Nouveau - Mar 2020 - [PATCH v7] pci: prevent putting nvidia GPUs into lower device states on certain intel bridges

Fixes the infamous 'runtime PM' bug many users are facing on Laptops
with
Nvidia Pascal GPUs by skipping said PCI power state changes on the GPU.

Depending on the used kernel there might be messages like those in demsg:

"nouveau 0000:01:00.0: Refused to change power state, currently in D3"
"nouveau 0000:01:00.0: can't change power state from D3cold to D0
(config
space inaccessible)"
followed by backtraces of kernel crashes or timeouts within nouveau.

It's still unkown why this issue exists, but this is a reliable workaround
and solves a very annoying issue for user having to choose between a
crashing kernel or higher power consumption of their Laptops.

Signed-off-by: Karol Herbst <kherbst at redhat.com>
Cc: Bjorn Helgaas <bhelgaas at google.com>
Cc: Lyude Paul <lyude at redhat.com>
Cc: Rafael J. Wysocki <rjw at rjwysocki.net>
Cc: Mika Westerberg <mika.westerberg at intel.com>
Cc: linux-pci at vger.kernel.org
Cc: linux-pm at vger.kernel.org
Cc: dri-devel at lists.freedesktop.org
Cc: nouveau at lists.freedesktop.org
Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=205623
Reviewed-by: Mika Westerberg <mika.westerberg at linux.intel.com>

---
v2: convert to pci_dev quirk
    put a proper technical explanation of the issue as a in-code comment
v3: disable it only for certain combinations of intel and nvidia hardware
v4: simplify quirk by setting flag on the GPU itself
v5: restructure quirk to make it easier to add new IDs
    fix whitespace issues
    fix potential NULL pointer access
    update the quirk documentation
v6: move quirk into nouveau
v7: fix typos and commit message

 drivers/gpu/drm/nouveau/nouveau_drm.c | 57 +++++++++++++++++++++++++++
 drivers/pci/pci.c                     |  8 ++++
 include/linux/pci.h                   |  1 +
 3 files changed, 66 insertions(+)

diff --git a/drivers/gpu/drm/nouveau/nouveau_drm.c
b/drivers/gpu/drm/nouveau/nouveau_drm.c
index b65ae817eabf..2c86f0248305 100644
--- a/drivers/gpu/drm/nouveau/nouveau_drm.c
+++ b/drivers/gpu/drm/nouveau/nouveau_drm.c
@@ -618,6 +618,60 @@ nouveau_drm_device_fini(struct drm_device *dev)
 	kfree(drm);
 }
 
+/*
+ * On some Intel PCIe bridge controllers doing a
+ * D0 -> D3hot -> D3cold -> D0 sequence causes Nvidia GPUs to not
reappear.
+ * Skipping the intermediate D3hot step seems to make it work again. This is
+ * probably caused by not meeting the expectation the involved AML code has
+ * when the GPU is put into D3hot state before invoking it.
+ *
+ * This leads to various manifestations of this issue:
+ *  - AML code execution to power on the GPU hits an infinite loop (as the
+ *    code waits on device memory to change).
+ *  - kernel crashes, as all PCI reads return -1, which most code isn't
able
+ *    to handle well enough.
+ *
+ * In all cases dmesg will contain at least one line like this:
+ * 'nouveau 0000:01:00.0: Refused to change power state, currently in
D3'
+ * followed by a lot of nouveau timeouts.
+ *
+ * In the \_SB.PCI0.PEG0.PG00._OFF code deeper down writes bit 0x80 to the not
+ * documented PCI config space register 0x248 of the Intel PCIe bridge
+ * controller (0x1901) in order to change the state of the PCIe link between
+ * the PCIe port and the GPU. There are alternative code paths using other
+ * registers, which seem to work fine (executed pre Windows 8):
+ *  - 0xbc bit 0x20 (publicly available documentation claims
'reserved')
+ *  - 0xb0 bit 0x10 (link disable)
+ * Changing the conditions inside the firmware by poking into the relevant
+ * addresses does resolve the issue, but it seemed to be ACPI private memory
+ * and not any device accessible memory at all, so there is no portable way of
+ * changing the conditions.
+ * On a XPS 9560 that means bits [0,3] on \CPEX need to be cleared.
+ *
+ * The only systems where this behavior can be seen are hybrid graphics laptops
+ * with a secondary Nvidia Maxwell, Pascal or Turing GPU. It's unclear
whether
+ * this issue only occurs in combination with listed Intel PCIe bridge
+ * controllers and the mentioned GPUs or other devices as well.
+ *
+ * documentation on the PCIe bridge controller can be found in the
+ * "7th Generation Intel? Processor Families for H Platforms Datasheet
Volume 2"
+ * Section "12 PCI Express* Controller (x16) Registers"
+ */
+
+static void quirk_broken_nv_runpm(struct pci_dev *dev)
+{
+	struct pci_dev *bridge = pci_upstream_bridge(dev);
+
+	if (!bridge || bridge->vendor != PCI_VENDOR_ID_INTEL)
+		return;
+
+	switch (bridge->device) {
+	case 0x1901:
+		dev->parent_d3cold = 1;
+		break;
+	}
+}
+
 static int nouveau_drm_probe(struct pci_dev *pdev,
 			     const struct pci_device_id *pent)
 {
@@ -699,6 +753,7 @@ static int nouveau_drm_probe(struct pci_dev *pdev,
 	if (ret)
 		goto fail_drm_dev_init;
 
