1806 lines · c
1// SPDX-License-Identifier: (GPL-2.0-only OR BSD-3-Clause)2/* QLogic qede NIC Driver3 * Copyright (c) 2015-2017 QLogic Corporation4 * Copyright (c) 2019-2020 Marvell International Ltd.5 */6 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>12#include <linux/ip.h>13#include <net/gro.h>14#include <net/ipv6.h>15#include <net/tcp.h>16#include <linux/if_ether.h>17#include <linux/if_vlan.h>18#include <net/ip6_checksum.h>19#include "qede_ptp.h"20 21#include <linux/qed/qed_if.h>22#include "qede.h"23/*********************************24 * Content also used by slowpath *25 *********************************/26 27int qede_alloc_rx_buffer(struct qede_rx_queue *rxq, bool allow_lazy)28{29 struct sw_rx_data *sw_rx_data;30 struct eth_rx_bd *rx_bd;31 dma_addr_t mapping;32 struct page *data;33 34 /* In case lazy-allocation is allowed, postpone allocation until the35 * end of the NAPI run. We'd still need to make sure the Rx ring has36 * sufficient buffers to guarantee an additional Rx interrupt.37 */38 if (allow_lazy && likely(rxq->filled_buffers > 12)) {39 rxq->filled_buffers--;40 return 0;41 }42 43 data = alloc_pages(GFP_ATOMIC, 0);44 if (unlikely(!data))45 return -ENOMEM;46 47 /* Map the entire page as it would be used48 * for multiple RX buffer segment size mapping.49 */50 mapping = dma_map_page(rxq->dev, data, 0,51 PAGE_SIZE, rxq->data_direction);52 if (unlikely(dma_mapping_error(rxq->dev, mapping))) {53 __free_page(data);54 return -ENOMEM;55 }56 57 sw_rx_data = &rxq->sw_rx_ring[rxq->sw_rx_prod & NUM_RX_BDS_MAX];58 sw_rx_data->page_offset = 0;59 sw_rx_data->data = data;60 sw_rx_data->mapping = mapping;61 62 /* Advance PROD and get BD pointer */63 rx_bd = (struct eth_rx_bd *)qed_chain_produce(&rxq->rx_bd_ring);64 WARN_ON(!rx_bd);65 rx_bd->addr.hi = cpu_to_le32(upper_32_bits(mapping));66 rx_bd->addr.lo = cpu_to_le32(lower_32_bits(mapping) +67 rxq->rx_headroom);68 69 rxq->sw_rx_prod++;70 rxq->filled_buffers++;71 72 return 0;73}74 75/* Unmap the data and free skb */76int qede_free_tx_pkt(struct qede_dev *edev, struct qede_tx_queue *txq, int *len)77{78 u16 idx = txq->sw_tx_cons;79 struct sk_buff *skb = txq->sw_tx_ring.skbs[idx].skb;80 struct eth_tx_1st_bd *first_bd;81 struct eth_tx_bd *tx_data_bd;82 int bds_consumed = 0;83 int nbds;84 bool data_split = txq->sw_tx_ring.skbs[idx].flags & QEDE_TSO_SPLIT_BD;85 int i, split_bd_len = 0;86 87 if (unlikely(!skb)) {88 DP_ERR(edev,89 "skb is null for txq idx=%d txq->sw_tx_cons=%d txq->sw_tx_prod=%d\n",90 idx, txq->sw_tx_cons, txq->sw_tx_prod);91 return -1;92 }93 94 *len = skb->len;95 96 first_bd = (struct eth_tx_1st_bd *)qed_chain_consume(&txq->tx_pbl);97 98 bds_consumed++;99 100 nbds = first_bd->data.nbds;101 102 if (data_split) {103 struct eth_tx_bd *split = (struct eth_tx_bd *)104 qed_chain_consume(&txq->tx_pbl);105 split_bd_len = BD_UNMAP_LEN(split);106 bds_consumed++;107 }108 dma_unmap_single(&edev->pdev->dev, BD_UNMAP_ADDR(first_bd),109 BD_UNMAP_LEN(first_bd) + split_bd_len, DMA_TO_DEVICE);110 111 /* Unmap the data of the skb frags */112 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++, bds_consumed++) {113 tx_data_bd = (struct eth_tx_bd *)114 qed_chain_consume(&txq->tx_pbl);115 dma_unmap_page(&edev->pdev->dev, BD_UNMAP_ADDR(tx_data_bd),116 BD_UNMAP_LEN(tx_data_bd), DMA_TO_DEVICE);117 }118 119 while (bds_consumed++ < nbds)120 qed_chain_consume(&txq->tx_pbl);121 122 /* Free skb */123 dev_kfree_skb_any(skb);124 txq->sw_tx_ring.skbs[idx].skb = NULL;125 txq->sw_tx_ring.skbs[idx].flags = 0;126 127 return 0;128}129 130/* Unmap the data and free skb when mapping failed during start_xmit */131static void qede_free_failed_tx_pkt(struct qede_tx_queue *txq,132 struct eth_tx_1st_bd *first_bd,133 int nbd, bool data_split)134{135 u16 idx = txq->sw_tx_prod;136 struct sk_buff *skb = txq->sw_tx_ring.skbs[idx].skb;137 struct eth_tx_bd *tx_data_bd;138 int i, split_bd_len = 0;139 140 /* Return prod to its position before this skb was handled */141 qed_chain_set_prod(&txq->tx_pbl,142 le16_to_cpu(txq->tx_db.data.bd_prod), first_bd);143 144 first_bd = (struct eth_tx_1st_bd *)qed_chain_produce(&txq->tx_pbl);145 146 if (data_split) {147 struct eth_tx_bd *split = (struct eth_tx_bd *)148 qed_chain_produce(&txq->tx_pbl);149 split_bd_len = BD_UNMAP_LEN(split);150 nbd--;151 }152 153 dma_unmap_single(txq->dev, BD_UNMAP_ADDR(first_bd),154 BD_UNMAP_LEN(first_bd) + split_bd_len, DMA_TO_DEVICE);155 156 /* Unmap the data of the skb frags */157 for (i = 0; i < nbd; i++) {158 tx_data_bd = (struct eth_tx_bd *)159 qed_chain_produce(&txq->tx_pbl);160 if (tx_data_bd->nbytes)161 dma_unmap_page(txq->dev,162 BD_UNMAP_ADDR(tx_data_bd),163 BD_UNMAP_LEN(tx_data_bd), DMA_TO_DEVICE);164 }165 166 /* Return again prod to its position before this skb was handled */167 qed_chain_set_prod(&txq->tx_pbl,168 le16_to_cpu(txq->tx_db.data.bd_prod), first_bd);169 170 /* Free skb */171 dev_kfree_skb_any(skb);172 txq->sw_tx_ring.skbs[idx].skb = NULL;173 txq->sw_tx_ring.skbs[idx].flags = 0;174}175 176static u32 qede_xmit_type(struct sk_buff *skb, int *ipv6_ext)177{178 u32 rc = XMIT_L4_CSUM;179 __be16 l3_proto;180 181 if (skb->ip_summed != CHECKSUM_PARTIAL)182 return XMIT_PLAIN;183 184 l3_proto = vlan_get_protocol(skb);185 if (l3_proto == htons(ETH_P_IPV6) &&186 (ipv6_hdr(skb)->nexthdr == NEXTHDR_IPV6))187 *ipv6_ext = 1;188 189 if (skb->encapsulation) {190 rc |= XMIT_ENC;191 if (skb_is_gso(skb)) {192 unsigned short gso_type = skb_shinfo(skb)->gso_type;193 194 if ((gso_type & SKB_GSO_UDP_TUNNEL_CSUM) ||195 (gso_type & SKB_GSO_GRE_CSUM))196 rc |= XMIT_ENC_GSO_L4_CSUM;197 198 rc |= XMIT_LSO;199 return rc;200 }201 }202 203 if (skb_is_gso(skb))204 rc |= XMIT_LSO;205 206 return rc;207}208 209static void qede_set_params_for_ipv6_ext(struct sk_buff *skb,210 struct eth_tx_2nd_bd *second_bd,211 struct eth_tx_3rd_bd *third_bd)212{213 u8 l4_proto;214 u16 bd2_bits1 = 0, bd2_bits2 = 0;215 216 bd2_bits1 |= (1 << ETH_TX_DATA_2ND_BD_IPV6_EXT_SHIFT);217 218 bd2_bits2 |= ((skb_transport_offset(skb) >> 1) &219 ETH_TX_DATA_2ND_BD_L4_HDR_START_OFFSET_W_MASK)220 << ETH_TX_DATA_2ND_BD_L4_HDR_START_OFFSET_W_SHIFT;221 222 bd2_bits1 |= (ETH_L4_PSEUDO_CSUM_CORRECT_LENGTH <<223 ETH_TX_DATA_2ND_BD_L4_PSEUDO_CSUM_MODE_SHIFT);224 225 if (vlan_get_protocol(skb) == htons(ETH_P_IPV6))226 l4_proto = ipv6_hdr(skb)->nexthdr;227 else228 l4_proto = ip_hdr(skb)->protocol;229 230 if (l4_proto == IPPROTO_UDP)231 bd2_bits1 |= 1 << ETH_TX_DATA_2ND_BD_L4_UDP_SHIFT;232 233 if (third_bd)234 third_bd->data.bitfields |=235 cpu_to_le16(((tcp_hdrlen(skb) / 4) &236 ETH_TX_DATA_3RD_BD_TCP_HDR_LEN_DW_MASK) <<237 