1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright (c) 2016 - 2018 Cavium Inc.
10 static inline int qede_alloc_rx_buffer(struct qede_rx_queue *rxq)
12 struct rte_mbuf *new_mb = NULL;
13 struct eth_rx_bd *rx_bd;
15 uint16_t idx = rxq->sw_rx_prod & NUM_RX_BDS(rxq);
17 new_mb = rte_mbuf_raw_alloc(rxq->mb_pool);
18 if (unlikely(!new_mb)) {
20 "Failed to allocate rx buffer "
21 "sw_rx_prod %u sw_rx_cons %u mp entries %u free %u",
22 idx, rxq->sw_rx_cons & NUM_RX_BDS(rxq),
23 rte_mempool_avail_count(rxq->mb_pool),
24 rte_mempool_in_use_count(rxq->mb_pool));
27 rxq->sw_rx_ring[idx].mbuf = new_mb;
28 rxq->sw_rx_ring[idx].page_offset = 0;
29 mapping = rte_mbuf_data_iova_default(new_mb);
30 /* Advance PROD and get BD pointer */
31 rx_bd = (struct eth_rx_bd *)ecore_chain_produce(&rxq->rx_bd_ring);
32 rx_bd->addr.hi = rte_cpu_to_le_32(U64_HI(mapping));
33 rx_bd->addr.lo = rte_cpu_to_le_32(U64_LO(mapping));
38 #define QEDE_MAX_BULK_ALLOC_COUNT 512
40 static inline int qede_alloc_rx_bulk_mbufs(struct qede_rx_queue *rxq, int count)
42 void *obj_p[QEDE_MAX_BULK_ALLOC_COUNT] __rte_cache_aligned;
43 struct rte_mbuf *mbuf = NULL;
44 struct eth_rx_bd *rx_bd;
49 if (count > QEDE_MAX_BULK_ALLOC_COUNT)
50 count = QEDE_MAX_BULK_ALLOC_COUNT;
52 ret = rte_mempool_get_bulk(rxq->mb_pool, obj_p, count);
55 "Failed to allocate %d rx buffers "
56 "sw_rx_prod %u sw_rx_cons %u mp entries %u free %u",
58 rxq->sw_rx_prod & NUM_RX_BDS(rxq),
59 rxq->sw_rx_cons & NUM_RX_BDS(rxq),
60 rte_mempool_avail_count(rxq->mb_pool),
61 rte_mempool_in_use_count(rxq->mb_pool));
65 for (i = 0; i < count; i++) {
67 if (likely(i < count - 1))
68 rte_prefetch0(obj_p[i + 1]);
70 idx = rxq->sw_rx_prod & NUM_RX_BDS(rxq);
71 rxq->sw_rx_ring[idx].mbuf = mbuf;
72 rxq->sw_rx_ring[idx].page_offset = 0;
73 mapping = rte_mbuf_data_iova_default(mbuf);
74 rx_bd = (struct eth_rx_bd *)
75 ecore_chain_produce(&rxq->rx_bd_ring);
76 rx_bd->addr.hi = rte_cpu_to_le_32(U64_HI(mapping));
77 rx_bd->addr.lo = rte_cpu_to_le_32(U64_LO(mapping));
84 /* Criterias for calculating Rx buffer size -
85 * 1) rx_buf_size should not exceed the size of mbuf
86 * 2) In scattered_rx mode - minimum rx_buf_size should be
87 * (MTU + Maximum L2 Header Size + 2) / ETH_RX_MAX_BUFF_PER_PKT
88 * 3) In regular mode - minimum rx_buf_size should be
89 * (MTU + Maximum L2 Header Size + 2)
90 * In above cases +2 corrosponds to 2 bytes padding in front of L2
92 * 4) rx_buf_size should be cacheline-size aligned. So considering
93 * criteria 1, we need to adjust the size to floor instead of ceil,
94 * so that we don't exceed mbuf size while ceiling rx_buf_size.
97 qede_calc_rx_buf_size(struct rte_eth_dev *dev, uint16_t mbufsz,
98 uint16_t max_frame_size)
100 struct qede_dev *qdev = QEDE_INIT_QDEV(dev);
101 struct ecore_dev *edev = QEDE_INIT_EDEV(qdev);
104 if (dev->data->scattered_rx) {
105 /* per HW limitation, only ETH_RX_MAX_BUFF_PER_PKT number of
106 * bufferes can be used for single packet. So need to make sure
107 * mbuf size is sufficient enough for this.
109 if ((mbufsz * ETH_RX_MAX_BUFF_PER_PKT) <
110 (max_frame_size + QEDE_ETH_OVERHEAD)) {
111 DP_ERR(edev, "mbuf %d size is not enough to hold max fragments (%d) for max rx packet length (%d)\n",
112 mbufsz, ETH_RX_MAX_BUFF_PER_PKT, max_frame_size);
116 rx_buf_size = RTE_MAX(mbufsz,
117 (max_frame_size + QEDE_ETH_OVERHEAD) /
118 ETH_RX_MAX_BUFF_PER_PKT);
120 rx_buf_size = max_frame_size + QEDE_ETH_OVERHEAD;
123 /* Align to cache-line size if needed */
124 return QEDE_FLOOR_TO_CACHE_LINE_SIZE(rx_buf_size);
127 static struct qede_rx_queue *
128 qede_alloc_rx_queue_mem(struct rte_eth_dev *dev,
131 unsigned int socket_id,
132 struct rte_mempool *mp,
135 struct qede_dev *qdev = QEDE_INIT_QDEV(dev);
136 struct ecore_dev *edev = QEDE_INIT_EDEV(qdev);
137 struct qede_rx_queue *rxq;
141 /* First allocate the rx queue data structure */
142 rxq = rte_zmalloc_socket("qede_rx_queue", sizeof(struct qede_rx_queue),
143 RTE_CACHE_LINE_SIZE, socket_id);
146 DP_ERR(edev, "Unable to allocate memory for rxq on socket %u",
153 rxq->nb_rx_desc = nb_desc;
154 rxq->queue_id = queue_idx;
155 rxq->port_id = dev->data->port_id;
158 rxq->rx_buf_size = bufsz;
160 DP_INFO(edev, "mtu %u mbufsz %u bd_max_bytes %u scatter_mode %d\n",
161 qdev->mtu, bufsz, rxq->rx_buf_size, dev->data->scattered_rx);
163 /* Allocate the parallel driver ring for Rx buffers */
164 size = sizeof(*rxq->sw_rx_ring) * rxq->nb_rx_desc;
165 rxq->sw_rx_ring = rte_zmalloc_socket("sw_rx_ring", size,
166 RTE_CACHE_LINE_SIZE, socket_id);
167 if (!rxq->sw_rx_ring) {
168 DP_ERR(edev, "Memory allocation fails for sw_rx_ring on"
169 " socket %u\n", socket_id);
174 /* Allocate FW Rx ring */
175 rc = qdev->ops->common->chain_alloc(edev,
176 ECORE_CHAIN_USE_TO_CONSUME_PRODUCE,
177 ECORE_CHAIN_MODE_NEXT_PTR,
178 ECORE_CHAIN_CNT_TYPE_U16,
180 sizeof(struct eth_rx_bd),
184 if (rc != ECORE_SUCCESS) {
185 DP_ERR(edev, "Memory allocation fails for RX BD ring"
186 " on socket %u\n", socket_id);
187 rte_free(rxq->sw_rx_ring);
192 /* Allocate FW completion ring */
193 rc = qdev->ops->common->chain_alloc(edev,
194 ECORE_CHAIN_USE_TO_CONSUME,
195 ECORE_CHAIN_MODE_PBL,
196 ECORE_CHAIN_CNT_TYPE_U16,
198 sizeof(union eth_rx_cqe),
202 if (rc != ECORE_SUCCESS) {
203 DP_ERR(edev, "Memory allocation fails for RX CQE ring"
204 " on socket %u\n", socket_id);
205 qdev->ops->common->chain_free(edev, &rxq->rx_bd_ring);
206 rte_free(rxq->sw_rx_ring);
215 qede_rx_queue_setup(struct rte_eth_dev *dev, uint16_t qid,
216 uint16_t nb_desc, unsigned int socket_id,
217 __rte_unused const struct rte_eth_rxconf *rx_conf,
218 struct rte_mempool *mp)
220 struct qede_dev *qdev = QEDE_INIT_QDEV(dev);
221 struct ecore_dev *edev = QEDE_INIT_EDEV(qdev);
222 struct rte_eth_rxmode *rxmode = &dev->data->dev_conf.rxmode;
223 struct qede_rx_queue *rxq;
224 uint16_t max_rx_pkt_len;
228 PMD_INIT_FUNC_TRACE(edev);
230 /* Note: Ring size/align is controlled by struct rte_eth_desc_lim */
231 if (!rte_is_power_of_2(nb_desc)) {
232 DP_ERR(edev, "Ring size %u is not power of 2\n",
237 /* Free memory prior to re-allocation if needed... */
238 if (dev->data->rx_queues[qid] != NULL) {
239 qede_rx_queue_release(dev->data->rx_queues[qid]);
240 dev->data->rx_queues[qid] = NULL;
243 max_rx_pkt_len = (uint16_t)rxmode->max_rx_pkt_len;
245 /* Fix up RX buffer size */
246 bufsz = (uint16_t)rte_pktmbuf_data_room_size(mp) - RTE_PKTMBUF_HEADROOM;
247 /* cache align the mbuf size to simplfy rx_buf_size calculation */
248 bufsz = QEDE_FLOOR_TO_CACHE_LINE_SIZE(bufsz);
249 if ((rxmode->offloads & DEV_RX_OFFLOAD_SCATTER) ||
250 (max_rx_pkt_len + QEDE_ETH_OVERHEAD) > bufsz) {
251 if (!dev->data->scattered_rx) {
252 DP_INFO(edev, "Forcing scatter-gather mode\n");
253 dev->data->scattered_rx = 1;
257 rc = qede_calc_rx_buf_size(dev, bufsz, max_rx_pkt_len);
263 if (ECORE_IS_CMT(edev)) {
264 rxq = qede_alloc_rx_queue_mem(dev, qid * 2, nb_desc,
265 socket_id, mp, bufsz);
269 qdev->fp_array[qid * 2].rxq = rxq;
270 rxq = qede_alloc_rx_queue_mem(dev, qid * 2 + 1, nb_desc,
271 socket_id, mp, bufsz);
275 qdev->fp_array[qid * 2 + 1].rxq = rxq;
276 /* provide per engine fp struct as rx queue */
277 dev->data->rx_queues[qid] = &qdev->fp_array_cmt[qid];
279 rxq = qede_alloc_rx_queue_mem(dev, qid, nb_desc,
280 socket_id, mp, bufsz);
284 dev->data->rx_queues[qid] = rxq;
285 qdev->fp_array[qid].rxq = rxq;
288 DP_INFO(edev, "rxq %d num_desc %u rx_buf_size=%u socket %u\n",
289 qid, nb_desc, rxq->rx_buf_size, socket_id);
295 qede_rx_queue_reset(__rte_unused struct qede_dev *qdev,
296 struct qede_rx_queue *rxq)
298 DP_INFO(&qdev->edev, "Reset RX queue %u\n", rxq->queue_id);
299 ecore_chain_reset(&rxq->rx_bd_ring);
300 ecore_chain_reset(&rxq->rx_comp_ring);
303 *rxq->hw_cons_ptr = 0;
306 static void qede_rx_queue_release_mbufs(struct qede_rx_queue *rxq)
310 if (rxq->sw_rx_ring) {
311 for (i = 0; i < rxq->nb_rx_desc; i++) {
312 if (rxq->sw_rx_ring[i].mbuf) {
313 rte_pktmbuf_free(rxq->sw_rx_ring[i].mbuf);
314 rxq->sw_rx_ring[i].mbuf = NULL;
320 static void _qede_rx_queue_release(struct qede_dev *qdev,
321 struct ecore_dev *edev,
322 struct qede_rx_queue *rxq)
324 qede_rx_queue_release_mbufs(rxq);
325 qdev->ops->common->chain_free(edev, &rxq->rx_bd_ring);
326 qdev->ops->common->chain_free(edev, &rxq->rx_comp_ring);
327 rte_free(rxq->sw_rx_ring);
331 void qede_rx_queue_release(void *rx_queue)
333 struct qede_rx_queue *rxq = rx_queue;
334 struct qede_fastpath_cmt *fp_cmt;
335 struct qede_dev *qdev;
336 struct ecore_dev *edev;
340 edev = QEDE_INIT_EDEV(qdev);
341 PMD_INIT_FUNC_TRACE(edev);
342 if (ECORE_IS_CMT(edev)) {
344 _qede_rx_queue_release(qdev, edev, fp_cmt->fp0->rxq);
345 _qede_rx_queue_release(qdev, edev, fp_cmt->fp1->rxq);
347 _qede_rx_queue_release(qdev, edev, rxq);
352 /* Stops a given RX queue in the HW */
353 static int qede_rx_queue_stop(struct rte_eth_dev *eth_dev, uint16_t rx_queue_id)
