[PATCH RESEND v2 4/6] mm/page_alloc: sort out the alloc_contig_range() gfp flags mess

[David Hildenbrand]

On 04.12.24 11:04, Oscar Salvador wrote:
> On Wed, Dec 04, 2024 at 10:28:39AM +0100, David Hildenbrand wrote:
>> On 04.12.24 10:15, Oscar Salvador wrote:
>>> On Wed, Dec 04, 2024 at 10:03:28AM +0100, Vlastimil Babka wrote:
>>>> On 12/4/24 09:59, Oscar Salvador wrote:
>>>>> On Tue, Dec 03, 2024 at 08:19:02PM +0100, David Hildenbrand wrote:
>>>>>> It was always set using "GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL",
>>>>>> and I removed the same flag combination in #2 from memory offline code, and
>>>>>> we do have the exact same thing in do_migrate_range() in
>>>>>> mm/memory_hotplug.c.
>>>>>>
>>>>>> We should investigate if__GFP_HARDWALL is the right thing to use here, and
>>>>>> if we can get rid of that by switching to GFP_KERNEL in all these places.
>>>>>
>>>>> Why would not we want __GFP_HARDWALL set?
>>>>> Without it, we could potentially migrate the page to a node which is not
>>>>> part of the cpuset of the task that originally allocated it, thus violating the
>>>>> policy? Is not that a problem?
>>>>
>>>> The task doing the alloc_contig_range() will likely not be the same task as
>>>> the one that originally allocated the page, so its policy would be
>>>> different, no? So even with __GFP_HARDWALL we might be already migrating
>>>> outside the original tasks's constraint? Am I missing something?
>>>
>>> Yes, that is right, I thought we derive the policy from the old page
>>> somehow when migrating it, but reading the code does not seem to be the
>>> case.
>>>
>>> Looking at prepare_alloc_pages(), if !ac->nodemask, which would be the
>>> case here, we would get the policy from the current task
>>> (alloc_contig_range()) when cpusets are enabled.
>>>
>>> So yes, I am a bit puzzled why __GFP_HARDWALL was chosen in the first
>>> place.
>>
>> I suspect because "GFP_USER" felt like the appropriate thing to do.
> 
> Looking back at when the whole contiguous allocator patchset was posted,
> it seems that it kinda copied what memory-offline code was doing, which
> was migrating pages with GFP_HIGHUSER_MOVABLE (hotremove_migrate_alloc()).
> 
> Then, the HIGHMEM modifier was dropped due to HIGHMEM restrictions on
> some systems, ending up with GFP_USER.

Looking at some other migration_target_control users, some of them also 
shouldn't be setting GFP_USER->HARDWALL either I think. Essentially, 
whenever we are migrating a page that is not guaranteed to be "ours" in 
the context of the caller.

mm/damon/paddr.c:__damon_pa_migrate_folio_list() for example, which 
obtained the addresses by scanning a chunk of physical address space.

For others it's less clear: soft_offline_in_use_page() may be called 
either using madvise() from process context, but also using sysfs using 
a PFN.


-- 
Cheers,

David / dhildenb