Re: [PATCH] mmc: Move the mmc driver init earlier

[Feng Tang]

On Tue, Jun 12, 2018 at 12:29:50PM +0200, Ulf Hansson wrote:
> On 12 June 2018 at 10:42, Feng Tang <feng.tang@xxxxxxxxx> wrote:
> > Hi Ulf,
> >
> > Thanks for the review.
> >
> > On Tue, Jun 12, 2018 at 08:25:44AM +0200, Ulf Hansson wrote:
> >> On 8 June 2018 at 11:51, Feng Tang <feng.tang@xxxxxxxxx> wrote:
> >> > When doing some boot time optimization for an eMMC rootfs NUCs,
> >> > we found the rootfs may spend around 100 microseconds waiting
> >> > for eMMC card to be initialized, then the rootfs could be
> >> > mounted.
> >> >         [    1.216561] Waiting for root device /dev/mmcblk1p1...
> >> >         [    1.289262] mmc1: new HS400 MMC card at address 0001
> >> >         [    1.289667] mmcblk1: mmc1:0001 R1J56L 14.7 GiB
> >> >         [    1.289772] mmcblk1boot0: mmc1:0001 R1J56L partition 1 8.00 MiB
> >> >         [    1.289869] mmcblk1boot1: mmc1:0001 R1J56L partition 2 8.00 MiB
> >> >         [    1.289967] mmcblk1rpmb: mmc1:0001 R1J56L partition 3 4.00 MiB
> >> >         [    1.292798]  mmcblk1: p1 p2 p3
> >> >         [    1.300576] EXT4-fs (mmcblk1p1): couldn't mount as ext3 due to feature incompatibilities
> >> >         [    1.300912] EXT4-fs (mmcblk1p1): couldn't mount as ext2 due to feature incompatibilities
> >> >
> >> > And this is a common problem for smartphones, tablets, embedded
> >> > devices and automotive products. This patch will make the eMMC/SD
> >> > card  start initializing earlier, by changing its order in drivers/Makefile.
> >> >
> >> > On our platform, the waiting for eMMC card is almost eliminated with the patch,
> >> > which is critical to boot time.
> >>
> >> I am wondering what kernel version you are running here. There have
> >> been some changes to the mmc initialization path, which perhaps can
> >> help.
> > These logs in commit msg are based on kernel 4.14, and the patch is generated
> > against kernel 4.17.
> 
> Right. So it's quite recent, even if lot's of changes have been made
> to the mmc core since then.
> 
> A few things (old/new) that is important.
> 1) Check if your mmc host driver support MMC_CAP_WAIT_WHILE_BUSY. That
> should have some effect, avoiding unnecessary polling.
> 
> 2) Since 4.18 rc1, you will be able to configure an over estimated
> "power on" delay (via DT as well). Look at commit
> 6d796c68cd15234a33a4bd2ef7231125fea2dc6c.
> 
> 3) If you use a DT based platform, I think what people do is to
> re-organize the order of device nodes, such that as many as possible
> -EPROBE_DEFER is avoided to be returned by drivers. This is also not a
> good solution, but the best we have at this moment.

Thanks for the detailed info!

Will try this out, but it may take a while as we need to migrate the
whole BSP to latest kernel first.
 
> >
> >>
> >> >
> >> > Signed-off-by: Feng Tang <feng.tang@xxxxxxxxx>
> >> > ---
> >> >  drivers/Makefile | 4 +++-
> >> >  1 file changed, 3 insertions(+), 1 deletion(-)
> >> >
> >> > diff --git a/drivers/Makefile b/drivers/Makefile
> >> > index 24cd47014657..c473afd3c688 100644
> >> > --- a/drivers/Makefile
> >> > +++ b/drivers/Makefile
> >> > @@ -50,6 +50,9 @@ obj-$(CONFIG_REGULATOR)               += regulator/
> >> >  # reset controllers early, since gpu drivers might rely on them to initialize
> >> >  obj-$(CONFIG_RESET_CONTROLLER) += reset/
> >> >
> >> > +# put mmc early as many morden devices use emm/sd card as rootfs storage
> >> > +obj-y                          += mmc/
> >> > +
> >>
> >> Your suggested approach isn't really a solution, as it may work for
> >> your particular case but not for everybody else.
> >
> > Do you mean the patch may break some platforms? Yes, I only tested on
> > some IA based NUCs, and I did think about other architectures, things
> > that may affect MMC are gpio/clk/pinctrl, and those are still earlier
> > than mmc after change.
> 
> I don't know if it breaks things, potentially it could, if drivers
> don't implement support for -EPROBE_DEFER properly.
> 
> However, more importantly, it's not real fix to the problem, just
> something that seems to work for you.

Got your point. Acutally I profiled the time for every function of mmc 
init, for example, mmc_init_card() will take about 40ms, mostly due to
its mmc operation model: send a CMD and wait for completion, which
time is HW bound and doesn't have much space to optimize.

Thanks,
Feng