1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2018-2019 Hisilicon Limited.
11 #include <rte_bus_pci.h>
12 #include <rte_byteorder.h>
13 #include <rte_common.h>
14 #include <rte_cycles.h>
17 #include <rte_ether.h>
18 #include <rte_vxlan.h>
19 #include <rte_ethdev_driver.h>
24 #include <rte_malloc.h>
26 #if defined(RTE_ARCH_ARM64) && defined(CC_SVE_SUPPORT)
27 #include <rte_cpuflags.h>
30 #include "hns3_ethdev.h"
31 #include "hns3_rxtx.h"
32 #include "hns3_regs.h"
33 #include "hns3_logs.h"
35 #define HNS3_CFG_DESC_NUM(num) ((num) / 8 - 1)
36 #define HNS3_RX_RING_PREFETCTH_MASK 3
39 hns3_rx_queue_release_mbufs(struct hns3_rx_queue *rxq)
43 /* Note: Fake rx queue will not enter here */
44 if (rxq->sw_ring == NULL)
47 if (rxq->rx_rearm_nb == 0) {
48 for (i = 0; i < rxq->nb_rx_desc; i++) {
49 if (rxq->sw_ring[i].mbuf != NULL) {
50 rte_pktmbuf_free_seg(rxq->sw_ring[i].mbuf);
51 rxq->sw_ring[i].mbuf = NULL;
55 for (i = rxq->next_to_use;
56 i != rxq->rx_rearm_start;
57 i = (i + 1) % rxq->nb_rx_desc) {
58 if (rxq->sw_ring[i].mbuf != NULL) {
59 rte_pktmbuf_free_seg(rxq->sw_ring[i].mbuf);
60 rxq->sw_ring[i].mbuf = NULL;
65 for (i = 0; i < rxq->bulk_mbuf_num; i++)
66 rte_pktmbuf_free_seg(rxq->bulk_mbuf[i]);
67 rxq->bulk_mbuf_num = 0;
69 if (rxq->pkt_first_seg) {
70 rte_pktmbuf_free(rxq->pkt_first_seg);
71 rxq->pkt_first_seg = NULL;
76 hns3_tx_queue_release_mbufs(struct hns3_tx_queue *txq)
80 /* Note: Fake tx queue will not enter here */
82 for (i = 0; i < txq->nb_tx_desc; i++) {
83 if (txq->sw_ring[i].mbuf) {
84 rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
85 txq->sw_ring[i].mbuf = NULL;
92 hns3_rx_queue_release(void *queue)
94 struct hns3_rx_queue *rxq = queue;
96 hns3_rx_queue_release_mbufs(rxq);
98 rte_memzone_free(rxq->mz);
100 rte_free(rxq->sw_ring);
106 hns3_tx_queue_release(void *queue)
108 struct hns3_tx_queue *txq = queue;
110 hns3_tx_queue_release_mbufs(txq);
112 rte_memzone_free(txq->mz);
114 rte_free(txq->sw_ring);
122 hns3_dev_rx_queue_release(void *queue)
124 struct hns3_rx_queue *rxq = queue;
125 struct hns3_adapter *hns;
131 rte_spinlock_lock(&hns->hw.lock);
132 hns3_rx_queue_release(queue);
133 rte_spinlock_unlock(&hns->hw.lock);
137 hns3_dev_tx_queue_release(void *queue)
139 struct hns3_tx_queue *txq = queue;
140 struct hns3_adapter *hns;
146 rte_spinlock_lock(&hns->hw.lock);
147 hns3_tx_queue_release(queue);
148 rte_spinlock_unlock(&hns->hw.lock);
152 hns3_fake_rx_queue_release(struct hns3_rx_queue *queue)
154 struct hns3_rx_queue *rxq = queue;
155 struct hns3_adapter *hns;
165 if (hw->fkq_data.rx_queues[idx]) {
166 hns3_rx_queue_release(hw->fkq_data.rx_queues[idx]);
167 hw->fkq_data.rx_queues[idx] = NULL;
170 /* free fake rx queue arrays */
171 if (idx == (hw->fkq_data.nb_fake_rx_queues - 1)) {
172 hw->fkq_data.nb_fake_rx_queues = 0;
173 rte_free(hw->fkq_data.rx_queues);
174 hw->fkq_data.rx_queues = NULL;
179 hns3_fake_tx_queue_release(struct hns3_tx_queue *queue)
181 struct hns3_tx_queue *txq = queue;
182 struct hns3_adapter *hns;
192 if (hw->fkq_data.tx_queues[idx]) {
193 hns3_tx_queue_release(hw->fkq_data.tx_queues[idx]);
194 hw->fkq_data.tx_queues[idx] = NULL;
197 /* free fake tx queue arrays */
198 if (idx == (hw->fkq_data.nb_fake_tx_queues - 1)) {
199 hw->fkq_data.nb_fake_tx_queues = 0;
200 rte_free(hw->fkq_data.tx_queues);
201 hw->fkq_data.tx_queues = NULL;
206 hns3_free_rx_queues(struct rte_eth_dev *dev)
208 struct hns3_adapter *hns = dev->data->dev_private;
209 struct hns3_fake_queue_data *fkq_data;
210 struct hns3_hw *hw = &hns->hw;
214 nb_rx_q = hw->data->nb_rx_queues;
215 for (i = 0; i < nb_rx_q; i++) {
216 if (dev->data->rx_queues[i]) {
217 hns3_rx_queue_release(dev->data->rx_queues[i]);
218 dev->data->rx_queues[i] = NULL;
222 /* Free fake Rx queues */
223 fkq_data = &hw->fkq_data;
224 for (i = 0; i < fkq_data->nb_fake_rx_queues; i++) {
225 if (fkq_data->rx_queues[i])
226 hns3_fake_rx_queue_release(fkq_data->rx_queues[i]);
231 hns3_free_tx_queues(struct rte_eth_dev *dev)
233 struct hns3_adapter *hns = dev->data->dev_private;
234 struct hns3_fake_queue_data *fkq_data;
235 struct hns3_hw *hw = &hns->hw;
239 nb_tx_q = hw->data->nb_tx_queues;
240 for (i = 0; i < nb_tx_q; i++) {
241 if (dev->data->tx_queues[i]) {
242 hns3_tx_queue_release(dev->data->tx_queues[i]);
243 dev->data->tx_queues[i] = NULL;
247 /* Free fake Tx queues */
248 fkq_data = &hw->fkq_data;
249 for (i = 0; i < fkq_data->nb_fake_tx_queues; i++) {
250 if (fkq_data->tx_queues[i])
251 hns3_fake_tx_queue_release(fkq_data->tx_queues[i]);
256 hns3_free_all_queues(struct rte_eth_dev *dev)
258 hns3_free_rx_queues(dev);
259 hns3_free_tx_queues(dev);
263 hns3_alloc_rx_queue_mbufs(struct hns3_hw *hw, struct hns3_rx_queue *rxq)
265 struct rte_mbuf *mbuf;
269 for (i = 0; i < rxq->nb_rx_desc; i++) {
270 mbuf = rte_mbuf_raw_alloc(rxq->mb_pool);
271 if (unlikely(mbuf == NULL)) {
272 hns3_err(hw, "Failed to allocate RXD[%d] for rx queue!",
274 hns3_rx_queue_release_mbufs(rxq);
278 rte_mbuf_refcnt_set(mbuf, 1);
280 mbuf->data_off = RTE_PKTMBUF_HEADROOM;
282 mbuf->port = rxq->port_id;
284 rxq->sw_ring[i].mbuf = mbuf;
285 dma_addr = rte_cpu_to_le_64(rte_mbuf_data_iova_default(mbuf));
286 rxq->rx_ring[i].addr = dma_addr;
287 rxq->rx_ring[i].rx.bd_base_info = 0;
294 hns3_buf_size2type(uint32_t buf_size)
300 bd_size_type = HNS3_BD_SIZE_512_TYPE;
303 bd_size_type = HNS3_BD_SIZE_1024_TYPE;
306 bd_size_type = HNS3_BD_SIZE_4096_TYPE;
309 bd_size_type = HNS3_BD_SIZE_2048_TYPE;
316 hns3_init_rx_queue_hw(struct hns3_rx_queue *rxq)
318 uint32_t rx_buf_len = rxq->rx_buf_len;
319 uint64_t dma_addr = rxq->rx_ring_phys_addr;
321 hns3_write_dev(rxq, HNS3_RING_RX_BASEADDR_L_REG, (uint32_t)dma_addr);
322 hns3_write_dev(rxq, HNS3_RING_RX_BASEADDR_H_REG,
323 (uint32_t)((dma_addr >> 31) >> 1));
325 hns3_write_dev(rxq, HNS3_RING_RX_BD_LEN_REG,
326 hns3_buf_size2type(rx_buf_len));
327 hns3_write_dev(rxq, HNS3_RING_RX_BD_NUM_REG,
328 HNS3_CFG_DESC_NUM(rxq->nb_rx_desc));
332 hns3_init_tx_queue_hw(struct hns3_tx_queue *txq)
334 uint64_t dma_addr = txq->tx_ring_phys_addr;
336 hns3_write_dev(txq, HNS3_RING_TX_BASEADDR_L_REG, (uint32_t)dma_addr);
337 hns3_write_dev(txq, HNS3_RING_TX_BASEADDR_H_REG,
338 (uint32_t)((dma_addr >> 31) >> 1));
340 hns3_write_dev(txq, HNS3_RING_TX_BD_NUM_REG,
341 HNS3_CFG_DESC_NUM(txq->nb_tx_desc));
345 hns3_update_all_queues_pvid_proc_en(struct hns3_hw *hw)
347 uint16_t nb_rx_q = hw->data->nb_rx_queues;
348 uint16_t nb_tx_q = hw->data->nb_tx_queues;
349 struct hns3_rx_queue *rxq;
350 struct hns3_tx_queue *txq;
354 pvid_en = hw->port_base_vlan_cfg.state == HNS3_PORT_BASE_VLAN_ENABLE;
355 for (i = 0; i < hw->cfg_max_queues; i++) {
357 rxq = hw->data->rx_queues[i];
359 rxq->pvid_sw_discard_en = pvid_en;
362 txq = hw->data->tx_queues[i];
364 txq->pvid_sw_shift_en = pvid_en;
370 hns3_enable_all_queues(struct hns3_hw *hw, bool en)
372 uint16_t nb_rx_q = hw->data->nb_rx_queues;
373 uint16_t nb_tx_q = hw->data->nb_tx_queues;
374 struct hns3_rx_queue *rxq;
375 struct hns3_tx_queue *txq;
380 for (i = 0; i < hw->cfg_max_queues; i++) {
381 if (hns3_dev_indep_txrx_supported(hw)) {
382 rxq = i < nb_rx_q ? hw->data->rx_queues[i] : NULL;
383 txq = i < nb_tx_q ? hw->data->tx_queues[i] : NULL;
385 * After initialization, rxq and txq won't be NULL at
389 tqp_base = rxq->io_base;
390 else if (txq != NULL)
391 tqp_base = txq->io_base;
395 rxq = i < nb_rx_q ? hw->data->rx_queues[i] :
396 hw->fkq_data.rx_queues[i - nb_rx_q];
398 tqp_base = rxq->io_base;
401 * This is the master switch that used to control the enabling
402 * of a pair of Tx and Rx queues. Both the Rx and Tx point to
405 rcb_reg = hns3_read_reg(tqp_base, HNS3_RING_EN_REG);
407 rcb_reg |= BIT(HNS3_RING_EN_B);
409 rcb_reg &= ~BIT(HNS3_RING_EN_B);
410 hns3_write_reg(tqp_base, HNS3_RING_EN_REG, rcb_reg);
415 hns3_enable_txq(struct hns3_tx_queue *txq, bool en)
417 struct hns3_hw *hw = &txq->hns->hw;
420 if (hns3_dev_indep_txrx_supported(hw)) {
421 reg = hns3_read_dev(txq, HNS3_RING_TX_EN_REG);
423 reg |= BIT(HNS3_RING_EN_B);
425 reg &= ~BIT(HNS3_RING_EN_B);
426 hns3_write_dev(txq, HNS3_RING_TX_EN_REG, reg);
432 hns3_enable_rxq(struct hns3_rx_queue *rxq, bool en)
434 struct hns3_hw *hw = &rxq->hns->hw;
437 if (hns3_dev_indep_txrx_supported(hw)) {
438 reg = hns3_read_dev(rxq, HNS3_RING_RX_EN_REG);
440 reg |= BIT(HNS3_RING_EN_B);
442 reg &= ~BIT(HNS3_RING_EN_B);
443 hns3_write_dev(rxq, HNS3_RING_RX_EN_REG, reg);
449 hns3_start_all_txqs(struct rte_eth_dev *dev)
451 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
452 struct hns3_tx_queue *txq;
455 for (i = 0; i < dev->data->nb_tx_queues; i++) {
456 txq = hw->data->tx_queues[i];
458 hns3_err(hw, "Tx queue %u not available or setup.", i);
459 goto start_txqs_fail;
462 * Tx queue is enabled by default. Therefore, the Tx queues
463 * needs to be disabled when deferred_start is set. There is
464 * another master switch used to control the enabling of a pair
465 * of Tx and Rx queues. And the master switch is disabled by
468 if (txq->tx_deferred_start)
469 hns3_enable_txq(txq, false);
471 hns3_enable_txq(txq, true);
476 for (j = 0; j < i; j++) {
477 txq = hw->data->tx_queues[j];
478 hns3_enable_txq(txq, false);
484 hns3_start_all_rxqs(struct rte_eth_dev *dev)
486 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
487 struct hns3_rx_queue *rxq;
490 for (i = 0; i < dev->data->nb_rx_queues; i++) {
491 rxq = hw->data->rx_queues[i];
493 hns3_err(hw, "Rx queue %u not available or setup.", i);
494 goto start_rxqs_fail;
497 * Rx queue is enabled by default. Therefore, the Rx queues
498 * needs to be disabled when deferred_start is set. There is
499 * another master switch used to control the enabling of a pair
500 * of Tx and Rx queues. And the master switch is disabled by
503 if (rxq->rx_deferred_start)
504 hns3_enable_rxq(rxq, false);
506 hns3_enable_rxq(rxq, true);
511 for (j = 0; j < i; j++) {
512 rxq = hw->data->rx_queues[j];
513 hns3_enable_rxq(rxq, false);
519 hns3_stop_all_txqs(struct rte_eth_dev *dev)
521 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
522 struct hns3_tx_queue *txq;
525 for (i = 0; i < dev->data->nb_tx_queues; i++) {
526 txq = hw->data->tx_queues[i];
529 hns3_enable_txq(txq, false);
534 hns3_tqp_enable(struct hns3_hw *hw, uint16_t queue_id, bool enable)
536 struct hns3_cfg_com_tqp_queue_cmd *req;
537 struct hns3_cmd_desc desc;
540 req = (struct hns3_cfg_com_tqp_queue_cmd *)desc.data;
542 hns3_cmd_setup_basic_desc(&desc, HNS3_OPC_CFG_COM_TQP_QUEUE, false);
543 req->tqp_id = rte_cpu_to_le_16(queue_id);
545 hns3_set_bit(req->enable, HNS3_TQP_ENABLE_B, enable ? 1 : 0);
547 ret = hns3_cmd_send(hw, &desc, 1);
549 hns3_err(hw, "TQP enable fail, ret = %d", ret);
555 hns3_send_reset_tqp_cmd(struct hns3_hw *hw, uint16_t queue_id, bool enable)
557 struct hns3_reset_tqp_queue_cmd *req;
558 struct hns3_cmd_desc desc;
561 hns3_cmd_setup_basic_desc(&desc, HNS3_OPC_RESET_TQP_QUEUE, false);
563 req = (struct hns3_reset_tqp_queue_cmd *)desc.data;
564 req->tqp_id = rte_cpu_to_le_16(queue_id);
565 hns3_set_bit(req->reset_req, HNS3_TQP_RESET_B, enable ? 1 : 0);
566 ret = hns3_cmd_send(hw, &desc, 1);
568 hns3_err(hw, "send tqp reset cmd error, queue_id = %u, "
569 "ret = %d", queue_id, ret);
575 hns3_get_tqp_reset_status(struct hns3_hw *hw, uint16_t queue_id,
576 uint8_t *reset_status)
578 struct hns3_reset_tqp_queue_cmd *req;
579 struct hns3_cmd_desc desc;
582 hns3_cmd_setup_basic_desc(&desc, HNS3_OPC_RESET_TQP_QUEUE, true);
584 req = (struct hns3_reset_tqp_queue_cmd *)desc.data;
585 req->tqp_id = rte_cpu_to_le_16(queue_id);
587 ret = hns3_cmd_send(hw, &desc, 1);
589 hns3_err(hw, "get tqp reset status error, queue_id = %u, "
590 "ret = %d.", queue_id, ret);
593 *reset_status = hns3_get_bit(req->ready_to_reset, HNS3_TQP_RESET_B);
598 hns3pf_reset_tqp(struct hns3_hw *hw, uint16_t queue_id)
600 #define HNS3_TQP_RESET_TRY_MS 200
601 uint8_t reset_status;
605 ret = hns3_tqp_enable(hw, queue_id, false);
610 * In current version VF is not supported when PF is driven by DPDK
611 * driver, all task queue pairs are mapped to PF function, so PF's queue
612 * id is equals to the global queue id in PF range.
