On Mon, 11 Oct 2021 at 07:23, Tony Lindgren <tony@xxxxxxxxxxx> wrote: > > * Ulf Hansson <ulf.hansson@xxxxxxxxxx> [211008 14:44]: > > On Thu, 30 Sept 2021 at 08:57, Tony Lindgren <tony@xxxxxxxxxxx> wrote: > > > > > > Implement PM runtime functions and enable MMC_CAP_AGGRESSIVE_PM. > > > > I suggest you split this change into two pieces. MMC_CAP_AGGRESSIVE_PM > > is about enabling runtime PM management for the eMMC/SD card device, > > which is perfectly fine to use independently of whether runtime PM is > > supported for the host device. > > OK > > > > @@ -1350,6 +1357,11 @@ static int sdhci_omap_probe(struct platform_device *pdev) > > > if (ret) > > > goto err_cleanup_host; > > > > > > + sdhci_omap_context_save(omap_host); > > > + omap_host->context_valid = 1; > > > > Looks like you can remove this flag, it's not being used. > > Hmm I think it is needed as otherwise we end up trying to restore > an invalid context on probe on the first pm_runtime_get(). Do you > have some nicer solution for that in mind? Right, I didn't notice that, my apologies. In any case, an option is to bring the device into full power, without calling pm_runtime_resume_and_get() from ->probe(). In principle, running the same operations as the ->runtime_resume() callback does, except for restoring the context then. When this is done, the following calls to runtime PM should do the trick (I extended it to support autosuspend as well): pm_runtime_get_noresume() pm_runtime_set_active() pm_runtime_set_autosuspend_delay() pm_runtime_use_autosuspend() pm_runtime_enable() Note that, this means that the omaps PM domain's ->runtime_resume() callback doesn't get invoked when powering on the device for the first time. Can this be a problem? > > > > + > > > + pm_runtime_put_sync(dev); > > > > I recommend to use the PM runtime autosuspend feature, as to avoid an > > initial latency for every I/O request to the host driver. The mmc core > > already supports that, see mmc_release_host(). > > > > The typical default timeout value for autosuspend, is usually set > > ~50-200ms, by host drivers (if I recall correctly). > > OK I have a patch to also enable autosuspend too, I'll add that > too for the next revision. > > > > @@ -1371,6 +1383,7 @@ static int sdhci_omap_remove(struct platform_device *pdev) > > > struct device *dev = &pdev->dev; > > > struct sdhci_host *host = platform_get_drvdata(pdev); > > > > > > + pm_runtime_get_sync(dev); > > > sdhci_remove_host(host, true); > > > pm_runtime_put_sync(dev); > > > > There is no guarantee that this triggers a call to > > ->sdhci_omap_runtime_suspend(), which I guess is what we want. > > Userspace via sysfs may have increase the RPM usage count > > (pm_runtime_forbid(), for example. > > > > To address this, I would call pm_runtime_disable() first and then > > explicitly put the device into low power state, rather than relying on > > runtime PM to do it. Another option could be to use > > pm_runtime_force_suspend(). > > OK I'll take a look. > > > > @@ -1402,42 +1415,75 @@ static void sdhci_omap_context_restore(struct sdhci_omap_host *omap_host) > > > sdhci_omap_writel(omap_host, SDHCI_OMAP_ISE, omap_host->ise); > > > } > > > > > > -static int __maybe_unused sdhci_omap_suspend(struct device *dev) > > > +static int __maybe_unused sdhci_omap_runtime_suspend(struct device *dev) > > > { > > > struct sdhci_host *host = dev_get_drvdata(dev); > > > struct sdhci_pltfm_host *pltfm_host = sdhci_priv(host); > > > struct sdhci_omap_host *omap_host = sdhci_pltfm_priv(pltfm_host); > > > > > > - sdhci_suspend_host(host); > > > - > > > > Shouldn't you call sdhci_runtime_suspend_host() somewhere here? > > I'm pretty sure I tried, but runtime resume did not seem to work after > doing that.. I'll take a look again. > > > > +static int __maybe_unused sdhci_omap_suspend(struct device *dev) > > > +{ > > > + struct sdhci_host *host = dev_get_drvdata(dev); > > > + struct sdhci_pltfm_host *pltfm_host = sdhci_priv(host); > > > + struct sdhci_omap_host *omap_host = sdhci_pltfm_priv(pltfm_host); > > > + > > > + if (omap_host->is_runtime_suspended) > > > + return 0; > > > > So if the host is already runtime suspended, it's okay to just leave it as is? > > Ideally yeah there should not be anything left to do for suspesnd at > that point. But sounds like I may be missing something. > > > In a way that sounds like you could call pm_runtime_force_suspend() > > instead, assuming the sdhci_omap_runtime_suspend() can be extended to > > do the right thing for system suspend as well. > > OK I'll check. > > > It looks a bit odd that sdhci_suspend_host() is called only when the > > host is runtime resumed. Perhaps you can elaborate a bit more on why > > this is, so I can understand better what you want to achieve here. > > I guess I'm not clear on what's left for sdhci_suspend_host() to do if > the host is already runtime suspended :) I think what boils down to that is that, sdhci_suspend|resume_host() adds some special treatment for system wakeups (for SDIO irqs). I am not sure whether you may need that. Some host drivers doesn't use sdhci_suspend|resume_host, but sticks to the sdhci_runtime_suspend|resume()_host() functions. Like sdhci-sprd.c, for example. Kind regards Uffe