Commit 80eabba7 authored by Linus Torvalds's avatar Linus Torvalds
Browse files

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
parents 852d21ae ce98321b

fs/buffer.c

+73 −31
Original line number Diff line number Diff line
@@ -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

+7 −21
Original line number Diff line number Diff line
@@ -842,24 +842,6 @@ out: 
	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

+3 −6
Original line number Diff line number Diff line
@@ -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

+2 −11
Original line number Diff line number Diff line
@@ -3777,14 +3777,6 @@ out: 
	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 @@ out: 
	 * 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

+5 −13
Original line number Diff line number Diff line
@@ -661,12 +661,8 @@ found: 
		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;

}

CRA Git | Maintained and supported by SUSTech CRA and CCSE