355 struct qede_dev *qdev = QEDE_INIT_QDEV(eth_dev);
356 struct ecore_dev *edev = QEDE_INIT_EDEV(qdev);
357 struct ecore_hwfn *p_hwfn;
358 struct qede_rx_queue *rxq;
362 if (rx_queue_id < qdev->num_rx_queues) {
363 rxq = qdev->fp_array[rx_queue_id].rxq;
364 hwfn_index = rx_queue_id % edev->num_hwfns;
365 p_hwfn = &edev->hwfns[hwfn_index];
366 rc = ecore_eth_rx_queue_stop(p_hwfn, rxq->handle,
368 if (rc != ECORE_SUCCESS) {
369 DP_ERR(edev, "RX queue %u stop fails\n", rx_queue_id);
372 qede_rx_queue_release_mbufs(rxq);
373 qede_rx_queue_reset(qdev, rxq);
374 eth_dev->data->rx_queue_state[rx_queue_id] =
375 RTE_ETH_QUEUE_STATE_STOPPED;
376 DP_INFO(edev, "RX queue %u stopped\n", rx_queue_id);
378 DP_ERR(edev, "RX queue %u is not in range\n", rx_queue_id);
385 static struct qede_tx_queue *
386 qede_alloc_tx_queue_mem(struct rte_eth_dev *dev,
389 unsigned int socket_id,
390 const struct rte_eth_txconf *tx_conf)
392 struct qede_dev *qdev = dev->data->dev_private;
393 struct ecore_dev *edev = &qdev->edev;
394 struct qede_tx_queue *txq;
396 size_t sw_tx_ring_size;
398 txq = rte_zmalloc_socket("qede_tx_queue", sizeof(struct qede_tx_queue),
399 RTE_CACHE_LINE_SIZE, socket_id);
403 "Unable to allocate memory for txq on socket %u",
408 txq->nb_tx_desc = nb_desc;
410 txq->port_id = dev->data->port_id;
412 rc = qdev->ops->common->chain_alloc(edev,
413 ECORE_CHAIN_USE_TO_CONSUME_PRODUCE,
414 ECORE_CHAIN_MODE_PBL,
415 ECORE_CHAIN_CNT_TYPE_U16,
417 sizeof(union eth_tx_bd_types),
420 if (rc != ECORE_SUCCESS) {
422 "Unable to allocate memory for txbd ring on socket %u",
424 qede_tx_queue_release(txq);
428 /* Allocate software ring */
429 sw_tx_ring_size = sizeof(txq->sw_tx_ring) * txq->nb_tx_desc;
430 txq->sw_tx_ring = rte_zmalloc_socket("txq->sw_tx_ring",
432 RTE_CACHE_LINE_SIZE, socket_id);
434 if (!txq->sw_tx_ring) {
436 "Unable to allocate memory for txbd ring on socket %u",
438 qdev->ops->common->chain_free(edev, &txq->tx_pbl);
439 qede_tx_queue_release(txq);
443 txq->queue_id = queue_idx;
445 txq->nb_tx_avail = txq->nb_tx_desc;
447 txq->tx_free_thresh =
448 tx_conf->tx_free_thresh ? tx_conf->tx_free_thresh :
449 (txq->nb_tx_desc - QEDE_DEFAULT_TX_FREE_THRESH);
452 "txq %u num_desc %u tx_free_thresh %u socket %u\n",
453 queue_idx, nb_desc, txq->tx_free_thresh, socket_id);
458 qede_tx_queue_setup(struct rte_eth_dev *dev,
461 unsigned int socket_id,
462 const struct rte_eth_txconf *tx_conf)
464 struct qede_dev *qdev = dev->data->dev_private;
465 struct ecore_dev *edev = &qdev->edev;
466 struct qede_tx_queue *txq;
468 PMD_INIT_FUNC_TRACE(edev);
470 if (!rte_is_power_of_2(nb_desc)) {
471 DP_ERR(edev, "Ring size %u is not power of 2\n",
476 /* Free memory prior to re-allocation if needed... */
477 if (dev->data->tx_queues[queue_idx] != NULL) {
478 qede_tx_queue_release(dev->data->tx_queues[queue_idx]);
479 dev->data->tx_queues[queue_idx] = NULL;
482 if (ECORE_IS_CMT(edev)) {
483 txq = qede_alloc_tx_queue_mem(dev, queue_idx * 2, nb_desc,
488 qdev->fp_array[queue_idx * 2].txq = txq;
489 txq = qede_alloc_tx_queue_mem(dev, (queue_idx * 2) + 1, nb_desc,
494 qdev->fp_array[(queue_idx * 2) + 1].txq = txq;
495 dev->data->tx_queues[queue_idx] =
496 &qdev->fp_array_cmt[queue_idx];
498 txq = qede_alloc_tx_queue_mem(dev, queue_idx, nb_desc,
503 dev->data->tx_queues[queue_idx] = txq;
504 qdev->fp_array[queue_idx].txq = txq;
511 qede_tx_queue_reset(__rte_unused struct qede_dev *qdev,
512 struct qede_tx_queue *txq)
514 DP_INFO(&qdev->edev, "Reset TX queue %u\n", txq->queue_id);
515 ecore_chain_reset(&txq->tx_pbl);
518 *txq->hw_cons_ptr = 0;
521 static void qede_tx_queue_release_mbufs(struct qede_tx_queue *txq)
525 if (txq->sw_tx_ring) {
526 for (i = 0; i < txq->nb_tx_desc; i++) {
527 if (txq->sw_tx_ring[i]) {
528 rte_pktmbuf_free(txq->sw_tx_ring[i]);
529 txq->sw_tx_ring[i] = NULL;
535 static void _qede_tx_queue_release(struct qede_dev *qdev,
536 struct ecore_dev *edev,
537 struct qede_tx_queue *txq)
539 qede_tx_queue_release_mbufs(txq);
540 qdev->ops->common->chain_free(edev, &txq->tx_pbl);
541 rte_free(txq->sw_tx_ring);
545 void qede_tx_queue_release(void *tx_queue)
547 struct qede_tx_queue *txq = tx_queue;
548 struct qede_fastpath_cmt *fp_cmt;
549 struct qede_dev *qdev;
550 struct ecore_dev *edev;
554 edev = QEDE_INIT_EDEV(qdev);
555 PMD_INIT_FUNC_TRACE(edev);
557 if (ECORE_IS_CMT(edev)) {
559 _qede_tx_queue_release(qdev, edev, fp_cmt->fp0->txq);
560 _qede_tx_queue_release(qdev, edev, fp_cmt->fp1->txq);
562 _qede_tx_queue_release(qdev, edev, txq);
567 /* This function allocates fast-path status block memory */
569 qede_alloc_mem_sb(struct qede_dev *qdev, struct ecore_sb_info *sb_info,
572 struct ecore_dev *edev = QEDE_INIT_EDEV(qdev);
573 struct status_block *sb_virt;
577 sb_virt = OSAL_DMA_ALLOC_COHERENT(edev, &sb_phys,
578 sizeof(struct status_block));
580 DP_ERR(edev, "Status block allocation failed\n");
583 rc = qdev->ops->common->sb_init(edev, sb_info, sb_virt,
586 DP_ERR(edev, "Status block initialization failed\n");
587 OSAL_DMA_FREE_COHERENT(edev, sb_virt, sb_phys,
588 sizeof(struct status_block));
595 int qede_alloc_fp_resc(struct qede_dev *qdev)
597 struct ecore_dev *edev = QEDE_INIT_EDEV(qdev);
598 struct qede_fastpath *fp;
603 PMD_INIT_FUNC_TRACE(edev);
606 ecore_vf_get_num_sbs(ECORE_LEADING_HWFN(edev), &num_sbs);
608 num_sbs = ecore_cxt_get_proto_cid_count
609 (ECORE_LEADING_HWFN(edev), PROTOCOLID_ETH, NULL);
612 DP_ERR(edev, "No status blocks available\n");
616 qdev->fp_array = rte_calloc("fp", QEDE_RXTX_MAX(qdev),
617 sizeof(*qdev->fp_array), RTE_CACHE_LINE_SIZE);
619 if (!qdev->fp_array) {
620 DP_ERR(edev, "fp array allocation failed\n");
624 memset((void *)qdev->fp_array, 0, QEDE_RXTX_MAX(qdev) *
625 sizeof(*qdev->fp_array));
627 if (ECORE_IS_CMT(edev)) {
628 qdev->fp_array_cmt = rte_calloc("fp_cmt",
629 QEDE_RXTX_MAX(qdev) / 2,
630 sizeof(*qdev->fp_array_cmt),
631 RTE_CACHE_LINE_SIZE);
633 if (!qdev->fp_array_cmt) {
634 DP_ERR(edev, "fp array for CMT allocation failed\n");
638 memset((void *)qdev->fp_array_cmt, 0,
639 (QEDE_RXTX_MAX(qdev) / 2) * sizeof(*qdev->fp_array_cmt));
641 /* Establish the mapping of fp_array with fp_array_cmt */
642 for (i = 0; i < QEDE_RXTX_MAX(qdev) / 2; i++) {
643 qdev->fp_array_cmt[i].qdev = qdev;
644 qdev->fp_array_cmt[i].fp0 = &qdev->fp_array[i * 2];
645 qdev->fp_array_cmt[i].fp1 = &qdev->fp_array[i * 2 + 1];
649 for (sb_idx = 0; sb_idx < QEDE_RXTX_MAX(qdev); sb_idx++) {
650 fp = &qdev->fp_array[sb_idx];
651 fp->sb_info = rte_calloc("sb", 1, sizeof(struct ecore_sb_info),
652 RTE_CACHE_LINE_SIZE);
654 DP_ERR(edev, "FP sb_info allocation fails\n");
657 if (qede_alloc_mem_sb(qdev, fp->sb_info, sb_idx)) {
658 DP_ERR(edev, "FP status block allocation fails\n");
661 DP_INFO(edev, "sb_info idx 0x%x initialized\n",
662 fp->sb_info->igu_sb_id);
668 void qede_dealloc_fp_resc(struct rte_eth_dev *eth_dev)
670 struct qede_dev *qdev = QEDE_INIT_QDEV(eth_dev);
671 struct ecore_dev *edev = QEDE_INIT_EDEV(qdev);
672 struct qede_fastpath *fp;
676 PMD_INIT_FUNC_TRACE(edev);
678 for (sb_idx = 0; sb_idx < QEDE_RXTX_MAX(qdev); sb_idx++) {
679 fp = &qdev->fp_array[sb_idx];
681 DP_INFO(edev, "Free sb_info index 0x%x\n",
682 fp->sb_info->igu_sb_id);
683 OSAL_DMA_FREE_COHERENT(edev, fp->sb_info->sb_virt,
684 fp->sb_info->sb_phys,
685 sizeof(struct status_block));
686 rte_free(fp->sb_info);
691 /* Free packet buffers and ring memories */
692 for (i = 0; i < eth_dev->data->nb_rx_queues; i++) {
693 if (eth_dev->data->rx_queues[i]) {
694 qede_rx_queue_release(eth_dev->data->rx_queues[i]);
695 eth_dev->data->rx_queues[i] = NULL;
699 for (i = 0; i < eth_dev->data->nb_tx_queues; i++) {
700 if (eth_dev->data->tx_queues[i]) {
701 qede_tx_queue_release(eth_dev->data->tx_queues[i]);
702 eth_dev->data->tx_queues[i] = NULL;
707 rte_free(qdev->fp_array);
708 qdev->fp_array = NULL;
710 if (qdev->fp_array_cmt)
711 rte_free(qdev->fp_array_cmt);
712 qdev->fp_array_cmt = NULL;
716 qede_update_rx_prod(__rte_unused struct qede_dev *edev,
717 struct qede_rx_queue *rxq)
719 uint16_t bd_prod = ecore_chain_get_prod_idx(&rxq->rx_bd_ring);
720 uint16_t cqe_prod = ecore_chain_get_prod_idx(&rxq->rx_comp_ring);
721 struct eth_rx_prod_data rx_prods = { 0 };
723 /* Update producers */
724 rx_prods.bd_prod = rte_cpu_to_le_16(bd_prod);
725 rx_prods.cqe_prod = rte_cpu_to_le_16(cqe_prod);
727 /* Make sure that the BD and SGE data is updated before updating the
728 * producers since FW might read the BD/SGE right after the producer
733 internal_ram_wr(rxq->hw_rxq_prod_addr, sizeof(rx_prods),
734 (uint32_t *)&rx_prods);
736 /* mmiowb is needed to synchronize doorbell writes from more than one
737 * processor. It guarantees that the write arrives to the device before
738 * the napi lock is released and another qede_poll is called (possibly
739 * on another CPU). Without this barrier, the next doorbell can bypass
740 * this doorbell. This is applicable to IA64/Altix systems.
