[2/5] btrfs: introduce the inode->file_extent_tree
diff mbox series

Message ID 20191230213118.7532-3-josef@toxicpanda.com
State New
Headers show
Series
  • btrfs: fix hole corruption issue with !NO_HOLES
Related show

Commit Message

Josef Bacik Dec. 30, 2019, 9:31 p.m. UTC
In order to keep track of where we have file extents on disk, and thus
where it is safe to adjust the i_size to, we need to have a tree in
place to keep track of the contiguous areas we have file extents for.
Add helpers to use this tree, as it's not required for NO_HOLES file
systems.  We will use this by setting DIRTY for areas we know we have
file extent item's set, and clearing it when we remove file extent items
for truncation.

Signed-off-by: Josef Bacik <josef@toxicpanda.com>
---
 fs/btrfs/btrfs_inode.h    |  5 +++
 fs/btrfs/ctree.h          |  5 +++
 fs/btrfs/extent-io-tree.h |  1 +
 fs/btrfs/extent_io.c      | 11 +++++
 fs/btrfs/file-item.c      | 91 +++++++++++++++++++++++++++++++++++++++
 fs/btrfs/inode.c          |  6 +++
 6 files changed, 119 insertions(+)

Comments

David Sterba Jan. 6, 2020, 5:22 p.m. UTC | #1
On Mon, Dec 30, 2019 at 04:31:15PM -0500, Josef Bacik wrote:
> @@ -60,6 +60,11 @@ struct btrfs_inode {
>  	 */
>  	struct extent_io_tree io_failure_tree;
>  
> +	/* keeps track of where we have extent items mapped in order to make
> +	 * sure our i_size adjustments are accurate.
> +	 */
> +	struct extent_io_tree file_extent_tree;

This is not exactly lightweight and cut to the minimum needed, the size
is 40 bytes and contains struct members that are unused. At least the
file extents tree seems to be in use unlike that io_failure_tree wasting
the bytes almost 100% of time.
Josef Bacik Jan. 6, 2020, 7:29 p.m. UTC | #2
On 1/6/20 12:22 PM, David Sterba wrote:
> On Mon, Dec 30, 2019 at 04:31:15PM -0500, Josef Bacik wrote:
>> @@ -60,6 +60,11 @@ struct btrfs_inode {
>>   	 */
>>   	struct extent_io_tree io_failure_tree;
>>   
>> +	/* keeps track of where we have extent items mapped in order to make
>> +	 * sure our i_size adjustments are accurate.
>> +	 */
>> +	struct extent_io_tree file_extent_tree;
> 
> This is not exactly lightweight and cut to the minimum needed, the size
> is 40 bytes and contains struct members that are unused. At least the
> file extents tree seems to be in use unlike that io_failure_tree wasting
> the bytes almost 100% of time.
>

I'm not in love with it either, but I don't want to invent some lighter weight 
range thingy that I'm going to end up messing up in other ways.

Don't take this series yet, there's still something fishy going on that I have 
to figure out.  Thanks,

Josef
David Sterba Jan. 7, 2020, 4:17 p.m. UTC | #3
On Mon, Jan 06, 2020 at 02:29:01PM -0500, Josef Bacik wrote:
> On 1/6/20 12:22 PM, David Sterba wrote:
> > On Mon, Dec 30, 2019 at 04:31:15PM -0500, Josef Bacik wrote:
> >> @@ -60,6 +60,11 @@ struct btrfs_inode {
> >>   	 */
> >>   	struct extent_io_tree io_failure_tree;
> >>   
> >> +	/* keeps track of where we have extent items mapped in order to make
> >> +	 * sure our i_size adjustments are accurate.
> >> +	 */
> >> +	struct extent_io_tree file_extent_tree;
> > 
> > This is not exactly lightweight and cut to the minimum needed, the size
> > is 40 bytes and contains struct members that are unused. At least the
> > file extents tree seems to be in use unlike that io_failure_tree wasting
> > the bytes almost 100% of time.
> >
> 
> I'm not in love with it either, but I don't want to invent some lighter weight 
> range thingy that I'm going to end up messing up in other ways.