614 ret = hns3_send_reset_tqp_cmd(hw, queue_id, true);
616 hns3_err(hw, "Send reset tqp cmd fail, ret = %d", ret);
619 end = get_timeofday_ms() + HNS3_TQP_RESET_TRY_MS;
621 /* Wait for tqp hw reset */
622 rte_delay_ms(HNS3_POLL_RESPONE_MS);
623 ret = hns3_get_tqp_reset_status(hw, queue_id, &reset_status);
629 } while (get_timeofday_ms() < end);
633 hns3_err(hw, "reset tqp timeout, queue_id = %u, ret = %d",
638 ret = hns3_send_reset_tqp_cmd(hw, queue_id, false);
640 hns3_err(hw, "Deassert the soft reset fail, ret = %d", ret);
645 hns3_send_reset_tqp_cmd(hw, queue_id, false);
650 hns3vf_reset_tqp(struct hns3_hw *hw, uint16_t queue_id)
655 /* Disable VF's queue before send queue reset msg to PF */
656 ret = hns3_tqp_enable(hw, queue_id, false);
660 memcpy(msg_data, &queue_id, sizeof(uint16_t));
662 ret = hns3_send_mbx_msg(hw, HNS3_MBX_QUEUE_RESET, 0, msg_data,
663 sizeof(msg_data), true, NULL, 0);
665 hns3_err(hw, "fail to reset tqp, queue_id = %u, ret = %d.",
671 hns3_reset_tqp(struct hns3_adapter *hns, uint16_t queue_id)
673 struct hns3_hw *hw = &hns->hw;
676 return hns3vf_reset_tqp(hw, queue_id);
678 return hns3pf_reset_tqp(hw, queue_id);
682 hns3_reset_all_tqps(struct hns3_adapter *hns)
684 struct hns3_hw *hw = &hns->hw;
687 for (i = 0; i < hw->cfg_max_queues; i++) {
688 ret = hns3_reset_tqp(hns, i);
690 hns3_err(hw, "Failed to reset No.%d queue: %d", i, ret);
698 hns3_send_reset_queue_cmd(struct hns3_hw *hw, uint16_t queue_id,
699 enum hns3_ring_type queue_type, bool enable)
701 struct hns3_reset_tqp_queue_cmd *req;
702 struct hns3_cmd_desc desc;
706 hns3_cmd_setup_basic_desc(&desc, HNS3_OPC_RESET_TQP_QUEUE_INDEP, false);
708 req = (struct hns3_reset_tqp_queue_cmd *)desc.data;
709 req->tqp_id = rte_cpu_to_le_16(queue_id);
710 queue_direction = queue_type == HNS3_RING_TYPE_TX ? 0 : 1;
711 req->queue_direction = rte_cpu_to_le_16(queue_direction);
712 hns3_set_bit(req->reset_req, HNS3_TQP_RESET_B, enable ? 1 : 0);
714 ret = hns3_cmd_send(hw, &desc, 1);
716 hns3_err(hw, "send queue reset cmd error, queue_id = %u, "
717 "queue_type = %s, ret = %d.", queue_id,
718 queue_type == HNS3_RING_TYPE_TX ? "Tx" : "Rx", ret);
723 hns3_get_queue_reset_status(struct hns3_hw *hw, uint16_t queue_id,
724 enum hns3_ring_type queue_type,
725 uint8_t *reset_status)
727 struct hns3_reset_tqp_queue_cmd *req;
728 struct hns3_cmd_desc desc;
732 hns3_cmd_setup_basic_desc(&desc, HNS3_OPC_RESET_TQP_QUEUE_INDEP, true);
734 req = (struct hns3_reset_tqp_queue_cmd *)desc.data;
735 req->tqp_id = rte_cpu_to_le_16(queue_id);
736 queue_direction = queue_type == HNS3_RING_TYPE_TX ? 0 : 1;
737 req->queue_direction = rte_cpu_to_le_16(queue_direction);
739 ret = hns3_cmd_send(hw, &desc, 1);
741 hns3_err(hw, "get queue reset status error, queue_id = %u "
742 "queue_type = %s, ret = %d.", queue_id,
743 queue_type == HNS3_RING_TYPE_TX ? "Tx" : "Rx", ret);
747 *reset_status = hns3_get_bit(req->ready_to_reset, HNS3_TQP_RESET_B);
752 hns3_reset_queue(struct hns3_hw *hw, uint16_t queue_id,
753 enum hns3_ring_type queue_type)
755 #define HNS3_QUEUE_RESET_TRY_MS 200
756 struct hns3_tx_queue *txq;
757 struct hns3_rx_queue *rxq;
758 uint32_t reset_wait_times;
759 uint32_t max_wait_times;
760 uint8_t reset_status;
763 if (queue_type == HNS3_RING_TYPE_TX) {
764 txq = hw->data->tx_queues[queue_id];
765 hns3_enable_txq(txq, false);
767 rxq = hw->data->rx_queues[queue_id];
768 hns3_enable_rxq(rxq, false);
771 ret = hns3_send_reset_queue_cmd(hw, queue_id, queue_type, true);
773 hns3_err(hw, "send reset queue cmd fail, ret = %d.", ret);
777 reset_wait_times = 0;
778 max_wait_times = HNS3_QUEUE_RESET_TRY_MS / HNS3_POLL_RESPONE_MS;
779 while (reset_wait_times < max_wait_times) {
780 /* Wait for queue hw reset */
781 rte_delay_ms(HNS3_POLL_RESPONE_MS);
782 ret = hns3_get_queue_reset_status(hw, queue_id,
783 queue_type, &reset_status);
785 goto queue_reset_fail;
793 hns3_err(hw, "reset queue timeout, queue_id = %u, "
794 "queue_type = %s", queue_id,
795 queue_type == HNS3_RING_TYPE_TX ? "Tx" : "Rx");
797 goto queue_reset_fail;
800 ret = hns3_send_reset_queue_cmd(hw, queue_id, queue_type, false);
802 hns3_err(hw, "deassert queue reset fail, ret = %d.", ret);
807 hns3_send_reset_queue_cmd(hw, queue_id, queue_type, false);
813 hns3_set_queue_intr_gl(struct hns3_hw *hw, uint16_t queue_id,
814 uint8_t gl_idx, uint16_t gl_value)
816 uint32_t offset[] = {HNS3_TQP_INTR_GL0_REG,
817 HNS3_TQP_INTR_GL1_REG,
818 HNS3_TQP_INTR_GL2_REG};
819 uint32_t addr, value;
821 if (gl_idx >= RTE_DIM(offset) || gl_value > HNS3_TQP_INTR_GL_MAX)
824 addr = offset[gl_idx] + queue_id * HNS3_TQP_INTR_REG_SIZE;
825 if (hw->intr.gl_unit == HNS3_INTR_COALESCE_GL_UINT_1US)
826 value = gl_value | HNS3_TQP_INTR_GL_UNIT_1US;
828 value = HNS3_GL_USEC_TO_REG(gl_value);
830 hns3_write_dev(hw, addr, value);
834 hns3_set_queue_intr_rl(struct hns3_hw *hw, uint16_t queue_id, uint16_t rl_value)
836 uint32_t addr, value;
838 if (rl_value > HNS3_TQP_INTR_RL_MAX)
841 addr = HNS3_TQP_INTR_RL_REG + queue_id * HNS3_TQP_INTR_REG_SIZE;
842 value = HNS3_RL_USEC_TO_REG(rl_value);
844 value |= HNS3_TQP_INTR_RL_ENABLE_MASK;
846 hns3_write_dev(hw, addr, value);
850 hns3_set_queue_intr_ql(struct hns3_hw *hw, uint16_t queue_id, uint16_t ql_value)
855 * int_ql_max == 0 means the hardware does not support QL,
856 * QL regs config is not permitted if QL is not supported,
859 if (hw->intr.int_ql_max == HNS3_INTR_QL_NONE)
862 addr = HNS3_TQP_INTR_TX_QL_REG + queue_id * HNS3_TQP_INTR_REG_SIZE;
863 hns3_write_dev(hw, addr, ql_value);
865 addr = HNS3_TQP_INTR_RX_QL_REG + queue_id * HNS3_TQP_INTR_REG_SIZE;
866 hns3_write_dev(hw, addr, ql_value);
870 hns3_queue_intr_enable(struct hns3_hw *hw, uint16_t queue_id, bool en)
872 uint32_t addr, value;
874 addr = HNS3_TQP_INTR_CTRL_REG + queue_id * HNS3_TQP_INTR_REG_SIZE;
877 hns3_write_dev(hw, addr, value);
881 * Enable all rx queue interrupt when in interrupt rx mode.
882 * This api was called before enable queue rx&tx (in normal start or reset
883 * recover scenes), used to fix hardware rx queue interrupt enable was clear
887 hns3_dev_all_rx_queue_intr_enable(struct hns3_hw *hw, bool en)
889 struct rte_eth_dev *dev = &rte_eth_devices[hw->data->port_id];
890 uint16_t nb_rx_q = hw->data->nb_rx_queues;
893 if (dev->data->dev_conf.intr_conf.rxq == 0)
896 for (i = 0; i < nb_rx_q; i++)
897 hns3_queue_intr_enable(hw, i, en);
901 hns3_dev_rx_queue_intr_enable(struct rte_eth_dev *dev, uint16_t queue_id)
903 struct rte_pci_device *pci_dev = RTE_ETH_DEV_TO_PCI(dev);
904 struct rte_intr_handle *intr_handle = &pci_dev->intr_handle;
905 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
907 if (dev->data->dev_conf.intr_conf.rxq == 0)
910 hns3_queue_intr_enable(hw, queue_id, true);
912 return rte_intr_ack(intr_handle);
916 hns3_dev_rx_queue_intr_disable(struct rte_eth_dev *dev, uint16_t queue_id)
918 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
920 if (dev->data->dev_conf.intr_conf.rxq == 0)
923 hns3_queue_intr_enable(hw, queue_id, false);
929 hns3_init_rxq(struct hns3_adapter *hns, uint16_t idx)
931 struct hns3_hw *hw = &hns->hw;
932 struct hns3_rx_queue *rxq;
935 PMD_INIT_FUNC_TRACE();
937 rxq = (struct hns3_rx_queue *)hw->data->rx_queues[idx];
938 ret = hns3_alloc_rx_queue_mbufs(hw, rxq);
940 hns3_err(hw, "fail to alloc mbuf for Rx queue %u, ret = %d.",
945 rxq->next_to_use = 0;
946 rxq->rx_rearm_start = 0;
947 rxq->rx_free_hold = 0;
948 rxq->rx_rearm_nb = 0;
949 rxq->pkt_first_seg = NULL;
950 rxq->pkt_last_seg = NULL;
951 hns3_init_rx_queue_hw(rxq);
952 hns3_rxq_vec_setup(rxq);
958 hns3_init_fake_rxq(struct hns3_adapter *hns, uint16_t idx)
960 struct hns3_hw *hw = &hns->hw;
961 struct hns3_rx_queue *rxq;
963 rxq = (struct hns3_rx_queue *)hw->fkq_data.rx_queues[idx];
964 rxq->next_to_use = 0;
965 rxq->rx_free_hold = 0;
966 rxq->rx_rearm_start = 0;
967 rxq->rx_rearm_nb = 0;
968 hns3_init_rx_queue_hw(rxq);
972 hns3_init_txq(struct hns3_tx_queue *txq)
974 struct hns3_desc *desc;
979 for (i = 0; i < txq->nb_tx_desc; i++) {
980 desc->tx.tp_fe_sc_vld_ra_ri = 0;
984 txq->next_to_use = 0;
985 txq->next_to_clean = 0;
986 txq->tx_bd_ready = txq->nb_tx_desc - 1;
987 hns3_init_tx_queue_hw(txq);
991 hns3_init_tx_ring_tc(struct hns3_adapter *hns)
993 struct hns3_hw *hw = &hns->hw;
994 struct hns3_tx_queue *txq;
997 for (i = 0; i < HNS3_MAX_TC_NUM; i++) {
998 struct hns3_tc_queue_info *tc_queue = &hw->tc_queue[i];
1001 if (!tc_queue->enable)
1004 for (j = 0; j < tc_queue->tqp_count; j++) {
1005 num = tc_queue->tqp_offset + j;
1006 txq = (struct hns3_tx_queue *)hw->data->tx_queues[num];
1010 hns3_write_dev(txq, HNS3_RING_TX_TC_REG, tc_queue->tc);
1016 hns3_init_rx_queues(struct hns3_adapter *hns)
1018 struct hns3_hw *hw = &hns->hw;
1019 struct hns3_rx_queue *rxq;
1023 /* Initialize RSS for queues */
1024 ret = hns3_config_rss(hns);
1026 hns3_err(hw, "failed to configure rss, ret = %d.", ret);
1030 for (i = 0; i < hw->data->nb_rx_queues; i++) {
1031 rxq = (struct hns3_rx_queue *)hw->data->rx_queues[i];
1033 hns3_err(hw, "Rx queue %u not available or setup.", i);
1037 if (rxq->rx_deferred_start)
1040 ret = hns3_init_rxq(hns, i);
1042 hns3_err(hw, "failed to init Rx queue %u, ret = %d.", i,
1048 for (i = 0; i < hw->fkq_data.nb_fake_rx_queues; i++)
1049 hns3_init_fake_rxq(hns, i);
1054 for (j = 0; j < i; j++) {
1055 rxq = (struct hns3_rx_queue *)hw->data->rx_queues[j];
1056 hns3_rx_queue_release_mbufs(rxq);
1063 hns3_init_tx_queues(struct hns3_adapter *hns)
1065 struct hns3_hw *hw = &hns->hw;
1066 struct hns3_tx_queue *txq;
1069 for (i = 0; i < hw->data->nb_tx_queues; i++) {
1070 txq = (struct hns3_tx_queue *)hw->data->tx_queues[i];
1072 hns3_err(hw, "Tx queue %u not available or setup.", i);
1076 if (txq->tx_deferred_start)
1081 for (i = 0; i < hw->fkq_data.nb_fake_tx_queues; i++) {
1082 txq = (struct hns3_tx_queue *)hw->fkq_data.tx_queues[i];
1085 hns3_init_tx_ring_tc(hns);
1092 * Note: just init and setup queues, and don't enable tqps.
1095 hns3_init_queues(struct hns3_adapter *hns, bool reset_queue)
1097 struct hns3_hw *hw = &hns->hw;
1101 ret = hns3_reset_all_tqps(hns);
1103 hns3_err(hw, "failed to reset all queues, ret = %d.",
1109 ret = hns3_init_rx_queues(hns);
1111 hns3_err(hw, "failed to init rx queues, ret = %d.", ret);
1115 ret = hns3_init_tx_queues(hns);
1117 hns3_dev_release_mbufs(hns);
1118 hns3_err(hw, "failed to init tx queues, ret = %d.", ret);
1125 hns3_start_tqps(struct hns3_hw *hw)
1127 struct hns3_tx_queue *txq;
1128 struct hns3_rx_queue *rxq;
1131 hns3_enable_all_queues(hw, true);
1133 for (i = 0; i < hw->data->nb_tx_queues; i++) {
1134 txq = hw->data->tx_queues[i];
1136 hw->data->tx_queue_state[i] =
1137 RTE_ETH_QUEUE_STATE_STARTED;
1140 for (i = 0; i < hw->data->nb_rx_queues; i++) {
1141 rxq = hw->data->rx_queues[i];
1143 hw->data->rx_queue_state[i] =
1144 RTE_ETH_QUEUE_STATE_STARTED;
1149 hns3_stop_tqps(struct hns3_hw *hw)
1153 hns3_enable_all_queues(hw, false);
1155 for (i = 0; i < hw->data->nb_tx_queues; i++)
1156 hw->data->tx_queue_state[i] = RTE_ETH_QUEUE_STATE_STOPPED;
1158 for (i = 0; i < hw->data->nb_rx_queues; i++)
1159 hw->data->rx_queue_state[i] = RTE_ETH_QUEUE_STATE_STOPPED;
1163 * Iterate over all Rx Queue, and call the callback() function for each Rx
1167 * The target eth dev.
1168 * @param[in] callback
1169 * The function to call for each queue.
1170 * if callback function return nonzero will stop iterate and return it's value
1172 * The arguments to provide the callback function with.
1175 * 0 on success, otherwise with errno set.
