[16:16] *** benjiG has left [18:11] *** awalls has left [20:16] <ndufresne> hverkuil: pinchartl: I think we have a serious performance regression on mapping vmalloc dmabufs [20:17] <ndufresne> kernel 4.4, what I'm testing is a UVC camera, giving a dmabuf/jpeg, for each frame produce, I mmap, decode, unmmap [20:17] <ndufresne> the unmap call uses a full core [20:18] <ndufresne> I do expect an overhead over keeping those mapped at all time, but a full i7 core, that way too much [20:19] <ndufresne> (720p here, and 640 uses half a core) [20:24] <ndufresne> wait, maybe I'm wrong [20:36] <ndufresne> hverkuil: pinchartl: nevermind, sorry for the noise [20:41] <pinchartl> ndufresne: no worries, happy to have helped :-) [20:43] <ndufresne> Someone broke the code that detect packetized data, and endup parsing the data, + the parser is completly broken [20:43] <ndufresne> it maps/unmap unside the parsing loop [21:03] <pinchartl> ouch [21:05] <ndufresne> (hundreds of time per frame, a terrible regression imho) [21:05] * ndufresne think Samsung should test their patches ...