(resend because gmail did not format to plain text...)
On Thu, Feb 20, 2020 at 8:45 PM Chia-I Wu <olvaffe@xxxxxxxxx> wrote:
>
>
>
> On Thu, Feb 20, 2020 at 4:23 PM Tian, Kevin <kevin.tian@xxxxxxxxx> wrote:
>>
>> > From: Chia-I Wu <olvaffe@xxxxxxxxx>
>> > Sent: Friday, February 21, 2020 6:24 AM
>> >
>> > On Wed, Feb 19, 2020 at 6:38 PM Tian, Kevin <kevin.tian@xxxxxxxxx> wrote:
>> > >
>> > > > From: Tian, Kevin
>> > > > Sent: Thursday, February 20, 2020 10:05 AM
>> > > >
>> > > > > From: Chia-I Wu <olvaffe@xxxxxxxxx>
>> > > > > Sent: Thursday, February 20, 2020 3:37 AM
>> > > > >
>> > > > > On Wed, Feb 19, 2020 at 1:52 AM Tian, Kevin <kevin.tian@xxxxxxxxx>
>> > wrote:
>> > > > > >
>> > > > > > > From: Paolo Bonzini
>> > > > > > > Sent: Wednesday, February 19, 2020 12:29 AM
>> > > > > > >
>> > > > > > > On 14/02/20 23:03, Sean Christopherson wrote:
>> > > > > > > >> On Fri, Feb 14, 2020 at 1:47 PM Chia-I Wu <olvaffe@xxxxxxxxx>
>> > > > wrote:
>> > > > > > > >>> AFAICT, it is currently allowed on ARM (verified) and AMD (not
>> > > > > > > >>> verified, but svm_get_mt_mask returns 0 which supposedly
>> > means
>> > > > > the
>> > > > > > > NPT
>> > > > > > > >>> does not restrict what the guest PAT can do). This diff would do
>> > the
>> > > > > > > >>> trick for Intel without needing any uapi change:
>> > > > > > > >> I would be concerned about Intel CPU errata such as SKX40 and
>> > > > SKX59.
>> > > > > > > > The part KVM cares about, #MC, is already addressed by forcing
>> > UC
>> > > > for
>> > > > > > > MMIO.
>> > > > > > > > The data corruption issue is on the guest kernel to correctly use
>> > WC
>> > > > > > > > and/or non-temporal writes.
>> > > > > > >
>> > > > > > > What about coherency across live migration? The userspace
>> > process
>> > > > > would
>> > > > > > > use cached accesses, and also a WBINVD could potentially corrupt
>> > guest
>> > > > > > > memory.
>> > > > > > >
>> > > > > >
>> > > > > > In such case the userspace process possibly should conservatively use
>> > > > > > UC mapping, as if for MMIO regions on a passthrough device.
>> > However
>> > > > > > there remains a problem. the definition of KVM_MEM_DMA implies
>> > > > > > favoring guest setting, which could be whatever type in concept. Then
>> > > > > > assuming UC is also problematic. I'm not sure whether inventing
>> > another
>> > > > > > interface to query effective memory type from KVM is a good idea.
>> > There
>> > > > > > is no guarantee that the guest will use same type for every page in the
>> > > > > > same slot, then such interface might be messy. Alternatively, maybe
>> > > > > > we could just have an interface for KVM userspace to force memory
>> > type
>> > > > > > for a given slot, if it is mainly used in para-virtualized scenarios (e.g.
>> > > > > > virtio-gpu) where the guest is enlightened to use a forced type (e.g.
>> > WC)?
>> > > > > KVM forcing the memory type for a given slot should work too. But the
>> > > > > ignore-guest-pat bit seems to be Intel-specific. We will need to
>> > > > > define how the second-level page attributes combine with the guest
>> > > > > page attributes somehow.
>> > > >
>> > > > oh, I'm not aware of that difference. without an ipat-equivalent
>> > > > capability, I'm not sure how to forcing random type here. If you look at
>> > > > table 11-7 in Intel SDM, none of MTRR (EPT) memory type can lead to
>> > > > consistent effective type when combining with random PAT value. So
>> > > > it is definitely a dead end.
