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));
93 rxq->filled_buffers++;
98 /* Unmap the data and free skb */
99 int qede_free_tx_pkt(struct qede_dev *edev, struct qede_tx_queue *txq, int *len)
101 u16 idx = txq->sw_tx_cons & NUM_TX_BDS_MAX;
102 struct sk_buff *skb = txq->sw_tx_ring.skbs[idx].skb;
103 struct eth_tx_1st_bd *first_bd;
104 struct eth_tx_bd *tx_data_bd;
105 int bds_consumed = 0;
107 bool data_split = txq->sw_tx_ring.skbs[idx].flags & QEDE_TSO_SPLIT_BD;
108 int i, split_bd_len = 0;
110 if (unlikely(!skb)) {
112 "skb is null for txq idx=%d txq->sw_tx_cons=%d txq->sw_tx_prod=%d\n",
113 idx, txq->sw_tx_cons, txq->sw_tx_prod);
119 first_bd = (struct eth_tx_1st_bd *)qed_chain_consume(&txq->tx_pbl);
123 nbds = first_bd->data.nbds;
126 struct eth_tx_bd *split = (struct eth_tx_bd *)
127 qed_chain_consume(&txq->tx_pbl);
128 split_bd_len = BD_UNMAP_LEN(split);
131 dma_unmap_single(&edev->pdev->dev, BD_UNMAP_ADDR(first_bd),
132 BD_UNMAP_LEN(first_bd) + split_bd_len, DMA_TO_DEVICE);
134 /* Unmap the data of the skb frags */
135 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++, bds_consumed++) {
136 tx_data_bd = (struct eth_tx_bd *)
137 qed_chain_consume(&txq->tx_pbl);
138 dma_unmap_page(&edev->pdev->dev, BD_UNMAP_ADDR(tx_data_bd),
139 BD_UNMAP_LEN(tx_data_bd), DMA_TO_DEVICE);
142 while (bds_consumed++ < nbds)
143 qed_chain_consume(&txq->tx_pbl);
146 dev_kfree_skb_any(skb);
147 txq->sw_tx_ring.skbs[idx].skb = NULL;
148 txq->sw_tx_ring.skbs[idx].flags = 0;
153 /* Unmap the data and free skb when mapping failed during start_xmit */
154 static void qede_free_failed_tx_pkt(struct qede_tx_queue *txq,
155 struct eth_tx_1st_bd *first_bd,
156 int nbd, bool data_split)
158 u16 idx = txq->sw_tx_prod & NUM_TX_BDS_MAX;
159 struct sk_buff *skb = txq->sw_tx_ring.skbs[idx].skb;
160 struct eth_tx_bd *tx_data_bd;
161 int i, split_bd_len = 0;
163 /* Return prod to its position before this skb was handled */
164 qed_chain_set_prod(&txq->tx_pbl,
165 le16_to_cpu(txq->tx_db.data.bd_prod), first_bd);
167 first_bd = (struct eth_tx_1st_bd *)qed_chain_produce(&txq->tx_pbl);
170 struct eth_tx_bd *split = (struct eth_tx_bd *)
171 qed_chain_produce(&txq->tx_pbl);
172 split_bd_len = BD_UNMAP_LEN(split);
176 dma_unmap_single(txq->dev, BD_UNMAP_ADDR(first_bd),
177 BD_UNMAP_LEN(first_bd) + split_bd_len, DMA_TO_DEVICE);
179 /* Unmap the data of the skb frags */
180 for (i = 0; i < nbd; i++) {
181 tx_data_bd = (struct eth_tx_bd *)
182 qed_chain_produce(&txq->tx_pbl);
183 if (tx_data_bd->nbytes)
184 dma_unmap_page(txq->dev,
185 BD_UNMAP_ADDR(tx_data_bd),
186 BD_UNMAP_LEN(tx_data_bd), DMA_TO_DEVICE);
189 /* Return again prod to its position before this skb was handled */
190 qed_chain_set_prod(&txq->tx_pbl,
191 le16_to_cpu(txq->tx_db.data.bd_prod), first_bd);
194 dev_kfree_skb_any(skb);
195 txq->sw_tx_ring.skbs[idx].skb = NULL;
196 txq->sw_tx_ring.skbs[idx].flags = 0;
199 static u32 qede_xmit_type(struct sk_buff *skb, int *ipv6_ext)
201 u32 rc = XMIT_L4_CSUM;
204 if (skb->ip_summed != CHECKSUM_PARTIAL)
207 l3_proto = vlan_get_protocol(skb);
208 if (l3_proto == htons(ETH_P_IPV6) &&
209 (ipv6_hdr(skb)->nexthdr == NEXTHDR_IPV6))
212 if (skb->encapsulation) {
214 if (skb_is_gso(skb)) {
215 unsigned short gso_type = skb_shinfo(skb)->gso_type;
217 if ((gso_type & SKB_GSO_UDP_TUNNEL_CSUM) ||
218 (gso_type & SKB_GSO_GRE_CSUM))
219 rc |= XMIT_ENC_GSO_L4_CSUM;
232 static void qede_set_params_for_ipv6_ext(struct sk_buff *skb,
233 struct eth_tx_2nd_bd *second_bd,
234 struct eth_tx_3rd_bd *third_bd)
237 u16 bd2_bits1 = 0, bd2_bits2 = 0;
239 bd2_bits1 |= (1 << ETH_TX_DATA_2ND_BD_IPV6_EXT_SHIFT);
241 bd2_bits2 |= ((((u8 *)skb_transport_header(skb) - skb->data) >> 1) &
242 ETH_TX_DATA_2ND_BD_L4_HDR_START_OFFSET_W_MASK)
243 << ETH_TX_DATA_2ND_BD_L4_HDR_START_OFFSET_W_SHIFT;
245 bd2_bits1 |= (ETH_L4_PSEUDO_CSUM_CORRECT_LENGTH <<
246 ETH_TX_DATA_2ND_BD_L4_PSEUDO_CSUM_MODE_SHIFT);
248 if (vlan_get_protocol(skb) == htons(ETH_P_IPV6))
249 l4_proto = ipv6_hdr(skb)->nexthdr;
251 l4_proto = ip_hdr(skb)->protocol;
253 if (l4_proto == IPPROTO_UDP)
254 bd2_bits1 |= 1 << ETH_TX_DATA_2ND_BD_L4_UDP_SHIFT;
257 third_bd->data.bitfields |=
258 cpu_to_le16(((tcp_hdrlen(skb) / 4) &
259 ETH_TX_DATA_3RD_BD_TCP_HDR_LEN_DW_MASK) <<
260 ETH_TX_DATA_3RD_BD_TCP_HDR_LEN_DW_SHIFT);
262 second_bd->data.bitfields1 = cpu_to_le16(bd2_bits1);
263 second_bd->data.bitfields2 = cpu_to_le16(bd2_bits2);
266 static int map_frag_to_bd(struct qede_tx_queue *txq,
267 skb_frag_t *frag, struct eth_tx_bd *bd)
271 /* Map skb non-linear frag data for DMA */
272 mapping = skb_frag_dma_map(txq->dev, frag, 0,
273 skb_frag_size(frag), DMA_TO_DEVICE);
274 if (unlikely(dma_mapping_error(txq->dev, mapping)))
277 /* Setup the data pointer of the frag data */
278 BD_SET_UNMAP_ADDR_LEN(bd, mapping, skb_frag_size(frag));
283 static u16 qede_get_skb_hlen(struct sk_buff *skb, bool is_encap_pkt)
286 return (skb_inner_transport_header(skb) +
287 inner_tcp_hdrlen(skb) - skb->data);
289 return (skb_transport_header(skb) +
290 tcp_hdrlen(skb) - skb->data);
293 /* +2 for 1st BD for headers and 2nd BD for headlen (if required) */
294 #if ((MAX_SKB_FRAGS + 2) > ETH_TX_MAX_BDS_PER_NON_LSO_PACKET)
295 static bool qede_pkt_req_lin(struct sk_buff *skb, u8 xmit_type)
297 int allowed_frags = ETH_TX_MAX_BDS_PER_NON_LSO_PACKET - 1;
299 if (xmit_type & XMIT_LSO) {
302 hlen = qede_get_skb_hlen(skb, xmit_type & XMIT_ENC);
304 /* linear payload would require its own BD */
305 if (skb_headlen(skb) > hlen)
309 return (skb_shinfo(skb)->nr_frags > allowed_frags);
313 static inline void qede_update_tx_producer(struct qede_tx_queue *txq)
315 /* wmb makes sure that the BDs data is updated before updating the
316 * producer, otherwise FW may read old data from the BDs.
