Re: [PATCH v3] pvcalls-front: Avoid __get_free_pages(GFP_KERNEL) under spinlock

[Stefano Stabellini]

On Fri, 30 Nov 2018, Boris Ostrovsky wrote:
> On 11/30/18 6:01 AM, Wen Yang wrote:
> > The problem is that we call this with a spin lock held.
> > The call tree is:
> > pvcalls_front_accept() holds bedata->socket_lock.
> >     -> create_active()
> >         -> __get_free_pages() uses GFP_KERNEL
> >
> > The create_active() function is only called from pvcalls_front_accept()
> > with a spin_lock held, The allocation is not allowed to sleep and
> > GFP_KERNEL is not sufficient.
> >
> > This issue was detected by using the Coccinelle software.
> >
> > v2: Add a function doing the allocations which is called
> >     outside the lock and passing the allocated data to
> >     create_active().
> > v3: Use the matching deallocators i.e., free_page()
> >     and free_pages(), respectively.
> >
> > Suggested-by: Juergen Gross <jgross@xxxxxxxx>
> > Signed-off-by: Wen Yang <wen.yang99@xxxxxxxxxx>
> > CC: Julia Lawall <julia.lawall@xxxxxxx>
> > CC: Boris Ostrovsky <boris.ostrovsky@xxxxxxxxxx>
> > CC: Juergen Gross <jgross@xxxxxxxx>
> > CC: Stefano Stabellini <sstabellini@xxxxxxxxxx>
> > CC: xen-devel@xxxxxxxxxxxxxxxxxxxx
> > CC: linux-kernel@xxxxxxxxxxxxxxx
> > ---
> >  drivers/xen/pvcalls-front.c | 71 ++++++++++++++++++++++++++++++-------
> >  1 file changed, 59 insertions(+), 12 deletions(-)
> >
> > diff --git a/drivers/xen/pvcalls-front.c b/drivers/xen/pvcalls-front.c
> > index 2f11ca72a281..a26f416daf46 100644
> > --- a/drivers/xen/pvcalls-front.c
> > +++ b/drivers/xen/pvcalls-front.c
> > @@ -335,7 +335,43 @@ int pvcalls_front_socket(struct socket *sock)
> >  	return ret;
> >  }
> >  
> > -static int create_active(struct sock_mapping *map, int *evtchn)
> > +struct sock_mapping_active_ring {
> > +	struct pvcalls_data_intf *ring;
> > +	RING_IDX ring_order;
> > +	void *bytes;
> > +};
> > +
> > +static int alloc_active_ring(struct sock_mapping_active_ring *active_ring)
> > +{
> > +	active_ring->ring = NULL;
> 
> This is not necessary.
> 
> > +	active_ring->bytes = NULL;
> > +
> > +	active_ring->ring = (struct pvcalls_data_intf *)
> > +		__get_free_page(GFP_KERNEL | __GFP_ZERO);
> > +	if (active_ring->ring == NULL)
> > +		goto out_error;
> > +	active_ring->ring_order = PVCALLS_RING_ORDER;
> > +	active_ring->bytes = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
> > +			PVCALLS_RING_ORDER);
> > +	if (active_ring->bytes == NULL)
> > +		goto out_error;
> > +
> > +	return 0;
> > +
> > +out_error:
> > +	free_pages((unsigned long)active_ring->bytes, active_ring->ring_order);
> > +	free_page((unsigned long)active_ring->ring);
> > +	return -ENOMEM;
> > +}
> > +
> 
> 
> 
> > @@ -397,6 +427,7 @@ int pvcalls_front_connect(struct socket *sock, struct sockaddr *addr,
> >  	struct sock_mapping *map = NULL;
> >  	struct xen_pvcalls_request *req;
> >  	int notify, req_id, ret, evtchn;
> > +	struct sock_mapping_active_ring active_ring;
> >  
> >  	if (addr->sa_family != AF_INET || sock->type != SOCK_STREAM)
> >  		return -EOPNOTSUPP;
> > @@ -406,15 +437,21 @@ int pvcalls_front_connect(struct socket *sock, struct sockaddr *addr,
> >  		return PTR_ERR(map);
> >  
> >  	bedata = dev_get_drvdata(&pvcalls_front_dev->dev);
> > +	ret = alloc_active_ring(&active_ring);
> 
> Why not just alloc_active_ring(map)?

Yes, I think it would be better to pre-populate map (struct
sock_mapping), rather than introducing one more new struct (struct
sock_mapping_active_ring).