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>
8 #include <rte_geneve.h>
10 #include <ethdev_driver.h>
13 #include <rte_malloc.h>
14 #if defined(RTE_ARCH_ARM64) && defined(__ARM_FEATURE_SVE)
15 #include <rte_cpuflags.h>
18 #include "hns3_ethdev.h"
19 #include "hns3_rxtx.h"
20 #include "hns3_regs.h"
21 #include "hns3_logs.h"
23 #define HNS3_CFG_DESC_NUM(num) ((num) / 8 - 1)
24 #define HNS3_RX_RING_PREFETCTH_MASK 3
27 hns3_rx_queue_release_mbufs(struct hns3_rx_queue *rxq)
31 /* Note: Fake rx queue will not enter here */
32 if (rxq->sw_ring == NULL)
35 if (rxq->rx_rearm_nb == 0) {
36 for (i = 0; i < rxq->nb_rx_desc; i++) {
37 if (rxq->sw_ring[i].mbuf != NULL) {
38 rte_pktmbuf_free_seg(rxq->sw_ring[i].mbuf);
39 rxq->sw_ring[i].mbuf = NULL;
43 for (i = rxq->next_to_use;
44 i != rxq->rx_rearm_start;
45 i = (i + 1) % rxq->nb_rx_desc) {
46 if (rxq->sw_ring[i].mbuf != NULL) {
47 rte_pktmbuf_free_seg(rxq->sw_ring[i].mbuf);
48 rxq->sw_ring[i].mbuf = NULL;
53 for (i = 0; i < rxq->bulk_mbuf_num; i++)
54 rte_pktmbuf_free_seg(rxq->bulk_mbuf[i]);
55 rxq->bulk_mbuf_num = 0;
57 if (rxq->pkt_first_seg) {
58 rte_pktmbuf_free(rxq->pkt_first_seg);
59 rxq->pkt_first_seg = NULL;
64 hns3_tx_queue_release_mbufs(struct hns3_tx_queue *txq)
68 /* Note: Fake tx queue will not enter here */
70 for (i = 0; i < txq->nb_tx_desc; i++) {
71 if (txq->sw_ring[i].mbuf) {
72 rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
73 txq->sw_ring[i].mbuf = NULL;
80 hns3_rx_queue_release(void *queue)
82 struct hns3_rx_queue *rxq = queue;
84 hns3_rx_queue_release_mbufs(rxq);
86 rte_memzone_free(rxq->mz);
88 rte_free(rxq->sw_ring);
94 hns3_tx_queue_release(void *queue)
96 struct hns3_tx_queue *txq = queue;
98 hns3_tx_queue_release_mbufs(txq);
100 rte_memzone_free(txq->mz);
102 rte_free(txq->sw_ring);
110 hns3_dev_rx_queue_release(void *queue)
112 struct hns3_rx_queue *rxq = queue;
113 struct hns3_adapter *hns;
119 rte_spinlock_lock(&hns->hw.lock);
120 hns3_rx_queue_release(queue);
121 rte_spinlock_unlock(&hns->hw.lock);
125 hns3_dev_tx_queue_release(void *queue)
127 struct hns3_tx_queue *txq = queue;
128 struct hns3_adapter *hns;
134 rte_spinlock_lock(&hns->hw.lock);
135 hns3_tx_queue_release(queue);
136 rte_spinlock_unlock(&hns->hw.lock);
140 hns3_fake_rx_queue_release(struct hns3_rx_queue *queue)
142 struct hns3_rx_queue *rxq = queue;
143 struct hns3_adapter *hns;
153 if (hw->fkq_data.rx_queues[idx]) {
154 hns3_rx_queue_release(hw->fkq_data.rx_queues[idx]);
155 hw->fkq_data.rx_queues[idx] = NULL;
158 /* free fake rx queue arrays */
159 if (idx == (hw->fkq_data.nb_fake_rx_queues - 1)) {
160 hw->fkq_data.nb_fake_rx_queues = 0;
161 rte_free(hw->fkq_data.rx_queues);
162 hw->fkq_data.rx_queues = NULL;
167 hns3_fake_tx_queue_release(struct hns3_tx_queue *queue)
169 struct hns3_tx_queue *txq = queue;
170 struct hns3_adapter *hns;
180 if (hw->fkq_data.tx_queues[idx]) {
181 hns3_tx_queue_release(hw->fkq_data.tx_queues[idx]);
182 hw->fkq_data.tx_queues[idx] = NULL;
185 /* free fake tx queue arrays */
186 if (idx == (hw->fkq_data.nb_fake_tx_queues - 1)) {
187 hw->fkq_data.nb_fake_tx_queues = 0;
188 rte_free(hw->fkq_data.tx_queues);
189 hw->fkq_data.tx_queues = NULL;
194 hns3_free_rx_queues(struct rte_eth_dev *dev)
196 struct hns3_adapter *hns = dev->data->dev_private;
197 struct hns3_fake_queue_data *fkq_data;
198 struct hns3_hw *hw = &hns->hw;
202 nb_rx_q = hw->data->nb_rx_queues;
203 for (i = 0; i < nb_rx_q; i++) {
204 if (dev->data->rx_queues[i]) {
205 hns3_rx_queue_release(dev->data->rx_queues[i]);
206 dev->data->rx_queues[i] = NULL;
210 /* Free fake Rx queues */
211 fkq_data = &hw->fkq_data;
212 for (i = 0; i < fkq_data->nb_fake_rx_queues; i++) {
213 if (fkq_data->rx_queues[i])
214 hns3_fake_rx_queue_release(fkq_data->rx_queues[i]);
219 hns3_free_tx_queues(struct rte_eth_dev *dev)
221 struct hns3_adapter *hns = dev->data->dev_private;
222 struct hns3_fake_queue_data *fkq_data;
223 struct hns3_hw *hw = &hns->hw;
227 nb_tx_q = hw->data->nb_tx_queues;
228 for (i = 0; i < nb_tx_q; i++) {
229 if (dev->data->tx_queues[i]) {
230 hns3_tx_queue_release(dev->data->tx_queues[i]);
231 dev->data->tx_queues[i] = NULL;
235 /* Free fake Tx queues */
236 fkq_data = &hw->fkq_data;
237 for (i = 0; i < fkq_data->nb_fake_tx_queues; i++) {
238 if (fkq_data->tx_queues[i])
239 hns3_fake_tx_queue_release(fkq_data->tx_queues[i]);
244 hns3_free_all_queues(struct rte_eth_dev *dev)
246 hns3_free_rx_queues(dev);
247 hns3_free_tx_queues(dev);
251 hns3_alloc_rx_queue_mbufs(struct hns3_hw *hw, struct hns3_rx_queue *rxq)
253 struct rte_mbuf *mbuf;
257 for (i = 0; i < rxq->nb_rx_desc; i++) {
258 mbuf = rte_mbuf_raw_alloc(rxq->mb_pool);
259 if (unlikely(mbuf == NULL)) {
260 hns3_err(hw, "Failed to allocate RXD[%u] for rx queue!",
262 hns3_rx_queue_release_mbufs(rxq);
266 rte_mbuf_refcnt_set(mbuf, 1);
268 mbuf->data_off = RTE_PKTMBUF_HEADROOM;
270 mbuf->port = rxq->port_id;
272 rxq->sw_ring[i].mbuf = mbuf;
273 dma_addr = rte_cpu_to_le_64(rte_mbuf_data_iova_default(mbuf));
274 rxq->rx_ring[i].addr = dma_addr;
275 rxq->rx_ring[i].rx.bd_base_info = 0;
282 hns3_buf_size2type(uint32_t buf_size)
288 bd_size_type = HNS3_BD_SIZE_512_TYPE;
291 bd_size_type = HNS3_BD_SIZE_1024_TYPE;
294 bd_size_type = HNS3_BD_SIZE_4096_TYPE;
297 bd_size_type = HNS3_BD_SIZE_2048_TYPE;
304 hns3_init_rx_queue_hw(struct hns3_rx_queue *rxq)
306 uint32_t rx_buf_len = rxq->rx_buf_len;
307 uint64_t dma_addr = rxq->rx_ring_phys_addr;
309 hns3_write_dev(rxq, HNS3_RING_RX_BASEADDR_L_REG, (uint32_t)dma_addr);
310 hns3_write_dev(rxq, HNS3_RING_RX_BASEADDR_H_REG,
311 (uint32_t)((dma_addr >> 31) >> 1));
313 hns3_write_dev(rxq, HNS3_RING_RX_BD_LEN_REG,
314 hns3_buf_size2type(rx_buf_len));
315 hns3_write_dev(rxq, HNS3_RING_RX_BD_NUM_REG,
316 HNS3_CFG_DESC_NUM(rxq->nb_rx_desc));
320 hns3_init_tx_queue_hw(struct hns3_tx_queue *txq)
322 uint64_t dma_addr = txq->tx_ring_phys_addr;
324 hns3_write_dev(txq, HNS3_RING_TX_BASEADDR_L_REG, (uint32_t)dma_addr);
325 hns3_write_dev(txq, HNS3_RING_TX_BASEADDR_H_REG,
326 (uint32_t)((dma_addr >> 31) >> 1));
328 hns3_write_dev(txq, HNS3_RING_TX_BD_NUM_REG,
329 HNS3_CFG_DESC_NUM(txq->nb_tx_desc));
333 hns3_update_all_queues_pvid_proc_en(struct hns3_hw *hw)
335 uint16_t nb_rx_q = hw->data->nb_rx_queues;
336 uint16_t nb_tx_q = hw->data->nb_tx_queues;
337 struct hns3_rx_queue *rxq;
338 struct hns3_tx_queue *txq;
342 pvid_en = hw->port_base_vlan_cfg.state == HNS3_PORT_BASE_VLAN_ENABLE;
343 for (i = 0; i < hw->cfg_max_queues; i++) {
345 rxq = hw->data->rx_queues[i];
347 rxq->pvid_sw_discard_en = pvid_en;
350 txq = hw->data->tx_queues[i];
352 txq->pvid_sw_shift_en = pvid_en;
358 hns3_stop_unused_queue(void *tqp_base, enum hns3_ring_type queue_type)
363 reg_offset = queue_type == HNS3_RING_TYPE_TX ?
364 HNS3_RING_TX_EN_REG : HNS3_RING_RX_EN_REG;
365 reg = hns3_read_reg(tqp_base, reg_offset);
366 reg &= ~BIT(HNS3_RING_EN_B);
367 hns3_write_reg(tqp_base, reg_offset, reg);
371 hns3_enable_all_queues(struct hns3_hw *hw, bool en)
373 uint16_t nb_rx_q = hw->data->nb_rx_queues;
374 uint16_t nb_tx_q = hw->data->nb_tx_queues;
375 struct hns3_rx_queue *rxq;
376 struct hns3_tx_queue *txq;
381 for (i = 0; i < hw->cfg_max_queues; i++) {
382 if (hns3_dev_indep_txrx_supported(hw)) {
383 rxq = i < nb_rx_q ? hw->data->rx_queues[i] : NULL;
384 txq = i < nb_tx_q ? hw->data->tx_queues[i] : NULL;
386 tqp_base = (void *)((char *)hw->io_base +
387 hns3_get_tqp_reg_offset(i));
389 * If queue struct is not initialized, it means the
390 * related HW ring has not been initialized yet.
391 * So, these queues should be disabled before enable
392 * the tqps to avoid a HW exception since the queues
393 * are enabled by default.
396 hns3_stop_unused_queue(tqp_base,
399 hns3_stop_unused_queue(tqp_base,
402 rxq = i < nb_rx_q ? hw->data->rx_queues[i] :
403 hw->fkq_data.rx_queues[i - nb_rx_q];
405 tqp_base = rxq->io_base;
408 * This is the master switch that used to control the enabling
409 * of a pair of Tx and Rx queues. Both the Rx and Tx point to
412 rcb_reg = hns3_read_reg(tqp_base, HNS3_RING_EN_REG);
414 rcb_reg |= BIT(HNS3_RING_EN_B);
416 rcb_reg &= ~BIT(HNS3_RING_EN_B);
417 hns3_write_reg(tqp_base, HNS3_RING_EN_REG, rcb_reg);
422 hns3_enable_txq(struct hns3_tx_queue *txq, bool en)
424 struct hns3_hw *hw = &txq->hns->hw;
427 if (hns3_dev_indep_txrx_supported(hw)) {
428 reg = hns3_read_dev(txq, HNS3_RING_TX_EN_REG);
430 reg |= BIT(HNS3_RING_EN_B);
432 reg &= ~BIT(HNS3_RING_EN_B);
433 hns3_write_dev(txq, HNS3_RING_TX_EN_REG, reg);
439 hns3_enable_rxq(struct hns3_rx_queue *rxq, bool en)
441 struct hns3_hw *hw = &rxq->hns->hw;
444 if (hns3_dev_indep_txrx_supported(hw)) {
445 reg = hns3_read_dev(rxq, HNS3_RING_RX_EN_REG);
447 reg |= BIT(HNS3_RING_EN_B);
449 reg &= ~BIT(HNS3_RING_EN_B);
450 hns3_write_dev(rxq, HNS3_RING_RX_EN_REG, reg);
456 hns3_start_all_txqs(struct rte_eth_dev *dev)
458 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
459 struct hns3_tx_queue *txq;
462 for (i = 0; i < dev->data->nb_tx_queues; i++) {
463 txq = hw->data->tx_queues[i];
465 hns3_err(hw, "Tx queue %u not available or setup.", i);
466 goto start_txqs_fail;
469 * Tx queue is enabled by default. Therefore, the Tx queues
470 * needs to be disabled when deferred_start is set. There is
471 * another master switch used to control the enabling of a pair
472 * of Tx and Rx queues. And the master switch is disabled by
475 if (txq->tx_deferred_start)
476 hns3_enable_txq(txq, false);
478 hns3_enable_txq(txq, true);
483 for (j = 0; j < i; j++) {
484 txq = hw->data->tx_queues[j];
485 hns3_enable_txq(txq, false);
491 hns3_start_all_rxqs(struct rte_eth_dev *dev)
493 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
494 struct hns3_rx_queue *rxq;
497 for (i = 0; i < dev->data->nb_rx_queues; i++) {
498 rxq = hw->data->rx_queues[i];
500 hns3_err(hw, "Rx queue %u not available or setup.", i);
501 goto start_rxqs_fail;
504 * Rx queue is enabled by default. Therefore, the Rx queues
505 * needs to be disabled when deferred_start is set. There is
506 * another master switch used to control the enabling of a pair
507 * of Tx and Rx queues. And the master switch is disabled by
510 if (rxq->rx_deferred_start)
511 hns3_enable_rxq(rxq, false);
513 hns3_enable_rxq(rxq, true);
518 for (j = 0; j < i; j++) {
519 rxq = hw->data->rx_queues[j];
520 hns3_enable_rxq(rxq, false);
526 hns3_restore_tqp_enable_state(struct hns3_hw *hw)
528 struct hns3_rx_queue *rxq;
529 struct hns3_tx_queue *txq;
532 for (i = 0; i < hw->data->nb_rx_queues; i++) {
533 rxq = hw->data->rx_queues[i];
535 hns3_enable_rxq(rxq, rxq->enabled);
538 for (i = 0; i < hw->data->nb_tx_queues; i++) {
539 txq = hw->data->tx_queues[i];
541 hns3_enable_txq(txq, txq->enabled);
546 hns3_stop_all_txqs(struct rte_eth_dev *dev)
548 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
549 struct hns3_tx_queue *txq;
552 for (i = 0; i < dev->data->nb_tx_queues; i++) {
553 txq = hw->data->tx_queues[i];
556 hns3_enable_txq(txq, false);
561 hns3_tqp_enable(struct hns3_hw *hw, uint16_t queue_id, bool enable)
563 struct hns3_cfg_com_tqp_queue_cmd *req;
564 struct hns3_cmd_desc desc;
567 req = (struct hns3_cfg_com_tqp_queue_cmd *)desc.data;
569 hns3_cmd_setup_basic_desc(&desc, HNS3_OPC_CFG_COM_TQP_QUEUE, false);
570 req->tqp_id = rte_cpu_to_le_16(queue_id);
572 hns3_set_bit(req->enable, HNS3_TQP_ENABLE_B, enable ? 1 : 0);
574 ret = hns3_cmd_send(hw, &desc, 1);
576 hns3_err(hw, "TQP enable fail, ret = %d", ret);
582 hns3_send_reset_tqp_cmd(struct hns3_hw *hw, uint16_t queue_id, bool enable)
584 struct hns3_reset_tqp_queue_cmd *req;
585 struct hns3_cmd_desc desc;
588 hns3_cmd_setup_basic_desc(&desc, HNS3_OPC_RESET_TQP_QUEUE, false);
590 req = (struct hns3_reset_tqp_queue_cmd *)desc.data;
591 req->tqp_id = rte_cpu_to_le_16(queue_id);
592 hns3_set_bit(req->reset_req, HNS3_TQP_RESET_B, enable ? 1 : 0);
593 ret = hns3_cmd_send(hw, &desc, 1);
595 hns3_err(hw, "send tqp reset cmd error, queue_id = %u, "
596 "ret = %d", queue_id, ret);
602 hns3_get_tqp_reset_status(struct hns3_hw *hw, uint16_t queue_id,
603 uint8_t *reset_status)
605 struct hns3_reset_tqp_queue_cmd *req;
606 struct hns3_cmd_desc desc;
609 hns3_cmd_setup_basic_desc(&desc, HNS3_OPC_RESET_TQP_QUEUE, true);
611 req = (struct hns3_reset_tqp_queue_cmd *)desc.data;
612 req->tqp_id = rte_cpu_to_le_16(queue_id);
614 ret = hns3_cmd_send(hw, &desc, 1);
616 hns3_err(hw, "get tqp reset status error, queue_id = %u, "
617 "ret = %d.", queue_id, ret);
620 *reset_status = hns3_get_bit(req->ready_to_reset, HNS3_TQP_RESET_B);
625 hns3pf_reset_tqp(struct hns3_hw *hw, uint16_t queue_id)
627 #define HNS3_TQP_RESET_TRY_MS 200
628 uint8_t reset_status;
632 ret = hns3_tqp_enable(hw, queue_id, false);
637 * In current version VF is not supported when PF is driven by DPDK
638 * driver, all task queue pairs are mapped to PF function, so PF's queue
639 * id is equals to the global queue id in PF range.
