On Wed, 2012-06-06 at 21:09 +0100, Matthew Garrett wrote: > On Wed, Jun 06, 2012 at 12:42:07PM -0700, Eric W. Biederman wrote: > > > pci_device_shutdown calls drv->shutdown before calling > > pci_device_disable. Which means that only devices that have trouble > > with this bit being flipped while DMA is ongoing and don't bother > > to stop their own DMA will have a problem. > > drv->shutdown should already be quiescing the hardware. If it isn't, it > should be. If it is, what does this patch fix? Many drivers > call pci_enable_device() early enough that they clearly expect the > hardware to be quiescent when they do so, so this really does seem to > simply handle the kexec case without handling any other cases that could > be similarly problematic. > I agree drv->shutdown should quiesce the hardware (although anecdotal evidence suggests that is not happening), so turning Bus Master bit off is an additional safety measure. As Alan pointed out, doing that can be fraught with danger. Clearing Bus Master bit seems like the right thing to do, but due to buggy hardware/firmware it can cause problems, although those problems happen at predictable times and thus become easier to diagnose. I am considering other options. Andi mentioned command line option which does allow for some control instead of blindly clearing Bus Master bit on all system. I am also wondering if clearing Bus Master bit on just the PCI bridges is an option. If Bus Master bit is clear on a bridge, it is not supposed to allow DMA transactions through it. That can prevent random memory corruption by broken devices and also not upset the devices that do not like their Bus Master bit cleared. Any opinions on this option? I have also acquired a broadcom NIC that is causing what looks like random DMAs into kernel memory on a kexec, although this is with a kernel that does not have the Bus Master reset patch. I will run some experiments on this NIC and see what I can figure out. ==================================================================== Khalid Aziz Unix Systems Lab (970)898-9214 Hewlett-Packard khalid.aziz@xxxxxx Fort Collins, CO -- 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