1178 hns3_rxq_iterate(struct rte_eth_dev *dev,
1179 int (*callback)(struct hns3_rx_queue *, void *), void *arg)
1184 if (dev->data->rx_queues == NULL)
1187 for (i = 0; i < dev->data->nb_rx_queues; i++) {
1188 ret = callback(dev->data->rx_queues[i], arg);
1197 hns3_alloc_rxq_and_dma_zone(struct rte_eth_dev *dev,
1198 struct hns3_queue_info *q_info)
1200 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1201 const struct rte_memzone *rx_mz;
1202 struct hns3_rx_queue *rxq;
1203 unsigned int rx_desc;
1205 rxq = rte_zmalloc_socket(q_info->type, sizeof(struct hns3_rx_queue),
1206 RTE_CACHE_LINE_SIZE, q_info->socket_id);
1208 hns3_err(hw, "Failed to allocate memory for No.%d rx ring!",
1213 /* Allocate rx ring hardware descriptors. */
1214 rxq->queue_id = q_info->idx;
1215 rxq->nb_rx_desc = q_info->nb_desc;
1218 * Allocate a litter more memory because rx vector functions
1219 * don't check boundaries each time.
1221 rx_desc = (rxq->nb_rx_desc + HNS3_DEFAULT_RX_BURST) *
1222 sizeof(struct hns3_desc);
1223 rx_mz = rte_eth_dma_zone_reserve(dev, q_info->ring_name, q_info->idx,
1224 rx_desc, HNS3_RING_BASE_ALIGN,
1226 if (rx_mz == NULL) {
1227 hns3_err(hw, "Failed to reserve DMA memory for No.%d rx ring!",
1229 hns3_rx_queue_release(rxq);
1233 rxq->rx_ring = (struct hns3_desc *)rx_mz->addr;
1234 rxq->rx_ring_phys_addr = rx_mz->iova;
1236 hns3_dbg(hw, "No.%d rx descriptors iova 0x%" PRIx64, q_info->idx,
1237 rxq->rx_ring_phys_addr);
1243 hns3_fake_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx,
1244 uint16_t nb_desc, unsigned int socket_id)
1246 struct hns3_adapter *hns = dev->data->dev_private;
1247 struct hns3_hw *hw = &hns->hw;
1248 struct hns3_queue_info q_info;
1249 struct hns3_rx_queue *rxq;
1252 if (hw->fkq_data.rx_queues[idx]) {
1253 hns3_rx_queue_release(hw->fkq_data.rx_queues[idx]);
1254 hw->fkq_data.rx_queues[idx] = NULL;
1258 q_info.socket_id = socket_id;
1259 q_info.nb_desc = nb_desc;
1260 q_info.type = "hns3 fake RX queue";
1261 q_info.ring_name = "rx_fake_ring";
1262 rxq = hns3_alloc_rxq_and_dma_zone(dev, &q_info);
1264 hns3_err(hw, "Failed to setup No.%d fake rx ring.", idx);
1268 /* Don't need alloc sw_ring, because upper applications don't use it */
1269 rxq->sw_ring = NULL;
1272 rxq->rx_deferred_start = false;
1273 rxq->port_id = dev->data->port_id;
1274 rxq->configured = true;
1275 nb_rx_q = dev->data->nb_rx_queues;
1276 rxq->io_base = (void *)((char *)hw->io_base + HNS3_TQP_REG_OFFSET +
1277 (nb_rx_q + idx) * HNS3_TQP_REG_SIZE);
1278 rxq->rx_buf_len = HNS3_MIN_BD_BUF_SIZE;
1280 rte_spinlock_lock(&hw->lock);
1281 hw->fkq_data.rx_queues[idx] = rxq;
1282 rte_spinlock_unlock(&hw->lock);
1288 hns3_alloc_txq_and_dma_zone(struct rte_eth_dev *dev,
1289 struct hns3_queue_info *q_info)
1291 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1292 const struct rte_memzone *tx_mz;
1293 struct hns3_tx_queue *txq;
1294 struct hns3_desc *desc;
1295 unsigned int tx_desc;
1298 txq = rte_zmalloc_socket(q_info->type, sizeof(struct hns3_tx_queue),
1299 RTE_CACHE_LINE_SIZE, q_info->socket_id);
1301 hns3_err(hw, "Failed to allocate memory for No.%d tx ring!",
1306 /* Allocate tx ring hardware descriptors. */
1307 txq->queue_id = q_info->idx;
1308 txq->nb_tx_desc = q_info->nb_desc;
1309 tx_desc = txq->nb_tx_desc * sizeof(struct hns3_desc);
1310 tx_mz = rte_eth_dma_zone_reserve(dev, q_info->ring_name, q_info->idx,
1311 tx_desc, HNS3_RING_BASE_ALIGN,
1313 if (tx_mz == NULL) {
1314 hns3_err(hw, "Failed to reserve DMA memory for No.%d tx ring!",
1316 hns3_tx_queue_release(txq);
1320 txq->tx_ring = (struct hns3_desc *)tx_mz->addr;
1321 txq->tx_ring_phys_addr = tx_mz->iova;
1323 hns3_dbg(hw, "No.%d tx descriptors iova 0x%" PRIx64, q_info->idx,
1324 txq->tx_ring_phys_addr);
1327 desc = txq->tx_ring;
1328 for (i = 0; i < txq->nb_tx_desc; i++) {
1329 desc->tx.tp_fe_sc_vld_ra_ri = 0;
1337 hns3_fake_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx,
1338 uint16_t nb_desc, unsigned int socket_id)
1340 struct hns3_adapter *hns = dev->data->dev_private;
1341 struct hns3_hw *hw = &hns->hw;
1342 struct hns3_queue_info q_info;
1343 struct hns3_tx_queue *txq;
1346 if (hw->fkq_data.tx_queues[idx] != NULL) {
1347 hns3_tx_queue_release(hw->fkq_data.tx_queues[idx]);
1348 hw->fkq_data.tx_queues[idx] = NULL;
1352 q_info.socket_id = socket_id;
1353 q_info.nb_desc = nb_desc;
1354 q_info.type = "hns3 fake TX queue";
1355 q_info.ring_name = "tx_fake_ring";
1356 txq = hns3_alloc_txq_and_dma_zone(dev, &q_info);
1358 hns3_err(hw, "Failed to setup No.%d fake tx ring.", idx);
1362 /* Don't need alloc sw_ring, because upper applications don't use it */
1363 txq->sw_ring = NULL;
1367 txq->tx_deferred_start = false;
1368 txq->port_id = dev->data->port_id;
1369 txq->configured = true;
1370 nb_tx_q = dev->data->nb_tx_queues;
1371 txq->io_base = (void *)((char *)hw->io_base + HNS3_TQP_REG_OFFSET +
1372 (nb_tx_q + idx) * HNS3_TQP_REG_SIZE);
1374 rte_spinlock_lock(&hw->lock);
1375 hw->fkq_data.tx_queues[idx] = txq;
1376 rte_spinlock_unlock(&hw->lock);
1382 hns3_fake_rx_queue_config(struct hns3_hw *hw, uint16_t nb_queues)
1384 uint16_t old_nb_queues = hw->fkq_data.nb_fake_rx_queues;
1388 if (hw->fkq_data.rx_queues == NULL && nb_queues != 0) {
1389 /* first time configuration */
1391 size = sizeof(hw->fkq_data.rx_queues[0]) * nb_queues;
1392 hw->fkq_data.rx_queues = rte_zmalloc("fake_rx_queues", size,
1393 RTE_CACHE_LINE_SIZE);
1394 if (hw->fkq_data.rx_queues == NULL) {
1395 hw->fkq_data.nb_fake_rx_queues = 0;
1398 } else if (hw->fkq_data.rx_queues != NULL && nb_queues != 0) {
1400 rxq = hw->fkq_data.rx_queues;
1401 for (i = nb_queues; i < old_nb_queues; i++)
1402 hns3_dev_rx_queue_release(rxq[i]);
1404 rxq = rte_realloc(rxq, sizeof(rxq[0]) * nb_queues,
1405 RTE_CACHE_LINE_SIZE);
1408 if (nb_queues > old_nb_queues) {
1409 uint16_t new_qs = nb_queues - old_nb_queues;
1410 memset(rxq + old_nb_queues, 0, sizeof(rxq[0]) * new_qs);
1413 hw->fkq_data.rx_queues = rxq;
1414 } else if (hw->fkq_data.rx_queues != NULL && nb_queues == 0) {
1415 rxq = hw->fkq_data.rx_queues;
1416 for (i = nb_queues; i < old_nb_queues; i++)
1417 hns3_dev_rx_queue_release(rxq[i]);
1419 rte_free(hw->fkq_data.rx_queues);
1420 hw->fkq_data.rx_queues = NULL;
1423 hw->fkq_data.nb_fake_rx_queues = nb_queues;
1429 hns3_fake_tx_queue_config(struct hns3_hw *hw, uint16_t nb_queues)
1431 uint16_t old_nb_queues = hw->fkq_data.nb_fake_tx_queues;
1435 if (hw->fkq_data.tx_queues == NULL && nb_queues != 0) {
1436 /* first time configuration */
1438 size = sizeof(hw->fkq_data.tx_queues[0]) * nb_queues;
1439 hw->fkq_data.tx_queues = rte_zmalloc("fake_tx_queues", size,
1440 RTE_CACHE_LINE_SIZE);
1441 if (hw->fkq_data.tx_queues == NULL) {
1442 hw->fkq_data.nb_fake_tx_queues = 0;
1445 } else if (hw->fkq_data.tx_queues != NULL && nb_queues != 0) {
1447 txq = hw->fkq_data.tx_queues;
1448 for (i = nb_queues; i < old_nb_queues; i++)
1449 hns3_dev_tx_queue_release(txq[i]);
1450 txq = rte_realloc(txq, sizeof(txq[0]) * nb_queues,
1451 RTE_CACHE_LINE_SIZE);
1454 if (nb_queues > old_nb_queues) {
1455 uint16_t new_qs = nb_queues - old_nb_queues;
1456 memset(txq + old_nb_queues, 0, sizeof(txq[0]) * new_qs);
1459 hw->fkq_data.tx_queues = txq;
1460 } else if (hw->fkq_data.tx_queues != NULL && nb_queues == 0) {
1461 txq = hw->fkq_data.tx_queues;
1462 for (i = nb_queues; i < old_nb_queues; i++)
1463 hns3_dev_tx_queue_release(txq[i]);
1465 rte_free(hw->fkq_data.tx_queues);
1466 hw->fkq_data.tx_queues = NULL;
1468 hw->fkq_data.nb_fake_tx_queues = nb_queues;
1474 hns3_set_fake_rx_or_tx_queues(struct rte_eth_dev *dev, uint16_t nb_rx_q,
1477 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1478 uint16_t rx_need_add_nb_q;
1479 uint16_t tx_need_add_nb_q;
1484 /* Setup new number of fake RX/TX queues and reconfigure device. */
1485 rx_need_add_nb_q = hw->cfg_max_queues - nb_rx_q;
1486 tx_need_add_nb_q = hw->cfg_max_queues - nb_tx_q;
1487 ret = hns3_fake_rx_queue_config(hw, rx_need_add_nb_q);
1489 hns3_err(hw, "Fail to configure fake rx queues: %d", ret);
1493 ret = hns3_fake_tx_queue_config(hw, tx_need_add_nb_q);
1495 hns3_err(hw, "Fail to configure fake rx queues: %d", ret);
1496 goto cfg_fake_tx_q_fail;
1499 /* Allocate and set up fake RX queue per Ethernet port. */
1500 port_id = hw->data->port_id;
1501 for (q = 0; q < rx_need_add_nb_q; q++) {
1502 ret = hns3_fake_rx_queue_setup(dev, q, HNS3_MIN_RING_DESC,
1503 rte_eth_dev_socket_id(port_id));
1505 goto setup_fake_rx_q_fail;
1508 /* Allocate and set up fake TX queue per Ethernet port. */
1509 for (q = 0; q < tx_need_add_nb_q; q++) {
1510 ret = hns3_fake_tx_queue_setup(dev, q, HNS3_MIN_RING_DESC,
1511 rte_eth_dev_socket_id(port_id));
1513 goto setup_fake_tx_q_fail;
1518 setup_fake_tx_q_fail:
1519 setup_fake_rx_q_fail:
1520 (void)hns3_fake_tx_queue_config(hw, 0);
1522 (void)hns3_fake_rx_queue_config(hw, 0);
1528 hns3_dev_release_mbufs(struct hns3_adapter *hns)
1530 struct rte_eth_dev_data *dev_data = hns->hw.data;
1531 struct hns3_rx_queue *rxq;
1532 struct hns3_tx_queue *txq;
1535 if (dev_data->rx_queues)
1536 for (i = 0; i < dev_data->nb_rx_queues; i++) {
1537 rxq = dev_data->rx_queues[i];
1540 hns3_rx_queue_release_mbufs(rxq);
1543 if (dev_data->tx_queues)
1544 for (i = 0; i < dev_data->nb_tx_queues; i++) {
1545 txq = dev_data->tx_queues[i];
1548 hns3_tx_queue_release_mbufs(txq);
1553 hns3_rx_buf_len_calc(struct rte_mempool *mp, uint16_t *rx_buf_len)
1555 uint16_t vld_buf_size;
1556 uint16_t num_hw_specs;
1560 * hns3 network engine only support to set 4 typical specification, and
1561 * different buffer size will affect the max packet_len and the max
1562 * number of segmentation when hw gro is turned on in receive side. The
1563 * relationship between them is as follows:
1564 * rx_buf_size | max_gro_pkt_len | max_gro_nb_seg
1565 * ---------------------|-------------------|----------------
1566 * HNS3_4K_BD_BUF_SIZE | 60KB | 15
1567 * HNS3_2K_BD_BUF_SIZE | 62KB | 31
1568 * HNS3_1K_BD_BUF_SIZE | 63KB | 63
1569 * HNS3_512_BD_BUF_SIZE | 31.5KB | 63
1571 static const uint16_t hw_rx_buf_size[] = {
1572 HNS3_4K_BD_BUF_SIZE,
1573 HNS3_2K_BD_BUF_SIZE,
1574 HNS3_1K_BD_BUF_SIZE,
1575 HNS3_512_BD_BUF_SIZE
1578 vld_buf_size = (uint16_t)(rte_pktmbuf_data_room_size(mp) -
1579 RTE_PKTMBUF_HEADROOM);
1581 if (vld_buf_size < HNS3_MIN_BD_BUF_SIZE)
1584 num_hw_specs = RTE_DIM(hw_rx_buf_size);
1585 for (i = 0; i < num_hw_specs; i++) {
1586 if (vld_buf_size >= hw_rx_buf_size[i]) {
1587 *rx_buf_len = hw_rx_buf_size[i];
1595 hns3_rxq_conf_runtime_check(struct hns3_hw *hw, uint16_t buf_size,
1598 struct rte_eth_dev *dev = &rte_eth_devices[hw->data->port_id];
1599 struct rte_eth_rxmode *rxmode = &hw->data->dev_conf.rxmode;
1600 eth_rx_burst_t pkt_burst = dev->rx_pkt_burst;
1601 uint16_t min_vec_bds;
1604 * HNS3 hardware network engine set scattered as default. If the driver
1605 * is not work in scattered mode and the pkts greater than buf_size
1606 * but smaller than max_rx_pkt_len will be distributed to multiple BDs.
1607 * Driver cannot handle this situation.
1609 if (!hw->data->scattered_rx && rxmode->max_rx_pkt_len > buf_size) {
1610 hns3_err(hw, "max_rx_pkt_len is not allowed to be set greater "
1611 "than rx_buf_len if scattered is off.");
1615 if (pkt_burst == hns3_recv_pkts_vec) {
1616 min_vec_bds = HNS3_DEFAULT_RXQ_REARM_THRESH +
1617 HNS3_DEFAULT_RX_BURST;
1618 if (nb_desc < min_vec_bds ||
1619 nb_desc % HNS3_DEFAULT_RXQ_REARM_THRESH) {
1620 hns3_err(hw, "if Rx burst mode is vector, "
1621 "number of descriptor is required to be "
1622 "bigger than min vector bds:%u, and could be "
1623 "divided by rxq rearm thresh:%u.",
1624 min_vec_bds, HNS3_DEFAULT_RXQ_REARM_THRESH);
1632 hns3_rx_queue_conf_check(struct hns3_hw *hw, const struct rte_eth_rxconf *conf,
1633 struct rte_mempool *mp, uint16_t nb_desc,
1638 if (nb_desc > HNS3_MAX_RING_DESC || nb_desc < HNS3_MIN_RING_DESC ||
1639 nb_desc % HNS3_ALIGN_RING_DESC) {
1640 hns3_err(hw, "Number (%u) of rx descriptors is invalid",
1645 if (conf->rx_drop_en == 0)
1646 hns3_warn(hw, "if no descriptors available, packets are always "
1647 "dropped and rx_drop_en (1) is fixed on");
1649 if (hns3_rx_buf_len_calc(mp, buf_size)) {
1650 hns3_err(hw, "rxq mbufs' data room size (%u) is not enough! "
1651 "minimal data room size (%u).",
1652 rte_pktmbuf_data_room_size(mp),
1653 HNS3_MIN_BD_BUF_SIZE + RTE_PKTMBUF_HEADROOM);
1657 if (hw->data->dev_started) {
1658 ret = hns3_rxq_conf_runtime_check(hw, *buf_size, nb_desc);
1660 hns3_err(hw, "Rx queue runtime setup fail.");
1669 hns3_get_tqp_reg_offset(uint16_t queue_id)
1671 uint32_t reg_offset;
1673 /* Need an extend offset to config queue > 1024 */
1674 if (queue_id < HNS3_MIN_EXTEND_QUEUE_ID)
1675 reg_offset = HNS3_TQP_REG_OFFSET + queue_id * HNS3_TQP_REG_SIZE;
1677 reg_offset = HNS3_TQP_REG_OFFSET + HNS3_TQP_EXT_REG_OFFSET +
1678 (queue_id - HNS3_MIN_EXTEND_QUEUE_ID) *
1685 hns3_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t nb_desc,
1686 unsigned int socket_id, const struct rte_eth_rxconf *conf,
1687 struct rte_mempool *mp)
1689 struct hns3_adapter *hns = dev->data->dev_private;
1690 struct hns3_hw *hw = &hns->hw;
1691 struct hns3_queue_info q_info;
1692 struct hns3_rx_queue *rxq;
1693 uint16_t rx_buf_size;
1697 ret = hns3_rx_queue_conf_check(hw, conf, mp, nb_desc, &rx_buf_size);
1701 if (dev->data->rx_queues[idx]) {
1702 hns3_rx_queue_release(dev->data->rx_queues[idx]);
1703 dev->data->rx_queues[idx] = NULL;
1707 q_info.socket_id = socket_id;
1708 q_info.nb_desc = nb_desc;
1709 q_info.type = "hns3 RX queue";
1710 q_info.ring_name = "rx_ring";
1712 rxq = hns3_alloc_rxq_and_dma_zone(dev, &q_info);
1715 "Failed to alloc mem and reserve DMA mem for rx ring!");
1720 rxq->ptype_tbl = &hns->ptype_tbl;
1722 rxq->rx_free_thresh = (conf->rx_free_thresh > 0) ?
