1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2018-2019 Hisilicon Limited.
5 #include <rte_bus_pci.h>
6 #include <rte_common.h>
7 #include <rte_cycles.h>
9 #include <rte_ethdev_driver.h>
12 #include <rte_malloc.h>
13 #if defined(RTE_ARCH_ARM64) && defined(CC_SVE_SUPPORT)
14 #include <rte_cpuflags.h>
17 #include "hns3_ethdev.h"
18 #include "hns3_rxtx.h"
19 #include "hns3_regs.h"
20 #include "hns3_logs.h"
22 #define HNS3_CFG_DESC_NUM(num) ((num) / 8 - 1)
23 #define HNS3_RX_RING_PREFETCTH_MASK 3
26 hns3_rx_queue_release_mbufs(struct hns3_rx_queue *rxq)
30 /* Note: Fake rx queue will not enter here */
31 if (rxq->sw_ring == NULL)
34 if (rxq->rx_rearm_nb == 0) {
35 for (i = 0; i < rxq->nb_rx_desc; i++) {
36 if (rxq->sw_ring[i].mbuf != NULL) {
37 rte_pktmbuf_free_seg(rxq->sw_ring[i].mbuf);
38 rxq->sw_ring[i].mbuf = NULL;
42 for (i = rxq->next_to_use;
43 i != rxq->rx_rearm_start;
44 i = (i + 1) % rxq->nb_rx_desc) {
45 if (rxq->sw_ring[i].mbuf != NULL) {
46 rte_pktmbuf_free_seg(rxq->sw_ring[i].mbuf);
47 rxq->sw_ring[i].mbuf = NULL;
52 for (i = 0; i < rxq->bulk_mbuf_num; i++)
53 rte_pktmbuf_free_seg(rxq->bulk_mbuf[i]);
54 rxq->bulk_mbuf_num = 0;
56 if (rxq->pkt_first_seg) {
57 rte_pktmbuf_free(rxq->pkt_first_seg);
58 rxq->pkt_first_seg = NULL;
63 hns3_tx_queue_release_mbufs(struct hns3_tx_queue *txq)
67 /* Note: Fake tx queue will not enter here */
69 for (i = 0; i < txq->nb_tx_desc; i++) {
70 if (txq->sw_ring[i].mbuf) {
71 rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
72 txq->sw_ring[i].mbuf = NULL;
79 hns3_rx_queue_release(void *queue)
81 struct hns3_rx_queue *rxq = queue;
83 hns3_rx_queue_release_mbufs(rxq);
85 rte_memzone_free(rxq->mz);
87 rte_free(rxq->sw_ring);
93 hns3_tx_queue_release(void *queue)
95 struct hns3_tx_queue *txq = queue;
97 hns3_tx_queue_release_mbufs(txq);
99 rte_memzone_free(txq->mz);
101 rte_free(txq->sw_ring);
109 hns3_dev_rx_queue_release(void *queue)
111 struct hns3_rx_queue *rxq = queue;
112 struct hns3_adapter *hns;
118 rte_spinlock_lock(&hns->hw.lock);
119 hns3_rx_queue_release(queue);
120 rte_spinlock_unlock(&hns->hw.lock);
124 hns3_dev_tx_queue_release(void *queue)
126 struct hns3_tx_queue *txq = queue;
127 struct hns3_adapter *hns;
133 rte_spinlock_lock(&hns->hw.lock);
134 hns3_tx_queue_release(queue);
135 rte_spinlock_unlock(&hns->hw.lock);
139 hns3_fake_rx_queue_release(struct hns3_rx_queue *queue)
141 struct hns3_rx_queue *rxq = queue;
142 struct hns3_adapter *hns;
152 if (hw->fkq_data.rx_queues[idx]) {
153 hns3_rx_queue_release(hw->fkq_data.rx_queues[idx]);
154 hw->fkq_data.rx_queues[idx] = NULL;
157 /* free fake rx queue arrays */
158 if (idx == (hw->fkq_data.nb_fake_rx_queues - 1)) {
159 hw->fkq_data.nb_fake_rx_queues = 0;
160 rte_free(hw->fkq_data.rx_queues);
161 hw->fkq_data.rx_queues = NULL;
166 hns3_fake_tx_queue_release(struct hns3_tx_queue *queue)
168 struct hns3_tx_queue *txq = queue;
169 struct hns3_adapter *hns;
179 if (hw->fkq_data.tx_queues[idx]) {
180 hns3_tx_queue_release(hw->fkq_data.tx_queues[idx]);
181 hw->fkq_data.tx_queues[idx] = NULL;
184 /* free fake tx queue arrays */
185 if (idx == (hw->fkq_data.nb_fake_tx_queues - 1)) {
186 hw->fkq_data.nb_fake_tx_queues = 0;
187 rte_free(hw->fkq_data.tx_queues);
188 hw->fkq_data.tx_queues = NULL;
193 hns3_free_rx_queues(struct rte_eth_dev *dev)
195 struct hns3_adapter *hns = dev->data->dev_private;
196 struct hns3_fake_queue_data *fkq_data;
197 struct hns3_hw *hw = &hns->hw;
201 nb_rx_q = hw->data->nb_rx_queues;
202 for (i = 0; i < nb_rx_q; i++) {
203 if (dev->data->rx_queues[i]) {
204 hns3_rx_queue_release(dev->data->rx_queues[i]);
205 dev->data->rx_queues[i] = NULL;
209 /* Free fake Rx queues */
210 fkq_data = &hw->fkq_data;
211 for (i = 0; i < fkq_data->nb_fake_rx_queues; i++) {
212 if (fkq_data->rx_queues[i])
213 hns3_fake_rx_queue_release(fkq_data->rx_queues[i]);
218 hns3_free_tx_queues(struct rte_eth_dev *dev)
220 struct hns3_adapter *hns = dev->data->dev_private;
221 struct hns3_fake_queue_data *fkq_data;
222 struct hns3_hw *hw = &hns->hw;
226 nb_tx_q = hw->data->nb_tx_queues;
227 for (i = 0; i < nb_tx_q; i++) {
228 if (dev->data->tx_queues[i]) {
229 hns3_tx_queue_release(dev->data->tx_queues[i]);
230 dev->data->tx_queues[i] = NULL;
234 /* Free fake Tx queues */
235 fkq_data = &hw->fkq_data;
236 for (i = 0; i < fkq_data->nb_fake_tx_queues; i++) {
237 if (fkq_data->tx_queues[i])
238 hns3_fake_tx_queue_release(fkq_data->tx_queues[i]);
243 hns3_free_all_queues(struct rte_eth_dev *dev)
245 hns3_free_rx_queues(dev);
246 hns3_free_tx_queues(dev);
250 hns3_alloc_rx_queue_mbufs(struct hns3_hw *hw, struct hns3_rx_queue *rxq)
252 struct rte_mbuf *mbuf;
256 for (i = 0; i < rxq->nb_rx_desc; i++) {
257 mbuf = rte_mbuf_raw_alloc(rxq->mb_pool);
258 if (unlikely(mbuf == NULL)) {
259 hns3_err(hw, "Failed to allocate RXD[%u] for rx queue!",
261 hns3_rx_queue_release_mbufs(rxq);
265 rte_mbuf_refcnt_set(mbuf, 1);
267 mbuf->data_off = RTE_PKTMBUF_HEADROOM;
269 mbuf->port = rxq->port_id;
271 rxq->sw_ring[i].mbuf = mbuf;
272 dma_addr = rte_cpu_to_le_64(rte_mbuf_data_iova_default(mbuf));
273 rxq->rx_ring[i].addr = dma_addr;
274 rxq->rx_ring[i].rx.bd_base_info = 0;
281 hns3_buf_size2type(uint32_t buf_size)
287 bd_size_type = HNS3_BD_SIZE_512_TYPE;
290 bd_size_type = HNS3_BD_SIZE_1024_TYPE;
293 bd_size_type = HNS3_BD_SIZE_4096_TYPE;
296 bd_size_type = HNS3_BD_SIZE_2048_TYPE;
303 hns3_init_rx_queue_hw(struct hns3_rx_queue *rxq)
305 uint32_t rx_buf_len = rxq->rx_buf_len;
306 uint64_t dma_addr = rxq->rx_ring_phys_addr;
308 hns3_write_dev(rxq, HNS3_RING_RX_BASEADDR_L_REG, (uint32_t)dma_addr);
309 hns3_write_dev(rxq, HNS3_RING_RX_BASEADDR_H_REG,
310 (uint32_t)((dma_addr >> 31) >> 1));
312 hns3_write_dev(rxq, HNS3_RING_RX_BD_LEN_REG,
313 hns3_buf_size2type(rx_buf_len));
314 hns3_write_dev(rxq, HNS3_RING_RX_BD_NUM_REG,
315 HNS3_CFG_DESC_NUM(rxq->nb_rx_desc));
319 hns3_init_tx_queue_hw(struct hns3_tx_queue *txq)
321 uint64_t dma_addr = txq->tx_ring_phys_addr;
323 hns3_write_dev(txq, HNS3_RING_TX_BASEADDR_L_REG, (uint32_t)dma_addr);
324 hns3_write_dev(txq, HNS3_RING_TX_BASEADDR_H_REG,
325 (uint32_t)((dma_addr >> 31) >> 1));
327 hns3_write_dev(txq, HNS3_RING_TX_BD_NUM_REG,
328 HNS3_CFG_DESC_NUM(txq->nb_tx_desc));
332 hns3_update_all_queues_pvid_proc_en(struct hns3_hw *hw)
334 uint16_t nb_rx_q = hw->data->nb_rx_queues;
335 uint16_t nb_tx_q = hw->data->nb_tx_queues;
336 struct hns3_rx_queue *rxq;
337 struct hns3_tx_queue *txq;
341 pvid_en = hw->port_base_vlan_cfg.state == HNS3_PORT_BASE_VLAN_ENABLE;
342 for (i = 0; i < hw->cfg_max_queues; i++) {
344 rxq = hw->data->rx_queues[i];
346 rxq->pvid_sw_discard_en = pvid_en;
349 txq = hw->data->tx_queues[i];
351 txq->pvid_sw_shift_en = pvid_en;
357 hns3_enable_all_queues(struct hns3_hw *hw, bool en)
359 uint16_t nb_rx_q = hw->data->nb_rx_queues;
360 uint16_t nb_tx_q = hw->data->nb_tx_queues;
361 struct hns3_rx_queue *rxq;
362 struct hns3_tx_queue *txq;
367 for (i = 0; i < hw->cfg_max_queues; i++) {
368 if (hns3_dev_indep_txrx_supported(hw)) {
369 rxq = i < nb_rx_q ? hw->data->rx_queues[i] : NULL;
370 txq = i < nb_tx_q ? hw->data->tx_queues[i] : NULL;
372 * After initialization, rxq and txq won't be NULL at
376 tqp_base = rxq->io_base;
377 else if (txq != NULL)
378 tqp_base = txq->io_base;
382 rxq = i < nb_rx_q ? hw->data->rx_queues[i] :
383 hw->fkq_data.rx_queues[i - nb_rx_q];
385 tqp_base = rxq->io_base;
388 * This is the master switch that used to control the enabling
389 * of a pair of Tx and Rx queues. Both the Rx and Tx point to
392 rcb_reg = hns3_read_reg(tqp_base, HNS3_RING_EN_REG);
394 rcb_reg |= BIT(HNS3_RING_EN_B);
396 rcb_reg &= ~BIT(HNS3_RING_EN_B);
397 hns3_write_reg(tqp_base, HNS3_RING_EN_REG, rcb_reg);
402 hns3_enable_txq(struct hns3_tx_queue *txq, bool en)
404 struct hns3_hw *hw = &txq->hns->hw;
407 if (hns3_dev_indep_txrx_supported(hw)) {
408 reg = hns3_read_dev(txq, HNS3_RING_TX_EN_REG);
410 reg |= BIT(HNS3_RING_EN_B);
412 reg &= ~BIT(HNS3_RING_EN_B);
413 hns3_write_dev(txq, HNS3_RING_TX_EN_REG, reg);
419 hns3_enable_rxq(struct hns3_rx_queue *rxq, bool en)
421 struct hns3_hw *hw = &rxq->hns->hw;
424 if (hns3_dev_indep_txrx_supported(hw)) {
425 reg = hns3_read_dev(rxq, HNS3_RING_RX_EN_REG);
427 reg |= BIT(HNS3_RING_EN_B);
429 reg &= ~BIT(HNS3_RING_EN_B);
430 hns3_write_dev(rxq, HNS3_RING_RX_EN_REG, reg);
436 hns3_start_all_txqs(struct rte_eth_dev *dev)
438 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
439 struct hns3_tx_queue *txq;
442 for (i = 0; i < dev->data->nb_tx_queues; i++) {
443 txq = hw->data->tx_queues[i];
445 hns3_err(hw, "Tx queue %u not available or setup.", i);
446 goto start_txqs_fail;
449 * Tx queue is enabled by default. Therefore, the Tx queues
450 * needs to be disabled when deferred_start is set. There is
451 * another master switch used to control the enabling of a pair
452 * of Tx and Rx queues. And the master switch is disabled by
455 if (txq->tx_deferred_start)
456 hns3_enable_txq(txq, false);
458 hns3_enable_txq(txq, true);
463 for (j = 0; j < i; j++) {
464 txq = hw->data->tx_queues[j];
465 hns3_enable_txq(txq, false);
471 hns3_start_all_rxqs(struct rte_eth_dev *dev)
473 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
474 struct hns3_rx_queue *rxq;
477 for (i = 0; i < dev->data->nb_rx_queues; i++) {
478 rxq = hw->data->rx_queues[i];
480 hns3_err(hw, "Rx queue %u not available or setup.", i);
481 goto start_rxqs_fail;
484 * Rx queue is enabled by default. Therefore, the Rx queues
485 * needs to be disabled when deferred_start is set. There is
486 * another master switch used to control the enabling of a pair
487 * of Tx and Rx queues. And the master switch is disabled by
490 if (rxq->rx_deferred_start)
491 hns3_enable_rxq(rxq, false);
493 hns3_enable_rxq(rxq, true);
498 for (j = 0; j < i; j++) {
499 rxq = hw->data->rx_queues[j];
500 hns3_enable_rxq(rxq, false);
506 hns3_stop_all_txqs(struct rte_eth_dev *dev)
508 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
509 struct hns3_tx_queue *txq;
512 for (i = 0; i < dev->data->nb_tx_queues; i++) {
513 txq = hw->data->tx_queues[i];
516 hns3_enable_txq(txq, false);
521 hns3_tqp_enable(struct hns3_hw *hw, uint16_t queue_id, bool enable)
523 struct hns3_cfg_com_tqp_queue_cmd *req;
524 struct hns3_cmd_desc desc;
527 req = (struct hns3_cfg_com_tqp_queue_cmd *)desc.data;
529 hns3_cmd_setup_basic_desc(&desc, HNS3_OPC_CFG_COM_TQP_QUEUE, false);
530 req->tqp_id = rte_cpu_to_le_16(queue_id);
532 hns3_set_bit(req->enable, HNS3_TQP_ENABLE_B, enable ? 1 : 0);
534 ret = hns3_cmd_send(hw, &desc, 1);
536 hns3_err(hw, "TQP enable fail, ret = %d", ret);
542 hns3_send_reset_tqp_cmd(struct hns3_hw *hw, uint16_t queue_id, bool enable)
544 struct hns3_reset_tqp_queue_cmd *req;
545 struct hns3_cmd_desc desc;
548 hns3_cmd_setup_basic_desc(&desc, HNS3_OPC_RESET_TQP_QUEUE, false);
550 req = (struct hns3_reset_tqp_queue_cmd *)desc.data;
551 req->tqp_id = rte_cpu_to_le_16(queue_id);
552 hns3_set_bit(req->reset_req, HNS3_TQP_RESET_B, enable ? 1 : 0);
553 ret = hns3_cmd_send(hw, &desc, 1);
555 hns3_err(hw, "send tqp reset cmd error, queue_id = %u, "
556 "ret = %d", queue_id, ret);
562 hns3_get_tqp_reset_status(struct hns3_hw *hw, uint16_t queue_id,
563 uint8_t *reset_status)
565 struct hns3_reset_tqp_queue_cmd *req;
566 struct hns3_cmd_desc desc;
569 hns3_cmd_setup_basic_desc(&desc, HNS3_OPC_RESET_TQP_QUEUE, true);
571 req = (struct hns3_reset_tqp_queue_cmd *)desc.data;
572 req->tqp_id = rte_cpu_to_le_16(queue_id);
574 ret = hns3_cmd_send(hw, &desc, 1);
576 hns3_err(hw, "get tqp reset status error, queue_id = %u, "
577 "ret = %d.", queue_id, ret);
580 *reset_status = hns3_get_bit(req->ready_to_reset, HNS3_TQP_RESET_B);
585 hns3pf_reset_tqp(struct hns3_hw *hw, uint16_t queue_id)
587 #define HNS3_TQP_RESET_TRY_MS 200
588 uint8_t reset_status;
592 ret = hns3_tqp_enable(hw, queue_id, false);
597 * In current version VF is not supported when PF is driven by DPDK
598 * driver, all task queue pairs are mapped to PF function, so PF's queue
599 * id is equals to the global queue id in PF range.
