[PATCH v2] net/intel: optimize for fast-free hint

[Bruce Richardson]

On Wed, Apr 08, 2026 at 09:27:11PM +0200, Morten Brørup wrote:
> > From: Bruce Richardson [mailto:bruce.richardson at intel.com]
> > Sent: Wednesday, 8 April 2026 15.25
> > 
> > When the fast-free hint is provided to the driver we know that the
> > mbufs
> > have refcnt of 1 and are from the same mempool. Therefore, we can
> > optimize a bit for this case by:
> > 
> > * resetting the necessary mbuf fields, ie. nb_seg and next pointer when
> >   we are accessing the mbuf on writing the descriptor.
> > * on cleanup of buffers after transmit, we can just write those buffers
> >   straight to the mempool without accessing them.
> > 
> > Signed-off-by: Bruce Richardson <bruce.richardson at intel.com>
> 
> A bunch of review thoughts inline below.
> The ones regarding instrumentation should be fixed.
> The rest might be irrelevant and/or nonsense.
>

See inline below. Summary:
* ack for mempool functions for instrumentation
* ack for referencing the mp pointer rather than the offload flags
* nak for setting the mbuf pointers for null, it's unnecessary with this
  design.
* ack for removing the prefetches

Thanks for the detailed review. Preparing v3 now.

