[RFC] basic delayed allocation in VFS

[Alex Tomas <alex@clusterfs.com>]

Good day,

please review ...

thanks, Alex


basic delayed allocation in VFS:

  * block_prepare_write() can be passed special ->get_block() which
    doesn't allocate blocks, but reserve them and mark bh delayed
  * a filesystem can use mpage_da_writepages() with other ->get_block()
    which doesn't defer allocation. mpage_da_writepages() finds all
    non-allocated blocks and try to allocate them with minimal calls
    to ->get_block(), then submit IO using __mpage_writepage()


Signed-off-by: Alex Tomas <alex@clusterfs.com>


Index: linux-2.6.22-rc4/fs/buffer.c
===================================================================
--- linux-2.6.22-rc4.orig/fs/buffer.c	2007-06-05 04:57:25.000000000 +0400
+++ linux-2.6.22-rc4/fs/buffer.c	2007-07-17 22:40:49.000000000 +0400
@@ -1630,7 +1630,8 @@ static int __block_write_full_page(struc
  			 */
  			clear_buffer_dirty(bh);
  			set_buffer_uptodate(bh);
-		} else if (!buffer_mapped(bh) && buffer_dirty(bh)) {
+		} else if ((!buffer_mapped(bh) || buffer_delay(bh))
+				&& buffer_dirty(bh)) {
  			WARN_ON(bh->b_size != blocksize);
  			err = get_block(inode, block, bh, 1);
  			if (err)
Index: linux-2.6.22-rc4/include/linux/mpage.h
===================================================================
--- linux-2.6.22-rc4.orig/include/linux/mpage.h	2007-06-05 04:57:25.000000000 +0400
+++ linux-2.6.22-rc4/include/linux/mpage.h	2007-07-25 09:48:44.000000000 +0400
@@ -20,5 +20,7 @@ int mpage_writepages(struct address_spac
  		struct writeback_control *wbc, get_block_t get_block);
  int mpage_writepage(struct page *page, get_block_t *get_block,
  		struct writeback_control *wbc);
+int mpage_da_writepages(struct address_space *mapping,
+		struct writeback_control *wbc, get_block_t get_block);

  #endif
Index: linux-2.6.22-rc4/fs/mpage.c
===================================================================
--- linux-2.6.22-rc4.orig/fs/mpage.c	2007-06-05 04:57:25.000000000 +0400
+++ linux-2.6.22-rc4/fs/mpage.c	2007-07-26 11:59:39.000000000 +0400
@@ -10,6 +10,8 @@
   *		Initial version
   * 27Jun2002	axboe@suse.de
   *		use bio_add_page() to build bio's just the right size
+ * 26Jul2007	alex@clusterfs.com AKA bzzz
+ *		basic delayed allocation support
   */

