1 // SPDX-License-Identifier: (GPL-2.0-only OR BSD-3-Clause)
2 /* QLogic qede NIC Driver
3 * Copyright (c) 2015-2017 QLogic Corporation
4 * Copyright (c) 2019-2020 Marvell International Ltd.
7 #include <linux/netdevice.h>
8 #include <linux/etherdevice.h>
9 #include <linux/skbuff.h>
10 #include <linux/bpf_trace.h>
11 #include <net/udp_tunnel.h>
15 #include <linux/if_ether.h>
16 #include <linux/if_vlan.h>
17 #include <net/ip6_checksum.h>
20 #include <linux/qed/qed_if.h>
22 /*********************************
23 * Content also used by slowpath *
24 *********************************/
26 int qede_alloc_rx_buffer(struct qede_rx_queue *rxq, bool allow_lazy)
28 struct sw_rx_data *sw_rx_data;
29 struct eth_rx_bd *rx_bd;
33 /* In case lazy-allocation is allowed, postpone allocation until the
34 * end of the NAPI run. We'd still need to make sure the Rx ring has
35 * sufficient buffers to guarantee an additional Rx interrupt.
37 if (allow_lazy && likely(rxq->filled_buffers > 12)) {
38 rxq->filled_buffers--;
42 data = alloc_pages(GFP_ATOMIC, 0);
46 /* Map the entire page as it would be used
47 * for multiple RX buffer segment size mapping.
49 mapping = dma_map_page(rxq->dev, data, 0,
50 PAGE_SIZE, rxq->data_direction);
51 if (unlikely(dma_mapping_error(rxq->dev, mapping))) {
56 sw_rx_data = &rxq->sw_rx_ring[rxq->sw_rx_prod & NUM_RX_BDS_MAX];
57 sw_rx_data->page_offset = 0;
58 sw_rx_data->data = data;
59 sw_rx_data->mapping = mapping;
61 /* Advance PROD and get BD pointer */
62 rx_bd = (struct eth_rx_bd *)qed_chain_produce(&rxq->rx_bd_ring);
64 rx_bd->addr.hi = cpu_to_le32(upper_32_bits(mapping));
65 rx_bd->addr.lo = cpu_to_le32(lower_32_bits(mapping) +
69 rxq->filled_buffers++;
74 /* Unmap the data and free skb */
75 int qede_free_tx_pkt(struct qede_dev *edev, struct qede_tx_queue *txq, int *len)
77 u16 idx = txq->sw_tx_cons;
78 struct sk_buff *skb = txq->sw_tx_ring.skbs[idx].skb;
79 struct eth_tx_1st_bd *first_bd;
80 struct eth_tx_bd *tx_data_bd;
83 bool data_split = txq->sw_tx_ring.skbs[idx].flags & QEDE_TSO_SPLIT_BD;
84 int i, split_bd_len = 0;
88 "skb is null for txq idx=%d txq->sw_tx_cons=%d txq->sw_tx_prod=%d\n",
89 idx, txq->sw_tx_cons, txq->sw_tx_prod);
95 first_bd = (struct eth_tx_1st_bd *)qed_chain_consume(&txq->tx_pbl);
99 nbds = first_bd->data.nbds;
102 struct eth_tx_bd *split = (struct eth_tx_bd *)
103 qed_chain_consume(&txq->tx_pbl);
104 split_bd_len = BD_UNMAP_LEN(split);
107 dma_unmap_single(&edev->pdev->dev, BD_UNMAP_ADDR(first_bd),
108 BD_UNMAP_LEN(first_bd) + split_bd_len, DMA_TO_DEVICE);
110 /* Unmap the data of the skb frags */
111 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++, bds_consumed++) {
112 tx_data_bd = (struct eth_tx_bd *)
113 qed_chain_consume(&txq->tx_pbl);
114 dma_unmap_page(&edev->pdev->dev, BD_UNMAP_ADDR(tx_data_bd),
115 BD_UNMAP_LEN(tx_data_bd), DMA_TO_DEVICE);
118 while (bds_consumed++ < nbds)
119 qed_chain_consume(&txq->tx_pbl);
122 dev_kfree_skb_any(skb);
123 txq->sw_tx_ring.skbs[idx].skb = NULL;
124 txq->sw_tx_ring.skbs[idx].flags = 0;
129 /* Unmap the data and free skb when mapping failed during start_xmit */
130 static void qede_free_failed_tx_pkt(struct qede_tx_queue *txq,
131 struct eth_tx_1st_bd *first_bd,
132 int nbd, bool data_split)
134 u16 idx = txq->sw_tx_prod;
135 struct sk_buff *skb = txq->sw_tx_ring.skbs[idx].skb;
136 struct eth_tx_bd *tx_data_bd;
137 int i, split_bd_len = 0;
139 /* Return prod to its position before this skb was handled */
140 qed_chain_set_prod(&txq->tx_pbl,
141 le16_to_cpu(txq->tx_db.data.bd_prod), first_bd);
143 first_bd = (struct eth_tx_1st_bd *)qed_chain_produce(&txq->tx_pbl);
146 struct eth_tx_bd *split = (struct eth_tx_bd *)
147 qed_chain_produce(&txq->tx_pbl);
148 split_bd_len = BD_UNMAP_LEN(split);
152 dma_unmap_single(txq->dev, BD_UNMAP_ADDR(first_bd),
153 BD_UNMAP_LEN(first_bd) + split_bd_len, DMA_TO_DEVICE);
155 /* Unmap the data of the skb frags */
156 for (i = 0; i < nbd; i++) {
157 tx_data_bd = (struct eth_tx_bd *)
158 qed_chain_produce(&txq->tx_pbl);
159 if (tx_data_bd->nbytes)
160 dma_unmap_page(txq->dev,
161 BD_UNMAP_ADDR(tx_data_bd),
162 BD_UNMAP_LEN(tx_data_bd), DMA_TO_DEVICE);
165 /* Return again prod to its position before this skb was handled */
166 qed_chain_set_prod(&txq->tx_pbl,
167 le16_to_cpu(txq->tx_db.data.bd_prod), first_bd);
170 dev_kfree_skb_any(skb);
171 txq->sw_tx_ring.skbs[idx].skb = NULL;
172 txq->sw_tx_ring.skbs[idx].flags = 0;
175 static u32 qede_xmit_type(struct sk_buff *skb, int *ipv6_ext)
177 u32 rc = XMIT_L4_CSUM;
180 if (skb->ip_summed != CHECKSUM_PARTIAL)
183 l3_proto = vlan_get_protocol(skb);
184 if (l3_proto == htons(ETH_P_IPV6) &&
185 (ipv6_hdr(skb)->nexthdr == NEXTHDR_IPV6))
188 if (skb->encapsulation) {
190 if (skb_is_gso(skb)) {
191 unsigned short gso_type = skb_shinfo(skb)->gso_type;
193 if ((gso_type & SKB_GSO_UDP_TUNNEL_CSUM) ||
194 (gso_type & SKB_GSO_GRE_CSUM))
195 rc |= XMIT_ENC_GSO_L4_CSUM;
208 static void qede_set_params_for_ipv6_ext(struct sk_buff *skb,
209 struct eth_tx_2nd_bd *second_bd,
210 struct eth_tx_3rd_bd *third_bd)
213 u16 bd2_bits1 = 0, bd2_bits2 = 0;
215 bd2_bits1 |= (1 << ETH_TX_DATA_2ND_BD_IPV6_EXT_SHIFT);
217 bd2_bits2 |= ((((u8 *)skb_transport_header(skb) - skb->data) >> 1) &
218 ETH_TX_DATA_2ND_BD_L4_HDR_START_OFFSET_W_MASK)
219 << ETH_TX_DATA_2ND_BD_L4_HDR_START_OFFSET_W_SHIFT;
221 bd2_bits1 |= (ETH_L4_PSEUDO_CSUM_CORRECT_LENGTH <<
222 ETH_TX_DATA_2ND_BD_L4_PSEUDO_CSUM_MODE_SHIFT);
224 if (vlan_get_protocol(skb) == htons(ETH_P_IPV6))
225 l4_proto = ipv6_hdr(skb)->nexthdr;
227 l4_proto = ip_hdr(skb)->protocol;
229 if (l4_proto == IPPROTO_UDP)
230 bd2_bits1 |= 1 << ETH_TX_DATA_2ND_BD_L4_UDP_SHIFT;
233 third_bd->data.bitfields |=