+	quirk_broken_nv_runpm(pdev);
 	return 0;
 
 fail_drm_dev_init:
@@ -735,6 +790,8 @@ nouveau_drm_remove(struct pci_dev *pdev)
 {
 	struct drm_device *dev = pci_get_drvdata(pdev);
 
+	/* revert our workaround */
+	pdev->parent_d3cold = false;
 	nouveau_drm_device_remove(dev);
 	pci_disable_device(pdev);
 }
diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
index d828ca835a98..9c4044fc2553 100644
--- a/drivers/pci/pci.c
+++ b/drivers/pci/pci.c
@@ -861,6 +861,14 @@ static int pci_raw_set_power_state(struct pci_dev *dev,
pci_power_t state)
 	   || (state == PCI_D2 && !dev->d2_support))
 		return -EIO;
 
+	/*
+	 * Power management can be disabled for certain devices as they don't
+	 * come back up later on runtime_resume. We rely on platform means to
+	 * cut power consumption instead (e.g. ACPI).
+	 */
+	if (state != PCI_D0 && dev->parent_d3cold)
+		return 0;
+
 	pci_read_config_word(dev, dev->pm_cap + PCI_PM_CTRL, &pmcsr);
 	if (pmcsr == (u16) ~0) {
 		pci_err(dev, "can't change power state from %s to %s (config space
inaccessible)\n",
diff --git a/include/linux/pci.h b/include/linux/pci.h
index 3840a541a9de..3c01f043519a 100644
--- a/include/linux/pci.h
+++ b/include/linux/pci.h
@@ -340,6 +340,7 @@ struct pci_dev {
 	unsigned int	no_d3cold:1;	/* D3cold is forbidden */
 	unsigned int	bridge_d3:1;	/* Allow D3 for bridge */
 	unsigned int	d3cold_allowed:1;	/* D3cold is allowed by user */
+	unsigned int	parent_d3cold:1;	/* Power manage the parent instead */
 	unsigned int	mmio_always_on:1;	/* Disallow turning off io/mem
 						   decoding during BAR sizing */
 	unsigned int	wakeup_prepared:1;
-- 
2.24.1

On Tue, Mar 10, 2020 at 08:26:27PM +0100, Karol Herbst
wrote:
> Fixes the infamous 'runtime PM' bug many users are facing on
Laptops with
> Nvidia Pascal GPUs by skipping said PCI power state changes on the GPU.
> 
> Depending on the used kernel there might be messages like those in demsg:
> 
> "nouveau 0000:01:00.0: Refused to change power state, currently in
D3"
> "nouveau 0000:01:00.0: can't change power state from D3cold to D0
(config
> space inaccessible)"
> followed by backtraces of kernel crashes or timeouts within nouveau.
> 
> It's still unkown why this issue exists, but this is a reliable
workaround
> and solves a very annoying issue for user having to choose between a
> crashing kernel or higher power consumption of their Laptops.
Thanks for the bugzilla link.  The bugzilla mentions lots of mailing
list discussion.  Can you include links to some of that?

IIUC this basically just turns off PCI power management for the GPU.
Can you do that with something like the following?  I don't know
anything about DRM, so I don't know where you could save the pm_cap,
but I'm sure the driver could keep it somewhere.


diff --git a/drivers/gpu/drm/nouveau/nouveau_drm.c
b/drivers/gpu/drm/nouveau/nouveau_drm.c
index b65ae817eabf..2ad825e8891c 100644
--- a/drivers/gpu/drm/nouveau/nouveau_drm.c
+++ b/drivers/gpu/drm/nouveau/nouveau_drm.c
@@ -618,6 +618,23 @@ nouveau_drm_device_fini(struct drm_device *dev)
 	kfree(drm);
 }
 
