Re: [PATCH 2/2] mm: forcely swapout when we are out of page cache

[Sonny Rao <sonnyrao@google.com>]

On Tue, Jan 15, 2013 at 4:50 PM, Andrew Morton
<akpm@linux-foundation.org> wrote:
> On Tue, 15 Jan 2013 16:32:38 -0800
> Sonny Rao <sonnyrao@google.com> wrote:
>
>> >> It's for saving the power to increase batter life.
>> >
>> > It might well have that effect, dunno.  That wasn't my intent.  Testing
>> > needed!
>> >
>>
>> Power saving is certainly why we had it on originally for ChromeOS,
>> but we turned it off due to misbehavior.
>>
>> Specifically, we saw a pathological behavior where we'd end up writing
>> to the disk every few seconds when laptop mode was turned on.  This
>> turned out to be because laptop-mode sets a timer which is used to
>> check for new dirty data after the initial flush and writes that out
>> before spinning the disk down, and on ChromeOS various chatty daemons
>> on the system were logging and dirtying data more or less constantly
>> so there was almost always something there to be written out.  So what
>> ended up happening was that we'd need to do a read, then wake up the
>> disk, and then keep writing every few seconds for a long period of
>> time, which had the opposite effect from what we wanted.
>
> So after the read, the disk would chatter away doing a dribble of
> writes?  That sounds like plain brokenness (and why did the chrome guys
> not tell anyone about it?!?!?).

Yes, either read or fsync.  I ranted about it a little (here:
http://marc.info/?l=linux-mm&m=135422986220016&w=4), but mostly
assumed it was working as expected, and that ChromeOS was just
dirtying data at an absurd pace.  Might have been a bad assumption and
I could have been more explicit about reporting it, sorry about that.

> The idea is that when the physical
> read occurs, we should opportunistically flush out all pending writes,
> while the disk is running.  Then go back into
> buffer-writes-for-a-long-time mode.
>

See the comment in page-writeback.c above laptop_io_completion():

/*
 * We've spun up the disk and we're in laptop mode: schedule writeback
 * of all dirty data a few seconds from now.  If the flush is already
scheduled
 * then push it back - the user is still using the disk.
 */
void laptop_io_completion(struct backing_dev_info *info)

What ends up happening fairly often is that there's always something
dirty with that few seconds (or even one second) on our system.

> I forget what we did with fsync() and friends.  Quite a lot of
> pestiferous applications like to do fsync quite frequently.  I had a
> special kernel in which fsync() consisted of "return 0;", but ISTR
> there being some resistance to productizing that idea.
>

Yeah, we have this problem and we try to fix up users of fsync() as we
find them but it's a bit of a never-ending battle.  Such a feature
would be useful.

>>  The issues
>> with zram swap just confirmed that we didn't want laptop mode.
>>
>> Most of our devices have had SSDs rather than spinning disks, so noise
>> wasn't an issue, although when we finally did support an official
>> device with a spinning disk people certainly complained when the disk
>> started clicking all the time
>
> hm, it's interesting that the general idea still has vailidity.  It
> would be a fun project for someone to sniff out all the requirements,
> fixup/enhance/rewrite the current implementation and generally make it
> all spiffy and nice.
>
>> (due to the underflow in the writeback code).
>
> To what underflow do you refer?
>
http://git.kernel.org/?p=linux/kernel/git/torvalds/linux.git;a=commit;h=c8b74c2f6604923de91f8aa6539f8bb934736754

That particular bug caused writes to happen almost instantly after the
underflow ocurred, and consequently slowed write throughput to a crawl
because there was no chance for contiguous writes to gather.

>> We do know that current SSDs save a significant amount of
>> power when they go into standby, so minimizing disk writes is still
>> useful on these devices.
>>
>> A very simple laptop mode which only does a single sync when we spin
>> up the disk, and didn't bother with the timer behavior or muck with
>> swap behavior might be something that is more useful for us, and I
>> suspect it might simplify the writeback code somewhat as well.
>
> I don't think I understand the problem with the timer.  My original RFC
> said
>
> : laptop_writeback_centisecs
> : --------------------------
> :
> : This tunable determines the maximum age of dirty data when the machine
> : is operating in Laptop mode.  The default value is 30000 - five
> : minutes.  This means that if applications are generating a small amount
> : of write traffic, the disk will spin up once per five minutes.
> :
> : If the disk is spun up for any other reason (such as for a read) then
> : all dirty data will be flushed anyway, and this timer is reset to zero.
>
> which all sounds very sensible and shouldn't exhibit the behavior you
> observed.
>

The laptop-mode timer get re-armed after each writeback (see above
laptop_io_completion function), even if it was caused by laptop-mode
itself.  So, if something is continually dirtying a little bit of
data, we end up getting a chain of small writes which keeps the disk
awake for long periods of time.