Btrfs: Search data ordered extents first for checksums on read
[linux-2.6-block.git] / fs / btrfs / disk-io.c
index b9a53646ceb2e8dd90a1f8b7dee205a68c36de6f..7ce3f83c5dd633bab76b3a535361b28013b134d3 100644 (file)
@@ -16,6 +16,7 @@
  * Boston, MA 021110-1307, USA.
  */
 
+#include <linux/version.h>
 #include <linux/fs.h>
 #include <linux/blkdev.h>
 #include <linux/scatterlist.h>
 #include <linux/writeback.h>
 #include <linux/buffer_head.h> // for block_sync_page
 #include <linux/workqueue.h>
+#include <linux/kthread.h>
+#if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,20)
+# include <linux/freezer.h>
+#else
+# include <linux/sched.h>
+#endif
 #include "crc32c.h"
 #include "ctree.h"
 #include "disk-io.h"
@@ -31,6 +38,8 @@
 #include "btrfs_inode.h"
 #include "volumes.h"
 #include "print-tree.h"
+#include "async-thread.h"
+#include "locking.h"
 
 #if 0
 static int check_tree_block(struct btrfs_root *root, struct extent_buffer *buf)
@@ -46,8 +55,7 @@ static int check_tree_block(struct btrfs_root *root, struct extent_buffer *buf)
 #endif
 
 static struct extent_io_ops btree_extent_io_ops;
-static struct workqueue_struct *end_io_workqueue;
-static struct workqueue_struct *async_submit_workqueue;
+static void end_workqueue_fn(struct btrfs_work *work);
 
 struct end_io_wq {
        struct bio *bio;
@@ -57,6 +65,7 @@ struct end_io_wq {
        int error;
        int metadata;
        struct list_head list;
+       struct btrfs_work work;
 };
 
 struct async_submit_bio {
@@ -66,6 +75,7 @@ struct async_submit_bio {
        extent_submit_bio_hook_t *submit_bio_hook;
        int rw;
        int mirror_num;
+       struct btrfs_work work;
 };
 
 struct extent_map *btree_get_extent(struct inode *inode, struct page *page,
@@ -285,7 +295,6 @@ int csum_dirty_buffer(struct btrfs_root *root, struct page *page)
        ret = btree_read_extent_buffer_pages(root, eb, start + PAGE_CACHE_SIZE,
                                             btrfs_header_generation(eb));
        BUG_ON(ret);
-       btrfs_clear_buffer_defrag(eb);
        found_start = btrfs_header_bytenr(eb);
        if (found_start != start) {
                printk("warning: eb start incorrect %Lu buffer %Lu len %lu\n",
@@ -345,7 +354,6 @@ int btree_readpage_end_io_hook(struct page *page, u64 start, u64 end,
        }
        eb = alloc_extent_buffer(tree, start, len, page, GFP_NOFS);
 
-       btrfs_clear_buffer_defrag(eb);
        found_start = btrfs_header_bytenr(eb);
        if (found_start != start) {
                ret = -EIO;
@@ -373,7 +381,6 @@ int btree_readpage_end_io_hook(struct page *page, u64 start, u64 end,
 
        end = min_t(u64, eb->len, PAGE_CACHE_SIZE);
        end = eb->start + end - 1;
-       release_extent_buffer_tail_pages(eb);
 err:
        free_extent_buffer(eb);
 out:
@@ -389,7 +396,6 @@ static int end_workqueue_bio(struct bio *bio,
 {
        struct end_io_wq *end_io_wq = bio->bi_private;
        struct btrfs_fs_info *fs_info;
-       unsigned long flags;
 
 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
        if (bio->bi_size)
@@ -397,11 +403,14 @@ static int end_workqueue_bio(struct bio *bio,
 #endif
 
        fs_info = end_io_wq->info;
-       spin_lock_irqsave(&fs_info->end_io_work_lock, flags);
        end_io_wq->error = err;
-       list_add_tail(&end_io_wq->list, &fs_info->end_io_work_list);
-       spin_unlock_irqrestore(&fs_info->end_io_work_lock, flags);
-       queue_work(end_io_workqueue, &fs_info->end_io_work);
+       end_io_wq->work.func = end_workqueue_fn;
+       end_io_wq->work.flags = 0;
+       if (bio->bi_rw & (1 << BIO_RW))
+               btrfs_queue_worker(&fs_info->endio_write_workers,
+                                  &end_io_wq->work);
+       else
+               btrfs_queue_worker(&fs_info->endio_workers, &end_io_wq->work);
 
 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
        return 0;
@@ -428,6 +437,19 @@ int btrfs_bio_wq_end_io(struct btrfs_fs_info *info, struct bio *bio,
        return 0;
 }
 
+static void run_one_async_submit(struct btrfs_work *work)
+{
+       struct btrfs_fs_info *fs_info;
+       struct async_submit_bio *async;
+
+       async = container_of(work, struct  async_submit_bio, work);
+       fs_info = BTRFS_I(async->inode)->root->fs_info;
+       atomic_dec(&fs_info->nr_async_submits);
+       async->submit_bio_hook(async->inode, async->rw, async->bio,
+                              async->mirror_num);
+       kfree(async);
+}
+
 int btrfs_wq_submit_bio(struct btrfs_fs_info *fs_info, struct inode *inode,
                        int rw, struct bio *bio, int mirror_num,
                        extent_submit_bio_hook_t *submit_bio_hook)
@@ -443,13 +465,10 @@ int btrfs_wq_submit_bio(struct btrfs_fs_info *fs_info, struct inode *inode,
        async->bio = bio;
        async->mirror_num = mirror_num;
        async->submit_bio_hook = submit_bio_hook;
-
-       spin_lock(&fs_info->async_submit_work_lock);
-       list_add_tail(&async->list, &fs_info->async_submit_work_list);
+       async->work.func = run_one_async_submit;
+       async->work.flags = 0;
        atomic_inc(&fs_info->nr_async_submits);
-       spin_unlock(&fs_info->async_submit_work_lock);
-
-       queue_work(async_submit_workqueue, &fs_info->async_submit_work);
+       btrfs_queue_worker(&fs_info->workers, &async->work);
        return 0;
 }
 
@@ -462,19 +481,32 @@ static int __btree_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
 
        offset = bio->bi_sector << 9;
 
+       /*
+        * when we're called for a write, we're already in the async
+        * submission context.  Just jump ingo btrfs_map_bio
+        */
        if (rw & (1 << BIO_RW)) {
-               return btrfs_map_bio(BTRFS_I(inode)->root, rw, bio, mirror_num);
+               return btrfs_map_bio(BTRFS_I(inode)->root, rw, bio,
+                                    mirror_num, 0);
        }
 
+       /*
+        * called for a read, do the setup so that checksum validation
+        * can happen in the async kernel threads
+        */
        ret = btrfs_bio_wq_end_io(root->fs_info, bio, 1);
        BUG_ON(ret);
 
-       return btrfs_map_bio(BTRFS_I(inode)->root, rw, bio, mirror_num);
+       return btrfs_map_bio(BTRFS_I(inode)->root, rw, bio, mirror_num, 1);
 }
 
 static int btree_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
                                 int mirror_num)
 {
+       /*
+        * kthread helpers are used to submit writes so that checksumming
+        * can happen in parallel across all CPUs
+        */
        if (!(rw & (1 << BIO_RW))) {
                return __btree_submit_bio_hook(inode, rw, bio, mirror_num);
        }
@@ -530,21 +562,21 @@ static int btree_releasepage(struct page *page, gfp_t gfp_flags)
        struct extent_map_tree *map;
        int ret;
 
-       if (page_count(page) > 3) {
-               /* once for page->private, once for the caller, once
-                * once for the page cache
-                */
-               return 0;
-       }
        tree = &BTRFS_I(page->mapping->host)->io_tree;
        map = &BTRFS_I(page->mapping->host)->extent_tree;
+
        ret = try_release_extent_state(map, tree, page, gfp_flags);
+       if (!ret) {
+               return 0;
+       }
+
+       ret = try_release_extent_buffer(tree, page);
        if (ret == 1) {
-               invalidate_extent_lru(tree, page_offset(page), PAGE_CACHE_SIZE);
                ClearPagePrivate(page);
                set_page_private(page, 0);
                page_cache_release(page);
        }
+
        return ret;
 }
 
