RE: [PATCH] net/i40e: Fast release optimizations

[Konstantin Ananyev <konstantin.ananyev@huawei.com>]

> > > > When fast releasing mbufs, the mbufs are not accessed, so do not
> > prefetch
> > > > them.
> > > > This saves a mbuf load operation for each fast released TX mbuf.
> > > >
> > > > When fast release of mbufs is enabled for a TX queue, cache the mbuf
> > > > mempool pointer in the TX queue structure.
> > > > This saves one mbuf load operation for each burst of fast released
> > TX
> > > > mbufs.
> > > >
> > > > The txep->mbuf pointer is not used after the mbuf has been freed, so
> > do
> > > > not reset the pointer.
> > > > This saves a txep store operation for each TX mbuf freed.
> > > >
> > > > Signed-off-by: Morten Brørup <mb@smartsharesystems.com>
> > > > ---
> > > >  drivers/net/intel/common/tx.h                 |  5 +++
> > > >  .../i40e/i40e_recycle_mbufs_vec_common.c      |  4 +-
> > > >  drivers/net/intel/i40e/i40e_rxtx.c            | 39 ++++++++++------
> > ---
> > > >  3 files changed, 28 insertions(+), 20 deletions(-)
> > > >
> > > > diff --git a/drivers/net/intel/common/tx.h
> > b/drivers/net/intel/common/tx.h
> > > > index b0a68bae44..54c9b845f7 100644
> > > > --- a/drivers/net/intel/common/tx.h
> > > > +++ b/drivers/net/intel/common/tx.h
> > > > @@ -62,6 +62,11 @@ struct ci_tx_queue {
> > > >  	uint16_t tx_next_dd;
> > > >  	uint16_t tx_next_rs;
> > > >  	uint64_t offloads;
> > > > +	/* Mempool pointer for fast release of mbufs.
> > > > +	 * NULL if disabled, UINTPTR_MAX if enabled and not yet known.
> > > > +	 * Initialized at first use.
> > > > +	 */
> > > > +	struct rte_mempool *fast_free_mp;
> > > >  	uint64_t mbuf_errors;
> > > >  	rte_iova_t tx_ring_dma;        /* TX ring DMA address */
> > > >  	bool tx_deferred_start; /* don't start this queue in dev start */
> > > > diff --git a/drivers/net/intel/i40e/i40e_recycle_mbufs_vec_common.c
> > > b/drivers/net/intel/i40e/i40e_recycle_mbufs_vec_common.c
> > > > index 2875c578af..a46605cee9 100644
> > > > --- a/drivers/net/intel/i40e/i40e_recycle_mbufs_vec_common.c
> > > > +++ b/drivers/net/intel/i40e/i40e_recycle_mbufs_vec_common.c
> > > > @@ -106,7 +106,9 @@ i40e_recycle_tx_mbufs_reuse_vec(void *tx_queue,
> > > >  	if (txq->offloads & RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE) {
> > > >  		/* Avoid txq contains buffers from unexpected mempool. */
> > > >  		if (unlikely(recycle_rxq_info->mp
> > > > -					!= txep[0].mbuf->pool))
> > > > +				!= (likely(txq->fast_free_mp != (void
> > *)UINTPTR_MAX)
> > > ?