641 ret = hns3_send_reset_tqp_cmd(hw, queue_id, true);
643 hns3_err(hw, "Send reset tqp cmd fail, ret = %d", ret);
646 end = get_timeofday_ms() + HNS3_TQP_RESET_TRY_MS;
648 /* Wait for tqp hw reset */
649 rte_delay_ms(HNS3_POLL_RESPONE_MS);
650 ret = hns3_get_tqp_reset_status(hw, queue_id, &reset_status);
656 } while (get_timeofday_ms() < end);
660 hns3_err(hw, "reset tqp timeout, queue_id = %u, ret = %d",
665 ret = hns3_send_reset_tqp_cmd(hw, queue_id, false);
667 hns3_err(hw, "Deassert the soft reset fail, ret = %d", ret);
672 hns3_send_reset_tqp_cmd(hw, queue_id, false);
677 hns3vf_reset_tqp(struct hns3_hw *hw, uint16_t queue_id)
682 /* Disable VF's queue before send queue reset msg to PF */
683 ret = hns3_tqp_enable(hw, queue_id, false);
687 memcpy(msg_data, &queue_id, sizeof(uint16_t));
689 ret = hns3_send_mbx_msg(hw, HNS3_MBX_QUEUE_RESET, 0, msg_data,
690 sizeof(msg_data), true, NULL, 0);
692 hns3_err(hw, "fail to reset tqp, queue_id = %u, ret = %d.",
698 hns3_reset_tqp(struct hns3_adapter *hns, uint16_t queue_id)
700 struct hns3_hw *hw = &hns->hw;
703 return hns3vf_reset_tqp(hw, queue_id);
705 return hns3pf_reset_tqp(hw, queue_id);
709 hns3_reset_all_tqps(struct hns3_adapter *hns)
711 struct hns3_hw *hw = &hns->hw;
714 for (i = 0; i < hw->cfg_max_queues; i++) {
715 ret = hns3_reset_tqp(hns, i);
717 hns3_err(hw, "Failed to reset No.%d queue: %d", i, ret);
725 hns3_send_reset_queue_cmd(struct hns3_hw *hw, uint16_t queue_id,
726 enum hns3_ring_type queue_type, bool enable)
728 struct hns3_reset_tqp_queue_cmd *req;
729 struct hns3_cmd_desc desc;
733 hns3_cmd_setup_basic_desc(&desc, HNS3_OPC_RESET_TQP_QUEUE_INDEP, false);
735 req = (struct hns3_reset_tqp_queue_cmd *)desc.data;
736 req->tqp_id = rte_cpu_to_le_16(queue_id);
737 queue_direction = queue_type == HNS3_RING_TYPE_TX ? 0 : 1;
738 req->queue_direction = rte_cpu_to_le_16(queue_direction);
739 hns3_set_bit(req->reset_req, HNS3_TQP_RESET_B, enable ? 1 : 0);
741 ret = hns3_cmd_send(hw, &desc, 1);
743 hns3_err(hw, "send queue reset cmd error, queue_id = %u, "
744 "queue_type = %s, ret = %d.", queue_id,
745 queue_type == HNS3_RING_TYPE_TX ? "Tx" : "Rx", ret);
750 hns3_get_queue_reset_status(struct hns3_hw *hw, uint16_t queue_id,
751 enum hns3_ring_type queue_type,
752 uint8_t *reset_status)
754 struct hns3_reset_tqp_queue_cmd *req;
755 struct hns3_cmd_desc desc;
759 hns3_cmd_setup_basic_desc(&desc, HNS3_OPC_RESET_TQP_QUEUE_INDEP, true);
761 req = (struct hns3_reset_tqp_queue_cmd *)desc.data;
762 req->tqp_id = rte_cpu_to_le_16(queue_id);
763 queue_direction = queue_type == HNS3_RING_TYPE_TX ? 0 : 1;
764 req->queue_direction = rte_cpu_to_le_16(queue_direction);
766 ret = hns3_cmd_send(hw, &desc, 1);
768 hns3_err(hw, "get queue reset status error, queue_id = %u "
769 "queue_type = %s, ret = %d.", queue_id,
770 queue_type == HNS3_RING_TYPE_TX ? "Tx" : "Rx", ret);
774 *reset_status = hns3_get_bit(req->ready_to_reset, HNS3_TQP_RESET_B);
779 hns3_reset_queue(struct hns3_hw *hw, uint16_t queue_id,
780 enum hns3_ring_type queue_type)
782 #define HNS3_QUEUE_RESET_TRY_MS 200
783 struct hns3_tx_queue *txq;
784 struct hns3_rx_queue *rxq;
785 uint32_t reset_wait_times;
786 uint32_t max_wait_times;
787 uint8_t reset_status;
790 if (queue_type == HNS3_RING_TYPE_TX) {
791 txq = hw->data->tx_queues[queue_id];
792 hns3_enable_txq(txq, false);
794 rxq = hw->data->rx_queues[queue_id];
795 hns3_enable_rxq(rxq, false);
798 ret = hns3_send_reset_queue_cmd(hw, queue_id, queue_type, true);
800 hns3_err(hw, "send reset queue cmd fail, ret = %d.", ret);
804 reset_wait_times = 0;
805 max_wait_times = HNS3_QUEUE_RESET_TRY_MS / HNS3_POLL_RESPONE_MS;
806 while (reset_wait_times < max_wait_times) {
807 /* Wait for queue hw reset */
808 rte_delay_ms(HNS3_POLL_RESPONE_MS);
809 ret = hns3_get_queue_reset_status(hw, queue_id,
810 queue_type, &reset_status);
812 goto queue_reset_fail;
820 hns3_err(hw, "reset queue timeout, queue_id = %u, "
821 "queue_type = %s", queue_id,
822 queue_type == HNS3_RING_TYPE_TX ? "Tx" : "Rx");
824 goto queue_reset_fail;
827 ret = hns3_send_reset_queue_cmd(hw, queue_id, queue_type, false);
829 hns3_err(hw, "deassert queue reset fail, ret = %d.", ret);
834 hns3_send_reset_queue_cmd(hw, queue_id, queue_type, false);
839 hns3_get_tqp_intr_reg_offset(uint16_t tqp_intr_id)
843 /* Need an extend offset to config queues > 64 */
844 if (tqp_intr_id < HNS3_MIN_EXT_TQP_INTR_ID)
845 reg_offset = HNS3_TQP_INTR_REG_BASE +
846 tqp_intr_id * HNS3_TQP_INTR_LOW_ORDER_OFFSET;
848 reg_offset = HNS3_TQP_INTR_EXT_REG_BASE +
849 tqp_intr_id / HNS3_MIN_EXT_TQP_INTR_ID *
850 HNS3_TQP_INTR_HIGH_ORDER_OFFSET +
851 tqp_intr_id % HNS3_MIN_EXT_TQP_INTR_ID *
852 HNS3_TQP_INTR_LOW_ORDER_OFFSET;
858 hns3_set_queue_intr_gl(struct hns3_hw *hw, uint16_t queue_id,
859 uint8_t gl_idx, uint16_t gl_value)
861 uint32_t offset[] = {HNS3_TQP_INTR_GL0_REG,
862 HNS3_TQP_INTR_GL1_REG,
863 HNS3_TQP_INTR_GL2_REG};
864 uint32_t addr, value;
866 if (gl_idx >= RTE_DIM(offset) || gl_value > HNS3_TQP_INTR_GL_MAX)
869 addr = offset[gl_idx] + hns3_get_tqp_intr_reg_offset(queue_id);
870 if (hw->intr.gl_unit == HNS3_INTR_COALESCE_GL_UINT_1US)
871 value = gl_value | HNS3_TQP_INTR_GL_UNIT_1US;
873 value = HNS3_GL_USEC_TO_REG(gl_value);
875 hns3_write_dev(hw, addr, value);
879 hns3_set_queue_intr_rl(struct hns3_hw *hw, uint16_t queue_id, uint16_t rl_value)
881 uint32_t addr, value;
883 if (rl_value > HNS3_TQP_INTR_RL_MAX)
886 addr = HNS3_TQP_INTR_RL_REG + hns3_get_tqp_intr_reg_offset(queue_id);
887 value = HNS3_RL_USEC_TO_REG(rl_value);
889 value |= HNS3_TQP_INTR_RL_ENABLE_MASK;
891 hns3_write_dev(hw, addr, value);
895 hns3_set_queue_intr_ql(struct hns3_hw *hw, uint16_t queue_id, uint16_t ql_value)
900 * int_ql_max == 0 means the hardware does not support QL,
901 * QL regs config is not permitted if QL is not supported,
904 if (hw->intr.int_ql_max == HNS3_INTR_QL_NONE)
907 addr = HNS3_TQP_INTR_TX_QL_REG + hns3_get_tqp_intr_reg_offset(queue_id);
908 hns3_write_dev(hw, addr, ql_value);
910 addr = HNS3_TQP_INTR_RX_QL_REG + hns3_get_tqp_intr_reg_offset(queue_id);
911 hns3_write_dev(hw, addr, ql_value);
915 hns3_queue_intr_enable(struct hns3_hw *hw, uint16_t queue_id, bool en)
917 uint32_t addr, value;
919 addr = HNS3_TQP_INTR_CTRL_REG + hns3_get_tqp_intr_reg_offset(queue_id);
922 hns3_write_dev(hw, addr, value);
926 * Enable all rx queue interrupt when in interrupt rx mode.
927 * This api was called before enable queue rx&tx (in normal start or reset
928 * recover scenes), used to fix hardware rx queue interrupt enable was clear
932 hns3_dev_all_rx_queue_intr_enable(struct hns3_hw *hw, bool en)
934 struct rte_eth_dev *dev = &rte_eth_devices[hw->data->port_id];
935 uint16_t nb_rx_q = hw->data->nb_rx_queues;
938 if (dev->data->dev_conf.intr_conf.rxq == 0)
941 for (i = 0; i < nb_rx_q; i++)
942 hns3_queue_intr_enable(hw, i, en);
946 hns3_dev_rx_queue_intr_enable(struct rte_eth_dev *dev, uint16_t queue_id)
948 struct rte_pci_device *pci_dev = RTE_ETH_DEV_TO_PCI(dev);
949 struct rte_intr_handle *intr_handle = &pci_dev->intr_handle;
950 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
952 if (dev->data->dev_conf.intr_conf.rxq == 0)
955 hns3_queue_intr_enable(hw, queue_id, true);
957 return rte_intr_ack(intr_handle);
961 hns3_dev_rx_queue_intr_disable(struct rte_eth_dev *dev, uint16_t queue_id)
963 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
965 if (dev->data->dev_conf.intr_conf.rxq == 0)
968 hns3_queue_intr_enable(hw, queue_id, false);
974 hns3_init_rxq(struct hns3_adapter *hns, uint16_t idx)
976 struct hns3_hw *hw = &hns->hw;
977 struct hns3_rx_queue *rxq;
980 PMD_INIT_FUNC_TRACE();
982 rxq = (struct hns3_rx_queue *)hw->data->rx_queues[idx];
983 ret = hns3_alloc_rx_queue_mbufs(hw, rxq);
985 hns3_err(hw, "fail to alloc mbuf for Rx queue %u, ret = %d.",
990 rxq->next_to_use = 0;
991 rxq->rx_rearm_start = 0;
992 rxq->rx_free_hold = 0;
993 rxq->rx_rearm_nb = 0;
994 rxq->pkt_first_seg = NULL;
995 rxq->pkt_last_seg = NULL;
996 hns3_init_rx_queue_hw(rxq);
997 hns3_rxq_vec_setup(rxq);
1003 hns3_init_fake_rxq(struct hns3_adapter *hns, uint16_t idx)
1005 struct hns3_hw *hw = &hns->hw;
1006 struct hns3_rx_queue *rxq;
1008 rxq = (struct hns3_rx_queue *)hw->fkq_data.rx_queues[idx];
1009 rxq->next_to_use = 0;
1010 rxq->rx_free_hold = 0;
1011 rxq->rx_rearm_start = 0;
1012 rxq->rx_rearm_nb = 0;
1013 hns3_init_rx_queue_hw(rxq);
1017 hns3_init_txq(struct hns3_tx_queue *txq)
1019 struct hns3_desc *desc;
1023 desc = txq->tx_ring;
1024 for (i = 0; i < txq->nb_tx_desc; i++) {
1025 desc->tx.tp_fe_sc_vld_ra_ri = 0;
1029 txq->next_to_use = 0;
1030 txq->next_to_clean = 0;
1031 txq->tx_bd_ready = txq->nb_tx_desc - 1;
1032 hns3_init_tx_queue_hw(txq);
1036 hns3_init_tx_ring_tc(struct hns3_adapter *hns)
1038 struct hns3_hw *hw = &hns->hw;
1039 struct hns3_tx_queue *txq;
1042 for (i = 0; i < HNS3_MAX_TC_NUM; i++) {
1043 struct hns3_tc_queue_info *tc_queue = &hw->tc_queue[i];
1046 if (!tc_queue->enable)
1049 for (j = 0; j < tc_queue->tqp_count; j++) {
1050 num = tc_queue->tqp_offset + j;
1051 txq = (struct hns3_tx_queue *)hw->data->tx_queues[num];
1055 hns3_write_dev(txq, HNS3_RING_TX_TC_REG, tc_queue->tc);
1061 hns3_init_rx_queues(struct hns3_adapter *hns)
1063 struct hns3_hw *hw = &hns->hw;
1064 struct hns3_rx_queue *rxq;
1068 /* Initialize RSS for queues */
1069 ret = hns3_config_rss(hns);
1071 hns3_err(hw, "failed to configure rss, ret = %d.", ret);
1075 for (i = 0; i < hw->data->nb_rx_queues; i++) {
1076 rxq = (struct hns3_rx_queue *)hw->data->rx_queues[i];
1078 hns3_err(hw, "Rx queue %u not available or setup.", i);
1082 if (rxq->rx_deferred_start)
1085 ret = hns3_init_rxq(hns, i);
1087 hns3_err(hw, "failed to init Rx queue %u, ret = %d.", i,
1093 for (i = 0; i < hw->fkq_data.nb_fake_rx_queues; i++)
1094 hns3_init_fake_rxq(hns, i);
1099 for (j = 0; j < i; j++) {
1100 rxq = (struct hns3_rx_queue *)hw->data->rx_queues[j];
1101 hns3_rx_queue_release_mbufs(rxq);
1108 hns3_init_tx_queues(struct hns3_adapter *hns)
1110 struct hns3_hw *hw = &hns->hw;
1111 struct hns3_tx_queue *txq;
1114 for (i = 0; i < hw->data->nb_tx_queues; i++) {
1115 txq = (struct hns3_tx_queue *)hw->data->tx_queues[i];
1117 hns3_err(hw, "Tx queue %u not available or setup.", i);
1121 if (txq->tx_deferred_start)
1126 for (i = 0; i < hw->fkq_data.nb_fake_tx_queues; i++) {
1127 txq = (struct hns3_tx_queue *)hw->fkq_data.tx_queues[i];
1130 hns3_init_tx_ring_tc(hns);
1137 * Note: just init and setup queues, and don't enable tqps.
1140 hns3_init_queues(struct hns3_adapter *hns, bool reset_queue)
1142 struct hns3_hw *hw = &hns->hw;
1146 ret = hns3_reset_all_tqps(hns);
1148 hns3_err(hw, "failed to reset all queues, ret = %d.",
1154 ret = hns3_init_rx_queues(hns);
1156 hns3_err(hw, "failed to init rx queues, ret = %d.", ret);
1160 ret = hns3_init_tx_queues(hns);
1162 hns3_dev_release_mbufs(hns);
1163 hns3_err(hw, "failed to init tx queues, ret = %d.", ret);
1170 hns3_start_tqps(struct hns3_hw *hw)
1172 struct hns3_tx_queue *txq;
1173 struct hns3_rx_queue *rxq;
1176 hns3_enable_all_queues(hw, true);
1178 for (i = 0; i < hw->data->nb_tx_queues; i++) {
1179 txq = hw->data->tx_queues[i];
1181 hw->data->tx_queue_state[i] =
1182 RTE_ETH_QUEUE_STATE_STARTED;
1185 for (i = 0; i < hw->data->nb_rx_queues; i++) {
1186 rxq = hw->data->rx_queues[i];
1188 hw->data->rx_queue_state[i] =
1189 RTE_ETH_QUEUE_STATE_STARTED;
1194 hns3_stop_tqps(struct hns3_hw *hw)
1198 hns3_enable_all_queues(hw, false);
1200 for (i = 0; i < hw->data->nb_tx_queues; i++)
1201 hw->data->tx_queue_state[i] = RTE_ETH_QUEUE_STATE_STOPPED;
1203 for (i = 0; i < hw->data->nb_rx_queues; i++)
1204 hw->data->rx_queue_state[i] = RTE_ETH_QUEUE_STATE_STOPPED;
1208 * Iterate over all Rx Queue, and call the callback() function for each Rx
1212 * The target eth dev.
1213 * @param[in] callback
1214 * The function to call for each queue.
1215 * if callback function return nonzero will stop iterate and return it's value
1217 * The arguments to provide the callback function with.
1220 * 0 on success, otherwise with errno set.
1223 hns3_rxq_iterate(struct rte_eth_dev *dev,
1224 int (*callback)(struct hns3_rx_queue *, void *), void *arg)
1229 if (dev->data->rx_queues == NULL)
1232 for (i = 0; i < dev->data->nb_rx_queues; i++) {
1233 ret = callback(dev->data->rx_queues[i], arg);
1242 hns3_alloc_rxq_and_dma_zone(struct rte_eth_dev *dev,
1243 struct hns3_queue_info *q_info)
1245 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1246 const struct rte_memzone *rx_mz;
1247 struct hns3_rx_queue *rxq;
1248 unsigned int rx_desc;
1250 rxq = rte_zmalloc_socket(q_info->type, sizeof(struct hns3_rx_queue),
1251 RTE_CACHE_LINE_SIZE, q_info->socket_id);
1253 hns3_err(hw, "Failed to allocate memory for No.%u rx ring!",
1258 /* Allocate rx ring hardware descriptors. */
1259 rxq->queue_id = q_info->idx;
1260 rxq->nb_rx_desc = q_info->nb_desc;
1263 * Allocate a litter more memory because rx vector functions
1264 * don't check boundaries each time.
