[Differential] [Request, 33 lines] D5416: buf_ring/drbr: Add buf_ring_peek_clear_sc and use it in drbr_peek

[sepherosa_gmail.com (Sepherosa Ziehau)]

sepherosa_gmail.com created this revision.
sepherosa_gmail.com added reviewers: network, adrian, delphij, rrs, glebius, kmacy, decui_microsoft.com, honzhan_microsoft.com, howard0su_gmail.com.
sepherosa_gmail.com added a subscriber: freebsd-net-list.

REVISION SUMMARY
  Unlike buf_ring_peek, it only supports single consumer mode, and it clears the cons_head if DEBUG_BUFRING/INVARIANTS is defined.
  
  The normal use case of drbr_peek for network drivers is:
  
    m = drbr_peek(br);
    err = hw_spec_encap(&m); /* could m_defrag/m_collapse */
    (*)
    if (err) {
        if (m == NULL)
            drbr_advance(br);
        else
            drbr_putback(br, m);
        /* break the loop */
    }
    drbr_advance(br);
  
  The race is:
  If hw_spec_encap() m_defrag or m_collapse the mbuf, i.e. the old mbuf was freed, or like the Hyper-V's network driver, that transmission-done does not even require the TX lock; then on the other CPU at the (*) time, the freed mbuf could be recycled and being drbr_enqueue even before the current CPU had the chance to call drbr_{advance,putback}.  This triggers a panic in drbr_enqueue duplicated element check, if DEBUG_BUFRING/INVARIANTS is defined.
  
  Use buf_ring_peek_clear_sc() in drbr_peek() to fix the above race.
  
  This change is a NO-OP, if neither DEBUG_BUFRING nor INVARIANTS are defined.

REVISION DETAIL
  https://reviews.freebsd.org/D5416

AFFECTED FILES
  sys/net/ifq.h
  sys/sys/buf_ring.h

CHANGE DETAILS
  diff --git a/sys/sys/buf_ring.h b/sys/sys/buf_ring.h
  --- a/sys/sys/buf_ring.h
  +++ b/sys/sys/buf_ring.h
  @@ -268,6 +268,37 @@
   	return (br->br_ring[br->br_cons_head]);
   }
   
  +static __inline void *
  +buf_ring_peek_clear_sc(struct buf_ring *br)
  +{
  +#ifdef DEBUG_BUFRING
  +	void *ret;
  +
  +	if (!mtx_owned(br->br_lock))
  +		panic("lock not held on single consumer dequeue");
  +#endif	
  +	/*
  +	 * I believe it is safe to not have a memory barrier
  +	 * here because we control cons and tail is worst case
  +	 * a lagging indicator so we worst case we might
  +	 * return NULL immediately after a buffer has been enqueued
  +	 */
  +	if (br->br_cons_head == br->br_prod_tail)
  +		return (NULL);
  +
  +#ifdef DEBUG_BUFRING
  +	/*
  +	 * Single consumer, i.e. cons_head will not move while we are
  +	 * running, so atomic_swap_ptr() is not necessary here.
  +	 */
  +	ret = br->br_ring[br->br_cons_head];
  +	br->br_ring[br->br_cons_head] = NULL;
  +	return (ret);
  +#else
  +	return (br->br_ring[br->br_cons_head]);
  +#endif
  +}
  +
   static __inline int
   buf_ring_full(struct buf_ring *br)
   {
  diff --git a/sys/net/ifq.h b/sys/net/ifq.h
  --- a/sys/net/ifq.h
  +++ b/sys/net/ifq.h
  @@ -369,7 +369,7 @@
   		return (m);
   	}
   #endif
  -	return(buf_ring_peek(br));
  +	return(buf_ring_peek_clear_sc(br));
   }
   
   static __inline void

EMAIL PREFERENCES
  https://reviews.freebsd.org/settings/panel/emailpreferences/

To: sepherosa_gmail.com, network, adrian, delphij, rrs, glebius, kmacy, decui_microsoft.com, honzhan_microsoft.com, howard0su_gmail.com
Cc: freebsd-net-list
-------------- next part --------------
A non-text attachment was scrubbed...
Name: D5416.13669.patch
Type: text/x-patch
Size: 1271 bytes
Desc: not available
URL: <http://lists.freebsd.org/pipermail/freebsd-net/attachments/20160224/fc46288d/attachment.bin>