601 ret = hns3_send_reset_tqp_cmd(hw, queue_id, true);
603 hns3_err(hw, "Send reset tqp cmd fail, ret = %d", ret);
606 end = get_timeofday_ms() + HNS3_TQP_RESET_TRY_MS;
608 /* Wait for tqp hw reset */
609 rte_delay_ms(HNS3_POLL_RESPONE_MS);
610 ret = hns3_get_tqp_reset_status(hw, queue_id, &reset_status);
616 } while (get_timeofday_ms() < end);
620 hns3_err(hw, "reset tqp timeout, queue_id = %u, ret = %d",
625 ret = hns3_send_reset_tqp_cmd(hw, queue_id, false);
627 hns3_err(hw, "Deassert the soft reset fail, ret = %d", ret);
632 hns3_send_reset_tqp_cmd(hw, queue_id, false);
637 hns3vf_reset_tqp(struct hns3_hw *hw, uint16_t queue_id)
642 /* Disable VF's queue before send queue reset msg to PF */
643 ret = hns3_tqp_enable(hw, queue_id, false);
647 memcpy(msg_data, &queue_id, sizeof(uint16_t));
649 ret = hns3_send_mbx_msg(hw, HNS3_MBX_QUEUE_RESET, 0, msg_data,
650 sizeof(msg_data), true, NULL, 0);
652 hns3_err(hw, "fail to reset tqp, queue_id = %u, ret = %d.",
658 hns3_reset_tqp(struct hns3_adapter *hns, uint16_t queue_id)
660 struct hns3_hw *hw = &hns->hw;
663 return hns3vf_reset_tqp(hw, queue_id);
665 return hns3pf_reset_tqp(hw, queue_id);
669 hns3_reset_all_tqps(struct hns3_adapter *hns)
671 struct hns3_hw *hw = &hns->hw;
674 for (i = 0; i < hw->cfg_max_queues; i++) {
675 ret = hns3_reset_tqp(hns, i);
677 hns3_err(hw, "Failed to reset No.%d queue: %d", i, ret);
685 hns3_send_reset_queue_cmd(struct hns3_hw *hw, uint16_t queue_id,
686 enum hns3_ring_type queue_type, bool enable)
688 struct hns3_reset_tqp_queue_cmd *req;
689 struct hns3_cmd_desc desc;
693 hns3_cmd_setup_basic_desc(&desc, HNS3_OPC_RESET_TQP_QUEUE_INDEP, false);
695 req = (struct hns3_reset_tqp_queue_cmd *)desc.data;
696 req->tqp_id = rte_cpu_to_le_16(queue_id);
697 queue_direction = queue_type == HNS3_RING_TYPE_TX ? 0 : 1;
698 req->queue_direction = rte_cpu_to_le_16(queue_direction);
699 hns3_set_bit(req->reset_req, HNS3_TQP_RESET_B, enable ? 1 : 0);
701 ret = hns3_cmd_send(hw, &desc, 1);
703 hns3_err(hw, "send queue reset cmd error, queue_id = %u, "
704 "queue_type = %s, ret = %d.", queue_id,
705 queue_type == HNS3_RING_TYPE_TX ? "Tx" : "Rx", ret);
710 hns3_get_queue_reset_status(struct hns3_hw *hw, uint16_t queue_id,
711 enum hns3_ring_type queue_type,
712 uint8_t *reset_status)
714 struct hns3_reset_tqp_queue_cmd *req;
715 struct hns3_cmd_desc desc;
719 hns3_cmd_setup_basic_desc(&desc, HNS3_OPC_RESET_TQP_QUEUE_INDEP, true);
721 req = (struct hns3_reset_tqp_queue_cmd *)desc.data;
722 req->tqp_id = rte_cpu_to_le_16(queue_id);
723 queue_direction = queue_type == HNS3_RING_TYPE_TX ? 0 : 1;
724 req->queue_direction = rte_cpu_to_le_16(queue_direction);
726 ret = hns3_cmd_send(hw, &desc, 1);
728 hns3_err(hw, "get queue reset status error, queue_id = %u "
729 "queue_type = %s, ret = %d.", queue_id,
730 queue_type == HNS3_RING_TYPE_TX ? "Tx" : "Rx", ret);
734 *reset_status = hns3_get_bit(req->ready_to_reset, HNS3_TQP_RESET_B);
739 hns3_reset_queue(struct hns3_hw *hw, uint16_t queue_id,
740 enum hns3_ring_type queue_type)
742 #define HNS3_QUEUE_RESET_TRY_MS 200
743 struct hns3_tx_queue *txq;
744 struct hns3_rx_queue *rxq;
745 uint32_t reset_wait_times;
746 uint32_t max_wait_times;
747 uint8_t reset_status;
750 if (queue_type == HNS3_RING_TYPE_TX) {
751 txq = hw->data->tx_queues[queue_id];
752 hns3_enable_txq(txq, false);
754 rxq = hw->data->rx_queues[queue_id];
755 hns3_enable_rxq(rxq, false);
758 ret = hns3_send_reset_queue_cmd(hw, queue_id, queue_type, true);
760 hns3_err(hw, "send reset queue cmd fail, ret = %d.", ret);
764 reset_wait_times = 0;
765 max_wait_times = HNS3_QUEUE_RESET_TRY_MS / HNS3_POLL_RESPONE_MS;
766 while (reset_wait_times < max_wait_times) {
767 /* Wait for queue hw reset */
768 rte_delay_ms(HNS3_POLL_RESPONE_MS);
769 ret = hns3_get_queue_reset_status(hw, queue_id,
770 queue_type, &reset_status);
772 goto queue_reset_fail;
780 hns3_err(hw, "reset queue timeout, queue_id = %u, "
781 "queue_type = %s", queue_id,
782 queue_type == HNS3_RING_TYPE_TX ? "Tx" : "Rx");
784 goto queue_reset_fail;
787 ret = hns3_send_reset_queue_cmd(hw, queue_id, queue_type, false);
789 hns3_err(hw, "deassert queue reset fail, ret = %d.", ret);
794 hns3_send_reset_queue_cmd(hw, queue_id, queue_type, false);
800 hns3_set_queue_intr_gl(struct hns3_hw *hw, uint16_t queue_id,
801 uint8_t gl_idx, uint16_t gl_value)
803 uint32_t offset[] = {HNS3_TQP_INTR_GL0_REG,
804 HNS3_TQP_INTR_GL1_REG,
805 HNS3_TQP_INTR_GL2_REG};
806 uint32_t addr, value;
808 if (gl_idx >= RTE_DIM(offset) || gl_value > HNS3_TQP_INTR_GL_MAX)
811 addr = offset[gl_idx] + queue_id * HNS3_TQP_INTR_REG_SIZE;
812 if (hw->intr.gl_unit == HNS3_INTR_COALESCE_GL_UINT_1US)
813 value = gl_value | HNS3_TQP_INTR_GL_UNIT_1US;
815 value = HNS3_GL_USEC_TO_REG(gl_value);
817 hns3_write_dev(hw, addr, value);
821 hns3_set_queue_intr_rl(struct hns3_hw *hw, uint16_t queue_id, uint16_t rl_value)
823 uint32_t addr, value;
825 if (rl_value > HNS3_TQP_INTR_RL_MAX)
828 addr = HNS3_TQP_INTR_RL_REG + queue_id * HNS3_TQP_INTR_REG_SIZE;
829 value = HNS3_RL_USEC_TO_REG(rl_value);
831 value |= HNS3_TQP_INTR_RL_ENABLE_MASK;
833 hns3_write_dev(hw, addr, value);
837 hns3_set_queue_intr_ql(struct hns3_hw *hw, uint16_t queue_id, uint16_t ql_value)
842 * int_ql_max == 0 means the hardware does not support QL,
843 * QL regs config is not permitted if QL is not supported,
846 if (hw->intr.int_ql_max == HNS3_INTR_QL_NONE)
849 addr = HNS3_TQP_INTR_TX_QL_REG + queue_id * HNS3_TQP_INTR_REG_SIZE;
850 hns3_write_dev(hw, addr, ql_value);
852 addr = HNS3_TQP_INTR_RX_QL_REG + queue_id * HNS3_TQP_INTR_REG_SIZE;
853 hns3_write_dev(hw, addr, ql_value);
857 hns3_queue_intr_enable(struct hns3_hw *hw, uint16_t queue_id, bool en)
859 uint32_t addr, value;
861 addr = HNS3_TQP_INTR_CTRL_REG + queue_id * HNS3_TQP_INTR_REG_SIZE;
864 hns3_write_dev(hw, addr, value);
868 * Enable all rx queue interrupt when in interrupt rx mode.
869 * This api was called before enable queue rx&tx (in normal start or reset
870 * recover scenes), used to fix hardware rx queue interrupt enable was clear
874 hns3_dev_all_rx_queue_intr_enable(struct hns3_hw *hw, bool en)
876 struct rte_eth_dev *dev = &rte_eth_devices[hw->data->port_id];
877 uint16_t nb_rx_q = hw->data->nb_rx_queues;
880 if (dev->data->dev_conf.intr_conf.rxq == 0)
883 for (i = 0; i < nb_rx_q; i++)
884 hns3_queue_intr_enable(hw, i, en);
888 hns3_dev_rx_queue_intr_enable(struct rte_eth_dev *dev, uint16_t queue_id)
890 struct rte_pci_device *pci_dev = RTE_ETH_DEV_TO_PCI(dev);
891 struct rte_intr_handle *intr_handle = &pci_dev->intr_handle;
892 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
894 if (dev->data->dev_conf.intr_conf.rxq == 0)
897 hns3_queue_intr_enable(hw, queue_id, true);
899 return rte_intr_ack(intr_handle);
903 hns3_dev_rx_queue_intr_disable(struct rte_eth_dev *dev, uint16_t queue_id)
905 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
907 if (dev->data->dev_conf.intr_conf.rxq == 0)
910 hns3_queue_intr_enable(hw, queue_id, false);
916 hns3_init_rxq(struct hns3_adapter *hns, uint16_t idx)
918 struct hns3_hw *hw = &hns->hw;
919 struct hns3_rx_queue *rxq;
922 PMD_INIT_FUNC_TRACE();
924 rxq = (struct hns3_rx_queue *)hw->data->rx_queues[idx];
925 ret = hns3_alloc_rx_queue_mbufs(hw, rxq);
927 hns3_err(hw, "fail to alloc mbuf for Rx queue %u, ret = %d.",
932 rxq->next_to_use = 0;
933 rxq->rx_rearm_start = 0;
934 rxq->rx_free_hold = 0;
935 rxq->rx_rearm_nb = 0;
936 rxq->pkt_first_seg = NULL;
937 rxq->pkt_last_seg = NULL;
938 hns3_init_rx_queue_hw(rxq);
939 hns3_rxq_vec_setup(rxq);
945 hns3_init_fake_rxq(struct hns3_adapter *hns, uint16_t idx)
947 struct hns3_hw *hw = &hns->hw;
948 struct hns3_rx_queue *rxq;
950 rxq = (struct hns3_rx_queue *)hw->fkq_data.rx_queues[idx];
951 rxq->next_to_use = 0;
952 rxq->rx_free_hold = 0;
953 rxq->rx_rearm_start = 0;
954 rxq->rx_rearm_nb = 0;
955 hns3_init_rx_queue_hw(rxq);
959 hns3_init_txq(struct hns3_tx_queue *txq)
961 struct hns3_desc *desc;
966 for (i = 0; i < txq->nb_tx_desc; i++) {
967 desc->tx.tp_fe_sc_vld_ra_ri = 0;
971 txq->next_to_use = 0;
972 txq->next_to_clean = 0;
973 txq->tx_bd_ready = txq->nb_tx_desc - 1;
974 hns3_init_tx_queue_hw(txq);
978 hns3_init_tx_ring_tc(struct hns3_adapter *hns)
980 struct hns3_hw *hw = &hns->hw;
981 struct hns3_tx_queue *txq;
984 for (i = 0; i < HNS3_MAX_TC_NUM; i++) {
985 struct hns3_tc_queue_info *tc_queue = &hw->tc_queue[i];
988 if (!tc_queue->enable)
991 for (j = 0; j < tc_queue->tqp_count; j++) {
992 num = tc_queue->tqp_offset + j;
993 txq = (struct hns3_tx_queue *)hw->data->tx_queues[num];
997 hns3_write_dev(txq, HNS3_RING_TX_TC_REG, tc_queue->tc);
1003 hns3_init_rx_queues(struct hns3_adapter *hns)
1005 struct hns3_hw *hw = &hns->hw;
1006 struct hns3_rx_queue *rxq;
1010 /* Initialize RSS for queues */
1011 ret = hns3_config_rss(hns);
1013 hns3_err(hw, "failed to configure rss, ret = %d.", ret);
1017 for (i = 0; i < hw->data->nb_rx_queues; i++) {
1018 rxq = (struct hns3_rx_queue *)hw->data->rx_queues[i];
1020 hns3_err(hw, "Rx queue %u not available or setup.", i);
1024 if (rxq->rx_deferred_start)
1027 ret = hns3_init_rxq(hns, i);
1029 hns3_err(hw, "failed to init Rx queue %u, ret = %d.", i,
1035 for (i = 0; i < hw->fkq_data.nb_fake_rx_queues; i++)
1036 hns3_init_fake_rxq(hns, i);
1041 for (j = 0; j < i; j++) {
1042 rxq = (struct hns3_rx_queue *)hw->data->rx_queues[j];
1043 hns3_rx_queue_release_mbufs(rxq);
1050 hns3_init_tx_queues(struct hns3_adapter *hns)
1052 struct hns3_hw *hw = &hns->hw;
1053 struct hns3_tx_queue *txq;
1056 for (i = 0; i < hw->data->nb_tx_queues; i++) {
1057 txq = (struct hns3_tx_queue *)hw->data->tx_queues[i];
1059 hns3_err(hw, "Tx queue %u not available or setup.", i);
1063 if (txq->tx_deferred_start)
1068 for (i = 0; i < hw->fkq_data.nb_fake_tx_queues; i++) {
1069 txq = (struct hns3_tx_queue *)hw->fkq_data.tx_queues[i];
1072 hns3_init_tx_ring_tc(hns);
1079 * Note: just init and setup queues, and don't enable tqps.
1082 hns3_init_queues(struct hns3_adapter *hns, bool reset_queue)
1084 struct hns3_hw *hw = &hns->hw;
1088 ret = hns3_reset_all_tqps(hns);
1090 hns3_err(hw, "failed to reset all queues, ret = %d.",
1096 ret = hns3_init_rx_queues(hns);
1098 hns3_err(hw, "failed to init rx queues, ret = %d.", ret);
1102 ret = hns3_init_tx_queues(hns);
1104 hns3_dev_release_mbufs(hns);
1105 hns3_err(hw, "failed to init tx queues, ret = %d.", ret);
1112 hns3_start_tqps(struct hns3_hw *hw)
1114 struct hns3_tx_queue *txq;
1115 struct hns3_rx_queue *rxq;
1118 hns3_enable_all_queues(hw, true);
1120 for (i = 0; i < hw->data->nb_tx_queues; i++) {
1121 txq = hw->data->tx_queues[i];
1123 hw->data->tx_queue_state[i] =
1124 RTE_ETH_QUEUE_STATE_STARTED;
1127 for (i = 0; i < hw->data->nb_rx_queues; i++) {
1128 rxq = hw->data->rx_queues[i];
1130 hw->data->rx_queue_state[i] =
1131 RTE_ETH_QUEUE_STATE_STARTED;
1136 hns3_stop_tqps(struct hns3_hw *hw)
1140 hns3_enable_all_queues(hw, false);
1142 for (i = 0; i < hw->data->nb_tx_queues; i++)
1143 hw->data->tx_queue_state[i] = RTE_ETH_QUEUE_STATE_STOPPED;
1145 for (i = 0; i < hw->data->nb_rx_queues; i++)
1146 hw->data->rx_queue_state[i] = RTE_ETH_QUEUE_STATE_STOPPED;
1150 * Iterate over all Rx Queue, and call the callback() function for each Rx
1154 * The target eth dev.
1155 * @param[in] callback
1156 * The function to call for each queue.
1157 * if callback function return nonzero will stop iterate and return it's value
1159 * The arguments to provide the callback function with.
1162 * 0 on success, otherwise with errno set.
1165 hns3_rxq_iterate(struct rte_eth_dev *dev,
1166 int (*callback)(struct hns3_rx_queue *, void *), void *arg)
1171 if (dev->data->rx_queues == NULL)
1174 for (i = 0; i < dev->data->nb_rx_queues; i++) {
1175 ret = callback(dev->data->rx_queues[i], arg);
1184 hns3_alloc_rxq_and_dma_zone(struct rte_eth_dev *dev,
1185 struct hns3_queue_info *q_info)
1187 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1188 const struct rte_memzone *rx_mz;
1189 struct hns3_rx_queue *rxq;
1190 unsigned int rx_desc;
1192 rxq = rte_zmalloc_socket(q_info->type, sizeof(struct hns3_rx_queue),
1193 RTE_CACHE_LINE_SIZE, q_info->socket_id);
1195 hns3_err(hw, "Failed to allocate memory for No.%u rx ring!",
1200 /* Allocate rx ring hardware descriptors. */
1201 rxq->queue_id = q_info->idx;
1202 rxq->nb_rx_desc = q_info->nb_desc;
1205 * Allocate a litter more memory because rx vector functions
1206 * don't check boundaries each time.
