[RFC PATCH] pci: prevent putting pcie devices into lower device states on certain intel bridges

[Karol Herbst]

Fixes state transitions of Nvidia Pascal GPUs from D3cold into higher device
states.

v2: convert to pci_dev quirk
    put a proper technical explenation of the issue as a in-code comment

RFC comment (copied from last sent):
We are quite sure that there is a higher amount of bridges affected by this,
but I was only testing it on my own machine for now.

I've stresstested runpm by doing 5000 runpm cycles with that patch applied
and never saw it fail.

I mainly wanted to get a discussion going on if that's a feasable workaround
indeed or if we need something better.

I am also sure, that the nouveau driver itself isn't at fault as I am able
to reproduce the same issue by poking into some PCI registers on the PCIe
bridge to put the GPU into D3cold as it's done in ACPI code.

I've written a little python script to reproduce this issue without the need
of loading nouveau:
https://raw.githubusercontent.com/karolherbst/pci-stub-runpm/master/nv_runpm_bug_test.py

Signed-off-by: Karol Herbst <kherbst@xxxxxxxxxx>
Cc: Bjorn Helgaas <bhelgaas@xxxxxxxxxx>
Cc: Lyude Paul <lyude@xxxxxxxxxx>
Cc: Rafael J. Wysocki <rjw@xxxxxxxxxxxxx>
Cc: Mika Westerberg <mika.westerberg@xxxxxxxxx>
Cc: linux-pci@xxxxxxxxxxxxxxx
Cc: linux-pm@xxxxxxxxxxxxxxx
Cc: dri-devel@xxxxxxxxxxxxxxxxxxxxx
Cc: nouveau@xxxxxxxxxxxxxxxxxxxxx
---
 drivers/pci/pci.c    | 11 ++++++++++
 drivers/pci/quirks.c | 50 ++++++++++++++++++++++++++++++++++++++++++++
 include/linux/pci.h  |  1 +
 3 files changed, 62 insertions(+)

diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
index 088fcdc8d2b4..65516b024ee5 100644
--- a/drivers/pci/pci.c
+++ b/drivers/pci/pci.c
@@ -799,6 +799,13 @@ static inline bool platform_pci_bridge_d3(struct pci_dev *dev)
 	return pci_platform_pm ? pci_platform_pm->bridge_d3(dev) : false;
 }
 
+static inline bool parent_broken_child_pm(struct pci_dev *dev)
+{
+	if (!dev->bus || !dev->bus->self)
+		return false;
+	return dev->bus->self->broken_child_pm;
+}
+
 /**
  * pci_raw_set_power_state - Use PCI PM registers to set the power state of
  *			     given PCI device
@@ -844,6 +851,10 @@ static int pci_raw_set_power_state(struct pci_dev *dev, pci_power_t state)
 	   || (state == PCI_D2 && !dev->d2_support))
 		return -EIO;
 
+	/* check if the bus controller causes issues */
+	if (state != PCI_D0 && parent_broken_child_pm(dev))
+		return 0;
+
 	pci_read_config_word(dev, dev->pm_cap + PCI_PM_CTRL, &pmcsr);
 
 	/*
diff --git a/drivers/pci/quirks.c b/drivers/pci/quirks.c
index 0f16acc323c6..2be0deec2c3d 100644
--- a/drivers/pci/quirks.c
+++ b/drivers/pci/quirks.c
@@ -5198,3 +5198,53 @@ static void quirk_reset_lenovo_thinkpad_p50_nvgpu(struct pci_dev *pdev)
 DECLARE_PCI_FIXUP_CLASS_FINAL(PCI_VENDOR_ID_NVIDIA, 0x13b1,
 			      PCI_CLASS_DISPLAY_VGA, 8,
 			      quirk_reset_lenovo_thinkpad_p50_nvgpu);
+
+/*
+ * Some Intel PCIe bridges cause devices to disappear from the PCIe bus after
+ * those were put into D3cold state if they were put into a non D0 PCI PM
+ * device state before doing so.
+ *
+ * This leads to various issue different issues which all manifest differently,
+ * but have the same root cause:
+ *  - AIML code execution hits an infinite loop (as the coe waits on device
+ *    memory to change).
+ *  - kernel crashes, as all pci reads return -1, which most code isn't able
+ *    to handle well enough.
+ *  - sudden shutdowns, as the kernel identified an unrecoverable error after
+ *    userspace tries to access the GPU.
+ *
+ * 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.
+ *
+ * ACPI code writes bit 0x80 to the not documented PCI register 0x248 of the
+ * PCIe bridge controller in order to power down the GPU.
+ * Nonetheless, there are other code paths inside the ACPI firmware which use
+ * other registers, which seem to work fine:
+ *  - 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.
+ *
+ * The only systems where this behavior can be seen are hybrid graphics laptops
+ * with a secondary Nvidia Pascal GPU. It cannot be ruled out that this issue
+ * only occurs in combination with listed Intel PCIe bridge controllers and
+ * the mentioned GPUs or if it's only a hw bug in the bridge controller.
+ *
+ * But because this issue was NOT seen on laptops with an Nvidia Pascal GPU
+ * and an Intel Coffee Lake SoC, there is a higher chance of there being a bug
+ * in the bridge controller rather than in the GPU.
+ *
+ * This issue was not able to be reproduced on non laptop systems.
+ */
+
+static void quirk_broken_child_pm(struct pci_dev *dev)
+{
+	dev->broken_child_pm = 1;
+	printk("applied broken child pm quirk!\n");
+}
+/* kaby lake */
+DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL, 0x1901,
+			quirk_broken_child_pm);
diff --git a/include/linux/pci.h b/include/linux/pci.h
index dd436da7eccc..01eb0a9e6c13 100644
--- a/include/linux/pci.h
+++ b/include/linux/pci.h
@@ -413,6 +413,7 @@ struct pci_dev {
 	unsigned int	__aer_firmware_first_valid:1;
 	unsigned int	__aer_firmware_first:1;
 	unsigned int	broken_intx_masking:1;	/* INTx masking can't be used */
+	unsigned int	broken_child_pm:1;	/* children put into lower PCI PM states won't recover after D3cold transition */
 	unsigned int	io_window_1k:1;		/* Intel bridge 1K I/O windows */
 	unsigned int	irq_managed:1;
 	unsigned int	has_secondary_link:1;
-- 
2.21.0