On Tue, 13 Oct 2020 16:23:36 +0300 Adrian Hunter wrote: > > On 13/10/20 3:32 pm, Ulf Hansson wrote: > > + Adrian > > > > On Tue, 13 Oct 2020 at 13:09, Jisheng Zhang <Jisheng.Zhang@xxxxxxxxxxxxx> wrote: > >> > >> Hi, > >> > >> I found an emmc performance regression with below simple benchmark cmd: > >> > >> dd if=/dev/mmcblk0 of=/dev/null bs=8192 count=100000 > >> > >> This regression has been bisected to commit > >> 427b6514d0953bf (“mmc: sdhci: Add Auto CMD Auto Select support”) > >> > >> If I revert this commit, the performance is good now. > >> > >> I’m not sure whether this is common issue or not. > > > > Can you provide some data, like how much is the regression and how > > many tests did you run? before the commit, 120MB/s after the commit, 51.3MB/s tested 10 times > > > > Did you try different block sizes of the dd command? > > Also what is the SDHCI version? The patch in question is for version >= 4.1 V4.2 > > What driver? sdhci-of-dwcmshc.c, but some phy tuning code isn't mainlined yet. > > Can you tell if the patch changed which auto command is used? autocmd23 Now I have a fix, the Auto CMD auto Select can only be used when v4 mode is enabled. Although the HW supports v4.1, but drivers may haven't enabled it due to various reasons. I just sent out the fix a few minutes ago. Thanks