From: Eric Sandeen Subject: Re: [PATCH] ext4: fix extent tree corruption that incurred by hole punch Date: Thu, 06 Dec 2012 09:45:38 -0600 Message-ID: <50C0BDA2.4090203@redhat.com> References: <1354623069-8124-1-git-send-email-forrestl@synology.com> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1 Content-Transfer-Encoding: 7bit Cc: "Theodore Ts'o" , ext4 development To: Forrest Liu Return-path: Received: from mx1.redhat.com ([209.132.183.28]:14750 "EHLO mx1.redhat.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1422692Ab2LFPpd (ORCPT ); Thu, 6 Dec 2012 10:45:33 -0500 In-Reply-To: <1354623069-8124-1-git-send-email-forrestl@synology.com> Sender: linux-ext4-owner@vger.kernel.org List-ID: On 12/4/12 6:11 AM, Forrest Liu wrote: > Extent indexes didn't update correctly in ext4_ext_rm_idx, when depth > of extent tree is greater than 1. This is interesting; we had 2 reports of similar corruption on the list, I wonder if the application in question was doing hole punching. I didn't expect that they were, so TBH I was pretty much ignoring the hole-punch cases for parent index updates. Hm. I'll have to look into that. Could you turn your testcase into an xfstest regression test? -Eric > Signed-off-by: Forrest Liu > --- > fs/ext4/extents.c | 24 ++++++++++++++++++++---- > 1 files changed, 20 insertions(+), 4 deletions(-) > > diff --git a/fs/ext4/extents.c b/fs/ext4/extents.c > index d3dd618..b10b8c0 100644 > --- a/fs/ext4/extents.c > +++ b/fs/ext4/extents.c > @@ -2190,13 +2190,15 @@ errout: > * removes index from the index block. > */ > static int ext4_ext_rm_idx(handle_t *handle, struct inode *inode, > - struct ext4_ext_path *path) > + struct ext4_ext_path *path, int depth) > { > int err; > ext4_fsblk_t leaf; > + __le32 border; > > /* free index block */ > - path--; > + depth--; > + path = path + depth; > leaf = ext4_idx_pblock(path->p_idx); > if (unlikely(path->p_hdr->eh_entries == 0)) { > EXT4_ERROR_INODE(inode, "path->p_hdr->eh_entries == 0"); > @@ -2221,6 +2223,20 @@ static int ext4_ext_rm_idx(handle_t *handle, struct inode *inode, > > ext4_free_blocks(handle, inode, NULL, leaf, 1, > EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET); > + > + border = path->p_idx->ei_block; > + while (--depth >= 0) { > + if (path->p_idx != EXT_FIRST_INDEX(path->p_hdr)) > + break; > + path--; > + err = ext4_ext_get_access(handle, inode, path); > + if (err) > + break; > + path->p_idx->ei_block = border; > + err = ext4_ext_dirty(handle, inode, path); > + if (err) > + break; > + } > return err; > } > > @@ -2557,7 +2573,7 @@ ext4_ext_rm_leaf(handle_t *handle, struct inode *inode, > /* if this leaf is free, then we should > * remove it from index block above */ > if (err == 0 && eh->eh_entries == 0 && path[depth].p_bh != NULL) > - err = ext4_ext_rm_idx(handle, inode, path + depth); > + err = ext4_ext_rm_idx(handle, inode, path, depth); > > out: > return err; > @@ -2760,7 +2776,7 @@ again: > /* index is empty, remove it; > * handle must be already prepared by the > * truncatei_leaf() */ > - err = ext4_ext_rm_idx(handle, inode, path + i); > + err = ext4_ext_rm_idx(handle, inode, path, i); > } > /* root level has p_bh == NULL, brelse() eats this */ > brelse(path[i].p_bh); >