On Sunday 06 May 2007, Markus Rechberger wrote: [snip] > On 5/6/07, Laurent Pinchart <[EMAIL PROTECTED]> wrote: > > On Sunday 06 May 2007, Markus Rechberger wrote: > > > On 5/6/07, Laurent Pinchart <[EMAIL PROTECTED]> wrote: > > > > On Sunday 06 May 2007, Markus Rechberger wrote: > > > > > On 5/6/07, Laurent Pinchart <[EMAIL PROTECTED]> wrote: > > > > > dwMaxVideoFrameBufferSize 2752512 (which is used for > > > > > usb_buffer_alloc, seems to be too much already) > > > > > > > > The bulk buffers are allocated using dwMaxPayloadTransferSize. > > > > Isochronous buffers are allocated using dwMaxVideoFrameSize, which is > > > > clearly a bug. I'm working on a fix. > > > > > > > > The stack trace you posted shows your device uses bulk transfers. > > > > Could you tell me how big dwMaxPayloadTransferSize is ? If the > > > > devices insists on using > > > > huge buffers, there's not much I can do about it. > > > > > > I simply wouldn't allocate that much memory for the urbs, this way > > > seems to work reliable with the driver I have. > > > I added some debug at the allocation it is 2.7 MB/URB > > > > I suspect that the dwMaxPayloadTransferSize value is thus equal to > > 2752512 as > > well. > > > > For bulk transfers, you are right, I could allocate smaller URBs. As bulk > > transfers are guaranteed, I can split payloads across multiple URBs. For > > isochronous transfers that won't be possible. > > it is also possible for isochronous transfers, if the framebuffer > doesn't get fully filled just drop the frame. By experience this works very > well with em28xx based devices.
After having a closer look at the spec, isochronous transfers won't be an issue. Each isochronous packet starts with a UVC header. I already submit isochronous URBs of no more than 40 packets. For bulk transfers, the situation seems a little more complex than I thought. Bulk payloads can be made of lots of packets, up to a whole frame per payload. The maximum payload size is reported by the device. As the device sends a header for each payload and not for each URB, I have to find a way detect payload boundaries (keep in mind that payloads can have a variable size). When using payload-sized URBs, each URB will contain a single payload. This is what the Linux UVC driver currently does, and it works fine as long as the device doesn't have a payload size bigger than what the host can allocate. When splitting payloads across multiple URBs, I need another way to detect payload boundaries. As the payload size isn't known in advance (only the maximum size is known), I can't rely on a byte count. A possible solution would have been to detect URBs with an actual length smaller than the transfer buffer length, as such URBs should indicate the end of a payload. Unfortunately, this won't work as the payload size can be an exact multiple of the URB size. Can you think of a way to detect payload boundaries properly when splitting payloads across multiple URBs ? [snip] Laurent Pinchart _______________________________________________ Linux-uvc-devel mailing list Linux-uvc-devel@lists.berlios.de https://lists.berlios.de/mailman/listinfo/linux-uvc-devel