744 PMD_RX_LOG(DEBUG, rxq, "bd_prod %u cqe_prod %u", bd_prod, cqe_prod);
747 /* Starts a given RX queue in HW */
749 qede_rx_queue_start(struct rte_eth_dev *eth_dev, uint16_t rx_queue_id)
751 struct qede_dev *qdev = QEDE_INIT_QDEV(eth_dev);
752 struct ecore_dev *edev = QEDE_INIT_EDEV(qdev);
753 struct ecore_queue_start_common_params params;
754 struct ecore_rxq_start_ret_params ret_params;
755 struct qede_rx_queue *rxq;
756 struct qede_fastpath *fp;
757 struct ecore_hwfn *p_hwfn;
758 dma_addr_t p_phys_table;
764 if (rx_queue_id < qdev->num_rx_queues) {
765 fp = &qdev->fp_array[rx_queue_id];
767 /* Allocate buffers for the Rx ring */
768 for (j = 0; j < rxq->nb_rx_desc; j++) {
769 rc = qede_alloc_rx_buffer(rxq);
771 DP_ERR(edev, "RX buffer allocation failed"
772 " for rxq = %u\n", rx_queue_id);
776 /* disable interrupts */
777 ecore_sb_ack(fp->sb_info, IGU_INT_DISABLE, 0);
779 memset(¶ms, 0, sizeof(params));
780 params.queue_id = rx_queue_id / edev->num_hwfns;
782 params.stats_id = params.vport_id;
783 params.p_sb = fp->sb_info;
784 DP_INFO(edev, "rxq %u igu_sb_id 0x%x\n",
785 fp->rxq->queue_id, fp->sb_info->igu_sb_id);
786 params.sb_idx = RX_PI;
787 hwfn_index = rx_queue_id % edev->num_hwfns;
788 p_hwfn = &edev->hwfns[hwfn_index];
789 p_phys_table = ecore_chain_get_pbl_phys(&fp->rxq->rx_comp_ring);
790 page_cnt = ecore_chain_get_page_cnt(&fp->rxq->rx_comp_ring);
791 memset(&ret_params, 0, sizeof(ret_params));
792 rc = ecore_eth_rx_queue_start(p_hwfn,
793 p_hwfn->hw_info.opaque_fid,
794 ¶ms, fp->rxq->rx_buf_size,
795 fp->rxq->rx_bd_ring.p_phys_addr,
796 p_phys_table, page_cnt,
799 DP_ERR(edev, "RX queue %u could not be started, rc = %d\n",
803 /* Update with the returned parameters */
804 fp->rxq->hw_rxq_prod_addr = ret_params.p_prod;
805 fp->rxq->handle = ret_params.p_handle;
807 fp->rxq->hw_cons_ptr = &fp->sb_info->sb_pi_array[RX_PI];
808 qede_update_rx_prod(qdev, fp->rxq);
809 eth_dev->data->rx_queue_state[rx_queue_id] =
810 RTE_ETH_QUEUE_STATE_STARTED;
811 DP_INFO(edev, "RX queue %u started\n", rx_queue_id);
813 DP_ERR(edev, "RX queue %u is not in range\n", rx_queue_id);
821 qede_tx_queue_start(struct rte_eth_dev *eth_dev, uint16_t tx_queue_id)
823 struct qede_dev *qdev = QEDE_INIT_QDEV(eth_dev);
824 struct ecore_dev *edev = QEDE_INIT_EDEV(qdev);
825 struct ecore_queue_start_common_params params;
826 struct ecore_txq_start_ret_params ret_params;
827 struct ecore_hwfn *p_hwfn;
828 dma_addr_t p_phys_table;
829 struct qede_tx_queue *txq;
830 struct qede_fastpath *fp;
835 if (tx_queue_id < qdev->num_tx_queues) {
836 fp = &qdev->fp_array[tx_queue_id];
838 memset(¶ms, 0, sizeof(params));
839 params.queue_id = tx_queue_id / edev->num_hwfns;
841 params.stats_id = params.vport_id;
842 params.p_sb = fp->sb_info;
843 DP_INFO(edev, "txq %u igu_sb_id 0x%x\n",
844 fp->txq->queue_id, fp->sb_info->igu_sb_id);
845 params.sb_idx = TX_PI(0); /* tc = 0 */
846 p_phys_table = ecore_chain_get_pbl_phys(&txq->tx_pbl);
847 page_cnt = ecore_chain_get_page_cnt(&txq->tx_pbl);
848 hwfn_index = tx_queue_id % edev->num_hwfns;
849 p_hwfn = &edev->hwfns[hwfn_index];
850 if (qdev->dev_info.is_legacy)
851 fp->txq->is_legacy = true;
852 rc = ecore_eth_tx_queue_start(p_hwfn,
853 p_hwfn->hw_info.opaque_fid,
855 p_phys_table, page_cnt,
857 if (rc != ECORE_SUCCESS) {
858 DP_ERR(edev, "TX queue %u couldn't be started, rc=%d\n",
862 txq->doorbell_addr = ret_params.p_doorbell;
863 txq->handle = ret_params.p_handle;
865 txq->hw_cons_ptr = &fp->sb_info->sb_pi_array[TX_PI(0)];
866 SET_FIELD(txq->tx_db.data.params, ETH_DB_DATA_DEST,
868 SET_FIELD(txq->tx_db.data.params, ETH_DB_DATA_AGG_CMD,
870 SET_FIELD(txq->tx_db.data.params,
871 ETH_DB_DATA_AGG_VAL_SEL,
872 DQ_XCM_ETH_TX_BD_PROD_CMD);
873 txq->tx_db.data.agg_flags = DQ_XCM_ETH_DQ_CF_CMD;
874 eth_dev->data->tx_queue_state[tx_queue_id] =
875 RTE_ETH_QUEUE_STATE_STARTED;
876 DP_INFO(edev, "TX queue %u started\n", tx_queue_id);
878 DP_ERR(edev, "TX queue %u is not in range\n", tx_queue_id);
886 qede_process_tx_compl(__rte_unused struct ecore_dev *edev,
887 struct qede_tx_queue *txq)
893 struct rte_mbuf *mbuf;
897 rte_compiler_barrier();
898 sw_tx_cons = ecore_chain_get_cons_idx(&txq->tx_pbl);
899 hw_bd_cons = rte_le_to_cpu_16(*txq->hw_cons_ptr);
900 #ifdef RTE_LIBRTE_QEDE_DEBUG_TX
901 PMD_TX_LOG(DEBUG, txq, "Tx Completions = %u\n",
902 abs(hw_bd_cons - sw_tx_cons));
905 mask = NUM_TX_BDS(txq);
906 idx = txq->sw_tx_cons & mask;
908 remaining = hw_bd_cons - sw_tx_cons;
909 txq->nb_tx_avail += remaining;
912 mbuf = txq->sw_tx_ring[idx];
914 nb_segs = mbuf->nb_segs;
915 remaining -= nb_segs;
917 PMD_TX_LOG(DEBUG, txq, "nb_segs to free %u\n", nb_segs);
920 ecore_chain_consume(&txq->tx_pbl);
924 rte_pktmbuf_free(mbuf);
925 idx = (idx + 1) & mask;
926 PMD_TX_LOG(DEBUG, txq, "Freed tx packet\n");
928 txq->sw_tx_cons = idx;
931 static int qede_drain_txq(struct qede_dev *qdev,
932 struct qede_tx_queue *txq, bool allow_drain)
934 struct ecore_dev *edev = &qdev->edev;
937 while (txq->sw_tx_cons != txq->sw_tx_prod) {
938 qede_process_tx_compl(edev, txq);
941 DP_ERR(edev, "Tx queue[%u] is stuck,"
942 "requesting MCP to drain\n",
944 rc = qdev->ops->common->drain(edev);
947 return qede_drain_txq(qdev, txq, false);
949 DP_ERR(edev, "Timeout waiting for tx queue[%d]:"
950 "PROD=%d, CONS=%d\n",
951 txq->queue_id, txq->sw_tx_prod,
957 rte_compiler_barrier();
960 /* FW finished processing, wait for HW to transmit all tx packets */
966 /* Stops a given TX queue in the HW */
967 static int qede_tx_queue_stop(struct rte_eth_dev *eth_dev, uint16_t tx_queue_id)
969 struct qede_dev *qdev = QEDE_INIT_QDEV(eth_dev);
970 struct ecore_dev *edev = QEDE_INIT_EDEV(qdev);
971 struct ecore_hwfn *p_hwfn;
972 struct qede_tx_queue *txq;
976 if (tx_queue_id < qdev->num_tx_queues) {
977 txq = qdev->fp_array[tx_queue_id].txq;
979 if (qede_drain_txq(qdev, txq, true))
980 return -1; /* For the lack of retcodes */
982 hwfn_index = tx_queue_id % edev->num_hwfns;
983 p_hwfn = &edev->hwfns[hwfn_index];
984 rc = ecore_eth_tx_queue_stop(p_hwfn, txq->handle);
985 if (rc != ECORE_SUCCESS) {
986 DP_ERR(edev, "TX queue %u stop fails\n", tx_queue_id);
989 qede_tx_queue_release_mbufs(txq);
990 qede_tx_queue_reset(qdev, txq);
991 eth_dev->data->tx_queue_state[tx_queue_id] =
992 RTE_ETH_QUEUE_STATE_STOPPED;
993 DP_INFO(edev, "TX queue %u stopped\n", tx_queue_id);
995 DP_ERR(edev, "TX queue %u is not in range\n", tx_queue_id);
1002 int qede_start_queues(struct rte_eth_dev *eth_dev)
1004 struct qede_dev *qdev = QEDE_INIT_QDEV(eth_dev);
1008 for (id = 0; id < qdev->num_rx_queues; id++) {
1009 rc = qede_rx_queue_start(eth_dev, id);
1010 if (rc != ECORE_SUCCESS)
1014 for (id = 0; id < qdev->num_tx_queues; id++) {
1015 rc = qede_tx_queue_start(eth_dev, id);
1016 if (rc != ECORE_SUCCESS)
1023 void qede_stop_queues(struct rte_eth_dev *eth_dev)
1025 struct qede_dev *qdev = QEDE_INIT_QDEV(eth_dev);
1028 /* Stopping RX/TX queues */
1029 for (id = 0; id < qdev->num_tx_queues; id++)
1030 qede_tx_queue_stop(eth_dev, id);
1032 for (id = 0; id < qdev->num_rx_queues; id++)
1033 qede_rx_queue_stop(eth_dev, id);
1036 static inline bool qede_tunn_exist(uint16_t flag)
1038 return !!((PARSING_AND_ERR_FLAGS_TUNNELEXIST_MASK <<
1039 PARSING_AND_ERR_FLAGS_TUNNELEXIST_SHIFT) & flag);
1042 static inline uint8_t qede_check_tunn_csum_l3(uint16_t flag)
1044 return !!((PARSING_AND_ERR_FLAGS_TUNNELIPHDRERROR_MASK <<
1045 PARSING_AND_ERR_FLAGS_TUNNELIPHDRERROR_SHIFT) & flag);
1049 * qede_check_tunn_csum_l4:
1051 * 1 : If L4 csum is enabled AND if the validation has failed.
1054 static inline uint8_t qede_check_tunn_csum_l4(uint16_t flag)
1056 if ((PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMWASCALCULATED_MASK <<
1057 PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMWASCALCULATED_SHIFT) & flag)
1058 return !!((PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMERROR_MASK <<
1059 PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMERROR_SHIFT) & flag);
1064 static inline uint8_t qede_check_notunn_csum_l4(uint16_t flag)
1066 if ((PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_MASK <<
1067 PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_SHIFT) & flag)
1068 return !!((PARSING_AND_ERR_FLAGS_L4CHKSMERROR_MASK <<
1069 PARSING_AND_ERR_FLAGS_L4CHKSMERROR_SHIFT) & flag);
1074 /* Returns outer L2, L3 and L4 packet_type for tunneled packets */
1075 static inline uint32_t qede_rx_cqe_to_pkt_type_outer(struct rte_mbuf *m)
1077 uint32_t packet_type = RTE_PTYPE_UNKNOWN;
1078 struct rte_ether_hdr *eth_hdr;
1079 struct rte_ipv4_hdr *ipv4_hdr;
1080 struct rte_ipv6_hdr *ipv6_hdr;
1081 struct rte_vlan_hdr *vlan_hdr;
1083 bool vlan_tagged = 0;
1086 eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
