[Intel-gfx] [PATCH 1/2] drm/vgem: Do not allocate backing shmemfs file for an import dmabuf object

Thu Jul 9 08:49:37 UTC 2020

Am 08.07.20 um 18:11 schrieb Daniel Vetter:
> On Wed, Jul 8, 2020 at 5:05 PM Christian König <christian.koenig at amd.com> wrote:
>> Am 08.07.20 um 17:01 schrieb Daniel Vetter:
>>> On Wed, Jul 8, 2020 at 4:37 PM Christian König <christian.koenig at amd.com> wrote:
>>>> Am 08.07.20 um 11:54 schrieb Daniel Vetter:
>>>>> On Wed, Jul 08, 2020 at 11:22:00AM +0200, Christian König wrote:
>>>>>> Am 07.07.20 um 20:35 schrieb Chris Wilson:
>>>>>>> Quoting lepton (2020-07-07 19:17:51)
>>>>>>>> On Tue, Jul 7, 2020 at 10:20 AM Chris Wilson <chris at chris-wilson.co.uk> wrote:
>>>>>>>>> Quoting lepton (2020-07-07 18:05:21)
>>>>>>>>>> On Tue, Jul 7, 2020 at 9:00 AM Chris Wilson <chris at chris-wilson.co.uk> wrote:
>>>>>>>>>>> If we assign obj->filp, we believe that the create vgem bo is native and
>>>>>>>>>>> allow direct operations like mmap() assuming it behaves as backed by a
>>>>>>>>>>> shmemfs inode. When imported from a dmabuf, the obj->pages are
>>>>>>>>>>> not always meaningful and the shmemfs backing store misleading.
>>>>>>>>>>>
>>>>>>>>>>> Note, that regular mmap access to a vgem bo is via the dumb buffer API,
>>>>>>>>>>> and that rejects attempts to mmap an imported dmabuf,
>>>>>>>>>> What do you mean by "regular mmap access" here?  It looks like vgem is
>>>>>>>>>> using vgem_gem_dumb_map as .dumb_map_offset callback then it doesn't call
>>>>>>>>>> drm_gem_dumb_map_offset
>>>>>>>>> As I too found out, and so had to correct my story telling.
>>>>>>>>>
>>>>>>>>> By regular mmap() access I mean mmap on the vgem bo [via the dumb buffer
>>>>>>>>> API] as opposed to mmap() via an exported dma-buf fd. I had to look at
>>>>>>>>> igt to see how it was being used.
>>>>>>>> Now it seems your fix is to disable "regular mmap" on imported dma buf
>>>>>>>> for vgem. I am not really a graphic guy, but then the api looks like:
>>>>>>>> for a gem handle, user space has to guess to find out the way to mmap
>>>>>>>> it. If user space guess wrong, then it will fail to mmap. Is this the
>>>>>>>> expected way
>>>>>>>> for people to handle gpu buffer?
>>>>>>> You either have a dumb buffer handle, or a dma-buf fd. If you have the
>>>>>>> handle, you have to use the dumb buffer API, there's no other way to
>>>>>>> mmap it. If you have the dma-buf fd, you should mmap it directly. Those
>>>>>>> two are clear.
>>>>>>>
>>>>>>> It's when you import the dma-buf into vgem and create a handle out of
>>>>>>> it, that's when the handle is no longer first class and certain uAPI
>>>>>>> [the dumb buffer API in particular] fail.
>>>>>>>
>>>>>>> It's not brilliant, as you say, it requires the user to remember the
>>>>>>> difference between the handles, but at the same time it does prevent
>>>>>>> them falling into coherency traps by forcing them to use the right
>>>>>>> driver to handle the object, and have to consider the additional ioctls
>>>>>>> that go along with that access.
>>>>>> Yes, Chris is right. Mapping DMA-buf through the mmap() APIs of an importer
>>>>>> is illegal.
>>>>>>
>>>>>> What we could maybe try to do is to redirect this mmap() API call on the
>>>>>> importer to the exporter, but I'm pretty sure that the fs layer wouldn't
>>>>>> like that without changes.
>>>>> We already do that, there's a full helper-ified path from I think shmem
>>>>> helpers through prime helpers to forward this all. Including handling
>>>>> buffer offsets and all the other lolz back&forth.
>>>> Oh, that most likely won't work correctly with unpinned DMA-bufs and
>>>> needs to be avoided.
>>>>
>>>> Each file descriptor is associated with an struct address_space. And
>>>> when you mmap() through the importer by redirecting the system call to
>>>> the exporter you end up with the wrong struct address_space in your VMA.
>>>>
>>>> That in turn can go up easily in flames when the exporter tries to
>>>> invalidate the CPU mappings for its DMA-buf while moving it.
>>>>
>>>> Where are we doing this? My last status was that this is forbidden.
>>> Hm I thought we're doing all that already, but looking at the code
>>> again we're only doing this when opening a new drm fd or dma-buf fd.
>>> So the right file->f_mapping is always set at file creation time.
>>>
>>> And we indeed don't frob this more when going another indirection ...
>>> Maybe we screwed up something somewhere :-/
>>>
>>> Also I thought the mapping is only taken after the vma is instatiated,
>>> otherwise the tricks we're playing with dma-buf already wouldn't work:
>>> dma-buf has the buffer always at offset 0, whereas gem drm_fd mmap has
>>> it somewhere else. We already adjust vma->vm_pgoff, so I'm wondering
>>> whether we could adjust vm_file too. Or is that the thing that's
>>> forbidden?
>> Yes, exactly. Modifying vm_pgoff is harmless, tons of code does that.
>>
>> But changing vma->vm_file, that's something I haven't seen before and
>> most likely could blow up badly.
> Ok, I read the shmem helpers again, I think those are the only ones
> which do the importer mmap -> dma_buf_mmap() forwarding, and hence
> break stuff all around here.
>
> They also remove the vma->vm_pgoff offset, which means
> unmap_mapping_range wont work anyway. I guess for drivers which use
> shmem helpers the hard assumption is that a) can't use p2p dma and b)
> pin everything into system memory.
>
> So not a problem. But something to keep in mind. I'll try to do a
> kerneldoc patch to note this somewhere. btw on that, did the
> timeline/syncobj documentation patch land by now? Just trying to make
> sure that doesn't get lost for another few months or so :-/

I still haven't found time to double check the documentation and most 
likely won't in quite a while.

Sorry for that, but yeah you know :)

Christian.

>
> Cheers, Daniel
>
>> Christian.
>>
>>> -Daniel
>>>
>>>> Christian.
>>>>
>>>>> Of course there's still the problem that many drivers don't forward the
>>>>> cache coherency calls for begin/end cpu access, so in a bunch of cases
>>>>> you'll cache cacheline dirt soup. But that's kinda standard procedure for
>>>>> dma-buf :-P
>>>>>
>>>>> But yeah trying to handle the mmap as an importer, bypassing the export:
>>>>> nope. The one exception is if you have some kind of fancy gart with
>>>>> cpu-visible pci bar (like at least integrated intel gpus have). But in
>>>>> that case the mmap very much looks&acts like device access in every way.
>>>>>
>>>>> Cheers, Daniel
>>>>>
>>>>>> Regards,
>>>>>> Christian.
>>>>>>
>>>>>>
>>>>>>> -Chris
>