@@ -555,7 +587,8 @@ static void btree_invalidatepage(struct page *page, unsigned long offset)
        extent_invalidatepage(tree, page, offset);
        btree_releasepage(page, GFP_NOFS);
        if (PagePrivate(page)) {
-               invalidate_extent_lru(tree, page_offset(page), PAGE_CACHE_SIZE);
+               printk("warning page private not zero on page %Lu\n",
+                      page_offset(page));
                ClearPagePrivate(page);
                set_page_private(page, 0);
                page_cache_release(page);
@@ -658,9 +691,11 @@ int clean_tree_block(struct btrfs_trans_handle *trans, struct btrfs_root *root,
 {
        struct inode *btree_inode = root->fs_info->btree_inode;
        if (btrfs_header_generation(buf) ==
-           root->fs_info->running_transaction->transid)
+           root->fs_info->running_transaction->transid) {
+               WARN_ON(!btrfs_tree_locked(buf));
                clear_extent_buffer_dirty(&BTRFS_I(btree_inode)->io_tree,
                                          buf);
+       }
        return 0;
 }
 
@@ -697,10 +732,13 @@ static int __setup_root(u32 nodesize, u32 leafsize, u32 sectorsize,
        root->in_sysfs = 0;
 
        INIT_LIST_HEAD(&root->dirty_list);
+       spin_lock_init(&root->node_lock);
+       mutex_init(&root->objectid_mutex);
        memset(&root->root_key, 0, sizeof(root->root_key));
        memset(&root->root_item, 0, sizeof(root->root_item));
        memset(&root->defrag_progress, 0, sizeof(root->defrag_progress));
        memset(&root->root_kobj, 0, sizeof(root->root_kobj));
+       root->defrag_trans_start = fs_info->generation;
        init_completion(&root->kobj_unregister);
        root->defrag_running = 0;
        root->defrag_level = 0;
@@ -973,9 +1011,16 @@ void btrfs_unplug_io_fn(struct backing_dev_info *bdi, struct page *page)
        spin_lock(&em_tree->lock);
        em = lookup_extent_mapping(em_tree, offset, PAGE_CACHE_SIZE);
        spin_unlock(&em_tree->lock);
-       if (!em)
+       if (!em) {
+               __unplug_io_fn(bdi, page);
                return;
+       }
 
+       if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
+               free_extent_map(em);
+               __unplug_io_fn(bdi, page);
+               return;
+       }
        offset = offset - em->start;
        btrfs_unplug_page(&BTRFS_I(inode)->root->fs_info->mapping_tree,
                          em->block_start + offset, page);
@@ -1036,95 +1081,118 @@ static int bio_ready_for_csum(struct bio *bio)
        return ret;
 }
 
-#if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,18)
-static void btrfs_end_io_csum(void *p)
-#else
-static void btrfs_end_io_csum(struct work_struct *work)
-#endif
+/*
+ * called by the kthread helper functions to finally call the bio end_io
+ * functions.  This is where read checksum verification actually happens
+ */
+static void end_workqueue_fn(struct btrfs_work *work)
 {
-#if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,18)
-       struct btrfs_fs_info *fs_info = p;
-#else
-       struct btrfs_fs_info *fs_info = container_of(work,
-                                                    struct btrfs_fs_info,
-                                                    end_io_work);
-#endif
-       unsigned long flags;
-       struct end_io_wq *end_io_wq;
        struct bio *bio;
-       struct list_head *next;
+       struct end_io_wq *end_io_wq;
+       struct btrfs_fs_info *fs_info;
        int error;
-       int was_empty;
 
