1 /* QLogic qede NIC Driver
2 * Copyright (c) 2015-2017 QLogic Corporation
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and /or other materials
21 * provided with the distribution.
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32 #include <linux/netdevice.h>
33 #include <linux/etherdevice.h>
34 #include <linux/skbuff.h>
35 #include <linux/bpf_trace.h>
36 #include <net/udp_tunnel.h>
40 #include <linux/if_ether.h>
41 #include <linux/if_vlan.h>
42 #include <net/ip6_checksum.h>
45 #include <linux/qed/qed_if.h>
47 /*********************************
48 * Content also used by slowpath *
49 *********************************/
51 int qede_alloc_rx_buffer(struct qede_rx_queue *rxq, bool allow_lazy)
53 struct sw_rx_data *sw_rx_data;
54 struct eth_rx_bd *rx_bd;
58 /* In case lazy-allocation is allowed, postpone allocation until the
59 * end of the NAPI run. We'd still need to make sure the Rx ring has
60 * sufficient buffers to guarantee an additional Rx interrupt.
62 if (allow_lazy && likely(rxq->filled_buffers > 12)) {
63 rxq->filled_buffers--;
67 data = alloc_pages(GFP_ATOMIC, 0);
71 /* Map the entire page as it would be used
72 * for multiple RX buffer segment size mapping.
74 mapping = dma_map_page(rxq->dev, data, 0,
75 PAGE_SIZE, rxq->data_direction);
76 if (unlikely(dma_mapping_error(rxq->dev, mapping))) {
81 sw_rx_data = &rxq->sw_rx_ring[rxq->sw_rx_prod & NUM_RX_BDS_MAX];
82 sw_rx_data->page_offset = 0;
83 sw_rx_data->data = data;
84 sw_rx_data->mapping = mapping;
86 /* Advance PROD and get BD pointer */
87 rx_bd = (struct eth_rx_bd *)qed_chain_produce(&rxq->rx_bd_ring);
89 rx_bd->addr.hi = cpu_to_le32(upper_32_bits(mapping));
90 rx_bd->addr.lo = cpu_to_le32(lower_32_bits(mapping) +
94 rxq->filled_buffers++;
99 /* Unmap the data and free skb */
100 int qede_free_tx_pkt(struct qede_dev *edev, struct qede_tx_queue *txq, int *len)
102 u16 idx = txq->sw_tx_cons;
103 struct sk_buff *skb = txq->sw_tx_ring.skbs[idx].skb;
104 struct eth_tx_1st_bd *first_bd;
105 struct eth_tx_bd *tx_data_bd;
106 int bds_consumed = 0;
108 bool data_split = txq->sw_tx_ring.skbs[idx].flags & QEDE_TSO_SPLIT_BD;
109 int i, split_bd_len = 0;
111 if (unlikely(!skb)) {
113 "skb is null for txq idx=%d txq->sw_tx_cons=%d txq->sw_tx_prod=%d\n",
114 idx, txq->sw_tx_cons, txq->sw_tx_prod);
120 first_bd = (struct eth_tx_1st_bd *)qed_chain_consume(&txq->tx_pbl);
124 nbds = first_bd->data.nbds;
127 struct eth_tx_bd *split = (struct eth_tx_bd *)
128 qed_chain_consume(&txq->tx_pbl);
129 split_bd_len = BD_UNMAP_LEN(split);
132 dma_unmap_single(&edev->pdev->dev, BD_UNMAP_ADDR(first_bd),
133 BD_UNMAP_LEN(first_bd) + split_bd_len, DMA_TO_DEVICE);
135 /* Unmap the data of the skb frags */
136 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++, bds_consumed++) {
137 tx_data_bd = (struct eth_tx_bd *)
138 qed_chain_consume(&txq->tx_pbl);
139 dma_unmap_page(&edev->pdev->dev, BD_UNMAP_ADDR(tx_data_bd),
140 BD_UNMAP_LEN(tx_data_bd), DMA_TO_DEVICE);
143 while (bds_consumed++ < nbds)
144 qed_chain_consume(&txq->tx_pbl);
147 dev_kfree_skb_any(skb);
148 txq->sw_tx_ring.skbs[idx].skb = NULL;
149 txq->sw_tx_ring.skbs[idx].flags = 0;
154 /* Unmap the data and free skb when mapping failed during start_xmit */
155 static void qede_free_failed_tx_pkt(struct qede_tx_queue *txq,
156 struct eth_tx_1st_bd *first_bd,
157 int nbd, bool data_split)
159 u16 idx = txq->sw_tx_prod;
160 struct sk_buff *skb = txq->sw_tx_ring.skbs[idx].skb;
161 struct eth_tx_bd *tx_data_bd;
162 int i, split_bd_len = 0;
164 /* Return prod to its position before this skb was handled */
165 qed_chain_set_prod(&txq->tx_pbl,
166 le16_to_cpu(txq->tx_db.data.bd_prod), first_bd);
168 first_bd = (struct eth_tx_1st_bd *)qed_chain_produce(&txq->tx_pbl);
171 struct eth_tx_bd *split = (struct eth_tx_bd *)
172 qed_chain_produce(&txq->tx_pbl);
173 split_bd_len = BD_UNMAP_LEN(split);
177 dma_unmap_single(txq->dev, BD_UNMAP_ADDR(first_bd),
178 BD_UNMAP_LEN(first_bd) + split_bd_len, DMA_TO_DEVICE);
180 /* Unmap the data of the skb frags */
181 for (i = 0; i < nbd; i++) {
182 tx_data_bd = (struct eth_tx_bd *)
183 qed_chain_produce(&txq->tx_pbl);
184 if (tx_data_bd->nbytes)
185 dma_unmap_page(txq->dev,
186 BD_UNMAP_ADDR(tx_data_bd),
187 BD_UNMAP_LEN(tx_data_bd), DMA_TO_DEVICE);
190 /* Return again prod to its position before this skb was handled */
191 qed_chain_set_prod(&txq->tx_pbl,
192 le16_to_cpu(txq->tx_db.data.bd_prod), first_bd);
195 dev_kfree_skb_any(skb);
196 txq->sw_tx_ring.skbs[idx].skb = NULL;
197 txq->sw_tx_ring.skbs[idx].flags = 0;
200 static u32 qede_xmit_type(struct sk_buff *skb, int *ipv6_ext)
202 u32 rc = XMIT_L4_CSUM;
205 if (skb->ip_summed != CHECKSUM_PARTIAL)
208 l3_proto = vlan_get_protocol(skb);
209 if (l3_proto == htons(ETH_P_IPV6) &&
210 (ipv6_hdr(skb)->nexthdr == NEXTHDR_IPV6))
213 if (skb->encapsulation) {
215 if (skb_is_gso(skb)) {
216 unsigned short gso_type = skb_shinfo(skb)->gso_type;
218 if ((gso_type & SKB_GSO_UDP_TUNNEL_CSUM) ||
219 (gso_type & SKB_GSO_GRE_CSUM))
220 rc |= XMIT_ENC_GSO_L4_CSUM;
233 static void qede_set_params_for_ipv6_ext(struct sk_buff *skb,
234 struct eth_tx_2nd_bd *second_bd,
235 struct eth_tx_3rd_bd *third_bd)
238 u16 bd2_bits1 = 0, bd2_bits2 = 0;
240 bd2_bits1 |= (1 << ETH_TX_DATA_2ND_BD_IPV6_EXT_SHIFT);
242 bd2_bits2 |= ((((u8 *)skb_transport_header(skb) - skb->data) >> 1) &