1723 conf->rx_free_thresh : HNS3_DEFAULT_RX_FREE_THRESH;
1725 rxq->rx_deferred_start = conf->rx_deferred_start;
1726 if (rxq->rx_deferred_start && !hns3_dev_indep_txrx_supported(hw)) {
1727 hns3_warn(hw, "deferred start is not supported.");
1728 rxq->rx_deferred_start = false;
1731 rx_entry_len = (rxq->nb_rx_desc + HNS3_DEFAULT_RX_BURST) *
1732 sizeof(struct hns3_entry);
1733 rxq->sw_ring = rte_zmalloc_socket("hns3 RX sw ring", rx_entry_len,
1734 RTE_CACHE_LINE_SIZE, socket_id);
1735 if (rxq->sw_ring == NULL) {
1736 hns3_err(hw, "Failed to allocate memory for rx sw ring!");
1737 hns3_rx_queue_release(rxq);
1741 rxq->next_to_use = 0;
1742 rxq->rx_free_hold = 0;
1743 rxq->rx_rearm_start = 0;
1744 rxq->rx_rearm_nb = 0;
1745 rxq->pkt_first_seg = NULL;
1746 rxq->pkt_last_seg = NULL;
1747 rxq->port_id = dev->data->port_id;
1749 * For hns3 PF device, if the VLAN mode is HW_SHIFT_AND_DISCARD_MODE,
1750 * the pvid_sw_discard_en in the queue struct should not be changed,
1751 * because PVID-related operations do not need to be processed by PMD
1752 * driver. For hns3 VF device, whether it needs to process PVID depends
1753 * on the configuration of PF kernel mode netdevice driver. And the
1754 * related PF configuration is delivered through the mailbox and finally
1755 * reflectd in port_base_vlan_cfg.
1757 if (hns->is_vf || hw->vlan_mode == HNS3_SW_SHIFT_AND_DISCARD_MODE)
1758 rxq->pvid_sw_discard_en = hw->port_base_vlan_cfg.state ==
1759 HNS3_PORT_BASE_VLAN_ENABLE;
1761 rxq->pvid_sw_discard_en = false;
1762 rxq->configured = true;
1763 rxq->io_base = (void *)((char *)hw->io_base + HNS3_TQP_REG_OFFSET +
1764 idx * HNS3_TQP_REG_SIZE);
1765 rxq->io_base = (void *)((char *)hw->io_base +
1766 hns3_get_tqp_reg_offset(idx));
1767 rxq->io_head_reg = (volatile void *)((char *)rxq->io_base +
1768 HNS3_RING_RX_HEAD_REG);
1769 rxq->rx_buf_len = rx_buf_size;
1771 rxq->pkt_len_errors = 0;
1772 rxq->l3_csum_errors = 0;
1773 rxq->l4_csum_errors = 0;
1774 rxq->ol3_csum_errors = 0;
1775 rxq->ol4_csum_errors = 0;
1777 /* CRC len set here is used for amending packet length */
1778 if (dev->data->dev_conf.rxmode.offloads & DEV_RX_OFFLOAD_KEEP_CRC)
1779 rxq->crc_len = RTE_ETHER_CRC_LEN;
1783 rxq->bulk_mbuf_num = 0;
1785 rte_spinlock_lock(&hw->lock);
1786 dev->data->rx_queues[idx] = rxq;
1787 rte_spinlock_unlock(&hw->lock);
1793 hns3_rx_scattered_reset(struct rte_eth_dev *dev)
1795 struct hns3_adapter *hns = dev->data->dev_private;
1796 struct hns3_hw *hw = &hns->hw;
1799 dev->data->scattered_rx = false;
1803 hns3_rx_scattered_calc(struct rte_eth_dev *dev)
1805 struct rte_eth_conf *dev_conf = &dev->data->dev_conf;
1806 struct hns3_adapter *hns = dev->data->dev_private;
1807 struct hns3_hw *hw = &hns->hw;
1808 struct hns3_rx_queue *rxq;
1811 if (dev->data->rx_queues == NULL)
1814 for (queue_id = 0; queue_id < dev->data->nb_rx_queues; queue_id++) {
1815 rxq = dev->data->rx_queues[queue_id];
1816 if (hw->rx_buf_len == 0)
1817 hw->rx_buf_len = rxq->rx_buf_len;
1819 hw->rx_buf_len = RTE_MIN(hw->rx_buf_len,
1823 if (dev_conf->rxmode.offloads & DEV_RX_OFFLOAD_SCATTER ||
1824 dev_conf->rxmode.max_rx_pkt_len > hw->rx_buf_len)
1825 dev->data->scattered_rx = true;
1829 hns3_dev_supported_ptypes_get(struct rte_eth_dev *dev)
1831 static const uint32_t ptypes[] = {
1833 RTE_PTYPE_L2_ETHER_VLAN,
1834 RTE_PTYPE_L2_ETHER_QINQ,
1835 RTE_PTYPE_L2_ETHER_LLDP,
1836 RTE_PTYPE_L2_ETHER_ARP,
1838 RTE_PTYPE_L3_IPV4_EXT,
1840 RTE_PTYPE_L3_IPV6_EXT,
1846 RTE_PTYPE_TUNNEL_GRE,
1847 RTE_PTYPE_INNER_L2_ETHER,
1848 RTE_PTYPE_INNER_L2_ETHER_VLAN,
1849 RTE_PTYPE_INNER_L2_ETHER_QINQ,
1850 RTE_PTYPE_INNER_L3_IPV4,
1851 RTE_PTYPE_INNER_L3_IPV6,
1852 RTE_PTYPE_INNER_L3_IPV4_EXT,
1853 RTE_PTYPE_INNER_L3_IPV6_EXT,
1854 RTE_PTYPE_INNER_L4_UDP,
1855 RTE_PTYPE_INNER_L4_TCP,
1856 RTE_PTYPE_INNER_L4_SCTP,
1857 RTE_PTYPE_INNER_L4_ICMP,
1858 RTE_PTYPE_TUNNEL_VXLAN,
1859 RTE_PTYPE_TUNNEL_NVGRE,
1863 if (dev->rx_pkt_burst == hns3_recv_pkts ||
1864 dev->rx_pkt_burst == hns3_recv_scattered_pkts ||
1865 dev->rx_pkt_burst == hns3_recv_pkts_vec ||
1866 dev->rx_pkt_burst == hns3_recv_pkts_vec_sve)
1873 hns3_init_non_tunnel_ptype_tbl(struct hns3_ptype_table *tbl)
1875 tbl->l2l3table[0][0] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4;
1876 tbl->l2l3table[0][1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6;
1877 tbl->l2l3table[0][2] = RTE_PTYPE_L2_ETHER_ARP;
1878 tbl->l2l3table[0][3] = RTE_PTYPE_L2_ETHER;
1879 tbl->l2l3table[0][4] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT;
1880 tbl->l2l3table[0][5] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT;
1881 tbl->l2l3table[0][6] = RTE_PTYPE_L2_ETHER_LLDP;
1882 tbl->l2l3table[0][15] = RTE_PTYPE_L2_ETHER;
1884 tbl->l2l3table[1][0] = RTE_PTYPE_L2_ETHER_VLAN | RTE_PTYPE_L3_IPV4;
1885 tbl->l2l3table[1][1] = RTE_PTYPE_L2_ETHER_VLAN | RTE_PTYPE_L3_IPV6;
1886 tbl->l2l3table[1][2] = RTE_PTYPE_L2_ETHER_ARP;
1887 tbl->l2l3table[1][3] = RTE_PTYPE_L2_ETHER_VLAN;
1888 tbl->l2l3table[1][4] = RTE_PTYPE_L2_ETHER_VLAN | RTE_PTYPE_L3_IPV4_EXT;
1889 tbl->l2l3table[1][5] = RTE_PTYPE_L2_ETHER_VLAN | RTE_PTYPE_L3_IPV6_EXT;
1890 tbl->l2l3table[1][6] = RTE_PTYPE_L2_ETHER_LLDP;
1891 tbl->l2l3table[1][15] = RTE_PTYPE_L2_ETHER_VLAN;
1893 tbl->l2l3table[2][0] = RTE_PTYPE_L2_ETHER_QINQ | RTE_PTYPE_L3_IPV4;
1894 tbl->l2l3table[2][1] = RTE_PTYPE_L2_ETHER_QINQ | RTE_PTYPE_L3_IPV6;
1895 tbl->l2l3table[2][2] = RTE_PTYPE_L2_ETHER_ARP;
1896 tbl->l2l3table[2][3] = RTE_PTYPE_L2_ETHER_QINQ;
1897 tbl->l2l3table[2][4] = RTE_PTYPE_L2_ETHER_QINQ | RTE_PTYPE_L3_IPV4_EXT;
1898 tbl->l2l3table[2][5] = RTE_PTYPE_L2_ETHER_QINQ | RTE_PTYPE_L3_IPV6_EXT;
1899 tbl->l2l3table[2][6] = RTE_PTYPE_L2_ETHER_LLDP;
1900 tbl->l2l3table[2][15] = RTE_PTYPE_L2_ETHER_QINQ;
1902 tbl->l4table[0] = RTE_PTYPE_L4_UDP;
1903 tbl->l4table[1] = RTE_PTYPE_L4_TCP;
1904 tbl->l4table[2] = RTE_PTYPE_TUNNEL_GRE;
1905 tbl->l4table[3] = RTE_PTYPE_L4_SCTP;
1906 tbl->l4table[4] = RTE_PTYPE_L4_IGMP;
1907 tbl->l4table[5] = RTE_PTYPE_L4_ICMP;
1911 hns3_init_tunnel_ptype_tbl(struct hns3_ptype_table *tbl)
1913 tbl->inner_l2table[0] = RTE_PTYPE_INNER_L2_ETHER;
1914 tbl->inner_l2table[1] = RTE_PTYPE_INNER_L2_ETHER_VLAN;
1915 tbl->inner_l2table[2] = RTE_PTYPE_INNER_L2_ETHER_QINQ;
1917 tbl->inner_l3table[0] = RTE_PTYPE_INNER_L3_IPV4;
1918 tbl->inner_l3table[1] = RTE_PTYPE_INNER_L3_IPV6;
1919 /* There is not a ptype for inner ARP/RARP */
1920 tbl->inner_l3table[2] = RTE_PTYPE_UNKNOWN;
1921 tbl->inner_l3table[3] = RTE_PTYPE_UNKNOWN;
1922 tbl->inner_l3table[4] = RTE_PTYPE_INNER_L3_IPV4_EXT;
1923 tbl->inner_l3table[5] = RTE_PTYPE_INNER_L3_IPV6_EXT;
1925 tbl->inner_l4table[0] = RTE_PTYPE_INNER_L4_UDP;
1926 tbl->inner_l4table[1] = RTE_PTYPE_INNER_L4_TCP;
1927 /* There is not a ptype for inner GRE */
1928 tbl->inner_l4table[2] = RTE_PTYPE_UNKNOWN;
1929 tbl->inner_l4table[3] = RTE_PTYPE_INNER_L4_SCTP;
1930 /* There is not a ptype for inner IGMP */
1931 tbl->inner_l4table[4] = RTE_PTYPE_UNKNOWN;
1932 tbl->inner_l4table[5] = RTE_PTYPE_INNER_L4_ICMP;
1934 tbl->ol2table[0] = RTE_PTYPE_L2_ETHER;
1935 tbl->ol2table[1] = RTE_PTYPE_L2_ETHER_VLAN;
1936 tbl->ol2table[2] = RTE_PTYPE_L2_ETHER_QINQ;
1938 tbl->ol3table[0] = RTE_PTYPE_L3_IPV4;
1939 tbl->ol3table[1] = RTE_PTYPE_L3_IPV6;
1940 tbl->ol3table[2] = RTE_PTYPE_UNKNOWN;
1941 tbl->ol3table[3] = RTE_PTYPE_UNKNOWN;
1942 tbl->ol3table[4] = RTE_PTYPE_L3_IPV4_EXT;
1943 tbl->ol3table[5] = RTE_PTYPE_L3_IPV6_EXT;
1945 tbl->ol4table[0] = RTE_PTYPE_UNKNOWN;
1946 tbl->ol4table[1] = RTE_PTYPE_TUNNEL_VXLAN;
1947 tbl->ol4table[2] = RTE_PTYPE_TUNNEL_NVGRE;
1951 hns3_init_rx_ptype_tble(struct rte_eth_dev *dev)
1953 struct hns3_adapter *hns = dev->data->dev_private;
1954 struct hns3_ptype_table *tbl = &hns->ptype_tbl;
1956 memset(tbl, 0, sizeof(*tbl));
1958 hns3_init_non_tunnel_ptype_tbl(tbl);
1959 hns3_init_tunnel_ptype_tbl(tbl);
1963 hns3_rxd_to_vlan_tci(struct hns3_rx_queue *rxq, struct rte_mbuf *mb,
1964 uint32_t l234_info, const struct hns3_desc *rxd)
1966 #define HNS3_STRP_STATUS_NUM 0x4
1968 #define HNS3_NO_STRP_VLAN_VLD 0x0
1969 #define HNS3_INNER_STRP_VLAN_VLD 0x1
1970 #define HNS3_OUTER_STRP_VLAN_VLD 0x2
1971 uint32_t strip_status;
1972 uint32_t report_mode;
1975 * Since HW limitation, the vlan tag will always be inserted into RX
1976 * descriptor when strip the tag from packet, driver needs to determine
1977 * reporting which tag to mbuf according to the PVID configuration
1978 * and vlan striped status.