1087 len = sizeof(struct rte_ether_hdr);
1088 ethertype = rte_cpu_to_be_16(eth_hdr->ether_type);
1090 /* Note: Valid only if VLAN stripping is disabled */
1091 if (ethertype == RTE_ETHER_TYPE_VLAN) {
1093 vlan_hdr = (struct rte_vlan_hdr *)(eth_hdr + 1);
1094 len += sizeof(struct rte_vlan_hdr);
1095 ethertype = rte_cpu_to_be_16(vlan_hdr->eth_proto);
1098 if (ethertype == RTE_ETHER_TYPE_IPV4) {
1099 packet_type |= RTE_PTYPE_L3_IPV4;
1100 ipv4_hdr = rte_pktmbuf_mtod_offset(m,
1101 struct rte_ipv4_hdr *, len);
1102 if (ipv4_hdr->next_proto_id == IPPROTO_TCP)
1103 packet_type |= RTE_PTYPE_L4_TCP;
1104 else if (ipv4_hdr->next_proto_id == IPPROTO_UDP)
1105 packet_type |= RTE_PTYPE_L4_UDP;
1106 } else if (ethertype == RTE_ETHER_TYPE_IPV6) {
1107 packet_type |= RTE_PTYPE_L3_IPV6;
1108 ipv6_hdr = rte_pktmbuf_mtod_offset(m,
1109 struct rte_ipv6_hdr *, len);
1110 if (ipv6_hdr->proto == IPPROTO_TCP)
1111 packet_type |= RTE_PTYPE_L4_TCP;
1112 else if (ipv6_hdr->proto == IPPROTO_UDP)
1113 packet_type |= RTE_PTYPE_L4_UDP;
1117 packet_type |= RTE_PTYPE_L2_ETHER_VLAN;
1119 packet_type |= RTE_PTYPE_L2_ETHER;
1124 static inline uint32_t qede_rx_cqe_to_pkt_type_inner(uint16_t flags)
1129 static const uint32_t
1130 ptype_lkup_tbl[QEDE_PKT_TYPE_MAX] __rte_cache_aligned = {
1131 [QEDE_PKT_TYPE_IPV4] = RTE_PTYPE_INNER_L3_IPV4 |
1132 RTE_PTYPE_INNER_L2_ETHER,
1133 [QEDE_PKT_TYPE_IPV6] = RTE_PTYPE_INNER_L3_IPV6 |
1134 RTE_PTYPE_INNER_L2_ETHER,
1135 [QEDE_PKT_TYPE_IPV4_TCP] = RTE_PTYPE_INNER_L3_IPV4 |
1136 RTE_PTYPE_INNER_L4_TCP |
1137 RTE_PTYPE_INNER_L2_ETHER,
1138 [QEDE_PKT_TYPE_IPV6_TCP] = RTE_PTYPE_INNER_L3_IPV6 |
1139 RTE_PTYPE_INNER_L4_TCP |
1140 RTE_PTYPE_INNER_L2_ETHER,
1141 [QEDE_PKT_TYPE_IPV4_UDP] = RTE_PTYPE_INNER_L3_IPV4 |
1142 RTE_PTYPE_INNER_L4_UDP |
1143 RTE_PTYPE_INNER_L2_ETHER,
1144 [QEDE_PKT_TYPE_IPV6_UDP] = RTE_PTYPE_INNER_L3_IPV6 |
1145 RTE_PTYPE_INNER_L4_UDP |
1146 RTE_PTYPE_INNER_L2_ETHER,
1147 /* Frags with no VLAN */
1148 [QEDE_PKT_TYPE_IPV4_FRAG] = RTE_PTYPE_INNER_L3_IPV4 |
1149 RTE_PTYPE_INNER_L4_FRAG |
1150 RTE_PTYPE_INNER_L2_ETHER,
1151 [QEDE_PKT_TYPE_IPV6_FRAG] = RTE_PTYPE_INNER_L3_IPV6 |
1152 RTE_PTYPE_INNER_L4_FRAG |
1153 RTE_PTYPE_INNER_L2_ETHER,
1155 [QEDE_PKT_TYPE_IPV4_VLAN] = RTE_PTYPE_INNER_L3_IPV4 |
1156 RTE_PTYPE_INNER_L2_ETHER_VLAN,
1157 [QEDE_PKT_TYPE_IPV6_VLAN] = RTE_PTYPE_INNER_L3_IPV6 |
1158 RTE_PTYPE_INNER_L2_ETHER_VLAN,
1159 [QEDE_PKT_TYPE_IPV4_TCP_VLAN] = RTE_PTYPE_INNER_L3_IPV4 |
1160 RTE_PTYPE_INNER_L4_TCP |
1161 RTE_PTYPE_INNER_L2_ETHER_VLAN,
1162 [QEDE_PKT_TYPE_IPV6_TCP_VLAN] = RTE_PTYPE_INNER_L3_IPV6 |
1163 RTE_PTYPE_INNER_L4_TCP |
1164 RTE_PTYPE_INNER_L2_ETHER_VLAN,
1165 [QEDE_PKT_TYPE_IPV4_UDP_VLAN] = RTE_PTYPE_INNER_L3_IPV4 |
1166 RTE_PTYPE_INNER_L4_UDP |
1167 RTE_PTYPE_INNER_L2_ETHER_VLAN,
1168 [QEDE_PKT_TYPE_IPV6_UDP_VLAN] = RTE_PTYPE_INNER_L3_IPV6 |
1169 RTE_PTYPE_INNER_L4_UDP |
1170 RTE_PTYPE_INNER_L2_ETHER_VLAN,
1171 /* Frags with VLAN */
1172 [QEDE_PKT_TYPE_IPV4_VLAN_FRAG] = RTE_PTYPE_INNER_L3_IPV4 |
1173 RTE_PTYPE_INNER_L4_FRAG |
1174 RTE_PTYPE_INNER_L2_ETHER_VLAN,
1175 [QEDE_PKT_TYPE_IPV6_VLAN_FRAG] = RTE_PTYPE_INNER_L3_IPV6 |
1176 RTE_PTYPE_INNER_L4_FRAG |
1177 RTE_PTYPE_INNER_L2_ETHER_VLAN,
1180 /* Bits (0..3) provides L3/L4 protocol type */
1181 /* Bits (4,5) provides frag and VLAN info */
1182 val = ((PARSING_AND_ERR_FLAGS_L3TYPE_MASK <<
1183 PARSING_AND_ERR_FLAGS_L3TYPE_SHIFT) |
1184 (PARSING_AND_ERR_FLAGS_L4PROTOCOL_MASK <<
1185 PARSING_AND_ERR_FLAGS_L4PROTOCOL_SHIFT) |
1186 (PARSING_AND_ERR_FLAGS_IPV4FRAG_MASK <<
1187 PARSING_AND_ERR_FLAGS_IPV4FRAG_SHIFT) |
1188 (PARSING_AND_ERR_FLAGS_TAG8021QEXIST_MASK <<
1189 PARSING_AND_ERR_FLAGS_TAG8021QEXIST_SHIFT)) & flags;
1191 if (val < QEDE_PKT_TYPE_MAX)
1192 return ptype_lkup_tbl[val];
1194 return RTE_PTYPE_UNKNOWN;
1197 static inline uint32_t qede_rx_cqe_to_pkt_type(uint16_t flags)
1202 static const uint32_t
1203 ptype_lkup_tbl[QEDE_PKT_TYPE_MAX] __rte_cache_aligned = {
1204 [QEDE_PKT_TYPE_IPV4] = RTE_PTYPE_L3_IPV4 | RTE_PTYPE_L2_ETHER,
1205 [QEDE_PKT_TYPE_IPV6] = RTE_PTYPE_L3_IPV6 | RTE_PTYPE_L2_ETHER,
1206 [QEDE_PKT_TYPE_IPV4_TCP] = RTE_PTYPE_L3_IPV4 |
1209 [QEDE_PKT_TYPE_IPV6_TCP] = RTE_PTYPE_L3_IPV6 |
1212 [QEDE_PKT_TYPE_IPV4_UDP] = RTE_PTYPE_L3_IPV4 |
1215 [QEDE_PKT_TYPE_IPV6_UDP] = RTE_PTYPE_L3_IPV6 |
1218 /* Frags with no VLAN */
1219 [QEDE_PKT_TYPE_IPV4_FRAG] = RTE_PTYPE_L3_IPV4 |
1222 [QEDE_PKT_TYPE_IPV6_FRAG] = RTE_PTYPE_L3_IPV6 |
1226 [QEDE_PKT_TYPE_IPV4_VLAN] = RTE_PTYPE_L3_IPV4 |
1227 RTE_PTYPE_L2_ETHER_VLAN,
1228 [QEDE_PKT_TYPE_IPV6_VLAN] = RTE_PTYPE_L3_IPV6 |
1229 RTE_PTYPE_L2_ETHER_VLAN,
1230 [QEDE_PKT_TYPE_IPV4_TCP_VLAN] = RTE_PTYPE_L3_IPV4 |
1232 RTE_PTYPE_L2_ETHER_VLAN,
1233 [QEDE_PKT_TYPE_IPV6_TCP_VLAN] = RTE_PTYPE_L3_IPV6 |
1235 RTE_PTYPE_L2_ETHER_VLAN,
1236 [QEDE_PKT_TYPE_IPV4_UDP_VLAN] = RTE_PTYPE_L3_IPV4 |
1238 RTE_PTYPE_L2_ETHER_VLAN,
1239 [QEDE_PKT_TYPE_IPV6_UDP_VLAN] = RTE_PTYPE_L3_IPV6 |
1241 RTE_PTYPE_L2_ETHER_VLAN,
1242 /* Frags with VLAN */
1243 [QEDE_PKT_TYPE_IPV4_VLAN_FRAG] = RTE_PTYPE_L3_IPV4 |
1245 RTE_PTYPE_L2_ETHER_VLAN,
1246 [QEDE_PKT_TYPE_IPV6_VLAN_FRAG] = RTE_PTYPE_L3_IPV6 |
1248 RTE_PTYPE_L2_ETHER_VLAN,
1251 /* Bits (0..3) provides L3/L4 protocol type */
1252 /* Bits (4,5) provides frag and VLAN info */
1253 val = ((PARSING_AND_ERR_FLAGS_L3TYPE_MASK <<
1254 PARSING_AND_ERR_FLAGS_L3TYPE_SHIFT) |
1255 (PARSING_AND_ERR_FLAGS_L4PROTOCOL_MASK <<
1256 PARSING_AND_ERR_FLAGS_L4PROTOCOL_SHIFT) |
1257 (PARSING_AND_ERR_FLAGS_IPV4FRAG_MASK <<
1258 PARSING_AND_ERR_FLAGS_IPV4FRAG_SHIFT) |
1259 (PARSING_AND_ERR_FLAGS_TAG8021QEXIST_MASK <<
1260 PARSING_AND_ERR_FLAGS_TAG8021QEXIST_SHIFT)) & flags;
1262 if (val < QEDE_PKT_TYPE_MAX)
1263 return ptype_lkup_tbl[val];
1265 return RTE_PTYPE_UNKNOWN;
1268 static inline uint8_t
1269 qede_check_notunn_csum_l3(struct rte_mbuf *m, uint16_t flag)
1271 struct rte_ipv4_hdr *ip;
1276 val = ((PARSING_AND_ERR_FLAGS_IPHDRERROR_MASK <<
1277 PARSING_AND_ERR_FLAGS_IPHDRERROR_SHIFT) & flag);
1279 if (unlikely(val)) {
1280 m->packet_type = qede_rx_cqe_to_pkt_type(flag);
1281 if (RTE_ETH_IS_IPV4_HDR(m->packet_type)) {
1282 ip = rte_pktmbuf_mtod_offset(m, struct rte_ipv4_hdr *,
1283 sizeof(struct rte_ether_hdr));
1284 pkt_csum = ip->hdr_checksum;
1285 ip->hdr_checksum = 0;
1286 calc_csum = rte_ipv4_cksum(ip);
1287 ip->hdr_checksum = pkt_csum;
1288 return (calc_csum != pkt_csum);
1289 } else if (RTE_ETH_IS_IPV6_HDR(m->packet_type)) {
1296 static inline void qede_rx_bd_ring_consume(struct qede_rx_queue *rxq)
1298 ecore_chain_consume(&rxq->rx_bd_ring);
1303 qede_reuse_page(__rte_unused struct qede_dev *qdev,
1304 struct qede_rx_queue *rxq, struct qede_rx_entry *curr_cons)
1306 struct eth_rx_bd *rx_bd_prod = ecore_chain_produce(&rxq->rx_bd_ring);
1307 uint16_t idx = rxq->sw_rx_prod & NUM_RX_BDS(rxq);
1308 struct qede_rx_entry *curr_prod;
1309 dma_addr_t new_mapping;
1311 curr_prod = &rxq->sw_rx_ring[idx];
1312 *curr_prod = *curr_cons;
1314 new_mapping = rte_mbuf_data_iova_default(curr_prod->mbuf) +
1315 curr_prod->page_offset;
1317 rx_bd_prod->addr.hi = rte_cpu_to_le_32(U64_HI(new_mapping));
1318 rx_bd_prod->addr.lo = rte_cpu_to_le_32(U64_LO(new_mapping));
1324 qede_recycle_rx_bd_ring(struct qede_rx_queue *rxq,
1325 struct qede_dev *qdev, uint8_t count)
1327 struct qede_rx_entry *curr_cons;
1329 for (; count > 0; count--) {
1330 curr_cons = &rxq->sw_rx_ring[rxq->sw_rx_cons & NUM_RX_BDS(rxq)];
1331 qede_reuse_page(qdev, rxq, curr_cons);
1332 qede_rx_bd_ring_consume(rxq);
1337 qede_rx_process_tpa_cmn_cont_end_cqe(__rte_unused struct qede_dev *qdev,
1338 struct qede_rx_queue *rxq,
1339 uint8_t agg_index, uint16_t len)
1341 struct qede_agg_info *tpa_info;
1342 struct rte_mbuf *curr_frag; /* Pointer to currently filled TPA seg */
1345 /* Under certain conditions it is possible that FW may not consume
1346 * additional or new BD. So decision to consume the BD must be made
1347 * based on len_list[0].
1349 if (rte_le_to_cpu_16(len)) {
1350 tpa_info = &rxq->tpa_info[agg_index];
1351 cons_idx = rxq->sw_rx_cons & NUM_RX_BDS(rxq);
1352 curr_frag = rxq->sw_rx_ring[cons_idx].mbuf;
1354 curr_frag->nb_segs = 1;
1355 curr_frag->pkt_len = rte_le_to_cpu_16(len);
1356 curr_frag->data_len = curr_frag->pkt_len;
1357 tpa_info->tpa_tail->next = curr_frag;
1358 tpa_info->tpa_tail = curr_frag;
1359 qede_rx_bd_ring_consume(rxq);
1360 if (unlikely(qede_alloc_rx_buffer(rxq) != 0)) {
1361 PMD_RX_LOG(ERR, rxq, "mbuf allocation fails\n");
1362 rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed++;
1363 rxq->rx_alloc_errors++;
1369 qede_rx_process_tpa_cont_cqe(struct qede_dev *qdev,
1370 struct qede_rx_queue *rxq,
1371 struct eth_fast_path_rx_tpa_cont_cqe *cqe)
1373 PMD_RX_LOG(INFO, rxq, "TPA cont[%d] - len [%d]\n",