234 cpu_to_le16(((tcp_hdrlen(skb) / 4) &
235 ETH_TX_DATA_3RD_BD_TCP_HDR_LEN_DW_MASK) <<
236 ETH_TX_DATA_3RD_BD_TCP_HDR_LEN_DW_SHIFT);
238 second_bd->data.bitfields1 = cpu_to_le16(bd2_bits1);
239 second_bd->data.bitfields2 = cpu_to_le16(bd2_bits2);
242 static int map_frag_to_bd(struct qede_tx_queue *txq,
243 skb_frag_t *frag, struct eth_tx_bd *bd)
247 /* Map skb non-linear frag data for DMA */
248 mapping = skb_frag_dma_map(txq->dev, frag, 0,
249 skb_frag_size(frag), DMA_TO_DEVICE);
250 if (unlikely(dma_mapping_error(txq->dev, mapping)))
253 /* Setup the data pointer of the frag data */
254 BD_SET_UNMAP_ADDR_LEN(bd, mapping, skb_frag_size(frag));
259 static u16 qede_get_skb_hlen(struct sk_buff *skb, bool is_encap_pkt)
262 return (skb_inner_transport_header(skb) +
263 inner_tcp_hdrlen(skb) - skb->data);
265 return (skb_transport_header(skb) +
266 tcp_hdrlen(skb) - skb->data);
269 /* +2 for 1st BD for headers and 2nd BD for headlen (if required) */
270 #if ((MAX_SKB_FRAGS + 2) > ETH_TX_MAX_BDS_PER_NON_LSO_PACKET)
271 static bool qede_pkt_req_lin(struct sk_buff *skb, u8 xmit_type)
273 int allowed_frags = ETH_TX_MAX_BDS_PER_NON_LSO_PACKET - 1;
275 if (xmit_type & XMIT_LSO) {
278 hlen = qede_get_skb_hlen(skb, xmit_type & XMIT_ENC);
280 /* linear payload would require its own BD */
281 if (skb_headlen(skb) > hlen)
285 return (skb_shinfo(skb)->nr_frags > allowed_frags);
289 static inline void qede_update_tx_producer(struct qede_tx_queue *txq)
291 /* wmb makes sure that the BDs data is updated before updating the
292 * producer, otherwise FW may read old data from the BDs.
296 writel(txq->tx_db.raw, txq->doorbell_addr);
298 /* Fence required to flush the write combined buffer, since another
299 * CPU may write to the same doorbell address and data may be lost
300 * due to relaxed order nature of write combined bar.
305 static int qede_xdp_xmit(struct qede_tx_queue *txq, dma_addr_t dma, u16 pad,
306 u16 len, struct page *page, struct xdp_frame *xdpf)
308 struct eth_tx_1st_bd *bd;
309 struct sw_tx_xdp *xdp;
312 if (unlikely(qed_chain_get_elem_used(&txq->tx_pbl) >=
313 txq->num_tx_buffers)) {
318 bd = qed_chain_produce(&txq->tx_pbl);
320 bd->data.bd_flags.bitfields = BIT(ETH_TX_1ST_BD_FLAGS_START_BD_SHIFT);
322 val = (len & ETH_TX_DATA_1ST_BD_PKT_LEN_MASK) <<
323 ETH_TX_DATA_1ST_BD_PKT_LEN_SHIFT;
325 bd->data.bitfields = cpu_to_le16(val);
327 /* We can safely ignore the offset, as it's 0 for XDP */
328 BD_SET_UNMAP_ADDR_LEN(bd, dma + pad, len);
330 xdp = txq->sw_tx_ring.xdp + txq->sw_tx_prod;
335 txq->sw_tx_prod = (txq->sw_tx_prod + 1) % txq->num_tx_buffers;
340 int qede_xdp_transmit(struct net_device *dev, int n_frames,
341 struct xdp_frame **frames, u32 flags)
343 struct qede_dev *edev = netdev_priv(dev);
344 struct device *dmadev = &edev->pdev->dev;
345 struct qede_tx_queue *xdp_tx;
346 struct xdp_frame *xdpf;
351 if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK))
354 if (unlikely(!netif_running(dev)))
357 i = smp_processor_id() % edev->total_xdp_queues;
358 xdp_tx = edev->fp_array[i].xdp_tx;
360 spin_lock(&xdp_tx->xdp_tx_lock);
362 for (i = 0; i < n_frames; i++) {
365 mapping = dma_map_single(dmadev, xdpf->data, xdpf->len,
367 if (unlikely(dma_mapping_error(dmadev, mapping))) {
368 xdp_return_frame_rx_napi(xdpf);
374 if (unlikely(qede_xdp_xmit(xdp_tx, mapping, 0, xdpf->len,
376 xdp_return_frame_rx_napi(xdpf);
381 if (flags & XDP_XMIT_FLUSH) {
382 xdp_prod = qed_chain_get_prod_idx(&xdp_tx->tx_pbl);
384 xdp_tx->tx_db.data.bd_prod = cpu_to_le16(xdp_prod);
385 qede_update_tx_producer(xdp_tx);
388 spin_unlock(&xdp_tx->xdp_tx_lock);
390 return n_frames - drops;
393 int qede_txq_has_work(struct qede_tx_queue *txq)
397 /* Tell compiler that consumer and producer can change */
399 hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr);
400 if (qed_chain_get_cons_idx(&txq->tx_pbl) == hw_bd_cons + 1)
403 return hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl);
406 static void qede_xdp_tx_int(struct qede_dev *edev, struct qede_tx_queue *txq)
408 struct sw_tx_xdp *xdp_info, *xdp_arr = txq->sw_tx_ring.xdp;
409 struct device *dev = &edev->pdev->dev;
410 struct xdp_frame *xdpf;
413 hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr);
416 while (hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl)) {
417 xdp_info = xdp_arr + txq->sw_tx_cons;
418 xdpf = xdp_info->xdpf;
421 dma_unmap_single(dev, xdp_info->mapping, xdpf->len,
423 xdp_return_frame(xdpf);
425 xdp_info->xdpf = NULL;
427 dma_unmap_page(dev, xdp_info->mapping, PAGE_SIZE,
429 __free_page(xdp_info->page);
432 qed_chain_consume(&txq->tx_pbl);
433 txq->sw_tx_cons = (txq->sw_tx_cons + 1) % txq->num_tx_buffers;
438 static int qede_tx_int(struct qede_dev *edev, struct qede_tx_queue *txq)
440 unsigned int pkts_compl = 0, bytes_compl = 0;
441 struct netdev_queue *netdev_txq;
445 netdev_txq = netdev_get_tx_queue(edev->ndev, txq->ndev_txq_id);
447 hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr);
450 while (hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl)) {
453 rc = qede_free_tx_pkt(edev, txq, &len);
455 DP_NOTICE(edev, "hw_bd_cons = %d, chain_cons=%d\n",
457 qed_chain_get_cons_idx(&txq->tx_pbl));
463 txq->sw_tx_cons = (txq->sw_tx_cons + 1) % txq->num_tx_buffers;
467 netdev_tx_completed_queue(netdev_txq, pkts_compl, bytes_compl);
469 /* Need to make the tx_bd_cons update visible to start_xmit()
470 * before checking for netif_tx_queue_stopped(). Without the
471 * memory barrier, there is a small possibility that
472 * start_xmit() will miss it and cause the queue to be stopped
474 * On the other hand we need an rmb() here to ensure the proper
475 * ordering of bit testing in the following
476 * netif_tx_queue_stopped(txq) call.
