On Tue, Oct 08, 2019 at 05:29:15PM +0200, Vlastimil Babka wrote:

Florian and Dave reported [1] a NULL pointer dereference in __reset_isolation_pfn(). While the exact cause is unclear, staring at the code revealed two bugs, which might be related.

I think the fix is a good fit. Even if the problem still occurs, it eliminates an important possibility.

One bug is that if zone starts in the middle of pageblock, block_page might correspond to different pfn than block_pfn, and then the pfn_valid_within() checks will check different pfn's than those accessed via struct page. This might result in acessing an unitialized page in CONFIG_HOLES_IN_ZONE configs.

s/acessing/accessing/

Aside from HOLES_IN_ZONE, the patch addresses an issue if the start of the zone is not pageblock-aligned. While this is common, it's not guaranteed. I don't think this needs to be clarified in the changelog as your example is valid. I'm commenting in case someone decides not to try the patch because they feel HOLES_IN_ZONE is required.

The other bug is that end_page refers to the first page of next pageblock and not last page of current pageblock. The online and valid check is then wrong and with sections, the while (page < end_page) loop might wander off actual struct page arrays.

[1] https://lore.kernel.org/linux-xfs/87o8z1fvqu.fsf@mid.deneb.enyo.de/

Reported-by: Florian Weimer fw@deneb.enyo.de Reported-by: Dave Chinner david@fromorbit.com Fixes: 6b0868c820ff ("mm/compaction.c: correct zone boundary handling when resetting pageblock skip hints") Cc: stable@vger.kernel.org Signed-off-by: Vlastimil Babka vbabka@suse.cz

Acked-by: Mel Gorman mgorman@techsingularity.net

Just one minor irrelevant note below.

---

mm/compaction.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/mm/compaction.c b/mm/compaction.c index ce08b39d85d4..672d3c78c6ab 100644 --- a/mm/compaction.c +++ b/mm/compaction.c @@ -270,14 +270,15 @@ __reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source, /* Ensure the start of the pageblock or zone is online and valid */ block_pfn = pageblock_start_pfn(pfn);

• block_page = pfn_to_online_page(max(block_pfn, zone->zone_start_pfn));

• block_pfn = max(block_pfn, zone->zone_start_pfn);
• block_page = pfn_to_online_page(block_pfn); if (block_page) { page = block_page; pfn = block_pfn; }

/* Ensure the end of the pageblock or zone is online and valid */

• block_pfn += pageblock_nr_pages;

• block_pfn = pageblock_end_pfn(pfn) - 1; block_pfn = min(block_pfn, zone_end_pfn(zone) - 1); end_page = pfn_to_online_page(block_pfn); if (!end_page)

This is fine and is definetly fixing a potential issue.

@@ -303,7 +304,7 @@ __reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source, page += (1 << PAGE_ALLOC_COSTLY_ORDER); pfn += (1 << PAGE_ALLOC_COSTLY_ORDER);

• } while (page < end_page);

• } while (page <= end_page);

return false; }

I think this is also ok as it's appropriate for PFN walkers in general of this style. However, I think it's unlikely to fix anything given that we are walking in steps of (1 << PAGE_ALLOC_COSTLY_ORDER) and the final page is not necessarily aligned on that boundary. Still, it's an improvement.

Thanks