Hi Rafael, Could you please kindly comment on the above? Thank you in advance, Grzegorz śr., 20 lip 2022 o 15:15 Grzegorz Jaszczyk <jaz@xxxxxxxxxxxx> napisał(a): > > wt., 19 lip 2022 o 20:09 Rafael J. Wysocki <rafael@xxxxxxxxxx> napisał(a): > > > > On Thu, Jul 7, 2022 at 2:56 PM Grzegorz Jaszczyk <jaz@xxxxxxxxxxxx> wrote: > > > > > > Currently the LPS0 prepare_late callback is aimed to run as the very > > > last thing before entering the S2Idle state from LPS0 perspective, > > > nevertheless between this call and the system actually entering the > > > S2Idle state there are several places where the suspension process could > > > be canceled. > > > > And why is this a problem? > > > > The cancellation will occur only if there is a wakeup signal that > > would otherwise cause one of the CPUs to exit the idle state. Such a > > wakeup signal can appear after calling the new notifier as well, so > > why does it make a difference? > > It could also occur due to suspend_test. Additionally with new > notifier we could get notification when the system wakes up from > s2idle_loop and immediately goes to sleep again (due to e.g. > acpi_s2idle_wake condition not being met) - in this case relying on > prepare_late callback is not possible since it is not called in this > path. > > > > > > In order to notify VMM about guest entering suspend, extend the S2Idle > > > ops by new notify callback, which will be really invoked as a very last > > > thing before guest actually enters S2Idle state. > > > > It is not guaranteed that "suspend" (defined as all CPUs entering idle > > states) will be actually entered even after this "last step". > > Since this whole patchset is aimed at notifying the host about a guest > entering s2idle state, reaching this step can be considered as a > suspend "entry point" for VM IMO. It is because we are talking about > the vCPU not the real CPU. Therefore it seems to me, that even if some > other vCPUs could still get some wakeup signal they will not be able > to kick (through s2idle_wake->swake_up_one(&s2idle_wait_head);) the > original vCPU which entered s2idle_loop, triggered the new notifier > and is halted due to handling vCPU exit (and was about to trigger > swait_event_exclusive). So it will prevent the VM's resume process > from being started. > > > > > > Additionally extend the acpi_s2idle_dev_ops by notify() callback so > > > any driver can hook into it and allow to implement its own notification. > > > > > > Taking advantage of e.g. existing acpi_s2idle_dev_ops's prepare/restore > > > hooks is not an option since it will not allow to prevent race > > > conditions: > > > - VM0 enters s2idle > > > - host notes about VM0 is in s2idle > > > - host continues with system suspension but in the meantime VM0 exits > > > s2idle and sends notification but it is already too late (VM could not > > > even send notification on time). > > > > Too late for what? > > Too late to cancel the host suspend process, which thinks that the VM > is in s2idle state while it isn't. > > > > > > Introducing notify() as a very last step before the system enters S2Idle > > > together with an assumption that the VMM has control over guest > > > resumption allows preventing mentioned races. > > > > How does it do that? > > At the moment when VM triggers this new notifier we trap on MMIO > access and the VMM handles vCPU exit (so the vCPU is "halted"). > Therefore the VMM could control when it finishes such handling and > releases the vCPU again. > > Maybe adding some more context will be helpful. This patchset was > aimed for two different scenarios actually: > 1) Host is about to enter the suspend state and needs first to suspend > VM with all pass-through devices. In this case the host waits for > s2idle notification from the guest and when it receives it, it > continues with its own suspend process. > 2) Guest could be a "privileged" one (in terms of VMM) and when the > guest enters s2idle state it notifies the host, which in turn triggers > the suspend process of the host. > > > > > It looks like you want suspend-to-idle to behave like S3 and it won't. > > In a way, yes, we compensate for the lack of something like PM1_CNT to > trap on for detecting that the guest is suspending. > We could instead force the guest to use S3 but IMO it is undesirable, > since it generally does make a difference which suspend mode is used > in the guest, s2idle or S3, e.g some drivers check which suspend type > is used and based on that behaves differently during suspend. One of > the example is: > https://elixir.bootlin.com/linux/v5.18.12/source/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c#L2323 > https://elixir.bootlin.com/linux/v5.18.12/source/drivers/gpu/drm/amd/amdgpu/amdgpu_acpi.c#L1069 > https://elixir.bootlin.com/linux/v5.18.12/source/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c#L583 > > Thank you, > Grzegorz