/Bruce
 
> > ---
> > V2: Fix issues with original submission:
> > * missed check for NULL mbufs
> > * fixed issue with freeing directly from sw_ring in scalar path which
> >   doesn't work as thats not a flag array of pointers
> > * fixed missing null assignment in case of large segments for TSO
> > ---
> >  drivers/net/intel/common/tx.h        | 21 ++++--
> >  drivers/net/intel/common/tx_scalar.h | 95 ++++++++++++++++++++++------
> >  2 files changed, 90 insertions(+), 26 deletions(-)
> > 
> > diff --git a/drivers/net/intel/common/tx.h
> > b/drivers/net/intel/common/tx.h
> > index 283bd58d5d..f2123f069c 100644
> > --- a/drivers/net/intel/common/tx.h
> > +++ b/drivers/net/intel/common/tx.h
> > @@ -363,13 +363,22 @@ ci_txq_release_all_mbufs(struct ci_tx_queue *txq,
> > bool use_ctx)
> >  		return;
> > 
> >  	if (!txq->use_vec_entry) {
> > -		/* Regular scalar path uses sw_ring with ci_tx_entry */
> > -		for (uint16_t i = 0; i < txq->nb_tx_desc; i++) {
> > -			if (txq->sw_ring[i].mbuf != NULL) {
> > -				rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
> > -				txq->sw_ring[i].mbuf = NULL;
> > -			}
> > +		/* Free mbufs from (last_desc_cleaned + 1) to (tx_tail -
> > 1). */
> > +		const uint16_t start = (txq->last_desc_cleaned + 1) % txq-
> > >nb_tx_desc;
> > +		const uint16_t nb_desc = txq->nb_tx_desc;
> > +		const uint16_t end = txq->tx_tail;
> > +
> > +		uint16_t i = start;
> > +		if (end < i) {
> > +			for (; i < nb_desc; i++)
> > +				if (txq->sw_ring[i].mbuf != NULL)
> > +					rte_pktmbuf_free_seg(txq-
> > >sw_ring[i].mbuf);
> > +			i = 0;
> >  		}
> > +		for (; i < end; i++)
> > +			if (txq->sw_ring[i].mbuf != NULL)
> > +				rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
> > +		memset(txq->sw_ring, 0, sizeof(txq->sw_ring[0]) * nb_desc);
> >  		return;
> >  	}
> 
> The above LGTM.
> IIRC, we already discussed it - or something very similar.
> 
> > 
> > diff --git a/drivers/net/intel/common/tx_scalar.h
> > b/drivers/net/intel/common/tx_scalar.h
> > index 9fcd2e4733..adbc4bafee 100644
> > --- a/drivers/net/intel/common/tx_scalar.h
> > +++ b/drivers/net/intel/common/tx_scalar.h
> > @@ -197,16 +197,63 @@ ci_tx_xmit_cleanup(struct ci_tx_queue *txq)
> >  	const uint16_t rs_idx = (last_desc_cleaned == nb_tx_desc - 1) ?
> >  			0 :
> >  			(last_desc_cleaned + 1) >> txq->log2_rs_thresh;
> > -	uint16_t desc_to_clean_to = (rs_idx << txq->log2_rs_thresh) +
> > (txq->tx_rs_thresh - 1);
> > +	const uint16_t dd_idx = txq->rs_last_id[rs_idx];
> > +	const uint16_t first_to_clean = rs_idx << txq->log2_rs_thresh;
> > 
> > -	/* Check if descriptor is done  */
> > -	if ((txd[txq->rs_last_id[rs_idx]].cmd_type_offset_bsz &
> > -			rte_cpu_to_le_64(CI_TXD_QW1_DTYPE_M)) !=
> > -				rte_cpu_to_le_64(CI_TX_DESC_DTYPE_DESC_DONE))
> > +	/* Check if descriptor is done - all drivers use 0xF as done
> > value in bits 3:0 */
> > +	if ((txd[dd_idx].cmd_type_offset_bsz &
> > rte_cpu_to_le_64(CI_TXD_QW1_DTYPE_M)) !=
> > +			rte_cpu_to_le_64(CI_TX_DESC_DTYPE_DESC_DONE))
> > +		/* Descriptor not yet processed by hardware */
> >  		return -1;
> > 
> > +	/* DD bit is set, descriptors are done. Now free the mbufs. */
> > +	/* Note: nb_tx_desc is guaranteed to be a multiple of
> > tx_rs_thresh,
> > +	 * validated during queue setup. This means cleanup never wraps
> > around
> > +	 * the ring within a single burst (e.g., ring=256, rs_thresh=32
> > gives
> > +	 * bursts of 0-31, 32-63, ..., 224-255).
> > +	 */
> > +	const uint16_t nb_to_clean = txq->tx_rs_thresh;
> > +	struct ci_tx_entry *sw_ring = txq->sw_ring;
> > +
> > +	if (txq->offloads & RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE) {
> > +		/* FAST_FREE path: mbufs are already reset, just return to
> > pool */
> 
> Depending on which cache lines from txq have already been loaded, unless txq->offloads is hot in the CPU cache and txq->fast_free_mp is not, consider testing (mp != NULL) instead of (txq->offloads & RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE).
> Like here:
> https://elixir.bootlin.com/dpdk/v26.03/source/drivers/net/intel/common/tx.h#L281
> 

Ack. The offloads are on the second cacheline of the txq, I think, while
the fast_free_mp is on the first one. Therefore, checking the mp pointer
for null is slightly better.

> > +		void *free[CI_TX_MAX_FREE_BUF_SZ];
> > +		uint16_t nb_free = 0;
> > +
> > +		/* Get cached mempool pointer, or cache it on first use */
> > +		struct rte_mempool *mp =
> > +			likely(txq->fast_free_mp != (void *)UINTPTR_MAX) ?
> > +			txq->fast_free_mp :
> > +			(txq->fast_free_mp = sw_ring[dd_idx].mbuf->pool);
> > +
> > +		/* Pack non-NULL mbufs in-place at start of sw_ring range.
> > +		 * No modulo needed in loop since we're guaranteed not to
> > wrap.
> > +		 */
> > +		for (uint16_t i = 0; i < nb_to_clean; i++) {
> > +			struct rte_mbuf *m = sw_ring[first_to_clean +
> > i].mbuf;
> > +			if (m == NULL)
> > +				continue;
> > +			free[nb_free++] = m;
> 
> Should sw_ring[first_to_clean + i].mbuf be set to NULL here, instead of in ci_xmit_pkts()?
> I don't know, just want you to consider it.

