Nouveau - Apr 2019 - [PATCH] pci/quirks: Add quirk to reset nvgpu at boot for the Lenovo ThinkPad P50

On Mon, Apr 15, 2019 at 02:07:18PM -0400, Lyude Paul
wrote:
> On Thu, 2019-04-04 at 09:17 -0500, Bjorn Helgaas wrote:
> > [+cc Hans, author of 0b2fe6594fa2 ("drm/nouveau: Queue hpd_work
on (runtime)
> > resume")]
> > 
> > On Fri, Mar 22, 2019 at 06:30:15AM -0500, Bjorn Helgaas wrote:
> > > On Thu, Mar 21, 2019 at 05:48:19PM -0500, Bjorn Helgaas wrote:
> > > > On Wed, Mar 13, 2019 at 06:25:02PM -0400, Lyude Paul wrote:
> > > > > On Fri, 2019-02-15 at 16:17 -0500, Lyude Paul wrote:
> > > > > > On Thu, 2019-02-14 at 18:43 -0600, Bjorn Helgaas
wrote:
> > > > > > > On Tue, Feb 12, 2019 at 05:02:30PM -0500,
Lyude Paul wrote:
> > > > > > > > On a very specific subset of ThinkPad
P50 SKUs, particularly
> > > > > > > > ones that come with a Quadro M1000M chip
instead of the M2000M
> > > > > > > > variant, the BIOS seems to have a very
nasty habit of not
> > > > > > > > always resetting the secondary Nvidia
GPU between full reboots
> > > > > > > > if the laptop is configured in Hybrid
Graphics mode. The
> > > > > > > > reason for this happening is unknown,
but the following steps
> > > > > > > > and possibly a good bit of patience will
reproduce the issue:
> > > > > > > > 
> > > > > > > > 1. Boot up the laptop normally in Hybrid
graphics mode
> > > > > > > > 2. Make sure nouveau is loaded and that
the GPU is awake
> > > > > > > > 2. Allow the nvidia GPU to runtime
suspend itself after being
> > > > > > > > idle
> > > > > > > > 3. Reboot the machine, the more sudden
the better (e.g sysrq-b
> > > > > > > > may help)
> > > > > > > > 4. If nouveau loads up properly, reboot
the machine again and go
> > > > > > > > back to
> > > > > > > > step 2 until you reproduce the issue
> > > > > > > > 
> > > > > > > > This results in some very strange
behavior: the GPU will quite
> > > > > > > > literally be left in exactly the same
state it was in when the
> > > > > > > > previously booted kernel started the
reboot. This has all
> > > > > > > > sorts of bad sideaffects: for starters,
this completely breaks
> > > > > > > > nouveau starting with a mysterious EVO
channel failure that
> > > > > > > > happens well before we've actually
used the EVO channel for
> > > > > > > > anything:
> > > > 
> > > > Thanks for the hybrid tutorial (snipped from this response).
IIUC,
> > > > what you said was that in hybrid mode, the Intel GPU drives
the
> > > > built-in display and the Nvidia GPU drives any external
displays and
> > > > may be used for DRI PRIME rendering (whatever that is).  But
since you
> > > > say the Nvidia device gets runtime suspended, I assume
there's no
> > > > external display here and you're not using DRI PRIME.
> > > > 
> > > > I wonder if it's related to the fact that the Nvidia GPU
has been
> > > > runtime suspended before you do the reboot.  Can you try
turning of
> > > > runtime power management for the GPU by setting the runpm
module
> > > > parameter to 0?  I *think* this would be booting with
> > > > "nouveau.runpm=0".
> > > 
> > > Sorry, I wasn't really thinking here.  You already *said*
this is
> > > related to runtime suspend.  It only happens when the Nvidia GPU
has
> > > been suspended.
> > > 
> > > I don't know that much about suspend, but ISTR seeing
comments about
> > > resuming devices before we shutdown.  If we do that, maybe
there's
> > > some kind of race between that resume and the reboot?
> > 
> > I think we do in fact resume PCI devices before shutdown.  Here's
the
> > path I'm looking at:
> > 
> >   device_shutdown
> >     pm_runtime_get_noresume
> >     pm_runtime_barrier
> >     dev->bus->shutdown
> >       pci_device_shutdown
> >         pm_runtime_resume
> >           __pm_runtime_resume(dev, 0)
> >             rpm_resume(dev, 0)
> >               __update_runtime_status(dev, RPM_RESUMING)
> >               callback = RPM_GET_CALLBACK(dev, runtime_resume)
> >               rpm_callback(callback, dev)
> >                 __rpm_callback
> >                   pci_pm_runtime_resume
> >                     drv->pm->runtime_resume
> >                       nouveau_pmops_runtime_resume
> >                         nouveau_do_resume
> >                         schedule_work(hpd_work)   # <---
> >                         ...
> >                         nouveau_display_hpd_work
> >                           pm_runtime_get_sync
> >                           drm_helper_hpd_irq_event
> >                           pm_runtime_mark_last_busy
> >                           pm_runtime_put_sync
> > 
> > I'm curious about that "schedule_work(hpd_work)" near
the end because
> > no other drivers seem to use schedule_work() in the runtime_resume
> > path, and I don't know how that synchronizes with the shutdown
> > process.  I don't see anything that waits for
> > nouveau_display_hpd_work() to complete, so it seems like something
> > that could be a race.
> > 
> > I wonder this problem would be easier to reproduce if you added a
> > sleep in nouveau_display_hpd_work() as in the first hunk below, and I
> > wonder if the problem would then go away if you stopped scheduling
> > hpd_work as in the second hunk?  Obviously the second hunk isn't a
> > solution, it's just an attempt to figure out if I'm looking in
the
> > right area.
> 
> So I did try making it so that we cancel hpd_work from
> pci_driver->shutdown and wait for it to complete, however that
> didn't really seem to make any difference. 
Not being a scheduled work expert, I was unsure if this experiment was
equivalent to what I proposed.