243 ETH_TX_DATA_2ND_BD_L4_HDR_START_OFFSET_W_MASK)
244 << ETH_TX_DATA_2ND_BD_L4_HDR_START_OFFSET_W_SHIFT;
246 bd2_bits1 |= (ETH_L4_PSEUDO_CSUM_CORRECT_LENGTH <<
247 ETH_TX_DATA_2ND_BD_L4_PSEUDO_CSUM_MODE_SHIFT);
249 if (vlan_get_protocol(skb) == htons(ETH_P_IPV6))
250 l4_proto = ipv6_hdr(skb)->nexthdr;
252 l4_proto = ip_hdr(skb)->protocol;
254 if (l4_proto == IPPROTO_UDP)
255 bd2_bits1 |= 1 << ETH_TX_DATA_2ND_BD_L4_UDP_SHIFT;
258 third_bd->data.bitfields |=
259 cpu_to_le16(((tcp_hdrlen(skb) / 4) &
260 ETH_TX_DATA_3RD_BD_TCP_HDR_LEN_DW_MASK) <<
261 ETH_TX_DATA_3RD_BD_TCP_HDR_LEN_DW_SHIFT);
263 second_bd->data.bitfields1 = cpu_to_le16(bd2_bits1);
264 second_bd->data.bitfields2 = cpu_to_le16(bd2_bits2);
267 static int map_frag_to_bd(struct qede_tx_queue *txq,
268 skb_frag_t *frag, struct eth_tx_bd *bd)
272 /* Map skb non-linear frag data for DMA */
273 mapping = skb_frag_dma_map(txq->dev, frag, 0,
274 skb_frag_size(frag), DMA_TO_DEVICE);
275 if (unlikely(dma_mapping_error(txq->dev, mapping)))
278 /* Setup the data pointer of the frag data */
279 BD_SET_UNMAP_ADDR_LEN(bd, mapping, skb_frag_size(frag));
284 static u16 qede_get_skb_hlen(struct sk_buff *skb, bool is_encap_pkt)
287 return (skb_inner_transport_header(skb) +
288 inner_tcp_hdrlen(skb) - skb->data);
290 return (skb_transport_header(skb) +
291 tcp_hdrlen(skb) - skb->data);
294 /* +2 for 1st BD for headers and 2nd BD for headlen (if required) */
295 #if ((MAX_SKB_FRAGS + 2) > ETH_TX_MAX_BDS_PER_NON_LSO_PACKET)
296 static bool qede_pkt_req_lin(struct sk_buff *skb, u8 xmit_type)
298 int allowed_frags = ETH_TX_MAX_BDS_PER_NON_LSO_PACKET - 1;
300 if (xmit_type & XMIT_LSO) {
303 hlen = qede_get_skb_hlen(skb, xmit_type & XMIT_ENC);
305 /* linear payload would require its own BD */
306 if (skb_headlen(skb) > hlen)
310 return (skb_shinfo(skb)->nr_frags > allowed_frags);
314 static inline void qede_update_tx_producer(struct qede_tx_queue *txq)
316 /* wmb makes sure that the BDs data is updated before updating the
317 * producer, otherwise FW may read old data from the BDs.
321 writel(txq->tx_db.raw, txq->doorbell_addr);
323 /* Fence required to flush the write combined buffer, since another
324 * CPU may write to the same doorbell address and data may be lost
325 * due to relaxed order nature of write combined bar.
330 static int qede_xdp_xmit(struct qede_dev *edev, struct qede_fastpath *fp,
331 struct sw_rx_data *metadata, u16 padding, u16 length)
333 struct qede_tx_queue *txq = fp->xdp_tx;
334 struct eth_tx_1st_bd *first_bd;
335 u16 idx = txq->sw_tx_prod;
338 if (!qed_chain_get_elem_left(&txq->tx_pbl)) {
343 first_bd = (struct eth_tx_1st_bd *)qed_chain_produce(&txq->tx_pbl);
345 memset(first_bd, 0, sizeof(*first_bd));
346 first_bd->data.bd_flags.bitfields =
347 BIT(ETH_TX_1ST_BD_FLAGS_START_BD_SHIFT);
349 val = (length & ETH_TX_DATA_1ST_BD_PKT_LEN_MASK) <<
350 ETH_TX_DATA_1ST_BD_PKT_LEN_SHIFT;
352 first_bd->data.bitfields |= cpu_to_le16(val);
353 first_bd->data.nbds = 1;
355 /* We can safely ignore the offset, as it's 0 for XDP */
356 BD_SET_UNMAP_ADDR_LEN(first_bd, metadata->mapping + padding, length);
358 /* Synchronize the buffer back to device, as program [probably]
361 dma_sync_single_for_device(&edev->pdev->dev,
362 metadata->mapping + padding,
363 length, PCI_DMA_TODEVICE);
365 txq->sw_tx_ring.xdp[idx].page = metadata->data;
366 txq->sw_tx_ring.xdp[idx].mapping = metadata->mapping;
367 txq->sw_tx_prod = (txq->sw_tx_prod + 1) % txq->num_tx_buffers;
369 /* Mark the fastpath for future XDP doorbell */
375 int qede_txq_has_work(struct qede_tx_queue *txq)
379 /* Tell compiler that consumer and producer can change */
381 hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr);
382 if (qed_chain_get_cons_idx(&txq->tx_pbl) == hw_bd_cons + 1)
385 return hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl);
388 static void qede_xdp_tx_int(struct qede_dev *edev, struct qede_tx_queue *txq)
392 hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr);
395 while (hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl)) {
396 qed_chain_consume(&txq->tx_pbl);
397 idx = txq->sw_tx_cons;
399 dma_unmap_page(&edev->pdev->dev,
400 txq->sw_tx_ring.xdp[idx].mapping,
401 PAGE_SIZE, DMA_BIDIRECTIONAL);
402 __free_page(txq->sw_tx_ring.xdp[idx].page);
404 txq->sw_tx_cons = (txq->sw_tx_cons + 1) % txq->num_tx_buffers;
409 static int qede_tx_int(struct qede_dev *edev, struct qede_tx_queue *txq)
411 unsigned int pkts_compl = 0, bytes_compl = 0;
412 struct netdev_queue *netdev_txq;
416 netdev_txq = netdev_get_tx_queue(edev->ndev, txq->ndev_txq_id);
418 hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr);
421 while (hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl)) {
424 rc = qede_free_tx_pkt(edev, txq, &len);
426 DP_NOTICE(edev, "hw_bd_cons = %d, chain_cons=%d\n",
428 qed_chain_get_cons_idx(&txq->tx_pbl));
434 txq->sw_tx_cons = (txq->sw_tx_cons + 1) % txq->num_tx_buffers;
438 netdev_tx_completed_queue(netdev_txq, pkts_compl, bytes_compl);
440 /* Need to make the tx_bd_cons update visible to start_xmit()
441 * before checking for netif_tx_queue_stopped(). Without the
442 * memory barrier, there is a small possibility that
443 * start_xmit() will miss it and cause the queue to be stopped
445 * On the other hand we need an rmb() here to ensure the proper
446 * ordering of bit testing in the following
447 * netif_tx_queue_stopped(txq) call.
