| 1 |
// SPDX-License-Identifier: GPL-2.0 |
| 2 |
/* Marvell RVU Ethernet driver |
| 3 |
* |
| 4 |
* Copyright (C) 2020 Marvell. |
| 5 |
* |
| 6 |
*/ |
| 7 |
|
| 8 |
#include <linux/etherdevice.h> |
| 9 |
#include <net/ip.h> |
| 10 |
#include <net/tso.h> |
| 11 |
#include <linux/bpf.h> |
| 12 |
#include <linux/bpf_trace.h> |
| 13 |
#include <net/ip6_checksum.h> |
| 14 |
#include <net/xfrm.h> |
| 15 |
#include <net/xdp.h> |
| 16 |
|
| 17 |
#include "otx2_reg.h" |
| 18 |
#include "otx2_common.h" |
| 19 |
#include "otx2_struct.h" |
| 20 |
#include "otx2_txrx.h" |
| 21 |
#include "otx2_ptp.h" |
| 22 |
#include "cn10k.h" |
| 23 |
#include "otx2_xsk.h" |
| 24 |
|
| 25 |
#define CQE_ADDR(CQ, idx) ((CQ)->cqe_base + ((CQ)->cqe_size * (idx))) |
| 26 |
#define PTP_PORT 0x13F |
| 27 |
/* PTPv2 header Original Timestamp starts at byte offset 34 and |
| 28 |
* contains 6 byte seconds field and 4 byte nano seconds field. |
| 29 |
*/ |
| 30 |
#define PTP_SYNC_SEC_OFFSET 34 |
| 31 |
|
| 32 |
DEFINE_STATIC_KEY_FALSE(cn10k_ipsec_sa_enabled); |
| 33 |
|
| 34 |
static int otx2_get_free_sqe(struct otx2_snd_queue *sq) |
| 35 |
{ |
| 36 |
return (sq->cons_head - sq->head - 1 + sq->sqe_cnt) |
| 37 |
& (sq->sqe_cnt - 1); |
| 38 |
} |
| 39 |
|
| 40 |
static bool otx2_xdp_rcv_pkt_handler(struct otx2_nic *pfvf, |
| 41 |
struct bpf_prog *prog, |
| 42 |
struct nix_cqe_rx_s *cqe, |
| 43 |
struct otx2_cq_queue *cq, |
| 44 |
u32 *metasize, bool *need_xdp_flush); |
| 45 |
|
| 46 |
static void otx2_sq_set_sqe_base(struct otx2_snd_queue *sq, |
| 47 |
struct sk_buff *skb) |
| 48 |
{ |
| 49 |
if (static_branch_unlikely(&cn10k_ipsec_sa_enabled) && |
| 50 |
(xfrm_offload(skb))) |
| 51 |
sq->sqe_base = sq->sqe_ring->base + sq->sqe_size + |
| 52 |
(sq->head * (sq->sqe_size * 2)); |
| 53 |
else |
| 54 |
sq->sqe_base = sq->sqe->base; |
| 55 |
} |
| 56 |
|
| 57 |
static int otx2_nix_cq_op_status(struct otx2_nic *pfvf, |
| 58 |
struct otx2_cq_queue *cq) |
| 59 |
{ |
| 60 |
u64 incr = (u64)(cq->cq_idx) << 32; |
| 61 |
u64 status; |
| 62 |
|
| 63 |
status = otx2_atomic64_fetch_add(incr, pfvf->cq_op_addr); |
| 64 |
|
| 65 |
if (unlikely(status & BIT_ULL(CQ_OP_STAT_OP_ERR) || |
| 66 |
status & BIT_ULL(CQ_OP_STAT_CQ_ERR))) { |
| 67 |
dev_err(pfvf->dev, "CQ stopped due to error"); |
| 68 |
return -EINVAL; |
| 69 |
} |
| 70 |
|
| 71 |
cq->cq_tail = status & 0xFFFFF; |
| 72 |
cq->cq_head = (status >> 20) & 0xFFFFF; |
| 73 |
if (cq->cq_tail < cq->cq_head) |
| 74 |
cq->pend_cqe = (cq->cqe_cnt - cq->cq_head) + |
| 75 |
cq->cq_tail; |
| 76 |
else |
| 77 |
cq->pend_cqe = cq->cq_tail - cq->cq_head; |
| 78 |
|
| 79 |
return 0; |
| 80 |
} |
| 81 |
|
| 82 |
static struct nix_cqe_hdr_s *otx2_get_next_cqe(struct otx2_cq_queue *cq) |
| 83 |
{ |
| 84 |
struct nix_cqe_hdr_s *cqe_hdr; |
| 85 |
|
| 86 |
cqe_hdr = (struct nix_cqe_hdr_s *)CQE_ADDR(cq, cq->cq_head); |
| 87 |
if (cqe_hdr->cqe_type == NIX_XQE_TYPE_INVALID) |
| 88 |
return NULL; |
| 89 |
|
| 90 |
cq->cq_head++; |
| 91 |
cq->cq_head &= (cq->cqe_cnt - 1); |
| 92 |
|
| 93 |
return cqe_hdr; |
| 94 |
} |
| 95 |
|
| 96 |
static unsigned int frag_num(unsigned int i) |
| 97 |
{ |
| 98 |
#ifdef __BIG_ENDIAN |
| 99 |
return (i & ~3) + 3 - (i & 3); |
| 100 |
#else |
| 101 |
return i; |
| 102 |
#endif |
| 103 |
} |
| 104 |
|
| 105 |
static void otx2_xdp_snd_pkt_handler(struct otx2_nic *pfvf, |
| 106 |
struct otx2_snd_queue *sq, |
| 107 |
struct nix_cqe_tx_s *cqe, |
| 108 |
int *xsk_frames) |
| 109 |
{ |
| 110 |
struct nix_send_comp_s *snd_comp = &cqe->comp; |
| 111 |
struct sg_list *sg; |
| 112 |
|
| 113 |
sg = &sq->sg[snd_comp->sqe_id]; |
| 114 |
if (sg->flags & OTX2_AF_XDP_FRAME) { |
| 115 |
(*xsk_frames)++; |
| 116 |
return; |
| 117 |
} |
| 118 |
|
| 119 |
if (sg->flags & OTX2_XDP_REDIRECT) |
| 120 |
otx2_dma_unmap_page(pfvf, sg->dma_addr[0], sg->size[0], DMA_TO_DEVICE); |
| 121 |
xdp_return_frame((struct xdp_frame *)sg->skb); |
| 122 |
sg->skb = (u64)NULL; |
| 123 |
} |
| 124 |
|
| 125 |
static void otx2_snd_pkt_handler(struct otx2_nic *pfvf, |
| 126 |
struct otx2_cq_queue *cq, |
| 127 |
struct otx2_snd_queue *sq, |
| 128 |
struct nix_cqe_tx_s *cqe, |
| 129 |
int budget, int *tx_pkts, int *tx_bytes) |
| 130 |
{ |
| 131 |
struct nix_send_comp_s *snd_comp = &cqe->comp; |
| 132 |
struct skb_shared_hwtstamps ts; |
| 133 |
struct sk_buff *skb = NULL; |
| 134 |
u64 timestamp, tsns; |
| 135 |
struct sg_list *sg; |
| 136 |
int err; |
| 137 |
|
| 138 |
if (unlikely(snd_comp->status) && netif_msg_tx_err(pfvf)) |
| 139 |
net_err_ratelimited("%s: TX%d: Error in send CQ status:%x\n", |
| 140 |
pfvf->netdev->name, cq->cint_idx, |
| 141 |
snd_comp->status); |
| 142 |
|
| 143 |
sg = &sq->sg[snd_comp->sqe_id]; |
| 144 |
skb = (struct sk_buff *)sg->skb; |
| 145 |
if (unlikely(!skb)) |
| 146 |
return; |
| 147 |
|
| 148 |
if (skb_shinfo(skb)->tx_flags & SKBTX_IN_PROGRESS) { |
| 149 |
timestamp = ((u64 *)sq->timestamps->base)[snd_comp->sqe_id]; |
| 150 |
if (timestamp != 1) { |
| 151 |
timestamp = pfvf->ptp->convert_tx_ptp_tstmp(timestamp); |
| 152 |
err = otx2_ptp_tstamp2time(pfvf, timestamp, &tsns); |
| 153 |
if (!err) { |
| 154 |
memset(&ts, 0, sizeof(ts)); |
| 155 |
ts.hwtstamp = ns_to_ktime(tsns); |
| 156 |
skb_tstamp_tx(skb, &ts); |
| 157 |
} |
| 158 |
} |
| 159 |
} |
| 160 |
|
| 161 |
*tx_bytes += skb->len; |
| 162 |
(*tx_pkts)++; |
| 163 |
otx2_dma_unmap_skb_frags(pfvf, sg); |
| 164 |
napi_consume_skb(skb, budget); |
| 165 |
sg->skb = (u64)NULL; |
| 166 |
} |
| 167 |
|
| 168 |
static void otx2_set_rxtstamp(struct otx2_nic *pfvf, |
| 169 |
struct sk_buff *skb, void *data) |
| 170 |
{ |
| 171 |
u64 timestamp, tsns; |
| 172 |
int err; |
| 173 |
|
| 174 |
if (!(pfvf->flags & OTX2_FLAG_RX_TSTAMP_ENABLED)) |
| 175 |
return; |
| 176 |
|
| 177 |
timestamp = pfvf->ptp->convert_rx_ptp_tstmp(*(u64 *)data); |
| 178 |
/* The first 8 bytes is the timestamp */ |
| 179 |
err = otx2_ptp_tstamp2time(pfvf, timestamp, &tsns); |
| 180 |
if (err) |
| 181 |
return; |
| 182 |
|
| 183 |
skb_hwtstamps(skb)->hwtstamp = ns_to_ktime(tsns); |
| 184 |
} |
| 185 |
|
| 186 |
static bool otx2_skb_add_frag(struct otx2_nic *pfvf, struct sk_buff *skb, |
| 187 |
u64 iova, int len, struct nix_rx_parse_s *parse, |
| 188 |
int qidx) |
| 189 |
{ |
| 190 |
struct page *page; |
| 191 |
int off = 0; |
| 192 |
void *va; |
| 193 |
|
| 194 |
va = phys_to_virt(otx2_iova_to_phys(pfvf->iommu_domain, iova)); |
| 195 |
|
| 196 |
if (likely(!skb_shinfo(skb)->nr_frags)) { |
| 197 |
/* Check if data starts at some nonzero offset |
| 198 |
