lkml.org 
[lkml]   [2010]   [Oct]   [17]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [RFC][PATCH 3/3] alloc contig pages with migration.
From
On Wed, Oct 13, 2010 at 12:18 PM, KAMEZAWA Hiroyuki
<kamezawa.hiroyu@jp.fujitsu.com> wrote:
> From: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
>
> Add an function to allocate contigous memory larger than MAX_ORDER.
> The main difference between usual page allocater is that this uses
> memory offline techiqueue (Isoalte pages and migrate remaining pages.).
>
> I think this is not 100% solution because we can't avoid fragmentation,
> but we have kernelcore= boot option and can create MOVABLE zone. That
> helps us to allow allocate a contigous range on demand.
>
> Maybe drivers can alloc contig pages by bootmem or hiding some memory
> from the kernel at boot. But if contig pages are necessary only in some
> situation, kernelcore= boot option and using page migration is a choice.
>
> Anyway, to allocate a contiguous chunk larger than MAX_ORDER, we need to
> add an overlay allocator on buddy allocator. This can be a 1st step.
>
> Note:
> This function is heavy if there are tons of memory requesters. So, maybe
> not good for 1GB pages for x86's usual use. It will requires some other
> tricks than migration.

I got found many typos but I don't pointed out each by each. :)
Please, correct typos in next version.

>
> TODO:
>  - allows the caller to specify the migration target pages.
>  - reduce the number of lru_add_drain_all()..etc...system wide heavy calls.
>  - Pass gfp_t for some purpose...
>
> Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
> ---
>  include/linux/page-isolation.h |    8 ++
>  mm/page_alloc.c                |   29 ++++++++
>  mm/page_isolation.c            |  136 +++++++++++++++++++++++++++++++++++++++++
>  3 files changed, 173 insertions(+)
>
> Index: mmotm-1008/mm/page_isolation.c
> ===================================================================
> --- mmotm-1008.orig/mm/page_isolation.c
> +++ mmotm-1008/mm/page_isolation.c
> @@ -7,6 +7,7 @@
>  #include <linux/mm.h>
>  #include <linux/page-isolation.h>
>  #include <linux/pageblock-flags.h>
> +#include <linux/swap.h>
>  #include <linux/memcontrol.h>
>  #include <linux/migrate.h>
>  #include <linux/memory_hotplug.h>
> @@ -384,3 +385,138 @@ retry:
>        }
>        return 0;
>  }
> +
> +/**
> + * alloc_contig_pages - allocate a contigous physical pages
> + * @hint:      the base address of searching free space(in pfn)
> + * @size:      size of requested area (in # of pages)

Could you add _range_ which can be specified by user into your TODO list?
Maybe some embedded system have a requirement to allocate contiguous
pages in some bank.
so guys try to allocate pages in some base address and if it fails, he
can try to next offset in same bank.
But it's very annoying. So let's add feature that user can specify
_range_ where user want to allocate.

> + * @node:       the node from which memory is allocated. "-1" means anywhere.
> + * @no_search: if true, "hint" is not a hint, requirement.

As I said previous, how about "strict" or "ALLOC_FIXED" like MAP_FIXED?

> + *
> + * Search an area of @size in the physical memory map and checks wheter

Typo
whether

> + * we can create a contigous free space. If it seems possible, try to
> + * create contigous space with page migration. If no_search==true, we just try
> + * to allocate [hint, hint+size) range of pages as contigous block.
> + *
> + * Returns a page of the beginning of contiguous block. At failure, NULL
> + * is returned. Each page in the area is set to page_count() = 1. Because

Why do you mention page_count() = 1?
Do users of this function have to know it?

> + * this function does page migration, this function is very heavy and

Nitpick.
page migration is implementation, too. Do we need to mention it in here?
We might add page reclaim/or new feature in future or page migration
might be very light function although it is not a easy. :)
Let's not show the implementation for users.

> + * sleeps some time. Caller must be aware that "NULL returned" is not a
> + * special case.

I think this information is enough to users.

> + *
> + * Now, returned range is aligned to MAX_ORDER. (So "hint" must be aligned
> + * if no_search==true.)

