Message ID | 20231109233341.2461129-1-m.grzeschik@pengutronix.de (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | usb: gadget: uvc: reduce the request size to increase the throughput | expand |
Hi, On Fri, Nov 10, 2023, Michael Grzeschik wrote: > One misconception of queueing request to the usb gadget controller is, > that the amount of data that one usb_request is representing is the same > as the hardware is able to transfer in one interval. > > This exact idea applies to the uvc_video gadget that assumes that the > request needs to have the size as the gadgets bandwidth parameters are > configured with. (maxpacket, multiplier and burst) > > Although it is actually no problem for the hardware to queue a big > request, in the case of isochronous transfers it is impractical to queue > big amount of data per request to the hardware. Especially if the > necessary bandwidth was increased on purpose to maintain high amounts of > data. > > E.g. in the dwc3-controller it has the negative impact that the endpoint > FIFO will not be filled fast enough by the internal hardware engine. > Since each transfer buffer (TRB) represents one big request, the > hardware will need a long time to prefill the hardware FIFO. This can be > avoided by queueing more smaller requests which will then lead to > smaller TRBs which the hardware engine can keep up with. Just want to clarify here to avoid any confusion, the hardware TX FIFO size is relatively small, usually can be smaller than the TRB. It should be fine whether the TRB is larger or smaller than the FIFO size. The hardware does not "prefill" the FIFO. It just fills whichever TRB it's currently processing (I think you may be mixing up with TRB cache). The performance impact from this change is to reduce the USB bandwidth usage. The current calculation in uvc function can use 48KB/uframe for each request in SuperSpeed, the max size for isoc in SuperSpeed. I know many hosts can't handle this kind of transfer rate from their hardware. (e.g. It gets worse when scheduling transfers for multiple endpoints and under multiple tier hubs). The bandwidth can be impacted by multiple factors and not just from the gadget side as noted in the discussion before. > > This patch is simply dropping the maxburst as an multiplier for the size > calculation and therefor decreases the overall maximum request size. > > Signed-off-by: Michael Grzeschik <m.grzeschik@pengutronix.de> > --- > This patch is created as an result from the discussion in the following thread: > > https://urldefense.com/v3/__https://lore.kernel.org/linux-usb/20231031231841.dbhtrgqounu2rvgn@synopsys.com/__;!!A4F2R9G_pg!fTaIo4tDljSbEvUY5SZLkNvKWcz0YeN0Ekzs0CPWyD73RGRmErRC2frODFgnMB1M4Nse0oKKgwxC65gePhGAtauKJq1Vnzlj$ > > drivers/usb/gadget/function/uvc_queue.c | 1 - > drivers/usb/gadget/function/uvc_video.c | 1 - > 2 files changed, 2 deletions(-) > > diff --git a/drivers/usb/gadget/function/uvc_queue.c b/drivers/usb/gadget/function/uvc_queue.c > index 0aa3d7e1f3cc32..1d3c3c09ff97cb 100644 > --- a/drivers/usb/gadget/function/uvc_queue.c > +++ b/drivers/usb/gadget/function/uvc_queue.c > @@ -55,7 +55,6 @@ static int uvc_queue_setup(struct vb2_queue *vq, > sizes[0] = video->imagesize; > > req_size = video->ep->maxpacket > - * max_t(unsigned int, video->ep->maxburst, 1) I think you're reducing a bit too much here? Also, take advantage of burst. So, perhaps keep request size to max(16K, MPS * maxburst)? Can be more or less depending on how much video data is needed to transfer over a video frame. BR, Thinh > * (video->ep->mult); > > /* We divide by two, to increase the chance to run > diff --git a/drivers/usb/gadget/function/uvc_video.c b/drivers/usb/gadget/function/uvc_video.c > index 91af3b1ef0d412..c6b61843bad3d7 100644 > --- a/drivers/usb/gadget/function/uvc_video.c > +++ b/drivers/usb/gadget/function/uvc_video.c > @@ -329,7 +329,6 @@ uvc_video_alloc_requests(struct uvc_video *video) > BUG_ON(video->req_size); > > req_size = video->ep->maxpacket > - * max_t(unsigned int, video->ep->maxburst, 1) > * (video->ep->mult); > > video->ureq = kcalloc(video->uvc_num_requests, sizeof(struct uvc_request), GFP_KERNEL); > -- > 2.39.2 > >
