From: Surbhi Palande Subject: Re: [RFC][PATCH] Re: [BUG] ext4: cannot unfreeze a filesystem due to a deadlock Date: Tue, 03 May 2011 10:27:00 +0300 Message-ID: <4DBFAE44.2080000@ubuntu.com> References: <20110331234050.GD2904@dastard> <20110401140856.GA5311@quack.suse.cz> <20110406054005.GD31057@dastard> <20110406061856.GC23285@quack.suse.cz> <20110406112135.GE31057@dastard> <4DBE746F.3090707@ubuntu.com> <20110502105629.GA4556@quack.suse.cz> <4DBE9537.4050708@ubuntu.com> <20110502122055.GB5855@quack.suse.cz> <4DBEA3DF.1060306@ubuntu.com> <20110502131619.GC5855@quack.suse.cz> <4DBEB025.90704@ubuntu.com> <4DBEB9E7.9070407@redhat.com> Reply-To: surbhi.palande@ubuntu.com Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: 7bit Cc: Jan Kara , Dave Chinner , Toshiyuki Okajima , Ted Ts'o , Masayoshi MIZUMA , Andreas Dilger , linux-ext4@vger.kernel.org, linux-fsdevel@vger.kernel.org, Christoph Hellwig To: Eric Sandeen Return-path: Received: from adelie.canonical.com ([91.189.90.139]:44492 "EHLO adelie.canonical.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1750907Ab1ECH1T (ORCPT ); Tue, 3 May 2011 03:27:19 -0400 In-Reply-To: <4DBEB9E7.9070407@redhat.com> Sender: linux-ext4-owner@vger.kernel.org List-ID: On 05/02/2011 05:04 PM, Eric Sandeen wrote: > On 5/2/11 8:22 AM, Surbhi Palande wrote: >> On 05/02/2011 04:16 PM, Jan Kara wrote: >>> On Mon 02-05-11 15:30:23, Surbhi Palande wrote: >>>> On 05/02/2011 03:20 PM, Jan Kara wrote: >>>>> On Mon 02-05-11 14:27:51, Surbhi Palande wrote: >>>>>> On 05/02/2011 01:56 PM, Jan Kara wrote: >>>>>>> On Mon 02-05-11 12:07:59, Surbhi Palande wrote: >>>>>>>> On 04/06/2011 02:21 PM, Dave Chinner wrote: >>>>>>>>> On Wed, Apr 06, 2011 at 08:18:56AM +0200, Jan Kara wrote: >>>>>>>>>> On Wed 06-04-11 15:40:05, Dave Chinner wrote: >>>>>>>>>>> On Fri, Apr 01, 2011 at 04:08:56PM +0200, Jan Kara wrote: >>>>>>>>>>>> On Fri 01-04-11 10:40:50, Dave Chinner wrote: >>>>>>>>>>>>> If you don't allow the page to be dirtied in the fist place, then >>>>>>>>>>>>> nothing needs to be done to the writeback path because there is >>>>>>>>>>>>> nothing dirty for it to write back. >>>>>>>>>>>> Sure but that's only the problem he was able to hit. But generally, >>>>>>>>>>>> there's a problem with needing s_umount for unfreezing because it isn't >>>>>>>>>>>> clear there aren't other code paths which can block with s_umount held >>>>>>>>>>>> waiting for fs to get unfrozen. And these code paths would cause the same >>>>>>>>>>>> deadlock. That's why I chose to get rid of s_umount during thawing. >>>>>>>>>>> Holding the s_umount lock while checking if frozen and sleeping >>>>>>>>>>> is essentially an ABBA lock inversion bug that can bite in many more >>>>>>>>>>> places that just thawing the filesystem. Any where this is done should >>>>>>>>>>> be fixed, so I don't think just removing the s_umount lock from the thaw >>>>>>>>>>> path is sufficient to avoid problems. >>>>>>>>>> That's easily said but hard to do - any transaction start in ext3/4 may >>>>>>>>>> block on filesystem being frozen (this seems to be similar for XFS as I'm >>>>>>>>>> looking into the code) and transaction start traditionally nests inside >>>>>>>>>> s_umount (and basically there's no way around that since sync() calls your >>>>>>>>>> fs code with s_umount held). >>>>>>>>> Sure, but the question must be asked - why is ext3/4 even starting a >>>>>>>>> transaction on a clean filesystem during sync? A frozen filesystem, >>>>>>>>> by definition, is a clean filesytem, and therefore sync calls of any >>>>>>>>> kind should not be trying to write to the FS or start transactions. >>>>>>>>> XFS does this just fine, so I'd consider such behaviour on a frozen >>>>>>>>> filesystem a bug in ext3/4... >>>>>>>> I had a look at the xfs code for seeing how this is done. >>>>>>>> xfs_file_aio_write() >>>>>>>> xfs_wait_for_freeze() >>>>>>>> vfs_check_frozen() >>>>>>>> So xfs_file_aio_write() writes to buffers when the FS is not frozen. >>>>>>>> >>>>>>>> Now, I want to know what stops the following scenario from happening: >>>>>>>> -------------------- >>>>>>>> xfs_file_aio_write() >>>>>>>> xfs_wait_for_freeze() >>>>>>>> vfs_check_frozen() >>>>>>>> At this point F.S was not frozen, so the next instruction in the >>>>>>>> xfs_file_aio_write() will be executed next. >>>>>>>> However at this point (i.e after checking if F.S is frozen) the >>>>>>>> write process gets pre-empted