480 if (unlikely(netif_tx_queue_stopped(netdev_txq))) {
481 /* Taking tx_lock is needed to prevent reenabling the queue
482 * while it's empty. This could have happen if rx_action() gets
483 * suspended in qede_tx_int() after the condition before
484 * netif_tx_wake_queue(), while tx_action (qede_start_xmit()):
486 * stops the queue->sees fresh tx_bd_cons->releases the queue->
487 * sends some packets consuming the whole queue again->
491 __netif_tx_lock(netdev_txq, smp_processor_id());
493 if ((netif_tx_queue_stopped(netdev_txq)) &&
494 (edev->state == QEDE_STATE_OPEN) &&
495 (qed_chain_get_elem_left(&txq->tx_pbl)
496 >= (MAX_SKB_FRAGS + 1))) {
497 netif_tx_wake_queue(netdev_txq);
498 DP_VERBOSE(edev, NETIF_MSG_TX_DONE,
499 "Wake queue was called\n");
502 __netif_tx_unlock(netdev_txq);
508 bool qede_has_rx_work(struct qede_rx_queue *rxq)
510 u16 hw_comp_cons, sw_comp_cons;
512 /* Tell compiler that status block fields can change */
515 hw_comp_cons = le16_to_cpu(*rxq->hw_cons_ptr);
516 sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring);
518 return hw_comp_cons != sw_comp_cons;
521 static inline void qede_rx_bd_ring_consume(struct qede_rx_queue *rxq)
523 qed_chain_consume(&rxq->rx_bd_ring);
527 /* This function reuses the buffer(from an offset) from
528 * consumer index to producer index in the bd ring
530 static inline void qede_reuse_page(struct qede_rx_queue *rxq,
531 struct sw_rx_data *curr_cons)
533 struct eth_rx_bd *rx_bd_prod = qed_chain_produce(&rxq->rx_bd_ring);
534 struct sw_rx_data *curr_prod;
535 dma_addr_t new_mapping;
537 curr_prod = &rxq->sw_rx_ring[rxq->sw_rx_prod & NUM_RX_BDS_MAX];
538 *curr_prod = *curr_cons;
540 new_mapping = curr_prod->mapping + curr_prod->page_offset;
542 rx_bd_prod->addr.hi = cpu_to_le32(upper_32_bits(new_mapping));
543 rx_bd_prod->addr.lo = cpu_to_le32(lower_32_bits(new_mapping) +
547 curr_cons->data = NULL;
550 /* In case of allocation failures reuse buffers
551 * from consumer index to produce buffers for firmware
553 void qede_recycle_rx_bd_ring(struct qede_rx_queue *rxq, u8 count)
555 struct sw_rx_data *curr_cons;
557 for (; count > 0; count--) {
558 curr_cons = &rxq->sw_rx_ring[rxq->sw_rx_cons & NUM_RX_BDS_MAX];
559 qede_reuse_page(rxq, curr_cons);
560 qede_rx_bd_ring_consume(rxq);
564 static inline int qede_realloc_rx_buffer(struct qede_rx_queue *rxq,
565 struct sw_rx_data *curr_cons)
567 /* Move to the next segment in the page */
568 curr_cons->page_offset += rxq->rx_buf_seg_size;
570 if (curr_cons->page_offset == PAGE_SIZE) {
571 if (unlikely(qede_alloc_rx_buffer(rxq, true))) {
572 /* Since we failed to allocate new buffer
573 * current buffer can be used again.
575 curr_cons->page_offset -= rxq->rx_buf_seg_size;
580 dma_unmap_page(rxq->dev, curr_cons->mapping,
581 PAGE_SIZE, rxq->data_direction);
583 /* Increment refcount of the page as we don't want
584 * network stack to take the ownership of the page
585 * which can be recycled multiple times by the driver.
587 page_ref_inc(curr_cons->data);
588 qede_reuse_page(rxq, curr_cons);
594 void qede_update_rx_prod(struct qede_dev *edev, struct qede_rx_queue *rxq)
596 u16 bd_prod = qed_chain_get_prod_idx(&rxq->rx_bd_ring);
597 u16 cqe_prod = qed_chain_get_prod_idx(&rxq->rx_comp_ring);
598 struct eth_rx_prod_data rx_prods = {0};
600 /* Update producers */
601 rx_prods.bd_prod = cpu_to_le16(bd_prod);
602 rx_prods.cqe_prod = cpu_to_le16(cqe_prod);
604 /* Make sure that the BD and SGE data is updated before updating the
605 * producers since FW might read the BD/SGE right after the producer
610 internal_ram_wr(rxq->hw_rxq_prod_addr, sizeof(rx_prods),
614 static void qede_get_rxhash(struct sk_buff *skb, u8 bitfields, __le32 rss_hash)
616 enum pkt_hash_types hash_type = PKT_HASH_TYPE_NONE;
617 enum rss_hash_type htype;
620 htype = GET_FIELD(bitfields, ETH_FAST_PATH_RX_REG_CQE_RSS_HASH_TYPE);
622 hash_type = ((htype == RSS_HASH_TYPE_IPV4) ||
623 (htype == RSS_HASH_TYPE_IPV6)) ?
624 PKT_HASH_TYPE_L3 : PKT_HASH_TYPE_L4;
625 hash = le32_to_cpu(rss_hash);
627 skb_set_hash(skb, hash, hash_type);
630 static void qede_set_skb_csum(struct sk_buff *skb, u8 csum_flag)
632 skb_checksum_none_assert(skb);
634 if (csum_flag & QEDE_CSUM_UNNECESSARY)
635 skb->ip_summed = CHECKSUM_UNNECESSARY;
637 if (csum_flag & QEDE_TUNN_CSUM_UNNECESSARY) {
639 skb->encapsulation = 1;
643 static inline void qede_skb_receive(struct qede_dev *edev,
644 struct qede_fastpath *fp,
645 struct qede_rx_queue *rxq,
646 struct sk_buff *skb, u16 vlan_tag)
649 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tag);
651 napi_gro_receive(&fp->napi, skb);
654 static void qede_set_gro_params(struct qede_dev *edev,
656 struct eth_fast_path_rx_tpa_start_cqe *cqe)
658 u16 parsing_flags = le16_to_cpu(cqe->pars_flags.flags);
660 if (((parsing_flags >> PARSING_AND_ERR_FLAGS_L3TYPE_SHIFT) &
661 PARSING_AND_ERR_FLAGS_L3TYPE_MASK) == 2)
662 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
664 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
666 skb_shinfo(skb)->gso_size = __le16_to_cpu(cqe->len_on_first_bd) -
670 static int qede_fill_frag_skb(struct qede_dev *edev,
671 struct qede_rx_queue *rxq,
672 u8 tpa_agg_index, u16 len_on_bd)
674 struct sw_rx_data *current_bd = &rxq->sw_rx_ring[rxq->sw_rx_cons &
676 struct qede_agg_info *tpa_info = &rxq->tpa_info[tpa_agg_index];
677 struct sk_buff *skb = tpa_info->skb;
679 if (unlikely(tpa_info->state != QEDE_AGG_STATE_START))
682 /* Add one frag and update the appropriate fields in the skb */
683 skb_fill_page_desc(skb, tpa_info->frag_id++,
685 current_bd->page_offset + rxq->rx_headroom,
688 if (unlikely(qede_realloc_rx_buffer(rxq, current_bd))) {
689 /* Incr page ref count to reuse on allocation failure
690 * so that it doesn't get freed while freeing SKB.
