Re: lockdep warning in urb.c:363 usb_submit_urb

[Rafael J. Wysocki]

On Friday, April 3, 2020 6:41:05 PM CEST Alan Stern wrote:
> On Fri, 3 Apr 2020, Rafael J. Wysocki wrote:
> 
> > I'll address the other points in your message separately.
> > 
> > The rules for SMART_SUSPEND are as follows:
> 
> These rules are a sort of high-level overview.  They don't (for the
> most part) say exactly what will happen, in terms of which callbacks
> will be issued and under what circumstances.  I have tried to provide 
> such a description inline below.
> 
> > (a) If SMART_SUSPEND is set and the device is runtime-suspended during system
> >     suspend, it is not expected to be resumed by the core or the middle layer
> >     (subsystem) code unless the latter has a specific reason to do that (e.g.
> >     it knows that the device needs to be reconfigured which cannot be done
> >     without resuming it).
> > 
> >     The device can still be resumed when it is needed to suspend a dependent
> >     device, but that cannot happen before the "late suspend" phase.
> 
> Don't you mean it cannot happen during or after the "late suspend"  
> phase?

Yes, I do.

> (More precisely, it cannot happen after the time when the core
> would issue the device's ->suspend_late callback, but it can happen
> between that time and the time when the "late suspend" phase began --
> for example, from within the dependent device's ->suspend_late
> callback.)  After all, __device_suspend_late() calls
> __pm_runtime_disable(), following which the device _cannot_ be runtime
> resumed.

Right.

> Putting this in operational terms, it seems to say that if
> SMART_SUSPEND is set and the device is runtime-suspended at the times
> when the core would normally issue the driver's ->suspend_late or
> ->suspend_noirq callback, then the core will skip the callback (with a
> similar requirement for subsystems).  Correct?

Yes.

> > (b) Drivers that set SMART_SUSPEND are allowed to reuse their PM-runtime
> >     callbacks for system-wide suspend and resume.
> > 
> >     That is, they can point either the ->suspend_late or the ->suspend_noirq
> >     callback pointer to the same function as ->runtime_suspend and they can
> >     point either the ->resume_noirq or ->the resume_early callback to the'
> >     same function as ->runtime_resume.
> 
> Well, in theory any driver or subsystem can do this whenever it wants
> to, regardless of any flag settings.

Not exactly.

Say the driver wants to point both ->runtime_suspend and ->suspend_late to
the same function.

If the bus type doesn't provide system-wide PM callbacks at all (which is
the case for some bus types), that only works if the device is never
runtime-suspended when ->suspend_late is about to run, because otherwise
the function in question needs to check the context in which it is running
(PM-runtime vs system-wide and runtime-suspended vs runtime-active in the
latter case) which at least is awkward and hard to get right.

> It's then up to the driver or
> subsystem to make sure the callback "does the right thing".

In theory.

> What I'm concerned about now is: What guarantees can the core give to 
> the driver and subsystem, so that they will know what is necessary in 
> order to "do the right thing"?

I'm not sure what you mean.

If the subsystem provides callbacks, the core will run them regardless.

If it does not provide callbacks, the core will skip ->suspend_late and
->suspend_noirq for the driver and the device will remain suspended.

If SMART_SUSPEND is not set, the core will execute all of the callbacks
that are present.

> > (c) Drivers that set SMART_SUSPEND are alwo allowed to provide special
> >     simplified callbacks for the "freeze" and "thaw" transitions during
> >     hibernation (and restore) and (if they do so) special callbacks for the
> >     "restore" phase.
> 
> What do you mean by "simplified"?

Avoiding actual PM.

> As I see it, the suspend-side callbacks are generally responsible for 
> four things:
> 
> 	1. Quiesce the device (finish ongoing I/O and do not allow any
> 	   more to start).
> 
> 	2. Save the current device state.
> 
> 	3. Install the appropriate wakeup settings.
> 
> 	4. Put the device into low-power mode.
> 
> (Not explicitly listed: Perform a runtime-resume if needed in order to
> carry out these four items.)
> 
> During a SUSPEND transition, we usually expect all four to happen.  
> During a FREEZE transition, we only require 1.

That's what I mean by "simplified".

> During a POWEROFF
> transition we require 1 and 3, and possibly 4 (depending on how the
> platform handles poweroff).

But doing 2 is not a bug AFAICS.

> Similar requirements apply to the resume-side callbacks.  (But note 
> that RESTORE is not the inverse of POWEROFF; it is more like an inverse 
> of FREEZE with the added complication that the device's initial state 
> is unknown.)

It actually isn't even an inverse of FREEZE.  It is like RESUME with the
additional requirements that (a) it can never be skipped and (b) the
device need not be in a low-power state when it runs (the initial state
of it is unknown if you will).

> What changes to this analysis would SMART_SUSPEND allow?  None if the 
> device is runtime-active.  But if the device is runtime-suspended and 
> the wakeup settings don't need to be changed, then presumably none of 
> the four items are necessary.
> 
> Is this what you mean?

