Hi,
On 12-08-15 14:23, [email protected] wrote:
> The 250ms timeout is too short.
>
> On my system enabling the oclk takes under 50ms and disabling slightly
> over 100ms when idle. Under load disabling the clock can take over
> 350ms.
>
> This does not make mmc clock gating look like good option to have on
> sunxi but the system should not crash with mmc clock gating enabled
> nonetheless.
>
> This patch sets the timeout to 750ms and adds debug prints which show
> how long enabling/disabling the clock took so more data can be collected
> from other systems.
>
> Signed-off-by: Michal Suchanek <[email protected]>
This is a big patch for just changing a timeout, most of this is in
extra verbosity which IMHO has little value, in the error path w
know we will have waited aprox 750 ms, so printing the waiting
time there is not worth all the extra code.
As for adding the warning I'm even less of a big fan of that,
if we need higher timeouts, we need higher timeouts, spamming the
kernel logs with that we are actually hitting the higher timeouts
is not productive IMHO.
Can you please resend this as a one-liner just changing the timeout ?
Regards,
Hans
>
> --
>
> - fix formatting
> ---
> drivers/mmc/host/sunxi-mmc.c | 22 +++++++++++++++++++---
> 1 file changed, 19 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/mmc/host/sunxi-mmc.c b/drivers/mmc/host/sunxi-mmc.c
> index 4d3e1ff..f808a02 100644
> --- a/drivers/mmc/host/sunxi-mmc.c
> +++ b/drivers/mmc/host/sunxi-mmc.c
> @@ -595,7 +595,7 @@ static irqreturn_t sunxi_mmc_handle_manual_stop(int irq, void *dev_id)
>
> static int sunxi_mmc_oclk_onoff(struct sunxi_mmc_host *host, u32 oclk_en)
> {
> - unsigned long expire = jiffies + msecs_to_jiffies(250);
> + unsigned long start, end;
> u32 rval;
>
> rval = mmc_readl(host, REG_CLKCR);
> @@ -604,6 +604,8 @@ static int sunxi_mmc_oclk_onoff(struct sunxi_mmc_host *host, u32 oclk_en)
> if (oclk_en)
> rval |= SDXC_CARD_CLOCK_ON;
>
> + start = jiffies;
> + end = start + msecs_to_jiffies(750);
> mmc_writel(host, REG_CLKCR, rval);
>
> rval = SDXC_START | SDXC_UPCLK_ONLY | SDXC_WAIT_PRE_OVER;
> @@ -611,15 +613,29 @@ static int sunxi_mmc_oclk_onoff(struct sunxi_mmc_host *host, u32 oclk_en)
>
> do {
> rval = mmc_readl(host, REG_CMDR);
> - } while (time_before(jiffies, expire) && (rval & SDXC_START));
> + } while (time_before(jiffies, end) && (rval & SDXC_START));
> + end = jiffies;
>
> /* clear irq status bits set by the command */
> mmc_writel(host, REG_RINTR,
> mmc_readl(host, REG_RINTR) & ~SDXC_SDIO_INTERRUPT);
>
> if (rval & SDXC_START) {
> - dev_err(mmc_dev(host->mmc), "fatal err update clk timeout\n");
> + dev_err(mmc_dev(host->mmc),
> + "fatal err update oclk timeout. Could not %s in %ims.\n",
> + oclk_en ? "enable" : "disable",
> + jiffies_to_msecs(end - start));
> return -EIO;
> + } else {
> + int msecs = jiffies_to_msecs(end - start);
> + const char *ing = oclk_en ? "enabling" : "disabling";
> +
> + if ((msecs > 150) || (oclk_en && (msecs > 50)))
> + dev_warn(mmc_dev(host->mmc),
> + "%s oclk took %ims", ing, msecs);
> + else
> + dev_dbg(mmc_dev(host->mmc),
> + "%s oclk took %ims", ing, msecs);
> }
>
> return 0;
>
While I can't speak for Michal obviously,
I left the debugging bit (in my v2 that i sent 2 minutes ago) as both
you and Hans where content with it back then and both acked it.
Michal, feel free to send the v3 without the debug info, unless you want
me to do it ;)
Olliver
On 12-08-15 14:32, Hans de Goede wrote:
> Hi,
>
> On 12-08-15 14:23, [email protected] wrote:
>> The 250ms timeout is too short.
>>
>> On my system enabling the oclk takes under 50ms and disabling slightly
>> over 100ms when idle. Under load disabling the clock can take over
>> 350ms.
>>
>> This does not make mmc clock gating look like good option to have on
>> sunxi but the system should not crash with mmc clock gating enabled
>> nonetheless.
>>
>> This patch sets the timeout to 750ms and adds debug prints which show
>> how long enabling/disabling the clock took so more data can be collected
>> from other systems.
>>
>> Signed-off-by: Michal Suchanek <[email protected]>
>
> This is a big patch for just changing a timeout, most of this is in
> extra verbosity which IMHO has little value, in the error path w
> know we will have waited aprox 750 ms, so printing the waiting
> time there is not worth all the extra code.
>
> As for adding the warning I'm even less of a big fan of that,
> if we need higher timeouts, we need higher timeouts, spamming the
> kernel logs with that we are actually hitting the higher timeouts
> is not productive IMHO.