  #include <linux/kernel.h>
@@ -732,3 +734,404 @@ int mpage_writepage(struct page *page, g
  	return ret;
  }
  EXPORT_SYMBOL(mpage_writepage);
+
+/*
+ * Delayed allocation stuff
+ */
+
+struct mpage_da_data {
+	struct inode *inode;
+	struct buffer_head lbh;			/* extent of blocks */
+	unsigned long first_page, next_page;	/* extent of pages */
+	get_block_t *get_block;
+	struct writeback_control *wbc;
+};
+
+
+/*
+ * mpage_da_submit_io - walks through extent of pages and try to write
+ * them with __mpage_writepage()
+ *
+ * @mpd->inode: inode
+ * @mpd->first_page: first page of the extent
+ * @mpd->next_page: page after the last page of the extent
+ * @mpd->get_block: the filesystem's block mapper function
+ *
+ * By the time mpage_da_submit_io() is called we expect all blocks
+ * to be allocated. this may be wrong if allocation failed.
+ *
+ * As pages are already locked by write_cache_pages(), we can't use it
+ */
+static int mpage_da_submit_io(struct mpage_da_data *mpd)
+{
+	struct address_space *mapping = mpd->inode->i_mapping;
+	struct mpage_data mpd_pp = {
+		.bio = NULL,
+		.last_block_in_bio = 0,
+		.get_block = mpd->get_block,
+		.use_writepage = 1,
+	};
+	int ret = 0, err, nr_pages, i;
+	unsigned long index, end;
+	struct pagevec pvec;
+
+	BUG_ON(mpd->next_page <= mpd->first_page);
+
+	pagevec_init(&pvec, 0);
+	index = mpd->first_page;
+	end = mpd->next_page - 1;
+
+	while (index <= end) {
+		/* XXX: optimize tail */
+		nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
+		if (nr_pages == 0)
+			break;for (i = 0; i < nr_pages; i++) {
+			struct page *page = pvec.pages[i];
+
+			index = page->index;
+			if (index > end)
+				break;
+			index++;
+
+			err = __mpage_writepage(page, mpd->wbc, &mpd_pp);
+
+			/*
+			 * In error case, we have to continue because
+			 * remaining pages are still locked
+			 * XXX: unlock and re-dirty them?
+			 */
+			if (ret == 0)
+				ret = err;
+		}
+		pagevec_release(&pvec);
+	}
+	if (mpd_pp.bio)
+		mpage_bio_submit(WRITE, mpd_pp.bio);
+
+	return ret;
+}
+
+/*
+ * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
+ *
+ * @mpd->inode - inode to walk through
+ * @exbh->b_blocknr - first block on a disk
+ * @exbh->b_size - amount of space in bytes
+ * @logical - first logical block to start assignment with
+ *
+ * the function goes through all passed space and put actual disk
+ * block numbers into buffer heads, dropping BH_Delay
+ */
+static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd, sector_t logical,
+				struct buffer_head *exbh)
+{
+	struct inode *inode = mpd->inode;
+	struct address_space *mapping = inode->i_mapping;
+	int blocks = exbh->b_size >> inode->i_blkbits;
+	sector_t pblock = exbh->b_blocknr, cur_logical;
+	struct buffer_head *head, *bh;
+	unsigned long index, end;
+	struct pagevec pvec;
+	int nr_pages, i;
+
+	index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
+	end = (logical + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
+	cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
+
+	pagevec_init(&pvec, 0);
+
+	while (index <= end) {
+		/* XXX: optimize tail */
+		nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
+		if (nr_pages == 0)
+			break;
+		for (i = 0; i < nr_pages; i++) {
+			struct page *page = pvec.pages[i];
+
+			index = page->index;
+			if (index > end)
+				break;
+			index++;
+
+			BUG_ON(!PageLocked(page));
+			BUG_ON(PageWriteback(page));
+			BUG_ON(!page_has_buffers(page));
+
+			head = bh = page_buffers(page);
+
+			/* skip blocks out of the range */
