Btrfs: rename the scrub context structure
[linux-3.10.git] / fs / btrfs / ordered-data.c
index 2eb6cab..f107312 100644 (file)
@@ -16,7 +16,6 @@
  * Boston, MA 021110-1307, USA.
  */
 
-#include <linux/gfp.h>
 #include <linux/slab.h>
 #include <linux/blkdev.h>
 #include <linux/writeback.h>
@@ -26,6 +25,8 @@
 #include "btrfs_inode.h"
 #include "extent_io.h"
 
+static struct kmem_cache *btrfs_ordered_extent_cache;
+
 static u64 entry_end(struct btrfs_ordered_extent *entry)
 {
        if (entry->file_offset + entry->len < entry->file_offset)
@@ -39,11 +40,11 @@ static u64 entry_end(struct btrfs_ordered_extent *entry)
 static struct rb_node *tree_insert(struct rb_root *root, u64 file_offset,
                                   struct rb_node *node)
 {
-       struct rb_node ** p = &root->rb_node;
-       struct rb_node * parent = NULL;
+       struct rb_node **p = &root->rb_node;
+       struct rb_node *parent = NULL;
        struct btrfs_ordered_extent *entry;
 
-       while(*p) {
+       while (*p) {
                parent = *p;
                entry = rb_entry(parent, struct btrfs_ordered_extent, rb_node);
 
@@ -60,6 +61,14 @@ static struct rb_node *tree_insert(struct rb_root *root, u64 file_offset,
        return NULL;
 }
 
+static void ordered_data_tree_panic(struct inode *inode, int errno,
+                                              u64 offset)
+{
+       struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
+       btrfs_panic(fs_info, errno, "Inconsistency in ordered tree at offset "
+                   "%llu\n", (unsigned long long)offset);
+}
+
 /*
  * look for a given offset in the tree, and if it can't be found return the
  * first lesser offset
@@ -67,13 +76,13 @@ static struct rb_node *tree_insert(struct rb_root *root, u64 file_offset,
 static struct rb_node *__tree_search(struct rb_root *root, u64 file_offset,
                                     struct rb_node **prev_ret)
 {
-       struct rb_node * n = root->rb_node;
+       struct rb_node *n = root->rb_node;
        struct rb_node *prev = NULL;
        struct rb_node *test;
        struct btrfs_ordered_extent *entry;
        struct btrfs_ordered_extent *prev_entry = NULL;
 
-       while(n) {
+       while (n) {
                entry = rb_entry(n, struct btrfs_ordered_extent, rb_node);
                prev = n;
                prev_entry = entry;
@@ -88,7 +97,7 @@ static struct rb_node *__tree_search(struct rb_root *root, u64 file_offset,
        if (!prev_ret)
                return NULL;
 
-       while(prev && file_offset >= entry_end(prev_entry)) {
+       while (prev && file_offset >= entry_end(prev_entry)) {
                test = rb_next(prev);
                if (!test)
                        break;
@@ -102,7 +111,7 @@ static struct rb_node *__tree_search(struct rb_root *root, u64 file_offset,
        if (prev)
                prev_entry = rb_entry(prev, struct btrfs_ordered_extent,
                                      rb_node);
-       while(prev && file_offset < entry_end(prev_entry)) {
+       while (prev && file_offset < entry_end(prev_entry)) {
                test = rb_prev(prev);
                if (!test)
                        break;
@@ -125,6 +134,15 @@ static int offset_in_entry(struct btrfs_ordered_extent *entry, u64 file_offset)
        return 1;
 }
 
+static int range_overlaps(struct btrfs_ordered_extent *entry, u64 file_offset,
+                         u64 len)
+{
+       if (file_offset + len <= entry->file_offset ||
+           entry->file_offset + entry->len <= file_offset)
+               return 0;
+       return 1;
+}
+
 /*
  * look find the first ordered struct that has this offset, otherwise
  * the first one less than this offset
@@ -133,7 +151,7 @@ static inline struct rb_node *tree_search(struct btrfs_ordered_inode_tree *tree,
                                          u64 file_offset)
 {
        struct rb_root *root = &tree->tree;
-       struct rb_node *prev;
+       struct rb_node *prev = NULL;
        struct rb_node *ret;
        struct btrfs_ordered_extent *entry;
 
@@ -159,72 +177,171 @@ static inline struct rb_node *tree_search(struct btrfs_ordered_inode_tree *tree,
  *
  * len is the length of the extent
  *
- * This also sets the EXTENT_ORDERED bit on the range in the inode.
- *
  * The tree is given a single reference on the ordered extent that was
  * inserted.
  */
-int btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
-                            u64 start, u64 len, int nocow)
+static int __btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
+                                     u64 start, u64 len, u64 disk_len,
+                                     int type, int dio, int compress_type)
 {
        struct btrfs_ordered_inode_tree *tree;
        struct rb_node *node;
        struct btrfs_ordered_extent *entry;
 
        tree = &BTRFS_I(inode)->ordered_tree;
-       entry = kzalloc(sizeof(*entry), GFP_NOFS);
+       entry = kmem_cache_zalloc(btrfs_ordered_extent_cache, GFP_NOFS);
        if (!entry)
                return -ENOMEM;
 