1208 rx_desc = (rxq->nb_rx_desc + HNS3_DEFAULT_RX_BURST) *
1209 sizeof(struct hns3_desc);
1210 rx_mz = rte_eth_dma_zone_reserve(dev, q_info->ring_name, q_info->idx,
1211 rx_desc, HNS3_RING_BASE_ALIGN,
1213 if (rx_mz == NULL) {
1214 hns3_err(hw, "Failed to reserve DMA memory for No.%u rx ring!",
1216 hns3_rx_queue_release(rxq);
1220 rxq->rx_ring = (struct hns3_desc *)rx_mz->addr;
1221 rxq->rx_ring_phys_addr = rx_mz->iova;
1223 hns3_dbg(hw, "No.%u rx descriptors iova 0x%" PRIx64, q_info->idx,
1224 rxq->rx_ring_phys_addr);
1230 hns3_fake_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx,
1231 uint16_t nb_desc, unsigned int socket_id)
1233 struct hns3_adapter *hns = dev->data->dev_private;
1234 struct hns3_hw *hw = &hns->hw;
1235 struct hns3_queue_info q_info;
1236 struct hns3_rx_queue *rxq;
1239 if (hw->fkq_data.rx_queues[idx]) {
1240 hns3_rx_queue_release(hw->fkq_data.rx_queues[idx]);
1241 hw->fkq_data.rx_queues[idx] = NULL;
1245 q_info.socket_id = socket_id;
1246 q_info.nb_desc = nb_desc;
1247 q_info.type = "hns3 fake RX queue";
1248 q_info.ring_name = "rx_fake_ring";
1249 rxq = hns3_alloc_rxq_and_dma_zone(dev, &q_info);
1251 hns3_err(hw, "Failed to setup No.%u fake rx ring.", idx);
1255 /* Don't need alloc sw_ring, because upper applications don't use it */
1256 rxq->sw_ring = NULL;
1259 rxq->rx_deferred_start = false;
1260 rxq->port_id = dev->data->port_id;
1261 rxq->configured = true;
1262 nb_rx_q = dev->data->nb_rx_queues;
1263 rxq->io_base = (void *)((char *)hw->io_base + HNS3_TQP_REG_OFFSET +
1264 (nb_rx_q + idx) * HNS3_TQP_REG_SIZE);
1265 rxq->rx_buf_len = HNS3_MIN_BD_BUF_SIZE;
1267 rte_spinlock_lock(&hw->lock);
1268 hw->fkq_data.rx_queues[idx] = rxq;
1269 rte_spinlock_unlock(&hw->lock);
1275 hns3_alloc_txq_and_dma_zone(struct rte_eth_dev *dev,
1276 struct hns3_queue_info *q_info)
1278 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1279 const struct rte_memzone *tx_mz;
1280 struct hns3_tx_queue *txq;
1281 struct hns3_desc *desc;
1282 unsigned int tx_desc;
1285 txq = rte_zmalloc_socket(q_info->type, sizeof(struct hns3_tx_queue),
1286 RTE_CACHE_LINE_SIZE, q_info->socket_id);
1288 hns3_err(hw, "Failed to allocate memory for No.%u tx ring!",
1293 /* Allocate tx ring hardware descriptors. */
1294 txq->queue_id = q_info->idx;
1295 txq->nb_tx_desc = q_info->nb_desc;
1296 tx_desc = txq->nb_tx_desc * sizeof(struct hns3_desc);
1297 tx_mz = rte_eth_dma_zone_reserve(dev, q_info->ring_name, q_info->idx,
1298 tx_desc, HNS3_RING_BASE_ALIGN,
1300 if (tx_mz == NULL) {
1301 hns3_err(hw, "Failed to reserve DMA memory for No.%u tx ring!",
1303 hns3_tx_queue_release(txq);
1307 txq->tx_ring = (struct hns3_desc *)tx_mz->addr;
1308 txq->tx_ring_phys_addr = tx_mz->iova;
1310 hns3_dbg(hw, "No.%u tx descriptors iova 0x%" PRIx64, q_info->idx,
1311 txq->tx_ring_phys_addr);
1314 desc = txq->tx_ring;
1315 for (i = 0; i < txq->nb_tx_desc; i++) {
1316 desc->tx.tp_fe_sc_vld_ra_ri = 0;
1324 hns3_fake_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx,
1325 uint16_t nb_desc, unsigned int socket_id)
1327 struct hns3_adapter *hns = dev->data->dev_private;
1328 struct hns3_hw *hw = &hns->hw;
1329 struct hns3_queue_info q_info;
1330 struct hns3_tx_queue *txq;
1333 if (hw->fkq_data.tx_queues[idx] != NULL) {
1334 hns3_tx_queue_release(hw->fkq_data.tx_queues[idx]);
1335 hw->fkq_data.tx_queues[idx] = NULL;
1339 q_info.socket_id = socket_id;
1340 q_info.nb_desc = nb_desc;
1341 q_info.type = "hns3 fake TX queue";
1342 q_info.ring_name = "tx_fake_ring";
1343 txq = hns3_alloc_txq_and_dma_zone(dev, &q_info);
1345 hns3_err(hw, "Failed to setup No.%u fake tx ring.", idx);
1349 /* Don't need alloc sw_ring, because upper applications don't use it */
1350 txq->sw_ring = NULL;
1354 txq->tx_deferred_start = false;
1355 txq->port_id = dev->data->port_id;
1356 txq->configured = true;
1357 nb_tx_q = dev->data->nb_tx_queues;
1358 txq->io_base = (void *)((char *)hw->io_base + HNS3_TQP_REG_OFFSET +
1359 (nb_tx_q + idx) * HNS3_TQP_REG_SIZE);
1361 rte_spinlock_lock(&hw->lock);
1362 hw->fkq_data.tx_queues[idx] = txq;
1363 rte_spinlock_unlock(&hw->lock);
1369 hns3_fake_rx_queue_config(struct hns3_hw *hw, uint16_t nb_queues)
1371 uint16_t old_nb_queues = hw->fkq_data.nb_fake_rx_queues;
1375 if (hw->fkq_data.rx_queues == NULL && nb_queues != 0) {
1376 /* first time configuration */
1378 size = sizeof(hw->fkq_data.rx_queues[0]) * nb_queues;
1379 hw->fkq_data.rx_queues = rte_zmalloc("fake_rx_queues", size,
1380 RTE_CACHE_LINE_SIZE);
1381 if (hw->fkq_data.rx_queues == NULL) {
1382 hw->fkq_data.nb_fake_rx_queues = 0;
1385 } else if (hw->fkq_data.rx_queues != NULL && nb_queues != 0) {
1387 rxq = hw->fkq_data.rx_queues;
1388 for (i = nb_queues; i < old_nb_queues; i++)
1389 hns3_dev_rx_queue_release(rxq[i]);
1391 rxq = rte_realloc(rxq, sizeof(rxq[0]) * nb_queues,
1392 RTE_CACHE_LINE_SIZE);
1395 if (nb_queues > old_nb_queues) {
1396 uint16_t new_qs = nb_queues - old_nb_queues;
1397 memset(rxq + old_nb_queues, 0, sizeof(rxq[0]) * new_qs);
1400 hw->fkq_data.rx_queues = rxq;
1401 } else if (hw->fkq_data.rx_queues != NULL && nb_queues == 0) {
1402 rxq = hw->fkq_data.rx_queues;
1403 for (i = nb_queues; i < old_nb_queues; i++)
1404 hns3_dev_rx_queue_release(rxq[i]);
1406 rte_free(hw->fkq_data.rx_queues);
1407 hw->fkq_data.rx_queues = NULL;
1410 hw->fkq_data.nb_fake_rx_queues = nb_queues;
1416 hns3_fake_tx_queue_config(struct hns3_hw *hw, uint16_t nb_queues)
1418 uint16_t old_nb_queues = hw->fkq_data.nb_fake_tx_queues;
1422 if (hw->fkq_data.tx_queues == NULL && nb_queues != 0) {
1423 /* first time configuration */
1425 size = sizeof(hw->fkq_data.tx_queues[0]) * nb_queues;
1426 hw->fkq_data.tx_queues = rte_zmalloc("fake_tx_queues", size,
1427 RTE_CACHE_LINE_SIZE);
1428 if (hw->fkq_data.tx_queues == NULL) {
1429 hw->fkq_data.nb_fake_tx_queues = 0;
1432 } else if (hw->fkq_data.tx_queues != NULL && nb_queues != 0) {
1434 txq = hw->fkq_data.tx_queues;
1435 for (i = nb_queues; i < old_nb_queues; i++)
1436 hns3_dev_tx_queue_release(txq[i]);
1437 txq = rte_realloc(txq, sizeof(txq[0]) * nb_queues,
1438 RTE_CACHE_LINE_SIZE);
1441 if (nb_queues > old_nb_queues) {
1442 uint16_t new_qs = nb_queues - old_nb_queues;
1443 memset(txq + old_nb_queues, 0, sizeof(txq[0]) * new_qs);
1446 hw->fkq_data.tx_queues = txq;
1447 } else if (hw->fkq_data.tx_queues != NULL && nb_queues == 0) {
1448 txq = hw->fkq_data.tx_queues;
1449 for (i = nb_queues; i < old_nb_queues; i++)
1450 hns3_dev_tx_queue_release(txq[i]);
1452 rte_free(hw->fkq_data.tx_queues);
1453 hw->fkq_data.tx_queues = NULL;
1455 hw->fkq_data.nb_fake_tx_queues = nb_queues;
1461 hns3_set_fake_rx_or_tx_queues(struct rte_eth_dev *dev, uint16_t nb_rx_q,
1464 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1465 uint16_t rx_need_add_nb_q;
1466 uint16_t tx_need_add_nb_q;
1471 /* Setup new number of fake RX/TX queues and reconfigure device. */
1472 rx_need_add_nb_q = hw->cfg_max_queues - nb_rx_q;
1473 tx_need_add_nb_q = hw->cfg_max_queues - nb_tx_q;
1474 ret = hns3_fake_rx_queue_config(hw, rx_need_add_nb_q);
1476 hns3_err(hw, "Fail to configure fake rx queues: %d", ret);
1480 ret = hns3_fake_tx_queue_config(hw, tx_need_add_nb_q);
1482 hns3_err(hw, "Fail to configure fake rx queues: %d", ret);
1483 goto cfg_fake_tx_q_fail;
1486 /* Allocate and set up fake RX queue per Ethernet port. */
1487 port_id = hw->data->port_id;
1488 for (q = 0; q < rx_need_add_nb_q; q++) {
1489 ret = hns3_fake_rx_queue_setup(dev, q, HNS3_MIN_RING_DESC,
1490 rte_eth_dev_socket_id(port_id));
1492 goto setup_fake_rx_q_fail;
1495 /* Allocate and set up fake TX queue per Ethernet port. */
1496 for (q = 0; q < tx_need_add_nb_q; q++) {
1497 ret = hns3_fake_tx_queue_setup(dev, q, HNS3_MIN_RING_DESC,
1498 rte_eth_dev_socket_id(port_id));
1500 goto setup_fake_tx_q_fail;
1505 setup_fake_tx_q_fail:
1506 setup_fake_rx_q_fail:
1507 (void)hns3_fake_tx_queue_config(hw, 0);
1509 (void)hns3_fake_rx_queue_config(hw, 0);
1515 hns3_dev_release_mbufs(struct hns3_adapter *hns)
1517 struct rte_eth_dev_data *dev_data = hns->hw.data;
1518 struct hns3_rx_queue *rxq;
1519 struct hns3_tx_queue *txq;
1522 if (dev_data->rx_queues)
1523 for (i = 0; i < dev_data->nb_rx_queues; i++) {
1524 rxq = dev_data->rx_queues[i];
1527 hns3_rx_queue_release_mbufs(rxq);
1530 if (dev_data->tx_queues)
1531 for (i = 0; i < dev_data->nb_tx_queues; i++) {
1532 txq = dev_data->tx_queues[i];
1535 hns3_tx_queue_release_mbufs(txq);
1540 hns3_rx_buf_len_calc(struct rte_mempool *mp, uint16_t *rx_buf_len)
1542 uint16_t vld_buf_size;
1543 uint16_t num_hw_specs;
1547 * hns3 network engine only support to set 4 typical specification, and
1548 * different buffer size will affect the max packet_len and the max
1549 * number of segmentation when hw gro is turned on in receive side. The
1550 * relationship between them is as follows:
1551 * rx_buf_size | max_gro_pkt_len | max_gro_nb_seg
1552 * ---------------------|-------------------|----------------
1553 * HNS3_4K_BD_BUF_SIZE | 60KB | 15
1554 * HNS3_2K_BD_BUF_SIZE | 62KB | 31
1555 * HNS3_1K_BD_BUF_SIZE | 63KB | 63
1556 * HNS3_512_BD_BUF_SIZE | 31.5KB | 63
1558 static const uint16_t hw_rx_buf_size[] = {
1559 HNS3_4K_BD_BUF_SIZE,
1560 HNS3_2K_BD_BUF_SIZE,
1561 HNS3_1K_BD_BUF_SIZE,
1562 HNS3_512_BD_BUF_SIZE
1565 vld_buf_size = (uint16_t)(rte_pktmbuf_data_room_size(mp) -
1566 RTE_PKTMBUF_HEADROOM);
1568 if (vld_buf_size < HNS3_MIN_BD_BUF_SIZE)
1571 num_hw_specs = RTE_DIM(hw_rx_buf_size);
1572 for (i = 0; i < num_hw_specs; i++) {
1573 if (vld_buf_size >= hw_rx_buf_size[i]) {
1574 *rx_buf_len = hw_rx_buf_size[i];
1582 hns3_rxq_conf_runtime_check(struct hns3_hw *hw, uint16_t buf_size,
1585 struct rte_eth_dev *dev = &rte_eth_devices[hw->data->port_id];
1586 struct rte_eth_rxmode *rxmode = &hw->data->dev_conf.rxmode;
1587 eth_rx_burst_t pkt_burst = dev->rx_pkt_burst;
1588 uint16_t min_vec_bds;
1591 * HNS3 hardware network engine set scattered as default. If the driver
1592 * is not work in scattered mode and the pkts greater than buf_size
1593 * but smaller than max_rx_pkt_len will be distributed to multiple BDs.
1594 * Driver cannot handle this situation.
1596 if (!hw->data->scattered_rx && rxmode->max_rx_pkt_len > buf_size) {
1597 hns3_err(hw, "max_rx_pkt_len is not allowed to be set greater "
1598 "than rx_buf_len if scattered is off.");
1602 if (pkt_burst == hns3_recv_pkts_vec) {
1603 min_vec_bds = HNS3_DEFAULT_RXQ_REARM_THRESH +
1604 HNS3_DEFAULT_RX_BURST;
1605 if (nb_desc < min_vec_bds ||
1606 nb_desc % HNS3_DEFAULT_RXQ_REARM_THRESH) {
1607 hns3_err(hw, "if Rx burst mode is vector, "
1608 "number of descriptor is required to be "
1609 "bigger than min vector bds:%u, and could be "
1610 "divided by rxq rearm thresh:%u.",
1611 min_vec_bds, HNS3_DEFAULT_RXQ_REARM_THRESH);
1619 hns3_rx_queue_conf_check(struct hns3_hw *hw, const struct rte_eth_rxconf *conf,
1620 struct rte_mempool *mp, uint16_t nb_desc,
1625 if (nb_desc > HNS3_MAX_RING_DESC || nb_desc < HNS3_MIN_RING_DESC ||
1626 nb_desc % HNS3_ALIGN_RING_DESC) {
1627 hns3_err(hw, "Number (%u) of rx descriptors is invalid",
1632 if (conf->rx_drop_en == 0)
1633 hns3_warn(hw, "if no descriptors available, packets are always "
1634 "dropped and rx_drop_en (1) is fixed on");
1636 if (hns3_rx_buf_len_calc(mp, buf_size)) {
1637 hns3_err(hw, "rxq mbufs' data room size (%u) is not enough! "
1638 "minimal data room size (%u).",
1639 rte_pktmbuf_data_room_size(mp),
1640 HNS3_MIN_BD_BUF_SIZE + RTE_PKTMBUF_HEADROOM);
1644 if (hw->data->dev_started) {
1645 ret = hns3_rxq_conf_runtime_check(hw, *buf_size, nb_desc);
1647 hns3_err(hw, "Rx queue runtime setup fail.");
1656 hns3_get_tqp_reg_offset(uint16_t queue_id)
1658 uint32_t reg_offset;
1660 /* Need an extend offset to config queue > 1024 */
1661 if (queue_id < HNS3_MIN_EXTEND_QUEUE_ID)
1662 reg_offset = HNS3_TQP_REG_OFFSET + queue_id * HNS3_TQP_REG_SIZE;
1664 reg_offset = HNS3_TQP_REG_OFFSET + HNS3_TQP_EXT_REG_OFFSET +
1665 (queue_id - HNS3_MIN_EXTEND_QUEUE_ID) *
1672 hns3_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t nb_desc,
1673 unsigned int socket_id, const struct rte_eth_rxconf *conf,
1674 struct rte_mempool *mp)
1676 struct hns3_adapter *hns = dev->data->dev_private;
1677 struct hns3_hw *hw = &hns->hw;
1678 struct hns3_queue_info q_info;
1679 struct hns3_rx_queue *rxq;
1680 uint16_t rx_buf_size;
1684 ret = hns3_rx_queue_conf_check(hw, conf, mp, nb_desc, &rx_buf_size);
1688 if (dev->data->rx_queues[idx]) {
1689 hns3_rx_queue_release(dev->data->rx_queues[idx]);
1690 dev->data->rx_queues[idx] = NULL;
1694 q_info.socket_id = socket_id;
1695 q_info.nb_desc = nb_desc;
1696 q_info.type = "hns3 RX queue";
1697 q_info.ring_name = "rx_ring";
1699 rxq = hns3_alloc_rxq_and_dma_zone(dev, &q_info);
1702 "Failed to alloc mem and reserve DMA mem for rx ring!");
1707 rxq->ptype_tbl = &hns->ptype_tbl;
1709 rxq->rx_free_thresh = (conf->rx_free_thresh > 0) ?
