On 06.09.21 13:50, David Hildenbrand wrote:
On 06.09.21 13:32, Miaohe Lin wrote:
On 2021/9/6 17:40, David Hildenbrand wrote:
On 04.09.21 11:18, Miaohe Lin wrote:
If __isolate_free_page() failed, due to zone watermark check, the page is
still on the free list. But this page will be put back to free list again
via __putback_isolated_page() now. This may trigger page->flags checks in
__free_one_page() if PageReported is set. Or we will corrupt the free list
because list_add() will be called for pages already on another list.
Fixes: 3c605096d315 ("mm/page_alloc: restrict max order of merging on isolated pageblock")
Signed-off-by: Miaohe Lin <linmiaohe@xxxxxxxxxx>
---
mm/page_isolation.c | 6 ++----
1 file changed, 2 insertions(+), 4 deletions(-)
diff --git a/mm/page_isolation.c b/mm/page_isolation.c
index 9bb562d5d194..7d70d772525c 100644
--- a/mm/page_isolation.c
+++ b/mm/page_isolation.c
@@ -93,10 +93,8 @@ static void unset_migratetype_isolate(struct page *page, unsigned migratetype)
buddy_pfn = __find_buddy_pfn(pfn, order);
buddy = page + (buddy_pfn - pfn);
- if (!is_migrate_isolate_page(buddy)) {
- __isolate_free_page(page, order);
- isolated_page = true;
- }
+ if (!is_migrate_isolate_page(buddy))
+ isolated_page = !!__isolate_free_page(page, order);
}
}
Shouldn't we much rather force to ignore watermarks here and make sure __isolate_free_page() never fails?
It seems it is not easy to force to ignore watermarks here. And it's not a problem
if __isolate_free_page() fails because we can do move_freepages_block() anyway.
What do you think? Many thanks.
I'm wondering if all this complexity in this function is even required. What about something like this: (completely untested)
diff --git a/mm/page_isolation.c b/mm/page_isolation.c
index bddf788f45bf..29ff2fcb339c 100644
--- a/mm/page_isolation.c
+++ b/mm/page_isolation.c
@@ -66,12 +66,10 @@ static int set_migratetype_isolate(struct page *page, int migratetype, int isol_
static void unset_migratetype_isolate(struct page *page, unsigned migratetype)
{
+ bool buddy_merge_possible = false;
struct zone *zone;
unsigned long flags, nr_pages;
- bool isolated_page = false;
unsigned int order;
- unsigned long pfn, buddy_pfn;
- struct page *buddy;
zone = page_zone(page);
spin_lock_irqsave(&zone->lock, flags);
@@ -79,26 +77,15 @@ static void unset_migratetype_isolate(struct page *page, unsigned migratetype)
goto out;
/*
- * Because freepage with more than pageblock_order on isolated
- * pageblock is restricted to merge due to freepage counting problem,
- * it is possible that there is free buddy page.
- * move_freepages_block() doesn't care of merge so we need other
- * approach in order to merge them. Isolation and free will make
- * these pages to be merged.
+ * If our free page spans at least this whole pageblock and could
+ * eventually get merged into an even bigger page, go via
+ * __putback_isolated_page(), because move_freepages_block() won't
+ * trigger merging of free pages.
*/
if (PageBuddy(page)) {
order = buddy_order(page);
- if (order >= pageblock_order && order < MAX_ORDER - 1) {
- pfn = page_to_pfn(page);
- buddy_pfn = __find_buddy_pfn(pfn, order);
- buddy = page + (buddy_pfn - pfn);
-
- if (pfn_valid_within(buddy_pfn) &&
- !is_migrate_isolate_page(buddy)) {
- __isolate_free_page(page, order);
- isolated_page = true;
- }
- }
+ if (order >= pageblock_order && order < MAX_ORDER - 1)
+ buddy_merge_possible = true;
}
/*
@@ -111,12 +98,12 @@ static void unset_migratetype_isolate(struct page *page, unsigned migratetype)
* onlining - just onlined memory won't immediately be considered for
* allocation.
*/
- if (!isolated_page) {
+ if (!buddy_merge_possible) {
nr_pages = move_freepages_block(zone, page, migratetype, NULL);
__mod_zone_freepage_state(zone, nr_pages, migratetype);
}
set_pageblock_migratetype(page, migratetype);
- if (isolated_page)
+ if (buddy_merge_possible)
__putback_isolated_page(page, order, migratetype);
zone->nr_isolate_pageblock--;
out:
Okay, I just had another look -- that won't work because as you
correctly said, it still is on the freelist ...
So your fix is certainly correct :)
--
Thanks,
David / dhildenb
