[PATCH v1 0/2] udmabuf: Add back support for mapping hugetlb pages

Tue Jun 27 06:37:27 UTC 2023

Hi David,

> On 26.06.23 19:52, Peter Xu wrote:
> > On Mon, Jun 26, 2023 at 07:45:37AM +0000, Kasireddy, Vivek wrote:
> >> Hi Peter,
> >>
> >>>
> >>> On Fri, Jun 23, 2023 at 06:13:02AM +0000, Kasireddy, Vivek wrote:
> >>>> Hi David,
> >>>>
> >>>>>> The first patch ensures that the mappings needed for handling mmap
> >>>>>> operation would be managed by using the pfn instead of struct page.
> >>>>>> The second patch restores support for mapping hugetlb pages where
> >>>>>> subpages of a hugepage are not directly used anymore (main reason
> >>>>>> for revert) and instead the hugetlb pages and the relevant offsets
> >>>>>> are used to populate the scatterlist for dma-buf export and for
> >>>>>> mmap operation.
> >>>>>>
> >>>>>> Testcase: default_hugepagesz=2M hugepagesz=2M hugepages=2500
> >>>>> options
> >>>>>> were passed to the Host kernel and Qemu was launched with these
> >>>>>> relevant options: qemu-system-x86_64 -m 4096m....
> >>>>>> -device virtio-gpu-pci,max_outputs=1,blob=true,xres=1920,yres=1080
> >>>>>> -display gtk,gl=on
> >>>>>> -object memory-backend-memfd,hugetlb=on,id=mem1,size=4096M
> >>>>>> -machine memory-backend=mem1
> >>>>>>
> >>>>>> Replacing -display gtk,gl=on with -display gtk,gl=off above would
> >>>>>> exercise the mmap handler.
> >>>>>>
> >>>>>
> >>>>> While I think the VM_PFNMAP approach is much better and should fix
> >>> that
> >>>>> issue at hand, I thought more about missing memlock support and
> >>> realized
> >>>>> that we might have to fix something else. SO I'm going to raise the
> >>>>> issue here.
> >>>>>
> >>>>> I think udmabuf chose the wrong interface to do what it's doing, that
> >>>>> makes it harder to fix it eventually.
> >>>>>
> >>>>> Instead of accepting a range in a memfd, it should just have accepted a
> >>>>> user space address range and then used
> >>>>> pin_user_pages(FOLL_WRITE|FOLL_LONGTERM) to longterm-pin the
> >>> pages
> >>>>> "officially".
> >>>> Udmabuf indeed started off by using user space address range and GUP
> >>> but
> >>>> the dma-buf subsystem maintainer had concerns with that approach in
> v2.
> >>>> It also had support for mlock in that version. Here is v2 and the relevant
> >>>> conversation:
> >>>>
> https://patchwork.freedesktop.org/patch/210992/?series=39879&rev=2
> >>>>
> >>>>>
> >>>>> So what's the issue? Udma effectively pins pages longterm ("possibly
> >>>>> forever") simply by grabbing a reference on them. These pages might
> >>>>> easily reside in ZONE_MOVABLE or in MIGRATE_CMA pageblocks.
> >>>>>
> >>>>> So what udmabuf does is break memory hotunplug and CMA, because
> it
> >>>>> turns
> >>>>> pages that have to remain movable unmovable.
> >>>>>
> >>>>> In the pin_user_pages(FOLL_LONGTERM) case we make sure to
> migrate
> >>>>> these
> >>>>> pages. See mm/gup.c:check_and_migrate_movable_pages() and
> >>> especially
> >>>>> folio_is_longterm_pinnable(). We'd probably have to implement
> >>> something
> >>>>> similar for udmabuf, where we detect such unpinnable pages and
> >>> migrate
> >>>>> them.
> >>>> The pages udmabuf pins are only those associated with Guest (GPU
> >>> driver/virtio-gpu)
> >>>> resources (or buffers allocated and pinned from shmem via drm GEM).
> >>> Some
> >>>> resources are short-lived, and some are long-lived and whenever a
> >>> resource
> >>>> gets destroyed, the pages are unpinned. And, not all resources have
> their
> >>> pages
> >>>> pinned. The resource that is pinned for the longest duration is the FB
> and
> >>> that's
> >>>> because it is updated every ~16ms (assuming 1920x1080 at 60) by the
> Guest
> >>>> GPU driver. We can certainly pin/unpin the FB after it is accessed on the
> >>> Host
> >>>> as a workaround, but I guess that may not be very efficient given the
> >>> amount
> >>>> of churn it would create.
> >>>>
> >>>> Also, as far as migration or S3/S4 is concerned, my understanding is
> that all
> >>>> the Guest resources are destroyed and recreated again. So, wouldn't
> >>> something
> >>>> similar happen during memory hotunplug?
> >>>>
> >>>>>
> >>>>>
> >>>>> For example, pairing udmabuf with vfio (which pins pages using
> >>>>> pin_user_pages(FOLL_LONGTERM)) in QEMU will most probably not
> work
> >>> in
> >>>>> all cases: if udmabuf longterm pinned the pages "the wrong way", vfio
> >>>>> will fail to migrate them during FOLL_LONGTERM and consequently fail
> >>>>> pin_user_pages(). As long as udmabuf holds a reference on these
> pages,
> >>>>> that will never succeed.
> >>>> Dma-buf rules (for exporters) indicate that the pages only need to be
> >>> pinned
> >>>> during the map_attachment phase (and until unmap attachment
> happens).
> >>>> In other words, only when the sg_table is created by udmabuf. I guess
> one
> >>>> option would be to not hold any references during UDMABUF_CREATE
> and
> >>>> only grab references to the pages (as and when it gets used) during this
> >>> step.
> >>>> Would this help?
> >>>
> >>> IIUC the refcount is needed, otherwise I don't see what to protect the
> page
> >>> from being freed and even reused elsewhere before map_attachment().
> >>>
> >>> It seems the previous concern on using gup was majorly fork(), if this is
> it:
> >>>
> >>>
> https://patchwork.freedesktop.org/patch/210992/?series=39879&rev=2#co
> >>> mment_414213
> >>>
> >>> Could it also be guarded by just making sure the pages are MAP_SHARED
> >>> when
> >>> creating the udmabuf, if fork() is a requirement of the feature?
> >>>
> >>> I had a feeling that the userspace still needs to always do the right thing
> >>> to make it work, even using pure PFN mappings.
> >>>
> >>> For instance, what if the userapp just punchs a hole in the
> shmem/hugetlbfs
> >>> file after creating the udmabuf (I see that F_SEAL_SHRINK is required, but
> >>> at least not F_SEAL_WRITE with current impl), and fault a new page into
> the
> >>> page cache?
> >> IIUC, Qemu creates and owns the memfd that is associated with Guest
> memory.
> >> And if it punches a hole in its own memfd that goes through Guest pinned
> pages
> >> associated with buffers/resources, then I think the proper way to fix this is
> to
> >> somehow notify the Guest driver (virtio-gpu in this case) that the backing
> store
> >> of the affected resources is no longer valid and that the resources need to
> be
> >> destroyed and re-created again.
> >>
> >> Having said that, one option I could think of is to probably install a
> mmu_notifier
> >> associated with the relevant pages in udmabuf and once we get notified
> about
> >> any invalidation event concerning any of the pages, we'd fail any
> subsequent
> >> attempt to access these pages and propagate the error across the stack.
> >
> > Sounds right, maybe it needs to go back to the old GUP solution, though, as
> > mmu notifiers are also mm-based not fd-based. Or to be explicit, I think
> > it'll be pin_user_pages(FOLL_LONGTERM) with the new API.  It'll also solve
> > the movable pages issue on pinning.
> 
> It better should be pin_user_pages(FOLL_LONGTERM). But I'm afraid we
> cannot achieve that without breaking the existing kernel interface ...
Yeah, as you suggest, we unfortunately cannot go back to using GUP
without breaking udmabuf_create UAPI that expects memfds and file
offsets. 