Yeah, the extent_io_tree is now being used for generic range tree, some
cleanups could remove the members that were added for the first specific
use (like the dirty_bytes, or track_updates). For correctness of NOHOLES
the inode size increase shouldn't be a blocker but needs to be addressed
later.

> Don't take this series yet, there's still something fishy going on that I have 
> to figure out.  Thanks,

Ok.
Filipe Manana Jan. 7, 2020, 4:45 p.m. UTC | #4
On Tue, Jan 7, 2020 at 4:18 PM David Sterba <dsterba@suse.cz> wrote:
>
> On Mon, Jan 06, 2020 at 02:29:01PM -0500, Josef Bacik wrote:
> > On 1/6/20 12:22 PM, David Sterba wrote:
> > > On Mon, Dec 30, 2019 at 04:31:15PM -0500, Josef Bacik wrote:
> > >> @@ -60,6 +60,11 @@ struct btrfs_inode {
> > >>     */
> > >>    struct extent_io_tree io_failure_tree;
> > >>
> > >> +  /* keeps track of where we have extent items mapped in order to make
> > >> +   * sure our i_size adjustments are accurate.
> > >> +   */
> > >> +  struct extent_io_tree file_extent_tree;
> > >
> > > This is not exactly lightweight and cut to the minimum needed, the size
> > > is 40 bytes and contains struct members that are unused. At least the
> > > file extents tree seems to be in use unlike that io_failure_tree wasting
> > > the bytes almost 100% of time.
> > >
> >
> > I'm not in love with it either, but I don't want to invent some lighter weight
> > range thingy that I'm going to end up messing up in other ways.
>
> Yeah, the extent_io_tree is now being used for generic range tree, some
> cleanups could remove the members that were added for the first specific
> use (like the dirty_bytes, or track_updates). For correctness of NOHOLES
> the inode size increase shouldn't be a blocker but needs to be addressed
> later.

Hum?
This isn't about the NO_HOLES case, it's about unordered i_size update
which can often lead to missing extent items for a file - when using
NO_HOLES fsck simply doesn't report any issue, but when not using it,
it does report about missing extents.

>
> > Don't take this series yet, there's still something fishy going on that I have
> > to figure out.  Thanks,
>
> Ok.
David Sterba Jan. 7, 2020, 4:46 p.m. UTC | #5
On Mon, Dec 30, 2019 at 04:31:15PM -0500, Josef Bacik wrote:
> +int btrfs_inode_set_file_extent_range(struct btrfs_inode *inode, u64 start,
> +				      u64 len)
> +{
> +	if (len == 0)
> +		return 0;
> +
> +	ASSERT(IS_ALIGNED(start + len, inode->root->fs_info->sectorsize));
> +
> +	if (btrfs_fs_incompat(inode->root->fs_info, NO_HOLES))
> +		return 0;
> +	return set_extent_bits(&inode->file_extent_tree, start, start + len - 1,
> +			       EXTENT_DIRTY);

set_extent_bits and friends do the allocations for range splits so this
is going to decrease performance. For the allocations and tree
traversals.  With enabled no-holes it's not going to be a problem, but
there are still many filesystems without the feature enabled so this
needs some evaluation.

Patch
diff mbox series

diff --git a/fs/btrfs/btrfs_inode.h b/fs/btrfs/btrfs_inode.h
index 4e12a477d32e..d9dcbac513ed 100644
--- a/fs/btrfs/btrfs_inode.h
+++ b/fs/btrfs/btrfs_inode.h
@@ -60,6 +60,11 @@  struct btrfs_inode {
 	 */
 	struct extent_io_tree io_failure_tree;
 
+	/* keeps track of where we have extent items mapped in order to make
+	 * sure our i_size adjustments are accurate.
+	 */
+	struct extent_io_tree file_extent_tree;
+
 	/* held while logging the inode in tree-log.c */
 	struct mutex log_mutex;
 
diff --git a/fs/btrfs/ctree.h b/fs/btrfs/ctree.h
index a01ff3e0ead4..7142124d03c5 100644
--- a/fs/btrfs/ctree.h
+++ b/fs/btrfs/ctree.h
@@ -2807,6 +2807,11 @@  void btrfs_extent_item_to_extent_map(struct btrfs_inode *inode,
 				     struct btrfs_file_extent_item *fi,
 				     const bool new_inline,
 				     struct extent_map *em);
+int btrfs_inode_clear_file_extent_range(struct btrfs_inode *inode, u64 start,
+					u64 len);
+int btrfs_inode_set_file_extent_range(struct btrfs_inode *inode, u64 start,
+				      u64 len);
+void btrfs_inode_safe_disk_i_size_write(struct inode *inode, u64 new_isize);
 