1980 static const uint32_t report_type[][HNS3_STRP_STATUS_NUM] = {
1982 HNS3_NO_STRP_VLAN_VLD,
1983 HNS3_OUTER_STRP_VLAN_VLD,
1984 HNS3_INNER_STRP_VLAN_VLD,
1985 HNS3_OUTER_STRP_VLAN_VLD
1988 HNS3_NO_STRP_VLAN_VLD,
1989 HNS3_NO_STRP_VLAN_VLD,
1990 HNS3_NO_STRP_VLAN_VLD,
1991 HNS3_INNER_STRP_VLAN_VLD
1994 strip_status = hns3_get_field(l234_info, HNS3_RXD_STRP_TAGP_M,
1995 HNS3_RXD_STRP_TAGP_S);
1996 report_mode = report_type[rxq->pvid_sw_discard_en][strip_status];
1997 switch (report_mode) {
1998 case HNS3_NO_STRP_VLAN_VLD:
2001 case HNS3_INNER_STRP_VLAN_VLD:
2002 mb->ol_flags |= PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
2003 mb->vlan_tci = rte_le_to_cpu_16(rxd->rx.vlan_tag);
2005 case HNS3_OUTER_STRP_VLAN_VLD:
2006 mb->ol_flags |= PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
2007 mb->vlan_tci = rte_le_to_cpu_16(rxd->rx.ot_vlan_tag);
2016 recalculate_data_len(struct rte_mbuf *first_seg, struct rte_mbuf *last_seg,
2017 struct rte_mbuf *rxm, struct hns3_rx_queue *rxq,
2020 uint8_t crc_len = rxq->crc_len;
2022 if (data_len <= crc_len) {
2023 rte_pktmbuf_free_seg(rxm);
2024 first_seg->nb_segs--;
2025 last_seg->data_len = (uint16_t)(last_seg->data_len -
2026 (crc_len - data_len));
2027 last_seg->next = NULL;
2029 rxm->data_len = (uint16_t)(data_len - crc_len);
2032 static inline struct rte_mbuf *
2033 hns3_rx_alloc_buffer(struct hns3_rx_queue *rxq)
2037 if (likely(rxq->bulk_mbuf_num > 0))
2038 return rxq->bulk_mbuf[--rxq->bulk_mbuf_num];
2040 ret = rte_mempool_get_bulk(rxq->mb_pool, (void **)rxq->bulk_mbuf,
2041 HNS3_BULK_ALLOC_MBUF_NUM);
2042 if (likely(ret == 0)) {
2043 rxq->bulk_mbuf_num = HNS3_BULK_ALLOC_MBUF_NUM;
2044 return rxq->bulk_mbuf[--rxq->bulk_mbuf_num];
2046 return rte_mbuf_raw_alloc(rxq->mb_pool);
2050 hns3_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
2052 volatile struct hns3_desc *rx_ring; /* RX ring (desc) */
2053 volatile struct hns3_desc *rxdp; /* pointer of the current desc */
2054 struct hns3_rx_queue *rxq; /* RX queue */
2055 struct hns3_entry *sw_ring;
2056 struct hns3_entry *rxe;
2057 struct hns3_desc rxd;
2058 struct rte_mbuf *nmb; /* pointer of the new mbuf */
2059 struct rte_mbuf *rxm;
2060 uint32_t bd_base_info;
2073 rx_ring = rxq->rx_ring;
2074 sw_ring = rxq->sw_ring;
2075 rx_id = rxq->next_to_use;
2077 while (nb_rx < nb_pkts) {
2078 rxdp = &rx_ring[rx_id];
2079 bd_base_info = rte_le_to_cpu_32(rxdp->rx.bd_base_info);
2080 if (unlikely(!(bd_base_info & BIT(HNS3_RXD_VLD_B))))
2083 rxd = rxdp[(bd_base_info & (1u << HNS3_RXD_VLD_B)) -
2084 (1u << HNS3_RXD_VLD_B)];
2086 nmb = hns3_rx_alloc_buffer(rxq);
2087 if (unlikely(nmb == NULL)) {
2090 port_id = rxq->port_id;
2091 rte_eth_devices[port_id].data->rx_mbuf_alloc_failed++;
2096 rxe = &sw_ring[rx_id];
2098 if (unlikely(rx_id == rxq->nb_rx_desc))
2101 rte_prefetch0(sw_ring[rx_id].mbuf);
2102 if ((rx_id & HNS3_RX_RING_PREFETCTH_MASK) == 0) {
2103 rte_prefetch0(&rx_ring[rx_id]);
2104 rte_prefetch0(&sw_ring[rx_id]);
2110 dma_addr = rte_mbuf_data_iova_default(nmb);
2111 rxdp->addr = rte_cpu_to_le_64(dma_addr);
2112 rxdp->rx.bd_base_info = 0;
2114 rxm->data_off = RTE_PKTMBUF_HEADROOM;
2115 rxm->pkt_len = (uint16_t)(rte_le_to_cpu_16(rxd.rx.pkt_len)) -
2117 rxm->data_len = rxm->pkt_len;
2118 rxm->port = rxq->port_id;
2119 rxm->hash.rss = rte_le_to_cpu_32(rxd.rx.rss_hash);
2120 rxm->ol_flags = PKT_RX_RSS_HASH;
2121 if (unlikely(bd_base_info & BIT(HNS3_RXD_LUM_B))) {
2123 rte_le_to_cpu_16(rxd.rx.fd_id);
2124 rxm->ol_flags |= PKT_RX_FDIR | PKT_RX_FDIR_ID;
2129 /* Load remained descriptor data and extract necessary fields */
2130 l234_info = rte_le_to_cpu_32(rxd.rx.l234_info);
2131 ol_info = rte_le_to_cpu_32(rxd.rx.ol_info);
2132 ret = hns3_handle_bdinfo(rxq, rxm, bd_base_info,
2133 l234_info, &cksum_err);
2137 rxm->packet_type = hns3_rx_calc_ptype(rxq, l234_info, ol_info);
2139 if (likely(bd_base_info & BIT(HNS3_RXD_L3L4P_B)))
2140 hns3_rx_set_cksum_flag(rxm, rxm->packet_type,
2142 hns3_rxd_to_vlan_tci(rxq, rxm, l234_info, &rxd);
2144 rx_pkts[nb_rx++] = rxm;
2147 rte_pktmbuf_free(rxm);
2150 rxq->next_to_use = rx_id;
2151 rxq->rx_free_hold += nb_rx_bd;
2152 if (rxq->rx_free_hold > rxq->rx_free_thresh) {
2153 hns3_write_reg_opt(rxq->io_head_reg, rxq->rx_free_hold);
2154 rxq->rx_free_hold = 0;
2161 hns3_recv_scattered_pkts(void *rx_queue,
2162 struct rte_mbuf **rx_pkts,
2165 volatile struct hns3_desc *rx_ring; /* RX ring (desc) */
2166 volatile struct hns3_desc *rxdp; /* pointer of the current desc */
2167 struct hns3_rx_queue *rxq; /* RX queue */
2168 struct hns3_entry *sw_ring;
2169 struct hns3_entry *rxe;
2170 struct rte_mbuf *first_seg;
2171 struct rte_mbuf *last_seg;
2172 struct hns3_desc rxd;
2173 struct rte_mbuf *nmb; /* pointer of the new mbuf */
2174 struct rte_mbuf *rxm;
2175 struct rte_eth_dev *dev;
2176 uint32_t bd_base_info;
2191 rx_id = rxq->next_to_use;
2192 rx_ring = rxq->rx_ring;
2193 sw_ring = rxq->sw_ring;
2194 first_seg = rxq->pkt_first_seg;
2195 last_seg = rxq->pkt_last_seg;
2197 while (nb_rx < nb_pkts) {
2198 rxdp = &rx_ring[rx_id];
2199 bd_base_info = rte_le_to_cpu_32(rxdp->rx.bd_base_info);
2200 if (unlikely(!(bd_base_info & BIT(HNS3_RXD_VLD_B))))
2204 * The interactive process between software and hardware of
2205 * receiving a new packet in hns3 network engine:
2206 * 1. Hardware network engine firstly writes the packet content
2207 * to the memory pointed by the 'addr' field of the Rx Buffer
2208 * Descriptor, secondly fills the result of parsing the
2209 * packet include the valid field into the Rx Buffer
2210 * Descriptor in one write operation.
2211 * 2. Driver reads the Rx BD's valid field in the loop to check
2212 * whether it's valid, if valid then assign a new address to
2213 * the addr field, clear the valid field, get the other
2214 * information of the packet by parsing Rx BD's other fields,
2215 * finally write back the number of Rx BDs processed by the
2216 * driver to the HNS3_RING_RX_HEAD_REG register to inform
2218 * In the above process, the ordering is very important. We must
2219 * make sure that CPU read Rx BD's other fields only after the
2222 * There are two type of re-ordering: compiler re-ordering and
2223 * CPU re-ordering under the ARMv8 architecture.
2224 * 1. we use volatile to deal with compiler re-ordering, so you
2225 * can see that rx_ring/rxdp defined with volatile.
2226 * 2. we commonly use memory barrier to deal with CPU
2227 * re-ordering, but the cost is high.
2229 * In order to solve the high cost of using memory barrier, we
2230 * use the data dependency order under the ARMv8 architecture,
2233 * instr02: load B <- A
2234 * the instr02 will always execute after instr01.
2236 * To construct the data dependency ordering, we use the
2237 * following assignment:
2238 * rxd = rxdp[(bd_base_info & (1u << HNS3_RXD_VLD_B)) -
2239 * (1u<<HNS3_RXD_VLD_B)]
2240 * Using gcc compiler under the ARMv8 architecture, the related
2241 * assembly code example as follows:
2242 * note: (1u << HNS3_RXD_VLD_B) equal 0x10
2243 * instr01: ldr w26, [x22, #28] --read bd_base_info
2244 * instr02: and w0, w26, #0x10 --calc bd_base_info & 0x10
2245 * instr03: sub w0, w0, #0x10 --calc (bd_base_info &
2247 * instr04: add x0, x22, x0, lsl #5 --calc copy source addr
2248 * instr05: ldp x2, x3, [x0]
2249 * instr06: stp x2, x3, [x29, #256] --copy BD's [0 ~ 15]B
2250 * instr07: ldp x4, x5, [x0, #16]
2251 * instr08: stp x4, x5, [x29, #272] --copy BD's [16 ~ 31]B
2252 * the instr05~08 depend on x0's value, x0 depent on w26's
2253 * value, the w26 is the bd_base_info, this form the data
2254 * dependency ordering.
2255 * note: if BD is valid, (bd_base_info & (1u<<HNS3_RXD_VLD_B)) -
2256 * (1u<<HNS3_RXD_VLD_B) will always zero, so the
2257 * assignment is correct.
2259 * So we use the data dependency ordering instead of memory
2260 * barrier to improve receive performance.
2262 rxd = rxdp[(bd_base_info & (1u << HNS3_RXD_VLD_B)) -
2263 (1u << HNS3_RXD_VLD_B)];
2265 nmb = hns3_rx_alloc_buffer(rxq);
2266 if (unlikely(nmb == NULL)) {
2267 dev = &rte_eth_devices[rxq->port_id];
2268 dev->data->rx_mbuf_alloc_failed++;
2273 rxe = &sw_ring[rx_id];
2275 if (unlikely(rx_id == rxq->nb_rx_desc))
2278 rte_prefetch0(sw_ring[rx_id].mbuf);
2279 if ((rx_id & HNS3_RX_RING_PREFETCTH_MASK) == 0) {
2280 rte_prefetch0(&rx_ring[rx_id]);
2281 rte_prefetch0(&sw_ring[rx_id]);
2287 dma_addr = rte_cpu_to_le_64(rte_mbuf_data_iova_default(nmb));
2288 rxdp->rx.bd_base_info = 0;
2289 rxdp->addr = dma_addr;
2291 if (first_seg == NULL) {
2293 first_seg->nb_segs = 1;
2295 first_seg->nb_segs++;
2296 last_seg->next = rxm;
2299 rxm->data_off = RTE_PKTMBUF_HEADROOM;
2300 rxm->data_len = rte_le_to_cpu_16(rxd.rx.size);
2302 if (!(bd_base_info & BIT(HNS3_RXD_FE_B))) {
2309 * The last buffer of the received packet. packet len from
2310 * buffer description may contains CRC len, packet len should
2311 * subtract it, same as data len.
2313 first_seg->pkt_len = rte_le_to_cpu_16(rxd.rx.pkt_len);
2316 * This is the last buffer of the received packet. If the CRC
2317 * is not stripped by the hardware:
2318 * - Subtract the CRC length from the total packet length.
2319 * - If the last buffer only contains the whole CRC or a part
2320 * of it, free the mbuf associated to the last buffer. If part
2321 * of the CRC is also contained in the previous mbuf, subtract
2322 * the length of that CRC part from the data length of the
2326 if (unlikely(rxq->crc_len > 0)) {
2327 first_seg->pkt_len -= rxq->crc_len;
2328 recalculate_data_len(first_seg, last_seg, rxm, rxq,
2332 first_seg->port = rxq->port_id;
2333 first_seg->hash.rss = rte_le_to_cpu_32(rxd.rx.rss_hash);
2334 first_seg->ol_flags = PKT_RX_RSS_HASH;
2335 if (unlikely(bd_base_info & BIT(HNS3_RXD_LUM_B))) {
2336 first_seg->hash.fdir.hi =
2337 rte_le_to_cpu_16(rxd.rx.fd_id);
2338 first_seg->ol_flags |= PKT_RX_FDIR | PKT_RX_FDIR_ID;
2341 gro_size = hns3_get_field(bd_base_info, HNS3_RXD_GRO_SIZE_M,
2342 HNS3_RXD_GRO_SIZE_S);
2343 if (gro_size != 0) {
2344 first_seg->ol_flags |= PKT_RX_LRO;
2345 first_seg->tso_segsz = gro_size;
2348 l234_info = rte_le_to_cpu_32(rxd.rx.l234_info);
2349 ol_info = rte_le_to_cpu_32(rxd.rx.ol_info);
2350 ret = hns3_handle_bdinfo(rxq, first_seg, bd_base_info,
2351 l234_info, &cksum_err);
2355 first_seg->packet_type = hns3_rx_calc_ptype(rxq,
2356 l234_info, ol_info);
2358 if (bd_base_info & BIT(HNS3_RXD_L3L4P_B))
2359 hns3_rx_set_cksum_flag(first_seg,
2360 first_seg->packet_type,
2362 hns3_rxd_to_vlan_tci(rxq, first_seg, l234_info, &rxd);
2364 rx_pkts[nb_rx++] = first_seg;
2368 rte_pktmbuf_free(first_seg);
2372 rxq->next_to_use = rx_id;
2373 rxq->pkt_first_seg = first_seg;
2374 rxq->pkt_last_seg = last_seg;
2376 rxq->rx_free_hold += nb_rx_bd;
2377 if (rxq->rx_free_hold > rxq->rx_free_thresh) {
2378 hns3_write_reg_opt(rxq->io_head_reg, rxq->rx_free_hold);
2379 rxq->rx_free_hold = 0;
2386 hns3_rxq_vec_setup(__rte_unused struct hns3_rx_queue *rxq)
2391 hns3_rx_check_vec_support(__rte_unused struct rte_eth_dev *dev)
2397 hns3_recv_pkts_vec(__rte_unused void *tx_queue,
2398 __rte_unused struct rte_mbuf **rx_pkts,
2399 __rte_unused uint16_t nb_pkts)
2405 hns3_recv_pkts_vec_sve(__rte_unused void *tx_queue,
2406 __rte_unused struct rte_mbuf **rx_pkts,
2407 __rte_unused uint16_t nb_pkts)
2413 hns3_rx_burst_mode_get(struct rte_eth_dev *dev, __rte_unused uint16_t queue_id,
2414 struct rte_eth_burst_mode *mode)
2416 static const struct {
2417 eth_rx_burst_t pkt_burst;
2420 { hns3_recv_pkts, "Scalar" },
2421 { hns3_recv_scattered_pkts, "Scalar Scattered" },
2422 { hns3_recv_pkts_vec, "Vector Neon" },
2423 { hns3_recv_pkts_vec_sve, "Vector Sve" },
2426 eth_rx_burst_t pkt_burst = dev->rx_pkt_burst;
2430 for (i = 0; i < RTE_DIM(burst_infos); i++) {
2431 if (pkt_burst == burst_infos[i].pkt_burst) {
2432 snprintf(mode->info, sizeof(mode->info), "%s",
2433 burst_infos[i].info);
2443 hns3_check_sve_support(void)
2445 #if defined(RTE_ARCH_ARM64) && defined(CC_SVE_SUPPORT)
2446 if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_SVE))
2452 static eth_rx_burst_t
2453 hns3_get_rx_function(struct rte_eth_dev *dev)
2455 struct hns3_adapter *hns = dev->data->dev_private;
2456 uint64_t offloads = dev->data->dev_conf.rxmode.offloads;
2458 if (hns->rx_vec_allowed && hns3_rx_check_vec_support(dev) == 0)
2459 return hns3_check_sve_support() ? hns3_recv_pkts_vec_sve :
2462 if (hns->rx_simple_allowed && !dev->data->scattered_rx &&
2463 (offloads & DEV_RX_OFFLOAD_TCP_LRO) == 0)
2464 return hns3_recv_pkts;
2466 return hns3_recv_scattered_pkts;
2470 hns3_tx_queue_conf_check(struct hns3_hw *hw, const struct rte_eth_txconf *conf,
2471 uint16_t nb_desc, uint16_t *tx_rs_thresh,
2472 uint16_t *tx_free_thresh, uint16_t idx)
2474 #define HNS3_TX_RS_FREE_THRESH_GAP 8
2475 uint16_t rs_thresh, free_thresh, fast_free_thresh;
2477 if (nb_desc > HNS3_MAX_RING_DESC || nb_desc < HNS3_MIN_RING_DESC ||
2478 nb_desc % HNS3_ALIGN_RING_DESC) {
2479 hns3_err(hw, "number (%u) of tx descriptors is invalid",
2484 rs_thresh = (conf->tx_rs_thresh > 0) ?
2485 conf->tx_rs_thresh : HNS3_DEFAULT_TX_RS_THRESH;
2486 free_thresh = (conf->tx_free_thresh > 0) ?