-       while(1) {
-               spin_lock_irqsave(&fs_info->end_io_work_lock, flags);
-               if (list_empty(&fs_info->end_io_work_list)) {
-                       spin_unlock_irqrestore(&fs_info->end_io_work_lock,
-                                              flags);
-                       return;
-               }
-               next = fs_info->end_io_work_list.next;
-               list_del(next);
-               spin_unlock_irqrestore(&fs_info->end_io_work_lock, flags);
-
-               end_io_wq = list_entry(next, struct end_io_wq, list);
-
-               bio = end_io_wq->bio;
-               if (end_io_wq->metadata && !bio_ready_for_csum(bio)) {
-                       spin_lock_irqsave(&fs_info->end_io_work_lock, flags);
-                       was_empty = list_empty(&fs_info->end_io_work_list);
-                       list_add_tail(&end_io_wq->list,
-                                     &fs_info->end_io_work_list);
-                       spin_unlock_irqrestore(&fs_info->end_io_work_lock,
-                                              flags);
-                       if (was_empty)
-                               return;
-                       continue;
-               }
-               error = end_io_wq->error;
-               bio->bi_private = end_io_wq->private;
-               bio->bi_end_io = end_io_wq->end_io;
-               kfree(end_io_wq);
+       end_io_wq = container_of(work, struct end_io_wq, work);
+       bio = end_io_wq->bio;
+       fs_info = end_io_wq->info;
+
+       /* metadata bios are special because the whole tree block must
+        * be checksummed at once.  This makes sure the entire block is in
+        * ram and up to date before trying to verify things.  For
+        * blocksize <= pagesize, it is basically a noop
+        */
+       if (end_io_wq->metadata && !bio_ready_for_csum(bio)) {
+               btrfs_queue_worker(&fs_info->endio_workers,
+                                  &end_io_wq->work);
+               return;
+       }
+       error = end_io_wq->error;
+       bio->bi_private = end_io_wq->private;
+       bio->bi_end_io = end_io_wq->end_io;
+       kfree(end_io_wq);
 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
-               bio_endio(bio, bio->bi_size, error);
+       bio_endio(bio, bio->bi_size, error);
 #else
-               bio_endio(bio, error);
+       bio_endio(bio, error);
 #endif
-       }
 }
 
-#if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,18)
-static void btrfs_async_submit_work(void *p)
-#else
-static void btrfs_async_submit_work(struct work_struct *work)
-#endif
+static int cleaner_kthread(void *arg)
 {
-#if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,18)
-       struct btrfs_fs_info *fs_info = p;
-#else
-       struct btrfs_fs_info *fs_info = container_of(work,
-                                                    struct btrfs_fs_info,
-                                                    async_submit_work);
-#endif
-       struct async_submit_bio *async;
-       struct list_head *next;
+       struct btrfs_root *root = arg;
 
-       while(1) {
-               spin_lock(&fs_info->async_submit_work_lock);
-               if (list_empty(&fs_info->async_submit_work_list)) {
-                       spin_unlock(&fs_info->async_submit_work_lock);
-                       return;
+       do {
+               smp_mb();
+               if (root->fs_info->closing)
+                       break;
+
+               vfs_check_frozen(root->fs_info->sb, SB_FREEZE_WRITE);
+               mutex_lock(&root->fs_info->cleaner_mutex);
+               btrfs_clean_old_snapshots(root);
+               mutex_unlock(&root->fs_info->cleaner_mutex);
+
+               if (freezing(current)) {
+                       refrigerator();
+               } else {
+                       smp_mb();
+                       if (root->fs_info->closing)
+                               break;
+                       set_current_state(TASK_INTERRUPTIBLE);
+                       schedule();
+                       __set_current_state(TASK_RUNNING);
                }
-               next = fs_info->async_submit_work_list.next;
-               list_del(next);
-               atomic_dec(&fs_info->nr_async_submits);
-               spin_unlock(&fs_info->async_submit_work_lock);
-
-               async = list_entry(next, struct async_submit_bio, list);
-               async->submit_bio_hook(async->inode, async->rw, async->bio,
-                                      async->mirror_num);
-               kfree(async);
-       }
+       } while (!kthread_should_stop());
+       return 0;
+}
+
+static int transaction_kthread(void *arg)
+{
+       struct btrfs_root *root = arg;
+       struct btrfs_trans_handle *trans;
+       struct btrfs_transaction *cur;
+       unsigned long now;
+       unsigned long delay;
+       int ret;
+
+       do {
+               smp_mb();
+               if (root->fs_info->closing)
+                       break;
+
+               delay = HZ * 30;
+               vfs_check_frozen(root->fs_info->sb, SB_FREEZE_WRITE);
+               mutex_lock(&root->fs_info->transaction_kthread_mutex);
+
+               mutex_lock(&root->fs_info->trans_mutex);
+               cur = root->fs_info->running_transaction;
+               if (!cur) {
+                       mutex_unlock(&root->fs_info->trans_mutex);
+                       goto sleep;
+               }
+               now = get_seconds();
+               if (now < cur->start_time || now - cur->start_time < 30) {
+                       mutex_unlock(&root->fs_info->trans_mutex);
+                       delay = HZ * 5;
+                       goto sleep;
+               }
+               mutex_unlock(&root->fs_info->trans_mutex);
+               trans = btrfs_start_transaction(root, 1);
+               ret = btrfs_commit_transaction(trans, root);
+sleep:
+               wake_up_process(root->fs_info->cleaner_kthread);
+               mutex_unlock(&root->fs_info->transaction_kthread_mutex);
+
+               if (freezing(current)) {
+                       refrigerator();
+               } else {
+                       if (root->fs_info->closing)
+                               break;
+                       set_current_state(TASK_INTERRUPTIBLE);
+                       schedule_timeout(delay);
+                       __set_current_state(TASK_RUNNING);
+               }
+       } while (!kthread_should_stop());
+       return 0;
 }
 
