Btrfs: Lower contention on the csum mutex
authorChris Mason <chris.mason@oracle.com>
Fri, 15 Aug 2008 19:34:18 +0000 (15:34 -0400)
committerChris Mason <chris.mason@oracle.com>
Thu, 25 Sep 2008 15:04:06 +0000 (11:04 -0400)
This takes the csum mutex deeper in the call chain and releases it
more often.

Signed-off-by: Chris Mason <chris.mason@oracle.com>
fs/btrfs/async-thread.c
fs/btrfs/disk-io.c
fs/btrfs/file-item.c
fs/btrfs/inode.c

index 958cd8b5f0d7fa77bf087a72f16e7a511df14827..2ee301740195a0d04c6040958d6ce8e6851a7a3e 100644 (file)
@@ -48,6 +48,7 @@ struct btrfs_worker_thread {
 
        /* number of things on the pending list */
        atomic_t num_pending;
+
        unsigned long sequence;
 
        /* protects the pending list. */
@@ -242,7 +243,7 @@ static struct btrfs_worker_thread *next_worker(struct btrfs_workers *workers)
        worker = list_entry(next, struct btrfs_worker_thread, worker_list);
        atomic_inc(&worker->num_pending);
        worker->sequence++;
-       if (worker->sequence % 4 == 0)
+       if (worker->sequence % workers->idle_thresh == 0)
                list_move_tail(next, &workers->worker_list);
        return worker;
 }
index 1aed1f4616b623a9e66e75510095a25239f7da12..92e14dd9bddbe673d582b90202e2135f0fc02e66 100644 (file)
@@ -1386,7 +1386,14 @@ struct btrfs_root *open_ctree(struct super_block *sb,
         * devices
         */
        fs_info->submit_workers.idle_thresh = 64;
-       fs_info->workers.idle_thresh = 32;
+
+       /* fs_info->workers is responsible for checksumming file data
+        * blocks and metadata.  Using a larger idle thresh allows each
+        * worker thread to operate on things in roughly the order they
+        * were sent by the writeback daemons, improving overall locality
+        * of the IO going down the pipe.
+        */
+       fs_info->workers.idle_thresh = 128;
 
        btrfs_init_workers(&fs_info->fixup_workers, "fixup", 1);
        btrfs_init_workers(&fs_info->endio_workers, "endio",
index 51aba8cee7ce3cc7add24c0bfe52e40cde09d12c..9454e0a07c893492439b9016b741aae378ba2b0c 100644 (file)
@@ -321,6 +321,7 @@ again:
        file_key.offset = offset;
        btrfs_set_key_type(&file_key, BTRFS_CSUM_ITEM_KEY);
 
+       mutex_lock(&BTRFS_I(inode)->csum_mutex);
        item = btrfs_lookup_csum(trans, root, path, objectid, offset, 1);
        if (!IS_ERR(item)) {
                leaf = path->nodes[0];
@@ -367,7 +368,7 @@ again:
        ret = btrfs_search_slot(trans, root, &file_key, path,
                                BTRFS_CRC32_SIZE, 1);
        if (ret < 0)
-               goto fail;
+               goto fail_unlock;
        if (ret == 0) {
                BUG();
        }
@@ -411,10 +412,10 @@ insert:
        ret = btrfs_insert_empty_item(trans, root, path, &file_key,
                                      ins_size);
        if (ret < 0)
-               goto fail;
+               goto fail_unlock;
        if (ret != 0) {
                WARN_ON(1);
-               goto fail;
+               goto fail_unlock;
        }
 csum:
        leaf = path->nodes[0];
@@ -427,6 +428,8 @@ found:
        item_end = (struct btrfs_csum_item *)((unsigned char *)item_end +
                                      btrfs_item_size_nr(leaf, path->slots[0]));
        eb_token = NULL;
+       mutex_unlock(&BTRFS_I(inode)->csum_mutex);
+       cond_resched();
 next_sector:
 
        if (!eb_token ||
@@ -467,13 +470,18 @@ next_sector:
                eb_token = NULL;
        }
        btrfs_mark_buffer_dirty(path->nodes[0]);
+       cond_resched();
        if (total_bytes < sums->len) {
                btrfs_release_path(root, path);
                goto again;
        }
-fail:
+out:
        btrfs_free_path(path);
        return ret;
+
+fail_unlock:
+       mutex_unlock(&BTRFS_I(inode)->csum_mutex);
+       goto out;
 }
 
 int btrfs_csum_truncate(struct btrfs_trans_handle *trans,
index 33b990878d84f40a2095786c4b9b04f620431867..65107894a5bbd6181fb7c00ba335850d84d28516 100644 (file)
@@ -415,10 +415,8 @@ static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
        btrfs_set_trans_block_group(trans, inode);
        list_for_each(cur, list) {
                sum = list_entry(cur, struct btrfs_ordered_sum, list);
-               mutex_lock(&BTRFS_I(inode)->csum_mutex);
                btrfs_csum_file_blocks(trans, BTRFS_I(inode)->root,
                                       inode, sum);
-               mutex_unlock(&BTRFS_I(inode)->csum_mutex);
        }
        return 0;
 }