I'm always suspicious of singleton solutions like this (using
schedule_work() in runtime_resume()) because usually they seem to be
solving a generic problem that should happen on many kinds of
hardware.  The 0b2fe6594fa2 ("drm/nouveau: Queue hpd_work on (runtime)
resume") commit log says:

  We need to call drm_helper_hpd_irq_event() on resume to properly
  detect monitor connection / disconnection on some laptops, use
  hpd_work for this to avoid deadlocks.

The situation of a monitor being connected or disconnected during
suspend can happen to *any* GPU, but the commit only changes nouveau,
which of course raises the question of how we deal with that in other
drivers.  If the Nvidia GPU has some unique behavior related to
monitor connection, that would explain special-case code there, but
the commit doesn't mention anything like that.

It should be simple to revert 0b2fe6594fa2 and see whether it changes
the behavior at all (well, simple except for the fact that this
problem isn't 100% reproducible in the first place).
> Do we want to have this discussion on the bz btw, or is this email
> thread fine?
Email is fine.
> > diff --git a/drivers/gpu/drm/nouveau/nouveau_display.c
> > b/drivers/gpu/drm/nouveau/nouveau_display.c
> > index 55c0fa451163..e50806012d41 100644
> > --- a/drivers/gpu/drm/nouveau/nouveau_display.c
> > +++ b/drivers/gpu/drm/nouveau/nouveau_display.c
> > @@ -350,6 +350,7 @@ nouveau_display_hpd_work(struct work_struct *work)
> >  
> >  	pm_runtime_get_sync(drm->dev->dev);
> >  
> > +	msleep(2000);
> >  	drm_helper_hpd_irq_event(drm->dev);
> >  
> >  	pm_runtime_mark_last_busy(drm->dev->dev);
> > 
> > 
> > 
> > diff --git a/drivers/gpu/drm/nouveau/nouveau_drm.c
> > b/drivers/gpu/drm/nouveau/nouveau_drm.c
> > index 5020265bfbd9..48da72caa017 100644
> > --- a/drivers/gpu/drm/nouveau/nouveau_drm.c
> > +++ b/drivers/gpu/drm/nouveau/nouveau_drm.c
> > @@ -946,9 +946,6 @@ nouveau_pmops_runtime_resume(struct device *dev)
> >  	nvif_mask(&device->object, 0x088488, (1 << 25), (1
<< 25));
> >  	drm_dev->switch_power_state = DRM_SWITCH_POWER_ON;
> >  
> > -	/* Monitors may have been connected / disconnected during suspend */
> > -	schedule_work(&nouveau_drm(drm_dev)->hpd_work);
> > -
> >  	return ret;
> >  }
> >  
> -- 
> Cheers,
> 	Lyude Paul
>