451 if (unlikely(netif_tx_queue_stopped(netdev_txq))) {
452 /* Taking tx_lock is needed to prevent reenabling the queue
453 * while it's empty. This could have happen if rx_action() gets
454 * suspended in qede_tx_int() after the condition before
455 * netif_tx_wake_queue(), while tx_action (qede_start_xmit()):
457 * stops the queue->sees fresh tx_bd_cons->releases the queue->
458 * sends some packets consuming the whole queue again->
462 __netif_tx_lock(netdev_txq, smp_processor_id());
464 if ((netif_tx_queue_stopped(netdev_txq)) &&
465 (edev->state == QEDE_STATE_OPEN) &&
466 (qed_chain_get_elem_left(&txq->tx_pbl)
467 >= (MAX_SKB_FRAGS + 1))) {
468 netif_tx_wake_queue(netdev_txq);
469 DP_VERBOSE(edev, NETIF_MSG_TX_DONE,
470 "Wake queue was called\n");
473 __netif_tx_unlock(netdev_txq);
479 bool qede_has_rx_work(struct qede_rx_queue *rxq)
481 u16 hw_comp_cons, sw_comp_cons;
483 /* Tell compiler that status block fields can change */
486 hw_comp_cons = le16_to_cpu(*rxq->hw_cons_ptr);
487 sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring);
489 return hw_comp_cons != sw_comp_cons;
492 static inline void qede_rx_bd_ring_consume(struct qede_rx_queue *rxq)
494 qed_chain_consume(&rxq->rx_bd_ring);
498 /* This function reuses the buffer(from an offset) from
499 * consumer index to producer index in the bd ring
501 static inline void qede_reuse_page(struct qede_rx_queue *rxq,
502 struct sw_rx_data *curr_cons)
504 struct eth_rx_bd *rx_bd_prod = qed_chain_produce(&rxq->rx_bd_ring);
505 struct sw_rx_data *curr_prod;
506 dma_addr_t new_mapping;
508 curr_prod = &rxq->sw_rx_ring[rxq->sw_rx_prod & NUM_RX_BDS_MAX];
509 *curr_prod = *curr_cons;
511 new_mapping = curr_prod->mapping + curr_prod->page_offset;
513 rx_bd_prod->addr.hi = cpu_to_le32(upper_32_bits(new_mapping));
514 rx_bd_prod->addr.lo = cpu_to_le32(lower_32_bits(new_mapping) +
518 curr_cons->data = NULL;
521 /* In case of allocation failures reuse buffers
522 * from consumer index to produce buffers for firmware
524 void qede_recycle_rx_bd_ring(struct qede_rx_queue *rxq, u8 count)
526 struct sw_rx_data *curr_cons;
528 for (; count > 0; count--) {
529 curr_cons = &rxq->sw_rx_ring[rxq->sw_rx_cons & NUM_RX_BDS_MAX];
530 qede_reuse_page(rxq, curr_cons);
531 qede_rx_bd_ring_consume(rxq);
535 static inline int qede_realloc_rx_buffer(struct qede_rx_queue *rxq,
536 struct sw_rx_data *curr_cons)
538 /* Move to the next segment in the page */
539 curr_cons->page_offset += rxq->rx_buf_seg_size;
541 if (curr_cons->page_offset == PAGE_SIZE) {
542 if (unlikely(qede_alloc_rx_buffer(rxq, true))) {
543 /* Since we failed to allocate new buffer
544 * current buffer can be used again.
546 curr_cons->page_offset -= rxq->rx_buf_seg_size;
551 dma_unmap_page(rxq->dev, curr_cons->mapping,
552 PAGE_SIZE, rxq->data_direction);
554 /* Increment refcount of the page as we don't want
555 * network stack to take the ownership of the page
556 * which can be recycled multiple times by the driver.
558 page_ref_inc(curr_cons->data);
559 qede_reuse_page(rxq, curr_cons);
565 void qede_update_rx_prod(struct qede_dev *edev, struct qede_rx_queue *rxq)
567 u16 bd_prod = qed_chain_get_prod_idx(&rxq->rx_bd_ring);
568 u16 cqe_prod = qed_chain_get_prod_idx(&rxq->rx_comp_ring);
569 struct eth_rx_prod_data rx_prods = {0};
571 /* Update producers */
572 rx_prods.bd_prod = cpu_to_le16(bd_prod);
573 rx_prods.cqe_prod = cpu_to_le16(cqe_prod);
575 /* Make sure that the BD and SGE data is updated before updating the
576 * producers since FW might read the BD/SGE right after the producer
581 internal_ram_wr(rxq->hw_rxq_prod_addr, sizeof(rx_prods),
584 /* mmiowb is needed to synchronize doorbell writes from more than one
585 * processor. It guarantees that the write arrives to the device before
586 * the napi lock is released and another qede_poll is called (possibly
587 * on another CPU). Without this barrier, the next doorbell can bypass
588 * this doorbell. This is applicable to IA64/Altix systems.
593 static void qede_get_rxhash(struct sk_buff *skb, u8 bitfields, __le32 rss_hash)
595 enum pkt_hash_types hash_type = PKT_HASH_TYPE_NONE;
596 enum rss_hash_type htype;
599 htype = GET_FIELD(bitfields, ETH_FAST_PATH_RX_REG_CQE_RSS_HASH_TYPE);
601 hash_type = ((htype == RSS_HASH_TYPE_IPV4) ||
602 (htype == RSS_HASH_TYPE_IPV6)) ?
603 PKT_HASH_TYPE_L3 : PKT_HASH_TYPE_L4;
604 hash = le32_to_cpu(rss_hash);
606 skb_set_hash(skb, hash, hash_type);
609 static void qede_set_skb_csum(struct sk_buff *skb, u8 csum_flag)
611 skb_checksum_none_assert(skb);
613 if (csum_flag & QEDE_CSUM_UNNECESSARY)
614 skb->ip_summed = CHECKSUM_UNNECESSARY;
616 if (csum_flag & QEDE_TUNN_CSUM_UNNECESSARY) {
618 skb->encapsulation = 1;
622 static inline void qede_skb_receive(struct qede_dev *edev,
623 struct qede_fastpath *fp,
624 struct qede_rx_queue *rxq,
625 struct sk_buff *skb, u16 vlan_tag)
628 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tag);
630 napi_gro_receive(&fp->napi, skb);
633 static void qede_set_gro_params(struct qede_dev *edev,
635 struct eth_fast_path_rx_tpa_start_cqe *cqe)
637 u16 parsing_flags = le16_to_cpu(cqe->pars_flags.flags);
639 if (((parsing_flags >> PARSING_AND_ERR_FLAGS_L3TYPE_SHIFT) &
640 PARSING_AND_ERR_FLAGS_L3TYPE_MASK) == 2)
641 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
643 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
645 skb_shinfo(skb)->gso_size = __le16_to_cpu(cqe->len_on_first_bd) -
649 static int qede_fill_frag_skb(struct qede_dev *edev,
650 struct qede_rx_queue *rxq,
651 u8 tpa_agg_index, u16 len_on_bd)
653 struct sw_rx_data *current_bd = &rxq->sw_rx_ring[rxq->sw_rx_cons &
655 struct qede_agg_info *tpa_info = &rxq->tpa_info[tpa_agg_index];
656 struct sk_buff *skb = tpa_info->skb;
658 if (unlikely(tpa_info->state != QEDE_AGG_STATE_START))
661 /* Add one frag and update the appropriate fields in the skb */
662 skb_fill_page_desc(skb, tpa_info->frag_id++,
664 current_bd->page_offset + rxq->rx_headroom,
667 if (unlikely(qede_realloc_rx_buffer(rxq, current_bd))) {
668 /* Incr page ref count to reuse on allocation failure
669 * so that it doesn't get freed while freeing SKB.
