[PATCH v6 2/4] net/i40e: implement mbufs recycle mode

[Feifei Wang]

> -----Original Message-----
> From: Konstantin Ananyev <konstantin.ananyev at huawei.com>
> Sent: Tuesday, June 6, 2023 4:27 PM
> To: Feifei Wang <Feifei.Wang2 at arm.com>; Константин Ананьев
> <konstantin.v.ananyev at yandex.ru>; Yuying Zhang
> <yuying.zhang at intel.com>; Beilei Xing <beilei.xing at intel.com>
> Cc: dev at dpdk.org; nd <nd at arm.com>; Honnappa Nagarahalli
> <Honnappa.Nagarahalli at arm.com>; Ruifeng Wang
> <Ruifeng.Wang at arm.com>; nd <nd at arm.com>; nd <nd at arm.com>
> Subject: RE: [PATCH v6 2/4] net/i40e: implement mbufs recycle mode
> 
> 
> 
> > > > Define specific function implementation for i40e driver.
> > > > Currently, mbufs recycle mode can support 128bit vector path and
> > > > avx2 path. And can be enabled both in fast free and no fast free mode.
> > > >
> > > > Suggested-by: Honnappa Nagarahalli
> > > > <mailto:honnappa.nagarahalli at arm.com>
> > > > Signed-off-by: Feifei Wang <mailto:feifei.wang2 at arm.com>
> > > > Reviewed-by: Ruifeng Wang <mailto:ruifeng.wang at arm.com>
> > > > Reviewed-by: Honnappa Nagarahalli
> > > > <mailto:honnappa.nagarahalli at arm.com>
> > > > ---
> > > >  drivers/net/i40e/i40e_ethdev.c | 1 +
> > > > drivers/net/i40e/i40e_ethdev.h | 2 +
> > > > .../net/i40e/i40e_recycle_mbufs_vec_common.c | 140
> > > ++++++++++++++++++
> > > > drivers/net/i40e/i40e_rxtx.c | 32 ++++
> > > > drivers/net/i40e/i40e_rxtx.h |
> > > > 4 +  drivers/net/i40e/meson.build | 2 +
> > > >  6 files changed, 181 insertions(+)  create mode 100644
> > > > drivers/net/i40e/i40e_recycle_mbufs_vec_common.c
> > > >
> > > > diff --git a/drivers/net/i40e/i40e_ethdev.c
> > > > b/drivers/net/i40e/i40e_ethdev.c index f9d8f9791f..d4eecd16cf
> > > > 100644
> > > > --- a/drivers/net/i40e/i40e_ethdev.c
> > > > +++ b/drivers/net/i40e/i40e_ethdev.c
> > > > @@ -496,6 +496,7 @@ static const struct eth_dev_ops
> > > > i40e_eth_dev_ops =
> > > {
> > > >          .flow_ops_get = i40e_dev_flow_ops_get,
> > > >          .rxq_info_get = i40e_rxq_info_get,
> > > >          .txq_info_get = i40e_txq_info_get,
> > > > + .recycle_rxq_info_get = i40e_recycle_rxq_info_get,
> > > >          .rx_burst_mode_get = i40e_rx_burst_mode_get,
> > > >          .tx_burst_mode_get = i40e_tx_burst_mode_get,
> > > >          .timesync_enable = i40e_timesync_enable, diff --git
> > > > a/drivers/net/i40e/i40e_ethdev.h b/drivers/net/i40e/i40e_ethdev.h
> > > > index 9b806d130e..b5b2d6cf2b 100644
> > > > --- a/drivers/net/i40e/i40e_ethdev.h
> > > > +++ b/drivers/net/i40e/i40e_ethdev.h
> > > > @@ -1355,6 +1355,8 @@ void i40e_rxq_info_get(struct rte_eth_dev
> > > > *dev,
> > > uint16_t queue_id,
> > > >          struct rte_eth_rxq_info *qinfo);  void
> > > > i40e_txq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
> > > >          struct rte_eth_txq_info *qinfo);
> > > > +void i40e_recycle_rxq_info_get(struct rte_eth_dev *dev, uint16_t
> > > > +queue_id,  struct rte_eth_recycle_rxq_info *recycle_rxq_info);
> > > >  int i40e_rx_burst_mode_get(struct rte_eth_dev *dev, uint16_t
> queue_id,
> > > >                             struct rte_eth_burst_mode *mode);  int
> > > > i40e_tx_burst_mode_get(struct rte_eth_dev *dev, uint16_t queue_id,
> > > > diff --git a/drivers/net/i40e/i40e_recycle_mbufs_vec_common.c
> > > > b/drivers/net/i40e/i40e_recycle_mbufs_vec_common.c
> > > > new file mode 100644
> > > > index 0000000000..08d708fd7d
> > > > --- /dev/null
> > > > +++ b/drivers/net/i40e/i40e_recycle_mbufs_vec_common.c
> > > > @@ -0,0 +1,140 @@
> > > > +/* SPDX-License-Identifier: BSD-3-Clause
> > > > + * Copyright (c) 2023 Arm Limited.
> > > > + */