1710 conf->rx_free_thresh : HNS3_DEFAULT_RX_FREE_THRESH;
1712 rxq->rx_deferred_start = conf->rx_deferred_start;
1713 if (rxq->rx_deferred_start && !hns3_dev_indep_txrx_supported(hw)) {
1714 hns3_warn(hw, "deferred start is not supported.");
1715 rxq->rx_deferred_start = false;
1718 rx_entry_len = (rxq->nb_rx_desc + HNS3_DEFAULT_RX_BURST) *
1719 sizeof(struct hns3_entry);
1720 rxq->sw_ring = rte_zmalloc_socket("hns3 RX sw ring", rx_entry_len,
1721 RTE_CACHE_LINE_SIZE, socket_id);
1722 if (rxq->sw_ring == NULL) {
1723 hns3_err(hw, "Failed to allocate memory for rx sw ring!");
1724 hns3_rx_queue_release(rxq);
1728 rxq->next_to_use = 0;
1729 rxq->rx_free_hold = 0;
1730 rxq->rx_rearm_start = 0;
1731 rxq->rx_rearm_nb = 0;
1732 rxq->pkt_first_seg = NULL;
1733 rxq->pkt_last_seg = NULL;
1734 rxq->port_id = dev->data->port_id;
1736 * For hns3 PF device, if the VLAN mode is HW_SHIFT_AND_DISCARD_MODE,
1737 * the pvid_sw_discard_en in the queue struct should not be changed,
1738 * because PVID-related operations do not need to be processed by PMD
1739 * driver. For hns3 VF device, whether it needs to process PVID depends
1740 * on the configuration of PF kernel mode netdevice driver. And the
1741 * related PF configuration is delivered through the mailbox and finally
1742 * reflectd in port_base_vlan_cfg.
1744 if (hns->is_vf || hw->vlan_mode == HNS3_SW_SHIFT_AND_DISCARD_MODE)
1745 rxq->pvid_sw_discard_en = hw->port_base_vlan_cfg.state ==
1746 HNS3_PORT_BASE_VLAN_ENABLE;
1748 rxq->pvid_sw_discard_en = false;
1749 rxq->configured = true;
1750 rxq->io_base = (void *)((char *)hw->io_base + HNS3_TQP_REG_OFFSET +
1751 idx * HNS3_TQP_REG_SIZE);
1752 rxq->io_base = (void *)((char *)hw->io_base +
1753 hns3_get_tqp_reg_offset(idx));
1754 rxq->io_head_reg = (volatile void *)((char *)rxq->io_base +
1755 HNS3_RING_RX_HEAD_REG);
1756 rxq->rx_buf_len = rx_buf_size;
1758 rxq->pkt_len_errors = 0;
1759 rxq->l3_csum_errors = 0;
1760 rxq->l4_csum_errors = 0;
1761 rxq->ol3_csum_errors = 0;
1762 rxq->ol4_csum_errors = 0;
1764 /* CRC len set here is used for amending packet length */
1765 if (dev->data->dev_conf.rxmode.offloads & DEV_RX_OFFLOAD_KEEP_CRC)
1766 rxq->crc_len = RTE_ETHER_CRC_LEN;
1770 rxq->bulk_mbuf_num = 0;
1772 rte_spinlock_lock(&hw->lock);
1773 dev->data->rx_queues[idx] = rxq;
1774 rte_spinlock_unlock(&hw->lock);
1780 hns3_rx_scattered_reset(struct rte_eth_dev *dev)
1782 struct hns3_adapter *hns = dev->data->dev_private;
1783 struct hns3_hw *hw = &hns->hw;
1786 dev->data->scattered_rx = false;
1790 hns3_rx_scattered_calc(struct rte_eth_dev *dev)
1792 struct rte_eth_conf *dev_conf = &dev->data->dev_conf;
1793 struct hns3_adapter *hns = dev->data->dev_private;
1794 struct hns3_hw *hw = &hns->hw;
1795 struct hns3_rx_queue *rxq;
1798 if (dev->data->rx_queues == NULL)
1801 for (queue_id = 0; queue_id < dev->data->nb_rx_queues; queue_id++) {
1802 rxq = dev->data->rx_queues[queue_id];
1803 if (hw->rx_buf_len == 0)
1804 hw->rx_buf_len = rxq->rx_buf_len;
1806 hw->rx_buf_len = RTE_MIN(hw->rx_buf_len,
1810 if (dev_conf->rxmode.offloads & DEV_RX_OFFLOAD_SCATTER ||
1811 dev_conf->rxmode.max_rx_pkt_len > hw->rx_buf_len)
1812 dev->data->scattered_rx = true;
1816 hns3_dev_supported_ptypes_get(struct rte_eth_dev *dev)
1818 static const uint32_t ptypes[] = {
1820 RTE_PTYPE_L2_ETHER_VLAN,
1821 RTE_PTYPE_L2_ETHER_QINQ,
1822 RTE_PTYPE_L2_ETHER_LLDP,
1823 RTE_PTYPE_L2_ETHER_ARP,
1825 RTE_PTYPE_L3_IPV4_EXT,
1827 RTE_PTYPE_L3_IPV6_EXT,
1833 RTE_PTYPE_TUNNEL_GRE,
1834 RTE_PTYPE_INNER_L2_ETHER,
1835 RTE_PTYPE_INNER_L2_ETHER_VLAN,
1836 RTE_PTYPE_INNER_L2_ETHER_QINQ,
1837 RTE_PTYPE_INNER_L3_IPV4,
1838 RTE_PTYPE_INNER_L3_IPV6,
1839 RTE_PTYPE_INNER_L3_IPV4_EXT,
1840 RTE_PTYPE_INNER_L3_IPV6_EXT,
1841 RTE_PTYPE_INNER_L4_UDP,
1842 RTE_PTYPE_INNER_L4_TCP,
1843 RTE_PTYPE_INNER_L4_SCTP,
1844 RTE_PTYPE_INNER_L4_ICMP,
1845 RTE_PTYPE_TUNNEL_VXLAN,
1846 RTE_PTYPE_TUNNEL_NVGRE,
1850 if (dev->rx_pkt_burst == hns3_recv_pkts ||
1851 dev->rx_pkt_burst == hns3_recv_scattered_pkts ||
1852 dev->rx_pkt_burst == hns3_recv_pkts_vec ||
1853 dev->rx_pkt_burst == hns3_recv_pkts_vec_sve)
1860 hns3_init_non_tunnel_ptype_tbl(struct hns3_ptype_table *tbl)
1862 tbl->l2l3table[0][0] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4;
1863 tbl->l2l3table[0][1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6;
1864 tbl->l2l3table[0][2] = RTE_PTYPE_L2_ETHER_ARP;
1865 tbl->l2l3table[0][3] = RTE_PTYPE_L2_ETHER;
1866 tbl->l2l3table[0][4] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT;
1867 tbl->l2l3table[0][5] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT;
1868 tbl->l2l3table[0][6] = RTE_PTYPE_L2_ETHER_LLDP;
1869 tbl->l2l3table[0][15] = RTE_PTYPE_L2_ETHER;
1871 tbl->l2l3table[1][0] = RTE_PTYPE_L2_ETHER_VLAN | RTE_PTYPE_L3_IPV4;
1872 tbl->l2l3table[1][1] = RTE_PTYPE_L2_ETHER_VLAN | RTE_PTYPE_L3_IPV6;
1873 tbl->l2l3table[1][2] = RTE_PTYPE_L2_ETHER_ARP;
1874 tbl->l2l3table[1][3] = RTE_PTYPE_L2_ETHER_VLAN;
1875 tbl->l2l3table[1][4] = RTE_PTYPE_L2_ETHER_VLAN | RTE_PTYPE_L3_IPV4_EXT;
1876 tbl->l2l3table[1][5] = RTE_PTYPE_L2_ETHER_VLAN | RTE_PTYPE_L3_IPV6_EXT;
1877 tbl->l2l3table[1][6] = RTE_PTYPE_L2_ETHER_LLDP;
1878 tbl->l2l3table[1][15] = RTE_PTYPE_L2_ETHER_VLAN;
1880 tbl->l2l3table[2][0] = RTE_PTYPE_L2_ETHER_QINQ | RTE_PTYPE_L3_IPV4;
1881 tbl->l2l3table[2][1] = RTE_PTYPE_L2_ETHER_QINQ | RTE_PTYPE_L3_IPV6;
1882 tbl->l2l3table[2][2] = RTE_PTYPE_L2_ETHER_ARP;
1883 tbl->l2l3table[2][3] = RTE_PTYPE_L2_ETHER_QINQ;
1884 tbl->l2l3table[2][4] = RTE_PTYPE_L2_ETHER_QINQ | RTE_PTYPE_L3_IPV4_EXT;
1885 tbl->l2l3table[2][5] = RTE_PTYPE_L2_ETHER_QINQ | RTE_PTYPE_L3_IPV6_EXT;
1886 tbl->l2l3table[2][6] = RTE_PTYPE_L2_ETHER_LLDP;
1887 tbl->l2l3table[2][15] = RTE_PTYPE_L2_ETHER_QINQ;
1889 tbl->l4table[0] = RTE_PTYPE_L4_UDP;
1890 tbl->l4table[1] = RTE_PTYPE_L4_TCP;
1891 tbl->l4table[2] = RTE_PTYPE_TUNNEL_GRE;
1892 tbl->l4table[3] = RTE_PTYPE_L4_SCTP;
1893 tbl->l4table[4] = RTE_PTYPE_L4_IGMP;
1894 tbl->l4table[5] = RTE_PTYPE_L4_ICMP;
1898 hns3_init_tunnel_ptype_tbl(struct hns3_ptype_table *tbl)
1900 tbl->inner_l2table[0] = RTE_PTYPE_INNER_L2_ETHER;
1901 tbl->inner_l2table[1] = RTE_PTYPE_INNER_L2_ETHER_VLAN;
1902 tbl->inner_l2table[2] = RTE_PTYPE_INNER_L2_ETHER_QINQ;
1904 tbl->inner_l3table[0] = RTE_PTYPE_INNER_L3_IPV4;
1905 tbl->inner_l3table[1] = RTE_PTYPE_INNER_L3_IPV6;
1906 /* There is not a ptype for inner ARP/RARP */
1907 tbl->inner_l3table[2] = RTE_PTYPE_UNKNOWN;
1908 tbl->inner_l3table[3] = RTE_PTYPE_UNKNOWN;
1909 tbl->inner_l3table[4] = RTE_PTYPE_INNER_L3_IPV4_EXT;
1910 tbl->inner_l3table[5] = RTE_PTYPE_INNER_L3_IPV6_EXT;
1912 tbl->inner_l4table[0] = RTE_PTYPE_INNER_L4_UDP;
1913 tbl->inner_l4table[1] = RTE_PTYPE_INNER_L4_TCP;
1914 /* There is not a ptype for inner GRE */
1915 tbl->inner_l4table[2] = RTE_PTYPE_UNKNOWN;
1916 tbl->inner_l4table[3] = RTE_PTYPE_INNER_L4_SCTP;
1917 /* There is not a ptype for inner IGMP */
1918 tbl->inner_l4table[4] = RTE_PTYPE_UNKNOWN;
1919 tbl->inner_l4table[5] = RTE_PTYPE_INNER_L4_ICMP;
1921 tbl->ol2table[0] = RTE_PTYPE_L2_ETHER;
1922 tbl->ol2table[1] = RTE_PTYPE_L2_ETHER_VLAN;
1923 tbl->ol2table[2] = RTE_PTYPE_L2_ETHER_QINQ;
1925 tbl->ol3table[0] = RTE_PTYPE_L3_IPV4;
1926 tbl->ol3table[1] = RTE_PTYPE_L3_IPV6;
1927 tbl->ol3table[2] = RTE_PTYPE_UNKNOWN;
1928 tbl->ol3table[3] = RTE_PTYPE_UNKNOWN;
1929 tbl->ol3table[4] = RTE_PTYPE_L3_IPV4_EXT;
1930 tbl->ol3table[5] = RTE_PTYPE_L3_IPV6_EXT;
1932 tbl->ol4table[0] = RTE_PTYPE_UNKNOWN;
1933 tbl->ol4table[1] = RTE_PTYPE_TUNNEL_VXLAN;
1934 tbl->ol4table[2] = RTE_PTYPE_TUNNEL_NVGRE;
1938 hns3_init_rx_ptype_tble(struct rte_eth_dev *dev)
1940 struct hns3_adapter *hns = dev->data->dev_private;
1941 struct hns3_ptype_table *tbl = &hns->ptype_tbl;
1943 memset(tbl, 0, sizeof(*tbl));
1945 hns3_init_non_tunnel_ptype_tbl(tbl);
1946 hns3_init_tunnel_ptype_tbl(tbl);
1950 hns3_rxd_to_vlan_tci(struct hns3_rx_queue *rxq, struct rte_mbuf *mb,
1951 uint32_t l234_info, const struct hns3_desc *rxd)
1953 #define HNS3_STRP_STATUS_NUM 0x4
1955 #define HNS3_NO_STRP_VLAN_VLD 0x0
1956 #define HNS3_INNER_STRP_VLAN_VLD 0x1
1957 #define HNS3_OUTER_STRP_VLAN_VLD 0x2
1958 uint32_t strip_status;
1959 uint32_t report_mode;
1962 * Since HW limitation, the vlan tag will always be inserted into RX
1963 * descriptor when strip the tag from packet, driver needs to determine
1964 * reporting which tag to mbuf according to the PVID configuration
1965 * and vlan striped status.
