On Wed, 11 Nov 2026 at 23:21, Thomas Zimmermann <[email protected]> wrote: > > Hi > > Am 10.11.26 um 16:49 schrieb Jie Feng: > > On Tue, 10 Nov 2026 at 22:23, Thomas Zimmermann <[email protected]> wrote: > >> Hi, > >> > >> sorry for letting you wait for so long. You'll find a review below. > > Thanks for the review! Since this driver shares many similarities > > with the ms91xx driver I am submitting too, I will be making > > the changes in that driver too. > > I've briefly seen that other driver. Let's focus on one of them first. > The other will then be easier, I guess. Sure, let's focus on the review here, this is a much smaller driver. > > > > >>> + ret = drm_gem_fb_begin_cpu_access(state->fb, DMA_FROM_DEVICE); > >>> + if (ret < 0) { > >>> + complete(¤t_transfer->frame_complete); > >>> + goto exit_save_pending; > >>> + } > >>> + > >>> + drm_fb_xrgb8888_to_rgb888(&data_map, NULL, > >>> + &shadow_plane_state->data[0], > >>> + state->fb, ¤t_rect, > >>> + &shadow_plane_state->fmtcnv_state); > >>> + > >>> + drm_gem_fb_end_cpu_access(state->fb, DMA_FROM_DEVICE); > >>> + > >>> + flush_kernel_vmap_range(current_transfer->buf.data, frame_len); > >>> + > >>> + queue_work(trigger5->transfer_wq, ¤t_transfer->transfer_work); > >>> + trigger5->current_transfer = !trigger5->current_transfer; > >> Why do you need a separate work item for the transfer. Can't you just > >> send from here? We do this in udl. > > I want to make the send async, so drm_fb_xrgb8888_to_rgb888 can run > > in parallel with the usb transfer. This is done since drm_fb_*_to_* > > functions are slow when copying on write-combined memory due to > > dma-buf sharing between the main GPU and this device. > > We do this in other drivers as well. Neither is supposed to be fast. > There's always this memcpy from the imported buffer. > > > > > > udl does usb_submit_urb, which is async, but the usb_sg_wait has no async > > counterpart, so I opted for putting it in a worker. > > Why do you use usb_sg_wait() in the first place? It appears to be a loop This part is referenced from gud. I thought it makes sense since I need a large vmalloc buffer, so using usb_sg_* looked like the right approach instead of asking for kmalloc(8MB).
Alternative I considered are: - usb_alloc_noncoherent: but that requires implementing a fallback if a controller doesn't support scatter gather DMA, while usb_sg_* does handle all of these cases. - pool of fixed kmalloc buffers: 64K seems too small, too many URBs for a 1080p frame, and tuning the number so that kmalloc will likely to succeed on a fragmented system while keeping the number of URBs per 1080p frame small doesn't seem like the right approach to me. Are there other possible approaches to this? > around usb_submit_urb() plus the waiting. (The comment at [1] is > actually quite funny. > > [1] > https://elixir.bootlin.com/linux/v7.3.1/source/drivers/usb/core/message.c#L761 Indeed! > > > > I saw there's a TODO within the format helpers for io memory, so > > maybe in a future patch I could change it to use drm_memcpy_from_wc > > to speed it up. Currently it couldn't run at full 1080p at 60fps on > > bare-metal, > > but on a VM it is fine due to not copying from WC memory. > > > > There's also the other issue of some compositors calling _update > > in the render loop or something that causes visible stutter when a device's > > _update callback is slow. > > > > The ideal way to solve this, with both low latency and low memory footprint > > is to take inspiration from gm12u320 and store the fb and update rect in > > the device struct, and let a worker obtain the latest one from the struct. > > It's not. None of our USB-based drivers are well optimized. You can take > some inspiration from them, but don't take them as they are. > > At the end of each page flip, DRM sends out a FLIP_COMPLETE event, [2] > telling the compositor that it's done with the display update. This > gives the compositor an idea about the performance of the device. It > can adapt to that and end out less updates. By moving out the actual > work into a separate thread, you break that feedback. The compositor > will see an early complete from the page flip and queue up the next > frame immediately. The driver will pile up updates for the worker to > send out. Hence, leave the copying and sending operations in the regular > update code. If you build this with usb_submit_urb(), there's no > waiting necessary. In that case, assuming I keep the usb_sg_wait, doing something like wait_for_completion(previous_transfer) would be more correct. > [2] > https://elixir.bootlin.com/linux/v7.3.1/source/include/uapi/drm/drm.h#L1395 > > What I have in mind for udl is this logic: > > 1. send the current frame's updates as usual using usb_submit_urb() > - transfer runs in parallel to current display update > - count the number of submitted urbs > 2. in each complete handler, decrement the urb count. > 3. in the next frame's atomic_begin, wait until the urb counter reaches zero > - most parallelism > - max out USB > - maintains correct transfer statistics and events > 4. if the second frame takes longer than expected, the FLIP_COMPLETE > event wil be late; the compositor can throttle the page-flip rate For this one, wait_for_completion(previous_transfer) looks to be similar in behaviour. > I made a nicely working prototype from this. > > Best regards > Thomas > > > > The transfer is then sent in blocks, while overlapping the drm copy > > and usb transfer in parallel. I think this is a bit too much complexity > > for a drm/tiny driver though. > > > > > > ...