No, we are ok just setting it there. The changes in this patch to the
ci_txq_release_all_mbufs() function mean that we assume all buffers between
tx_tail and last_desc_cleaned are invalid/freed (i.e. we free from
last-cleaned to tx_tail only), and so we don't need to set the invalid
slots to NULL.  The NULL sentinel is only used where we have a valid slot
but which does not have an associated mbuf.

> 
> > +			if (unlikely(nb_free == CI_TX_MAX_FREE_BUF_SZ)) {
> > +				rte_mempool_put_bulk(mp, free, nb_free);
> 
> rte_mempool_put_bulk() -> rte_mbuf_raw_free_bulk(),
> for instrumentation.
> 

Ack

> > +				nb_free = 0;
> > +			}
> > +		}
> > +
> > +		/* Bulk return to mempool using packed sw_ring entries
> > directly */
> > +		if (nb_free > 0)
> > +			rte_mempool_put_bulk(mp, free, nb_free);
> 
> rte_mempool_put_bulk() -> rte_mbuf_raw_free_bulk(),
> for instrumentation.

Ack

> 
> > +	} else {
> > +		/* Non-FAST_FREE path: use prefree_seg for refcount checks
> > */
> > +		for (uint16_t i = 0; i < nb_to_clean; i++) {
> > +			struct rte_mbuf *m = sw_ring[first_to_clean +
> > i].mbuf;
> > +			if (m != NULL)
> > +				rte_pktmbuf_free_seg(m);
> 
> Should sw_ring[first_to_clean + i].mbuf be set to NULL here, instead of in ci_xmit_pkts()?
> I don't know, just want you to consider it.
> 

As above, reason should hold for both fast-free and non-fast-free paths.

> > +		}
> > +	}
> > +
> >  	/* Update the txq to reflect the last descriptor that was cleaned
> > */
> > -	txq->last_desc_cleaned = desc_to_clean_to;
> > +	txq->last_desc_cleaned = first_to_clean + txq->tx_rs_thresh - 1;
> >  	txq->nb_tx_free += txq->tx_rs_thresh;
> > 
> >  	return 0;
> > @@ -450,8 +497,6 @@ ci_xmit_pkts(struct ci_tx_queue *txq,
> >  			txd = &ci_tx_ring[tx_id];
> >  			tx_id = txe->next_id;
> > 
> > -			if (txe->mbuf)
> > -				rte_pktmbuf_free_seg(txe->mbuf);
> >  			txe->mbuf = tx_pkt;
> >  			/* Setup TX Descriptor */
> >  			td_cmd |= CI_TX_DESC_CMD_EOP;
> > @@ -472,10 +517,7 @@ ci_xmit_pkts(struct ci_tx_queue *txq,
> > 
> >  			txn = &sw_ring[txe->next_id];
> >  			RTE_MBUF_PREFETCH_TO_FREE(txn->mbuf);
> 
> RTE_MBUF_PREFETCH_TO_FREE() doesn't seem relevant here anymore.
> I don't know if it fits into ci_tx_xmit_cleanup() instead.

Yes, dropping.