320 writel(txq->tx_db.raw, txq->doorbell_addr);
322 /* mmiowb is needed to synchronize doorbell writes from more than one
323 * processor. It guarantees that the write arrives to the device before
324 * the queue lock is released and another start_xmit is called (possibly
325 * on another CPU). Without this barrier, the next doorbell can bypass
326 * this doorbell. This is applicable to IA64/Altix systems.
331 static int qede_xdp_xmit(struct qede_dev *edev, struct qede_fastpath *fp,
332 struct sw_rx_data *metadata, u16 padding, u16 length)
334 struct qede_tx_queue *txq = fp->xdp_tx;
335 u16 idx = txq->sw_tx_prod & NUM_TX_BDS_MAX;
336 struct eth_tx_1st_bd *first_bd;
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);
348 first_bd->data.bitfields |=
349 (length & ETH_TX_DATA_1ST_BD_PKT_LEN_MASK) <<
350 ETH_TX_DATA_1ST_BD_PKT_LEN_SHIFT;
351 first_bd->data.nbds = 1;
353 /* We can safely ignore the offset, as it's 0 for XDP */
354 BD_SET_UNMAP_ADDR_LEN(first_bd, metadata->mapping + padding, length);
356 /* Synchronize the buffer back to device, as program [probably]
359 dma_sync_single_for_device(&edev->pdev->dev,
360 metadata->mapping + padding,
361 length, PCI_DMA_TODEVICE);
363 txq->sw_tx_ring.pages[idx] = metadata->data;
366 /* Mark the fastpath for future XDP doorbell */
372 int qede_txq_has_work(struct qede_tx_queue *txq)
376 /* Tell compiler that consumer and producer can change */
378 hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr);
379 if (qed_chain_get_cons_idx(&txq->tx_pbl) == hw_bd_cons + 1)
382 return hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl);
385 static void qede_xdp_tx_int(struct qede_dev *edev, struct qede_tx_queue *txq)
387 struct eth_tx_1st_bd *bd;
390 hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr);
393 while (hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl)) {
394 bd = (struct eth_tx_1st_bd *)qed_chain_consume(&txq->tx_pbl);
396 dma_unmap_single(&edev->pdev->dev, BD_UNMAP_ADDR(bd),
397 PAGE_SIZE, DMA_BIDIRECTIONAL);
398 __free_page(txq->sw_tx_ring.pages[txq->sw_tx_cons &
406 static int qede_tx_int(struct qede_dev *edev, struct qede_tx_queue *txq)
408 struct netdev_queue *netdev_txq;
410 unsigned int pkts_compl = 0, bytes_compl = 0;
413 netdev_txq = netdev_get_tx_queue(edev->ndev, txq->index);
415 hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr);
418 while (hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl)) {
421 rc = qede_free_tx_pkt(edev, txq, &len);
423 DP_NOTICE(edev, "hw_bd_cons = %d, chain_cons=%d\n",
425 qed_chain_get_cons_idx(&txq->tx_pbl));
435 netdev_tx_completed_queue(netdev_txq, pkts_compl, bytes_compl);
437 /* Need to make the tx_bd_cons update visible to start_xmit()
438 * before checking for netif_tx_queue_stopped(). Without the
439 * memory barrier, there is a small possibility that
440 * start_xmit() will miss it and cause the queue to be stopped
442 * On the other hand we need an rmb() here to ensure the proper
443 * ordering of bit testing in the following
444 * netif_tx_queue_stopped(txq) call.
448 if (unlikely(netif_tx_queue_stopped(netdev_txq))) {
449 /* Taking tx_lock is needed to prevent reenabling the queue
450 * while it's empty. This could have happen if rx_action() gets
451 * suspended in qede_tx_int() after the condition before
452 * netif_tx_wake_queue(), while tx_action (qede_start_xmit()):
454 * stops the queue->sees fresh tx_bd_cons->releases the queue->
455 * sends some packets consuming the whole queue again->
459 __netif_tx_lock(netdev_txq, smp_processor_id());
461 if ((netif_tx_queue_stopped(netdev_txq)) &&
462 (edev->state == QEDE_STATE_OPEN) &&
463 (qed_chain_get_elem_left(&txq->tx_pbl)
464 >= (MAX_SKB_FRAGS + 1))) {
465 netif_tx_wake_queue(netdev_txq);
466 DP_VERBOSE(edev, NETIF_MSG_TX_DONE,
467 "Wake queue was called\n");
470 __netif_tx_unlock(netdev_txq);
476 bool qede_has_rx_work(struct qede_rx_queue *rxq)
478 u16 hw_comp_cons, sw_comp_cons;
480 /* Tell compiler that status block fields can change */
483 hw_comp_cons = le16_to_cpu(*rxq->hw_cons_ptr);
484 sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring);
486 return hw_comp_cons != sw_comp_cons;
489 static inline void qede_rx_bd_ring_consume(struct qede_rx_queue *rxq)
491 qed_chain_consume(&rxq->rx_bd_ring);
495 /* This function reuses the buffer(from an offset) from
496 * consumer index to producer index in the bd ring
498 static inline void qede_reuse_page(struct qede_rx_queue *rxq,
499 struct sw_rx_data *curr_cons)
501 struct eth_rx_bd *rx_bd_prod = qed_chain_produce(&rxq->rx_bd_ring);
502 struct sw_rx_data *curr_prod;
503 dma_addr_t new_mapping;
505 curr_prod = &rxq->sw_rx_ring[rxq->sw_rx_prod & NUM_RX_BDS_MAX];
506 *curr_prod = *curr_cons;
508 new_mapping = curr_prod->mapping + curr_prod->page_offset;
510 rx_bd_prod->addr.hi = cpu_to_le32(upper_32_bits(new_mapping));
511 rx_bd_prod->addr.lo = cpu_to_le32(lower_32_bits(new_mapping));
514 curr_cons->data = NULL;
517 /* In case of allocation failures reuse buffers
518 * from consumer index to produce buffers for firmware
520 void qede_recycle_rx_bd_ring(struct qede_rx_queue *rxq, u8 count)
522 struct sw_rx_data *curr_cons;
524 for (; count > 0; count--) {
525 curr_cons = &rxq->sw_rx_ring[rxq->sw_rx_cons & NUM_RX_BDS_MAX];
526 qede_reuse_page(rxq, curr_cons);
527 qede_rx_bd_ring_consume(rxq);
531 static inline int qede_realloc_rx_buffer(struct qede_rx_queue *rxq,
532 struct sw_rx_data *curr_cons)
534 /* Move to the next segment in the page */
535 curr_cons->page_offset += rxq->rx_buf_seg_size;
537 if (curr_cons->page_offset == PAGE_SIZE) {
538 if (unlikely(qede_alloc_rx_buffer(rxq, true))) {
539 /* Since we failed to allocate new buffer
540 * current buffer can be used again.