On Fri, Nov 10, 2023, Thinh Nguyen wrote: > Hi, > > On Fri, Nov 10, 2023, Michael Grzeschik wrote: > > One misconception of queueing request to the usb gadget controller is, > > that the amount of data that one usb_request is representing is the same > > as the hardware is able to transfer in one interval. > > > > This exact idea applies to the uvc_video gadget that assumes that the > > request needs to have the size as the gadgets bandwidth parameters are > > configured with. (maxpacket, multiplier and burst) > > > > Although it is actually no problem for the hardware to queue a big > > request, in the case of isochronous transfers it is impractical to queue > > big amount of data per request to the hardware. Especially if the > > necessary bandwidth was increased on purpose to maintain high amounts of > > data. > > > > E.g. in the dwc3-controller it has the negative impact that the endpoint > > FIFO will not be filled fast enough by the internal hardware engine. > > Since each transfer buffer (TRB) represents one big request, the > > hardware will need a long time to prefill the hardware FIFO. This can be > > avoided by queueing more smaller requests which will then lead to > > smaller TRBs which the hardware engine can keep up with. > > Just want to clarify here to avoid any confusion, the hardware TX FIFO > size is relatively small, usually can be smaller than the TRB. It should > be fine whether the TRB is larger or smaller than the FIFO size. The > hardware does not "prefill" the FIFO. It just fills whichever TRB it's > currently processing (I think you may be mixing up with TRB cache). > > The performance impact from this change is to reduce the USB bandwidth > usage. The current calculation in uvc function can use 48KB/uframe for > each request in SuperSpeed, the max size for isoc in SuperSpeed. I know > many hosts can't handle this kind of transfer rate from their hardware. > (e.g. It gets worse when scheduling transfers for multiple endpoints and > under multiple tier hubs). > > The bandwidth can be impacted by multiple factors and not just from the > gadget side as noted in the discussion before. > > > > > This patch is simply dropping the maxburst as an multiplier for the size > > calculation and therefor decreases the overall maximum request size. > > > > Signed-off-by: Michael Grzeschik <m.grzeschik@pengutronix.de> > > --- > > This patch is created as an result from the discussion in the following thread: > > > > https://urldefense.com/v3/__https://lore.kernel.org/linux-usb/20231031231841.dbhtrgqounu2rvgn@synopsys.com/__;!!A4F2R9G_pg!fTaIo4tDljSbEvUY5SZLkNvKWcz0YeN0Ekzs0CPWyD73RGRmErRC2frODFgnMB1M4Nse0oKKgwxC65gePhGAtauKJq1Vnzlj$ > > > > drivers/usb/gadget/function/uvc_queue.c | 1 - > > drivers/usb/gadget/function/uvc_video.c | 1 - > > 2 files changed, 2 deletions(-) > > > > diff --git a/drivers/usb/gadget/function/uvc_queue.c b/drivers/usb/gadget/function/uvc_queue.c > > index 0aa3d7e1f3cc32..1d3c3c09ff97cb 100644 > > --- a/drivers/usb/gadget/function/uvc_queue.c > > +++ b/drivers/usb/gadget/function/uvc_queue.c > > @@ -55,7 +55,6 @@ static int uvc_queue_setup(struct vb2_queue *vq, > > sizes[0] = video->imagesize; > > > > req_size = video->ep->maxpacket > > - * max_t(unsigned int, video->ep->maxburst, 1) > > I think you're reducing a bit too much here? Also, take advantage of > burst. So, perhaps keep request size to max(16K, MPS * maxburst)? I mean to potentially cap at 16K, it didn't translate correctly to code.... I meant something like this: max_size = mult * maxburst * MPS; clamp(max_size, mult * MPS, 16K); > > Can be more or less depending on how much video data is needed to > transfer over a video frame. Can the uvc gadget figure out the max video data over a video frame? Or is it not available at this layer? I figure different setup requires more video data payload. If we have this info, then you can properly cap the request size. Thanks, Thinh
