Merge branch 'for-4.10/fs-unmap' of git://git.kernel.dk/linux-block

Pull fs meta data unmap optimization from Jens Axboe:
 "A series from Jan Kara, providing a more efficient way for unmapping
  meta data from in the buffer cache than doing it block-by-block.

  Provide a general helper that existing callers can use"

* 'for-4.10/fs-unmap' of git://git.kernel.dk/linux-block:
  fs: Remove unmap_underlying_metadata
  fs: Add helper to clean bdev aliases under a bh and use it
  ext2: Use clean_bdev_aliases() instead of iteration
  ext4: Use clean_bdev_aliases() instead of iteration
  direct-io: Use clean_bdev_aliases() instead of handmade iteration
  fs: Provide function to unmap metadata for a range of blocks

fs/buffer.c

@@ -43,6 +43,7 @@
 #include <linux/bitops.h>
 #include <linux/mpage.h>
 #include <linux/bit_spinlock.h>
+#include <linux/pagevec.h>
 #include <trace/events/block.h>
 
 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
@@ -1604,37 +1605,80 @@ void create_empty_buffers(struct page *page,
 }
 EXPORT_SYMBOL(create_empty_buffers);
 
-/*
- * We are taking a block for data and we don't want any output from any
- * buffer-cache aliases starting from return from that function and
- * until the moment when something will explicitly mark the buffer
- * dirty (hopefully that will not happen until we will free that block ;-)
- * We don't even need to mark it not-uptodate - nobody can expect
- * anything from a newly allocated buffer anyway. We used to used
- * unmap_buffer() for such invalidation, but that was wrong. We definitely
- * don't want to mark the alias unmapped, for example - it would confuse
- * anyone who might pick it with bread() afterwards...
- *
- * Also..  Note that bforget() doesn't lock the buffer.  So there can
- * be writeout I/O going on against recently-freed buffers.  We don't
- * wait on that I/O in bforget() - it's more efficient to wait on the I/O
- * only if we really need to.  That happens here.
- */
-void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
+/**
+ * clean_bdev_aliases: clean a range of buffers in block device
+ * @bdev: Block device to clean buffers in
+ * @block: Start of a range of blocks to clean
+ * @len: Number of blocks to clean
+ *
+ * We are taking a range of blocks for data and we don't want writeback of any
+ * buffer-cache aliases starting from return from this function and until the
+ * moment when something will explicitly mark the buffer dirty (hopefully that
+ * will not happen until we will free that block ;-) We don't even need to mark
+ * it not-uptodate - nobody can expect anything from a newly allocated buffer
+ * anyway. We used to use unmap_buffer() for such invalidation, but that was
+ * wrong. We definitely don't want to mark the alias unmapped, for example - it
+ * would confuse anyone who might pick it with bread() afterwards...
+ *
+ * Also..  Note that bforget() doesn't lock the buffer.  So there can be
+ * writeout I/O going on against recently-freed buffers.  We don't wait on that
+ * I/O in bforget() - it's more efficient to wait on the I/O only if we really
+ * need to.  That happens here.
+ */
+void clean_bdev_aliases(struct block_device *bdev, sector_t block, sector_t len)
 {
-	struct buffer_head *old_bh;
+	struct inode *bd_inode = bdev->bd_inode;
+	struct address_space *bd_mapping = bd_inode->i_mapping;
+	struct pagevec pvec;
+	pgoff_t index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
+	pgoff_t end;
+	int i;
+	struct buffer_head *bh;
+	struct buffer_head *head;
 
-	might_sleep();
+	end = (block + len - 1) >> (PAGE_SHIFT - bd_inode->i_blkbits);
+	pagevec_init(&pvec, 0);
+	while (index <= end && pagevec_lookup(&pvec, bd_mapping, index,
+			min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
+		for (i = 0; i < pagevec_count(&pvec); i++) {
+			struct page *page = pvec.pages[i];
 
