lkml.org 
[lkml]   [2018]   [Oct]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [RFC v4 PATCH 2/5] mm/__free_one_page: skip merge for order-0 page unless compaction failed
From
Date
On 10/18/18 8:48 AM, Aaron Lu wrote:
> On Wed, Oct 17, 2018 at 07:03:30PM +0200, Vlastimil Babka wrote:
>> On 10/17/18 3:58 PM, Mel Gorman wrote:
>>> Again, as compaction is not guaranteed to find the pageblocks, it would
>>> be important to consider whether a) that matters or b) find an
>>> alternative way of keeping unmerged buddies on separate lists so they
>>> can be quickly discovered when a high-order allocation fails.
>>
>> Agree, unmerged buddies could be on separate freelist from regular
>> order-0 freelist. That list could be also preferred to allocations
>> before the regular one. Then one could e.g. try "direct merging" via
>> this list when compaction fails, or prefer direct merging to compaction
>> for non-costly-order allocations, do direct merging when allocation
>> context doesn't even allow compaction (atomic etc).
>
> One concern regarding "direct merging" these unmerged pages via this
> separate freelist(let's call it unmerged_free_list) is: adjacent
> unmerged pages on the unmerged_free_list could be far away from each
> other regarding their physical positions, so during the process of
> merging them, the needed high order page may not be able to be formed
> in a short time. Actually, the time could be unbound in a bad condition
> when:
> 1 unmerged pages adjacent on the unmerged_free_list happen to be far
> away from each other regarding their physical positions; and

I'm not sure I understand. Why should it matter for merging if pages are
adjacent on the unmerged_free_list? The buddy for merging is found the
usual way, no?

> 2 there are a lot of unmerged pages on unmerged_free_list.

That will affect allocation latency, yeah. Still might be faster than
direct compaction. And possible to do in GFP_ATOMIC context, unlike
direct compaction.

> That's the reason I hooked the merging of unmerged pages in compaction
> when isolate_migratepages_block() is scanning every page of a pageblock
> in PFN order.
>
> OTOH, if there is a kernel thread trying to reduce fragmentation by
> doing merges for these unmerged pages, I think it's perfect fine to let
> it iterate all unmerged pages of that list and do_merge() for all of
> them.
>
> So what about this: if kcompactd is running, let it handle these
> unmerged pages on the list and after that, do its usual job of
> compaction. If direct compaction is running, do not handle unmerged
> pages on that list but rely on isolate_migratepages_block() to do the
> merging as is done in this patchset.
>
> This of course has the effect of tying compaction with 'lazy merging'.
> If it is not desirable, what about creating a new kernel thread to do
> the merging of unmerged pages on the list while keeping the behaviour of
> isolate_migratepages_block() in this patchset to improve compaction
> success rate.

Note that anything based on daemons will seem like reducing latency for
allocations, but if we delay merging and then later do it from a daemon,
the overall zone lock times will be essentially the same, right? The
reduced zone lock benefits only happen when the unmerged pages get
reallocated.

>> Also I would definitely consider always merging pages freed to
>> non-MOVABLE pageblocks. We really don't want to increase the
>> fragmentation in those. However that means it probably won't help the
>> netperf case?
>
> Yes, that would be unfortunate for all in-kernel users of page
> allocator...

In that case there should definitely be a direct merging possibility
IMHO, even if only as a last resort before stealing from another pageblock.

\
 
 \ /
  Last update: 2018-10-18 10:23    [W:0.078 / U:9.264 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site