-       mutex_lock(&tree->mutex);
        entry->file_offset = file_offset;
        entry->start = start;
        entry->len = len;
-       entry->inode = inode;
-       if (nocow)
-               set_bit(BTRFS_ORDERED_NOCOW, &entry->flags);
+       entry->disk_len = disk_len;
+       entry->bytes_left = len;
+       entry->inode = igrab(inode);
+       entry->compress_type = compress_type;
+       if (type != BTRFS_ORDERED_IO_DONE && type != BTRFS_ORDERED_COMPLETE)
+               set_bit(type, &entry->flags);
+
+       if (dio)
+               set_bit(BTRFS_ORDERED_DIRECT, &entry->flags);
 
        /* one ref for the tree */
        atomic_set(&entry->refs, 1);
        init_waitqueue_head(&entry->wait);
        INIT_LIST_HEAD(&entry->list);
        INIT_LIST_HEAD(&entry->root_extent_list);
+       INIT_LIST_HEAD(&entry->work_list);
+       init_completion(&entry->completion);
 
+       trace_btrfs_ordered_extent_add(inode, entry);
+
+       spin_lock_irq(&tree->lock);
        node = tree_insert(&tree->tree, file_offset,
                           &entry->rb_node);
-       if (node) {
-               printk("warning dup entry from add_ordered_extent\n");
-               BUG();
-       }
-       set_extent_ordered(&BTRFS_I(inode)->io_tree, file_offset,
-                          entry_end(entry) - 1, GFP_NOFS);
+       if (node)
+               ordered_data_tree_panic(inode, -EEXIST, file_offset);
+       spin_unlock_irq(&tree->lock);
 
        spin_lock(&BTRFS_I(inode)->root->fs_info->ordered_extent_lock);
        list_add_tail(&entry->root_extent_list,
                      &BTRFS_I(inode)->root->fs_info->ordered_extents);
        spin_unlock(&BTRFS_I(inode)->root->fs_info->ordered_extent_lock);
 
-       mutex_unlock(&tree->mutex);
-       BUG_ON(node);
        return 0;
 }
 
+int btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
+                            u64 start, u64 len, u64 disk_len, int type)
+{
+       return __btrfs_add_ordered_extent(inode, file_offset, start, len,
+                                         disk_len, type, 0,
+                                         BTRFS_COMPRESS_NONE);
+}
+
+int btrfs_add_ordered_extent_dio(struct inode *inode, u64 file_offset,
+                                u64 start, u64 len, u64 disk_len, int type)
+{
+       return __btrfs_add_ordered_extent(inode, file_offset, start, len,
+                                         disk_len, type, 1,
+                                         BTRFS_COMPRESS_NONE);
+}
+
+int btrfs_add_ordered_extent_compress(struct inode *inode, u64 file_offset,
+                                     u64 start, u64 len, u64 disk_len,
+                                     int type, int compress_type)
+{
+       return __btrfs_add_ordered_extent(inode, file_offset, start, len,
+                                         disk_len, type, 0,
+                                         compress_type);
+}
+
 /*
  * Add a struct btrfs_ordered_sum into the list of checksums to be inserted
  * when an ordered extent is finished.  If the list covers more than one
  * ordered extent, it is split across multiples.
  */
-int btrfs_add_ordered_sum(struct inode *inode,
-                         struct btrfs_ordered_extent *entry,
-                         struct btrfs_ordered_sum *sum)
+void btrfs_add_ordered_sum(struct inode *inode,
+                          struct btrfs_ordered_extent *entry,
+                          struct btrfs_ordered_sum *sum)
 {
        struct btrfs_ordered_inode_tree *tree;
 
        tree = &BTRFS_I(inode)->ordered_tree;
-       mutex_lock(&tree->mutex);
+       spin_lock_irq(&tree->lock);
        list_add_tail(&sum->list, &entry->list);
-       mutex_unlock(&tree->mutex);
-       return 0;
+       spin_unlock_irq(&tree->lock);
+}
+
+/*
+ * this is used to account for finished IO across a given range
+ * of the file.  The IO may span ordered extents.  If
+ * a given ordered_extent is completely done, 1 is returned, otherwise
+ * 0.
+ *
+ * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
+ * to make sure this function only returns 1 once for a given ordered extent.
+ *
+ * file_offset is updated to one byte past the range that is recorded as
+ * complete.  This allows you to walk forward in the file.
+ */
+int btrfs_dec_test_first_ordered_pending(struct inode *inode,
+                                  struct btrfs_ordered_extent **cached,
+                                  u64 *file_offset, u64 io_size, int uptodate)
+{
+       struct btrfs_ordered_inode_tree *tree;
+       struct rb_node *node;
+       struct btrfs_ordered_extent *entry = NULL;
+       int ret;
+       unsigned long flags;
+       u64 dec_end;
+       u64 dec_start;
+       u64 to_dec;
+
+       tree = &BTRFS_I(inode)->ordered_tree;
+       spin_lock_irqsave(&tree->lock, flags);
+       node = tree_search(tree, *file_offset);
+       if (!node) {
+               ret = 1;
+               goto out;
+       }
+
+       entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
+       if (!offset_in_entry(entry, *file_offset)) {
+               ret = 1;
+               goto out;
+       }
+
+       dec_start = max(*file_offset, entry->file_offset);
+       dec_end = min(*file_offset + io_size, entry->file_offset +
+                     entry->len);
+       *file_offset = dec_end;
+       if (dec_start > dec_end) {
+               printk(KERN_CRIT "bad ordering dec_start %llu end %llu\n",
+                      (unsigned long long)dec_start,
+                      (unsigned long long)dec_end);
+       }
+       to_dec = dec_end - dec_start;
+       if (to_dec > entry->bytes_left) {
+               printk(KERN_CRIT "bad ordered accounting left %llu size %llu\n",
+                      (unsigned long long)entry->bytes_left,
+                      (unsigned long long)to_dec);
+       }
+       entry->bytes_left -= to_dec;
+       if (!uptodate)
+               set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
+
+       if (entry->bytes_left == 0)
+               ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
+       else
+               ret = 1;
+out:
+       if (!ret && cached && entry) {
+               *cached = entry;
+               atomic_inc(&entry->refs);
+       }
+       spin_unlock_irqrestore(&tree->lock, flags);
+       return ret == 0;
 }
 
