On Wed, Mar 31, 2010 at 11:54:30AM -0400, Vivek Goyal wrote: > On Wed, Mar 31, 2010 at 11:24:17AM -0400, Neil Horman wrote: > > Flush iommu during shutdown > > > > When using an iommu, its possible, if a kdump kernel boot follows a primary > > kernel crash, that dma operations might still be in flight from the previous > > kernel during the kdump kernel boot. This can lead to memory corruption, > > crashes, and other erroneous behavior, specifically I've seen it manifest during > > a kdump boot as endless iommu error log entries of the form: > > AMD-Vi: Event logged [IO_PAGE_FAULT device=00:14.1 domain=0x000d > > address=0x000000000245a0c0 flags=0x0070] > > > > Followed by an inability to access hard drives, and various other resources. > > > > I've written this fix for it. In short it just forces a flush of the in flight > > dma operations on shutdown, so that the new kernel is certain not to have any > > in-flight dmas trying to complete after we've reset all the iommu page tables, > > causing the above errors. I've tested it and it fixes the problem for me quite > > well. > > CCing Eric also. > > Neil, this is interesting. In the past we noticed similar issues, > especially on PPC. But I was told that we could not clear the iommu > mapping entries as we had no control on in flight DMA and if a DMA comes > later after clearing an entry and entry is not present, it is an error. > Yes, the problem is (as I understand it) is that the triggering of DMA operations to/from a device doesn't have synchronization with the iommu itself. I.e. to conduct a dma you have to: 1) map the in-memory buffer to a dma address using something like pci_map_single. This results (in systems with an iommu) getting page table space allocated in the iommu for the translation. 2) triggering the dma to/from the device by tickling whatever hardware the device has mapped. 3) completing the dma by calling pci_unmap_single (or other function) which frees the page table space in the iommu The problem, exactly as you indicate is that on a kdump panic, we might boot the new kernel and re-enable the iommu with these dmas still in flight. If we start messing about with the iommu page tables then, we start getting all sorts of errors, and other various failures. > Hence one of the suggestions was not to clear iommu mapping entries but > reserve some for kdump operation and use those in kdump kernel. > Yeah, thats a solution, but it seems awfully complex to me. To do that, we need to teach every iommu we support about kdump, by telling it how much space to reserve, and when to use it and when not to (i.e. we'd have to tell it to use the kdump space, vs the normal space dependent on the status of the reset_devices flag, or something equally unpleasant). Actually, thinking about it, I'm not sure that will even work, as IIRC the iommu only has one page table base pointer. So we would either need to re-write that pointer to point into the kdump kernels memory space (invalidating the old table entries, which perpetuates this bug), or we would need to further enhance the iommu code to be able to access the old page tables via read_from_oldmem/write_to_oldmem when booting a kdump kernel, wouldn't we? Using this method, all we really do is try to ensure that, prior to disabling the iommu, we make sure that any pending dmas are complete. That way, when we re-enable the iommu in the kdump kernel, we can safely maniuplate the new page tables, knowing that no pending dma is using them In fairness to this debate, my proposal does have a small race condition. In the above sequence, because the cpu triggers a dma independently of the setup of the mapping in the iommu, it is possible that a dma might be triggered immediately after we flush the iotlb, which may leave an in-flight dma pending while we boot the kdump kernel. In practice though, this will never happen. By the time we arrive at this code, we've already executed native_machine_crash_shutdown which: 1) halts all the other cpus in the system 2) disables local interrupts Because of those two events, we're effectively on a path that we can't be preempted-from. So as long as we don't trigger any dma operations between our return from iommu_shutdown and machine_kexec (which is the next call), we're safe. > So this call amd_iommu_flush_all_devices() will be able to tell devices > that don't do any more DMAs and hence it is safe to reprogram iommu > mapping entries. > It blocks the cpu until any pending DMA operations are complete. Hmm, as I think about it, there is still a small possibility that a device like a NIC which has several buffers pre-dma-mapped could start a new dma before we completely disabled the iommu, althought thats small. I never saw that in my testing, but hitting that would be fairly difficult I think, since its literally just a few hundred cycles between the flush and the actual hardware disable operation. According to this though: http://support.amd.com/us/Processor_TechDocs/34434-IOMMU-Rev_1.26_2-11-09.pdf That window could be closed fairly easily, but simply disabling read and write permissions for each device table entry prior to calling flush. If we do that, then flush the device table, any subsequently started dma operation would just get noted in the error log, which we could ignore, since we're abot to boot to the kdump kernel anyway. Would you like me to respin w/ that modification? Neil >