1374 cqe->tpa_agg_index, rte_le_to_cpu_16(cqe->len_list[0]));
1375 /* only len_list[0] will have value */
1376 qede_rx_process_tpa_cmn_cont_end_cqe(qdev, rxq, cqe->tpa_agg_index,
1381 qede_rx_process_tpa_end_cqe(struct qede_dev *qdev,
1382 struct qede_rx_queue *rxq,
1383 struct eth_fast_path_rx_tpa_end_cqe *cqe)
1385 struct rte_mbuf *rx_mb; /* Pointer to head of the chained agg */
1387 qede_rx_process_tpa_cmn_cont_end_cqe(qdev, rxq, cqe->tpa_agg_index,
1389 /* Update total length and frags based on end TPA */
1390 rx_mb = rxq->tpa_info[cqe->tpa_agg_index].tpa_head;
1391 /* TODO: Add Sanity Checks */
1392 rx_mb->nb_segs = cqe->num_of_bds;
1393 rx_mb->pkt_len = cqe->total_packet_len;
1395 PMD_RX_LOG(INFO, rxq, "TPA End[%d] reason %d cqe_len %d nb_segs %d"
1396 " pkt_len %d\n", cqe->tpa_agg_index, cqe->end_reason,
1397 rte_le_to_cpu_16(cqe->len_list[0]), rx_mb->nb_segs,
1401 static inline uint32_t qede_rx_cqe_to_tunn_pkt_type(uint16_t flags)
1406 static const uint32_t
1407 ptype_tunn_lkup_tbl[QEDE_PKT_TYPE_TUNN_MAX_TYPE] __rte_cache_aligned = {
1408 [QEDE_PKT_TYPE_UNKNOWN] = RTE_PTYPE_UNKNOWN,
1409 [QEDE_PKT_TYPE_TUNN_GENEVE] = RTE_PTYPE_TUNNEL_GENEVE,
1410 [QEDE_PKT_TYPE_TUNN_GRE] = RTE_PTYPE_TUNNEL_GRE,
1411 [QEDE_PKT_TYPE_TUNN_VXLAN] = RTE_PTYPE_TUNNEL_VXLAN,
1412 [QEDE_PKT_TYPE_TUNN_L2_TENID_NOEXIST_GENEVE] =
1413 RTE_PTYPE_TUNNEL_GENEVE,
1414 [QEDE_PKT_TYPE_TUNN_L2_TENID_NOEXIST_GRE] =
1415 RTE_PTYPE_TUNNEL_GRE,
1416 [QEDE_PKT_TYPE_TUNN_L2_TENID_NOEXIST_VXLAN] =
1417 RTE_PTYPE_TUNNEL_VXLAN,
1418 [QEDE_PKT_TYPE_TUNN_L2_TENID_EXIST_GENEVE] =
1419 RTE_PTYPE_TUNNEL_GENEVE,
1420 [QEDE_PKT_TYPE_TUNN_L2_TENID_EXIST_GRE] =
1421 RTE_PTYPE_TUNNEL_GRE,
1422 [QEDE_PKT_TYPE_TUNN_L2_TENID_EXIST_VXLAN] =
1423 RTE_PTYPE_TUNNEL_VXLAN,
1424 [QEDE_PKT_TYPE_TUNN_IPV4_TENID_NOEXIST_GENEVE] =
1425 RTE_PTYPE_TUNNEL_GENEVE | RTE_PTYPE_L3_IPV4,
1426 [QEDE_PKT_TYPE_TUNN_IPV4_TENID_NOEXIST_GRE] =
1427 RTE_PTYPE_TUNNEL_GRE | RTE_PTYPE_L3_IPV4,
1428 [QEDE_PKT_TYPE_TUNN_IPV4_TENID_NOEXIST_VXLAN] =
1429 RTE_PTYPE_TUNNEL_VXLAN | RTE_PTYPE_L3_IPV4,
1430 [QEDE_PKT_TYPE_TUNN_IPV4_TENID_EXIST_GENEVE] =
1431 RTE_PTYPE_TUNNEL_GENEVE | RTE_PTYPE_L3_IPV4,
1432 [QEDE_PKT_TYPE_TUNN_IPV4_TENID_EXIST_GRE] =
1433 RTE_PTYPE_TUNNEL_GRE | RTE_PTYPE_L3_IPV4,
1434 [QEDE_PKT_TYPE_TUNN_IPV4_TENID_EXIST_VXLAN] =
1435 RTE_PTYPE_TUNNEL_VXLAN | RTE_PTYPE_L3_IPV4,
1436 [QEDE_PKT_TYPE_TUNN_IPV6_TENID_NOEXIST_GENEVE] =
1437 RTE_PTYPE_TUNNEL_GENEVE | RTE_PTYPE_L3_IPV6,
1438 [QEDE_PKT_TYPE_TUNN_IPV6_TENID_NOEXIST_GRE] =
1439 RTE_PTYPE_TUNNEL_GRE | RTE_PTYPE_L3_IPV6,
1440 [QEDE_PKT_TYPE_TUNN_IPV6_TENID_NOEXIST_VXLAN] =
1441 RTE_PTYPE_TUNNEL_VXLAN | RTE_PTYPE_L3_IPV6,
1442 [QEDE_PKT_TYPE_TUNN_IPV6_TENID_EXIST_GENEVE] =
1443 RTE_PTYPE_TUNNEL_GENEVE | RTE_PTYPE_L3_IPV6,
1444 [QEDE_PKT_TYPE_TUNN_IPV6_TENID_EXIST_GRE] =
1445 RTE_PTYPE_TUNNEL_GRE | RTE_PTYPE_L3_IPV6,
1446 [QEDE_PKT_TYPE_TUNN_IPV6_TENID_EXIST_VXLAN] =
1447 RTE_PTYPE_TUNNEL_VXLAN | RTE_PTYPE_L3_IPV6,
1450 /* Cover bits[4-0] to include tunn_type and next protocol */
1451 val = ((ETH_TUNNEL_PARSING_FLAGS_TYPE_MASK <<
1452 ETH_TUNNEL_PARSING_FLAGS_TYPE_SHIFT) |
1453 (ETH_TUNNEL_PARSING_FLAGS_NEXT_PROTOCOL_MASK <<
1454 ETH_TUNNEL_PARSING_FLAGS_NEXT_PROTOCOL_SHIFT)) & flags;
1456 if (val < QEDE_PKT_TYPE_TUNN_MAX_TYPE)
1457 return ptype_tunn_lkup_tbl[val];
1459 return RTE_PTYPE_UNKNOWN;
1463 qede_process_sg_pkts(void *p_rxq, struct rte_mbuf *rx_mb,
1464 uint8_t num_segs, uint16_t pkt_len)
1466 struct qede_rx_queue *rxq = p_rxq;
1467 struct qede_dev *qdev = rxq->qdev;
1468 register struct rte_mbuf *seg1 = NULL;
1469 register struct rte_mbuf *seg2 = NULL;
1470 uint16_t sw_rx_index;
1475 cur_size = pkt_len > rxq->rx_buf_size ? rxq->rx_buf_size :
1477 if (unlikely(!cur_size)) {
1478 PMD_RX_LOG(ERR, rxq, "Length is 0 while %u BDs"
1479 " left for mapping jumbo\n", num_segs);
1480 qede_recycle_rx_bd_ring(rxq, qdev, num_segs);
1483 sw_rx_index = rxq->sw_rx_cons & NUM_RX_BDS(rxq);
1484 seg2 = rxq->sw_rx_ring[sw_rx_index].mbuf;
1485 qede_rx_bd_ring_consume(rxq);
1486 pkt_len -= cur_size;
1487 seg2->data_len = cur_size;
1497 #ifdef RTE_LIBRTE_QEDE_DEBUG_RX
1499 print_rx_bd_info(struct rte_mbuf *m, struct qede_rx_queue *rxq,
1502 PMD_RX_LOG(INFO, rxq,
1503 "len 0x%04x bf 0x%04x hash_val 0x%x"
1504 " ol_flags 0x%04lx l2=%s l3=%s l4=%s tunn=%s"
1505 " inner_l2=%s inner_l3=%s inner_l4=%s\n",
1506 m->data_len, bitfield, m->hash.rss,
1507 (unsigned long)m->ol_flags,
1508 rte_get_ptype_l2_name(m->packet_type),
1509 rte_get_ptype_l3_name(m->packet_type),
1510 rte_get_ptype_l4_name(m->packet_type),
1511 rte_get_ptype_tunnel_name(m->packet_type),
1512 rte_get_ptype_inner_l2_name(m->packet_type),
1513 rte_get_ptype_inner_l3_name(m->packet_type),
1514 rte_get_ptype_inner_l4_name(m->packet_type));
1519 qede_recv_pkts_regular(void *p_rxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
1521 struct eth_fast_path_rx_reg_cqe *fp_cqe = NULL;
1522 register struct rte_mbuf *rx_mb = NULL;
1523 struct qede_rx_queue *rxq = p_rxq;
1524 struct qede_dev *qdev = rxq->qdev;
1525 struct ecore_dev *edev = &qdev->edev;
1526 union eth_rx_cqe *cqe;
1528 enum eth_rx_cqe_type cqe_type;
1529 int rss_enable = qdev->rss_enable;
1530 int rx_alloc_count = 0;
1531 uint32_t packet_type;
1533 uint16_t vlan_tci, port_id;
1534 uint16_t hw_comp_cons, sw_comp_cons, sw_rx_index, num_rx_bds;
1535 uint16_t rx_pkt = 0;
1536 uint16_t pkt_len = 0;
1537 uint16_t len; /* Length of first BD */
1538 uint16_t preload_idx;
1539 uint16_t parse_flag;
1540 #ifdef RTE_LIBRTE_QEDE_DEBUG_RX
1541 uint8_t bitfield_val;
1543 uint8_t offset, flags, bd_num;
1546 /* Allocate buffers that we used in previous loop */
1547 if (rxq->rx_alloc_count) {
1548 if (unlikely(qede_alloc_rx_bulk_mbufs(rxq,
1549 rxq->rx_alloc_count))) {
1550 struct rte_eth_dev *dev;
1552 PMD_RX_LOG(ERR, rxq,
1553 "New buffer allocation failed,"
1554 "dropping incoming packetn");
1555 dev = &rte_eth_devices[rxq->port_id];
1556 dev->data->rx_mbuf_alloc_failed +=
1557 rxq->rx_alloc_count;
1558 rxq->rx_alloc_errors += rxq->rx_alloc_count;
1561 qede_update_rx_prod(qdev, rxq);
1562 rxq->rx_alloc_count = 0;
1565 hw_comp_cons = rte_le_to_cpu_16(*rxq->hw_cons_ptr);
1566 sw_comp_cons = ecore_chain_get_cons_idx(&rxq->rx_comp_ring);
1570 if (hw_comp_cons == sw_comp_cons)
1573 num_rx_bds = NUM_RX_BDS(rxq);
1574 port_id = rxq->port_id;
1576 while (sw_comp_cons != hw_comp_cons) {
1578 packet_type = RTE_PTYPE_UNKNOWN;
1582 /* Get the CQE from the completion ring */
1584 (union eth_rx_cqe *)ecore_chain_consume(&rxq->rx_comp_ring);
1585 cqe_type = cqe->fast_path_regular.type;
1586 PMD_RX_LOG(INFO, rxq, "Rx CQE type %d\n", cqe_type);
1588 if (likely(cqe_type == ETH_RX_CQE_TYPE_REGULAR)) {
1589 fp_cqe = &cqe->fast_path_regular;
1591 if (cqe_type == ETH_RX_CQE_TYPE_SLOW_PATH) {
1592 PMD_RX_LOG(INFO, rxq, "Got unexpected slowpath CQE\n");
1593 ecore_eth_cqe_completion
1594 (&edev->hwfns[rxq->queue_id %
1596 (struct eth_slow_path_rx_cqe *)cqe);
1601 /* Get the data from the SW ring */
1602 sw_rx_index = rxq->sw_rx_cons & num_rx_bds;
1603 rx_mb = rxq->sw_rx_ring[sw_rx_index].mbuf;
1604 assert(rx_mb != NULL);
1606 parse_flag = rte_le_to_cpu_16(fp_cqe->pars_flags.flags);
1607 offset = fp_cqe->placement_offset;
1608 len = rte_le_to_cpu_16(fp_cqe->len_on_first_bd);
1609 pkt_len = rte_le_to_cpu_16(fp_cqe->pkt_len);
1610 vlan_tci = rte_le_to_cpu_16(fp_cqe->vlan_tag);
1611 rss_hash = rte_le_to_cpu_32(fp_cqe->rss_hash);
1612 bd_num = fp_cqe->bd_num;
1613 #ifdef RTE_LIBRTE_QEDE_DEBUG_RX
1614 bitfield_val = fp_cqe->bitfields;
1617 if (unlikely(qede_tunn_exist(parse_flag))) {
1618 PMD_RX_LOG(INFO, rxq, "Rx tunneled packet\n");
1619 if (unlikely(qede_check_tunn_csum_l4(parse_flag))) {
1620 PMD_RX_LOG(ERR, rxq,
1621 "L4 csum failed, flags = 0x%x\n",
1623 rxq->rx_hw_errors++;
1624 ol_flags |= PKT_RX_L4_CKSUM_BAD;
1626 ol_flags |= PKT_RX_L4_CKSUM_GOOD;
1629 if (unlikely(qede_check_tunn_csum_l3(parse_flag))) {
1630 PMD_RX_LOG(ERR, rxq,
1631 "Outer L3 csum failed, flags = 0x%x\n",
1633 rxq->rx_hw_errors++;
1634 ol_flags |= PKT_RX_OUTER_IP_CKSUM_BAD;
1636 ol_flags |= PKT_RX_IP_CKSUM_GOOD;
1639 flags = fp_cqe->tunnel_pars_flags.flags;
1643 qede_rx_cqe_to_tunn_pkt_type(flags);
1647 qede_rx_cqe_to_pkt_type_inner(parse_flag);
1649 /* Outer L3/L4 types is not available in CQE */
1650 packet_type |= qede_rx_cqe_to_pkt_type_outer(rx_mb);
1652 /* Outer L3/L4 types is not available in CQE.
1653 * Need to add offset to parse correctly,
1655 rx_mb->data_off = offset + RTE_PKTMBUF_HEADROOM;
1656 packet_type |= qede_rx_cqe_to_pkt_type_outer(rx_mb);
1658 packet_type |= qede_rx_cqe_to_pkt_type(parse_flag);
1661 /* Common handling for non-tunnel packets and for inner
1662 * headers in the case of tunnel.
