Re: [dpdk-dev] [PATCH] mbuf: add comment explaining confusing code

[Neil Horman <nhorman@tuxdriver.com>]

On Mon, Mar 30, 2015 at 09:39:13PM +0200, Marc Sune wrote:
> 
> 
> On 27/03/15 15:30, Bruce Richardson wrote:
> >On Fri, Mar 27, 2015 at 10:07:35AM -0400, Neil Horman wrote:
> >>On Fri, Mar 27, 2015 at 11:32:38AM +0000, Bruce Richardson wrote:
> >>>On Fri, Mar 27, 2015 at 06:29:56AM -0400, Neil Horman wrote:
> >>>>On Thu, Mar 26, 2015 at 09:14:54PM +0000, Bruce Richardson wrote:
> >>>>>The logic used in the condition check before freeing an mbuf is
> >>>>>sometimes confusing, so explain it in a proper comment.
> >>>>>
> >>>>>Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> >>>>>---
> >>>>>  lib/librte_mbuf/rte_mbuf.h | 10 ++++++++++
> >>>>>  1 file changed, 10 insertions(+)
> >>>>>
> >>>>>diff --git a/lib/librte_mbuf/rte_mbuf.h b/lib/librte_mbuf/rte_mbuf.h
> >>>>>index 17ba791..0265172 100644
> >>>>>--- a/lib/librte_mbuf/rte_mbuf.h
> >>>>>+++ b/lib/librte_mbuf/rte_mbuf.h
> >>>>>@@ -764,6 +764,16 @@ __rte_pktmbuf_prefree_seg(struct rte_mbuf *m)
> >>>>>  {
> >>>>>  	__rte_mbuf_sanity_check(m, 0);
> >>>>>+	/*
> >>>>>+	 * Check to see if this is the last reference to the mbuf.
> >>>>>+	 * Note: the double check here is deliberate. If the ref_cnt is "atomic"
> >>>>>+	 * the call to "refcnt_update" is a very expensive operation, so we
> >>>>>+	 * don't want to call it in the case where we know we are the holder
> >>>>>+	 * of the last reference to this mbuf i.e. ref_cnt == 1.
> >>>>>+	 * If however, ref_cnt != 1, it's still possible that we may still be
> >>>>>+	 * the final decrementer of the count, so we need to check that
> >>>>>+	 * result also, to make sure the mbuf is freed properly.
> >>>>>+	 */
> >>>>>  	if (likely (rte_mbuf_refcnt_read(m) == 1) ||
> >>>>>  			likely (rte_mbuf_refcnt_update(m, -1) == 0)) {
> >>>>>-- 
> >>>>>2.1.0
> >>>>>
> >>>>>
> >>>>NAK
> >>>>  the comment is incorrect, a return code of 1 from rte_mbuf_refcnt_read doesn't
> >>>>guarantee you are the last holder of the buffer if two contexts have a pointer
> >>>>to it.
> >>>If two threads have pointers to it, and are both going to free it, the refcnt
> >>>must be 2 not one, otherwise the refcnt is meaningless.
> >>>
> >>What about the other concrete case that I illustrated, where one context is
> >>attempting to increment the refcount, while the other is decrementing it with
> >>the intention to free?  By making the read and set operation disctinct here
> >>you've broken the atomicity of the read and update logic that atomics are there
> >>for and created a race condition.  I don't know how else to explain this to you.
> >>if(atomic_read == 1) then atomic_set(0), breaks the entire notion of what
> >>atomics are meant to do (namely update and read state as an atomic unit), you
> >>just can't get away with not having that atomicity here.  If you could, you
> >>might as well be using plain integers for the reference count, as you're not
> >>using the atomic properties of the type.
> >>
> >>Neil
> >I disagree.
> >
> >A value of one, indicates that there is only one owner of the mbuf,
> >and therefore since we are in the free routine, we are that owner.
> 
> This is true if you assume the application is bug-free and two threads will
> never hold mistakenly the same mbuf pointer with ref_cnt=1.  That's ok in
> general, but if it doesn't, debugging it can be pretty painful.
> 
> Marc
> 
The use case for the above is situations in which mbufs are stored/queued and
must be looked up prior to taking a hold on the refcnt (i.e. you have to get the
pointer so that you can pass it in to rte_pktmbuf_refcnt_inc or whtever
increases the hold count.  I don't think that use case is considered here
though, in that it presumes that in a provider/receiver context relationship,
the provider always increases the refcount before giving it to the receiver.  I
don't care for it either, but it is what it is.
Neil

> >  If there
> >are to be two owners, the refcnt must be incremented before handing over the
> >pointer to the other thread - to get to the example you make. If that does not
> >occur, we can also have the situation where the "sending" thread calls
> >free - and therefore this function - before the other thread receives the
> >pointer. In that case, we will have the receiving thread getting a pointer
> >to an mbuf which is now invalid as it has been put back into the mempool
> >
> >Again, in short, if refcnt == 1, there is only one mbuf owner. If refcnt == 1
> >and we are currently executing in prefree_seg, we are the owner and no other
> >thread is allow to muck about with the mbuf.
> >
> >/Bruce
> >
> 
>