692 page_ref_inc(current_bd->data);
696 qede_rx_bd_ring_consume(rxq);
698 skb->data_len += len_on_bd;
699 skb->truesize += rxq->rx_buf_seg_size;
700 skb->len += len_on_bd;
705 tpa_info->state = QEDE_AGG_STATE_ERROR;
706 qede_recycle_rx_bd_ring(rxq, 1);
711 static bool qede_tunn_exist(u16 flag)
713 return !!(flag & (PARSING_AND_ERR_FLAGS_TUNNELEXIST_MASK <<
714 PARSING_AND_ERR_FLAGS_TUNNELEXIST_SHIFT));
717 static u8 qede_check_tunn_csum(u16 flag)
722 if (flag & (PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMWASCALCULATED_MASK <<
723 PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMWASCALCULATED_SHIFT))
724 csum_flag |= PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMERROR_MASK <<
725 PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMERROR_SHIFT;
727 if (flag & (PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_MASK <<
728 PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_SHIFT)) {
729 csum_flag |= PARSING_AND_ERR_FLAGS_L4CHKSMERROR_MASK <<
730 PARSING_AND_ERR_FLAGS_L4CHKSMERROR_SHIFT;
731 tcsum = QEDE_TUNN_CSUM_UNNECESSARY;
734 csum_flag |= PARSING_AND_ERR_FLAGS_TUNNELIPHDRERROR_MASK <<
735 PARSING_AND_ERR_FLAGS_TUNNELIPHDRERROR_SHIFT |
736 PARSING_AND_ERR_FLAGS_IPHDRERROR_MASK <<
737 PARSING_AND_ERR_FLAGS_IPHDRERROR_SHIFT;
739 if (csum_flag & flag)
740 return QEDE_CSUM_ERROR;
742 return QEDE_CSUM_UNNECESSARY | tcsum;
745 static inline struct sk_buff *
746 qede_build_skb(struct qede_rx_queue *rxq,
747 struct sw_rx_data *bd, u16 len, u16 pad)
752 buf = page_address(bd->data) + bd->page_offset;
753 skb = build_skb(buf, rxq->rx_buf_seg_size);
758 skb_reserve(skb, pad);
764 static struct sk_buff *
765 qede_tpa_rx_build_skb(struct qede_dev *edev,
766 struct qede_rx_queue *rxq,
767 struct sw_rx_data *bd, u16 len, u16 pad,
772 skb = qede_build_skb(rxq, bd, len, pad);
773 bd->page_offset += rxq->rx_buf_seg_size;
775 if (bd->page_offset == PAGE_SIZE) {
776 if (unlikely(qede_alloc_rx_buffer(rxq, true))) {
778 "Failed to allocate RX buffer for tpa start\n");
779 bd->page_offset -= rxq->rx_buf_seg_size;
780 page_ref_inc(bd->data);
781 dev_kfree_skb_any(skb);
785 page_ref_inc(bd->data);
786 qede_reuse_page(rxq, bd);
789 /* We've consumed the first BD and prepared an SKB */
790 qede_rx_bd_ring_consume(rxq);
795 static struct sk_buff *
796 qede_rx_build_skb(struct qede_dev *edev,
797 struct qede_rx_queue *rxq,
798 struct sw_rx_data *bd, u16 len, u16 pad)
800 struct sk_buff *skb = NULL;
802 /* For smaller frames still need to allocate skb, memcpy
803 * data and benefit in reusing the page segment instead of
806 if ((len + pad <= edev->rx_copybreak)) {
807 unsigned int offset = bd->page_offset + pad;
809 skb = netdev_alloc_skb(edev->ndev, QEDE_RX_HDR_SIZE);
813 skb_reserve(skb, pad);
814 skb_put_data(skb, page_address(bd->data) + offset, len);
815 qede_reuse_page(rxq, bd);
819 skb = qede_build_skb(rxq, bd, len, pad);
821 if (unlikely(qede_realloc_rx_buffer(rxq, bd))) {
822 /* Incr page ref count to reuse on allocation failure so
823 * that it doesn't get freed while freeing SKB [as its
824 * already mapped there].
826 page_ref_inc(bd->data);
827 dev_kfree_skb_any(skb);
831 /* We've consumed the first BD and prepared an SKB */
832 qede_rx_bd_ring_consume(rxq);
837 static void qede_tpa_start(struct qede_dev *edev,
838 struct qede_rx_queue *rxq,
839 struct eth_fast_path_rx_tpa_start_cqe *cqe)
841 struct qede_agg_info *tpa_info = &rxq->tpa_info[cqe->tpa_agg_index];
842 struct sw_rx_data *sw_rx_data_cons;
845 sw_rx_data_cons = &rxq->sw_rx_ring[rxq->sw_rx_cons & NUM_RX_BDS_MAX];
846 pad = cqe->placement_offset + rxq->rx_headroom;
848 tpa_info->skb = qede_tpa_rx_build_skb(edev, rxq, sw_rx_data_cons,
849 le16_to_cpu(cqe->len_on_first_bd),
851 tpa_info->buffer.page_offset = sw_rx_data_cons->page_offset;
852 tpa_info->buffer.mapping = sw_rx_data_cons->mapping;
854 if (unlikely(!tpa_info->skb)) {
855 DP_NOTICE(edev, "Failed to allocate SKB for gro\n");
857 /* Consume from ring but do not produce since
858 * this might be used by FW still, it will be re-used
861 tpa_info->tpa_start_fail = true;
862 qede_rx_bd_ring_consume(rxq);
863 tpa_info->state = QEDE_AGG_STATE_ERROR;
867 tpa_info->frag_id = 0;
868 tpa_info->state = QEDE_AGG_STATE_START;
870 if ((le16_to_cpu(cqe->pars_flags.flags) >>
871 PARSING_AND_ERR_FLAGS_TAG8021QEXIST_SHIFT) &
872 PARSING_AND_ERR_FLAGS_TAG8021QEXIST_MASK)
873 tpa_info->vlan_tag = le16_to_cpu(cqe->vlan_tag);
875 tpa_info->vlan_tag = 0;
877 qede_get_rxhash(tpa_info->skb, cqe->bitfields, cqe->rss_hash);
879 /* This is needed in order to enable forwarding support */
880 qede_set_gro_params(edev, tpa_info->skb, cqe);
882 cons_buf: /* We still need to handle bd_len_list to consume buffers */
883 if (likely(cqe->bw_ext_bd_len_list[0]))
884 qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index,
885 le16_to_cpu(cqe->bw_ext_bd_len_list[0]));
887 if (unlikely(cqe->bw_ext_bd_len_list[1])) {
889 "Unlikely - got a TPA aggregation with more than one bw_ext_bd_len_list entry in the TPA start\n");
890 tpa_info->state = QEDE_AGG_STATE_ERROR;
895 static void qede_gro_ip_csum(struct sk_buff *skb)
897 const struct iphdr *iph = ip_hdr(skb);
900 skb_set_transport_header(skb, sizeof(struct iphdr));
903 th->check = ~tcp_v4_check(skb->len - skb_transport_offset(skb),
904 iph->saddr, iph->daddr, 0);
906 tcp_gro_complete(skb);
909 static void qede_gro_ipv6_csum(struct sk_buff *skb)
911 struct ipv6hdr *iph = ipv6_hdr(skb);
914 skb_set_transport_header(skb, sizeof(struct ipv6hdr));
917 th->check = ~tcp_v6_check(skb->len - skb_transport_offset(skb),
918 &iph->saddr, &iph->daddr, 0);
919 tcp_gro_complete(skb);
923 static void qede_gro_receive(struct qede_dev *edev,
924 struct qede_fastpath *fp,
928 /* FW can send a single MTU sized packet from gro flow
929 * due to aggregation timeout/last segment etc. which
930 * is not expected to be a gro packet. If a skb has zero
931 * frags then simply push it in the stack as non gso skb.