 /* inode.c */
 struct extent_map *btrfs_get_extent_fiemap(struct btrfs_inode *inode,
diff --git a/fs/btrfs/extent-io-tree.h b/fs/btrfs/extent-io-tree.h
index a3febe746c79..c8bcd2e3184c 100644
--- a/fs/btrfs/extent-io-tree.h
+++ b/fs/btrfs/extent-io-tree.h
@@ -44,6 +44,7 @@  enum {
 	IO_TREE_TRANS_DIRTY_PAGES,
 	IO_TREE_ROOT_DIRTY_LOG_PAGES,
 	IO_TREE_SELFTEST,
+	IO_TREE_INODE_FILE_EXTENT,
 };
 
 struct extent_io_tree {
diff --git a/fs/btrfs/extent_io.c b/fs/btrfs/extent_io.c
index e374411ed08c..410f5a64d3a6 100644
--- a/fs/btrfs/extent_io.c
+++ b/fs/btrfs/extent_io.c
@@ -265,6 +265,15 @@  void __cold extent_io_exit(void)
 	bioset_exit(&btrfs_bioset);
 }
 
+/*
+ * For the file_extent_tree, we want to hold the inode lock when we lookup and
+ * update the disk_i_size, but lockdep will complain because our io_tree we hold
+ * the tree lock and get the inode lock when setting delalloc.  These two things
+ * are unrelated, so make a class for the file_extent_tree so we don't get the
+ * two locking patterns mixed up.
+ */
+static struct lock_class_key file_extent_tree_class;
+
 void extent_io_tree_init(struct btrfs_fs_info *fs_info,
 			 struct extent_io_tree *tree, unsigned int owner,
 			 void *private_data)
@@ -276,6 +285,8 @@  void extent_io_tree_init(struct btrfs_fs_info *fs_info,
 	spin_lock_init(&tree->lock);
 	tree->private_data = private_data;
 	tree->owner = owner;
+	if (owner == IO_TREE_INODE_FILE_EXTENT)
+		lockdep_set_class(&tree->lock, &file_extent_tree_class);
 }
 
 void extent_io_tree_release(struct extent_io_tree *tree)
diff --git a/fs/btrfs/file-item.c b/fs/btrfs/file-item.c
index 1a599f50837b..b733d85510ed 100644
--- a/fs/btrfs/file-item.c
+++ b/fs/btrfs/file-item.c
@@ -23,6 +23,97 @@ 
 #define MAX_CSUM_ITEMS(r, size) (min_t(u32, __MAX_CSUM_ITEMS(r, size), \
 				       PAGE_SIZE))
 
