Re: [PATCH] mm: page_alloc: Assume huge tail pages are valid when allocating contiguous pages

[David Hildenbrand <david@redhat.com>]

On 14.04.23 12:19, Michal Hocko wrote:
> On Fri 14-04-23 10:52:04, Mel Gorman wrote:
>> On Fri, Apr 14, 2023 at 10:55:04AM +0200, Michal Hocko wrote:
>>> On Fri 14-04-23 09:22:22, Mel Gorman wrote:
>>> [...]
>>>> +
>>>> +		/*
>>>> +		 * Do not migrate huge pages that span the size of the region
>>>> +		 * being allocated contiguous. e.g. Do not migrate a 1G page
>>>> +		 * for a 1G allocation request. CMA is an exception as the
>>>> +		 * region may be reserved for hardware that requires physical
>>>> +		 * memory without a MMU or scatter/gather capability.
>>>> +		 *
>>>> +		 * Note that the compound check is race-prone versus
>>>> +		 * free/split/collapse but it should be safe and result in
>>>> +		 * a premature skip or a useless migration attempt.
>>>> +		 */
>>>> +		if (PageHuge(page) && compound_nr(page) >= nr_pages &&
>>>> +		    !is_migrate_cma_page(page)) {
>>>> +			return false;
>>>
>>> Is the CMA check working as expected?
>>
>> I didn't test it as I don't have a good simulator for CMA contraints which
>> is still a mobile phone concern for devices like cameras.
>>
>>> The function sounds quite generic
>>> and I agree that it would make sense if it was generic but it is used
>>> only for GB pages in fact and unless I am missing something it would
>>> allow to migrate CMA pages and potentially allocate over that region
>>> without any possibility to migrate GB page out so the CMA region would
>>> be essentially unusable for CMA users.
>>
>> It's used primarily for 1G pages but does have other users (debugging
>> mostly, low priority). As it's advertised as a general API, I decided to
>> treat it as such and that meant being nice to CMA if possible. If CMA pages
>> migrate but can still use the target location then it should be fine. If a
>> CMA can migrate to an usable location that breaks a device then that's a bug.
>>
>>> GB pages already have their CMA
>>> allocator path before we get to alloc_contig_pages. Or do I miss
>>> something?
>>
>> I don't think you missed anything. The CMA check is, at best, an effort
>> to have a potentially useful semantic but it's very doubtful anyone will
>> notice or care. I'm perfectly happy just to drop the CMA check because it's a
>> straight-forward fix and more suitable as a -stable backport.  I'm also happy
>> to just go with a PageHuge check and ignore any possibility that a 2M page
>> could be migrated to satisfy a 1G allocation.  1G allocation requests after
>> significant uptime is a crapshoot at best and relying on them succeeding is
>> unwise. There is a non-zero possibility that the latency incurred migrating
>> 2M pages and still failing a 1G allocation could itself be classed as a
>> bug with users preferring fast-failure of 1G allocation attempts.
> 
> Yes, the simpler the better. If we encounter a real usecase where couple
> of 2MB hugetlb pages stand in the way to GB pages then we can add the
> check so I would just go with reintroducing the PageHuge check alone.

alloc_contig_pages() -> __alloc_contig_pages() -> 
alloc_contig_range(MIGRATE_MOVABLE)

Should always fail when stumbling over MIGRATE_CMA pageblocks IIRC.

So we could bail out in that function early if we stumble over any CMA 
region.

-- 
Thanks,

David / dhildenb