ETH_TX_DATA_3RD_BD_TCP_HDR_LEN_DW_SHIFT);238 239 second_bd->data.bitfields1 = cpu_to_le16(bd2_bits1);240 second_bd->data.bitfields2 = cpu_to_le16(bd2_bits2);241}242 243static int map_frag_to_bd(struct qede_tx_queue *txq,244 skb_frag_t *frag, struct eth_tx_bd *bd)245{246 dma_addr_t mapping;247 248 /* Map skb non-linear frag data for DMA */249 mapping = skb_frag_dma_map(txq->dev, frag, 0,250 skb_frag_size(frag), DMA_TO_DEVICE);251 if (unlikely(dma_mapping_error(txq->dev, mapping)))252 return -ENOMEM;253 254 /* Setup the data pointer of the frag data */255 BD_SET_UNMAP_ADDR_LEN(bd, mapping, skb_frag_size(frag));256 257 return 0;258}259 260static u16 qede_get_skb_hlen(struct sk_buff *skb, bool is_encap_pkt)261{262 if (is_encap_pkt)263 return skb_inner_tcp_all_headers(skb);264 265 return skb_tcp_all_headers(skb);266}267 268/* +2 for 1st BD for headers and 2nd BD for headlen (if required) */269#if ((MAX_SKB_FRAGS + 2) > ETH_TX_MAX_BDS_PER_NON_LSO_PACKET)270static bool qede_pkt_req_lin(struct sk_buff *skb, u8 xmit_type)271{272 int allowed_frags = ETH_TX_MAX_BDS_PER_NON_LSO_PACKET - 1;273 274 if (xmit_type & XMIT_LSO) {275 int hlen;276 277 hlen = qede_get_skb_hlen(skb, xmit_type & XMIT_ENC);278 279 /* linear payload would require its own BD */280 if (skb_headlen(skb) > hlen)281 allowed_frags--;282 }283 284 return (skb_shinfo(skb)->nr_frags > allowed_frags);285}286#endif287 288static inline void qede_update_tx_producer(struct qede_tx_queue *txq)289{290 /* wmb makes sure that the BDs data is updated before updating the291 * producer, otherwise FW may read old data from the BDs.292 */293 wmb();294 barrier();295 writel(txq->tx_db.raw, txq->doorbell_addr);296 297 /* Fence required to flush the write combined buffer, since another298 * CPU may write to the same doorbell address and data may be lost299 * due to relaxed order nature of write combined bar.300 */301 wmb();302}303 304static int qede_xdp_xmit(struct qede_tx_queue *txq, dma_addr_t dma, u16 pad,305 u16 len, struct page *page, struct xdp_frame *xdpf)306{307 struct eth_tx_1st_bd *bd;308 struct sw_tx_xdp *xdp;309 u16 val;310 311 if (unlikely(qed_chain_get_elem_used(&txq->tx_pbl) >=312 txq->num_tx_buffers)) {313 txq->stopped_cnt++;314 return -ENOMEM;315 }316 317 bd = qed_chain_produce(&txq->tx_pbl);318 bd->data.nbds = 1;319 bd->data.bd_flags.bitfields = BIT(ETH_TX_1ST_BD_FLAGS_START_BD_SHIFT);320 321 val = (len & ETH_TX_DATA_1ST_BD_PKT_LEN_MASK) <<322 ETH_TX_DATA_1ST_BD_PKT_LEN_SHIFT;323 324 bd->data.bitfields = cpu_to_le16(val);325 326 /* We can safely ignore the offset, as it's 0 for XDP */327 BD_SET_UNMAP_ADDR_LEN(bd, dma + pad, len);328 329 xdp = txq->sw_tx_ring.xdp + txq->sw_tx_prod;330 xdp->mapping = dma;331 xdp->page = page;332 xdp->xdpf = xdpf;333 334 txq->sw_tx_prod = (txq->sw_tx_prod + 1) % txq->num_tx_buffers;335 336 return 0;337}338 339int qede_xdp_transmit(struct net_device *dev, int n_frames,340 struct xdp_frame **frames, u32 flags)341{342 struct qede_dev *edev = netdev_priv(dev);343 struct device *dmadev = &edev->pdev->dev;344 struct qede_tx_queue *xdp_tx;345 struct xdp_frame *xdpf;346 dma_addr_t mapping;347 int i, nxmit = 0;348 u16 xdp_prod;349 350 if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK))351 return -EINVAL;352 353 if (unlikely(!netif_running(dev)))354 return -ENETDOWN;355 356 i = smp_processor_id() % edev->total_xdp_queues;357 xdp_tx = edev->fp_array[i].xdp_tx;358 359 spin_lock(&xdp_tx->xdp_tx_lock);360 361 for (i = 0; i < n_frames; i++) {362 xdpf = frames[i];363 364 mapping = dma_map_single(dmadev, xdpf->data, xdpf->len,365 DMA_TO_DEVICE);366 if (unlikely(dma_mapping_error(dmadev, mapping)))367 break;368 369 if (unlikely(qede_xdp_xmit(xdp_tx, mapping, 0, xdpf->len,370 NULL, xdpf)))371 break;372 nxmit++;373 }374 375 if (flags & XDP_XMIT_FLUSH) {376 xdp_prod = qed_chain_get_prod_idx(&xdp_tx->tx_pbl);377 378 xdp_tx->tx_db.data.bd_prod = cpu_to_le16(xdp_prod);379 qede_update_tx_producer(xdp_tx);380 }381 382 spin_unlock(&xdp_tx->xdp_tx_lock);383 384 return nxmit;385}386 387int qede_txq_has_work(struct qede_tx_queue *txq)388{389 u16 hw_bd_cons;390 391 /* Tell compiler that consumer and producer can change */392 barrier();393 hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr);394 if (qed_chain_get_cons_idx(&txq->tx_pbl) == hw_bd_cons + 1)395 return 0;396 397 return hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl);398}399 400static void qede_xdp_tx_int(struct qede_dev *edev, struct qede_tx_queue *txq)401{402 struct sw_tx_xdp *xdp_info, *xdp_arr = txq->sw_tx_ring.xdp;403 struct device *dev = &edev->pdev->dev;404 struct xdp_frame *xdpf;405 u16 hw_bd_cons;406 407 hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr);408 barrier();409 410 while (hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl)) {411 xdp_info = xdp_arr + txq->sw_tx_cons;412 xdpf = xdp_info->xdpf;413 414 if (xdpf) {415 dma_unmap_single(dev, xdp_info->mapping, xdpf->len,416 DMA_TO_DEVICE);417 xdp_return_frame(xdpf);418 419 xdp_info->xdpf = NULL;420 } else {421 dma_unmap_page(dev, xdp_info->mapping, PAGE_SIZE,422 DMA_BIDIRECTIONAL);423 __free_page(xdp_info->page);424 }425 426 qed_chain_consume(&txq->tx_pbl);427 txq->sw_tx_cons = (txq->sw_tx_cons + 1) % txq->num_tx_buffers;428 txq->xmit_pkts++;429 }430}431 432static int qede_tx_int(struct qede_dev *edev, struct qede_tx_queue *txq)433{434 unsigned int pkts_compl = 0, bytes_compl = 0;435 struct netdev_queue *netdev_txq;436 u16 hw_bd_cons;437 int rc;438 439 netdev_txq = netdev_get_tx_queue(edev->ndev, txq->ndev_txq_id);440 441 hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr);442 barrier();443 444 while (hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl)) {445 int len = 0;446 447 rc = qede_free_tx_pkt(edev, txq, &len);448 if (rc) {449 DP_NOTICE(edev, "hw_bd_cons = %d, chain_cons=%d\n",450 hw_bd_cons,451 qed_chain_get_cons_idx(&txq->tx_pbl));452 break;453 }454 455 bytes_compl += len;456 pkts_compl++;457 txq->sw_tx_cons = (txq->sw_tx_cons + 1) % txq->num_tx_buffers;458 txq->xmit_pkts++;459 }460 461 netdev_tx_completed_queue(netdev_txq, pkts_compl, bytes_compl);462 463 /* Need to make the tx_bd_cons update visible to start_xmit()464 * before checking for netif_tx_queue_stopped(). Without the465 * memory barrier, there is a small possibility that466 * start_xmit() will miss it and cause the queue to be stopped467 * forever.468 * On the other hand we need an rmb() here to ensure the proper469 * ordering of bit testing in the following470 * netif_tx_queue_stopped(txq) call.471 */472 smp_mb();473 474 if (unlikely(netif_tx_queue_stopped(netdev_txq))) {475 /* Taking tx_lock is needed to prevent reenabling the queue476 * while it's empty. This could have happen if rx_action() gets477 * suspended in qede_tx_int() after the condition before478 * netif_tx_wake_queue(), while tx_action (qede_start_xmit()):479 *480 * stops the queue->sees fresh tx_bd_cons->releases the queue->481 * sends some packets consuming the whole queue again->482 * stops the queue483 */484 485 __netif_tx_lock(netdev_txq, smp_processor_id());486 487 if ((netif_tx_queue_stopped(netdev_txq)) &&488 (edev->state == QEDE_STATE_OPEN) &&489 (qed_chain_get_elem_left(&txq->tx_pbl)490 >= (MAX_SKB_FRAGS + 1))) {491 netif_tx_wake_queue(netdev_txq);492 DP_VERBOSE(edev, NETIF_MSG_TX_DONE,493 "Wake queue was called\n");494 }495 496 __netif_tx_unlock(netdev_txq);497 }498 499 return 0;500}501 502bool qede_has_rx_work(struct qede_rx_queue *rxq)503{504 u16 hw_comp_cons, sw_comp_cons;505 506 /* Tell compiler that status block fields can change */507 barrier();508 509 hw_comp_cons = le16_to_cpu(*rxq->hw_cons_ptr);510 sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring);511 512 return hw_comp_cons != sw_comp_cons;513}514 515static inline void qede_rx_bd_ring_consume(struct qede_rx_queue *rxq)516{517 qed_chain_consume(&rxq->rx_bd_ring);518 rxq->sw_rx_cons++;519}520 521/* This function reuses the buffer(from an offset) from522 * consumer index to producer index in the bd ring523 */524static inline void qede_reuse_page(struct qede_rx_queue *rxq,525 struct sw_rx_data *curr_cons)526{527 struct eth_rx_bd *rx_bd_prod = qed_chain_produce(&rxq->rx_bd_ring);528 struct sw_rx_data *curr_prod;529 dma_addr_t new_mapping;530 531 curr_prod = &rxq->sw_rx_ring[rxq->sw_rx_prod & NUM_RX_BDS_MAX];532 *curr_prod = *curr_cons;533 534 new_mapping = curr_prod->mapping + curr_prod->page_offset;535 536 rx_bd_prod->addr.hi = cpu_to_le32(upper_32_bits(new_mapping));537 rx_bd_prod->addr.lo = cpu_to_le32(lower_32_bits(new_mapping) +538 rxq->rx_headroom);539 540 rxq->sw_rx_prod++;541 curr_cons->data = NULL;542}543 544/* In case of allocation failures reuse buffers545 * from consumer index to produce buffers for firmware546 */547void qede_recycle_rx_bd_ring(struct qede_rx_queue *rxq, u8 count)548{549 struct sw_rx_data *curr_cons;550 551 for (; count > 0; count--) {552 curr_cons = &rxq->sw_rx_ring[rxq->sw_rx_cons & NUM_RX_BDS_MAX];553 qede_reuse_page(rxq, curr_cons);554 qede_rx_bd_ring_consume(rxq);555 }556}557 558static inline int qede_realloc_rx_buffer(struct qede_rx_queue *rxq,559 struct sw_rx_data *curr_cons)560{561 /* Move to the next segment in the page */562 curr_cons->page_offset += rxq->rx_buf_seg_size;563 564 if (curr_cons->page_offset == PAGE_SIZE) {565 if (unlikely(qede_alloc_rx_buffer(rxq, true))) {566 /* Since we failed to allocate new buffer567 * current buffer can be used again.568 */569 curr_cons->page_offset -= rxq->rx_buf_seg_size;570 571 return -ENOMEM;572 }573 574 dma_unmap_page(rxq->dev, curr_cons->mapping,575 PAGE_SIZE, rxq->data_direction);576 } else {577 /* Increment refcount of the page as we don't want578 * network stack to take the ownership of the page579 * which can be recycled multiple times by the driver.580 */581 page_ref_inc(curr_cons->data);582 qede_reuse_page(rxq, curr_cons);583 }584 585 return 0;586}587 588void qede_update_rx_prod(struct qede_dev *edev, struct qede_rx_queue *rxq)589{590 u16 bd_prod = qed_chain_get_prod_idx(&rxq->rx_bd_ring);591 u16 cqe_prod = qed_chain_get_prod_idx(&rxq->rx_comp_ring);592 struct eth_rx_prod_data rx_prods = {0};593 594 /* Update producers */595 rx_prods.bd_prod = cpu_to_le16(bd_prod);596 rx_prods.cqe_prod = cpu_to_le16(cqe_prod);597 598 /* Make sure that the BD and SGE data is updated before updating the599 * producers since FW might read the BD/SGE right after the producer600 * is updated.601 */602 wmb();603 604 internal_ram_wr(rxq->hw_rxq_prod_addr, sizeof(rx_prods),605 (u32 *)&rx_prods);606}607 608static void qede_get_rxhash(struct sk_buff *skb, u8 bitfields, __le32 rss_hash)609{610 enum pkt_hash_types hash_type = PKT_HASH_TYPE_NONE;611 enum rss_hash_type htype;612 u32 hash = 0;613 614 htype = GET_FIELD(bitfields, ETH_FAST_PATH_RX_REG_CQE_RSS_HASH_TYPE);615 if (htype) {616 hash_type = ((htype == RSS_HASH_TYPE_IPV4) ||617 (htype == RSS_HASH_TYPE_IPV6)) ?618 PKT_HASH_TYPE_L3 : PKT_HASH_TYPE_L4;619 hash = le32_to_cpu(rss_hash);620 }621 skb_set_hash(skb, hash, hash_type);622}623 624static void qede_set_skb_csum(struct sk_buff *skb, u8 csum_flag)625{626 skb_checksum_none_assert(skb);627 628 if (csum_flag & QEDE_CSUM_UNNECESSARY)629 skb->ip_summed = CHECKSUM_UNNECESSARY;630 631 if (csum_flag & QEDE_TUNN_CSUM_UNNECESSARY) {632 skb->csum_level = 1;633 skb->encapsulation = 1;634 }635}636 637static inline void qede_skb_receive(struct qede_dev *edev,638 struct qede_fastpath *fp,639 struct qede_rx_queue *rxq,640 struct sk_buff *skb, u16 vlan_tag)641{642 if (vlan_tag)643 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tag);644 645 napi_gro_receive(&fp->napi, skb);646}647 648static void qede_set_gro_params(struct qede_dev *edev,649 struct sk_buff *skb,650 struct eth_fast_path_rx_tpa_start_cqe *cqe)651{652 u16 parsing_flags = le16_to_cpu(cqe->pars_flags.flags);653 654 if (((parsing_flags >> PARSING_AND_ERR_FLAGS_L3TYPE_SHIFT) &655 PARSING_AND_ERR_FLAGS_L3TYPE_MASK) == 2)656 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;657 else658 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;659 660 skb_shinfo(skb)->gso_size = __le16_to_cpu(cqe->len_on_first_bd) -661 cqe->header_len;662}663 664static int qede_fill_frag_skb(struct qede_dev *edev,665 struct qede_rx_queue *rxq,666 u8 tpa_agg_index, u16 len_on_bd)667{668 struct sw_rx_data *current_bd = &rxq->sw_rx_ring[rxq->sw_rx_cons &669 NUM_RX_BDS_MAX];670 struct qede_agg_info *tpa_info = &rxq->tpa_info[tpa_agg_index];671 struct sk_buff *skb = tpa_info->skb;672 673 if (unlikely(tpa_info->state != QEDE_AGG_STATE_START))674 goto out;675 676 /* Add one frag and update the appropriate fields in the skb */677 skb_fill_page_desc(skb, tpa_info->frag_id++,678 current_bd->data,679 current_bd->page_offset + rxq->rx_headroom,680 len_on_bd);681 682 if (unlikely(qede_realloc_rx_buffer(rxq, current_bd))) {683 /* Incr page ref count to reuse on allocation failure684 * so that it doesn't get freed while freeing SKB.685 */686 page_ref_inc(current_bd->data);687 goto out;688 }689 690 qede_rx_bd_ring_consume(rxq);691 692 skb->data_len += len_on_bd;693 skb->truesize += rxq->rx_buf_seg_size;694 skb->len += len_on_bd;695 696 return 0;697 698out:699 tpa_info->state = QEDE_AGG_STATE_ERROR;700 qede_recycle_rx_bd_ring(rxq, 1);701 702 return -ENOMEM;703}704 705static bool qede_tunn_exist(u16 flag)706{707 return !!