On 10 April 2017 6:51:21 pm AEST, "Michel Dänzer" <michel@xxxxxxxxxxx> wrote: >On 06/04/17 04:47 PM, Christopher James Halse Rogers wrote: >> On Wed, 5 Apr 2017 at 20:14 Lucas Stach <l.stach@xxxxxxxxxxxxxx >> <mailto:l.stach@xxxxxxxxxxxxxx>> wrote: >> >> Am Mittwoch, den 05.04.2017, 11:59 +0200 schrieb Daniel Vetter: >> > On Wed, Apr 05, 2017 at 10:15:44AM +0200, Lucas Stach wrote: >> > > Am Mittwoch, den 05.04.2017, 00:20 +0000 schrieb Christopher >> James Halse >> > > Rogers: >> > > > >> > > > >> > > > On Tue, Apr 4, 2017 at 9:53 PM Daniel Vetter ><daniel@xxxxxxxx >> <mailto:daniel@xxxxxxxx>> wrote: >> > > > >> > > > On Tue, Apr 4, 2017 at 12:43 PM, Lucas Stach >> > > > <l.stach@xxxxxxxxxxxxxx >> <mailto:l.stach@xxxxxxxxxxxxxx>> wrote: >> > > > >> If I could guarantee that I'd only ever run on >> > > > 4.13-or-later kernels >> > > > >> (I think that's when the previous patches will >> land?), then >> > > > this would >> > > > >> indeed be mostly unnecessary. It would save me a >> bunch of >> > > > addfb calls >> > > > >> that would predictably fail, but they're cheap. >> > > > > >> > > > > I don't think we ever had caps for "things are >> working now, >> > > > as they are >> > > > > supposed to". i915 wasn't properly synchronizing >on >> foreign >> > > > fences for a >> > > > > long time, yet we didn't gain a cap for "cross >> device sync >> > > > works now". >> > > > > >> > > > > If your distro use-case relies on those things >> working it's >> > > > probably >> > > > > best to just backport the relevant fixes. >> > > > >> > > > The even better solution for this is to push the >backports >> > > > through >> > > > upstream -stable kernels. This stuff here is simple >enough >> > > > that we can >> > > > do it. Same could have been done for the fairly >minimal >> > > > fencing fixes >> > > > for i915 (at least some of them, the ones in the >> page-flip). >> > > > >> > > > Otherwise we'll end up with tons IM_NOT_BUGGY and >> > > > IM_SLIGHTLY_LESS_BUGGY and >> > > > IM_NOT_BUGGY_EXCEPT_THIS_BOTCHED_BACKPORT >> > > > flags where no one at all knows what they mean, >usage >> between >> > > > different drivers and different userspace is >entirely >> > > > inconsistent and >> > > > they just all add to the confusion. They're just >bugs, >> lets >> > > > treat them >> > > > like that. >> > > > >> > > > >> > > > It's not *quite* DRM_CAP_PRIME_SCANOUT_NOT_BUGGY - while >the >> relevant >> > > > hardware allegedly supports it, nouveau/radeon/amdgpu don't >do >> scanout >> > > > of GTT, so the lack of this cap indicates that there's no >point in >> > > > trying to call addfb2. >> > > > >> > > > >> > > > But calling addfb2 and it failing is not expensive, so this >is >> rather >> > > > niche. >> > > > >> > > > >> > > > In practice I can just restrict attempting to scanout of >imported >> > > > buffers to i915, as that's the only driver that it'll work >on. >> By the >> > > > time nouveau/radeon/amdgpu get patches to scanout of GTT >the fixes >> > > > should be old enough that I don't need to care about >unfixed >> kernels. >> > > > >> > > So given that most discreet hardware won't ever be able to >> scanout from >> > > GTT (latency and iso requirements will be hard to meet), >can't >> we just >> > > fix the case of the broken prime sharing when migrating to >VRAM? >> > > >> > > I'm thinking about attaching an exclusive fence to the >dma-buf >> when the >> > > migration to VRAM happens, then when the GPU is done with the >> buffer we >> > > can either write back any changes to GTT, or just drop the >fence >> in case >> > > the GPU didn't modify the buffer. >> > >> > We could, but someone needs to type the code for it. There's >also the >> > problem that you need to migrate back, and then doing all that >behind >> > userspaces back might not be the best idea. >> >> Drivers with separate VRAM and GTT are already doing a lot of >migration >> behind the userspaces back. The only issue with dma-buf migration >to >> VRAM is that you probably don't want to migrate the pages, but >duplicate >> them in VRAM, doubling memory consumption with possible OOM. But >then >> you could alloc the memory on addfb where you are able to return >proper >> errors. >> >> >> Hm. So, on a first inspection, this looks like something I could >> actually cook up. >> >> Looking at amdgpu it seems like the thing to do would be to associate >a >> shadow-bo in VRAM for the imported dma-buf in the addfb call, then >> pin-and-copy-to the shadow bo in the places where the bo is currently >> pinned. >> >> Is this approach likely to be acceptable? > >It would break e.g. with DRI2 flips, because they replace the screen >pixmap buffer with the buffer we're flipping to. If the app stops >flipping while such a shadow BO is being scanned out, later draws to >the >screen pixmap won't become visible. This shadow BO would only ever be used for imported dma-bufs. This would change the behaviour from “addfb fails” to “you get a shadow BO”. (And, pre-patch, from “addfb succeeds but you never see any new rendering”). I don't think any DRI2 implementation hits this, because of it did it would already be broken. (On paternity leave, so I'll be intermittent in following this up) _______________________________________________ dri-devel mailing list dri-devel@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/dri-devel