542 curr_cons->page_offset -= rxq->rx_buf_seg_size;
547 dma_unmap_page(rxq->dev, curr_cons->mapping,
548 PAGE_SIZE, rxq->data_direction);
550 /* Increment refcount of the page as we don't want
551 * network stack to take the ownership of the page
552 * which can be recycled multiple times by the driver.
554 page_ref_inc(curr_cons->data);
555 qede_reuse_page(rxq, curr_cons);
561 void qede_update_rx_prod(struct qede_dev *edev, struct qede_rx_queue *rxq)
563 u16 bd_prod = qed_chain_get_prod_idx(&rxq->rx_bd_ring);
564 u16 cqe_prod = qed_chain_get_prod_idx(&rxq->rx_comp_ring);
565 struct eth_rx_prod_data rx_prods = {0};
567 /* Update producers */
568 rx_prods.bd_prod = cpu_to_le16(bd_prod);
569 rx_prods.cqe_prod = cpu_to_le16(cqe_prod);
571 /* Make sure that the BD and SGE data is updated before updating the
572 * producers since FW might read the BD/SGE right after the producer
577 internal_ram_wr(rxq->hw_rxq_prod_addr, sizeof(rx_prods),
580 /* mmiowb is needed to synchronize doorbell writes from more than one
581 * processor. It guarantees that the write arrives to the device before
582 * the napi lock is released and another qede_poll is called (possibly
583 * on another CPU). Without this barrier, the next doorbell can bypass
584 * this doorbell. This is applicable to IA64/Altix systems.
589 static void qede_get_rxhash(struct sk_buff *skb, u8 bitfields, __le32 rss_hash)
591 enum pkt_hash_types hash_type = PKT_HASH_TYPE_NONE;
592 enum rss_hash_type htype;
595 htype = GET_FIELD(bitfields, ETH_FAST_PATH_RX_REG_CQE_RSS_HASH_TYPE);
597 hash_type = ((htype == RSS_HASH_TYPE_IPV4) ||
598 (htype == RSS_HASH_TYPE_IPV6)) ?
599 PKT_HASH_TYPE_L3 : PKT_HASH_TYPE_L4;
600 hash = le32_to_cpu(rss_hash);
602 skb_set_hash(skb, hash, hash_type);
605 static void qede_set_skb_csum(struct sk_buff *skb, u8 csum_flag)
607 skb_checksum_none_assert(skb);
609 if (csum_flag & QEDE_CSUM_UNNECESSARY)
610 skb->ip_summed = CHECKSUM_UNNECESSARY;
612 if (csum_flag & QEDE_TUNN_CSUM_UNNECESSARY) {
614 skb->encapsulation = 1;
618 static inline void qede_skb_receive(struct qede_dev *edev,
619 struct qede_fastpath *fp,
620 struct qede_rx_queue *rxq,
621 struct sk_buff *skb, u16 vlan_tag)
624 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tag);
626 napi_gro_receive(&fp->napi, skb);
630 static void qede_set_gro_params(struct qede_dev *edev,
632 struct eth_fast_path_rx_tpa_start_cqe *cqe)
634 u16 parsing_flags = le16_to_cpu(cqe->pars_flags.flags);
636 if (((parsing_flags >> PARSING_AND_ERR_FLAGS_L3TYPE_SHIFT) &
637 PARSING_AND_ERR_FLAGS_L3TYPE_MASK) == 2)
638 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
640 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
642 skb_shinfo(skb)->gso_size = __le16_to_cpu(cqe->len_on_first_bd) -
646 static int qede_fill_frag_skb(struct qede_dev *edev,
647 struct qede_rx_queue *rxq,
648 u8 tpa_agg_index, u16 len_on_bd)
650 struct sw_rx_data *current_bd = &rxq->sw_rx_ring[rxq->sw_rx_cons &
652 struct qede_agg_info *tpa_info = &rxq->tpa_info[tpa_agg_index];
653 struct sk_buff *skb = tpa_info->skb;
655 if (unlikely(tpa_info->state != QEDE_AGG_STATE_START))
658 /* Add one frag and update the appropriate fields in the skb */
659 skb_fill_page_desc(skb, tpa_info->frag_id++,
660 current_bd->data, current_bd->page_offset,
663 if (unlikely(qede_realloc_rx_buffer(rxq, current_bd))) {
664 /* Incr page ref count to reuse on allocation failure
665 * so that it doesn't get freed while freeing SKB.
667 page_ref_inc(current_bd->data);
671 qed_chain_consume(&rxq->rx_bd_ring);
674 skb->data_len += len_on_bd;
675 skb->truesize += rxq->rx_buf_seg_size;
676 skb->len += len_on_bd;
681 tpa_info->state = QEDE_AGG_STATE_ERROR;
682 qede_recycle_rx_bd_ring(rxq, 1);
687 static bool qede_tunn_exist(u16 flag)
689 return !!(flag & (PARSING_AND_ERR_FLAGS_TUNNELEXIST_MASK <<
690 PARSING_AND_ERR_FLAGS_TUNNELEXIST_SHIFT));
693 static u8 qede_check_tunn_csum(u16 flag)
698 if (flag & (PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMWASCALCULATED_MASK <<
699 PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMWASCALCULATED_SHIFT))
700 csum_flag |= PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMERROR_MASK <<
701 PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMERROR_SHIFT;
703 if (flag & (PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_MASK <<
704 PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_SHIFT)) {
705 csum_flag |= PARSING_AND_ERR_FLAGS_L4CHKSMERROR_MASK <<
706 PARSING_AND_ERR_FLAGS_L4CHKSMERROR_SHIFT;
707 tcsum = QEDE_TUNN_CSUM_UNNECESSARY;
710 csum_flag |= PARSING_AND_ERR_FLAGS_TUNNELIPHDRERROR_MASK <<
711 PARSING_AND_ERR_FLAGS_TUNNELIPHDRERROR_SHIFT |
712 PARSING_AND_ERR_FLAGS_IPHDRERROR_MASK <<
713 PARSING_AND_ERR_FLAGS_IPHDRERROR_SHIFT;
715 if (csum_flag & flag)
716 return QEDE_CSUM_ERROR;
718 return QEDE_CSUM_UNNECESSARY | tcsum;
721 static void qede_tpa_start(struct qede_dev *edev,
722 struct qede_rx_queue *rxq,
723 struct eth_fast_path_rx_tpa_start_cqe *cqe)
725 struct qede_agg_info *tpa_info = &rxq->tpa_info[cqe->tpa_agg_index];
726 struct eth_rx_bd *rx_bd_cons = qed_chain_consume(&rxq->rx_bd_ring);
727 struct eth_rx_bd *rx_bd_prod = qed_chain_produce(&rxq->rx_bd_ring);
728 struct sw_rx_data *replace_buf = &tpa_info->buffer;
729 dma_addr_t mapping = tpa_info->buffer_mapping;
730 struct sw_rx_data *sw_rx_data_cons;
731 struct sw_rx_data *sw_rx_data_prod;
733 sw_rx_data_cons = &rxq->sw_rx_ring[rxq->sw_rx_cons & NUM_RX_BDS_MAX];
734 sw_rx_data_prod = &rxq->sw_rx_ring[rxq->sw_rx_prod & NUM_RX_BDS_MAX];
736 /* Use pre-allocated replacement buffer - we can't release the agg.