1967 static const uint32_t report_type[][HNS3_STRP_STATUS_NUM] = {
1969 HNS3_NO_STRP_VLAN_VLD,
1970 HNS3_OUTER_STRP_VLAN_VLD,
1971 HNS3_INNER_STRP_VLAN_VLD,
1972 HNS3_OUTER_STRP_VLAN_VLD
1975 HNS3_NO_STRP_VLAN_VLD,
1976 HNS3_NO_STRP_VLAN_VLD,
1977 HNS3_NO_STRP_VLAN_VLD,
1978 HNS3_INNER_STRP_VLAN_VLD
1981 strip_status = hns3_get_field(l234_info, HNS3_RXD_STRP_TAGP_M,
1982 HNS3_RXD_STRP_TAGP_S);
1983 report_mode = report_type[rxq->pvid_sw_discard_en][strip_status];
1984 switch (report_mode) {
1985 case HNS3_NO_STRP_VLAN_VLD:
1988 case HNS3_INNER_STRP_VLAN_VLD:
1989 mb->ol_flags |= PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
1990 mb->vlan_tci = rte_le_to_cpu_16(rxd->rx.vlan_tag);
1992 case HNS3_OUTER_STRP_VLAN_VLD:
1993 mb->ol_flags |= PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
1994 mb->vlan_tci = rte_le_to_cpu_16(rxd->rx.ot_vlan_tag);
2003 recalculate_data_len(struct rte_mbuf *first_seg, struct rte_mbuf *last_seg,
2004 struct rte_mbuf *rxm, struct hns3_rx_queue *rxq,
2007 uint8_t crc_len = rxq->crc_len;
2009 if (data_len <= crc_len) {
2010 rte_pktmbuf_free_seg(rxm);
2011 first_seg->nb_segs--;
2012 last_seg->data_len = (uint16_t)(last_seg->data_len -
2013 (crc_len - data_len));
2014 last_seg->next = NULL;
2016 rxm->data_len = (uint16_t)(data_len - crc_len);
2019 static inline struct rte_mbuf *
2020 hns3_rx_alloc_buffer(struct hns3_rx_queue *rxq)
2024 if (likely(rxq->bulk_mbuf_num > 0))
2025 return rxq->bulk_mbuf[--rxq->bulk_mbuf_num];
2027 ret = rte_mempool_get_bulk(rxq->mb_pool, (void **)rxq->bulk_mbuf,
2028 HNS3_BULK_ALLOC_MBUF_NUM);
2029 if (likely(ret == 0)) {
2030 rxq->bulk_mbuf_num = HNS3_BULK_ALLOC_MBUF_NUM;
2031 return rxq->bulk_mbuf[--rxq->bulk_mbuf_num];
2033 return rte_mbuf_raw_alloc(rxq->mb_pool);
2037 hns3_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
2039 volatile struct hns3_desc *rx_ring; /* RX ring (desc) */
2040 volatile struct hns3_desc *rxdp; /* pointer of the current desc */
2041 struct hns3_rx_queue *rxq; /* RX queue */
2042 struct hns3_entry *sw_ring;
2043 struct hns3_entry *rxe;
2044 struct hns3_desc rxd;
2045 struct rte_mbuf *nmb; /* pointer of the new mbuf */
2046 struct rte_mbuf *rxm;
2047 uint32_t bd_base_info;
2060 rx_ring = rxq->rx_ring;
2061 sw_ring = rxq->sw_ring;
2062 rx_id = rxq->next_to_use;
2064 while (nb_rx < nb_pkts) {
2065 rxdp = &rx_ring[rx_id];
2066 bd_base_info = rte_le_to_cpu_32(rxdp->rx.bd_base_info);
2067 if (unlikely(!(bd_base_info & BIT(HNS3_RXD_VLD_B))))
2070 rxd = rxdp[(bd_base_info & (1u << HNS3_RXD_VLD_B)) -
2071 (1u << HNS3_RXD_VLD_B)];
2073 nmb = hns3_rx_alloc_buffer(rxq);
2074 if (unlikely(nmb == NULL)) {
2077 port_id = rxq->port_id;
2078 rte_eth_devices[port_id].data->rx_mbuf_alloc_failed++;
2083 rxe = &sw_ring[rx_id];
2085 if (unlikely(rx_id == rxq->nb_rx_desc))
2088 rte_prefetch0(sw_ring[rx_id].mbuf);
2089 if ((rx_id & HNS3_RX_RING_PREFETCTH_MASK) == 0) {
2090 rte_prefetch0(&rx_ring[rx_id]);
2091 rte_prefetch0(&sw_ring[rx_id]);
2097 dma_addr = rte_mbuf_data_iova_default(nmb);
2098 rxdp->addr = rte_cpu_to_le_64(dma_addr);
2099 rxdp->rx.bd_base_info = 0;
2101 rxm->data_off = RTE_PKTMBUF_HEADROOM;
2102 rxm->pkt_len = (uint16_t)(rte_le_to_cpu_16(rxd.rx.pkt_len)) -
2104 rxm->data_len = rxm->pkt_len;
2105 rxm->port = rxq->port_id;
2106 rxm->hash.rss = rte_le_to_cpu_32(rxd.rx.rss_hash);
2107 rxm->ol_flags = PKT_RX_RSS_HASH;
2108 if (unlikely(bd_base_info & BIT(HNS3_RXD_LUM_B))) {
2110 rte_le_to_cpu_16(rxd.rx.fd_id);
2111 rxm->ol_flags |= PKT_RX_FDIR | PKT_RX_FDIR_ID;
2116 /* Load remained descriptor data and extract necessary fields */
2117 l234_info = rte_le_to_cpu_32(rxd.rx.l234_info);
2118 ol_info = rte_le_to_cpu_32(rxd.rx.ol_info);
2119 ret = hns3_handle_bdinfo(rxq, rxm, bd_base_info,
2120 l234_info, &cksum_err);
2124 rxm->packet_type = hns3_rx_calc_ptype(rxq, l234_info, ol_info);
2126 if (likely(bd_base_info & BIT(HNS3_RXD_L3L4P_B)))
2127 hns3_rx_set_cksum_flag(rxm, rxm->packet_type,
2129 hns3_rxd_to_vlan_tci(rxq, rxm, l234_info, &rxd);
2131 rx_pkts[nb_rx++] = rxm;
2134 rte_pktmbuf_free(rxm);
2137 rxq->next_to_use = rx_id;
2138 rxq->rx_free_hold += nb_rx_bd;
2139 if (rxq->rx_free_hold > rxq->rx_free_thresh) {
2140 hns3_write_reg_opt(rxq->io_head_reg, rxq->rx_free_hold);
2141 rxq->rx_free_hold = 0;
2148 hns3_recv_scattered_pkts(void *rx_queue,
2149 struct rte_mbuf **rx_pkts,
2152 volatile struct hns3_desc *rx_ring; /* RX ring (desc) */
2153 volatile struct hns3_desc *rxdp; /* pointer of the current desc */
2154 struct hns3_rx_queue *rxq; /* RX queue */
2155 struct hns3_entry *sw_ring;
2156 struct hns3_entry *rxe;
2157 struct rte_mbuf *first_seg;
2158 struct rte_mbuf *last_seg;
2159 struct hns3_desc rxd;
2160 struct rte_mbuf *nmb; /* pointer of the new mbuf */
2161 struct rte_mbuf *rxm;
2162 struct rte_eth_dev *dev;
2163 uint32_t bd_base_info;
2178 rx_id = rxq->next_to_use;
2179 rx_ring = rxq->rx_ring;
2180 sw_ring = rxq->sw_ring;
2181 first_seg = rxq->pkt_first_seg;
2182 last_seg = rxq->pkt_last_seg;
2184 while (nb_rx < nb_pkts) {
2185 rxdp = &rx_ring[rx_id];
2186 bd_base_info = rte_le_to_cpu_32(rxdp->rx.bd_base_info);
2187 if (unlikely(!(bd_base_info & BIT(HNS3_RXD_VLD_B))))
2191 * The interactive process between software and hardware of
2192 * receiving a new packet in hns3 network engine:
2193 * 1. Hardware network engine firstly writes the packet content
2194 * to the memory pointed by the 'addr' field of the Rx Buffer
2195 * Descriptor, secondly fills the result of parsing the
2196 * packet include the valid field into the Rx Buffer
2197 * Descriptor in one write operation.
2198 * 2. Driver reads the Rx BD's valid field in the loop to check
2199 * whether it's valid, if valid then assign a new address to
2200 * the addr field, clear the valid field, get the other
2201 * information of the packet by parsing Rx BD's other fields,
2202 * finally write back the number of Rx BDs processed by the
2203 * driver to the HNS3_RING_RX_HEAD_REG register to inform
2205 * In the above process, the ordering is very important. We must
2206 * make sure that CPU read Rx BD's other fields only after the
2209 * There are two type of re-ordering: compiler re-ordering and
2210 * CPU re-ordering under the ARMv8 architecture.
2211 * 1. we use volatile to deal with compiler re-ordering, so you
2212 * can see that rx_ring/rxdp defined with volatile.
2213 * 2. we commonly use memory barrier to deal with CPU
2214 * re-ordering, but the cost is high.
2216 * In order to solve the high cost of using memory barrier, we
2217 * use the data dependency order under the ARMv8 architecture,
2220 * instr02: load B <- A
2221 * the instr02 will always execute after instr01.
2223 * To construct the data dependency ordering, we use the
2224 * following assignment:
2225 * rxd = rxdp[(bd_base_info & (1u << HNS3_RXD_VLD_B)) -
2226 * (1u<<HNS3_RXD_VLD_B)]
2227 * Using gcc compiler under the ARMv8 architecture, the related
2228 * assembly code example as follows:
2229 * note: (1u << HNS3_RXD_VLD_B) equal 0x10
2230 * instr01: ldr w26, [x22, #28] --read bd_base_info
2231 * instr02: and w0, w26, #0x10 --calc bd_base_info & 0x10
2232 * instr03: sub w0, w0, #0x10 --calc (bd_base_info &
2234 * instr04: add x0, x22, x0, lsl #5 --calc copy source addr
2235 * instr05: ldp x2, x3, [x0]
2236 * instr06: stp x2, x3, [x29, #256] --copy BD's [0 ~ 15]B
2237 * instr07: ldp x4, x5, [x0, #16]
2238 * instr08: stp x4, x5, [x29, #272] --copy BD's [16 ~ 31]B
2239 * the instr05~08 depend on x0's value, x0 depent on w26's
2240 * value, the w26 is the bd_base_info, this form the data
2241 * dependency ordering.
2242 * note: if BD is valid, (bd_base_info & (1u<<HNS3_RXD_VLD_B)) -
2243 * (1u<<HNS3_RXD_VLD_B) will always zero, so the
2244 * assignment is correct.
2246 * So we use the data dependency ordering instead of memory
2247 * barrier to improve receive performance.
2249 rxd = rxdp[(bd_base_info & (1u << HNS3_RXD_VLD_B)) -
2250 (1u << HNS3_RXD_VLD_B)];
2252 nmb = hns3_rx_alloc_buffer(rxq);
2253 if (unlikely(nmb == NULL)) {
2254 dev = &rte_eth_devices[rxq->port_id];
2255 dev->data->rx_mbuf_alloc_failed++;
2260 rxe = &sw_ring[rx_id];
2262 if (unlikely(rx_id == rxq->nb_rx_desc))
2265 rte_prefetch0(sw_ring[rx_id].mbuf);
2266 if ((rx_id & HNS3_RX_RING_PREFETCTH_MASK) == 0) {
2267 rte_prefetch0(&rx_ring[rx_id]);
2268 rte_prefetch0(&sw_ring[rx_id]);
2274 dma_addr = rte_cpu_to_le_64(rte_mbuf_data_iova_default(nmb));
2275 rxdp->rx.bd_base_info = 0;
2276 rxdp->addr = dma_addr;
2278 if (first_seg == NULL) {
2280 first_seg->nb_segs = 1;
2282 first_seg->nb_segs++;
2283 last_seg->next = rxm;
2286 rxm->data_off = RTE_PKTMBUF_HEADROOM;
2287 rxm->data_len = rte_le_to_cpu_16(rxd.rx.size);
2289 if (!(bd_base_info & BIT(HNS3_RXD_FE_B))) {
2296 * The last buffer of the received packet. packet len from
2297 * buffer description may contains CRC len, packet len should
2298 * subtract it, same as data len.
2300 first_seg->pkt_len = rte_le_to_cpu_16(rxd.rx.pkt_len);
2303 * This is the last buffer of the received packet. If the CRC
2304 * is not stripped by the hardware:
2305 * - Subtract the CRC length from the total packet length.
2306 * - If the last buffer only contains the whole CRC or a part
2307 * of it, free the mbuf associated to the last buffer. If part
2308 * of the CRC is also contained in the previous mbuf, subtract
2309 * the length of that CRC part from the data length of the
2313 if (unlikely(rxq->crc_len > 0)) {
2314 first_seg->pkt_len -= rxq->crc_len;
2315 recalculate_data_len(first_seg, last_seg, rxm, rxq,
2319 first_seg->port = rxq->port_id;
2320 first_seg->hash.rss = rte_le_to_cpu_32(rxd.rx.rss_hash);
2321 first_seg->ol_flags = PKT_RX_RSS_HASH;
2322 if (unlikely(bd_base_info & BIT(HNS3_RXD_LUM_B))) {
2323 first_seg->hash.fdir.hi =
2324 rte_le_to_cpu_16(rxd.rx.fd_id);
2325 first_seg->ol_flags |= PKT_RX_FDIR | PKT_RX_FDIR_ID;
2328 gro_size = hns3_get_field(bd_base_info, HNS3_RXD_GRO_SIZE_M,
2329 HNS3_RXD_GRO_SIZE_S);
2330 if (gro_size != 0) {
2331 first_seg->ol_flags |= PKT_RX_LRO;
2332 first_seg->tso_segsz = gro_size;
2335 l234_info = rte_le_to_cpu_32(rxd.rx.l234_info);
2336 ol_info = rte_le_to_cpu_32(rxd.rx.ol_info);
2337 ret = hns3_handle_bdinfo(rxq, first_seg, bd_base_info,
2338 l234_info, &cksum_err);
2342 first_seg->packet_type = hns3_rx_calc_ptype(rxq,
2343 l234_info, ol_info);
2345 if (bd_base_info & BIT(HNS3_RXD_L3L4P_B))
2346 hns3_rx_set_cksum_flag(first_seg,
2347 first_seg->packet_type,
2349 hns3_rxd_to_vlan_tci(rxq, first_seg, l234_info, &rxd);
2351 rx_pkts[nb_rx++] = first_seg;
2355 rte_pktmbuf_free(first_seg);
2359 rxq->next_to_use = rx_id;
2360 rxq->pkt_first_seg = first_seg;
2361 rxq->pkt_last_seg = last_seg;
2363 rxq->rx_free_hold += nb_rx_bd;
2364 if (rxq->rx_free_hold > rxq->rx_free_thresh) {
2365 hns3_write_reg_opt(rxq->io_head_reg, rxq->rx_free_hold);
2366 rxq->rx_free_hold = 0;
2373 hns3_rxq_vec_setup(__rte_unused struct hns3_rx_queue *rxq)
2378 hns3_rx_check_vec_support(__rte_unused struct rte_eth_dev *dev)
2384 hns3_recv_pkts_vec(__rte_unused void *tx_queue,
2385 __rte_unused struct rte_mbuf **rx_pkts,
2386 __rte_unused uint16_t nb_pkts)
2392 hns3_recv_pkts_vec_sve(__rte_unused void *tx_queue,
2393 __rte_unused struct rte_mbuf **rx_pkts,
2394 __rte_unused uint16_t nb_pkts)
2400 hns3_rx_burst_mode_get(struct rte_eth_dev *dev, __rte_unused uint16_t queue_id,
2401 struct rte_eth_burst_mode *mode)
2403 static const struct {
2404 eth_rx_burst_t pkt_burst;
2407 { hns3_recv_pkts, "Scalar" },
2408 { hns3_recv_scattered_pkts, "Scalar Scattered" },
2409 { hns3_recv_pkts_vec, "Vector Neon" },
2410 { hns3_recv_pkts_vec_sve, "Vector Sve" },
2413 eth_rx_burst_t pkt_burst = dev->rx_pkt_burst;
2417 for (i = 0; i < RTE_DIM(burst_infos); i++) {
2418 if (pkt_burst == burst_infos[i].pkt_burst) {
2419 snprintf(mode->info, sizeof(mode->info), "%s",
2420 burst_infos[i].info);
2430 hns3_check_sve_support(void)
2432 #if defined(RTE_ARCH_ARM64) && defined(CC_SVE_SUPPORT)
2433 if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_SVE))
2439 static eth_rx_burst_t
2440 hns3_get_rx_function(struct rte_eth_dev *dev)
2442 struct hns3_adapter *hns = dev->data->dev_private;
2443 uint64_t offloads = dev->data->dev_conf.rxmode.offloads;
2445 if (hns->rx_vec_allowed && hns3_rx_check_vec_support(dev) == 0)
2446 return hns3_check_sve_support() ? hns3_recv_pkts_vec_sve :
2449 if (hns->rx_simple_allowed && !dev->data->scattered_rx &&
2450 (offloads & DEV_RX_OFFLOAD_TCP_LRO) == 0)
2451 return hns3_recv_pkts;
2453 return hns3_recv_scattered_pkts;
2457 hns3_tx_queue_conf_check(struct hns3_hw *hw, const struct rte_eth_txconf *conf,
2458 uint16_t nb_desc, uint16_t *tx_rs_thresh,
2459 uint16_t *tx_free_thresh, uint16_t idx)
2461 #define HNS3_TX_RS_FREE_THRESH_GAP 8
2462 uint16_t rs_thresh, free_thresh, fast_free_thresh;
2464 if (nb_desc > HNS3_MAX_RING_DESC || nb_desc < HNS3_MIN_RING_DESC ||
2465 nb_desc % HNS3_ALIGN_RING_DESC) {
2466 hns3_err(hw, "number (%u) of tx descriptors is invalid",
2471 rs_thresh = (conf->tx_rs_thresh > 0) ?
2472 conf->tx_rs_thresh : HNS3_DEFAULT_TX_RS_THRESH;
2473 free_thresh = (conf->tx_free_thresh > 0) ?
