Btrfs: two stage dirty block group writeout
authorChris Mason <clm@fb.com>
Sun, 5 Apr 2015 00:14:42 +0000 (17:14 -0700)
committerChris Mason <clm@fb.com>
Fri, 10 Apr 2015 21:07:11 +0000 (14:07 -0700)
Block group cache writeout is currently waiting on the pages for each
block group cache before moving on to writing the next one.  This commit
switches things around to send down all the caches and then wait on them
in batches.

The end result is much faster, since we're keeping the disk pipeline
full.

Signed-off-by: Chris Mason <clm@fb.com>
fs/btrfs/ctree.h
fs/btrfs/extent-tree.c
fs/btrfs/free-space-cache.c
fs/btrfs/free-space-cache.h

index e305ccd731fe22d00ec52bf9d1d76bb27f7c50b1..1df0d9db5332cbdbdbaf96e58eb4339bd4179e55 100644 (file)
@@ -1261,9 +1261,12 @@ struct btrfs_io_ctl {
        struct page *page;
        struct page **pages;
        struct btrfs_root *root;
+       struct inode *inode;
        unsigned long size;
        int index;
        int num_pages;
+       int entries;
+       int bitmaps;
        unsigned check_crcs:1;
 };
 
@@ -1332,6 +1335,9 @@ struct btrfs_block_group_cache {
 
        /* For dirty block groups */
        struct list_head dirty_list;
+       struct list_head io_list;
+
+       struct btrfs_io_ctl io_ctl;
 };
 
 /* delayed seq elem */
index 3d4b3d680feaa4819a8c215a4f2b0a2a45c9c8a7..40c95135d03717911a684f3e545b43cd47bdda08 100644 (file)
@@ -3388,7 +3388,11 @@ int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
        struct btrfs_block_group_cache *cache;
        struct btrfs_transaction *cur_trans = trans->transaction;
        int ret = 0;
+       int should_put;
        struct btrfs_path *path;
+       LIST_HEAD(io);
+       int num_started = 0;
+       int num_waited = 0;
 
        if (list_empty(&cur_trans->dirty_bgs))
                return 0;
@@ -3407,16 +3411,60 @@ int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
                cache = list_first_entry(&cur_trans->dirty_bgs,
                                         struct btrfs_block_group_cache,
                                         dirty_list);
+
+               /*
+                * this can happen if cache_save_setup re-dirties a block
+                * group that is already under IO.  Just wait for it to
+                * finish and then do it all again
+                */
+               if (!list_empty(&cache->io_list)) {
+                       list_del_init(&cache->io_list);
+                       btrfs_wait_cache_io(root, trans, cache,
+                                           &cache->io_ctl, path,
+                                           cache->key.objectid);
+                       btrfs_put_block_group(cache);
+                       num_waited++;
+               }
+
                list_del_init(&cache->dirty_list);
+               should_put = 1;
+
                if (cache->disk_cache_state == BTRFS_DC_CLEAR)
                        cache_save_setup(cache, trans, path);
+
                if (!ret)
-                       ret = btrfs_run_delayed_refs(trans, root,
-                                                    (unsigned long) -1);
-               if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP)
-                       btrfs_write_out_cache(root, trans, cache, path);
+                       ret = btrfs_run_delayed_refs(trans, root, (unsigned long) -1);
+
+               if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
+                       cache->io_ctl.inode = NULL;
+                       ret = btrfs_write_out_cache(root, trans, cache, path);
+                       if (ret == 0 && cache->io_ctl.inode) {
+                               num_started++;
+                               should_put = 0;
+                               list_add_tail(&cache->io_list, &io);
+                       } else {
+                               /*
+                                * if we failed to write the cache, the
+                                * generation will be bad and life goes on
+                                */
+                               ret = 0;
+                       }
+               }
                if (!ret)
                        ret = write_one_cache_group(trans, root, path, cache);
+
+               /* if its not on the io list, we need to put the block group */
+               if (should_put)
+                       btrfs_put_block_group(cache);
+       }
+
+       while (!list_empty(&io)) {
+               cache = list_first_entry(&io, struct btrfs_block_group_cache,
+                                        io_list);
+               list_del_init(&cache->io_list);
+               num_waited++;
+               btrfs_wait_cache_io(root, trans, cache,
+                                   &cache->io_ctl, path, cache->key.objectid);
                btrfs_put_block_group(cache);
        }
 