 struct btrfs_root *open_ctree(struct super_block *sb,
@@ -1149,25 +1217,18 @@ struct btrfs_root *open_ctree(struct super_block *sb,
                                              GFP_NOFS);
        int ret;
        int err = -EINVAL;
+
        struct btrfs_super_block *disk_super;
 
        if (!extent_root || !tree_root || !fs_info) {
                err = -ENOMEM;
                goto fail;
        }
-       end_io_workqueue = create_workqueue("btrfs-end-io");
-       BUG_ON(!end_io_workqueue);
-       async_submit_workqueue = create_workqueue("btrfs-async-submit");
-
        INIT_RADIX_TREE(&fs_info->fs_roots_radix, GFP_NOFS);
        INIT_LIST_HEAD(&fs_info->trans_list);
        INIT_LIST_HEAD(&fs_info->dead_roots);
        INIT_LIST_HEAD(&fs_info->hashers);
-       INIT_LIST_HEAD(&fs_info->end_io_work_list);
-       INIT_LIST_HEAD(&fs_info->async_submit_work_list);
        spin_lock_init(&fs_info->hash_lock);
-       spin_lock_init(&fs_info->end_io_work_lock);
-       spin_lock_init(&fs_info->async_submit_work_lock);
        spin_lock_init(&fs_info->delalloc_lock);
        spin_lock_init(&fs_info->new_trans_lock);
 
@@ -1181,6 +1242,7 @@ struct btrfs_root *open_ctree(struct super_block *sb,
        INIT_LIST_HEAD(&fs_info->space_info);
        btrfs_mapping_init(&fs_info->mapping_tree);
        atomic_set(&fs_info->nr_async_submits, 0);
+       atomic_set(&fs_info->throttles, 0);
        fs_info->sb = sb;
        fs_info->max_extent = (u64)-1;
        fs_info->max_inline = 8192 * 1024;
@@ -1188,6 +1250,7 @@ struct btrfs_root *open_ctree(struct super_block *sb,
        fs_info->btree_inode = new_inode(sb);
        fs_info->btree_inode->i_ino = 1;
        fs_info->btree_inode->i_nlink = 1;
+       fs_info->thread_pool_size = min(num_online_cpus() + 2, 8);
 
        sb->s_blocksize = 4096;
        sb->s_blocksize_bits = blksize_bits(4096);
@@ -1221,16 +1284,6 @@ struct btrfs_root *open_ctree(struct super_block *sb,
                             fs_info->btree_inode->i_mapping, GFP_NOFS);
        fs_info->do_barriers = 1;
 
