[Intel-gfx] BUG: sleeping function called from invalid context on 3.10.10-rt7

[Mario Kleiner]

On 09/17/2013 10:55 PM, Daniel Vetter wrote:
> On Tue, Sep 17, 2013 at 9:50 PM, Peter Hurley <peter at hurleysoftware.com> wrote:
>> On 09/11/2013 03:31 PM, Peter Hurley wrote:
>>>
>>> [+cc dri-devel]
>>>
>>> On 09/11/2013 11:38 AM, Steven Rostedt wrote:
>>>>
>>>> On Wed, 11 Sep 2013 11:16:43 -0400
>>>> Peter Hurley <peter at hurleysoftware.com> wrote:
>>>>
>>>>>> The funny part is, there's a comment there that shows that this was
>>>>>> done even for "PREEMPT_RT". Unfortunately, the call to
>>>>>> "get_scanout_position()" can call functions that use the rt-mutex
>>>>>> "sleeping spin locks" and it breaks there.
>>>>>>
>>>>>> I guess we need to ask the authors of the mainline patch exactly why
>>>>>> that preempt_disable() is needed?
>>>>>
>>>>>
>>>>> The drm core associates a timestamp with each vertical blank frame #.
>>>>> Drm drivers can optionally support a 'high resolution' hw timestamp.
>>>>> The vblank frame #/timestamp tuple is user-space visible.
>>>>>
>>>>> The i915 drm driver supports a hw timestamp via this drm helper function
>>>>> which computes the timestamp from the crtc scan position (based on the
>>>>> pixel clock).
>>>>>
>>>>> For mainline, the preempt_disable/_enable() isn't actually necessary
>>>>> because every call tree that leads here already has preemption disabled.
>>>>>
>>>>> For -RT, the maybe i915 register spinlock (uncore.lock) should be raw?
>>>>>
>>>>
>>>> No, it should not. Note, any other lock that can be held when it is
>>>> held would also need to be raw.
>>>
>>>
>>> By that, you mean "any other lock" that might be claimed "would also need
>>> to be raw"?  Hopefully not "any other lock" already held?
>>>
>>>> And by taking a quick audit of the code, I see this:
>>>>
>>>>      spin_lock_irqsave(&dev_priv->uncore.lock, irqflags);
>>>>
>>>>      /* Reset the chip */
>>>>
>>>>      /* GEN6_GDRST is not in the gt power well, no need to check
>>>>       * for fifo space for the write or forcewake the chip for
>>>>       * the read
>>>>       */
>>>>      __raw_i915_write32(dev_priv, GEN6_GDRST, GEN6_GRDOM_FULL);
>>>>
>>>>      /* Spin waiting for the device to ack the reset request */
>>>>      ret = wait_for((__raw_i915_read32(dev_priv, GEN6_GDRST) &
>>>> GEN6_GRDOM_FULL) == 0, 500);
>>>>
>>>> That spin is unacceptable in RT with preemption and interrupts disabled.
>>>
>>>
>>> Yep. That would be bad.
>>>
>>> AFAICT the registers read in i915_get_crtc_scanoutpos() aren't included
>>> in the force-wake set, so raw reads of the registers would
>>> probably be acceptable (thus obviating the need for claiming the
>>> uncore.lock).
>>>
>>> Except that _ALL_ register access is disabled with the uncore.lock
>>> during a gpu reset. Not sure if that's meant to include crtc registers
>>> or not, or what other synchronization/serialization issues are being
>>> handled/hidden by forcing all register accesses to wait during a gpu
>>> reset.
>>>
>>> Hopefully an i915 expert can weigh in here?
>>
>>
>>
>> Daniel,
>>
>> Can you shed some light on whether the i915+ crtc registers (specifically
>> those in i915_get_crtc_scanoutpos() and i915_/gm45_get_vblank_counter())
>> read as part of the vblank counter/timestamp handling need to
>> be prevented during gpu reset?
>
> The depency here in the locking is a recent addition:
>
> commit a7cd1b8fea2f341b626b255d9898a5ca5fabbf0a
> Author: Chris Wilson <chris at chris-wilson.co.uk>
> Date:   Fri Jul 19 20:36:51 2013 +0100
>
>      drm/i915: Serialize almost all register access
>
> It's a (slightly) oversized hammer to work around a hardware issue -
> we could break it down to register blocks, which can be accessed
> concurrently, but that tends to be more fragile. But the chip really
> dies if you access (even just reads) the same block concurrently :(
>
> We could try break the spinlock protected section a bit in the reset
> handler - register access on a hung gpu tends to be ill-defined
> anyway.
>
>> The implied wait with preemption and interrupts disabled is causing grief
>> in -RT, but also a 4ms wait inside an irq handler seems like a bad idea.
>
> Oops, the magic code in wait_for which is just there to make the imo
> totally misguided kgdb support work papered over the aweful long wait
> in atomic context ever since we've added this in
>
> commit b6e45f866465f42b53d803b0c574da0fc508a0e9
> Author: Keith Packard <keithp at keithp.com>
> Date:   Fri Jan 6 11:34:04 2012 -0800
>
>      drm/i915: Move reset forcewake processing to gen6_do_reset
>
> Reverting this change should be enough (code moved obviously a bit).
>
> Cheers, Daniel
>
>>
>> Regards,
>> Peter Hurley
>>
>>
>>
>>>> What's the real issue here?
>>>
>>>
>>> That the vblank timestamp needs to be an accurate measurement of a
>>> realtime event. Sleeping/servicing interrupts while reading
>>> the registers necessary to compute the timestamp would be bad too.
>>>
>>> (edit: which hopefully Mario Kleiner clarified in his reply)
>>>
>>> My point earlier was three-fold:
>>> 1. Don't need the preempt_disable() for mainline: all callers are already
>>>      holding interrupt-disabling spinlocks.
>>> 2. -RT still needs to prevent scheduling there.
>>> 3. the problem is i915-specific.
>>>
>>> [update: the radeon driver should also BUG like the i915 driver but
>>> probably
>>> should have mmio_idx_lock spinlock as raw]
>>