> 
> So we might have to implement the same page migration as gup does on
> FOLL_LONGTERM here ... maybe there are more such cases/drivers that
> actually require that handling when simply taking pages out of the
> memfd, believing they can hold on to them forever.
IIUC, I don't think just handling the page migration in udmabuf is going to
cut it. It might require active cooperation of the Guest GPU driver as well
if this is even feasible.

> 
> >
> >>
> >> However, it feels like udmabuf is not the right place to handle this issue
> because
> >> there are very limited options for taking proper corrective action if Qemu
> decides
> >> to punch a hole in Guest memory that takes out pages that are pinned.
> >
> > I'm not familiar with the use case that much, but IMHO it's fine if the
> > driver relies on proper behaviors of userapp to work.
> >
> > IIUC the worst case here is the udmabuf holds some pages that are not the
> > pages of the guest mem anymore, but it only happens on misbehaved
> > userspace, then it looks all fine as long as they can at least be properly
> > released when releasing the udmabuf.  It'll be better if the udmabuf can
> > fail hard when detecting this, but IMHO even that can be optional
> depending
> > on the need, while any corrective action will be even one step further.
> 
> For vfio the issue are e.g., VM reboots, not a misbehaving guest. If the
> old guest kernel inflated the balloon and we reboot the VM, VFIO will
> still reference the old pages and the new (unaware kernel) might use
> that previously inflated memory to communicate with the devices. Because
> of that, we disable balloon inflation when vfio is active in QEMU.
> 
> [there are more cases like unloading the balloon driver or deflating the
> balloon, and then using that memory for communicating with the device]
> 
> Maybe it's all fine with udmabuf because of the way it is setup/torn
> down by the guest driver. Unfortunately I can't tell.
Here are the functions used by virtio-gpu (Guest GPU driver) to allocate
pages for its resources:
__drm_gem_shmem_create: https://elixir.bootlin.com/linux/latest/source/drivers/gpu/drm/drm_gem_shmem_helper.c#L97
Interestingly, the comment in the above function says that the pages
should not be allocated from the MOVABLE zone.
The pages along with their dma addresses are then extracted and shared
with Qemu using these two functions:
drm_gem_get_pages: https://elixir.bootlin.com/linux/latest/source/drivers/gpu/drm/drm_gem.c#L534
virtio_gpu_object_shmem_init: https://elixir.bootlin.com/linux/latest/source/drivers/gpu/drm/virtio/virtgpu_object.c#L135
Qemu then translates the dma addresses into file offsets and creates
udmabufs -- as an optimization to avoid data copies only if blob is set
to true.

I'll need to run some experiments to understand the impact of memfd page
migration (or replacement) on virtio-gpu driver's resources -- and also on the
rest of the Guest graphics stack.

Thanks,
Vivek
> 
> --
> Cheers,
> 
> David / dhildenb