock);
@@ -3443,7 +3465,7 @@ again:
        if (unused <= space_info->total_bytes) {
                unused = space_info->total_bytes - unused;
                if (unused >= num_bytes) {
-                       space_info->bytes_reserved += orig_bytes;
+                       space_info->bytes_may_use += orig_bytes;
                        ret = 0;
                } else {
                        /*
@@ -3614,7 +3636,7 @@ static void block_rsv_release_bytes(struct btrfs_block_rsv *block_rsv,
                }
                if (num_bytes) {
                        spin_lock(&space_info->lock);
-                       space_info->bytes_reserved -= num_bytes;
+                       space_info->bytes_may_use -= num_bytes;
                        space_info->reservation_progress++;
                        spin_unlock(&space_info->lock);
                }
@@ -3825,12 +3847,12 @@ static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
        if (sinfo->total_bytes > num_bytes) {
                num_bytes = sinfo->total_bytes - num_bytes;
                block_rsv->reserved += num_bytes;
-               sinfo->bytes_reserved += num_bytes;
+               sinfo->bytes_may_use += num_bytes;
        }
 
        if (block_rsv->reserved >= block_rsv->size) {
                num_bytes = block_rsv->reserved - block_rsv->size;
-               sinfo->bytes_reserved -= num_bytes;
+               sinfo->bytes_may_use -= num_bytes;
                sinfo->reservation_progress++;
                block_rsv->reserved = block_rsv->size;
                block_rsv->full = 1;
@@ -4133,7 +4155,6 @@ static int update_block_group(struct btrfs_trans_handle *trans,
                        btrfs_set_block_group_used(&cache->item, old_val);
                        cache->reserved -= num_bytes;
                        cache->space_info->bytes_reserved -= num_bytes;
-                       cache->space_info->reservation_progress++;
                        cache->space_info->bytes_used += num_bytes;
                        cache->space_info->disk_used += num_bytes * factor;
                        spin_unlock(&cache->lock);
@@ -4185,7 +4206,6 @@ static int pin_down_extent(struct btrfs_root *root,
        if (reserved) {
                cache->reserved -= num_bytes;
                cache->space_info->bytes_reserved -= num_bytes;
-               cache->space_info->reservation_progress++;
        }
        spin_unlock(&cache->lock);
        spin_unlock(&cache->space_info->lock);
@@ -4212,46 +4232,55 @@ int btrfs_pin_extent(struct btrfs_root *root,
        return 0;
 }
 