@@ -9013,6 +9061,7 @@ btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
        INIT_LIST_HEAD(&cache->bg_list);
        INIT_LIST_HEAD(&cache->ro_list);
        INIT_LIST_HEAD(&cache->dirty_list);
+       INIT_LIST_HEAD(&cache->io_list);
        btrfs_init_free_space_ctl(cache);
        atomic_set(&cache->trimming, 0);
 
index 47c2adb3ddf339428c5e90281dfabd0b8f64067c..6886ae063483aecc84570221ac0aba3dea930ce3 100644 (file)
@@ -170,13 +170,13 @@ static int __create_free_space_inode(struct btrfs_root *root,
        key.objectid = BTRFS_FREE_SPACE_OBJECTID;
        key.offset = offset;
        key.type = 0;
-
        ret = btrfs_insert_empty_item(trans, root, path, &key,
                                      sizeof(struct btrfs_free_space_header));
        if (ret < 0) {
                btrfs_release_path(path);
                return ret;
        }
+
        leaf = path->nodes[0];
        header = btrfs_item_ptr(leaf, path->slots[0],
                                struct btrfs_free_space_header);
@@ -296,6 +296,7 @@ static int io_ctl_init(struct btrfs_io_ctl *io_ctl, struct inode *inode,
        io_ctl->num_pages = num_pages;
        io_ctl->root = root;
        io_ctl->check_crcs = check_crcs;
+       io_ctl->inode = inode;
 
        return 0;
 }
@@ -303,6 +304,7 @@ static int io_ctl_init(struct btrfs_io_ctl *io_ctl, struct inode *inode,
 static void io_ctl_free(struct btrfs_io_ctl *io_ctl)
 {
        kfree(io_ctl->pages);
+       io_ctl->pages = NULL;
 }
 
 static void io_ctl_unmap_page(struct btrfs_io_ctl *io_ctl)
@@ -1092,6 +1094,61 @@ cleanup_write_cache_enospc(struct inode *inode,
                             GFP_NOFS);
 }
 
+int btrfs_wait_cache_io(struct btrfs_root *root,
+                       struct btrfs_trans_handle *trans,
+                       struct btrfs_block_group_cache *block_group,
+                       struct btrfs_io_ctl *io_ctl,
+                       struct btrfs_path *path, u64 offset)
+{
+       int ret;
+       struct inode *inode = io_ctl->inode;
+
+       root = root->fs_info->tree_root;
+
+       /* Flush the dirty pages in the cache file. */
+       ret = flush_dirty_cache(inode);
+       if (ret)
+               goto out;
+
+       /* Update the cache item to tell everyone this cache file is valid. */
+       ret = update_cache_item(trans, root, inode, path, offset,
+                               io_ctl->entries, io_ctl->bitmaps);
+out:
+       io_ctl_free(io_ctl);
+       if (ret) {
+               invalidate_inode_pages2(inode->i_mapping);
+               BTRFS_I(inode)->generation = 0;
+               if (block_group) {
+#ifdef DEBUG
+                       btrfs_err(root->fs_info,
+                               "failed to write free space cache for block group %llu",
+                               block_group->key.objectid);
+#endif
+               }
+       }
+       btrfs_update_inode(trans, root, inode);
+
+       if (block_group) {
+               spin_lock(&block_group->lock);
+
+               /*
+                * only mark this as written if we didn't get put back on
+                * the dirty list while waiting for IO.
+                */
+               if (!ret && list_empty(&block_group->dirty_list))
+                       block_group->disk_cache_state = BTRFS_DC_WRITTEN;
+               else if (ret)
+                       block_group->disk_cache_state = BTRFS_DC_ERROR;
+
+               spin_unlock(&block_group->lock);
+               io_ctl->inode = NULL;
+               iput(inode);
+       }
+
+       return ret;
+
+}
+
 /**
  * __btrfs_write_out_cache - write out cached info to an inode
  * @root - the root the inode belongs to
@@ -1108,20 +1165,22 @@ cleanup_write_cache_enospc(struct inode *inode,
 static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
                                   struct btrfs_free_space_ctl *ctl,
                                   struct btrfs_block_group_cache *block_group,
+                                  struct btrfs_io_ctl *io_ctl,
                                   struct btrfs_trans_handle *trans,
                                   struct btrfs_path *path, u64 offset)
 {
        struct extent_state *cached_state = NULL;
-       struct btrfs_io_ctl io_ctl;
        LIST_HEAD(bitmap_list);
        int entries = 0;
        int bitmaps = 0;
        int ret;
+       int must_iput = 0;
 
        if (!i_size_read(inode))
                return -1;
 
