Replace urb buffers allocation to use the DMA noncontiguous API. Similarly to a commit sent to uvc: [1] This improves performance on Arm. The series was tested on rock-pi4 (Arm64) and shows about 10x improvements: NON_CONTIGUOUS uS/qty: 9231963/106048 sec: 9231.963472 avg: 87.054 min: 0.291 max: 216.423 (uS) raw decode speed: 9.217 Gbits/s bytes 9.971904900 G bytes/urb: 100290 COHERENT uS/qty: 92423219/98904 sec: 92423.219509 avg: 934.474 min: 0.583 max: 2196.583 (uS) raw decode speed: 787.698 Mbits/s bytes 8.510250828 G bytes/urb: 92010 On x86 there is no significant difference: NON_CONTIGUOUS uS/qty: 3203891/50096 sec: 3203.891724 avg: 63.955 min: 0.141 max: 11696.636 (uS) raw decode speed: 23.783 Gbits/s bytes 8.932543668 G bytes/urb: 190084 COHERENT uS/qty: 3081680/50112 sec: 3081.680438 avg: 61.495 min: 0.160 max: 482.099 (uS) raw decode speed: 21.557 Gbits/s bytes 7.786260844 G bytes/urb: 165677 [1] https://lkml.org/lkml/2021/3/12/1506 Dafna Hirschfeld (3): media: stk1160: fix number of buffers in case not all buffers are created media: stk1160: move transfer_buffer and urb to same struct 'stk1160_urb' media: stk1160: use dma_alloc_noncontiguous API drivers/media/usb/stk1160/stk1160-v4l.c | 5 +- drivers/media/usb/stk1160/stk1160-video.c | 137 +++++++++++----------- drivers/media/usb/stk1160/stk1160.h | 21 +++- 3 files changed, 91 insertions(+), 72 deletions(-) -- 2.17.1