-/*
- * update size of reserved extents. this function may return -EAGAIN
- * if 'reserve' is true or 'sinfo' is false.
+/**
+ * btrfs_update_reserved_bytes - update the block_group and space info counters
+ * @cache:     The cache we are manipulating
+ * @num_bytes: The number of bytes in question
+ * @reserve:   One of the reservation enums
+ *
+ * This is called by the allocator when it reserves space, or by somebody who is
+ * freeing space that was never actually used on disk.  For example if you
+ * reserve some space for a new leaf in transaction A and before transaction A
+ * commits you free that leaf, you call this with reserve set to 0 in order to
+ * clear the reservation.
+ *
+ * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
+ * ENOSPC accounting.  For data we handle the reservation through clearing the
+ * delalloc bits in the io_tree.  We have to do this since we could end up
+ * allocating less disk space for the amount of data we have reserved in the
+ * case of compression.
+ *
+ * If this is a reservation and the block group has become read only we cannot
+ * make the reservation and return -EAGAIN, otherwise this function always
+ * succeeds.
  */
-int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
-                               u64 num_bytes, int reserve, int sinfo)
+static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
+                                      u64 num_bytes, int reserve)
 {
+       struct btrfs_space_info *space_info = cache->space_info;
        int ret = 0;
-       if (sinfo) {
-               struct btrfs_space_info *space_info = cache->space_info;
-               spin_lock(&space_info->lock);
-               spin_lock(&cache->lock);
-               if (reserve) {
-                       if (cache->ro) {
-                               ret = -EAGAIN;
-                       } else {
-                               cache->reserved += num_bytes;
-                               space_info->bytes_reserved += num_bytes;
-                       }
-               } else {
-                       if (cache->ro)
-                               space_info->bytes_readonly += num_bytes;
-                       cache->reserved -= num_bytes;
-                       space_info->bytes_reserved -= num_bytes;
-                       space_info->reservation_progress++;
-               }
-               spin_unlock(&cache->lock);
-               spin_unlock(&space_info->lock);
-       } else {
-               spin_lock(&cache->lock);
+       spin_lock(&space_info->lock);
+       spin_lock(&cache->lock);
+       if (reserve != RESERVE_FREE) {
                if (cache->ro) {
                        ret = -EAGAIN;
                } else {
-                       if (reserve)
-                               cache->reserved += num_bytes;
-                       else
-                               cache->reserved -= num_bytes;
+                       cache->reserved += num_bytes;
+                       space_info->bytes_reserved += num_bytes;
+                       if (reserve == RESERVE_ALLOC) {
+                               BUG_ON(space_info->bytes_may_use < num_bytes);
+                               space_info->bytes_may_use -= num_bytes;
+                       }
                }
-               spin_unlock(&cache->lock);
+       } else {
+               if (cache->ro)
+                       space_info->bytes_readonly += num_bytes;
+               cache->reserved -= num_bytes;
+               space_info->bytes_reserved -= num_bytes;
+               space_info->reservation_progress++;
        }
+       spin_unlock(&cache->lock);
+       spin_unlock(&space_info->lock);
        return ret;
 }
 
@@ -4322,7 +4351,7 @@ static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
                } else if (cache->reserved_pinned > 0) {
                        len = min(len, cache->reserved_pinned);
                        cache->reserved_pinned -= len;
-                       cache->space_info->bytes_reserved += len;
+                       cache->space_info->bytes_may_use += len;
                }
                spin_unlock(&cache->lock);
                spin_unlock(&cache->space_info->lock);
@@ -4701,27 +4730,8 @@ void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
                WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
 
                btrfs_add_free_space(cache, buf->start, buf->len);
-               ret = btrfs_update_reserved_bytes(cache, buf->len, 0, 0);
-               if (ret == -EAGAIN) {
-                       /* block group became read-only */
-                       btrfs_update_reserved_bytes(cache, buf->len, 0, 1);
-                       goto out;
-               }
+               btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
 
-               ret = 1;
-               spin_lock(&block_rsv->lock);
-               if (block_rsv->reserved < block_rsv->size) {
-                       block_rsv->reserved += buf->len;
-                       ret = 0;
-               }
-               spin_unlock(&block_rsv->lock);
-
-               if (ret) {
-                       spin_lock(&cache->space_info->lock);
-                       cache->space_info->bytes_reserved -= buf->len;
-                       cache->space_info->reservation_progress++;
-                       spin_unlock(&cache->space_info->lock);
-               }
                goto out;
        }
 pin:
@@ -4881,6 +4891,8 @@ static noinline int find_free_extent(struct btrfs_trans_handle *trans,
        int last_ptr_loop = 0;
        int loop = 0;
        int index = 0;
+       int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
+               RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
        bool found_uncached_bg = false;
        bool failed_cluster_refill = false;
        bool failed_alloc = false;
@@ -5200,8 +5212,8 @@ checks:
                                             search_start - offset);
                BUG_ON(offset > search_start);
 