-	old_bh = __find_get_block_slow(bdev, block);
-	if (old_bh) {
-		clear_buffer_dirty(old_bh);
-		wait_on_buffer(old_bh);
-		clear_buffer_req(old_bh);
-		__brelse(old_bh);
+			index = page->index;
+			if (index > end)
+				break;
+			if (!page_has_buffers(page))
+				continue;
+			/*
+			 * We use page lock instead of bd_mapping->private_lock
+			 * to pin buffers here since we can afford to sleep and
+			 * it scales better than a global spinlock lock.
+			 */
+			lock_page(page);
+			/* Recheck when the page is locked which pins bhs */
+			if (!page_has_buffers(page))
+				goto unlock_page;
+			head = page_buffers(page);
+			bh = head;
+			do {
+				if (!buffer_mapped(bh))
+					goto next;
+				if (bh->b_blocknr >= block + len)
+					break;
+				clear_buffer_dirty(bh);
+				wait_on_buffer(bh);
+				clear_buffer_req(bh);
+next:
+				bh = bh->b_this_page;
+			} while (bh != head);
+unlock_page:
+			unlock_page(page);
+		}
+		pagevec_release(&pvec);
+		cond_resched();
+		index++;
 	}
 }
-EXPORT_SYMBOL(unmap_underlying_metadata);
+EXPORT_SYMBOL(clean_bdev_aliases);
 
 /*
  * Size is a power-of-two in the range 512..PAGE_SIZE,
@@ -1745,8 +1789,7 @@ int __block_write_full_page(struct inode *inode, struct page *page,
 			if (buffer_new(bh)) {
 				/* blockdev mappings never come here */
 				clear_buffer_new(bh);
-				unmap_underlying_metadata(bh->b_bdev,
-							bh->b_blocknr);
+				clean_bdev_bh_alias(bh);
 			}
 		}
 		bh = bh->b_this_page;
@@ -1992,8 +2035,7 @@ int __block_write_begin_int(struct page *page, loff_t pos, unsigned len,
 			}
 
 			if (buffer_new(bh)) {
-				unmap_underlying_metadata(bh->b_bdev,
-							bh->b_blocknr);
+				clean_bdev_bh_alias(bh);
 				if (PageUptodate(page)) {
 					clear_buffer_new(bh);
 					set_buffer_uptodate(bh);
@@ -2633,7 +2675,7 @@ int nobh_write_begin(struct address_space *mapping,
 		if (!buffer_mapped(bh))
 			is_mapped_to_disk = 0;
 		if (buffer_new(bh))
-			unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
+			clean_bdev_bh_alias(bh);
 		if (PageUptodate(page)) {
 			set_buffer_uptodate(bh);
 			continue;

fs/direct-io.c

@@ -842,24 +842,6 @@ submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
 	return ret;
 }
 
-/*
- * Clean any dirty buffers in the blockdev mapping which alias newly-created
- * file blocks.  Only called for S_ISREG files - blockdevs do not set
- * buffer_new
- */
-static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
-{
-	unsigned i;
-	unsigned nblocks;
-
-	nblocks = map_bh->b_size >> dio->inode->i_blkbits;
-
-	for (i = 0; i < nblocks; i++) {
-		unmap_underlying_metadata(map_bh->b_bdev,
-					  map_bh->b_blocknr + i);
-	}
-}
-
 /*
  * If we are not writing the entire block and get_block() allocated
  * the block for us, we need to fill-in the unused portion of the
@@ -960,11 +942,15 @@ static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
 					goto do_holes;
 
 				sdio->blocks_available =
-						map_bh->b_size >> sdio->blkbits;
+						map_bh->b_size >> blkbits;
 				sdio->next_block_for_io =
 					map_bh->b_blocknr << sdio->blkfactor;
-				if (buffer_new(map_bh))
-					clean_blockdev_aliases(dio, map_bh);
+				if (buffer_new(map_bh)) {
+					clean_bdev_aliases(
+						map_bh->b_bdev,
+						map_bh->b_blocknr,
+						map_bh->b_size >> blkbits);
+				}
 
 				if (!sdio->blkfactor)
 					goto do_holes;