On Fri, Nov 10, 2023 at 02:16:59AM +0000, Thinh Nguyen wrote: >On Fri, Nov 10, 2023, Michael Grzeschik wrote: >> One misconception of queueing request to the usb gadget controller is, >> that the amount of data that one usb_request is representing is the same >> as the hardware is able to transfer in one interval. >> >> This exact idea applies to the uvc_video gadget that assumes that the >> request needs to have the size as the gadgets bandwidth parameters are >> configured with. (maxpacket, multiplier and burst) >> >> Although it is actually no problem for the hardware to queue a big >> request, in the case of isochronous transfers it is impractical to queue >> big amount of data per request to the hardware. Especially if the >> necessary bandwidth was increased on purpose to maintain high amounts of >> data. >> >> E.g. in the dwc3-controller it has the negative impact that the endpoint >> FIFO will not be filled fast enough by the internal hardware engine. >> Since each transfer buffer (TRB) represents one big request, the >> hardware will need a long time to prefill the hardware FIFO. This can be >> avoided by queueing more smaller requests which will then lead to >> smaller TRBs which the hardware engine can keep up with. > >Just want to clarify here to avoid any confusion, the hardware TX FIFO >size is relatively small, usually can be smaller than the TRB. It should >be fine whether the TRB is larger or smaller than the FIFO size. The >hardware does not "prefill" the FIFO. It just fills whichever TRB it's >currently processing (I think you may be mixing up with TRB cache). What I see is, that by using bigger TRBs the hardware is not able to keep up with reading from the memory when the system is under heavy memory pressure. This is the main reason for this change. Since we found out that increasing the FIFO size had an effect to how high we are able to set the hardware endpoint configuration on our gadget side (params.param0), until we saw the issue reoccur. So the Idea here was to have a tweak on how the hardware handles the data from the memory to the Hardware-FIFO which seems not to underrun with smaller TRBs. >The performance impact from this change is to reduce the USB bandwidth >usage. The current calculation in uvc function can use 48KB/uframe for >each request in SuperSpeed, the max size for isoc in SuperSpeed. I know >many hosts can't handle this kind of transfer rate from their hardware. >(e.g. It gets worse when scheduling transfers for multiple endpoints and >under multiple tier hubs). I think I don't fully understand here. We change the overall buffersize of the usb_request and therefor limit the size of possible TRBs. This should even only have most effect on the trbsize for the memcopy path, since the scatter gather requests are already split into multiple trbs which is capped to the maximum mappable memory range of PAGE_SIZE (4k). Other then that, the parameterization of the endpoint on our gadget side is not changed by this patch. The endpoint configuration is set as follows: params.param0 |= DWC3_DEPCFG_BURST_SIZE(burst - 1) | DWC3_DEPCFG_MAX_PACKET_SIZE(usb_endpoint_maxp(desc)); So by changing the request_size there should not be any other change in the gadget side hardware configuration. How is the overall bandwidth usage affected by this change then other than we have more smaller potential trbs in the queue. If the Intervallength is not coupled to the amount of to be transfered TRBs in any case, it should not have an effect to the bandwidth. If I am mistaken here, can you point me to some code? >The bandwidth can be impacted by multiple factors and not just from the >gadget side as noted in the discussion before. Right. >> This patch is simply dropping the maxburst as an multiplier for the size >> calculation and therefor decreases the overall maximum request size. >> >> Signed-off-by: Michael Grzeschik <m.grzeschik@pengutronix.de> >> --- >> This patch is created as an result from the discussion in the following thread: >> >> https://urldefense.com/v3/__https://lore.kernel.org/linux-usb/20231031231841.dbhtrgqounu2rvgn@synopsys.com/__;!!A4F2R9G_pg!fTaIo4tDljSbEvUY5SZLkNvKWcz0YeN0Ekzs0CPWyD73RGRmErRC2frODFgnMB1M4Nse0oKKgwxC65gePhGAtauKJq1Vnzlj$ >> >> drivers/usb/gadget/function/uvc_queue.c | 1 - >> drivers/usb/gadget/function/uvc_video.c | 1 - >> 2 files changed, 2 deletions(-) >> >> diff --git a/drivers/usb/gadget/function/uvc_queue.c b/drivers/usb/gadget/function/uvc_queue.c >> index 0aa3d7e1f3cc32..1d3c3c09ff97cb 100644 >> --- a/drivers/usb/gadget/function/uvc_queue.c >> +++ b/drivers/usb/gadget/function/uvc_queue.c >> @@ -55,7 +55,6 @@ static int uvc_queue_setup(struct vb2_queue *vq, >> sizes[0] = video->imagesize; >> >> req_size = video->ep->maxpacket >> - * max_t(unsigned int, video->ep->maxburst, 1) > >I think you're reducing a bit too much here? Also, take advantage of >burst. So, perhaps keep request size to max(16K, MPS * maxburst)? > >Can be more or less depending on how much video data is needed to >transfer over a video frame. > >BR, >Thinh > >> * (video->ep->mult); >> >> /* We divide by two, to increase the chance to run >> diff --git a/drivers/usb/gadget/function/uvc_video.c b/drivers/usb/gadget/function/uvc_video.c >> index 91af3b1ef0d412..c6b61843bad3d7 100644 >> --- a/drivers/usb/gadget/function/uvc_video.c >> +++ b/drivers/usb/gadget/function/uvc_video.c >> @@ -329,7 +329,6 @@ uvc_video_alloc_requests(struct uvc_video *video) >> BUG_ON(video->req_size); >> >> req_size = video->ep->maxpacket >> - * max_t(unsigned int, video->ep->maxburst, 1) >> * (video->ep->mult); >> >> video->ureq = kcalloc(video->uvc_num_requests, sizeof(struct uvc_request), GFP_KERNEL); >> -- >> 2.39.2 >> >>
On Mon, Nov 13, 2023, Michael Grzeschik wrote: > On Fri, Nov 10, 2023 at 02:16:59AM +0000, Thinh Nguyen wrote: > > On Fri, Nov 10, 2023, Michael Grzeschik wrote: > > > One misconception of queueing request to the usb gadget controller is, > > > that the amount of data that one usb_request is representing is the same > > > as the hardware is able to transfer in one interval. > > > > > > This exact idea applies to the uvc_video gadget that assumes that the > > > request needs to have the size as the gadgets bandwidth parameters are > > > configured with. (maxpacket, multiplier and burst) > > > > > > Although it is actually no problem for the hardware to queue a big > > > request, in the case of isochronous transfers it is impractical to queue > > > big amount of data per request to the hardware. Especially if the > > > necessary bandwidth was increased on purpose to maintain high amounts of > > > data. > > > > > > E.g. in the dwc3-controller it has the negative impact that the endpoint > > > FIFO will not be filled fast enough by the internal hardware engine. > > > Since each transfer buffer (TRB) represents one big request, the > > > hardware will need a long time to prefill the hardware FIFO. This can be > > > avoided by queueing more smaller requests which will then lead to > > > smaller TRBs which the hardware engine can keep up with. > > > > Just want to clarify here to avoid any confusion, the hardware TX FIFO > > size is relatively small, usually can be smaller than the TRB. It should > > be fine whether the TRB is larger or smaller than the FIFO size. The > > hardware does not "prefill" the FIFO. It just fills whichever TRB it's > > currently processing (I think you may be mixing up with TRB cache). > > What I see is, that by using bigger TRBs the hardware is not able > to keep up with reading from the memory when the system is under > heavy memory pressure. This is the main reason for this change. > > Since we found out that increasing the FIFO size had an effect to how > high we are able to set the hardware endpoint configuration on our > gadget side (params.param0), until we saw the issue reoccur. > > So the Idea here was to have a tweak on how the hardware handles the > data from the memory to the Hardware-FIFO which seems not to underrun > with smaller TRBs. > > > The performance impact from this change is to reduce the USB bandwidth > > usage. The current calculation in uvc function can use 48KB/uframe for > > each request in SuperSpeed, the max size for isoc in SuperSpeed. I know > > many hosts can't handle this kind of transfer rate from their hardware. > > (e.g. It gets worse when scheduling transfers for multiple endpoints and > > under multiple tier hubs). > > I think I don't fully understand here. > > We change the overall buffersize of the usb_request and therefor limit > the size of possible TRBs. This should even only have most effect on the > trbsize for the memcopy path, since the scatter gather requests are > already split into multiple trbs which is capped to the maximum mappable > memory range of PAGE_SIZE (4k). > > Other then that, the parameterization