+			do {
+				if (cur_logical >= logical)
+					break;
+				cur_logical++;
+				pblock++;
+			} while ((bh = bh->b_this_page) != head);
+
+			do {
+				if (cur_logical >= logical + blocks)
+					break;
+
+				if (buffer_delay(bh)) {
+					bh->b_blocknr = pblock;
+					clear_buffer_delay(bh);
+				} else if (buffer_mapped(bh)) {
+					BUG_ON(bh->b_blocknr != pblock);
+				}
+
+				cur_logical++;
+				pblock++;
+			} while ((bh = bh->b_this_page) != head);
+		}
+		pagevec_release(&pvec);
+	}
+}
+
+
+/*
+ * __unmap_underlying_blocks - just a helper function to unmap
+ * set of blocks described by @bh
+ */
+static inline void __unmap_underlying_blocks(struct inode *inode,
+					struct buffer_head *bh)
+{
+	struct block_device *bdev = inode->i_sb->s_bdev;
+	int blocks, i;
+
+	blocks = bh->b_size >> inode->i_blkbits;
+	for (i = 0; i < blocks; i++)
+		unmap_underlying_metadata(bdev, bh->b_blocknr + i);
+}
+
+/*
+ * mpage_da_map_blocks - go through given space
+ *
+ * @mpd->lbh - bh describing space
+ * @mpd->get_block - the filesystem's block mapper function
+ *
+ * The function skips space we know is already mapped to disk blocks.
+ *
+ * The function ignores errors ->get_block() returns, thus real
+ * error handling is postponed to __mpage_writepage()
+ */
+static void mpage_da_map_blocks(struct mpage_da_data *mpd)
+{
+	struct buffer_head *lbh = &mpd->lbh;
+	int err = 0, remain = lbh->b_size;
+	sector_t next = lbh->b_blocknr;
+	struct buffer_head new;
+
+	/*
+	 * We consider only non-mapped and non-allocated blocks
+	 */
+	if (buffer_mapped(lbh) && !buffer_delay(lbh))
+		return;
+
+	while (remain) {
+		new.b_state = lbh->b_state;
+		new.b_blocknr = 0;
+		new.b_size = remain;
+		err = mpd->get_block(mpd->inode, next, &new, 1);
+		if (err) {
+			/*
+			 * Rather than implement own error handling
+			 * here, we just leave remaining blocks
+			 * unallocated and try again with ->writepage()
+			 */
+			break;
+		}
+		BUG_ON(new.b_size == 0);
+
+		if (buffer_new(&new))
+			__unmap_underlying_blocks(mpd->inode, &new);
+
+		/*
+		 * If blocks are delayed marked, we need to
+		 * put actual blocknr and drop delayed bit
+		 */
+		if (buffer_delay(lbh))
+			mpage_put_bnr_to_bhs(mpd, next, &new);
+
+		/* go for the remaining blocks */
+		next += new.b_size >> mpd->inode->i_blkbits;
+		remain -= new.b_size;
+	}
+}
+
+#define BH_FLAGS	((1 << BH_Uptodate) | (1 << BH_Mapped) | (1 << BH_Delay))
+
+/*
+ * mpage_add_bh_to_extent - try to add one more block to extent of blocks
+ *
+ * @mpd->lbh - extent of blocks
+ * @logical - logical number of the block in the file
+ * @bh - bh of the block (used to access block's state)
+ *
+ * the function is used to collect contig. blocks in same state
+ */
+static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
+				sector_t logical, struct buffer_head *bh)
+{
+	struct buffer_head *lbh = &mpd->lbh;
+	sector_t next;
+
+	next = lbh->b_blocknr + (lbh->b_size >> mpd->inode->i_blkbits);
+
+	/*
+	 * First block in the extent
+	 */
+	if (lbh->b_size == 0) {
+		lbh->b_blocknr = logical;
+		lbh->b_size = bh->b_size;
+		lbh->b_state = bh->b_state & BH_FLAGS;
+		return;
+	}
+
+	/*
+	 * Can we merge the block to our big extent?
+	 */
+	if (logical == next && (bh->b_state & BH_FLAGS) == lbh->b_state) {
+		lbh->b_size += bh->b_size;
+		return;
+	}
+
+	/*
+	 * We couldn't merge the block to our extent, so we
+	 * need to flush current  extent and start new one
+	 */
+	mpage_da_map_blocks(mpd);
+
+	/*
+	 * Now start a new extent