fs/ext2/inode.c

@@ -732,16 +732,13 @@ static int ext2_get_blocks(struct inode *inode,
 	}
 
 	if (IS_DAX(inode)) {
-		int i;
-
 		/*
 		 * We must unmap blocks before zeroing so that writeback cannot
 		 * overwrite zeros with stale data from block device page cache.
 		 */
-		for (i = 0; i < count; i++) {
-			unmap_underlying_metadata(inode->i_sb->s_bdev,
-					le32_to_cpu(chain[depth-1].key) + i);
-		}
+		clean_bdev_aliases(inode->i_sb->s_bdev,
+				   le32_to_cpu(chain[depth-1].key),
+				   count);
 		/*
 		 * block must be initialised before we put it in the tree
 		 * so that it's not found by another thread before it's

fs/ext4/extents.c

@@ -3777,14 +3777,6 @@ static int ext4_convert_unwritten_extents_endio(handle_t *handle,
 	return err;
 }
 
-static void unmap_underlying_metadata_blocks(struct block_device *bdev,
-			sector_t block, int count)
-{
-	int i;
-	for (i = 0; i < count; i++)
-                unmap_underlying_metadata(bdev, block + i);
-}
-
 /*
  * Handle EOFBLOCKS_FL flag, clearing it if necessary
  */
@@ -4121,8 +4113,7 @@ ext4_ext_handle_unwritten_extents(handle_t *handle, struct inode *inode,
 	 * new.
 	 */
 	if (allocated > map->m_len) {
-		unmap_underlying_metadata_blocks(inode->i_sb->s_bdev,
-					newblock + map->m_len,
+		clean_bdev_aliases(inode->i_sb->s_bdev, newblock + map->m_len,
 				   allocated - map->m_len);
 		allocated = map->m_len;
 	}

fs/ext4/inode.c

@@ -661,12 +661,8 @@ int ext4_map_blocks(handle_t *handle, struct inode *inode,
 		if (flags & EXT4_GET_BLOCKS_ZERO &&
 		    map->m_flags & EXT4_MAP_MAPPED &&
 		    map->m_flags & EXT4_MAP_NEW) {
-			ext4_lblk_t i;
-
-			for (i = 0; i < map->m_len; i++) {
-				unmap_underlying_metadata(inode->i_sb->s_bdev,
-							  map->m_pblk + i);
-			}
+			clean_bdev_aliases(inode->i_sb->s_bdev, map->m_pblk,
+					   map->m_len);
 			ret = ext4_issue_zeroout(inode, map->m_lblk,
 						 map->m_pblk, map->m_len);
 			if (ret) {
@@ -1137,8 +1133,7 @@ static int ext4_block_write_begin(struct page *page, loff_t pos, unsigned len,
 			if (err)
 				break;
 			if (buffer_new(bh)) {
-				unmap_underlying_metadata(bh->b_bdev,
-							  bh->b_blocknr);
+				clean_bdev_bh_alias(bh);
 				if (PageUptodate(page)) {
 					clear_buffer_new(bh);
 					set_buffer_uptodate(bh);
@@ -2371,11 +2366,8 @@ static int mpage_map_one_extent(handle_t *handle, struct mpage_da_data *mpd)
 
 	BUG_ON(map->m_len == 0);
 	if (map->m_flags & EXT4_MAP_NEW) {
-		struct block_device *bdev = inode->i_sb->s_bdev;
-		int i;
-
-		for (i = 0; i < map->m_len; i++)
-			unmap_underlying_metadata(bdev, map->m_pblk + i);
+		clean_bdev_aliases(inode->i_sb->s_bdev, map->m_pblk,
+				   map->m_len);
 	}
 	return 0;
 }

fs/ext4/page-io.c

@@ -457,7 +457,7 @@ int ext4_bio_write_page(struct ext4_io_submit *io,
 		}
 		if (buffer_new(bh)) {
 			clear_buffer_new(bh);
-			unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
+			clean_bdev_bh_alias(bh);
 		}
 		set_buffer_async_write(bh);
 		nr_to_submit++;