(flag & (PARSING_AND_ERR_FLAGS_TUNNELEXIST_MASK <<708 PARSING_AND_ERR_FLAGS_TUNNELEXIST_SHIFT));709}710 711static u8 qede_check_tunn_csum(u16 flag)712{713 u16 csum_flag = 0;714 u8 tcsum = 0;715 716 if (flag & (PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMWASCALCULATED_MASK <<717 PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMWASCALCULATED_SHIFT))718 csum_flag |= PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMERROR_MASK <<719 PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMERROR_SHIFT;720 721 if (flag & (PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_MASK <<722 PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_SHIFT)) {723 csum_flag |= PARSING_AND_ERR_FLAGS_L4CHKSMERROR_MASK <<724 PARSING_AND_ERR_FLAGS_L4CHKSMERROR_SHIFT;725 tcsum = QEDE_TUNN_CSUM_UNNECESSARY;726 }727 728 csum_flag |= PARSING_AND_ERR_FLAGS_TUNNELIPHDRERROR_MASK <<729 PARSING_AND_ERR_FLAGS_TUNNELIPHDRERROR_SHIFT |730 PARSING_AND_ERR_FLAGS_IPHDRERROR_MASK <<731 PARSING_AND_ERR_FLAGS_IPHDRERROR_SHIFT;732 733 if (csum_flag & flag)734 return QEDE_CSUM_ERROR;735 736 return QEDE_CSUM_UNNECESSARY | tcsum;737}738 739static inline struct sk_buff *740qede_build_skb(struct qede_rx_queue *rxq,741 struct sw_rx_data *bd, u16 len, u16 pad)742{743 struct sk_buff *skb;744 void *buf;745 746 buf = page_address(bd->data) + bd->page_offset;747 skb = build_skb(buf, rxq->rx_buf_seg_size);748 749 if (unlikely(!skb))750 return NULL;751 752 skb_reserve(skb, pad);753 skb_put(skb, len);754 755 return skb;756}757 758static struct sk_buff *759qede_tpa_rx_build_skb(struct qede_dev *edev,760 struct qede_rx_queue *rxq,761 struct sw_rx_data *bd, u16 len, u16 pad,762 bool alloc_skb)763{764 struct sk_buff *skb;765 766 skb = qede_build_skb(rxq, bd, len, pad);767 bd->page_offset += rxq->rx_buf_seg_size;768 769 if (bd->page_offset == PAGE_SIZE) {770 if (unlikely(qede_alloc_rx_buffer(rxq, true))) {771 DP_NOTICE(edev,772 "Failed to allocate RX buffer for tpa start\n");773 bd->page_offset -= rxq->rx_buf_seg_size;774 page_ref_inc(bd->data);775 dev_kfree_skb_any(skb);776 return NULL;777 }778 } else {779 page_ref_inc(bd->data);780 qede_reuse_page(rxq, bd);781 }782 783 /* We've consumed the first BD and prepared an SKB */784 qede_rx_bd_ring_consume(rxq);785 786 return skb;787}788 789static struct sk_buff *790qede_rx_build_skb(struct qede_dev *edev,791 struct qede_rx_queue *rxq,792 struct sw_rx_data *bd, u16 len, u16 pad)793{794 struct sk_buff *skb = NULL;795 796 /* For smaller frames still need to allocate skb, memcpy797 * data and benefit in reusing the page segment instead of798 * un-mapping it.799 */800 if ((len + pad <= edev->rx_copybreak)) {801 unsigned int offset = bd->page_offset + pad;802 803 skb = netdev_alloc_skb(edev->ndev, QEDE_RX_HDR_SIZE);804 if (unlikely(!skb))805 return NULL;806 807 skb_reserve(skb, pad);808 skb_put_data(skb, page_address(bd->data) + offset, len);809 qede_reuse_page(rxq, bd);810 goto out;811 }812 813 skb = qede_build_skb(rxq, bd, len, pad);814 815 if (unlikely(qede_realloc_rx_buffer(rxq, bd))) {816 /* Incr page ref count to reuse on allocation failure so817 * that it doesn't get freed while freeing SKB [as its818 * already mapped there].819 */820 page_ref_inc(bd->data);821 dev_kfree_skb_any(skb);822 return NULL;823 }824out:825 /* We've consumed the first BD and prepared an SKB */826 qede_rx_bd_ring_consume(rxq);827 828 return skb;829}830 831static void qede_tpa_start(struct qede_dev *edev,832 struct qede_rx_queue *rxq,833 struct eth_fast_path_rx_tpa_start_cqe *cqe)834{835 struct qede_agg_info *tpa_info = &rxq->tpa_info[cqe->tpa_agg_index];836 struct sw_rx_data *sw_rx_data_cons;837 u16 pad;838 839 sw_rx_data_cons = &rxq->sw_rx_ring[rxq->sw_rx_cons & NUM_RX_BDS_MAX];840 pad = cqe->placement_offset + rxq->rx_headroom;841 842 tpa_info->skb = qede_tpa_rx_build_skb(edev, rxq, sw_rx_data_cons,843 le16_to_cpu(cqe->len_on_first_bd),844 pad, false);845 tpa_info->buffer.page_offset = sw_rx_data_cons->page_offset;846 tpa_info->buffer.mapping = sw_rx_data_cons->mapping;847 848 if (unlikely(!tpa_info->skb)) {849 DP_NOTICE(edev, "Failed to allocate SKB for gro\n");850 851 /* Consume from ring but do not produce since852 * this might be used by FW still, it will be re-used853 * at TPA end.854 */855 tpa_info->tpa_start_fail = true;856 qede_rx_bd_ring_consume(rxq);857 tpa_info->state = QEDE_AGG_STATE_ERROR;858 goto cons_buf;859 }860 861 tpa_info->frag_id = 0;862 tpa_info->state = QEDE_AGG_STATE_START;863 864 if ((le16_to_cpu(cqe->pars_flags.flags) >>865 PARSING_AND_ERR_FLAGS_TAG8021QEXIST_SHIFT) &866 PARSING_AND_ERR_FLAGS_TAG8021QEXIST_MASK)867 tpa_info->vlan_tag = le16_to_cpu(cqe->vlan_tag);868 else869 tpa_info->vlan_tag = 0;870 871 qede_get_rxhash(tpa_info->skb, cqe->bitfields, cqe->rss_hash);872 873 /* This is needed in order to enable forwarding support */874 qede_set_gro_params(edev, tpa_info->skb, cqe);875 876cons_buf: /* We still need to handle bd_len_list to consume buffers */877 if (likely(cqe->bw_ext_bd_len_list[0]))878 qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index,879 le16_to_cpu(cqe->bw_ext_bd_len_list[0]));880 881 if (unlikely(cqe->bw_ext_bd_len_list[1])) {882 DP_ERR(edev,883 "Unlikely - got a TPA aggregation with more than one bw_ext_bd_len_list entry in the TPA start\n");884 tpa_info->state = QEDE_AGG_STATE_ERROR;885 }886}887 888#ifdef CONFIG_INET889static void qede_gro_ip_csum(struct sk_buff *skb)890{891 const struct iphdr *iph = ip_hdr(skb);892 struct tcphdr *th;893 894 skb_set_transport_header(skb, sizeof(struct iphdr));895 th = tcp_hdr(skb);896 897 th->check = ~tcp_v4_check(skb->len - skb_transport_offset(skb),898 iph->saddr, iph->daddr, 0);899 900 tcp_gro_complete(skb);901}902 903static void qede_gro_ipv6_csum(struct sk_buff *skb)904{905 struct ipv6hdr *iph = ipv6_hdr(skb);906 struct tcphdr *th;907 908 skb_set_transport_header(skb, sizeof(struct ipv6hdr));909 th = tcp_hdr(skb);910 911 th->check = ~tcp_v6_check(skb->len - skb_transport_offset(skb),912 &iph->saddr, &iph->daddr, 0);913 tcp_gro_complete(skb);914}915#endif916 917static void qede_gro_receive(struct qede_dev *edev,918 struct qede_fastpath *fp,919 struct sk_buff *skb,920 u16 vlan_tag)921{922 /* FW can send a single MTU sized packet from gro flow923 * due to aggregation timeout/last segment etc. which924 * is not expected to be a gro packet. If a skb has zero925 * frags then simply push