On 12 June 2018 at 12:25, Marek Szyprowski <m.szyprowski@xxxxxxxxxxx> wrote: > Hi Ulf, > > On 2018-06-12 12:07, Ulf Hansson wrote: >> On 12 June 2018 at 11:52, Marek Szyprowski <m.szyprowski@xxxxxxxxxxx> wrote: >>> On 2018-06-12 11:20, Ulf Hansson wrote: >>>> On 12 June 2018 at 10:28, Marek Szyprowski <m.szyprowski@xxxxxxxxxxx> wrote: >>>>> On 2018-06-11 14:24, Ulf Hansson wrote: >>>>>> On 11 June 2018 at 11:50, Marek Szyprowski <m.szyprowski@xxxxxxxxxxx> wrote: >>>>>>> On 2018-06-11 11:35, Ulf Hansson wrote: >>>>>>>> On 11 June 2018 at 08:48, Marek Szyprowski <m.szyprowski@xxxxxxxxxxx> wrote: >>>>>>>>> dw_mci_exynos_resume_noirq() performs DWMMC register access without >>>>>>>>> ensuring that respective clocks are enabled. This might cause external >>>>>>>>> abort on some systems (observed on Exynos5433 based boards). Fix this >>>>>>>>> by adding needed prepare_enable/disable_unprepare calls. >>>>>>>>> >>>>>>>>> Signed-off-by: Marek Szyprowski <m.szyprowski@xxxxxxxxxxx> >>>>>>>>> --- >>>>>>>>> drivers/mmc/host/dw_mmc-exynos.c | 6 ++++++ >>>>>>>>> 1 file changed, 6 insertions(+) >>>>>>>>> >>>>>>>>> diff --git a/drivers/mmc/host/dw_mmc-exynos.c b/drivers/mmc/host/dw_mmc-exynos.c >>>>>>>>> index 3164681108ae..6125b68726b0 100644 >>>>>>>>> --- a/drivers/mmc/host/dw_mmc-exynos.c >>>>>>>>> +++ b/drivers/mmc/host/dw_mmc-exynos.c >>>>>>>>> @@ -193,6 +193,9 @@ static int dw_mci_exynos_resume_noirq(struct device *dev) >>>>>>>>> struct dw_mci_exynos_priv_data *priv = host->priv; >>>>>>>>> u32 clksel; >>>>>>>>> >>>>>>>>> + clk_prepare_enable(host->biu_clk); >>>>>>>>> + clk_prepare_enable(host->ciu_clk); >>>>>>>>> + >>>>>>>>> if (priv->ctrl_type == DW_MCI_TYPE_EXYNOS7 || >>>>>>>>> priv->ctrl_type == DW_MCI_TYPE_EXYNOS7_SMU) >>>>>>>>> clksel = mci_readl(host, CLKSEL64); >>>>>>>>> @@ -207,6 +210,9 @@ static int dw_mci_exynos_resume_noirq(struct device *dev) >>>>>>>>> mci_writel(host, CLKSEL, clksel); >>>>>>>>> } >>>>>>>>> >>>>>>>>> + clk_disable_unprepare(host->biu_clk); >>>>>>>>> + clk_disable_unprepare(host->ciu_clk); >>>>>>>>> + >>>>>>>>> return 0; >>>>>>>>> } >>>>>>>>> #else >>>>>>>> I looked a little closer and I am wondering if it wouldn't be possible >>>>>>>> to use SET_NOIRQ_SYSTEM_SLEEP_PM_OPS() instead of >>>>>>>> SET_SYSTEM_SLEEP_PM_OPS()? >>>>>>>> >>>>>>>> Somelike this: >>>>>>>> SET_NOIRQ_SYSTEM_SLEEP_PM_OPS(pm_runtime_force_suspend, >>>>>>>> dw_mci_exynos_resume_noirq) >>>>>>>> >>>>>>>> Then from dw_mci_exynos_resume_noirq() call pm_runtime_force_resume(). >>>>>>>> >>>>>>>> I think it would simplify the code a bit, as you can rely on the >>>>>>>> runtime PM callbacks to deal with clk_prepare_enable() and >>>>>>>> clk_disable_unprepare(), unless I am mistaken. >>>>>>> This will not fix the problem, because mci_writel() calls in >>>>>>> dw_mci_exynos_resume_noirq are done unconditionally, regardless of the >>>>>>> controller's runtime pm state. Since commit 1d9174fbc55e after calling >>>>>>> pm_runtime_force_resume() there is no guarantee that device is in >>>>>>> runtime active state if it was runtime suspended state. >>>>>> Yes, because the runtime PM usage count is greater than 1. >>>>>> (pm_runtime_get_noresume() is called during probe). >>>>>> >>>>>> If you want to make this explicit (not relying on ->probe()), one can >>>>>> add a ->suspend_noirq() callback and call pm_runtime_get_noresume() in >>>>>> it. >>>>> Sorry, but I don't get how this would work. Exactly the same pattern as >>>>> you have proposed was already used in s3c-64xx SPI driver and it didn't >>>>> work properly (tested on the same SoC as this DW-MMC change). I had to >>>>> move register access to runtime resume callback to fix external abort >>>>> issue: >>>>> >>>>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=e935dba111621bd6a0c5d48e6511a4d9885103b4 >>>> Yep, that is a correct solution. >>>> >>>>> Here in DW-MMC driver such approach (moving all the code to runtime >>>>> resume callback) is not possible because of the potential interrupt storm >>>>> caused by the hw bug (that's the reason of using noirq resume callback). >>>> I understand. What you need is to run the runtime resume/suspend >>>> callbacks in the resume/suspend noirq phase. Moreover, you need to >>>> make sure that the runtime resume callback, really becomes invoked >>>> during the resume noirq phase, because of the HW bug. >>>> >>>> I think the below should work. Can you give it a try? >>>> >>>> It relies on the call pm_runtime_get_noresume(), done during >>>> ->probe(). Note that, the driver always keeps the RPM usage count >>>> increased, thus preventing runtime suspend during normal execution. >>>> >>>> Anyway, if this doesn't work, your suggested approach works fine as well. >>> Okay, finally I got it. I wasn't aware that dw_mmc-exynos keeps device >>> runtime active all the time between the driver probe() and remove(). >>> Right, this will fix this specific case, but it isn't a generic solution, >>> so I will also add a comment on that, so one would not need to debug it >>> again if he decides to change runtime pm usage scheme in dw_mmc-exynos >>> in the future. >> Seems reasonable! >> >> If you want the more generic solution, I would add a exynos specific >> suspend_noirq() callback, let it call pm_runtime_get_noresume() and >> them pm_runtime_force_suspend(). >> >> In the corresponding resume_noirq() callback, extend my suggested >> changes, with a call to pm_runtime_put_noidle() after all actions has >> been done in it. > > Okay, this finally looks like a proper and future-proof solution. Just > one more question: why pm_runtime_put_noidle()? Hypothetically, when the > dw_mmc-exynos driver gets proper runtime PM management and system suspend > will happen when the device is in runtime suspended state, this will leave > it in runtime active state with refcount = 0 after the system resume cycle. > Imho simple pm_runtime_put() will be better in this case. It doesn't really matter. Runtime PM is disabled for the device (done by driver core) and the PM workqueue is suspended, which means pm_runtime_put() will not immediately runtime suspend the device. The important part is to restore the usage count, such that the driver core can potentially runtime suspend the device when device_complete() is called for it. Kind regards Uffe -- To unsubscribe from this list: send the line "unsubscribe linux-mmc" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
