On Friday 26 Jun 2015 17:50:54 Liu Bo wrote: > On Mon, Jun 01, 2015 at 08:52:37PM +0530, Chandan Rajendra wrote: > > For the subpagesize-blocksize scenario, a page can contain multiple > > blocks. In such cases, this patch handles writing data to files. > > > > Also, When setting EXTENT_DELALLOC, we no longer set EXTENT_UPTODATE bit > > on > > the extent_io_tree since uptodate status is being tracked by the bitmap > > pointed to by page->private. > > To be honestly, I'm not sure why we set EXTENT_UPTODATE bit for data as we > don't check for that bit at all for now, correct me if I'm wrong.
Yes, I didn't find any code using EXTENT_UPTODATE flag. That is probably because we could get away by referring to the page's PG_uptodate flag in blocksize == Pagesize scenario. But for the subpagesize-blocksize scenario we need BLK_STATE_UPTODATE to determine if a page's PG_uptodate flag can be set. > > > Signed-off-by: Chandan Rajendra <[email protected]> > > --- > > > > fs/btrfs/extent_io.c | 141 > > +++++++++++++++++++++++---------------------------- fs/btrfs/file.c > > | 16 ++++++ > > fs/btrfs/inode.c | 58 ++++++++++++++++----- > > 3 files changed, 125 insertions(+), 90 deletions(-) > > > > diff --git a/fs/btrfs/extent_io.c b/fs/btrfs/extent_io.c > > index d37badb..3736ab5 100644 > > --- a/fs/btrfs/extent_io.c > > +++ b/fs/btrfs/extent_io.c > > @@ -1283,9 +1283,8 @@ int clear_extent_bits(struct extent_io_tree *tree, > > u64 start, u64 end,> > > int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end, > > > > struct extent_state **cached_state, gfp_t mask) > > > > { > > > > - return set_extent_bit(tree, start, end, > > - EXTENT_DELALLOC | EXTENT_UPTODATE, > > - NULL, cached_state, mask); > > + return set_extent_bit(tree, start, end, EXTENT_DELALLOC, > > + NULL, cached_state, mask); > > > > } > > > > int set_extent_defrag(struct extent_io_tree *tree, u64 start, u64 end, > > > > @@ -1498,25 +1497,6 @@ int extent_range_redirty_for_io(struct inode > > *inode, u64 start, u64 end)> > > return 0; > > > > } > > > > -/* > > - * helper function to set both pages and extents in the tree writeback > > - */ > > -static int set_range_writeback(struct extent_io_tree *tree, u64 start, > > u64 end) -{ > > - unsigned long index = start >> PAGE_CACHE_SHIFT; > > - unsigned long end_index = end >> PAGE_CACHE_SHIFT; > > - struct page *page; > > - > > - while (index <= end_index) { > > - page = find_get_page(tree->mapping, index); > > - BUG_ON(!page); /* Pages should be in the extent_io_tree */ > > - set_page_writeback(page); > > - page_cache_release(page); > > - index++; > > - } > > - return 0; > > -} > > - > > > > /* find the first state struct with 'bits' set after 'start', and > > > > * return it. tree->lock must be held. NULL will returned if > > * nothing was found after 'start' > > > > @@ -2080,6 +2060,14 @@ static int page_read_complete(struct page *page) > > > > return !test_page_blks_state(page, BLK_STATE_IO, start, end, 0); > > > > } > > > > +static int page_write_complete(struct page *page) > > +{ > > + u64 start = page_offset(page); > > + u64 end = start + PAGE_CACHE_SIZE - 1; > > + > > + return !test_page_blks_state(page, BLK_STATE_IO, start, end, 0); > > +} > > + > > > > int free_io_failure(struct inode *inode, struct io_failure_record *rec) > > { > > > > int ret; > > > > @@ -2575,38 +2563,37 @@ int end_extent_writepage(struct page *page, int > > err, u64 start, u64 end)> > > */ > > > > static void end_bio_extent_writepage(struct bio *bio, int err) > > { > > > > + struct btrfs_page_private *pg_private; > > > > struct bio_vec *bvec; > > > > + unsigned long flags; > > > > u64 start; > > u64 end; > > > > + int clear_writeback; > > > > int i; > > > > bio_for_each_segment_all(bvec, bio, i) { > > > > struct page *page = bvec->bv_page; > > > > - /* We always