Infiniband allows the mapping of PCI memory regions to userland to achieve faster speeds. Using this feature with mthca cards in a Xen dom0 kernel leads to a pretty bad crash, and the process being killed (see http://xen.1045712.n5.nabble.com/Infiniband-from-userland-in-dom0-process-killed-bad-pagetable-td3259124.html )
This patch fixes the issue by setting the VM_IO flag on the vma, and refreshing the vm_page_prot field. The vm_page_prot will then have the _PAGE_IOMAP bit set, and Xen will therefore map the actual PCI space rather than a piece of non-existent RAM. Signed-off-by: Vivien Bernet-Rollande <v...@soprive.net> --- This has been heavily tested on a Fedora Core modified 2.6.32 kernel (with pvops patches : http://fedorapeople.org/~myoung/dom0/src/ ).I have also checked that it does not break anything on a stock kernel (2.6.37-rc7 from kernel.org), although only outside of Xen. There has also been some discussion about this issue on the Xen mailing list : http://lists.xensource.com/archives/html/xen-devel/2010-11/msg01739.html diff -uNrp kernel-2.6.32.fc12.orig/drivers/infiniband/hw/mthca/mthca_provider.c kernel-2.6.32.fc12.new/drivers/infiniband/hw/mthca/mthca_provider.c --- kernel-2.6.32.fc12.orig/drivers/infiniband/hw/mthca/mthca_provider.c 2010-12-16 16:39:06.901130993 +0100 +++ kernel-2.6.32.fc12.new/drivers/infiniband/hw/mthca/mthca_provider.c 2010-12-16 17:02:07.944127489 +0100 @@ -391,6 +391,8 @@ static int mthca_mmap_uar(struct ib_ucon if (vma->vm_end - vma->vm_start != PAGE_SIZE) return -EINVAL; + vma->vm_flags |= VM_IO; + vma->vm_page_prot = vm_get_page_prot(vma->vm_flags); vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); if (io_remap_pfn_range(vma, vma->vm_start, -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
