Re: read_barrier_depends() usage in vhost.c

From: Will Deacon
Date: Wed Oct 16 2019 - 19:36:09 EST


[Bah: I typoed the LKML address, so I've fixed it for this one]

On Thu, Oct 17, 2019 at 12:33:40AM +0100, Will Deacon wrote:
> Hi all,
>
> In an attempt to remove the remaining traces of [smp_]read_barrier_depends()
> following my previous patches to strengthen READ_ONCE() for Alpha [1], I
> ended up trying to decipher the read_barrier_depends() usage in the vhost
> driver:
>
> --->8
>
> // drivers/vhost/vhost.c
> static int get_indirect(struct vhost_virtqueue *vq,
> struct iovec iov[], unsigned int iov_size,
> unsigned int *out_num, unsigned int *in_num,
> struct vhost_log *log, unsigned int *log_num,
> struct vring_desc *indirect)
> {
> [...]
>
> /* We will use the result as an address to read from, so most
> * architectures only need a compiler barrier here. */
> read_barrier_depends();
>
> --->8
>
> Unfortunately, although the barrier is commented (hurrah!), it's not
> particularly enlightening about the accesses making up the dependency
> chain, and I don't understand the supposed need for a compiler barrier
> either (read_barrier_depends() doesn't generally provide this).
>
> Does anybody know which accesses are being ordered here? Usually you'd need
> a READ_ONCE()/rcu_dereference() beginning the chain, but I haven't managed
> to find one...
>
> Thanks,
>
> Will
>
> [1] c2bc66082e10 ("locking/barriers: Add implicit smp_read_barrier_depends() to READ_ONCE()")