* from the start of the buffer. For now the |
| 199 |
* only possible offset is 8 bytes in the case |
| 200 |
* where packet is prepended by a timestamp. |
| 201 |
*/ |
| 202 |
if (parse->laptr) { |
| 203 |
otx2_set_rxtstamp(pfvf, skb, va); |
| 204 |
off = OTX2_HW_TIMESTAMP_LEN; |
| 205 |
} |
| 206 |
} |
| 207 |
|
| 208 |
page = virt_to_page(va); |
| 209 |
if (likely(skb_shinfo(skb)->nr_frags < MAX_SKB_FRAGS)) { |
| 210 |
skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, |
| 211 |
va - page_address(page) + off, |
| 212 |
len - off, pfvf->rbsize); |
| 213 |
return true; |
| 214 |
} |
| 215 |
|
| 216 |
/* If more than MAX_SKB_FRAGS fragments are received then |
| 217 |
* give back those buffer pointers to hardware for reuse. |
| 218 |
*/ |
| 219 |
pfvf->hw_ops->aura_freeptr(pfvf, qidx, iova & ~0x07ULL); |
| 220 |
|
| 221 |
return false; |
| 222 |
} |
| 223 |
|
| 224 |
static void otx2_set_rxhash(struct otx2_nic *pfvf, |
| 225 |
struct nix_cqe_rx_s *cqe, struct sk_buff *skb) |
| 226 |
{ |
| 227 |
enum pkt_hash_types hash_type = PKT_HASH_TYPE_NONE; |
| 228 |
struct otx2_rss_info *rss; |
| 229 |
u32 hash = 0; |
| 230 |
|
| 231 |
if (!(pfvf->netdev->features & NETIF_F_RXHASH)) |
| 232 |
return; |
| 233 |
|
| 234 |
rss = &pfvf->hw.rss_info; |
| 235 |
if (rss->flowkey_cfg) { |
| 236 |
if (rss->flowkey_cfg & |
| 237 |
~(NIX_FLOW_KEY_TYPE_IPV4 | NIX_FLOW_KEY_TYPE_IPV6)) |
| 238 |
hash_type = PKT_HASH_TYPE_L4; |
| 239 |
else |
| 240 |
hash_type = PKT_HASH_TYPE_L3; |
| 241 |
hash = cqe->hdr.flow_tag; |
| 242 |
} |
| 243 |
skb_set_hash(skb, hash, hash_type); |
| 244 |
} |
| 245 |
|
| 246 |
static void otx2_free_rcv_seg(struct otx2_nic *pfvf, struct nix_cqe_rx_s *cqe, |
| 247 |
int qidx) |
| 248 |
{ |
| 249 |
struct nix_rx_sg_s *sg = &cqe->sg; |
| 250 |
void *end, *start; |
| 251 |
u64 *seg_addr; |
| 252 |
int seg; |
| 253 |
|
| 254 |
start = (void *)sg; |
| 255 |
end = start + ((cqe->parse.desc_sizem1 + 1) * 16); |
| 256 |
while (start < end) { |
| 257 |
sg = (struct nix_rx_sg_s *)start; |
| 258 |
seg_addr = &sg->seg_addr; |
| 259 |
for (seg = 0; seg < sg->segs; seg++, seg_addr++) |
| 260 |
pfvf->hw_ops->aura_freeptr(pfvf, qidx, |
| 261 |
*seg_addr & ~0x07ULL); |
| 262 |
start += sizeof(*sg); |
| 263 |
} |
| 264 |
} |
| 265 |
|
| 266 |
static bool otx2_check_rcv_errors(struct otx2_nic *pfvf, |
| 267 |
struct nix_cqe_rx_s *cqe, int qidx) |
| 268 |
{ |
| 269 |
struct otx2_drv_stats *stats = &pfvf->hw.drv_stats; |
| 270 |
struct nix_rx_parse_s *parse = &cqe->parse; |
| 271 |
|
| 272 |
if (netif_msg_rx_err(pfvf)) |
| 273 |
netdev_err(pfvf->netdev, |
| 274 |
"RQ%d: Error pkt with errlev:0x%x errcode:0x%x\n", |
| 275 |
qidx, parse->errlev, parse->errcode); |
| 276 |
|
| 277 |
if (parse->errlev == NPC_ERRLVL_RE) { |
| 278 |
switch (parse->errcode) { |
| 279 |
case ERRCODE_FCS: |
| 280 |
case ERRCODE_FCS_RCV: |
| 281 |
atomic_inc(&stats->rx_fcs_errs); |
| 282 |
break; |
| 283 |
case ERRCODE_UNDERSIZE: |
| 284 |
atomic_inc(&stats->rx_undersize_errs); |
| 285 |
break; |
| 286 |
case ERRCODE_OVERSIZE: |
| 287 |
atomic_inc(&stats->rx_oversize_errs); |
| 288 |
break; |
| 289 |
case ERRCODE_OL2_LEN_MISMATCH: |
| 290 |
atomic_inc(&stats->rx_len_errs); |
| 291 |
break; |
| 292 |
default: |
| 293 |
atomic_inc(&stats->rx_other_errs); |
| 294 |
break; |
| 295 |
} |
| 296 |
} else if (parse->errlev == NPC_ERRLVL_NIX) { |
| 297 |
switch (parse->errcode) { |
| 298 |
case ERRCODE_OL3_LEN: |
| 299 |
case ERRCODE_OL4_LEN: |
| 300 |
case ERRCODE_IL3_LEN: |
| 301 |
case ERRCODE_IL4_LEN: |
| 302 |
atomic_inc(&stats->rx_len_errs); |
| 303 |
break; |
| 304 |
case ERRCODE_OL4_CSUM: |
| 305 |
case ERRCODE_IL4_CSUM: |
| 306 |
atomic_inc(&stats->rx_csum_errs); |
| 307 |
break; |
| 308 |
default: |
| 309 |
atomic_inc(&stats->rx_other_errs); |
| 310 |
break; |
| 311 |
} |
| 312 |
} else { |
| 313 |
atomic_inc(&stats->rx_other_errs); |
| 314 |
/* For now ignore all the NPC parser errors and |
| 315 |
* pass the packets to stack. |
| 316 |
*/ |
| 317 |
return false; |
| 318 |
} |
| 319 |
|
| 320 |
/* If RXALL is enabled pass on packets to stack. */ |
| 321 |
if (pfvf->netdev->features & NETIF_F_RXALL) |
| 322 |
return false; |
| 323 |
|
| 324 |
/* Free buffer back to pool */ |
| 325 |
if (cqe->sg.segs) |
| 326 |
otx2_free_rcv_seg(pfvf, cqe, qidx); |
| 327 |
return true; |
| 328 |
} |
| 329 |
|
| 330 |
static void otx2_rcv_pkt_handler(struct otx2_nic *pfvf, |
| 331 |
struct napi_struct *napi, |
| 332 |
struct otx2_cq_queue *cq, |
| 333 |
struct nix_cqe_rx_s *cqe, bool *need_xdp_flush) |
| 334 |
{ |
| 335 |
struct nix_rx_parse_s *parse = &cqe->parse; |
| 336 |
struct nix_rx_sg_s *sg = &cqe->sg; |
| 337 |
struct sk_buff *skb = NULL; |
| 338 |
u64 *word = (u64 *)parse; |
| 339 |
void *end, *start; |
| 340 |
u32 metasize = 0; |
| 341 |
u64 *seg_addr; |
| 342 |
u16 *seg_size; |
| 343 |
int seg; |
| 344 |
|
| 345 |
if (unlikely(parse->errlev || parse->errcode)) { |
| 346 |
if (otx2_check_rcv_errors(pfvf, cqe, cq->cq_idx)) { |
| 347 |
trace_otx2_parse_dump(pfvf->pdev, "Err:", word); |
| 348 |
return; |
| 349 |
} |
| 350 |
} |
| 351 |
trace_otx2_parse_dump(pfvf->pdev, "", word); |
| 352 |
|
| 353 |
if (pfvf->xdp_prog) |
| 354 |
if (otx2_xdp_rcv_pkt_handler(pfvf, pfvf->xdp_prog, cqe, cq, |
| 355 |
&metasize, need_xdp_flush)) |
| 356 |
return; |
| 357 |
|
| 358 |
skb = napi_get_frags(napi); |
| 359 |
if (unlikely(!skb)) |
| 360 |
return; |
| 361 |
|
| 362 |
start = (void *)sg; |
| 363 |
end = start + ((cqe->parse.desc_sizem1 + 1) * 16); |
| 364 |
while (start < end) { |
| 365 |
sg = (struct nix_rx_sg_s *)start; |
| 366 |
seg_addr = &sg->seg_addr; |
| 367 |
seg_size = (void *)sg; |
| 368 |
for (seg = 0; seg < sg->segs; seg++, seg_addr++) { |
| 369 |
if (otx2_skb_add_frag(pfvf, skb, *seg_addr, |
| 370 |
seg_size[seg], parse, cq->cq_idx)) |
| 371 |
cq->pool_ptrs++; |
| 372 |
} |
| 373 |
start += sizeof(*sg); |
| 374 |
} |
| 375 |
otx2_set_rxhash(pfvf, cqe, skb); |
| 376 |
|
| 377 |
if (!(pfvf->flags & OTX2_FLAG_REP_MODE_ENABLED)) { |
| 378 |
skb_record_rx_queue(skb, cq->cq_idx); |
| 379 |
if (pfvf->netdev->features & NETIF_F_RXCSUM) |
| 380 |
skb->ip_summed = CHECKSUM_UNNECESSARY; |
| 381 |
} |
| 382 |
|
| 383 |
if (pfvf->flags & OTX2_FLAG_TC_MARK_ENABLED) |
| 384 |
skb->mark = parse->match_id; |
| 385 |
|
| 386 |
skb_mark_for_recycle(skb); |
| 387 |
if (metasize) |
| 388 |
skb_metadata_set(skb, metasize); |
| 389 |
|
| 390 |
napi_gro_frags(napi); |
| 391 |
} |
| 392 |
|
| 393 |
static int otx2_rx_napi_handler(struct otx2_nic *pfvf, |
