> -----Original Message-----
> From: Chuanyu Xue <chuanyu....@uconn.edu>
> Sent: Sunday, December 31, 2023 12:35 AM
> To: Su, Simei <simei...@intel.com>; Lu, Wenzhuo <wenzhuo...@intel.com>;
> Zhang, Qi Z <qi.z.zh...@intel.com>; Xing, Beilei <beilei.x...@intel.com>
> Cc: dev@dpdk.org; Chuanyu Xue <chuanyu....@uconn.edu>
> Subject: [PATCH v2] net/e1000: support launchtime feature
>
> Enable the time-based scheduled Tx of packets based on the
> RTE_ETH_TX_OFFLOAD_SEND_ON_TIMESTAMP flag. The launchtime defines
> the packet transmission time based on PTP clock at MAC layer, which should
> be set to the advanced transmit descriptor.
>
> Signed-off-by: Chuanyu Xue <chuanyu....@uconn.edu>
> ---
> change log:
>
> v2:
> - Add delay compensation for i210 NIC by setting tx offset register.
> - Revise read_clock function.
>
> drivers/net/e1000/base/e1000_regs.h | 1 +
> drivers/net/e1000/e1000_ethdev.h | 14 +++++++
> drivers/net/e1000/igb_ethdev.c | 63
> ++++++++++++++++++++++++++++-
> drivers/net/e1000/igb_rxtx.c | 42 +++++++++++++++----
> 4 files changed, 112 insertions(+), 8 deletions(-)
>
> diff --git a/drivers/net/e1000/base/e1000_regs.h
> b/drivers/net/e1000/base/e1000_regs.h
> index d44de59c29..092d9d71e6 100644
> --- a/drivers/net/e1000/base/e1000_regs.h
> +++ b/drivers/net/e1000/base/e1000_regs.h
> @@ -162,6 +162,7 @@
>
> /* QAV Tx mode control register */
> #define E1000_I210_TQAVCTRL 0x3570
> +#define E1000_I210_LAUNCH_OS0 0x3578
>
> /* QAV Tx mode control register bitfields masks */
> /* QAV enable */
> diff --git a/drivers/net/e1000/e1000_ethdev.h
> b/drivers/net/e1000/e1000_ethdev.h
> index 718a9746ed..339ae1f4b6 100644
> --- a/drivers/net/e1000/e1000_ethdev.h
> +++ b/drivers/net/e1000/e1000_ethdev.h
> @@ -382,6 +382,20 @@ extern struct igb_rss_filter_list igb_filter_rss_list;
> TAILQ_HEAD(igb_flow_mem_list, igb_flow_mem); extern struct
> igb_flow_mem_list igb_flow_list;
>
> +/*
> + * Macros to compensate the constant latency observed in i210 for
> +launch time
> + *
> + * launch time = (offset_speed - offset_base + txtime) * 32
> + * offset_speed is speed dependent, set in E1000_I210_LAUNCH_OS0 */
> +#define IGB_I210_TX_OFFSET_BASE 0xffe0
> +#define IGB_I210_TX_OFFSET_SPEED_10 0xc7a0
> +#define IGB_I210_TX_OFFSET_SPEED_100 0x86e0
> +#define IGB_I210_TX_OFFSET_SPEED_1000 0xbe00
> +
> +extern uint64_t igb_tx_timestamp_dynflag; extern int
> +igb_tx_timestamp_dynfield_offset;
> +
> extern const struct rte_flow_ops igb_flow_ops;
>
> /*
> diff --git a/drivers/net/e1000/igb_ethdev.c b/drivers/net/e1000/igb_ethdev.c
> index 8858f975f8..2262035710 100644
> --- a/drivers/net/e1000/igb_ethdev.c
> +++ b/drivers/net/e1000/igb_ethdev.c
> @@ -223,6 +223,7 @@ static int igb_timesync_read_time(struct rte_eth_dev
> *dev,
> struct timespec *timestamp);
> static int igb_timesync_write_time(struct rte_eth_dev *dev,
> const struct timespec *timestamp);
> +static int eth_igb_read_clock(struct rte_eth_dev *dev, uint64_t
> +*clock);
> static int eth_igb_rx_queue_intr_enable(struct rte_eth_dev *dev,
> uint16_t queue_id);
> static int eth_igb_rx_queue_intr_disable(struct rte_eth_dev *dev, @@ -313,6
> +314,9 @@ static const struct rte_pci_id pci_id_igbvf_map[] = {
> { .vendor_id = 0, /* sentinel */ },
> };
>
> +uint64_t igb_tx_timestamp_dynflag;
> +int igb_tx_timestamp_dynfield_offset = -1;
> +
> static const struct rte_eth_desc_lim rx_desc_lim = {