+/**
+ * @inode - the inode we want to update the disk_i_size for
+ * @new_isize - the isize we want to set to, 0 if we use i_size
+ *
+ * With NO_HOLES set this simply sets the disk_is_size to whatever i_size_read()
+ * returns as it is perfectly fine with a file that has holes without hole file
+ * extent items.
+ *
+ * However for !NO_HOLES we need to only return the area that is contiguous from
+ * the 0 offset of the file.  Otherwise we could end up adjust i_size up to an
+ * extent that has a gap in between.
+ *
+ * Finally new_isize should only be set in the case of truncate where we're not
+ * ready to use i_size_read() as the limiter yet.
+ */
+void btrfs_inode_safe_disk_i_size_write(struct inode *inode, u64 new_isize)
+{
+	struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
+	u64 start, end, isize;
+	int ret;
+
+	isize = new_isize ? new_isize : i_size_read(inode);
+	if (btrfs_fs_incompat(fs_info, NO_HOLES)) {
+		BTRFS_I(inode)->disk_i_size = isize;
+		return;
+	}
+
+	spin_lock(&BTRFS_I(inode)->lock);
+	ret = find_first_extent_bit(&BTRFS_I(inode)->file_extent_tree, 0,
+				    &start, &end, EXTENT_DIRTY, NULL);
+	if (!ret && start == 0)
+		isize = min(isize, end + 1);
+	else
+		isize = 0;
+	BTRFS_I(inode)->disk_i_size = isize;
+	spin_unlock(&BTRFS_I(inode)->lock);
+}
+
+/**
+ * @inode - the inode we're modifying
+ * @start - the start file offset of the file extent we've inserted
+ * @len - the logical length of the file extent item
+ *
+ * Call when we are insering a new file extent where there was none before.
+ * Does not need to call this in the case where we're replacing an existing file
+ * extent, however if you're unsure it's fine to call this multiple times.
+ *
+ * The start and len must match the file extent item, so thus must be sectorsize
+ * aligned.
+ */
+int btrfs_inode_set_file_extent_range(struct btrfs_inode *inode, u64 start,
+				      u64 len)
+{
+	if (len == 0)
+		return 0;
+
+	ASSERT(IS_ALIGNED(start + len, inode->root->fs_info->sectorsize));
+
+	if (btrfs_fs_incompat(inode->root->fs_info, NO_HOLES))
+		return 0;
+	return set_extent_bits(&inode->file_extent_tree, start, start + len - 1,
+			       EXTENT_DIRTY);
+}
+
+/**
+ * @inode - the inode we're modifying
+ * @start - the start file offset of the file extent we've inserted
+ * @len - the logical length of the file extent item
+ *
+ * Called when we drop a file extent, for example when we truncate.  Doesn't
+ * need to be called for cases where we're replacing a file extent, like when
+ * we've cow'ed a file extent.
+ *
+ * The start and len must match the file extent item, so thus must be sectorsize
+ * aligned.
+ */
+int btrfs_inode_clear_file_extent_range(struct btrfs_inode *inode, u64 start,
+					u64 len)
+{
+	if (len == 0)
+		return 0;
+
+	ASSERT(IS_ALIGNED(start + len, inode->root->fs_info->sectorsize) ||
+	       len == (u64)-1);
+
+	if (btrfs_fs_incompat(inode->root->fs_info, NO_HOLES))
+		return 0;
+	return clear_extent_bit(&inode->file_extent_tree, start,
+				start + len - 1, EXTENT_DIRTY, 0, 0, NULL);
+}
+
 static inline u32 max_ordered_sum_bytes(struct btrfs_fs_info *fs_info,
 					u16 csum_size)
 {
diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c
index db67e1984c91..ab8b972863b1 100644
--- a/fs/btrfs/inode.c
+++ b/fs/btrfs/inode.c
@@ -3784,6 +3784,9 @@  static int btrfs_read_locked_inode(struct inode *inode,
 	i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
 	i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
 	btrfs_i_size_write(BTRFS_I(inode), btrfs_inode_size(leaf, inode_item));
+	btrfs_inode_set_file_extent_range(BTRFS_I(inode), 0,
+					  round_up(i_size_read(inode),
+						   fs_info->sectorsize));
 
 	inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
 	inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
@@ -9363,6 +9366,8 @@  struct inode *btrfs_alloc_inode(struct super_block *sb)
 	extent_io_tree_init(fs_info, &ei->io_tree, IO_TREE_INODE_IO, inode);
 	extent_io_tree_init(fs_info, &ei->io_failure_tree,
 			    IO_TREE_INODE_IO_FAILURE, inode);
+	extent_io_tree_init(fs_info, &ei->file_extent_tree,
+			    IO_TREE_INODE_FILE_EXTENT, inode);
 	ei->io_tree.track_uptodate = true;
 	ei->io_failure_tree.track_uptodate = true;
 	atomic_set(&ei->sync_writers, 0);
@@ -9429,6 +9434,7 @@  void btrfs_destroy_inode(struct inode *inode)
 	btrfs_qgroup_check_reserved_leak(inode);
 	inode_tree_del(inode);
 	btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
+	btrfs_inode_clear_file_extent_range(BTRFS_I(inode), 0, (u64)-1);
 	btrfs_put_root(BTRFS_I(inode)->root);
 }