Re: [PATCH v4 2/6] vfio: Add a new device feature for the power management

[Alex Williamson <alex.williamson@redhat.com>]

On Mon, 11 Jul 2022 15:13:13 +0530
Abhishek Sahu <abhsahu@nvidia.com> wrote:

> On 7/8/2022 10:06 PM, Alex Williamson wrote:
> > On Fri, 8 Jul 2022 15:09:22 +0530
> > Abhishek Sahu <abhsahu@nvidia.com> wrote:
> >   
> >> On 7/6/2022 9:09 PM, Alex Williamson wrote:  
> >>> On Fri, 1 Jul 2022 16:38:10 +0530
> >>> Abhishek Sahu <abhsahu@nvidia.com> wrote:
> >>>     
> >>>> This patch adds the new feature VFIO_DEVICE_FEATURE_POWER_MANAGEMENT
> >>>> for the power management in the header file. The implementation for the
> >>>> same will be added in the subsequent patches.
> >>>>
> >>>> With the standard registers, all power states cannot be achieved. The
> >>>> platform-based power management needs to be involved to go into the
> >>>> lowest power state. For all the platform-based power management, this
> >>>> device feature can be used.
> >>>>
> >>>> This device feature uses flags to specify the different operations. In
> >>>> the future, if any more power management functionality is needed then
> >>>> a new flag can be added to it. It supports both GET and SET operations.
> >>>>
> >>>> Signed-off-by: Abhishek Sahu <abhsahu@nvidia.com>
> >>>> ---
> >>>>  include/uapi/linux/vfio.h | 55 +++++++++++++++++++++++++++++++++++++++
> >>>>  1 file changed, 55 insertions(+)
> >>>>
> >>>> diff --git a/include/uapi/linux/vfio.h b/include/uapi/linux/vfio.h
> >>>> index 733a1cddde30..7e00de5c21ea 100644
> >>>> --- a/include/uapi/linux/vfio.h
> >>>> +++ b/include/uapi/linux/vfio.h
> >>>> @@ -986,6 +986,61 @@ enum vfio_device_mig_state {
> >>>>  	VFIO_DEVICE_STATE_RUNNING_P2P = 5,
> >>>>  };
> >>>>  
> >>>> +/*
> >>>> + * Perform power management-related operations for the VFIO device.
> >>>> + *
> >>>> + * The low power feature uses platform-based power management to move the
> >>>> + * device into the low power state.  This low power state is device-specific.
> >>>> + *
> >>>> + * This device feature uses flags to specify the different operations.
> >>>> + * It supports both the GET and SET operations.
> >>>> + *
> >>>> + * - VFIO_PM_LOW_POWER_ENTER flag moves the VFIO device into the low power
> >>>> + *   state with platform-based power management.  This low power state will be
> >>>> + *   internal to the VFIO driver and the user will not come to know which power
> >>>> + *   state is chosen.  Once the user has moved the VFIO device into the low
> >>>> + *   power state, then the user should not do any device access without moving
> >>>> + *   the device out of the low power state.    
> >>>
> >>> Except we're wrapping device accesses to make this possible.  This
> >>> should probably describe how any discrete access will wake the device
> >>> but ongoing access through mmaps will generate user faults.
> >>>     
> >>
> >>  Sure. I will add that details also.
> >>  
> >>>> + *
> >>>> + * - VFIO_PM_LOW_POWER_EXIT flag moves the VFIO device out of the low power
> >>>> + *    state.  This flag should only be set if the user has previously put the
> >>>> + *    device into low power state with the VFIO_PM_LOW_POWER_ENTER flag.    
> >>>
> >>> Indenting.
> >>>     
> >>  
> >>  I will fix this.
> >>  
> >>>> + *
> >>>> + * - VFIO_PM_LOW_POWER_ENTER and VFIO_PM_LOW_POWER_EXIT are mutually exclusive.
> >>>> + *
> >>>> + * - VFIO_PM_LOW_POWER_REENTERY_DISABLE flag is only valid with
> >>>> + *   VFIO_PM_LOW_POWER_ENTER.  If there is any access for the VFIO device on
> >>>> + *   the host side, then the device will be moved out of the low power state
> >>>> + *   without the user's guest driver involvement.  Some devices require the
> >>>> + *   user's guest driver involvement for each low-power entry.  If this flag is
> >>>> + *   set, then the re-entry to the low power state will be disabled, and the
> >>>> + *   host kernel will not move the device again into the low power state.
> >>>> + *   The VFIO driver internally maintains a list of devices for which low
> >>>> + *   power re-entry is disabled by default and for those devices, the
> >>>> + *   re-entry will be disabled even if the user has not set this flag
> >>>> + *   explicitly.    
> >>>
> >>> Wrong polarity.  The kernel should not maintain the policy.  By default
> >>> every wakeup, whether from host kernel accesses or via user accesses
> >>> that do a pm-get should signal a wakeup to userspace.  Userspace needs
> >>> to opt-out of that wakeup to let the kernel automatically re-enter low
> >>> power and userspace needs to maintain the policy for which devices it
> >>> wants that to occur.
> >>>     
> >>  
> >>  Okay. So that means, in the kernel side, we don’t have to maintain
> >>  the list which currently contains NVIDIA device ID. Also, in our
> >>  updated approach, this opt-out of that wake-up means that user
> >>  has not provided eventfd in the feature SET ioctl. Correct ?  
> > 
> > Yes, I'm imagining that if the user hasn't provided a one-shot wake-up
> > eventfd, that's the opt-out for being notified of device wakes.  For
> > example, pm-resume would have something like:
> > 
> > 	
> > 	if (vdev->pm_wake_eventfd) {
> > 		eventfd_signal(vdev->pm_wake_eventfd, 1);
> > 		vdev->pm_wake_eventfd = NULL;
> > 		pm_runtime_get_noresume(dev);
> > 	}
> > 
> > (eventfd pseudo handling substantially simplified)
> > 
> > So w/o a wake-up eventfd, the user would need to call the pm feature
> > exit ioctl to elevate the pm reference to prevent it going back to low
> > power.  The pm feature exit ioctl would be optional if a wake eventfd is
> > provided, so some piece of the eventfd context would need to remain to
> > determine whether a pm-get is necessary.
> >   
> >>>> + *
> >>>> + * For the IOCTL call with VFIO_DEVICE_FEATURE_GET:
> >>>> + *
> >>>> + * - VFIO_PM_LOW_POWER_ENTER will be set if the user has put the device into
> >>>> + *   the low power state, otherwise, VFIO_PM_LOW_POWER_EXIT will be set.
> >>>> + *
> >>>> + * - If the device is in a normal power state currently, then
> >>>> + *   VFIO_PM_LOW_POWER_REENTERY_DISABLE will be set for the devices where low
> >>>> + *   power re-entry is disabled by default.  If the device is in the low power
> >>>> + *   state currently, then VFIO_PM_LOW_POWER_REENTERY_DISABLE will be set
> >>>> + *   according to the current transition.    
> >>>
> >>> Very confusing semantics.
> >>>
> >>> What if the feature SET ioctl took an eventfd and that eventfd was one
> >>> time use.  Calling the ioctl would setup the eventfd to notify the user
> >>> on wakeup and call pm-put.  Any access to the device via host, ioctl,
> >>> or region would be wrapped in pm-get/put and the pm-resume handler
> >>> would perform the matching pm-get to balance the feature SET and signal
> >>> the eventfd.     
> >>
> >>  This seems a better option. It will help in making the ioctl simpler
> >>  and we don’t have to add a separate index for PME which I added in
> >>  patch 6. 
> >>  
> >>> If the user opts-out by not providing a wakeup eventfd,
> >>> then the pm-resume handler does not perform a pm-get. Possibly we
> >>> could even allow mmap access if a wake-up eventfd is provided.    
> >>
> >>  Sorry. I am not clear on this mmap part. We currently invalidates
> >>  mapping before going into runtime-suspend. Now, if use tries do
> >>  mmap then do we need some extra handling in the fault handler ?
> >>  Need your help in understanding this part.  
> > 
> > The option that I'm thinking about is if the mmap fault handler is
> > wrapped in a pm-get/put then we could actually populate the mmap.  In
> > the case where the pm-get triggers the wake-eventfd in pm-resume, the
> > device doesn't return to low power when the mmap fault handler calls
> > pm-put.  This possibly allows that we could actually invalidate mmaps on
> > pm-suspend rather than in the pm feature enter ioctl, essentially the
> > same as we're doing for intx.  I wonder though if this allows the
> > possibility that we just bounce between mmap fault and pm-suspend.  So
> > long as some work can be done, for instance the pm-suspend occurs
> > asynchronously to the pm-put, this might be ok.
> >   
> 
>  We can do this. But in the normal use case, the situation should
>  never arise where user should access any mmaped region when user has
>  already put the device into D3 (D3hot or D3cold). This can only happen
>  if there is some bug in the guest driver or user is doing wrong
>  sequence. Do we need to add handling to officially support this part ?