1266 rx_desc = (rxq->nb_rx_desc + HNS3_DEFAULT_RX_BURST) *
1267 sizeof(struct hns3_desc);
1268 rx_mz = rte_eth_dma_zone_reserve(dev, q_info->ring_name, q_info->idx,
1269 rx_desc, HNS3_RING_BASE_ALIGN,
1271 if (rx_mz == NULL) {
1272 hns3_err(hw, "Failed to reserve DMA memory for No.%u rx ring!",
1274 hns3_rx_queue_release(rxq);
1278 rxq->rx_ring = (struct hns3_desc *)rx_mz->addr;
1279 rxq->rx_ring_phys_addr = rx_mz->iova;
1281 hns3_dbg(hw, "No.%u rx descriptors iova 0x%" PRIx64, q_info->idx,
1282 rxq->rx_ring_phys_addr);
1288 hns3_fake_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx,
1289 uint16_t nb_desc, unsigned int socket_id)
1291 struct hns3_adapter *hns = dev->data->dev_private;
1292 struct hns3_hw *hw = &hns->hw;
1293 struct hns3_queue_info q_info;
1294 struct hns3_rx_queue *rxq;
1297 if (hw->fkq_data.rx_queues[idx]) {
1298 hns3_rx_queue_release(hw->fkq_data.rx_queues[idx]);
1299 hw->fkq_data.rx_queues[idx] = NULL;
1303 q_info.socket_id = socket_id;
1304 q_info.nb_desc = nb_desc;
1305 q_info.type = "hns3 fake RX queue";
1306 q_info.ring_name = "rx_fake_ring";
1307 rxq = hns3_alloc_rxq_and_dma_zone(dev, &q_info);
1309 hns3_err(hw, "Failed to setup No.%u fake rx ring.", idx);
1313 /* Don't need alloc sw_ring, because upper applications don't use it */
1314 rxq->sw_ring = NULL;
1317 rxq->rx_deferred_start = false;
1318 rxq->port_id = dev->data->port_id;
1319 rxq->configured = true;
1320 nb_rx_q = dev->data->nb_rx_queues;
1321 rxq->io_base = (void *)((char *)hw->io_base + HNS3_TQP_REG_OFFSET +
1322 (nb_rx_q + idx) * HNS3_TQP_REG_SIZE);
1323 rxq->rx_buf_len = HNS3_MIN_BD_BUF_SIZE;
1325 rte_spinlock_lock(&hw->lock);
1326 hw->fkq_data.rx_queues[idx] = rxq;
1327 rte_spinlock_unlock(&hw->lock);
1333 hns3_alloc_txq_and_dma_zone(struct rte_eth_dev *dev,
1334 struct hns3_queue_info *q_info)
1336 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1337 const struct rte_memzone *tx_mz;
1338 struct hns3_tx_queue *txq;
1339 struct hns3_desc *desc;
1340 unsigned int tx_desc;
1343 txq = rte_zmalloc_socket(q_info->type, sizeof(struct hns3_tx_queue),
1344 RTE_CACHE_LINE_SIZE, q_info->socket_id);
1346 hns3_err(hw, "Failed to allocate memory for No.%u tx ring!",
1351 /* Allocate tx ring hardware descriptors. */
1352 txq->queue_id = q_info->idx;
1353 txq->nb_tx_desc = q_info->nb_desc;
1354 tx_desc = txq->nb_tx_desc * sizeof(struct hns3_desc);
1355 tx_mz = rte_eth_dma_zone_reserve(dev, q_info->ring_name, q_info->idx,
1356 tx_desc, HNS3_RING_BASE_ALIGN,
1358 if (tx_mz == NULL) {
1359 hns3_err(hw, "Failed to reserve DMA memory for No.%u tx ring!",
1361 hns3_tx_queue_release(txq);
1365 txq->tx_ring = (struct hns3_desc *)tx_mz->addr;
1366 txq->tx_ring_phys_addr = tx_mz->iova;
1368 hns3_dbg(hw, "No.%u tx descriptors iova 0x%" PRIx64, q_info->idx,
1369 txq->tx_ring_phys_addr);
1372 desc = txq->tx_ring;
1373 for (i = 0; i < txq->nb_tx_desc; i++) {
1374 desc->tx.tp_fe_sc_vld_ra_ri = 0;
1382 hns3_fake_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx,
1383 uint16_t nb_desc, unsigned int socket_id)
1385 struct hns3_adapter *hns = dev->data->dev_private;
1386 struct hns3_hw *hw = &hns->hw;
1387 struct hns3_queue_info q_info;
1388 struct hns3_tx_queue *txq;
1391 if (hw->fkq_data.tx_queues[idx] != NULL) {
1392 hns3_tx_queue_release(hw->fkq_data.tx_queues[idx]);
1393 hw->fkq_data.tx_queues[idx] = NULL;
1397 q_info.socket_id = socket_id;
1398 q_info.nb_desc = nb_desc;
1399 q_info.type = "hns3 fake TX queue";
1400 q_info.ring_name = "tx_fake_ring";
1401 txq = hns3_alloc_txq_and_dma_zone(dev, &q_info);
1403 hns3_err(hw, "Failed to setup No.%u fake tx ring.", idx);
1407 /* Don't need alloc sw_ring, because upper applications don't use it */
1408 txq->sw_ring = NULL;
1412 txq->tx_deferred_start = false;
1413 txq->port_id = dev->data->port_id;
1414 txq->configured = true;
1415 nb_tx_q = dev->data->nb_tx_queues;
1416 txq->io_base = (void *)((char *)hw->io_base + HNS3_TQP_REG_OFFSET +
1417 (nb_tx_q + idx) * HNS3_TQP_REG_SIZE);
1419 rte_spinlock_lock(&hw->lock);
1420 hw->fkq_data.tx_queues[idx] = txq;
1421 rte_spinlock_unlock(&hw->lock);
1427 hns3_fake_rx_queue_config(struct hns3_hw *hw, uint16_t nb_queues)
1429 uint16_t old_nb_queues = hw->fkq_data.nb_fake_rx_queues;
1433 if (hw->fkq_data.rx_queues == NULL && nb_queues != 0) {
1434 /* first time configuration */
1436 size = sizeof(hw->fkq_data.rx_queues[0]) * nb_queues;
1437 hw->fkq_data.rx_queues = rte_zmalloc("fake_rx_queues", size,
1438 RTE_CACHE_LINE_SIZE);
1439 if (hw->fkq_data.rx_queues == NULL) {
1440 hw->fkq_data.nb_fake_rx_queues = 0;
1443 } else if (hw->fkq_data.rx_queues != NULL && nb_queues != 0) {
1445 rxq = hw->fkq_data.rx_queues;
1446 for (i = nb_queues; i < old_nb_queues; i++)
1447 hns3_dev_rx_queue_release(rxq[i]);
1449 rxq = rte_realloc(rxq, sizeof(rxq[0]) * nb_queues,
1450 RTE_CACHE_LINE_SIZE);
1453 if (nb_queues > old_nb_queues) {
1454 uint16_t new_qs = nb_queues - old_nb_queues;
1455 memset(rxq + old_nb_queues, 0, sizeof(rxq[0]) * new_qs);
1458 hw->fkq_data.rx_queues = rxq;
1459 } else if (hw->fkq_data.rx_queues != NULL && nb_queues == 0) {
1460 rxq = hw->fkq_data.rx_queues;
1461 for (i = nb_queues; i < old_nb_queues; i++)
1462 hns3_dev_rx_queue_release(rxq[i]);
1464 rte_free(hw->fkq_data.rx_queues);
1465 hw->fkq_data.rx_queues = NULL;
1468 hw->fkq_data.nb_fake_rx_queues = nb_queues;
1474 hns3_fake_tx_queue_config(struct hns3_hw *hw, uint16_t nb_queues)
1476 uint16_t old_nb_queues = hw->fkq_data.nb_fake_tx_queues;
1480 if (hw->fkq_data.tx_queues == NULL && nb_queues != 0) {
1481 /* first time configuration */
1483 size = sizeof(hw->fkq_data.tx_queues[0]) * nb_queues;
1484 hw->fkq_data.tx_queues = rte_zmalloc("fake_tx_queues", size,
1485 RTE_CACHE_LINE_SIZE);
1486 if (hw->fkq_data.tx_queues == NULL) {
1487 hw->fkq_data.nb_fake_tx_queues = 0;
1490 } else if (hw->fkq_data.tx_queues != NULL && nb_queues != 0) {
1492 txq = hw->fkq_data.tx_queues;
1493 for (i = nb_queues; i < old_nb_queues; i++)
1494 hns3_dev_tx_queue_release(txq[i]);
1495 txq = rte_realloc(txq, sizeof(txq[0]) * nb_queues,
1496 RTE_CACHE_LINE_SIZE);
1499 if (nb_queues > old_nb_queues) {
1500 uint16_t new_qs = nb_queues - old_nb_queues;
1501 memset(txq + old_nb_queues, 0, sizeof(txq[0]) * new_qs);
1504 hw->fkq_data.tx_queues = txq;
1505 } else if (hw->fkq_data.tx_queues != NULL && nb_queues == 0) {
1506 txq = hw->fkq_data.tx_queues;
1507 for (i = nb_queues; i < old_nb_queues; i++)
1508 hns3_dev_tx_queue_release(txq[i]);
1510 rte_free(hw->fkq_data.tx_queues);
1511 hw->fkq_data.tx_queues = NULL;
1513 hw->fkq_data.nb_fake_tx_queues = nb_queues;
1519 hns3_set_fake_rx_or_tx_queues(struct rte_eth_dev *dev, uint16_t nb_rx_q,
1522 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1523 uint16_t rx_need_add_nb_q;
1524 uint16_t tx_need_add_nb_q;
1529 /* Setup new number of fake RX/TX queues and reconfigure device. */
1530 rx_need_add_nb_q = hw->cfg_max_queues - nb_rx_q;
1531 tx_need_add_nb_q = hw->cfg_max_queues - nb_tx_q;
1532 ret = hns3_fake_rx_queue_config(hw, rx_need_add_nb_q);
1534 hns3_err(hw, "Fail to configure fake rx queues: %d", ret);
1538 ret = hns3_fake_tx_queue_config(hw, tx_need_add_nb_q);
1540 hns3_err(hw, "Fail to configure fake rx queues: %d", ret);
1541 goto cfg_fake_tx_q_fail;
1544 /* Allocate and set up fake RX queue per Ethernet port. */
1545 port_id = hw->data->port_id;
1546 for (q = 0; q < rx_need_add_nb_q; q++) {
1547 ret = hns3_fake_rx_queue_setup(dev, q, HNS3_MIN_RING_DESC,
1548 rte_eth_dev_socket_id(port_id));
1550 goto setup_fake_rx_q_fail;
1553 /* Allocate and set up fake TX queue per Ethernet port. */
1554 for (q = 0; q < tx_need_add_nb_q; q++) {
1555 ret = hns3_fake_tx_queue_setup(dev, q, HNS3_MIN_RING_DESC,
1556 rte_eth_dev_socket_id(port_id));
1558 goto setup_fake_tx_q_fail;
1563 setup_fake_tx_q_fail:
1564 setup_fake_rx_q_fail:
1565 (void)hns3_fake_tx_queue_config(hw, 0);
1567 (void)hns3_fake_rx_queue_config(hw, 0);
1573 hns3_dev_release_mbufs(struct hns3_adapter *hns)
1575 struct rte_eth_dev_data *dev_data = hns->hw.data;
1576 struct hns3_rx_queue *rxq;
1577 struct hns3_tx_queue *txq;
1580 if (dev_data->rx_queues)
1581 for (i = 0; i < dev_data->nb_rx_queues; i++) {
1582 rxq = dev_data->rx_queues[i];
1585 hns3_rx_queue_release_mbufs(rxq);
1588 if (dev_data->tx_queues)
1589 for (i = 0; i < dev_data->nb_tx_queues; i++) {
1590 txq = dev_data->tx_queues[i];
1593 hns3_tx_queue_release_mbufs(txq);
1598 hns3_rx_buf_len_calc(struct rte_mempool *mp, uint16_t *rx_buf_len)
1600 uint16_t vld_buf_size;
1601 uint16_t num_hw_specs;
1605 * hns3 network engine only support to set 4 typical specification, and
1606 * different buffer size will affect the max packet_len and the max
1607 * number of segmentation when hw gro is turned on in receive side. The
1608 * relationship between them is as follows:
1609 * rx_buf_size | max_gro_pkt_len | max_gro_nb_seg
1610 * ---------------------|-------------------|----------------
1611 * HNS3_4K_BD_BUF_SIZE | 60KB | 15
1612 * HNS3_2K_BD_BUF_SIZE | 62KB | 31
1613 * HNS3_1K_BD_BUF_SIZE | 63KB | 63
1614 * HNS3_512_BD_BUF_SIZE | 31.5KB | 63
1616 static const uint16_t hw_rx_buf_size[] = {
1617 HNS3_4K_BD_BUF_SIZE,
1618 HNS3_2K_BD_BUF_SIZE,
1619 HNS3_1K_BD_BUF_SIZE,
1620 HNS3_512_BD_BUF_SIZE
1623 vld_buf_size = (uint16_t)(rte_pktmbuf_data_room_size(mp) -
1624 RTE_PKTMBUF_HEADROOM);
1625 if (vld_buf_size < HNS3_MIN_BD_BUF_SIZE)
1628 num_hw_specs = RTE_DIM(hw_rx_buf_size);
1629 for (i = 0; i < num_hw_specs; i++) {
1630 if (vld_buf_size >= hw_rx_buf_size[i]) {
1631 *rx_buf_len = hw_rx_buf_size[i];
1639 hns3_rxq_conf_runtime_check(struct hns3_hw *hw, uint16_t buf_size,
1642 struct rte_eth_dev *dev = &rte_eth_devices[hw->data->port_id];
1643 struct rte_eth_rxmode *rxmode = &hw->data->dev_conf.rxmode;
1644 eth_rx_burst_t pkt_burst = dev->rx_pkt_burst;
1645 uint16_t min_vec_bds;
1648 * HNS3 hardware network engine set scattered as default. If the driver
1649 * is not work in scattered mode and the pkts greater than buf_size
1650 * but smaller than max_rx_pkt_len will be distributed to multiple BDs.
1651 * Driver cannot handle this situation.
1653 if (!hw->data->scattered_rx && rxmode->max_rx_pkt_len > buf_size) {
1654 hns3_err(hw, "max_rx_pkt_len is not allowed to be set greater "
1655 "than rx_buf_len if scattered is off.");
1659 if (pkt_burst == hns3_recv_pkts_vec) {
1660 min_vec_bds = HNS3_DEFAULT_RXQ_REARM_THRESH +
1661 HNS3_DEFAULT_RX_BURST;
1662 if (nb_desc < min_vec_bds ||
1663 nb_desc % HNS3_DEFAULT_RXQ_REARM_THRESH) {
1664 hns3_err(hw, "if Rx burst mode is vector, "
1665 "number of descriptor is required to be "
1666 "bigger than min vector bds:%u, and could be "
1667 "divided by rxq rearm thresh:%u.",
1668 min_vec_bds, HNS3_DEFAULT_RXQ_REARM_THRESH);
1676 hns3_rx_queue_conf_check(struct hns3_hw *hw, const struct rte_eth_rxconf *conf,
1677 struct rte_mempool *mp, uint16_t nb_desc,
1682 if (nb_desc > HNS3_MAX_RING_DESC || nb_desc < HNS3_MIN_RING_DESC ||
1683 nb_desc % HNS3_ALIGN_RING_DESC) {
1684 hns3_err(hw, "Number (%u) of rx descriptors is invalid",
1689 if (conf->rx_drop_en == 0)
1690 hns3_warn(hw, "if no descriptors available, packets are always "
1691 "dropped and rx_drop_en (1) is fixed on");
1693 if (hns3_rx_buf_len_calc(mp, buf_size)) {
1694 hns3_err(hw, "rxq mbufs' data room size (%u) is not enough! "
1695 "minimal data room size (%u).",
1696 rte_pktmbuf_data_room_size(mp),
1697 HNS3_MIN_BD_BUF_SIZE + RTE_PKTMBUF_HEADROOM);
1701 if (hw->data->dev_started) {
1702 ret = hns3_rxq_conf_runtime_check(hw, *buf_size, nb_desc);
1704 hns3_err(hw, "Rx queue runtime setup fail.");
1713 hns3_get_tqp_reg_offset(uint16_t queue_id)
1715 uint32_t reg_offset;
1717 /* Need an extend offset to config queue > 1024 */
1718 if (queue_id < HNS3_MIN_EXTEND_QUEUE_ID)
1719 reg_offset = HNS3_TQP_REG_OFFSET + queue_id * HNS3_TQP_REG_SIZE;
1721 reg_offset = HNS3_TQP_REG_OFFSET + HNS3_TQP_EXT_REG_OFFSET +
1722 (queue_id - HNS3_MIN_EXTEND_QUEUE_ID) *
1729 hns3_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t nb_desc,
1730 unsigned int socket_id, const struct rte_eth_rxconf *conf,
1731 struct rte_mempool *mp)
1733 struct hns3_adapter *hns = dev->data->dev_private;
1734 struct hns3_hw *hw = &hns->hw;
1735 struct hns3_queue_info q_info;
1736 struct hns3_rx_queue *rxq;
1737 uint16_t rx_buf_size;
1741 ret = hns3_rx_queue_conf_check(hw, conf, mp, nb_desc, &rx_buf_size);
1745 if (dev->data->rx_queues[idx]) {
1746 hns3_rx_queue_release(dev->data->rx_queues[idx]);
1747 dev->data->rx_queues[idx] = NULL;
1751 q_info.socket_id = socket_id;
1752 q_info.nb_desc = nb_desc;
1753 q_info.type = "hns3 RX queue";
1754 q_info.ring_name = "rx_ring";
1756 rxq = hns3_alloc_rxq_and_dma_zone(dev, &q_info);
1759 "Failed to alloc mem and reserve DMA mem for rx ring!");
1764 rxq->ptype_tbl = &hns->ptype_tbl;
1766 rxq->rx_free_thresh = (conf->rx_free_thresh > 0) ?
1767 conf->rx_free_thresh : HNS3_DEFAULT_RX_FREE_THRESH;
1769 rxq->rx_deferred_start = conf->rx_deferred_start;
1770 if (rxq->rx_deferred_start && !hns3_dev_indep_txrx_supported(hw)) {
1771 hns3_warn(hw, "deferred start is not supported.");
1772 rxq->rx_deferred_start = false;
1775 rx_entry_len = (rxq->nb_rx_desc + HNS3_DEFAULT_RX_BURST) *
1776 sizeof(struct hns3_entry);
1777 rxq->sw_ring = rte_zmalloc_socket("hns3 RX sw ring", rx_entry_len,
1778 RTE_CACHE_LINE_SIZE, socket_id);
1779 if (rxq->sw_ring == NULL) {
1780 hns3_err(hw, "Failed to allocate memory for rx sw ring!");
1781 hns3_rx_queue_release(rxq);
1785 rxq->next_to_use = 0;
1786 rxq->rx_free_hold = 0;
1787 rxq->rx_rearm_start = 0;
1788 rxq->rx_rearm_nb = 0;
1789 rxq->pkt_first_seg = NULL;
1790 rxq->pkt_last_seg = NULL;
1791 rxq->port_id = dev->data->port_id;
1793 * For hns3 PF device, if the VLAN mode is HW_SHIFT_AND_DISCARD_MODE,
1794 * the pvid_sw_discard_en in the queue struct should not be changed,
1795 * because PVID-related operations do not need to be processed by PMD
1796 * driver. For hns3 VF device, whether it needs to process PVID depends
1797 * on the configuration of PF kernel mode netdevice driver. And the
1798 * related PF configuration is delivered through the mailbox and finally
1799 * reflectd in port_base_vlan_cfg.