1664 if (unlikely(qede_check_notunn_csum_l4(parse_flag))) {
1665 PMD_RX_LOG(ERR, rxq,
1666 "L4 csum failed, flags = 0x%x\n",
1668 rxq->rx_hw_errors++;
1669 ol_flags |= PKT_RX_L4_CKSUM_BAD;
1671 ol_flags |= PKT_RX_L4_CKSUM_GOOD;
1673 if (unlikely(qede_check_notunn_csum_l3(rx_mb, parse_flag))) {
1674 PMD_RX_LOG(ERR, rxq, "IP csum failed, flags = 0x%x\n",
1676 rxq->rx_hw_errors++;
1677 ol_flags |= PKT_RX_IP_CKSUM_BAD;
1679 ol_flags |= PKT_RX_IP_CKSUM_GOOD;
1682 if (unlikely(CQE_HAS_VLAN(parse_flag) ||
1683 CQE_HAS_OUTER_VLAN(parse_flag))) {
1684 /* Note: FW doesn't indicate Q-in-Q packet */
1685 ol_flags |= PKT_RX_VLAN;
1686 if (qdev->vlan_strip_flg) {
1687 ol_flags |= PKT_RX_VLAN_STRIPPED;
1688 rx_mb->vlan_tci = vlan_tci;
1693 ol_flags |= PKT_RX_RSS_HASH;
1694 rx_mb->hash.rss = rss_hash;
1698 qede_rx_bd_ring_consume(rxq);
1700 /* Prefetch next mbuf while processing current one. */
1701 preload_idx = rxq->sw_rx_cons & num_rx_bds;
1702 rte_prefetch0(rxq->sw_rx_ring[preload_idx].mbuf);
1704 /* Update rest of the MBUF fields */
1705 rx_mb->data_off = offset + RTE_PKTMBUF_HEADROOM;
1706 rx_mb->port = port_id;
1707 rx_mb->ol_flags = ol_flags;
1708 rx_mb->data_len = len;
1709 rx_mb->packet_type = packet_type;
1710 #ifdef RTE_LIBRTE_QEDE_DEBUG_RX
1711 print_rx_bd_info(rx_mb, rxq, bitfield_val);
1713 rx_mb->nb_segs = bd_num;
1714 rx_mb->pkt_len = pkt_len;
1716 rx_pkts[rx_pkt] = rx_mb;
1720 ecore_chain_recycle_consumed(&rxq->rx_comp_ring);
1721 sw_comp_cons = ecore_chain_get_cons_idx(&rxq->rx_comp_ring);
1722 if (rx_pkt == nb_pkts) {
1723 PMD_RX_LOG(DEBUG, rxq,
1724 "Budget reached nb_pkts=%u received=%u",
1730 /* Request number of bufferes to be allocated in next loop */
1731 rxq->rx_alloc_count = rx_alloc_count;
1733 rxq->rcv_pkts += rx_pkt;
1734 rxq->rx_segs += rx_pkt;
1735 PMD_RX_LOG(DEBUG, rxq, "rx_pkts=%u core=%d", rx_pkt, rte_lcore_id());
1741 qede_recv_pkts(void *p_rxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
1743 struct qede_rx_queue *rxq = p_rxq;
1744 struct qede_dev *qdev = rxq->qdev;
1745 struct ecore_dev *edev = &qdev->edev;
1746 uint16_t hw_comp_cons, sw_comp_cons, sw_rx_index;
1747 uint16_t rx_pkt = 0;
1748 union eth_rx_cqe *cqe;
1749 struct eth_fast_path_rx_reg_cqe *fp_cqe = NULL;
1750 register struct rte_mbuf *rx_mb = NULL;
1751 register struct rte_mbuf *seg1 = NULL;
1752 enum eth_rx_cqe_type cqe_type;
1753 uint16_t pkt_len = 0; /* Sum of all BD segments */
1754 uint16_t len; /* Length of first BD */
1755 uint8_t num_segs = 1;
1756 uint16_t preload_idx;
1757 uint16_t parse_flag;
1758 #ifdef RTE_LIBRTE_QEDE_DEBUG_RX
1759 uint8_t bitfield_val;
1761 uint8_t tunn_parse_flag;
1762 struct eth_fast_path_rx_tpa_start_cqe *cqe_start_tpa;
1764 uint32_t packet_type;
1767 uint8_t offset, tpa_agg_idx, flags;
1768 struct qede_agg_info *tpa_info = NULL;
1770 int rx_alloc_count = 0;
1773 /* Allocate buffers that we used in previous loop */
1774 if (rxq->rx_alloc_count) {
1775 if (unlikely(qede_alloc_rx_bulk_mbufs(rxq,
1776 rxq->rx_alloc_count))) {
1777 struct rte_eth_dev *dev;
1779 PMD_RX_LOG(ERR, rxq,
1780 "New buffer allocation failed,"
1781 "dropping incoming packetn");
1782 dev = &rte_eth_devices[rxq->port_id];
1783 dev->data->rx_mbuf_alloc_failed +=
1784 rxq->rx_alloc_count;
1785 rxq->rx_alloc_errors += rxq->rx_alloc_count;
1788 qede_update_rx_prod(qdev, rxq);
1789 rxq->rx_alloc_count = 0;
1792 hw_comp_cons = rte_le_to_cpu_16(*rxq->hw_cons_ptr);
1793 sw_comp_cons = ecore_chain_get_cons_idx(&rxq->rx_comp_ring);
1797 if (hw_comp_cons == sw_comp_cons)
1800 while (sw_comp_cons != hw_comp_cons) {
1802 packet_type = RTE_PTYPE_UNKNOWN;
1804 tpa_start_flg = false;
1807 /* Get the CQE from the completion ring */
1809 (union eth_rx_cqe *)ecore_chain_consume(&rxq->rx_comp_ring);
1810 cqe_type = cqe->fast_path_regular.type;
1811 PMD_RX_LOG(INFO, rxq, "Rx CQE type %d\n", cqe_type);
1814 case ETH_RX_CQE_TYPE_REGULAR:
1815 fp_cqe = &cqe->fast_path_regular;
1817 case ETH_RX_CQE_TYPE_TPA_START:
1818 cqe_start_tpa = &cqe->fast_path_tpa_start;
1819 tpa_info = &rxq->tpa_info[cqe_start_tpa->tpa_agg_index];
1820 tpa_start_flg = true;
1821 /* Mark it as LRO packet */
1822 ol_flags |= PKT_RX_LRO;
1823 /* In split mode, seg_len is same as len_on_first_bd
1824 * and bw_ext_bd_len_list will be empty since there are
1825 * no additional buffers
1827 PMD_RX_LOG(INFO, rxq,
1828 "TPA start[%d] - len_on_first_bd %d header %d"
1829 " [bd_list[0] %d], [seg_len %d]\n",
1830 cqe_start_tpa->tpa_agg_index,
1831 rte_le_to_cpu_16(cqe_start_tpa->len_on_first_bd),
1832 cqe_start_tpa->header_len,
1833 rte_le_to_cpu_16(cqe_start_tpa->bw_ext_bd_len_list[0]),
1834 rte_le_to_cpu_16(cqe_start_tpa->seg_len));
1837 case ETH_RX_CQE_TYPE_TPA_CONT:
1838 qede_rx_process_tpa_cont_cqe(qdev, rxq,
1839 &cqe->fast_path_tpa_cont);
1841 case ETH_RX_CQE_TYPE_TPA_END:
1842 qede_rx_process_tpa_end_cqe(qdev, rxq,
1843 &cqe->fast_path_tpa_end);
1844 tpa_agg_idx = cqe->fast_path_tpa_end.tpa_agg_index;
1845 tpa_info = &rxq->tpa_info[tpa_agg_idx];
1846 rx_mb = rxq->tpa_info[tpa_agg_idx].tpa_head;
1848 case ETH_RX_CQE_TYPE_SLOW_PATH:
1849 PMD_RX_LOG(INFO, rxq, "Got unexpected slowpath CQE\n");
1850 ecore_eth_cqe_completion(
1851 &edev->hwfns[rxq->queue_id % edev->num_hwfns],
1852 (struct eth_slow_path_rx_cqe *)cqe);
1858 /* Get the data from the SW ring */
1859 sw_rx_index = rxq->sw_rx_cons & NUM_RX_BDS(rxq);
1860 rx_mb = rxq->sw_rx_ring[sw_rx_index].mbuf;
1861 assert(rx_mb != NULL);
1863 /* Handle regular CQE or TPA start CQE */
1864 if (!tpa_start_flg) {
1865 parse_flag = rte_le_to_cpu_16(fp_cqe->pars_flags.flags);
1866 offset = fp_cqe->placement_offset;
1867 len = rte_le_to_cpu_16(fp_cqe->len_on_first_bd);
1868 pkt_len = rte_le_to_cpu_16(fp_cqe->pkt_len);
1869 vlan_tci = rte_le_to_cpu_16(fp_cqe->vlan_tag);
1870 rss_hash = rte_le_to_cpu_32(fp_cqe->rss_hash);
1871 #ifdef RTE_LIBRTE_QEDE_DEBUG_RX
1872 bitfield_val = fp_cqe->bitfields;
1876 rte_le_to_cpu_16(cqe_start_tpa->pars_flags.flags);
1877 offset = cqe_start_tpa->placement_offset;
1878 /* seg_len = len_on_first_bd */
1879 len = rte_le_to_cpu_16(cqe_start_tpa->len_on_first_bd);
1880 vlan_tci = rte_le_to_cpu_16(cqe_start_tpa->vlan_tag);
1881 #ifdef RTE_LIBRTE_QEDE_DEBUG_RX
1882 bitfield_val = cqe_start_tpa->bitfields;
1884 rss_hash = rte_le_to_cpu_32(cqe_start_tpa->rss_hash);
1886 if (qede_tunn_exist(parse_flag)) {
1887 PMD_RX_LOG(INFO, rxq, "Rx tunneled packet\n");
1888 if (unlikely(qede_check_tunn_csum_l4(parse_flag))) {
1889 PMD_RX_LOG(ERR, rxq,
1890 "L4 csum failed, flags = 0x%x\n",
1892 rxq->rx_hw_errors++;
1893 ol_flags |= PKT_RX_L4_CKSUM_BAD;
1895 ol_flags |= PKT_RX_L4_CKSUM_GOOD;
1898 if (unlikely(qede_check_tunn_csum_l3(parse_flag))) {
1899 PMD_RX_LOG(ERR, rxq,
1900 "Outer L3 csum failed, flags = 0x%x\n",
1902 rxq->rx_hw_errors++;
1903 ol_flags |= PKT_RX_OUTER_IP_CKSUM_BAD;
1905 ol_flags |= PKT_RX_IP_CKSUM_GOOD;
1909 flags = cqe_start_tpa->tunnel_pars_flags.flags;
1911 flags = fp_cqe->tunnel_pars_flags.flags;
1912 tunn_parse_flag = flags;
1916 qede_rx_cqe_to_tunn_pkt_type(tunn_parse_flag);
1920 qede_rx_cqe_to_pkt_type_inner(parse_flag);
1922 /* Outer L3/L4 types is not available in CQE */
1923 packet_type |= qede_rx_cqe_to_pkt_type_outer(rx_mb);
1925 /* Outer L3/L4 types is not available in CQE.
1926 * Need to add offset to parse correctly,
1928 rx_mb->data_off = offset + RTE_PKTMBUF_HEADROOM;
1929 packet_type |= qede_rx_cqe_to_pkt_type_outer(rx_mb);
1931 packet_type |= qede_rx_cqe_to_pkt_type(parse_flag);
1934 /* Common handling for non-tunnel packets and for inner
1935 * headers in the case of tunnel.
1937 if (unlikely(qede_check_notunn_csum_l4(parse_flag))) {
1938 PMD_RX_LOG(ERR, rxq,
1939 "L4 csum failed, flags = 0x%x\n",
1941 rxq->rx_hw_errors++;
1942 ol_flags |= PKT_RX_L4_CKSUM_BAD;
1944 ol_flags |= PKT_RX_L4_CKSUM_GOOD;
1946 if (unlikely(qede_check_notunn_csum_l3(rx_mb, parse_flag))) {
1947 PMD_RX_LOG(ERR, rxq, "IP csum failed, flags = 0x%x\n",
1949 rxq->rx_hw_errors++;
1950 ol_flags |= PKT_RX_IP_CKSUM_BAD;
1952 ol_flags |= PKT_RX_IP_CKSUM_GOOD;
1955 if (CQE_HAS_VLAN(parse_flag) ||
1956 CQE_HAS_OUTER_VLAN(parse_flag)) {
1957 /* Note: FW doesn't indicate Q-in-Q packet */
1958 ol_flags |= PKT_RX_VLAN;
1959 if (qdev->vlan_strip_flg) {
1960 ol_flags |= PKT_RX_VLAN_STRIPPED;
1961 rx_mb->vlan_tci = vlan_tci;
1966 if (qdev->rss_enable) {
1967 ol_flags |= PKT_RX_RSS_HASH;
1968 rx_mb->hash.rss = rss_hash;
1972 qede_rx_bd_ring_consume(rxq);
1974 if (!tpa_start_flg && fp_cqe->bd_num > 1) {
1975 PMD_RX_LOG(DEBUG, rxq, "Jumbo-over-BD packet: %02x BDs"
1976 " len on first: %04x Total Len: %04x",
1977 fp_cqe->bd_num, len, pkt_len);
1978 num_segs = fp_cqe->bd_num - 1;
1980 if (qede_process_sg_pkts(p_rxq, seg1, num_segs,
1984 rx_alloc_count += num_segs;
1985 rxq->rx_segs += num_segs;
1987 rxq->rx_segs++; /* for the first segment */
1989 /* Prefetch next mbuf while processing current one. */
1990 preload_idx = rxq->sw_rx_cons & NUM_RX_BDS(rxq);
1991 rte_prefetch0(rxq->sw_rx_ring[preload_idx].mbuf);
1993 /* Update rest of the MBUF fields */
1994 rx_mb->data_off = offset + RTE_PKTMBUF_HEADROOM;
1995 rx_mb->port = rxq->port_id;
1996 rx_mb->ol_flags = ol_flags;
1997 rx_mb->data_len = len;
1998 rx_mb->packet_type = packet_type;
1999 #ifdef RTE_LIBRTE_QEDE_DEBUG_RX
2000 print_rx_bd_info(rx_mb, rxq, bitfield_val);
2002 if (!tpa_start_flg) {
2003 rx_mb->nb_segs = fp_cqe->bd_num;
2004 rx_mb->pkt_len = pkt_len;
2006 /* store ref to the updated mbuf */
2007 tpa_info->tpa_head = rx_mb;
2008 tpa_info->tpa_tail = tpa_info->tpa_head;
2010 rte_prefetch1(rte_pktmbuf_mtod(rx_mb, void *));
2012 if (!tpa_start_flg) {
2013 rx_pkts[rx_pkt] = rx_mb;
2017 ecore_chain_recycle_consumed(&rxq->rx_comp_ring);
2018 sw_comp_cons = ecore_chain_get_cons_idx(&rxq->rx_comp_ring);
2019 if (rx_pkt == nb_pkts) {
2020 PMD_RX_LOG(DEBUG, rxq,
2021 "Budget reached nb_pkts=%u received=%u",
2027 /* Request number of bufferes to be allocated in next loop */
2028 rxq->rx_alloc_count = rx_alloc_count;
2030 rxq->rcv_pkts += rx_pkt;