>> > > >
>> > > > >
>> > > > > KVM should in theory be able to tell that the userspace region is
>> > > > > mapped with a certain memory type and can force the same memory
>> > type
>> > > > > onto the guest. The userspace does not need to be involved. But that
>> > > > > sounds very slow? This may be a dumb question, but would it help to
>> > > > > add KVM_SET_DMA_BUF and let KVM negotiate the memory type with
>> > the
>> > > > > in-kernel GPU drivers?
>> > > > >
>> > > > >
>> > > >
>> > > > KVM_SET_DMA_BUF looks more reasonable. But I guess we don't need
>> > > > KVM to be aware of such negotiation. We can continue your original
>> > > > proposal to have KVM simply favor guest memory type (maybe still call
>> > > > KVM_MEM_DMA). On the other hand, Qemu should just mmap on the
>> > > > fd handle of the dmabuf passed from the virtio-gpu device backend, e.g.
>> > > > to conduct migration. That way the mmap request is finally served by
>> > > > DRM and underlying GPU drivers, with proper type enforced
>> > automatically.
>> > > >
>> > >
>> > > Thinking more possibly we don't need introduce new interface to KVM.
>> > > As long as Qemu uses dmabuf interface to mmap the specific region,
>> > > KVM can simply check memory type in host page table given hva of a
>> > > memslot. If the type is UC or WC, it implies that userspace wants a
>> > > non-coherent mapping which should be reflected in the guest side too.
>> > > In such case, KVM can go to non-cohenrent DMA path and favor guest
>> > > memory type automatically.
>> > Sorry, I mixed two things together.
>> >
>> > Userspace access to dmabuf mmap must be guarded by
>> > DMA_BUF_SYNC_{START,END} ioctls. It is possible that the GPU driver
>> > always picks a WB mapping and let the ioctls flush/invalidate CPU
>> > caches. We actually want the guest memory type to match vkMapMemory's
>> > memory type, which can be different from dmabuf mmap's memory type.
>> > It is not enough for KVM to inspect the hva's memory type.
>>
>> I'm not familiar with dmabuf and what is the difference between
>> vkMapMemory and mmap. Just a simple thought that whatever
>> memory type/synchronization enforced on the host userspace should
>> ideally be applied to guest userspace too. e.g. in above example we
>> possibly want the guest to use WB and issue flush/invalidate hypercalls
>> to guard with other potential parallel operations in the host side.
>> otherwise I cannot see how synchronization can be done when one
>> use WB with sync primitives while the other simply use WC w/o such
>> primitives.
>
I am reasonably familiar with the GPU stacks, but I am not familiar with KVM
:)
When allocating a GPU memory, the userspace can specify whether it wants a
coherent one or an incoherent one. vkMapMemory returns a coherent or a
incoherent mapping respectively. Indeed we also want the guest userspace to
have a coherent or a incoherent mapping respectively.
The GPU memory can be exported as a dmabuf to share with another device or
process. For security, we allocate the GPU memory in a GPU process and we
export the dmabuf to the hypervisor. mmap of dmabuf semantically returns an
incoherent mapping. As a result, the guest will set up a mapping that has the
same memory type as the vkMapMemory mapping does, but the hva in
KVM_SET_USER_MEMORY_REGION points to the dmabuf's incoherent mapping.
If you think it is the best for KVM to inspect hva to determine the memory
type with page granularity, that is reasonable and should work for us too.
The userspace can do something (e.g., add a GPU driver dependency to the
hypervisor such that the dma-buf is imported as a GPU memory and mapped using
vkMapMemory) or I can work with dma-buf maintainers to see if dma-buf's
semantics can be changed.
>
>>
>> >
>> > KVM_SET_DMA_BUF, if supported, is a signal to KVM that the guest
>> > memory type should be honored (or forced if there is a new op in
>> > dma_buf_ops that tells KVM which memory type to force). KVM_MEM_DMA
>> > flag in this RFC sends the same signal. Unless KVM_SET_DMA_BUF gives
>> > the userspace other features such as setting unlimited number of
>> > dmabufs to subregions of a memslot, it is not very useful.
>>
>> the good part of a new interface is its simplicity, but only in slot
>> granularity. instead having KVM to inspect hva can support page
>> granularity, but adding run-time overhead. Let's see how Paolo
>> thinks.
>>
>> >
>> > If uapi change is to be avoided, it is the easiest that guest memory
>> > type is always honored unless it causes #MC (i.e.,is_mmio==true).
>> >
>>
>> I feel this goes too far...
>>
>> Thanks
>> Kevin