| 394 |
struct napi_struct *napi, |
| 395 |
struct otx2_cq_queue *cq, int budget) |
| 396 |
{ |
| 397 |
bool need_xdp_flush = false; |
| 398 |
struct nix_cqe_rx_s *cqe; |
| 399 |
int processed_cqe = 0; |
| 400 |
|
| 401 |
if (cq->pend_cqe >= budget) |
| 402 |
goto process_cqe; |
| 403 |
|
| 404 |
if (otx2_nix_cq_op_status(pfvf, cq) || !cq->pend_cqe) |
| 405 |
return 0; |
| 406 |
|
| 407 |
process_cqe: |
| 408 |
while (likely(processed_cqe < budget) && cq->pend_cqe) { |
| 409 |
cqe = (struct nix_cqe_rx_s *)CQE_ADDR(cq, cq->cq_head); |
| 410 |
if (cqe->hdr.cqe_type == NIX_XQE_TYPE_INVALID || |
| 411 |
!cqe->sg.seg_addr) { |
| 412 |
if (!processed_cqe) |
| 413 |
return 0; |
| 414 |
break; |
| 415 |
} |
| 416 |
cq->cq_head++; |
| 417 |
cq->cq_head &= (cq->cqe_cnt - 1); |
| 418 |
|
| 419 |
otx2_rcv_pkt_handler(pfvf, napi, cq, cqe, &need_xdp_flush); |
| 420 |
|
| 421 |
cqe->hdr.cqe_type = NIX_XQE_TYPE_INVALID; |
| 422 |
cqe->sg.seg_addr = 0x00; |
| 423 |
processed_cqe++; |
| 424 |
cq->pend_cqe--; |
| 425 |
} |
| 426 |
if (need_xdp_flush) |
| 427 |
xdp_do_flush(); |
| 428 |
|
| 429 |
/* Free CQEs to HW */ |
| 430 |
otx2_write64(pfvf, NIX_LF_CQ_OP_DOOR, |
| 431 |
((u64)cq->cq_idx << 32) | processed_cqe); |
| 432 |
|
| 433 |
return processed_cqe; |
| 434 |
} |
| 435 |
|
| 436 |
int otx2_refill_pool_ptrs(void *dev, struct otx2_cq_queue *cq) |
| 437 |
{ |
| 438 |
struct otx2_nic *pfvf = dev; |
| 439 |
int cnt = cq->pool_ptrs; |
| 440 |
dma_addr_t bufptr; |
| 441 |
|
| 442 |
while (cq->pool_ptrs) { |
| 443 |
if (otx2_alloc_buffer(pfvf, cq, &bufptr)) |
| 444 |
break; |
| 445 |
otx2_aura_freeptr(pfvf, cq->cq_idx, bufptr + OTX2_HEAD_ROOM); |
| 446 |
cq->pool_ptrs--; |
| 447 |
} |
| 448 |
|
| 449 |
return cnt - cq->pool_ptrs; |
| 450 |
} |
| 451 |
|
| 452 |
static void otx2_zc_submit_pkts(struct otx2_nic *pfvf, struct xsk_buff_pool *xsk_pool, |
| 453 |
int *xsk_frames, int qidx, int budget) |
| 454 |
{ |
| 455 |
if (*xsk_frames) |
| 456 |
xsk_tx_completed(xsk_pool, *xsk_frames); |
| 457 |
|
| 458 |
if (xsk_uses_need_wakeup(xsk_pool)) |
| 459 |
xsk_set_tx_need_wakeup(xsk_pool); |
| 460 |
|
| 461 |
otx2_zc_napi_handler(pfvf, xsk_pool, qidx, budget); |
| 462 |
} |
| 463 |
|
| 464 |
static int otx2_tx_napi_handler(struct otx2_nic *pfvf, |
| 465 |
struct otx2_cq_queue *cq, int budget) |
| 466 |
{ |
| 467 |
int tx_pkts = 0, tx_bytes = 0, qidx; |
| 468 |
struct otx2_snd_queue *sq; |
| 469 |
struct nix_cqe_tx_s *cqe; |
| 470 |
struct net_device *ndev; |
| 471 |
int processed_cqe = 0; |
| 472 |
int xsk_frames = 0; |
| 473 |
|
| 474 |
qidx = cq->cq_idx - pfvf->hw.rx_queues; |
| 475 |
sq = &pfvf->qset.sq[qidx]; |
| 476 |
|
| 477 |
if (cq->pend_cqe >= budget) |
| 478 |
goto process_cqe; |
| 479 |
|
| 480 |
if (otx2_nix_cq_op_status(pfvf, cq) || !cq->pend_cqe) { |
| 481 |
if (sq->xsk_pool) |
| 482 |
otx2_zc_submit_pkts(pfvf, sq->xsk_pool, &xsk_frames, |
| 483 |
qidx, budget); |
| 484 |
return 0; |
| 485 |
} |
| 486 |
|
| 487 |
process_cqe: |
| 488 |
|
| 489 |
while (likely(processed_cqe < budget) && cq->pend_cqe) { |
| 490 |
cqe = (struct nix_cqe_tx_s *)otx2_get_next_cqe(cq); |
| 491 |
if (unlikely(!cqe)) { |
| 492 |
if (!processed_cqe) |
| 493 |
return 0; |
| 494 |
break; |
| 495 |
} |
| 496 |
|
| 497 |
if (cq->cq_type == CQ_XDP) |
| 498 |
otx2_xdp_snd_pkt_handler(pfvf, sq, cqe, &xsk_frames); |
| 499 |
else |
| 500 |
otx2_snd_pkt_handler(pfvf, cq, &pfvf->qset.sq[qidx], |
| 501 |
cqe, budget, &tx_pkts, &tx_bytes); |
| 502 |
|
| 503 |
cqe->hdr.cqe_type = NIX_XQE_TYPE_INVALID; |
| 504 |
processed_cqe++; |
| 505 |
cq->pend_cqe--; |
| 506 |
|
| 507 |
sq->cons_head++; |
| 508 |
sq->cons_head &= (sq->sqe_cnt - 1); |
| 509 |
} |
| 510 |
|
| 511 |
/* Free CQEs to HW */ |
| 512 |
otx2_write64(pfvf, NIX_LF_CQ_OP_DOOR, |
| 513 |
((u64)cq->cq_idx << 32) | processed_cqe); |
| 514 |
|
| 515 |
#if IS_ENABLED(CONFIG_RVU_ESWITCH) |
| 516 |
if (pfvf->flags & OTX2_FLAG_REP_MODE_ENABLED) |
| 517 |
ndev = pfvf->reps[qidx]->netdev; |
| 518 |
else |
| 519 |
#endif |
| 520 |
ndev = pfvf->netdev; |
| 521 |
|
| 522 |
if (likely(tx_pkts)) { |
| 523 |
struct netdev_queue *txq; |
| 524 |
|
| 525 |
qidx = cq->cq_idx - pfvf->hw.rx_queues; |
| 526 |
|
| 527 |
if (qidx >= pfvf->hw.tx_queues) |
| 528 |
qidx -= pfvf->hw.xdp_queues; |
| 529 |
if (pfvf->flags & OTX2_FLAG_REP_MODE_ENABLED) |
| 530 |
qidx = 0; |
| 531 |
txq = netdev_get_tx_queue(ndev, qidx); |
| 532 |
netdev_tx_completed_queue(txq, tx_pkts, tx_bytes); |
| 533 |
/* Check if queue was stopped earlier due to ring full */ |
| 534 |
smp_mb(); |
| 535 |
if (netif_tx_queue_stopped(txq) && |
| 536 |
netif_carrier_ok(ndev)) |
| 537 |
netif_tx_wake_queue(txq); |
| 538 |
} |
| 539 |
|
| 540 |
if (sq->xsk_pool) |
| 541 |
otx2_zc_submit_pkts(pfvf, sq->xsk_pool, &xsk_frames, qidx, budget); |
| 542 |
|
| 543 |
return 0; |
| 544 |
} |
| 545 |
|
| 546 |
static void otx2_adjust_adaptive_coalese(struct otx2_nic *pfvf, struct otx2_cq_poll *cq_poll) |
| 547 |
{ |
| 548 |
struct dim_sample dim_sample = { 0 }; |
| 549 |
u64 rx_frames, rx_bytes; |
| 550 |
u64 tx_frames, tx_bytes; |
| 551 |
|
| 552 |
rx_frames = OTX2_GET_RX_STATS(RX_BCAST) + OTX2_GET_RX_STATS(RX_MCAST) + |
| 553 |
OTX2_GET_RX_STATS(RX_UCAST); |
| 554 |
rx_bytes = OTX2_GET_RX_STATS(RX_OCTS); |
| 555 |
tx_bytes = OTX2_GET_TX_STATS(TX_OCTS); |
| 556 |
tx_frames = OTX2_GET_TX_STATS(TX_UCAST); |
| 557 |
|
| 558 |
dim_update_sample(pfvf->napi_events, |
| 559 |
rx_frames + tx_frames, |
| 560 |
rx_bytes + tx_bytes, |
| 561 |
&dim_sample); |
| 562 |
net_dim(&cq_poll->dim, &dim_sample); |
| 563 |
} |
| 564 |
|
| 565 |
int otx2_napi_handler(struct napi_struct *napi, int budget) |
| 566 |
{ |
| 567 |
struct otx2_cq_queue *rx_cq = NULL; |
| 568 |
struct otx2_cq_queue *cq = NULL; |
| 569 |
struct otx2_pool *pool = NULL; |
| 570 |
struct otx2_cq_poll *cq_poll; |
| 571 |
int workdone = 0, cq_idx, i; |
| 572 |
struct otx2_qset *qset; |
| 573 |
struct otx2_nic *pfvf; |
| 574 |
int filled_cnt = -1; |
| 575 |
|
| 576 |
cq_poll = container_of(napi, struct otx2_cq_poll, napi); |
| 577 |
pfvf = (struct otx2_nic *)cq_poll->dev; |
| 578 |
qset = &pfvf->qset; |
| 579 |
|
| 580 |
for (i = 0; i < CQS_PER_CINT; i++) { |
| 581 |
cq_idx = cq_poll->cq_ids[i]; |
| 582 |
if (unlikely(cq_idx == CINT_INVALID_CQ)) |
| 583 |
continue; |
| 584 |
cq = &qset->cq[cq_idx]; |
| 585 |
if (cq->cq_type == CQ_RX) { |
| 586 |
rx_cq = cq; |
| 587 |
workdone += otx2_rx_napi_handler(pfvf, napi, |
| 588 |
cq, budget); |
| 589 |
} else { |
| 590 |
workdone += otx2_tx_napi_handler(pfvf, cq, budget); |