We cannot rely on userspace drivers to be bug free or non-malicious,
but if we want to impose that an mmap access while low power is
enabled always triggers a fault, that's ok.

>  pm-get can take more than a second for resume for some devices and
>  will doing this in fault handler be safe ?
> 
>  Also, we will add this support only when wake-eventfd is provided so
>  still w/o wake-eventfd case, the mmap access will still generate fault.
>  So, we will have different behavior. Will that be acceptable ?

Let's keep it simple, generate a fault for all cases.

> >>> The
> >>> feature GET ioctl would be used to exit low power behavior and would be
> >>> a no-op if the wakeup eventfd had already been signaled.  Thanks,
> >>>    
> >>  
> >>  I will use the GET ioctl for low power exit instead of returning the
> >>  current status.  
> > 
> > Note that Yishai is proposing a device DMA dirty logging feature where
> > the stop and start are exposed via SET on separate features, rather
> > than SET/GET.  We should probably maintain some consistency between
> > these use cases.  Possibly we might even want two separate pm enter
> > ioctls, one with the wake eventfd and one without.  I think this is the
> > sort of thing Jason is describing for future expansion of the dirty
> > tracking uAPI.  Thanks,
> > 
> > Alex
> >   
> 
>  Okay. So, we need to add 3 device features in total.
> 
>  VFIO_DEVICE_FEATURE_PM_ENTRY
>  VFIO_DEVICE_FEATURE_PM_ENTRY_WITH_WAKEUP
>  VFIO_DEVICE_FEATURE_PM_EXIT
> 
>  And only the second one need structure which will have only one field
>  for eventfd and we need to return error if wakeup-eventfd is not
>  provided in the second feature ?

Yes, we'd use eventfd_ctx and fail on a bad fileget.

>  Do we need to support GET operation also for these ?
>  We can skip GET operation since that won’t be very useful.

What would they do?  Thanks,

Alex