There are numerous ATI/AMD GPUs available that report that they support a PM reset (NoSoftRst-) but for which such a reset has no apparent effect on the device. These devices continue to display the same framebuffer across PM reset and the fan speed remains constant, while a proper bus reset causes the display to lose sync and the fan to reset to high speed. Create a device specific reset for ATI vendor devices that tries to catch these devices and report that pci_reset_function() is not supported.
Signed-off-by: Alex Williamson <[email protected]> --- This patch makes the series "vfio-pci: Reset improvements" far more useful for users with one of these GPUs. If pci_reset_function() indicates that it's supported and successful, then I have no reason to resort to a bus/slot reset in the vfio-pci code. Since it doesn't seem to do anything anyway, let's just forget that PM reset exists for these devices. drivers/pci/quirks.c | 53 ++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 53 insertions(+) diff --git a/drivers/pci/quirks.c b/drivers/pci/quirks.c index 460c354..bed9c63 100644 --- a/drivers/pci/quirks.c +++ b/drivers/pci/quirks.c @@ -3289,6 +3289,57 @@ static int reset_chelsio_generic_dev(struct pci_dev *dev, int probe) return 0; } +/* + * Numerous AMD/ATI GPUs report that they're capable of PM reset (NoSoftRst-) + * and pci_reset_function() reports the device as successfully reset, but + * there's no apparent effect from the reset. Test for these, being sure to + * allow FLR should it ever exist, and use the device specific reset to + * disable any sort of function-local reset if only PM reset is available. + */ +static int reset_ati_gpu(struct pci_dev *dev, int probe) +{ + u16 pm_csr; + u32 devcap; + int af_pos; + + /* + * VGA class devices, not on the root bus, PCI function 0 of a + * multifunction device with PM capabilities + */ + if ((dev->class >> 8) != PCI_CLASS_DISPLAY_VGA || + pci_is_root_bus(dev->bus) || PCI_FUNC(dev->devfn) || + !dev->multifunction || !dev->pm_cap) + return -ENOTTY; + + /* PM reports NoSoftRst- */ + pci_read_config_word(dev, dev->pm_cap + PCI_PM_CTRL, &pm_csr); + if (pm_csr & PCI_PM_CTRL_NO_SOFT_RESET) + return -ENOTTY; + + /* No PCIe FLR */ + pcie_capability_read_dword(dev, PCI_EXP_DEVCAP, &devcap); + if (devcap & PCI_EXP_DEVCAP_FLR) + return -ENOTTY; + + /* No AF FLR */ + af_pos = pci_find_capability(dev, PCI_CAP_ID_AF); + if (af_pos) { + u8 af_cap; + + pci_read_config_byte(dev, af_pos + PCI_AF_CAP, &af_cap); + if ((af_cap && PCI_AF_CAP_TP) && (af_cap && PCI_AF_CAP_FLR)) + return -ENOTTY; + } + + /* + * We could attempt a singleton bus/slot reset here to override + * PM reset priority over these, but the devices we're interested + * in are multifunction GPU + audio devices in their known configs. + */ + + return -EINVAL; +} + #define PCI_DEVICE_ID_INTEL_82599_SFP_VF 0x10ed #define PCI_DEVICE_ID_INTEL_IVB_M_VGA 0x0156 #define PCI_DEVICE_ID_INTEL_IVB_M2_VGA 0x0166 @@ -3304,6 +3355,8 @@ static const struct pci_dev_reset_methods pci_dev_reset_methods[] = { reset_intel_generic_dev }, { PCI_VENDOR_ID_CHELSIO, PCI_ANY_ID, reset_chelsio_generic_dev }, + { PCI_VENDOR_ID_ATI, PCI_ANY_ID, + reset_ati_gpu }, { 0 } }; -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