issue full-page reads, but if some block > > - * in a page fails to read, blk_update_request() will > > - * advance bv_offset and adjust bv_len to compensate. > > - * Print a warning for nonzero offsets, and an error > > - * if they don't add up to a full page. */ > > - if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) { > > - if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE) > > - btrfs_err(BTRFS_I(page->mapping->host)->root- >fs_info, > > - "partial page write in btrfs with offset %u and length %u", > > - bvec->bv_offset, bvec->bv_len); > > - else > > - btrfs_info(BTRFS_I(page->mapping->host)->root- >fs_info, > > - "incomplete page write in btrfs with offset %u and " > > - "length %u", > > - bvec->bv_offset, bvec->bv_len); > > - } > > + start = page_offset(page) + bvec->bv_offset; > > + end = start + bvec->bv_len - 1; > > > > - start = page_offset(page); > > - end = start + bvec->bv_offset + bvec->bv_len - 1; > > + pg_private = (struct btrfs_page_private *)page->private; > > + > > + spin_lock_irqsave(&pg_private->io_lock, flags); > > > > - if (end_extent_writepage(page, err, start, end)) > > + if (end_extent_writepage(page, err, start, end)) { > > + spin_unlock_irqrestore(&pg_private->io_lock, flags); > > > > continue; > > > > + } > > > > - end_page_writeback(page); > > + clear_page_blks_state(page, 1 << BLK_STATE_IO, start, end); > > + > > + clear_writeback = page_write_complete(page); > > + > > + spin_unlock_irqrestore(&pg_private->io_lock, flags); > > + > > + if (clear_writeback) > > + end_page_writeback(page); > > > > } > > > > bio_put(bio); > > > > @@ -3417,10 +3404,9 @@ static noinline_for_stack int > > __extent_writepage_io(struct inode *inode,> > > u64 block_start; > > u64 iosize; > > sector_t sector; > > > > - struct extent_state *cached_state = NULL; > > > > struct extent_map *em; > > struct block_device *bdev; > > > > - size_t pg_offset = 0; > > + size_t pg_offset; > > > > size_t blocksize; > > int ret = 0; > > int nr = 0; > > > > @@ -3467,8 +3453,16 @@ static noinline_for_stack int > > __extent_writepage_io(struct inode *inode,> > > page_end, NULL, 1); > > > > break; > > > > } > > > > - em = epd->get_extent(inode, page, pg_offset, cur, > > - end - cur + 1, 1); > > + > > + pg_offset = cur & (PAGE_CACHE_SIZE - 1); > > + > > + if (!test_page_blks_state(page, BLK_STATE_DIRTY, cur, > > + cur + blocksize - 1, 1)) { > > + cur += blocksize; > > + continue; > > + } > > If we don't check this, the below get_extent() will return a HOLE > (block_start == EXTENT_MAP_HOLE) and we can still go on to the next block, > then we don't need to maintain this BLK_STATE_DIRTY bit all the while. Sorry, I am not sure if I understood your comment correctly. Are you suggesting that *page blocks* that are not dirty are always holes? Let's assume a 64k page whose contents are within i_size and none of the blocks of the page map to a file hole. Also assume 4k as the block size. Say, the userspace writes to the "block 0" of the page. The corresponding code in __btrfs_buffered_write() reads up the complete page into the inode's page cache and then marks "block 0" of the page as BLK_STATE_DIRTY. Next, the userspace seeks and writes to "block 4" of the page. In this case, since the page has PG_uptodate flag already set we don't read the data from the disk again. We simply go ahead and mark "block 4" as BLK_STATE_DIRTY. As can be seen in the example scenario, the blocks 1, 2 and 3 are not holes and hence btrfs_get_extent() would end up returning values other than EXTENT_MAP_HOLE for em->block_start. > > > + > > + em = epd->get_extent(inode, page, pg_offset, cur, blocksize, 1); > > > > if (IS_ERR_OR_NULL(em)) { > > > > SetPageError(page); > > ret = PTR_ERR_OR_ZERO(em); > > > > @@ -3479,7 +3473,7 @@ static noinline_for_stack int > > __extent_writepage_io(struct inode *inode,> > > em_end = extent_map_end(em); > > BUG_ON(em_end <= cur); > > BUG_ON(end < cur); > > > > - iosize = min(em_end - cur, end - cur + 1); > > + iosize = min_t(u64, em_end - cur, blocksize); > > > > iosize = ALIGN(iosize, blocksize); > > This limits us to do one block per loop, if two blocks are contiguous, > it should be fine to write them along. Yes, I agree. I will fix this up in one of the next versions of the patchset. Thanks for pointing it out. > > > sector = (em->block_start + extent_offset) >> 9; > > bdev = em->bdev; > > > > @@ -3488,32 +3482,20 @@ static noinline_for_stack int > > __extent_writepage_io(struct inode *inode,> > > free_extent_map(em); > > em = NULL; > > > > - /* > > - * compressed and inline extents are written through other > > - * paths in the FS > > - */ > > - if (compressed || block_start == EXTENT_MAP_HOLE || > > - block_start == EXTENT_MAP_INLINE) { > > - /* > > - * end_io notification does not happen here for > > - * compressed extents > > - */ > > - if (!compressed && tree->ops && > > - tree->ops->writepage_end_io_hook) > > - tree->ops->writepage_end_io_hook(page, cur, > > - cur + iosize - 1, > > - NULL, 1); > > - else if (compressed) { > > - /* we don't want to end_page_writeback on > > - * a compressed extent. this happens > > - * elsewhere > > - */ > > - nr++; > > - } > > + BUG_ON(compressed); > > + BUG_ON(block_start == EXTENT_MAP_INLINE); > > > > - cur += iosize; > > - pg_offset += iosize; > > - continue; > > + if (block_start == EXTENT_MAP_HOLE) { > > + if (test_page_blks_state(page, BLK_STATE_UPTODATE, cur, > > + cur + iosize - 1, 1)) { > > + clear_page_blks_state(page, > > + 1 << BLK_STATE_DIRTY, cur, > > + cur + iosize - 1); > > + cur += iosize; > > + continue; > > + } else { > > + BUG(); > > + } > > > > } > > > > if (tree->ops && tree->ops->writepage_io_hook) { > > > > @@ -3527,7 +3509,13 @@ static noinline_for_stack int > > __extent_writepage_io(struct inode *inode,> > > } else { > > > > unsigned long max_nr = (i_size >> PAGE_CACHE_SHIFT) + 1; > > > > - set_range_writeback(tree, cur, cur + iosize - 1); > > + clear_page_blks_state(page, 1 << BLK_STATE_DIRTY, cur, > > + cur + iosize - 1); > > + set_page_writeback(page); > > + > > + set_page_blks_state(page, 1 << BLK_STATE_IO, cur, > > + cur + iosize - 1); > > + > > > > if (!PageWriteback(page)) { > > > > btrfs_err(BTRFS_I(inode)->root->fs_info, > > > > "page %lu not writeback, cur %llu end %llu", > > > > @@ -3542,17 +3530,14 @@ static noinline_for_stack int > > __extent_writepage_io(struct inode *inode,> > > if (ret) > > > > SetPageError(page); > > > > } > > > > - cur = cur + iosize; > > - pg_offset += iosize; > > + > > + cur += iosize; > > > > nr++; > > > > } > > > > done: > > *nr_ret = nr; > > > > done_unlocked: > > - > > - /* drop our reference on any cached states */ > > - free_extent_state(cached_state); > > > > return ret; > > > > } > > > > diff --git a/fs/btrfs/file.c b/fs/btrfs/file.c > > index 23b6e03..cbe6381 100644 > > --- a/fs/btrfs/file.c > > +++ b/fs/btrfs/file.c > > @@ -495,6 +495,9 @@ int btrfs_dirty_pages(struct btrfs_root *root, struct > > inode *inode,> > > u64 num_bytes; > > u64 start_pos; > > u64 end_of_last_block; > > > > + u64 start; > > + u64 end; > > + u64 page_end; > > > > u64 end_pos = pos + write_bytes; > > loff_t isize = i_size_read(inode); > > > > @@ -507,11 +510,24 @@ int btrfs_dirty_pages(struct btrfs_root *root, > > struct inode *inode,> > > if (err) > > > > return err; > > > > + start = start_pos; > > + > > > > for (i = 0; i < num_pages; i++) { > > > > struct page *p = pages[i]; > > SetPageUptodate(p); > > ClearPageChecked(p); > > > > + > > + end = page_end = page_offset(p) + PAGE_CACHE_SIZE - 1; > > + > > + if (i == num_pages - 1) > > + end = min_t(u64, page_end, end_of_last_block); > > + > > + set_page_blks_state(p, > > + 