| Message ID | 20170419003130.5302-1-yinghai@kernel.org (mailing list archive) |
|---|---|
| State | New, archived |
| Delegated to: | Bjorn Helgaas |
| Headers | show |
On 4/18/2017 8:31 PM, Yinghai Lu wrote: > * pci_setup_device - fill in class and map information of a device > * @dev: the device structure to fill > @@ -1572,6 +1592,9 @@ int pci_setup_device(struct pci_dev *dev > /* device class may be changed after fixup */ > class = dev->class >> 8; > > + if (is_kdump_kernel()) > + pci_disable_serr(dev); > + This sounds like something that needs to be done while shutting down the first kernel as part of the kdump procedure rather than boot of the kdump kernel in pci setup.
On Thu, Apr 20, 2017 at 12:14 PM, Sinan Kaya <okaya@codeaurora.org> wrote: > On 4/18/2017 8:31 PM, Yinghai Lu wrote: >> * pci_setup_device - fill in class and map information of a device >> * @dev: the device structure to fill >> @@ -1572,6 +1592,9 @@ int pci_setup_device(struct pci_dev *dev >> /* device class may be changed after fixup */ >> class = dev->class >> 8; >> >> + if (is_kdump_kernel()) >> + pci_disable_serr(dev); >> + > > This sounds like something that needs to be done while shutting down > the first kernel as part of the kdump procedure rather than boot of > the kdump kernel in pci setup. In general, I would rather make the new kernel more tolerant than make assumptions about how the old kernel shut down. I don't know if there's an explicit statement of kexec philosophy on this (it'd be nice if there were), but it seems like a more robust strategy, e.g., less prone to revlock issues between the old/new kernels. Bjorn
On 4/20/2017 2:38 PM, Bjorn Helgaas wrote: > On Thu, Apr 20, 2017 at 12:14 PM, Sinan Kaya <okaya@codeaurora.org> wrote: >> On 4/18/2017 8:31 PM, Yinghai Lu wrote: >>> * pci_setup_device - fill in class and map information of a device >>> * @dev: the device structure to fill >>> @@ -1572,6 +1592,9 @@ int pci_setup_device(struct pci_dev *dev >>> /* device class may be changed after fixup */ >>> class = dev->class >> 8; >>> >>> + if (is_kdump_kernel()) >>> + pci_disable_serr(dev); >>> + >> >> This sounds like something that needs to be done while shutting down >> the first kernel as part of the kdump procedure rather than boot of >> the kdump kernel in pci setup. > > In general, I would rather make the new kernel more tolerant than make > assumptions about how the old kernel shut down. I don't know if > there's an explicit statement of kexec philosophy on this (it'd be > nice if there were), but it seems like a more robust strategy, e.g., > less prone to revlock issues between the old/new kernels. > What if the secondary kernel never gets a chance to boot due to excessive errors? Code might not even make to the point where PCI driver is executed. If I remember this right, kexec is already doing PCI cleanup operation during shutdown and it is also calling the shutdown hook of device drivers. (I recently added a shutdown hook to my own HIDMA driver for the very same reason) The requirement for the second kernel boot is not to have any pending DMA and IRQs so that secondary kernel can boot safely. Maybe, the right thing is to look for a way to put PCI into some safe mode. There should be some code there disabling the COMMAND enable bits if not already. This code could be added to the same place. http://lxr.free-electrons.com/ident?i=pci_device_shutdown 469 /* 470 * If this is a kexec reboot, turn off Bus Master bit on the 471 * device to tell it to not continue to do DMA. Don't touch 472 * devices in D3cold or unknown states. 473 * If it is not a kexec reboot, firmware will hit the PCI 474 * devices with big hammer and stop their DMA any way. 475 */ 476 if (kexec_in_progress && (pci_dev->current_state <= PCI_D3hot)) 477 pci_clear_master(pci_dev); > Bjorn >