-#if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,18)
-       INIT_WORK(&fs_info->end_io_work, btrfs_end_io_csum, fs_info);
-       INIT_WORK(&fs_info->async_submit_work, btrfs_async_submit_work,
-                 fs_info);
-       INIT_WORK(&fs_info->trans_work, btrfs_transaction_cleaner, fs_info);
-#else
-       INIT_WORK(&fs_info->end_io_work, btrfs_end_io_csum);
-       INIT_WORK(&fs_info->async_submit_work, btrfs_async_submit_work);
-       INIT_DELAYED_WORK(&fs_info->trans_work, btrfs_transaction_cleaner);
-#endif
        BTRFS_I(fs_info->btree_inode)->root = tree_root;
        memset(&BTRFS_I(fs_info->btree_inode)->location, 0,
               sizeof(struct btrfs_key));
@@ -1238,7 +1291,14 @@ struct btrfs_root *open_ctree(struct super_block *sb,
        mapping_set_gfp_mask(fs_info->btree_inode->i_mapping, GFP_NOFS);
 
        mutex_init(&fs_info->trans_mutex);
-       mutex_init(&fs_info->fs_mutex);
+       mutex_init(&fs_info->drop_mutex);
+       mutex_init(&fs_info->alloc_mutex);
+       mutex_init(&fs_info->chunk_mutex);
+       mutex_init(&fs_info->transaction_kthread_mutex);
+       mutex_init(&fs_info->cleaner_mutex);
+       mutex_init(&fs_info->volume_mutex);
+       init_waitqueue_head(&fs_info->transaction_throttle);
+       init_waitqueue_head(&fs_info->transaction_wait);
 
 #if 0
        ret = add_hasher(fs_info, "crc32c");
@@ -1270,6 +1330,24 @@ struct btrfs_root *open_ctree(struct super_block *sb,
        if (err)
                goto fail_sb_buffer;
 
+       /*
+        * we need to start all the end_io workers up front because the
+        * queue work function gets called at interrupt time, and so it
+        * cannot dynamically grow.
+        */
+       btrfs_init_workers(&fs_info->workers, fs_info->thread_pool_size);
+       btrfs_init_workers(&fs_info->submit_workers, fs_info->thread_pool_size);
+       btrfs_init_workers(&fs_info->fixup_workers, 1);
+       btrfs_init_workers(&fs_info->endio_workers, fs_info->thread_pool_size);
+       btrfs_init_workers(&fs_info->endio_write_workers,
+                          fs_info->thread_pool_size);
+       btrfs_start_workers(&fs_info->workers, 1);
+       btrfs_start_workers(&fs_info->submit_workers, 1);
+       btrfs_start_workers(&fs_info->fixup_workers, 1);
+       btrfs_start_workers(&fs_info->endio_workers, fs_info->thread_pool_size);
+       btrfs_start_workers(&fs_info->endio_write_workers,
+                           fs_info->thread_pool_size);
+
        err = -EINVAL;
        if (btrfs_super_num_devices(disk_super) > fs_devices->open_devices) {
                printk("Btrfs: wanted %llu devices, but found %llu\n",
@@ -1302,9 +1380,9 @@ struct btrfs_root *open_ctree(struct super_block *sb,
                goto fail_sb_buffer;
        }
 
-       mutex_lock(&fs_info->fs_mutex);
-
+       mutex_lock(&fs_info->chunk_mutex);
        ret = btrfs_read_sys_array(tree_root);
+       mutex_unlock(&fs_info->chunk_mutex);
        if (ret) {
                printk("btrfs: failed to read the system array on %s\n",
                       sb->s_id);
@@ -1326,7 +1404,9 @@ struct btrfs_root *open_ctree(struct super_block *sb,
                 (unsigned long)btrfs_header_chunk_tree_uuid(chunk_root->node),
                 BTRFS_UUID_SIZE);
 
+       mutex_lock(&fs_info->chunk_mutex);
        ret = btrfs_read_chunk_tree(chunk_root);
+       mutex_unlock(&fs_info->chunk_mutex);
        BUG_ON(ret);
 
        btrfs_close_extra_devices(fs_devices);
@@ -1361,18 +1441,33 @@ struct btrfs_root *open_ctree(struct super_block *sb,
        fs_info->data_alloc_profile = (u64)-1;
        fs_info->metadata_alloc_profile = (u64)-1;
        fs_info->system_alloc_profile = fs_info->metadata_alloc_profile;
+       fs_info->cleaner_kthread = kthread_run(cleaner_kthread, tree_root,
+                                              "btrfs-cleaner");
+       if (!fs_info->cleaner_kthread)
+               goto fail_extent_root;
+
+       fs_info->transaction_kthread = kthread_run(transaction_kthread,
+                                                  tree_root,
+                                                  "btrfs-transaction");
+       if (!fs_info->transaction_kthread)
+               goto fail_cleaner;
+
 