> > > > +				txq->fast_free_mp :
> > > > +				(txq->fast_free_mp = txep[0].mbuf->pool))))
> > > >  			return 0;
> > > >
> > > >  		/* Directly put mbufs from Tx to Rx. */
> > > > diff --git a/drivers/net/intel/i40e/i40e_rxtx.c
> > > b/drivers/net/intel/i40e/i40e_rxtx.c
> > > > index c3ff2e05c3..679c1340b8 100644
> > > > --- a/drivers/net/intel/i40e/i40e_rxtx.c
> > > > +++ b/drivers/net/intel/i40e/i40e_rxtx.c
> > > > @@ -1332,7 +1332,7 @@ static __rte_always_inline int
> > > >  i40e_tx_free_bufs(struct ci_tx_queue *txq)
> > > >  {
> > > >  	struct ci_tx_entry *txep;
> > > > -	uint16_t tx_rs_thresh = txq->tx_rs_thresh;
> > > > +	const uint16_t tx_rs_thresh = txq->tx_rs_thresh;
> > > >  	uint16_t i = 0, j = 0;
> > > >  	struct rte_mbuf *free[RTE_I40E_TX_MAX_FREE_BUF_SZ];
> > > >  	const uint16_t k = RTE_ALIGN_FLOOR(tx_rs_thresh,
> > > RTE_I40E_TX_MAX_FREE_BUF_SZ);
> > > > @@ -1345,41 +1345,40 @@ i40e_tx_free_bufs(struct ci_tx_queue *txq)
> > > >
> > > >  	txep = &txq->sw_ring[txq->tx_next_dd - (tx_rs_thresh - 1)];
> > > >
> > > > -	for (i = 0; i < tx_rs_thresh; i++)
> > > > -		rte_prefetch0((txep + i)->mbuf);
> > > > -
> > > >  	if (txq->offloads & RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE) {
> > > > +		struct rte_mempool * const fast_free_mp =
> > > > +				likely(txq->fast_free_mp != (void
> > *)UINTPTR_MAX) ?
> > > > +				txq->fast_free_mp :
> > > > +				(txq->fast_free_mp = txep[0].mbuf->pool);
> > > > +
> 
> Speaking about optimizations, these once-in-a-lifetime initializations of txq->fast_free_mp are the perfect candidates for a new
> superlikely() macro in <rte_branch_prediction.h>, which BTW is not only about branch prediction, but also about letting the compiler
> optimize the likely code path, e.g. by moving unlikely code away from it, thereby reducing the instruction cache pressure:
> 
> #define superlikely(x)	__builtin_expect_with_probability(!!(x), 1, 1.0)
> #define superunlikely(x)	__builtin_expect_with_probability(!!(x), 0, 1.0)
> 
> 	if (txq->offloads & RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE) {
> +		struct rte_mempool * const fast_free_mp =
> +				superlikely(txq->fast_free_mp != (void *)UINTPTR_MAX) ?
> +				txq->fast_free_mp :
> +				(txq->fast_free_mp = txep[0].mbuf->pool);
> 
> > >
> > > Nit idea.
> > > Acked-by: Konstantin Ananyev <konstantin.ananyev@huawei.com>
> > >
> > > Just as a suggestion for further improvement:
> > > can we update (& check) txq->fast_free_mp not at tx_free_bufs() time,
> > > but when we fill txep[] and filling txd[] based on mbuf values?
> > > In theory it should allow to remove the check above.
> > > Also, again in theory, it opens opportunity (with some extra effort)
> > to use
> > > similar optimization rte_mempool_put_bulk)
> > > even for cases when  RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE is not set.
> >
> > Checking the TX mbufs to determine if rte_mempool_put_bulk() can be used
> > even when the RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE offload flag is not
> > set...
> >
> > That would require that the tx_burst() function checks all the TX mbufs
> > for the RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE requirements to set the txq-
> > >fast_free_mp pointer.
> > And if the requirements are not met, it must never set the txq-
> > >fast_free_mp pointer again. Otherwise, some previously transmitted
> > mbufs, waiting to be freed, might not meet the
> > RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE requirements.

I am talking about different thing:
I think with some extra effort driver can use (in some cases)
rte_mbuf_raw_free_bulk()  even  when   RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE 
is not specified.
Let say we can make txq->fast_free_mp[] an array with the same size as txq->txep[].
At tx_burst() when filling txep[] we can do pre_free() checks for that mbuf,
and in case of success store it's mempool pointer in corresponding  txq->fast_free_mp[],
otherwise put NULL there.
Then at tx_free() we can scan fast_free_mp[] and invoke   raw_free() for non-NULL entries.
Again, for now it is just an idea probably worth to think about.  