1801 if (hns->is_vf || hw->vlan_mode == HNS3_SW_SHIFT_AND_DISCARD_MODE)
1802 rxq->pvid_sw_discard_en = hw->port_base_vlan_cfg.state ==
1803 HNS3_PORT_BASE_VLAN_ENABLE;
1805 rxq->pvid_sw_discard_en = false;
1806 rxq->ptype_en = hns3_dev_rxd_adv_layout_supported(hw) ? true : false;
1807 rxq->configured = true;
1808 rxq->io_base = (void *)((char *)hw->io_base + HNS3_TQP_REG_OFFSET +
1809 idx * HNS3_TQP_REG_SIZE);
1810 rxq->io_base = (void *)((char *)hw->io_base +
1811 hns3_get_tqp_reg_offset(idx));
1812 rxq->io_head_reg = (volatile void *)((char *)rxq->io_base +
1813 HNS3_RING_RX_HEAD_REG);
1814 rxq->rx_buf_len = rx_buf_size;
1815 memset(&rxq->basic_stats, 0, sizeof(struct hns3_rx_basic_stats));
1816 memset(&rxq->err_stats, 0, sizeof(struct hns3_rx_bd_errors_stats));
1817 memset(&rxq->dfx_stats, 0, sizeof(struct hns3_rx_dfx_stats));
1819 /* CRC len set here is used for amending packet length */
1820 if (dev->data->dev_conf.rxmode.offloads & DEV_RX_OFFLOAD_KEEP_CRC)
1821 rxq->crc_len = RTE_ETHER_CRC_LEN;
1825 rxq->bulk_mbuf_num = 0;
1827 rte_spinlock_lock(&hw->lock);
1828 dev->data->rx_queues[idx] = rxq;
1829 rte_spinlock_unlock(&hw->lock);
1835 hns3_rx_scattered_reset(struct rte_eth_dev *dev)
1837 struct hns3_adapter *hns = dev->data->dev_private;
1838 struct hns3_hw *hw = &hns->hw;
1841 dev->data->scattered_rx = false;
1845 hns3_rx_scattered_calc(struct rte_eth_dev *dev)
1847 struct rte_eth_conf *dev_conf = &dev->data->dev_conf;
1848 struct hns3_adapter *hns = dev->data->dev_private;
1849 struct hns3_hw *hw = &hns->hw;
1850 struct hns3_rx_queue *rxq;
1853 if (dev->data->rx_queues == NULL)
1856 for (queue_id = 0; queue_id < dev->data->nb_rx_queues; queue_id++) {
1857 rxq = dev->data->rx_queues[queue_id];
1858 if (hw->rx_buf_len == 0)
1859 hw->rx_buf_len = rxq->rx_buf_len;
1861 hw->rx_buf_len = RTE_MIN(hw->rx_buf_len,
1865 if (dev_conf->rxmode.offloads & DEV_RX_OFFLOAD_SCATTER ||
1866 dev_conf->rxmode.max_rx_pkt_len > hw->rx_buf_len)
1867 dev->data->scattered_rx = true;
1871 hns3_dev_supported_ptypes_get(struct rte_eth_dev *dev)
1873 static const uint32_t ptypes[] = {
1875 RTE_PTYPE_L2_ETHER_VLAN,
1876 RTE_PTYPE_L2_ETHER_QINQ,
1877 RTE_PTYPE_L2_ETHER_LLDP,
1878 RTE_PTYPE_L2_ETHER_ARP,
1880 RTE_PTYPE_L3_IPV4_EXT,
1882 RTE_PTYPE_L3_IPV6_EXT,
1888 RTE_PTYPE_TUNNEL_GRE,
1889 RTE_PTYPE_INNER_L2_ETHER,
1890 RTE_PTYPE_INNER_L2_ETHER_VLAN,
1891 RTE_PTYPE_INNER_L2_ETHER_QINQ,
1892 RTE_PTYPE_INNER_L3_IPV4,
1893 RTE_PTYPE_INNER_L3_IPV6,
1894 RTE_PTYPE_INNER_L3_IPV4_EXT,
1895 RTE_PTYPE_INNER_L3_IPV6_EXT,
1896 RTE_PTYPE_INNER_L4_UDP,
1897 RTE_PTYPE_INNER_L4_TCP,
1898 RTE_PTYPE_INNER_L4_SCTP,
1899 RTE_PTYPE_INNER_L4_ICMP,
1900 RTE_PTYPE_TUNNEL_VXLAN,
1901 RTE_PTYPE_TUNNEL_NVGRE,
1905 if (dev->rx_pkt_burst == hns3_recv_pkts ||
1906 dev->rx_pkt_burst == hns3_recv_scattered_pkts ||
1907 dev->rx_pkt_burst == hns3_recv_pkts_vec ||
1908 dev->rx_pkt_burst == hns3_recv_pkts_vec_sve)
1915 hns3_init_non_tunnel_ptype_tbl(struct hns3_ptype_table *tbl)
1917 tbl->l2l3table[0][0] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4;
1918 tbl->l2l3table[0][1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6;
1919 tbl->l2l3table[0][2] = RTE_PTYPE_L2_ETHER_ARP;
1920 tbl->l2l3table[0][3] = RTE_PTYPE_L2_ETHER;
1921 tbl->l2l3table[0][4] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT;
1922 tbl->l2l3table[0][5] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT;
1923 tbl->l2l3table[0][6] = RTE_PTYPE_L2_ETHER_LLDP;
1924 tbl->l2l3table[0][15] = RTE_PTYPE_L2_ETHER;
1926 tbl->l2l3table[1][0] = RTE_PTYPE_L2_ETHER_VLAN | RTE_PTYPE_L3_IPV4;
1927 tbl->l2l3table[1][1] = RTE_PTYPE_L2_ETHER_VLAN | RTE_PTYPE_L3_IPV6;
1928 tbl->l2l3table[1][2] = RTE_PTYPE_L2_ETHER_ARP;
1929 tbl->l2l3table[1][3] = RTE_PTYPE_L2_ETHER_VLAN;
1930 tbl->l2l3table[1][4] = RTE_PTYPE_L2_ETHER_VLAN | RTE_PTYPE_L3_IPV4_EXT;
1931 tbl->l2l3table[1][5] = RTE_PTYPE_L2_ETHER_VLAN | RTE_PTYPE_L3_IPV6_EXT;
1932 tbl->l2l3table[1][6] = RTE_PTYPE_L2_ETHER_LLDP;
1933 tbl->l2l3table[1][15] = RTE_PTYPE_L2_ETHER_VLAN;
1935 tbl->l2l3table[2][0] = RTE_PTYPE_L2_ETHER_QINQ | RTE_PTYPE_L3_IPV4;
1936 tbl->l2l3table[2][1] = RTE_PTYPE_L2_ETHER_QINQ | RTE_PTYPE_L3_IPV6;
1937 tbl->l2l3table[2][2] = RTE_PTYPE_L2_ETHER_ARP;
1938 tbl->l2l3table[2][3] = RTE_PTYPE_L2_ETHER_QINQ;
1939 tbl->l2l3table[2][4] = RTE_PTYPE_L2_ETHER_QINQ | RTE_PTYPE_L3_IPV4_EXT;
1940 tbl->l2l3table[2][5] = RTE_PTYPE_L2_ETHER_QINQ | RTE_PTYPE_L3_IPV6_EXT;
1941 tbl->l2l3table[2][6] = RTE_PTYPE_L2_ETHER_LLDP;
1942 tbl->l2l3table[2][15] = RTE_PTYPE_L2_ETHER_QINQ;
1944 tbl->l4table[0] = RTE_PTYPE_L4_UDP;
1945 tbl->l4table[1] = RTE_PTYPE_L4_TCP;
1946 tbl->l4table[2] = RTE_PTYPE_TUNNEL_GRE;
1947 tbl->l4table[3] = RTE_PTYPE_L4_SCTP;
1948 tbl->l4table[4] = RTE_PTYPE_L4_IGMP;
1949 tbl->l4table[5] = RTE_PTYPE_L4_ICMP;
1953 hns3_init_tunnel_ptype_tbl(struct hns3_ptype_table *tbl)
1955 tbl->inner_l2table[0] = RTE_PTYPE_INNER_L2_ETHER;
1956 tbl->inner_l2table[1] = RTE_PTYPE_INNER_L2_ETHER_VLAN;
1957 tbl->inner_l2table[2] = RTE_PTYPE_INNER_L2_ETHER_QINQ;
1959 tbl->inner_l3table[0] = RTE_PTYPE_INNER_L3_IPV4;
1960 tbl->inner_l3table[1] = RTE_PTYPE_INNER_L3_IPV6;
1961 /* There is not a ptype for inner ARP/RARP */
1962 tbl->inner_l3table[2] = RTE_PTYPE_UNKNOWN;
1963 tbl->inner_l3table[3] = RTE_PTYPE_UNKNOWN;
1964 tbl->inner_l3table[4] = RTE_PTYPE_INNER_L3_IPV4_EXT;
1965 tbl->inner_l3table[5] = RTE_PTYPE_INNER_L3_IPV6_EXT;
1967 tbl->inner_l4table[0] = RTE_PTYPE_INNER_L4_UDP;
1968 tbl->inner_l4table[1] = RTE_PTYPE_INNER_L4_TCP;
1969 /* There is not a ptype for inner GRE */
1970 tbl->inner_l4table[2] = RTE_PTYPE_UNKNOWN;
1971 tbl->inner_l4table[3] = RTE_PTYPE_INNER_L4_SCTP;
1972 /* There is not a ptype for inner IGMP */
1973 tbl->inner_l4table[4] = RTE_PTYPE_UNKNOWN;
1974 tbl->inner_l4table[5] = RTE_PTYPE_INNER_L4_ICMP;
1976 tbl->ol2table[0] = RTE_PTYPE_L2_ETHER;
1977 tbl->ol2table[1] = RTE_PTYPE_L2_ETHER_VLAN;
1978 tbl->ol2table[2] = RTE_PTYPE_L2_ETHER_QINQ;
1980 tbl->ol3table[0] = RTE_PTYPE_L3_IPV4;
1981 tbl->ol3table[1] = RTE_PTYPE_L3_IPV6;
1982 tbl->ol3table[2] = RTE_PTYPE_UNKNOWN;
1983 tbl->ol3table[3] = RTE_PTYPE_UNKNOWN;
1984 tbl->ol3table[4] = RTE_PTYPE_L3_IPV4_EXT;
1985 tbl->ol3table[5] = RTE_PTYPE_L3_IPV6_EXT;
1987 tbl->ol4table[0] = RTE_PTYPE_UNKNOWN;
1988 tbl->ol4table[1] = RTE_PTYPE_TUNNEL_VXLAN;
1989 tbl->ol4table[2] = RTE_PTYPE_TUNNEL_NVGRE;
1993 hns3_init_adv_layout_ptype(struct hns3_ptype_table *tbl)
1995 uint32_t *ptype = tbl->ptype;
1998 ptype[1] = RTE_PTYPE_L2_ETHER_ARP;
1999 ptype[3] = RTE_PTYPE_L2_ETHER_LLDP;
2000 ptype[8] = RTE_PTYPE_L2_ETHER_TIMESYNC;
2002 /* Non-tunnel IPv4 */
2003 ptype[17] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
2005 ptype[18] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
2006 RTE_PTYPE_L4_NONFRAG;
2007 ptype[19] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
2009 ptype[20] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
2011 /* The next ptype is GRE over IPv4 */
2012 ptype[21] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN;
2013 ptype[22] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
2015 ptype[23] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
2017 ptype[24] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
2019 /* The next ptype is PTP over IPv4 + UDP */
2020 ptype[25] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
2023 /* IPv4 --> GRE/Teredo/VXLAN */
2024 ptype[29] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
2025 RTE_PTYPE_TUNNEL_GRENAT;
2026 /* IPv4 --> GRE/Teredo/VXLAN --> MAC */
2027 ptype[30] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
2028 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER;
2030 /* IPv4 --> GRE/Teredo/VXLAN --> MAC --> IPv4 */
2031 ptype[31] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
2032 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2033 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
2034 RTE_PTYPE_INNER_L4_FRAG;
2035 ptype[32] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
2036 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2037 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
2038 RTE_PTYPE_INNER_L4_NONFRAG;
2039 ptype[33] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
2040 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2041 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
2042 RTE_PTYPE_INNER_L4_UDP;
2043 ptype[34] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
2044 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2045 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
2046 RTE_PTYPE_INNER_L4_TCP;
2047 ptype[35] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
2048 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2049 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
2050 RTE_PTYPE_INNER_L4_SCTP;
2051 /* The next ptype's inner L4 is IGMP */
2052 ptype[36] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
2053 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2054 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN;
2055 ptype[37] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
2056 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2057 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
2058 RTE_PTYPE_INNER_L4_ICMP;
2060 /* IPv4 --> GRE/Teredo/VXLAN --> MAC --> IPv6 */
2061 ptype[39] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
2062 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2063 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
2064 RTE_PTYPE_INNER_L4_FRAG;
2065 ptype[40] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
2066 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2067 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
2068 RTE_PTYPE_INNER_L4_NONFRAG;
2069 ptype[41] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
2070 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2071 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
2072 RTE_PTYPE_INNER_L4_UDP;
2073 ptype[42] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
2074 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2075 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
2076 RTE_PTYPE_INNER_L4_TCP;
2077 ptype[43] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
2078 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2079 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
2080 RTE_PTYPE_INNER_L4_SCTP;
2081 /* The next ptype's inner L4 is IGMP */
2082 ptype[44] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
2083 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2084 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN;
2085 ptype[45] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
2086 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2087 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
2088 RTE_PTYPE_INNER_L4_ICMP;
2090 /* Non-tunnel IPv6 */
2091 ptype[111] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
2093 ptype[112] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
2094 RTE_PTYPE_L4_NONFRAG;
2095 ptype[113] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
2097 ptype[114] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
2099 /* The next ptype is GRE over IPv6 */
2100 ptype[115] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN;
2101 ptype[116] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
2103 ptype[117] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
2105 ptype[118] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
2107 /* Special for PTP over IPv6 + UDP */
2108 ptype[119] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
2111 /* IPv6 --> GRE/Teredo/VXLAN */
2112 ptype[123] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
2113 RTE_PTYPE_TUNNEL_GRENAT;
2114 /* IPv6 --> GRE/Teredo/VXLAN --> MAC */
2115 ptype[124] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
2116 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER;
2118 /* IPv6 --> GRE/Teredo/VXLAN --> MAC --> IPv4 */
2119 ptype[125] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
2120 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2121 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
2122 RTE_PTYPE_INNER_L4_FRAG;
2123 ptype[126] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
2124 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2125 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
2126 RTE_PTYPE_INNER_L4_NONFRAG;
2127 ptype[127] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
2128 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2129 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
2130 RTE_PTYPE_INNER_L4_UDP;
2131 ptype[128] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
2132 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2133 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
2134 RTE_PTYPE_INNER_L4_TCP;
2135 ptype[129] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
2136 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2137 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
2138 RTE_PTYPE_INNER_L4_SCTP;
2139 /* The next ptype's inner L4 is IGMP */
2140 ptype[130] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
2141 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2142 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN;
2143 ptype[131] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
2144 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2145 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
2146 RTE_PTYPE_INNER_L4_ICMP;
2148 /* IPv6 --> GRE/Teredo/VXLAN --> MAC --> IPv6 */
2149 ptype[133] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
2150 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2151 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
2152 RTE_PTYPE_INNER_L4_FRAG;
2153 ptype[134] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
2154 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2155 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
2156 RTE_PTYPE_INNER_L4_NONFRAG;
2157 ptype[135] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
2158 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2159 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
2160 RTE_PTYPE_INNER_L4_UDP;
2161 ptype[136] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
2162 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2163 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
2164 RTE_PTYPE_INNER_L4_TCP;
2165 ptype[137] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
2166 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2167 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
2168 RTE_PTYPE_INNER_L4_SCTP;
2169 /* The next ptype's inner L4 is IGMP */
2170 ptype[138] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
2171 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2172 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN;
2173 ptype[139] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
2174 RTE_PTYPE_TUNNEL_GRENAT | RTE_PTYPE_INNER_L2_ETHER |
2175 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
2176 RTE_PTYPE_INNER_L4_ICMP;
2180 hns3_init_rx_ptype_tble(struct rte_eth_dev *dev)
2182 struct hns3_adapter *hns = dev->data->dev_private;
2183 struct hns3_ptype_table *tbl = &hns->ptype_tbl;
2185 memset(tbl, 0, sizeof(*tbl));
2187 hns3_init_non_tunnel_ptype_tbl(tbl);
2188 hns3_init_tunnel_ptype_tbl(tbl);
2189 hns3_init_adv_layout_ptype(tbl);
2193 hns3_rxd_to_vlan_tci(struct hns3_rx_queue *rxq, struct rte_mbuf *mb,
2194 uint32_t l234_info, const struct hns3_desc *rxd)
2196 #define HNS3_STRP_STATUS_NUM 0x4
2198 #define HNS3_NO_STRP_VLAN_VLD 0x0
2199 #define HNS3_INNER_STRP_VLAN_VLD 0x1
2200 #define HNS3_OUTER_STRP_VLAN_VLD 0x2
2201 uint32_t strip_status;
2202 uint32_t report_mode;
2205 * Since HW limitation, the vlan tag will always be inserted into RX
2206 * descriptor when strip the tag from packet, driver needs to determine
2207 * reporting which tag to mbuf according to the PVID configuration
2208 * and vlan striped status.
