On 9 Jan 2026, at 15:03, Matthew Brost wrote: > On Fri, Jan 09, 2026 at 02:23:49PM -0500, Zi Yan wrote: >> On 9 Jan 2026, at 14:08, Matthew Brost wrote: >> >>> On Fri, Jan 09, 2026 at 01:53:33PM -0500, Zi Yan wrote: >>>> On 9 Jan 2026, at 13:26, Matthew Brost wrote: >>>> >>>>> On Fri, Jan 09, 2026 at 12:28:22PM -0500, Zi Yan wrote: >>>>>> On 9 Jan 2026, at 6:09, Mika Penttilä wrote: >>>>>> >>>>>>> Hi, >>>>>>> >>>>>>> On 1/9/26 10:54, Francois Dugast wrote: >>>>>>> >>>>>>>> From: Matthew Brost <[email protected]> >>>>>>>> >>>>>>>> Split device-private and coherent folios into individual pages before >>>>>>>> freeing so that any order folio can be formed upon the next use of the >>>>>>>> pages. >>>>>>>> >>>>>>>> Cc: Balbir Singh <[email protected]> >>>>>>>> Cc: Alistair Popple <[email protected]> >>>>>>>> Cc: Zi Yan <[email protected]> >>>>>>>> Cc: David Hildenbrand <[email protected]> >>>>>>>> Cc: Oscar Salvador <[email protected]> >>>>>>>> Cc: Andrew Morton <[email protected]> >>>>>>>> Cc: [email protected] >>>>>>>> Cc: [email protected] >>>>>>>> Cc: [email protected] >>>>>>>> Signed-off-by: Matthew Brost <[email protected]> >>>>>>>> Signed-off-by: Francois Dugast <[email protected]> >>>>>>>> --- >>>>>>>> mm/memremap.c | 2 ++ >>>>>>>> 1 file changed, 2 insertions(+) >>>>>>>> >>>>>>>> diff --git a/mm/memremap.c b/mm/memremap.c >>>>>>>> index 63c6ab4fdf08..7289cdd6862f 100644 >>>>>>>> --- a/mm/memremap.c >>>>>>>> +++ b/mm/memremap.c >>>>>>>> @@ -453,6 +453,8 @@ void free_zone_device_folio(struct folio *folio) >>>>>>>> case MEMORY_DEVICE_COHERENT: >>>>>>>> if (WARN_ON_ONCE(!pgmap->ops || >>>>>>>> !pgmap->ops->folio_free)) >>>>>>>> break; >>>>>>>> + >>>>>>>> + folio_split_unref(folio); >>>>>>>> pgmap->ops->folio_free(folio); >>>>>>>> percpu_ref_put_many(&folio->pgmap->ref, nr); >>>>>>>> break; >>>>>>> >>>>>>> This breaks folio_free implementations like nouveau_dmem_folio_free >>>>>>> which checks the folio order and act upon that. >>>>>>> Maybe add an order parameter to folio_free or let the driver handle the >>>>>>> split? >>>>> >>>>> 'let the driver handle the split?' - I had consisder this as an option. >>>>> >>>>>> >>>>>> Passing an order parameter might be better to avoid exposing core MM >>>>>> internals >>>>>> by asking drivers to undo compound pages. >>>>>> >>>>> >>>>> It looks like Nouveau tracks free folios and free pages—something Xe’s >>>>> device memory allocator (DRM Buddy) cannot do. I guess this answers my >>>>> earlier question of how Nouveau avoids hitting the same bug as Xe / GPU >>>>> SVM with respect to reusing folios. It appears Nouveau prefers not to >>>>> split the folio, so I’m leaning toward moving this call into the >>>>> driver’s folio_free function. >>>> >>>> No, that creates asymmetric page handling and is error prone. >>>> >>> >>> I agree it is asymmetric and symmetric is likely better. >>> >>>> In addition, looking at nouveau’s implementation in >>>> nouveau_dmem_page_alloc_locked(), it gets a folio from >>>> drm->dmem->free_folios, >>>> which is never split, and passes it to zone_device_folio_init(). This >>>> is wrong, since if the folio is large, it will go through >>>> prep_compound_page() >>>> again. The bug has not manifested because there is only order-9 large >>>> folios. >>>> Once mTHP support is added, how is nouveau going to allocate a order-4 >>>> folio >>>> from a free order-9 folio? Maintain a per-order free folio list and >>>> reimplement a buddy allocator? Nevertheless, nouveau’s implementation >>> >>> The way Nouveau handles memory allocations here looks wrong to me—it >>> should probably use DRM Buddy and convert a block buddy to pages rather >>> than tracking a free folio list and free page list. But this is not my >>> driver. >>> >>>> is wrong by calling prep_compound_page() on a folio (already compound >>>> page). >>>> >>> >>> I don’t disagree that this implementation is questionable. >>> >>> So what’s the suggestion here—add folio order to folio_free just to >>> accommodate Nouveau’s rather odd memory allocation algorithm? That >>> doesn’t seem right to me either. >> >> Splitting the folio in free_zone_device_folio() and passing folio order >> to folio_free() make sense to me, since after the split, the folio passed > > If this is concensous / direction - I can do this but a tree wide > change. > > I do have another question for everyone here - do we think this spliting > implementation should be considered a Fixes so this can go into 6.19?
IMHO, this should be a fix, since it is wrong to call prep_compound_page() on a large folio. IIUC this seems to only affect nouveau now, I will let them to decide. > >> to folio_free() contains no order information, but just the used-to-be >> head page and the remaining 511 pages are free. How does Intel Xe driver >> handle it without knowing folio order? >> > > It’s a bit convoluted, but folio/page->zone_device_data points to a > reference-counted object in GPU SVM. When the object’s reference count > drops to zero, we callback into the driver layer to release the memory. > In Xe, this is a TTM BO that resolves to a DRM Buddy allocation, which > is then released. If it’s not clear, our original allocation size > determines the granularity at which we free the backing store. > >> Do we really need the order info in ->folio_free() if the folio is split >> in free_zone_device_folio()? free_zone_device_folio() should just call >> ->folio_free() 2^order times to free individual page. >> > > No. If it’s a higher-order folio—let’s say a 2MB folio—we have one > reference to our GPU SVM object, so we can free the backing in a single > ->folio_free call. > > Now, if that folio gets split at some point into 4KB pages, then we’d > have 512 references to this object set up in the ->folio_split calls. > We’d then expect 512 ->folio_free() calls. Same case here: if, for > whatever reason, we can’t create a 2MB device page during a 2MB > migration and need to create 512 4KB pages so we'd have 512 references > to our GPU SVM object. Thank you for the explanation. Adding folio order to ->folio_free() makes sense to me now. Best Regards, Yan, Zi