933 if (unlikely(!skb->data_len)) {
934 skb_shinfo(skb)->gso_type = 0;
935 skb_shinfo(skb)->gso_size = 0;
940 if (skb_shinfo(skb)->gso_size) {
941 skb_reset_network_header(skb);
943 switch (skb->protocol) {
944 case htons(ETH_P_IP):
945 qede_gro_ip_csum(skb);
947 case htons(ETH_P_IPV6):
948 qede_gro_ipv6_csum(skb);
952 "Error: FW GRO supports only IPv4/IPv6, not 0x%04x\n",
953 ntohs(skb->protocol));
959 skb_record_rx_queue(skb, fp->rxq->rxq_id);
960 qede_skb_receive(edev, fp, fp->rxq, skb, vlan_tag);
963 static inline void qede_tpa_cont(struct qede_dev *edev,
964 struct qede_rx_queue *rxq,
965 struct eth_fast_path_rx_tpa_cont_cqe *cqe)
969 for (i = 0; cqe->len_list[i]; i++)
970 qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index,
971 le16_to_cpu(cqe->len_list[i]));
975 "Strange - TPA cont with more than a single len_list entry\n");
978 static int qede_tpa_end(struct qede_dev *edev,
979 struct qede_fastpath *fp,
980 struct eth_fast_path_rx_tpa_end_cqe *cqe)
982 struct qede_rx_queue *rxq = fp->rxq;
983 struct qede_agg_info *tpa_info;
987 tpa_info = &rxq->tpa_info[cqe->tpa_agg_index];
990 if (tpa_info->buffer.page_offset == PAGE_SIZE)
991 dma_unmap_page(rxq->dev, tpa_info->buffer.mapping,
992 PAGE_SIZE, rxq->data_direction);
994 for (i = 0; cqe->len_list[i]; i++)
995 qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index,
996 le16_to_cpu(cqe->len_list[i]));
999 "Strange - TPA emd with more than a single len_list entry\n");
1001 if (unlikely(tpa_info->state != QEDE_AGG_STATE_START))
1005 if (unlikely(cqe->num_of_bds != tpa_info->frag_id + 1))
1007 "Strange - TPA had %02x BDs, but SKB has only %d frags\n",
1008 cqe->num_of_bds, tpa_info->frag_id);
1009 if (unlikely(skb->len != le16_to_cpu(cqe->total_packet_len)))
1011 "Strange - total packet len [cqe] is %4x but SKB has len %04x\n",
1012 le16_to_cpu(cqe->total_packet_len), skb->len);
1014 /* Finalize the SKB */
1015 skb->protocol = eth_type_trans(skb, edev->ndev);
1016 skb->ip_summed = CHECKSUM_UNNECESSARY;
1018 /* tcp_gro_complete() will copy NAPI_GRO_CB(skb)->count
1019 * to skb_shinfo(skb)->gso_segs
1021 NAPI_GRO_CB(skb)->count = le16_to_cpu(cqe->num_of_coalesced_segs);
1023 qede_gro_receive(edev, fp, skb, tpa_info->vlan_tag);
1025 tpa_info->state = QEDE_AGG_STATE_NONE;
1029 tpa_info->state = QEDE_AGG_STATE_NONE;
1031 if (tpa_info->tpa_start_fail) {
1032 qede_reuse_page(rxq, &tpa_info->buffer);
1033 tpa_info->tpa_start_fail = false;
1036 dev_kfree_skb_any(tpa_info->skb);
1037 tpa_info->skb = NULL;
1041 static u8 qede_check_notunn_csum(u16 flag)
1046 if (flag & (PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_MASK <<
1047 PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_SHIFT)) {
1048 csum_flag |= PARSING_AND_ERR_FLAGS_L4CHKSMERROR_MASK <<
1049 PARSING_AND_ERR_FLAGS_L4CHKSMERROR_SHIFT;
1050 csum = QEDE_CSUM_UNNECESSARY;
1053 csum_flag |= PARSING_AND_ERR_FLAGS_IPHDRERROR_MASK <<
1054 PARSING_AND_ERR_FLAGS_IPHDRERROR_SHIFT;
1056 if (csum_flag & flag)
1057 return QEDE_CSUM_ERROR;
1062 static u8 qede_check_csum(u16 flag)
1064 if (!qede_tunn_exist(flag))
1065 return qede_check_notunn_csum(flag);
1067 return qede_check_tunn_csum(flag);
1070 static bool qede_pkt_is_ip_fragmented(struct eth_fast_path_rx_reg_cqe *cqe,
1073 u8 tun_pars_flg = cqe->tunnel_pars_flags.flags;
1075 if ((tun_pars_flg & (ETH_TUNNEL_PARSING_FLAGS_IPV4_FRAGMENT_MASK <<
1076 ETH_TUNNEL_PARSING_FLAGS_IPV4_FRAGMENT_SHIFT)) ||
1077 (flag & (PARSING_AND_ERR_FLAGS_IPV4FRAG_MASK <<
1078 PARSING_AND_ERR_FLAGS_IPV4FRAG_SHIFT)))
1084 /* Return true iff packet is to be passed to stack */
1085 static bool qede_rx_xdp(struct qede_dev *edev,
1086 struct qede_fastpath *fp,
1087 struct qede_rx_queue *rxq,
1088 struct bpf_prog *prog,
1089 struct sw_rx_data *bd,
1090 struct eth_fast_path_rx_reg_cqe *cqe,
1091 u16 *data_offset, u16 *len)
1093 struct xdp_buff xdp;
1094 enum xdp_action act;
1096 xdp.data_hard_start = page_address(bd->data);
1097 xdp.data = xdp.data_hard_start + *data_offset;
1098 xdp_set_data_meta_invalid(&xdp);
1099 xdp.data_end = xdp.data + *len;
1100 xdp.rxq = &rxq->xdp_rxq;
1101 xdp.frame_sz = rxq->rx_buf_seg_size; /* PAGE_SIZE when XDP enabled */
1103 /* Queues always have a full reset currently, so for the time
1104 * being until there's atomic program replace just mark read
1105 * side for map helpers.
1108 act = bpf_prog_run_xdp(prog, &xdp);
1111 /* Recalculate, as XDP might have changed the headers */
1112 *data_offset = xdp.data - xdp.data_hard_start;
1113 *len = xdp.data_end - xdp.data;
1115 if (act == XDP_PASS)
1118 /* Count number of packets not to be passed to stack */
1123 /* We need the replacement buffer before transmit. */
1124 if (unlikely(qede_alloc_rx_buffer(rxq, true))) {
1125 qede_recycle_rx_bd_ring(rxq, 1);
1127 trace_xdp_exception(edev->ndev, prog, act);
1131 /* Now if there's a transmission problem, we'd still have to
1132 * throw current buffer, as replacement was already allocated.
1134 if (unlikely(qede_xdp_xmit(fp->xdp_tx, bd->mapping,
1135 *data_offset, *len, bd->data,
1137 dma_unmap_page(rxq->dev, bd->mapping, PAGE_SIZE,
1138 rxq->data_direction);
1139 __free_page(bd->data);
1141 trace_xdp_exception(edev->ndev, prog, act);
1143 dma_sync_single_for_device(rxq->dev,
1144 bd->mapping + *data_offset,
1145 *len, rxq->data_direction);
1146 fp->xdp_xmit |= QEDE_XDP_TX;
1149 /* Regardless, we've consumed an Rx BD */
1150 qede_rx_bd_ring_consume(rxq);
1153 /* We need the replacement buffer before transmit. */
1154 if (unlikely(qede_alloc_rx_buffer(rxq, true))) {
1155 qede_recycle_rx_bd_ring(rxq, 1);
1157 trace_xdp_exception(edev->ndev, prog, act);
1161 dma_unmap_page(rxq->dev, bd->mapping, PAGE_SIZE,
1162 rxq->data_direction);
1164 if (unlikely(xdp_do_redirect(edev->ndev, &xdp, prog)))
1165 DP_NOTICE(edev, "Failed to redirect the packet\n");
1167 fp->xdp_xmit |= QEDE_XDP_REDIRECT;
1169 qede_rx_bd_ring_consume(rxq);
1172 bpf_warn_invalid_xdp_action(act);
1175 trace_xdp_exception(edev->ndev, prog, act);
1178 qede_recycle_rx_bd_ring(rxq, cqe->bd_num);
1184 static int qede_rx_build_jumbo(struct qede_dev *edev,
1185 struct qede_rx_queue *rxq,
1186 struct sk_buff *skb,
1187 struct eth_fast_path_rx_reg_cqe *cqe,
1190 u16 pkt_len = le16_to_cpu(cqe->pkt_len);
1191 struct sw_rx_data *bd;
1195 pkt_len -= first_bd_len;
1197 /* We've already used one BD for the SKB. Now take care of the rest */
1198 for (num_frags = cqe->bd_num - 1; num_frags > 0; num_frags--) {
1199 u16 cur_size = pkt_len > rxq->rx_buf_size ? rxq->rx_buf_size :
1202 if (unlikely(!cur_size)) {
1204 "Still got %d BDs for mapping jumbo, but length became 0\n",
1209 /* We need a replacement buffer for each BD */
1210 if (unlikely(qede_alloc_rx_buffer(rxq, true)))
1213 /* Now that we've allocated the replacement buffer,
1214 * we can safely consume the next BD and map it to the SKB.
