From: Eric Sandeen <sandeen@redhat.com>
Subject: Re: fstrim has no effect on a just-mounted filesystem
Date: Tue, 11 Mar 2014 17:08:19 -0500
Message-ID: <531F8953.1030702@redhat.com>
References: <20140311213932.GA19176@redhat.com> <531F8456.2020404@redhat.com> <20140311220013.GV1346@redhat.com>
Mime-Version: 1.0
Content-Type: text/plain; charset=ISO-8859-1
Content-Transfer-Encoding: 7bit
Cc: linux-ext4@vger.kernel.org
To: "Richard W.M. Jones" <rjones@redhat.com>
Return-path: <linux-ext4-owner@vger.kernel.org>
Received: from mx1.redhat.com ([209.132.183.28]:10933 "EHLO mx1.redhat.com"
	rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP
	id S1754433AbaCKWIV (ORCPT <rfc822;linux-ext4@vger.kernel.org>);
	Tue, 11 Mar 2014 18:08:21 -0400
Received: from int-mx02.intmail.prod.int.phx2.redhat.com (int-mx02.intmail.prod.int.phx2.redhat.com [10.5.11.12])
	by mx1.redhat.com (8.14.4/8.14.4) with ESMTP id s2BM8K38014606
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=OK)
	for <linux-ext4@vger.kernel.org>; Tue, 11 Mar 2014 18:08:21 -0400
In-Reply-To: <20140311220013.GV1346@redhat.com>
Sender: linux-ext4-owner@vger.kernel.org
List-ID: <linux-ext4.vger.kernel.org>

On 3/11/14, 5:00 PM, Richard W.M. Jones wrote:
> [The context of this is trying to get virt-sparsify to work in place
> on disks.]
> 
> On Tue, Mar 11, 2014 at 04:47:02PM -0500, Eric Sandeen wrote:
>> On 3/11/14, 4:39 PM, Richard W.M. Jones wrote:
>>>
>>> Here's a problem I can't work out:
>>>
>>> I have a filesystem (in a VM) that I know has at least 100MB of
>>> deleted files on it.
>>
>> Was it mounted with -o discard at the time the files were deleted?
> 
> No, it was not.

Ok, worth asking.  :)

> I know that the original 'rm' command didn't recover any space because
> the disk image grew by ~100 MB.

>> If so, then the trim is already done during the unlink process,
>> and there's no more work to do.
>>
>> So that's my first thought, but ...
>>
>>>  Doing this in a script:
>>>
>>>   mount -o discard /dev/sda1 /mnt
>>>   fstrim /mnt
>>>
>>> ... does nothing.  Also the fstrim is almost instantaneous -- there's
>>> no way it could be scanning the disk.
>>
>> blktrace would be a better tool to find out whether or not discards
>> are actually getting issued to storage...

blktrace is probably the place to start.  Do you see discard
requests?  then ext4 is doing its job.  If not, we can trace
ext4 to see why it's not issuing them, assuming there really
is work to do.

>> And if you strace it what does the ioctl return?
> 
> I'll try that in a few minutes.
> 
> In the mean time I captured the fstrim -v output:
> 
>   fstrim -v /
>   /: 124 MiB (130039808 bytes) trimmed
> 
> 124 MB is (within 25%) the amount of data I would expect needs to be
> trimmed.

Ok, so it says that it did do what you expected...

>> Enabling the trace_ext4_trim_all_free tracepoint might be interesting too.
> 
> That a systemtap thing?  It's tricky to get systemtap working in a
> virtual machine, but I guess I can try if nothing else works.

# trace-cmd record -e &
# <run test>
# fg
<ctrl-c>
# trace-cmd report > trace_report.txt

should do it.

>>> However, if I start with the same filesystem, mounted with -o discard,
>>> and create and rm large files, while observing the size of the
>>> underlying virtual disk, then discard is obviously working fine.  'rm'
>>> of large files makes the underlying disk shrink.

(backing up, that's the "-o discard" option at work)

>>> Any ideas here?
>>
>> first of all, I should point out that "-o discard" is not necessary for
>> fstrim / FITRIM ioctl to work.  The former tries to trim as soon
>> as files are unlinked; FITRIM goes looking for free blocks to trim.
>>
>> If you're mounting with -o discard, then fstrim should never find any
>> work to do.
> 
> Useful to know.
> 
> I thought I had to use -o discard in order for the ext4 module to send
> discard commands at all to the block layer.

nope.  That just makes it do it every time a block is freed, instead
of in batches via fstrim:

discard                 Controls whether ext4 should issue discard/TRIM
nodiscard(*)            commands to the underlying block device when
                        blocks are freed.

the FITRIM ioctl works fine w/o the mount option, and in fact as I said,
should have no work to do if the mount option is there - every freed block
shouid get discarded  (well, maybe modulo some size thresholds, I don't
remember for sure).

-Eric

> Thanks,
> 
> Rich.
>