2210 static const uint32_t report_type[][HNS3_STRP_STATUS_NUM] = {
2212 HNS3_NO_STRP_VLAN_VLD,
2213 HNS3_OUTER_STRP_VLAN_VLD,
2214 HNS3_INNER_STRP_VLAN_VLD,
2215 HNS3_OUTER_STRP_VLAN_VLD
2218 HNS3_NO_STRP_VLAN_VLD,
2219 HNS3_NO_STRP_VLAN_VLD,
2220 HNS3_NO_STRP_VLAN_VLD,
2221 HNS3_INNER_STRP_VLAN_VLD
2224 strip_status = hns3_get_field(l234_info, HNS3_RXD_STRP_TAGP_M,
2225 HNS3_RXD_STRP_TAGP_S);
2226 report_mode = report_type[rxq->pvid_sw_discard_en][strip_status];
2227 switch (report_mode) {
2228 case HNS3_NO_STRP_VLAN_VLD:
2231 case HNS3_INNER_STRP_VLAN_VLD:
2232 mb->ol_flags |= PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
2233 mb->vlan_tci = rte_le_to_cpu_16(rxd->rx.vlan_tag);
2235 case HNS3_OUTER_STRP_VLAN_VLD:
2236 mb->ol_flags |= PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
2237 mb->vlan_tci = rte_le_to_cpu_16(rxd->rx.ot_vlan_tag);
2246 recalculate_data_len(struct rte_mbuf *first_seg, struct rte_mbuf *last_seg,
2247 struct rte_mbuf *rxm, struct hns3_rx_queue *rxq,
2250 uint8_t crc_len = rxq->crc_len;
2252 if (data_len <= crc_len) {
2253 rte_pktmbuf_free_seg(rxm);
2254 first_seg->nb_segs--;
2255 last_seg->data_len = (uint16_t)(last_seg->data_len -
2256 (crc_len - data_len));
2257 last_seg->next = NULL;
2259 rxm->data_len = (uint16_t)(data_len - crc_len);
2262 static inline struct rte_mbuf *
2263 hns3_rx_alloc_buffer(struct hns3_rx_queue *rxq)
2267 if (likely(rxq->bulk_mbuf_num > 0))
2268 return rxq->bulk_mbuf[--rxq->bulk_mbuf_num];
2270 ret = rte_mempool_get_bulk(rxq->mb_pool, (void **)rxq->bulk_mbuf,
2271 HNS3_BULK_ALLOC_MBUF_NUM);
2272 if (likely(ret == 0)) {
2273 rxq->bulk_mbuf_num = HNS3_BULK_ALLOC_MBUF_NUM;
2274 return rxq->bulk_mbuf[--rxq->bulk_mbuf_num];
2276 return rte_mbuf_raw_alloc(rxq->mb_pool);
2280 hns3_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
2282 volatile struct hns3_desc *rx_ring; /* RX ring (desc) */
2283 volatile struct hns3_desc *rxdp; /* pointer of the current desc */
2284 struct hns3_rx_queue *rxq; /* RX queue */
2285 struct hns3_entry *sw_ring;
2286 struct hns3_entry *rxe;
2287 struct hns3_desc rxd;
2288 struct rte_mbuf *nmb; /* pointer of the new mbuf */
2289 struct rte_mbuf *rxm;
2290 uint32_t bd_base_info;
2303 rx_ring = rxq->rx_ring;
2304 sw_ring = rxq->sw_ring;
2305 rx_id = rxq->next_to_use;
2307 while (nb_rx < nb_pkts) {
2308 rxdp = &rx_ring[rx_id];
2309 bd_base_info = rte_le_to_cpu_32(rxdp->rx.bd_base_info);
2310 if (unlikely(!(bd_base_info & BIT(HNS3_RXD_VLD_B))))
2313 rxd = rxdp[(bd_base_info & (1u << HNS3_RXD_VLD_B)) -
2314 (1u << HNS3_RXD_VLD_B)];
2316 nmb = hns3_rx_alloc_buffer(rxq);
2317 if (unlikely(nmb == NULL)) {
2320 port_id = rxq->port_id;
2321 rte_eth_devices[port_id].data->rx_mbuf_alloc_failed++;
2326 rxe = &sw_ring[rx_id];
2328 if (unlikely(rx_id == rxq->nb_rx_desc))
2331 rte_prefetch0(sw_ring[rx_id].mbuf);
2332 if ((rx_id & HNS3_RX_RING_PREFETCTH_MASK) == 0) {
2333 rte_prefetch0(&rx_ring[rx_id]);
2334 rte_prefetch0(&sw_ring[rx_id]);
2340 dma_addr = rte_mbuf_data_iova_default(nmb);
2341 rxdp->addr = rte_cpu_to_le_64(dma_addr);
2342 rxdp->rx.bd_base_info = 0;
2344 rxm->data_off = RTE_PKTMBUF_HEADROOM;
2345 rxm->pkt_len = (uint16_t)(rte_le_to_cpu_16(rxd.rx.pkt_len)) -
2347 rxm->data_len = rxm->pkt_len;
2348 rxm->port = rxq->port_id;
2349 rxm->hash.rss = rte_le_to_cpu_32(rxd.rx.rss_hash);
2350 rxm->ol_flags = PKT_RX_RSS_HASH;
2351 if (unlikely(bd_base_info & BIT(HNS3_RXD_LUM_B))) {
2353 rte_le_to_cpu_16(rxd.rx.fd_id);
2354 rxm->ol_flags |= PKT_RX_FDIR | PKT_RX_FDIR_ID;
2359 /* Load remained descriptor data and extract necessary fields */
2360 l234_info = rte_le_to_cpu_32(rxd.rx.l234_info);
2361 ol_info = rte_le_to_cpu_32(rxd.rx.ol_info);
2362 ret = hns3_handle_bdinfo(rxq, rxm, bd_base_info,
2363 l234_info, &cksum_err);
2367 rxm->packet_type = hns3_rx_calc_ptype(rxq, l234_info, ol_info);
2369 if (likely(bd_base_info & BIT(HNS3_RXD_L3L4P_B)))
2370 hns3_rx_set_cksum_flag(rxm, rxm->packet_type,
2372 hns3_rxd_to_vlan_tci(rxq, rxm, l234_info, &rxd);
2374 /* Increment bytes counter */
2375 rxq->basic_stats.bytes += rxm->pkt_len;
2377 rx_pkts[nb_rx++] = rxm;
2380 rte_pktmbuf_free(rxm);
2383 rxq->next_to_use = rx_id;
2384 rxq->rx_free_hold += nb_rx_bd;
2385 if (rxq->rx_free_hold > rxq->rx_free_thresh) {
2386 hns3_write_reg_opt(rxq->io_head_reg, rxq->rx_free_hold);
2387 rxq->rx_free_hold = 0;
2394 hns3_recv_scattered_pkts(void *rx_queue,
2395 struct rte_mbuf **rx_pkts,
2398 volatile struct hns3_desc *rx_ring; /* RX ring (desc) */
2399 volatile struct hns3_desc *rxdp; /* pointer of the current desc */
2400 struct hns3_rx_queue *rxq; /* RX queue */
2401 struct hns3_entry *sw_ring;
2402 struct hns3_entry *rxe;
2403 struct rte_mbuf *first_seg;
2404 struct rte_mbuf *last_seg;
2405 struct hns3_desc rxd;
2406 struct rte_mbuf *nmb; /* pointer of the new mbuf */
2407 struct rte_mbuf *rxm;
2408 struct rte_eth_dev *dev;
2409 uint32_t bd_base_info;
2424 rx_id = rxq->next_to_use;
2425 rx_ring = rxq->rx_ring;
2426 sw_ring = rxq->sw_ring;
2427 first_seg = rxq->pkt_first_seg;
2428 last_seg = rxq->pkt_last_seg;
2430 while (nb_rx < nb_pkts) {
2431 rxdp = &rx_ring[rx_id];
2432 bd_base_info = rte_le_to_cpu_32(rxdp->rx.bd_base_info);
2433 if (unlikely(!(bd_base_info & BIT(HNS3_RXD_VLD_B))))
2437 * The interactive process between software and hardware of
2438 * receiving a new packet in hns3 network engine:
2439 * 1. Hardware network engine firstly writes the packet content
2440 * to the memory pointed by the 'addr' field of the Rx Buffer
2441 * Descriptor, secondly fills the result of parsing the
2442 * packet include the valid field into the Rx Buffer
2443 * Descriptor in one write operation.
2444 * 2. Driver reads the Rx BD's valid field in the loop to check
2445 * whether it's valid, if valid then assign a new address to
2446 * the addr field, clear the valid field, get the other
2447 * information of the packet by parsing Rx BD's other fields,
2448 * finally write back the number of Rx BDs processed by the
2449 * driver to the HNS3_RING_RX_HEAD_REG register to inform
2451 * In the above process, the ordering is very important. We must
2452 * make sure that CPU read Rx BD's other fields only after the
2455 * There are two type of re-ordering: compiler re-ordering and
2456 * CPU re-ordering under the ARMv8 architecture.
2457 * 1. we use volatile to deal with compiler re-ordering, so you
2458 * can see that rx_ring/rxdp defined with volatile.
2459 * 2. we commonly use memory barrier to deal with CPU
2460 * re-ordering, but the cost is high.
2462 * In order to solve the high cost of using memory barrier, we
2463 * use the data dependency order under the ARMv8 architecture,
2466 * instr02: load B <- A
2467 * the instr02 will always execute after instr01.
2469 * To construct the data dependency ordering, we use the
2470 * following assignment:
2471 * rxd = rxdp[(bd_base_info & (1u << HNS3_RXD_VLD_B)) -
2472 * (1u<<HNS3_RXD_VLD_B)]
2473 * Using gcc compiler under the ARMv8 architecture, the related
2474 * assembly code example as follows:
2475 * note: (1u << HNS3_RXD_VLD_B) equal 0x10
2476 * instr01: ldr w26, [x22, #28] --read bd_base_info
2477 * instr02: and w0, w26, #0x10 --calc bd_base_info & 0x10
2478 * instr03: sub w0, w0, #0x10 --calc (bd_base_info &
2480 * instr04: add x0, x22, x0, lsl #5 --calc copy source addr
2481 * instr05: ldp x2, x3, [x0]
2482 * instr06: stp x2, x3, [x29, #256] --copy BD's [0 ~ 15]B
2483 * instr07: ldp x4, x5, [x0, #16]
2484 * instr08: stp x4, x5, [x29, #272] --copy BD's [16 ~ 31]B
2485 * the instr05~08 depend on x0's value, x0 depent on w26's
2486 * value, the w26 is the bd_base_info, this form the data
2487 * dependency ordering.
2488 * note: if BD is valid, (bd_base_info & (1u<<HNS3_RXD_VLD_B)) -
2489 * (1u<<HNS3_RXD_VLD_B) will always zero, so the
2490 * assignment is correct.
2492 * So we use the data dependency ordering instead of memory
2493 * barrier to improve receive performance.
2495 rxd = rxdp[(bd_base_info & (1u << HNS3_RXD_VLD_B)) -
2496 (1u << HNS3_RXD_VLD_B)];
2498 nmb = hns3_rx_alloc_buffer(rxq);
2499 if (unlikely(nmb == NULL)) {
2500 dev = &rte_eth_devices[rxq->port_id];
2501 dev->data->rx_mbuf_alloc_failed++;
2506 rxe = &sw_ring[rx_id];
2508 if (unlikely(rx_id == rxq->nb_rx_desc))
2511 rte_prefetch0(sw_ring[rx_id].mbuf);
2512 if ((rx_id & HNS3_RX_RING_PREFETCTH_MASK) == 0) {
2513 rte_prefetch0(&rx_ring[rx_id]);
2514 rte_prefetch0(&sw_ring[rx_id]);
2520 dma_addr = rte_cpu_to_le_64(rte_mbuf_data_iova_default(nmb));
2521 rxdp->rx.bd_base_info = 0;
2522 rxdp->addr = dma_addr;
2524 if (first_seg == NULL) {
2526 first_seg->nb_segs = 1;
2528 first_seg->nb_segs++;
2529 last_seg->next = rxm;
2532 rxm->data_off = RTE_PKTMBUF_HEADROOM;
2533 rxm->data_len = rte_le_to_cpu_16(rxd.rx.size);
2535 if (!(bd_base_info & BIT(HNS3_RXD_FE_B))) {
2542 * The last buffer of the received packet. packet len from
2543 * buffer description may contains CRC len, packet len should
2544 * subtract it, same as data len.
2546 first_seg->pkt_len = rte_le_to_cpu_16(rxd.rx.pkt_len);
2549 * This is the last buffer of the received packet. If the CRC
2550 * is not stripped by the hardware:
2551 * - Subtract the CRC length from the total packet length.
2552 * - If the last buffer only contains the whole CRC or a part
2553 * of it, free the mbuf associated to the last buffer. If part
2554 * of the CRC is also contained in the previous mbuf, subtract
2555 * the length of that CRC part from the data length of the
2559 if (unlikely(rxq->crc_len > 0)) {
2560 first_seg->pkt_len -= rxq->crc_len;
2561 recalculate_data_len(first_seg, last_seg, rxm, rxq,
2565 first_seg->port = rxq->port_id;
2566 first_seg->hash.rss = rte_le_to_cpu_32(rxd.rx.rss_hash);
2567 first_seg->ol_flags = PKT_RX_RSS_HASH;
2568 if (unlikely(bd_base_info & BIT(HNS3_RXD_LUM_B))) {
2569 first_seg->hash.fdir.hi =
2570 rte_le_to_cpu_16(rxd.rx.fd_id);
2571 first_seg->ol_flags |= PKT_RX_FDIR | PKT_RX_FDIR_ID;
2574 gro_size = hns3_get_field(bd_base_info, HNS3_RXD_GRO_SIZE_M,
2575 HNS3_RXD_GRO_SIZE_S);
2576 if (gro_size != 0) {
2577 first_seg->ol_flags |= PKT_RX_LRO;
2578 first_seg->tso_segsz = gro_size;
2581 l234_info = rte_le_to_cpu_32(rxd.rx.l234_info);
2582 ol_info = rte_le_to_cpu_32(rxd.rx.ol_info);
2583 ret = hns3_handle_bdinfo(rxq, first_seg, bd_base_info,
2584 l234_info, &cksum_err);
2588 first_seg->packet_type = hns3_rx_calc_ptype(rxq,
2589 l234_info, ol_info);
2591 if (bd_base_info & BIT(HNS3_RXD_L3L4P_B))
2592 hns3_rx_set_cksum_flag(first_seg,
2593 first_seg->packet_type,
2595 hns3_rxd_to_vlan_tci(rxq, first_seg, l234_info, &rxd);
2597 /* Increment bytes counter */
2598 rxq->basic_stats.bytes += first_seg->pkt_len;
2600 rx_pkts[nb_rx++] = first_seg;
2604 rte_pktmbuf_free(first_seg);
2608 rxq->next_to_use = rx_id;
2609 rxq->pkt_first_seg = first_seg;
2610 rxq->pkt_last_seg = last_seg;
2612 rxq->rx_free_hold += nb_rx_bd;
2613 if (rxq->rx_free_hold > rxq->rx_free_thresh) {
2614 hns3_write_reg_opt(rxq->io_head_reg, rxq->rx_free_hold);
2615 rxq->rx_free_hold = 0;
2622 hns3_rxq_vec_setup(__rte_unused struct hns3_rx_queue *rxq)
2627 hns3_rx_check_vec_support(__rte_unused struct rte_eth_dev *dev)
2633 hns3_recv_pkts_vec(__rte_unused void *tx_queue,
2634 __rte_unused struct rte_mbuf **rx_pkts,
2635 __rte_unused uint16_t nb_pkts)
2641 hns3_recv_pkts_vec_sve(__rte_unused void *tx_queue,
2642 __rte_unused struct rte_mbuf **rx_pkts,
2643 __rte_unused uint16_t nb_pkts)
2649 hns3_rx_burst_mode_get(struct rte_eth_dev *dev, __rte_unused uint16_t queue_id,
2650 struct rte_eth_burst_mode *mode)
2652 static const struct {
2653 eth_rx_burst_t pkt_burst;
2656 { hns3_recv_pkts, "Scalar" },
2657 { hns3_recv_scattered_pkts, "Scalar Scattered" },
2658 { hns3_recv_pkts_vec, "Vector Neon" },
2659 { hns3_recv_pkts_vec_sve, "Vector Sve" },
2662 eth_rx_burst_t pkt_burst = dev->rx_pkt_burst;
2666 for (i = 0; i < RTE_DIM(burst_infos); i++) {
2667 if (pkt_burst == burst_infos[i].pkt_burst) {
2668 snprintf(mode->info, sizeof(mode->info), "%s",
2669 burst_infos[i].info);
2679 hns3_check_sve_support(void)
2681 #if defined(RTE_ARCH_ARM64) && defined(__ARM_FEATURE_SVE)
2682 if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_SVE))
2688 static eth_rx_burst_t
2689 hns3_get_rx_function(struct rte_eth_dev *dev)
2691 struct hns3_adapter *hns = dev->data->dev_private;
2692 uint64_t offloads = dev->data->dev_conf.rxmode.offloads;
2693 bool vec_allowed, sve_allowed, simple_allowed;
2695 vec_allowed = hns->rx_vec_allowed &&
2696 hns3_rx_check_vec_support(dev) == 0;
2697 sve_allowed = vec_allowed && hns3_check_sve_support();
2698 simple_allowed = hns->rx_simple_allowed && !dev->data->scattered_rx &&
2699 (offloads & DEV_RX_OFFLOAD_TCP_LRO) == 0;
2701 if (hns->rx_func_hint == HNS3_IO_FUNC_HINT_VEC && vec_allowed)
2702 return hns3_recv_pkts_vec;
2703 if (hns->rx_func_hint == HNS3_IO_FUNC_HINT_SVE && sve_allowed)
2704 return hns3_recv_pkts_vec_sve;
2705 if (hns->rx_func_hint == HNS3_IO_FUNC_HINT_SIMPLE && simple_allowed)
2706 return hns3_recv_pkts;
2707 if (hns->rx_func_hint == HNS3_IO_FUNC_HINT_COMMON)
2708 return hns3_recv_scattered_pkts;
2711 return hns3_recv_pkts_vec;
2713 return hns3_recv_pkts;
2715 return hns3_recv_scattered_pkts;
2719 hns3_tx_queue_conf_check(struct hns3_hw *hw, const struct rte_eth_txconf *conf,
2720 uint16_t nb_desc, uint16_t *tx_rs_thresh,
2721 uint16_t *tx_free_thresh, uint16_t idx)
2723 #define HNS3_TX_RS_FREE_THRESH_GAP 8
2724 uint16_t rs_thresh, free_thresh, fast_free_thresh;
2726 if (nb_desc > HNS3_MAX_RING_DESC || nb_desc < HNS3_MIN_RING_DESC ||
2727 nb_desc % HNS3_ALIGN_RING_DESC) {
2728 hns3_err(hw, "number (%u) of tx descriptors is invalid",
2733 rs_thresh = (conf->tx_rs_thresh > 0) ?
2734 conf->tx_rs_thresh : HNS3_DEFAULT_TX_RS_THRESH;
2735 free_thresh = (conf->tx_free_thresh > 0) ?
2736 conf->tx_free_thresh : HNS3_DEFAULT_TX_FREE_THRESH;
2737 if (rs_thresh + free_thresh > nb_desc || nb_desc % rs_thresh ||
2738 rs_thresh >= nb_desc - HNS3_TX_RS_FREE_THRESH_GAP ||
2739 free_thresh >= nb_desc - HNS3_TX_RS_FREE_THRESH_GAP) {
2740 hns3_err(hw, "tx_rs_thresh (%u) tx_free_thresh (%u) nb_desc "
2741 "(%u) of tx descriptors for port=%u queue=%u check "
2743 rs_thresh, free_thresh, nb_desc, hw->data->port_id,
2748 if (conf->tx_free_thresh == 0) {
2749 /* Fast free Tx memory buffer to improve cache hit rate */
2750 fast_free_thresh = nb_desc - rs_thresh;
2751 if (fast_free_thresh >=
2752 HNS3_TX_FAST_FREE_AHEAD + HNS3_DEFAULT_TX_FREE_THRESH)
2753 free_thresh = fast_free_thresh -
2754 HNS3_TX_FAST_FREE_AHEAD;
2757 *tx_rs_thresh = rs_thresh;
2758 *tx_free_thresh = free_thresh;
2763 hns3_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t nb_desc,
2764 unsigned int socket_id, const struct rte_eth_txconf *conf)
2766 struct hns3_adapter *hns = dev->data->dev_private;
2767 uint16_t tx_rs_thresh, tx_free_thresh;
2768 struct hns3_hw *hw = &hns->hw;
2769 struct hns3_queue_info q_info;
2770 struct hns3_tx_queue *txq;
2774 ret = hns3_tx_queue_conf_check(hw, conf, nb_desc,
2775 &tx_rs_thresh, &tx_free_thresh, idx);
2779 if (dev->data->tx_queues[idx] != NULL) {
2780 hns3_tx_queue_release(dev->data->tx_queues[idx]);
2781 dev->data->tx_queues[idx] = NULL;
2785 q_info.socket_id = socket_id;
2786 q_info.nb_desc = nb_desc;
2787 q_info.type = "hns3 TX queue";
2788 q_info.ring_name = "tx_ring";
2789 txq = hns3_alloc_txq_and_dma_zone(dev, &q_info);
2792 "Failed to alloc mem and reserve DMA mem for tx ring!");
2796 txq->tx_deferred_start = conf->tx_deferred_start;
2797 if (txq->tx_deferred_start && !hns3_dev_indep_txrx_supported(hw)) {
2798 hns3_warn(hw, "deferred start is not supported.");
2799 txq->tx_deferred_start = false;
2802 tx_entry_len = sizeof(struct hns3_entry) * txq->nb_tx_desc;
2803 txq->sw_ring = rte_zmalloc_socket("hns3 TX sw ring", tx_entry_len,
2804 RTE_CACHE_LINE_SIZE, socket_id);
2805 if (txq->sw_ring == NULL) {
2806 hns3_err(hw, "Failed to allocate memory for tx sw ring!");
2807 hns3_tx_queue_release(txq);
2812 txq->next_to_use = 0;
2813 txq->next_to_clean = 0;
2814 txq->tx_bd_ready = txq->nb_tx_desc - 1;
2815 txq->tx_free_thresh = tx_free_thresh;
2816 txq->tx_rs_thresh = tx_rs_thresh;
2817 txq->free = rte_zmalloc_socket("hns3 TX mbuf free array",
2818 sizeof(struct rte_mbuf *) * txq->tx_rs_thresh,
2819 RTE_CACHE_LINE_SIZE, socket_id);
2821 hns3_err(hw, "failed to allocate tx mbuf free array!");
2822 hns3_tx_queue_release(txq);
2826 txq->port_id = dev->data->port_id;
2828 * For hns3 PF device, if the VLAN mode is HW_SHIFT_AND_DISCARD_MODE,
2829 * the pvid_sw_shift_en in the queue struct should not be changed,
2830 * because PVID-related operations do not need to be processed by PMD
2831 * driver. For hns3 VF device, whether it needs to process PVID depends
2832 * on the configuration of PF kernel mode netdev driver. And the
2833 * related PF configuration is delivered through the mailbox and finally
2834 * reflectd in port_base_vlan_cfg.