-               ret = btrfs_update_reserved_bytes(block_group, num_bytes, 1,
-                                           (data & BTRFS_BLOCK_GROUP_DATA));
+               ret = btrfs_update_reserved_bytes(block_group, num_bytes,
+                                                 alloc_type);
                if (ret == -EAGAIN) {
                        btrfs_add_free_space(block_group, offset, num_bytes);
                        goto loop;
@@ -5323,7 +5335,8 @@ static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
        int index = 0;
 
        spin_lock(&info->lock);
-       printk(KERN_INFO "space_info has %llu free, is %sfull\n",
+       printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
+              (unsigned long long)info->flags,
               (unsigned long long)(info->total_bytes - info->bytes_used -
                                    info->bytes_pinned - info->bytes_reserved -
                                    info->bytes_readonly),
@@ -5425,7 +5438,7 @@ int btrfs_free_reserved_extent(struct btrfs_root *root, u64 start, u64 len)
                ret = btrfs_discard_extent(root, start, len, NULL);
 
        btrfs_add_free_space(cache, start, len);
-       btrfs_update_reserved_bytes(cache, len, 0, 1);
+       btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
        btrfs_put_block_group(cache);
 
        trace_btrfs_reserved_extent_free(root, start, len);
@@ -5628,7 +5641,8 @@ int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
                put_caching_control(caching_ctl);
        }
 
-       ret = btrfs_update_reserved_bytes(block_group, ins->offset, 1, 1);
+       ret = btrfs_update_reserved_bytes(block_group, ins->offset,
+                                         RESERVE_ALLOC_NO_ACCOUNT);
        BUG_ON(ret);
        btrfs_put_block_group(block_group);
        ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
@@ -6594,7 +6608,7 @@ static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
            cache->reserved_pinned + num_bytes + min_allocable_bytes <=
            sinfo->total_bytes) {
                sinfo->bytes_readonly += num_bytes;
-               sinfo->bytes_reserved += cache->reserved_pinned;
+               sinfo->bytes_may_use += cache->reserved_pinned;
                cache->reserved_pinned = 0;
                cache->ro = 1;
                ret = 0;
@@ -6962,7 +6976,8 @@ int btrfs_free_block_groups(struct btrfs_fs_info *info)
                                        struct btrfs_space_info,
                                        list);
                if (space_info->bytes_pinned > 0 ||
-                   space_info->bytes_reserved > 0) {
+                   space_info->bytes_reserved > 0 ||
+                   space_info->bytes_may_use > 0) {
                        WARN_ON(1);
                        dump_space_info(space_info, 0, 0);
                }
index 41ac927..79c16a6 100644 (file)
@@ -2472,9 +2472,19 @@ int btrfs_trim_block_group(struct btrfs_block_group_cache *block_group,
                spin_unlock(&ctl->tree_lock);
 
                if (bytes >= minlen) {
-                       int update_ret;
-                       update_ret = btrfs_update_reserved_bytes(block_group,
-                                                                bytes, 1, 1);
+                       struct btrfs_space_info *space_info;
+                       int update = 0;
+
+                       space_info = block_group->space_info;
+                       spin_lock(&space_info->lock);
+                       spin_lock(&block_group->lock);
+                       if (!block_group->ro) {
+                               block_group->reserved += bytes;
+                               space_info->bytes_reserved += bytes;
+                               update = 1;
+                       }
+                       spin_unlock(&block_group->lock);
+                       spin_unlock(&space_info->lock);
 
                        ret = btrfs_error_discard_extent(fs_info->extent_root,
                                                         start,
@@ -2482,9 +2492,16 @@ int btrfs_trim_block_group(struct btrfs_block_group_cache *block_group,
                                                         &actually_trimmed);
 
                        btrfs_add_free_space(block_group, start, bytes);
-                       if (!update_ret)
-                               btrfs_update_reserved_bytes(block_group,
-                                                           bytes, 0, 1);
+                       if (update) {
+                               spin_lock(&space_info->lock);
+                               spin_lock(&block_group->lock);
+                               if (block_group->ro)
+                                       space_info->bytes_readonly += bytes;
+                               block_group->reserved -= bytes;
+                               space_info->bytes_reserved -= bytes;
+                               spin_unlock(&space_info->lock);
+                               spin_unlock(&block_group->lock);
+                       }
 
                        if (ret)
                                break;