On Fri, May 15, 2020 at 02:07:06PM +0900, David Stevens wrote:
On Thu, May 14, 2020 at 9:30 PM Daniel Vetter daniel@ffwll.ch wrote:
On Thu, May 14, 2020 at 05:19:40PM +0900, David Stevens wrote:
Sorry for the duplicate reply, didn't notice this until now.
Just storing the uuid should be doable (assuming this doesn't change during the lifetime of the buffer), so no need for a callback.
Directly storing the uuid doesn't work that well because of synchronization issues. The uuid needs to be shared between multiple virtio devices with independent command streams, so to prevent races between importing and exporting, the exporting driver can't share the uuid with other drivers until it knows that the device has finished registering the uuid. That requires a round trip to and then back from the device. Using a callback allows the latency from that round trip registration to be hidden.
Uh, that means you actually do something and there's locking involved. Makes stuff more complicated, invariant attributes are a lot easier generally. Registering that uuid just always doesn't work, and blocking when you're exporting?
Registering the id at creation and blocking in gem export is feasible, but it doesn't work well for systems with a centralized buffer allocator that doesn't support batch allocations (e.g. gralloc). In such a system, the round trip latency would almost certainly be included in the buffer allocation time. At least on the system I'm working on, I suspect that would add 10s of milliseconds of startup latency to video pipelines (although I haven't benchmarked the difference). Doing the blocking as late as possible means most or all of the latency can be hidden behind other pipeline setup work.
In terms of complexity, I think the synchronization would be basically the same in either approach, just in different locations. All it would do is alleviate the need for a callback to fetch the UUID.
Hm ok. I guess if we go with the older patch, where this all is a lot more just code in virtio, doing an extra function to allocate the uuid sounds fine. Then synchronization is entirely up to the virtio subsystem and not a dma-buf problem (and hence not mine). You can use dma_resv_lock or so, but no need to. But with callbacks potentially going both ways things always get a bit interesting wrt locking - this is what makes peer2peer dma-buf so painful right now. Hence I'd like to avoid that if needed, at least at the dma-buf level. virtio code I don't mind what you do there :-)
Cheers, Daniel
Hello David,
On Fri, 15 May 2020 at 19:33, Daniel Vetter daniel@ffwll.ch wrote:
On Fri, May 15, 2020 at 02:07:06PM +0900, David Stevens wrote:
On Thu, May 14, 2020 at 9:30 PM Daniel Vetter daniel@ffwll.ch wrote:
On Thu, May 14, 2020 at 05:19:40PM +0900, David Stevens wrote:
Sorry for the duplicate reply, didn't notice this until now.
Just storing the uuid should be doable (assuming this doesn't change during the lifetime of the buffer), so no need for a callback.
Directly storing the uuid doesn't work that well because of synchronization issues. The uuid needs to be shared between multiple virtio devices with independent command streams, so to prevent races between importing and exporting, the exporting driver can't share the uuid with other drivers until it knows that the device has finished registering the uuid. That requires a round trip to and then back from the device. Using a callback allows the latency from that round trip registration to be hidden.
Uh, that means you actually do something and there's locking involved. Makes stuff more complicated, invariant attributes are a lot easier generally. Registering that uuid just always doesn't work, and blocking when you're exporting?
Registering the id at creation and blocking in gem export is feasible, but it doesn't work well for systems with a centralized buffer allocator that doesn't support batch allocations (e.g. gralloc). In such a system, the round trip latency would almost certainly be included in the buffer allocation time. At least on the system I'm working on, I suspect that would add 10s of milliseconds of startup latency to video pipelines (although I haven't benchmarked the difference). Doing the blocking as late as possible means most or all of the latency can be hidden behind other pipeline setup work.
In terms of complexity, I think the synchronization would be basically the same in either approach, just in different locations. All it would do is alleviate the need for a callback to fetch the UUID.
I think I agree with Daniel there - this seems best suited for code within virtio.
Hm ok. I guess if we go with the older patch, where this all is a lot more just code in virtio, doing an extra function to allocate the uuid sounds fine. Then synchronization is entirely up to the virtio subsystem and not a dma-buf problem (and hence not mine). You can use dma_resv_lock or so, but no need to. But with callbacks potentially going both ways things always get a bit interesting wrt locking - this is what makes peer2peer dma-buf so painful right now. Hence I'd like to avoid that if needed, at least at the dma-buf level. virtio code I don't mind what you do there :-)
Cheers, Daniel
Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch
Best, Sumit.
linaro-mm-sig@lists.linaro.org
-
Daniel Vetter -
Sumit Semwal