it in the stack as non gso skb.926 */927 if (unlikely(!skb->data_len)) {928 skb_shinfo(skb)->gso_type = 0;929 skb_shinfo(skb)->gso_size = 0;930 goto send_skb;931 }932 933#ifdef CONFIG_INET934 if (skb_shinfo(skb)->gso_size) {935 skb_reset_network_header(skb);936 937 switch (skb->protocol) {938 case htons(ETH_P_IP):939 qede_gro_ip_csum(skb);940 break;941 case htons(ETH_P_IPV6):942 qede_gro_ipv6_csum(skb);943 break;944 default:945 DP_ERR(edev,946 "Error: FW GRO supports only IPv4/IPv6, not 0x%04x\n",947 ntohs(skb->protocol));948 }949 }950#endif951 952send_skb:953 skb_record_rx_queue(skb, fp->rxq->rxq_id);954 qede_skb_receive(edev, fp, fp->rxq, skb, vlan_tag);955}956 957static inline void qede_tpa_cont(struct qede_dev *edev,958 struct qede_rx_queue *rxq,959 struct eth_fast_path_rx_tpa_cont_cqe *cqe)960{961 int i;962 963 for (i = 0; cqe->len_list[i]; i++)964 qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index,965 le16_to_cpu(cqe->len_list[i]));966 967 if (unlikely(i > 1))968 DP_ERR(edev,969 "Strange - TPA cont with more than a single len_list entry\n");970}971 972static int qede_tpa_end(struct qede_dev *edev,973 struct qede_fastpath *fp,974 struct eth_fast_path_rx_tpa_end_cqe *cqe)975{976 struct qede_rx_queue *rxq = fp->rxq;977 struct qede_agg_info *tpa_info;978 struct sk_buff *skb;979 int i;980 981 tpa_info = &rxq->tpa_info[cqe->tpa_agg_index];982 skb = tpa_info->skb;983 984 if (tpa_info->buffer.page_offset == PAGE_SIZE)985 dma_unmap_page(rxq->dev, tpa_info->buffer.mapping,986 PAGE_SIZE, rxq->data_direction);987 988 for (i = 0; cqe->len_list[i]; i++)989 qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index,990 le16_to_cpu(cqe->len_list[i]));991 if (unlikely(i > 1))992 DP_ERR(edev,993 "Strange - TPA emd with more than a single len_list entry\n");994 995 if (unlikely(tpa_info->state != QEDE_AGG_STATE_START))996 goto err;997 998 /* Sanity */999 if (unlikely(cqe->num_of_bds != tpa_info->frag_id + 1))1000 DP_ERR(edev,1001 "Strange - TPA had %02x BDs, but SKB has only %d frags\n",1002 cqe->num_of_bds, tpa_info->frag_id);1003 if (unlikely(skb->len != le16_to_cpu(cqe->total_packet_len)))1004 DP_ERR(edev,1005 "Strange - total packet len [cqe] is %4x but SKB has len %04x\n",1006 le16_to_cpu(cqe->total_packet_len), skb->len);1007 1008 /* Finalize the SKB */1009 skb->protocol = eth_type_trans(skb, edev->ndev);1010 skb->ip_summed = CHECKSUM_UNNECESSARY;1011 1012 /* tcp_gro_complete() will copy NAPI_GRO_CB(skb)->count1013 * to skb_shinfo(skb)->gso_segs1014 */1015 NAPI_GRO_CB(skb)->count = le16_to_cpu(cqe->num_of_coalesced_segs);1016 1017 qede_gro_receive(edev, fp, skb, tpa_info->vlan_tag);1018 1019 tpa_info->state = QEDE_AGG_STATE_NONE;1020 1021 return 1;1022err:1023 tpa_info->state = QEDE_AGG_STATE_NONE;1024 1025 if (tpa_info->tpa_start_fail) {1026 qede_reuse_page(rxq, &tpa_info->buffer);1027 tpa_info->tpa_start_fail = false;1028 }1029 1030 dev_kfree_skb_any(tpa_info->skb);1031 tpa_info->skb = NULL;1032 return 0;1033}1034 1035static u8 qede_check_notunn_csum(u16 flag)1036{1037 u16 csum_flag = 0;1038 u8 csum = 0;1039 1040 if (flag & (PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_MASK <<1041 PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_SHIFT)) {1042 csum_flag |= PARSING_AND_ERR_FLAGS_L4CHKSMERROR_MASK <<1043 PARSING_AND_ERR_FLAGS_L4CHKSMERROR_SHIFT;1044 csum = QEDE_CSUM_UNNECESSARY;1045 }1046 1047 csum_flag |= PARSING_AND_ERR_FLAGS_IPHDRERROR_MASK <<1048 PARSING_AND_ERR_FLAGS_IPHDRERROR_SHIFT;1049 1050 if (csum_flag & flag)1051 return QEDE_CSUM_ERROR;1052 1053 return csum;1054}1055 1056static u8 qede_check_csum(u16 flag)1057{1058 if (!qede_tunn_exist(flag))1059 return qede_check_notunn_csum(flag);1060 else1061 return qede_check_tunn_csum(flag);1062}1063 1064static bool qede_pkt_is_ip_fragmented(struct eth_fast_path_rx_reg_cqe *cqe,1065 u16 flag)1066{1067 u8 tun_pars_flg = cqe->tunnel_pars_flags.flags;1068 1069 if ((tun_pars_flg & (ETH_TUNNEL_PARSING_FLAGS_IPV4_FRAGMENT_MASK <<1070 ETH_TUNNEL_PARSING_FLAGS_IPV4_FRAGMENT_SHIFT)) ||1071 (flag & (PARSING_AND_ERR_FLAGS_IPV4FRAG_MASK <<1072 PARSING_AND_ERR_FLAGS_IPV4FRAG_SHIFT)))1073 return true;1074 1075 return false;1076}1077 1078/* Return true iff packet is to be passed to stack */1079static bool qede_rx_xdp(struct qede_dev *edev,1080 struct qede_fastpath *fp,1081 struct qede_rx_queue *rxq,1082 struct bpf_prog *prog,1083 struct sw_rx_data *bd,1084 struct eth_fast_path_rx_reg_cqe *cqe,1085 u16 *data_offset, u16 *len)1086{1087 struct xdp_buff xdp;1088 enum xdp_action act;1089 1090 xdp_init_buff(&xdp, rxq->rx_buf_seg_size, &rxq->xdp_rxq);1091 xdp_prepare_buff(&xdp, page_address(bd->data), *data_offset,1092 *len, false);1093 1094 act = bpf_prog_run_xdp(prog, &xdp);1095 1096 /* Recalculate, as XDP might have changed the headers */1097 *data_offset = xdp.data - xdp.data_hard_start;1098 *len = xdp.data_end - xdp.data;1099 1100 if (act == XDP_PASS)1101 return true;1102 1103 /* Count number of packets not to be passed to stack */1104 rxq->xdp_no_pass++;1105 1106 switch (act) {1107 case XDP_TX:1108 /* We need the replacement buffer before transmit. */1109 if (unlikely(qede_alloc_rx_buffer(rxq, true))) {1110 qede_recycle_rx_bd_ring(rxq, 1);1111 1112 trace_xdp_exception(edev->ndev, prog, act);1113 break;1114 }1115 1116 /* Now if there's a transmission problem, we'd still have to1117 * throw current buffer, as replacement was already allocated.1118 */1119 if (unlikely(qede_xdp_xmit(fp->xdp_tx, bd->mapping,1120 *data_offset, *len, bd->data,1121 NULL))) {1122 dma_unmap_page(rxq->dev, bd->mapping, PAGE_SIZE,1123 rxq->data_direction);1124 __free_page(bd->data);1125 1126 trace_xdp_exception(edev->ndev, prog, act);1127 } else {1128 dma_sync_single_for_device(rxq->dev,1129 bd->mapping + *data_offset,1130 *len, rxq->data_direction);1131 fp->xdp_xmit |= QEDE_XDP_TX;1132 }1133 1134 /* Regardless, we've consumed an Rx BD */1135 qede_rx_bd_ring_consume(rxq);1136 break;1137 case XDP_REDIRECT:1138 /* We need the replacement buffer before transmit. */1139 if (unlikely(qede_alloc_rx_buffer(rxq, true))) {1140 qede_recycle_rx_bd_ring(rxq, 1);1141 1142 trace_xdp_exception(edev->ndev, prog, act);1143 break;1144 }1145 1146 dma_unmap_page(rxq->dev, bd->mapping, PAGE_SIZE,1147 rxq->data_direction);1148 1149 if (unlikely(xdp_do_redirect(edev->ndev, &xdp, prog)))1150 DP_NOTICE(edev, "Failed to redirect the packet\n");1151 else1152 fp->xdp_xmit |= QEDE_XDP_REDIRECT;1153 1154 qede_rx_bd_ring_consume(rxq);1155 break;1156 default:1157 bpf_warn_invalid_xdp_action(edev->ndev, prog, act);1158 fallthrough;1159 case