2836 if (hns->is_vf || hw->vlan_mode == HNS3_SW_SHIFT_AND_DISCARD_MODE)
2837 txq->pvid_sw_shift_en = hw->port_base_vlan_cfg.state ==
2838 HNS3_PORT_BASE_VLAN_ENABLE;
2840 txq->pvid_sw_shift_en = false;
2841 txq->max_non_tso_bd_num = hw->max_non_tso_bd_num;
2842 txq->configured = true;
2843 txq->io_base = (void *)((char *)hw->io_base +
2844 hns3_get_tqp_reg_offset(idx));
2845 txq->io_tail_reg = (volatile void *)((char *)txq->io_base +
2846 HNS3_RING_TX_TAIL_REG);
2847 txq->min_tx_pkt_len = hw->min_tx_pkt_len;
2848 txq->tso_mode = hw->tso_mode;
2849 txq->udp_cksum_mode = hw->udp_cksum_mode;
2850 memset(&txq->basic_stats, 0, sizeof(struct hns3_tx_basic_stats));
2851 memset(&txq->dfx_stats, 0, sizeof(struct hns3_tx_dfx_stats));
2853 rte_spinlock_lock(&hw->lock);
2854 dev->data->tx_queues[idx] = txq;
2855 rte_spinlock_unlock(&hw->lock);
2861 hns3_tx_free_useless_buffer(struct hns3_tx_queue *txq)
2863 uint16_t tx_next_clean = txq->next_to_clean;
2864 uint16_t tx_next_use = txq->next_to_use;
2865 uint16_t tx_bd_ready = txq->tx_bd_ready;
2866 uint16_t tx_bd_max = txq->nb_tx_desc;
2867 struct hns3_entry *tx_bak_pkt = &txq->sw_ring[tx_next_clean];
2868 struct hns3_desc *desc = &txq->tx_ring[tx_next_clean];
2869 struct rte_mbuf *mbuf;
2871 while ((!(desc->tx.tp_fe_sc_vld_ra_ri &
2872 rte_cpu_to_le_16(BIT(HNS3_TXD_VLD_B)))) &&
2873 tx_next_use != tx_next_clean) {
2874 mbuf = tx_bak_pkt->mbuf;
2876 rte_pktmbuf_free_seg(mbuf);
2877 tx_bak_pkt->mbuf = NULL;
2885 if (tx_next_clean >= tx_bd_max) {
2887 desc = txq->tx_ring;
2888 tx_bak_pkt = txq->sw_ring;
2892 txq->next_to_clean = tx_next_clean;
2893 txq->tx_bd_ready = tx_bd_ready;
2897 hns3_config_gro(struct hns3_hw *hw, bool en)
2899 struct hns3_cfg_gro_status_cmd *req;
2900 struct hns3_cmd_desc desc;
2903 hns3_cmd_setup_basic_desc(&desc, HNS3_OPC_GRO_GENERIC_CONFIG, false);
2904 req = (struct hns3_cfg_gro_status_cmd *)desc.data;
2906 req->gro_en = rte_cpu_to_le_16(en ? 1 : 0);
2908 ret = hns3_cmd_send(hw, &desc, 1);
2910 hns3_err(hw, "%s hardware GRO failed, ret = %d",
2911 en ? "enable" : "disable", ret);
2917 hns3_restore_gro_conf(struct hns3_hw *hw)
2923 offloads = hw->data->dev_conf.rxmode.offloads;
2924 gro_en = offloads & DEV_RX_OFFLOAD_TCP_LRO ? true : false;
2925 ret = hns3_config_gro(hw, gro_en);
2927 hns3_err(hw, "restore hardware GRO to %s failed, ret = %d",
2928 gro_en ? "enabled" : "disabled", ret);
2934 hns3_pkt_is_tso(struct rte_mbuf *m)
2936 return (m->tso_segsz != 0 && m->ol_flags & PKT_TX_TCP_SEG);
2940 hns3_set_tso(struct hns3_desc *desc, uint32_t paylen, struct rte_mbuf *rxm)
2942 if (!hns3_pkt_is_tso(rxm))
2945 if (paylen <= rxm->tso_segsz)
2948 desc->tx.type_cs_vlan_tso_len |= rte_cpu_to_le_32(BIT(HNS3_TXD_TSO_B));
2949 desc->tx.mss = rte_cpu_to_le_16(rxm->tso_segsz);
2953 hns3_fill_per_desc(struct hns3_desc *desc, struct rte_mbuf *rxm)
2955 desc->addr = rte_mbuf_data_iova(rxm);
2956 desc->tx.send_size = rte_cpu_to_le_16(rte_pktmbuf_data_len(rxm));
2957 desc->tx.tp_fe_sc_vld_ra_ri = rte_cpu_to_le_16(BIT(HNS3_TXD_VLD_B));
2961 hns3_fill_first_desc(struct hns3_tx_queue *txq, struct hns3_desc *desc,
2962 struct rte_mbuf *rxm)
2964 uint64_t ol_flags = rxm->ol_flags;
2968 hdr_len = rxm->l2_len + rxm->l3_len + rxm->l4_len;
2969 hdr_len += (ol_flags & PKT_TX_TUNNEL_MASK) ?
2970 rxm->outer_l2_len + rxm->outer_l3_len : 0;
2971 paylen = rxm->pkt_len - hdr_len;
2972 desc->tx.paylen_fd_dop_ol4cs |= rte_cpu_to_le_32(paylen);
2973 hns3_set_tso(desc, paylen, rxm);
2976 * Currently, hardware doesn't support more than two layers VLAN offload
2977 * in Tx direction based on hns3 network engine. So when the number of
2978 * VLANs in the packets represented by rxm plus the number of VLAN
2979 * offload by hardware such as PVID etc, exceeds two, the packets will
2980 * be discarded or the original VLAN of the packets will be overwitted
2981 * by hardware. When the PF PVID is enabled by calling the API function
2982 * named rte_eth_dev_set_vlan_pvid or the VF PVID is enabled by the hns3
2983 * PF kernel ether driver, the outer VLAN tag will always be the PVID.
2984 * To avoid the VLAN of Tx descriptor is overwritten by PVID, it should
2985 * be added to the position close to the IP header when PVID is enabled.
2987 if (!txq->pvid_sw_shift_en && ol_flags & (PKT_TX_VLAN_PKT |
2989 desc->tx.ol_type_vlan_len_msec |=
2990 rte_cpu_to_le_32(BIT(HNS3_TXD_OVLAN_B));
2991 if (ol_flags & PKT_TX_QINQ_PKT)
2992 desc->tx.outer_vlan_tag =
2993 rte_cpu_to_le_16(rxm->vlan_tci_outer);
2995 desc->tx.outer_vlan_tag =
2996 rte_cpu_to_le_16(rxm->vlan_tci);
2999 if (ol_flags & PKT_TX_QINQ_PKT ||
3000 ((ol_flags & PKT_TX_VLAN_PKT) && txq->pvid_sw_shift_en)) {
3001 desc->tx.type_cs_vlan_tso_len |=
3002 rte_cpu_to_le_32(BIT(HNS3_TXD_VLAN_B));
3003 desc->tx.vlan_tag = rte_cpu_to_le_16(rxm->vlan_tci);
3008 hns3_tx_alloc_mbufs(struct rte_mempool *mb_pool, uint16_t nb_new_buf,
3009 struct rte_mbuf **alloc_mbuf)
3011 #define MAX_NON_TSO_BD_PER_PKT 18
3012 struct rte_mbuf *pkt_segs[MAX_NON_TSO_BD_PER_PKT];
3015 /* Allocate enough mbufs */
3016 if (rte_mempool_get_bulk(mb_pool, (void **)pkt_segs, nb_new_buf))
3019 for (i = 0; i < nb_new_buf - 1; i++)
3020 pkt_segs[i]->next = pkt_segs[i + 1];
3022 pkt_segs[nb_new_buf - 1]->next = NULL;
3023 pkt_segs[0]->nb_segs = nb_new_buf;
3024 *alloc_mbuf = pkt_segs[0];
3030 hns3_pktmbuf_copy_hdr(struct rte_mbuf *new_pkt, struct rte_mbuf *old_pkt)
3032 new_pkt->ol_flags = old_pkt->ol_flags;
3033 new_pkt->pkt_len = rte_pktmbuf_pkt_len(old_pkt);
3034 new_pkt->outer_l2_len = old_pkt->outer_l2_len;
3035 new_pkt->outer_l3_len = old_pkt->outer_l3_len;
3036 new_pkt->l2_len = old_pkt->l2_len;
3037 new_pkt->l3_len = old_pkt->l3_len;
3038 new_pkt->l4_len = old_pkt->l4_len;
3039 new_pkt->vlan_tci_outer = old_pkt->vlan_tci_outer;
3040 new_pkt->vlan_tci = old_pkt->vlan_tci;
3044 hns3_reassemble_tx_pkts(struct rte_mbuf *tx_pkt, struct rte_mbuf **new_pkt,
3045 uint8_t max_non_tso_bd_num)
3047 struct rte_mempool *mb_pool;
3048 struct rte_mbuf *new_mbuf;
3049 struct rte_mbuf *temp_new;
3050 struct rte_mbuf *temp;
3051 uint16_t last_buf_len;
3052 uint16_t nb_new_buf;
3062 mb_pool = tx_pkt->pool;
3063 buf_size = tx_pkt->buf_len - RTE_PKTMBUF_HEADROOM;
3064 nb_new_buf = (rte_pktmbuf_pkt_len(tx_pkt) - 1) / buf_size + 1;
3065 if (nb_new_buf > max_non_tso_bd_num)
3068 last_buf_len = rte_pktmbuf_pkt_len(tx_pkt) % buf_size;
3069 if (last_buf_len == 0)
3070 last_buf_len = buf_size;
3072 /* Allocate enough mbufs */
3073 ret = hns3_tx_alloc_mbufs(mb_pool, nb_new_buf, &new_mbuf);
3077 /* Copy the original packet content to the new mbufs */
3079 s = rte_pktmbuf_mtod(temp, char *);
3080 len_s = rte_pktmbuf_data_len(temp);
3081 temp_new = new_mbuf;
3082 while (temp != NULL && temp_new != NULL) {
3083 d = rte_pktmbuf_mtod(temp_new, char *);
3084 buf_len = temp_new->next == NULL ? last_buf_len : buf_size;
3088 len = RTE_MIN(len_s, len_d);
3092 len_d = len_d - len;
3093 len_s = len_s - len;
3099 s = rte_pktmbuf_mtod(temp, char *);
3100 len_s = rte_pktmbuf_data_len(temp);
3104 temp_new->data_len = buf_len;
3105 temp_new = temp_new->next;
3107 hns3_pktmbuf_copy_hdr(new_mbuf, tx_pkt);
3109 /* free original mbufs */
3110 rte_pktmbuf_free(tx_pkt);
3112 *new_pkt = new_mbuf;
3118 hns3_parse_outer_params(struct rte_mbuf *m, uint32_t *ol_type_vlan_len_msec)
3120 uint32_t tmp = *ol_type_vlan_len_msec;
3121 uint64_t ol_flags = m->ol_flags;
3123 /* (outer) IP header type */
3124 if (ol_flags & PKT_TX_OUTER_IPV4) {
3125 if (ol_flags & PKT_TX_OUTER_IP_CKSUM)
3126 tmp |= hns3_gen_field_val(HNS3_TXD_OL3T_M,
3127 HNS3_TXD_OL3T_S, HNS3_OL3T_IPV4_CSUM);
3129 tmp |= hns3_gen_field_val(HNS3_TXD_OL3T_M,
3130 HNS3_TXD_OL3T_S, HNS3_OL3T_IPV4_NO_CSUM);
3131 } else if (ol_flags & PKT_TX_OUTER_IPV6) {
3132 tmp |= hns3_gen_field_val(HNS3_TXD_OL3T_M, HNS3_TXD_OL3T_S,
3135 /* OL3 header size, defined in 4 bytes */
3136 tmp |= hns3_gen_field_val(HNS3_TXD_L3LEN_M, HNS3_TXD_L3LEN_S,
3137 m->outer_l3_len >> HNS3_L3_LEN_UNIT);
3138 *ol_type_vlan_len_msec = tmp;
3142 hns3_parse_inner_params(struct rte_mbuf *m, uint32_t *ol_type_vlan_len_msec,
3143 uint32_t *type_cs_vlan_tso_len)
3145 #define HNS3_NVGRE_HLEN 8
3146 uint32_t tmp_outer = *ol_type_vlan_len_msec;
3147 uint32_t tmp_inner = *type_cs_vlan_tso_len;
3148 uint64_t ol_flags = m->ol_flags;
3149 uint16_t inner_l2_len;
3151 switch (ol_flags & PKT_TX_TUNNEL_MASK) {
3152 case PKT_TX_TUNNEL_VXLAN_GPE:
3153 case PKT_TX_TUNNEL_GENEVE:
3154 case PKT_TX_TUNNEL_VXLAN:
3155 /* MAC in UDP tunnelling packet, include VxLAN and GENEVE */
3156 tmp_outer |= hns3_gen_field_val(HNS3_TXD_TUNTYPE_M,
3157 HNS3_TXD_TUNTYPE_S, HNS3_TUN_MAC_IN_UDP);
3159 * The inner l2 length of mbuf is the sum of outer l4 length,
3160 * tunneling header length and inner l2 length for a tunnel
3161 * packect. But in hns3 tx descriptor, the tunneling header
3162 * length is contained in the field of outer L4 length.
3163 * Therefore, driver need to calculate the outer L4 length and
3166 tmp_outer |= hns3_gen_field_val(HNS3_TXD_L4LEN_M,
3168 (uint8_t)RTE_ETHER_VXLAN_HLEN >>
3171 inner_l2_len = m->l2_len - RTE_ETHER_VXLAN_HLEN;
3173 case PKT_TX_TUNNEL_GRE:
3174 tmp_outer |= hns3_gen_field_val(HNS3_TXD_TUNTYPE_M,
3175 HNS3_TXD_TUNTYPE_S, HNS3_TUN_NVGRE);
3177 * For NVGRE tunnel packect, the outer L4 is empty. So only
3178 * fill the NVGRE header length to the outer L4 field.
3180 tmp_outer |= hns3_gen_field_val(HNS3_TXD_L4LEN_M,
3182 (uint8_t)HNS3_NVGRE_HLEN >> HNS3_L4_LEN_UNIT);
3184 inner_l2_len = m->l2_len - HNS3_NVGRE_HLEN;
3187 /* For non UDP / GRE tunneling, drop the tunnel packet */
3191 tmp_inner |= hns3_gen_field_val(HNS3_TXD_L2LEN_M, HNS3_TXD_L2LEN_S,
3192 inner_l2_len >> HNS3_L2_LEN_UNIT);
3193 /* OL2 header size, defined in 2 bytes */
3194 tmp_outer |= hns3_gen_field_val(HNS3_TXD_L2LEN_M, HNS3_TXD_L2LEN_S,
3195 m->outer_l2_len >> HNS3_L2_LEN_UNIT);
3197 *type_cs_vlan_tso_len = tmp_inner;
3198 *ol_type_vlan_len_msec = tmp_outer;
3204 hns3_parse_tunneling_params(struct hns3_tx_queue *txq, struct rte_mbuf *m,
3205 uint16_t tx_desc_id)
3207 struct hns3_desc *tx_ring = txq->tx_ring;
3208 struct hns3_desc *desc = &tx_ring[tx_desc_id];
3209 uint64_t ol_flags = m->ol_flags;
3210 uint32_t tmp_outer = 0;
3211 uint32_t tmp_inner = 0;
3216 * The tunnel header is contained in the inner L2 header field of the
3217 * mbuf, but for hns3 descriptor, it is contained in the outer L4. So,
3218 * there is a need that switching between them. To avoid multiple
3219 * calculations, the length of the L2 header include the outer and
3220 * inner, will be filled during the parsing of tunnel packects.
3222 if (!(ol_flags & PKT_TX_TUNNEL_MASK)) {
3224 * For non tunnel type the tunnel type id is 0, so no need to
3225 * assign a value to it. Only the inner(normal) L2 header length
3228 tmp_inner |= hns3_gen_field_val(HNS3_TXD_L2LEN_M,
3229 HNS3_TXD_L2LEN_S, m->l2_len >> HNS3_L2_LEN_UNIT);
3232 * If outer csum is not offload, the outer length may be filled
3233 * with 0. And the length of the outer header is added to the
3234 * inner l2_len. It would lead a cksum error. So driver has to
3235 * calculate the header length.
3237 if (unlikely(!(ol_flags &
3238 (PKT_TX_OUTER_IP_CKSUM | PKT_TX_OUTER_UDP_CKSUM)) &&
3239 m->outer_l2_len == 0)) {
3240 struct rte_net_hdr_lens hdr_len;
3241 (void)rte_net_get_ptype(m, &hdr_len,
3242 RTE_PTYPE_L2_MASK | RTE_PTYPE_L3_MASK);
3243 m->outer_l3_len = hdr_len.l3_len;
3244 m->outer_l2_len = hdr_len.l2_len;
3245 m->l2_len = m->l2_len - hdr_len.l2_len - hdr_len.l3_len;
3247 hns3_parse_outer_params(m, &tmp_outer);
3248 ret = hns3_parse_inner_params(m, &tmp_outer, &tmp_inner);
3253 desc->tx.ol_type_vlan_len_msec = rte_cpu_to_le_32(tmp_outer);
3254 desc->tx.type_cs_vlan_tso_len = rte_cpu_to_le_32(tmp_inner);
3255 tmp_ol4cs = ol_flags & PKT_TX_OUTER_UDP_CKSUM ?
3256 BIT(HNS3_TXD_OL4CS_B) : 0;
3257 desc->tx.paylen_fd_dop_ol4cs = rte_cpu_to_le_32(tmp_ol4cs);
3263 hns3_parse_l3_cksum_params(struct rte_mbuf *m, uint32_t *type_cs_vlan_tso_len)
3265 uint64_t ol_flags = m->ol_flags;
3269 tmp = *type_cs_vlan_tso_len;
3270 if (ol_flags & PKT_TX_IPV4)
3271 l3_type = HNS3_L3T_IPV4;
3272 else if (ol_flags & PKT_TX_IPV6)
3273 l3_type = HNS3_L3T_IPV6;
3275 l3_type = HNS3_L3T_NONE;
3277 /* inner(/normal) L3 header size, defined in 4 bytes */
3278 tmp |= hns3_gen_field_val(HNS3_TXD_L3LEN_M, HNS3_TXD_L3LEN_S,
3279 m->l3_len >> HNS3_L3_LEN_UNIT);
3281 tmp |= hns3_gen_field_val(HNS3_TXD_L3T_M, HNS3_TXD_L3T_S, l3_type);
3283 /* Enable L3 checksum offloads */
3284 if (ol_flags & PKT_TX_IP_CKSUM)
3285 tmp |= BIT(HNS3_TXD_L3CS_B);
3286 *type_cs_vlan_tso_len = tmp;
3290 hns3_parse_l4_cksum_params(struct rte_mbuf *m, uint32_t *type_cs_vlan_tso_len)
3292 uint64_t ol_flags = m->ol_flags;
3294 /* Enable L4 checksum offloads */
3295 switch (ol_flags & (PKT_TX_L4_MASK | PKT_TX_TCP_SEG)) {
3296 case PKT_TX_TCP_CKSUM | PKT_TX_TCP_SEG:
3297 case PKT_TX_TCP_CKSUM:
3298 case PKT_TX_TCP_SEG:
3299 tmp = *type_cs_vlan_tso_len;
3300 tmp |= hns3_gen_field_val(HNS3_TXD_L4T_M, HNS3_TXD_L4T_S,
3303 case PKT_TX_UDP_CKSUM:
3304 tmp = *type_cs_vlan_tso_len;
3305 tmp |= hns3_gen_field_val(HNS3_TXD_L4T_M, HNS3_TXD_L4T_S,
3308 case PKT_TX_SCTP_CKSUM:
3309 tmp = *type_cs_vlan_tso_len;
3310 tmp |= hns3_gen_field_val(HNS3_TXD_L4T_M, HNS3_TXD_L4T_S,
3316 tmp |= BIT(HNS3_TXD_L4CS_B);
3317 tmp |= hns3_gen_field_val(HNS3_TXD_L4LEN_M, HNS3_TXD_L4LEN_S,
3318 m->l4_len >> HNS3_L4_LEN_UNIT);
3319 *type_cs_vlan_tso_len = tmp;
3323 hns3_txd_enable_checksum(struct hns3_tx_queue *txq, struct rte_mbuf *m,
3324 uint16_t tx_desc_id)
3326 struct hns3_desc *tx_ring = txq->tx_ring;
3327 struct hns3_desc *desc = &tx_ring[tx_desc_id];
3330 hns3_parse_l3_cksum_params(m, &value);
3331 hns3_parse_l4_cksum_params(m, &value);
3333 desc->tx.type_cs_vlan_tso_len |= rte_cpu_to_le_32(value);
3337 hns3_pkt_need_linearized(struct rte_mbuf *tx_pkts, uint32_t bd_num,
3338 uint32_t max_non_tso_bd_num)
3340 struct rte_mbuf *m_first = tx_pkts;
3341 struct rte_mbuf *m_last = tx_pkts;
3342 uint32_t tot_len = 0;
3347 * Hardware requires that the sum of the data length of every 8
3348 * consecutive buffers is greater than MSS in hns3 network engine.
