Re: [dpdk-dev] [PATCH 2/2] ixgbe: don't override mbuf buffer length

[Bruce Richardson <bruce.richardson@intel.com>]

On Thu, Dec 04, 2014 at 05:17:16PM +0000, Ananyev, Konstantin wrote:
> 
> 
> > -----Original Message-----
> > From: Richardson, Bruce
> > Sent: Thursday, December 04, 2014 4:59 PM
> > To: Ananyev, Konstantin
> > Cc: Thomas Monjalon; dev@dpdk.org
> > Subject: Re: [dpdk-dev] [PATCH 2/2] ixgbe: don't override mbuf buffer length
> > 
> > On Thu, Dec 04, 2014 at 04:18:03PM +0000, Ananyev, Konstantin wrote:
> > >
> > >
> > > > -----Original Message-----
> > > > From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Ananyev, Konstantin
> > > > Sent: Thursday, December 04, 2014 4:05 PM
> > > > To: Thomas Monjalon
> > > > Cc: dev@dpdk.org
> > > > Subject: Re: [dpdk-dev] [PATCH 2/2] ixgbe: don't override mbuf buffer length
> > > >
> > > >
> > > >
> > > > > -----Original Message-----
> > > > > From: Thomas Monjalon [mailto:thomas.monjalon@6wind.com]
> > > > > Sent: Thursday, December 04, 2014 3:48 PM
> > > > > To: Ananyev, Konstantin
> > > > > Cc: dev@dpdk.org; Richardson, Bruce
> > > > > Subject: Re: [dpdk-dev] [PATCH 2/2] ixgbe: don't override mbuf buffer length
> > > > >
> > > > > 2014-12-04 15:29, Ananyev, Konstantin:
> > > > > > From: Richardson, Bruce
> > > > > > > On Thu, Dec 04, 2014 at 02:50:11PM +0000, Ananyev, Konstantin wrote:
> > > > > > > > From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Jean-Mickael Guerin
> > > > > > > > > The template mbuf_initializer is hard coded with a buflen which
> > > > > > > > > might have been set differently by the application at the time of
> > > > > > > > > mbuf pool creation.
> > > > > > > > >
> > > > > > > > > Switch to a mbuf allocation, to fetch the correct default values.
> > > > > > > > > There is no performance impact because this is not a data-plane API.
> > > > > > > > >
> > > > > > > > > Signed-off-by: Jean-Mickael Guerin <jean-mickael.guerin@6wind.com>
> > > > > > > > > Acked-by: David Marchand <david.marchand@6wind.com>
> > > > > > > > > Fixes: 0ff3324da2 ("ixgbe: rework vector pmd following mbuf changes")
> > > > > > > > > ---
> > > > > > > > >  lib/librte_pmd_ixgbe/ixgbe_rxtx_vec.c | 19 ++++++++++++-------
> > > > > > > > >  1 file changed, 12 insertions(+), 7 deletions(-)
> > > > > > > > >
> > > > > > > > > diff --git a/lib/librte_pmd_ixgbe/ixgbe_rxtx_vec.c b/lib/librte_pmd_ixgbe/ixgbe_rxtx_vec.c
> > > > > > > > > index c1b5a78..f7b02f5 100644
> > > > > > > > > --- a/lib/librte_pmd_ixgbe/ixgbe_rxtx_vec.c
> > > > > > > > > +++ b/lib/librte_pmd_ixgbe/ixgbe_rxtx_vec.c
> > > > > > > > > @@ -732,17 +732,22 @@ static struct ixgbe_txq_ops vec_txq_ops = {
> > > > > > > > >  int
> > > > > > > > >  ixgbe_rxq_vec_setup(struct igb_rx_queue *rxq)
> > > > > > > > >  {
> > > > > > > > > -	struct rte_mbuf mb_def = { .buf_addr = 0 }; /* zeroed mbuf */
> > > > > > > > > +	struct rte_mbuf *mb_def;
> > > > > > > > >
> > > > > > > > > -	mb_def.nb_segs = 1;
> > > > > > > > > -	mb_def.data_off = RTE_PKTMBUF_HEADROOM;
> > > > > > > > > -	mb_def.buf_len = rxq->mb_pool->elt_size - sizeof(struct rte_mbuf);
> > > > > > > > > -	mb_def.port = rxq->port_id;
> > > > > > > > > -	rte_mbuf_refcnt_set(&mb_def, 1);
> > > > > > > > > +	mb_def = rte_pktmbuf_alloc(rxq->mb_pool);
> > > > > > > >
> > > > > > > > Could you explain to me, what is an advantage of using dynamic allocation vs local struct here?
> > > > > > > > I don't see any.
> > > > > > >
> > > > > > > It means that we get an mbuf that is initialized as done by the initialization
> > > > > > > function passed to the mempool_create call. The static variable method assumes
> > > > > > > that we configure the mbuf using default setting for buf_len etc.
> > > > > >
> > > > > > I understand that, but  why it can't be done in some other way?
> > > > > > Without allocating/freeing?
> > > > > > Let say, at mempool_create() store obj_init() and then add ability to call it again?
> > > > > > Anyway, it doesn't look to me like a critical problem, that requires an urgent patch for 1.8.
> > > > >
> > > > > Konstantin, when a bug is seen, it must be fixed ASAP.
> > > >
> > > > Well, it will be exposed only if someone will use a custom mbufs right?
> > > > I.e, the se 2 lines would not be correct:
> > > > mb_def.data_off = RTE_PKTMBUF_HEADROOM;
> > > > mb_def.buf_len = rxq->mb_pool->elt_size - sizeof(struct rte_mbuf);
> > > >
> > > > Thoug we setup same data_off  like that in all other PMDs as well.
> > > > Something like that:
> > > > m->data_off = RTE_PKTMBUF_HEADROOM;
> > > > could be seen across all RX functions we have for different PMDs.
> > > >
> > > > The only difference is buf_len, but in theory even with dynamic allocation,
> > > > the fix would be totally correct.
> > > > As no one can guarantee, that with custom mbufs, all buffers inside the pool will have the same length.
> > >
> > > Which makes me think, that we probably shouldn't overwrite buf_len by rxq->mbuf_initializer.
> > >
> > I believe that it is perfectly safe to do so. All buffers from a mempool are meant
> > to be the same size, therefore reading the length of one buffer should tell you
> > what size all buffers are.
> 
> Yes, objects in the mempool are the same size
> But nothing prevents you, in your custom obj_init() to setup mbuf->buf_len to some other value,
> that could be smaller, then mempool element size.
> Let say you'd like all your mbufs from particular mempool to be at least 2K long and 1K aligned.
> So you set RTE_MBUF_REFCNT_ATOMIC=n, and call  rte_mempool_create(...,elst_size=0xc00,...);
> Then at you custom obj_init() you do:
> 
> struct rte_mbuf *m =  RTE_ALIGN_CEIL(_m, 1024);
> buf_len = mp->elt_size - sizeof(struct rte_mbuf) - (m - _m);
> ...
> 
> From my point, nothing wrong is done here, and we have a mempool where mbufs might have different buf_len.
> 
> Another example, is attachment of external buffer to the mbuf.
> We are doing it to support zero-copy inside our vhost app.
> Right now we don't allow external buffer length be bigger then mbuf buf_len, but again some people may like to allow that.
> 
Would the originally proposed solution not work in all these cases - assuming it's
enhanced to catch and properly handle failure cases?

> > If we do hit a scenario where we do need to support
> > variable size buffers from a single mempool, we can do that via the older unoptimized
> > code paths, I think, since it's a definite edge case.
> > 
> > /Bruce
>