1 << BLK_STATE_DIRTY | 1 << BLK_STATE_UPTODATE, > > + start, end); > > > > set_page_dirty(p); > > > > + > > + start = page_end + 1; > > This is not the usual way, page_end is unnecessary, (start += > PAGE_CACHE_SIZE) should work. "start" may not always be set to a file offset that is a multiple of page size. If the userspace dirties say "block 4" of 64k page, then start will be set to 16384. Hence in such cases, "start += PAGE_CACHE_SIZE" would yield an incorrect value. > > } > > > > /* > > > > diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c > > index 8262f83..ac6a3f3 100644 > > --- a/fs/btrfs/inode.c > > +++ b/fs/btrfs/inode.c > > > > @@ -1995,6 +1995,11 @@ again: > > } > > > > btrfs_set_extent_delalloc(inode, page_start, page_end, &cached_state); > > > > + > > + set_page_blks_state(page, > > + 1 << BLK_STATE_DIRTY | 1 << BLK_STATE_UPTODATE, > > + page_start, page_end); > > + > > > > ClearPageChecked(page); > > set_page_dirty(page); > > > > out: > > @@ -2984,26 +2989,48 @@ static int btrfs_writepage_end_io_hook(struct page > > *page, u64 start, u64 end,> > > struct btrfs_ordered_extent *ordered_extent = NULL; > > struct btrfs_workqueue *wq; > > btrfs_work_func_t func; > > > > + u64 ordered_start, ordered_end; > > + int done; > > > > trace_btrfs_writepage_end_io_hook(page, start, end, uptodate); > > > > ClearPagePrivate2(page); > > > > - if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start, > > - end - start + 1, uptodate)) > > - return 0; > > +loop: > > + ordered_extent = btrfs_lookup_ordered_range(inode, start, > > + end - start + 1); > > + if (!ordered_extent) > > + goto out; > > > > - if (btrfs_is_free_space_inode(inode)) { > > - wq = root->fs_info->endio_freespace_worker; > > - func = btrfs_freespace_write_helper; > > - } else { > > - wq = root->fs_info->endio_write_workers; > > - func = btrfs_endio_write_helper; > > + ordered_start = max_t(u64, start, ordered_extent->file_offset); > > + ordered_end = min_t(u64, end, > > + ordered_extent->file_offset + ordered_extent->len - 1); > > + > > + done = btrfs_dec_test_ordered_pending(inode, &ordered_extent, > > + ordered_start, > > + ordered_end - ordered_start + 1, > > + uptodate); > > + if (done) { > > + if (btrfs_is_free_space_inode(inode)) { > > + wq = root->fs_info->endio_freespace_worker; > > + func = btrfs_freespace_write_helper; > > + } else { > > + wq = root->fs_info->endio_write_workers; > > + func = btrfs_endio_write_helper; > > + } > > + > > + btrfs_init_work(&ordered_extent->work, func, > > + finish_ordered_fn, NULL, NULL); > > + btrfs_queue_work(wq, &ordered_extent->work); > > > > } > > > > - btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL, > > - NULL); > > - btrfs_queue_work(wq, &ordered_extent->work); > > + btrfs_put_ordered_extent(ordered_extent); > > + > > + start = ordered_end + 1; > > + > > + if (start < end) > > + goto loop; > > > +out: > I saw this's put a BUG_ON(block_start == EXTENT_MAP_INLINE); in writepage(), > but I didn't see the code of disabling inline data in patch 01 and patch > 02, but anyway I think we can avoid above searching for ordered_extents in > a single page if we enable inline data. For inline extents, The call to __extent_writepage => writepage_delalloc => run_delalloc_range => cow_file_range => cow_file_range_inline should write the block's content into the appropriate location in the btree leaf. Hence __extent_writepage_io() should never get invoked for files with inline extents. The call to BUG_ON(block_start == EXTENT_MAP_INLINE) just makes this explicit and also helps in debugging. Liu, However I am not sure if we could avoid looping across ordered extents in the above code. Could you please elaborate on that? -- chandan -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