3349 * We simplify it by ensuring pkt_headlen + the first 8 consecutive
3350 * frags greater than gso header len + mss, and the remaining 7
3351 * consecutive frags greater than MSS except the last 7 frags.
3353 if (bd_num <= max_non_tso_bd_num)
3356 for (i = 0; m_last && i < max_non_tso_bd_num - 1;
3357 i++, m_last = m_last->next)
3358 tot_len += m_last->data_len;
3363 /* ensure the first 8 frags is greater than mss + header */
3364 hdr_len = tx_pkts->l2_len + tx_pkts->l3_len + tx_pkts->l4_len;
3365 hdr_len += (tx_pkts->ol_flags & PKT_TX_TUNNEL_MASK) ?
3366 tx_pkts->outer_l2_len + tx_pkts->outer_l3_len : 0;
3367 if (tot_len + m_last->data_len < tx_pkts->tso_segsz + hdr_len)
3371 * ensure the sum of the data length of every 7 consecutive buffer
3372 * is greater than mss except the last one.
3374 for (i = 0; m_last && i < bd_num - max_non_tso_bd_num; i++) {
3375 tot_len -= m_first->data_len;
3376 tot_len += m_last->data_len;
3378 if (tot_len < tx_pkts->tso_segsz)
3381 m_first = m_first->next;
3382 m_last = m_last->next;
3389 hns3_outer_ipv4_cksum_prepared(struct rte_mbuf *m, uint64_t ol_flags,
3392 struct rte_ipv4_hdr *ipv4_hdr;
3393 ipv4_hdr = rte_pktmbuf_mtod_offset(m, struct rte_ipv4_hdr *,
3395 if (ol_flags & PKT_TX_OUTER_IP_CKSUM)
3396 ipv4_hdr->hdr_checksum = 0;
3397 if (ol_flags & PKT_TX_OUTER_UDP_CKSUM) {
3398 struct rte_udp_hdr *udp_hdr;
3400 * If OUTER_UDP_CKSUM is support, HW can caclulate the pseudo
3401 * header for TSO packets
3403 if (ol_flags & PKT_TX_TCP_SEG)
3405 udp_hdr = rte_pktmbuf_mtod_offset(m, struct rte_udp_hdr *,
3406 m->outer_l2_len + m->outer_l3_len);
3407 udp_hdr->dgram_cksum = rte_ipv4_phdr_cksum(ipv4_hdr, ol_flags);
3411 *l4_proto = ipv4_hdr->next_proto_id;
3416 hns3_outer_ipv6_cksum_prepared(struct rte_mbuf *m, uint64_t ol_flags,
3419 struct rte_ipv6_hdr *ipv6_hdr;
3420 ipv6_hdr = rte_pktmbuf_mtod_offset(m, struct rte_ipv6_hdr *,
3422 if (ol_flags & PKT_TX_OUTER_UDP_CKSUM) {
3423 struct rte_udp_hdr *udp_hdr;
3425 * If OUTER_UDP_CKSUM is support, HW can caclulate the pseudo
3426 * header for TSO packets
3428 if (ol_flags & PKT_TX_TCP_SEG)
3430 udp_hdr = rte_pktmbuf_mtod_offset(m, struct rte_udp_hdr *,
3431 m->outer_l2_len + m->outer_l3_len);
3432 udp_hdr->dgram_cksum = rte_ipv6_phdr_cksum(ipv6_hdr, ol_flags);
3436 *l4_proto = ipv6_hdr->proto;
3441 hns3_outer_header_cksum_prepare(struct rte_mbuf *m)
3443 uint64_t ol_flags = m->ol_flags;
3444 uint32_t paylen, hdr_len, l4_proto;
3445 struct rte_udp_hdr *udp_hdr;
3447 if (!(ol_flags & (PKT_TX_OUTER_IPV4 | PKT_TX_OUTER_IPV6)))
3450 if (ol_flags & PKT_TX_OUTER_IPV4) {
3451 if (hns3_outer_ipv4_cksum_prepared(m, ol_flags, &l4_proto))
3454 if (hns3_outer_ipv6_cksum_prepared(m, ol_flags, &l4_proto))
3458 /* driver should ensure the outer udp cksum is 0 for TUNNEL TSO */
3459 if (l4_proto == IPPROTO_UDP && (ol_flags & PKT_TX_TCP_SEG)) {
3460 hdr_len = m->l2_len + m->l3_len + m->l4_len;
3461 hdr_len += m->outer_l2_len + m->outer_l3_len;
3462 paylen = m->pkt_len - hdr_len;
3463 if (paylen <= m->tso_segsz)
3465 udp_hdr = rte_pktmbuf_mtod_offset(m, struct rte_udp_hdr *,
3468 udp_hdr->dgram_cksum = 0;
3473 hns3_check_tso_pkt_valid(struct rte_mbuf *m)
3475 uint32_t tmp_data_len_sum = 0;
3476 uint16_t nb_buf = m->nb_segs;
3477 uint32_t paylen, hdr_len;
3478 struct rte_mbuf *m_seg;
3481 if (nb_buf > HNS3_MAX_TSO_BD_PER_PKT)
3484 hdr_len = m->l2_len + m->l3_len + m->l4_len;
3485 hdr_len += (m->ol_flags & PKT_TX_TUNNEL_MASK) ?
3486 m->outer_l2_len + m->outer_l3_len : 0;
3487 if (hdr_len > HNS3_MAX_TSO_HDR_SIZE)
3490 paylen = m->pkt_len - hdr_len;
3491 if (paylen > HNS3_MAX_BD_PAYLEN)
3495 * The TSO header (include outer and inner L2, L3 and L4 header)
3496 * should be provided by three descriptors in maximum in hns3 network
3500 for (i = 0; m_seg != NULL && i < HNS3_MAX_TSO_HDR_BD_NUM && i < nb_buf;
3501 i++, m_seg = m_seg->next) {
3502 tmp_data_len_sum += m_seg->data_len;
3505 if (hdr_len > tmp_data_len_sum)
3511 #ifdef RTE_LIBRTE_ETHDEV_DEBUG
3513 hns3_vld_vlan_chk(struct hns3_tx_queue *txq, struct rte_mbuf *m)
3515 struct rte_ether_hdr *eh;
3516 struct rte_vlan_hdr *vh;
3518 if (!txq->pvid_sw_shift_en)
3522 * Due to hardware limitations, we only support two-layer VLAN hardware
3523 * offload in Tx direction based on hns3 network engine, so when PVID is
3524 * enabled, QinQ insert is no longer supported.
3525 * And when PVID is enabled, in the following two cases:
3526 * i) packets with more than two VLAN tags.
3527 * ii) packets with one VLAN tag while the hardware VLAN insert is
3529 * The packets will be regarded as abnormal packets and discarded by
3530 * hardware in Tx direction. For debugging purposes, a validation check
3531 * for these types of packets is added to the '.tx_pkt_prepare' ops
3532 * implementation function named hns3_prep_pkts to inform users that
3533 * these packets will be discarded.
3535 if (m->ol_flags & PKT_TX_QINQ_PKT)
3538 eh = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
3539 if (eh->ether_type == rte_cpu_to_be_16(RTE_ETHER_TYPE_VLAN)) {
3540 if (m->ol_flags & PKT_TX_VLAN_PKT)
3543 /* Ensure the incoming packet is not a QinQ packet */
3544 vh = (struct rte_vlan_hdr *)(eh + 1);
3545 if (vh->eth_proto == rte_cpu_to_be_16(RTE_ETHER_TYPE_VLAN))
3554 hns3_udp_cksum_help(struct rte_mbuf *m)
3556 uint64_t ol_flags = m->ol_flags;
3560 if (ol_flags & PKT_TX_IPV4) {
3561 struct rte_ipv4_hdr *ipv4_hdr = rte_pktmbuf_mtod_offset(m,
3562 struct rte_ipv4_hdr *, m->l2_len);
3563 l4_len = rte_be_to_cpu_16(ipv4_hdr->total_length) - m->l3_len;
3565 struct rte_ipv6_hdr *ipv6_hdr = rte_pktmbuf_mtod_offset(m,
3566 struct rte_ipv6_hdr *, m->l2_len);
3567 l4_len = rte_be_to_cpu_16(ipv6_hdr->payload_len);
3570 rte_raw_cksum_mbuf(m, m->l2_len + m->l3_len, l4_len, &cksum);
3574 * RFC 768:If the computed checksum is zero for UDP, it is transmitted
3580 return (uint16_t)cksum;
3584 hns3_validate_tunnel_cksum(struct hns3_tx_queue *tx_queue, struct rte_mbuf *m)
3586 uint64_t ol_flags = m->ol_flags;
3587 struct rte_udp_hdr *udp_hdr;
3590 if (tx_queue->udp_cksum_mode == HNS3_SPECIAL_PORT_HW_CKSUM_MODE ||
3591 ol_flags & PKT_TX_TUNNEL_MASK ||
3592 (ol_flags & PKT_TX_L4_MASK) != PKT_TX_UDP_CKSUM)
3595 * A UDP packet with the same dst_port as VXLAN\VXLAN_GPE\GENEVE will
3596 * be recognized as a tunnel packet in HW. In this case, if UDP CKSUM
3597 * offload is set and the tunnel mask has not been set, the CKSUM will
3598 * be wrong since the header length is wrong and driver should complete
3599 * the CKSUM to avoid CKSUM error.
3601 udp_hdr = rte_pktmbuf_mtod_offset(m, struct rte_udp_hdr *,
3602 m->l2_len + m->l3_len);
3603 dst_port = rte_be_to_cpu_16(udp_hdr->dst_port);
3605 case RTE_VXLAN_DEFAULT_PORT:
3606 case RTE_VXLAN_GPE_DEFAULT_PORT:
3607 case RTE_GENEVE_DEFAULT_PORT:
3608 udp_hdr->dgram_cksum = hns3_udp_cksum_help(m);
3609 m->ol_flags = ol_flags & ~PKT_TX_L4_MASK;
3617 hns3_prep_pkt_proc(struct hns3_tx_queue *tx_queue, struct rte_mbuf *m)
3621 #ifdef RTE_LIBRTE_ETHDEV_DEBUG
3622 ret = rte_validate_tx_offload(m);
3628 ret = hns3_vld_vlan_chk(tx_queue, m);
3634 if (hns3_pkt_is_tso(m)) {
3635 if (hns3_pkt_need_linearized(m, m->nb_segs,
3636 tx_queue->max_non_tso_bd_num) ||
3637 hns3_check_tso_pkt_valid(m)) {
3642 if (tx_queue->tso_mode != HNS3_TSO_SW_CAL_PSEUDO_H_CSUM) {
3644 * (tso mode != HNS3_TSO_SW_CAL_PSEUDO_H_CSUM) means
3645 * hardware support recalculate the TCP pseudo header
3646 * checksum of packets that need TSO, so network driver
3647 * software not need to recalculate it.
3649 hns3_outer_header_cksum_prepare(m);
3654 ret = rte_net_intel_cksum_prepare(m);
3660 if (!hns3_validate_tunnel_cksum(tx_queue, m))
3663 hns3_outer_header_cksum_prepare(m);
3669 hns3_prep_pkts(__rte_unused void *tx_queue, struct rte_mbuf **tx_pkts,
3675 for (i = 0; i < nb_pkts; i++) {
3677 if (hns3_prep_pkt_proc(tx_queue, m))
3685 hns3_parse_cksum(struct hns3_tx_queue *txq, uint16_t tx_desc_id,
3688 struct hns3_desc *tx_ring = txq->tx_ring;
3689 struct hns3_desc *desc = &tx_ring[tx_desc_id];
3691 /* Enable checksum offloading */
3692 if (m->ol_flags & HNS3_TX_CKSUM_OFFLOAD_MASK) {
3693 /* Fill in tunneling parameters if necessary */
3694 if (hns3_parse_tunneling_params(txq, m, tx_desc_id)) {
3695 txq->dfx_stats.unsupported_tunnel_pkt_cnt++;
3699 hns3_txd_enable_checksum(txq, m, tx_desc_id);
3701 /* clear the control bit */
3702 desc->tx.type_cs_vlan_tso_len = 0;
3703 desc->tx.ol_type_vlan_len_msec = 0;
3710 hns3_check_non_tso_pkt(uint16_t nb_buf, struct rte_mbuf **m_seg,
3711 struct rte_mbuf *tx_pkt, struct hns3_tx_queue *txq)
3713 uint8_t max_non_tso_bd_num;
3714 struct rte_mbuf *new_pkt;
3717 if (hns3_pkt_is_tso(*m_seg))
3721 * If packet length is greater than HNS3_MAX_FRAME_LEN
3722 * driver support, the packet will be ignored.
3724 if (unlikely(rte_pktmbuf_pkt_len(tx_pkt) > HNS3_MAX_FRAME_LEN)) {
3725 txq->dfx_stats.over_length_pkt_cnt++;
3729 max_non_tso_bd_num = txq->max_non_tso_bd_num;
3730 if (unlikely(nb_buf > max_non_tso_bd_num)) {
3731 txq->dfx_stats.exceed_limit_bd_pkt_cnt++;
3732 ret = hns3_reassemble_tx_pkts(tx_pkt, &new_pkt,
3733 max_non_tso_bd_num);
3735 txq->dfx_stats.exceed_limit_bd_reassem_fail++;
3745 hns3_tx_free_buffer_simple(struct hns3_tx_queue *txq)
3747 struct hns3_entry *tx_entry;
3748 struct hns3_desc *desc;
3749 uint16_t tx_next_clean;
3753 if (HNS3_GET_TX_QUEUE_PEND_BD_NUM(txq) < txq->tx_rs_thresh)
3757 * All mbufs can be released only when the VLD bits of all
3758 * descriptors in a batch are cleared.
3760 tx_next_clean = (txq->next_to_clean + txq->tx_rs_thresh - 1) %
3762 desc = &txq->tx_ring[tx_next_clean];
3763 for (i = 0; i < txq->tx_rs_thresh; i++) {
3764 if (rte_le_to_cpu_16(desc->tx.tp_fe_sc_vld_ra_ri) &
3765 BIT(HNS3_TXD_VLD_B))
3770 tx_entry = &txq->sw_ring[txq->next_to_clean];
3772 for (i = 0; i < txq->tx_rs_thresh; i++)
3773 rte_prefetch0((tx_entry + i)->mbuf);
3774 for (i = 0; i < txq->tx_rs_thresh; i++, tx_entry++) {
3775 rte_mempool_put(tx_entry->mbuf->pool, tx_entry->mbuf);
3776 tx_entry->mbuf = NULL;
3779 txq->next_to_clean = (tx_next_clean + 1) % txq->nb_tx_desc;
3780 txq->tx_bd_ready += txq->tx_rs_thresh;
3785 hns3_tx_backup_1mbuf(struct hns3_entry *tx_entry, struct rte_mbuf **pkts)
3787 tx_entry->mbuf = pkts[0];
3791 hns3_tx_backup_4mbuf(struct hns3_entry *tx_entry, struct rte_mbuf **pkts)
3793 hns3_tx_backup_1mbuf(&tx_entry[0], &pkts[0]);
3794 hns3_tx_backup_1mbuf(&tx_entry[1], &pkts[1]);
3795 hns3_tx_backup_1mbuf(&tx_entry[2], &pkts[2]);
3796 hns3_tx_backup_1mbuf(&tx_entry[3], &pkts[3]);
3800 hns3_tx_setup_4bd(struct hns3_desc *txdp, struct rte_mbuf **pkts)
3802 #define PER_LOOP_NUM 4
3803 const uint16_t bd_flag = BIT(HNS3_TXD_VLD_B) | BIT(HNS3_TXD_FE_B);
3807 for (i = 0; i < PER_LOOP_NUM; i++, txdp++, pkts++) {
3808 dma_addr = rte_mbuf_data_iova(*pkts);
3809 txdp->addr = rte_cpu_to_le_64(dma_addr);
3810 txdp->tx.send_size = rte_cpu_to_le_16((*pkts)->data_len);
3811 txdp->tx.paylen_fd_dop_ol4cs = 0;
3812 txdp->tx.type_cs_vlan_tso_len = 0;
3813 txdp->tx.ol_type_vlan_len_msec = 0;
3814 txdp->tx.tp_fe_sc_vld_ra_ri = rte_cpu_to_le_16(bd_flag);
3819 hns3_tx_setup_1bd(struct hns3_desc *txdp, struct rte_mbuf **pkts)
3821 const uint16_t bd_flag = BIT(HNS3_TXD_VLD_B) | BIT(HNS3_TXD_FE_B);
3824 dma_addr = rte_mbuf_data_iova(*pkts);
3825 txdp->addr = rte_cpu_to_le_64(dma_addr);
3826 txdp->tx.send_size = rte_cpu_to_le_16((*pkts)->data_len);
3827 txdp->tx.paylen_fd_dop_ol4cs = 0;
3828 txdp->tx.type_cs_vlan_tso_len = 0;
3829 txdp->tx.ol_type_vlan_len_msec = 0;
3830 txdp->tx.tp_fe_sc_vld_ra_ri = rte_cpu_to_le_16(bd_flag);
3834 hns3_tx_fill_hw_ring(struct hns3_tx_queue *txq,
3835 struct rte_mbuf **pkts,
3838 #define PER_LOOP_NUM 4
3839 #define PER_LOOP_MASK (PER_LOOP_NUM - 1)
3840 struct hns3_desc *txdp = &txq->tx_ring[txq->next_to_use];
3841 struct hns3_entry *tx_entry = &txq->sw_ring[txq->next_to_use];
3842 const uint32_t mainpart = (nb_pkts & ((uint32_t)~PER_LOOP_MASK));
3843 const uint32_t leftover = (nb_pkts & ((uint32_t)PER_LOOP_MASK));
3846 for (i = 0; i < mainpart; i += PER_LOOP_NUM) {
3847 hns3_tx_backup_4mbuf(tx_entry + i, pkts + i);
3848 hns3_tx_setup_4bd(txdp + i, pkts + i);
3850 /* Increment bytes counter */
3852 for (j = 0; j < PER_LOOP_NUM; j++)
3853 txq->basic_stats.bytes += pkts[i + j]->pkt_len;
3855 if (unlikely(leftover > 0)) {
3856 for (i = 0; i < leftover; i++) {
3857 hns3_tx_backup_1mbuf(tx_entry + mainpart + i,
3858 pkts + mainpart + i);
3859 hns3_tx_setup_1bd(txdp + mainpart + i,
3860 pkts + mainpart + i);
3862 /* Increment bytes counter */
3863 txq->basic_stats.bytes += pkts[mainpart + i]->pkt_len;
3869 hns3_xmit_pkts_simple(void *tx_queue,
3870 struct rte_mbuf **tx_pkts,
3873 struct hns3_tx_queue *txq = tx_queue;
3876 hns3_tx_free_buffer_simple(txq);
3878 nb_pkts = RTE_MIN(txq->tx_bd_ready, nb_pkts);
3879 if (unlikely(nb_pkts == 0)) {
3880 if (txq->tx_bd_ready == 0)
3881 txq->dfx_stats.queue_full_cnt++;
3885 txq->tx_bd_ready -= nb_pkts;
3886 if (txq->next_to_use + nb_pkts > txq->nb_tx_desc) {
3887 nb_tx = txq->nb_tx_desc - txq->next_to_use;
3888 hns3_tx_fill_hw_ring(txq, tx_pkts, nb_tx);
3889 txq->next_to_use = 0;
3892 hns3_tx_fill_hw_ring(txq, tx_pkts + nb_tx, nb_pkts - nb_tx);
3893 txq->next_to_use += nb_pkts - nb_tx;
3895 hns3_write_reg_opt(txq->io_tail_reg, nb_pkts);
3901 hns3_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
3903 struct hns3_tx_queue *txq = tx_queue;
3904 struct hns3_entry *tx_bak_pkt;
3905 struct hns3_desc *tx_ring;
3906 struct rte_mbuf *tx_pkt;
3907 struct rte_mbuf *m_seg;
3908 struct hns3_desc *desc;
3909 uint32_t nb_hold = 0;
3910 uint16_t tx_next_use;
3911 uint16_t tx_pkt_num;
3917 /* free useless buffer */
3918 hns3_tx_free_useless_buffer(txq);
3920 tx_next_use = txq->next_to_use;
3921 tx_bd_max = txq->nb_tx_desc;
3922 tx_pkt_num = nb_pkts;
3923 tx_ring = txq->tx_ring;
3926 tx_bak_pkt = &txq->sw_ring[tx_next_use];
3927 for (nb_tx = 0; nb_tx < tx_pkt_num; nb_tx++) {
3928 tx_pkt = *tx_pkts++;
3930 nb_buf = tx_pkt->nb_segs;
3932 if (nb_buf > txq->tx_bd_ready) {
3933 txq->dfx_stats.queue_full_cnt++;
3941 * If packet length is less than minimum packet length supported
3942 * by hardware in Tx direction, driver need to pad it to avoid
3945 if (unlikely(rte_pktmbuf_pkt_len(tx_pkt) <
3946 txq->min_tx_pkt_len)) {
3950 add_len = txq->min_tx_pkt_len -
3951 rte_pktmbuf_pkt_len(tx_pkt);
3952 appended = rte_pktmbuf_append(tx_pkt, add_len);
3953 if (appended == NULL) {
3954 txq->dfx_stats.pkt_padding_fail_cnt++;
3958 memset(appended, 0, add_len);
3963 if (hns3_check_non_tso_pkt(nb_buf, &m_seg, tx_pkt, txq))
3966 if (hns3_parse_cksum(txq, tx_next_use, m_seg))
3970 desc = &tx_ring[tx_next_use];
3973 * If the packet is divided into multiple Tx Buffer Descriptors,
3974 * only need to fill vlan, paylen and tso into the first Tx
3975 * Buffer Descriptor.