-       ret = io_ctl_init(&io_ctl, inode, root, 1);
+       WARN_ON(io_ctl->pages);
+       ret = io_ctl_init(io_ctl, inode, root, 1);
        if (ret)
                return -1;
 
@@ -1134,22 +1193,23 @@ static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
                        up_write(&block_group->data_rwsem);
                        BTRFS_I(inode)->generation = 0;
                        ret = 0;
+                       must_iput = 1;
                        goto out;
                }
                spin_unlock(&block_group->lock);
        }
 
        /* Lock all pages first so we can lock the extent safely. */
-       io_ctl_prepare_pages(&io_ctl, inode, 0);
+       io_ctl_prepare_pages(io_ctl, inode, 0);
 
        lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
                         0, &cached_state);
 
-       io_ctl_set_generation(&io_ctl, trans->transid);
+       io_ctl_set_generation(io_ctl, trans->transid);
 
        mutex_lock(&ctl->cache_writeout_mutex);
        /* Write out the extent entries in the free space cache */
-       ret = write_cache_extent_entries(&io_ctl, ctl,
+       ret = write_cache_extent_entries(io_ctl, ctl,
                                         block_group, &entries, &bitmaps,
                                         &bitmap_list);
        if (ret) {
@@ -1162,7 +1222,7 @@ static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
         * they will be added into free space cache after the transaction is
         * committed, we shouldn't lose them.
         */
-       ret = write_pinned_extent_entries(root, block_group, &io_ctl, &entries);
+       ret = write_pinned_extent_entries(root, block_group, io_ctl, &entries);
        if (ret) {
                mutex_unlock(&ctl->cache_writeout_mutex);
                goto out_nospc;
@@ -1173,16 +1233,16 @@ static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
         * locked while doing it because a concurrent trim can be manipulating
         * or freeing the bitmap.
         */
-       ret = write_bitmap_entries(&io_ctl, &bitmap_list);
+       ret = write_bitmap_entries(io_ctl, &bitmap_list);
        mutex_unlock(&ctl->cache_writeout_mutex);
        if (ret)
                goto out_nospc;
 
        /* Zero out the rest of the pages just to make sure */
-       io_ctl_zero_remaining_pages(&io_ctl);
+       io_ctl_zero_remaining_pages(io_ctl);
 
        /* Everything is written out, now we dirty the pages in the file. */
-       ret = btrfs_dirty_pages(root, inode, io_ctl.pages, io_ctl.num_pages,
+       ret = btrfs_dirty_pages(root, inode, io_ctl->pages, io_ctl->num_pages,
                                0, i_size_read(inode), &cached_state);
        if (ret)
                goto out_nospc;
@@ -1193,30 +1253,39 @@ static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
         * Release the pages and unlock the extent, we will flush
         * them out later
         */
-       io_ctl_drop_pages(&io_ctl);
+       io_ctl_drop_pages(io_ctl);
 
        unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
                             i_size_read(inode) - 1, &cached_state, GFP_NOFS);
 
-       /* Flush the dirty pages in the cache file. */
-       ret = flush_dirty_cache(inode);
+       /*
+        * at this point the pages are under IO and we're happy,
+        * The caller is responsible for waiting on them and updating the
+        * the cache and the inode
+        */
+       io_ctl->entries = entries;
+       io_ctl->bitmaps = bitmaps;
+
+       ret = btrfs_fdatawrite_range(inode, 0, (u64)-1);
        if (ret)
                goto out;
 