of the endpoint on our gadget side > is not changed by this patch. The endpoint configuration is set as follows: > > params.param0 |= DWC3_DEPCFG_BURST_SIZE(burst - 1) | > DWC3_DEPCFG_MAX_PACKET_SIZE(usb_endpoint_maxp(desc)); > > So by changing the request_size there should not be any other change in the > gadget side hardware configuration. > > How is the overall bandwidth usage affected by this change then other > than we have more smaller potential trbs in the queue. > > If the Intervallength is not coupled to the amount of to be transfered > TRBs in any case, it should not have an effect to the bandwidth. > > If I am mistaken here, can you point me to some code? > My point was to clarify that the reduction of request size would mean more time available for the transfer to go out within the given usb bandwidth. Both the host and the device sides can affect when and how much data can be sent within an interval. The host has it own calculation on how much usb bandwidth is allocated for the device and endpoint. Some hosts give more, some give less. The more data is transferred, the less time it has to send out all the data. Your tweaking of the FIFO size and the bus burst help improve the performance on the device side. I also want to note that the transfer size of the usb request is what matter since it accounts for transfer over an interval. If you use SG request, then you may have many smaller TRBs. For newer controllers, the performance should be equivalent to a request using a large TRB. Regarding the question earlier, can you cap the size of the request depending on how much video data is expected over a given video frame instead? Thanks, Thinh
diff --git a/drivers/usb/gadget/function/uvc_queue.c b/drivers/usb/gadget/function/uvc_queue.c index 0aa3d7e1f3cc32..1d3c3c09ff97cb 100644 --- a/drivers/usb/gadget/function/uvc_queue.c +++ b/drivers/usb/gadget/function/uvc_queue.c @@ -55,7 +55,6 @@ static int uvc_queue_setup(struct vb2_queue *vq, sizes[0] = video->imagesize; req_size = video->ep->maxpacket - * max_t(unsigned int, video->ep->maxburst, 1) * (video->ep->mult); /* We divide by two, to increase the chance to run diff --git a/drivers/usb/gadget/function/uvc_video.c b/drivers/usb/gadget/function/uvc_video.c index 91af3b1ef0d412..c6b61843bad3d7 100644 --- a/drivers/usb/gadget/function/uvc_video.c +++ b/drivers/usb/gadget/function/uvc_video.c @@ -329,7 +329,6 @@ uvc_video_alloc_requests(struct uvc_video *video) BUG_ON(video->req_size); req_size = video->ep->maxpacket - * max_t(unsigned int, video->ep->maxburst, 1) * (video->ep->mult); video->ureq = kcalloc(video->uvc_num_requests, sizeof(struct uvc_request), GFP_KERNEL);
One misconception of queueing request to the usb gadget controller is, that the amount of data that one usb_request is representing is the same as the hardware is able to transfer in one interval. This exact idea applies to the uvc_video gadget that assumes that the request needs to have the size as the gadgets bandwidth parameters are configured with. (maxpacket, multiplier and burst) Although it is actually no problem for the hardware to queue a big request, in the case of isochronous transfers it is impractical to queue big amount of data per request to the hardware. Especially if the necessary bandwidth was increased on purpose to maintain high amounts of data. E.g. in the dwc3-controller it has the negative impact that the endpoint FIFO will not be filled fast enough by the internal hardware engine. Since each transfer buffer (TRB) represents one big request, the hardware will need a long time to prefill the hardware FIFO. This can be avoided by queueing more smaller requests which will then lead to smaller TRBs which the hardware engine can keep up with. This patch is simply dropping the maxburst as an multiplier for the size calculation and therefor decreases the overall maximum request size. Signed-off-by: Michael Grzeschik <m.grzeschik@pengutronix.de> --- This patch is created as an result from the discussion in the following thread: https://lore.kernel.org/linux-usb/20231031231841.dbhtrgqounu2rvgn@synopsys.com/ drivers/usb/gadget/function/uvc_queue.c | 1 - drivers/usb/gadget/function/uvc_video.c | 1 - 2 files changed, 2 deletions(-)