1216 bd_cons_idx = rxq->sw_rx_cons & NUM_RX_BDS_MAX;
1217 bd = &rxq->sw_rx_ring[bd_cons_idx];
1218 qede_rx_bd_ring_consume(rxq);
1220 dma_unmap_page(rxq->dev, bd->mapping,
1221 PAGE_SIZE, DMA_FROM_DEVICE);
1223 skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags++,
1224 bd->data, rxq->rx_headroom, cur_size);
1226 skb->truesize += PAGE_SIZE;
1227 skb->data_len += cur_size;
1228 skb->len += cur_size;
1229 pkt_len -= cur_size;
1232 if (unlikely(pkt_len))
1234 "Mapped all BDs of jumbo, but still have %d bytes\n",
1241 static int qede_rx_process_tpa_cqe(struct qede_dev *edev,
1242 struct qede_fastpath *fp,
1243 struct qede_rx_queue *rxq,
1244 union eth_rx_cqe *cqe,
1245 enum eth_rx_cqe_type type)
1248 case ETH_RX_CQE_TYPE_TPA_START:
1249 qede_tpa_start(edev, rxq, &cqe->fast_path_tpa_start);
1251 case ETH_RX_CQE_TYPE_TPA_CONT:
1252 qede_tpa_cont(edev, rxq, &cqe->fast_path_tpa_cont);
1254 case ETH_RX_CQE_TYPE_TPA_END:
1255 return qede_tpa_end(edev, fp, &cqe->fast_path_tpa_end);
1261 static int qede_rx_process_cqe(struct qede_dev *edev,
1262 struct qede_fastpath *fp,
1263 struct qede_rx_queue *rxq)
1265 struct bpf_prog *xdp_prog = READ_ONCE(rxq->xdp_prog);
1266 struct eth_fast_path_rx_reg_cqe *fp_cqe;
1267 u16 len, pad, bd_cons_idx, parse_flag;
1268 enum eth_rx_cqe_type cqe_type;
1269 union eth_rx_cqe *cqe;
1270 struct sw_rx_data *bd;
1271 struct sk_buff *skb;
1275 /* Get the CQE from the completion ring */
1276 cqe = (union eth_rx_cqe *)qed_chain_consume(&rxq->rx_comp_ring);
1277 cqe_type = cqe->fast_path_regular.type;
1279 /* Process an unlikely slowpath event */
1280 if (unlikely(cqe_type == ETH_RX_CQE_TYPE_SLOW_PATH)) {
1281 struct eth_slow_path_rx_cqe *sp_cqe;
1283 sp_cqe = (struct eth_slow_path_rx_cqe *)cqe;
1284 edev->ops->eth_cqe_completion(edev->cdev, fp->id, sp_cqe);
1288 /* Handle TPA cqes */
1289 if (cqe_type != ETH_RX_CQE_TYPE_REGULAR)
1290 return qede_rx_process_tpa_cqe(edev, fp, rxq, cqe, cqe_type);
1292 /* Get the data from the SW ring; Consume it only after it's evident
1293 * we wouldn't recycle it.
1295 bd_cons_idx = rxq->sw_rx_cons & NUM_RX_BDS_MAX;
1296 bd = &rxq->sw_rx_ring[bd_cons_idx];
1298 fp_cqe = &cqe->fast_path_regular;
1299 len = le16_to_cpu(fp_cqe->len_on_first_bd);
1300 pad = fp_cqe->placement_offset + rxq->rx_headroom;
1302 /* Run eBPF program if one is attached */
1304 if (!qede_rx_xdp(edev, fp, rxq, xdp_prog, bd, fp_cqe,
1308 /* If this is an error packet then drop it */
1309 flags = cqe->fast_path_regular.pars_flags.flags;
1310 parse_flag = le16_to_cpu(flags);
1312 csum_flag = qede_check_csum(parse_flag);
1313 if (unlikely(csum_flag == QEDE_CSUM_ERROR)) {
1314 if (qede_pkt_is_ip_fragmented(fp_cqe, parse_flag))
1317 rxq->rx_hw_errors++;
1320 /* Basic validation passed; Need to prepare an SKB. This would also
1321 * guarantee to finally consume the first BD upon success.
1323 skb = qede_rx_build_skb(edev, rxq, bd, len, pad);
1325 rxq->rx_alloc_errors++;
1326 qede_recycle_rx_bd_ring(rxq, fp_cqe->bd_num);
1330 /* In case of Jumbo packet, several PAGE_SIZEd buffers will be pointed
1333 if (fp_cqe->bd_num > 1) {
1334 u16 unmapped_frags = qede_rx_build_jumbo(edev, rxq, skb,
1337 if (unlikely(unmapped_frags > 0)) {
1338 qede_recycle_rx_bd_ring(rxq, unmapped_frags);
1339 dev_kfree_skb_any(skb);
1344 /* The SKB contains all the data. Now prepare meta-magic */
1345 skb->protocol = eth_type_trans(skb, edev->ndev);
1346 qede_get_rxhash(skb, fp_cqe->bitfields, fp_cqe->rss_hash);
1347 qede_set_skb_csum(skb, csum_flag);
1348 skb_record_rx_queue(skb, rxq->rxq_id);
1349 qede_ptp_record_rx_ts(edev, cqe, skb);
1351 /* SKB is prepared - pass it to stack */
1352 qede_skb_receive(edev, fp, rxq, skb, le16_to_cpu(fp_cqe->vlan_tag));
1357 static int qede_rx_int(struct qede_fastpath *fp, int budget)
1359 struct qede_rx_queue *rxq = fp->rxq;
1360 struct qede_dev *edev = fp->edev;
1361 int work_done = 0, rcv_pkts = 0;
1362 u16 hw_comp_cons, sw_comp_cons;
1364 hw_comp_cons = le16_to_cpu(*rxq->hw_cons_ptr);
1365 sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring);
1367 /* Memory barrier to prevent the CPU from doing speculative reads of CQE
1368 * / BD in the while-loop before reading hw_comp_cons. If the CQE is
1369 * read before it is written by FW, then FW writes CQE and SB, and then
1370 * the CPU reads the hw_comp_cons, it will use an old CQE.
1374 /* Loop to complete all indicated BDs */
1375 while ((sw_comp_cons != hw_comp_cons) && (work_done < budget)) {
1376 rcv_pkts += qede_rx_process_cqe(edev, fp, rxq);
1377 qed_chain_recycle_consumed(&rxq->rx_comp_ring);
1378 sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring);
1382 rxq->rcv_pkts += rcv_pkts;
1384 /* Allocate replacement buffers */
1385 while (rxq->num_rx_buffers - rxq->filled_buffers)
1386 if (qede_alloc_rx_buffer(rxq, false))
1389 /* Update producers */
1390 qede_update_rx_prod(edev, rxq);
1395 static bool qede_poll_is_more_work(struct qede_fastpath *fp)
1397 qed_sb_update_sb_idx(fp->sb_info);
1399 /* *_has_*_work() reads the status block, thus we need to ensure that
1400 * status block indices have been actually read (qed_sb_update_sb_idx)
1401 * prior to this check (*_has_*_work) so that we won't write the
1402 * "newer" value of the status block to HW (if there was a DMA right
1403 * after qede_has_rx_work and if there is no rmb, the memory reading
1404 * (qed_sb_update_sb_idx) may be postponed to right before *_ack_sb).