| 591 |
} |
| 592 |
} |
| 593 |
|
| 594 |
if (rx_cq && rx_cq->pool_ptrs) |
| 595 |
filled_cnt = pfvf->hw_ops->refill_pool_ptrs(pfvf, rx_cq); |
| 596 |
|
| 597 |
/* Clear the IRQ */ |
| 598 |
otx2_write64(pfvf, NIX_LF_CINTX_INT(cq_poll->cint_idx), BIT_ULL(0)); |
| 599 |
|
| 600 |
if (workdone < budget && napi_complete_done(napi, workdone)) { |
| 601 |
/* If interface is going down, don't re-enable IRQ */ |
| 602 |
if (pfvf->flags & OTX2_FLAG_INTF_DOWN) |
| 603 |
return workdone; |
| 604 |
|
| 605 |
/* Adjust irq coalese using net_dim */ |
| 606 |
if (pfvf->flags & OTX2_FLAG_ADPTV_INT_COAL_ENABLED) |
| 607 |
otx2_adjust_adaptive_coalese(pfvf, cq_poll); |
| 608 |
|
| 609 |
if (likely(cq)) |
| 610 |
pool = &pfvf->qset.pool[cq->cq_idx]; |
| 611 |
|
| 612 |
if (unlikely(!filled_cnt)) { |
| 613 |
struct refill_work *work; |
| 614 |
struct delayed_work *dwork; |
| 615 |
|
| 616 |
if (likely(cq)) { |
| 617 |
work = &pfvf->refill_wrk[cq->cq_idx]; |
| 618 |
dwork = &work->pool_refill_work; |
| 619 |
/* Schedule a task if no other task is running */ |
| 620 |
if (!cq->refill_task_sched) { |
| 621 |
work->napi = napi; |
| 622 |
cq->refill_task_sched = true; |
| 623 |
schedule_delayed_work(dwork, |
| 624 |
msecs_to_jiffies(100)); |
| 625 |
} |
| 626 |
/* Call wake-up for not able to fill buffers */ |
| 627 |
if (pool->xsk_pool) |
| 628 |
xsk_set_rx_need_wakeup(pool->xsk_pool); |
| 629 |
} |
| 630 |
} else { |
| 631 |
/* Clear wake-up, since buffers are filled successfully */ |
| 632 |
if (pool && pool->xsk_pool) |
| 633 |
xsk_clear_rx_need_wakeup(pool->xsk_pool); |
| 634 |
/* Re-enable interrupts */ |
| 635 |
otx2_write64(pfvf, |
| 636 |
NIX_LF_CINTX_ENA_W1S(cq_poll->cint_idx), |
| 637 |
BIT_ULL(0)); |
| 638 |
} |
| 639 |
} |
| 640 |
return workdone; |
| 641 |
} |
| 642 |
EXPORT_SYMBOL(otx2_napi_handler); |
| 643 |
|
| 644 |
void otx2_sqe_flush(void *dev, struct otx2_snd_queue *sq, |
| 645 |
int size, int qidx) |
| 646 |
{ |
| 647 |
u64 status; |
| 648 |
|
| 649 |
/* Packet data stores should finish before SQE is flushed to HW */ |
| 650 |
dma_wmb(); |
| 651 |
|
| 652 |
do { |
| 653 |
memcpy(sq->lmt_addr, sq->sqe_base, size); |
| 654 |
status = otx2_lmt_flush(sq->io_addr); |
| 655 |
} while (status == 0); |
| 656 |
|
| 657 |
sq->head++; |
| 658 |
sq->head &= (sq->sqe_cnt - 1); |
| 659 |
} |
| 660 |
|
| 661 |
/* Add SQE scatter/gather subdescriptor structure */ |
| 662 |
static bool otx2_sqe_add_sg(struct otx2_nic *pfvf, struct otx2_snd_queue *sq, |
| 663 |
struct sk_buff *skb, int num_segs, int *offset) |
| 664 |
{ |
| 665 |
struct nix_sqe_sg_s *sg = NULL; |
| 666 |
u64 dma_addr, *iova = NULL; |
| 667 |
u16 *sg_lens = NULL; |
| 668 |
int seg, len; |
| 669 |
|
| 670 |
sq->sg[sq->head].num_segs = 0; |
| 671 |
|
| 672 |
for (seg = 0; seg < num_segs; seg++) { |
| 673 |
if ((seg % MAX_SEGS_PER_SG) == 0) { |
| 674 |
sg = (struct nix_sqe_sg_s *)(sq->sqe_base + *offset); |
| 675 |
sg->ld_type = NIX_SEND_LDTYPE_LDD; |
| 676 |
sg->subdc = NIX_SUBDC_SG; |
| 677 |
sg->segs = 0; |
| 678 |
sg_lens = (void *)sg; |
| 679 |
iova = (void *)sg + sizeof(*sg); |
| 680 |
/* Next subdc always starts at a 16byte boundary. |
| 681 |
* So if sg->segs is whether 2 or 3, offset += 16bytes. |
| 682 |
*/ |
| 683 |
if ((num_segs - seg) >= (MAX_SEGS_PER_SG - 1)) |
| 684 |
*offset += sizeof(*sg) + (3 * sizeof(u64)); |
| 685 |
else |
| 686 |
*offset += sizeof(*sg) + sizeof(u64); |
| 687 |
} |
| 688 |
dma_addr = otx2_dma_map_skb_frag(pfvf, skb, seg, &len); |
| 689 |
if (dma_mapping_error(pfvf->dev, dma_addr)) |
| 690 |
return false; |
| 691 |
|
| 692 |
sg_lens[frag_num(seg % MAX_SEGS_PER_SG)] = len; |
| 693 |
sg->segs++; |
| 694 |
*iova++ = dma_addr; |
| 695 |
|
| 696 |
/* Save DMA mapping info for later unmapping */ |
| 697 |
sq->sg[sq->head].dma_addr[seg] = dma_addr; |
| 698 |
sq->sg[sq->head].size[seg] = len; |
| 699 |
sq->sg[sq->head].num_segs++; |
| 700 |
} |
| 701 |
|
| 702 |
sq->sg[sq->head].skb = (u64)skb; |
| 703 |
return true; |
| 704 |
} |
| 705 |
|
| 706 |
/* Add SQE extended header subdescriptor */ |
| 707 |
static void otx2_sqe_add_ext(struct otx2_nic *pfvf, struct otx2_snd_queue *sq, |
| 708 |
struct sk_buff *skb, int *offset) |
| 709 |
{ |
| 710 |
struct nix_sqe_ext_s *ext; |
| 711 |
|
| 712 |
ext = (struct nix_sqe_ext_s *)(sq->sqe_base + *offset); |
| 713 |
ext->subdc = NIX_SUBDC_EXT; |
| 714 |
if (skb_shinfo(skb)->gso_size) { |
| 715 |
ext->lso = 1; |
| 716 |
ext->lso_sb = skb_tcp_all_headers(skb); |
| 717 |
ext->lso_mps = skb_shinfo(skb)->gso_size; |
| 718 |
|
| 719 |
/* Only TSOv4 and TSOv6 GSO offloads are supported */ |
| 720 |
if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) { |
| 721 |
ext->lso_format = pfvf->hw.lso_tsov4_idx; |
| 722 |
|
| 723 |
/* HW adds payload size to 'ip_hdr->tot_len' while |
| 724 |
* sending TSO segment, hence set payload length |
| 725 |
* in IP header of the packet to just header length. |
| 726 |
*/ |
| 727 |
ip_hdr(skb)->tot_len = |
| 728 |
htons(ext->lso_sb - skb_network_offset(skb)); |
| 729 |
} else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6) { |
| 730 |
ext->lso_format = pfvf->hw.lso_tsov6_idx; |
| 731 |
ipv6_hdr(skb)->payload_len = htons(tcp_hdrlen(skb)); |
| 732 |
} else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) { |
| 733 |
__be16 l3_proto = vlan_get_protocol(skb); |
| 734 |
struct udphdr *udph = udp_hdr(skb); |
| 735 |
__be16 iplen; |
| 736 |
|
| 737 |
ext->lso_sb = skb_transport_offset(skb) + |
| 738 |
sizeof(struct udphdr); |
| 739 |
|
| 740 |
/* HW adds payload size to length fields in IP and |
| 741 |
* UDP headers while segmentation, hence adjust the |
| 742 |
* lengths to just header sizes. |
| 743 |
*/ |
| 744 |
iplen = htons(ext->lso_sb - skb_network_offset(skb)); |
| 745 |
if (l3_proto == htons(ETH_P_IP)) { |
| 746 |
ip_hdr(skb)->tot_len = iplen; |
| 747 |
ext->lso_format = pfvf->hw.lso_udpv4_idx; |
| 748 |
} else { |
| 749 |
ipv6_hdr(skb)->payload_len = iplen; |
| 750 |
ext->lso_format = pfvf->hw.lso_udpv6_idx; |
| 751 |
} |
| 752 |
|
| 753 |
udph->len = htons(sizeof(struct udphdr)); |
| 754 |
} |
| 755 |
} else if (skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP) { |
| 756 |
ext->tstmp = 1; |
| 757 |
} |
| 758 |
|
| 759 |
#define OTX2_VLAN_PTR_OFFSET (ETH_HLEN - ETH_TLEN) |
| 760 |
if (skb_vlan_tag_present(skb)) { |
| 761 |
if (skb->vlan_proto == htons(ETH_P_8021Q)) { |
| 762 |
ext->vlan1_ins_ena = 1; |
| 763 |
ext->vlan1_ins_ptr = OTX2_VLAN_PTR_OFFSET; |
| 764 |