On Thu, Apr 20, 2017 at 10:14 AM, Sinan Kaya <okaya@codeaurora.org> wrote: > On 4/18/2017 8:31 PM, Yinghai Lu wrote: >> * pci_setup_device - fill in class and map information of a device >> * @dev: the device structure to fill >> @@ -1572,6 +1592,9 @@ int pci_setup_device(struct pci_dev *dev >> /* device class may be changed after fixup */ >> class = dev->class >> 8; >> >> + if (is_kdump_kernel()) >> + pci_disable_serr(dev); >> + > > This sounds like something that needs to be done while shutting down > the first kernel as part of the kdump procedure rather than boot of > the kdump kernel in pci setup. For kdump path, first kernel shutdown path is not called. We have to do sth in second kernel instead. Thanks Yinghai
On 4/20/2017 7:37 PM, Yinghai Lu wrote: > On Thu, Apr 20, 2017 at 10:14 AM, Sinan Kaya <okaya@codeaurora.org> wrote: >> On 4/18/2017 8:31 PM, Yinghai Lu wrote: >>> * pci_setup_device - fill in class and map information of a device >>> * @dev: the device structure to fill >>> @@ -1572,6 +1592,9 @@ int pci_setup_device(struct pci_dev *dev >>> /* device class may be changed after fixup */ >>> class = dev->class >> 8; >>> >>> + if (is_kdump_kernel()) >>> + pci_disable_serr(dev); >>> + >> >> This sounds like something that needs to be done while shutting down >> the first kernel as part of the kdump procedure rather than boot of >> the kdump kernel in pci setup. > > For kdump path, first kernel shutdown path is not called. > > We have to do sth in second kernel instead. > I didn't know that. Bjorn's compatibility point is also a concern. Ideally you want something in both places. > Thanks > > Yinghai >
Index: linux-2.6/drivers/pci/probe.c =================================================================== --- linux-2.6.orig/drivers/pci/probe.c +++ linux-2.6/drivers/pci/probe.c @@ -17,6 +17,8 @@ #include <linux/acpi.h> #include <linux/irqdomain.h> #include <linux/pm_runtime.h> +#include <linux/crash_dump.h> + #include "pci.h" #define CARDBUS_LATENCY_TIMER 176 /* secondary latency timer */ @@ -1515,6 +1517,24 @@ static void pci_msi_setup_pci_dev(struct pci_msix_clear_and_set_ctrl(dev, PCI_MSIX_FLAGS_ENABLE, 0); } +static void pci_disable_serr(struct pci_dev *dev) +{ + u16 pci_cmd, pci_bctl; + + pci_read_config_word(dev, PCI_COMMAND, &pci_cmd); + pci_cmd &= ~PCI_COMMAND_SERR; + pci_write_config_word(dev, PCI_COMMAND, pci_cmd); + dev_printk(KERN_DEBUG, &dev->dev, "SERR cleared\n"); + + /* Program bridge control value */ + if ((dev->class >> 8) == PCI_CLASS_BRIDGE_PCI) { + pci_read_config_word(dev, PCI_BRIDGE_CONTROL, &pci_bctl); + pci_bctl &= ~PCI_BRIDGE_CTL_SERR; + pci_write_config_word(dev, PCI_BRIDGE_CONTROL, pci_bctl); + dev_printk(KERN_DEBUG, &dev->dev, "BRIDGE SERR cleared\n"); + } +} + /** * pci_setup_device - fill in class and map information of a device * @dev: the device structure to fill @@ -1572,6 +1592,9 @@ int pci_setup_device(struct pci_dev *dev /* device class may be changed after fixup */ class = dev->class >> 8; + if (is_kdump_kernel()) + pci_disable_serr(dev); + if (dev->non_compliant_bars) { pci_read_config_word(dev, PCI_COMMAND, &cmd); if (cmd & (PCI_COMMAND_IO | PCI_COMMAND_MEMORY)) {
Found one system with infiniband with SRIOV enabled, kdump kernel SRIOV BAR probing trigger one pci fatal error. That assert error pin, and host get reset by BMC. We can just ignore that error to let kernel go on and kdump to create vmcore. -v2: add debug print out Signed-off-by: Yinghai Lu <yinghai@kernel.org> --- drivers/pci/probe.c | 23 +++++++++++++++++++++++ 1 file changed, 23 insertions(+)