Re: [PATCH v3 11/13] xen/pvcalls: implement poll command

[Stefano Stabellini <sstabellini@kernel.org>]

On Tue, 12 Sep 2017, Boris Ostrovsky wrote:
> On 09/12/2017 06:17 PM, Stefano Stabellini wrote:
> > On Tue, 12 Sep 2017, Boris Ostrovsky wrote:
> >>>>> +
> >>>>> +unsigned int pvcalls_front_poll(struct file *file, struct socket *sock,
> >>>>> +			       poll_table *wait)
> >>>>> +{
> >>>>> +	struct pvcalls_bedata *bedata;
> >>>>> +	struct sock_mapping *map;
> >>>>> +
> >>>>> +	if (!pvcalls_front_dev)
> >>>>> +		return POLLNVAL;
> >>>>> +	bedata = dev_get_drvdata(&pvcalls_front_dev->dev);
> >>>>> +
> >>>>> +	map = (struct sock_mapping *) READ_ONCE(sock->sk->sk_send_head);
> >>>> I just noticed this --- why is it READ_ONCE? Are you concerned that
> >>>> sk_send_head may change?
> >>> No, but I wanted to avoid partial reads. A caller could call
> >>> pvcalls_front_accept and pvcalls_front_poll on newsock almost at the
> >>> same time (it is probably not the correct way to use the API), I wanted
> >>> to make sure that "map" is either read correctly, or not read at all.
> >> How can you have a partial read on a pointer?
> > I don't think that the compiler makes any promises on translating a
> > pointer read into a single read instruction. Of couse, I expect gcc to
> > actually do it without any need for READ/WRITE_ONCE.
> 
> READ_ONCE() only guarantees ordering but not atomicity. It resolves (for
> 64-bit pointers) to
> 
>         case 8: *(__u64 *)res = *(volatile __u64 *)p; break;
> 
> so if compiler was breaking accesses into two then nothing would have
> prevented it from breaking them here (I don't think volatile declaration
> would affect this). Moreover, for sizes >8 bytes  READ_ONCE() is
> __builtin_memcpy() which is definitely not atomic.
> 
> So you can't rely on READ_ONCE being atomic from that perspective.

I thought that READ_ONCE guaranteed atomicity for sizes less or equal to
the machine word size. It doesn't make any atomicity guarantees for
sizes >8 bytes.


> OTOH, I am pretty sure pointer accesses are guaranteed to be atomic. For
> example, atomic64_read() is READ_ONCE(u64), which (per above) is
> dereferencing of a 64-bit pointer in C.

I am happy to remove the READ_ONCE and WRITE_ONCE, if we all think it is
safe.