Date:   Tue, 10 Mar 2020 10:46:55 +0100
From:   Ulf Hansson <ulf.hansson@...aro.org>
To:     Sowjanya Komatineni <skomatineni@...dia.com>
Cc:     Jon Hunter <jonathanh@...dia.com>,
        Bitan Biswas <bbiswas@...dia.com>,
        Adrian Hunter <adrian.hunter@...el.com>,
        Naresh Kamboju <naresh.kamboju@...aro.org>,
        Jens Axboe <axboe@...nel.dk>,
        Alexei Starovoitov <ast@...nel.org>,
        linux-block <linux-block@...r.kernel.org>,
        lkft-triage@...ts.linaro.org,
        open list <linux-kernel@...r.kernel.org>,
        "linux-mmc@...r.kernel.org" <linux-mmc@...r.kernel.org>,
        Arnd Bergmann <arnd@...db.de>,
        John Stultz <john.stultz@...aro.org>,
        Faiz Abbas <faiz_abbas@...com>,
        Thierry Reding <treding@...dia.com>,
        Anders Roxell <anders.roxell@...aro.org>,
        Kishon <kishon@...com>
Subject: Re: LKFT: arm x15: mmc1: cache flush error -110

On Mon, 9 Mar 2020 at 18:33, Sowjanya Komatineni <skomatineni@...dia.com> wrote:
>
>
> On 3/6/20 3:14 AM, Ulf Hansson wrote:
> > External email: Use caution opening links or attachments
> >
> >
> > [...]
> >
> >>>>>>>>>>> Actually we always use R1B with CMD6 as per spec.
> >>>>>>>>>> I fully agree that R1B is preferable, but it's not against the
> >>>>>>>>>> spec to
> >>>>>>>>>> send CMD13 to poll for busy.
> >>>>>>>>>>
> >>>>>>>>>> Moreover, we need to cope with the scenario when the host has
> >>>>>>>>>> specified a maximum timeout that isn't sufficiently long enough for
> >>>>>>>>>> the requested operation. Do you have another proposal for how to
> >>>>>>>>>> manage this, but disabling MMC_RSP_BUSY?
> >>>>>>>>>>
> >>>>>>>>>> Let's assume you driver would get a R1B for the CMD6 (we force it),
> >>>>>>>>>> then what timeout would the driver be using if we would set
> >>>>>>>>>> cmd.busy_timeout to 30ms?
> >>>>>>>>>>
> >>>> Sorry didn't understood clearly. Are you asking with 30s timeout, whats
> >>>> the data timeout counter used?
> >>> Yes. It seems like it will pick the maximum, which is 11s?
> >> yes
> > Okay, thanks!
> >
> >>>> Because of above mentioned issue on our host where CMD interrupt happens
> >>>> after busy state, poll for busy returns right away as not busy.
> >>> I see.
> >>>
> >>>> So issuing CMD13 after CMD6-R1 followed by busy poll should be working.
> >>>> But weird that with small delay of 1ms or debug print before CMD13 it
> >>>> doesn't timeout and works all the time.
> >>> I have digested the information you provided in these emails. Let me
> >>> summarize it, to see if I have understood correctly.
> >>>
> >>> 1.
> >>> Your controller can't distinguish between R1 and R1B because of a
> >>> limitation in the HW. So, in both cases you need to wait for the card
> >>> to stop signal busy, before the controller can give an IRQ to notify
> >>> that the R1 response has been received. Correct?
> >>>
> >>> In this context, I am wondering if sdhci_send_command(), really
> >>> conforms to these requirements. For example, depending on if the CMD6
> >>> has MMC_RSP_BUSY or not, it may pick either SDHCI_CMD_RESP_SHORT or
> >>> SDHCI_CMD_RESP_SHORT_BUSY.
> >>>
> >>> Does this work as expected for your case?
> >> Design team re-verified internally and bug where HW waits for busy state
> >> before IRQ is only for R1B and R1 is spec compliant.
> >>
> >> So, with R1, CMD complete is generated after response received.
> > Okay.
> >
> > So, the issue we see for CMD6 with R1, is a software problem that we
> > should be able to fix.
> >
> >> With R1B, CMD complete and xfer complete both are generated after
> >> response received + device busy (max timeout of 11s)
> >> DATA timeout interrupt will be asserted incase if HW busy detection fails.
> >>
> >> With R1B we may see DATA Timeout if operation takes more than max busy
> >> timeout of 11s.
> > Okay, I see.
> >
> >>> 2.
> >>> Assuming my interpretation of the above is somewhat correct. Then you
> >>> always need to set a busy timeout for R1/R1B responses in the
> >>> controller. The maximum timeout seems to be 11s long. Obviously, this
> >>> isn't enough for all cases, such as cache flushing and erase, for
> >>> example. So, what can we do to support a longer timeouts than 11s?
> >>> Would it be possible to disable the HW timeout, if the requested
> >>> timeout is longer than 11s and use a SW timeout instead?
> >>>
> >>> Kind regards
> >>> Uffe
> >> For erase long operations we have register bit to enable for infinite
> >> busy wait mode where host controller would be monitoring until card is busy.
> > Alright, that sounds great!
> >
> >> But so far for emmc devices we used on our platforms, we haven't seen
> >> cache flush taking more than 11s.
> > I understand that 11s is probably fine to use, for most cases.
> >
> > However, it's not spec compliant, as for some operations there are
> > simply no timeout specified. BKOPS, cache flush, sanitize are cases
> > like this - and then 11s is definitely not sufficient.
> >
> >> Will get back on possibility of disabling HW timeout and using SW timeout..
> > Thanks!
> >
> > I would like to get the regression fixed asap, but I also would like
> > to avoid reverting patches, unless really necessary. May I propose the
> > following two options.
> >
> > 1. Find out why polling with ->card_busy() or CMD13, for a CMD6 with
> > an R1 response doesn't work - and then fix that behaviour.
> >
> > 2. Set the mmc->max_busy_timeout to zero for sdhci-tegra, which makes
> > the core to always use R1B for CMD6 (and erase). This also means that
> > when the cmd->busy_timeout becomes longer than 11s, sdhci-tegra must
> > disable the HW busy timeout and just wait "forever".
> >
> > If you decide for 2, you can add the software timeout support on top,
> > but make that can be considered as a next step of an improvement,
> > rather than needed as fix. Note that, I believe there are some support
> > for software timeout already in the sdhci core, maybe you need to
> > tweak it a bit for your case, I don't know.
> >
> > Kind regards
> > Uffe
>
> Hi Uffe
>
> Will go with 2nd option and will send patches out when ready.