671 page_ref_inc(current_bd->data);
675 qede_rx_bd_ring_consume(rxq);
677 skb->data_len += len_on_bd;
678 skb->truesize += rxq->rx_buf_seg_size;
679 skb->len += len_on_bd;
684 tpa_info->state = QEDE_AGG_STATE_ERROR;
685 qede_recycle_rx_bd_ring(rxq, 1);
690 static bool qede_tunn_exist(u16 flag)
692 return !!(flag & (PARSING_AND_ERR_FLAGS_TUNNELEXIST_MASK <<
693 PARSING_AND_ERR_FLAGS_TUNNELEXIST_SHIFT));
696 static u8 qede_check_tunn_csum(u16 flag)
701 if (flag & (PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMWASCALCULATED_MASK <<
702 PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMWASCALCULATED_SHIFT))
703 csum_flag |= PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMERROR_MASK <<
704 PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMERROR_SHIFT;
706 if (flag & (PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_MASK <<
707 PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_SHIFT)) {
708 csum_flag |= PARSING_AND_ERR_FLAGS_L4CHKSMERROR_MASK <<
709 PARSING_AND_ERR_FLAGS_L4CHKSMERROR_SHIFT;
710 tcsum = QEDE_TUNN_CSUM_UNNECESSARY;
713 csum_flag |= PARSING_AND_ERR_FLAGS_TUNNELIPHDRERROR_MASK <<
714 PARSING_AND_ERR_FLAGS_TUNNELIPHDRERROR_SHIFT |
715 PARSING_AND_ERR_FLAGS_IPHDRERROR_MASK <<
716 PARSING_AND_ERR_FLAGS_IPHDRERROR_SHIFT;
718 if (csum_flag & flag)
719 return QEDE_CSUM_ERROR;
721 return QEDE_CSUM_UNNECESSARY | tcsum;
724 static inline struct sk_buff *
725 qede_build_skb(struct qede_rx_queue *rxq,
726 struct sw_rx_data *bd, u16 len, u16 pad)
731 buf = page_address(bd->data) + bd->page_offset;
732 skb = build_skb(buf, rxq->rx_buf_seg_size);
737 skb_reserve(skb, pad);
743 static struct sk_buff *
744 qede_tpa_rx_build_skb(struct qede_dev *edev,
745 struct qede_rx_queue *rxq,
746 struct sw_rx_data *bd, u16 len, u16 pad,
751 skb = qede_build_skb(rxq, bd, len, pad);
752 bd->page_offset += rxq->rx_buf_seg_size;
754 if (bd->page_offset == PAGE_SIZE) {
755 if (unlikely(qede_alloc_rx_buffer(rxq, true))) {
757 "Failed to allocate RX buffer for tpa start\n");
758 bd->page_offset -= rxq->rx_buf_seg_size;
759 page_ref_inc(bd->data);
760 dev_kfree_skb_any(skb);
764 page_ref_inc(bd->data);
765 qede_reuse_page(rxq, bd);
768 /* We've consumed the first BD and prepared an SKB */
769 qede_rx_bd_ring_consume(rxq);
774 static struct sk_buff *
775 qede_rx_build_skb(struct qede_dev *edev,
776 struct qede_rx_queue *rxq,
777 struct sw_rx_data *bd, u16 len, u16 pad)
779 struct sk_buff *skb = NULL;
781 /* For smaller frames still need to allocate skb, memcpy
782 * data and benefit in reusing the page segment instead of
785 if ((len + pad <= edev->rx_copybreak)) {
786 unsigned int offset = bd->page_offset + pad;
788 skb = netdev_alloc_skb(edev->ndev, QEDE_RX_HDR_SIZE);
792 skb_reserve(skb, pad);
793 memcpy(skb_put(skb, len),
794 page_address(bd->data) + offset, len);
795 qede_reuse_page(rxq, bd);
799 skb = qede_build_skb(rxq, bd, len, pad);
801 if (unlikely(qede_realloc_rx_buffer(rxq, bd))) {
802 /* Incr page ref count to reuse on allocation failure so
803 * that it doesn't get freed while freeing SKB [as its
804 * already mapped there].
806 page_ref_inc(bd->data);
807 dev_kfree_skb_any(skb);
811 /* We've consumed the first BD and prepared an SKB */
812 qede_rx_bd_ring_consume(rxq);
817 static void qede_tpa_start(struct qede_dev *edev,
818 struct qede_rx_queue *rxq,
819 struct eth_fast_path_rx_tpa_start_cqe *cqe)
821 struct qede_agg_info *tpa_info = &rxq->tpa_info[cqe->tpa_agg_index];
822 struct sw_rx_data *sw_rx_data_cons;
825 sw_rx_data_cons = &rxq->sw_rx_ring[rxq->sw_rx_cons & NUM_RX_BDS_MAX];
826 pad = cqe->placement_offset + rxq->rx_headroom;
828 tpa_info->skb = qede_tpa_rx_build_skb(edev, rxq, sw_rx_data_cons,
829 le16_to_cpu(cqe->len_on_first_bd),
831 tpa_info->buffer.page_offset = sw_rx_data_cons->page_offset;
832 tpa_info->buffer.mapping = sw_rx_data_cons->mapping;
834 if (unlikely(!tpa_info->skb)) {
835 DP_NOTICE(edev, "Failed to allocate SKB for gro\n");
837 /* Consume from ring but do not produce since
838 * this might be used by FW still, it will be re-used
841 tpa_info->tpa_start_fail = true;
842 qede_rx_bd_ring_consume(rxq);
843 tpa_info->state = QEDE_AGG_STATE_ERROR;
847 tpa_info->frag_id = 0;
848 tpa_info->state = QEDE_AGG_STATE_START;
850 if ((le16_to_cpu(cqe->pars_flags.flags) >>
851 PARSING_AND_ERR_FLAGS_TAG8021QEXIST_SHIFT) &
852 PARSING_AND_ERR_FLAGS_TAG8021QEXIST_MASK)
853 tpa_info->vlan_tag = le16_to_cpu(cqe->vlan_tag);
855 tpa_info->vlan_tag = 0;
857 qede_get_rxhash(tpa_info->skb, cqe->bitfields, cqe->rss_hash);
859 /* This is needed in order to enable forwarding support */
860 qede_set_gro_params(edev, tpa_info->skb, cqe);
862 cons_buf: /* We still need to handle bd_len_list to consume buffers */
863 if (likely(cqe->ext_bd_len_list[0]))
864 qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index,
865 le16_to_cpu(cqe->ext_bd_len_list[0]));
867 if (unlikely(cqe->ext_bd_len_list[1])) {
869 "Unlikely - got a TPA aggregation with more than one ext_bd_len_list entry in the TPA start\n");
870 tpa_info->state = QEDE_AGG_STATE_ERROR;
875 static void qede_gro_ip_csum(struct sk_buff *skb)
877 const struct iphdr *iph = ip_hdr(skb);
880 skb_set_transport_header(skb, sizeof(struct iphdr));
883 th->check = ~tcp_v4_check(skb->len - skb_transport_offset(skb),
884 iph->saddr, iph->daddr, 0);
886 tcp_gro_complete(skb);
889 static void qede_gro_ipv6_csum(struct sk_buff *skb)
891 struct ipv6hdr *iph = ipv6_hdr(skb);
894 skb_set_transport_header(skb, sizeof(struct ipv6hdr));
897 th->check = ~tcp_v6_check(skb->len - skb_transport_offset(skb),
898 &iph->saddr, &iph->daddr, 0);
899 tcp_gro_complete(skb);
903 static void qede_gro_receive(struct qede_dev *edev,
904 struct qede_fastpath *fp,
908 /* FW can send a single MTU sized packet from gro flow
909 * due to aggregation timeout/last segment etc. which
910 * is not expected to be a gro packet. If a skb has zero
911 * frags then simply push it in the stack as non gso skb.