2032 PMD_RX_LOG(DEBUG, rxq, "rx_pkts=%u core=%d", rx_pkt, rte_lcore_id());
2038 qede_recv_pkts_cmt(void *p_fp_cmt, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
2040 struct qede_fastpath_cmt *fp_cmt = p_fp_cmt;
2041 uint16_t eng0_pkts, eng1_pkts;
2043 eng0_pkts = nb_pkts / 2;
2045 eng0_pkts = qede_recv_pkts(fp_cmt->fp0->rxq, rx_pkts, eng0_pkts);
2047 eng1_pkts = nb_pkts - eng0_pkts;
2049 eng1_pkts = qede_recv_pkts(fp_cmt->fp1->rxq, rx_pkts + eng0_pkts,
2052 return eng0_pkts + eng1_pkts;
2055 /* Populate scatter gather buffer descriptor fields */
2056 static inline uint16_t
2057 qede_encode_sg_bd(struct qede_tx_queue *p_txq, struct rte_mbuf *m_seg,
2058 struct eth_tx_2nd_bd **bd2, struct eth_tx_3rd_bd **bd3,
2061 struct qede_tx_queue *txq = p_txq;
2062 struct eth_tx_bd *tx_bd = NULL;
2064 uint16_t nb_segs = 0;
2066 /* Check for scattered buffers */
2068 if (start_seg == 0) {
2070 *bd2 = (struct eth_tx_2nd_bd *)
2071 ecore_chain_produce(&txq->tx_pbl);
2072 memset(*bd2, 0, sizeof(struct eth_tx_2nd_bd));
2075 mapping = rte_mbuf_data_iova(m_seg);
2076 QEDE_BD_SET_ADDR_LEN(*bd2, mapping, m_seg->data_len);
2077 PMD_TX_LOG(DEBUG, txq, "BD2 len %04x", m_seg->data_len);
2078 } else if (start_seg == 1) {
2080 *bd3 = (struct eth_tx_3rd_bd *)
2081 ecore_chain_produce(&txq->tx_pbl);
2082 memset(*bd3, 0, sizeof(struct eth_tx_3rd_bd));
2085 mapping = rte_mbuf_data_iova(m_seg);
2086 QEDE_BD_SET_ADDR_LEN(*bd3, mapping, m_seg->data_len);
2087 PMD_TX_LOG(DEBUG, txq, "BD3 len %04x", m_seg->data_len);
2089 tx_bd = (struct eth_tx_bd *)
2090 ecore_chain_produce(&txq->tx_pbl);
2091 memset(tx_bd, 0, sizeof(*tx_bd));
2093 mapping = rte_mbuf_data_iova(m_seg);
2094 QEDE_BD_SET_ADDR_LEN(tx_bd, mapping, m_seg->data_len);
2095 PMD_TX_LOG(DEBUG, txq, "BD len %04x", m_seg->data_len);
2098 m_seg = m_seg->next;
2101 /* Return total scattered buffers */
2105 #ifdef RTE_LIBRTE_QEDE_DEBUG_TX
2107 print_tx_bd_info(struct qede_tx_queue *txq,
2108 struct eth_tx_1st_bd *bd1,
2109 struct eth_tx_2nd_bd *bd2,
2110 struct eth_tx_3rd_bd *bd3,
2111 uint64_t tx_ol_flags)
2113 char ol_buf[256] = { 0 }; /* for verbose prints */
2116 PMD_TX_LOG(INFO, txq,
2117 "BD1: nbytes=0x%04x nbds=0x%04x bd_flags=0x%04x bf=0x%04x",
2118 rte_cpu_to_le_16(bd1->nbytes), bd1->data.nbds,
2119 bd1->data.bd_flags.bitfields,
2120 rte_cpu_to_le_16(bd1->data.bitfields));
2122 PMD_TX_LOG(INFO, txq,
2123 "BD2: nbytes=0x%04x bf1=0x%04x bf2=0x%04x tunn_ip=0x%04x\n",
2124 rte_cpu_to_le_16(bd2->nbytes), bd2->data.bitfields1,
2125 bd2->data.bitfields2, bd2->data.tunn_ip_size);
2127 PMD_TX_LOG(INFO, txq,
2128 "BD3: nbytes=0x%04x bf=0x%04x MSS=0x%04x "
2129 "tunn_l4_hdr_start_offset_w=0x%04x tunn_hdr_size=0x%04x\n",
2130 rte_cpu_to_le_16(bd3->nbytes),
2131 rte_cpu_to_le_16(bd3->data.bitfields),
2132 rte_cpu_to_le_16(bd3->data.lso_mss),
2133 bd3->data.tunn_l4_hdr_start_offset_w,
2134 bd3->data.tunn_hdr_size_w);
2136 rte_get_tx_ol_flag_list(tx_ol_flags, ol_buf, sizeof(ol_buf));
2137 PMD_TX_LOG(INFO, txq, "TX offloads = %s\n", ol_buf);
2141 /* TX prepare to check packets meets TX conditions */
2143 #ifdef RTE_LIBRTE_QEDE_DEBUG_TX
2144 qede_xmit_prep_pkts(void *p_txq, struct rte_mbuf **tx_pkts,
2147 struct qede_tx_queue *txq = p_txq;
2149 qede_xmit_prep_pkts(__rte_unused void *p_txq, struct rte_mbuf **tx_pkts,
2156 #ifdef RTE_LIBRTE_ETHDEV_DEBUG
2160 for (i = 0; i < nb_pkts; i++) {
2162 ol_flags = m->ol_flags;
2163 if (ol_flags & PKT_TX_TCP_SEG) {
2164 if (m->nb_segs >= ETH_TX_MAX_BDS_PER_LSO_PACKET) {
2168 /* TBD: confirm its ~9700B for both ? */
2169 if (m->tso_segsz > ETH_TX_MAX_NON_LSO_PKT_LEN) {
2174 if (m->nb_segs >= ETH_TX_MAX_BDS_PER_NON_LSO_PACKET) {
2179 if (ol_flags & QEDE_TX_OFFLOAD_NOTSUP_MASK) {
2180 /* We support only limited tunnel protocols */
2181 if (ol_flags & PKT_TX_TUNNEL_MASK) {
2184 temp = ol_flags & PKT_TX_TUNNEL_MASK;
2185 if (temp == PKT_TX_TUNNEL_VXLAN ||
2186 temp == PKT_TX_TUNNEL_GENEVE ||
2187 temp == PKT_TX_TUNNEL_MPLSINUDP ||
2188 temp == PKT_TX_TUNNEL_GRE)
2192 rte_errno = ENOTSUP;
2196 #ifdef RTE_LIBRTE_ETHDEV_DEBUG
2197 ret = rte_validate_tx_offload(m);
2205 #ifdef RTE_LIBRTE_QEDE_DEBUG_TX
2206 if (unlikely(i != nb_pkts))
2207 PMD_TX_LOG(ERR, txq, "TX prepare failed for %u\n",
2213 #define MPLSINUDP_HDR_SIZE (12)
2215 #ifdef RTE_LIBRTE_QEDE_DEBUG_TX
2217 qede_mpls_tunn_tx_sanity_check(struct rte_mbuf *mbuf,
2218 struct qede_tx_queue *txq)
2220 if (((mbuf->outer_l2_len + mbuf->outer_l3_len) / 2) > 0xff)
2221 PMD_TX_LOG(ERR, txq, "tunn_l4_hdr_start_offset overflow\n");
2222 if (((mbuf->outer_l2_len + mbuf->outer_l3_len +
2223 MPLSINUDP_HDR_SIZE) / 2) > 0xff)
2224 PMD_TX_LOG(ERR, txq, "tunn_hdr_size overflow\n");
2225 if (((mbuf->l2_len - MPLSINUDP_HDR_SIZE) / 2) >
2226 ETH_TX_DATA_2ND_BD_TUNN_INNER_L2_HDR_SIZE_W_MASK)
2227 PMD_TX_LOG(ERR, txq, "inner_l2_hdr_size overflow\n");
2228 if (((mbuf->l2_len - MPLSINUDP_HDR_SIZE + mbuf->l3_len) / 2) >
2229 ETH_TX_DATA_2ND_BD_L4_HDR_START_OFFSET_W_MASK)
2230 PMD_TX_LOG(ERR, txq, "inner_l2_hdr_size overflow\n");
2235 qede_xmit_pkts_regular(void *p_txq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
2237 struct qede_tx_queue *txq = p_txq;
2238 struct qede_dev *qdev = txq->qdev;
2239 struct ecore_dev *edev = &qdev->edev;
2240 struct eth_tx_1st_bd *bd1;
2241 struct eth_tx_2nd_bd *bd2;
2242 struct eth_tx_3rd_bd *bd3;
2243 struct rte_mbuf *m_seg = NULL;
2244 struct rte_mbuf *mbuf;
2245 struct rte_mbuf **sw_tx_ring;
2246 uint16_t nb_tx_pkts;
2249 uint16_t nb_frags = 0;
2250 uint16_t nb_pkt_sent = 0;
2252 uint64_t tx_ol_flags;
2255 uint8_t bd1_bd_flags_bf;
2257 if (unlikely(txq->nb_tx_avail < txq->tx_free_thresh)) {
2258 PMD_TX_LOG(DEBUG, txq, "send=%u avail=%u free_thresh=%u",
2259 nb_pkts, txq->nb_tx_avail, txq->tx_free_thresh);
2260 qede_process_tx_compl(edev, txq);
2263 nb_tx_pkts = nb_pkts;
2264 bd_prod = rte_cpu_to_le_16(ecore_chain_get_prod_idx(&txq->tx_pbl));
2265 sw_tx_ring = txq->sw_tx_ring;
2267 while (nb_tx_pkts--) {
2268 /* Init flags/values */
2274 bd1_bd_flags_bf = 0;
2281 /* Check minimum TX BDS availability against available BDs */
2282 if (unlikely(txq->nb_tx_avail < mbuf->nb_segs))
2285 tx_ol_flags = mbuf->ol_flags;
2286 bd1_bd_flags_bf |= 1 << ETH_TX_1ST_BD_FLAGS_START_BD_SHIFT;
2288 if (unlikely(txq->nb_tx_avail <
2289 ETH_TX_MIN_BDS_PER_NON_LSO_PKT))
2292 (mbuf->pkt_len & ETH_TX_DATA_1ST_BD_PKT_LEN_MASK)
2293 << ETH_TX_DATA_1ST_BD_PKT_LEN_SHIFT;
2295 /* Offload the IP checksum in the hardware */
2296 if (tx_ol_flags & PKT_TX_IP_CKSUM)
2298 1 << ETH_TX_1ST_BD_FLAGS_IP_CSUM_SHIFT;
2300 /* L4 checksum offload (tcp or udp) */
2301 if ((tx_ol_flags & (PKT_TX_IPV4 | PKT_TX_IPV6)) &&
2302 (tx_ol_flags & (PKT_TX_UDP_CKSUM | PKT_TX_TCP_CKSUM)))
2304 1 << ETH_TX_1ST_BD_FLAGS_L4_CSUM_SHIFT;
2306 /* Fill the entry in the SW ring and the BDs in the FW ring */
2308 sw_tx_ring[idx] = mbuf;
2311 bd1 = (struct eth_tx_1st_bd *)ecore_chain_produce(&txq->tx_pbl);
2312 memset(bd1, 0, sizeof(struct eth_tx_1st_bd));
2315 /* Map MBUF linear data for DMA and set in the BD1 */
2316 QEDE_BD_SET_ADDR_LEN(bd1, rte_mbuf_data_iova(mbuf),
2318 bd1->data.bitfields = rte_cpu_to_le_16(bd1_bf);
2319 bd1->data.bd_flags.bitfields = bd1_bd_flags_bf;
2321 /* Handle fragmented MBUF */
2322 if (unlikely(mbuf->nb_segs > 1)) {
2325 /* Encode scatter gather buffer descriptors */
2326 nb_frags = qede_encode_sg_bd(txq, m_seg, &bd2, &bd3,
2330 bd1->data.nbds = nbds + nb_frags;
2332 txq->nb_tx_avail -= bd1->data.nbds;
2335 rte_cpu_to_le_16(ecore_chain_get_prod_idx(&txq->tx_pbl));
2336 #ifdef RTE_LIBRTE_QEDE_DEBUG_TX
2337 print_tx_bd_info(txq, bd1, bd2, bd3, tx_ol_flags);
2343 /* Write value of prod idx into bd_prod */
2344 txq->tx_db.data.bd_prod = bd_prod;
2346 rte_compiler_barrier();
2347 DIRECT_REG_WR_RELAXED(edev, txq->doorbell_addr, txq->tx_db.raw);
2350 /* Check again for Tx completions */
2351 qede_process_tx_compl(edev, txq);
2353 PMD_TX_LOG(DEBUG, txq, "to_send=%u sent=%u bd_prod=%u core=%d",
2354 nb_pkts, nb_pkt_sent, TX_PROD(txq), rte_lcore_id());
2360 qede_xmit_pkts(void *p_txq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
2362 struct qede_tx_queue *txq = p_txq;
2363 struct qede_dev *qdev = txq->qdev;
2364 struct ecore_dev *edev = &qdev->edev;
2365 struct rte_mbuf *mbuf;
2366 struct rte_mbuf *m_seg = NULL;
2367 uint16_t nb_tx_pkts;
2371 uint16_t nb_pkt_sent = 0;
2375 __rte_unused bool tunn_flg;
2376 bool tunn_ipv6_ext_flg;
2377 struct eth_tx_1st_bd *bd1;
2378 struct eth_tx_2nd_bd *bd2;
2379 struct eth_tx_3rd_bd *bd3;
2380 uint64_t tx_ol_flags;
2384 uint8_t bd1_bd_flags_bf;
2393 uint8_t tunn_l4_hdr_start_offset;
2394 uint8_t tunn_hdr_size;
2395 uint8_t inner_l2_hdr_size;
2396 uint16_t inner_l4_hdr_offset;
2398 if (unlikely(txq->nb_tx_avail < txq->tx_free_thresh)) {
2399 PMD_TX_LOG(DEBUG, txq, "send=%u avail=%u free_thresh=%u",
2400 nb_pkts, txq->nb_tx_avail, txq->tx_free_thresh);
2401 qede_process_tx_compl(edev, txq);
2404 nb_tx_pkts = nb_pkts;
2405 bd_prod = rte_cpu_to_le_16(ecore_chain_get_prod_idx(&txq->tx_pbl));
2406 while (nb_tx_pkts--) {
2407 /* Init flags/values */
2417 bd1_bd_flags_bf = 0;
2422 mplsoudp_flg = false;
2423 tunn_ipv6_ext_flg = false;
2425 tunn_l4_hdr_start_offset = 0;
2430 /* Check minimum TX BDS availability against available BDs */
2431 if (unlikely(txq->nb_tx_avail < mbuf->nb_segs))
2434 tx_ol_flags = mbuf->ol_flags;
2435 bd1_bd_flags_bf |= 1 << ETH_TX_1ST_BD_FLAGS_START_BD_SHIFT;
2437 /* TX prepare would have already checked supported tunnel Tx
2438 * offloads. Don't rely on pkt_type marked by Rx, instead use
2439 * tx_ol_flags to decide.
2441 tunn_flg = !!(tx_ol_flags & PKT_TX_TUNNEL_MASK);
2444 /* Check against max which is Tunnel IPv6 + ext */
2445 if (unlikely(txq->nb_tx_avail <
2446 ETH_TX_MIN_BDS_PER_TUNN_IPV6_WITH_EXT_PKT))
2449 /* First indicate its a tunnel pkt */
2450 bd1_bf |= ETH_TX_DATA_1ST_BD_TUNN_FLAG_MASK <<
2451 ETH_TX_DATA_1ST_BD_TUNN_FLAG_SHIFT;
2452 /* Legacy FW had flipped behavior in regard to this bit
2453 * i.e. it needed to set to prevent FW from touching
2454 * encapsulated packets when it didn't need to.