 /*
@@ -237,18 +354,22 @@ int btrfs_add_ordered_sum(struct inode *inode,
  * to make sure this function only returns 1 once for a given ordered extent.
  */
 int btrfs_dec_test_ordered_pending(struct inode *inode,
-                                  u64 file_offset, u64 io_size)
+                                  struct btrfs_ordered_extent **cached,
+                                  u64 file_offset, u64 io_size, int uptodate)
 {
        struct btrfs_ordered_inode_tree *tree;
        struct rb_node *node;
-       struct btrfs_ordered_extent *entry;
-       struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
+       struct btrfs_ordered_extent *entry = NULL;
+       unsigned long flags;
        int ret;
 
        tree = &BTRFS_I(inode)->ordered_tree;
-       mutex_lock(&tree->mutex);
-       clear_extent_ordered(io_tree, file_offset, file_offset + io_size - 1,
-                            GFP_NOFS);
+       spin_lock_irqsave(&tree->lock, flags);
+       if (cached && *cached) {
+               entry = *cached;
+               goto have_entry;
+       }
+
        node = tree_search(tree, file_offset);
        if (!node) {
                ret = 1;
@@ -256,18 +377,31 @@ int btrfs_dec_test_ordered_pending(struct inode *inode,
        }
 
        entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
+have_entry:
        if (!offset_in_entry(entry, file_offset)) {
                ret = 1;
                goto out;
        }
 
-       ret = test_range_bit(io_tree, entry->file_offset,
-                            entry->file_offset + entry->len - 1,
-                            EXTENT_ORDERED, 0);
-       if (ret == 0)
+       if (io_size > entry->bytes_left) {
+               printk(KERN_CRIT "bad ordered accounting left %llu size %llu\n",
+                      (unsigned long long)entry->bytes_left,
+                      (unsigned long long)io_size);
+       }
+       entry->bytes_left -= io_size;
+       if (!uptodate)
+               set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
+
+       if (entry->bytes_left == 0)
                ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
+       else
+               ret = 1;
 out:
-       mutex_unlock(&tree->mutex);
+       if (!ret && cached && entry) {
+               *cached = entry;
+               atomic_inc(&entry->refs);
+       }
+       spin_unlock_irqrestore(&tree->lock, flags);
        return ret == 0;
 }
 
@@ -275,61 +409,85 @@ out:
  * used to drop a reference on an ordered extent.  This will free
  * the extent if the last reference is dropped
  */
-int btrfs_put_ordered_extent(struct btrfs_ordered_extent *entry)
+void btrfs_put_ordered_extent(struct btrfs_ordered_extent *entry)
 {
        struct list_head *cur;
        struct btrfs_ordered_sum *sum;
 
+       trace_btrfs_ordered_extent_put(entry->inode, entry);
+
        if (atomic_dec_and_test(&entry->refs)) {
-               while(!list_empty(&entry->list)) {
+               if (entry->inode)
+                       btrfs_add_delayed_iput(entry->inode);
+               while (!list_empty(&entry->list)) {
                        cur = entry->list.next;
                        sum = list_entry(cur, struct btrfs_ordered_sum, list);
                        list_del(&sum->list);
                        kfree(sum);
                }
-               kfree(entry);
+               kmem_cache_free(btrfs_ordered_extent_cache, entry);
        }
-       return 0;
 }
 
 /*
  * remove an ordered extent from the tree.  No references are dropped
- * but, anyone waiting on this extent is woken up.
+ * and waiters are woken up.
  */
-int btrfs_remove_ordered_extent(struct inode *inode,
-                               struct btrfs_ordered_extent *entry)
+void btrfs_remove_ordered_extent(struct inode *inode,
+                                struct btrfs_ordered_extent *entry)
 {
        struct btrfs_ordered_inode_tree *tree;
+       struct btrfs_root *root = BTRFS_I(inode)->root;
        struct rb_node *node;
 
        tree = &BTRFS_I(inode)->ordered_tree;
-       mutex_lock(&tree->mutex);
+       spin_lock_irq(&tree->lock);
        node = &entry->rb_node;
        rb_erase(node, &tree->tree);
        tree->last = NULL;
        set_bit(BTRFS_ORDERED_COMPLETE, &entry->flags);
+       spin_unlock_irq(&tree->lock);
 