737 * start until its over and we don't want to risk allocation failing
738 * here, so re-allocate when aggregation will be over.
740 sw_rx_data_prod->mapping = replace_buf->mapping;
742 sw_rx_data_prod->data = replace_buf->data;
743 rx_bd_prod->addr.hi = cpu_to_le32(upper_32_bits(mapping));
744 rx_bd_prod->addr.lo = cpu_to_le32(lower_32_bits(mapping));
745 sw_rx_data_prod->page_offset = replace_buf->page_offset;
749 /* move partial skb from cons to pool (don't unmap yet)
750 * save mapping, incase we drop the packet later on.
752 tpa_info->buffer = *sw_rx_data_cons;
753 mapping = HILO_U64(le32_to_cpu(rx_bd_cons->addr.hi),
754 le32_to_cpu(rx_bd_cons->addr.lo));
756 tpa_info->buffer_mapping = mapping;
759 /* set tpa state to start only if we are able to allocate skb
760 * for this aggregation, otherwise mark as error and aggregation will
763 tpa_info->skb = netdev_alloc_skb(edev->ndev,
764 le16_to_cpu(cqe->len_on_first_bd));
765 if (unlikely(!tpa_info->skb)) {
766 DP_NOTICE(edev, "Failed to allocate SKB for gro\n");
767 tpa_info->state = QEDE_AGG_STATE_ERROR;
771 /* Start filling in the aggregation info */
772 skb_put(tpa_info->skb, le16_to_cpu(cqe->len_on_first_bd));
773 tpa_info->frag_id = 0;
774 tpa_info->state = QEDE_AGG_STATE_START;
776 /* Store some information from first CQE */
777 tpa_info->start_cqe_placement_offset = cqe->placement_offset;
778 tpa_info->start_cqe_bd_len = le16_to_cpu(cqe->len_on_first_bd);
779 if ((le16_to_cpu(cqe->pars_flags.flags) >>
780 PARSING_AND_ERR_FLAGS_TAG8021QEXIST_SHIFT) &
781 PARSING_AND_ERR_FLAGS_TAG8021QEXIST_MASK)
782 tpa_info->vlan_tag = le16_to_cpu(cqe->vlan_tag);
784 tpa_info->vlan_tag = 0;
786 qede_get_rxhash(tpa_info->skb, cqe->bitfields, cqe->rss_hash);
788 /* This is needed in order to enable forwarding support */
789 qede_set_gro_params(edev, tpa_info->skb, cqe);
791 cons_buf: /* We still need to handle bd_len_list to consume buffers */
792 if (likely(cqe->ext_bd_len_list[0]))
793 qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index,
794 le16_to_cpu(cqe->ext_bd_len_list[0]));
796 if (unlikely(cqe->ext_bd_len_list[1])) {
798 "Unlikely - got a TPA aggregation with more than one ext_bd_len_list entry in the TPA start\n");
799 tpa_info->state = QEDE_AGG_STATE_ERROR;
804 static void qede_gro_ip_csum(struct sk_buff *skb)
806 const struct iphdr *iph = ip_hdr(skb);
809 skb_set_transport_header(skb, sizeof(struct iphdr));
812 th->check = ~tcp_v4_check(skb->len - skb_transport_offset(skb),
813 iph->saddr, iph->daddr, 0);
815 tcp_gro_complete(skb);
818 static void qede_gro_ipv6_csum(struct sk_buff *skb)
820 struct ipv6hdr *iph = ipv6_hdr(skb);
823 skb_set_transport_header(skb, sizeof(struct ipv6hdr));
826 th->check = ~tcp_v6_check(skb->len - skb_transport_offset(skb),
827 &iph->saddr, &iph->daddr, 0);
828 tcp_gro_complete(skb);
832 static void qede_gro_receive(struct qede_dev *edev,
833 struct qede_fastpath *fp,
837 /* FW can send a single MTU sized packet from gro flow
838 * due to aggregation timeout/last segment etc. which
839 * is not expected to be a gro packet. If a skb has zero
840 * frags then simply push it in the stack as non gso skb.