2487 conf->tx_free_thresh : HNS3_DEFAULT_TX_FREE_THRESH;
2488 if (rs_thresh + free_thresh > nb_desc || nb_desc % rs_thresh ||
2489 rs_thresh >= nb_desc - HNS3_TX_RS_FREE_THRESH_GAP ||
2490 free_thresh >= nb_desc - HNS3_TX_RS_FREE_THRESH_GAP) {
2491 hns3_err(hw, "tx_rs_thresh (%d) tx_free_thresh (%d) nb_desc "
2492 "(%d) of tx descriptors for port=%d queue=%d check "
2494 rs_thresh, free_thresh, nb_desc, hw->data->port_id,
2499 if (conf->tx_free_thresh == 0) {
2500 /* Fast free Tx memory buffer to improve cache hit rate */
2501 fast_free_thresh = nb_desc - rs_thresh;
2502 if (fast_free_thresh >=
2503 HNS3_TX_FAST_FREE_AHEAD + HNS3_DEFAULT_TX_FREE_THRESH)
2504 free_thresh = fast_free_thresh -
2505 HNS3_TX_FAST_FREE_AHEAD;
2508 *tx_rs_thresh = rs_thresh;
2509 *tx_free_thresh = free_thresh;
2514 hns3_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t nb_desc,
2515 unsigned int socket_id, const struct rte_eth_txconf *conf)
2517 struct hns3_adapter *hns = dev->data->dev_private;
2518 uint16_t tx_rs_thresh, tx_free_thresh;
2519 struct hns3_hw *hw = &hns->hw;
2520 struct hns3_queue_info q_info;
2521 struct hns3_tx_queue *txq;
2525 ret = hns3_tx_queue_conf_check(hw, conf, nb_desc,
2526 &tx_rs_thresh, &tx_free_thresh, idx);
2530 if (dev->data->tx_queues[idx] != NULL) {
2531 hns3_tx_queue_release(dev->data->tx_queues[idx]);
2532 dev->data->tx_queues[idx] = NULL;
2536 q_info.socket_id = socket_id;
2537 q_info.nb_desc = nb_desc;
2538 q_info.type = "hns3 TX queue";
2539 q_info.ring_name = "tx_ring";
2540 txq = hns3_alloc_txq_and_dma_zone(dev, &q_info);
2543 "Failed to alloc mem and reserve DMA mem for tx ring!");
2547 txq->tx_deferred_start = conf->tx_deferred_start;
2548 if (txq->tx_deferred_start && !hns3_dev_indep_txrx_supported(hw)) {
2549 hns3_warn(hw, "deferred start is not supported.");
2550 txq->tx_deferred_start = false;
2553 tx_entry_len = sizeof(struct hns3_entry) * txq->nb_tx_desc;
2554 txq->sw_ring = rte_zmalloc_socket("hns3 TX sw ring", tx_entry_len,
2555 RTE_CACHE_LINE_SIZE, socket_id);
2556 if (txq->sw_ring == NULL) {
2557 hns3_err(hw, "Failed to allocate memory for tx sw ring!");
2558 hns3_tx_queue_release(txq);
2563 txq->next_to_use = 0;
2564 txq->next_to_clean = 0;
2565 txq->tx_bd_ready = txq->nb_tx_desc - 1;
2566 txq->tx_free_thresh = tx_free_thresh;
2567 txq->tx_rs_thresh = tx_rs_thresh;
2568 txq->free = rte_zmalloc_socket("hns3 TX mbuf free array",
2569 sizeof(struct rte_mbuf *) * txq->tx_rs_thresh,
2570 RTE_CACHE_LINE_SIZE, socket_id);
2572 hns3_err(hw, "failed to allocate tx mbuf free array!");
2573 hns3_tx_queue_release(txq);
2577 txq->port_id = dev->data->port_id;
2579 * For hns3 PF device, if the VLAN mode is HW_SHIFT_AND_DISCARD_MODE,
2580 * the pvid_sw_shift_en in the queue struct should not be changed,
2581 * because PVID-related operations do not need to be processed by PMD
2582 * driver. For hns3 VF device, whether it needs to process PVID depends
2583 * on the configuration of PF kernel mode netdev driver. And the
2584 * related PF configuration is delivered through the mailbox and finally
2585 * reflectd in port_base_vlan_cfg.
2587 if (hns->is_vf || hw->vlan_mode == HNS3_SW_SHIFT_AND_DISCARD_MODE)
2588 txq->pvid_sw_shift_en = hw->port_base_vlan_cfg.state ==
2589 HNS3_PORT_BASE_VLAN_ENABLE;
2591 txq->pvid_sw_shift_en = false;
2592 txq->max_non_tso_bd_num = hw->max_non_tso_bd_num;
2593 txq->configured = true;
2594 txq->io_base = (void *)((char *)hw->io_base +
2595 hns3_get_tqp_reg_offset(idx));
2596 txq->io_tail_reg = (volatile void *)((char *)txq->io_base +
2597 HNS3_RING_TX_TAIL_REG);
2598 txq->min_tx_pkt_len = hw->min_tx_pkt_len;
2599 txq->tso_mode = hw->tso_mode;
2600 txq->over_length_pkt_cnt = 0;
2601 txq->exceed_limit_bd_pkt_cnt = 0;
2602 txq->exceed_limit_bd_reassem_fail = 0;
2603 txq->unsupported_tunnel_pkt_cnt = 0;
2604 txq->queue_full_cnt = 0;
2605 txq->pkt_padding_fail_cnt = 0;
2606 rte_spinlock_lock(&hw->lock);
2607 dev->data->tx_queues[idx] = txq;
2608 rte_spinlock_unlock(&hw->lock);
2614 hns3_tx_free_useless_buffer(struct hns3_tx_queue *txq)
2616 uint16_t tx_next_clean = txq->next_to_clean;
2617 uint16_t tx_next_use = txq->next_to_use;
2618 uint16_t tx_bd_ready = txq->tx_bd_ready;
2619 uint16_t tx_bd_max = txq->nb_tx_desc;
2620 struct hns3_entry *tx_bak_pkt = &txq->sw_ring[tx_next_clean];
2621 struct hns3_desc *desc = &txq->tx_ring[tx_next_clean];
2622 struct rte_mbuf *mbuf;
2624 while ((!(desc->tx.tp_fe_sc_vld_ra_ri &
2625 rte_cpu_to_le_16(BIT(HNS3_TXD_VLD_B)))) &&
2626 tx_next_use != tx_next_clean) {
2627 mbuf = tx_bak_pkt->mbuf;
2629 rte_pktmbuf_free_seg(mbuf);
2630 tx_bak_pkt->mbuf = NULL;
2638 if (tx_next_clean >= tx_bd_max) {
2640 desc = txq->tx_ring;
2641 tx_bak_pkt = txq->sw_ring;
2645 txq->next_to_clean = tx_next_clean;
2646 txq->tx_bd_ready = tx_bd_ready;
2650 hns3_config_gro(struct hns3_hw *hw, bool en)
2652 struct hns3_cfg_gro_status_cmd *req;
2653 struct hns3_cmd_desc desc;
2656 hns3_cmd_setup_basic_desc(&desc, HNS3_OPC_GRO_GENERIC_CONFIG, false);
2657 req = (struct hns3_cfg_gro_status_cmd *)desc.data;
2659 req->gro_en = rte_cpu_to_le_16(en ? 1 : 0);
2661 ret = hns3_cmd_send(hw, &desc, 1);
2663 hns3_err(hw, "%s hardware GRO failed, ret = %d",
2664 en ? "enable" : "disable", ret);
2670 hns3_restore_gro_conf(struct hns3_hw *hw)
2676 offloads = hw->data->dev_conf.rxmode.offloads;
2677 gro_en = offloads & DEV_RX_OFFLOAD_TCP_LRO ? true : false;
2678 ret = hns3_config_gro(hw, gro_en);
2680 hns3_err(hw, "restore hardware GRO to %s failed, ret = %d",
2681 gro_en ? "enabled" : "disabled", ret);
2687 hns3_pkt_is_tso(struct rte_mbuf *m)
2689 return (m->tso_segsz != 0 && m->ol_flags & PKT_TX_TCP_SEG);
2693 hns3_set_tso(struct hns3_desc *desc, uint32_t paylen, struct rte_mbuf *rxm)
2695 if (!hns3_pkt_is_tso(rxm))
2698 if (paylen <= rxm->tso_segsz)
2701 desc->tx.type_cs_vlan_tso_len |= rte_cpu_to_le_32(BIT(HNS3_TXD_TSO_B));
2702 desc->tx.mss = rte_cpu_to_le_16(rxm->tso_segsz);
2706 hns3_fill_per_desc(struct hns3_desc *desc, struct rte_mbuf *rxm)
2708 desc->addr = rte_mbuf_data_iova(rxm);
2709 desc->tx.send_size = rte_cpu_to_le_16(rte_pktmbuf_data_len(rxm));
2710 desc->tx.tp_fe_sc_vld_ra_ri = rte_cpu_to_le_16(BIT(HNS3_TXD_VLD_B));
2714 hns3_fill_first_desc(struct hns3_tx_queue *txq, struct hns3_desc *desc,
2715 struct rte_mbuf *rxm)
2717 uint64_t ol_flags = rxm->ol_flags;
2721 hdr_len = rxm->l2_len + rxm->l3_len + rxm->l4_len;
2722 hdr_len += (ol_flags & PKT_TX_TUNNEL_MASK) ?
2723 rxm->outer_l2_len + rxm->outer_l3_len : 0;
2724 paylen = rxm->pkt_len - hdr_len;
2725 desc->tx.paylen = rte_cpu_to_le_32(paylen);
2726 hns3_set_tso(desc, paylen, rxm);
2729 * Currently, hardware doesn't support more than two layers VLAN offload
2730 * in Tx direction based on hns3 network engine. So when the number of
2731 * VLANs in the packets represented by rxm plus the number of VLAN
2732 * offload by hardware such as PVID etc, exceeds two, the packets will
2733 * be discarded or the original VLAN of the packets will be overwitted
2734 * by hardware. When the PF PVID is enabled by calling the API function
2735 * named rte_eth_dev_set_vlan_pvid or the VF PVID is enabled by the hns3
2736 * PF kernel ether driver, the outer VLAN tag will always be the PVID.
2737 * To avoid the VLAN of Tx descriptor is overwritten by PVID, it should
2738 * be added to the position close to the IP header when PVID is enabled.
2740 if (!txq->pvid_sw_shift_en && ol_flags & (PKT_TX_VLAN_PKT |
2742 desc->tx.ol_type_vlan_len_msec |=
2743 rte_cpu_to_le_32(BIT(HNS3_TXD_OVLAN_B));
2744 if (ol_flags & PKT_TX_QINQ_PKT)
2745 desc->tx.outer_vlan_tag =
2746 rte_cpu_to_le_16(rxm->vlan_tci_outer);
2748 desc->tx.outer_vlan_tag =
2749 rte_cpu_to_le_16(rxm->vlan_tci);
2752 if (ol_flags & PKT_TX_QINQ_PKT ||
2753 ((ol_flags & PKT_TX_VLAN_PKT) && txq->pvid_sw_shift_en)) {
2754 desc->tx.type_cs_vlan_tso_len |=
2755 rte_cpu_to_le_32(BIT(HNS3_TXD_VLAN_B));
2756 desc->tx.vlan_tag = rte_cpu_to_le_16(rxm->vlan_tci);
2761 hns3_tx_alloc_mbufs(struct rte_mempool *mb_pool, uint16_t nb_new_buf,
2762 struct rte_mbuf **alloc_mbuf)
2764 #define MAX_NON_TSO_BD_PER_PKT 18
2765 struct rte_mbuf *pkt_segs[MAX_NON_TSO_BD_PER_PKT];
2768 /* Allocate enough mbufs */
2769 if (rte_mempool_get_bulk(mb_pool, (void **)pkt_segs, nb_new_buf))
2772 for (i = 0; i < nb_new_buf - 1; i++)
2773 pkt_segs[i]->next = pkt_segs[i + 1];
2775 pkt_segs[nb_new_buf - 1]->next = NULL;
2776 pkt_segs[0]->nb_segs = nb_new_buf;
2777 *alloc_mbuf = pkt_segs[0];
2783 hns3_pktmbuf_copy_hdr(struct rte_mbuf *new_pkt, struct rte_mbuf *old_pkt)
2785 new_pkt->ol_flags = old_pkt->ol_flags;
2786 new_pkt->pkt_len = rte_pktmbuf_pkt_len(old_pkt);
2787 new_pkt->outer_l2_len = old_pkt->outer_l2_len;
2788 new_pkt->outer_l3_len = old_pkt->outer_l3_len;
2789 new_pkt->l2_len = old_pkt->l2_len;
2790 new_pkt->l3_len = old_pkt->l3_len;
2791 new_pkt->l4_len = old_pkt->l4_len;
2792 new_pkt->vlan_tci_outer = old_pkt->vlan_tci_outer;
2793 new_pkt->vlan_tci = old_pkt->vlan_tci;
2797 hns3_reassemble_tx_pkts(struct rte_mbuf *tx_pkt, struct rte_mbuf **new_pkt,
2798 uint8_t max_non_tso_bd_num)
2800 struct rte_mempool *mb_pool;
2801 struct rte_mbuf *new_mbuf;
2802 struct rte_mbuf *temp_new;
2803 struct rte_mbuf *temp;
2804 uint16_t last_buf_len;
2805 uint16_t nb_new_buf;
2815 mb_pool = tx_pkt->pool;
2816 buf_size = tx_pkt->buf_len - RTE_PKTMBUF_HEADROOM;
2817 nb_new_buf = (rte_pktmbuf_pkt_len(tx_pkt) - 1) / buf_size + 1;
2818 if (nb_new_buf > max_non_tso_bd_num)
2821 last_buf_len = rte_pktmbuf_pkt_len(tx_pkt) % buf_size;
2822 if (last_buf_len == 0)
2823 last_buf_len = buf_size;
2825 /* Allocate enough mbufs */
2826 ret = hns3_tx_alloc_mbufs(mb_pool, nb_new_buf, &new_mbuf);
2830 /* Copy the original packet content to the new mbufs */
2832 s = rte_pktmbuf_mtod(temp, char *);
2833 len_s = rte_pktmbuf_data_len(temp);
2834 temp_new = new_mbuf;
2835 while (temp != NULL && temp_new != NULL) {
2836 d = rte_pktmbuf_mtod(temp_new, char *);
2837 buf_len = temp_new->next == NULL ? last_buf_len : buf_size;
2841 len = RTE_MIN(len_s, len_d);
2845 len_d = len_d - len;
2846 len_s = len_s - len;
2852 s = rte_pktmbuf_mtod(temp, char *);
2853 len_s = rte_pktmbuf_data_len(temp);
2857 temp_new->data_len = buf_len;
2858 temp_new = temp_new->next;
2860 hns3_pktmbuf_copy_hdr(new_mbuf, tx_pkt);
2862 /* free original mbufs */
2863 rte_pktmbuf_free(tx_pkt);
2865 *new_pkt = new_mbuf;
2871 hns3_parse_outer_params(struct rte_mbuf *m, uint32_t *ol_type_vlan_len_msec)
2873 uint32_t tmp = *ol_type_vlan_len_msec;
2874 uint64_t ol_flags = m->ol_flags;
2876 /* (outer) IP header type */
2877 if (ol_flags & PKT_TX_OUTER_IPV4) {
2878 if (ol_flags & PKT_TX_OUTER_IP_CKSUM)
2879 tmp |= hns3_gen_field_val(HNS3_TXD_OL3T_M,
2880 HNS3_TXD_OL3T_S, HNS3_OL3T_IPV4_CSUM);
2882 tmp |= hns3_gen_field_val(HNS3_TXD_OL3T_M,
2883 HNS3_TXD_OL3T_S, HNS3_OL3T_IPV4_NO_CSUM);
2884 } else if (ol_flags & PKT_TX_OUTER_IPV6) {
2885 tmp |= hns3_gen_field_val(HNS3_TXD_OL3T_M, HNS3_TXD_OL3T_S,
2888 /* OL3 header size, defined in 4 bytes */
2889 tmp |= hns3_gen_field_val(HNS3_TXD_L3LEN_M, HNS3_TXD_L3LEN_S,
2890 m->outer_l3_len >> HNS3_L3_LEN_UNIT);
2891 *ol_type_vlan_len_msec = tmp;
2895 hns3_parse_inner_params(struct rte_mbuf *m, uint32_t *ol_type_vlan_len_msec,
2896 uint32_t *type_cs_vlan_tso_len)
2898 #define HNS3_NVGRE_HLEN 8
2899 uint32_t tmp_outer = *ol_type_vlan_len_msec;
2900 uint32_t tmp_inner = *type_cs_vlan_tso_len;
2901 uint64_t ol_flags = m->ol_flags;
2902 uint16_t inner_l2_len;
2904 switch (ol_flags & PKT_TX_TUNNEL_MASK) {
2905 case PKT_TX_TUNNEL_VXLAN_GPE:
2906 case PKT_TX_TUNNEL_GENEVE:
2907 case PKT_TX_TUNNEL_VXLAN:
2908 /* MAC in UDP tunnelling packet, include VxLAN and GENEVE */
2909 tmp_outer |= hns3_gen_field_val(HNS3_TXD_TUNTYPE_M,
2910 HNS3_TXD_TUNTYPE_S, HNS3_TUN_MAC_IN_UDP);
2912 * The inner l2 length of mbuf is the sum of outer l4 length,
2913 * tunneling header length and inner l2 length for a tunnel
2914 * packect. But in hns3 tx descriptor, the tunneling header
2915 * length is contained in the field of outer L4 length.