> > > > +
> > > > +#include <stdint.h>
> > > > +#include <ethdev_driver.h>
> > > > +
> > > > +#include "base/i40e_prototype.h"
> > > > +#include "base/i40e_type.h"
> > > > +#include "i40e_ethdev.h"
> > > > +#include "i40e_rxtx.h"
> > > > +
> > > > +#pragma GCC diagnostic ignored "-Wcast-qual"
> > > > +
> > > > +void
> > > > +i40e_recycle_rx_descriptors_refill_vec(void *rx_queue, uint16_t
> > > > +nb_mbufs) {  struct i40e_rx_queue *rxq = rx_queue;  struct
> > > > +i40e_rx_entry *rxep;  volatile union i40e_rx_desc *rxdp;
> > > > +uint16_t rx_id;  uint64_t paddr;  uint64_t dma_addr;  uint16_t i;
> > > > +
> > > > + rxdp = rxq->rx_ring + rxq->rxrearm_start; rxep =
> > > > + &rxq->sw_ring[rxq->rxrearm_start];
> > > > +
> > > > + for (i = 0; i < nb_mbufs; i++) {
> > > > + /* Initialize rxdp descs. */
> > > > + paddr = (rxep[i].mbuf)->buf_iova + RTE_PKTMBUF_HEADROOM;
> > > dma_addr =
> > > > + rte_cpu_to_le_64(paddr);
> > > > + /* flush desc with pa dma_addr */ rxdp[i].read.hdr_addr = 0;
> > > > + rxdp[i].read.pkt_addr = dma_addr; }
> > > > +
> > > > + /* Update the descriptor initializer index */
> > > > + rxq->rxrearm_start += nb_mbufs;
> > > > + rx_id = rxq->rxrearm_start - 1;
> > > > +
> > > > + if (unlikely(rxq->rxrearm_start >= rxq->nb_rx_desc)) {
> > > > + rxq->rxrearm_start = 0;
> > > > + rx_id = rxq->nb_rx_desc - 1;
> > > > + }
> > > > +
> > > > + rxq->rxrearm_nb -= nb_mbufs;
> > > > +
> > > > + rte_io_wmb();
> > > > + /* Update the tail pointer on the NIC */
> > > > +I40E_PCI_REG_WRITE_RELAXED(rxq->qrx_tail, rx_id); }
> > > > +
> > > > +uint16_t
> > > > +i40e_recycle_tx_mbufs_reuse_vec(void *tx_queue,  struct
> > > > +rte_eth_recycle_rxq_info *recycle_rxq_info) {  struct
> > > > +i40e_tx_queue *txq = tx_queue;  struct i40e_tx_entry *txep;
> > > > +struct rte_mbuf **rxep;  struct rte_mbuf
> > > > +*m[RTE_I40E_TX_MAX_FREE_BUF_SZ];  int i, j, n;  uint16_t avail =
> > > > +0;  uint16_t mbuf_ring_size = recycle_rxq_info->mbuf_ring_size;
> > > > +uint16_t mask = recycle_rxq_info->mbuf_ring_size - 1;  uint16_t
> > > > +refill_requirement = recycle_rxq_info->refill_requirement;
> > > > + uint16_t refill_head = *recycle_rxq_info->refill_head;  uint16_t
> > > > +receive_tail = *recycle_rxq_info->receive_tail;
> > > > +
> > > > + /* Get available recycling Rx buffers. */ avail =
> > > > + (mbuf_ring_size - (refill_head - receive_tail)) & mask;
> > > > +
> > > > + /* Check Tx free thresh and Rx available space. */ if
> > > > + (txq->nb_tx_free > txq->tx_free_thresh || avail <=
> > > > + txq->tx_rs_thresh) return 0;
> > > > +
> > > > + /* check DD bits on threshold descriptor */ if
> > > > + ((txq->tx_ring[txq->tx_next_dd].cmd_type_offset_bsz &
> > > > + rte_cpu_to_le_64(I40E_TXD_QW1_DTYPE_MASK)) !=
> > > > + rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DESC_DONE))
> > > > + return 0;
> > > > +
> > > > + n = txq->tx_rs_thresh;
> > > > +
> > > > + /* Mbufs recycle mode can only support no ring buffer wrapping
> around.
> > > > + * Two case for this:
> > > > + *
> > > > + * case 1: The refill head of Rx buffer ring needs to be aligned
> > > > + with
> > > > + * mbuf ring size. In this case, the number of Tx freeing buffers
> > > > + * should be equal to refill_requirement.
> > > > + *
> > > > + * case 2: The refill head of Rx ring buffer does not need to be
> > > > + aligned
> > > > + * with mbuf ring size. In this case, the update of refill head
> > > > + can not
> > > > + * exceed the Rx mbuf ring size.
> > > > + */
> > > > + if (refill_requirement != n ||
> > > > + (!refill_requirement && (refill_head + n > mbuf_ring_size)))
> > > > + return 0;
> > > > +
> > > > + /* First buffer to free from S/W ring is at index
