Hi Ben, Thanks for reviewing our patch. On Mon, Aug 28, 2023 at 12:03 PM Ben Chuang <benchuanggli@xxxxxxxxx> wrote: > > Hi, > > On 24/08/23 20:18, Adrian Hunter wrote: > > On 24/08/23 14:50, Stanisław Kardach wrote: > > > Hi Adrian, > > > > > > Thanks for reviewing our patches. > > > > > > On Thu, Aug 24, 2023 at 1:47 PM Adrian Hunter <adrian.hunter@xxxxxxxxx <mailto:adrian.hunter@xxxxxxxxx>> wrote: > > > > > > Hi > > > > > > Looks OK - a few minor comments below > > > > > > On 23/08/23 20:41, Sven van Ashbrook wrote: > > > > To improve the r/w performance of GL9763E, the current driver inhibits LPM > > > > negotiation while the device is active. > > > > > > > > This prevents a large number of SoCs from suspending, notably x86 systems > > > > > > If possible, can you give example of which SoCs / products > > > > > > > which use S0ix as the suspend mechanism: > > > > 1. Userspace initiates s2idle suspend (e.g. via writing to > > > > /sys/power/state) > > > > 2. This switches the runtime_pm device state to active, which disables > > > > LPM negotiation, then calls the "regular" suspend callback > > > > 3. With LPM negotiation disabled, the bus cannot enter low-power state > > > > 4. On a large number of SoCs, if the bus not in a low-power state, S0ix > > > > cannot be entered, which in turn prevents the SoC from entering > > > > suspend. > > > > > > > > Fix by re-enabling LPM negotiation in the device's suspend callback. > > > > > > > > Suggested-by: Stanislaw Kardach <skardach@xxxxxxxxxx <mailto:skardach@xxxxxxxxxx>> > > > > Fixes: f9e5b33934ce ("mmc: host: Improve I/O read/write performance for GL9763E") > > > > Cc: stable@xxxxxxxxxxxxxxx <mailto:stable@xxxxxxxxxxxxxxx> > > > > Signed-off-by: Sven van Ashbrook <svenva@xxxxxxxxxxxx <mailto:svenva@xxxxxxxxxxxx>> > > > > # on gladios device > > > > # on 15590.0.0 with v5.10 and upstream (v6.4) kernels > > > > > > > > > > 3 extraneous lines here - please remove > > > > > > > --- > > > > > > > > Changes in v2: > > > > - improved symmetry and error path in s2idle suspend callback (internal review) > > > > > > > > drivers/mmc/host/sdhci-pci-gli.c | 102 +++++++++++++++++++------------ > > > > 1 file changed, 64 insertions(+), 38 deletions(-) > > > > > > > > diff --git a/drivers/mmc/host/sdhci-pci-gli.c b/drivers/mmc/host/sdhci-pci-gli.c > > > > index 1792665c9494a..19f577cc8bceb 100644 > > > > --- a/drivers/mmc/host/sdhci-pci-gli.c > > > > +++ b/drivers/mmc/host/sdhci-pci-gli.c > > > > @@ -745,42 +745,6 @@ static u32 sdhci_gl9750_readl(struct sdhci_host *host, int reg) > > > > return value; > > > > } > > > > > > > > -#ifdef CONFIG_PM_SLEEP > > > > -static int sdhci_pci_gli_resume(struct sdhci_pci_chip *chip) > > > > -{ > > > > - struct sdhci_pci_slot *slot = chip->slots[0]; > > > > - > > > > - pci_free_irq_vectors(slot->chip->pdev); > > > > - gli_pcie_enable_msi(slot); > > > > - > > > > - return sdhci_pci_resume_host(chip); > > > > -} > > > > - > > > > -static int sdhci_cqhci_gli_resume(struct sdhci_pci_chip *chip) > > > > -{ > > > > - struct sdhci_pci_slot *slot = chip->slots[0]; > > > > - int ret; > > > > - > > > > - ret = sdhci_pci_gli_resume(chip); > > > > - if (ret) > > > > - return ret; > > > > - > > > > - return cqhci_resume(slot->host->mmc); > > > > -} > > > > - > > > > -static int sdhci_cqhci_gli_suspend(struct sdhci_pci_chip *chip) > > > > -{ > > > > - struct sdhci_pci_slot *slot = chip->slots[0]; > > > > - int ret; > > > > - > > > > - ret = cqhci_suspend(slot->host->mmc); > > > > - if (ret) > > > > - return ret; > > > > - > > > > - return sdhci_suspend_host(slot->host); > > > > -} > > > > -#endif > > > > - > > > > static void gl9763e_hs400_enhanced_strobe(struct mmc_host *mmc, > > > > struct mmc_ios *ios) > > > > { > > > > @@ -1029,6 +993,68 @@ static int gl9763e_runtime_resume(struct sdhci_pci_chip *chip) > > > > } > > > > #endif > > > > > > > > +#ifdef CONFIG_PM_SLEEP > > > > +static int sdhci_pci_gli_resume(struct sdhci_pci_chip *chip) > > > > +{ > > > > + struct sdhci_pci_slot *slot = chip->slots[0]; > > > > + > > > > + pci_free_irq_vectors(slot->chip->pdev); > > > > + gli_pcie_enable_msi(slot); > > > > + > > > > + return sdhci_pci_resume_host(chip); > > > > +} > > sdhci_pci_gli_resume() is the same as before. Is there any reason to move it here? To avoid having multiple #ifdef CONFIG_PM_SLEEP blocks. We can leave it, where it was if you prefer. > > > > > + > > > > +static int gl9763e_resume(struct sdhci_pci_chip *chip) > > > > +{ > > > > + struct sdhci_pci_slot *slot = chip->slots[0]; > > > > + int ret; > > > > + > > > > + ret = sdhci_pci_gli_resume(chip); > > > > + if (ret) > > > > + return ret; > > > > + > > > > + ret = cqhci_resume(slot->host->mmc); > > > > + if (ret) > > > > + return ret; > > > > + > > > > + /* Disable LPM negotiation to bring device back