-       /* Update the cache item to tell everyone this cache file is valid. */
-       ret = update_cache_item(trans, root, inode, path, offset,
-                               entries, bitmaps);
+       return 0;
+
 out:
-       io_ctl_free(&io_ctl);
+       io_ctl->inode = NULL;
+       io_ctl_free(io_ctl);
        if (ret) {
                invalidate_inode_pages2(inode->i_mapping);
                BTRFS_I(inode)->generation = 0;
        }
        btrfs_update_inode(trans, root, inode);
+       if (must_iput)
+               iput(inode);
        return ret;
 
 out_nospc:
-       cleanup_write_cache_enospc(inode, &io_ctl, &cached_state, &bitmap_list);
+       cleanup_write_cache_enospc(inode, io_ctl, &cached_state, &bitmap_list);
 
        if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
                up_write(&block_group->data_rwsem);
@@ -1232,7 +1301,6 @@ int btrfs_write_out_cache(struct btrfs_root *root,
        struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
        struct inode *inode;
        int ret = 0;
-       enum btrfs_disk_cache_state dcs = BTRFS_DC_WRITTEN;
 
        root = root->fs_info->tree_root;
 
@@ -1253,22 +1321,28 @@ int btrfs_write_out_cache(struct btrfs_root *root,
        if (IS_ERR(inode))
                return 0;
 
-       ret = __btrfs_write_out_cache(root, inode, ctl, block_group, trans,
+       ret = __btrfs_write_out_cache(root, inode, ctl, block_group,
+                                     &block_group->io_ctl, trans,
                                      path, block_group->key.objectid);
        if (ret) {
-               dcs = BTRFS_DC_ERROR;
-               ret = 0;
 #ifdef DEBUG
                btrfs_err(root->fs_info,
                        "failed to write free space cache for block group %llu",
                        block_group->key.objectid);
 #endif
+               spin_lock(&block_group->lock);
+               block_group->disk_cache_state = BTRFS_DC_ERROR;
+               spin_unlock(&block_group->lock);
+
+               block_group->io_ctl.inode = NULL;
+               iput(inode);
        }
 
-       spin_lock(&block_group->lock);
-       block_group->disk_cache_state = dcs;
-       spin_unlock(&block_group->lock);
-       iput(inode);
+       /*
+        * if ret == 0 the caller is expected to call btrfs_wait_cache_io
+        * to wait for IO and put the inode
+        */
+
        return ret;
 }
 
@@ -3331,11 +3405,14 @@ int btrfs_write_out_ino_cache(struct btrfs_root *root,
 {
        struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
        int ret;
+       struct btrfs_io_ctl io_ctl;
 
        if (!btrfs_test_opt(root, INODE_MAP_CACHE))
                return 0;
 
-       ret = __btrfs_write_out_cache(root, inode, ctl, NULL, trans, path, 0);
+       ret = __btrfs_write_out_cache(root, inode, ctl, NULL, &io_ctl,
+                                     trans, path, 0) ||
+               btrfs_wait_cache_io(root, trans, NULL, &io_ctl, path, 0);
        if (ret) {
                btrfs_delalloc_release_metadata(inode, inode->i_size);
 #ifdef DEBUG
index 88b2238a0aeda1cf4c91f7a0d32de09b19c2174c..c4339863af05d4b6b214d7bc7399fa4e691fff18 100644 (file)
@@ -48,6 +48,8 @@ struct btrfs_free_space_op {
                           struct btrfs_free_space *info);
 };
 
+struct btrfs_io_ctl;
+
 struct inode *lookup_free_space_inode(struct btrfs_root *root,
                                      struct btrfs_block_group_cache
                                      *block_group, struct btrfs_path *path);
@@ -63,11 +65,15 @@ int btrfs_truncate_free_space_cache(struct btrfs_root *root,
                                    struct inode *inode);
 int load_free_space_cache(struct btrfs_fs_info *fs_info,
                          struct btrfs_block_group_cache *block_group);
+int btrfs_wait_cache_io(struct btrfs_root *root,
+                       struct btrfs_trans_handle *trans,
+                       struct btrfs_block_group_cache *block_group,
+                       struct btrfs_io_ctl *io_ctl,
+                       struct btrfs_path *path, u64 offset);
 int btrfs_write_out_cache(struct btrfs_root *root,
                          struct btrfs_trans_handle *trans,
                          struct btrfs_block_group_cache *block_group,
                          struct btrfs_path *path);
-
 struct inode *lookup_free_ino_inode(struct btrfs_root *root,
                                    struct btrfs_path *path);
 int create_free_ino_inode(struct btrfs_root *root,