+	 */
+	lbh->b_size = bh->b_size;
+	lbh->b_state = bh->b_state & BH_FLAGS;
+	lbh->b_blocknr = logical;
+}
+
+/*
+ * __mpage_da_writepage - finds extent of pages and blocks
+ *
+ * @page: page to consider
+ * @wbc: not used, we just follow rules
+ * @data: context
+ *
+ * The function finds extents of pages and scan them for all blocks.
+ */
+static int __mpage_da_writepage(struct page *page,
+			struct writeback_control *wbc, void *data)
+{
+	struct mpage_da_data *mpd = data;
+	struct inode *inode = mpd->inode;
+	struct buffer_head *bh, *head, fake;
+	sector_t logical;
+
+	/*
+	 * Can we merge this page to current extent?
+	 */
+	if (mpd->next_page != page->index) {
+		/*
+		 * Nope, we can't. So, we map non-allocated blocks
+		 * and start IO on them using __mpage_writepage()
+		 */
+		if (mpd->next_page != mpd->first_page) {
+			mpage_da_map_blocks(mpd);
+			mpage_da_submit_io(mpd);
+		}
+
+		/*
+		 * Start next extent of pages ...
+		 */
+		mpd->first_page = page->index;
+
+		/*
+		 * ... and blocks
+		 */
+		mpd->lbh.b_size = 0;
+		mpd->lbh.b_state = 0;
+		mpd->lbh.b_blocknr = 0;
+	}
+
+	mpd->next_page = page->index + 1;
+	logical = (sector_t) page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
+
+	if (!page_has_buffers(page)) {
+		/*
+		 * There is no attached buffer heads yet (mmap?)
+		 * we treat the page asfull of dirty blocks
+		 */
+		bh = &fake;
+		bh->b_size = PAGE_CACHE_SIZE;
+		bh->b_state = 0;
+		set_buffer_dirty(bh);
+		set_buffer_uptodate(bh);
+		mpage_add_bh_to_extent(mpd, logical, bh);
+	} else {
+		/*
+		 * Page with regular buffer heads, just add all dirty ones
+		 */
+		bh = head = page_buffers(page);
+		do {
+			BUG_ON(buffer_locked(bh));
+			if (buffer_dirty(bh))
+				mpage_add_bh_to_extent(mpd, logical, bh);
+			logical++;
+		} while ((bh = bh->b_this_page) != head);
+	}
+
+	return 0;
+}
+
+/*
+ * mpage_da_writepages - walk the list of dirty pages of the given
+ * address space, allocates non-allocated blocks, maps newly-allocated
+ * blocks to existing bhs and issue IO them
+ *
+ * @mapping: address space structure to write
+ * @wbc: subtract the number of written pages from *@wbc->nr_to_write
+ * @get_block: the filesystem's block mapper function.
+ *
+ * This is a library function, which implements the writepages()
+ * address_space_operation.
+ *
+ * In order to avoid duplication of logic that deals with partial pages,
+ * multiple bio per page, etc, we find non-allocated blocks, allocate
+ * them with minimal calls to ->get_block() and re-use __mpage_writepage()
+ *
+ * It's important that we call __mpage_writepage() only once for each
+ * involved page, otherwise we'd have to implement more complicated logic
+ * to deal with pages w/o PG_lock or w/ PG_writeback and so on.
+ *
+ * See comments to mpage_writepages()
+ */
+int mpage_da_writepages(struct address_space *mapping,
+		struct writeback_control *wbc, get_block_t get_block)
+{
+	struct mpage_da_data mpd;
+	int ret;
+
+	if (!get_block)
+		return generic_writepages(mapping, wbc);
+
+	mpd.wbc = wbc;
+	mpd.inode = mapping->host;
+	mpd.lbh.b_size = 0;
+	mpd.lbh.b_state = 0;
+	mpd.lbh.b_blocknr = 0;
+	mpd.first_page = 0;
+	mpd.next_page = 0;
+	mpd.get_block = get_block;
+
+	ret = write_cache_pages(mapping, wbc, __mpage_da_writepage, &mpd);
+
+	/*
+	 * Handle last extent of pages
+	 */
+	if (mpd.next_page != mpd.first_page) {
+		mpage_da_map_blocks(&mpd);
+		mpage_da_submit_io(&mpd);
+	}
+
+	return ret;
+}
+EXPORT_SYMBOL(mpage_da_writepages);
+