fs/mpage.c

@@ -555,8 +555,7 @@ static int __mpage_writepage(struct page *page, struct writeback_control *wbc,
 		if (mpd->get_block(inode, block_in_file, &map_bh, 1))
 			goto confused;
 		if (buffer_new(&map_bh))
-			unmap_underlying_metadata(map_bh.b_bdev,
-						map_bh.b_blocknr);
+			clean_bdev_bh_alias(&map_bh);
 		if (buffer_boundary(&map_bh)) {
 			boundary_block = map_bh.b_blocknr;
 			boundary_bdev = map_bh.b_bdev;

fs/ntfs/aops.c

@@ -765,7 +765,7 @@ static int ntfs_write_block(struct page *page, struct writeback_control *wbc)
 			}
 			// TODO: Instantiate the hole.
 			// clear_buffer_new(bh);
-			// unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
+			// clean_bdev_bh_alias(bh);
 			ntfs_error(vol->sb, "Writing into sparse regions is "
 					"not supported yet. Sorry.");
 			err = -EOPNOTSUPP;

fs/ntfs/file.c

@@ -740,8 +740,7 @@ static int ntfs_prepare_pages_for_non_resident_write(struct page **pages,
 					set_buffer_uptodate(bh);
 				if (unlikely(was_hole)) {
 					/* We allocated the buffer. */
-					unmap_underlying_metadata(bh->b_bdev,
-							bh->b_blocknr);
+					clean_bdev_bh_alias(bh);
 					if (bh_end <= pos || bh_pos >= end)
 						mark_buffer_dirty(bh);
 					else
@@ -784,7 +783,7 @@ static int ntfs_prepare_pages_for_non_resident_write(struct page **pages,
 				continue;
 			}
 			/* We allocated the buffer. */
-			unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
+			clean_bdev_bh_alias(bh);
 			/*
 			 * If the buffer is fully outside the write, zero it,
 			 * set it uptodate, and mark it dirty so it gets

fs/ocfs2/aops.c

@@ -630,7 +630,7 @@ int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
 
 		if (!buffer_mapped(bh)) {
 			map_bh(bh, inode->i_sb, *p_blkno);
-			unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
+			clean_bdev_bh_alias(bh);
 		}
 
 		if (PageUptodate(page)) {

fs/ufs/balloc.c

@@ -307,8 +307,7 @@ static void ufs_change_blocknr(struct inode *inode, sector_t beg,
 			     (unsigned long long)(pos + newb), pos);
 
 			bh->b_blocknr = newb + pos;
-			unmap_underlying_metadata(bh->b_bdev,
-						  bh->b_blocknr);
+			clean_bdev_bh_alias(bh);
 			mark_buffer_dirty(bh);
 			++j;
 			bh = bh->b_this_page;

fs/ufs/inode.c

@@ -1070,8 +1070,7 @@ static int ufs_alloc_lastblock(struct inode *inode, loff_t size)
 
        if (buffer_new(bh)) {
 	       clear_buffer_new(bh);
-	       unmap_underlying_metadata(bh->b_bdev,
-					 bh->b_blocknr);
+	       clean_bdev_bh_alias(bh);
 	       /*
 		* we do not zeroize fragment, because of
 		* if it maped to hole, it already contains zeroes

include/linux/buffer_head.h

@@ -168,7 +168,12 @@ int inode_has_buffers(struct inode *);
 void invalidate_inode_buffers(struct inode *);
 int remove_inode_buffers(struct inode *inode);
 int sync_mapping_buffers(struct address_space *mapping);
-void unmap_underlying_metadata(struct block_device *bdev, sector_t block);
+void clean_bdev_aliases(struct block_device *bdev, sector_t block,
+			sector_t len);
+static inline void clean_bdev_bh_alias(struct buffer_head *bh)
+{
+	clean_bdev_aliases(bh->b_bdev, bh->b_blocknr, 1);
+}
 
 void mark_buffer_async_write(struct buffer_head *bh);
 void __wait_on_buffer(struct buffer_head *);