-       spin_lock(&BTRFS_I(inode)->root->fs_info->ordered_extent_lock);
+       spin_lock(&root->fs_info->ordered_extent_lock);
        list_del_init(&entry->root_extent_list);
-       spin_unlock(&BTRFS_I(inode)->root->fs_info->ordered_extent_lock);
 
-       mutex_unlock(&tree->mutex);
+       trace_btrfs_ordered_extent_remove(inode, entry);
+
+       /*
+        * we have no more ordered extents for this inode and
+        * no dirty pages.  We can safely remove it from the
+        * list of ordered extents
+        */
+       if (RB_EMPTY_ROOT(&tree->tree) &&
+           !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
+               list_del_init(&BTRFS_I(inode)->ordered_operations);
+       }
+       spin_unlock(&root->fs_info->ordered_extent_lock);
        wake_up(&entry->wait);
-       return 0;
+}
+
+static void btrfs_run_ordered_extent_work(struct btrfs_work *work)
+{
+       struct btrfs_ordered_extent *ordered;
+
+       ordered = container_of(work, struct btrfs_ordered_extent, flush_work);
+       btrfs_start_ordered_extent(ordered->inode, ordered, 1);
+       complete(&ordered->completion);
 }
 
 /*
  * wait for all the ordered extents in a root.  This is done when balancing
  * space between drives.
  */
-int btrfs_wait_ordered_extents(struct btrfs_root *root, int nocow_only)
+void btrfs_wait_ordered_extents(struct btrfs_root *root, int delay_iput)
 {
-       struct list_head splice;
+       struct list_head splice, works;
        struct list_head *cur;
-       struct btrfs_ordered_extent *ordered;
+       struct btrfs_ordered_extent *ordered, *next;
        struct inode *inode;
 
        INIT_LIST_HEAD(&splice);
+       INIT_LIST_HEAD(&works);
 
        spin_lock(&root->fs_info->ordered_extent_lock);
        list_splice_init(&root->fs_info->ordered_extents, &splice);
@@ -337,14 +495,6 @@ int btrfs_wait_ordered_extents(struct btrfs_root *root, int nocow_only)
                cur = splice.next;
                ordered = list_entry(cur, struct btrfs_ordered_extent,
                                     root_extent_list);
-               if (nocow_only &&
-                   !test_bit(BTRFS_ORDERED_NOCOW, &ordered->flags)) {
-                       list_move(&ordered->root_extent_list,
-                                 &root->fs_info->ordered_extents);
-                       cond_resched_lock(&root->fs_info->ordered_extent_lock);
-                       continue;
-               }
-
                list_del_init(&ordered->root_extent_list);
                atomic_inc(&ordered->refs);
 
@@ -356,17 +506,114 @@ int btrfs_wait_ordered_extents(struct btrfs_root *root, int nocow_only)
                spin_unlock(&root->fs_info->ordered_extent_lock);
 
                if (inode) {
-                       btrfs_start_ordered_extent(inode, ordered, 1);
-                       btrfs_put_ordered_extent(ordered);
-                       iput(inode);
+                       ordered->flush_work.func = btrfs_run_ordered_extent_work;
+                       list_add_tail(&ordered->work_list, &works);
+                       btrfs_queue_worker(&root->fs_info->flush_workers,
+                                          &ordered->flush_work);
                } else {
                        btrfs_put_ordered_extent(ordered);
                }
 