in sync > > > > + * with its runtime_pm state. > > > > + */ > > > > > > I would prefer the comment style: > > > > > > /* > > > * Blah, blah ... > > > * Blah, blah, blah. > > > */ > > > > > > > + gl9763e_set_low_power_negotiation(slot, false); > > There is a situation for your reference. > If `allow_runtime_pm' is set to false and the system resumes from suspend, GL9763E > LPM negotiation will be always disabled on S0. GL9763E will stay L0 and never > enter L1 because GL9763E LPM negotiation is disabled. > > This patch enables allow_runtime_pm. The simple flow is > gl9763e_suspend() -> LPM enabled -> gl9763e_resume() -> LPM disabled -> (a) > (a) -+--> idle --> gl9763e_runtime_suspend() -> LPM enabled > | > +--> no idle -> gl9763e_runtime_resume() -> LPM disabled Is the lower branch of this sequence possible? I mean please correct me if I'm wrong but after the s2idle resume, devices are considered runtime active unless deemed otherwise, so gl_9763e_resume() -> fl9763e_runtime_resume() should not happen, right? > > This patch disables allow_runtime_pm. The simple flow is > gl9763e_suspend() -> LPM enabled -> gl9763e_resume() -> LPM disabled (no runtime_pm) > > Although that may not be the case with the current configuration, it's only a > possibility. Actually last year there was a patch to improve R/W performance of GL9763E: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=f9e5b33934cec24b8c024add5c5d65d2f93ade05 If I'm reading the code right we're not modifying allow_runtime_pm, rather we're disabling the L1 negotiation in the same manner as the abovementioned patch does in the runtime PM flow. So in a way we're restoring whatever setting was there except a situation where the runtime PM was disabled from the start on the device and we do not restore the original state of the LPM negotiation fields. Not sure if such case is really possible. > > > > > + > > > > + return 0; > > > > +} > > > > + > > > > +static int gl9763e_suspend(struct sdhci_pci_chip *chip) > > > > +{ > > > > + struct sdhci_pci_slot *slot = chip->slots[0]; > > > > + int ret; > > > > + > > > > + /* Certain SoCs can suspend only with the bus in low- > > > > > > Ditto re comment style > > > > > > > + * power state, notably x86 SoCs when using S0ix. > > > > + * Re-enable LPM negotiation to allow entering L1 state > > > > + * and entering system suspend. > > > > + */ > > > > + gl9763e_set_low_power_negotiation(slot, true); > > > > > > Couldn't this be at the end of the function, save > > > an error path > > > > > > Please correct me if I'm wrong but writing to device config > > > space could trigger a side effect, so it's probably better to > > > do it before calling functions suspending the device? > > > > sdhci doesn't know anything about the bus. It is independent > > of PCI, so I can't see how it would make any difference. > > One of the people cc'ed might know more. Jason Lai (cc'ed) > > added it for runtime PM. > > > > As far as I know, when disabling LPM negotiation, the GL9763E will stop entering > L1. It doesn't other side effect. Does Jason.Lai and Victor.Shih have any comments > or suggestions? The reason we've put the LPM negotiation handling at the start of suspend and finish of resume was mostly for semantical reasons - first do device logic, then call base framework/module logic which might expect all the device-specific steps to be performed already. Maybe it does not make any difference in the real world for SDHCI controllers but it just seemed to look better. Also suspend and resume callbacks already have such ordering when it comes to cqhci_* and sdhci_* functions. > > Best regards, > Ben Chuang > > > > > > > > > > > + > > > > + ret = cqhci_suspend(slot->host->mmc); > > > > + if (ret) > > > > + goto err_suspend; > > > > + > > > > + ret = sdhci_suspend_host(slot->host); > > > > + if (ret) > > > > + goto err_suspend_host; > > > > + > > > > + return 0; > > > > + > > > > +err_suspend_host: > > > > + cqhci_resume(slot->host->mmc); > > > > +err_suspend: > > > > + gl9763e_set_low_power_negotiation(slot, false); > > > > + return ret; > > > > +} > > > > +#endif > > > > + > > > > static int gli_probe_slot_gl9763e(struct sdhci_pci_slot *slot) > > > > { > > > > struct pci_dev *pdev = slot->chip->pdev; > > > > @@ -1113,8 +1139,8 @@ const struct sdhci_pci_fixes sdhci_gl9763e = { > > > > .probe_slot = gli_probe_slot_gl9763e, > > > > .ops = &sdhci_gl9763e_ops, > > > > #ifdef CONFIG_PM_SLEEP > > > > - .resume = sdhci_cqhci_gli_resume, > > > > - .suspend = sdhci_cqhci_gli_suspend, > > > > + .resume = gl9763e_resume, > > > > + .suspend = gl9763e_suspend, > > > > #endif > > > > #ifdef CONFIG_PM > > > > .runtime_suspend = gl9763e_runtime_suspend, > > > > > > > > > > > > -- > > > Best Regards, > > > Stanisław Kardach > > > > -- Best Regards, Stanisław Kardach