2456 if (unlikely(txq->is_legacy)) {
2458 ETH_TX_DATA_1ST_BD_TUNN_FLAG_SHIFT;
2461 /* Outer IP checksum offload */
2462 if (tx_ol_flags & (PKT_TX_OUTER_IP_CKSUM |
2463 PKT_TX_OUTER_IPV4)) {
2465 ETH_TX_1ST_BD_FLAGS_TUNN_IP_CSUM_MASK <<
2466 ETH_TX_1ST_BD_FLAGS_TUNN_IP_CSUM_SHIFT;
2470 * Currently, only inner checksum offload in MPLS-in-UDP
2471 * tunnel with one MPLS label is supported. Both outer
2472 * and inner layers lengths need to be provided in
2475 if ((tx_ol_flags & PKT_TX_TUNNEL_MASK) ==
2476 PKT_TX_TUNNEL_MPLSINUDP) {
2477 mplsoudp_flg = true;
2478 #ifdef RTE_LIBRTE_QEDE_DEBUG_TX
2479 qede_mpls_tunn_tx_sanity_check(mbuf, txq);
2481 /* Outer L4 offset in two byte words */
2482 tunn_l4_hdr_start_offset =
2483 (mbuf->outer_l2_len + mbuf->outer_l3_len) / 2;
2484 /* Tunnel header size in two byte words */
2485 tunn_hdr_size = (mbuf->outer_l2_len +
2486 mbuf->outer_l3_len +
2487 MPLSINUDP_HDR_SIZE) / 2;
2488 /* Inner L2 header size in two byte words */
2489 inner_l2_hdr_size = (mbuf->l2_len -
2490 MPLSINUDP_HDR_SIZE) / 2;
2491 /* Inner L4 header offset from the beggining
2492 * of inner packet in two byte words
2494 inner_l4_hdr_offset = (mbuf->l2_len -
2495 MPLSINUDP_HDR_SIZE + mbuf->l3_len) / 2;
2497 /* Inner L2 size and address type */
2498 bd2_bf1 |= (inner_l2_hdr_size &
2499 ETH_TX_DATA_2ND_BD_TUNN_INNER_L2_HDR_SIZE_W_MASK) <<
2500 ETH_TX_DATA_2ND_BD_TUNN_INNER_L2_HDR_SIZE_W_SHIFT;
2501 bd2_bf1 |= (UNICAST_ADDRESS &
2502 ETH_TX_DATA_2ND_BD_TUNN_INNER_ETH_TYPE_MASK) <<
2503 ETH_TX_DATA_2ND_BD_TUNN_INNER_ETH_TYPE_SHIFT;
2504 /* Treated as IPv6+Ext */
2506 1 << ETH_TX_DATA_2ND_BD_TUNN_IPV6_EXT_SHIFT;
2508 /* Mark inner IPv6 if present */
2509 if (tx_ol_flags & PKT_TX_IPV6)
2511 1 << ETH_TX_DATA_2ND_BD_TUNN_INNER_IPV6_SHIFT;
2513 /* Inner L4 offsets */
2514 if ((tx_ol_flags & (PKT_TX_IPV4 | PKT_TX_IPV6)) &&
2515 (tx_ol_flags & (PKT_TX_UDP_CKSUM |
2516 PKT_TX_TCP_CKSUM))) {
2517 /* Determines if BD3 is needed */
2518 tunn_ipv6_ext_flg = true;
2519 if ((tx_ol_flags & PKT_TX_L4_MASK) ==
2522 1 << ETH_TX_DATA_2ND_BD_L4_UDP_SHIFT;
2525 /* TODO other pseudo checksum modes are
2529 ETH_L4_PSEUDO_CSUM_CORRECT_LENGTH <<
2530 ETH_TX_DATA_2ND_BD_L4_PSEUDO_CSUM_MODE_SHIFT;
2531 bd2_bf2 |= (inner_l4_hdr_offset &
2532 ETH_TX_DATA_2ND_BD_L4_HDR_START_OFFSET_W_MASK) <<
2533 ETH_TX_DATA_2ND_BD_L4_HDR_START_OFFSET_W_SHIFT;
2535 } /* End MPLSoUDP */
2536 } /* End Tunnel handling */
2538 if (tx_ol_flags & PKT_TX_TCP_SEG) {
2540 if (unlikely(txq->nb_tx_avail <
2541 ETH_TX_MIN_BDS_PER_LSO_PKT))
2543 /* For LSO, packet header and payload must reside on
2544 * buffers pointed by different BDs. Using BD1 for HDR
2545 * and BD2 onwards for data.
2547 hdr_size = mbuf->l2_len + mbuf->l3_len + mbuf->l4_len;
2549 hdr_size += mbuf->outer_l2_len +
2552 bd1_bd_flags_bf |= 1 << ETH_TX_1ST_BD_FLAGS_LSO_SHIFT;
2554 1 << ETH_TX_1ST_BD_FLAGS_IP_CSUM_SHIFT;
2555 /* PKT_TX_TCP_SEG implies PKT_TX_TCP_CKSUM */
2557 1 << ETH_TX_1ST_BD_FLAGS_L4_CSUM_SHIFT;
2558 mss = rte_cpu_to_le_16(mbuf->tso_segsz);
2559 /* Using one header BD */
2560 bd3_bf |= rte_cpu_to_le_16(1 <<
2561 ETH_TX_DATA_3RD_BD_HDR_NBD_SHIFT);
2563 if (unlikely(txq->nb_tx_avail <
2564 ETH_TX_MIN_BDS_PER_NON_LSO_PKT))
2567 (mbuf->pkt_len & ETH_TX_DATA_1ST_BD_PKT_LEN_MASK)
2568 << ETH_TX_DATA_1ST_BD_PKT_LEN_SHIFT;
2571 /* Descriptor based VLAN insertion */
2572 if (tx_ol_flags & PKT_TX_VLAN_PKT) {
2573 vlan = rte_cpu_to_le_16(mbuf->vlan_tci);
2575 1 << ETH_TX_1ST_BD_FLAGS_VLAN_INSERTION_SHIFT;
2578 /* Offload the IP checksum in the hardware */
2579 if (tx_ol_flags & PKT_TX_IP_CKSUM) {
2581 1 << ETH_TX_1ST_BD_FLAGS_IP_CSUM_SHIFT;
2582 /* There's no DPDK flag to request outer-L4 csum
2583 * offload. But in the case of tunnel if inner L3 or L4
2584 * csum offload is requested then we need to force
2585 * recalculation of L4 tunnel header csum also.
2587 if (tunn_flg && ((tx_ol_flags & PKT_TX_TUNNEL_MASK) !=
2588 PKT_TX_TUNNEL_GRE)) {
2590 ETH_TX_1ST_BD_FLAGS_TUNN_L4_CSUM_MASK <<
2591 ETH_TX_1ST_BD_FLAGS_TUNN_L4_CSUM_SHIFT;
2595 /* L4 checksum offload (tcp or udp) */
2596 if ((tx_ol_flags & (PKT_TX_IPV4 | PKT_TX_IPV6)) &&
2597 (tx_ol_flags & (PKT_TX_UDP_CKSUM | PKT_TX_TCP_CKSUM))) {
2599 1 << ETH_TX_1ST_BD_FLAGS_L4_CSUM_SHIFT;
2600 /* There's no DPDK flag to request outer-L4 csum
2601 * offload. But in the case of tunnel if inner L3 or L4
2602 * csum offload is requested then we need to force
2603 * recalculation of L4 tunnel header csum also.
2607 ETH_TX_1ST_BD_FLAGS_TUNN_L4_CSUM_MASK <<
2608 ETH_TX_1ST_BD_FLAGS_TUNN_L4_CSUM_SHIFT;
2612 /* Fill the entry in the SW ring and the BDs in the FW ring */
2614 txq->sw_tx_ring[idx] = mbuf;
2617 bd1 = (struct eth_tx_1st_bd *)ecore_chain_produce(&txq->tx_pbl);
2618 memset(bd1, 0, sizeof(struct eth_tx_1st_bd));
2621 /* Map MBUF linear data for DMA and set in the BD1 */
2622 QEDE_BD_SET_ADDR_LEN(bd1, rte_mbuf_data_iova(mbuf),
2624 bd1->data.bitfields = rte_cpu_to_le_16(bd1_bf);
2625 bd1->data.bd_flags.bitfields = bd1_bd_flags_bf;
2626 bd1->data.vlan = vlan;
2628 if (lso_flg || mplsoudp_flg) {
2629 bd2 = (struct eth_tx_2nd_bd *)ecore_chain_produce
2631 memset(bd2, 0, sizeof(struct eth_tx_2nd_bd));
2635 QEDE_BD_SET_ADDR_LEN(bd1, rte_mbuf_data_iova(mbuf),
2638 QEDE_BD_SET_ADDR_LEN(bd2, (hdr_size +
2639 rte_mbuf_data_iova(mbuf)),
2640 mbuf->data_len - hdr_size);
2641 bd2->data.bitfields1 = rte_cpu_to_le_16(bd2_bf1);
2643 bd2->data.bitfields2 =
2644 rte_cpu_to_le_16(bd2_bf2);
2646 bd2->data.tunn_ip_size =
2647 rte_cpu_to_le_16(mbuf->outer_l3_len);
2650 if (lso_flg || (mplsoudp_flg && tunn_ipv6_ext_flg)) {
2651 bd3 = (struct eth_tx_3rd_bd *)
2652 ecore_chain_produce(&txq->tx_pbl);
2653 memset(bd3, 0, sizeof(struct eth_tx_3rd_bd));
2655 bd3->data.bitfields = rte_cpu_to_le_16(bd3_bf);
2657 bd3->data.lso_mss = mss;
2659 bd3->data.tunn_l4_hdr_start_offset_w =
2660 tunn_l4_hdr_start_offset;
2661 bd3->data.tunn_hdr_size_w =
2667 /* Handle fragmented MBUF */
2670 /* Encode scatter gather buffer descriptors if required */
2671 nb_frags = qede_encode_sg_bd(txq, m_seg, &bd2, &bd3, nbds - 1);
2672 bd1->data.nbds = nbds + nb_frags;
2674 txq->nb_tx_avail -= bd1->data.nbds;
2677 rte_cpu_to_le_16(ecore_chain_get_prod_idx(&txq->tx_pbl));
2678 #ifdef RTE_LIBRTE_QEDE_DEBUG_TX
2679 print_tx_bd_info(txq, bd1, bd2, bd3, tx_ol_flags);
2685 /* Write value of prod idx into bd_prod */
2686 txq->tx_db.data.bd_prod = bd_prod;
2688 rte_compiler_barrier();
2689 DIRECT_REG_WR_RELAXED(edev, txq->doorbell_addr, txq->tx_db.raw);
2692 /* Check again for Tx completions */
2693 qede_process_tx_compl(edev, txq);
2695 PMD_TX_LOG(DEBUG, txq, "to_send=%u sent=%u bd_prod=%u core=%d",
2696 nb_pkts, nb_pkt_sent, TX_PROD(txq), rte_lcore_id());
2702 qede_xmit_pkts_cmt(void *p_fp_cmt, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
2704 struct qede_fastpath_cmt *fp_cmt = p_fp_cmt;
2705 uint16_t eng0_pkts, eng1_pkts;
2707 eng0_pkts = nb_pkts / 2;
2709 eng0_pkts = qede_xmit_pkts(fp_cmt->fp0->txq, tx_pkts, eng0_pkts);
2711 eng1_pkts = nb_pkts - eng0_pkts;
2713 eng1_pkts = qede_xmit_pkts(fp_cmt->fp1->txq, tx_pkts + eng0_pkts,
2716 return eng0_pkts + eng1_pkts;
2720 qede_rxtx_pkts_dummy(__rte_unused void *p_rxq,
2721 __rte_unused struct rte_mbuf **pkts,
2722 __rte_unused uint16_t nb_pkts)
2728 /* this function does a fake walk through over completion queue
2729 * to calculate number of BDs used by HW.
2730 * At the end, it restores the state of completion queue.
2733 qede_parse_fp_cqe(struct qede_rx_queue *rxq)
2735 uint16_t hw_comp_cons, sw_comp_cons, bd_count = 0;
2736 union eth_rx_cqe *cqe, *orig_cqe = NULL;
2738 hw_comp_cons = rte_le_to_cpu_16(*rxq->hw_cons_ptr);
2739 sw_comp_cons = ecore_chain_get_cons_idx(&rxq->rx_comp_ring);
2741 if (hw_comp_cons == sw_comp_cons)
2744 /* Get the CQE from the completion ring */
2745 cqe = (union eth_rx_cqe *)ecore_chain_consume(&rxq->rx_comp_ring);
2748 while (sw_comp_cons != hw_comp_cons) {
2749 switch (cqe->fast_path_regular.type) {
2750 case ETH_RX_CQE_TYPE_REGULAR:
2751 bd_count += cqe->fast_path_regular.bd_num;
2753 case ETH_RX_CQE_TYPE_TPA_END:
2754 bd_count += cqe->fast_path_tpa_end.num_of_bds;
2761 (union eth_rx_cqe *)ecore_chain_consume(&rxq->rx_comp_ring);
2762 sw_comp_cons = ecore_chain_get_cons_idx(&rxq->rx_comp_ring);
2765 /* revert comp_ring to original state */
2766 ecore_chain_set_cons(&rxq->rx_comp_ring, sw_comp_cons, orig_cqe);
2772 qede_rx_descriptor_status(void *p_rxq, uint16_t offset)
2774 uint16_t hw_bd_cons, sw_bd_cons, sw_bd_prod;
2775 uint16_t produced, consumed;
2776 struct qede_rx_queue *rxq = p_rxq;
2778 if (offset > rxq->nb_rx_desc)
2781 sw_bd_cons = ecore_chain_get_cons_idx(&rxq->rx_bd_ring);
2782 sw_bd_prod = ecore_chain_get_prod_idx(&rxq->rx_bd_ring);
2784 /* find BDs used by HW from completion queue elements */
2785 hw_bd_cons = sw_bd_cons + qede_parse_fp_cqe(rxq);
2787 if (hw_bd_cons < sw_bd_cons)
2788 /* wraparound case */
2789 consumed = (0xffff - sw_bd_cons) + hw_bd_cons;
2791 consumed = hw_bd_cons - sw_bd_cons;
2793 if (offset <= consumed)
2794 return RTE_ETH_RX_DESC_DONE;
2796 if (sw_bd_prod < sw_bd_cons)
2797 /* wraparound case */
2798 produced = (0xffff - sw_bd_cons) + sw_bd_prod;
2800 produced = sw_bd_prod - sw_bd_cons;
2802 if (offset <= produced)
2803 return RTE_ETH_RX_DESC_AVAIL;
2805 return RTE_ETH_RX_DESC_UNAVAIL;