+static void quirk_broken_nv_runpm(struct drm_device *drm_dev)
+{
+	struct pci_dev *pdev = drm_dev->pdev;
+	struct pci_dev *bridge = pci_upstream_bridge(pdev);
+
+	if (!bridge || bridge->vendor != PCI_VENDOR_ID_INTEL)
+		return;
+
+	switch (bridge->device) {
+	case 0x1901:
+		STASH->pm_cap = pdev->pm_cap;
+		pdev->pm_cap = 0;
+		NV_INFO(drm_dev, "Disabling PCI power management to avoid bug\n");
+		break;
+	}
+}
+
 static int nouveau_drm_probe(struct pci_dev *pdev,
 			     const struct pci_device_id *pent)
 {
@@ -699,6 +716,7 @@ static int nouveau_drm_probe(struct pci_dev *pdev,
 	if (ret)
 		goto fail_drm_dev_init;
 
+	quirk_broken_nv_runpm(drm_dev);
 	return 0;
 
 fail_drm_dev_init:
@@ -735,6 +753,9 @@ nouveau_drm_remove(struct pci_dev *pdev)
 {
 	struct drm_device *dev = pci_get_drvdata(pdev);
 
+	/* If we disabled PCI power management, restore it */
+	if (STASH->pm_cap)
+		pdev->pm_cap = STASH->pm_cap;
 	nouveau_drm_device_remove(dev);
 	pci_disable_device(pdev);
 }

On Fri, Mar 20, 2020 at 11:19 PM Bjorn Helgaas <helgaas at kernel.org>
wrote:
>
> On Tue, Mar 10, 2020 at 08:26:27PM +0100, Karol Herbst wrote:
> > Fixes the infamous 'runtime PM' bug many users are facing on
Laptops with
> > Nvidia Pascal GPUs by skipping said PCI power state changes on the
GPU.
> >
> > Depending on the used kernel there might be messages like those in
demsg:
> >
> > "nouveau 0000:01:00.0: Refused to change power state, currently
in D3"
> > "nouveau 0000:01:00.0: can't change power state from D3cold
to D0 (config
> > space inaccessible)"
> > followed by backtraces of kernel crashes or timeouts within nouveau.
> >
> > It's still unkown why this issue exists, but this is a reliable
workaround
> > and solves a very annoying issue for user having to choose between a
> > crashing kernel or higher power consumption of their Laptops.
>
> Thanks for the bugzilla link.  The bugzilla mentions lots of mailing
> list discussion.  Can you include links to some of that?
>
> IIUC this basically just turns off PCI power management for the GPU.
> Can you do that with something like the following?  I don't know
> anything about DRM, so I don't know where you could save the pm_cap,
> but I'm sure the driver could keep it somewhere.
>
Sure this would work? From a quick look over the pci code, it looks
like a of code would be skipped we really need, like the platform code
to turn off the GPU via ACPI. But I could also remember incorrectly on
how all of that worked again. I can of course try and see what the
effect of this patch would be. And would the parent bus even go into
D3hot if it knows one of its children is still at D0? Because that's
what the result of that would be as well, no? And I know that if the
bus stays in D0, that it has a negative impact on power consumption.

Anyway, I will try that out, I am just not seeing how that would help.
>
> diff --git a/drivers/gpu/drm/nouveau/nouveau_drm.c
b/drivers/gpu/drm/nouveau/nouveau_drm.c
> index b65ae817eabf..2ad825e8891c 100644
> --- a/drivers/gpu/drm/nouveau/nouveau_drm.c
> +++ b/drivers/gpu/drm/nouveau/nouveau_drm.c
> @@ -618,6 +618,23 @@ nouveau_drm_device_fini(struct drm_device *dev)
>         kfree(drm);
>  }
>
> +static void quirk_broken_nv_runpm(struct drm_device *drm_dev)
> +{
> +       struct pci_dev *pdev = drm_dev->pdev;
> +       struct pci_dev *bridge = pci_upstream_bridge(pdev);
> +
> +       if (!bridge || bridge->vendor != PCI_VENDOR_ID_INTEL)
> +               return;
> +
> +       switch (bridge->device) {
> +       case 0x1901:
> +               STASH->pm_cap = pdev->pm_cap;
> +               pdev->pm_cap = 0;
> +               NV_INFO(drm_dev, "Disabling PCI power management to
avoid bug\n");
> +               break;
> +       }
> +}
> +
>  static int nouveau_drm_probe(struct pci_dev *pdev,
>                              const struct pci_device_id *pent)
>  {
> @@ -699,6 +716,7 @@ static int nouveau_drm_probe(struct pci_dev *pdev,
>         if (ret)
>                 goto fail_drm_dev_init;
>
> +       quirk_broken_nv_runpm(drm_dev);
>         return 0;
>
>  fail_drm_dev_init:
> @@ -735,6 +753,9 @@ nouveau_drm_remove(struct pci_dev *pdev)
>  {
>         struct drm_device *dev = pci_get_drvdata(pdev);
>
> +       /* If we disabled PCI power management, restore it */
> +       if (STASH->pm_cap)
> +               pdev->pm_cap = STASH->pm_cap;
>         nouveau_drm_device_remove(dev);
>         pci_disable_device(pdev);
>  }
>