Re: [PATCH] fuse: virtio_fs: add checks for FUSE protocol compliance

On Tue, Oct 28, 2025 at 11:01:08PM -0700, Darrick J. Wong wrote:
> On Wed, Oct 29, 2025 at 08:58:30AM +0530, Brajesh Patil wrote:
> > On Tue, Oct 28, 2025 at 01:07:55PM -0700, Darrick J. Wong wrote:
> > > On Wed, Oct 29, 2025 at 01:33:11AM +0530, Brajesh Patil wrote:
> > > > Add validation in virtio-fs to ensure the server follows the FUSE
> > > > protocol for response headers, addressing the existing TODO for
> > > > verifying protocol compliance.
> > > > 
> > > > Add checks for fuse_out_header to verify:
> > > >  - oh->unique matches req->in.h.unique
> > > >  - FUSE_INT_REQ_BIT is not set
> > > >  - error codes are valid
> > > >  - oh->len does not exceed the expected size
> > > > 
> > > > Signed-off-by: Brajesh Patil <[email protected]>
> > > > ---
> > > >  fs/fuse/virtio_fs.c | 30 +++++++++++++++++++++++++-----
> > > >  1 file changed, 25 insertions(+), 5 deletions(-)
> > > > 
> > > > diff --git a/fs/fuse/virtio_fs.c b/fs/fuse/virtio_fs.c
> > > > index 6bc7c97b017d..52e8338bf436 100644
> > > > --- a/fs/fuse/virtio_fs.c
> > > > +++ b/fs/fuse/virtio_fs.c
> > > > @@ -764,14 +764,34 @@ static void virtio_fs_request_complete(struct fuse_req *req,
> > > >  {
> > > >  	struct fuse_args *args;
> > > >  	struct fuse_args_pages *ap;
> > > > -	unsigned int len, i, thislen;
> > > > +	struct fuse_out_header *oh;
> > > > +	unsigned int len, i, thislen, expected_len = 0;
> > > >  	struct folio *folio;
> > > >  
> > > > -	/*
> > > > -	 * TODO verify that server properly follows FUSE protocol
> > > > -	 * (oh.uniq, oh.len)
> > > > -	 */
> > > > +	oh = &req->out.h;
> > > > +
> > > > +	if (oh->unique == 0)
> > > > +		pr_warn_once("notify through fuse-virtio-fs not supported");
> > > > +
> > > > +	if ((oh->unique & ~FUSE_INT_REQ_BIT) != req->in.h.unique)
> > > > +		pr_warn_ratelimited("virtio-fs: unique mismatch, expected: %llu got %llu\n",
> > > > +				    req->in.h.unique, oh->unique & ~FUSE_INT_REQ_BIT);
> > > 
> > > Er... shouldn't these be rejecting the response somehow?  Instead of
> > > warning that something's amiss but continuing with known bad data?
> > > 
> > > --D
> > >
> > 
> > Right, continuing here is unsafe.
> > 
> > I plan to update the code so that in case of any header validation
> > failure (e.g. unique mismatch, invalid error, length mismatch), it
> > should skip copying data and jump directly to the section that marks
> > request as complete
> > 
> > Does this seem like a feasible approach?
> 
> Yeah, I think you can just set req->out.h.error to some errno (EIO?) and
> jump to fuse_request_end, sort of like what fuse_dev_do_write sort of
> does.  I think that sends the errno back to whatever code initiated the
> request.  I don't know if virtiofs should be throwing an error back to
> the server?
> 
> --D
> 

I think it is okay to set oh.error in fuse_dev_do_write as it is a server side
reply. But as virtio_fs is on the client side and oh.error has been set by
virtiofsd, I think so we should not overwrite oh.error. Instead, if we
encounter an error in any of the if conditions, skip copying arguments and jump
to the line clear_bit(FR_SENT, &req->flags).

> > > > +
> > > > +	WARN_ON_ONCE(oh->unique & FUSE_INT_REQ_BIT);
> > > > +
> > > > +	if (oh->error <= -ERESTARTSYS || oh->error > 0)
> > > > +		pr_warn_ratelimited("virtio-fs: invalid error code from server: %d\n",
> > > > +				    oh->error);
> > > > +
> > > >  	args = req->args;
> > > > +
> > > > +	for (i = 0; i < args->out_numargs; i++)
> > > > +		expected_len += args->out_args[i].size;
> > > > +
> > > > +	if (oh->len > sizeof(*oh) + expected_len)
> > > > +		pr_warn("FUSE reply too long! got=%u expected<=%u\n",
> > > > +			oh->len, (unsigned int)(sizeof(*oh) + expected_len));
> > > > +
> > > >  	copy_args_from_argbuf(args, req);
> > > >  
> > > >  	if (args->out_pages && args->page_zeroing) {
> > > > -- 
> > > > 2.43.0
> > > > 
> > > >