913 if (unlikely(!skb->data_len)) {
914 skb_shinfo(skb)->gso_type = 0;
915 skb_shinfo(skb)->gso_size = 0;
920 if (skb_shinfo(skb)->gso_size) {
921 skb_reset_network_header(skb);
923 switch (skb->protocol) {
924 case htons(ETH_P_IP):
925 qede_gro_ip_csum(skb);
927 case htons(ETH_P_IPV6):
928 qede_gro_ipv6_csum(skb);
932 "Error: FW GRO supports only IPv4/IPv6, not 0x%04x\n",
933 ntohs(skb->protocol));
939 skb_record_rx_queue(skb, fp->rxq->rxq_id);
940 qede_skb_receive(edev, fp, fp->rxq, skb, vlan_tag);
943 static inline void qede_tpa_cont(struct qede_dev *edev,
944 struct qede_rx_queue *rxq,
945 struct eth_fast_path_rx_tpa_cont_cqe *cqe)
949 for (i = 0; cqe->len_list[i]; i++)
950 qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index,
951 le16_to_cpu(cqe->len_list[i]));
955 "Strange - TPA cont with more than a single len_list entry\n");
958 static int qede_tpa_end(struct qede_dev *edev,
959 struct qede_fastpath *fp,
960 struct eth_fast_path_rx_tpa_end_cqe *cqe)
962 struct qede_rx_queue *rxq = fp->rxq;
963 struct qede_agg_info *tpa_info;
967 tpa_info = &rxq->tpa_info[cqe->tpa_agg_index];
970 if (tpa_info->buffer.page_offset == PAGE_SIZE)
971 dma_unmap_page(rxq->dev, tpa_info->buffer.mapping,
972 PAGE_SIZE, rxq->data_direction);
974 for (i = 0; cqe->len_list[i]; i++)
975 qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index,
976 le16_to_cpu(cqe->len_list[i]));
979 "Strange - TPA emd with more than a single len_list entry\n");
981 if (unlikely(tpa_info->state != QEDE_AGG_STATE_START))
985 if (unlikely(cqe->num_of_bds != tpa_info->frag_id + 1))
987 "Strange - TPA had %02x BDs, but SKB has only %d frags\n",
988 cqe->num_of_bds, tpa_info->frag_id);
989 if (unlikely(skb->len != le16_to_cpu(cqe->total_packet_len)))
991 "Strange - total packet len [cqe] is %4x but SKB has len %04x\n",
992 le16_to_cpu(cqe->total_packet_len), skb->len);
994 /* Finalize the SKB */
995 skb->protocol = eth_type_trans(skb, edev->ndev);
996 skb->ip_summed = CHECKSUM_UNNECESSARY;
998 /* tcp_gro_complete() will copy NAPI_GRO_CB(skb)->count
999 * to skb_shinfo(skb)->gso_segs
1001 NAPI_GRO_CB(skb)->count = le16_to_cpu(cqe->num_of_coalesced_segs);
1003 qede_gro_receive(edev, fp, skb, tpa_info->vlan_tag);
1005 tpa_info->state = QEDE_AGG_STATE_NONE;
1009 tpa_info->state = QEDE_AGG_STATE_NONE;
1011 if (tpa_info->tpa_start_fail) {
1012 qede_reuse_page(rxq, &tpa_info->buffer);
1013 tpa_info->tpa_start_fail = false;
1016 dev_kfree_skb_any(tpa_info->skb);
1017 tpa_info->skb = NULL;
1021 static u8 qede_check_notunn_csum(u16 flag)
1026 if (flag & (PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_MASK <<
1027 PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_SHIFT)) {
1028 csum_flag |= PARSING_AND_ERR_FLAGS_L4CHKSMERROR_MASK <<
1029 PARSING_AND_ERR_FLAGS_L4CHKSMERROR_SHIFT;
1030 csum = QEDE_CSUM_UNNECESSARY;
1033 csum_flag |= PARSING_AND_ERR_FLAGS_IPHDRERROR_MASK <<
1034 PARSING_AND_ERR_FLAGS_IPHDRERROR_SHIFT;
1036 if (csum_flag & flag)
1037 return QEDE_CSUM_ERROR;
1042 static u8 qede_check_csum(u16 flag)
1044 if (!qede_tunn_exist(flag))
1045 return qede_check_notunn_csum(flag);
1047 return qede_check_tunn_csum(flag);
1050 static bool qede_pkt_is_ip_fragmented(struct eth_fast_path_rx_reg_cqe *cqe,
1053 u8 tun_pars_flg = cqe->tunnel_pars_flags.flags;
1055 if ((tun_pars_flg & (ETH_TUNNEL_PARSING_FLAGS_IPV4_FRAGMENT_MASK <<
1056 ETH_TUNNEL_PARSING_FLAGS_IPV4_FRAGMENT_SHIFT)) ||
1057 (flag & (PARSING_AND_ERR_FLAGS_IPV4FRAG_MASK <<
1058 PARSING_AND_ERR_FLAGS_IPV4FRAG_SHIFT)))
1064 /* Return true iff packet is to be passed to stack */
1065 static bool qede_rx_xdp(struct qede_dev *edev,
1066 struct qede_fastpath *fp,
1067 struct qede_rx_queue *rxq,
1068 struct bpf_prog *prog,
1069 struct sw_rx_data *bd,
1070 struct eth_fast_path_rx_reg_cqe *cqe,
1071 u16 *data_offset, u16 *len)
1073 struct xdp_buff xdp;
1074 enum xdp_action act;
1076 xdp.data_hard_start = page_address(bd->data);
1077 xdp.data = xdp.data_hard_start + *data_offset;
1078 xdp_set_data_meta_invalid(&xdp);
1079 xdp.data_end = xdp.data + *len;
1080 xdp.rxq = &rxq->xdp_rxq;
1082 /* Queues always have a full reset currently, so for the time
1083 * being until there's atomic program replace just mark read
1084 * side for map helpers.
1087 act = bpf_prog_run_xdp(prog, &xdp);
1090 /* Recalculate, as XDP might have changed the headers */
1091 *data_offset = xdp.data - xdp.data_hard_start;
1092 *len = xdp.data_end - xdp.data;
1094 if (act == XDP_PASS)
1097 /* Count number of packets not to be passed to stack */
1102 /* We need the replacement buffer before transmit. */
1103 if (qede_alloc_rx_buffer(rxq, true)) {
1104 qede_recycle_rx_bd_ring(rxq, 1);
1105 trace_xdp_exception(edev->ndev, prog, act);
1109 /* Now if there's a transmission problem, we'd still have to
1110 * throw current buffer, as replacement was already allocated.
1112 if (qede_xdp_xmit(edev, fp, bd, *data_offset, *len)) {
1113 dma_unmap_page(rxq->dev, bd->mapping,
1114 PAGE_SIZE, DMA_BIDIRECTIONAL);
1115 __free_page(bd->data);
1116 trace_xdp_exception(edev->ndev, prog, act);
1119 /* Regardless, we've consumed an Rx BD */
1120 qede_rx_bd_ring_consume(rxq);
1124 bpf_warn_invalid_xdp_action(act);
1127 trace_xdp_exception(edev->ndev, prog, act);
1130 qede_recycle_rx_bd_ring(rxq, cqe->bd_num);
1136 static int qede_rx_build_jumbo(struct qede_dev *edev,
1137 struct qede_rx_queue *rxq,
1138 struct sk_buff *skb,
1139 struct eth_fast_path_rx_reg_cqe *cqe,
1142 u16 pkt_len = le16_to_cpu(cqe->pkt_len);
1143 struct sw_rx_data *bd;
1147 pkt_len -= first_bd_len;
1149 /* We've already used one BD for the SKB. Now take care of the rest */
1150 for (num_frags = cqe->bd_num - 1; num_frags > 0; num_frags--) {
1151 u16 cur_size = pkt_len > rxq->rx_buf_size ? rxq->rx_buf_size :
1154 if (unlikely(!cur_size)) {
1156 "Still got %d BDs for mapping jumbo, but length became 0\n",
1161 /* We need a replacement buffer for each BD */
1162 if (unlikely(qede_alloc_rx_buffer(rxq, true)))
1165 /* Now that we've allocated the replacement buffer,
1166 * we can safely consume the next BD and map it to the SKB.