ext->vlan1_ins_tci = skb_vlan_tag_get(skb); |
| 765 |
} else if (skb->vlan_proto == htons(ETH_P_8021AD)) { |
| 766 |
ext->vlan0_ins_ena = 1; |
| 767 |
ext->vlan0_ins_ptr = OTX2_VLAN_PTR_OFFSET; |
| 768 |
ext->vlan0_ins_tci = skb_vlan_tag_get(skb); |
| 769 |
} |
| 770 |
} |
| 771 |
|
| 772 |
*offset += sizeof(*ext); |
| 773 |
} |
| 774 |
|
| 775 |
static void otx2_sqe_add_mem(struct otx2_snd_queue *sq, int *offset, |
| 776 |
int alg, u64 iova, int ptp_offset, |
| 777 |
u64 base_ns, bool udp_csum_crt) |
| 778 |
{ |
| 779 |
struct nix_sqe_mem_s *mem; |
| 780 |
|
| 781 |
mem = (struct nix_sqe_mem_s *)(sq->sqe_base + *offset); |
| 782 |
mem->subdc = NIX_SUBDC_MEM; |
| 783 |
mem->alg = alg; |
| 784 |
mem->wmem = 1; /* wait for the memory operation */ |
| 785 |
mem->addr = iova; |
| 786 |
|
| 787 |
if (ptp_offset) { |
| 788 |
mem->start_offset = ptp_offset; |
| 789 |
mem->udp_csum_crt = !!udp_csum_crt; |
| 790 |
mem->base_ns = base_ns; |
| 791 |
mem->step_type = 1; |
| 792 |
} |
| 793 |
|
| 794 |
*offset += sizeof(*mem); |
| 795 |
} |
| 796 |
|
| 797 |
/* Add SQE header subdescriptor structure */ |
| 798 |
static void otx2_sqe_add_hdr(struct otx2_nic *pfvf, struct otx2_snd_queue *sq, |
| 799 |
struct nix_sqe_hdr_s *sqe_hdr, |
| 800 |
struct sk_buff *skb, u16 qidx) |
| 801 |
{ |
| 802 |
int proto = 0; |
| 803 |
|
| 804 |
/* Check if SQE was framed before, if yes then no need to |
| 805 |
* set these constants again and again. |
| 806 |
*/ |
| 807 |
if (!sqe_hdr->total) { |
| 808 |
/* Don't free Tx buffers to Aura */ |
| 809 |
sqe_hdr->df = 1; |
| 810 |
sqe_hdr->aura = sq->aura_id; |
| 811 |
/* Post a CQE Tx after pkt transmission */ |
| 812 |
sqe_hdr->pnc = 1; |
| 813 |
sqe_hdr->sq = (qidx >= pfvf->hw.tx_queues) ? |
| 814 |
qidx + pfvf->hw.xdp_queues : qidx; |
| 815 |
} |
| 816 |
sqe_hdr->total = skb->len; |
| 817 |
/* Set SQE identifier which will be used later for freeing SKB */ |
| 818 |
sqe_hdr->sqe_id = sq->head; |
| 819 |
|
| 820 |
/* Offload TCP/UDP checksum to HW */ |
| 821 |
if (skb->ip_summed == CHECKSUM_PARTIAL) { |
| 822 |
sqe_hdr->ol3ptr = skb_network_offset(skb); |
| 823 |
sqe_hdr->ol4ptr = skb_transport_offset(skb); |
| 824 |
/* get vlan protocol Ethertype */ |
| 825 |
if (eth_type_vlan(skb->protocol)) |
| 826 |
skb->protocol = vlan_get_protocol(skb); |
| 827 |
|
| 828 |
if (skb->protocol == htons(ETH_P_IP)) { |
| 829 |
proto = ip_hdr(skb)->protocol; |
| 830 |
/* In case of TSO, HW needs this to be explicitly set. |
| 831 |
* So set this always, instead of adding a check. |
| 832 |
*/ |
| 833 |
sqe_hdr->ol3type = NIX_SENDL3TYPE_IP4_CKSUM; |
| 834 |
} else if (skb->protocol == htons(ETH_P_IPV6)) { |
| 835 |
proto = ipv6_hdr(skb)->nexthdr; |
| 836 |
sqe_hdr->ol3type = NIX_SENDL3TYPE_IP6; |
| 837 |
} |
| 838 |
|
| 839 |
if (proto == IPPROTO_TCP) |
| 840 |
sqe_hdr->ol4type = NIX_SENDL4TYPE_TCP_CKSUM; |
| 841 |
else if (proto == IPPROTO_UDP) |
| 842 |
sqe_hdr->ol4type = NIX_SENDL4TYPE_UDP_CKSUM; |
| 843 |
} |
| 844 |
} |
| 845 |
|
| 846 |
static int otx2_dma_map_tso_skb(struct otx2_nic *pfvf, |
| 847 |
struct otx2_snd_queue *sq, |
| 848 |
struct sk_buff *skb, int sqe, int hdr_len) |
| 849 |
{ |
| 850 |
int num_segs = skb_shinfo(skb)->nr_frags + 1; |
| 851 |
struct sg_list *sg = &sq->sg[sqe]; |
| 852 |
u64 dma_addr; |
| 853 |
int seg, len; |
| 854 |
|
| 855 |
sg->num_segs = 0; |
| 856 |
|
| 857 |
/* Get payload length at skb->data */ |
| 858 |
len = skb_headlen(skb) - hdr_len; |
| 859 |
|
| 860 |
for (seg = 0; seg < num_segs; seg++) { |
| 861 |
/* Skip skb->data, if there is no payload */ |
| 862 |
if (!seg && !len) |
| 863 |
continue; |
| 864 |
dma_addr = otx2_dma_map_skb_frag(pfvf, skb, seg, &len); |
| 865 |
if (dma_mapping_error(pfvf->dev, dma_addr)) |
| 866 |
goto unmap; |
| 867 |
|
| 868 |
/* Save DMA mapping info for later unmapping */ |
| 869 |
sg->dma_addr[sg->num_segs] = dma_addr; |
| 870 |
sg->size[sg->num_segs] = len; |
| 871 |
sg->num_segs++; |
| 872 |
} |
| 873 |
return 0; |
| 874 |
unmap: |
| 875 |
otx2_dma_unmap_skb_frags(pfvf, sg); |
| 876 |
return -EINVAL; |
| 877 |
} |
| 878 |
|
| 879 |
static u64 otx2_tso_frag_dma_addr(struct otx2_snd_queue *sq, |
| 880 |
struct sk_buff *skb, int seg, |
| 881 |
u64 seg_addr, int hdr_len, int sqe) |
| 882 |
{ |
| 883 |
struct sg_list *sg = &sq->sg[sqe]; |
| 884 |
const skb_frag_t *frag; |
| 885 |
int offset; |
| 886 |
|
| 887 |
if (seg < 0) |
| 888 |
return sg->dma_addr[0] + (seg_addr - (u64)skb->data); |
| 889 |
|
| 890 |
frag = &skb_shinfo(skb)->frags[seg]; |
| 891 |
offset = seg_addr - (u64)skb_frag_address(frag); |
| 892 |
if (skb_headlen(skb) - hdr_len) |
| 893 |
seg++; |
| 894 |
return sg->dma_addr[seg] + offset; |
| 895 |
} |
| 896 |
|
| 897 |
static void otx2_sqe_tso_add_sg(struct otx2_snd_queue *sq, |
| 898 |
struct sg_list *list, int *offset) |
| 899 |
{ |
| 900 |
struct nix_sqe_sg_s *sg = NULL; |
| 901 |
u16 *sg_lens = NULL; |
| 902 |
u64 *iova = NULL; |
| 903 |
int seg; |
| 904 |
|
| 905 |
/* Add SG descriptors with buffer addresses */ |
| 906 |
for (seg = 0; seg < list->num_segs; seg++) { |
| 907 |
if ((seg % MAX_SEGS_PER_SG) == 0) { |
| 908 |
sg = (struct nix_sqe_sg_s *)(sq->sqe_base + *offset); |
| 909 |
sg->ld_type = NIX_SEND_LDTYPE_LDD; |
| 910 |
sg->subdc = NIX_SUBDC_SG; |
| 911 |
sg->segs = 0; |
| 912 |
sg_lens = (void *)sg; |
| 913 |
iova = (void *)sg + sizeof(*sg); |
| 914 |
/* Next subdc always starts at a 16byte boundary. |
| 915 |
* So if sg->segs is whether 2 or 3, offset += 16bytes. |
| 916 |
*/ |
| 917 |
if ((list->num_segs - seg) >= (MAX_SEGS_PER_SG - 1)) |
| 918 |
*offset += sizeof(*sg) + (3 * sizeof(u64)); |
| 919 |
else |
| 920 |
*offset += sizeof(*sg) + sizeof(u64); |
| 921 |
} |
| 922 |
sg_lens[frag_num(seg % MAX_SEGS_PER_SG)] = list->size[seg]; |
| 923 |
*iova++ = list->dma_addr[seg]; |
| 924 |
sg->segs++; |
| 925 |
} |
| 926 |
} |
| 927 |
|
| 928 |
static void otx2_sq_append_tso(struct otx2_nic *pfvf, struct otx2_snd_queue *sq, |
| 929 |
struct sk_buff *skb, u16 qidx) |
| 930 |
{ |
| 931 |
struct netdev_queue *txq = netdev_get_tx_queue(pfvf->netdev, qidx); |
| 932 |
int hdr_len, tcp_data, seg_len, pkt_len, offset; |
| 933 |
struct nix_sqe_hdr_s *sqe_hdr; |
| 934 |
int first_sqe = sq->head; |
| 935 |
struct sg_list list; |
| 936 |
struct tso_t tso; |
| 937 |
|
| 938 |
hdr_len = tso_start(skb, &tso); |
| 939 |
|
| 940 |
/* Map SKB's fragments to DMA. |
| 941 |
* It's done here to avoid mapping for every TSO segment's packet. |
| 942 |
*/ |
| 943 |
if (otx2_dma_map_tso_skb(pfvf, sq, skb, first_sqe, hdr_len)) { |