-       mutex_unlock(&fs_info->fs_mutex);
        return tree_root;
 
+fail_cleaner:
+       kthread_stop(fs_info->cleaner_kthread);
 fail_extent_root:
        free_extent_buffer(extent_root->node);
 fail_tree_root:
        free_extent_buffer(tree_root->node);
 fail_sys_array:
-       mutex_unlock(&fs_info->fs_mutex);
 fail_sb_buffer:
-       extent_io_tree_empty_lru(&BTRFS_I(fs_info->btree_inode)->io_tree);
+       btrfs_stop_workers(&fs_info->fixup_workers);
+       btrfs_stop_workers(&fs_info->workers);
+       btrfs_stop_workers(&fs_info->endio_workers);
+       btrfs_stop_workers(&fs_info->endio_write_workers);
+       btrfs_stop_workers(&fs_info->submit_workers);
 fail_iput:
        iput(fs_info->btree_inode);
 fail:
@@ -1579,9 +1674,12 @@ int close_ctree(struct btrfs_root *root)
        struct btrfs_fs_info *fs_info = root->fs_info;
 
        fs_info->closing = 1;
-       btrfs_transaction_flush_work(root);
-       mutex_lock(&fs_info->fs_mutex);
-       btrfs_defrag_dirty_roots(root->fs_info);
+       smp_mb();
+
+       kthread_stop(root->fs_info->transaction_kthread);
+       kthread_stop(root->fs_info->cleaner_kthread);
+
+       btrfs_clean_old_snapshots(root);
        trans = btrfs_start_transaction(root, 1);
        ret = btrfs_commit_transaction(trans, root);
        /* run commit again to  drop the original snapshot */
@@ -1591,9 +1689,6 @@ int close_ctree(struct btrfs_root *root)
        BUG_ON(ret);
 
        write_ctree_super(NULL, root);
-       mutex_unlock(&fs_info->fs_mutex);
-
-       btrfs_transaction_flush_work(root);
 
        if (fs_info->delalloc_bytes) {
                printk("btrfs: at unmount delalloc count %Lu\n",
@@ -1616,23 +1711,13 @@ int close_ctree(struct btrfs_root *root)
 
        filemap_write_and_wait(fs_info->btree_inode->i_mapping);
 
-       extent_io_tree_empty_lru(&fs_info->free_space_cache);
-       extent_io_tree_empty_lru(&fs_info->block_group_cache);
-       extent_io_tree_empty_lru(&fs_info->pinned_extents);
-       extent_io_tree_empty_lru(&fs_info->pending_del);
-       extent_io_tree_empty_lru(&fs_info->extent_ins);
-       extent_io_tree_empty_lru(&BTRFS_I(fs_info->btree_inode)->io_tree);
-
-       flush_workqueue(async_submit_workqueue);
-       flush_workqueue(end_io_workqueue);
-
        truncate_inode_pages(fs_info->btree_inode->i_mapping, 0);
 
-       flush_workqueue(async_submit_workqueue);
-       destroy_workqueue(async_submit_workqueue);
-
-       flush_workqueue(end_io_workqueue);
-       destroy_workqueue(end_io_workqueue);
+       btrfs_stop_workers(&fs_info->fixup_workers);
+       btrfs_stop_workers(&fs_info->workers);
+       btrfs_stop_workers(&fs_info->endio_workers);
+       btrfs_stop_workers(&fs_info->endio_write_workers);
+       btrfs_stop_workers(&fs_info->submit_workers);
 
        iput(fs_info->btree_inode);
 #if 0
@@ -1686,6 +1771,7 @@ void btrfs_mark_buffer_dirty(struct extent_buffer *buf)
        u64 transid = btrfs_header_generation(buf);
        struct inode *btree_inode = root->fs_info->btree_inode;
 