2916 * Therefore, driver need to calculate the outer L4 length and
2919 tmp_outer |= hns3_gen_field_val(HNS3_TXD_L4LEN_M,
2921 (uint8_t)RTE_ETHER_VXLAN_HLEN >>
2924 inner_l2_len = m->l2_len - RTE_ETHER_VXLAN_HLEN;
2926 case PKT_TX_TUNNEL_GRE:
2927 tmp_outer |= hns3_gen_field_val(HNS3_TXD_TUNTYPE_M,
2928 HNS3_TXD_TUNTYPE_S, HNS3_TUN_NVGRE);
2930 * For NVGRE tunnel packect, the outer L4 is empty. So only
2931 * fill the NVGRE header length to the outer L4 field.
2933 tmp_outer |= hns3_gen_field_val(HNS3_TXD_L4LEN_M,
2935 (uint8_t)HNS3_NVGRE_HLEN >> HNS3_L4_LEN_UNIT);
2937 inner_l2_len = m->l2_len - HNS3_NVGRE_HLEN;
2940 /* For non UDP / GRE tunneling, drop the tunnel packet */
2944 tmp_inner |= hns3_gen_field_val(HNS3_TXD_L2LEN_M, HNS3_TXD_L2LEN_S,
2945 inner_l2_len >> HNS3_L2_LEN_UNIT);
2946 /* OL2 header size, defined in 2 bytes */
2947 tmp_outer |= hns3_gen_field_val(HNS3_TXD_L2LEN_M, HNS3_TXD_L2LEN_S,
2948 m->outer_l2_len >> HNS3_L2_LEN_UNIT);
2950 *type_cs_vlan_tso_len = tmp_inner;
2951 *ol_type_vlan_len_msec = tmp_outer;
2957 hns3_parse_tunneling_params(struct hns3_tx_queue *txq, struct rte_mbuf *m,
2958 uint16_t tx_desc_id)
2960 struct hns3_desc *tx_ring = txq->tx_ring;
2961 struct hns3_desc *desc = &tx_ring[tx_desc_id];
2962 uint32_t tmp_outer = 0;
2963 uint32_t tmp_inner = 0;
2967 * The tunnel header is contained in the inner L2 header field of the
2968 * mbuf, but for hns3 descriptor, it is contained in the outer L4. So,
2969 * there is a need that switching between them. To avoid multiple
2970 * calculations, the length of the L2 header include the outer and
2971 * inner, will be filled during the parsing of tunnel packects.
2973 if (!(m->ol_flags & PKT_TX_TUNNEL_MASK)) {
2975 * For non tunnel type the tunnel type id is 0, so no need to
2976 * assign a value to it. Only the inner(normal) L2 header length
2979 tmp_inner |= hns3_gen_field_val(HNS3_TXD_L2LEN_M,
2980 HNS3_TXD_L2LEN_S, m->l2_len >> HNS3_L2_LEN_UNIT);
2983 * If outer csum is not offload, the outer length may be filled
2984 * with 0. And the length of the outer header is added to the
2985 * inner l2_len. It would lead a cksum error. So driver has to
2986 * calculate the header length.
2988 if (unlikely(!(m->ol_flags & PKT_TX_OUTER_IP_CKSUM) &&
2989 m->outer_l2_len == 0)) {
2990 struct rte_net_hdr_lens hdr_len;
2991 (void)rte_net_get_ptype(m, &hdr_len,
2992 RTE_PTYPE_L2_MASK | RTE_PTYPE_L3_MASK);
2993 m->outer_l3_len = hdr_len.l3_len;
2994 m->outer_l2_len = hdr_len.l2_len;
2995 m->l2_len = m->l2_len - hdr_len.l2_len - hdr_len.l3_len;
2997 hns3_parse_outer_params(m, &tmp_outer);
2998 ret = hns3_parse_inner_params(m, &tmp_outer, &tmp_inner);
3003 desc->tx.ol_type_vlan_len_msec = rte_cpu_to_le_32(tmp_outer);
3004 desc->tx.type_cs_vlan_tso_len = rte_cpu_to_le_32(tmp_inner);
3010 hns3_parse_l3_cksum_params(struct rte_mbuf *m, uint32_t *type_cs_vlan_tso_len)
3012 uint64_t ol_flags = m->ol_flags;
3016 tmp = *type_cs_vlan_tso_len;
3017 if (ol_flags & PKT_TX_IPV4)
3018 l3_type = HNS3_L3T_IPV4;
3019 else if (ol_flags & PKT_TX_IPV6)
3020 l3_type = HNS3_L3T_IPV6;
3022 l3_type = HNS3_L3T_NONE;
3024 /* inner(/normal) L3 header size, defined in 4 bytes */
3025 tmp |= hns3_gen_field_val(HNS3_TXD_L3LEN_M, HNS3_TXD_L3LEN_S,
3026 m->l3_len >> HNS3_L3_LEN_UNIT);
3028 tmp |= hns3_gen_field_val(HNS3_TXD_L3T_M, HNS3_TXD_L3T_S, l3_type);
3030 /* Enable L3 checksum offloads */
3031 if (ol_flags & PKT_TX_IP_CKSUM)
3032 tmp |= BIT(HNS3_TXD_L3CS_B);
3033 *type_cs_vlan_tso_len = tmp;
3037 hns3_parse_l4_cksum_params(struct rte_mbuf *m, uint32_t *type_cs_vlan_tso_len)
3039 uint64_t ol_flags = m->ol_flags;
3041 /* Enable L4 checksum offloads */
3042 switch (ol_flags & (PKT_TX_L4_MASK | PKT_TX_TCP_SEG)) {
3043 case PKT_TX_TCP_CKSUM:
3044 case PKT_TX_TCP_SEG:
3045 tmp = *type_cs_vlan_tso_len;
3046 tmp |= hns3_gen_field_val(HNS3_TXD_L4T_M, HNS3_TXD_L4T_S,
3049 case PKT_TX_UDP_CKSUM:
3050 tmp = *type_cs_vlan_tso_len;
3051 tmp |= hns3_gen_field_val(HNS3_TXD_L4T_M, HNS3_TXD_L4T_S,
3054 case PKT_TX_SCTP_CKSUM:
3055 tmp = *type_cs_vlan_tso_len;
3056 tmp |= hns3_gen_field_val(HNS3_TXD_L4T_M, HNS3_TXD_L4T_S,
3062 tmp |= BIT(HNS3_TXD_L4CS_B);
3063 tmp |= hns3_gen_field_val(HNS3_TXD_L4LEN_M, HNS3_TXD_L4LEN_S,
3064 m->l4_len >> HNS3_L4_LEN_UNIT);
3065 *type_cs_vlan_tso_len = tmp;
3069 hns3_txd_enable_checksum(struct hns3_tx_queue *txq, struct rte_mbuf *m,
3070 uint16_t tx_desc_id)
3072 struct hns3_desc *tx_ring = txq->tx_ring;
3073 struct hns3_desc *desc = &tx_ring[tx_desc_id];
3076 hns3_parse_l3_cksum_params(m, &value);
3077 hns3_parse_l4_cksum_params(m, &value);
3079 desc->tx.type_cs_vlan_tso_len |= rte_cpu_to_le_32(value);
3083 hns3_pkt_need_linearized(struct rte_mbuf *tx_pkts, uint32_t bd_num,
3084 uint32_t max_non_tso_bd_num)
3086 struct rte_mbuf *m_first = tx_pkts;
3087 struct rte_mbuf *m_last = tx_pkts;
3088 uint32_t tot_len = 0;
3093 * Hardware requires that the sum of the data length of every 8
3094 * consecutive buffers is greater than MSS in hns3 network engine.
3095 * We simplify it by ensuring pkt_headlen + the first 8 consecutive
3096 * frags greater than gso header len + mss, and the remaining 7
3097 * consecutive frags greater than MSS except the last 7 frags.
3099 if (bd_num <= max_non_tso_bd_num)
3102 for (i = 0; m_last && i < max_non_tso_bd_num - 1;
3103 i++, m_last = m_last->next)
3104 tot_len += m_last->data_len;
3109 /* ensure the first 8 frags is greater than mss + header */
3110 hdr_len = tx_pkts->l2_len + tx_pkts->l3_len + tx_pkts->l4_len;
3111 hdr_len += (tx_pkts->ol_flags & PKT_TX_TUNNEL_MASK) ?
3112 tx_pkts->outer_l2_len + tx_pkts->outer_l3_len : 0;
3113 if (tot_len + m_last->data_len < tx_pkts->tso_segsz + hdr_len)
3117 * ensure the sum of the data length of every 7 consecutive buffer
3118 * is greater than mss except the last one.
3120 for (i = 0; m_last && i < bd_num - max_non_tso_bd_num; i++) {
3121 tot_len -= m_first->data_len;
3122 tot_len += m_last->data_len;
3124 if (tot_len < tx_pkts->tso_segsz)
3127 m_first = m_first->next;
3128 m_last = m_last->next;
3135 hns3_outer_header_cksum_prepare(struct rte_mbuf *m)
3137 uint64_t ol_flags = m->ol_flags;
3138 uint32_t paylen, hdr_len, l4_proto;
3140 if (!(ol_flags & (PKT_TX_OUTER_IPV4 | PKT_TX_OUTER_IPV6)))
3143 if (ol_flags & PKT_TX_OUTER_IPV4) {
3144 struct rte_ipv4_hdr *ipv4_hdr;
3145 ipv4_hdr = rte_pktmbuf_mtod_offset(m, struct rte_ipv4_hdr *,
3147 l4_proto = ipv4_hdr->next_proto_id;
3148 if (ol_flags & PKT_TX_OUTER_IP_CKSUM)
3149 ipv4_hdr->hdr_checksum = 0;
3151 struct rte_ipv6_hdr *ipv6_hdr;
3152 ipv6_hdr = rte_pktmbuf_mtod_offset(m, struct rte_ipv6_hdr *,
3154 l4_proto = ipv6_hdr->proto;
3156 /* driver should ensure the outer udp cksum is 0 for TUNNEL TSO */
3157 if (l4_proto == IPPROTO_UDP && (ol_flags & PKT_TX_TCP_SEG)) {
3158 struct rte_udp_hdr *udp_hdr;
3159 hdr_len = m->l2_len + m->l3_len + m->l4_len;
3160 hdr_len += m->outer_l2_len + m->outer_l3_len;
3161 paylen = m->pkt_len - hdr_len;
3162 if (paylen <= m->tso_segsz)
3164 udp_hdr = rte_pktmbuf_mtod_offset(m, struct rte_udp_hdr *,
3167 udp_hdr->dgram_cksum = 0;
3172 hns3_check_tso_pkt_valid(struct rte_mbuf *m)
3174 uint32_t tmp_data_len_sum = 0;
3175 uint16_t nb_buf = m->nb_segs;
3176 uint32_t paylen, hdr_len;
3177 struct rte_mbuf *m_seg;
3180 if (nb_buf > HNS3_MAX_TSO_BD_PER_PKT)
3183 hdr_len = m->l2_len + m->l3_len + m->l4_len;
3184 hdr_len += (m->ol_flags & PKT_TX_TUNNEL_MASK) ?
3185 m->outer_l2_len + m->outer_l3_len : 0;
3186 if (hdr_len > HNS3_MAX_TSO_HDR_SIZE)
3189 paylen = m->pkt_len - hdr_len;
3190 if (paylen > HNS3_MAX_BD_PAYLEN)
3194 * The TSO header (include outer and inner L2, L3 and L4 header)
3195 * should be provided by three descriptors in maximum in hns3 network
3199 for (i = 0; m_seg != NULL && i < HNS3_MAX_TSO_HDR_BD_NUM && i < nb_buf;
3200 i++, m_seg = m_seg->next) {
3201 tmp_data_len_sum += m_seg->data_len;
3204 if (hdr_len > tmp_data_len_sum)
3210 #ifdef RTE_LIBRTE_ETHDEV_DEBUG
3212 hns3_vld_vlan_chk(struct hns3_tx_queue *txq, struct rte_mbuf *m)
3214 struct rte_ether_hdr *eh;
3215 struct rte_vlan_hdr *vh;
3217 if (!txq->pvid_sw_shift_en)
3221 * Due to hardware limitations, we only support two-layer VLAN hardware
3222 * offload in Tx direction based on hns3 network engine, so when PVID is
3223 * enabled, QinQ insert is no longer supported.
3224 * And when PVID is enabled, in the following two cases:
3225 * i) packets with more than two VLAN tags.
3226 * ii) packets with one VLAN tag while the hardware VLAN insert is
3228 * The packets will be regarded as abnormal packets and discarded by
3229 * hardware in Tx direction. For debugging purposes, a validation check
3230 * for these types of packets is added to the '.tx_pkt_prepare' ops
3231 * implementation function named hns3_prep_pkts to inform users that
3232 * these packets will be discarded.
3234 if (m->ol_flags & PKT_TX_QINQ_PKT)
3237 eh = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
3238 if (eh->ether_type == rte_cpu_to_be_16(RTE_ETHER_TYPE_VLAN)) {
3239 if (m->ol_flags & PKT_TX_VLAN_PKT)
3242 /* Ensure the incoming packet is not a QinQ packet */
3243 vh = (struct rte_vlan_hdr *)(eh + 1);
3244 if (vh->eth_proto == rte_cpu_to_be_16(RTE_ETHER_TYPE_VLAN))
3253 hns3_prep_pkt_proc(struct hns3_tx_queue *tx_queue, struct rte_mbuf *m)
3257 #ifdef RTE_LIBRTE_ETHDEV_DEBUG
3258 ret = rte_validate_tx_offload(m);
3264 ret = hns3_vld_vlan_chk(tx_queue, m);
3270 if (hns3_pkt_is_tso(m)) {
3271 if (hns3_pkt_need_linearized(m, m->nb_segs,
3272 tx_queue->max_non_tso_bd_num) ||
3273 hns3_check_tso_pkt_valid(m)) {
3278 if (tx_queue->tso_mode != HNS3_TSO_SW_CAL_PSEUDO_H_CSUM) {
3280 * (tso mode != HNS3_TSO_SW_CAL_PSEUDO_H_CSUM) means
3281 * hardware support recalculate the TCP pseudo header
3282 * checksum of packets that need TSO, so network driver
3283 * software not need to recalculate it.
3285 hns3_outer_header_cksum_prepare(m);
3290 ret = rte_net_intel_cksum_prepare(m);
3296 hns3_outer_header_cksum_prepare(m);
3302 hns3_prep_pkts(__rte_unused void *tx_queue, struct rte_mbuf **tx_pkts,
3308 for (i = 0; i < nb_pkts; i++) {
3310 if (hns3_prep_pkt_proc(tx_queue, m))
3318 hns3_parse_cksum(struct hns3_tx_queue *txq, uint16_t tx_desc_id,
3321 struct hns3_desc *tx_ring = txq->tx_ring;
3322 struct hns3_desc *desc = &tx_ring[tx_desc_id];
3324 /* Enable checksum offloading */
3325 if (m->ol_flags & HNS3_TX_CKSUM_OFFLOAD_MASK) {
3326 /* Fill in tunneling parameters if necessary */
3327 if (hns3_parse_tunneling_params(txq, m, tx_desc_id)) {
3328 txq->unsupported_tunnel_pkt_cnt++;
3332 hns3_txd_enable_checksum(txq, m, tx_desc_id);
3334 /* clear the control bit */
3335 desc->tx.type_cs_vlan_tso_len = 0;
3336 desc->tx.ol_type_vlan_len_msec = 0;
3343 hns3_check_non_tso_pkt(uint16_t nb_buf, struct rte_mbuf **m_seg,
3344 struct rte_mbuf *tx_pkt, struct hns3_tx_queue *txq)
3346 uint8_t max_non_tso_bd_num;
3347 struct rte_mbuf *new_pkt;
3350 if (hns3_pkt_is_tso(*m_seg))
3354 * If packet length is greater than HNS3_MAX_FRAME_LEN
3355 * driver support, the packet will be ignored.
3357 if (unlikely(rte_pktmbuf_pkt_len(tx_pkt) > HNS3_MAX_FRAME_LEN)) {
3358 txq->over_length_pkt_cnt++;
3362 max_non_tso_bd_num = txq->max_non_tso_bd_num;
3363 if (unlikely(nb_buf > max_non_tso_bd_num)) {
3364 txq->exceed_limit_bd_pkt_cnt++;
3365 ret = hns3_reassemble_tx_pkts(tx_pkt, &new_pkt,
3366 max_non_tso_bd_num);
3368 txq->exceed_limit_bd_reassem_fail++;
3378 hns3_tx_free_buffer_simple(struct hns3_tx_queue *txq)
3380 struct hns3_entry *tx_entry;
3381 struct hns3_desc *desc;
3382 uint16_t tx_next_clean;
3386 if (HNS3_GET_TX_QUEUE_PEND_BD_NUM(txq) < txq->tx_rs_thresh)
3390 * All mbufs can be released only when the VLD bits of all
3391 * descriptors in a batch are cleared.