| 944 |
dev_kfree_skb_any(skb); |
| 945 |
return; |
| 946 |
} |
| 947 |
|
| 948 |
netdev_tx_sent_queue(txq, skb->len); |
| 949 |
|
| 950 |
tcp_data = skb->len - hdr_len; |
| 951 |
while (tcp_data > 0) { |
| 952 |
char *hdr; |
| 953 |
|
| 954 |
seg_len = min_t(int, skb_shinfo(skb)->gso_size, tcp_data); |
| 955 |
tcp_data -= seg_len; |
| 956 |
|
| 957 |
/* Set SQE's SEND_HDR */ |
| 958 |
memset(sq->sqe_base, 0, sq->sqe_size); |
| 959 |
sqe_hdr = (struct nix_sqe_hdr_s *)(sq->sqe_base); |
| 960 |
otx2_sqe_add_hdr(pfvf, sq, sqe_hdr, skb, qidx); |
| 961 |
offset = sizeof(*sqe_hdr); |
| 962 |
|
| 963 |
/* Add TSO segment's pkt header */ |
| 964 |
hdr = sq->tso_hdrs->base + (sq->head * TSO_HEADER_SIZE); |
| 965 |
tso_build_hdr(skb, hdr, &tso, seg_len, tcp_data == 0); |
| 966 |
list.dma_addr[0] = |
| 967 |
sq->tso_hdrs->iova + (sq->head * TSO_HEADER_SIZE); |
| 968 |
list.size[0] = hdr_len; |
| 969 |
list.num_segs = 1; |
| 970 |
|
| 971 |
/* Add TSO segment's payload data fragments */ |
| 972 |
pkt_len = hdr_len; |
| 973 |
while (seg_len > 0) { |
| 974 |
int size; |
| 975 |
|
| 976 |
size = min_t(int, tso.size, seg_len); |
| 977 |
|
| 978 |
list.size[list.num_segs] = size; |
| 979 |
list.dma_addr[list.num_segs] = |
| 980 |
otx2_tso_frag_dma_addr(sq, skb, |
| 981 |
tso.next_frag_idx - 1, |
| 982 |
(u64)tso.data, hdr_len, |
| 983 |
first_sqe); |
| 984 |
list.num_segs++; |
| 985 |
pkt_len += size; |
| 986 |
seg_len -= size; |
| 987 |
tso_build_data(skb, &tso, size); |
| 988 |
} |
| 989 |
sqe_hdr->total = pkt_len; |
| 990 |
otx2_sqe_tso_add_sg(sq, &list, &offset); |
| 991 |
|
| 992 |
/* DMA mappings and skb needs to be freed only after last |
| 993 |
* TSO segment is transmitted out. So set 'PNC' only for |
| 994 |
* last segment. Also point last segment's sqe_id to first |
| 995 |
* segment's SQE index where skb address and DMA mappings |
| 996 |
* are saved. |
| 997 |
*/ |
| 998 |
if (!tcp_data) { |
| 999 |
sqe_hdr->pnc = 1; |
| 1000 |
sqe_hdr->sqe_id = first_sqe; |
| 1001 |
sq->sg[first_sqe].skb = (u64)skb; |
| 1002 |
} else { |
| 1003 |
sqe_hdr->pnc = 0; |
| 1004 |
} |
| 1005 |
|
| 1006 |
sqe_hdr->sizem1 = (offset / 16) - 1; |
| 1007 |
|
| 1008 |
/* Flush SQE to HW */ |
| 1009 |
pfvf->hw_ops->sqe_flush(pfvf, sq, offset, qidx); |
| 1010 |
} |
| 1011 |
} |
| 1012 |
|
| 1013 |
static bool is_hw_tso_supported(struct otx2_nic *pfvf, |
| 1014 |
struct sk_buff *skb) |
| 1015 |
{ |
| 1016 |
int payload_len, last_seg_size; |
| 1017 |
|
| 1018 |
if (test_bit(HW_TSO, &pfvf->hw.cap_flag)) |
| 1019 |
return true; |
| 1020 |
|
| 1021 |
/* On 96xx A0, HW TSO not supported */ |
| 1022 |
if (!is_96xx_B0(pfvf->pdev)) |
| 1023 |
return false; |
| 1024 |
|
| 1025 |
/* HW has an issue due to which when the payload of the last LSO |
| 1026 |
* segment is shorter than 16 bytes, some header fields may not |
| 1027 |
* be correctly modified, hence don't offload such TSO segments. |
| 1028 |
*/ |
| 1029 |
|
| 1030 |
payload_len = skb->len - skb_tcp_all_headers(skb); |
| 1031 |
last_seg_size = payload_len % skb_shinfo(skb)->gso_size; |
| 1032 |
if (last_seg_size && last_seg_size < 16) |
| 1033 |
return false; |
| 1034 |
|
| 1035 |
return true; |
| 1036 |
} |
| 1037 |
|
| 1038 |
static int otx2_get_sqe_count(struct otx2_nic *pfvf, struct sk_buff *skb) |
| 1039 |
{ |
| 1040 |
if (!skb_shinfo(skb)->gso_size) |
| 1041 |
return 1; |
| 1042 |
|
| 1043 |
/* HW TSO */ |
| 1044 |
if (is_hw_tso_supported(pfvf, skb)) |
| 1045 |
return 1; |
| 1046 |
|
| 1047 |
/* SW TSO */ |
| 1048 |
return skb_shinfo(skb)->gso_segs; |
| 1049 |
} |
| 1050 |
|
| 1051 |
static bool otx2_validate_network_transport(struct sk_buff *skb) |
| 1052 |
{ |
| 1053 |
if ((ip_hdr(skb)->protocol == IPPROTO_UDP) || |
| 1054 |
(ipv6_hdr(skb)->nexthdr == IPPROTO_UDP)) { |
| 1055 |
struct udphdr *udph = udp_hdr(skb); |
| 1056 |
|
| 1057 |
if (udph->source == htons(PTP_PORT) && |
| 1058 |
udph->dest == htons(PTP_PORT)) |
| 1059 |
return true; |
| 1060 |
} |
| 1061 |
|
| 1062 |
return false; |
| 1063 |
} |
| 1064 |
|
| 1065 |
static bool otx2_ptp_is_sync(struct sk_buff *skb, int *offset, bool *udp_csum_crt) |
| 1066 |
{ |
| 1067 |
struct ethhdr *eth = (struct ethhdr *)(skb->data); |
| 1068 |
u16 nix_offload_hlen = 0, inner_vhlen = 0; |
| 1069 |
bool udp_hdr_present = false, is_sync; |
| 1070 |
u8 *data = skb->data, *msgtype; |
| 1071 |
__be16 proto = eth->h_proto; |
| 1072 |
int network_depth = 0; |
| 1073 |
|
| 1074 |
/* NIX is programmed to offload outer VLAN header |
| 1075 |
* in case of single vlan protocol field holds Network header ETH_IP/V6 |
| 1076 |
* in case of stacked vlan protocol field holds Inner vlan (8100) |
| 1077 |
*/ |
| 1078 |
if (skb->dev->features & NETIF_F_HW_VLAN_CTAG_TX && |
| 1079 |
skb->dev->features & NETIF_F_HW_VLAN_STAG_TX) { |
| 1080 |
if (skb->vlan_proto == htons(ETH_P_8021AD)) { |
| 1081 |
/* Get vlan protocol */ |
| 1082 |
proto = __vlan_get_protocol(skb, eth->h_proto, NULL); |
| 1083 |
/* SKB APIs like skb_transport_offset does not include |
| 1084 |
* offloaded vlan header length. Need to explicitly add |
| 1085 |
* the length |
| 1086 |
*/ |
| 1087 |
nix_offload_hlen = VLAN_HLEN; |
| 1088 |
inner_vhlen = VLAN_HLEN; |
| 1089 |
} else if (skb->vlan_proto == htons(ETH_P_8021Q)) { |
| 1090 |
nix_offload_hlen = VLAN_HLEN; |
| 1091 |
} |
| 1092 |
} else if (eth_type_vlan(eth->h_proto)) { |
| 1093 |
proto = __vlan_get_protocol(skb, eth->h_proto, &network_depth); |
| 1094 |
} |
| 1095 |
|
| 1096 |
switch (ntohs(proto)) { |
| 1097 |
case ETH_P_1588: |
| 1098 |
if (network_depth) |
| 1099 |
*offset = network_depth; |
| 1100 |
else |
| 1101 |
*offset = ETH_HLEN + nix_offload_hlen + |
| 1102 |
inner_vhlen; |
| 1103 |
break; |
| 1104 |
case ETH_P_IP: |
| 1105 |
case ETH_P_IPV6: |
| 1106 |
if (!otx2_validate_network_transport(skb)) |
| 1107 |
return false; |
| 1108 |
|
| 1109 |
*offset = nix_offload_hlen + skb_transport_offset(skb) + |
| 1110 |
sizeof(struct udphdr); |
| 1111 |
udp_hdr_present = true; |
| 1112 |
|
| 1113 |
} |
| 1114 |
|
| 1115 |
msgtype = data + *offset; |
| 1116 |
/* Check PTP messageId is SYNC or not */ |
| 1117 |
is_sync = !(*msgtype & 0xf); |
| 1118 |
if (is_sync) |
| 1119 |
*udp_csum_crt = udp_hdr_present; |
| 1120 |
else |
| 1121 |
*offset = 0; |
| 1122 |
|
| 1123 |
return is_sync; |
| 1124 |
} |
| 1125 |
|
| 1126 |
static void otx2_set_txtstamp(struct otx2_nic *pfvf, struct sk_buff *skb, |
| 1127 |
struct otx2_snd_queue *sq, int *offset) |
| 1128 |
{ |
| 1129 |
struct ethhdr *eth = (struct ethhdr *)(skb->data); |