XDP_ABORTED:1160 trace_xdp_exception(edev->ndev, prog, act);1161 fallthrough;1162 case XDP_DROP:1163 qede_recycle_rx_bd_ring(rxq, cqe->bd_num);1164 }1165 1166 return false;1167}1168 1169static int qede_rx_build_jumbo(struct qede_dev *edev,1170 struct qede_rx_queue *rxq,1171 struct sk_buff *skb,1172 struct eth_fast_path_rx_reg_cqe *cqe,1173 u16 first_bd_len)1174{1175 u16 pkt_len = le16_to_cpu(cqe->pkt_len);1176 struct sw_rx_data *bd;1177 u16 bd_cons_idx;1178 u8 num_frags;1179 1180 pkt_len -= first_bd_len;1181 1182 /* We've already used one BD for the SKB. Now take care of the rest */1183 for (num_frags = cqe->bd_num - 1; num_frags > 0; num_frags--) {1184 u16 cur_size = pkt_len > rxq->rx_buf_size ? rxq->rx_buf_size :1185 pkt_len;1186 1187 if (unlikely(!cur_size)) {1188 DP_ERR(edev,1189 "Still got %d BDs for mapping jumbo, but length became 0\n",1190 num_frags);1191 goto out;1192 }1193 1194 /* We need a replacement buffer for each BD */1195 if (unlikely(qede_alloc_rx_buffer(rxq, true)))1196 goto out;1197 1198 /* Now that we've allocated the replacement buffer,1199 * we can safely consume the next BD and map it to the SKB.1200 */1201 bd_cons_idx = rxq->sw_rx_cons & NUM_RX_BDS_MAX;1202 bd = &rxq->sw_rx_ring[bd_cons_idx];1203 qede_rx_bd_ring_consume(rxq);1204 1205 dma_unmap_page(rxq->dev, bd->mapping,1206 PAGE_SIZE, DMA_FROM_DEVICE);1207 1208 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, bd->data,1209 rxq->rx_headroom, cur_size, PAGE_SIZE);1210 1211 pkt_len -= cur_size;1212 }1213 1214 if (unlikely(pkt_len))1215 DP_ERR(edev,1216 "Mapped all BDs of jumbo, but still have %d bytes\n",1217 pkt_len);1218 1219out:1220 return num_frags;1221}1222 1223static int qede_rx_process_tpa_cqe(struct qede_dev *edev,1224 struct qede_fastpath *fp,1225 struct qede_rx_queue *rxq,1226 union eth_rx_cqe *cqe,1227 enum eth_rx_cqe_type type)1228{1229 switch (type) {1230 case ETH_RX_CQE_TYPE_TPA_START:1231 qede_tpa_start(edev, rxq, &cqe->fast_path_tpa_start);1232 return 0;1233 case ETH_RX_CQE_TYPE_TPA_CONT:1234 qede_tpa_cont(edev, rxq, &cqe->fast_path_tpa_cont);1235 return 0;1236 case ETH_RX_CQE_TYPE_TPA_END:1237 return qede_tpa_end(edev, fp, &cqe->fast_path_tpa_end);1238 default:1239 return 0;1240 }1241}1242 1243static int qede_rx_process_cqe(struct qede_dev *edev,1244 struct qede_fastpath *fp,1245 struct qede_rx_queue *rxq)1246{1247 struct bpf_prog *xdp_prog = READ_ONCE(rxq->xdp_prog);1248 struct eth_fast_path_rx_reg_cqe *fp_cqe;1249 u16 len, pad, bd_cons_idx, parse_flag;1250 enum eth_rx_cqe_type cqe_type;1251 union eth_rx_cqe *cqe;1252 struct sw_rx_data *bd;1253 struct sk_buff *skb;1254 __le16 flags;1255 u8 csum_flag;1256 1257 /* Get the CQE from the completion ring */1258 cqe = (union eth_rx_cqe *)qed_chain_consume(&rxq->rx_comp_ring);1259 cqe_type = cqe->fast_path_regular.type;1260 1261 /* Process an unlikely slowpath event */1262 if (unlikely(cqe_type == ETH_RX_CQE_TYPE_SLOW_PATH)) {1263 struct eth_slow_path_rx_cqe *sp_cqe;1264 1265 sp_cqe = (struct eth_slow_path_rx_cqe *)cqe;1266 edev->ops->eth_cqe_completion(edev->cdev, fp->id, sp_cqe);1267 return 0;1268 }1269 1270 /* Handle TPA cqes */1271 if (cqe_type != ETH_RX_CQE_TYPE_REGULAR)1272 return qede_rx_process_tpa_cqe(edev, fp, rxq, cqe, cqe_type);1273 1274 /* Get the data from the SW ring; Consume it only after it's evident1275 * we wouldn't recycle it.1276 */1277 bd_cons_idx = rxq->sw_rx_cons & NUM_RX_BDS_MAX;1278 bd = &rxq->sw_rx_ring[bd_cons_idx];1279 1280 fp_cqe = &cqe->fast_path_regular;1281 len = le16_to_cpu(fp_cqe->len_on_first_bd);1282 pad = fp_cqe->placement_offset + rxq->rx_headroom;1283 1284 /* Run eBPF program if one is attached */1285 if (xdp_prog)1286 if (!qede_rx_xdp(edev, fp, rxq, xdp_prog, bd, fp_cqe,1287 &pad, &len))1288 return 0;1289 1290 /* If this is an error packet then drop it */1291 flags = cqe->fast_path_regular.pars_flags.flags;1292 parse_flag = le16_to_cpu(flags);1293 1294 csum_flag = qede_check_csum(parse_flag);1295 if (unlikely(csum_flag == QEDE_CSUM_ERROR)) {1296 if (qede_pkt_is_ip_fragmented(fp_cqe, parse_flag))1297 rxq->rx_ip_frags++;1298 else1299 rxq->rx_hw_errors++;1300 }1301 1302 /* Basic validation passed; Need to prepare an SKB. This would also1303 * guarantee to finally consume the first BD upon success.1304 */1305 skb = qede_rx_build_skb(edev, rxq, bd, len, pad);1306 if (!skb) {1307 rxq->rx_alloc_errors++;1308 qede_recycle_rx_bd_ring(rxq, fp_cqe->bd_num);1309 return 0;1310 }1311 1312 /* In case of Jumbo packet, several PAGE_SIZEd buffers will be pointed1313 * by a single cqe.1314 */1315 if (fp_cqe->bd_num > 1) {1316 u16 unmapped_frags = qede_rx_build_jumbo(edev, rxq, skb,1317 fp_cqe, len);1318 1319 if (unlikely(unmapped_frags > 0)) {1320 qede_recycle_rx_bd_ring(rxq, unmapped_frags);1321 dev_kfree_skb_any(skb);1322 return 0;1323 }1324 }1325 1326 /* The SKB contains all the data. Now prepare meta-magic */1327 skb->protocol = eth_type_trans(skb, edev->ndev);1328 qede_get_rxhash(skb, fp_cqe->bitfields, fp_cqe->rss_hash);1329 qede_set_skb_csum(skb, csum_flag);1330 skb_record_rx_queue(skb, rxq->rxq_id);1331 qede_ptp_record_rx_ts(edev, cqe, skb);1332 1333 /* SKB is prepared - pass it to stack */1334 qede_skb_receive(edev, fp, rxq, skb, le16_to_cpu(fp_cqe->vlan_tag));1335 1336 return 1;1337}1338 1339static int qede_rx_int(struct qede_fastpath *fp, int budget)1340{1341 struct qede_rx_queue *rxq = fp->rxq;1342 struct qede_dev *edev = fp->edev;1343 int work_done = 0, rcv_pkts = 0;1344 u16 hw_comp_cons, sw_comp_cons;1345 1346 hw_comp_cons = le16_to_cpu(*rxq->hw_cons_ptr);1347 sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring);1348 1349 /* Memory barrier to prevent the CPU from doing speculative reads of CQE1350 * / BD in the while-loop before reading hw_comp_cons. If the CQE is1351 * read before it is written by FW, then FW writes CQE and SB, and then1352 * the CPU reads the hw_comp_cons, it will use an old CQE.1353 */1354 rmb();1355 1356 /* Loop to complete all indicated BDs */1357 while ((sw_comp_cons != hw_comp_cons) && (work_done < budget)) {1358 rcv_pkts += qede_rx_process_cqe(edev, fp, rxq);1359 qed_chain_recycle_consumed(&rxq->rx_comp_ring);1360 sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring);1361 work_done++;1362 }1363 1364 rxq->rcv_pkts += rcv_pkts;1365 1366 /* Allocate replacement buffers */1367 while (rxq->num_rx_buffers - rxq->filled_buffers)1368 if (qede_alloc_rx_buffer(rxq, false))1369 break;1370 1371 /* Update producers */1372 qede_update_rx_prod(edev, rxq);1373 1374 return work_done;1375}1376 1377static bool qede_poll_is_more_work(struct