1405 * In this case there will never be another interrupt until there is
1406 * another update of the status block, while there is still unhandled
1411 if (likely(fp->type & QEDE_FASTPATH_RX))
1412 if (qede_has_rx_work(fp->rxq))
1415 if (fp->type & QEDE_FASTPATH_XDP)
1416 if (qede_txq_has_work(fp->xdp_tx))
1419 if (likely(fp->type & QEDE_FASTPATH_TX)) {
1422 for_each_cos_in_txq(fp->edev, cos) {
1423 if (qede_txq_has_work(&fp->txq[cos]))
1431 /*********************
1432 * NDO & API related *
1433 *********************/
1434 int qede_poll(struct napi_struct *napi, int budget)
1436 struct qede_fastpath *fp = container_of(napi, struct qede_fastpath,
1438 struct qede_dev *edev = fp->edev;
1439 int rx_work_done = 0;
1444 if (likely(fp->type & QEDE_FASTPATH_TX)) {
1447 for_each_cos_in_txq(fp->edev, cos) {
1448 if (qede_txq_has_work(&fp->txq[cos]))
1449 qede_tx_int(edev, &fp->txq[cos]);
1453 if ((fp->type & QEDE_FASTPATH_XDP) && qede_txq_has_work(fp->xdp_tx))
1454 qede_xdp_tx_int(edev, fp->xdp_tx);
1456 rx_work_done = (likely(fp->type & QEDE_FASTPATH_RX) &&
1457 qede_has_rx_work(fp->rxq)) ?
1458 qede_rx_int(fp, budget) : 0;
1460 if (fp->xdp_xmit & QEDE_XDP_REDIRECT)
1463 /* Handle case where we are called by netpoll with a budget of 0 */
1464 if (rx_work_done < budget || !budget) {
1465 if (!qede_poll_is_more_work(fp)) {
1466 napi_complete_done(napi, rx_work_done);
1468 /* Update and reenable interrupts */
1469 qed_sb_ack(fp->sb_info, IGU_INT_ENABLE, 1);
1471 rx_work_done = budget;
1475 if (fp->xdp_xmit & QEDE_XDP_TX) {
1476 xdp_prod = qed_chain_get_prod_idx(&fp->xdp_tx->tx_pbl);
1478 fp->xdp_tx->tx_db.data.bd_prod = cpu_to_le16(xdp_prod);
1479 qede_update_tx_producer(fp->xdp_tx);
1482 return rx_work_done;
1485 irqreturn_t qede_msix_fp_int(int irq, void *fp_cookie)
1487 struct qede_fastpath *fp = fp_cookie;
1489 qed_sb_ack(fp->sb_info, IGU_INT_DISABLE, 0 /*do not update*/);
1491 napi_schedule_irqoff(&fp->napi);
1495 /* Main transmit function */
1496 netdev_tx_t qede_start_xmit(struct sk_buff *skb, struct net_device *ndev)
1498 struct qede_dev *edev = netdev_priv(ndev);
1499 struct netdev_queue *netdev_txq;
1500 struct qede_tx_queue *txq;
1501 struct eth_tx_1st_bd *first_bd;
1502 struct eth_tx_2nd_bd *second_bd = NULL;
1503 struct eth_tx_3rd_bd *third_bd = NULL;
1504 struct eth_tx_bd *tx_data_bd = NULL;
1505 u16 txq_index, val = 0;
1508 int rc, frag_idx = 0, ipv6_ext = 0;
1512 bool data_split = false;
1514 /* Get tx-queue context and netdev index */
1515 txq_index = skb_get_queue_mapping(skb);
1516 WARN_ON(txq_index >= QEDE_TSS_COUNT(edev) * edev->dev_info.num_tc);
1517 txq = QEDE_NDEV_TXQ_ID_TO_TXQ(edev, txq_index);
1518 netdev_txq = netdev_get_tx_queue(ndev, txq_index);
1520 WARN_ON(qed_chain_get_elem_left(&txq->tx_pbl) < (MAX_SKB_FRAGS + 1));
1522 xmit_type = qede_xmit_type(skb, &ipv6_ext);
1524 #if ((MAX_SKB_FRAGS + 2) > ETH_TX_MAX_BDS_PER_NON_LSO_PACKET)
1525 if (qede_pkt_req_lin(skb, xmit_type)) {
1526 if (skb_linearize(skb)) {
1527 txq->tx_mem_alloc_err++;
1529 dev_kfree_skb_any(skb);
1530 return NETDEV_TX_OK;
1535 /* Fill the entry in the SW ring and the BDs in the FW ring */
1536 idx = txq->sw_tx_prod;
1537 txq->sw_tx_ring.skbs[idx].skb = skb;
1538 first_bd = (struct eth_tx_1st_bd *)
1539 qed_chain_produce(&txq->tx_pbl);
1540 memset(first_bd, 0, sizeof(*first_bd));
1541 first_bd->data.bd_flags.bitfields =
1542 1 << ETH_TX_1ST_BD_FLAGS_START_BD_SHIFT;
1544 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP))
1545 qede_ptp_tx_ts(edev, skb);
1547 /* Map skb linear data for DMA and set in the first BD */
1548 mapping = dma_map_single(txq->dev, skb->data,
1549 skb_headlen(skb), DMA_TO_DEVICE);
1550 if (unlikely(dma_mapping_error(txq->dev, mapping))) {
1551 DP_NOTICE(edev, "SKB mapping failed\n");
1552 qede_free_failed_tx_pkt(txq, first_bd, 0, false);
1553 qede_update_tx_producer(txq);
1554 return NETDEV_TX_OK;
1557 BD_SET_UNMAP_ADDR_LEN(first_bd, mapping, skb_headlen(skb));
1559 /* In case there is IPv6 with extension headers or LSO we need 2nd and
1562 if (unlikely((xmit_type & XMIT_LSO) | ipv6_ext)) {
1563 second_bd = (struct eth_tx_2nd_bd *)
1564 qed_chain_produce(&txq->tx_pbl);
1565 memset(second_bd, 0, sizeof(*second_bd));
1568 third_bd = (struct eth_tx_3rd_bd *)
1569 qed_chain_produce(&txq->tx_pbl);
1570 memset(third_bd, 0, sizeof(*third_bd));
1573 /* We need to fill in additional data in second_bd... */
1574 tx_data_bd = (struct eth_tx_bd *)second_bd;
1577 if (skb_vlan_tag_present(skb)) {
1578 first_bd->data.vlan = cpu_to_le16(skb_vlan_tag_get(skb));
1579 first_bd->data.bd_flags.bitfields |=
1580 1 << ETH_TX_1ST_BD_FLAGS_VLAN_INSERTION_SHIFT;
1583 /* Fill the parsing flags & params according to the requested offload */
1584 if (xmit_type & XMIT_L4_CSUM) {
1585 /* We don't re-calculate IP checksum as it is already done by
1588 first_bd->data.bd_flags.bitfields |=
1589 1 << ETH_TX_1ST_BD_FLAGS_L4_CSUM_SHIFT;
1591 if (xmit_type & XMIT_ENC) {
1592 first_bd->data.bd_flags.bitfields |=
1593 1 << ETH_TX_1ST_BD_FLAGS_IP_CSUM_SHIFT;
1595 val |= (1 << ETH_TX_DATA_1ST_BD_TUNN_FLAG_SHIFT);
1598 /* Legacy FW had flipped behavior in regard to this bit -
1599 * I.e., needed to set to prevent FW from touching encapsulated
1600 * packets when it didn't need to.
1602 if (unlikely(txq->is_legacy))
1603 val ^= (1 << ETH_TX_DATA_1ST_BD_TUNN_FLAG_SHIFT);
1605 /* If the packet is IPv6 with extension header, indicate that
1606 * to FW and pass few params, since the device cracker doesn't
1607 * support parsing IPv6 with extension header/s.