and say the _freeze_ process gets >>>>>>>> control. >>>>>>>> >>>>>>>> Now the F.S freezes and the write process gets the control back. And >>>>>>>> so we end up writing to the page cache when the F.S is frozen. >>>>>>>> -------------------- >>>>>>>> >>>>>>>> Can anyone please enlighten me on how& why this premption is _not_ >>>>>>>> possible? >>>>>> Thanks for your reply. >>>>>>> XFS works similarly as ext4 in this regard I believe. They have the log >>>>>>> frozen in xfs_freeze() so if the race you describe above happens, either >>>>>>> the writing process gets caught waiting for log to unfreeze >>>>>> Agreed. >>>>>>> or it manages >>>>>>> to start a transaction and then freezing process waits for transaction to >>>>>>> finish before it can proceed with freezing. I'm not sure why is there the >>>>>>> check in xfs_file_aio_write()... >>>>>>> >>>>>>> >>>>>> I am sorry, but I don't understand how this will happen - i.e I >>>>>> can't understand what stops freeze_super() (or ext4_freeze) from >>>>>> freezing a superblock (as the write process stopped just before >>>>>> writing anything for this transaction and has not taken any locks?) >>>>> So ext4_freeze() does >>>>> jbd2_journal_lock_updates(journal) >>>>> which waits for all running transactions to finish and updates >>>>> j_barrier_count which stops any news ones from proceeding (check >>>>> function start_this_handle()). >>>>> >>>> Yes, but ext4_freeze() also calls >>>> jbd2_journal_unlock_updates(journal) which decrements the >>>> j_barrier_count (which was previously updated/incremented in >>>> jbd2_journal_lock_updates) ? before it returns. So after this call a >>>> new transaction/handle can be accepted/started. >>>> >>>> A comment in ext4_freeze() says: >>>> /* we rely on s_frozen to stop further updates */ >>>> (before calling jbd2_journal_unlock_updates()) >>> Ah, drat, you're right. I've missed this other part. It's the problem >>> that if you expect to see something, you'll see it regardless of the real >>> code ;). >>> >>> The fact is we do vfs_check_frozen() in ext4_journal_start_sb() but indeed >>> it's still racy (although the race window is relatively small) because the >>> filesystem can become frozen the instant after we check vfs_check_frozen(). >>> Commit 6b0310fb broke it for ext4. >>> >>> I guess the code was mostly copied from XFS which seems to have the same >>> problem in xfs_trans_alloc() since the git history beginning. I see two >>> ways to fix this - either fix ext4/xfs to check s_frozen after starting >>> a transaction and if the filesystem is being frozen, we stop the >>> transaction, wait for fs to get unfrozen, and restart. Another option is >>> to create an analogous logic using a atomic counter of write ops in vfs >>> that could be used by all filesystems. We'd just have to replace >>> vfs_check_frozen() with vfs_start_write() and add vfs_stop_write() at >>> appropriate places... >> How about calling jbd2_journal_unlock_updates(EXT4_SB(sb)->s_journal); >> from ext4_unfreeze()? > we used to have that, but holding it locked until then means we exit the kernel > with a mutex held, which is pretty icky. > > ================================================ > [ BUG: lock held when returning to user space! ] > ------------------------------------------------ > lvcreate/1075 is leaving the kernel with locks still held! > 1 lock held by lvcreate/1075: > #0: (&journal->j_barrier){+.+...}, at: [] > jbd2_journal_lock_updates+0xe1/0xf0 > > > -Eric Should this not be reverted? I think that its a lot easier to stop a transaction between a freeze and a thaw that way! If you agree, can I send a patch for the same? Thanks! Warm Regards, Surbhi. >> So that indeed no transactions can be started before unfreeze is called. >> >> This has another advantage, that it rightfully does not let you update the access time when the F.S is frozen (touch_atime called from a read path when the F.S is frozen) Otherwise we also need to fix this path. >> >> Warm Regards, >> Surbhi. >> >>> Dave, Christoph, any opinions on this? >>> Honza >> -- >> To unsubscribe from this list: send the line "unsubscribe linux-ext4" in >> the body of a message to majordomo@vger.kernel.org >> More majordomo info at http://vger.kernel.org/majordomo-info.html > -- > To unsubscribe from this list: send the line "unsubscribe linux-ext4" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html