On 15/01/18 12:54, Adrian Hunter wrote: > On 15/01/18 11:53, Linus Walleij wrote: >> On Mon, Jan 15, 2018 at 7:07 AM, Benjamin Beckmeyer >> <beckmeyer.b@xxxxxxxxx> wrote: >> >>> Good morning, >>> >>>> - Requesting a test of this on i.MX. (Sorry Benjamin.) >>> >>> No worries. I did another test with your v5 patch, but it doesn't work. >> >> Wow that is weird. >> >> So v3 works. >> >> And it's not the code that bails out if we have nr_segs == 1 >> because that code path is not taken. >> >> It is the 64K instead of 512K buffer that makes this happen, >> then. > > That might be because you are not updating max_req_size. > >> I guess I just respin v3 with the change that we fall back to >> 1 segment if we can's allocate coherent memory for the >> bounce buffer. > > I would like to support pre_req() and post_req(). I have some code for that > if you are interested. > Then again it depends on the amount of memory allocated. In some cases, one big buffer will probably be better than 2 half the size.