+               cond_resched();
                spin_lock(&root->fs_info->ordered_extent_lock);
        }
        spin_unlock(&root->fs_info->ordered_extent_lock);
-       return 0;
+
+       list_for_each_entry_safe(ordered, next, &works, work_list) {
+               list_del_init(&ordered->work_list);
+               wait_for_completion(&ordered->completion);
+
+               inode = ordered->inode;
+               btrfs_put_ordered_extent(ordered);
+               if (delay_iput)
+                       btrfs_add_delayed_iput(inode);
+               else
+                       iput(inode);
+
+               cond_resched();
+       }
+}
+
+/*
+ * this is used during transaction commit to write all the inodes
+ * added to the ordered operation list.  These files must be fully on
+ * disk before the transaction commits.
+ *
+ * we have two modes here, one is to just start the IO via filemap_flush
+ * and the other is to wait for all the io.  When we wait, we have an
+ * extra check to make sure the ordered operation list really is empty
+ * before we return
+ */
+int btrfs_run_ordered_operations(struct btrfs_root *root, int wait)
+{
+       struct btrfs_inode *btrfs_inode;
+       struct inode *inode;
+       struct list_head splice;
+       struct list_head works;
+       struct btrfs_delalloc_work *work, *next;
+       int ret = 0;
+
+       INIT_LIST_HEAD(&splice);
+       INIT_LIST_HEAD(&works);
+
+       mutex_lock(&root->fs_info->ordered_operations_mutex);
+       spin_lock(&root->fs_info->ordered_extent_lock);
+again:
+       list_splice_init(&root->fs_info->ordered_operations, &splice);
+
+       while (!list_empty(&splice)) {
+
+               btrfs_inode = list_entry(splice.next, struct btrfs_inode,
+                                  ordered_operations);
+
+               inode = &btrfs_inode->vfs_inode;
+
+               list_del_init(&btrfs_inode->ordered_operations);
+
+               /*
+                * the inode may be getting freed (in sys_unlink path).
+                */
+               inode = igrab(inode);
+
+               if (!wait && inode) {
+                       list_add_tail(&BTRFS_I(inode)->ordered_operations,
+                             &root->fs_info->ordered_operations);
+               }
+
+               if (!inode)
+                       continue;
+               spin_unlock(&root->fs_info->ordered_extent_lock);
+
+               work = btrfs_alloc_delalloc_work(inode, wait, 1);
+               if (!work) {
+                       if (list_empty(&BTRFS_I(inode)->ordered_operations))
+                               list_add_tail(&btrfs_inode->ordered_operations,
+                                             &splice);
+                       spin_lock(&root->fs_info->ordered_extent_lock);
+                       list_splice_tail(&splice,
+                                        &root->fs_info->ordered_operations);
+                       spin_unlock(&root->fs_info->ordered_extent_lock);
+                       ret = -ENOMEM;
+                       goto out;
+               }
+               list_add_tail(&work->list, &works);
+               btrfs_queue_worker(&root->fs_info->flush_workers,
+                                  &work->work);
+
+               cond_resched();
+               spin_lock(&root->fs_info->ordered_extent_lock);
+       }
+       if (wait && !list_empty(&root->fs_info->ordered_operations))
+               goto again;
+
+       spin_unlock(&root->fs_info->ordered_extent_lock);
+out:
+       list_for_each_entry_safe(work, next, &works, list) {
+               list_del_init(&work->list);
+               btrfs_wait_and_free_delalloc_work(work);
+       }
+       mutex_unlock(&root->fs_info->ordered_operations_mutex);
+       return ret;
 }
 
 /*
@@ -383,25 +630,28 @@ void btrfs_start_ordered_extent(struct inode *inode,
        u64 start = entry->file_offset;
        u64 end = start + entry->len - 1;
 
+       trace_btrfs_ordered_extent_start(inode, entry);
+
        /*
         * pages in the range can be dirty, clean or writeback.  We
         * start IO on any dirty ones so the wait doesn't stall waiting
-        * for pdflush to find them
+        * for the flusher thread to find them
         */
-       btrfs_fdatawrite_range(inode->i_mapping, start, end, WB_SYNC_NONE);
-       if (wait)
+       if (!test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
+               filemap_fdatawrite_range(inode->i_mapping, start, end);
+       if (wait) {
                wait_event(entry->wait, test_bit(BTRFS_ORDERED_COMPLETE,
                                                 &entry->flags));
+       }
 }
 
 /*
  * Used to wait on ordered extents across a large range of bytes.
  */
-int btrfs_wait_ordered_range(struct inode *inode, u64 start, u64 len)
+void btrfs_wait_ordered_range(struct inode *inode, u64 start, u64 len)
 {
        u64 end;
        u64 orig_end;
-       u64 wait_end;
        struct btrfs_ordered_extent *ordered;
 
        if (start + len < start) {
@@ -411,23 +661,37 @@ int btrfs_wait_ordered_range(struct inode *inode, u64 start, u64 len)
                if (orig_end > INT_LIMIT(loff_t))
                        orig_end = INT_LIMIT(loff_t);
        }
-       wait_end = orig_end;
-again:
+
        /* start IO across the range first to instantiate any delalloc
         * extents
         */
-       btrfs_fdatawrite_range(inode->i_mapping, start, orig_end, WB_SYNC_NONE);
+       filemap_fdatawrite_range(inode->i_mapping, start, orig_end);
 
-       btrfs_wait_on_page_writeback_range(inode->i_mapping,
-                                          start >> PAGE_CACHE_SHIFT,
-                                          orig_end >> PAGE_CACHE_SHIFT);
+       /*
+        * So with compression we will find and lock a dirty page and clear the
+        * first one as dirty, setup an async extent, and immediately return
+        * with the entire range locked but with nobody actually marked with
+        * writeback.  So we can't just filemap_write_and_wait_range() and
+        * expect it to work since it will just kick off a thread to do the
+        * actual work.  So we need to call filemap_fdatawrite_range _again_
+        * since it will wait on the page lock, which won't be unlocked until
+        * after the pages have been marked as writeback and so we're good to go
+        * from there.  We have to do this otherwise we'll miss the ordered
+        * extents and that results in badness.  Please Josef, do not think you
+        * know better and pull this out at some point in the future, it is
+        * right and you are wrong.
+        */
+       if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
+                    &BTRFS_I(inode)->runtime_flags))
+               filemap_fdatawrite_range(inode->i_mapping, start, orig_end);
+
+       filemap_fdatawait_range(inode->i_mapping, start, orig_end);
 
        end = orig_end;
-       while(1) {
+       while (1) {
                ordered = btrfs_lookup_first_ordered_extent(inode, end);
-               if (!ordered) {
+               if (!ordered)
                        break;
-               }
                if (ordered->file_offset > orig_end) {
                        btrfs_put_ordered_extent(ordered);
                        break;
@@ -443,15 +707,6 @@ again:
                        break;
                end--;
        }
-       if (test_range_bit(&BTRFS_I(inode)->io_tree, start, orig_end,
-                          EXTENT_ORDERED | EXTENT_DELALLOC, 0)) {
-               printk("inode %lu still ordered or delalloc after wait "
-                      "%llu %llu\n", inode->i_ino,
-                      (unsigned long long)start,
-                      (unsigned long long)orig_end);
-               goto again;
-       }
-       return 0;
 }
 
