Re: [RFC PATCH 1/3] RDMA/umem: Change for rdma devices has not dma device

From: Daniel Vetter
Date: Mon Sep 13 2021 - 15:23:07 EST


On Fri, Sep 10, 2021 at 3:46 AM Shunsuke Mie <mie@xxxxxxxxxx> wrote:
>
> 2021年9月9日(木) 18:26 Daniel Vetter <daniel.vetter@xxxxxxxx>:
> >
> > On Thu, Sep 9, 2021 at 1:33 AM Jason Gunthorpe <jgg@xxxxxxxx> wrote:
> > > On Wed, Sep 08, 2021 at 09:22:37PM +0200, Daniel Vetter wrote:
> > > > On Wed, Sep 8, 2021 at 3:33 PM Christian König <christian.koenig@xxxxxxx> wrote:
> > > > > Am 08.09.21 um 13:18 schrieb Jason Gunthorpe:
> > > > > > On Wed, Sep 08, 2021 at 05:41:39PM +0900, Shunsuke Mie wrote:
> > > > > >> 2021年9月8日(水) 16:20 Christoph Hellwig <hch@xxxxxxxxxxxxx>:
> > > > > >>> On Wed, Sep 08, 2021 at 04:01:14PM +0900, Shunsuke Mie wrote:
> > > > > >>>> Thank you for your comment.
> > > > > >>>>> On Wed, Sep 08, 2021 at 03:16:09PM +0900, Shunsuke Mie wrote:
> > > > > >>>>>> To share memory space using dma-buf, a API of the dma-buf requires dma
> > > > > >>>>>> device, but devices such as rxe do not have a dma device. For those case,
> > > > > >>>>>> change to specify a device of struct ib instead of the dma device.
> > > > > >>>>> So if dma-buf doesn't actually need a device to dma map why do we ever
> > > > > >>>>> pass the dma_device here? Something does not add up.
> > > > > >>>> As described in the dma-buf api guide [1], the dma_device is used by dma-buf
> > > > > >>>> exporter to know the device buffer constraints of importer.
> > > > > >>>> [1] https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flwn.net%2FArticles%2F489703%2F&amp;data=04%7C01%7Cchristian.koenig%40amd.com%7C4d18470a94df4ed24c8108d972ba5591%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637666967356417448%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C2000&amp;sdata=ARwQyo%2BCjMohaNbyREofToHIj2bndL5L0HaU9cOrYq4%3D&amp;reserved=0
> > > > > >>> Which means for rxe you'd also have to pass the one for the underlying
> > > > > >>> net device.
> > > > > >> I thought of that way too. In that case, the memory region is constrained by the
> > > > > >> net device, but rxe driver copies data using CPU. To avoid the constraints, I
> > > > > >> decided to use the ib device.
> > > > > > Well, that is the whole problem.
> > > > > >
> > > > > > We can't mix the dmabuf stuff people are doing that doesn't fill in
> > > > > > the CPU pages in the SGL with RXE - it is simply impossible as things
> > > > > > currently are for RXE to acess this non-struct page memory.
> > > > >
> > > > > Yeah, agree that doesn't make much sense.
> > > > >
> > > > > When you want to access the data with the CPU then why do you want to
> > > > > use DMA-buf in the first place?
> > > > >
> > > > > Please keep in mind that there is work ongoing to replace the sg table
> > > > > with an DMA address array and so make the underlying struct page
> > > > > inaccessible for importers.
> > > >
> > > > Also if you do have a dma-buf, you can just dma_buf_vmap() the buffer
> > > > for cpu access. Which intentionally does not require any device. No
> > > > idea why there's a dma_buf_attach involved. Now not all exporters
> > > > support this, but that's fixable, and you must call
> > > > dma_buf_begin/end_cpu_access for cache management if the allocation
> > > > isn't cpu coherent. But it's all there, no need to apply hacks of
> > > > allowing a wrong device or other fun things.
> > >
> > > Can rxe leave the vmap in place potentially forever?
> >
> > Yeah, it's like perma-pinning the buffer into system memory for
> > non-p2p dma-buf sharing. We just squint and pretend that can't be
> > abused too badly :-) On 32bit you'll run out of vmap space rather
> > quickly, but that's not something anyone cares about here either. We
> > have a bunch of more sw modesetting drivers in drm which use
> > dma_buf_vmap() like this, so it's all fine.
> > -Daniel
> > --
> > Daniel Vetter
> > Software Engineer, Intel Corporation
> > http://blog.ffwll.ch
>
> Thanks for your comments.
>
> In the first place, the CMA region cannot be used for RDMA because the
> region has no struct page. In addition, some GPU drivers use CMA and share
> the region as dma-buf. As a result, RDMA cannot transfer for the region. To
> solve this problem, rxe dma-buf support is better I thought.
>
> I'll consider and redesign the rxe dma-buf support using the dma_buf_vmap()
> instead of the dma_buf_dynamic_attach().

btw for next version please cc dri-devel. get_maintainers.pl should
pick it up for these patches.
-Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch