Re: [PATCH] vfio/pci: make the vfio_pci_mmap_fault reentrant

[Alex Williamson]

On Tue, 9 Mar 2021 14:48:24 -0500
Peter Xu <peterx@xxxxxxxxxx> wrote:

> On Tue, Mar 09, 2021 at 12:26:07PM -0700, Alex Williamson wrote:
> > On Tue, 9 Mar 2021 13:47:39 -0500
> > Peter Xu <peterx@xxxxxxxxxx> wrote:
> >   
> > > On Tue, Mar 09, 2021 at 12:40:04PM -0400, Jason Gunthorpe wrote:  
> > > > On Tue, Mar 09, 2021 at 08:29:51AM -0700, Alex Williamson wrote:    
> > > > > On Tue, 9 Mar 2021 08:46:09 -0400
> > > > > Jason Gunthorpe <jgg@xxxxxxxxxx> wrote:
> > > > >     
> > > > > > On Tue, Mar 09, 2021 at 03:49:09AM +0000, Zengtao (B) wrote:    
> > > > > > > Hi guys:
> > > > > > > 
> > > > > > > Thanks for the helpful comments, after rethinking the issue, I have proposed
> > > > > > >  the following change: 
> > > > > > > 1. follow_pte instead of follow_pfn.      
> > > > > > 
> > > > > > Still no on follow_pfn, you don't need it once you use vmf_insert_pfn    
> > > > > 
> > > > > vmf_insert_pfn() only solves the BUG_ON, follow_pte() is being used
> > > > > here to determine whether the translation is already present to avoid
> > > > > both duplicate work in inserting the translation and allocating a
> > > > > duplicate vma tracking structure.    
> > > >  
> > > > Oh.. Doing something stateful in fault is not nice at all
> > > > 
> > > > I would rather see __vfio_pci_add_vma() search the vma_list for dups
> > > > than call follow_pfn/pte..    
> > > 
> > > It seems to me that searching vma list is still the simplest way to fix the
> > > problem for the current code base.  I see io_remap_pfn_range() is also used in
> > > the new series - maybe that'll need to be moved to where PCI_COMMAND_MEMORY got
> > > turned on/off in the new series (I just noticed remap_pfn_range modifies vma
> > > flags..), as you suggested in the other email.  
> > 
> > 
> > In the new series, I think the fault handler becomes (untested):
> > 
> > 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;
> >         unsigned long base_pfn, pgoff;
> >         vm_fault_t ret = VM_FAULT_SIGBUS;
> > 
> >         if (vfio_pci_bar_vma_to_pfn(vma, &base_pfn))
> >                 return ret;
> > 
> >         pgoff = (vmf->address - vma->vm_start) >> PAGE_SHIFT;
> > 
> >         down_read(&vdev->memory_lock);
> > 
> >         if (__vfio_pci_memory_enabled(vdev))
> >                 ret = vmf_insert_pfn(vma, vmf->address, pgoff + base_pfn);
> > 
> >         up_read(&vdev->memory_lock);
> > 
> >         return ret;
> > }  
> 
> It's just that the initial MMIO access delay would be spread to the 1st access
> of each mmio page access rather than using the previous pre-fault scheme.  I
> think an userspace cares the delay enough should pre-fault all pages anyway,
> but just raise this up.  Otherwise looks sane.

Yep, this is a concern.  Is it safe to have loops concurrently and fully
populating the same vma with vmf_insert_pfn()?  If it is then we could
just ignore that we're doing duplicate work when we hit this race
condition.  Otherwise we'd need to serialize again, perhaps via a lock
and flag stored in a struct linked from vm_private_data, along with
tracking to free that object :-\  Thanks,

Alex