 /*
@@ -466,7 +721,7 @@ struct btrfs_ordered_extent *btrfs_lookup_ordered_extent(struct inode *inode,
        struct btrfs_ordered_extent *entry = NULL;
 
        tree = &BTRFS_I(inode)->ordered_tree;
-       mutex_lock(&tree->mutex);
+       spin_lock_irq(&tree->lock);
        node = tree_search(tree, file_offset);
        if (!node)
                goto out;
@@ -477,7 +732,48 @@ struct btrfs_ordered_extent *btrfs_lookup_ordered_extent(struct inode *inode,
        if (entry)
                atomic_inc(&entry->refs);
 out:
-       mutex_unlock(&tree->mutex);
+       spin_unlock_irq(&tree->lock);
+       return entry;
+}
+
+/* Since the DIO code tries to lock a wide area we need to look for any ordered
+ * extents that exist in the range, rather than just the start of the range.
+ */
+struct btrfs_ordered_extent *btrfs_lookup_ordered_range(struct inode *inode,
+                                                       u64 file_offset,
+                                                       u64 len)
+{
+       struct btrfs_ordered_inode_tree *tree;
+       struct rb_node *node;
+       struct btrfs_ordered_extent *entry = NULL;
+
+       tree = &BTRFS_I(inode)->ordered_tree;
+       spin_lock_irq(&tree->lock);
+       node = tree_search(tree, file_offset);
+       if (!node) {
+               node = tree_search(tree, file_offset + len);
+               if (!node)
+                       goto out;
+       }
+
+       while (1) {
+               entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
+               if (range_overlaps(entry, file_offset, len))
+                       break;
+
+               if (entry->file_offset >= file_offset + len) {
+                       entry = NULL;
+                       break;
+               }
+               entry = NULL;
+               node = rb_next(node);
+               if (!node)
+                       break;
+       }
+out:
+       if (entry)
+               atomic_inc(&entry->refs);
+       spin_unlock_irq(&tree->lock);
        return entry;
 }
 
@@ -486,14 +782,14 @@ out:
  * if none is found
  */
 struct btrfs_ordered_extent *
-btrfs_lookup_first_ordered_extent(struct inode * inode, u64 file_offset)
+btrfs_lookup_first_ordered_extent(struct inode *inode, u64 file_offset)
 {
        struct btrfs_ordered_inode_tree *tree;
        struct rb_node *node;
        struct btrfs_ordered_extent *entry = NULL;
 
        tree = &BTRFS_I(inode)->ordered_tree;
-       mutex_lock(&tree->mutex);
+       spin_lock_irq(&tree->lock);
        node = tree_search(tree, file_offset);
        if (!node)
                goto out;
@@ -501,7 +797,7 @@ btrfs_lookup_first_ordered_extent(struct inode * inode, u64 file_offset)
        entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
        atomic_inc(&entry->refs);
 out:
-       mutex_unlock(&tree->mutex);
+       spin_unlock_irq(&tree->lock);
        return entry;
 }
 
@@ -509,94 +805,109 @@ out:
  * After an extent is done, call this to conditionally update the on disk
  * i_size.  i_size is updated to cover any fully written part of the file.
  */
-int btrfs_ordered_update_i_size(struct inode *inode,
+int btrfs_ordered_update_i_size(struct inode *inode, u64 offset,
                                struct btrfs_ordered_extent *ordered)
 {
        struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
-       struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
        u64 disk_i_size;
        u64 new_i_size;
-       u64 i_size_test;
+       u64 i_size = i_size_read(inode);
        struct rb_node *node;
+       struct rb_node *prev = NULL;
        struct btrfs_ordered_extent *test;
+       int ret = 1;
 
-       mutex_lock(&tree->mutex);
+       if (ordered)
+               offset = entry_end(ordered);
+       else
+               offset = ALIGN(offset, BTRFS_I(inode)->root->sectorsize);
+
+       spin_lock_irq(&tree->lock);
        disk_i_size = BTRFS_I(inode)->disk_i_size;
 
-       /*
-        * if the disk i_size is already at the inode->i_size, or
-        * this ordered extent is inside the disk i_size, we're done
-        */
-       if (disk_i_size >= inode->i_size ||
-           ordered->file_offset + ordered->len <= disk_i_size) {
+       /* truncate file */
+       if (disk_i_size > i_size) {
+               BTRFS_I(inode)->disk_i_size = i_size;
+               ret = 0;
                goto out;
        }
 