842 if (unlikely(!skb->data_len)) {
843 skb_shinfo(skb)->gso_type = 0;
844 skb_shinfo(skb)->gso_size = 0;
849 if (skb_shinfo(skb)->gso_size) {
850 skb_reset_network_header(skb);
852 switch (skb->protocol) {
853 case htons(ETH_P_IP):
854 qede_gro_ip_csum(skb);
856 case htons(ETH_P_IPV6):
857 qede_gro_ipv6_csum(skb);
861 "Error: FW GRO supports only IPv4/IPv6, not 0x%04x\n",
862 ntohs(skb->protocol));
868 skb_record_rx_queue(skb, fp->rxq->rxq_id);
869 qede_skb_receive(edev, fp, fp->rxq, skb, vlan_tag);
872 static inline void qede_tpa_cont(struct qede_dev *edev,
873 struct qede_rx_queue *rxq,
874 struct eth_fast_path_rx_tpa_cont_cqe *cqe)
878 for (i = 0; cqe->len_list[i]; i++)
879 qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index,
880 le16_to_cpu(cqe->len_list[i]));
884 "Strange - TPA cont with more than a single len_list entry\n");
887 static void qede_tpa_end(struct qede_dev *edev,
888 struct qede_fastpath *fp,
889 struct eth_fast_path_rx_tpa_end_cqe *cqe)
891 struct qede_rx_queue *rxq = fp->rxq;
892 struct qede_agg_info *tpa_info;
896 tpa_info = &rxq->tpa_info[cqe->tpa_agg_index];
899 for (i = 0; cqe->len_list[i]; i++)
900 qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index,
901 le16_to_cpu(cqe->len_list[i]));
904 "Strange - TPA emd with more than a single len_list entry\n");
906 if (unlikely(tpa_info->state != QEDE_AGG_STATE_START))
910 if (unlikely(cqe->num_of_bds != tpa_info->frag_id + 1))
912 "Strange - TPA had %02x BDs, but SKB has only %d frags\n",
913 cqe->num_of_bds, tpa_info->frag_id);
914 if (unlikely(skb->len != le16_to_cpu(cqe->total_packet_len)))
916 "Strange - total packet len [cqe] is %4x but SKB has len %04x\n",
917 le16_to_cpu(cqe->total_packet_len), skb->len);
920 page_address(tpa_info->buffer.data) +
921 tpa_info->start_cqe_placement_offset +
922 tpa_info->buffer.page_offset, tpa_info->start_cqe_bd_len);
924 /* Finalize the SKB */
925 skb->protocol = eth_type_trans(skb, edev->ndev);
926 skb->ip_summed = CHECKSUM_UNNECESSARY;
928 /* tcp_gro_complete() will copy NAPI_GRO_CB(skb)->count
929 * to skb_shinfo(skb)->gso_segs
931 NAPI_GRO_CB(skb)->count = le16_to_cpu(cqe->num_of_coalesced_segs);
933 qede_gro_receive(edev, fp, skb, tpa_info->vlan_tag);
935 tpa_info->state = QEDE_AGG_STATE_NONE;
939 tpa_info->state = QEDE_AGG_STATE_NONE;
940 dev_kfree_skb_any(tpa_info->skb);
941 tpa_info->skb = NULL;
944 static u8 qede_check_notunn_csum(u16 flag)
949 if (flag & (PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_MASK <<
950 PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_SHIFT)) {
951 csum_flag |= PARSING_AND_ERR_FLAGS_L4CHKSMERROR_MASK <<
952 PARSING_AND_ERR_FLAGS_L4CHKSMERROR_SHIFT;
953 csum = QEDE_CSUM_UNNECESSARY;
956 csum_flag |= PARSING_AND_ERR_FLAGS_IPHDRERROR_MASK <<
957 PARSING_AND_ERR_FLAGS_IPHDRERROR_SHIFT;
959 if (csum_flag & flag)
960 return QEDE_CSUM_ERROR;
965 static u8 qede_check_csum(u16 flag)
967 if (!qede_tunn_exist(flag))
968 return qede_check_notunn_csum(flag);
970 return qede_check_tunn_csum(flag);
973 static bool qede_pkt_is_ip_fragmented(struct eth_fast_path_rx_reg_cqe *cqe,
976 u8 tun_pars_flg = cqe->tunnel_pars_flags.flags;
978 if ((tun_pars_flg & (ETH_TUNNEL_PARSING_FLAGS_IPV4_FRAGMENT_MASK <<
979 ETH_TUNNEL_PARSING_FLAGS_IPV4_FRAGMENT_SHIFT)) ||
980 (flag & (PARSING_AND_ERR_FLAGS_IPV4FRAG_MASK <<
981 PARSING_AND_ERR_FLAGS_IPV4FRAG_SHIFT)))
987 /* Return true iff packet is to be passed to stack */
988 static bool qede_rx_xdp(struct qede_dev *edev,
989 struct qede_fastpath *fp,
990 struct qede_rx_queue *rxq,
991 struct bpf_prog *prog,
992 struct sw_rx_data *bd,
993 struct eth_fast_path_rx_reg_cqe *cqe)
995 u16 len = le16_to_cpu(cqe->len_on_first_bd);
999 xdp.data = page_address(bd->data) + cqe->placement_offset;
1000 xdp.data_end = xdp.data + len;
1002 /* Queues always have a full reset currently, so for the time
1003 * being until there's atomic program replace just mark read
1004 * side for map helpers.
1007 act = bpf_prog_run_xdp(prog, &xdp);
1010 if (act == XDP_PASS)
1013 /* Count number of packets not to be passed to stack */
1018 /* We need the replacement buffer before transmit. */
1019 if (qede_alloc_rx_buffer(rxq, true)) {
1020 qede_recycle_rx_bd_ring(rxq, 1);
1021 trace_xdp_exception(edev->ndev, prog, act);
1025 /* Now if there's a transmission problem, we'd still have to
1026 * throw current buffer, as replacement was already allocated.
1028 if (qede_xdp_xmit(edev, fp, bd, cqe->placement_offset, len)) {
1029 dma_unmap_page(rxq->dev, bd->mapping,
1030 PAGE_SIZE, DMA_BIDIRECTIONAL);
1031 __free_page(bd->data);
1032 trace_xdp_exception(edev->ndev, prog, act);
1035 /* Regardless, we've consumed an Rx BD */
1036 qede_rx_bd_ring_consume(rxq);
1040 bpf_warn_invalid_xdp_action(act);
1042 trace_xdp_exception(edev->ndev, prog, act);
1044 qede_recycle_rx_bd_ring(rxq, cqe->bd_num);
1050 static struct sk_buff *qede_rx_allocate_skb(struct qede_dev *edev,
1051 struct qede_rx_queue *rxq,
1052 struct sw_rx_data *bd, u16 len,
1055 unsigned int offset = bd->page_offset;
1056 struct skb_frag_struct *frag;
1057 struct page *page = bd->data;
1058 unsigned int pull_len;
1059 struct sk_buff *skb;
1062 /* Allocate a new SKB with a sufficient large header len */
1063 skb = netdev_alloc_skb(edev->ndev, QEDE_RX_HDR_SIZE);
1067 /* Copy data into SKB - if it's small, we can simply copy it and
1068 * re-use the already allcoated & mapped memory.
1070 if (len + pad <= edev->rx_copybreak) {
1071 memcpy(skb_put(skb, len),
1072 page_address(page) + pad + offset, len);
1073 qede_reuse_page(rxq, bd);
1077 frag = &skb_shinfo(skb)->frags[0];
1079 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags,
1080 page, pad + offset, len, rxq->rx_buf_seg_size);
1082 va = skb_frag_address(frag);
1083 pull_len = eth_get_headlen(va, QEDE_RX_HDR_SIZE);
1085 /* Align the pull_len to optimize memcpy */
1086 memcpy(skb->data, va, ALIGN(pull_len, sizeof(long)));
1088 /* Correct the skb & frag sizes offset after the pull */
1089 skb_frag_size_sub(frag, pull_len);
1090 frag->page_offset += pull_len;
1091 skb->data_len -= pull_len;
1092 skb->tail += pull_len;
1094 if (unlikely(qede_realloc_rx_buffer(rxq, bd))) {
1095 /* Incr page ref count to reuse on allocation failure so
1096 * that it doesn't get freed while freeing SKB [as its
1097 * already mapped there].
1100 dev_kfree_skb_any(skb);
1105 /* We've consumed the first BD and prepared an SKB */
1106 qede_rx_bd_ring_consume(rxq);
1110 static int qede_rx_build_jumbo(struct qede_dev *edev,
1111 struct qede_rx_queue *rxq,
1112 struct sk_buff *skb,
1113 struct eth_fast_path_rx_reg_cqe *cqe,
1116 u16 pkt_len = le16_to_cpu(cqe->pkt_len);
1117 struct sw_rx_data *bd;
1121 pkt_len -= first_bd_len;
1123 /* We've already used one BD for the SKB. Now take care of the rest */
1124 for (num_frags = cqe->bd_num - 1; num_frags > 0; num_frags--) {
1125 u16 cur_size = pkt_len > rxq->rx_buf_size ? rxq->rx_buf_size :
1128 if (unlikely(!cur_size)) {
1130 "Still got %d BDs for mapping jumbo, but length became 0\n",
1135 /* We need a replacement buffer for each BD */
1136 if (unlikely(qede_alloc_rx_buffer(rxq, true)))
1139 /* Now that we've allocated the replacement buffer,
1140 * we can safely consume the next BD and map it to the SKB.