> .nb_max = E1000_MAX_RING_DESC,
> .nb_min = E1000_MIN_RING_DESC,
> @@ -389,6 +393,7 @@ static const struct eth_dev_ops eth_igb_ops = {
> .timesync_adjust_time = igb_timesync_adjust_time,
> .timesync_read_time = igb_timesync_read_time,
> .timesync_write_time = igb_timesync_write_time,
> + .read_clock = eth_igb_read_clock,
> };
>
> /*
> @@ -1188,6 +1193,40 @@ eth_igb_rxtx_control(struct rte_eth_dev *dev,
> E1000_WRITE_FLUSH(hw);
> }
>
> +
> +static uint32_t igb_tx_offset(struct rte_eth_dev *dev) {
> + struct e1000_hw *hw =
> + E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
> +
> + uint16_t duplex, speed;
> + hw->mac.ops.get_link_up_info(hw, &speed, &duplex);
> +
> + uint32_t launch_os0 = E1000_READ_REG(hw, E1000_I210_LAUNCH_OS0);
> + if (hw->mac.type != e1000_i210) {
> + /* Set launch offset to base, no compensation */
> + launch_os0 |= IGB_I210_TX_OFFSET_BASE;
> + } else {
> + /* Set launch offset depend on link speeds */
> + switch (speed) {
> + case SPEED_10:
> + launch_os0 |= IGB_I210_TX_OFFSET_SPEED_10;
> + break;
> + case SPEED_100:
> + launch_os0 |= IGB_I210_TX_OFFSET_SPEED_100;
> + break;
> + case SPEED_1000:
> + launch_os0 |= IGB_I210_TX_OFFSET_SPEED_1000;
> + break;
> + default:
> + launch_os0 |= IGB_I210_TX_OFFSET_BASE;
> + break;
> + }
> + }
> +
> + return launch_os0;
> +}
> +
> static int
> eth_igb_start(struct rte_eth_dev *dev)
> {
> @@ -1198,6 +1237,7 @@ eth_igb_start(struct rte_eth_dev *dev)
> struct rte_pci_device *pci_dev = RTE_ETH_DEV_TO_PCI(dev);
> struct rte_intr_handle *intr_handle = pci_dev->intr_handle;
> int ret, mask;
> + uint32_t tqavctrl;
> uint32_t intr_vector = 0;
> uint32_t ctrl_ext;
> uint32_t *speeds;
> @@ -1273,6 +1313,15 @@ eth_igb_start(struct rte_eth_dev *dev)
>
> eth_igb_tx_init(dev);
>
> + if (igb_tx_timestamp_dynflag > 0) {
> + tqavctrl = E1000_READ_REG(hw, E1000_I210_TQAVCTRL);
> + tqavctrl |= E1000_TQAVCTRL_MODE; /* Enable Qav mode */
> + tqavctrl |= E1000_TQAVCTRL_FETCH_ARB; /* ARB fetch, no Round
> Robin*/
> + tqavctrl |= E1000_TQAVCTRL_LAUNCH_TIMER_ENABLE; /* Enable Tx
> launch time*/
> + E1000_WRITE_REG(hw, E1000_I210_TQAVCTRL, tqavctrl);
> + E1000_WRITE_REG(hw, E1000_I210_LAUNCH_OS0,
> igb_tx_offset(dev));
> + }
> +
> /* This can fail when allocating mbufs for descriptor rings */
> ret = eth_igb_rx_init(dev);
> if (ret) {
> @@ -1393,7 +1442,6 @@ eth_igb_start(struct rte_eth_dev *dev)
>
> eth_igb_rxtx_control(dev, true);
> eth_igb_link_update(dev, 0);
> -
> PMD_INIT_LOG(DEBUG, "<<");
>
> return 0;
> @@ -4882,6 +4930,19 @@ igb_timesync_read_tx_timestamp(struct
> rte_eth_dev *dev,
> return 0;
> }
>
> +static int
> +eth_igb_read_clock(struct rte_eth_dev *dev, uint64_t *clock) {
> + struct e1000_adapter *adapter = dev->data->dev_private;
> + struct rte_timecounter *tc = &adapter->systime_tc;
> + uint64_t cycles;
> +
> + cycles = igb_read_systime_cyclecounter(dev);
> + *clock = rte_timecounter_update(tc, cycles);
> +
> + return 0;
> +}
> +
> static int
> eth_igb_get_reg_length(struct rte_eth_dev *dev __rte_unused) { diff --git
> a/drivers/net/e1000/igb_rxtx.c b/drivers/net/e1000/igb_rxtx.c index
> 448c4b7d9d..5cafd6f1ce 100644
> --- a/drivers/net/e1000/igb_rxtx.c
> +++ b/drivers/net/e1000/igb_rxtx.c
> @@ -244,12 +244,13 @@ check_tso_para(uint64_t ol_req, union
> igb_tx_offload ol_para) static inline void igbe_set_xmit_ctx(struct