        /*
-        * we can't update the disk_isize if there are delalloc bytes
-        * between disk_i_size and  this ordered extent
+        * if the disk i_size is already at the inode->i_size, or
+        * this ordered extent is inside the disk i_size, we're done
         */
-       if (test_range_bit(io_tree, disk_i_size,
-                          ordered->file_offset + ordered->len - 1,
-                          EXTENT_DELALLOC, 0)) {
+       if (disk_i_size == i_size || offset <= disk_i_size) {
                goto out;
        }
+
        /*
         * walk backward from this ordered extent to disk_i_size.
         * if we find an ordered extent then we can't update disk i_size
         * yet
         */
-       node = &ordered->rb_node;
-       while(1) {
-               node = rb_prev(node);
-               if (!node)
-                       break;
+       if (ordered) {
+               node = rb_prev(&ordered->rb_node);
+       } else {
+               prev = tree_search(tree, offset);
+               /*
+                * we insert file extents without involving ordered struct,
+                * so there should be no ordered struct cover this offset
+                */
+               if (prev) {
+                       test = rb_entry(prev, struct btrfs_ordered_extent,
+                                       rb_node);
+                       BUG_ON(offset_in_entry(test, offset));
+               }
+               node = prev;
+       }
+       for (; node; node = rb_prev(node)) {
                test = rb_entry(node, struct btrfs_ordered_extent, rb_node);
+
+               /* We treat this entry as if it doesnt exist */
+               if (test_bit(BTRFS_ORDERED_UPDATED_ISIZE, &test->flags))
+                       continue;
                if (test->file_offset + test->len <= disk_i_size)
                        break;
-               if (test->file_offset >= inode->i_size)
+               if (test->file_offset >= i_size)
                        break;
-               if (test->file_offset >= disk_i_size)
+               if (test->file_offset >= disk_i_size) {
+                       /*
+                        * we don't update disk_i_size now, so record this
+                        * undealt i_size. Or we will not know the real
+                        * i_size.
+                        */
+                       if (test->outstanding_isize < offset)
+                               test->outstanding_isize = offset;
+                       if (ordered &&
+                           ordered->outstanding_isize >
+                           test->outstanding_isize)
+                               test->outstanding_isize =
+                                               ordered->outstanding_isize;
                        goto out;
-       }
-       new_i_size = min_t(u64, entry_end(ordered), i_size_read(inode));
-
-       /*
-        * at this point, we know we can safely update i_size to at least
-        * the offset from this ordered extent.  But, we need to
-        * walk forward and see if ios from higher up in the file have
-        * finished.
-        */
-       node = rb_next(&ordered->rb_node);
-       i_size_test = 0;
-       if (node) {
-               /*
-                * do we have an area where IO might have finished
-                * between our ordered extent and the next one.
-                */
-               test = rb_entry(node, struct btrfs_ordered_extent, rb_node);
-               if (test->file_offset > entry_end(ordered)) {
-                       i_size_test = test->file_offset;
                }
-       } else {
-               i_size_test = i_size_read(inode);
        }
+       new_i_size = min_t(u64, offset, i_size);
 
        /*
-        * i_size_test is the end of a region after this ordered
-        * extent where there are no ordered extents.  As long as there
-        * are no delalloc bytes in this area, it is safe to update
-        * disk_i_size to the end of the region.
+        * Some ordered extents may completed before the current one, and
+        * we hold the real i_size in ->outstanding_isize.
         */
-       if (i_size_test > entry_end(ordered) &&
-           !test_range_bit(io_tree, entry_end(ordered), i_size_test - 1,
-                          EXTENT_DELALLOC, 0)) {
-               new_i_size = min_t(u64, i_size_test, i_size_read(inode));
-       }
+       if (ordered && ordered->outstanding_isize > new_i_size)
+               new_i_size = min_t(u64, ordered->outstanding_isize, i_size);
        BTRFS_I(inode)->disk_i_size = new_i_size;
+       ret = 0;
 out:
-       mutex_unlock(&tree->mutex);
-       return 0;
+       /*
+        * We need to do this because we can't remove ordered extents until
+        * after the i_disk_size has been updated and then the inode has been
+        * updated to reflect the change, so we need to tell anybody who finds
+        * this ordered extent that we've already done all the real work, we
+        * just haven't completed all the other work.
+        */
+       if (ordered)
+               set_bit(BTRFS_ORDERED_UPDATED_ISIZE, &ordered->flags);
+       spin_unlock_irq(&tree->lock);
+       return ret;
 }
 
 /*
@@ -604,13 +915,13 @@ out:
  * try to find a checksum.  This is used because we allow pages to
  * be reclaimed before their checksum is actually put into the btree
  */
-int btrfs_find_ordered_sum(struct inode *inode, u64 offset, u32 *sum)
+int btrfs_find_ordered_sum(struct inode *inode, u64 offset, u64 disk_bytenr,
+                          u32 *sum)
 {
        struct btrfs_ordered_sum *ordered_sum;
        struct btrfs_sector_sum *sector_sums;
        struct btrfs_ordered_extent *ordered;
        struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
-       struct list_head *cur;
        unsigned long num_sectors;
        unsigned long i;
        u32 sectorsize = BTRFS_I(inode)->root->sectorsize;
@@ -620,14 +931,13 @@ int btrfs_find_ordered_sum(struct inode *inode, u64 offset, u32 *sum)
        if (!ordered)
                return 1;
 