>
> Can you please resend this as a one-liner just changing the timeout ?
>
> Regards,
>
> Hans
>
>
>>
>> --
>>
>> - fix formatting
>> ---
>> drivers/mmc/host/sunxi-mmc.c | 22 +++++++++++++++++++---
>> 1 file changed, 19 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/mmc/host/sunxi-mmc.c b/drivers/mmc/host/sunxi-mmc.c
>> index 4d3e1ff..f808a02 100644
>> --- a/drivers/mmc/host/sunxi-mmc.c
>> +++ b/drivers/mmc/host/sunxi-mmc.c
>> @@ -595,7 +595,7 @@ static irqreturn_t
>> sunxi_mmc_handle_manual_stop(int irq, void *dev_id)
>>
>> static int sunxi_mmc_oclk_onoff(struct sunxi_mmc_host *host, u32
>> oclk_en)
>> {
>> - unsigned long expire = jiffies + msecs_to_jiffies(250);
>> + unsigned long start, end;
>> u32 rval;
>>
>> rval = mmc_readl(host, REG_CLKCR);
>> @@ -604,6 +604,8 @@ static int sunxi_mmc_oclk_onoff(struct
>> sunxi_mmc_host *host, u32 oclk_en)
>> if (oclk_en)
>> rval |= SDXC_CARD_CLOCK_ON;
>>
>> + start = jiffies;
>> + end = start + msecs_to_jiffies(750);
>> mmc_writel(host, REG_CLKCR, rval);
>>
>> rval = SDXC_START | SDXC_UPCLK_ONLY | SDXC_WAIT_PRE_OVER;
>> @@ -611,15 +613,29 @@ static int sunxi_mmc_oclk_onoff(struct
>> sunxi_mmc_host *host, u32 oclk_en)
>>
>> do {
>> rval = mmc_readl(host, REG_CMDR);
>> - } while (time_before(jiffies, expire) && (rval & SDXC_START));
>> + } while (time_before(jiffies, end) && (rval & SDXC_START));
>> + end = jiffies;
>>
>> /* clear irq status bits set by the command */
>> mmc_writel(host, REG_RINTR,
>> mmc_readl(host, REG_RINTR) & ~SDXC_SDIO_INTERRUPT);
>>
>> if (rval & SDXC_START) {
>> - dev_err(mmc_dev(host->mmc), "fatal err update clk timeout\n");
>> + dev_err(mmc_dev(host->mmc),
>> + "fatal err update oclk timeout. Could not %s in %ims.\n",
>> + oclk_en ? "enable" : "disable",
>> + jiffies_to_msecs(end - start));
>> return -EIO;
>> + } else {
>> + int msecs = jiffies_to_msecs(end - start);
>> + const char *ing = oclk_en ? "enabling" : "disabling";
>> +
>> + if ((msecs > 150) || (oclk_en && (msecs > 50)))
>> + dev_warn(mmc_dev(host->mmc),
>> + "%s oclk took %ims", ing, msecs);
>> + else
>> + dev_dbg(mmc_dev(host->mmc),
>> + "%s oclk took %ims", ing, msecs);
>> }
>>
>> return 0;
>>
>
On 12 August 2015 at 14:40, Olliver Schinagl <[email protected]> wrote:
>
> On 12-08-15 14:32, Hans de Goede wrote:
>>
>> Hi,
>>
>> On 12-08-15 14:23, [email protected] wrote:
>>>
>>> The 250ms timeout is too short.
>>>
>>> On my system enabling the oclk takes under 50ms and disabling slightly
>>> over 100ms when idle. Under load disabling the clock can take over
>>> 350ms.
>>>
>>> This does not make mmc clock gating look like good option to have on
>>> sunxi but the system should not crash with mmc clock gating enabled
>>> nonetheless.
>>>
>>> This patch sets the timeout to 750ms and adds debug prints which show
>>> how long enabling/disabling the clock took so more data can be collected
>>> from other systems.
>>>
>>> Signed-off-by: Michal Suchanek <[email protected]>
>>
>>
>> This is a big patch for just changing a timeout, most of this is in
>> extra verbosity which IMHO has little value, in the error path w
>> know we will have waited aprox 750 ms, so printing the waiting
>> time there is not worth all the extra code.
>>
>> As for adding the warning I'm even less of a big fan of that,
>> if we need higher timeouts, we need higher timeouts, spamming the
>> kernel logs with that we are actually hitting the higher timeouts
>> is not productive IMHO.
>>
>> Can you please resend this as a one-liner just changing the timeout ?
>>
If we expect the timeout to be short and it isn't that means something is wrong.
The user is probably experiencing degraded performance in that case so
it's good to have a diagnostic for it IMHO.
Yes, printing of the timeout in the error case does not have that much
value but since it's calculated anyway I just pasted it there.
> While I can't speak for Michal obviously,
>
> I left the debugging bit (in my v2 that i sent 2 minutes ago) as both you
> and Hans where content with it back then and both acked it.
>
> Michal, feel free to send the v3 without the debug info, unless you want me
> to do it ;)
>
I don't really care either way so long as the boards do not crash.
Thanks
Michal