3393 tx_next_clean = (txq->next_to_clean + txq->tx_rs_thresh - 1) %
3395 desc = &txq->tx_ring[tx_next_clean];
3396 for (i = 0; i < txq->tx_rs_thresh; i++) {
3397 if (rte_le_to_cpu_16(desc->tx.tp_fe_sc_vld_ra_ri) &
3398 BIT(HNS3_TXD_VLD_B))
3403 tx_entry = &txq->sw_ring[txq->next_to_clean];
3405 for (i = 0; i < txq->tx_rs_thresh; i++)
3406 rte_prefetch0((tx_entry + i)->mbuf);
3407 for (i = 0; i < txq->tx_rs_thresh; i++, tx_entry++) {
3408 rte_mempool_put(tx_entry->mbuf->pool, tx_entry->mbuf);
3409 tx_entry->mbuf = NULL;
3412 txq->next_to_clean = (tx_next_clean + 1) % txq->nb_tx_desc;
3413 txq->tx_bd_ready += txq->tx_rs_thresh;
3418 hns3_tx_backup_1mbuf(struct hns3_entry *tx_entry, struct rte_mbuf **pkts)
3420 tx_entry->mbuf = pkts[0];
3424 hns3_tx_backup_4mbuf(struct hns3_entry *tx_entry, struct rte_mbuf **pkts)
3426 hns3_tx_backup_1mbuf(&tx_entry[0], &pkts[0]);
3427 hns3_tx_backup_1mbuf(&tx_entry[1], &pkts[1]);
3428 hns3_tx_backup_1mbuf(&tx_entry[2], &pkts[2]);
3429 hns3_tx_backup_1mbuf(&tx_entry[3], &pkts[3]);
3433 hns3_tx_setup_4bd(struct hns3_desc *txdp, struct rte_mbuf **pkts)
3435 #define PER_LOOP_NUM 4
3436 const uint16_t bd_flag = BIT(HNS3_TXD_VLD_B) | BIT(HNS3_TXD_FE_B);
3440 for (i = 0; i < PER_LOOP_NUM; i++, txdp++, pkts++) {
3441 dma_addr = rte_mbuf_data_iova(*pkts);
3442 txdp->addr = rte_cpu_to_le_64(dma_addr);
3443 txdp->tx.send_size = rte_cpu_to_le_16((*pkts)->data_len);
3444 txdp->tx.paylen = 0;
3445 txdp->tx.type_cs_vlan_tso_len = 0;
3446 txdp->tx.ol_type_vlan_len_msec = 0;
3447 txdp->tx.tp_fe_sc_vld_ra_ri = rte_cpu_to_le_16(bd_flag);
3452 hns3_tx_setup_1bd(struct hns3_desc *txdp, struct rte_mbuf **pkts)
3454 const uint16_t bd_flag = BIT(HNS3_TXD_VLD_B) | BIT(HNS3_TXD_FE_B);
3457 dma_addr = rte_mbuf_data_iova(*pkts);
3458 txdp->addr = rte_cpu_to_le_64(dma_addr);
3459 txdp->tx.send_size = rte_cpu_to_le_16((*pkts)->data_len);
3460 txdp->tx.paylen = 0;
3461 txdp->tx.type_cs_vlan_tso_len = 0;
3462 txdp->tx.ol_type_vlan_len_msec = 0;
3463 txdp->tx.tp_fe_sc_vld_ra_ri = rte_cpu_to_le_16(bd_flag);
3467 hns3_tx_fill_hw_ring(struct hns3_tx_queue *txq,
3468 struct rte_mbuf **pkts,
3471 #define PER_LOOP_NUM 4
3472 #define PER_LOOP_MASK (PER_LOOP_NUM - 1)
3473 struct hns3_desc *txdp = &txq->tx_ring[txq->next_to_use];
3474 struct hns3_entry *tx_entry = &txq->sw_ring[txq->next_to_use];
3475 const uint32_t mainpart = (nb_pkts & ((uint32_t)~PER_LOOP_MASK));
3476 const uint32_t leftover = (nb_pkts & ((uint32_t)PER_LOOP_MASK));
3479 for (i = 0; i < mainpart; i += PER_LOOP_NUM) {
3480 hns3_tx_backup_4mbuf(tx_entry + i, pkts + i);
3481 hns3_tx_setup_4bd(txdp + i, pkts + i);
3483 if (unlikely(leftover > 0)) {
3484 for (i = 0; i < leftover; i++) {
3485 hns3_tx_backup_1mbuf(tx_entry + mainpart + i,
3486 pkts + mainpart + i);
3487 hns3_tx_setup_1bd(txdp + mainpart + i,
3488 pkts + mainpart + i);
3494 hns3_xmit_pkts_simple(void *tx_queue,
3495 struct rte_mbuf **tx_pkts,
3498 struct hns3_tx_queue *txq = tx_queue;
3501 hns3_tx_free_buffer_simple(txq);
3503 nb_pkts = RTE_MIN(txq->tx_bd_ready, nb_pkts);
3504 if (unlikely(nb_pkts == 0)) {
3505 if (txq->tx_bd_ready == 0)
3506 txq->queue_full_cnt++;
3510 txq->tx_bd_ready -= nb_pkts;
3511 if (txq->next_to_use + nb_pkts > txq->nb_tx_desc) {
3512 nb_tx = txq->nb_tx_desc - txq->next_to_use;
3513 hns3_tx_fill_hw_ring(txq, tx_pkts, nb_tx);
3514 txq->next_to_use = 0;
3517 hns3_tx_fill_hw_ring(txq, tx_pkts + nb_tx, nb_pkts - nb_tx);
3518 txq->next_to_use += nb_pkts - nb_tx;
3520 hns3_write_reg_opt(txq->io_tail_reg, nb_pkts);
3526 hns3_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
3528 struct hns3_tx_queue *txq = tx_queue;
3529 struct hns3_entry *tx_bak_pkt;
3530 struct hns3_desc *tx_ring;
3531 struct rte_mbuf *tx_pkt;
3532 struct rte_mbuf *m_seg;
3533 struct hns3_desc *desc;
3534 uint32_t nb_hold = 0;
3535 uint16_t tx_next_use;
3536 uint16_t tx_pkt_num;
3542 /* free useless buffer */
3543 hns3_tx_free_useless_buffer(txq);
3545 tx_next_use = txq->next_to_use;
3546 tx_bd_max = txq->nb_tx_desc;
3547 tx_pkt_num = nb_pkts;
3548 tx_ring = txq->tx_ring;
3551 tx_bak_pkt = &txq->sw_ring[tx_next_use];
3552 for (nb_tx = 0; nb_tx < tx_pkt_num; nb_tx++) {
3553 tx_pkt = *tx_pkts++;
3555 nb_buf = tx_pkt->nb_segs;
3557 if (nb_buf > txq->tx_bd_ready) {
3558 txq->queue_full_cnt++;
3566 * If packet length is less than minimum packet length supported
3567 * by hardware in Tx direction, driver need to pad it to avoid
3570 if (unlikely(rte_pktmbuf_pkt_len(tx_pkt) <
3571 txq->min_tx_pkt_len)) {
3575 add_len = txq->min_tx_pkt_len -
3576 rte_pktmbuf_pkt_len(tx_pkt);
3577 appended = rte_pktmbuf_append(tx_pkt, add_len);
3578 if (appended == NULL) {
3579 txq->pkt_padding_fail_cnt++;
3583 memset(appended, 0, add_len);
3588 if (hns3_check_non_tso_pkt(nb_buf, &m_seg, tx_pkt, txq))
3591 if (hns3_parse_cksum(txq, tx_next_use, m_seg))
3595 desc = &tx_ring[tx_next_use];
3598 * If the packet is divided into multiple Tx Buffer Descriptors,
3599 * only need to fill vlan, paylen and tso into the first Tx
3600 * Buffer Descriptor.
3602 hns3_fill_first_desc(txq, desc, m_seg);
3605 desc = &tx_ring[tx_next_use];
3607 * Fill valid bits, DMA address and data length for each
3608 * Tx Buffer Descriptor.
3610 hns3_fill_per_desc(desc, m_seg);
3611 tx_bak_pkt->mbuf = m_seg;
3612 m_seg = m_seg->next;
3615 if (tx_next_use >= tx_bd_max) {
3617 tx_bak_pkt = txq->sw_ring;
3621 } while (m_seg != NULL);
3623 /* Add end flag for the last Tx Buffer Descriptor */
3624 desc->tx.tp_fe_sc_vld_ra_ri |=
3625 rte_cpu_to_le_16(BIT(HNS3_TXD_FE_B));
3628 txq->next_to_use = tx_next_use;
3629 txq->tx_bd_ready -= i;
3635 hns3_write_reg_opt(txq->io_tail_reg, nb_hold);
3641 hns3_tx_check_vec_support(__rte_unused struct rte_eth_dev *dev)
3647 hns3_xmit_pkts_vec(__rte_unused void *tx_queue,
3648 __rte_unused struct rte_mbuf **tx_pkts,
3649 __rte_unused uint16_t nb_pkts)
3655 hns3_xmit_pkts_vec_sve(void __rte_unused * tx_queue,
3656 struct rte_mbuf __rte_unused **tx_pkts,
3657 uint16_t __rte_unused nb_pkts)
3663 hns3_tx_burst_mode_get(struct rte_eth_dev *dev, __rte_unused uint16_t queue_id,
3664 struct rte_eth_burst_mode *mode)
3666 eth_tx_burst_t pkt_burst = dev->tx_pkt_burst;
3667 const char *info = NULL;
3669 if (pkt_burst == hns3_xmit_pkts_simple)
3670 info = "Scalar Simple";
3671 else if (pkt_burst == hns3_xmit_pkts)
3673 else if (pkt_burst == hns3_xmit_pkts_vec)
3674 info = "Vector Neon";
3675 else if (pkt_burst == hns3_xmit_pkts_vec_sve)
3676 info = "Vector Sve";
3681 snprintf(mode->info, sizeof(mode->info), "%s", info);
3686 static eth_tx_burst_t
3687 hns3_get_tx_function(struct rte_eth_dev *dev, eth_tx_prep_t *prep)
3689 uint64_t offloads = dev->data->dev_conf.txmode.offloads;
3690 struct hns3_adapter *hns = dev->data->dev_private;
3692 if (hns->tx_vec_allowed && hns3_tx_check_vec_support(dev) == 0) {
3694 return hns3_check_sve_support() ? hns3_xmit_pkts_vec_sve :
3698 if (hns->tx_simple_allowed &&
3699 offloads == (offloads & DEV_TX_OFFLOAD_MBUF_FAST_FREE)) {
3701 return hns3_xmit_pkts_simple;
3704 *prep = hns3_prep_pkts;
3705 return hns3_xmit_pkts;
3709 hns3_dummy_rxtx_burst(void *dpdk_txq __rte_unused,
3710 struct rte_mbuf **pkts __rte_unused,
3711 uint16_t pkts_n __rte_unused)
3716 void hns3_set_rxtx_function(struct rte_eth_dev *eth_dev)
3718 struct hns3_adapter *hns = eth_dev->data->dev_private;
3719 eth_tx_prep_t prep = NULL;
3721 if (hns->hw.adapter_state == HNS3_NIC_STARTED &&
3722 rte_atomic16_read(&hns->hw.reset.resetting) == 0) {
3723 eth_dev->rx_pkt_burst = hns3_get_rx_function(eth_dev);
3724 eth_dev->tx_pkt_burst = hns3_get_tx_function(eth_dev, &prep);
3725 eth_dev->tx_pkt_prepare = prep;
3727 eth_dev->rx_pkt_burst = hns3_dummy_rxtx_burst;
3728 eth_dev->tx_pkt_burst = hns3_dummy_rxtx_burst;
3729 eth_dev->tx_pkt_prepare = hns3_dummy_rxtx_burst;
3734 hns3_rxq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
3735 struct rte_eth_rxq_info *qinfo)
3737 struct hns3_rx_queue *rxq = dev->data->rx_queues[queue_id];
3739 qinfo->mp = rxq->mb_pool;
3740 qinfo->nb_desc = rxq->nb_rx_desc;
3741 qinfo->scattered_rx = dev->data->scattered_rx;
3742 /* Report the HW Rx buffer length to user */
3743 qinfo->rx_buf_size = rxq->rx_buf_len;
3746 * If there are no available Rx buffer descriptors, incoming packets
3747 * are always dropped by hardware based on hns3 network engine.
3749 qinfo->conf.rx_drop_en = 1;
3750 qinfo->conf.offloads = dev->data->dev_conf.rxmode.offloads;
3751 qinfo->conf.rx_free_thresh = rxq->rx_free_thresh;
3752 qinfo->conf.rx_deferred_start = rxq->rx_deferred_start;
3756 hns3_txq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
3757 struct rte_eth_txq_info *qinfo)
3759 struct hns3_tx_queue *txq = dev->data->tx_queues[queue_id];
3761 qinfo->nb_desc = txq->nb_tx_desc;
3762 qinfo->conf.offloads = dev->data->dev_conf.txmode.offloads;
3763 qinfo->conf.tx_rs_thresh = txq->tx_rs_thresh;
3764 qinfo->conf.tx_free_thresh = txq->tx_free_thresh;
3765 qinfo->conf.tx_deferred_start = txq->tx_deferred_start;
3769 hns3_dev_rx_queue_start(struct rte_eth_dev *dev, uint16_t rx_queue_id)
3771 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3772 struct hns3_rx_queue *rxq = dev->data->rx_queues[rx_queue_id];
3773 struct hns3_adapter *hns = HNS3_DEV_HW_TO_ADAPTER(hw);
3776 if (!hns3_dev_indep_txrx_supported(hw))
3779 ret = hns3_reset_queue(hw, rx_queue_id, HNS3_RING_TYPE_RX);
3781 hns3_err(hw, "fail to reset Rx queue %u, ret = %d.",
3786 ret = hns3_init_rxq(hns, rx_queue_id);
3788 hns3_err(hw, "fail to init Rx queue %u, ret = %d.",
3793 hns3_enable_rxq(rxq, true);
3794 dev->data->rx_queue_state[rx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
3800 hns3_reset_sw_rxq(struct hns3_rx_queue *rxq)
3802 rxq->next_to_use = 0;
3803 rxq->rx_rearm_start = 0;
3804 rxq->rx_free_hold = 0;
3805 rxq->rx_rearm_nb = 0;
3806 rxq->pkt_first_seg = NULL;
3807 rxq->pkt_last_seg = NULL;
3808 memset(&rxq->rx_ring[0], 0, rxq->nb_rx_desc * sizeof(struct hns3_desc));
3809 hns3_rxq_vec_setup(rxq);
3813 hns3_dev_rx_queue_stop(struct rte_eth_dev *dev, uint16_t rx_queue_id)
3815 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3816 struct hns3_rx_queue *rxq = dev->data->rx_queues[rx_queue_id];
3818 if (!hns3_dev_indep_txrx_supported(hw))
3821 hns3_enable_rxq(rxq, false);
3823 hns3_rx_queue_release_mbufs(rxq);
3825 hns3_reset_sw_rxq(rxq);
3826 dev->data->rx_queue_state[rx_queue_id] = RTE_ETH_QUEUE_STATE_STOPPED;
3832 hns3_dev_tx_queue_start(struct rte_eth_dev *dev, uint16_t tx_queue_id)
3834 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3835 struct hns3_tx_queue *txq = dev->data->tx_queues[tx_queue_id];
3838 if (!hns3_dev_indep_txrx_supported(hw))
3841 ret = hns3_reset_queue(hw, tx_queue_id, HNS3_RING_TYPE_TX);
3843 hns3_err(hw, "fail to reset Tx queue %u, ret = %d.",
3849 hns3_enable_txq(txq, true);
3850 dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
3856 hns3_dev_tx_queue_stop(struct rte_eth_dev *dev, uint16_t tx_queue_id)
3858 struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3859 struct hns3_tx_queue *txq = dev->data->tx_queues[tx_queue_id];
3861 if (!hns3_dev_indep_txrx_supported(hw))
3864 hns3_enable_txq(txq, false);
3865 hns3_tx_queue_release_mbufs(txq);
3867 * All the mbufs in sw_ring are released and all the pointers in sw_ring
3868 * are set to NULL. If this queue is still called by upper layer,
3869 * residual SW status of this txq may cause these pointers in sw_ring
3870 * which have been set to NULL to be released again. To avoid it,
3874 dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STOPPED;
3880 hns3_rx_queue_count(struct rte_eth_dev *dev, uint16_t rx_queue_id)
3883 * Number of BDs that have been processed by the driver
3884 * but have not been notified to the hardware.
3886 uint32_t driver_hold_bd_num;
3887 struct hns3_rx_queue *rxq;
3890 rxq = dev->data->rx_queues[rx_queue_id];
3891 fbd_num = hns3_read_dev(rxq, HNS3_RING_RX_FBDNUM_REG);
3892 if (dev->rx_pkt_burst == hns3_recv_pkts_vec ||
3893 dev->rx_pkt_burst == hns3_recv_pkts_vec_sve)
3894 driver_hold_bd_num = rxq->rx_rearm_nb;
3896 driver_hold_bd_num = rxq->rx_free_hold;
3898 if (fbd_num <= driver_hold_bd_num)
3901 return fbd_num - driver_hold_bd_num;