Re: [PATCH] mmc: core: don't return 1 for max_discard

[Adrian Hunter <adrian.hunter-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>]

On 19/12/13 12:26, Ulf Hansson wrote:
> On 19 December 2013 10:42, Adrian Hunter <adrian.hunter-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org> wrote:
>> On 19/12/13 11:14, Vladimir Zapolskiy wrote:
>>> On 12/19/13 10:01, Adrian Hunter wrote:
>>>> On 19/12/13 01:00, Stephen Warren wrote:
>>>>> On 12/18/2013 03:27 PM, Stephen Warren wrote:
>>>>>> From: Stephen Warren<swarren-DDmLM1+adcrQT0dZR+AlfA@public.gmane.org>
>>>>>>
>>>>>> In mmc_do_calc_max_discard(), if only a single erase block can be
>>>>>> discarded within the host controller's timeout, don't allow discard
>>>>>> operations at all.
>>>>>>
>>>>>> Previously, the code allowed sector-at-a-time discard (rather than
>>>>>> erase-block-at-a-time), which was chronically slow.
>>>>>>
>>>>>> Without this patch, on the NVIDIA Tegra Cardhu board, the loops result
>>>>>> in qty == 1, which is immediately returned. This causes discard to
>>>>>> operate a single sector at a time, which is chronically slow. With this
>>>>>> patch in place, discard operates a single erase block at a time, which
>>>>>> is reasonably fast.
>>>>>
>>>>> Alternatively, is the real fix a revert of e056a1b5b67b "mmc: queue: let
>>>>> host controllers specify maximum discard timeout", followed by:
>>>>>
>>>>>> diff --git a/drivers/mmc/core/core.c b/drivers/mmc/core/core.c
>>>>>> index 050eb262485c..35c5b5d86c99 100644
>>>>>> --- a/drivers/mmc/core/core.c
>>>>>> +++ b/drivers/mmc/core/core.c
>>>>>> @@ -1950,7 +1950,6 @@ static int mmc_do_erase(struct mmc_card *card,
>>>>>> unsigned int from,
>>>>>>          cmd.opcode = MMC_ERASE;
>>>>>>          cmd.arg = arg;
>>>>>>          cmd.flags = MMC_RSP_SPI_R1B | MMC_RSP_R1B | MMC_CMD_AC;
>>>>>> -       cmd.cmd_timeout_ms = mmc_erase_timeout(card, arg, qty);
>>>>>>          err = mmc_wait_for_cmd(card->host,&cmd, 0);
>>>>>>          if (err) {
>>>>>>                  pr_err("mmc_erase: erase error %d, status %#x\n",
>>>>>> @@ -1962,7 +1961,7 @@ static int mmc_do_erase(struct mmc_card *card,
>>>>>> unsigned int from,
>>>>>>          if (mmc_host_is_spi(card->host))
>>>>>>                  goto out;
>>>>>>
>>>>>> -       timeout = jiffies + msecs_to_jiffies(MMC_CORE_TIMEOUT_MS);
>>>>>> +       timeout = jiffies + msecs_to_jiffies(mmc_erase_timeout(card,
>>>>>> arg, qty));
>>>>>>          do {
>>>>>>                  memset(&cmd, 0, sizeof(struct mmc_command));
>>>>>>                  cmd.opcode = MMC_SEND_STATUS;
>>>>>
>>>>> That certainly also seems to solve the problem on my board...
>>>>
>>>> But large erases will timeout when they should have been split into smaller
>>>> chunks.
>>>>
>>>> A generic solution needs to be able to explain what happens when the host
>>>> controller *does* timeout.
>>>
>>> Please correct me, but if Data Timeout Error is disabled, then this is not
>>> an issue for most of the host controllers.
>>
>> That is a very good point.  My experience with SDHCI was that masking the
>> "Data Timeout Error Status Enable" and "Data Timeout Error Signal Enable
>> " bits did not disable the timeout i.e. the host controller would not
>> deliver a TC interrupt if the erase exceeded the timeout.
>>
>> What happens on your board?
>>
> 
> I posted a response yesterday for "[PATCH] mmc: core: don't decrement
> qty when calculating max_discard", related to this. Please have a
> look.
> 
> I think the interesting case to consider here is how we can handle
> busy detection timeouts that is bigger than what the host hw can
> support.
> 
> Option 1)
> Should we tell the host to disable the timeout in this case? That
> potentially means hanging forever - if the card misbehaves. Like
> omap_hsmmc does for erase commands. Maybe that is an okay limitation?

sdhci anyway has a 10 second timer to catch unresponsive host controllers.
I recently sent a patch to use the cmd_timeout_ms if it is bigger than 10
seconds.

	http://permalink.gmane.org/gmane.linux.kernel.mmc/23557

> 
> Option 2)
> Use a R1 response instead if R1B to prevent the host from doing busy
> detection. Then rely on the CMD13 to poll for completion instead.
> Obviously we can then stop polling after some selected timeout is the
> card don't complete it's operations.

It would be nice to avoid polling when the timeout can be supported. Also
the polling should be periodic.

> 
> Would be very interesting to know what option you prefer!?

At least 1 of the host controllers I have seen does not support disabling
the timeout - so option 1) might not work in all cases.  Although it is the
nicer option i.e. replace the hardware timeout with a software timeout.

So I would probably allow both options to co-exist.

> 
> Kind regards
> Uffe
> 
>