1168 bd_cons_idx = rxq->sw_rx_cons & NUM_RX_BDS_MAX;
1169 bd = &rxq->sw_rx_ring[bd_cons_idx];
1170 qede_rx_bd_ring_consume(rxq);
1172 dma_unmap_page(rxq->dev, bd->mapping,
1173 PAGE_SIZE, DMA_FROM_DEVICE);
1175 skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags++,
1176 bd->data, rxq->rx_headroom, cur_size);
1178 skb->truesize += PAGE_SIZE;
1179 skb->data_len += cur_size;
1180 skb->len += cur_size;
1181 pkt_len -= cur_size;
1184 if (unlikely(pkt_len))
1186 "Mapped all BDs of jumbo, but still have %d bytes\n",
1193 static int qede_rx_process_tpa_cqe(struct qede_dev *edev,
1194 struct qede_fastpath *fp,
1195 struct qede_rx_queue *rxq,
1196 union eth_rx_cqe *cqe,
1197 enum eth_rx_cqe_type type)
1200 case ETH_RX_CQE_TYPE_TPA_START:
1201 qede_tpa_start(edev, rxq, &cqe->fast_path_tpa_start);
1203 case ETH_RX_CQE_TYPE_TPA_CONT:
1204 qede_tpa_cont(edev, rxq, &cqe->fast_path_tpa_cont);
1206 case ETH_RX_CQE_TYPE_TPA_END:
1207 return qede_tpa_end(edev, fp, &cqe->fast_path_tpa_end);
1213 static int qede_rx_process_cqe(struct qede_dev *edev,
1214 struct qede_fastpath *fp,
1215 struct qede_rx_queue *rxq)
1217 struct bpf_prog *xdp_prog = READ_ONCE(rxq->xdp_prog);
1218 struct eth_fast_path_rx_reg_cqe *fp_cqe;
1219 u16 len, pad, bd_cons_idx, parse_flag;
1220 enum eth_rx_cqe_type cqe_type;
1221 union eth_rx_cqe *cqe;
1222 struct sw_rx_data *bd;
1223 struct sk_buff *skb;
1227 /* Get the CQE from the completion ring */
1228 cqe = (union eth_rx_cqe *)qed_chain_consume(&rxq->rx_comp_ring);
1229 cqe_type = cqe->fast_path_regular.type;
1231 /* Process an unlikely slowpath event */
1232 if (unlikely(cqe_type == ETH_RX_CQE_TYPE_SLOW_PATH)) {
1233 struct eth_slow_path_rx_cqe *sp_cqe;
1235 sp_cqe = (struct eth_slow_path_rx_cqe *)cqe;
1236 edev->ops->eth_cqe_completion(edev->cdev, fp->id, sp_cqe);
1240 /* Handle TPA cqes */
1241 if (cqe_type != ETH_RX_CQE_TYPE_REGULAR)
1242 return qede_rx_process_tpa_cqe(edev, fp, rxq, cqe, cqe_type);
1244 /* Get the data from the SW ring; Consume it only after it's evident
1245 * we wouldn't recycle it.
1247 bd_cons_idx = rxq->sw_rx_cons & NUM_RX_BDS_MAX;
1248 bd = &rxq->sw_rx_ring[bd_cons_idx];
1250 fp_cqe = &cqe->fast_path_regular;
1251 len = le16_to_cpu(fp_cqe->len_on_first_bd);
1252 pad = fp_cqe->placement_offset + rxq->rx_headroom;
1254 /* Run eBPF program if one is attached */
1256 if (!qede_rx_xdp(edev, fp, rxq, xdp_prog, bd, fp_cqe,
1260 /* If this is an error packet then drop it */
1261 flags = cqe->fast_path_regular.pars_flags.flags;
1262 parse_flag = le16_to_cpu(flags);
1264 csum_flag = qede_check_csum(parse_flag);
1265 if (unlikely(csum_flag == QEDE_CSUM_ERROR)) {
1266 if (qede_pkt_is_ip_fragmented(fp_cqe, parse_flag))
1269 rxq->rx_hw_errors++;
1272 /* Basic validation passed; Need to prepare an SKB. This would also
1273 * guarantee to finally consume the first BD upon success.
1275 skb = qede_rx_build_skb(edev, rxq, bd, len, pad);
1277 rxq->rx_alloc_errors++;
1278 qede_recycle_rx_bd_ring(rxq, fp_cqe->bd_num);
1282 /* In case of Jumbo packet, several PAGE_SIZEd buffers will be pointed
1285 if (fp_cqe->bd_num > 1) {
1286 u16 unmapped_frags = qede_rx_build_jumbo(edev, rxq, skb,
1289 if (unlikely(unmapped_frags > 0)) {
1290 qede_recycle_rx_bd_ring(rxq, unmapped_frags);
1291 dev_kfree_skb_any(skb);
1296 /* The SKB contains all the data. Now prepare meta-magic */
1297 skb->protocol = eth_type_trans(skb, edev->ndev);
1298 qede_get_rxhash(skb, fp_cqe->bitfields, fp_cqe->rss_hash);
1299 qede_set_skb_csum(skb, csum_flag);
1300 skb_record_rx_queue(skb, rxq->rxq_id);
1301 qede_ptp_record_rx_ts(edev, cqe, skb);
1303 /* SKB is prepared - pass it to stack */
1304 qede_skb_receive(edev, fp, rxq, skb, le16_to_cpu(fp_cqe->vlan_tag));
1309 static int qede_rx_int(struct qede_fastpath *fp, int budget)
1311 struct qede_rx_queue *rxq = fp->rxq;
1312 struct qede_dev *edev = fp->edev;
1313 int work_done = 0, rcv_pkts = 0;
1314 u16 hw_comp_cons, sw_comp_cons;
1316 hw_comp_cons = le16_to_cpu(*rxq->hw_cons_ptr);
1317 sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring);
1319 /* Memory barrier to prevent the CPU from doing speculative reads of CQE
1320 * / BD in the while-loop before reading hw_comp_cons. If the CQE is
1321 * read before it is written by FW, then FW writes CQE and SB, and then
1322 * the CPU reads the hw_comp_cons, it will use an old CQE.
1326 /* Loop to complete all indicated BDs */
1327 while ((sw_comp_cons != hw_comp_cons) && (work_done < budget)) {
1328 rcv_pkts += qede_rx_process_cqe(edev, fp, rxq);
1329 qed_chain_recycle_consumed(&rxq->rx_comp_ring);
1330 sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring);
1334 rxq->rcv_pkts += rcv_pkts;
1336 /* Allocate replacement buffers */
1337 while (rxq->num_rx_buffers - rxq->filled_buffers)
1338 if (qede_alloc_rx_buffer(rxq, false))
1341 /* Update producers */
1342 qede_update_rx_prod(edev, rxq);
1347 static bool qede_poll_is_more_work(struct qede_fastpath *fp)
1349 qed_sb_update_sb_idx(fp->sb_info);
1351 /* *_has_*_work() reads the status block, thus we need to ensure that
1352 * status block indices have been actually read (qed_sb_update_sb_idx)
1353 * prior to this check (*_has_*_work) so that we won't write the
1354 * "newer" value of the status block to HW (if there was a DMA right
1355 * after qede_has_rx_work and if there is no rmb, the memory reading
1356 * (qed_sb_update_sb_idx) may be postponed to right before *_ack_sb).