2474 conf->tx_free_thresh : HNS3_DEFAULT_TX_FREE_THRESH;
2475 if (rs_thresh + free_thresh > nb_desc || nb_desc % rs_thresh ||
2476 rs_thresh >= nb_desc - HNS3_TX_RS_FREE_THRESH_GAP ||
2477 free_thresh >= nb_desc - HNS3_TX_RS_FREE_THRESH_GAP) {
2478 hns3_err(hw, "tx_rs_thresh (%u) tx_free_thresh (%u) nb_desc "
2479 "(%u) of tx descriptors for port=%u queue=%u check "
2481 rs_thresh, free_thresh, nb_desc, hw->data->port_id,
2486 if (conf->tx_free_thresh == 0) {
2487 /* Fast free Tx memory buffer to improve cache hit rate */
2488 fast_free_thresh = nb_desc - rs_thresh;
2489 if (fast_free_thresh >=
2490 HNS3_TX_FAST_FREE_AHEAD + HNS3_DEFAULT_TX_FREE_THRESH)
2491 free_thresh = fast_free_thresh -
2492 HNS3_TX_FAST_FREE_AHEAD;
2495 *tx_rs_thresh = rs_thresh;
2496 *tx_free_thresh = free_thresh;
2501 hns3_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t nb_desc,
2502 unsigned int socket_id, const struct rte_eth_txconf *conf)
2504 struct hns3_adapter *hns = dev->data->dev_private;
2505 uint16_t tx_rs_thresh, tx_free_thresh;
2506 struct hns3_hw *hw = &hns->hw;
2507 struct hns3_queue_info q_info;
2508 struct hns3_tx_queue *txq;
2512 ret = hns3_tx_queue_conf_check(hw, conf, nb_desc,
2513 &tx_rs_thresh, &tx_free_thresh, idx);
2517 if (dev->data->tx_queues[idx] != NULL) {
2518 hns3_tx_queue_release(dev->data->tx_queues[idx]);
2519 dev->data->tx_queues[idx] = NULL;
2523 q_info.socket_id = socket_id;
2524 q_info.nb_desc = nb_desc;
2525 q_info.type = "hns3 TX queue";
2526 q_info.ring_name = "tx_ring";
2527 txq = hns3_alloc_txq_and_dma_zone(dev, &q_info);
2530 "Failed to alloc mem and reserve DMA mem for tx ring!");
2534 txq->tx_deferred_start = conf->tx_deferred_start;
2535 if (txq->tx_deferred_start && !hns3_dev_indep_txrx_supported(hw)) {
2536 hns3_warn(hw, "deferred start is not supported.");
2537 txq->tx_deferred_start = false;
2540 tx_entry_len = sizeof(struct hns3_entry) * txq->nb_tx_desc;
2541 txq->sw_ring = rte_zmalloc_socket("hns3 TX sw ring", tx_entry_len,
2542 RTE_CACHE_LINE_SIZE, socket_id);
2543 if (txq->sw_ring == NULL) {
2544 hns3_err(hw, "Failed to allocate memory for tx sw ring!");
2545 hns3_tx_queue_release(txq);
2550 txq->next_to_use = 0;
2551 txq->next_to_clean = 0;
2552 txq->tx_bd_ready = txq->nb_tx_desc - 1;
2553 txq->tx_free_thresh = tx_free_thresh;
2554 txq->tx_rs_thresh = tx_rs_thresh;
2555 txq->free = rte_zmalloc_socket("hns3 TX mbuf free array",
2556 sizeof(struct rte_mbuf *) * txq->tx_rs_thresh,
2557 RTE_CACHE_LINE_SIZE, socket_id);
2559 hns3_err(hw, "failed to allocate tx mbuf free array!");
2560 hns3_tx_queue_release(txq);
2564 txq->port_id = dev->data->port_id;
2566 * For hns3 PF device, if the VLAN mode is HW_SHIFT_AND_DISCARD_MODE,
2567 * the pvid_sw_shift_en in the queue struct should not be changed,
2568 * because PVID-related operations do not need to be processed by PMD
2569 * driver. For hns3 VF device, whether it needs to process PVID depends
2570 * on the configuration of PF kernel mode netdev driver. And the
2571 * related PF configuration is delivered through the mailbox and finally
2572 * reflectd in port_base_vlan_cfg.
2574 if (hns->is_vf || hw->vlan_mode == HNS3_SW_SHIFT_AND_DISCARD_MODE)
2575 txq->pvid_sw_shift_en = hw->port_base_vlan_cfg.state ==
2576 HNS3_PORT_BASE_VLAN_ENABLE;
2578 txq->pvid_sw_shift_en = false;
2579 txq->max_non_tso_bd_num = hw->max_non_tso_bd_num;
2580 txq->configured = true;
2581 txq->io_base = (void *)((char *)hw->io_base +
2582 hns3_get_tqp_reg_offset(idx));
2583 txq->io_tail_reg = (volatile void *)((char *)txq->io_base +
2584 HNS3_RING_TX_TAIL_REG);
2585 txq->min_tx_pkt_len = hw->min_tx_pkt_len;
2586 txq->tso_mode = hw->tso_mode;
2587 txq->over_length_pkt_cnt = 0;
2588 txq->exceed_limit_bd_pkt_cnt = 0;
2589 txq->exceed_limit_bd_reassem_fail = 0;
2590 txq->unsupported_tunnel_pkt_cnt = 0;
2591 txq->queue_full_cnt = 0;
2592 txq->pkt_padding_fail_cnt = 0;
2593 rte_spinlock_lock(&hw->lock);
2594 dev->data->tx_queues[idx] = txq;
2595 rte_spinlock_unlock(&hw->lock);
2601 hns3_tx_free_useless_buffer(struct hns3_tx_queue *txq)
2603 uint16_t tx_next_clean = txq->next_to_clean;
2604 uint16_t tx_next_use = txq->next_to_use;
2605 uint16_t tx_bd_ready = txq->tx_bd_ready;
2606 uint16_t tx_bd_max = txq->nb_tx_desc;
2607 struct hns3_entry *tx_bak_pkt = &txq->sw_ring[tx_next_clean];
2608 struct hns3_desc *desc = &txq->tx_ring[tx_next_clean];
2609 struct rte_mbuf *mbuf;
2611 while ((!(desc->tx.tp_fe_sc_vld_ra_ri &
2612 rte_cpu_to_le_16(BIT(HNS3_TXD_VLD_B)))) &&
2613 tx_next_use != tx_next_clean) {
2614 mbuf = tx_bak_pkt->mbuf;
2616 rte_pktmbuf_free_seg(mbuf);
2617 tx_bak_pkt->mbuf = NULL;
2625 if (tx_next_clean >= tx_bd_max) {
2627 desc = txq->tx_ring;
2628 tx_bak_pkt = txq->sw_ring;
2632 txq->next_to_clean = tx_next_clean;
2633 txq->tx_bd_ready = tx_bd_ready;
2637 hns3_config_gro(struct hns3_hw *hw, bool en)
2639 struct hns3_cfg_gro_status_cmd *req;
2640 struct hns3_cmd_desc desc;
2643 hns3_cmd_setup_basic_desc(&desc, HNS3_OPC_GRO_GENERIC_CONFIG, false);
2644 req = (struct hns3_cfg_gro_status_cmd *)desc.data;
2646 req->gro_en = rte_cpu_to_le_16(en ? 1 : 0);
2648 ret = hns3_cmd_send(hw, &desc, 1);
2650 hns3_err(hw, "%s hardware GRO failed, ret = %d",
2651 en ? "enable" : "disable", ret);
2657 hns3_restore_gro_conf(struct hns3_hw *hw)
2663 offloads = hw->data->dev_conf.rxmode.offloads;
2664 gro_en = offloads & DEV_RX_OFFLOAD_TCP_LRO ? true : false;
2665 ret = hns3_config_gro(hw, gro_en);
2667 hns3_err(hw, "restore hardware GRO to %s failed, ret = %d",
2668 gro_en ? "enabled" : "disabled", ret);
2674 hns3_pkt_is_tso(struct rte_mbuf *m)
2676 return (m->tso_segsz != 0 && m->ol_flags & PKT_TX_TCP_SEG);
2680 hns3_set_tso(struct hns3_desc *desc, uint32_t paylen, struct rte_mbuf *rxm)
2682 if (!hns3_pkt_is_tso(rxm))
2685 if (paylen <= rxm->tso_segsz)
2688 desc->tx.type_cs_vlan_tso_len |= rte_cpu_to_le_32(BIT(HNS3_TXD_TSO_B));
2689 desc->tx.mss = rte_cpu_to_le_16(rxm->tso_segsz);
2693 hns3_fill_per_desc(struct hns3_desc *desc, struct rte_mbuf *rxm)
2695 desc->addr = rte_mbuf_data_iova(rxm);
2696 desc->tx.send_size = rte_cpu_to_le_16(rte_pktmbuf_data_len(rxm));
2697 desc->tx.tp_fe_sc_vld_ra_ri = rte_cpu_to_le_16(BIT(HNS3_TXD_VLD_B));
2701 hns3_fill_first_desc(struct hns3_tx_queue *txq, struct hns3_desc *desc,
2702 struct rte_mbuf *rxm)
2704 uint64_t ol_flags = rxm->ol_flags;
2708 hdr_len = rxm->l2_len + rxm->l3_len + rxm->l4_len;
2709 hdr_len += (ol_flags & PKT_TX_TUNNEL_MASK) ?
2710 rxm->outer_l2_len + rxm->outer_l3_len : 0;
2711 paylen = rxm->pkt_len - hdr_len;
2712 desc->tx.paylen = rte_cpu_to_le_32(paylen);
2713 hns3_set_tso(desc, paylen, rxm);
2716 * Currently, hardware doesn't support more than two layers VLAN offload
2717 * in Tx direction based on hns3 network engine. So when the number of
2718 * VLANs in the packets represented by rxm plus the number of VLAN
2719 * offload by hardware such as PVID etc, exceeds two, the packets will
2720 * be discarded or the original VLAN of the packets will be overwitted
2721 * by hardware. When the PF PVID is enabled by calling the API function
2722 * named rte_eth_dev_set_vlan_pvid or the VF PVID is enabled by the hns3
2723 * PF kernel ether driver, the outer VLAN tag will always be the PVID.
2724 * To avoid the VLAN of Tx descriptor is overwritten by PVID, it should
2725 * be added to the position close to the IP header when PVID is enabled.
2727 if (!txq->pvid_sw_shift_en && ol_flags & (PKT_TX_VLAN_PKT |
2729 desc->tx.ol_type_vlan_len_msec |=
2730 rte_cpu_to_le_32(BIT(HNS3_TXD_OVLAN_B));
2731 if (ol_flags & PKT_TX_QINQ_PKT)
2732 desc->tx.outer_vlan_tag =
2733 rte_cpu_to_le_16(rxm->vlan_tci_outer);
2735 desc->tx.outer_vlan_tag =
2736 rte_cpu_to_le_16(rxm->vlan_tci);
2739 if (ol_flags & PKT_TX_QINQ_PKT ||
2740 ((ol_flags & PKT_TX_VLAN_PKT) && txq->pvid_sw_shift_en)) {
2741 desc->tx.type_cs_vlan_tso_len |=
2742 rte_cpu_to_le_32(BIT(HNS3_TXD_VLAN_B));
2743 desc->tx.vlan_tag = rte_cpu_to_le_16(rxm->vlan_tci);
2748 hns3_tx_alloc_mbufs(struct rte_mempool *mb_pool, uint16_t nb_new_buf,
2749 struct rte_mbuf **alloc_mbuf)
2751 #define MAX_NON_TSO_BD_PER_PKT 18
2752 struct rte_mbuf *pkt_segs[MAX_NON_TSO_BD_PER_PKT];
2755 /* Allocate enough mbufs */
2756 if (rte_mempool_get_bulk(mb_pool, (void **)pkt_segs, nb_new_buf))
2759 for (i = 0; i < nb_new_buf - 1; i++)
2760 pkt_segs[i]->next = pkt_segs[i + 1];
2762 pkt_segs[nb_new_buf - 1]->next = NULL;
2763 pkt_segs[0]->nb_segs = nb_new_buf;
2764 *alloc_mbuf = pkt_segs[0];
2770 hns3_pktmbuf_copy_hdr(struct rte_mbuf *new_pkt, struct rte_mbuf *old_pkt)
2772 new_pkt->ol_flags = old_pkt->ol_flags;
2773 new_pkt->pkt_len = rte_pktmbuf_pkt_len(old_pkt);
2774 new_pkt->outer_l2_len = old_pkt->outer_l2_len;
2775 new_pkt->outer_l3_len = old_pkt->outer_l3_len;
2776 new_pkt->l2_len = old_pkt->l2_len;
2777 new_pkt->l3_len = old_pkt->l3_len;
2778 new_pkt->l4_len = old_pkt->l4_len;
2779 new_pkt->vlan_tci_outer = old_pkt->vlan_tci_outer;
2780 new_pkt->vlan_tci = old_pkt->vlan_tci;
2784 hns3_reassemble_tx_pkts(struct rte_mbuf *tx_pkt, struct rte_mbuf **new_pkt,
2785 uint8_t max_non_tso_bd_num)
2787 struct rte_mempool *mb_pool;
2788 struct rte_mbuf *new_mbuf;
2789 struct rte_mbuf *temp_new;
2790 struct rte_mbuf *temp;
2791 uint16_t last_buf_len;
2792 uint16_t nb_new_buf;
2802 mb_pool = tx_pkt->pool;
2803 buf_size = tx_pkt->buf_len - RTE_PKTMBUF_HEADROOM;
2804 nb_new_buf = (rte_pktmbuf_pkt_len(tx_pkt) - 1) / buf_size + 1;
2805 if (nb_new_buf > max_non_tso_bd_num)
2808 last_buf_len = rte_pktmbuf_pkt_len(tx_pkt) % buf_size;
2809 if (last_buf_len == 0)
2810 last_buf_len = buf_size;
2812 /* Allocate enough mbufs */
2813 ret = hns3_tx_alloc_mbufs(mb_pool, nb_new_buf, &new_mbuf);
2817 /* Copy the original packet content to the new mbufs */
2819 s = rte_pktmbuf_mtod(temp, char *);
2820 len_s = rte_pktmbuf_data_len(temp);
2821 temp_new = new_mbuf;
2822 while (temp != NULL && temp_new != NULL) {
2823 d = rte_pktmbuf_mtod(temp_new, char *);
2824 buf_len = temp_new->next == NULL ? last_buf_len : buf_size;
2828 len = RTE_MIN(len_s, len_d);
2832 len_d = len_d - len;
2833 len_s = len_s - len;
2839 s = rte_pktmbuf_mtod(temp, char *);
2840 len_s = rte_pktmbuf_data_len(temp);
2844 temp_new->data_len = buf_len;
2845 temp_new = temp_new->next;
2847 hns3_pktmbuf_copy_hdr(new_mbuf, tx_pkt);
2849 /* free original mbufs */
2850 rte_pktmbuf_free(tx_pkt);
2852 *new_pkt = new_mbuf;
2858 hns3_parse_outer_params(struct rte_mbuf *m, uint32_t *ol_type_vlan_len_msec)
2860 uint32_t tmp = *ol_type_vlan_len_msec;
2861 uint64_t ol_flags = m->ol_flags;
2863 /* (outer) IP header type */
2864 if (ol_flags & PKT_TX_OUTER_IPV4) {
2865 if (ol_flags & PKT_TX_OUTER_IP_CKSUM)
2866 tmp |= hns3_gen_field_val(HNS3_TXD_OL3T_M,
2867 HNS3_TXD_OL3T_S, HNS3_OL3T_IPV4_CSUM);
2869 tmp |= hns3_gen_field_val(HNS3_TXD_OL3T_M,
2870 HNS3_TXD_OL3T_S, HNS3_OL3T_IPV4_NO_CSUM);
2871 } else if (ol_flags & PKT_TX_OUTER_IPV6) {
2872 tmp |= hns3_gen_field_val(HNS3_TXD_OL3T_M, HNS3_TXD_OL3T_S,
2875 /* OL3 header size, defined in 4 bytes */
2876 tmp |= hns3_gen_field_val(HNS3_TXD_L3LEN_M, HNS3_TXD_L3LEN_S,
2877 m->outer_l3_len >> HNS3_L3_LEN_UNIT);
2878 *ol_type_vlan_len_msec = tmp;
2882 hns3_parse_inner_params(struct rte_mbuf *m, uint32_t *ol_type_vlan_len_msec,
2883 uint32_t *type_cs_vlan_tso_len)
2885 #define HNS3_NVGRE_HLEN 8
2886 uint32_t tmp_outer = *ol_type_vlan_len_msec;
2887 uint32_t tmp_inner = *type_cs_vlan_tso_len;
2888 uint64_t ol_flags = m->ol_flags;
2889 uint16_t inner_l2_len;
2891 switch (ol_flags & PKT_TX_TUNNEL_MASK) {
2892 case PKT_TX_TUNNEL_VXLAN_GPE:
2893 case PKT_TX_TUNNEL_GENEVE:
2894 case PKT_TX_TUNNEL_VXLAN:
2895 /* MAC in UDP tunnelling packet, include VxLAN and GENEVE */
2896 tmp_outer |= hns3_gen_field_val(HNS3_TXD_TUNTYPE_M,
2897 HNS3_TXD_TUNTYPE_S, HNS3_TUN_MAC_IN_UDP);
2899 * The inner l2 length of mbuf is the sum of outer l4 length,
2900 * tunneling header length and inner l2 length for a tunnel
2901 * packect. But in hns3 tx descriptor, the tunneling header
2902 * length is contained in the field of outer L4 length.