> 
> > -			if (txe->mbuf) {
> > -				rte_pktmbuf_free_seg(txe->mbuf);
> > -				txe->mbuf = NULL;
> > -			}
> > +			txe->mbuf = NULL;
> 
> Already mentioned: Should txe->mbuf be set to NULL in ci_tx_xmit_cleanup() instead of in ci_tx_xmit_pkts()?
> 
> > 
> >  			write_txd(ctx_txd, cd_qw0, cd_qw1);
> > 
> > @@ -489,10 +531,7 @@ ci_xmit_pkts(struct ci_tx_queue *txq,
> > 
> >  			txn = &sw_ring[txe->next_id];
> >  			RTE_MBUF_PREFETCH_TO_FREE(txn->mbuf);
> 
> RTE_MBUF_PREFETCH_TO_FREE() doesn't seem relevant here anymore.
> I don't know if it fits into ci_tx_xmit_cleanup() instead.
> 
> > -			if (txe->mbuf) {
> > -				rte_pktmbuf_free_seg(txe->mbuf);
> > -				txe->mbuf = NULL;
> > -			}
> > +			txe->mbuf = NULL;
> 
> Already mentioned: Should txe->mbuf be set to NULL in ci_tx_xmit_cleanup() instead of in ci_tx_xmit_pkts()?
> 
> > 
> >  			ipsec_txd[0] = ipsec_qw0;
> >  			ipsec_txd[1] = ipsec_qw1;
> > @@ -507,10 +546,21 @@ ci_xmit_pkts(struct ci_tx_queue *txq,
> >  			txd = &ci_tx_ring[tx_id];
> >  			txn = &sw_ring[txe->next_id];
> > 
> > -			if (txe->mbuf)
> > -				rte_pktmbuf_free_seg(txe->mbuf);
> >  			txe->mbuf = m_seg;
> > 
> > +			/* For FAST_FREE: reset mbuf fields while we have it
> > in cache.
> > +			 * FAST_FREE guarantees refcnt=1 and direct mbufs, so
> > we only
> > +			 * need to reset nb_segs and next pointer as per
> > rte_pktmbuf_prefree_seg.
> > +			 * Save next pointer before resetting since we need
> > it for loop iteration.
> > +			 */
> > +			struct rte_mbuf *next_seg = m_seg->next;
> > +			if (txq->offloads &
> > RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE) {
> 
> Similar to comment further above: Is txq->offloads or txq->fast_free_mp hotter in the CPU cache here?
> 

Updating as you suggest.

> > +				if (m_seg->nb_segs != 1)
> > +					m_seg->nb_segs = 1;
> > +				if (next_seg != NULL)
> > +					m_seg->next = NULL;
> > +			}
> > +
> >  			/* Setup TX Descriptor */
> >  			/* Calculate segment length, using IPsec callback if
> > provided */
> >  			if (ipsec_ops != NULL)
> > @@ -528,18 +578,23 @@ ci_xmit_pkts(struct ci_tx_queue *txq,
> >  					((uint64_t)CI_MAX_DATA_PER_TXD <<
> > CI_TXD_QW1_TX_BUF_SZ_S) |
> >  					((uint64_t)td_tag <<
> > CI_TXD_QW1_L2TAG1_S);
> >  				write_txd(txd, buf_dma_addr,
> > cmd_type_offset_bsz);
> > +				/* txe for this slot has already been written
> > (e.g. above outside
> > +				 * loop), so we write the extra NULL mbuf
> > pointer for this
> > +				 * descriptor after we increment txe below.
> > +				 */
> > 
> >  				buf_dma_addr += CI_MAX_DATA_PER_TXD;
> >  				slen -= CI_MAX_DATA_PER_TXD;
> > 
> >  				tx_id = txe->next_id;
> >  				txe = txn;
> > +				txe->mbuf = NULL;
> >  				txd = &ci_tx_ring[tx_id];
> >  				txn = &sw_ring[txe->next_id];
> >  			}
> > 
> >  			/* fill the last descriptor with End of Packet (EOP)
> > bit */
> > -			if (m_seg->next == NULL)
> > +			if (next_seg == NULL)
> >  				td_cmd |= CI_TX_DESC_CMD_EOP;
> > 
> >  			const uint64_t cmd_type_offset_bsz =
> > CI_TX_DESC_DTYPE_DATA |
> > @@ -551,7 +606,7 @@ ci_xmit_pkts(struct ci_tx_queue *txq,
> > 
> >  			tx_id = txe->next_id;
> >  			txe = txn;
> > -			m_seg = m_seg->next;
> > +			m_seg = next_seg;
> >  		} while (m_seg);
> >  end_pkt:
> >  		txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_used);
> > --
> > 2.51.0
>