Re: [RFC PATCH 1/4] mm: use current as mmu notifier's owner

["Mika Penttilä" <mpenttil@redhat.com>]

On 8/14/25 17:11, Jason Gunthorpe wrote:

> On Thu, Aug 14, 2025 at 04:20:42PM +0300, Mika Penttilä wrote:
>> On 8/14/25 16:04, Jason Gunthorpe wrote:
>>
>>> On Thu, Aug 14, 2025 at 03:53:00PM +0300, Mika Penttilä wrote:
>>>> On 8/14/25 15:40, Jason Gunthorpe wrote:
>>>>> On Thu, Aug 14, 2025 at 10:19:26AM +0300, Mika Penttilä wrote:
>>>>>> When doing migration in combination with device fault handling,
>>>>>> detect the case in the interval notifier.
>>>>>>
>>>>>> Without that, we would livelock with our own invalidations
>>>>>> while migrating and splitting pages during fault handling.
>>>>>>
>>>>>> Note, pgmap_owner, used in some other code paths as owner for filtering,
>>>>>> is not readily available for split path, so use current for this use case.
>>>>>> Also, current and pgmap_owner, both being pointers to memory, can not be
>>>>>> mis-interpreted to each other.
>>>>>>
>>>>>> Cc: David Hildenbrand <david@redhat.com>
>>>>>> Cc: Jason Gunthorpe <jgg@nvidia.com>
>>>>>> Cc: Leon Romanovsky <leonro@nvidia.com>
>>>>>> Cc: Alistair Popple <apopple@nvidia.com>
>>>>>> Cc: Balbir Singh <balbirs@nvidia.com>
>>>>>>
>>>>>> Signed-off-by: Mika Penttilä <mpenttil@redhat.com>
>>>>>> ---
>>>>>>  lib/test_hmm.c   | 5 +++++
>>>>>>  mm/huge_memory.c | 6 +++---
>>>>>>  mm/rmap.c        | 4 ++--
>>>>>>  3 files changed, 10 insertions(+), 5 deletions(-)
>>>>>>
>>>>>> diff --git a/lib/test_hmm.c b/lib/test_hmm.c
>>>>>> index 761725bc713c..cd5c139213be 100644
>>>>>> --- a/lib/test_hmm.c
>>>>>> +++ b/lib/test_hmm.c
>>>>>> @@ -269,6 +269,11 @@ static bool dmirror_interval_invalidate(struct mmu_interval_notifier *mni,
>>>>>>  	    range->owner == dmirror->mdevice)
>>>>>>  		return true;
>>>>>>  
>>>>>> +	if (range->event == MMU_NOTIFY_CLEAR &&
>>>>>> +	    range->owner == current) {
>>>>>> +		return true;
>>>>>> +	}
>>>>> I don't understand this, there is nothing in hmm that says only
>>>>> current can call hmm_range_fault, and indeed most applications won't
>>>>> even gurantee that.
>>>> No it's the opposite, if we are ourselves the invalidator, don't care.
>>> I think you've missed the point, you cannot use range->owner in any
>>> way to tell "we are ourselves the invalidator". It is simply not the
>>> meaning of range->owner.
>> Usually it is the device but used similarly, look for instance nouveau:
> Yes, dev is fine, but current or struct task is not reasonable.
>
>> Yes the commit message could be better. 
>> But this is essentially the same as nouveau is doing with 
>> MMU_NOTIFY_EXCLUSIVE handling, just not using the dev as the
>> qualifier,
> I wouldn't necesarily assume anything nouveau is correct, but this is
> certainly not the same thing. nouveau is trying to eliminate an
> unncessary invalidation for it's HW when it knows the memory is
> already only private to this local device.

Yes but it is omitting
  mmu_interval_set_seq()
in that case while returning early, which is ok.

as well as hmm test module with :

         * Ignore invalidation callbacks for device private pages since
         * the invalidation is handled as part of the migration process.
         */
        if (range->event == MMU_NOTIFY_MIGRATE &&
            range->owner == dmirror->mdevice)
                return true;


> This is not a statement about callchain or recursion.

In my case the unnecessary invalidation (self caused) causes hang,
so whatever the reason for unnecessary, the mechanism is the same
and used elsewhere in kernel also.

So technically I think that would work. But there are others ways to do that
without using owner, so will look that way in next version.

>
> Jason
>

--Mika