Re: [PATCH] mm: cma: free cma page to buddy instead of being cpu hotpage

[Minchan Kim]

Hello,

On Tue, Oct 29, 2013 at 09:14:03AM -0700, Laura Abbott wrote:
> On 10/29/2013 8:02 AM, Zhang Mingjun wrote:
> 
> >    It would move the cost to the CMA paths so I would complain less. Bear
> >    in mind as well that forcing everything to go through free_one_page()
> >    means that every free goes through the zone lock. I doubt you have any
> >    machine large enough but it is possible for simultaneous CMA allocations
> >    to now contend on the zone lock that would have been previously fine.
> >    Hence, I'm interesting in knowing the underlying cause of the
> >    problem you
> >    are experiencing.
> >
> >my platform uses CMA but disabled CMA's migration func by del MIGRATE_CMA
> >in fallbacks[MIGRATE_MOVEABLE]. But I find CMA pages can still used by
> >pagecache or page fault page request from PCP list and cma allocation has to
> >migrate these page. So I want to free these cma pages to buddy directly
> >not PCP..
> >
> >     > of course, it will waste the memory outside of the alloc range
> >    but in the
> >     > pageblocks.
> >     >
> >
> >    I would hope/expect that the loss would only last for the duration of
> >    the allocation attempt and a small amount of memory.
> >
> >     > > when a range of pages have been isolated and migrated. Is there any
> >     > > measurable benefit to this patch?
> >     > >
> >     > after applying this patch, the video player on my platform works more
> >     > fluent,
> >
> >    fluent almost always refers to ones command of a spoken language. I do
> >    not see how a video player can be fluent in anything. What is measurably
> >    better?
> >
> >    For example, are allocations faster? If so, why? What cost from another
> >    path is removed as a result of this patch? If the cost is in the PCP
> >    flush then can it be checked if the PCP flush was unnecessary and called
> >    unconditionally even though all the pages were freed already? We had
> >    problems in the past where drain_all_pages() or similar were called
> >    unnecessarily causing long sync stalls related to IPIs. I'm wondering if
> >    we are seeing a similar problem here.
> >
> >    Maybe the problem is the complete opposite. Are allocations failing
> >    because there are PCP pages in the way? In that case, it real fix might
> >    be to insert a  if the allocation is failing due to per-cpu
> >    pages.
> >
> >problem is not the allocation failing, but the unexpected cma migration
> >slows
> >down the allocation.
> >
> >
> >     > and the driver of video decoder on my test platform using cma
> >    alloc/free
> >     > frequently.
> >     >
> >
> >    CMA allocations are almost never used outside of these contexts. While I
> >    appreciate that embedded use is important I'm reluctant to see an impact
> >    in fast paths unless there is a good reason for every other use case. I
> >    also am a bit unhappy to see CMA allocations making the zone->lock
> >    hotter than necessary even if no embedded use case it likely to
> >    experience the problem in the short-term.
> >
> >    --
> >    Mel Gorman
> >    SUSE Labs
> >
> >
> 
> We've had a similar patch in our tree for a year and a half because
> of CMA migration failures, not just for a speedup in allocation
> time. I understand that CMA is not the fast case or the general use
> case but the problem is that the cost of CMA failure is very high
> (complete failure of the feature using CMA). Putting CMA on the PCP
> lists means they may be picked up by users who temporarily make the
> movable pages unmovable (page cache etc.) which prevents the
> allocation from succeeding. The problem still exists even if the CMA
> pages are not on the PCP list but the window gets slightly smaller.

I understand that I have seen many people want to use CMA have tweaked
their system to work well and although they do best effort, it doesn't
work well because CMA doesn't gaurantee to succeed in getting free
space since there are lots of hurdle. (get_user_pages, AIO ring buffer,
buffer cache, short of free memory for migration, no swap and so on).
Even, someone want to allocate CMA space with speedy. SIGH.

Yeah, at the moment, CMA is really SUCK.

> 
> This really highlights one of the biggest issues with CMA today.
> Movable pages make return -EBUSY for any number of reasons. For
> non-CMA pages this is mostly fine, another movable page may be
> substituted for the movable page that is busy. CMA is a restricted
> range though so any failure in that range is very costly because CMA
> regions are generally sized exactly for the use cases at hand which
> means there is very little extra space for retries.
> 
> To make CMA actually usable, we've had to go through and add in
> hacks/quirks that prevent CMA from being allocated in any path which
> may prevent migration. I've been mixed on if this is the right path
> or if the definition of MIGRATE_CMA needs to be changed to be more
> restrictive (can't prevent migration).

Fundamental problem is that every subsystem could grab a page anytime
and they doesn't gaurantee to release it soonish or within time CMA
user want so it turns out non-determisitic mess which just hook into
core MM system here and there.

Sometime, I see some people try to solve it case by case with ad-hoc
approach. I guess it would be never ending story as kernel evolves.

I suggest that we could make new wheel with frontswap/cleancache stuff.
The idea is that pages in frontswap/cleancache are evicted from kernel
POV so that we can gaurantee that there is no chance to grab a page
in CMA area and we could remove lots of hook from core MM which just
complicated MM without benefit.

As benefit, cleancache pages could drop easily so it would be fast
to get free space but frontswap cache pages should be move into somewhere.
If there are enough free pages, it could be migrated out there. Optionally
we could compress them. Otherwise, we could pageout them into backed device.
Yeah, it could be slow than migration but at least, we could estimate the time
by storage speed ideally so we could have tunable knob. If someone want
fast CMA, he could control it with ratio of cleancache:frontswap.
IOW, higher frontswap page ratio is, slower the speed would be.
Important thing is admin could have tuned control knob and it gaurantees to
get CMA free space with deterministic time.

As drawback, if we fail to tune the ratio, memeory efficieny would be
bad so that it ends up thrashing but you guys is saying we have been
used CMA without movable fallback which means that it's already static
reserved memory and it's never CMA so you already have lost memory
efficiency and even fail to get a space so I think it's good trade-off
for embedded people.

If anyone has interest the idea, I will move into that.
If it sounds crazy idea, feel free to ignore, please.

Thanks.

> 
> Thanks,
> Laura
> -- 
> Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
> hosted by The Linux Foundation
> 
> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to majordomo@xxxxxxxxxx  For more info on Linux MM,
> see: http://www.linux-mm.org/ .
> Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>

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