No.

What I meant was that even if the driver pointed ->runtime_suspend and
->suspend_late (say) to the same function and it pointed ->resume_early
and ->runtime_resume to the same function, it didn't have to point
->freeze_late and ->thaw_early to the same pair of functions, respectively.

It can point ->freeze_late and ->thaw_early to a pair of different functions
that only quiesce the device and reverse that, respectively.

> > [OK, I realize that (b) and (c) are not documented, see the notes below.]
> > 
> > Because of (a), if the device with SMART_SUSPEND set is still runtime-suspended
> > during the "late" phase of suspend, the core will not invoke the driver's
> > "late" and "noirq" suspend callbacks directly (*).  Middle layer (subsystem)
> > code is expected to behave accordingly.
> 
> Okay, this agrees with what I wrote above.
> 
> > Because of (b), if the "late" and "noirq" driver callbacks were skipped during
> > the "freeze" transition, the core will also avoid invoking the "noirq" and
> > "early" callbacks provided by the driver during the "thaw" transition and
> > the callbacks during the "restore" transition will be executed unconditionally
> > (**).  Middle layer code is expected to behave accordingly.
> 
> All right.  To summarize: If the driver's ->freeze_late callback is
> skipped then the driver's ->thaw-early will be skipped, and similarly
> for ->freeze_noirq and ->thaw_noirq.  But RESTORE callbacks are never
> skipped.  Correct?

Yes.

> However, the most difficult transitions are SUSPEND and RESUME.  Is it
> accurate to say that if the driver's ->suspend_late callback is skipped
> then the driver's ->resume_early will be skipped, and similarly for
> ->suspend_noirq and ->resume_noirq?

If LEAVE_SUSPENDED is set in addition to SMART_SUSPEND, then yes.

> > Notes:
> > 
> > 1. I have considered splitting SMART_SUSPEND into two or even three flags
> >    so that (a), (b) and (c) are each associated with a separate flag, but
> >    then I would expect the majority of users to use all of them anyway.
> > 
> > 2. LEAVE_SUSPENDED (which may be better renamed to SKIP_RESUME) is kind of
> >    expected to be used along with SMART_SUSPEND unless there is a good enough
> >    reason to avoid using it.  I admit that this isn't really straightforward,
> >    maybe the default behavior should be to skip the resume and there should be
> >    FORCE_RESUME instead of LEAVE_SUSPENDED.
> 
> One question not addressed above (in fact, the original reason for 
> getting you involved in this discussion): What about the device's 
> power.runtime_status?  Shall we say that that core will call 
> pm_runtime_set_active() at some point before issuing the ->complete 
> callback unless some combination of flags is set?  And what should that 
> combination be?
> 
> After all, we expect that most drivers will want their devices to be in 
> the runtime-active state at the end of a system sleep or hibernation.  
> It makes sense for the core to do the necessary housekeeping.

The core will set the PM-runtime status to "active" in device_resume_noirq()
if (a) the subsystem callbacks are not invoked (otherwise the subsystem is
responsible for doing that) and (b) if the driver's callback not skipped
(in which case its ->resume_early callback will not be skipped too).

> > 3. (*) Under the assumption that either ->suspend_late or ->suspend_noirq
> >    points to the same routine as ->runtime_suspend (and the other is NULL),
> >    invokig that callback for a runtime-suspended device is technically invalid.
> 
> Does this invalidate anything I wrote above?

I don't think so.  It is the reason why driver callbacks are skipped for
runtime-suspended devices.

> >    In turn, under the assumption that either ->resume_early or ->resume_noirq
> >    points to the same routine as ->runtime_resume (and the other is NULL), it is
> >    valid to invoke that callback if the late/noirq suspend was skipped.
> 
> In other words, it's okay for the core either to issue or skip those 
> callbacks.  Presumably the decision will be made based on some flag 
> setting?

Yes.  A flag combined with the PM-runtime status of the device in
device_suspend_noirq().

> > 4. (**) If the "freeze" and "thaw" callbacks are simplified, they cannot be
> >    run back-to-back with ->runtime_resume and ->runtime_suspend, respectively.
> >    Thus if "freeze" is skippend, "thaw" must be skipped too.  However,
> >    "restore" needs to be prepared to be invoked after "freeze" or
> >    ->runtime_suspend (and the state of the device may not match the
> >    callback that ran previously), so it must be special.
> > 
> > 5. I agree that skipping the driver level of callbacks depending on what is
> >    provided by the middle layer is inconsistent, but I wanted to take the
> >    users of pm_runtime_force_suspend/resume() into account by letting those
> >    things run.
> > 
> >    It would be more consistent to expect middle layer code (bus types, PM
> >    domains) to provide either all of the noirq/early/late callbacks, or none
> >    of them and make SMART_SUSPEND and pm_runtime_force_suspend/resume()
> >    mutually exclusive.
> 
> I don't have a clear idea of how pm_runtime_force_suspend/resume() gets 
> used.  Are we better off ignoring it for the time being?

Yes, we are.