On Sun, 1 Feb 2015, Chris J Arges wrote: > Date: Sun, 01 Feb 2015 15:46:06 -0600 > From: Chris J Arges <chris.j.arges@xxxxxxxxxxxxx> > To: Lukas Czerner <lczerner@xxxxxxxxxx>, Theodore Ts'o <tytso@xxxxxxx>, > Jan Kara <jack@xxxxxxx> > Cc: linux-fsdevel@xxxxxxxxxxxxxxx > Subject: Re: regression in ext4_ind_remove_space > > > > On 01/30/2015 03:14 PM, Chris J Arges wrote: > > > > > > On 01/30/2015 12:56 PM, Chris J Arges wrote: > >> Hi, > >> > >> Users of non-extent ext4 filesystems (ext4 ^extents, or ext3 w/ > >> CONFIG_EXT4_USE_FOR_EXT23=y) can encounter data corruption when using > >> fallocate with FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE flags. This > >> seems to be a regression in ext4_ind_remove_space introduced in > >> 4f579ae7, whereas commit 77ea2a4b passes the following test case. > >> > >> To reproduce this issue do the following: > >> 1) Setup ext4 ^extents, or ext3 filesystem with CONFIG_EXT4_USE_FOR_EXT23=y > >> 2) Create and install a VM using a qcow2 image and store the file on the > >> filesystem > >> 3) Snapshot the image with qemu-img > >> 4) Boot the image and do some disk operations (fio,etc) > >> 5) Shutdown image and delete snapshot > >> 6) Repeat 3-5 until VM no longer boots due to image corruption, > >> generally this takes a few iterations depending on disk operations. > >> > >> In addition, I've tested this with a single vCPU and single host CPU and > >> the problem persists. Running the same test on ext4 w/ extents exhibits > >> no failures. > >> > >> Any ideas for bug hunting here? Commit 4f579ae7 completely re-writes > >> ext4_ind_remove_space. A revert of that commit from master fixes the > >> issue, but I'm unsure if that un-fixes other things. > > > > Well, after a bit more extensive testing a revert of this patch still > > has failures. This may just be a novel bug. > > > > I'll run longer tests on 77ea2a4b as well. > > --chris > > 77ea2a4b eventually fails (~20 iterations) but seems to be a bit more > resilient. So this does not seem like a regression; but still > ext4_ind_remove_space is at the heart of the issue. If I use the > following code snippet below, I can run the test case for days (~400 > iterations so far) without failure: > > diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c > index 5653fa4..e14cdfe 100644 > --- a/fs/ext4/inode.c > +++ b/fs/ext4/inode.c > @@ -3367,6 +3367,10 @@ int ext4_punch_hole(struct inode *inode, loff_t > offset, loff_t length) > unsigned int credits; > int ret = 0; > > + /* EXTENTS required */ > + if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) > + return -EOPNOTSUPP; > + > if (!S_ISREG(inode->i_mode)) > return -EOPNOTSUPP; Hi Chris, unfortunately you're right. I've noticed this recently as well and I am looking into this, however the indirect block handling code is damn complicated (but maybe it's just me) and I was not able to find the root cause of the problem yet. Btw, this can also be very easily reproduced by running xfstest generic/270 on 1k file system. That way I am able to reproduce the problem in more than 90% runs. Now since Al Viro was the one who wrote most of the indirect code and since he is much smarter than me, maybe we can kindly ask him to take a look at the ext4_ind_remove_space() to figure out where the problem is ? Al, could you take a quick look at this ? Thanks! -Lukas > > > --chris > > > > > I'm happy to > >> continue debugging, or run any tests as necessary. > >> > >> Thanks, > >> --chris j arges > >> > -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