1142 bd_cons_idx = rxq->sw_rx_cons & NUM_RX_BDS_MAX;
1143 bd = &rxq->sw_rx_ring[bd_cons_idx];
1144 qede_rx_bd_ring_consume(rxq);
1146 dma_unmap_page(rxq->dev, bd->mapping,
1147 PAGE_SIZE, DMA_FROM_DEVICE);
1149 skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags++,
1150 bd->data, 0, cur_size);
1152 skb->truesize += PAGE_SIZE;
1153 skb->data_len += cur_size;
1154 skb->len += cur_size;
1155 pkt_len -= cur_size;
1158 if (unlikely(pkt_len))
1160 "Mapped all BDs of jumbo, but still have %d bytes\n",
1167 static int qede_rx_process_tpa_cqe(struct qede_dev *edev,
1168 struct qede_fastpath *fp,
1169 struct qede_rx_queue *rxq,
1170 union eth_rx_cqe *cqe,
1171 enum eth_rx_cqe_type type)
1174 case ETH_RX_CQE_TYPE_TPA_START:
1175 qede_tpa_start(edev, rxq, &cqe->fast_path_tpa_start);
1177 case ETH_RX_CQE_TYPE_TPA_CONT:
1178 qede_tpa_cont(edev, rxq, &cqe->fast_path_tpa_cont);
1180 case ETH_RX_CQE_TYPE_TPA_END:
1181 qede_tpa_end(edev, fp, &cqe->fast_path_tpa_end);
1188 static int qede_rx_process_cqe(struct qede_dev *edev,
1189 struct qede_fastpath *fp,
1190 struct qede_rx_queue *rxq)
1192 struct bpf_prog *xdp_prog = READ_ONCE(rxq->xdp_prog);
1193 struct eth_fast_path_rx_reg_cqe *fp_cqe;
1194 u16 len, pad, bd_cons_idx, parse_flag;
1195 enum eth_rx_cqe_type cqe_type;
1196 union eth_rx_cqe *cqe;
1197 struct sw_rx_data *bd;
1198 struct sk_buff *skb;
1202 /* Get the CQE from the completion ring */
1203 cqe = (union eth_rx_cqe *)qed_chain_consume(&rxq->rx_comp_ring);
1204 cqe_type = cqe->fast_path_regular.type;
1206 /* Process an unlikely slowpath event */
1207 if (unlikely(cqe_type == ETH_RX_CQE_TYPE_SLOW_PATH)) {
1208 struct eth_slow_path_rx_cqe *sp_cqe;
1210 sp_cqe = (struct eth_slow_path_rx_cqe *)cqe;
1211 edev->ops->eth_cqe_completion(edev->cdev, fp->id, sp_cqe);
1215 /* Handle TPA cqes */
1216 if (cqe_type != ETH_RX_CQE_TYPE_REGULAR)
1217 return qede_rx_process_tpa_cqe(edev, fp, rxq, cqe, cqe_type);
1219 /* Get the data from the SW ring; Consume it only after it's evident
1220 * we wouldn't recycle it.
1222 bd_cons_idx = rxq->sw_rx_cons & NUM_RX_BDS_MAX;
1223 bd = &rxq->sw_rx_ring[bd_cons_idx];
1225 fp_cqe = &cqe->fast_path_regular;
1226 len = le16_to_cpu(fp_cqe->len_on_first_bd);
1227 pad = fp_cqe->placement_offset;
1229 /* Run eBPF program if one is attached */
1231 if (!qede_rx_xdp(edev, fp, rxq, xdp_prog, bd, fp_cqe))
1234 /* If this is an error packet then drop it */
1235 flags = cqe->fast_path_regular.pars_flags.flags;
1236 parse_flag = le16_to_cpu(flags);
1238 csum_flag = qede_check_csum(parse_flag);
1239 if (unlikely(csum_flag == QEDE_CSUM_ERROR)) {
1240 if (qede_pkt_is_ip_fragmented(fp_cqe, parse_flag)) {
1244 "CQE has error, flags = %x, dropping incoming packet\n",
1246 rxq->rx_hw_errors++;
1247 qede_recycle_rx_bd_ring(rxq, fp_cqe->bd_num);
1252 /* Basic validation passed; Need to prepare an SKB. This would also
1253 * guarantee to finally consume the first BD upon success.
1255 skb = qede_rx_allocate_skb(edev, rxq, bd, len, pad);
1257 rxq->rx_alloc_errors++;
1258 qede_recycle_rx_bd_ring(rxq, fp_cqe->bd_num);
1262 /* In case of Jumbo packet, several PAGE_SIZEd buffers will be pointed
1265 if (fp_cqe->bd_num > 1) {
1266 u16 unmapped_frags = qede_rx_build_jumbo(edev, rxq, skb,
1269 if (unlikely(unmapped_frags > 0)) {
1270 qede_recycle_rx_bd_ring(rxq, unmapped_frags);
1271 dev_kfree_skb_any(skb);
1276 /* The SKB contains all the data. Now prepare meta-magic */
1277 skb->protocol = eth_type_trans(skb, edev->ndev);
1278 qede_get_rxhash(skb, fp_cqe->bitfields, fp_cqe->rss_hash);
1279 qede_set_skb_csum(skb, csum_flag);
1280 skb_record_rx_queue(skb, rxq->rxq_id);
1281 qede_ptp_record_rx_ts(edev, cqe, skb);
1283 /* SKB is prepared - pass it to stack */
1284 qede_skb_receive(edev, fp, rxq, skb, le16_to_cpu(fp_cqe->vlan_tag));
1289 static int qede_rx_int(struct qede_fastpath *fp, int budget)
1291 struct qede_rx_queue *rxq = fp->rxq;
1292 struct qede_dev *edev = fp->edev;
1293 u16 hw_comp_cons, sw_comp_cons;
1296 hw_comp_cons = le16_to_cpu(*rxq->hw_cons_ptr);
1297 sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring);
1299 /* Memory barrier to prevent the CPU from doing speculative reads of CQE
1300 * / BD in the while-loop before reading hw_comp_cons. If the CQE is
1301 * read before it is written by FW, then FW writes CQE and SB, and then
1302 * the CPU reads the hw_comp_cons, it will use an old CQE.
1306 /* Loop to complete all indicated BDs */
1307 while ((sw_comp_cons != hw_comp_cons) && (work_done < budget)) {
1308 qede_rx_process_cqe(edev, fp, rxq);
1309 qed_chain_recycle_consumed(&rxq->rx_comp_ring);
1310 sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring);
1314 /* Allocate replacement buffers */
1315 while (rxq->num_rx_buffers - rxq->filled_buffers)
1316 if (qede_alloc_rx_buffer(rxq, false))
1319 /* Update producers */
1320 qede_update_rx_prod(edev, rxq);
1325 static bool qede_poll_is_more_work(struct qede_fastpath *fp)
1327 qed_sb_update_sb_idx(fp->sb_info);
1329 /* *_has_*_work() reads the status block, thus we need to ensure that
1330 * status block indices have been actually read (qed_sb_update_sb_idx)
1331 * prior to this check (*_has_*_work) so that we won't write the
1332 * "newer" value of the status block to HW (if there was a DMA right
1333 * after qede_has_rx_work and if there is no rmb, the memory reading
1334 * (qed_sb_update_sb_idx) may be postponed to right before *_ack_sb).
