Re: [PATCH v2] vfio/pci: Handle concurrent vma faults

On Mon, 28 Jun 2021 20:26:25 -0300
Jason Gunthorpe [off-list ref] wrote:

On Mon, Jun 28, 2021 at 01:30:19PM -0600, Alex Williamson wrote:

quoted

On Mon, 28 Jun 2021 15:52:42 -0300
Jason Gunthorpe [off-list ref] wrote:
  

quoted

On Mon, Jun 28, 2021 at 12:36:21PM -0600, Alex Williamson wrote:  

quoted

On Mon, 28 Jun 2021 14:30:28 -0300
Jason Gunthorpe [off-list ref] wrote:
    

quoted

On Mon, Jun 28, 2021 at 10:46:53AM -0600, Alex Williamson wrote:    

quoted

On Wed, 10 Mar 2021 11:58:07 -0700
Alex Williamson [off-list ref] wrote:
      

quoted

vfio_pci_mmap_fault() incorrectly makes use of io_remap_pfn_range()
from within a vm_ops fault handler.  This function will trigger a
BUG_ON if it encounters a populated pte within the remapped range,
where any fault is meant to populate the entire vma.  Concurrent
inflight faults to the same vma will therefore hit this issue,
triggering traces such as:      

If it is just about concurrancy can the vma_lock enclose
io_remap_pfn_range() ?    

We could extend vma_lock around io_remap_pfn_range(), but that alone
would just block the concurrent faults to the same vma and once we
released them they'd still hit the BUG_ON in io_remap_pfn_range()
because the page is no longer pte_none().  We'd need to combine that
with something like __vfio_pci_add_vma() returning -EEXIST to skip the
io_remap_pfn_range(), but I've been advised that we shouldn't be
calling io_remap_pfn_range() from within the fault handler anyway, we
should be using something like vmf_insert_pfn() instead, which I
understand can be called safely in the same situation.  That's rather
the testing I was hoping someone who reproduced the issue previously
could validate.    

Yes, using the vmf_ stuff is 'righter' for sure, but there isn't
really a vmf for IO mappings..
  

quoted

quoted

I assume there is a reason why vm_lock can't be used here, so I
wouldn't object, though I don't especially like the loss of tracking
either.    

There's no loss of tracking here, we were only expecting a single fault
per vma to add the vma to our list.  This just skips adding duplicates
in these cases where we can have multiple faults in-flight.  Thanks,    

I mean the arch tracking of IO maps that is hidden inside ioremap_pfn  

Ok, so I take it you'd feel more comfortable with something like this,
right?  Thanks,  

I think so, it doesn't abuse the arch code, but it does abuse not
using vmf_* in a fault handler.

quoted

index 759dfb118712..74fc66cf9cf4 100644

+++ b/drivers/vfio/pci/vfio_pci.c

@@ -1584,6 +1584,7 @@ static vm_fault_t vfio_pci_mmap_fault(struct vm_fault *vmf)
 {
 	struct vm_area_struct *vma = vmf->vma;
 	struct vfio_pci_device *vdev = vma->vm_private_data;
+	struct vfio_pci_mmap_vma *mmap_vma;
 	vm_fault_t ret = VM_FAULT_NOPAGE;
 
 	mutex_lock(&vdev->vma_lock);

@@ -1591,24 +1592,33 @@ static vm_fault_t vfio_pci_mmap_fault(struct vm_fault *vmf)
 
 	if (!__vfio_pci_memory_enabled(vdev)) {
 		ret = VM_FAULT_SIGBUS;
-		mutex_unlock(&vdev->vma_lock);
 		goto up_out;
 	}
 
-	if (__vfio_pci_add_vma(vdev, vma)) {
-		ret = VM_FAULT_OOM;
-		mutex_unlock(&vdev->vma_lock);
-		goto up_out;  

quoted

+	/*
+	 * Skip existing vmas, assume concurrent in-flight faults to avoid
+	 * BUG_ON from io_remap_pfn_range() hitting !pte_none() pages.
+	 */
+	list_for_each_entry(mmap_vma, &vdev->vma_list, vma_next) {
+		if (mmap_vma->vma == vma)
+			goto up_out;
 	}
 
-	mutex_unlock(&vdev->vma_lock);
-
 	if (io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
-			       vma->vm_end - vma->vm_start, vma->vm_page_prot))
+			       vma->vm_end - vma->vm_start,
+			       vma->vm_page_prot)) {
 		ret = VM_FAULT_SIGBUS;
+		goto up_out;
+	}  

I suppose io_remap_pfn_range can fail inside after partially
populating the range, ie if it fails to allocate another pte table or
something.

Since partial allocations are not allowed we'd have to zap it here
too.

Yup, easy enough.
 

I suppose the other idea is to do the io_remap_pfn_range() when the
mmap becomes valid and the zap when it becomes invalid and just have
the fault handler always fail. That way we don't abuse anything.

That's coming, but it's a more substantial change, I'd rather fix this
within the current framework first.

Was there some tricky locking reason why this didn't work? Does it get
better with the address_space?

Yes it does, we can create the reverse of unmmap_mapping_range() using
the address space solution, which hugely simplifies zapping and
re-inserting BAR mappings.  This has been stalled due to lack of time
to work out the notifier issues for dropping IOMMU mappings, but it
also stands on its own, so I'll see if I can get that far in the series
reposted.  v3 w/ zap on io_remap_pfn_range() error path coming.  Thanks,

Alex