From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from smtp.kernel.org (aws-us-west-2-korg-mail-alma10-1.taild15c8.ts.net [100.103.45.18]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.subspace.kernel.org (Postfix) with ESMTPS id D12A3409635 for ; Thu, 27 Aug 2026 08:56:59 +0000 (UTC) Authentication-Results: smtp.subspace.kernel.org; arc=none smtp.client-ip=100.103.45.18 ARC-Seal:i=1; a=rsa-sha256; d=subspace.kernel.org; s=arc-20240116; t=1787821024; cv=none; b=WuEM14H1FUENTJOYLmLpQ+uZL6ac7ZPX5HlnRWWrWP1fVNq84x8ubNF4/XQ3wdqgkZdEeHOSoCeCaXGfsVWbpeKuSa21T0TNWyo8/7rDCAJTWDp9UWKzcqoQEtZKU5KLF+6SF0q95sXS8MBuI64ansj+HPYWZ01eXdmDKKZzrGY= ARC-Message-Signature:i=1; a=rsa-sha256; d=subspace.kernel.org; s=arc-20240116; t=1787821024; c=relaxed/simple; bh=8GLsPyCYSReJXtFnw/pED7QoBc7HdU/pN0zL/QOQEEM=; h=Message-ID:Date:MIME-Version:Cc:Subject:To:References:From: In-Reply-To:Content-Type; b=bLrK56gwA5+dCj9eQm9eqz8gh2tgyuvuTh2fiD3ux+KK4QW2m/m7JaK86F+62SFsUqE+uDcTFXIDuO3KRdvsqaKU/4/Q5+tLRpQlSsIolRTkVPzxHmx5Gm2ru0x5kuZYma+PuZnb0cGeGBArWdCu3LJR/6RTjq5SJwfyK85JyqM= ARC-Authentication-Results:i=1; smtp.subspace.kernel.org; dkim=pass (2048-bit key) header.d=kernel.org header.i=@kernel.org header.b=gp0g68Aa; arc=none smtp.client-ip=100.103.45.18 Authentication-Results: smtp.subspace.kernel.org; dkim=pass (2048-bit key) header.d=kernel.org header.i=@kernel.org header.b="gp0g68Aa" Received: by smtp.kernel.org (Postfix) with ESMTPSA id B202C1F000E9; Thu, 27 Aug 2026 08:56:55 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=kernel.org; s=k20260515; t=1787821018; bh=EOEJ4CK0mQUXac3rydzccVHddtIdGliR8fWdejy4oaI=; h=Date:Cc:Subject:To:References:From:In-Reply-To; b=gp0g68AaIhrthBM/KDS6XlNpnTpE4gg4BJ4XIiKjXsgCrxbNUGE6QJghvBBsmzlG6 qoW6PB6d1GHchgmay9SK2R8tPi9G0TyKesB8F1IXgbFoQgloXt23qtXgnLtHsXt3Qg xz6fWhjyxgMD+z7gu6ylS9AsHJZkqihljdLuofpByrylKpfEC1W2KBRfzc5EUU1ExQ t7cHAT//9DFfi87ZlPiOSu2iEE1m8UZ1yuD6oypN11ayv+ZWXWYBN8yQMwx0zC99Ip 3azGSVBIGcpcsRAJMuGXO0TyiAH12RoNOnvxE/YeqlF+MmzQ/p0GCh2upZTEKl4ATr iwjzEOSlYBeZw== Message-ID: <83fe187e-e3e8-4328-bbbf-a6751873b9e0@kernel.org> Date: Thu, 27 Aug 2026 16:56:53 +0800 Precedence: bulk X-Mailing-List: linux-kernel@vger.kernel.org List-Id: List-Subscribe: List-Unsubscribe: MIME-Version: 1.0 User-Agent: Mozilla Thunderbird Cc: chao@kernel.org, Barry Song , Nanzhe Zhao , Juan Yescas , Dev Jain , linux-kernel@vger.kernel.org, David Hildenbrand , Bo Zhang , Kalesh Singh , Ryan Roberts , Pengfei Li Subject: Re: [PATCH 03/14] f2fs: support regular file buffered writes on large folios To: Nanzhe Zhao , linux-f2fs-devel@lists.sourceforge.net, Jaegeuk Kim References: <20260826082641.2007658-1-zhaonanzhe@xiaomi.com> <20260826082641.2007658-4-zhaonanzhe@xiaomi.com> Content-Language: en-US From: Chao Yu In-Reply-To: <20260826082641.2007658-4-zhaonanzhe@xiaomi.com> Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 7bit On 8/26/26 16:26, Nanzhe Zhao wrote: > To avoid the complexity of unlocking a large folio in write_begin, > preallocate partial blocks for inodes that can use large folios. > During write_begin, read only the partial head and tail 4K subpages > that need read-before-write, and skip read I/O for the full middle > subpages covered by the write. > > Signed-off-by: Nanzhe Zhao > --- > fs/f2fs/data.c | 289 +++++++++++++++++++++++++++++++++++++++++++++++-- > fs/f2fs/f2fs.h | 1 + > fs/f2fs/file.c | 17 ++- > 3 files changed, 293 insertions(+), 14 deletions(-) > > diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c > index 8649d443ca74..7ce96ae02cfc 100644 > --- a/fs/f2fs/data.c > +++ b/fs/f2fs/data.c > @@ -1249,6 +1249,40 @@ static void f2fs_submit_page_read(struct inode *inode, struct fsverity_info *vi, > f2fs_submit_read_bio(sbi, bio, DATA); > } > > +/* > + * Synchronously read a single 4K subpage by reusing f2fs_submit_page_read() > + * so that iostat, trace, blk-crypto and post-read handling are all preserved. > + * The caller must have already allocated ffs for the folio. > + */ > +static int f2fs_submit_page_read_sync(struct inode *inode, struct folio *folio, > + pgoff_t index, block_t blkaddr) > +{ > + struct f2fs_folio_state *ffs = folio->private; > + unsigned long flags; > + > + /* Add bias so end_io does not call folio_end_read(). */ > + spin_lock_irqsave(&ffs->state_lock, flags); > + ffs->read_pages_pending++; > + spin_unlock_irqrestore(&ffs->state_lock, flags); > + > + f2fs_submit_page_read(inode, NULL, folio, index, blkaddr, > + REQ_OP_READ, false); > + > + /* Wait for pending drops back to bias which indicates all bio have completed. */ > + while (READ_ONCE(ffs->read_pages_pending) != 1) > + f2fs_io_schedule_timeout(DEFAULT_SCHEDULE_TIMEOUT); Suspect there will be potential performance issue, as timeout interval is large. > + > + /* Remove the bias. */ > + spin_lock_irqsave(&ffs->state_lock, flags); > + ffs->read_pages_pending--; > + spin_unlock_irqrestore(&ffs->state_lock, flags); Looks hacky. If large folio has 4 sub pages, first 3 pages are uptodate. Then we use f2fs_submit_page_read_sync() to read last page, in f2fs_finish_read_bio(), we update state bitmap of last page to 1, but as read_pages_pending is 1, so folio won't be set as uptodate. Oh, in f2fs_write_end(), we have another change to set folio uptodate though, so the flow is a little bit different here. > + > + if (!f2fs_ffs_test_blk_uptodate(folio, index)) > + return -EIO; > + > + return 0; > +} > + > static void __set_data_blkaddr(struct dnode_of_data *dn, block_t blkaddr) > { > __le32 *addr = get_dnode_addr(dn->inode, dn->node_folio); > @@ -2582,6 +2616,112 @@ static void ffs_detach_free(struct folio *folio) > kfree(ffs); > } > > +bool f2fs_ffs_test_blk_uptodate(const struct folio *folio, pgoff_t index) > +{ > + struct f2fs_folio_state *ffs; > + unsigned int idx; > + > + if (!f2fs_folio_has_ffs(folio)) > + return folio_test_uptodate(folio); > + > + if (!folio_contains(folio, index)) > + return false; Can we check this out of f2fs_ffs_test_blk_uptodate()? and add a sanity check here: f2fs_bug_on(sbi, !folio_contains(folio, index)); > + > + ffs = (struct f2fs_folio_state *)folio->private; > + idx = index - folio->index; > + return test_bit(idx, ffs->state); > +} > + > +static bool __ffs_mark_subrange_uptodate(struct folio *folio, > + struct f2fs_folio_state *ffs, size_t offset, size_t len) > +{ > + unsigned int nr_subpages = folio_nr_pages(folio); > + unsigned int start, end; > + > + start = offset >> PAGE_SHIFT; > + end = (offset + len + PAGE_SIZE - 1) >> PAGE_SHIFT; > + end = min(end, nr_subpages); > + > + bitmap_set(ffs->state, start, end - start); > + return bitmap_full(ffs->state, nr_subpages); > +} > + > +static void f2fs_ffs_mark_subrange_uptodate(struct folio *folio, size_t offset, > + size_t len) > +{ > + struct f2fs_folio_state *ffs; > + unsigned long flags; > + bool mark_uptodate = false; > + > + f2fs_bug_on(F2FS_F_SB(folio), offset + len > folio_size(folio)); > + > + if (!f2fs_folio_has_ffs(folio)) { > + folio_mark_uptodate(folio); > + return; > + } > + > + ffs = (struct f2fs_folio_state *)folio->private; > + spin_lock_irqsave(&ffs->state_lock, flags); > + mark_uptodate = __ffs_mark_subrange_uptodate(folio, ffs, offset, len) && > + !ffs->read_pages_pending; > + spin_unlock_irqrestore(&ffs->state_lock, flags); > + if (mark_uptodate) > + folio_mark_uptodate(folio); > +} > + > +static void f2fs_ffs_mark_subrange_dirty(struct folio *folio, > + size_t offset, size_t len) > +{ > + struct f2fs_folio_state *ffs; > + unsigned int nr_subpages, start, end; > + unsigned long flags; > + > + f2fs_bug_on(F2FS_F_SB(folio), offset + len > folio_size(folio)); > + > + if (!f2fs_folio_has_ffs(folio)) > + return; > + > + ffs = (struct f2fs_folio_state *)folio->private; > + nr_subpages = folio_nr_pages(folio); > + start = offset >> PAGE_SHIFT; > + end = (offset + len + PAGE_SIZE - 1) >> PAGE_SHIFT; > + end = min(end, nr_subpages); > + > + spin_lock_irqsave(&ffs->state_lock, flags); > + bitmap_set(ffs->state, nr_subpages + start, end - start); > + spin_unlock_irqrestore(&ffs->state_lock, flags); > +} > + > +static bool find_next_valid_block(const struct folio *folio, > + size_t orig_off, size_t *need_off, > + size_t len) > +{ > + size_t start = orig_off; > + size_t end = start + len; > + size_t head, tail; > + pgoff_t index; > + > + if (start & (PAGE_SIZE - 1)) { > + head = round_down(start, PAGE_SIZE); > + index = folio->index + (head >> PAGE_SHIFT); > + if (!f2fs_ffs_test_blk_uptodate(folio, index)) { > + *need_off = head; > + return true; > + } > + } > + > + if (end & (PAGE_SIZE - 1)) { > + tail = round_down(end - 1, PAGE_SIZE); > + index = folio->index + (tail >> PAGE_SHIFT); > + if (!f2fs_ffs_test_blk_uptodate(folio, index)) { > + *need_off = tail; > + return true; > + } > + } > + > + return false; > +} > + > static int f2fs_read_data_large_folio(struct inode *inode, > struct fsverity_info *vi, > struct readahead_control *rac, struct folio *folio) > @@ -3975,6 +4115,102 @@ static int prepare_atomic_write_begin(struct f2fs_sb_info *sbi, > return 0; > } > > +static int prepare_large_folio_write_begin(struct inode *inode, > + struct folio *folio, loff_t pos, > + unsigned int len) > +{ > + struct f2fs_sb_info *sbi = F2FS_I_SB(inode); > + struct f2fs_folio_state *ffs; > + size_t ori_off = offset_in_folio(folio, pos); > + size_t need_off = ori_off; > + int err = 0; > + > + len = min_t(unsigned int, len, folio_size(folio) - ori_off); > + > + /* > + * When folio minimum order is non-zero, the fsverity > + * page_cache_write() path enters f2fs_write_begin() via > + * aops->write_begin without going through f2fs_write_iter(), > + * so preallocation from f2fs_write_iter() is skipped. In that > + * case, if FI_PREALLOCATED_ALL is not set, we must preallocate > + * the write blocks here. > + */ > + if (!is_inode_flag_set(inode, FI_PREALLOCATED_ALL)) { > + struct f2fs_map_blocks map = {}; > + > + map.m_lblk = F2FS_BYTES_TO_BLK(pos); > + map.m_len = F2FS_BLK_ALIGN(pos + len) - map.m_lblk; > + > + if (!IS_DEVICE_ALIASING(inode)) > + map.m_may_create = true; > + map.m_seg_type = NO_CHECK_TYPE; > + > + err = f2fs_map_blocks(inode, &map, F2FS_GET_BLOCK_PRE_AIO); > + if (err) > + return err; > + } > + > + /* > + * Skip read and ffs allocation if the folio is already fully > + * uptodate or the write covers the entire folio. > + */ > + if (folio_test_uptodate(folio) || len == folio_size(folio)) > + return 0; Do we need to adopt fsverity like we did in f2fs_write_begin()? if (!