1609 if (unlikely(ipv6_ext))
1610 qede_set_params_for_ipv6_ext(skb, second_bd, third_bd);
1613 if (xmit_type & XMIT_LSO) {
1614 first_bd->data.bd_flags.bitfields |=
1615 (1 << ETH_TX_1ST_BD_FLAGS_LSO_SHIFT);
1616 third_bd->data.lso_mss =
1617 cpu_to_le16(skb_shinfo(skb)->gso_size);
1619 if (unlikely(xmit_type & XMIT_ENC)) {
1620 first_bd->data.bd_flags.bitfields |=
1621 1 << ETH_TX_1ST_BD_FLAGS_TUNN_IP_CSUM_SHIFT;
1623 if (xmit_type & XMIT_ENC_GSO_L4_CSUM) {
1624 u8 tmp = ETH_TX_1ST_BD_FLAGS_TUNN_L4_CSUM_SHIFT;
1626 first_bd->data.bd_flags.bitfields |= 1 << tmp;
1628 hlen = qede_get_skb_hlen(skb, true);
1630 first_bd->data.bd_flags.bitfields |=
1631 1 << ETH_TX_1ST_BD_FLAGS_IP_CSUM_SHIFT;
1632 hlen = qede_get_skb_hlen(skb, false);
1635 /* @@@TBD - if will not be removed need to check */
1636 third_bd->data.bitfields |=
1637 cpu_to_le16(1 << ETH_TX_DATA_3RD_BD_HDR_NBD_SHIFT);
1639 /* Make life easier for FW guys who can't deal with header and
1640 * data on same BD. If we need to split, use the second bd...
1642 if (unlikely(skb_headlen(skb) > hlen)) {
1643 DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED,
1644 "TSO split header size is %d (%x:%x)\n",
1645 first_bd->nbytes, first_bd->addr.hi,
1648 mapping = HILO_U64(le32_to_cpu(first_bd->addr.hi),
1649 le32_to_cpu(first_bd->addr.lo)) +
1652 BD_SET_UNMAP_ADDR_LEN(tx_data_bd, mapping,
1653 le16_to_cpu(first_bd->nbytes) -
1656 /* this marks the BD as one that has no
1657 * individual mapping
1659 txq->sw_tx_ring.skbs[idx].flags |= QEDE_TSO_SPLIT_BD;
1661 first_bd->nbytes = cpu_to_le16(hlen);
1663 tx_data_bd = (struct eth_tx_bd *)third_bd;
1667 if (unlikely(skb->len > ETH_TX_MAX_NON_LSO_PKT_LEN)) {
1668 DP_ERR(edev, "Unexpected non LSO skb length = 0x%x\n", skb->len);
1669 qede_free_failed_tx_pkt(txq, first_bd, 0, false);
1670 qede_update_tx_producer(txq);
1671 return NETDEV_TX_OK;
1674 val |= ((skb->len & ETH_TX_DATA_1ST_BD_PKT_LEN_MASK) <<
1675 ETH_TX_DATA_1ST_BD_PKT_LEN_SHIFT);
1678 first_bd->data.bitfields = cpu_to_le16(val);
1680 /* Handle fragmented skb */
1681 /* special handle for frags inside 2nd and 3rd bds.. */
1682 while (tx_data_bd && frag_idx < skb_shinfo(skb)->nr_frags) {
1683 rc = map_frag_to_bd(txq,
1684 &skb_shinfo(skb)->frags[frag_idx],
1687 qede_free_failed_tx_pkt(txq, first_bd, nbd, data_split);
1688 qede_update_tx_producer(txq);
1689 return NETDEV_TX_OK;
1692 if (tx_data_bd == (struct eth_tx_bd *)second_bd)
1693 tx_data_bd = (struct eth_tx_bd *)third_bd;
1700 /* map last frags into 4th, 5th .... */
1701 for (; frag_idx < skb_shinfo(skb)->nr_frags; frag_idx++, nbd++) {
1702 tx_data_bd = (struct eth_tx_bd *)
1703 qed_chain_produce(&txq->tx_pbl);
1705 memset(tx_data_bd, 0, sizeof(*tx_data_bd));
1707 rc = map_frag_to_bd(txq,
1708 &skb_shinfo(skb)->frags[frag_idx],
1711 qede_free_failed_tx_pkt(txq, first_bd, nbd, data_split);
1712 qede_update_tx_producer(txq);
1713 return NETDEV_TX_OK;
1717 /* update the first BD with the actual num BDs */
1718 first_bd->data.nbds = nbd;
1720 netdev_tx_sent_queue(netdev_txq, skb->len);
1722 skb_tx_timestamp(skb);
1724 /* Advance packet producer only before sending the packet since mapping
1725 * of pages may fail.
1727 txq->sw_tx_prod = (txq->sw_tx_prod + 1) % txq->num_tx_buffers;
1729 /* 'next page' entries are counted in the producer value */
1730 txq->tx_db.data.bd_prod =
1731 cpu_to_le16(qed_chain_get_prod_idx(&txq->tx_pbl));
1733 if (!netdev_xmit_more() || netif_xmit_stopped(netdev_txq))
1734 qede_update_tx_producer(txq);
1736 if (unlikely(qed_chain_get_elem_left(&txq->tx_pbl)
1737 < (MAX_SKB_FRAGS + 1))) {
1738 if (netdev_xmit_more())
1739 qede_update_tx_producer(txq);
1741 netif_tx_stop_queue(netdev_txq);
1743 DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED,
1744 "Stop queue was called\n");
1745 /* paired memory barrier is in qede_tx_int(), we have to keep
1746 * ordering of set_bit() in netif_tx_stop_queue() and read of
1751 if ((qed_chain_get_elem_left(&txq->tx_pbl) >=
1752 (MAX_SKB_FRAGS + 1)) &&
1753 (edev->state == QEDE_STATE_OPEN)) {
1754 netif_tx_wake_queue(netdev_txq);
1755 DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED,
1756 "Wake queue was called\n");
1760 return NETDEV_TX_OK;
1763 u16 qede_select_queue(struct net_device *dev, struct sk_buff *skb,
1764 struct net_device *sb_dev)
1766 struct qede_dev *edev = netdev_priv(dev);
1769 total_txq = QEDE_TSS_COUNT(edev) * edev->dev_info.num_tc;
1771 return QEDE_TSS_COUNT(edev) ?
1772 netdev_pick_tx(dev, skb, NULL) % total_txq : 0;
1775 /* 8B udp header + 8B base tunnel header + 32B option length */
1776 #define QEDE_MAX_TUN_HDR_LEN 48
1778 netdev_features_t qede_features_check(struct sk_buff *skb,
1779 struct net_device *dev,
1780 netdev_features_t features)
1782 if (skb->encapsulation) {
1785 switch (vlan_get_protocol(skb)) {
1786 case htons(ETH_P_IP):
1787 l4_proto = ip_hdr(skb)->protocol;
1789 case htons(ETH_P_IPV6):
1790 l4_proto = ipv6_hdr(skb)->nexthdr;
1796 /* Disable offloads for geneve tunnels, as HW can't parse
1797 * the geneve header which has option length greater than 32b
1798 * and disable offloads for the ports which are not offloaded.
1800 if (l4_proto == IPPROTO_UDP) {
1801 struct qede_dev *edev = netdev_priv(dev);
1802 u16 hdrlen, vxln_port, gnv_port;
1804 hdrlen = QEDE_MAX_TUN_HDR_LEN;
1805 vxln_port = edev->vxlan_dst_port;
1806 gnv_port = edev->geneve_dst_port;
1808 if ((skb_inner_mac_header(skb) -
1809 skb_transport_header(skb)) > hdrlen ||
1810 (ntohs(udp_hdr(skb)->dest) != vxln_port &&
1811 ntohs(udp_hdr(skb)->dest) != gnv_port))
1812 return features & ~(NETIF_F_CSUM_MASK |
1814 } else if (l4_proto == IPPROTO_IPIP) {
1815 /* IPIP tunnels are unknown to the device or at least unsupported natively,
1816 * offloads for them can't be done trivially, so disable them for such skb.
1818 return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);