Ok, so register access must be serialized, at least within a register 
block, no way around it. Thanks for explaining this. That makes me a bit 
depressed. Is this also true for older hw generations than gen6?

Daniel, could we move access to the display register block to a separate 
lock, so the I915_READ to PIPEDSL in  i915_get_crtc_scanoutpos() can 
avoid the uncore.lock? If the display registers don't need forcewake 
then i assume we could have much shorter lock hold times by avoiding the 
large up to 4 msecs waits in forcewake handling. Probably also much less 
contention in normal operation when no modesetting happens? I think i 
can get rid of the other register reads in that function. Those settings 
are already cached and accessible from the intel_crtc_config->adjusted_mode.

Steven, would it then be acceptable to convert that "faster" lock into a 
raw_spinlock_t or is this unacceptable? If so, the preempt_disable() 
could stay, right?

I really would like to keep those preempt_disable()/enable() calls where 
they are if somehow possible - it makes usage of the timestamping under 
any kernel nicely plug & play.

If that's not possible i think i could work around it for at least my 
application under rt kernel, but it would be somewhat ugly and high 
maintenance, and wouldn't help other apps with similar needs, which 
don't know what knobs to tune - or that any tuning is suddenly needed on 
the latest kernels.

In my case i could set the drm.vblankoffdelay module parameter to zero 
to keep vblank irq's running all the time. That would leave the irq 
kernel thread for the gpu interrupt as the only caller of the scanout 
timestamping. Then one could set the priority of that irq thread to 
something very high and hope for the best. How to locate the proper irq 
thread(s) for any given gpu(s), what priority(s) to choose etc. and how 
to do this portably across various versions of various distros, i don't 
know. Doesn't look very user friendly, compared to the current solution, 
but not impossible.

In any case, maybe the simple retry 3x loop in the DRM for measuring the 
timestamp is not good enough anymore to keep this reliable and accurate. 
Maybe we would need some loop that retries until an accurate measurement 
or a user configurable timeout is reached. Then users like  mine could 
set a very high timeout which rather totally degrades the system and 
causes severe stuttering of graphics updates than silently failing with 
inaccurate timestamps. The default could be something safe for normal 
desktop use.

Can i ask somebody with the hardware to give me an idea how big the 
error there is? At a high drm.debug setting, you should see this logged 
at each vblank interrupt, as seen in this DRM_DEBUG statement:

http://lxr.free-electrons.com/source/drivers/gpu/drm/drm_irq.c#L717

e.g., ....[e 4 us, 1 rep] ... for 1 repetitions with the lowest error 
being 4 usecs.

Especially while rendering a lot, so the uncore.lock gets something to do.

Thanks,
-mario