> > Relying on the explicit RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE offload flag
> > (here or elsewhere) only requires one check (of the offload flag and/or
> > the mempool pointer (whichever is hotter in cache)) per burst of
> > packets.
> >
> > PS: The drivers really should be using the new rte_mbuf_raw_free_bulk()
> > instead of rte_mempool_put_bulk(), so the freed mbufs are sanity checked
> > in RTE_LIBRTE_MBUF_DEBUG mode. But such changes belong in another patch
> > series.
> >
> > >
> > >
> > > >  		if (k) {
> > > >  			for (j = 0; j != k; j += RTE_I40E_TX_MAX_FREE_BUF_SZ)
> > {
> > > > -				for (i = 0; i < RTE_I40E_TX_MAX_FREE_BUF_SZ;
> > ++i,
> > > ++txep) {
> > > > +				for (i = 0; i < RTE_I40E_TX_MAX_FREE_BUF_SZ;
> > ++i,
> > > ++txep)
> > > >  					free[i] = txep->mbuf;
> > > > -					txep->mbuf = NULL;
> > > > -				}
> > > > -				rte_mempool_put_bulk(free[0]->pool, (void
> > **)free,
> > > > +				rte_mempool_put_bulk(fast_free_mp, (void
> > **)free,
> > > >  						RTE_I40E_TX_MAX_FREE_BUF_SZ);
> > > >  			}
> > > >  		}
> > > >
> > > >  		if (m) {
> > > > -			for (i = 0; i < m; ++i, ++txep) {
> > > > +			for (i = 0; i < m; ++i, ++txep)
> > > >  				free[i] = txep->mbuf;
> > > > -				txep->mbuf = NULL;
> > > > -			}
> > > > -			rte_mempool_put_bulk(free[0]->pool, (void **)free,
> > m);
> > > > +			rte_mempool_put_bulk(fast_free_mp, (void **)free, m);
> > > >  		}
> > > >  	} else {
> > > > -		for (i = 0; i < txq->tx_rs_thresh; ++i, ++txep) {
> > > > +		for (i = 0; i < tx_rs_thresh; i++)
> > > > +			rte_prefetch0((txep + i)->mbuf);
> > > > +
> > > > +		for (i = 0; i < tx_rs_thresh; ++i, ++txep)
> > > >  			rte_pktmbuf_free_seg(txep->mbuf);
> > > > -			txep->mbuf = NULL;
> > > > -		}
> > > >  	}
> > > >
> > > > -	txq->nb_tx_free = (uint16_t)(txq->nb_tx_free + txq->tx_rs_thresh);
> > > > -	txq->tx_next_dd = (uint16_t)(txq->tx_next_dd + txq->tx_rs_thresh);
> > > > +	txq->nb_tx_free = (uint16_t)(txq->nb_tx_free + tx_rs_thresh);
> > > > +	txq->tx_next_dd = (uint16_t)(txq->tx_next_dd + tx_rs_thresh);
> > > >  	if (txq->tx_next_dd >= txq->nb_tx_desc)
> > > > -		txq->tx_next_dd = (uint16_t)(txq->tx_rs_thresh - 1);
> > > > +		txq->tx_next_dd = (uint16_t)(tx_rs_thresh - 1);
> > > >
> > > > -	return txq->tx_rs_thresh;
> > > > +	return tx_rs_thresh;
> > > >  }
> > > >
> > > >  /* Populate 4 descriptors with data from 4 mbufs */
> > > > @@ -2546,6 +2545,8 @@ i40e_dev_tx_queue_setup(struct rte_eth_dev
> > *dev,
> > > >  	txq->reg_idx = reg_idx;
> > > >  	txq->port_id = dev->data->port_id;
> > > >  	txq->offloads = offloads;
> > > > +	txq->fast_free_mp = offloads & RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE ?
> > > > +			(void *)UINTPTR_MAX : NULL;
> > > >  	txq->i40e_vsi = vsi;
> > > >  	txq->tx_deferred_start = tx_conf->tx_deferred_start;
> > > >
> > > > --
> > > > 2.43.0