Re: Re: [RFC v2 09/13] vduse: Add support for processing vhost iotlb message

[Yongji Xie <xieyongji@bytedance.com>]

On Thu, Dec 31, 2020 at 1:50 PM Jason Wang <jasowang@redhat.com> wrote:
>
>
> On 2020/12/31 下午1:15, Yongji Xie wrote:
> > On Thu, Dec 31, 2020 at 10:49 AM Jason Wang <jasowang@redhat.com> wrote:
> >>
> >> On 2020/12/30 下午6:12, Yongji Xie wrote:
> >>> On Wed, Dec 30, 2020 at 4:41 PM Jason Wang <jasowang@redhat.com> wrote:
> >>>> On 2020/12/30 下午3:09, Yongji Xie wrote:
> >>>>> On Wed, Dec 30, 2020 at 2:11 PM Jason Wang <jasowang@redhat.com> wrote:
> >>>>>> On 2020/12/29 下午6:26, Yongji Xie wrote:
> >>>>>>> On Tue, Dec 29, 2020 at 5:11 PM Jason Wang <jasowang@redhat.com> wrote:
> >>>>>>>> ----- Original Message -----
> >>>>>>>>> On Mon, Dec 28, 2020 at 4:43 PM Jason Wang <jasowang@redhat.com> wrote:
> >>>>>>>>>> On 2020/12/28 下午4:14, Yongji Xie wrote:
> >>>>>>>>>>>> I see. So all the above two questions are because VHOST_IOTLB_INVALIDATE
> >>>>>>>>>>>> is expected to be synchronous. This need to be solved by tweaking the
> >>>>>>>>>>>> current VDUSE API or we can re-visit to go with descriptors relaying
> >>>>>>>>>>>> first.
> >>>>>>>>>>>>
> >>>>>>>>>>> Actually all vdpa related operations are synchronous in current
> >>>>>>>>>>> implementation. The ops.set_map/dma_map/dma_unmap should not return
> >>>>>>>>>>> until the VDUSE_UPDATE_IOTLB/VDUSE_INVALIDATE_IOTLB message is replied
> >>>>>>>>>>> by userspace. Could it solve this problem?
> >>>>>>>>>>       I was thinking whether or not we need to generate IOTLB_INVALIDATE
> >>>>>>>>>> message to VDUSE during dma_unmap (vduse_dev_unmap_page).
> >>>>>>>>>>
> >>>>>>>>>> If we don't, we're probably fine.
> >>>>>>>>>>
> >>>>>>>>> It seems not feasible. This message will be also used in the
> >>>>>>>>> virtio-vdpa case to notify userspace to unmap some pages during
> >>>>>>>>> consistent dma unmapping. Maybe we can document it to make sure the
> >>>>>>>>> users can handle the message correctly.
> >>>>>>>> Just to make sure I understand your point.
> >>>>>>>>
> >>>>>>>> Do you mean you plan to notify the unmap of 1) streaming DMA or 2)
> >>>>>>>> coherent DMA?
> >>>>>>>>
> >>>>>>>> For 1) you probably need a workqueue to do that since dma unmap can
> >>>>>>>> be done in irq or bh context. And if usrspace does't do the unmap, it
> >>>>>>>> can still access the bounce buffer (if you don't zap pte)?
> >>>>>>>>
> >>>>>>> I plan to do it in the coherent DMA case.
> >>>>>> Any reason for treating coherent DMA differently?
> >>>>>>
> >>>>> Now the memory of the bounce buffer is allocated page by page in the
> >>>>> page fault handler. So it can't be used in coherent DMA mapping case
> >>>>> which needs some memory with contiguous virtual addresses. I can use
> >>>>> vmalloc() to do allocation for the bounce buffer instead. But it might
> >>>>> cause some memory waste. Any suggestion?
> >>>> I may miss something. But I don't see a relationship between the
> >>>> IOTLB_UNMAP and vmalloc().
> >>>>
> >>> In the vmalloc() case, the coherent DMA page will be taken from the
> >>> memory allocated by vmalloc(). So IOTLB_UNMAP is not needed anymore
> >>> during coherent DMA unmapping because those vmalloc'ed memory which
> >>> has been mapped into userspace address space during initialization can
> >>> be reused. And userspace should not unmap the region until we destroy
> >>> the device.
> >>
> >> Just to make sure I understand. My understanding is that IOTLB_UNMAP is
> >> only needed when there's a change the mapping from IOVA to page.
> >>
> > Yes, that's true.
> >
> >> So if we stick to the mapping, e.g during dma_unmap, we just put IOVA to
> >> free list to be used by the next IOVA allocating. IOTLB_UNMAP could be
> >> avoided.
> >>
> >> So we are not limited by how the pages are actually allocated?
> >>
> > In coherent DMA cases, we need to return some memory with contiguous
> > kernel virtual addresses. That is the reason why we need vmalloc()
> > here. If we allocate the memory page by page, the corresponding kernel
> > virtual addresses in a contiguous IOVA range might not be contiguous.
>
>
> Yes, but we can do that as what has been done in the series
> (alloc_pages_exact()). Or do you mean it would be a little bit hard to
> recycle IOVA/pages here?
>

Yes, it might be hard to reuse the memory. For example, we firstly
allocate 1 IOVA/page during dma_map, then the IOVA is freed during
dma_unmap. Actually we can't reuse this single page if we need a
two-pages area in the next IOVA allocating. So the best way is using
IOTLB_UNMAP to free this single page during dma_unmap too.

Thanks,
Yongji