On Wed, 2019-04-24 at 13:59 -0500, Bjorn Helgaas wrote:
> Not being a scheduled work expert, I was unsure if this experiment was
> equivalent to what I proposed.
> 
> I'm always suspicious of singleton solutions like this (using
> schedule_work() in runtime_resume()) because usually they seem to be
> solving a generic problem that should happen on many kinds of
> hardware.  The 0b2fe6594fa2 ("drm/nouveau: Queue hpd_work on (runtime)
> resume") commit log says:
> 
>   We need to call drm_helper_hpd_irq_event() on resume to properly
>   detect monitor connection / disconnection on some laptops, use
>   hpd_work for this to avoid deadlocks.
> 
> The situation of a monitor being connected or disconnected during
> suspend can happen to *any* GPU, but the commit only changes nouveau,
> which of course raises the question of how we deal with that in other
> drivers.  If the Nvidia GPU has some unique behavior related to
> monitor connection, that would explain special-case code there, but
> the commit doesn't mention anything like that.
> 
> It should be simple to revert 0b2fe6594fa2 and see whether it changes
> the behavior at all (well, simple except for the fact that this
> problem isn't 100% reproducible in the first place).
It's not 100% reproducible, but it's at least 90% so it's not
difficult for me
to test at all.

Also, reverting this commit makes no difference either. Note that while that
commit only changed nouveau, scheduled_work() is exactly how a number of other
drivers (i915 for instance) handle reprobing like this as well. The reason
being that we can't do full connector reprobing in our runtime resume thread
because we could deadlock if someone else is holding a modesetting lock we
need and waiting on us to resume at the same time (there's a number of other
bug fixes in nouveau for other issues caused by the same deadlock scenario). 

I'm confused here though, it sounds like you're running under the
assumption
that PCI devices like this aren't reset into a clean state during a system
reboot, is that correct?
> 
> > Do we want to have this discussion on the bz btw, or is this email
> > thread fine?
> 
> Email is fine.
-- 
Cheers,
	Lyude Paul

On Wed, Apr 24, 2019 at 03:16:37PM -0400, Lyude Paul
wrote:
> On Wed, 2019-04-24 at 13:59 -0500, Bjorn Helgaas wrote:
> > Not being a scheduled work expert, I was unsure if this experiment was
> > equivalent to what I proposed.
> > 
> > I'm always suspicious of singleton solutions like this (using
> > schedule_work() in runtime_resume()) because usually they seem to be
> > solving a generic problem that should happen on many kinds of
> > hardware.  The 0b2fe6594fa2 ("drm/nouveau: Queue hpd_work on
(runtime)
> > resume") commit log says:
> > 
> >   We need to call drm_helper_hpd_irq_event() on resume to properly
> >   detect monitor connection / disconnection on some laptops, use
> >   hpd_work for this to avoid deadlocks.
> > 
> > The situation of a monitor being connected or disconnected during
> > suspend can happen to *any* GPU, but the commit only changes nouveau,
> > which of course raises the question of how we deal with that in other
> > drivers.  If the Nvidia GPU has some unique behavior related to
> > monitor connection, that would explain special-case code there, but
> > the commit doesn't mention anything like that.
> > 
> > It should be simple to revert 0b2fe6594fa2 and see whether it changes
> > the behavior at all (well, simple except for the fact that this
> > problem isn't 100% reproducible in the first place).
> 
> It's not 100% reproducible, but it's at least 90% so it's not
> difficult for me to test at all.
> 
> Also, reverting this commit makes no difference either. 
OK, great, that makes it crystal clear.  I didn't know you had
specifically tested that revert.  Thanks for doing that.
> Note that while that commit only changed nouveau, scheduled_work()
> is exactly how a number of other drivers (i915 for instance) handle
> reprobing like this as well.
OK.  The GPU code would be a lot more approachable if similar things
were done in similar ways.  I spent an hour or so looking for this
similar code in i915, but gave up.
> The reason being that we can't do full connector reprobing in our
> runtime resume thread because we could deadlock if someone else is
> holding a modesetting lock we need and waiting on us to resume at
> the same time (there's a number of other bug fixes in nouveau for
> other issues caused by the same deadlock scenario). 
You mention nouveau specifically here, but I assume this is a generic
deadlock scenario that applies to any GPU, and they all avoid the
deadlock in the same way.  Right?
> I'm confused here though, it sounds like you're running under the
> assumption that PCI devices like this aren't reset into a clean
> state during a system reboot, is that correct?
No, I wasn't trying to say anything about that.  My point here is
that:

  - you're reporting a problem that only happens with nouveau and
    only happens during shutdown/reboot
  - the behavior is similar to a race (not 100% reproducible, seems
    to happen more if shutdown is faster)
  - shutdown involves resuming the device (see pci_device_shutdown())
  - nouveau_pmops_runtime_resume() schedules asynchronous work, which
    (to my untrained eye) looks unusual
  - asynchronous work is inherently subject to races

So I think that's all somewhat suspicious.  But if the same problem
happens without the asynchronous work, obviously the issue is
elsewhere.

