On Tue, 24 May 2022 15:47:56 -0400 Zi Yan <[email protected]> wrote:
> From: Zi Yan <[email protected]>
>
> In isolate_single_pageblock() called by start_isolate_page_range(),
> there are some pageblock isolation issues causing a potential
> infinite loop when isolating a page range. This is reported by Qian Cai.
>
> 1. the pageblock was isolated by just changing pageblock migratetype
> without checking unmovable pages. Calling set_migratetype_isolate() to
> isolate pageblock properly.
> 2. an off-by-one error caused migrating pages unnecessarily, since the page
> is not crossing pageblock boundary.
> 3. migrating a compound page across pageblock boundary then splitting the
> free page later has a small race window that the free page might be
> allocated again, so that the code will try again, causing an potential
> infinite loop. Temporarily set the to-be-migrated page's pageblock to
> MIGRATE_ISOLATE to prevent that and bail out early if no free page is
> found after page migration.
>
> An additional fix to split_free_page() aims to avoid crashing in
> __free_one_page(). When the free page is split at the specified
> split_pfn_offset, free_page_order should check both the first bit of
> free_page_pfn and the last bit of split_pfn_offset and use the smaller one.
> For example, if free_page_pfn=0x10000, split_pfn_offset=0xc000,
> free_page_order should first be 0x8000 then 0x4000, instead of 0x4000 then
> 0x8000, which the original algorithm did.
>
> ...
>
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -1114,13 +1114,16 @@ void split_free_page(struct page *free_page,
> unsigned long flags;
> int free_page_order;
>
> + if (split_pfn_offset == 0)
> + return;
> +
> spin_lock_irqsave(&zone->lock, flags);
> del_page_from_free_list(free_page, zone, order);
> for (pfn = free_page_pfn;
> pfn < free_page_pfn + (1UL << order);) {
> int mt = get_pfnblock_migratetype(pfn_to_page(pfn), pfn);
>
> - free_page_order = ffs(split_pfn_offset) - 1;
> + free_page_order = min(pfn ? __ffs(pfn) : order, __fls(split_pfn_offset));
Why is it testing the zeroness of `pfn' here? Can pfn==0 even happen?
If so, it's a legitimate value so why does it get special-cased?
On 24 May 2022, at 16:23, Andrew Morton wrote:
> On Tue, 24 May 2022 15:47:56 -0400 Zi Yan <[email protected]> wrote:
>
>> From: Zi Yan <[email protected]>
>>
>> In isolate_single_pageblock() called by start_isolate_page_range(),
>> there are some pageblock isolation issues causing a potential
>> infinite loop when isolating a page range. This is reported by Qian Cai.
>>
>> 1. the pageblock was isolated by just changing pageblock migratetype
>> without checking unmovable pages. Calling set_migratetype_isolate() to
>> isolate pageblock properly.
>> 2. an off-by-one error caused migrating pages unnecessarily, since the page
>> is not crossing pageblock boundary.
>> 3. migrating a compound page across pageblock boundary then splitting the
>> free page later has a small race window that the free page might be
>> allocated again, so that the code will try again, causing an potential
>> infinite loop. Temporarily set the to-be-migrated page's pageblock to
>> MIGRATE_ISOLATE to prevent that and bail out early if no free page is
>> found after page migration.
>>
>> An additional fix to split_free_page() aims to avoid crashing in
>> __free_one_page(). When the free page is split at the specified
>> split_pfn_offset, free_page_order should check both the first bit of
>> free_page_pfn and the last bit of split_pfn_offset and use the smaller one.
>> For example, if free_page_pfn=0x10000, split_pfn_offset=0xc000,
>> free_page_order should first be 0x8000 then 0x4000, instead of 0x4000 then
>> 0x8000, which the original algorithm did.
>>
>> ...
>>
>> --- a/mm/page_alloc.c
>> +++ b/mm/page_alloc.c
>> @@ -1114,13 +1114,16 @@ void split_free_page(struct page *free_page,
>> unsigned long flags;
>> int free_page_order;
>>
>> + if (split_pfn_offset == 0)
>> + return;
>> +
>> spin_lock_irqsave(&zone->lock, flags);
>> del_page_from_free_list(free_page, zone, order);
>> for (pfn = free_page_pfn;
>> pfn < free_page_pfn + (1UL << order);) {
>> int mt = get_pfnblock_migratetype(pfn_to_page(pfn), pfn);
>>
>> - free_page_order = ffs(split_pfn_offset) - 1;
>> + free_page_order = min(pfn ? __ffs(pfn) : order, __fls(split_pfn_offset));
>
> Why is it testing the zeroness of `pfn' here? Can pfn==0 even happen?
> If so, it's a legitimate value so why does it get special-cased?
__ffs() and __fls() are undefined if no bit exists, based on their
comments. I checked both pfn and split_pfn_offset against 0
just in case, even if pfn most likely is not going to be 0.
--
Best Regards,
Yan, Zi