[lkml]   [2008]   [Nov]   [24]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH -mm] vmscan: bail out of page reclaim after swap_cluster_max pages
>> Rik, sorry, I nak current your patch. because it don't fix old akpm issue.
> You are right. We do need to keep pressure between zones
> equivalent to the size of the zones (or more precisely, to
> the number of pages the zones have on their LRU lists).

Oh, sorry.
you are right. but I talked about reverse thing.

1. shrink_zones() doesn't have any shortcut exiting way.
it always call all zone's shrink_zone()
2. balance_pgdat also doesn't have shortcut.

simple shrink_zone() shortcut and lite memory pressure cause following
bad scenario.

1. reclaim 32 page from ZONE_HIGHMEM
2. reclaim 32 page from ZONE_NORMAL
3. reclaim 32 page from ZONE_DMA
4. exit reclaim
5. another task call page alloc and it cause try_to_free_pages()
6. reclaim 32 page from ZONE_HIGHMEM
7. reclaim 32 page from ZONE_NORMAL
8. reclaim 32 page from ZONE_DMA

oops, all zone reclaimed the same pages although ZONE_HIGHMEM have
much memory than ZONE_DMA.
IOW, ZONE_DMA's reclaim scanning rate is much than ZONE_HIGHMEM largely.

it isn't intentionally.

Actually, try_to_free_pages don't need pressure fairness. it is the
role of the balance_pgdat().

> However, having dozens of direct reclaim tasks all getting
> to the lower priority levels can be disastrous, causing
> extraordinarily large amounts of memory to be swapped out
> and minutes-long stalls to applications.


> I think we can come up with a middle ground here:
> - always let kswapd continue its rounds


> - have direct reclaim tasks continue when priority == DEF_PRIORITY

it cause above bad scenario, I think.

> - break out of the loop for direct reclaim tasks, when
> priority < DEF_PRIORITY and enough pages have been freed
> Does that sound like it would mostly preserve memory pressure
> between zones, while avoiding the worst of the worst when it
> comes to excessive page eviction?

 \ /
  Last update: 2008-11-24 20:15    [W:0.058 / U:0.100 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site