Re: [PATCH] mmc: renesas_sdhi_internal_dmac: Add MMC_CAP_ERASE to Gen3 SoCs

[Ulf Hansson]

On Thu, 14 Nov 2019 at 12:37, Eugeniu Rosca <erosca@xxxxxxxxxxxxxx> wrote:
>
> Hi everyone,
>
> On Thu, Nov 14, 2019 at 11:56:23AM +0100, Ulf Hansson wrote:
> > On Tue, 12 Nov 2019 at 21:49, Wolfram Sang <wsa@xxxxxxxxxxxxx> wrote:
> > >
> > > On Tue, Nov 12, 2019 at 02:48:08PM +0100, Eugeniu Rosca wrote:
> > > > From: Harish Jenny K N <harish_kandiga@xxxxxxxxxx>
> > > >
> > > > Enable MMC_CAP_ERASE capability in the driver to allow
> > > > erase/discard/trim requests.
> > > >
> > > > Suggested-by: Andrew Gabbasov <andrew_gabbasov@xxxxxxxxxx>
> > > > Signed-off-by: Harish Jenny K N <harish_kandiga@xxxxxxxxxx>
> > > > [erosca: Forward-port and test on v5.4-rc7 using H3ULCB-KF:
> > > >          "blkdiscard /dev/mmcblk0" passes with this patch applied
> > > >          and complains otherwise:
> > > >        "BLKDISCARD ioctl failed: Operation not supported"]
> > > > Signed-off-by: Eugeniu Rosca <erosca@xxxxxxxxxxxxxx>
> > >
> > > Looks good to me. Just a generic question, probably more for Ulf:
> > >
> > > Why does this CAP_ERASE exist? As I understand, the driver only needs to
> > > set the flag and no further handling is required. Why would a driver not
> > > set this flag and not support erase/trim commands?
> >
> > I am working on removing the cap, altogether. Step by step, this is
> > getting closer now.
> >
> > The main problem has been about busy detect timeouts, as an erase
> > command may have a very long busy timeout. On the host side, they
> > typically need to respect the cmd->busy_timeout for the request, and
> > if it can't because of some HW limitation, it needs to set
> > mmc->max_busy_timeout.
>
> FWIW we've discussed such concerns internally, based on past commits
> which either disable [1-2] busy timeouts or increase their value [3].
>
> To get a feeling if this is relevant for R-Car3, I've run blkdiscard on
> a 64 GiB eMMC without noticing any issues on v5.4-rc7. Hopefully this
> is sufficient as testing?

Let's first take a step back, because I don't know how the HW busy
detection works for your controller.

I have noticed there is TMIO_STAT_CMD_BUSY bit being set for some
variants, which seems to cause renesas_sdhi_wait_idle() to loop for a
pre-defined number of loops/timeout. This looks scary, but I can't
tell if it's really a problem.

BTW, do you know what TMIO_STAT_CMD_BUSY actually is monitoring?

I have also noticed that MMC_CAP_WAIT_WHILE_BUSY isn't set for any of
the renesas/tmio variant hosts. Is that simply because the HW doesn't
support this? Or because implementation is missing?

If you want to run a test that stretches the behaviour on the timeout
path, I would rather use an SD-card (the older the better). For eMMCs
the erase likely translates to a trim/discard, which is far more
quicker than a real erase - as is what happens on an old SD card.

>
> >
> > Once that is fixed for all, we can drop CAP_ERASE.
> >
> > Kind regards
> > Uffe
>
> [1] 93caf8e69eac76 ("omap_hsmmc: add erase capability")
> [2] b13d1f0f9ad64b ("mmc: omap: Add erase capability")
> [3] ec30f11e821f2d ("mmc: rtsx_usb: Use the provided busy timeout from the mmc core")
>
> --
> Best Regards,
> Eugeniu

Kind regards
Uffe