But you *are* right that if the device were actually reset, none of
this should matter.  It certainly seems that the BIOS neglects to
reset it in some cases.

I can sort of imagine a BIOS engineer thinking that if the device
looks like it's in use, we shouldn't reset it, and it's still
conceivable that some sort of Linux shutdown race could leave it
looking like it's in use.  But you've been working with Lenovo on
this, and it seems like that would be pretty obvious to somebody with
the BIOS source (though I just demonstrated above that even with the
source it's easy to miss things).

I'm out of ideas, so I think your quirk is the best way forward.  I
trimmed out some of the commit log backtraces and such, added the
bugzilla, and tweaked the patch to use pci_iomap() instead of
ioremap().  Would the patch below work for you?


commit 18dc5b3c7ddc
Author: Lyude Paul <lyude at redhat.com>
Date:   Tue Feb 12 17:02:30 2019 -0500

    PCI: Reset Lenovo ThinkPad P50 nvgpu at boot if necessary
    
    On ThinkPad P50 SKUs with an Nvidia Quadro M1000M instead of the M2000M
    variant, the BIOS does not always reset the secondary Nvidia GPU during
    reboot if the laptop is configured in Hybrid Graphics mode.  The reason is
    unknown, but the following steps and possibly a good bit of patience will
    reproduce the issue:
    
      1. Boot up the laptop normally in Hybrid Graphics mode
      2. Make sure nouveau is loaded and that the GPU is awake
      2. Allow the Nvidia GPU to runtime suspend itself after being idle
      3. Reboot the machine, the more sudden the better (e.g sysrq-b may help)
      4. If nouveau loads up properly, reboot the machine again and go back to
         step 2 until you reproduce the issue
    
    This results in some very strange behavior: the GPU will be left in exactly
    the same state it was in when the previously booted kernel started the
    reboot.  This has all sorts of bad side effects: for starters, this
    completely breaks nouveau starting with a mysterious EVO channel failure
    that happens well before we've actually used the EVO channel for
anything:
    
      nouveau 0000:01:00.0: disp: chid 0 mthd 0000 data 00000400 00001000
00000002
    
    This causes a timeout trying to bring up the GR ctx:
    
      nouveau 0000:01:00.0: timeout
      WARNING: CPU: 0 PID: 12 at
drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgf100.c:1547
gf100_grctx_generate+0x7b2/0x850 [nouveau]
      Hardware name: LENOVO 20EQS64N0B/20EQS64N0B, BIOS N1EET82W (1.55 )
12/18/2018
      Workqueue: events_long drm_dp_mst_link_probe_work [drm_kms_helper]
      ...
      nouveau 0000:01:00.0: gr: wait for idle timeout (en: 1, ctxsw: 0, busy: 1)
      nouveau 0000:01:00.0: gr: wait for idle timeout (en: 1, ctxsw: 0, busy: 1)
      nouveau 0000:01:00.0: fifo: fault 01 [WRITE] at 0000000000008000 engine 00
[GR] client 15 [HUB/SCC_NB] reason c4 [] on channel -1 [0000000000 unknown]
    
    The GPU never manages to recover.  Booting without loading nouveau causes
    issues as well, since the GPU starts sending spurious interrupts that cause
    other device's IRQs to get disabled by the kernel:
    
      irq 16: nobody cared (try booting with the "irqpoll" option)
      ...
      handlers:
      [<000000007faa9e99>] i801_isr [i2c_i801]
      Disabling IRQ #16
      ...
      serio: RMI4 PS/2 pass-through port at rmi4-00.fn03
      i801_smbus 0000:00:1f.4: Timeout waiting for interrupt!
      i801_smbus 0000:00:1f.4: Transaction timeout
      rmi4_f03 rmi4-00.fn03: rmi_f03_pt_write: Failed to write to F03 TX
register (-110).
      i801_smbus 0000:00:1f.4: Timeout waiting for interrupt!
      i801_smbus 0000:00:1f.4: Transaction timeout
      rmi4_physical rmi4-00: rmi_driver_set_irq_bits: Failed to change enabled
interrupts!
    
