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

From: Laura Abbott
Date: Tue Oct 29 2013 - 12:14:11 EST

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

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

problem is not the allocation failing, but the unexpected cma migration
down the allocation.

> and the driver of video decoder on my test platform using cma
> 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

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.

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).

Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
hosted by The Linux Foundation
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at
Please read the FAQ at