On Sat, Jan 10, 2026 at 09:36:04AM +1100, Balbir Singh wrote: > On 1/10/26 08:14, Zi Yan wrote: > > On 9 Jan 2026, at 17:11, Balbir Singh wrote: > > > >> On 1/10/26 07:43, Zi Yan wrote: > >>> On 9 Jan 2026, at 16:34, Balbir Singh wrote: > >>> > >>>> On 1/10/26 06:15, Zi Yan wrote: > >>>>> 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. > >>>>> > >>>> > >>>> Agreed, free_zone_device_folio() needs to split the folio on put. > >>>> > >>>> > >>>>>> > >>>>>>> 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. > >>>>> > >>>> > >>>> I still don't follow why the folio_order does not capture the order of > >>>> the folio. > >>>> If the folio is split, we should now have 512 split folios for THP > >>> > >>> folio_order() should return 0 after the folio is split. > >>> > >>> In terms of the number of after-split folios, it is 512 for current code > >>> base > >>> since THP is only 2MB in zone devices, but not future proof if mTHP > >>> support > >>> is added. It also causes confusion in core MM, where folio can have > >>> all kinds of orders. > >>> > >>> > >> > >> I see that folio_split_unref() to see that there is no driver > >> callback during the split. Patch 3 controls the order of > >> > >> + folio_split_unref(folio); > >> pgmap->ops->folio_free(folio); > >> > >> @Matthew, is there a reason to do the split prior to free? > >> pgmap->ops->folio_free(folio) > >> shouldn't impact the folio itself, the backing memory can be freed and > >> then the > >> folio split? > > > > Quote Matthew from [1]: > > > > ... this step must be done before calling folio_free and include a barrier,
Actually, I think it’s fine without a barrier—I confused myself a bit there. But yes, it must be split before releasing the memory back to the pool from which it can be reallocated. > > as the page can be immediately reallocated. > > > > [1] https://lore.kernel.org/all/[email protected]/ > > > > Thanks, I am not a TTM/BO expert > > So that leaves us with > > 1. Pass the order to folio_free() > 2. Consider calling folio_free() callback for each split folio during > folio_split_unref(), > but that means the driver needs to consolidate all the relevant information > > #1 works, but the information there is stale, in the sense that we are > passing in the > old order information, the order is useful for the driver to know the size of > it's > backing allocation #1 is my preference here. We don’t need this information in GPU SVM for Xe, but Nouveau does, and I see a straightforward change in Nouveau. In this case, “order” means the folio plus the number of pages being released, with each individual page in an initialized state (i.e., not compound and with a proper pgmap value, they look the pages output from memremap_pages, etc...). I think this interface actually makes sense now that I’ve written it down. My next revision will implement this along with the renaming suggestions for s/folio_split_unref/free_zone_device_folio_prepare agreed upon in patch #1. I’ll also likely mark the relevant core MM patches wirh fixes tags in my next revision so 6.19 has the correct folio-splitting behavior - it would be a bit odd to have kernel floating around with different behavior here. Let me know if anyone has objections before I move forward with this. Matt > #2 should work too, but it means PMD_ORDER frees as opposed to 1 > > Balbir