qede_fastpath *fp)1378{1379 qed_sb_update_sb_idx(fp->sb_info);1380 1381 /* *_has_*_work() reads the status block, thus we need to ensure that1382 * status block indices have been actually read (qed_sb_update_sb_idx)1383 * prior to this check (*_has_*_work) so that we won't write the1384 * "newer" value of the status block to HW (if there was a DMA right1385 * after qede_has_rx_work and if there is no rmb, the memory reading1386 * (qed_sb_update_sb_idx) may be postponed to right before *_ack_sb).1387 * In this case there will never be another interrupt until there is1388 * another update of the status block, while there is still unhandled1389 * work.1390 */1391 rmb();1392 1393 if (likely(fp->type & QEDE_FASTPATH_RX))1394 if (qede_has_rx_work(fp->rxq))1395 return true;1396 1397 if (fp->type & QEDE_FASTPATH_XDP)1398 if (qede_txq_has_work(fp->xdp_tx))1399 return true;1400 1401 if (likely(fp->type & QEDE_FASTPATH_TX)) {1402 int cos;1403 1404 for_each_cos_in_txq(fp->edev, cos) {1405 if (qede_txq_has_work(&fp->txq[cos]))1406 return true;1407 }1408 }1409 1410 return false;1411}1412 1413/*********************1414 * NDO & API related *1415 *********************/1416int qede_poll(struct napi_struct *napi, int budget)1417{1418 struct qede_fastpath *fp = container_of(napi, struct qede_fastpath,1419 napi);1420 struct qede_dev *edev = fp->edev;1421 int rx_work_done = 0;1422 u16 xdp_prod;1423 1424 fp->xdp_xmit = 0;1425 1426 if (likely(fp->type & QEDE_FASTPATH_TX)) {1427 int cos;1428 1429 for_each_cos_in_txq(fp->edev, cos) {1430 if (qede_txq_has_work(&fp->txq[cos]))1431 qede_tx_int(edev, &fp->txq[cos]);1432 }1433 }1434 1435 if ((fp->type & QEDE_FASTPATH_XDP) && qede_txq_has_work(fp->xdp_tx))1436 qede_xdp_tx_int(edev, fp->xdp_tx);1437 1438 rx_work_done = (likely(fp->type & QEDE_FASTPATH_RX) &&1439 qede_has_rx_work(fp->rxq)) ?1440 qede_rx_int(fp, budget) : 0;1441 1442 if (fp->xdp_xmit & QEDE_XDP_REDIRECT)1443 xdp_do_flush();1444 1445 /* Handle case where we are called by netpoll with a budget of 0 */1446 if (rx_work_done < budget || !budget) {1447 if (!qede_poll_is_more_work(fp)) {1448 napi_complete_done(napi, rx_work_done);1449 1450 /* Update and reenable interrupts */1451 qed_sb_ack(fp->sb_info, IGU_INT_ENABLE, 1);1452 } else {1453 rx_work_done = budget;1454 }1455 }1456 1457 if (fp->xdp_xmit & QEDE_XDP_TX) {1458 xdp_prod = qed_chain_get_prod_idx(&fp->xdp_tx->tx_pbl);1459 1460 fp->xdp_tx->tx_db.data.bd_prod = cpu_to_le16(xdp_prod);1461 qede_update_tx_producer(fp->xdp_tx);1462 }1463 1464 return rx_work_done;1465}1466 1467irqreturn_t qede_msix_fp_int(int irq, void *fp_cookie)1468{1469 struct qede_fastpath *fp = fp_cookie;1470 1471 qed_sb_ack(fp->sb_info, IGU_INT_DISABLE, 0 /*do not update*/);1472 1473 napi_schedule_irqoff(&fp->napi);1474 return IRQ_HANDLED;1475}1476 1477/* Main transmit function */1478netdev_tx_t qede_start_xmit(struct sk_buff *skb, struct net_device *ndev)1479{1480 struct qede_dev *edev = netdev_priv(ndev);1481 struct netdev_queue *netdev_txq;1482 struct qede_tx_queue *txq;1483 struct eth_tx_1st_bd *first_bd;1484 struct eth_tx_2nd_bd *second_bd = NULL;1485 struct eth_tx_3rd_bd *third_bd = NULL;1486 struct eth_tx_bd *tx_data_bd = NULL;1487 u16 txq_index, val = 0;1488 u8 nbd = 0;1489 dma_addr_t mapping;1490 int rc, frag_idx = 0, ipv6_ext = 0;1491 u8 xmit_type;1492 u16 idx;1493 u16 hlen;1494 bool data_split = false;1495 1496 /* Get tx-queue context and netdev index */1497 txq_index = skb_get_queue_mapping(skb);1498 WARN_ON(txq_index >= QEDE_TSS_COUNT(edev) * edev->dev_info.num_tc);1499 txq = QEDE_NDEV_TXQ_ID_TO_TXQ(edev, txq_index);1500 netdev_txq = netdev_get_tx_queue(ndev, txq_index);1501 1502 WARN_ON(qed_chain_get_elem_left(&txq->tx_pbl) < (MAX_SKB_FRAGS + 1));1503 1504 xmit_type = qede_xmit_type(skb, &ipv6_ext);1505 1506#if ((MAX_SKB_FRAGS + 2) > ETH_TX_MAX_BDS_PER_NON_LSO_PACKET)1507 if (qede_pkt_req_lin(skb, xmit_type)) {1508 if (skb_linearize(skb)) {1509 txq->tx_mem_alloc_err++;1510 1511 dev_kfree_skb_any(skb);1512 return NETDEV_TX_OK;1513 }1514 }1515#endif1516 1517 /* Fill the entry in the SW ring and the BDs in the FW ring */1518 idx = txq->sw_tx_prod;1519 txq->sw_tx_ring.skbs[idx].skb = skb;1520 first_bd = (struct eth_tx_1st_bd *)1521 qed_chain_produce(&txq->tx_pbl);1522 memset(first_bd, 0, sizeof(*first_bd));1523 first_bd->data.bd_flags.bitfields =1524 1 << ETH_TX_1ST_BD_FLAGS_START_BD_SHIFT;1525 1526 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP))1527 qede_ptp_tx_ts(edev, skb);1528 1529 /* Map skb linear data for DMA and set in the first BD */1530 mapping = dma_map_single(txq->dev, skb->data,1531 skb_headlen(skb), DMA_TO_DEVICE);1532 if (unlikely(dma_mapping_error(txq->dev, mapping))) {1533 DP_NOTICE(edev, "SKB mapping failed\n");1534 qede_free_failed_tx_pkt(txq, first_bd, 0, false);1535 qede_update_tx_producer(txq);1536 return NETDEV_TX_OK;1537 }1538 nbd++;1539 BD_SET_UNMAP_ADDR_LEN(first_bd, mapping, skb_headlen(skb));1540 1541 /* In case there is IPv6 with extension headers or LSO we need 2nd and1542 * 3rd BDs.1543 */1544 if (unlikely((xmit_type & XMIT_LSO) | ipv6_ext)) {1545 second_bd = (struct eth_tx_2nd_bd *)1546 qed_chain_produce(&txq->tx_pbl);1547 memset(second_bd, 0, sizeof(*second_bd));1548 1549 nbd++;1550 third_bd = (struct eth_tx_3rd_bd *)1551 qed_chain_produce(&txq->tx_pbl);1552 memset(third_bd, 0, sizeof(*third_bd));1553 1554 nbd++;1555 /* We need to fill in additional data in second_bd... */1556 tx_data_bd = (struct eth_tx_bd *)second_bd;1557 }1558 1559 if (skb_vlan_tag_present(skb)) {1560 first_bd->data.vlan = cpu_to_le16(skb_vlan_tag_get(skb));1561 first_bd->data.bd_flags.bitfields |=1562 1 << ETH_TX_1ST_BD_FLAGS_VLAN_INSERTION_SHIFT;1563 }1564 1565 /* Fill the parsing flags & params according to the requested offload */1566 if (xmit_type & XMIT_L4_CSUM) {1567 /* We don't re-calculate IP checksum as it is already done by1568 * the upper stack1569 */1570 first_bd->data.bd_flags.bitfields |=1571 1 << ETH_TX_1ST_BD_FLAGS_L4_CSUM_SHIFT;1572 1573 if (xmit_type & XMIT_ENC) {1574 first_bd->data.bd_flags.bitfields |=1575 1 << ETH_TX_1ST_BD_FLAGS_IP_CSUM_SHIFT;1576 1577 val |= (1 << ETH_TX_DATA_1ST_BD_TUNN_FLAG_SHIFT);1578 }1579 1580 /* Legacy FW had flipped behavior in regard to this bit -1581 * I.e., needed to set to prevent FW from touching encapsulated1582 * packets when it didn't need to.1583 */1584 if (unlikely(txq->is_legacy))1585 val ^= (1 << ETH_TX_DATA_1ST_BD_TUNN_FLAG_SHIFT);1586 1587 /* If the packet is IPv6 with extension header, indicate that1588 * to FW and pass few params, since the device