1357 * In this case there will never be another interrupt until there is
1358 * another update of the status block, while there is still unhandled
1363 if (likely(fp->type & QEDE_FASTPATH_RX))
1364 if (qede_has_rx_work(fp->rxq))
1367 if (fp->type & QEDE_FASTPATH_XDP)
1368 if (qede_txq_has_work(fp->xdp_tx))
1371 if (likely(fp->type & QEDE_FASTPATH_TX)) {
1374 for_each_cos_in_txq(fp->edev, cos) {
1375 if (qede_txq_has_work(&fp->txq[cos]))
1383 /*********************
1384 * NDO & API related *
1385 *********************/
1386 int qede_poll(struct napi_struct *napi, int budget)
1388 struct qede_fastpath *fp = container_of(napi, struct qede_fastpath,
1390 struct qede_dev *edev = fp->edev;
1391 int rx_work_done = 0;
1393 if (likely(fp->type & QEDE_FASTPATH_TX)) {
1396 for_each_cos_in_txq(fp->edev, cos) {
1397 if (qede_txq_has_work(&fp->txq[cos]))
1398 qede_tx_int(edev, &fp->txq[cos]);
1402 if ((fp->type & QEDE_FASTPATH_XDP) && qede_txq_has_work(fp->xdp_tx))
1403 qede_xdp_tx_int(edev, fp->xdp_tx);
1405 rx_work_done = (likely(fp->type & QEDE_FASTPATH_RX) &&
1406 qede_has_rx_work(fp->rxq)) ?
1407 qede_rx_int(fp, budget) : 0;
1408 if (rx_work_done < budget) {
1409 if (!qede_poll_is_more_work(fp)) {
1410 napi_complete_done(napi, rx_work_done);
1412 /* Update and reenable interrupts */
1413 qed_sb_ack(fp->sb_info, IGU_INT_ENABLE, 1);
1415 rx_work_done = budget;
1420 u16 xdp_prod = qed_chain_get_prod_idx(&fp->xdp_tx->tx_pbl);
1423 fp->xdp_tx->tx_db.data.bd_prod = cpu_to_le16(xdp_prod);
1424 qede_update_tx_producer(fp->xdp_tx);
1427 return rx_work_done;
1430 irqreturn_t qede_msix_fp_int(int irq, void *fp_cookie)
1432 struct qede_fastpath *fp = fp_cookie;
1434 qed_sb_ack(fp->sb_info, IGU_INT_DISABLE, 0 /*do not update*/);
1436 napi_schedule_irqoff(&fp->napi);
1440 /* Main transmit function */
1441 netdev_tx_t qede_start_xmit(struct sk_buff *skb, struct net_device *ndev)
1443 struct qede_dev *edev = netdev_priv(ndev);
1444 struct netdev_queue *netdev_txq;
1445 struct qede_tx_queue *txq;
1446 struct eth_tx_1st_bd *first_bd;
1447 struct eth_tx_2nd_bd *second_bd = NULL;
1448 struct eth_tx_3rd_bd *third_bd = NULL;
1449 struct eth_tx_bd *tx_data_bd = NULL;
1450 u16 txq_index, val = 0;
1453 int rc, frag_idx = 0, ipv6_ext = 0;
1457 bool data_split = false;
1459 /* Get tx-queue context and netdev index */
1460 txq_index = skb_get_queue_mapping(skb);
1461 WARN_ON(txq_index >= QEDE_TSS_COUNT(edev) * edev->dev_info.num_tc);
1462 txq = QEDE_NDEV_TXQ_ID_TO_TXQ(edev, txq_index);
1463 netdev_txq = netdev_get_tx_queue(ndev, txq_index);
1465 WARN_ON(qed_chain_get_elem_left(&txq->tx_pbl) < (MAX_SKB_FRAGS + 1));
1467 xmit_type = qede_xmit_type(skb, &ipv6_ext);
1469 #if ((MAX_SKB_FRAGS + 2) > ETH_TX_MAX_BDS_PER_NON_LSO_PACKET)
1470 if (qede_pkt_req_lin(skb, xmit_type)) {
1471 if (skb_linearize(skb)) {
1473 "SKB linearization failed - silently dropping this SKB\n");
1474 dev_kfree_skb_any(skb);
1475 return NETDEV_TX_OK;
1480 /* Fill the entry in the SW ring and the BDs in the FW ring */
1481 idx = txq->sw_tx_prod;
1482 txq->sw_tx_ring.skbs[idx].skb = skb;
1483 first_bd = (struct eth_tx_1st_bd *)
1484 qed_chain_produce(&txq->tx_pbl);
1485 memset(first_bd, 0, sizeof(*first_bd));
1486 first_bd->data.bd_flags.bitfields =
1487 1 << ETH_TX_1ST_BD_FLAGS_START_BD_SHIFT;
1489 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP))
1490 qede_ptp_tx_ts(edev, skb);
1492 /* Map skb linear data for DMA and set in the first BD */
1493 mapping = dma_map_single(txq->dev, skb->data,
1494 skb_headlen(skb), DMA_TO_DEVICE);
1495 if (unlikely(dma_mapping_error(txq->dev, mapping))) {
1496 DP_NOTICE(edev, "SKB mapping failed\n");
1497 qede_free_failed_tx_pkt(txq, first_bd, 0, false);
1498 qede_update_tx_producer(txq);
1499 return NETDEV_TX_OK;
1502 BD_SET_UNMAP_ADDR_LEN(first_bd, mapping, skb_headlen(skb));
1504 /* In case there is IPv6 with extension headers or LSO we need 2nd and
1507 if (unlikely((xmit_type & XMIT_LSO) | ipv6_ext)) {
1508 second_bd = (struct eth_tx_2nd_bd *)
1509 qed_chain_produce(&txq->tx_pbl);
1510 memset(second_bd, 0, sizeof(*second_bd));
1513 third_bd = (struct eth_tx_3rd_bd *)
1514 qed_chain_produce(&txq->tx_pbl);
1515 memset(third_bd, 0, sizeof(*third_bd));
1518 /* We need to fill in additional data in second_bd... */
1519 tx_data_bd = (struct eth_tx_bd *)second_bd;
1522 if (skb_vlan_tag_present(skb)) {
1523 first_bd->data.vlan = cpu_to_le16(skb_vlan_tag_get(skb));
1524 first_bd->data.bd_flags.bitfields |=
1525 1 << ETH_TX_1ST_BD_FLAGS_VLAN_INSERTION_SHIFT;
1528 /* Fill the parsing flags & params according to the requested offload */
1529 if (xmit_type & XMIT_L4_CSUM) {
1530 /* We don't re-calculate IP checksum as it is already done by
1533 first_bd->data.bd_flags.bitfields |=
1534 1 << ETH_TX_1ST_BD_FLAGS_L4_CSUM_SHIFT;
1536 if (xmit_type & XMIT_ENC) {
1537 first_bd->data.bd_flags.bitfields |=
1538 1 << ETH_TX_1ST_BD_FLAGS_IP_CSUM_SHIFT;
1540 val |= (1 << ETH_TX_DATA_1ST_BD_TUNN_FLAG_SHIFT);
1543 /* Legacy FW had flipped behavior in regard to this bit -
1544 * I.e., needed to set to prevent FW from touching encapsulated
1545 * packets when it didn't need to.
1547 if (unlikely(txq->is_legacy))
1548 val ^= (1 << ETH_TX_DATA_1ST_BD_TUNN_FLAG_SHIFT);
1550 /* If the packet is IPv6 with extension header, indicate that
1551 * to FW and pass few params, since the device cracker doesn't
1552 * support parsing IPv6 with extension header/s.
