Re: [GIT PULL] ext4 updates for 3.9

[Eric Sandeen <sandeen@redhat.com>]

On 2/27/13 2:58 PM, Dmitry Monakhov wrote:
> On Wed, 27 Feb 2013 14:29:07 -0500, Theodore Ts'o <tytso@mit.edu> wrote:
>> On Wed, Feb 27, 2013 at 02:19:23PM -0500, Dave Jones wrote:
>>>
>>> Looks like it's fixed here too.
>>>
>>> How did this make it through -next without anyone hitting it ?
>>>
>>> I can't remember how many years ago I last bought a disk < 1TB,
>>> and I can't be alone.  Or is everyone all about SSDs these days?
>>
>> I use LVM, so I have a number of volues which are smaler than 512GB,
>> but very few which are actually larger than 1TB.  And none on my test
>> boxes.  I was running the bleeding edge ext4 code on my laptop as for
>> dogfooding purposes, but I have an 80GB mSATA SSD and a 500GB HDD on
>> my X230 laptop (it requires a thin laptop drive, and 7mm drives don't
>> come any bigger, alas).
>>
>>> Is anyone running xfstests or similar on linux-next regularly ?
>>
>> I run xfstests on the ext4 tree, and I ran it on ext4 plus Linus's tip
>> before I submitted a pull request.  The problem is that XFSTESTS is
>> S-L-O-W if you use large partitions, so typically I use a 5GB
> Indeed. That's why i give-up rotated disks and run xfstest only on SSD
> or brd module 
>> partition sizes for my test runs.  Normally we're worried about race
>> condition bugs, not something as bone-headed as a bitmasking problem,
>> so it makes sense to use a smaller disk for most of your testing.
>> (Some folks do their xfstests run on SSD's or tmpfs image files, again
>> for speed reasons, and it's unlikely they would be big enough.)
>>
>> So what we probably need to do is to have a separate set of tests
>> using a loopback mount, and perhaps an artificially created file
>> system which has a large percentage of the blocks in the middle of the
>> file system busied out, to make efficient testing of these sorts of
>> bugs more efficient.  As I said, I'm thinking about how's the best way
>> to improve our testing regime to catch such problems the next time around.
> Amazing idea. Something like:
> 
> #dd if=/dev/zero of=/tmp/fs.img bs=1M seek=2000000 count=1
> #mkfs.ext4 -m0 -i4096000 /tmp/fs.img
> #mount /tmp/fs.img /mnt/ -oloop
> #for ((i=0; i < 2000; i++));do   fallocate -l $((1024*1024*1024)) /mnt/f$i ;done
> #for ((i=0; i < 2000; i++));do   truncate -s $((1023*1024*1024)) /mnt/f$i ;done
> 
> As result file system image has 2gb of free space wich is fragmented to ~2000
> chunks 1Mb each. But image itself is quite small
> # df /mnt
> Filesystem      1K-blocks       Used Available Use% Mounted on
> /dev/loop0     2047678076 2045679228   1998848 100% /mnt
> # du -sch /tmp/fs.img 
> 242M     /tmp/fs.img
> 242M     total
> 
> Later we can simply run xfstest/fio/fsx on this image.
> I'll prepare new xfstest based on that idea. But the only disadvantage
> is that loop dev has bottleneck, all requests will be serialized on i_mutex.

Before anyone does too much work, it would be worth revisiting
dchinner's
	[PATCH 0/10] xfstests: rework large filesystem testing
series from July 2012 to see if it meets the needs already.

It almost got all reviews, with one sticking point left, AFAICT.

-Eric