Re: btrfs flooding the I/O subsystem and hanging the machine, with bcache cache turned off

[Michal Hocko <mhocko@kernel.org>]

On Wed 30-11-16 10:16:53, Marc MERLIN wrote:
> +folks from linux-mm thread for your suggestion
> 
> On Wed, Nov 30, 2016 at 01:00:45PM -0500, Austin S. Hemmelgarn wrote:
> > > swraid5 < bcache < dmcrypt < btrfs
> > > 
> > > Copying with btrfs send/receive causes massive hangs on the system.
> > > Please see this explanation from Linus on why the workaround was
> > > suggested:
> > > https://lkml.org/lkml/2016/11/29/667
> > And Linux' assessment is absolutely correct (at least, the general
> > assessment is, I have no idea about btrfs_start_shared_extent, but I'm more
> > than willing to bet he's correct that that's the culprit).
> 
> > > All of this mostly went away with Linus' suggestion:
> > > echo 2 > /proc/sys/vm/dirty_ratio
> > > echo 1 > /proc/sys/vm/dirty_background_ratio
> > > 
> > > But that's hiding the symptom which I think is that btrfs is piling up too many I/O
> > > requests during btrfs send/receive and btrfs scrub (probably balance too) and not
> > > looking at resulting impact to system health.
> 
> > I see pretty much identical behavior using any number of other storage
> > configurations on a USB 2.0 flash drive connected to a system with 16GB of
> > RAM with the default dirty ratios because it's trying to cache up to 3.2GB
> > of data for writeback.  While BTRFS is doing highly sub-optimal things here,
> > the ancient default writeback ratios are just as much a culprit.  I would
> > suggest that get changed to 200MB or 20% of RAM, whichever is smaller, which
> > would give overall almost identical behavior to x86-32, which in turn works
> > reasonably well for most cases.  I sadly don't have the time, patience, or
> > expertise to write up such a patch myself though.
> 
> Dear linux-mm folks, is that something you could consider (changing the
> dirty_ratio defaults) given that it affects at least bcache and btrfs
> (with or without bcache)?

As much as the dirty_*ratio defaults a major PITA this is not something
that would be _easy_ to change without high risks of regressions. This
topic has been discussed many times with many good ideas, nothing really
materialized from them though :/

To be honest I really do hate dirty_*ratio and have seen many issues on
very large machines and always suggested to use dirty_bytes instead but
a particular value has always been a challenge to get right. It has
always been very workload specific.

That being said this is something more for IO people than MM IMHO.

-- 
Michal Hocko
SUSE Labs

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>