Date: Fri, 08 May 2015 12:35:54 +0800 From: rhenjau <rhenjau@gmail.com> To: Ian Lepore <ian@freebsd.org> Cc: freebsd-hackers@freebsd.org Subject: Re: Memory barriers about buf_ring(9) Message-ID: <554C3D2A.7030909@gmail.com> In-Reply-To: <1431008458.6170.165.camel@freebsd.org> References: <CA%2BO7MXzRuy33tWQRWAnh3Z6ZsdLOsVvm0-1p8%2BjrfFOannjZbQ@mail.gmail.com> <1431008458.6170.165.camel@freebsd.org>
next in thread | previous in thread | raw e-mail | index | archive | help
Hi, Ian
Thank you for your information.
On 2015/5/7 22:20, Ian Lepore wrote:
> On Thu, 2015-05-07 at 18:10 +0800, rhenjau wrote:
>> Hi, hackers
>> I'm reading buf_ring(9) and have a question about the following
>> function. Is a memory barrier needed before br_cons_tail to prevent the
>> loading of br_ring[cons_head] load is reordered after br_cons_tail is set?
>> Producers may overwrite the area.
>>
>> /*
>> * single-consumer dequeue
>> * use where dequeue is protected by a lock
>> * e.g. a network driver's tx queue lock
>> */
>> static __inline void *
>> buf_ring_dequeue_sc(struct buf_ring *br)
>> {
>> uint32_t cons_head, cons_next;
>> #ifdef PREFETCH_DEFINED
>> uint32_t cons_next_next;
>> #endif
>> uint32_t prod_tail;
>> void *buf;
>>
>> cons_head = br->br_cons_head;
>> prod_tail = br->br_prod_tail;
>>
>> cons_next = (cons_head + 1) & br->br_cons_mask;
>> #ifdef PREFETCH_DEFINED
>> cons_next_next = (cons_head + 2) & br->br_cons_mask;
>> #endif
>>
>> if (cons_head == prod_tail)
>> return (NULL);
>>
>> #ifdef PREFETCH_DEFINED
>> if (cons_next != prod_tail) {
>> prefetch(br->br_ring[cons_next]);
>> if (cons_next_next != prod_tail)
>> prefetch(br->br_ring[cons_next_next]);
>> }
>> #endif
>> br->br_cons_head = cons_next;
>> buf = br->br_ring[cons_head];
>>
>> #ifdef DEBUG_BUFRING
>> br->br_ring[cons_head] = NULL;
>> if (!mtx_owned(br->br_lock))
>> panic("lock not held on single consumer dequeue");
>> if (br->br_cons_tail != cons_head)
>> panic("inconsistent list cons_tail=%d cons_head=%d",
>> br->br_cons_tail, cons_head);
>> #endif
>> ----------------------------------------------------- need memory
>> barrier?
>> br->br_cons_tail = cons_next;
>> return (buf);
>> }
>>
> There was some discussion and changes relating to barriers in buf_ring,
> but it looks like the changes never got committed (I don't know why).
>
> https://reviews.freebsd.org/D1945
>
> -- Ian
>
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?554C3D2A.7030909>