> > > > + * tx_next_dd - (tx_rs_thresh-1).
> > > > + */
> > > > + txep = &txq->sw_ring[txq->tx_next_dd - (n - 1)]; rxep =
> > > > + recycle_rxq_info->mbuf_ring; rxep += refill_head;
> > > > +
> > > > + if (txq->offloads & RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE) {
> > > > + /* Directly put mbufs from Tx to Rx. */ for (i = 0; i < n; i++,
> > > > + rxep++, txep++) *rxep = txep[0].mbuf; } else { for (i = 0, j =
> > > > + rxep++0; i
> > > > + < n; i++) {
> > > > + /* Avoid txq contains buffers from expected mempool. */ if
> > > > + (unlikely(recycle_rxq_info->mp != txep[i].mbuf->pool)) return 0;
> > > > I don't think that it is possible to simply return 0 here:
> > > > we might already have some mbufs inside rxep[], so we probably
> > > > need to return that number (j).
> > > >
> > > > No, here is just pre-free, not actually put mbufs into rxeq.
> > >
> > > I understand that.
> > > What I am saying: after you call pktmbuf_prefree_seg(mbuf), you
> > > can’t keep it in the txq anymore.
> > > You have either to put it into rxep[], or back into mempool.
> > > Also txq state (nb_tx_free, etc.) need to be updated.
> > >
> > > > After run out of the loop, we call rte_memcpy to actually copy
> > > > mbufs into rxep.
> > > > +
> > > > + m[j] = rte_pktmbuf_prefree_seg(txep[i].mbuf);
> > > > +
> > > > + /* In case 1, each of Tx buffers should be the
> > > > + * last reference.
> > > > + */
> > > > + if (unlikely(m[j] == NULL && refill_requirement)) return 0;
> > > >
> > > > same here, we can't simply return 0, it will introduce mbuf leakage.
> > > > + /* In case 2, the number of valid Tx free
> > > > + * buffers should be recorded.
> > > > + */
> > > > + j++;
> > > > + }
> > > > + rte_memcpy(rxep, m, sizeof(void *) * j);
> > > > Wonder why do you need intermediate buffer for released mbufs?
> > > > Why can't just:
> > > > ...
> > > > m = rte_pktmbuf_prefree_seg(txep[i].mbuf);
> > > > ...
> > > > rxep[j++] = m;
> > > > ?
> > > > Might save you few extra cycles.
> > > > Sometimes ‘rte_pktmbuf_prefree_seg’ can return NULL due to
> > > > mbuf->refcnt > 1. So we should firstly ensure all ‘m’ are valid
> > > > mbuf->and
> > > > then copy them into rxep.
> > >
> > > I understand that, but you can check is it NULL or not.
> > For i40e rxq, it must rearm ' RTE_I40E_RXQ_REARM_THRESH ' pkts once a
> > time based on its ring wrapping mechanism.
> >
> > For i40e txq, it must free ' txq->tx_rs_thresh' pkts once a time.
> >
> > So we need firstly ensure all tx free mbufs are valid, and then copy these into
> rxq.
> > If not enough valid mbufs, it will break rxq's ring wrapping mechanism.
> 
> I think you can still copy mbufs into rxep[], if there are not enough mbufs, you
> can still return 0 (or whatever is a proper value here), and that would mean all
> these new rxep[] entries will be considered as invalid.
> Anyway that's just a suggestion to avoid extra copy.

If I understand correctly, you means we can firstly copy mbufs into rxep.
If there are invalid buffers, previous copied buffers are also considered as invalid.

Thus, this can save CPU cycles in most of the correct cases. And just in few case,
We need to give up the copied rxep[] buffers.

That's a good comments, I agree that we can do like this.
> 
> >
> > >
> > > > + }
> > > > +
> > > > + /* Update counters for Tx. */
> > > > + txq->nb_tx_free = (uint16_t)(txq->nb_tx_free +
> > > > + txq->txq->tx_rs_thresh); tx_next_dd = (uint16_t)(txq->tx_next_dd
> > > > + txq->+ txq->tx_rs_thresh);
> > > > + if (txq->tx_next_dd >= txq->nb_tx_desc)
> > > > + txq->tx_next_dd = (uint16_t)(txq->tx_rs_thresh - 1);
> > > > +
> > > > + return n;
> > > > +}