On Fri, Dec 13, 2024 at 03:35:15PM +0100, Rafael J. Wysocki wrote: > On Thu, Dec 12, 2024 at 4:14 PM Christoph Hellwig <hch@xxxxxx> wrote: > > > > On Thu, Dec 12, 2024 at 01:49:15PM +0100, Ulf Hansson wrote: > > > Right. This seems to somewhat work for ACPI types of systems, because > > > ACPI is controlling the low power state for all the devices. Based on > > > the requested system wide low power state, ACPI can then decide to > > > call pm_set_suspend_via_firmware() or not. > > > > > > Still there is a problem with this for ACPI too. > > > > > > How does ACPI know whether it's actually a good idea to keep the NVMe > > > storage powered in s2idle (ACPI calls pm_set_suspend_via_firmware() > > > only for S2R and S2disk!?)? Especially when my laptop only supports > > > s2idle and that's what I will use when I close the lid. In this way, > > > the NMVe storage will certainly contribute to draining the battery, > > > especially when I won't be using my laptop for a couple of days. > > > > > > In my opinion, we need a better approach that is both flexible and > > > that dynamically adjusts based upon the use case. > > > > Agreed. I'd be happy to work with the PM maintainers to do this, > > but I don't really know enough about the PM core to drive it > > (as the reply from Rafael to my mail makes pretty clear :)) > > I'm here to help. > > Let me know what exactly you want to achieve and we'll see how to make it work. I'll try to summarize the requirement here since I started this thread: Problem statement ================= We need a PM core API that tells the device drivers when it is safe to powerdown the devices. The usecase here is with PCIe based NVMe devices but the problem is applicable to other devices as well. Drivers are relying on couple of options now: 1. If pm_suspend_via_firmware() returns true, then drivers will shutdown the device assuming that the firmware is going to handle the suspend. But this API is currently used only by ACPI. Even there, ACPI relies on S2R being supported by the platform and it sets pm_set_suspend_via_firmware() only when the suspend is S2R. But if the platform doesn't support S2R (current case of most of the Qcom SoCs), then pm_suspend_via_firmware() will return false and NVMe won't be powered down draining the battery. If the platform is using DT, then there is no entity setting pm_set_suspend_via_firmware(). So NVMe will be kept in low power state all the time (still draining the battery). There were attempts to set this flag from PSCI [1], but there were objections on setting this flag when PSCI_SUSPEND is not supported by the platform (again, the case with Qcom SoCs). Even if this approach succeeds, then there are concerns that if the platform is used in an OS like Android where the S2Idle cycle is far more high, NVMe will wear out very quickly. So this is where the forthcoming API need to "dynamically adjusts based upon the use case" as quoted by Ulf in his previous reply. One way to achieve would be by giving the flexibility to the userspace to choose the suspend state (if platform has options to select). UFS does something similar with 'spm_lvl' [2] sysfs attribute that I believe Android userspace itself makes use of. 2. Making use of pm_suspend_target_state to differentiate between suspend states and powering down the devices only during PM_SUSPEND_MEM (S2R). But this also suffers from the same issue as mentioned above (when platform doesn't support S2R). TLDR: We need a PM core API that that sets a sane default suspend state for the platform and also allows dynamically changing/overriding the state (by taking inputs from userspace etc...). This API should also forbid override, if the platform has limitations (like if it requires powering down the devices all the time (x13s laptops)). Finally, this API would be used by the device drivers to decide when to safely power down the devices during suspend. @Ulf/others: Please chime in if you don't agree with anything I said above. - Mani [1] https://lore.kernel.org/all/20241028-topic-cpu_suspend_s2ram-v1-0-9fdd9a04b75c@xxxxxxxxxxxxxxxx [2] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/Documentation/ABI/testing/sysfs-driver-ufs#n1041 -- மணிவண்ணன் சதாசிவம்
