Re: Regression due to commit 1161db6776bd1c35257e1e362e91bcb1b3fe4347 (media: usb: pwc: Don't use coherent DMA buffers for ISO transfer)
Sorry for the empty message. I would like to add Joerg Roedel and IOMMU related maillist in case they can help on this issue. пн, 23 нояб. 2020 г. в 16:32, Matwey V. Kornilov : > > пн, 23 нояб. 2020 г. в 14:19, Thimo Emmerich : > > > > Hi, > > > > thanks for your prompt answer. Unfortunately I was out of office on Friday. > > > > Here the updates: > > - Tracing showed the lines you mentioned. > > - Updating the kernel to the latest stable (v5.9.10) did not change > > anything. > > - Adding "amd_iommu=off" fixes the issue (of course including > > dysfunctional IOMMU) > > > > Best regards, > > Thimo > > > > On 20.11.20 08:55, Matwey V. Kornilov wrote: > > > Hi, > > > > > > As far as I understand, the driver itself works correctly. It > > > communicates with the device and receives the data. Only difference is > > > that the data is zeroed. If so, it means that the memory is allocated > > > without errors and the DMA is also mapped without errors. Otherwise, > > > you would get an error when trying to open the device/start stream. To > > > finally prove this, could you please use tracing points to see what is > > > going on while the video stream is on? > > > > > > mount -ttracefs tracefs /sys/kernel/tracing > > > cd /sys/kernel/tracing > > > echo 1 > events/pwc/enable > > > echo 1 > tracing_on > > > cat trace > > > > > > And then you should see lots of lines like the following when video > > > stream is running: > > > > > > -0 [002] ..s. 34645.185239: pwc_handler_enter: dev=Philips webcam > > > 1-1.4:1.0 (fbuf=303a809e filled=31581) urb=7e3783c4 > > > (status=0 actual_length=9570) > > > -0 [002] ..s. 34645.185244: pwc_handler_exit: dev=Philips webcam > > > 1-1.4:1.0 (fbuf=303a809e filled=41151) urb=7e3783c4 > > > > > > Use > > > > > > echo 0 > tracing_on > > > > > > to disable the recording. > > > > > > > > > I've found a commit that claims to fix an issue with similar wording: > > > https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_torvalds_linux_commit_27d2dcb1b95c8c39da417839c11eda2e665cd389=DwIFaQ=fP4tf--1dS0biCFlB0saz0I0kjO5v7-GLPtvShAo4cc=uorjfjHlvtqEyqWp5b8YlexJq-ujpFniwt87GU9IrXI=ZPgwadRZcYtXJwzoIM6ibrGC8jNcJkKloZFNKDK6PBE=pC6Xy7iT5Yf6nx7dSNiDOCvvCs7HENylTiquLeJ_0Og= > > > > > > Could you please try to boot with "amd_iommu=off" kernel command line > > > parameter and see what happens? > > > Is it also possible for you to check the latest kernel version on your > > > setup? > > > > > > > > > чт, 19 нояб. 2020 г. в 18:49, Thimo Emmerich : > > > > > > > > >> Hi, > > >> > > >> not sure about the application behaviour. > > >> I observed this green screen in MS Teams. For easy verification I used > > >> "cheese" which shows a live stream. Same behaviour. > > >> The issue is 100% reproducible: loading the stock Ubuntu (HWE) kernel > > >> module --> green screen, unloading and loading patched module --> > > >> picture. > > >> For this procedure it is even not necessary to disconnect the cam itself. > > >> > > >> I did now also another test excluding possible application magic: > > >> dd if=/dev/video0 of=/tmp/vid.raw bs=460800 count=20 > > >> If I interpret this data as a video (640x720), with the patched module, > > >> I see a running sequence whereas with the stock module there are just > > >> zeros. Of course, the zeros interpreted as YCbCr stream lead to a still > > >> green image. > > >> > > >> I am using the stock Ubuntu HWE kernel (linux-image-5.4.0-53-generic), > > >> config is attached. > > >> I have to admit that I did not use the Ubuntu-patched kernel sources but > > >> the v5.4 tagged sources from git.kernel.org since I guess there are no > > >> patches in this special section. I also verified building the vanilla > > >> module has the same behaviour as the Ubuntu kernel module (green screen). > > >> > > >> I am not using the virtualization feature (currently). Nevertheless, the > > >> IOMMU is enabled (by default) in the BIOS. > > >> Unfortunately, I do not have another webcam at hand (the other ones are > > >> notebook build-ins). > > >> I checked now with another PC (ancient Intel Xeon system) - the stock > > >> driver works without trouble. > > >> > > >> Do you use/have access to a Ryzen CPU? Maybe there is some > > >> issue/specialty with its IOMMU? > > >> > > >> Best regards, > > >>Thimo Emmerich > > >> > > >> On 19.11.20 14:26, Matwey V. Kornilov wrote: > > >>> Hi, > > >>> > > >>> I've just checked my PWC900 camera at kernel 5.9.8 and unfortunately > > >>> It works perfectly, so there issue is probably deeper. > > >>> > > >>> You say that you see the green image. Could you ensure that the green > > >>> image is what you actually get from the kernel driver, but not an > > >>> application behaviour when there are no frames. > > >>> Could you also provide your kernel build config? I see that you are > > >>> running some kind of hardware assisted virtualization, right? May I > > >>> also ask you to look for some other USB webcam