2903 * Therefore, driver need to calculate the outer L4 length and
2906 tmp_outer |= hns3_gen_field_val(HNS3_TXD_L4LEN_M,
2908 (uint8_t)RTE_ETHER_VXLAN_HLEN >>
2911 inner_l2_len = m->l2_len - RTE_ETHER_VXLAN_HLEN;
2913 case PKT_TX_TUNNEL_GRE:
2914 tmp_outer |= hns3_gen_field_val(HNS3_TXD_TUNTYPE_M,
2915 HNS3_TXD_TUNTYPE_S, HNS3_TUN_NVGRE);
2917 * For NVGRE tunnel packect, the outer L4 is empty. So only
2918 * fill the NVGRE header length to the outer L4 field.
2920 tmp_outer |= hns3_gen_field_val(HNS3_TXD_L4LEN_M,
2922 (uint8_t)HNS3_NVGRE_HLEN >> HNS3_L4_LEN_UNIT);
2924 inner_l2_len = m->l2_len - HNS3_NVGRE_HLEN;
2927 /* For non UDP / GRE tunneling, drop the tunnel packet */
2931 tmp_inner |= hns3_gen_field_val(HNS3_TXD_L2LEN_M, HNS3_TXD_L2LEN_S,
2932 inner_l2_len >> HNS3_L2_LEN_UNIT);
2933 /* OL2 header size, defined in 2 bytes */
2934 tmp_outer |= hns3_gen_field_val(HNS3_TXD_L2LEN_M, HNS3_TXD_L2LEN_S,
2935 m->outer_l2_len >> HNS3_L2_LEN_UNIT);
2937 *type_cs_vlan_tso_len = tmp_inner;
2938 *ol_type_vlan_len_msec = tmp_outer;
2944 hns3_parse_tunneling_params(struct hns3_tx_queue *txq, struct rte_mbuf *m,
2945 uint16_t tx_desc_id)
2947 struct hns3_desc *tx_ring = txq->tx_ring;
2948 struct hns3_desc *desc = &tx_ring[tx_desc_id];
2949 uint32_t tmp_outer = 0;
2950 uint32_t tmp_inner = 0;
2954 * The tunnel header is contained in the inner L2 header field of the
2955 * mbuf, but for hns3 descriptor, it is contained in the outer L4. So,
2956 * there is a need that switching between them. To avoid multiple
2957 * calculations, the length of the L2 header include the outer and
2958 * inner, will be filled during the parsing of tunnel packects.
2960 if (!(m->ol_flags & PKT_TX_TUNNEL_MASK)) {
2962 * For non tunnel type the tunnel type id is 0, so no need to
2963 * assign a value to it. Only the inner(normal) L2 header length
2966 tmp_inner |= hns3_gen_field_val(HNS3_TXD_L2LEN_M,
2967 HNS3_TXD_L2LEN_S, m->l2_len >> HNS3_L2_LEN_UNIT);
2970 * If outer csum is not offload, the outer length may be filled
2971 * with 0. And the length of the outer header is added to the
2972 * inner l2_len. It would lead a cksum error. So driver has to
2973 * calculate the header length.
2975 if (unlikely(!(m->ol_flags & PKT_TX_OUTER_IP_CKSUM) &&
2976 m->outer_l2_len == 0)) {
2977 struct rte_net_hdr_lens hdr_len;
2978 (void)rte_net_get_ptype(m, &hdr_len,
2979 RTE_PTYPE_L2_MASK | RTE_PTYPE_L3_MASK);
2980 m->outer_l3_len = hdr_len.l3_len;
2981 m->outer_l2_len = hdr_len.l2_len;
2982 m->l2_len = m->l2_len - hdr_len.l2_len - hdr_len.l3_len;
2984 hns3_parse_outer_params(m, &tmp_outer);
2985 ret = hns3_parse_inner_params(m, &tmp_outer, &tmp_inner);
2990 desc->tx.ol_type_vlan_len_msec = rte_cpu_to_le_32(tmp_outer);
2991 desc->tx.type_cs_vlan_tso_len = rte_cpu_to_le_32(tmp_inner);
2997 hns3_parse_l3_cksum_params(struct rte_mbuf *m, uint32_t *type_cs_vlan_tso_len)
2999 uint64_t ol_flags = m->ol_flags;
3003 tmp = *type_cs_vlan_tso_len;
3004 if (ol_flags & PKT_TX_IPV4)
3005 l3_type = HNS3_L3T_IPV4;
3006 else if (ol_flags & PKT_TX_IPV6)
3007 l3_type = HNS3_L3T_IPV6;
3009 l3_type = HNS3_L3T_NONE;
3011 /* inner(/normal) L3 header size, defined in 4 bytes */
3012 tmp |= hns3_gen_field_val(HNS3_TXD_L3LEN_M, HNS3_TXD_L3LEN_S,
3013 m->l3_len >> HNS3_L3_LEN_UNIT);
3015 tmp |= hns3_gen_field_val(HNS3_TXD_L3T_M, HNS3_TXD_L3T_S, l3_type);
3017 /* Enable L3 checksum offloads */
3018 if (ol_flags & PKT_TX_IP_CKSUM)
3019 tmp |= BIT(HNS3_TXD_L3CS_B);
3020 *type_cs_vlan_tso_len = tmp;
3024 hns3_parse_l4_cksum_params(struct rte_mbuf *m, uint32_t *type_cs_vlan_tso_len)
3026 uint64_t ol_flags = m->ol_flags;
3028 /* Enable L4 checksum offloads */
3029 switch (ol_flags & (PKT_TX_L4_MASK | PKT_TX_TCP_SEG)) {
3030 case PKT_TX_TCP_CKSUM:
3031 case PKT_TX_TCP_SEG:
3032 tmp = *type_cs_vlan_tso_len;
3033 tmp |= hns3_gen_field_val(HNS3_TXD_L4T_M, HNS3_TXD_L4T_S,
3036 case PKT_TX_UDP_CKSUM:
3037 tmp = *type_cs_vlan_tso_len;
3038 tmp |= hns3_gen_field_val(HNS3_TXD_L4T_M, HNS3_TXD_L4T_S,
3041 case PKT_TX_SCTP_CKSUM:
3042 tmp = *type_cs_vlan_tso_len;
3043 tmp |= hns3_gen_field_val(HNS3_TXD_L4T_M, HNS3_TXD_L4T_S,
3049 tmp |= BIT(HNS3_TXD_L4CS_B);
3050 tmp |= hns3_gen_field_val(HNS3_TXD_L4LEN_M, HNS3_TXD_L4LEN_S,
3051 m->l4_len >> HNS3_L4_LEN_UNIT);
3052 *type_cs_vlan_tso_len = tmp;
3056 hns3_txd_enable_checksum(struct hns3_tx_queue *txq, struct rte_mbuf *m,
3057 uint16_t tx_desc_id)
3059 struct hns3_desc *tx_ring = txq->tx_ring;
3060 struct hns3_desc *desc = &tx_ring[tx_desc_id];
3063 hns3_parse_l3_cksum_params(m, &value);
3064 hns3_parse_l4_cksum_params(m, &value);
3066 desc->tx.type_cs_vlan_tso_len |= rte_cpu_to_le_32(value);
3070 hns3_pkt_need_linearized(struct rte_mbuf *tx_pkts, uint32_t bd_num,
3071 uint32_t max_non_tso_bd_num)
3073 struct rte_mbuf *m_first = tx_pkts;
3074 struct rte_mbuf *m_last = tx_pkts;
3075 uint32_t tot_len = 0;
3080 * Hardware requires that the sum of the data length of every 8
3081 * consecutive buffers is greater than MSS in hns3 network engine.
3082 * We simplify it by ensuring pkt_headlen + the first 8 consecutive
3083 * frags greater than gso header len + mss, and the remaining 7
3084 * consecutive frags greater than MSS except the last 7 frags.
3086 if (bd_num <= max_non_tso_bd_num)
3089 for (i = 0; m_last && i < max_non_tso_bd_num - 1;
3090 i++, m_last = m_last->next)
3091 tot_len += m_last->data_len;
3096 /* ensure the first 8 frags is greater than mss + header */
3097 hdr_len = tx_pkts->l2_len + tx_pkts->l3_len + tx_pkts->l4_len;
3098 hdr_len += (tx_pkts->ol_flags & PKT_TX_TUNNEL_MASK) ?
3099 tx_pkts->outer_l2_len + tx_pkts->outer_l3_len : 0;
3100 if (tot_len + m_last->data_len < tx_pkts->tso_segsz + hdr_len)
3104 * ensure the sum of the data length of every 7 consecutive buffer
3105 * is greater than mss except the last one.
3107 for (i = 0; m_last && i < bd_num - max_non_tso_bd_num; i++) {
3108 tot_len -= m_first->data_len;
3109 tot_len += m_last->data_len;
3111 if (tot_len < tx_pkts->tso_segsz)
3114 m_first = m_first->next;
3115 m_last = m_last->next;
3122 hns3_outer_header_cksum_prepare(struct rte_mbuf *m)
3124 uint64_t ol_flags = m->ol_flags;
3125 uint32_t paylen, hdr_len, l4_proto;
3127 if (!(ol_flags & (PKT_TX_OUTER_IPV4 | PKT_TX_OUTER_IPV6)))
3130 if (ol_flags & PKT_TX_OUTER_IPV4) {
3131 struct rte_ipv4_hdr *ipv4_hdr;
3132 ipv4_hdr = rte_pktmbuf_mtod_offset(m, struct rte_ipv4_hdr *,
3134 l4_proto = ipv4_hdr->next_proto_id;
3135 if (ol_flags & PKT_TX_OUTER_IP_CKSUM)
3136 ipv4_hdr->hdr_checksum = 0;
3138 struct rte_ipv6_hdr *ipv6_hdr;
3139 ipv6_hdr = rte_pktmbuf_mtod_offset(m, struct rte_ipv6_hdr *,
3141 l4_proto = ipv6_hdr->proto;
3143 /* driver should ensure the outer udp cksum is 0 for TUNNEL TSO */
3144 if (l4_proto == IPPROTO_UDP && (ol_flags & PKT_TX_TCP_SEG)) {
3145 struct rte_udp_hdr *udp_hdr;
3146 hdr_len = m->l2_len + m->l3_len + m->l4_len;
3147 hdr_len += m->outer_l2_len + m->outer_l3_len;
3148 paylen = m->pkt_len - hdr_len;
3149 if (paylen <= m->tso_segsz)
3151 udp_hdr = rte_pktmbuf_mtod_offset(m, struct rte_udp_hdr *,
3154 udp_hdr->dgram_cksum = 0;
3159 hns3_check_tso_pkt_valid(struct rte_mbuf *m)
3161 uint32_t tmp_data_len_sum = 0;
3162 uint16_t nb_buf = m->nb_segs;
3163 uint32_t paylen, hdr_len;
3164 struct rte_mbuf *m_seg;
3167 if (nb_buf > HNS3_MAX_TSO_BD_PER_PKT)
3170 hdr_len = m->l2_len + m->l3_len + m->l4_len;
3171 hdr_len += (m->ol_flags & PKT_TX_TUNNEL_MASK) ?
3172 m->outer_l2_len + m->outer_l3_len : 0;
3173 if (hdr_len > HNS3_MAX_TSO_HDR_SIZE)
3176 paylen = m->pkt_len - hdr_len;
3177 if (paylen > HNS3_MAX_BD_PAYLEN)
3181 * The TSO header (include outer and inner L2, L3 and L4 header)
3182 * should be provided by three descriptors in maximum in hns3 network
3186 for (i = 0; m_seg != NULL && i < HNS3_MAX_TSO_HDR_BD_NUM && i < nb_buf;
3187 i++, m_seg = m_seg->next) {
3188 tmp_data_len_sum += m_seg->data_len;
3191 if (hdr_len > tmp_data_len_sum)
3197 #ifdef RTE_LIBRTE_ETHDEV_DEBUG
3199 hns3_vld_vlan_chk(struct hns3_tx_queue *txq, struct rte_mbuf *m)
3201 struct rte_ether_hdr *eh;
3202 struct rte_vlan_hdr *vh;
3204 if (!txq->pvid_sw_shift_en)
3208 * Due to hardware limitations, we only support two-layer VLAN hardware
3209 * offload in Tx direction based on hns3 network engine, so when PVID is
3210 * enabled, QinQ insert is no longer supported.
3211 * And when PVID is enabled, in the following two cases:
3212 * i) packets with more than two VLAN tags.
3213 * ii) packets with one VLAN tag while the hardware VLAN insert is
3215 * The packets will be regarded as abnormal packets and discarded by
3216 * hardware in Tx direction. For debugging purposes, a validation check
3217 * for these types of packets is added to the '.tx_pkt_prepare' ops
3218 * implementation function named hns3_prep_pkts to inform users that
3219 * these packets will be discarded.
3221 if (m->ol_flags & PKT_TX_QINQ_PKT)
3224 eh = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
3225 if (eh->ether_type == rte_cpu_to_be_16(RTE_ETHER_TYPE_VLAN)) {
3226 if (m->ol_flags & PKT_TX_VLAN_PKT)
3229 /* Ensure the incoming packet is not a QinQ packet */
3230 vh = (struct rte_vlan_hdr *)(eh + 1);
3231 if (vh->eth_proto == rte_cpu_to_be_16(RTE_ETHER_TYPE_VLAN))
3240 hns3_prep_pkt_proc(struct hns3_tx_queue *tx_queue, struct rte_mbuf *m)
3244 #ifdef RTE_LIBRTE_ETHDEV_DEBUG
3245 ret = rte_validate_tx_offload(m);
3251 ret = hns3_vld_vlan_chk(tx_queue, m);
3257 if (hns3_pkt_is_tso(m)) {
3258 if (hns3_pkt_need_linearized(m, m->nb_segs,
3259 tx_queue->max_non_tso_bd_num) ||
3260 hns3_check_tso_pkt_valid(m)) {
3265 if (tx_queue->tso_mode != HNS3_TSO_SW_CAL_PSEUDO_H_CSUM) {
3267 * (tso mode != HNS3_TSO_SW_CAL_PSEUDO_H_CSUM) means
3268 * hardware support recalculate the TCP pseudo header
3269 * checksum of packets that need TSO, so network driver
3270 * software not need to recalculate it.
3272 hns3_outer_header_cksum_prepare(m);
3277 ret = rte_net_intel_cksum_prepare(m);
3283 hns3_outer_header_cksum_prepare(m);
3289 hns3_prep_pkts(__rte_unused void *tx_queue, struct rte_mbuf **tx_pkts,
3295 for (i = 0; i < nb_pkts; i++) {
3297 if (hns3_prep_pkt_proc(tx_queue, m))
3305 hns3_parse_cksum(struct hns3_tx_queue *txq, uint16_t tx_desc_id,
3308 struct hns3_desc *tx_ring = txq->tx_ring;
3309 struct hns3_desc *desc = &tx_ring[tx_desc_id];
3311 /* Enable checksum offloading */
3312 if (m->ol_flags & HNS3_TX_CKSUM_OFFLOAD_MASK) {
3313 /* Fill in tunneling parameters if necessary */
3314 if (hns3_parse_tunneling_params(txq, m, tx_desc_id)) {
3315 txq->unsupported_tunnel_pkt_cnt++;
3319 hns3_txd_enable_checksum(txq, m, tx_desc_id);
3321 /* clear the control bit */
3322 desc->tx.type_cs_vlan_tso_len = 0;
3323 desc->tx.ol_type_vlan_len_msec = 0;
3330 hns3_check_non_tso_pkt(uint16_t nb_buf, struct rte_mbuf **m_seg,
3331 struct rte_mbuf *tx_pkt, struct hns3_tx_queue *txq)
3333 uint8_t max_non_tso_bd_num;
3334 struct rte_mbuf *new_pkt;
3337 if (hns3_pkt_is_tso(*m_seg))
3341 * If packet length is greater than HNS3_MAX_FRAME_LEN
3342 * driver support, the packet will be ignored.
3344 if (unlikely(rte_pktmbuf_pkt_len(tx_pkt) > HNS3_MAX_FRAME_LEN)) {
3345 txq->over_length_pkt_cnt++;
3349 max_non_tso_bd_num = txq->max_non_tso_bd_num;
3350 if (unlikely(nb_buf > max_non_tso_bd_num)) {
3351 txq->exceed_limit_bd_pkt_cnt++;
3352 ret = hns3_reassemble_tx_pkts(tx_pkt, &new_pkt,
3353 max_non_tso_bd_num);
3355 txq->exceed_limit_bd_reassem_fail++;
3365 hns3_tx_free_buffer_simple(struct hns3_tx_queue *txq)
3367 struct hns3_entry *tx_entry;
3368 struct hns3_desc *desc;
3369 uint16_t tx_next_clean;
3373 if (HNS3_GET_TX_QUEUE_PEND_BD_NUM(txq) < txq->tx_rs_thresh)
3377 * All mbufs can be released only when the VLD bits of all
3378 * descriptors in a batch are cleared.