cracker doesn't1589 * support parsing IPv6 with extension header/s.1590 */1591 if (unlikely(ipv6_ext))1592 qede_set_params_for_ipv6_ext(skb, second_bd, third_bd);1593 }1594 1595 if (xmit_type & XMIT_LSO) {1596 first_bd->data.bd_flags.bitfields |=1597 (1 << ETH_TX_1ST_BD_FLAGS_LSO_SHIFT);1598 third_bd->data.lso_mss =1599 cpu_to_le16(skb_shinfo(skb)->gso_size);1600 1601 if (unlikely(xmit_type & XMIT_ENC)) {1602 first_bd->data.bd_flags.bitfields |=1603 1 << ETH_TX_1ST_BD_FLAGS_TUNN_IP_CSUM_SHIFT;1604 1605 if (xmit_type & XMIT_ENC_GSO_L4_CSUM) {1606 u8 tmp = ETH_TX_1ST_BD_FLAGS_TUNN_L4_CSUM_SHIFT;1607 1608 first_bd->data.bd_flags.bitfields |= 1 << tmp;1609 }1610 hlen = qede_get_skb_hlen(skb, true);1611 } else {1612 first_bd->data.bd_flags.bitfields |=1613 1 << ETH_TX_1ST_BD_FLAGS_IP_CSUM_SHIFT;1614 hlen = qede_get_skb_hlen(skb, false);1615 }1616 1617 /* @@@TBD - if will not be removed need to check */1618 third_bd->data.bitfields |=1619 cpu_to_le16(1 << ETH_TX_DATA_3RD_BD_HDR_NBD_SHIFT);1620 1621 /* Make life easier for FW guys who can't deal with header and1622 * data on same BD. If we need to split, use the second bd...1623 */1624 if (unlikely(skb_headlen(skb) > hlen)) {1625 DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED,1626 "TSO split header size is %d (%x:%x)\n",1627 first_bd->nbytes, first_bd->addr.hi,1628 first_bd->addr.lo);1629 1630 mapping = HILO_U64(le32_to_cpu(first_bd->addr.hi),1631 le32_to_cpu(first_bd->addr.lo)) +1632 hlen;1633 1634 BD_SET_UNMAP_ADDR_LEN(tx_data_bd, mapping,1635 le16_to_cpu(first_bd->nbytes) -1636 hlen);1637 1638 /* this marks the BD as one that has no1639 * individual mapping1640 */1641 txq->sw_tx_ring.skbs[idx].flags |= QEDE_TSO_SPLIT_BD;1642 1643 first_bd->nbytes = cpu_to_le16(hlen);1644 1645 tx_data_bd = (struct eth_tx_bd *)third_bd;1646 data_split = true;1647 }1648 } else {1649 if (unlikely(skb->len > ETH_TX_MAX_NON_LSO_PKT_LEN)) {1650 DP_ERR(edev, "Unexpected non LSO skb length = 0x%x\n", skb->len);1651 qede_free_failed_tx_pkt(txq, first_bd, 0, false);1652 qede_update_tx_producer(txq);1653 return NETDEV_TX_OK;1654 }1655 1656 val |= ((skb->len & ETH_TX_DATA_1ST_BD_PKT_LEN_MASK) <<1657 ETH_TX_DATA_1ST_BD_PKT_LEN_SHIFT);1658 }1659 1660 first_bd->data.bitfields = cpu_to_le16(val);1661 1662 /* Handle fragmented skb */1663 /* special handle for frags inside 2nd and 3rd bds.. */1664 while (tx_data_bd && frag_idx < skb_shinfo(skb)->nr_frags) {1665 rc = map_frag_to_bd(txq,1666 &skb_shinfo(skb)->frags[frag_idx],1667 tx_data_bd);1668 if (rc) {1669 qede_free_failed_tx_pkt(txq, first_bd, nbd, data_split);1670 qede_update_tx_producer(txq);1671 return NETDEV_TX_OK;1672 }1673 1674 if (tx_data_bd == (struct eth_tx_bd *)second_bd)1675 tx_data_bd = (struct eth_tx_bd *)third_bd;1676 else1677 tx_data_bd = NULL;1678 1679 frag_idx++;1680 }1681 1682 /* map last frags into 4th, 5th .... */1683 for (; frag_idx < skb_shinfo(skb)->nr_frags; frag_idx++, nbd++) {1684 tx_data_bd = (struct eth_tx_bd *)1685 qed_chain_produce(&txq->tx_pbl);1686 1687 memset(tx_data_bd, 0, sizeof(*tx_data_bd));1688 1689 rc = map_frag_to_bd(txq,1690 &skb_shinfo(skb)->frags[frag_idx],1691 tx_data_bd);1692 if (rc) {1693 qede_free_failed_tx_pkt(txq, first_bd, nbd, data_split);1694 qede_update_tx_producer(txq);1695 return NETDEV_TX_OK;1696 }1697 }1698 1699 /* update the first BD with the actual num BDs */1700 first_bd->data.nbds = nbd;1701 1702 netdev_tx_sent_queue(netdev_txq, skb->len);1703 1704 skb_tx_timestamp(skb);1705 1706 /* Advance packet producer only before sending the packet since mapping1707 * of pages may fail.1708 */1709 txq->sw_tx_prod = (txq->sw_tx_prod + 1) % txq->num_tx_buffers;1710 1711 /* 'next page' entries are counted in the producer value */1712 txq->tx_db.data.bd_prod =1713 cpu_to_le16(qed_chain_get_prod_idx(&txq->tx_pbl));1714 1715 if (!netdev_xmit_more() || netif_xmit_stopped(netdev_txq))1716 qede_update_tx_producer(txq);1717 1718 if (unlikely(qed_chain_get_elem_left(&txq->tx_pbl)1719 < (MAX_SKB_FRAGS + 1))) {1720 if (netdev_xmit_more())1721 qede_update_tx_producer(txq);1722 1723 netif_tx_stop_queue(netdev_txq);1724 txq->stopped_cnt++;1725 DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED,1726 "Stop queue was called\n");1727 /* paired memory barrier is in qede_tx_int(), we have to keep1728 * ordering of set_bit() in netif_tx_stop_queue() and read of1729 * fp->bd_tx_cons1730 */1731 smp_mb();1732 1733 if ((qed_chain_get_elem_left(&txq->tx_pbl) >=1734 (MAX_SKB_FRAGS + 1)) &&1735 (edev->state == QEDE_STATE_OPEN)) {1736 netif_tx_wake_queue(netdev_txq);1737 DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED,1738 "Wake queue was called\n");1739 }1740 }1741 1742 return NETDEV_TX_OK;1743}1744 1745u16 qede_select_queue(struct net_device *dev, struct sk_buff *skb,1746 struct net_device *sb_dev)1747{1748 struct qede_dev *edev = netdev_priv(dev);1749 int total_txq;1750 1751 total_txq = QEDE_TSS_COUNT(edev) * edev->dev_info.num_tc;1752 1753 return QEDE_TSS_COUNT(edev) ?1754 netdev_pick_tx(dev, skb, NULL) % total_txq : 0;1755}1756 1757/* 8B udp header + 8B base tunnel header + 32B option length */1758#define QEDE_MAX_TUN_HDR_LEN 481759 1760netdev_features_t qede_features_check(struct sk_buff *skb,1761 struct net_device *dev,1762 netdev_features_t features)1763{1764 if (skb->encapsulation) {1765 u8 l4_proto = 0;1766 1767 switch (vlan_get_protocol(skb)) {1768 case htons(ETH_P_IP):1769 l4_proto = ip_hdr(skb)->protocol;1770 break;1771 case htons(ETH_P_IPV6):1772 l4_proto = ipv6_hdr(skb)->nexthdr;1773 break;1774 default:1775 return features;1776 }1777 1778 /* Disable offloads for geneve tunnels, as HW can't parse1779 * the geneve header which has option length greater than 32b1780 * and disable offloads for the ports which are not offloaded.1781 */1782 if (l4_proto == IPPROTO_UDP) {1783 struct qede_dev *edev = netdev_priv(dev);1784 u16 hdrlen, vxln_port, gnv_port;1785 1786 hdrlen = QEDE_MAX_TUN_HDR_LEN;1787 vxln_port = edev->vxlan_dst_port;1788 gnv_port = edev->geneve_dst_port;1789 1790 if ((skb_inner_mac_header(skb) -1791 skb_transport_header(skb)) > hdrlen ||1792 (ntohs(udp_hdr(skb)->dest) != vxln_port &&1793 ntohs(udp_hdr(skb)->dest) != gnv_port))1794 return features & ~(NETIF_F_CSUM_MASK |1795 NETIF_F_GSO_MASK);1796 } else if (l4_proto == IPPROTO_IPIP) {1797 /* IPIP tunnels are unknown to the device or at least unsupported natively,1798 * offloads for them can't be done trivially, so disable them for such skb.1799 */1800 return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);1801 }1802 }1803 1804 return features;1805}1806