Re: Regression due to commit 1161db6776bd1c35257e1e362e91bcb1b3fe4347 (media: usb: pwc: Don't use coherent DMA buffers for ISO transfer)
пн, 23 нояб. 2020 г. в 14:19, Thimo Emmerich : > > Hi, > > thanks for your prompt answer. Unfortunately I was out of office on Friday. > > Here the updates: > - Tracing showed the lines you mentioned. > - Updating the kernel to the latest stable (v5.9.10) did not change > anything. > - Adding "amd_iommu=off" fixes the issue (of course including > dysfunctional IOMMU) > > Best regards, > Thimo > > On 20.11.20 08:55, Matwey V. Kornilov wrote: > > Hi, > > > > As far as I understand, the driver itself works correctly. It > > communicates with the device and receives the data. Only difference is > > that the data is zeroed. If so, it means that the memory is allocated > > without errors and the DMA is also mapped without errors. Otherwise, > > you would get an error when trying to open the device/start stream. To > > finally prove this, could you please use tracing points to see what is > > going on while the video stream is on? > > > > mount -ttracefs tracefs /sys/kernel/tracing > > cd /sys/kernel/tracing > > echo 1 > events/pwc/enable > > echo 1 > tracing_on > > cat trace > > > > And then you should see lots of lines like the following when video > > stream is running: > > > > -0 [002] ..s. 34645.185239: pwc_handler_enter: dev=Philips webcam > > 1-1.4:1.0 (fbuf=303a809e filled=31581) urb=7e3783c4 > > (status=0 actual_length=9570) > > -0 [002] ..s. 34645.185244: pwc_handler_exit: dev=Philips webcam > > 1-1.4:1.0 (fbuf=303a809e filled=41151) urb=7e3783c4 > > > > Use > > > > echo 0 > tracing_on > > > > to disable the recording. > > > > > > I've found a commit that claims to fix an issue with similar wording: > > https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_torvalds_linux_commit_27d2dcb1b95c8c39da417839c11eda2e665cd389=DwIFaQ=fP4tf--1dS0biCFlB0saz0I0kjO5v7-GLPtvShAo4cc=uorjfjHlvtqEyqWp5b8YlexJq-ujpFniwt87GU9IrXI=ZPgwadRZcYtXJwzoIM6ibrGC8jNcJkKloZFNKDK6PBE=pC6Xy7iT5Yf6nx7dSNiDOCvvCs7HENylTiquLeJ_0Og= > > > > Could you please try to boot with "amd_iommu=off" kernel command line > > parameter and see what happens? > > Is it also possible for you to check the latest kernel version on your > > setup? > > > > > > чт, 19 нояб. 2020 г. в 18:49, Thimo Emmerich : > > > > > >> Hi, > >> > >> not sure about the application behaviour. > >> I observed this green screen in MS Teams. For easy verification I used > >> "cheese" which shows a live stream. Same behaviour. > >> The issue is 100% reproducible: loading the stock Ubuntu (HWE) kernel > >> module --> green screen, unloading and loading patched module --> picture. > >> For this procedure it is even not necessary to disconnect the cam itself. > >> > >> I did now also another test excluding possible application magic: > >> dd if=/dev/video0 of=/tmp/vid.raw bs=460800 count=20 > >> If I interpret this data as a video (640x720), with the patched module, > >> I see a running sequence whereas with the stock module there are just > >> zeros. Of course, the zeros interpreted as YCbCr stream lead to a still > >> green image. > >> > >> I am using the stock Ubuntu HWE kernel (linux-image-5.4.0-53-generic), > >> config is attached. > >> I have to admit that I did not use the Ubuntu-patched kernel sources but > >> the v5.4 tagged sources from git.kernel.org since I guess there are no > >> patches in this special section. I also verified building the vanilla > >> module has the same behaviour as the Ubuntu kernel module (green screen). > >> > >> I am not using the virtualization feature (currently). Nevertheless, the > >> IOMMU is enabled (by default) in the BIOS. > >> Unfortunately, I do not have another webcam at hand (the other ones are > >> notebook build-ins). > >> I checked now with another PC (ancient Intel Xeon system) - the stock > >> driver works without trouble. > >> > >> Do you use/have access to a Ryzen CPU? Maybe there is some > >> issue/specialty with its IOMMU? > >> > >> Best regards, > >>Thimo Emmerich > >> > >> On 19.11.20 14:26, Matwey V. Kornilov wrote: > >>> Hi, > >>> > >>> I've just checked my PWC900 camera at kernel 5.9.8 and unfortunately > >>> It works perfectly, so there issue is probably deeper. > >>> > >>> You say that you see the green image. Could you ensure that the green > >>> image is what you actually get from the kernel driver, but not an > >>> application behaviour when there are no frames. > >>> Could you also provide your kernel build config? I see that you are > >>> running some kind of hardware assisted virtualization, right? May I > >>> also ask you to look for some other USB webcam (other model) around > >>> you and check if it works at the same configuration? > >>> > >>> > >>> чт, 19 нояб. 2020 г. в 14:59, Thimo Emmerich: > >>> > Hi Matwey, > > I am using a Philips SPC 900NC USB webcam on my Ubuntu 18.04 system. > Recently I upgraded from the stock kernel (4.15) to the latest HWE > kernel (5.4). > > Unfortunately my camera was not working anymore.