1335 * In this case there will never be another interrupt until there is
1336 * another update of the status block, while there is still unhandled
1341 if (likely(fp->type & QEDE_FASTPATH_RX))
1342 if (qede_has_rx_work(fp->rxq))
1345 if (fp->type & QEDE_FASTPATH_XDP)
1346 if (qede_txq_has_work(fp->xdp_tx))
1349 if (likely(fp->type & QEDE_FASTPATH_TX))
1350 if (qede_txq_has_work(fp->txq))
1356 /*********************
1357 * NDO & API related *
1358 *********************/
1359 int qede_poll(struct napi_struct *napi, int budget)
1361 struct qede_fastpath *fp = container_of(napi, struct qede_fastpath,
1363 struct qede_dev *edev = fp->edev;
1364 int rx_work_done = 0;
1366 if (likely(fp->type & QEDE_FASTPATH_TX) && qede_txq_has_work(fp->txq))
1367 qede_tx_int(edev, fp->txq);
1369 if ((fp->type & QEDE_FASTPATH_XDP) && qede_txq_has_work(fp->xdp_tx))
1370 qede_xdp_tx_int(edev, fp->xdp_tx);
1372 rx_work_done = (likely(fp->type & QEDE_FASTPATH_RX) &&
1373 qede_has_rx_work(fp->rxq)) ?
1374 qede_rx_int(fp, budget) : 0;
1375 if (rx_work_done < budget) {
1376 if (!qede_poll_is_more_work(fp)) {
1377 napi_complete_done(napi, rx_work_done);
1379 /* Update and reenable interrupts */
1380 qed_sb_ack(fp->sb_info, IGU_INT_ENABLE, 1);
1382 rx_work_done = budget;
1387 u16 xdp_prod = qed_chain_get_prod_idx(&fp->xdp_tx->tx_pbl);
1390 fp->xdp_tx->tx_db.data.bd_prod = cpu_to_le16(xdp_prod);
1391 qede_update_tx_producer(fp->xdp_tx);
1394 return rx_work_done;
1397 irqreturn_t qede_msix_fp_int(int irq, void *fp_cookie)
1399 struct qede_fastpath *fp = fp_cookie;
1401 qed_sb_ack(fp->sb_info, IGU_INT_DISABLE, 0 /*do not update*/);
1403 napi_schedule_irqoff(&fp->napi);
1407 /* Main transmit function */
1408 netdev_tx_t qede_start_xmit(struct sk_buff *skb, struct net_device *ndev)
1410 struct qede_dev *edev = netdev_priv(ndev);
1411 struct netdev_queue *netdev_txq;
1412 struct qede_tx_queue *txq;
1413 struct eth_tx_1st_bd *first_bd;
1414 struct eth_tx_2nd_bd *second_bd = NULL;
1415 struct eth_tx_3rd_bd *third_bd = NULL;
1416 struct eth_tx_bd *tx_data_bd = NULL;
1420 int rc, frag_idx = 0, ipv6_ext = 0;
1424 bool data_split = false;
1426 /* Get tx-queue context and netdev index */
1427 txq_index = skb_get_queue_mapping(skb);
1428 WARN_ON(txq_index >= QEDE_TSS_COUNT(edev));
1429 txq = edev->fp_array[edev->fp_num_rx + txq_index].txq;
1430 netdev_txq = netdev_get_tx_queue(ndev, txq_index);
1432 WARN_ON(qed_chain_get_elem_left(&txq->tx_pbl) < (MAX_SKB_FRAGS + 1));
1434 xmit_type = qede_xmit_type(skb, &ipv6_ext);
1436 #if ((MAX_SKB_FRAGS + 2) > ETH_TX_MAX_BDS_PER_NON_LSO_PACKET)
1437 if (qede_pkt_req_lin(skb, xmit_type)) {
1438 if (skb_linearize(skb)) {
1440 "SKB linearization failed - silently dropping this SKB\n");
1441 dev_kfree_skb_any(skb);
1442 return NETDEV_TX_OK;
1447 /* Fill the entry in the SW ring and the BDs in the FW ring */
1448 idx = txq->sw_tx_prod & NUM_TX_BDS_MAX;
1449 txq->sw_tx_ring.skbs[idx].skb = skb;
1450 first_bd = (struct eth_tx_1st_bd *)
1451 qed_chain_produce(&txq->tx_pbl);
1452 memset(first_bd, 0, sizeof(*first_bd));
1453 first_bd->data.bd_flags.bitfields =
1454 1 << ETH_TX_1ST_BD_FLAGS_START_BD_SHIFT;
1456 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP))
1457 qede_ptp_tx_ts(edev, skb);
1459 /* Map skb linear data for DMA and set in the first BD */
1460 mapping = dma_map_single(txq->dev, skb->data,
1461 skb_headlen(skb), DMA_TO_DEVICE);
1462 if (unlikely(dma_mapping_error(txq->dev, mapping))) {
1463 DP_NOTICE(edev, "SKB mapping failed\n");
1464 qede_free_failed_tx_pkt(txq, first_bd, 0, false);
1465 qede_update_tx_producer(txq);
1466 return NETDEV_TX_OK;
1469 BD_SET_UNMAP_ADDR_LEN(first_bd, mapping, skb_headlen(skb));
1471 /* In case there is IPv6 with extension headers or LSO we need 2nd and
1474 if (unlikely((xmit_type & XMIT_LSO) | ipv6_ext)) {
1475 second_bd = (struct eth_tx_2nd_bd *)
1476 qed_chain_produce(&txq->tx_pbl);
1477 memset(second_bd, 0, sizeof(*second_bd));
1480 third_bd = (struct eth_tx_3rd_bd *)
1481 qed_chain_produce(&txq->tx_pbl);
1482 memset(third_bd, 0, sizeof(*third_bd));
1485 /* We need to fill in additional data in second_bd... */
1486 tx_data_bd = (struct eth_tx_bd *)second_bd;
1489 if (skb_vlan_tag_present(skb)) {
1490 first_bd->data.vlan = cpu_to_le16(skb_vlan_tag_get(skb));
1491 first_bd->data.bd_flags.bitfields |=
1492 1 << ETH_TX_1ST_BD_FLAGS_VLAN_INSERTION_SHIFT;
1495 /* Fill the parsing flags & params according to the requested offload */
1496 if (xmit_type & XMIT_L4_CSUM) {
1497 /* We don't re-calculate IP checksum as it is already done by
1500 first_bd->data.bd_flags.bitfields |=
1501 1 << ETH_TX_1ST_BD_FLAGS_L4_CSUM_SHIFT;
1503 if (xmit_type & XMIT_ENC) {
1504 first_bd->data.bd_flags.bitfields |=
1505 1 << ETH_TX_1ST_BD_FLAGS_IP_CSUM_SHIFT;
1506 first_bd->data.bitfields |=
1507 1 << ETH_TX_DATA_1ST_BD_TUNN_FLAG_SHIFT;
1510 /* Legacy FW had flipped behavior in regard to this bit -
1511 * I.e., needed to set to prevent FW from touching encapsulated
1512 * packets when it didn't need to.