3380 tx_next_clean = (txq->next_to_clean + txq->tx_rs_thresh - 1) %
3382 desc = &txq->tx_ring[tx_next_clean];
3383 for (i = 0; i < txq->tx_rs_thresh; i++) {
3384 if (rte_le_to_cpu_16(desc->tx.tp_fe_sc_vld_ra_ri) &
3385 BIT(HNS3_TXD_VLD_B))
3390 tx_entry = &txq->sw_ring[txq->next_to_clean];
3392 for (i = 0; i < txq->tx_rs_thresh; i++)
3393 rte_prefetch0((tx_entry + i)->mbuf);
3394 for (i = 0; i < txq->tx_rs_thresh; i++, tx_entry++) {
3395 rte_mempool_put(tx_entry->mbuf->pool, tx_entry->mbuf);
3396 tx_entry->mbuf = NULL;
3399 txq->next_to_clean = (tx_next_clean + 1) % txq->nb_tx_desc;
3400 txq->tx_bd_ready += txq->tx_rs_thresh;
3405 hns3_tx_backup_1mbuf(struct hns3_entry *tx_entry, struct rte_mbuf **pkts)
3407 tx_entry->mbuf = pkts[0];
3411 hns3_tx_backup_4mbuf(struct hns3_entry *tx_entry, struct rte_mbuf **pkts)
3413 hns3_tx_backup_1mbuf(&tx_entry[0], &pkts[0]);
3414 hns3_tx_backup_1mbuf(&tx_entry[1], &pkts[1]);
3415 hns3_tx_backup_1mbuf(&tx_entry[2], &pkts[2]);
3416 hns3_tx_backup_1mbuf(&tx_entry[3], &pkts[3]);
3420 hns3_tx_setup_4bd(struct hns3_desc *txdp, struct rte_mbuf **pkts)
3422 #define PER_LOOP_NUM 4
3423 const uint16_t bd_flag = BIT(HNS3_TXD_VLD_B) | BIT(HNS3_TXD_FE_B);
3427 for (i = 0; i < PER_LOOP_NUM; i++, txdp++, pkts++) {
3428 dma_addr = rte_mbuf_data_iova(*pkts);
3429 txdp->addr = rte_cpu_to_le_64(dma_addr);
3430 txdp->tx.send_size = rte_cpu_to_le_16((*pkts)->data_len);
3431 txdp->tx.paylen = 0;
3432 txdp->tx.type_cs_vlan_tso_len = 0;
3433 txdp->tx.ol_type_vlan_len_msec = 0;
3434 txdp->tx.tp_fe_sc_vld_ra_ri = rte_cpu_to_le_16(bd_flag);
3439 hns3_tx_setup_1bd(struct hns3_desc *txdp, struct rte_mbuf **pkts)
3441 const uint16_t bd_flag = BIT(HNS3_TXD_VLD_B) | BIT(HNS3_TXD_FE_B);
3444 dma_addr = rte_mbuf_data_iova(*pkts);
3445 txdp->addr = rte_cpu_to_le_64(dma_addr);
3446 txdp->tx.send_size = rte_cpu_to_le_16((*pkts)->data_len);
3447 txdp->tx.paylen = 0;
3448 txdp->tx.type_cs_vlan_tso_len = 0;
3449 txdp->tx.ol_type_vlan_len_msec = 0;
3450 txdp->tx.tp_fe_sc_vld_ra_ri = rte_cpu_to_le_16(bd_flag);
3454 hns3_tx_fill_hw_ring(struct hns3_tx_queue *txq,
3455 struct rte_mbuf **pkts,
3458 #define PER_LOOP_NUM 4
3459 #define PER_LOOP_MASK (PER_LOOP_NUM - 1)
3460 struct hns3_desc *txdp = &txq->tx_ring[txq->next_to_use];
3461 struct hns3_entry *tx_entry = &txq->sw_ring[txq->next_to_use];
3462 const uint32_t mainpart = (nb_pkts & ((uint32_t)~PER_LOOP_MASK));
3463 const uint32_t leftover = (nb_pkts & ((uint32_t)PER_LOOP_MASK));
3466 for (i = 0; i < mainpart; i += PER_LOOP_NUM) {
3467 hns3_tx_backup_4mbuf(tx_entry + i, pkts + i);
3468 hns3_tx_setup_4bd(txdp + i, pkts + i);
3470 if (unlikely(leftover > 0)) {
3471 for (i = 0; i < leftover; i++) {
3472 hns3_tx_backup_1mbuf(tx_entry + mainpart + i,
3473 pkts + mainpart + i);
3474 hns3_tx_setup_1bd(txdp + mainpart + i,
3475 pkts + mainpart + i);
3481 hns3_xmit_pkts_simple(void *tx_queue,
3482 struct rte_mbuf **tx_pkts,
3485 struct hns3_tx_queue *txq = tx_queue;
3488 hns3_tx_free_buffer_simple(txq);
3490 nb_pkts = RTE_MIN(txq->tx_bd_ready, nb_pkts);
3491 if (unlikely(nb_pkts == 0)) {
3492 if (txq->tx_bd_ready == 0)
3493 txq->queue_full_cnt++;
3497 txq->tx_bd_ready -= nb_pkts;
3498 if (txq->next_to_use + nb_pkts > txq->nb_tx_desc) {
3499 nb_tx = txq->nb_tx_desc - txq->next_to_use;
3500 hns3_tx_fill_hw_ring(txq, tx_pkts, nb_tx);
3501 txq->next_to_use = 0;
3504 hns3_tx_fill_hw_ring(txq, tx_pkts + nb_tx, nb_pkts - nb_tx);
3505 txq->next_to_use += nb_pkts - nb_tx;
3507 hns3_write_reg_opt(txq->io_tail_reg, nb_pkts);
3513 hns3_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
3515 struct hns3_tx_queue *txq = tx_queue;
3516 struct hns3_entry *tx_bak_pkt;
3517 struct hns3_desc *tx_ring;
3518 struct rte_mbuf *tx_pkt;
3519 struct rte_mbuf *m_seg;
3520 struct hns3_desc *desc;
3521 uint32_t nb_hold = 0;
3522 uint16_t tx_next_use;
3523 uint16_t tx_pkt_num;
3529 /* free useless buffer */
3530 hns3_tx_free_useless_buffer(txq);
3532 tx_next_use = txq->next_to_use;
3533 tx_bd_max = txq->nb_tx_desc;
3534 tx_pkt_num = nb_pkts;
3535 tx_ring = txq->tx_ring;
3538 tx_bak_pkt = &txq->sw_ring[tx_next_use];
3539 for (nb_tx = 0; nb_tx < tx_pkt_num; nb_tx++) {
3540 tx_pkt = *tx_pkts++;
3542 nb_buf = tx_pkt->nb_segs;
3544 if (nb_buf > txq->tx_bd_ready) {
3545 txq->queue_full_cnt++;
3553 * If packet length is less than minimum packet length supported
3554 * by hardware in Tx direction, driver need to pad it to avoid
3557 if (unlikely(rte_pktmbuf_pkt_len(tx_pkt) <
3558 txq->min_tx_pkt_len)) {
3562 add_len = txq->min_tx_pkt_len -
3563 rte_pktmbuf_pkt_len(tx_pkt);
3564 appended = rte_pktmbuf_append(tx_pkt, add_len);
3565 if (appended == NULL) {
3566 txq->pkt_padding_fail_cnt++;
3570 memset(appended, 0, add_len);
3575 if (hns3_check_non_tso_pkt(nb_buf, &m_seg, tx_pkt, txq))
3578 if (hns3_parse_cksum(txq, tx_next_use, m_seg))
3582 desc = &tx_ring[tx_next_use];
3585 * If the packet is divided into multiple Tx Buffer Descriptors,
3586 * only need to fill vlan, paylen and tso into the first Tx
3587 * Buffer Descriptor.
3589 hns3_fill_first_desc(txq, desc, m_seg);
3592 desc = &tx_ring[tx_next_use];
3594 * Fill valid bits, DMA address and data length for each
3595 * Tx Buffer Descriptor.
3597 hns3_fill_per_desc(desc, m_seg);
3598 tx_bak_pkt->mbuf = m_seg;
3599 m_seg = m_seg->next;
3602 if (tx_next_use >= tx_bd_max) {
3604 tx_bak_pkt = txq->sw_ring;
3608 } while (m_seg != NULL);
3610 /* Add end flag for the last Tx Buffer Descriptor */
3611 desc->tx.tp_fe_sc_vld_ra_ri |=
3612 rte_cpu_to_le_16(BIT(HNS3_TXD_FE_B));
3615 txq->next_to_use = tx_next_use;
3616 txq->tx_bd_ready -= i;
3622 hns3_write_reg_opt(txq->io_tail_reg, nb_hold);
3628 hns3_tx_check_vec_support(__rte_unused struct rte_eth_dev *dev)
3634 hns3_xmit_pkts_vec(__rte_unused void *tx_queue,
3635 __rte_unused struct rte_mbuf **tx_pkts,
3636 __rte_unused uint16_t nb_pkts)
3642 hns3_xmit_pkts_vec_sve(void __rte_unused * tx_queue,
3643 struct rte_mbuf __rte_unused **tx_pkts,
3644 uint16_t __rte_unused nb_pkts)
3650 hns3_tx_burst_mode_get(struct rte_eth_dev *dev, __rte_unused uint16_t queue_id,
3651 struct rte_eth_burst_mode *mode)
3653 eth_tx_burst_t pkt_burst = dev->tx_pkt_burst;
3654 const char *info = NULL;
3656 if (pkt_burst == hns3_xmit_pkts_simple)
3657 info = "Scalar Simple";
3658 else if (pkt_burst == hns3_xmit_pkts)
3660 else if (pkt_burst == hns3_xmit_pkts_vec)
3661 info = "Vector Neon";
3662 else if (pkt_burst == hns3_xmit_pkts_vec_sve)
3663 info = "Vector Sve";
3668 snprintf(mode->info, sizeof(mode->info), "%s", info);
3673 static eth_tx_burst_t
3674 hns3_get_tx_function(struct rte_eth_dev *dev, eth_tx_prep_t *prep)
3676 uint64_t offloads = dev->data->dev_conf.txmode.offloads;
3677 struct hns3_adapter *hns = dev->data->dev_private;
3679 if (hns->tx_vec_allowed && hns3_tx_check_vec_support(dev) == 0) {
3681 return hns3_check_sve_support() ? hns3_xmit_pkts_vec_sve :
3685 if (hns->tx_simple_allowed &&
3686 offloads == (offloads & DEV_TX_OFFLOAD_MBUF_FAST_FREE)) {
3688 return hns3_xmit_pkts_simple;
3691 *prep = hns3_prep_pkts;
3692 return hns3_xmit_pkts;
3696 hns3_dummy_rxtx_burst(void *dpdk_txq __rte_unused,
3697 struct rte_mbuf **pkts __rte_unused,
3698 uint16_t pkts_n __rte_unused)
3703 void hns3_set_rxtx_function(struct rte_eth_dev *eth_dev)
3705 struct hns3_adapter *hns = eth_dev->data->dev_private;
3706 eth_tx_prep_t prep = NULL;
3708 if (hns->hw.adapter_state == HNS3_NIC_STARTED &&
3709 rte_atomic16_read(&hns->hw.reset.resetting) == 0) {
3710 eth_dev->rx_pkt_burst = hns3_get_rx_function(eth_dev);
3711 eth_dev->tx_pkt_burst = hns3_get_tx_function(eth_dev, &prep);
3712 eth_dev->tx_pkt_prepare = prep;
3714 eth_dev->rx_pkt_burst = hns3_dummy_rxtx_burst;
3715 eth_dev->tx_pkt_burst = hns3_dummy_rxtx_burst;
3716 eth_dev->tx_pkt_prepare = hns3_dummy_rxtx_burst;
3721 hns3_rxq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
3722 struct rte_eth_rxq_info *qinfo)
3724 struct hns3_rx_queue *rxq = dev->data->rx_queues[queue_id];
3726 qinfo->mp = rxq->mb_pool;
3727 qinfo->nb_desc = rxq->nb_rx_desc;
3728 qinfo->scattered_rx = dev->data->scattered_rx;
3729 /* Report the HW Rx buffer length to user */
3730 qinfo->rx_buf_size = rxq->rx_buf_len;
3733 * If there are no available Rx buffer descriptors, incoming packets
3734 * are always dropped by hardware based on hns3 network engine.
3736 qinfo->conf.rx_drop_en = 1;
3737 qinfo->conf.offloads = dev->data->dev_conf.rxmode.offloads;
3738 qinfo->conf.rx_free_thresh = rxq->rx_free_thresh;
3739 qinfo->conf.rx_deferred_start = rxq->rx_deferred_start;
3743 hns3_txq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
3744 struct rte_eth_txq_info *qinfo)
3746 struct hns3_tx_queue *txq = dev->data->tx_queues[queue_id];
3748 qinfo->nb_desc = txq->nb_tx_desc;
3749 qinfo->conf.offloads = dev->data->dev_conf.txmode.offloads;
3750 qinfo->conf.tx_rs_thresh = txq->tx_rs_thresh;
3751 qinfo->conf.tx_free_thresh = txq->tx_free_thresh;
3752 qinfo->conf.tx_deferred_start = txq->tx_deferred_start;
3756 hns3_dev_rx_queue_start(struct rte_eth_dev *dev, uint16_t rx_queue_id)
3758 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3759 struct hns3_rx_queue *rxq = dev->data->rx_queues[rx_queue_id];
3760 struct hns3_adapter *hns = HNS3_DEV_HW_TO_ADAPTER(hw);
3763 if (!hns3_dev_indep_txrx_supported(hw))
3766 ret = hns3_reset_queue(hw, rx_queue_id, HNS3_RING_TYPE_RX);
3768 hns3_err(hw, "fail to reset Rx queue %u, ret = %d.",
3773 ret = hns3_init_rxq(hns, rx_queue_id);
3775 hns3_err(hw, "fail to init Rx queue %u, ret = %d.",
3780 hns3_enable_rxq(rxq, true);
3781 dev->data->rx_queue_state[rx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
3787 hns3_reset_sw_rxq(struct hns3_rx_queue *rxq)
3789 rxq->next_to_use = 0;
3790 rxq->rx_rearm_start = 0;
3791 rxq->rx_free_hold = 0;
3792 rxq->rx_rearm_nb = 0;
3793 rxq->pkt_first_seg = NULL;
3794 rxq->pkt_last_seg = NULL;
3795 memset(&rxq->rx_ring[0], 0, rxq->nb_rx_desc * sizeof(struct hns3_desc));
3796 hns3_rxq_vec_setup(rxq);
3800 hns3_dev_rx_queue_stop(struct rte_eth_dev *dev, uint16_t rx_queue_id)
3802 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3803 struct hns3_rx_queue *rxq = dev->data->rx_queues[rx_queue_id];
3805 if (!hns3_dev_indep_txrx_supported(hw))
3808 hns3_enable_rxq(rxq, false);
3810 hns3_rx_queue_release_mbufs(rxq);
3812 hns3_reset_sw_rxq(rxq);
3813 dev->data->rx_queue_state[rx_queue_id] = RTE_ETH_QUEUE_STATE_STOPPED;
3819 hns3_dev_tx_queue_start(struct rte_eth_dev *dev, uint16_t tx_queue_id)
3821 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3822 struct hns3_tx_queue *txq = dev->data->tx_queues[tx_queue_id];
3825 if (!hns3_dev_indep_txrx_supported(hw))
3828 ret = hns3_reset_queue(hw, tx_queue_id, HNS3_RING_TYPE_TX);
3830 hns3_err(hw, "fail to reset Tx queue %u, ret = %d.",
3836 hns3_enable_txq(txq, true);
3837 dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
3843 hns3_dev_tx_queue_stop(struct rte_eth_dev *dev, uint16_t tx_queue_id)
3845 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3846 struct hns3_tx_queue *txq = dev->data->tx_queues[tx_queue_id];
3848 if (!hns3_dev_indep_txrx_supported(hw))
3851 hns3_enable_txq(txq, false);
3852 hns3_tx_queue_release_mbufs(txq);
3854 * All the mbufs in sw_ring are released and all the pointers in sw_ring
3855 * are set to NULL. If this queue is still called by upper layer,
3856 * residual SW status of this txq may cause these pointers in sw_ring
3857 * which have been set to NULL to be released again. To avoid it,
3861 dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STOPPED;
3867 hns3_rx_queue_count(struct rte_eth_dev *dev, uint16_t rx_queue_id)
3870 * Number of BDs that have been processed by the driver
3871 * but have not been notified to the hardware.
3873 uint32_t driver_hold_bd_num;
3874 struct hns3_rx_queue *rxq;
3877 rxq = dev->data->rx_queues[rx_queue_id];
3878 fbd_num = hns3_read_dev(rxq, HNS3_RING_RX_FBDNUM_REG);
3879 if (dev->rx_pkt_burst == hns3_recv_pkts_vec ||
3880 dev->rx_pkt_burst == hns3_recv_pkts_vec_sve)
3881 driver_hold_bd_num = rxq->rx_rearm_nb;
3883 driver_hold_bd_num = rxq->rx_free_hold;
3885 if (fbd_num <= driver_hold_bd_num)
3888 return fbd_num - driver_hold_bd_num;