On Thu, Jul 15, 2021 at 1:21 PM Kenneth Graunke <kenneth@xxxxxxxxxxxxx> wrote:
>
> On Thursday, July 15, 2021 4:27:44 AM PDT Tvrtko Ursulin wrote:
> >
> > On 15/07/2021 12:07, Daniel Vetter wrote:
> > > On Thu, Jul 15, 2021 at 11:33:10AM +0100, Tvrtko Ursulin wrote:
> > >>
> > >> On 15/07/2021 11:15, Matthew Auld wrote:
> > >>> From: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
> > >>>
> > >>> Jason Ekstrand requested a more efficient method than userptr+set-domain
> > >>> to determine if the userptr object was backed by a complete set of pages
> > >>> upon creation. To be more efficient than simply populating the userptr
> > >>> using get_user_pages() (as done by the call to set-domain or execbuf),
> > >>> we can walk the tree of vm_area_struct and check for gaps or vma not
> > >>> backed by struct page (VM_PFNMAP). The question is how to handle
> > >>> VM_MIXEDMAP which may be either struct page or pfn backed...
> > >>>
> > >>> With discrete are going to drop support for set_domain(), so offering a
> > >>> way to probe the pages, without having to resort to dummy batches has
> > >>> been requested.
> > >>>
> > >>> v2:
> > >>> - add new query param for the PROPBE flag, so userspace can easily
> > >>> check if the kernel supports it(Jason).
> > >>> - use mmap_read_{lock, unlock}.
> > >>> - add some kernel-doc.
> > >>
> > >> 1)
> > >>
> > >> I think probing is too weak to be offered as part of the uapi. What probes
> > >> successfully at create time might not be there anymore at usage time. So if
> > >> the pointer is not trusted at one point, why should it be at a later stage?
> > >>
> > >> Only thing which works for me is populate (so get_pages) at create time. But
> > >> again with no guarantees they are still there at use time clearly
> > >> documented.
> > >
> > > Populate is exactly as racy as probe. We don't support pinned userptr
> > > anymore.
> >
> > Yes, wrote so myself - "..again with no guarantees they are still there
> > at use time..".
> >
> > Perhaps I don't understand what problem is probe supposed to solve. It
> > doesn't deal 1:1 with set_domain removal since that one actually did
> > get_pages so that would be populate. But fact remains regardless that if
> > userspace is given a pointer it doesn't trust, _and_ wants the check it
> > for this reason or that, then probe solves nothing. Unless there is
> > actually at minimum some protocol to reply to whoever sent the pointer
> > like "not that pointer please".
>
> That's exactly the point. GL_AMD_pinned_memory requires us the OpenGL
> implementation to return an error for "not that pointer, please", at the
> time when said pointer is supplied - not at first use.
>
> Sure, there can be reasons why it might seem fine up front, and not work
> later. But an early check of "just no, you're doing it totally wrong"
> at the right moment can be helpful for application developers. While it
> shouldn't really happen, if it ever did, it would be a lot more obvious
> to debug than "much later on, when something randomly flushed the GPU
> commands we were building, something went wrong, and we don't know why."
And before someone chimes in saying that Vulkan doesn't need this
because we can just VK_ERROR_DEVICE_LOST later, that's not true. We'd
like to be able to return VK_ERROR_INVALID_EXTERNAL_HANDLE in the
easily checkable cases like if they try to pass in a mmapped file.
--Jason