Couldn't we add handling of this exception?
If (hint != MAX_ORDER_ALIGH(hint) && no_search == true)
return 0 or WARN_ON?

> + */
> +
> +#define MIGRATION_RETRY        (5)
> +struct page *alloc_contig_pages(unsigned long hint, unsigned long size,
> +                               int node, bool no_search)
> +{
> +       unsigned long base, found, end, pages, start;
> +       struct page *ret = NULL;
> +       int migration_failed;
> +       struct zone *zone;
> +
> +       hint = MAX_ORDER_ALIGN(hint);
> +       /*
> +        * request size should be aligned to pageblock_order..but use
> +        * MAX_ORDER here for avoiding messy checks.
> +        */
> +       pages = MAX_ORDER_ALIGN(size);
> +       found = 0;
> +retry:
> +       for_each_populated_zone(zone) {
> +               unsigned long zone_end_pfn;
> +
> +               if (node >= 0 && node != zone_to_nid(zone))
> +                       continue;
> +               if (zone->present_pages < pages)
> +                       continue;
> +               base = MAX_ORDER_ALIGN(zone->zone_start_pfn);
> +               base = max(base, hint);
> +               zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
> +               if (base + pages > zone_end_pfn)
> +                       continue;
> +               found = find_contig_block(base, zone_end_pfn, pages, no_search);
> +               /* Next try will see the next block. */
> +               hint = base + MAX_ORDER_NR_PAGES;
> +               if (found)
> +                       break;
> +       }
> +
> +       if (!found)
> +               return NULL;
> +
> +       if (no_search && found != hint)

You increased hint before "break".
So if the no_search is true, this condition (found != hint) is always true.


> +               return NULL;
> +
> +       /*
> +        * Ok, here, we have contiguous pageblock marked as "isolated"
> +        * try migration.
> +        *
> +        * FIXME: permanent migration_failure detection logic is required.
> +        */
> +       lru_add_drain_all();
> +       flush_scheduled_work();
> +       drain_all_pages();
> +
> +       end = found + pages;
> +       /*
> +        * scan_lru_pages() finds the next PG_lru page in the range
> +        * scan_lru_pages() returns 0 when it reaches the end.
> +        */
> +       for (start = scan_lru_pages(found, end), migration_failed = 0;
> +            start && start < end;
> +            start = scan_lru_pages(start, end)) {
> +               if (do_migrate_range(start, end)) {
> +                       /* it's better to try another block ? */
> +                       if (++migration_failed >= MIGRATION_RETRY)
> +                               break;
> +                       /* take a rest and synchronize LRU etc. */
> +                       lru_add_drain_all();
> +                       flush_scheduled_work();
> +                       cond_resched();
> +                       drain_all_pages();
> +               } else /* reset migration_failure counter */
> +                       migration_failed = 0;
> +       }
> +
> +       lru_add_drain_all();
> +       flush_scheduled_work();
> +       drain_all_pages();

Hmm.. as you mentioned, It would be better to remove many flush lru/per-cpu.
But in embedded system, it couldn't be a big overhead.

> +       /* Check all pages are isolated */
> +       if (test_pages_isolated(found, end)) {
> +               undo_isolate_page_range(found, pages);
> +               /*
> +                * We failed at [start...end) migration.
> +                * FIXME: there may be better restaring point.
> +                */
> +               hint = MAX_ORDER_ALIGN(end + 1);
> +               goto retry; /* goto next chunk */
> +       }
> +       /*
> +        * Ok, here, [found...found+pages) memory are isolated.
> +        * All pages in the range will be moved into the list with
> +        * page_count(page)=1.
> +        */
> +       ret = pfn_to_page(found);
> +       alloc_contig_freed_pages(found, found + pages);
> +       /* unset ISOLATE */
> +       undo_isolate_page_range(found, pages);
> +       /* Free unnecessary pages in tail */
> +       for (start = found + size; start < found + pages; start++)
> +               __free_page(pfn_to_page(start));
> +       return ret;
> +
> +}

Thanks for the good patches, Kame.


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

\
 
 \ /
  Last update: 2010-10-17 06:15    [W:0.160 / U:0.232 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site