3977 hns3_fill_first_desc(txq, desc, m_seg);
3980 desc = &tx_ring[tx_next_use];
3982 * Fill valid bits, DMA address and data length for each
3983 * Tx Buffer Descriptor.
3985 hns3_fill_per_desc(desc, m_seg);
3986 tx_bak_pkt->mbuf = m_seg;
3987 m_seg = m_seg->next;
3990 if (tx_next_use >= tx_bd_max) {
3992 tx_bak_pkt = txq->sw_ring;
3996 } while (m_seg != NULL);
3998 /* Add end flag for the last Tx Buffer Descriptor */
3999 desc->tx.tp_fe_sc_vld_ra_ri |=
4000 rte_cpu_to_le_16(BIT(HNS3_TXD_FE_B));
4002 /* Increment bytes counter */
4003 txq->basic_stats.bytes += tx_pkt->pkt_len;
4005 txq->next_to_use = tx_next_use;
4006 txq->tx_bd_ready -= i;
4012 hns3_write_reg_opt(txq->io_tail_reg, nb_hold);
4018 hns3_tx_check_vec_support(__rte_unused struct rte_eth_dev *dev)
4024 hns3_xmit_pkts_vec(__rte_unused void *tx_queue,
4025 __rte_unused struct rte_mbuf **tx_pkts,
4026 __rte_unused uint16_t nb_pkts)
4032 hns3_xmit_pkts_vec_sve(void __rte_unused * tx_queue,
4033 struct rte_mbuf __rte_unused **tx_pkts,
4034 uint16_t __rte_unused nb_pkts)
4040 hns3_tx_burst_mode_get(struct rte_eth_dev *dev, __rte_unused uint16_t queue_id,
4041 struct rte_eth_burst_mode *mode)
4043 eth_tx_burst_t pkt_burst = dev->tx_pkt_burst;
4044 const char *info = NULL;
4046 if (pkt_burst == hns3_xmit_pkts_simple)
4047 info = "Scalar Simple";
4048 else if (pkt_burst == hns3_xmit_pkts)
4050 else if (pkt_burst == hns3_xmit_pkts_vec)
4051 info = "Vector Neon";
4052 else if (pkt_burst == hns3_xmit_pkts_vec_sve)
4053 info = "Vector Sve";
4058 snprintf(mode->info, sizeof(mode->info), "%s", info);
4063 static eth_tx_burst_t
4064 hns3_get_tx_function(struct rte_eth_dev *dev, eth_tx_prep_t *prep)
4066 uint64_t offloads = dev->data->dev_conf.txmode.offloads;
4067 struct hns3_adapter *hns = dev->data->dev_private;
4068 bool vec_allowed, sve_allowed, simple_allowed;
4070 vec_allowed = hns->tx_vec_allowed &&
4071 hns3_tx_check_vec_support(dev) == 0;
4072 sve_allowed = vec_allowed && hns3_check_sve_support();
4073 simple_allowed = hns->tx_simple_allowed &&
4074 offloads == (offloads & DEV_TX_OFFLOAD_MBUF_FAST_FREE);
4078 if (hns->tx_func_hint == HNS3_IO_FUNC_HINT_VEC && vec_allowed)
4079 return hns3_xmit_pkts_vec;
4080 if (hns->tx_func_hint == HNS3_IO_FUNC_HINT_SVE && sve_allowed)
4081 return hns3_xmit_pkts_vec_sve;
4082 if (hns->tx_func_hint == HNS3_IO_FUNC_HINT_SIMPLE && simple_allowed)
4083 return hns3_xmit_pkts_simple;
4084 if (hns->tx_func_hint == HNS3_IO_FUNC_HINT_COMMON) {
4085 *prep = hns3_prep_pkts;
4086 return hns3_xmit_pkts;
4090 return hns3_xmit_pkts_vec;
4092 return hns3_xmit_pkts_simple;
4094 *prep = hns3_prep_pkts;
4095 return hns3_xmit_pkts;
4099 hns3_dummy_rxtx_burst(void *dpdk_txq __rte_unused,
4100 struct rte_mbuf **pkts __rte_unused,
4101 uint16_t pkts_n __rte_unused)
4106 void hns3_set_rxtx_function(struct rte_eth_dev *eth_dev)
4108 struct hns3_adapter *hns = eth_dev->data->dev_private;
4109 eth_tx_prep_t prep = NULL;
4111 if (hns->hw.adapter_state == HNS3_NIC_STARTED &&
4112 __atomic_load_n(&hns->hw.reset.resetting, __ATOMIC_RELAXED) == 0) {
4113 eth_dev->rx_pkt_burst = hns3_get_rx_function(eth_dev);
4114 eth_dev->rx_descriptor_status = hns3_dev_rx_descriptor_status;
4115 eth_dev->tx_pkt_burst = hns3_get_tx_function(eth_dev, &prep);
4116 eth_dev->tx_pkt_prepare = prep;
4117 eth_dev->tx_descriptor_status = hns3_dev_tx_descriptor_status;
4119 eth_dev->rx_pkt_burst = hns3_dummy_rxtx_burst;
4120 eth_dev->tx_pkt_burst = hns3_dummy_rxtx_burst;
4121 eth_dev->tx_pkt_prepare = hns3_dummy_rxtx_burst;
4126 hns3_rxq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
4127 struct rte_eth_rxq_info *qinfo)
4129 struct hns3_rx_queue *rxq = dev->data->rx_queues[queue_id];
4131 qinfo->mp = rxq->mb_pool;
4132 qinfo->nb_desc = rxq->nb_rx_desc;
4133 qinfo->scattered_rx = dev->data->scattered_rx;
4134 /* Report the HW Rx buffer length to user */
4135 qinfo->rx_buf_size = rxq->rx_buf_len;
4138 * If there are no available Rx buffer descriptors, incoming packets
4139 * are always dropped by hardware based on hns3 network engine.
4141 qinfo->conf.rx_drop_en = 1;
4142 qinfo->conf.offloads = dev->data->dev_conf.rxmode.offloads;
4143 qinfo->conf.rx_free_thresh = rxq->rx_free_thresh;
4144 qinfo->conf.rx_deferred_start = rxq->rx_deferred_start;
4148 hns3_txq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
4149 struct rte_eth_txq_info *qinfo)
4151 struct hns3_tx_queue *txq = dev->data->tx_queues[queue_id];
4153 qinfo->nb_desc = txq->nb_tx_desc;
4154 qinfo->conf.offloads = dev->data->dev_conf.txmode.offloads;
4155 qinfo->conf.tx_rs_thresh = txq->tx_rs_thresh;
4156 qinfo->conf.tx_free_thresh = txq->tx_free_thresh;
4157 qinfo->conf.tx_deferred_start = txq->tx_deferred_start;
4161 hns3_dev_rx_queue_start(struct rte_eth_dev *dev, uint16_t rx_queue_id)
4163 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
4164 struct hns3_rx_queue *rxq = dev->data->rx_queues[rx_queue_id];
4165 struct hns3_adapter *hns = HNS3_DEV_HW_TO_ADAPTER(hw);
4168 if (!hns3_dev_indep_txrx_supported(hw))
4171 ret = hns3_reset_queue(hw, rx_queue_id, HNS3_RING_TYPE_RX);
4173 hns3_err(hw, "fail to reset Rx queue %u, ret = %d.",
4178 ret = hns3_init_rxq(hns, rx_queue_id);
4180 hns3_err(hw, "fail to init Rx queue %u, ret = %d.",
4185 hns3_enable_rxq(rxq, true);
4186 dev->data->rx_queue_state[rx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
4192 hns3_reset_sw_rxq(struct hns3_rx_queue *rxq)
4194 rxq->next_to_use = 0;
4195 rxq->rx_rearm_start = 0;
4196 rxq->rx_free_hold = 0;
4197 rxq->rx_rearm_nb = 0;
4198 rxq->pkt_first_seg = NULL;
4199 rxq->pkt_last_seg = NULL;
4200 memset(&rxq->rx_ring[0], 0, rxq->nb_rx_desc * sizeof(struct hns3_desc));
4201 hns3_rxq_vec_setup(rxq);
4205 hns3_dev_rx_queue_stop(struct rte_eth_dev *dev, uint16_t rx_queue_id)
4207 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
4208 struct hns3_rx_queue *rxq = dev->data->rx_queues[rx_queue_id];
4210 if (!hns3_dev_indep_txrx_supported(hw))
4213 hns3_enable_rxq(rxq, false);
4215 hns3_rx_queue_release_mbufs(rxq);
4217 hns3_reset_sw_rxq(rxq);
4218 dev->data->rx_queue_state[rx_queue_id] = RTE_ETH_QUEUE_STATE_STOPPED;
4224 hns3_dev_tx_queue_start(struct rte_eth_dev *dev, uint16_t tx_queue_id)
4226 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
4227 struct hns3_tx_queue *txq = dev->data->tx_queues[tx_queue_id];
4230 if (!hns3_dev_indep_txrx_supported(hw))
4233 ret = hns3_reset_queue(hw, tx_queue_id, HNS3_RING_TYPE_TX);
4235 hns3_err(hw, "fail to reset Tx queue %u, ret = %d.",
4241 hns3_enable_txq(txq, true);
4242 dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
4248 hns3_dev_tx_queue_stop(struct rte_eth_dev *dev, uint16_t tx_queue_id)
4250 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
4251 struct hns3_tx_queue *txq = dev->data->tx_queues[tx_queue_id];
4253 if (!hns3_dev_indep_txrx_supported(hw))
4256 hns3_enable_txq(txq, false);
4257 hns3_tx_queue_release_mbufs(txq);
4259 * All the mbufs in sw_ring are released and all the pointers in sw_ring
4260 * are set to NULL. If this queue is still called by upper layer,
4261 * residual SW status of this txq may cause these pointers in sw_ring
4262 * which have been set to NULL to be released again. To avoid it,
4266 dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STOPPED;
4272 hns3_tx_done_cleanup_full(struct hns3_tx_queue *txq, uint32_t free_cnt)
4274 uint16_t next_to_clean = txq->next_to_clean;
4275 uint16_t next_to_use = txq->next_to_use;
4276 uint16_t tx_bd_ready = txq->tx_bd_ready;
4277 struct hns3_entry *tx_pkt = &txq->sw_ring[next_to_clean];
4278 struct hns3_desc *desc = &txq->tx_ring[next_to_clean];
4281 if (free_cnt == 0 || free_cnt > txq->nb_tx_desc)
4282 free_cnt = txq->nb_tx_desc;
4284 for (idx = 0; idx < free_cnt; idx++) {
4285 if (next_to_clean == next_to_use)
4288 if (desc->tx.tp_fe_sc_vld_ra_ri &
4289 rte_cpu_to_le_16(BIT(HNS3_TXD_VLD_B)))
4292 if (tx_pkt->mbuf != NULL) {
4293 rte_pktmbuf_free_seg(tx_pkt->mbuf);
4294 tx_pkt->mbuf = NULL;
4301 if (next_to_clean == txq->nb_tx_desc) {
4302 tx_pkt = txq->sw_ring;
4303 desc = txq->tx_ring;
4309 txq->next_to_clean = next_to_clean;
4310 txq->tx_bd_ready = tx_bd_ready;
4317 hns3_tx_done_cleanup(void *txq, uint32_t free_cnt)
4319 struct hns3_tx_queue *q = (struct hns3_tx_queue *)txq;
4320 struct rte_eth_dev *dev = &rte_eth_devices[q->port_id];
4322 if (dev->tx_pkt_burst == hns3_xmit_pkts)
4323 return hns3_tx_done_cleanup_full(q, free_cnt);
4324 else if (dev->tx_pkt_burst == hns3_dummy_rxtx_burst)
4331 hns3_dev_rx_descriptor_status(void *rx_queue, uint16_t offset)
4333 volatile struct hns3_desc *rxdp;
4334 struct hns3_rx_queue *rxq;
4335 struct rte_eth_dev *dev;
4336 uint32_t bd_base_info;
4339 rxq = (struct hns3_rx_queue *)rx_queue;
4340 if (offset >= rxq->nb_rx_desc)
4343 desc_id = (rxq->next_to_use + offset) % rxq->nb_rx_desc;
4344 rxdp = &rxq->rx_ring[desc_id];
4345 bd_base_info = rte_le_to_cpu_32(rxdp->rx.bd_base_info);
4346 dev = &rte_eth_devices[rxq->port_id];
4347 if (dev->rx_pkt_burst == hns3_recv_pkts ||
4348 dev->rx_pkt_burst == hns3_recv_scattered_pkts) {
4349 if (offset >= rxq->nb_rx_desc - rxq->rx_free_hold)
4350 return RTE_ETH_RX_DESC_UNAVAIL;
4351 } else if (dev->rx_pkt_burst == hns3_recv_pkts_vec ||
4352 dev->rx_pkt_burst == hns3_recv_pkts_vec_sve){
4353 if (offset >= rxq->nb_rx_desc - rxq->rx_rearm_nb)
4354 return RTE_ETH_RX_DESC_UNAVAIL;
4356 return RTE_ETH_RX_DESC_UNAVAIL;
4359 if (!(bd_base_info & BIT(HNS3_RXD_VLD_B)))
4360 return RTE_ETH_RX_DESC_AVAIL;
4362 return RTE_ETH_RX_DESC_DONE;
4366 hns3_dev_tx_descriptor_status(void *tx_queue, uint16_t offset)
4368 volatile struct hns3_desc *txdp;
4369 struct hns3_tx_queue *txq;
4370 struct rte_eth_dev *dev;
4373 txq = (struct hns3_tx_queue *)tx_queue;
4374 if (offset >= txq->nb_tx_desc)
4377 dev = &rte_eth_devices[txq->port_id];
4378 if (dev->tx_pkt_burst != hns3_xmit_pkts_simple &&
4379 dev->tx_pkt_burst != hns3_xmit_pkts &&
4380 dev->tx_pkt_burst != hns3_xmit_pkts_vec_sve &&
4381 dev->tx_pkt_burst != hns3_xmit_pkts_vec)
4382 return RTE_ETH_TX_DESC_UNAVAIL;
4384 desc_id = (txq->next_to_use + offset) % txq->nb_tx_desc;
4385 txdp = &txq->tx_ring[desc_id];
4386 if (txdp->tx.tp_fe_sc_vld_ra_ri & rte_cpu_to_le_16(BIT(HNS3_TXD_VLD_B)))
4387 return RTE_ETH_TX_DESC_FULL;
4389 return RTE_ETH_TX_DESC_DONE;
4393 hns3_rx_queue_count(struct rte_eth_dev *dev, uint16_t rx_queue_id)
4396 * Number of BDs that have been processed by the driver
4397 * but have not been notified to the hardware.
4399 uint32_t driver_hold_bd_num;
4400 struct hns3_rx_queue *rxq;
4403 rxq = dev->data->rx_queues[rx_queue_id];
4404 fbd_num = hns3_read_dev(rxq, HNS3_RING_RX_FBDNUM_REG);
4405 if (dev->rx_pkt_burst == hns3_recv_pkts_vec ||
4406 dev->rx_pkt_burst == hns3_recv_pkts_vec_sve)
4407 driver_hold_bd_num = rxq->rx_rearm_nb;
4409 driver_hold_bd_num = rxq->rx_free_hold;
4411 if (fbd_num <= driver_hold_bd_num)
4414 return fbd_num - driver_hold_bd_num;
4418 hns3_enable_rxd_adv_layout(struct hns3_hw *hw)
4421 * If the hardware support rxd advanced layout, then driver enable it
4424 if (hns3_dev_rxd_adv_layout_supported(hw))
4425 hns3_write_dev(hw, HNS3_RXD_ADV_LAYOUT_EN_REG, 1);