| 1130 |
struct ptpv2_tstamp *origin_tstamp; |
| 1131 |
bool udp_csum_crt = false; |
| 1132 |
unsigned int udphoff; |
| 1133 |
struct timespec64 ts; |
| 1134 |
int ptp_offset = 0; |
| 1135 |
__wsum skb_csum; |
| 1136 |
u64 iova; |
| 1137 |
|
| 1138 |
if (unlikely(!skb_shinfo(skb)->gso_size && |
| 1139 |
(skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP))) { |
| 1140 |
if (unlikely(pfvf->flags & OTX2_FLAG_PTP_ONESTEP_SYNC && |
| 1141 |
otx2_ptp_is_sync(skb, &ptp_offset, &udp_csum_crt))) { |
| 1142 |
origin_tstamp = (struct ptpv2_tstamp *) |
| 1143 |
((u8 *)skb->data + ptp_offset + |
| 1144 |
PTP_SYNC_SEC_OFFSET); |
| 1145 |
ts = ns_to_timespec64(pfvf->ptp->tstamp); |
| 1146 |
origin_tstamp->seconds_msb = htons((ts.tv_sec >> 32) & 0xffff); |
| 1147 |
origin_tstamp->seconds_lsb = htonl(ts.tv_sec & 0xffffffff); |
| 1148 |
origin_tstamp->nanoseconds = htonl(ts.tv_nsec); |
| 1149 |
/* Point to correction field in PTP packet */ |
| 1150 |
ptp_offset += 8; |
| 1151 |
|
| 1152 |
/* When user disables hw checksum, stack calculates the csum, |
| 1153 |
* but it does not cover ptp timestamp which is added later. |
| 1154 |
* Recalculate the checksum manually considering the timestamp. |
| 1155 |
*/ |
| 1156 |
if (udp_csum_crt) { |
| 1157 |
struct udphdr *uh = udp_hdr(skb); |
| 1158 |
|
| 1159 |
if (skb->ip_summed != CHECKSUM_PARTIAL && uh->check != 0) { |
| 1160 |
udphoff = skb_transport_offset(skb); |
| 1161 |
uh->check = 0; |
| 1162 |
skb_csum = skb_checksum(skb, udphoff, skb->len - udphoff, |
| 1163 |
0); |
| 1164 |
if (ntohs(eth->h_proto) == ETH_P_IPV6) |
| 1165 |
uh->check = csum_ipv6_magic(&ipv6_hdr(skb)->saddr, |
| 1166 |
&ipv6_hdr(skb)->daddr, |
| 1167 |
skb->len - udphoff, |
| 1168 |
ipv6_hdr(skb)->nexthdr, |
| 1169 |
skb_csum); |
| 1170 |
else |
| 1171 |
uh->check = csum_tcpudp_magic(ip_hdr(skb)->saddr, |
| 1172 |
ip_hdr(skb)->daddr, |
| 1173 |
skb->len - udphoff, |
| 1174 |
IPPROTO_UDP, |
| 1175 |
skb_csum); |
| 1176 |
} |
| 1177 |
} |
| 1178 |
} else { |
| 1179 |
skb_shinfo(skb)->tx_flags |= SKBTX_IN_PROGRESS; |
| 1180 |
} |
| 1181 |
iova = sq->timestamps->iova + (sq->head * sizeof(u64)); |
| 1182 |
otx2_sqe_add_mem(sq, offset, NIX_SENDMEMALG_E_SETTSTMP, iova, |
| 1183 |
ptp_offset, pfvf->ptp->base_ns, udp_csum_crt); |
| 1184 |
} else { |
| 1185 |
skb_tx_timestamp(skb); |
| 1186 |
} |
| 1187 |
} |
| 1188 |
|
| 1189 |
bool otx2_sq_append_skb(void *dev, struct netdev_queue *txq, |
| 1190 |
struct otx2_snd_queue *sq, |
| 1191 |
struct sk_buff *skb, u16 qidx) |
| 1192 |
{ |
| 1193 |
int offset, num_segs, free_desc; |
| 1194 |
struct nix_sqe_hdr_s *sqe_hdr; |
| 1195 |
struct otx2_nic *pfvf = dev; |
| 1196 |
bool ret; |
| 1197 |
|
| 1198 |
/* Check if there is enough room between producer |
| 1199 |
* and consumer index. |
| 1200 |
*/ |
| 1201 |
free_desc = otx2_get_free_sqe(sq); |
| 1202 |
if (free_desc < sq->sqe_thresh) |
| 1203 |
return false; |
| 1204 |
|
| 1205 |
if (free_desc < otx2_get_sqe_count(pfvf, skb)) |
| 1206 |
return false; |
| 1207 |
|
| 1208 |
num_segs = skb_shinfo(skb)->nr_frags + 1; |
| 1209 |
|
| 1210 |
/* If SKB doesn't fit in a single SQE, linearize it. |
| 1211 |
* TODO: Consider adding JUMP descriptor instead. |
| 1212 |
*/ |
| 1213 |
|
| 1214 |
if (unlikely(num_segs > OTX2_MAX_FRAGS_IN_SQE)) { |
| 1215 |
if (__skb_linearize(skb)) { |
| 1216 |
dev_kfree_skb_any(skb); |
| 1217 |
return true; |
| 1218 |
} |
| 1219 |
num_segs = skb_shinfo(skb)->nr_frags + 1; |
| 1220 |
} |
| 1221 |
|
| 1222 |
if (skb_shinfo(skb)->gso_size && !is_hw_tso_supported(pfvf, skb)) { |
| 1223 |
/* Insert vlan tag before giving pkt to tso */ |
| 1224 |
if (skb_vlan_tag_present(skb)) { |
| 1225 |
skb = __vlan_hwaccel_push_inside(skb); |
| 1226 |
if (!skb) |
| 1227 |
return true; |
| 1228 |
} |
| 1229 |
otx2_sq_append_tso(pfvf, sq, skb, qidx); |
| 1230 |
return true; |
| 1231 |
} |
| 1232 |
|
| 1233 |
/* Set sqe base address */ |
| 1234 |
otx2_sq_set_sqe_base(sq, skb); |
| 1235 |
|
| 1236 |
/* Set SQE's SEND_HDR. |
| 1237 |
* Do not clear the first 64bit as it contains constant info. |
| 1238 |
*/ |
| 1239 |
memset(sq->sqe_base + 8, 0, sq->sqe_size - 8); |
| 1240 |
sqe_hdr = (struct nix_sqe_hdr_s *)(sq->sqe_base); |
| 1241 |
otx2_sqe_add_hdr(pfvf, sq, sqe_hdr, skb, qidx); |
| 1242 |
offset = sizeof(*sqe_hdr); |
| 1243 |
|
| 1244 |
/* Add extended header if needed */ |
| 1245 |
otx2_sqe_add_ext(pfvf, sq, skb, &offset); |
| 1246 |
|
| 1247 |
/* Add SG subdesc with data frags */ |
| 1248 |
if (static_branch_unlikely(&cn10k_ipsec_sa_enabled) && |
| 1249 |
(xfrm_offload(skb))) |
| 1250 |
ret = otx2_sqe_add_sg_ipsec(pfvf, sq, skb, num_segs, &offset); |
| 1251 |
else |
| 1252 |
ret = otx2_sqe_add_sg(pfvf, sq, skb, num_segs, &offset); |
| 1253 |
|
| 1254 |
if (!ret) { |
| 1255 |
otx2_dma_unmap_skb_frags(pfvf, &sq->sg[sq->head]); |
| 1256 |
return false; |
| 1257 |
} |
| 1258 |
|
| 1259 |
otx2_set_txtstamp(pfvf, skb, sq, &offset); |
| 1260 |
|
| 1261 |
sqe_hdr->sizem1 = (offset / 16) - 1; |
| 1262 |
|
| 1263 |
if (static_branch_unlikely(&cn10k_ipsec_sa_enabled) && |
| 1264 |
(xfrm_offload(skb))) |
| 1265 |
return cn10k_ipsec_transmit(pfvf, txq, sq, skb, num_segs, |
| 1266 |
offset); |
| 1267 |
|
| 1268 |
netdev_tx_sent_queue(txq, skb->len); |
| 1269 |
|
| 1270 |
/* Flush SQE to HW */ |
| 1271 |
pfvf->hw_ops->sqe_flush(pfvf, sq, offset, qidx); |
| 1272 |
return true; |
| 1273 |
} |
| 1274 |
EXPORT_SYMBOL(otx2_sq_append_skb); |
| 1275 |
|
| 1276 |
void otx2_cleanup_rx_cqes(struct otx2_nic *pfvf, struct otx2_cq_queue *cq, int qidx) |
| 1277 |
{ |
| 1278 |
struct nix_cqe_rx_s *cqe; |
| 1279 |
struct otx2_pool *pool; |
| 1280 |
int processed_cqe = 0; |
| 1281 |
u16 pool_id; |
| 1282 |
u64 iova; |
| 1283 |
|
| 1284 |
pool_id = otx2_get_pool_idx(pfvf, AURA_NIX_RQ, qidx); |
| 1285 |
pool = &pfvf->qset.pool[pool_id]; |
| 1286 |
|
| 1287 |
if (pfvf->xdp_prog) { |
| 1288 |
if (pool->page_pool) |
| 1289 |
xdp_rxq_info_unreg_mem_model(&cq->xdp_rxq); |
| 1290 |
|
| 1291 |
xdp_rxq_info_unreg(&cq->xdp_rxq); |
| 1292 |
} |
| 1293 |
|
| 1294 |
if (otx2_nix_cq_op_status(pfvf, cq) || !cq->pend_cqe) |
| 1295 |
return; |
| 1296 |
|
| 1297 |
while (cq->pend_cqe) { |
| 1298 |
cqe = (struct nix_cqe_rx_s *)otx2_get_next_cqe(cq); |
| 1299 |
processed_cqe++; |
| 1300 |
cq->pend_cqe--; |
| 1301 |
|
| 1302 |
if (!cqe) |
| 1303 |
continue; |
| 1304 |
if (cqe->sg.segs > 1) { |
| 1305 |
otx2_free_rcv_seg(pfvf, cqe, cq->cq_idx); |
| 1306 |
continue; |
| 1307 |
} |
| 1308 |
