Messages in this thread | | | Subject | Re: [PATCH] KVM: arm64: Correctly handle the mmio faulting | From | Gavin Shan <> | Date | Thu, 22 Apr 2021 12:02:00 +1000 |
| |
Hi Marc,
On 4/21/21 9:59 PM, Marc Zyngier wrote: > On Wed, 21 Apr 2021 07:17:44 +0100, > Keqian Zhu <zhukeqian1@huawei.com> wrote: >> On 2021/4/21 14:20, Gavin Shan wrote: >>> On 4/21/21 12:59 PM, Keqian Zhu wrote: >>>> On 2020/10/22 0:16, Santosh Shukla wrote: >>>>> The Commit:6d674e28 introduces a notion to detect and handle the >>>>> device mapping. The commit checks for the VM_PFNMAP flag is set >>>>> in vma->flags and if set then marks force_pte to true such that >>>>> if force_pte is true then ignore the THP function check >>>>> (/transparent_hugepage_adjust()). >>>>> >>>>> There could be an issue with the VM_PFNMAP flag setting and checking. >>>>> For example consider a case where the mdev vendor driver register's >>>>> the vma_fault handler named vma_mmio_fault(), which maps the >>>>> host MMIO region in-turn calls remap_pfn_range() and maps >>>>> the MMIO's vma space. Where, remap_pfn_range implicitly sets >>>>> the VM_PFNMAP flag into vma->flags. >>>> Could you give the name of the mdev vendor driver that triggers this issue? >>>> I failed to find one according to your description. Thanks. >>>> >>> >>> I think it would be fixed in driver side to set VM_PFNMAP in >>> its mmap() callback (call_mmap()), like vfio PCI driver does. >>> It means it won't be delayed until page fault is issued and >>> remap_pfn_range() is called. It's determined from the beginning >>> that the vma associated the mdev vendor driver is serving as >>> PFN remapping purpose. So the vma should be populated completely, >>> including the VM_PFNMAP flag before it becomes visible to user >>> space. > > Why should that be a requirement? Lazy populating of the VMA should be > perfectly acceptable if the fault can only happen on the CPU side. >
It isn't a requirement and the drivers needn't follow strictly. I checked several drivers before looking into the patch and found almost all the drivers have VM_PFNMAP set at mmap() time. In drivers/vfio/vfio-pci.c, there is a comment as below, but it doesn't reveal too much about why we can't set VM_PFNMAP at fault time.
static int vfio_pci_mmap(void *device_data, struct vm_area_struct *vma) { : /* * See remap_pfn_range(), called from vfio_pci_fault() but we can't * change vm_flags within the fault handler. Set them now. */ vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP; vma->vm_ops = &vfio_pci_mmap_ops;
return 0; }
To set these flags in advance does have advantages. For example, VM_DONTEXPAND prevents the vma to be merged with another one. VM_DONTDUMP make this vma isn't eligible for coredump. Otherwise, the address space, which is associated with the vma is accessed and unnecessary page faults are triggered on coredump. VM_IO and VM_PFNMAP avoids to walk the page frames associated with the vma since we don't have valid PFN in the mapping.
>>> >>> The example can be found from vfio driver in drivers/vfio/pci/vfio_pci.c: >>> vfio_pci_mmap: VM_PFNMAP is set for the vma >>> vfio_pci_mmap_fault: remap_pfn_range() is called >> Right. I have discussed the above with Marc. I want to find the driver >> to fix it. However, AFAICS, there is no driver matches the description... > > I have the feeling this is an out-of-tree driver (and Santosh email > address is bouncing, so I guess we won't have much information from > him). > > However, the simple fact that any odd driver can provide a fault > handler and populate it the VMA on demand makes me think that we need > to support this case. >
Yeah, Santosh's email isn't reachable. I think the VM_PFNMAP need to be rechecked after gup if we're going to support this case.
Thanks, Gavin
| |