    This causes the touchpad and sometimes other things to get disabled.
    
    Since this happens without nouveau, we can't fix this problem from
nouveau
    itself.
    
    Add a PCI quirk for the specific P50 variant of this GPU.  Make sure the
    GPU is advertising NoReset- so we don't reset the GPU when the machine
is
    in Dedicated graphics mode (where the GPU being initialized by the BIOS is
    normal and expected).  Map the GPU MMIO space and read the magic 0x2240c
    register, which will have bit 1 set if the device was POSTed during a
    previous boot.  Once we've confirmed all of this, reset the GPU and
    re-disable it - bringing it back to a healthy state.
    
    Link: https://bugzilla.kernel.org/show_bug.cgi?id=203003
    Link: https://lore.kernel.org/lkml/20190212220230.1568-1-lyude at redhat.com
    Signed-off-by: Lyude Paul <lyude at redhat.com>
    Signed-off-by: Bjorn Helgaas <bhelgaas at google.com>
    Cc: nouveau at lists.freedesktop.org
    Cc: dri-devel at lists.freedesktop.org
    Cc: Karol Herbst <kherbst at redhat.com>
    Cc: Ben Skeggs <skeggsb at gmail.com>
    Cc: stable at vger.kernel.org

diff --git a/drivers/pci/quirks.c b/drivers/pci/quirks.c
index a59ad09ce911..819a595b0b1d 100644
--- a/drivers/pci/quirks.c
+++ b/drivers/pci/quirks.c
@@ -5120,3 +5120,61 @@ SWITCHTEC_QUIRK(0x8573);  /* PFXI 48XG3 */
 SWITCHTEC_QUIRK(0x8574);  /* PFXI 64XG3 */
 SWITCHTEC_QUIRK(0x8575);  /* PFXI 80XG3 */
 SWITCHTEC_QUIRK(0x8576);  /* PFXI 96XG3 */
+
+/*
+ * On Lenovo Thinkpad P50 SKUs with a Nvidia Quadro M1000M, the BIOS does
+ * not always reset the secondary Nvidia GPU between reboots if the system
+ * is configured to use Hybrid Graphics mode.  This results in the GPU
+ * being left in whatever state it was in during the *previous* boot, which
+ * causes spurious interrupts from the GPU, which in turn causes us to
+ * disable the wrong IRQ and end up breaking the touchpad.  Unsurprisingly,
+ * this also completely breaks nouveau.
+ *
+ * Luckily, it seems a simple reset of the Nvidia GPU brings it back to a
+ * clean state and fixes all these issues.
+ *
+ * When the machine is configured in Dedicated display mode, the issue
+ * doesn't occur.  Fortunately the GPU advertises NoReset+ when in this
+ * mode, so we can detect that and avoid resetting it.
+ */
+static void quirk_reset_lenovo_thinkpad_p50_nvgpu(struct pci_dev *pdev)
+{
+	void __iomem *map;
+	int ret;
+
+	if (pdev->subsystem_vendor != PCI_VENDOR_ID_LENOVO ||
+	    pdev->subsystem_device != 0x222e ||
+	    !pdev->reset_fn)
+		return;
+
+	if (pci_enable_device_mem(pdev))
+		return;
+
+	/*
+	 * Based on nvkm_device_ctor() in
+	 * drivers/gpu/drm/nouveau/nvkm/engine/device/base.c
+	 */
+	map = pci_iomap(pdev, 0, 0x23000);
+	if (!map) {
+		pci_err(pdev, "Can't map MMIO space\n");
+		goto out_disable;
+	}
+
+	/*
+	 * Make sure the GPU looks like it's been POSTed before resetting
+	 * it.
+	 */
+	if (ioread32(map + 0x2240c) & 0x2) {
+		pci_info(pdev, FW_BUG "GPU left initialized by EFI, resetting\n");
+		ret = pci_reset_function(pdev);
+		if (ret < 0)
+			pci_err(pdev, "Failed to reset GPU: %d\n", ret);
+	}
+
+	iounmap(map);
+out_disable:
+	pci_disable_device(pdev);
+}
+DECLARE_PCI_FIXUP_CLASS_FINAL(PCI_VENDOR_ID_NVIDIA, 0x13b1,
+			      PCI_CLASS_DISPLAY_VGA, 8,
+			      quirk_reset_lenovo_thinkpad_p50_nvgpu);