lkml.org 
[lkml]   [2020]   [May]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] mm/compaction: Fix the incorrect hole in fast_isolate_freepages()
On Thu, May 21, 2020 at 09:44:07AM +0800, Baoquan He wrote:
> After investigation, it turns out that this is introduced by commit of
> linux-next: commit f6edbdb71877 ("mm: memmap_init: iterate over memblock
> regions rather that check each PFN").
>
> After investigation, it turns out that this is introduced by commit of
> linux-next, the patch subject is:
> "mm: memmap_init: iterate over memblock regions rather that check each PFN".
>

Some repetition here. I assume it's because the commit ID is not stable
because it's in linux-next.

> Qian added debugging code. The debugging log shows that the fault page is
> 0x2a800000. From the system e820 map which is pasted at bottom, the page
> is in e820 reserved range:
> BIOS-e820: [mem 0x0000000029ffe000-0x000000002a80afff] reserved
> And it's in section [0x28000000, 0x2fffffff]. In that secion, there are
> several usable ranges and some e820 reserved ranges.
>
> For this kind of e820 reserved range, it won't be added to memblock allocator.
> However, init_unavailable_mem() will initialize to add them into node 0,
> zone 0.

Why is it appropriate for init_unavailable_mem to add a e820 reserved
range to the zone at all? The bug being triggered indicates there is a
mismatch between the zone of a struct page and the PFN passed in.

> Before that commit, later, memmap_init() will add e820 reserved
> ranges into the zone where they are contained, because it can pass
> the checking of early_pfn_valid() and early_pfn_in_nid(). In this case,
> the e820 reserved range where fault page 0x2a800000 is located is added
> into DMA32 zone. After that commit, the e820 reserved rgions are kept
> in node 0, zone 0, since we iterate over memblock regions to iniatialize
> in memmap_init() instead, their node and zone won't be changed.
>

This implies that we have struct pages that should never be used (e820
reserved) but exist somehow in a zone range but with broken linkages to
their node and zone.

> Reported-by: Qian Cai <cai@lca.pw>
> Signed-off-by: Baoquan He <bhe@redhat.com>
> ---
> mm/compaction.c | 4 +++-
> 1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/mm/compaction.c b/mm/compaction.c
> index 67fd317f78db..9ce4cff4d407 100644
> --- a/mm/compaction.c
> +++ b/mm/compaction.c
> @@ -1418,7 +1418,9 @@ fast_isolate_freepages(struct compact_control *cc)
> cc->free_pfn = highest;
> } else {
> if (cc->direct_compaction && pfn_valid(min_pfn)) {
> - page = pfn_to_page(min_pfn);
> + page = pageblock_pfn_to_page(min_pfn,
> + pageblock_end_pfn(min_pfn),
> + cc->zone);
> cc->free_pfn = min_pfn;
> }
> }

Why is the correct fix not to avoid creating struct pages for e820
ranges and make sure that struct pages that are reachable have proper
node and zone linkages?

--
Mel Gorman
SUSE Labs

\
 
 \ /
  Last update: 2020-05-21 11:38    [W:0.078 / U:2.040 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site