lkml.org 
[lkml]   [2020]   [Nov]   [25]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: [PATCH 1/1] mm: compaction: avoid fast_isolate_around() to set pageblock_skip on reserved pages
Date

> Am 25.11.2020 um 06:34 schrieb Andrea Arcangeli <aarcange@redhat.com>:
>
> Hello,
>
>> On Mon, Nov 23, 2020 at 02:01:16PM +0100, Vlastimil Babka wrote:
>>> On 11/21/20 8:45 PM, Andrea Arcangeli wrote:
>>> A corollary issue was fixed in
>>> 39639000-39814fff : Unknown E820 type
>>>
>>> pfn 0x7a200 -> 0x7a200000 min_pfn hit non-RAM:
>>>
>>> 7a17b000-7a216fff : Unknown E820 type
>>
>> It would be nice to also provide a /proc/zoneinfo and how exactly the
>> "zone_spans_pfn" was violated. I assume we end up below zone's
>> start_pfn, but is it true?
>
> Agreed, I was about to grab that info along with all page struct
> around the pfn 0x7a200 and phys address 0x7a216fff.
>
> # grep -A1 E820 /proc/iomem
> 7a17b000-7a216fff : Unknown E820 type
> 7a217000-7bffffff : System RAM
>
> DMA zone_start_pfn 1 zone_end_pfn() 4096 contiguous 1
> DMA32 zone_start_pfn 4096 zone_end_pfn() 1048576 contiguous 0
> Normal zone_start_pfn 1048576 zone_end_pfn() 4715392 contiguous 1
> Movable zone_start_pfn 0 zone_end_pfn() 0 contiguous 0
>
> 500222 0x7a1fe000 0x1fff000000001000 reserved True
> 500223 0x7a1ff000 0x1fff000000001000 reserved True
>
> # I suspect "highest pfn" was somewhere in the RAM range
> # 0x7a217000-0x7a400000 and the pageblock_start_pfn(pfn)
> # made highest point to pfn 0x7a200 physaddr 0x7a200000
> # below, which is reserved indeed since it's non-RAM
> # first number is pfn hex(500224) == 0x7a200
>
> pfn physaddr page->flags
> 500224 0x7a200000 0x1fff000000001000 reserved True
> 500225 0x7a201000 0x1fff000000001000 reserved True
> *snip*
> 500245 0x7a215000 0x1fff000000001000 reserved True
> 500246 0x7a216000 0x1fff000000001000 reserved True
> 500247 0x7a217000 0x3fff000000000000 reserved False
> 500248 0x7a218000 0x3fff000000000000 reserved False
>
> All RAM pages non-reserved are starting at 0x7a217000 as expected.
>
> The non-RAM page_zonenum(pfn_to_page(0x7a200)) points to ZONE_DMA and
> page_zone(page) below was the DMA zone despite the pfn of 0x7a200 is
> in DMA32.
>
> VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
>
> So the patch I sent earlier should prevent the above BUG_ON by not
> setting highest to 0x7a200 when pfn is in the phys RAM range
> 0x7a217000-0x7a400000, because pageblock_pfn_to_page will notice that
> the zone is the wrong one.
>
> if (page_zone(start_page) != zone)
> return NULL;
>
> However the real bug seems that reserved pages have a zero zone_id in
> the page->flags when it should have the real zone id/nid. The patch I
> sent earlier to validate highest would only be needed to deal with
> pfn_valid.
>
> Something must have changed more recently than v5.1 that caused the
> zoneid of reserved pages to be wrong, a possible candidate for the
> real would be this change below:
>
> + __init_single_page(pfn_to_page(pfn), pfn, 0, 0);
>

Before that change, the memmap of memory holes were only zeroed out. So the zones/nid was 0, however, pages were not reserved and had a refcount of zero - resulting in other issues.

Most pfn walkers shouldn‘t mess with reserved pages and simply skip them. That would be the right fix here.

> Even if it may not be it, at the light of how the reserved page
> zoneid/nid initialized went wrong, the above line like it's too flakey
> to stay.
>
> It'd be preferable if the pfn_valid fails and the
> pfn_to_section_nr(pfn) returns an invalid section for the intermediate
> step. Even better memset 0xff over the whole page struct until the
> second stage comes around.

I recently discussed with Baoquan to
1. Using a new pagetype to mark memory holes
2. Using special nid/zid to mark memory holes

Setting the memmap to 0xff would be even more dangerous - pfn_zone() might simply BUG_ON.

>
> Whenever pfn_valid is true, it's better that the zoneid/nid is correct
> all times, otherwise if the second stage fails we end up in a bug with
> weird side effects.

Memory holes with a valid memmap might not have a zone/nid. For now, skipping reserved pages should be good enough, no?

>
> Maybe it's not the above that left a zero zoneid though, I haven't
> tried to bisect it yet to look how the page->flags looked like on a
> older kernel that didn't seem to reproduce this crash, I'm just
> guessing.
>
> Thanks,
> Andrea

\
 
 \ /
  Last update: 2020-11-25 07:46    [W:0.240 / U:11.992 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site