Hi Chang, I'm working on DMA problem of kdump and interested in your patch. What happens when clearing Bus Master bit? The system hangs up immediately after calling pci_device_shutdown()? Thanks, Takao Indoh (2013/10/30 20:22), Chang Liu wrote: > This patch adds a blacklist that prevents pci device shutdown code > from clearing Bus Master bit on certain hardware that cannot tolerate it. > > Clearing Bus Master bit was originally added in commit b566a22 to address > the issue of some PCI device breaking kexec by continuing to do DMA > after having been signaled to shutdown. However, this introduced a > regression (https://bugzilla.kernel.org/show_bug.cgi?id=63861) that > hangs the machine on kernel power off path. > > It has been pointed out previously (https://lkml.org/lkml/2012/6/6/545) > that clearing Bus Master bit during PCI shutdown may have surprising > consequences as some devices may not tolerate this, and may hang the > system indefinitely. However, not doing so may break kexec. Since only > one bug report has come up since the introduction of b566a22, therefore > indicating that these misbehaving devices are in the minority, the > logical way to fix the hang while not breaking kexec is to blacklist > these devices from having their Bus Master bit cleared during the PCI > shutdown routine. > > Signed-off-by: Chang Liu <cl91tp@xxxxxxxxx> > --- > This fixes the above mentioned bug https://bugzilla.kernel.org/show_bug.cgi?id=63861 > > As Alan Cox has warned in https://lkml.org/lkml/2012/6/6/545, some device > will break if we clear bus master bit on shutdown. So sooner or later > we will need to blacklist some device if we are to keep Bus Master being cleared > as the default behavior. Now with the first (as I have been aware) device > that breaks under this default behavior has surfaced, we might as well > add the infrastructure code now in case some other devices break down > in the future. These devices may not be many so a blacklist likely won't > add much maintainance overhead. And we should keep the blacklist here in > the pci shutdown code instead of in individual device drivers since this > is the most direct way and will likely aid future debug process. > > Regarding the kexec issue which the original commit b566a22c was trying to address: > kexec doesn't work on this machine, even before the patch was applied (ie. in > earlier kernels, I tested 3.11.6 and 3.12-rc7). Possibly because some devices > won't reinitialize themselves properly. On 3.11.6, kexec -e triggers an immediate > reboot. On 3.12-rc7, both with and without this patch applied, kexec -e can get it > halfway into booting the new kernel but then a reboot is also triggered. I'm not sure > if this patch has helped the kexec situation in this machine or not, but since it > has no effects on machines without this defective SATA controller (and/or its firmware), > I would say it is unlikely to negatively affect kexec on other machines. > > drivers/pci/pci-driver.c | 34 +++++++++++++++++++++++++++++++++- > 1 file changed, 33 insertions(+), 1 deletion(-) > > diff --git a/drivers/pci/pci-driver.c b/drivers/pci/pci-driver.c > index 98f7b9b..1744ebf 100644 > --- a/drivers/pci/pci-driver.c > +++ b/drivers/pci/pci-driver.c > @@ -376,6 +376,29 @@ static int pci_device_remove(struct device * dev) > return 0; > } > > +/* > + * Blacklisting certain hardware from having their Bus Master bit cleared > + * during device shutdown. This is to workaround certain hardware's issue > + * with clearing Bus Master bit that hangs the entire system. > + */ > +struct { > + unsigned short vendor; > + unsigned short device; > +} pci_bus_master_blacklist[] = { > + { 0x8086, 0x9c03 }, /* Intel Corporation Lynx Point-LP SATA Controller */ > +}; > + > +static bool pci_should_clear_master(struct pci_dev *pdev) > +{ > + int i; > + for (i = 0; i < ARRAY_SIZE(pci_bus_master_blacklist); i++) { > + if (pdev->vendor == pci_bus_master_blacklist[i].vendor > + && pdev->device == pci_bus_master_blacklist[i].device) > + return false; > + } > + return true; > +} > + > static void pci_device_shutdown(struct device *dev) > { > struct pci_dev *pci_dev = to_pci_dev(dev); > @@ -391,9 +414,18 @@ static void pci_device_shutdown(struct device *dev) > /* > * Turn off Bus Master bit on the device to tell it to not > * continue to do DMA. Don't touch devices in D3cold or unknown states. > + * This is useful for proper kexec. However, a number of hardware > + * aren't happy with this. At the slightest, some hardware simply > + * ignore the bus master bit. For some other hardware, clearing > + * the bit on shutdown path hangs the entire system. > + * This is likely to be a firmware or hardware problem, but > + * we can workaround it here by blacklisting these hardware > + * from having their bus master bit cleared during device shutdown. > */ > - if (pci_dev->current_state <= PCI_D3hot) > + if (pci_should_clear_master(pci_dev) > + && pci_dev->current_state <= PCI_D3hot) { > pci_clear_master(pci_dev); > + } > } > > #ifdef CONFIG_PM > -- To unsubscribe from this list: send the line "unsubscribe linux-pci" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
