lists.openwall.net   lists  /  announce  owl-users  owl-dev  john-users  john-dev  passwdqc-users  yescrypt  popa3d-users  /  oss-security  kernel-hardening  musl  sabotage  tlsify  passwords  /  crypt-dev  xvendor  /  Bugtraq  Full-Disclosure  linux-kernel  linux-netdev  linux-ext4  linux-hardening  linux-cve-announce  PHC 
Open Source and information security mailing list archives
 
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date:	Fri, 9 Jan 2015 16:23:36 -0500
From:	Tejun Heo <tj@...nel.org>
To:	Dave Chinner <david@...morbit.com>
Cc:	axboe@...nel.dk, linux-kernel@...r.kernel.org, jack@...e.cz,
	hch@...radead.org, hannes@...xchg.org,
	linux-fsdevel@...r.kernel.org, vgoyal@...hat.com,
	lizefan@...wei.com, cgroups@...r.kernel.org, linux-mm@...ck.org,
	mhocko@...e.cz, clm@...com, fengguang.wu@...el.com
Subject: Re: [PATCHSET RFC block/for-next] writeback: cgroup writeback support

Hello, Dave.

On Thu, Jan 08, 2015 at 10:45:32AM +1100, Dave Chinner wrote:
> > Complications mostly arise from filesystems and inodes having to deal
> > with multiple split bdi's instead of one, but those are mostly
> > straight-forward 1:N mapping issues.  It does get tedious here and
> > there but doesn't complicate the overall picture.
> 
> Some filesystems don't track metadata-dirty inode state in the bdi
> lists, and instead track that in their own lists (usually deep
> inside the journalling subsystem). i.e. I_DIRTY_PAGES are the only
> dirty state that is tracked in the VFS. i.e. inode metadata
> writeback will still be considered global, but pages won't be. Hence
> you might get pages written back quickly, but the inodes are going
> to remain dirty and unreclaimable until the filesystem flushes some
> time in the future after the journal is committed and the inode
> written...

I'm not sure I'm following.  What writeback layer provides is cgroup
awareness when dealing with I_DIRTY_PAGES.  Metadata writebacks will
become automatically cgroup-aware to the extent they go through
regular page dirtying mechanism.  If some don't go through that
channel (e.g. journals shouldn't to avoid priority inversion), it's
upto the specific filesystem to decide how to handle them.  In most
cases, I imagine they'd be sent down as originating from the root
cgroup, ie, as the system cost.  Specific filesystems can be more
sophisticated, I suppose.

Ultimately, the only thing which matters is with which cgroup a bio is
associated when issued.  What's implemented in this patchset is
propagation of memcg tags for pagecache pages.  If necessary, further
mechanisms can be added, but this should cover the basics.

> There has also been talk of allowing filesystems to directly track
> dirty page state as well - the discussion came out of the way tux3
> was tracking and committing delta changes to file data. Now that
> hasn't gone anywhere, but I'm wondering what impact this patch set
> would have on such proposals?

Would such a filesystem take over writeback mechanism too?  The
implemented mechanism is fairly modular and the counterparts in each
filesystem should be able to use them the same way the core writeback
code does.  I'm afraid I can't say much without knowing further
details.

> Similarly, I'm concerned about additional overhead in the writeback
> path - we can easily drive the flusher thread to be CPU bound on IO
> subsystems that have decent bandwidth (low GB/s), so adding more
> overhead to every page we have to flush is going to reduce
> performance on these systems. Do you have any idea what impact
> just enabling the memcg/blkcg tracking has on writeback performance
> and CPU consumption?

I measured avg sys+user time of 50 iterations of

  fs_mark -d /mnt/tmp/ -s 104857600 -n 32

on an ext2 on a ramdisk, which should put the hot path part - page
faulting and inode dirtying - under spotlight.  cgroup writeback
enabled but not used case consumes around 1% more cpu time - AVG 6.616
STDEV 0.050 w/o this patchset, AVG 6.682 STDEV 0.046 with.  This is an
extreme case and while it isn't free the overhead is fairly low.

> A further complication for data writeback is that some filesystems
> do their own adjacent page write clustering own inside their own
> ->writepages/->writepage implementations. Both ext4 and XFS do this,
> and it makes no sense from a system and filesystem performance
> perspective to turn sequential ranges of dirty pages into much
> slower, semi-random IO just because the pages belong to different
> memcgs. It's not a good idea to compromise bulk writeback
> throughput under memory pressure just because a different memcgs
> write to the same files, so what is going to be the impact of
> filesystems ignoring memcg ownership during writeback clustering?

I don't think that's a good idea.  Implementing that isn't hard.
Range writeback can simply avoid skipping pages from different
cgroups; however, different cgroups can have vastly different
characteristics.  One may be configured to have a majority of the
available bandwidth while another has to scrap by with few hundreds of
k's per sec.  Initiating write out on pages which belong to the former
from the writeback of the latter may cause serious priority inversion
issues.

Maybe we can think of optimizations down the road but I'd strongly
prefer to stick to simple and clear divisions among cgroups.  Also, a
file highly interleaved by multiple cgroups isn't a particularly
likely use case.

> Finally, distros are going to ship with this always enabled, so what
> is the overall increase in the size of the struct inode on a 64
> bit system with it enabled?

This was in the head message, but, to repeat, two more pointers if
!CONFIG_IMA, which is the case for fedora at least.  If CONFIG_IMA is
enabled, it becomes three pointers.  In my test setup, before the
patchset or CONFIG_CGROUP_WRITEBACK disabled, it's 544 bytes and w/
CONFIG_CGROUP_WRITEBACK 560.

Thanks.

-- 
tejun
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists