Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Mon, 28 Jun 2021 12:36:21 -0600
From:   Alex Williamson <alex.williamson@redhat.com>
To:     Jason Gunthorpe <jgg@nvidia.com>
Cc:     kvm@vger.kernel.org, linux-kernel@vger.kernel.org,
        peterx@redhat.com, prime.zeng@hisilicon.com, cohuck@redhat.com
Subject: Re: [PATCH v2] vfio/pci: Handle concurrent vma faults
Message-ID: <20210628123621.7fd36a1b.alex.williamson@redhat.com>
In-Reply-To: <20210628173028.GF4459@nvidia.com>
References: <161540257788.10151.6284852774772157400.stgit@gimli.home>
        <20210628104653.4ca65921.alex.williamson@redhat.com>
        <20210628173028.GF4459@nvidia.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit
Precedence: bulk

On Mon, 28 Jun 2021 14:30:28 -0300
Jason Gunthorpe <jgg@nvidia.com> wrote:

> On Mon, Jun 28, 2021 at 10:46:53AM -0600, Alex Williamson wrote:
> > On Wed, 10 Mar 2021 11:58:07 -0700
> > Alex Williamson <alex.williamson@redhat.com> wrote:
> >   
> > > 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.
 
> > IIRC, there were no blocking issues on this patch as an interim fix to
> > resolve the concurrent fault issues with io_remap_pfn_range().
> > Unfortunately it also got no Reviewed-by or Tested-by feedback.  I'd
> > like to put this in for v5.14 (should have gone in earlier).  Any final
> > comments?  Thanks,  
> 
> 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,

Alex