> igb_tx_queue* txq,
> volatile struct e1000_adv_tx_context_desc *ctx_txd,
> - uint64_t ol_flags, union igb_tx_offload tx_offload)
> + uint64_t ol_flags, union igb_tx_offload tx_offload, uint64_t
> txtime)
> {
> uint32_t type_tucmd_mlhl;
> uint32_t mss_l4len_idx;
> uint32_t ctx_idx, ctx_curr;
> uint32_t vlan_macip_lens;
> + uint32_t launch_time;
> union igb_tx_offload tx_offload_mask;
>
> ctx_curr = txq->ctx_curr;
> @@ -312,16 +313,25 @@ igbe_set_xmit_ctx(struct igb_tx_queue* txq,
> }
> }
>
> - txq->ctx_cache[ctx_curr].flags = ol_flags;
> - txq->ctx_cache[ctx_curr].tx_offload.data =
> - tx_offload_mask.data & tx_offload.data;
> - txq->ctx_cache[ctx_curr].tx_offload_mask = tx_offload_mask;
> + if (!txtime) {
> + txq->ctx_cache[ctx_curr].flags = ol_flags;
> + txq->ctx_cache[ctx_curr].tx_offload.data =
> + tx_offload_mask.data & tx_offload.data;
> + txq->ctx_cache[ctx_curr].tx_offload_mask = tx_offload_mask;
> + }
>
> ctx_txd->type_tucmd_mlhl = rte_cpu_to_le_32(type_tucmd_mlhl);
> vlan_macip_lens = (uint32_t)tx_offload.data;
> ctx_txd->vlan_macip_lens = rte_cpu_to_le_32(vlan_macip_lens);
> ctx_txd->mss_l4len_idx = rte_cpu_to_le_32(mss_l4len_idx);
> ctx_txd->u.seqnum_seed = 0;
> +
> + if (txtime) {
> + launch_time = (txtime - IGB_I210_TX_OFFSET_BASE) %
> NSEC_PER_SEC;
> + ctx_txd->u.launch_time = rte_cpu_to_le_32(launch_time / 32);
> + } else {
> + ctx_txd->u.launch_time = 0;
> + }
> }
>
> /*
> @@ -400,6 +410,7 @@ eth_igb_xmit_pkts(void *tx_queue, struct rte_mbuf
> **tx_pkts,
> uint32_t new_ctx = 0;
> uint32_t ctx = 0;
> union igb_tx_offload tx_offload = {0};
> + uint64_t ts;
>
> txq = tx_queue;
> sw_ring = txq->sw_ring;
> @@ -552,7 +563,13 @@ eth_igb_xmit_pkts(void *tx_queue, struct rte_mbuf
> **tx_pkts,
> txe->mbuf = NULL;
> }
>
> - igbe_set_xmit_ctx(txq, ctx_txd, tx_ol_req,
> tx_offload);
> + if (igb_tx_timestamp_dynflag > 0) {
> + ts = *RTE_MBUF_DYNFIELD(tx_pkt,
> +
> igb_tx_timestamp_dynfield_offset, uint64_t *);
> + igbe_set_xmit_ctx(txq, ctx_txd,
> tx_ol_req, tx_offload,
> ts);
> + } else {
> + igbe_set_xmit_ctx(txq, ctx_txd,
> tx_ol_req, tx_offload,
> 0);
> + }
>
> txe->last_id = tx_last;
> tx_id = txe->next_id;
> @@ -1464,7 +1481,8 @@ igb_get_tx_port_offloads_capa(struct rte_eth_dev
> *dev)
> RTE_ETH_TX_OFFLOAD_TCP_CKSUM |
> RTE_ETH_TX_OFFLOAD_SCTP_CKSUM |
> RTE_ETH_TX_OFFLOAD_TCP_TSO |
> - RTE_ETH_TX_OFFLOAD_MULTI_SEGS;
> + RTE_ETH_TX_OFFLOAD_MULTI_SEGS |
> + RTE_ETH_TX_OFFLOAD_SEND_ON_TIMESTAMP;
>
> return tx_offload_capa;
> }
> @@ -2579,9 +2597,11 @@ eth_igb_tx_init(struct rte_eth_dev *dev) {
> struct e1000_hw *hw;
> struct igb_tx_queue *txq;
> + uint64_t offloads = dev->data->dev_conf.txmode.offloads;
> uint32_t tctl;
> uint32_t txdctl;
> uint16_t i;
> + int err;
>
> hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
>
> @@ -2612,6 +2632,14 @@ eth_igb_tx_init(struct rte_eth_dev *dev)
> dev->data->tx_queue_state[i] = RTE_ETH_QUEUE_STATE_STARTED;
> }
>
> + if (offloads & RTE_ETH_TX_OFFLOAD_SEND_ON_TIMESTAMP) {
> + err = rte_mbuf_dyn_tx_timestamp_register(
> + &igb_tx_timestamp_dynfield_offset,
> + &igb_tx_timestamp_dynflag);
> + if (err)
> + PMD_DRV_LOG(ERR, "Failed to register tx timestamp
> dynamic
> field");
> + }
> +
> /* Program the Transmit Control Register. */
> tctl = E1000_READ_REG(hw, E1000_TCTL);
> tctl &= ~E1000_TCTL_CT;
> --
> 2.25.1
Reviewed-by: Simei Su <simei...@intel.com>
Thanks,
Simei
