Re: [PATCH v1 2/2] mm/migrate: move NUMA hinting fault folio isolation + checks under PTL

[David Hildenbrand <david@redhat.com>]

On 21.06.24 22:48, Zi Yan wrote:
> On 21 Jun 2024, at 16:18, David Hildenbrand wrote:
> 
>> On 21.06.24 15:44, Zi Yan wrote:
>>> On 20 Jun 2024, at 17:29, David Hildenbrand wrote:
>>>
>>>> Currently we always take a folio reference even if migration will not
>>>> even be tried or isolation failed, requiring us to grab+drop an additional
>>>> reference.
>>>>
>>>> Further, we end up calling folio_likely_mapped_shared() while the folio
>>>> might have already been unmapped, because after we dropped the PTL, that
>>>> can easily happen. We want to stop touching mapcounts and friends from
>>>> such context, and only call folio_likely_mapped_shared() while the folio
>>>> is still mapped: mapcount information is pretty much stale and unreliable
>>>> otherwise.
>>>>
>>>> So let's move checks into numamigrate_isolate_folio(), rename that
>>>> function to migrate_misplaced_folio_prepare(), and call that function
>>>> from callsites where we call migrate_misplaced_folio(), but still with
>>>> the PTL held.
>>>>
>>>> We can now stop taking temporary folio references, and really only take
>>>> a reference if folio isolation succeeded. Doing the
>>>> folio_likely_mapped_shared() + golio isolation under PT lock is now similar
>>>> to how we handle MADV_PAGEOUT.
>>>>
>>>> While at it, combine the folio_is_file_lru() checks.
>>>>
>>>> Signed-off-by: David Hildenbrand <david@redhat.com>
>>>> ---
>>>>    include/linux/migrate.h |  7 ++++
>>>>    mm/huge_memory.c        |  8 ++--
>>>>    mm/memory.c             |  9 +++--
>>>>    mm/migrate.c            | 81 +++++++++++++++++++----------------------
>>>>    4 files changed, 55 insertions(+), 50 deletions(-)
>>>
>>> LGTM. Reviewed-by: Zi Yan <ziy@nvidia.com>
>>>
>>> One nit below:
>>>
>>> <snip>
>>>
>>>> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
>>>> index fc27dabcd8e3..4b2817bb2c7d 100644
>>>> --- a/mm/huge_memory.c
>>>> +++ b/mm/huge_memory.c
>>>> @@ -1688,11 +1688,13 @@ vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
>>>>    	if (node_is_toptier(nid))
>>>>    		last_cpupid = folio_last_cpupid(folio);
>>>>    	target_nid = numa_migrate_prep(folio, vmf, haddr, nid, &flags);
>>>> -	if (target_nid == NUMA_NO_NODE) {
>>>> -		folio_put(folio);
>>>> +	if (target_nid == NUMA_NO_NODE)
>>>> +		goto out_map;
>>>> +	if (migrate_misplaced_folio_prepare(folio, vma, target_nid)) {
>>>> +		flags |= TNF_MIGRATE_FAIL;
>>>>    		goto out_map;
>>>>    	}
>>>> -
>>>> +	/* The folio is isolated and isolation code holds a folio reference. */
>>>>    	spin_unlock(vmf->ptl);
>>>>    	writable = false;
>>>>
>>>> diff --git a/mm/memory.c b/mm/memory.c
>>>> index 118660de5bcc..4fd1ecfced4d 100644
>>>> --- a/mm/memory.c
>>>> +++ b/mm/memory.c
>>>
>>> <snip>
>>>
>>>> @@ -5345,10 +5343,13 @@ static vm_fault_t do_numa_page(struct vm_fault *vmf)
>>>>    	else
>>>>    		last_cpupid = folio_last_cpupid(folio);
>>>>    	target_nid = numa_migrate_prep(folio, vmf, vmf->address, nid, &flags);
>>>> -	if (target_nid == NUMA_NO_NODE) {
>>>> -		folio_put(folio);
>>>> +	if (target_nid == NUMA_NO_NODE)
>>>> +		goto out_map;
>>>> +	if (migrate_misplaced_folio_prepare(folio, vma, target_nid)) {
>>>> +		flags |= TNF_MIGRATE_FAIL;
>>>>    		goto out_map;
>>>>    	}
>>>
>>> These two locations are repeated code, maybe just merge the ifs into
>>> numa_migrate_prep(). Feel free to ignore if you are not going to send
>>> another version. :)
>>
>> I went back and forth a couple of times and
>>
>> a) Didn't want to move numa_migrate_prep() into
>>     migrate_misplaced_folio_prepare(), because having that code in
>>     mm/migrate.c felt a bit odd.
> 
> I agree after checking the actual code, since the code is just
> updating NUMA fault stats and checking where the folio should be.
> 
>>
>> b) Didn't want to move migrate_misplaced_folio_prepare() because I enjoy
>>     seeing the migrate_misplaced_folio_prepare() and
>>     migrate_misplaced_folio() calls in the same callercontext.
>>
>> I also considered renaming numa_migrate_prep(), but wasn't really able to come up with a good name.
> 
> How about numa_migrate_check()? Since it tells whether a folio should be
> migrated or not.
> 
>>
>> But maybe a) is not too bad?
>>
>> We'd have migrate_misplaced_folio_prepare() consume &flags and &target_nid, and perform the "flags |= TNF_MIGRATE_FAIL;" internally.
>>
>> What would be your take?
> 
> I would either rename numa_migrate_prep() or just do nothing. I have to admit
> that the "prep" and "prepare" in both function names motivated me to propose
> the merge, but now the actual code tells me they should be separate.

Let's leave it like that for now. Renaming to numa_migrate_check() makes 
sense, and likely moving more numa handling stuff in there.

Bit I yet have to figure out why some of the memory.c vs. huge_memory.c 
code differences exist, so we can unify them.

For example, why did 33024536bafd9 introduce slightly different 
last_cpupid handling in do_huge_pmd_numa_page(), whereby it seems like 
some subtle difference in handling NUMA_BALANCING_MEMORY_TIERING? Maybe 
I am missing something obvious. :)

-- 
Cheers,

David / dhildenb