1554 if (unlikely(ipv6_ext))
1555 qede_set_params_for_ipv6_ext(skb, second_bd, third_bd);
1558 if (xmit_type & XMIT_LSO) {
1559 first_bd->data.bd_flags.bitfields |=
1560 (1 << ETH_TX_1ST_BD_FLAGS_LSO_SHIFT);
1561 third_bd->data.lso_mss =
1562 cpu_to_le16(skb_shinfo(skb)->gso_size);
1564 if (unlikely(xmit_type & XMIT_ENC)) {
1565 first_bd->data.bd_flags.bitfields |=
1566 1 << ETH_TX_1ST_BD_FLAGS_TUNN_IP_CSUM_SHIFT;
1568 if (xmit_type & XMIT_ENC_GSO_L4_CSUM) {
1569 u8 tmp = ETH_TX_1ST_BD_FLAGS_TUNN_L4_CSUM_SHIFT;
1571 first_bd->data.bd_flags.bitfields |= 1 << tmp;
1573 hlen = qede_get_skb_hlen(skb, true);
1575 first_bd->data.bd_flags.bitfields |=
1576 1 << ETH_TX_1ST_BD_FLAGS_IP_CSUM_SHIFT;
1577 hlen = qede_get_skb_hlen(skb, false);
1580 /* @@@TBD - if will not be removed need to check */
1581 third_bd->data.bitfields |=
1582 cpu_to_le16(1 << ETH_TX_DATA_3RD_BD_HDR_NBD_SHIFT);
1584 /* Make life easier for FW guys who can't deal with header and
1585 * data on same BD. If we need to split, use the second bd...
1587 if (unlikely(skb_headlen(skb) > hlen)) {
1588 DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED,
1589 "TSO split header size is %d (%x:%x)\n",
1590 first_bd->nbytes, first_bd->addr.hi,
1593 mapping = HILO_U64(le32_to_cpu(first_bd->addr.hi),
1594 le32_to_cpu(first_bd->addr.lo)) +
1597 BD_SET_UNMAP_ADDR_LEN(tx_data_bd, mapping,
1598 le16_to_cpu(first_bd->nbytes) -
1601 /* this marks the BD as one that has no
1602 * individual mapping
1604 txq->sw_tx_ring.skbs[idx].flags |= QEDE_TSO_SPLIT_BD;
1606 first_bd->nbytes = cpu_to_le16(hlen);
1608 tx_data_bd = (struct eth_tx_bd *)third_bd;
1612 if (unlikely(skb->len > ETH_TX_MAX_NON_LSO_PKT_LEN)) {
1613 DP_ERR(edev, "Unexpected non LSO skb length = 0x%x\n", skb->len);
1614 qede_free_failed_tx_pkt(txq, first_bd, 0, false);
1615 qede_update_tx_producer(txq);
1616 return NETDEV_TX_OK;
1619 val |= ((skb->len & ETH_TX_DATA_1ST_BD_PKT_LEN_MASK) <<
1620 ETH_TX_DATA_1ST_BD_PKT_LEN_SHIFT);
1623 first_bd->data.bitfields = cpu_to_le16(val);
1625 /* Handle fragmented skb */
1626 /* special handle for frags inside 2nd and 3rd bds.. */
1627 while (tx_data_bd && frag_idx < skb_shinfo(skb)->nr_frags) {
1628 rc = map_frag_to_bd(txq,
1629 &skb_shinfo(skb)->frags[frag_idx],
1632 qede_free_failed_tx_pkt(txq, first_bd, nbd, data_split);
1633 qede_update_tx_producer(txq);
1634 return NETDEV_TX_OK;
1637 if (tx_data_bd == (struct eth_tx_bd *)second_bd)
1638 tx_data_bd = (struct eth_tx_bd *)third_bd;
1645 /* map last frags into 4th, 5th .... */
1646 for (; frag_idx < skb_shinfo(skb)->nr_frags; frag_idx++, nbd++) {
1647 tx_data_bd = (struct eth_tx_bd *)
1648 qed_chain_produce(&txq->tx_pbl);
1650 memset(tx_data_bd, 0, sizeof(*tx_data_bd));
1652 rc = map_frag_to_bd(txq,
1653 &skb_shinfo(skb)->frags[frag_idx],
1656 qede_free_failed_tx_pkt(txq, first_bd, nbd, data_split);
1657 qede_update_tx_producer(txq);
1658 return NETDEV_TX_OK;
1662 /* update the first BD with the actual num BDs */
1663 first_bd->data.nbds = nbd;
1665 netdev_tx_sent_queue(netdev_txq, skb->len);
1667 skb_tx_timestamp(skb);
1669 /* Advance packet producer only before sending the packet since mapping
1670 * of pages may fail.
1672 txq->sw_tx_prod = (txq->sw_tx_prod + 1) % txq->num_tx_buffers;
1674 /* 'next page' entries are counted in the producer value */
1675 txq->tx_db.data.bd_prod =
1676 cpu_to_le16(qed_chain_get_prod_idx(&txq->tx_pbl));
1678 if (!skb->xmit_more || netif_xmit_stopped(netdev_txq))
1679 qede_update_tx_producer(txq);
1681 if (unlikely(qed_chain_get_elem_left(&txq->tx_pbl)
1682 < (MAX_SKB_FRAGS + 1))) {
1684 qede_update_tx_producer(txq);
1686 netif_tx_stop_queue(netdev_txq);
1688 DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED,
1689 "Stop queue was called\n");
1690 /* paired memory barrier is in qede_tx_int(), we have to keep
1691 * ordering of set_bit() in netif_tx_stop_queue() and read of
1696 if ((qed_chain_get_elem_left(&txq->tx_pbl) >=
1697 (MAX_SKB_FRAGS + 1)) &&
1698 (edev->state == QEDE_STATE_OPEN)) {
1699 netif_tx_wake_queue(netdev_txq);
1700 DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED,
1701 "Wake queue was called\n");
1705 return NETDEV_TX_OK;
1708 u16 qede_select_queue(struct net_device *dev, struct sk_buff *skb,
1709 struct net_device *sb_dev,
1710 select_queue_fallback_t fallback)
1712 struct qede_dev *edev = netdev_priv(dev);
1715 total_txq = QEDE_TSS_COUNT(edev) * edev->dev_info.num_tc;
1717 return QEDE_TSS_COUNT(edev) ?
1718 fallback(dev, skb, NULL) % total_txq : 0;
1721 /* 8B udp header + 8B base tunnel header + 32B option length */
1722 #define QEDE_MAX_TUN_HDR_LEN 48
1724 netdev_features_t qede_features_check(struct sk_buff *skb,
1725 struct net_device *dev,
1726 netdev_features_t features)
1728 if (skb->encapsulation) {
1731 switch (vlan_get_protocol(skb)) {
1732 case htons(ETH_P_IP):
1733 l4_proto = ip_hdr(skb)->protocol;
1735 case htons(ETH_P_IPV6):
1736 l4_proto = ipv6_hdr(skb)->nexthdr;
1742 /* Disable offloads for geneve tunnels, as HW can't parse
1743 * the geneve header which has option length greater than 32b
1744 * and disable offloads for the ports which are not offloaded.
1746 if (l4_proto == IPPROTO_UDP) {
1747 struct qede_dev *edev = netdev_priv(dev);
1748 u16 hdrlen, vxln_port, gnv_port;
1750 hdrlen = QEDE_MAX_TUN_HDR_LEN;
1751 vxln_port = edev->vxlan_dst_port;
1752 gnv_port = edev->geneve_dst_port;
1754 if ((skb_inner_mac_header(skb) -
1755 skb_transport_header(skb)) > hdrlen ||
1756 (ntohs(udp_hdr(skb)->dest) != vxln_port &&
1757 ntohs(udp_hdr(skb)->dest) != gnv_port))
1758 return features & ~(NETIF_F_CSUM_MASK |
1760 } else if (l4_proto == IPPROTO_IPIP) {
1761 /* IPIP tunnels are unknown to the device or at least unsupported natively,
1762 * offloads for them can't be done trivially, so disable them for such skb.
1764 return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);