Okay, good.

>
> BTW, Tegra host also supports SDHCI_QUIRK_DATA_TIMEOUT_USES_SDCLK for
> data timeout based on host clock when using finite mode (HW busy
> detection based on DATA TIMEOUT count value when cmd operation timeout
> is < 11s for tegra host).
>
> So, looks like we cant set host max_busy_timeout to 0 for Tegra host to
> force R1B during SWITCH and SLEEP_AWAKE.
>
> So, was thinking to introduce host capability MMC_CAP2_LONG_WAIT_HW_BUSY
> which can be used for hosts supporting long or infinite HW busy wait
> detection and will update mmc and mmc_ops drivers to not allow convert
> R1B to R1B for hosts with this capability during SLEEP_AWAKE and SWITCH.

That seems reasonable, it becomes probably both easier and clearer by
adding a new host cap.

In any case, let me help out and cook a patch for this for the core
part (I leave the sdhci change to you). It may be a bit tricky,
especially since I have currently queued a bunch of new changes for
v5.7, that enables more users of mmc_poll_for_busy() in the core.
Maybe I need to temporarily drop them, so we can fix these problems
first. I will check.

Probably, I would also name the cap MMC_CAP_HW_NEED_RSP_BUSY, as that
seems to be describing the common problem we have for sdhci
omap/tegra.

Finally, it seems like MMC_CAP_WAIT_WHILE_BUSY should be set for
sdhci- tegra, so while at it, perhaps you can cook a patch for that as
well.

Kind regards
Uffe

Powered by blists - more mailing lists