(pos & (PAGE_SIZE - 1)) && (pos + len) >= i_size_read(inode) && !f2fs_verity_in_progress(inode)) { folio_zero_segment(folio, len, folio_size(folio)); return 0; } > + > + ffs = f2fs_ffs_find_or_alloc(folio); f2fs_ffs_find_or_alloc() will never fail. > + if (!ffs) > + return -ENOMEM; > + > + /* Inline data must have been converted before reaching here. */ > + f2fs_bug_on(sbi, f2fs_has_inline_data(inode)); > + > + while (find_next_valid_block(folio, ori_off, > + &need_off, len)) { > + struct dnode_of_data dn; > + pgoff_t index = folio->index + (need_off >> PAGE_SHIFT); > + block_t blkaddr; > + bool get_dn = false; > + > + if (!f2fs_lookup_read_extent_cache_block(inode, index, > + &blkaddr)) { > + if (IS_DEVICE_ALIASING(inode)) > + return -ENODATA; > + > + set_new_dnode(&dn, inode, NULL, NULL, 0); > + err = f2fs_get_dnode_of_data(&dn, index, LOOKUP_NODE); > + if (err) > + return err; > + get_dn = true; > + blkaddr = dn.data_blkaddr; > + > + if (blkaddr == NEW_ADDR) { > + size_t off = offset_in_folio(folio, > + index << PAGE_SHIFT); > + > + folio_zero_segment(folio, off, off + PAGE_SIZE); > + f2fs_ffs_mark_subrange_uptodate(folio, off, > + PAGE_SIZE); > + goto out; > + } > + > + if (!f2fs_is_valid_blkaddr(sbi, blkaddr, > + DATA_GENERIC_ENHANCE_READ)) { > + err = -EFSCORRUPTED; > + goto out; > + } > + } > + > + err = f2fs_submit_page_read_sync(inode, folio, index, blkaddr); > +out: > + if (get_dn) > + f2fs_put_dnode(&dn); > + if (err) > + return err; > + } > + > + return 0; > +} > + > static int f2fs_write_begin(const struct kiocb *iocb, > struct address_space *mapping, > loff_t pos, unsigned len, struct folio **foliop, > @@ -3986,6 +4222,7 @@ static int f2fs_write_begin(const struct kiocb *iocb, > pgoff_t index = pos >> PAGE_SHIFT; > bool need_balance = false; > block_t blkaddr = NULL_ADDR; > + fgf_t fgp = FGP_LOCK | FGP_WRITE | FGP_CREAT; > int err = 0; > > trace_f2fs_write_begin(inode, pos, len); > @@ -4033,9 +4270,9 @@ static int f2fs_write_begin(const struct kiocb *iocb, > * Do not use FGP_STABLE to avoid deadlock. > * Will wait that below with our IO control. > */ > - folio = f2fs_filemap_get_folio(mapping, index, > - FGP_LOCK | FGP_WRITE | FGP_CREAT, > - mapping_gfp_mask(mapping)); > + fgp |= fgf_set_order(len); > + folio = __filemap_get_folio(mapping, index, fgp, > + mapping_gfp_mask(mapping)); It detaches from fault injection, why we need to change to use __filemap_get_folio()? > if (IS_ERR(folio)) { > err = PTR_ERR(folio); > goto fail; > @@ -4048,7 +4285,7 @@ static int f2fs_write_begin(const struct kiocb *iocb, > if (f2fs_is_atomic_file(inode)) > err = prepare_atomic_write_begin(sbi, folio, pos, len, > &blkaddr, &need_balance); > - else > + else if (!folio_test_large(folio)) > err = prepare_write_begin(sbi, folio, pos, len, > &blkaddr, &need_balance); > if (err) > @@ -4069,6 +4306,14 @@ static int f2fs_write_begin(const struct kiocb *iocb, > > f2fs_folio_wait_writeback(folio, DATA, false, true); > > + if (folio_test_large(folio)) { > + err = prepare_large_folio_write_begin(inode, > + folio, pos, len); > + if (!err) > + return 0; > + goto put_folio; > + } > + > if (len == folio_size(folio) || folio_test_uptodate(folio)) > return 0; > > @@ -4129,15 +4374,20 @@ static int f2fs_write_end(const struct kiocb *iocb, > trace_f2fs_write_end(inode, pos, len, copied); > > /* > - * This should be come from len == PAGE_SIZE, and we expect copied > - * should be PAGE_SIZE. Otherwise, we treat it with zero copied and > - * let generic_perform_write() try to copy data again through