+       WARN_ON(!btrfs_tree_locked(buf));
        if (transid != root->fs_info->generation) {
                printk(KERN_CRIT "transid mismatch buffer %llu, found %Lu running %Lu\n",
                        (unsigned long long)buf->start,
@@ -1695,20 +1781,6 @@ void btrfs_mark_buffer_dirty(struct extent_buffer *buf)
        set_extent_buffer_dirty(&BTRFS_I(btree_inode)->io_tree, buf);
 }
 
-void btrfs_throttle(struct btrfs_root *root)
-{
-       struct backing_dev_info *bdi;
-
-       bdi = &root->fs_info->bdi;
-       if (root->fs_info->throttles && bdi_write_congested(bdi)) {
-#if LINUX_VERSION_CODE > KERNEL_VERSION(2,6,18)
-               congestion_wait(WRITE, HZ/20);
-#else
-               blk_congestion_wait(WRITE, HZ/20);
-#endif
-       }
-}
-
 void btrfs_btree_balance_dirty(struct btrfs_root *root, unsigned long nr)
 {
        /*
@@ -1733,58 +1805,6 @@ void btrfs_btree_balance_dirty(struct btrfs_root *root, unsigned long nr)
        return;
 }
 
-void btrfs_set_buffer_defrag(struct extent_buffer *buf)
-{
-       struct btrfs_root *root = BTRFS_I(buf->first_page->mapping->host)->root;
-       struct inode *btree_inode = root->fs_info->btree_inode;
-       set_extent_bits(&BTRFS_I(btree_inode)->io_tree, buf->start,
-                       buf->start + buf->len - 1, EXTENT_DEFRAG, GFP_NOFS);
-}
-
-void btrfs_set_buffer_defrag_done(struct extent_buffer *buf)
-{
-       struct btrfs_root *root = BTRFS_I(buf->first_page->mapping->host)->root;
-       struct inode *btree_inode = root->fs_info->btree_inode;
-       set_extent_bits(&BTRFS_I(btree_inode)->io_tree, buf->start,
-                       buf->start + buf->len - 1, EXTENT_DEFRAG_DONE,
-                       GFP_NOFS);
-}
-
-int btrfs_buffer_defrag(struct extent_buffer *buf)
-{
-       struct btrfs_root *root = BTRFS_I(buf->first_page->mapping->host)->root;
-       struct inode *btree_inode = root->fs_info->btree_inode;
-       return test_range_bit(&BTRFS_I(btree_inode)->io_tree,
-                    buf->start, buf->start + buf->len - 1, EXTENT_DEFRAG, 0);
-}
-
-int btrfs_buffer_defrag_done(struct extent_buffer *buf)
-{
-       struct btrfs_root *root = BTRFS_I(buf->first_page->mapping->host)->root;
-       struct inode *btree_inode = root->fs_info->btree_inode;
-       return test_range_bit(&BTRFS_I(btree_inode)->io_tree,
-                    buf->start, buf->start + buf->len - 1,
-                    EXTENT_DEFRAG_DONE, 0);
-}
-
-int btrfs_clear_buffer_defrag_done(struct extent_buffer *buf)
-{
-       struct btrfs_root *root = BTRFS_I(buf->first_page->mapping->host)->root;
-       struct inode *btree_inode = root->fs_info->btree_inode;
-       return clear_extent_bits(&BTRFS_I(btree_inode)->io_tree,
-                    buf->start, buf->start + buf->len - 1,
-                    EXTENT_DEFRAG_DONE, GFP_NOFS);
-}
-
-int btrfs_clear_buffer_defrag(struct extent_buffer *buf)
-{
-       struct btrfs_root *root = BTRFS_I(buf->first_page->mapping->host)->root;
-       struct inode *btree_inode = root->fs_info->btree_inode;
-       return clear_extent_bits(&BTRFS_I(btree_inode)->io_tree,
-                    buf->start, buf->start + buf->len - 1,
-                    EXTENT_DEFRAG, GFP_NOFS);
-}
-
 int btrfs_read_buffer(struct extent_buffer *buf, u64 parent_transid)
 {
        struct btrfs_root *root = BTRFS_I(buf->first_page->mapping->host)->root;