-       mutex_lock(&tree->mutex);
-       list_for_each_prev(cur, &ordered->list) {
-               ordered_sum = list_entry(cur, struct btrfs_ordered_sum, list);
-               if (offset >= ordered_sum->file_offset) {
+       spin_lock_irq(&tree->lock);
+       list_for_each_entry_reverse(ordered_sum, &ordered->list, list) {
+               if (disk_bytenr >= ordered_sum->bytenr) {
                        num_sectors = ordered_sum->len / sectorsize;
                        sector_sums = ordered_sum->sums;
                        for (i = 0; i < num_sectors; i++) {
-                               if (sector_sums[i].offset == offset) {
+                               if (sector_sums[i].bytenr == disk_bytenr) {
                                        *sum = sector_sums[i].sum;
                                        ret = 0;
                                        goto out;
@@ -636,92 +946,60 @@ int btrfs_find_ordered_sum(struct inode *inode, u64 offset, u32 *sum)
                }
        }
 out:
-       mutex_unlock(&tree->mutex);
+       spin_unlock_irq(&tree->lock);
        btrfs_put_ordered_extent(ordered);
        return ret;
 }
 
 
-/**
- * taken from mm/filemap.c because it isn't exported
+/*
+ * add a given inode to the list of inodes that must be fully on
+ * disk before a transaction commit finishes.
  *
- * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
- * @mapping:   address space structure to write
- * @start:     offset in bytes where the range starts
- * @end:       offset in bytes where the range ends (inclusive)
- * @sync_mode: enable synchronous operation
+ * This basically gives us the ext3 style data=ordered mode, and it is mostly
+ * used to make sure renamed files are fully on disk.
  *
- * Start writeback against all of a mapping's dirty pages that lie
- * within the byte offsets <start, end> inclusive.
+ * It is a noop if the inode is already fully on disk.
  *
- * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
- * opposed to a regular memory cleansing writeback.  The difference between
- * these two operations is that if a dirty page/buffer is encountered, it must
- * be waited upon, and not just skipped over.
+ * If trans is not null, we'll do a friendly check for a transaction that
+ * is already flushing things and force the IO down ourselves.
  */
-int btrfs_fdatawrite_range(struct address_space *mapping, loff_t start,
-                          loff_t end, int sync_mode)
+void btrfs_add_ordered_operation(struct btrfs_trans_handle *trans,
+                                struct btrfs_root *root, struct inode *inode)
 {
-       struct writeback_control wbc = {
-               .sync_mode = sync_mode,
-               .nr_to_write = mapping->nrpages * 2,
-               .range_start = start,
-               .range_end = end,
-               .for_writepages = 1,
-       };
-       return btrfs_writepages(mapping, &wbc);
-}
-
-/**
- * taken from mm/filemap.c because it isn't exported
- *
- * wait_on_page_writeback_range - wait for writeback to complete
- * @mapping:   target address_space
- * @start:     beginning page index
- * @end:       ending page index
- *
- * Wait for writeback to complete against pages indexed by start->end
- * inclusive
- */
-int btrfs_wait_on_page_writeback_range(struct address_space *mapping,
-                                      pgoff_t start, pgoff_t end)
-{
-       struct pagevec pvec;
-       int nr_pages;
-       int ret = 0;
-       pgoff_t index;
-
-       if (end < start)
-               return 0;
+       u64 last_mod;
 
-       pagevec_init(&pvec, 0);
-       index = start;
-       while ((index <= end) &&
-                       (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
-                       PAGECACHE_TAG_WRITEBACK,
-                       min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
-               unsigned i;
+       last_mod = max(BTRFS_I(inode)->generation, BTRFS_I(inode)->last_trans);
 
-               for (i = 0; i < nr_pages; i++) {
-                       struct page *page = pvec.pages[i];
-
-                       /* until radix tree lookup accepts end_index */
-                       if (page->index > end)
-                               continue;
+       /*
+        * if this file hasn't been changed since the last transaction
+        * commit, we can safely return without doing anything
+        */
+       if (last_mod < root->fs_info->last_trans_committed)
+               return;
 
-                       wait_on_page_writeback(page);
-                       if (PageError(page))
-                               ret = -EIO;
-               }
-               pagevec_release(&pvec);
-               cond_resched();
+       spin_lock(&root->fs_info->ordered_extent_lock);
+       if (list_empty(&BTRFS_I(inode)->ordered_operations)) {
+               list_add_tail(&BTRFS_I(inode)->ordered_operations,
+                             &root->fs_info->ordered_operations);
        }
+       spin_unlock(&root->fs_info->ordered_extent_lock);
+}
 
-       /* Check for outstanding write errors */
-       if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
-               ret = -ENOSPC;
-       if (test_and_clear_bit(AS_EIO, &mapping->flags))
-               ret = -EIO;
+int __init ordered_data_init(void)
+{
+       btrfs_ordered_extent_cache = kmem_cache_create("btrfs_ordered_extent",
+                                    sizeof(struct btrfs_ordered_extent), 0,
+                                    SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD,
+                                    NULL);
+       if (!btrfs_ordered_extent_cache)
+               return -ENOMEM;
 
-       return ret;
+       return 0;
+}
+
+void ordered_data_exit(void)
+{
+       if (btrfs_ordered_extent_cache)
+               kmem_cache_destroy(btrfs_ordered_extent_cache);
 }