copied=0. > + * If a short copy happens on a folio that isn't uptodate, we treat > + * it with zero copied and let generic_perform_write() try to copy > + * data again through copied=0. > */ > if (!folio_test_uptodate(folio)) { > - if (unlikely(copied != len)) > + if (unlikely(copied != len)) { > copied = 0; > - else > + } else if (folio_test_large(folio)) { > + f2fs_ffs_mark_subrange_uptodate(folio, > + offset_in_folio(folio, pos), len); > + } else { > + /* This should be come from len == PAGE_SIZE */ > folio_mark_uptodate(folio); > + } > } > > #ifdef CONFIG_F2FS_FS_COMPRESSION > @@ -4156,6 +4406,9 @@ static int f2fs_write_end(const struct kiocb *iocb, > if (!copied) > goto unlock_out; > > + if (folio_test_large(folio)) > + f2fs_ffs_mark_subrange_dirty(folio, offset_in_folio(folio, pos), > + copied); > folio_mark_dirty(folio); > > if (f2fs_is_atomic_file(inode)) > @@ -4218,8 +4471,22 @@ static bool f2fs_dirty_data_folio(struct address_space *mapping, > > trace_f2fs_set_page_dirty(folio, DATA); > > - if (!folio_test_uptodate(folio)) > - folio_mark_uptodate(folio); > + if (!folio_test_uptodate(folio)) { > + bool uptodate = true; > + > + if (f2fs_folio_has_ffs(folio)) { > + struct f2fs_folio_state *ffs = > + (struct f2fs_folio_state *)folio->private; > + unsigned long flags; > + > + spin_lock_irqsave(&ffs->state_lock, flags); > + uptodate = bitmap_full(ffs->state, folio_nr_pages(folio)) && > + !ffs->read_pages_pending; > + spin_unlock_irqrestore(&ffs->state_lock, flags); f2fs_ffs_mark_subrange_uptodate(folio, 0, folio_size(folio)); ? > + } > + if (uptodate) > + folio_mark_uptodate(folio); > + } > BUG_ON(folio_test_swapcache(folio)); > > if (filemap_dirty_folio(mapping, folio)) { > diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h > index 31ddd321daf6..ae6031693700 100644 > --- a/fs/f2fs/f2fs.h > +++ b/fs/f2fs/f2fs.h > @@ -4342,6 +4342,7 @@ int f2fs_write_single_data_page(struct folio *folio, int *submitted, > struct writeback_control *wbc, > enum iostat_type io_type, > int compr_blocks, bool allow_balance); > +bool f2fs_ffs_test_blk_uptodate(const struct folio *folio, pgoff_t index); > void f2fs_write_failed(struct inode *inode, loff_t to); > void f2fs_invalidate_folio(struct folio *folio, size_t offset, size_t length); > bool f2fs_release_folio(struct folio *folio, gfp_t wait); > diff --git a/fs/f2fs/file.c b/fs/f2fs/file.c > index d440231b8cb9..e40285d4f9c2 100644 > --- a/fs/f2fs/file.c > +++ b/fs/f2fs/file.c > @@ -5405,9 +5405,20 @@ static int f2fs_preallocate_blocks(struct kiocb *iocb, struct iov_iter *iter, > return ret; > } > > - /* Do not preallocate blocks that will be written partially in 4KB. */ > - map.m_lblk = F2FS_BLK_ALIGN(pos); > - map.m_len = F2FS_BYTES_TO_BLK(pos + count); > + if (mapping_large_folio_support(inode->i_mapping)) { > + /* > + * Preallocate all blocks touched by a large-folio buffered write so > + * the regular write_begin path does not need to unlock the folio for > + * f2fs_balance_fs(). Rechecking large-folio state after unlock is > + * unreliable since partial truncation can split the folio. You mean buffered io case, right? so we need to check !dio as well? Thanks, > + */ > + map.m_lblk = F2FS_BYTES_TO_BLK(pos); > + map.m_len = F2FS_BLK_ALIGN(pos + count); > + } else { > + /* Do not preallocate blocks that will be written partially in 4KB. */ > + map.m_lblk = F2FS_BLK_ALIGN(pos); > + map.m_len = F2FS_BYTES_TO_BLK(pos + count); > + } > if (map.m_len > map.m_lblk) > map.m_len -= map.m_lblk; > else