1514 if (unlikely(txq->is_legacy))
1515 first_bd->data.bitfields ^=
1516 1 << ETH_TX_DATA_1ST_BD_TUNN_FLAG_SHIFT;
1518 /* If the packet is IPv6 with extension header, indicate that
1519 * to FW and pass few params, since the device cracker doesn't
1520 * support parsing IPv6 with extension header/s.
1522 if (unlikely(ipv6_ext))
1523 qede_set_params_for_ipv6_ext(skb, second_bd, third_bd);
1526 if (xmit_type & XMIT_LSO) {
1527 first_bd->data.bd_flags.bitfields |=
1528 (1 << ETH_TX_1ST_BD_FLAGS_LSO_SHIFT);
1529 third_bd->data.lso_mss =
1530 cpu_to_le16(skb_shinfo(skb)->gso_size);
1532 if (unlikely(xmit_type & XMIT_ENC)) {
1533 first_bd->data.bd_flags.bitfields |=
1534 1 << ETH_TX_1ST_BD_FLAGS_TUNN_IP_CSUM_SHIFT;
1536 if (xmit_type & XMIT_ENC_GSO_L4_CSUM) {
1537 u8 tmp = ETH_TX_1ST_BD_FLAGS_TUNN_L4_CSUM_SHIFT;
1539 first_bd->data.bd_flags.bitfields |= 1 << tmp;
1541 hlen = qede_get_skb_hlen(skb, true);
1543 first_bd->data.bd_flags.bitfields |=
1544 1 << ETH_TX_1ST_BD_FLAGS_IP_CSUM_SHIFT;
1545 hlen = qede_get_skb_hlen(skb, false);
1548 /* @@@TBD - if will not be removed need to check */
1549 third_bd->data.bitfields |=
1550 cpu_to_le16(1 << ETH_TX_DATA_3RD_BD_HDR_NBD_SHIFT);
1552 /* Make life easier for FW guys who can't deal with header and
1553 * data on same BD. If we need to split, use the second bd...
1555 if (unlikely(skb_headlen(skb) > hlen)) {
1556 DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED,
1557 "TSO split header size is %d (%x:%x)\n",
1558 first_bd->nbytes, first_bd->addr.hi,
1561 mapping = HILO_U64(le32_to_cpu(first_bd->addr.hi),
1562 le32_to_cpu(first_bd->addr.lo)) +
1565 BD_SET_UNMAP_ADDR_LEN(tx_data_bd, mapping,
1566 le16_to_cpu(first_bd->nbytes) -
1569 /* this marks the BD as one that has no
1570 * individual mapping
1572 txq->sw_tx_ring.skbs[idx].flags |= QEDE_TSO_SPLIT_BD;
1574 first_bd->nbytes = cpu_to_le16(hlen);
1576 tx_data_bd = (struct eth_tx_bd *)third_bd;
1580 first_bd->data.bitfields |=
1581 (skb->len & ETH_TX_DATA_1ST_BD_PKT_LEN_MASK) <<
1582 ETH_TX_DATA_1ST_BD_PKT_LEN_SHIFT;
1585 /* Handle fragmented skb */
1586 /* special handle for frags inside 2nd and 3rd bds.. */
1587 while (tx_data_bd && frag_idx < skb_shinfo(skb)->nr_frags) {
1588 rc = map_frag_to_bd(txq,
1589 &skb_shinfo(skb)->frags[frag_idx],
1592 qede_free_failed_tx_pkt(txq, first_bd, nbd, data_split);
1593 qede_update_tx_producer(txq);
1594 return NETDEV_TX_OK;
1597 if (tx_data_bd == (struct eth_tx_bd *)second_bd)
1598 tx_data_bd = (struct eth_tx_bd *)third_bd;
1605 /* map last frags into 4th, 5th .... */
1606 for (; frag_idx < skb_shinfo(skb)->nr_frags; frag_idx++, nbd++) {
1607 tx_data_bd = (struct eth_tx_bd *)
1608 qed_chain_produce(&txq->tx_pbl);
1610 memset(tx_data_bd, 0, sizeof(*tx_data_bd));
1612 rc = map_frag_to_bd(txq,
1613 &skb_shinfo(skb)->frags[frag_idx],
1616 qede_free_failed_tx_pkt(txq, first_bd, nbd, data_split);
1617 qede_update_tx_producer(txq);
1618 return NETDEV_TX_OK;
1622 /* update the first BD with the actual num BDs */
1623 first_bd->data.nbds = nbd;
1625 netdev_tx_sent_queue(netdev_txq, skb->len);
1627 skb_tx_timestamp(skb);
1629 /* Advance packet producer only before sending the packet since mapping
1630 * of pages may fail.
1634 /* 'next page' entries are counted in the producer value */
1635 txq->tx_db.data.bd_prod =
1636 cpu_to_le16(qed_chain_get_prod_idx(&txq->tx_pbl));
1638 if (!skb->xmit_more || netif_xmit_stopped(netdev_txq))
1639 qede_update_tx_producer(txq);
1641 if (unlikely(qed_chain_get_elem_left(&txq->tx_pbl)
1642 < (MAX_SKB_FRAGS + 1))) {
1644 qede_update_tx_producer(txq);
1646 netif_tx_stop_queue(netdev_txq);
1648 DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED,
1649 "Stop queue was called\n");
1650 /* paired memory barrier is in qede_tx_int(), we have to keep
1651 * ordering of set_bit() in netif_tx_stop_queue() and read of
1656 if ((qed_chain_get_elem_left(&txq->tx_pbl) >=
1657 (MAX_SKB_FRAGS + 1)) &&
1658 (edev->state == QEDE_STATE_OPEN)) {
1659 netif_tx_wake_queue(netdev_txq);
1660 DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED,
1661 "Wake queue was called\n");
1665 return NETDEV_TX_OK;
1668 /* 8B udp header + 8B base tunnel header + 32B option length */
1669 #define QEDE_MAX_TUN_HDR_LEN 48
1671 netdev_features_t qede_features_check(struct sk_buff *skb,
1672 struct net_device *dev,
1673 netdev_features_t features)
1675 if (skb->encapsulation) {
1678 switch (vlan_get_protocol(skb)) {
1679 case htons(ETH_P_IP):
1680 l4_proto = ip_hdr(skb)->protocol;
1682 case htons(ETH_P_IPV6):
1683 l4_proto = ipv6_hdr(skb)->nexthdr;
1689 /* Disable offloads for geneve tunnels, as HW can't parse
1690 * the geneve header which has option length greater than 32B.
1692 if ((l4_proto == IPPROTO_UDP) &&
1693 ((skb_inner_mac_header(skb) -
1694 skb_transport_header(skb)) > QEDE_MAX_TUN_HDR_LEN))
1695 return features & ~(NETIF_F_CSUM_MASK |