iova = cqe->sg.seg_addr - OTX2_HEAD_ROOM; |
| 1309 |
|
| 1310 |
otx2_free_bufs(pfvf, pool, iova, pfvf->rbsize); |
| 1311 |
} |
| 1312 |
|
| 1313 |
/* Free CQEs to HW */ |
| 1314 |
otx2_write64(pfvf, NIX_LF_CQ_OP_DOOR, |
| 1315 |
((u64)cq->cq_idx << 32) | processed_cqe); |
| 1316 |
} |
| 1317 |
|
| 1318 |
void otx2_cleanup_tx_cqes(struct otx2_nic *pfvf, struct otx2_cq_queue *cq) |
| 1319 |
{ |
| 1320 |
int tx_pkts = 0, tx_bytes = 0; |
| 1321 |
struct sk_buff *skb = NULL; |
| 1322 |
struct otx2_snd_queue *sq; |
| 1323 |
struct nix_cqe_tx_s *cqe; |
| 1324 |
struct netdev_queue *txq; |
| 1325 |
int processed_cqe = 0; |
| 1326 |
struct sg_list *sg; |
| 1327 |
int qidx; |
| 1328 |
|
| 1329 |
qidx = cq->cq_idx - pfvf->hw.rx_queues; |
| 1330 |
sq = &pfvf->qset.sq[qidx]; |
| 1331 |
|
| 1332 |
if (otx2_nix_cq_op_status(pfvf, cq) || !cq->pend_cqe) |
| 1333 |
return; |
| 1334 |
|
| 1335 |
while (cq->pend_cqe) { |
| 1336 |
cqe = (struct nix_cqe_tx_s *)otx2_get_next_cqe(cq); |
| 1337 |
processed_cqe++; |
| 1338 |
cq->pend_cqe--; |
| 1339 |
|
| 1340 |
if (!cqe) |
| 1341 |
continue; |
| 1342 |
sg = &sq->sg[cqe->comp.sqe_id]; |
| 1343 |
skb = (struct sk_buff *)sg->skb; |
| 1344 |
if (skb) { |
| 1345 |
tx_bytes += skb->len; |
| 1346 |
tx_pkts++; |
| 1347 |
otx2_dma_unmap_skb_frags(pfvf, sg); |
| 1348 |
dev_kfree_skb_any(skb); |
| 1349 |
sg->skb = (u64)NULL; |
| 1350 |
} |
| 1351 |
} |
| 1352 |
|
| 1353 |
if (likely(tx_pkts)) { |
| 1354 |
if (qidx >= pfvf->hw.tx_queues) |
| 1355 |
qidx -= pfvf->hw.xdp_queues; |
| 1356 |
txq = netdev_get_tx_queue(pfvf->netdev, qidx); |
| 1357 |
netdev_tx_completed_queue(txq, tx_pkts, tx_bytes); |
| 1358 |
} |
| 1359 |
/* Free CQEs to HW */ |
| 1360 |
otx2_write64(pfvf, NIX_LF_CQ_OP_DOOR, |
| 1361 |
((u64)cq->cq_idx << 32) | processed_cqe); |
| 1362 |
} |
| 1363 |
|
| 1364 |
int otx2_rxtx_enable(struct otx2_nic *pfvf, bool enable) |
| 1365 |
{ |
| 1366 |
struct msg_req *msg; |
| 1367 |
int err; |
| 1368 |
|
| 1369 |
mutex_lock(&pfvf->mbox.lock); |
| 1370 |
if (enable) |
| 1371 |
msg = otx2_mbox_alloc_msg_nix_lf_start_rx(&pfvf->mbox); |
| 1372 |
else |
| 1373 |
msg = otx2_mbox_alloc_msg_nix_lf_stop_rx(&pfvf->mbox); |
| 1374 |
|
| 1375 |
if (!msg) { |
| 1376 |
mutex_unlock(&pfvf->mbox.lock); |
| 1377 |
return -ENOMEM; |
| 1378 |
} |
| 1379 |
|
| 1380 |
err = otx2_sync_mbox_msg(&pfvf->mbox); |
| 1381 |
mutex_unlock(&pfvf->mbox.lock); |
| 1382 |
return err; |
| 1383 |
} |
| 1384 |
|
| 1385 |
void otx2_free_pending_sqe(struct otx2_nic *pfvf) |
| 1386 |
{ |
| 1387 |
int tx_pkts = 0, tx_bytes = 0; |
| 1388 |
struct sk_buff *skb = NULL; |
| 1389 |
struct otx2_snd_queue *sq; |
| 1390 |
struct netdev_queue *txq; |
| 1391 |
struct sg_list *sg; |
| 1392 |
int sq_idx, sqe; |
| 1393 |
|
| 1394 |
for (sq_idx = 0; sq_idx < pfvf->hw.tx_queues; sq_idx++) { |
| 1395 |
sq = &pfvf->qset.sq[sq_idx]; |
| 1396 |
for (sqe = 0; sqe < sq->sqe_cnt; sqe++) { |
| 1397 |
sg = &sq->sg[sqe]; |
| 1398 |
skb = (struct sk_buff *)sg->skb; |
| 1399 |
if (skb) { |
| 1400 |
tx_bytes += skb->len; |
| 1401 |
tx_pkts++; |
| 1402 |
otx2_dma_unmap_skb_frags(pfvf, sg); |
| 1403 |
dev_kfree_skb_any(skb); |
| 1404 |
sg->skb = (u64)NULL; |
| 1405 |
} |
| 1406 |
} |
| 1407 |
|
| 1408 |
if (!tx_pkts) |
| 1409 |
continue; |
| 1410 |
txq = netdev_get_tx_queue(pfvf->netdev, sq_idx); |
| 1411 |
netdev_tx_completed_queue(txq, tx_pkts, tx_bytes); |
| 1412 |
tx_pkts = 0; |
| 1413 |
tx_bytes = 0; |
| 1414 |
} |
| 1415 |
} |
| 1416 |
|
| 1417 |
void otx2_xdp_sqe_add_sg(struct otx2_snd_queue *sq, struct xdp_frame *xdpf, |
| 1418 |
u64 dma_addr, int len, int *offset, u16 flags) |
| 1419 |
{ |
| 1420 |
struct nix_sqe_sg_s *sg = NULL; |
| 1421 |
u64 *iova = NULL; |
| 1422 |
|
| 1423 |
sg = (struct nix_sqe_sg_s *)(sq->sqe_base + *offset); |
| 1424 |
sg->ld_type = NIX_SEND_LDTYPE_LDD; |
| 1425 |
sg->subdc = NIX_SUBDC_SG; |
| 1426 |
sg->segs = 1; |
| 1427 |
sg->seg1_size = len; |
| 1428 |
iova = (void *)sg + sizeof(*sg); |
| 1429 |
*iova = dma_addr; |
| 1430 |
*offset += sizeof(*sg) + sizeof(u64); |
| 1431 |
|
| 1432 |
sq->sg[sq->head].dma_addr[0] = dma_addr; |
| 1433 |
sq->sg[sq->head].size[0] = len; |
| 1434 |
sq->sg[sq->head].num_segs = 1; |
| 1435 |
sq->sg[sq->head].flags = flags; |
| 1436 |
sq->sg[sq->head].skb = (u64)xdpf; |
| 1437 |
} |
| 1438 |
|
| 1439 |
int otx2_read_free_sqe(struct otx2_nic *pfvf, u16 qidx) |
| 1440 |
{ |
| 1441 |
struct otx2_snd_queue *sq; |
| 1442 |
int free_sqe; |
| 1443 |
|
| 1444 |
sq = &pfvf->qset.sq[qidx]; |
| 1445 |
free_sqe = otx2_get_free_sqe(sq); |
| 1446 |
if (free_sqe < sq->sqe_thresh) { |
| 1447 |
netdev_warn(pfvf->netdev, "No free sqe for Send queue%d\n", qidx); |
| 1448 |
return 0; |
| 1449 |
} |
| 1450 |
|
| 1451 |
return free_sqe - sq->sqe_thresh; |
| 1452 |
} |
| 1453 |
|
| 1454 |
bool otx2_xdp_sq_append_pkt(struct otx2_nic *pfvf, struct xdp_frame *xdpf, |
| 1455 |
u64 iova, int len, u16 qidx, u16 flags) |
| 1456 |
{ |
| 1457 |
struct nix_sqe_hdr_s *sqe_hdr; |
| 1458 |
struct otx2_snd_queue *sq; |
| 1459 |
int offset, free_sqe; |
| 1460 |
|
| 1461 |
sq = &pfvf->qset.sq[qidx]; |
| 1462 |
free_sqe = otx2_get_free_sqe(sq); |
| 1463 |
if (free_sqe < sq->sqe_thresh) |
| 1464 |
return false; |
| 1465 |
|
| 1466 |
memset(sq->sqe_base + 8, 0, sq->sqe_size - 8); |
| 1467 |
|
| 1468 |
sqe_hdr = (struct nix_sqe_hdr_s *)(sq->sqe_base); |
| 1469 |
|
| 1470 |
if (!sqe_hdr->total) { |
| 1471 |
sqe_hdr->aura = sq->aura_id; |
| 1472 |
sqe_hdr->df = 1; |
| 1473 |
sqe_hdr->sq = qidx; |
| 1474 |
sqe_hdr->pnc = 1; |
| 1475 |
} |
| 1476 |
sqe_hdr->total = len; |
| 1477 |
sqe_hdr->sqe_id = sq->head; |
| 1478 |
|
| 1479 |
offset = sizeof(*sqe_hdr); |
| 1480 |
|
| 1481 |
otx2_xdp_sqe_add_sg(sq, xdpf, iova, len, &offset, flags); |
| 1482 |
sqe_hdr->sizem1 = (offset / 16) - 1; |
| 1483 |
pfvf->hw_ops->sqe_flush(pfvf, sq, offset, qidx); |
| 1484 |
|
| 1485 |
return true; |
| 1486 |
} |
| 1487 |
|
| 1488 |
static bool otx2_xdp_rcv_pkt_handler(struct otx2_nic *pfvf, |
| 1489 |
struct bpf_prog *prog, |
| 1490 |
struct nix_cqe_rx_s *cqe, |
| 1491 |
struct otx2_cq_queue *cq, |
| 1492 |
u32 *metasize, bool *need_xdp_flush) |
| 1493 |
{ |
| 1494 |
struct xdp_buff xdp, *xsk_buff = NULL; |
| 1495 |
unsigned char *hard_start; |
| 1496 |
struct otx2_pool *pool; |
| 1497 |
struct xdp_frame *xdpf; |
| 1498 |
int qidx = cq->cq_idx; |
| 1499 |
struct page *page; |
| 1500 |
u64 iova, pa; |
| 1501 |
u32 act; |
| 1502 |
int err; |
| 1503 |
|
| 1504 |
pool = &pfvf->qset.pool[qidx]; |
| 1505 |
|
| 1506 |
if (pool->xsk_pool) { |
| 1507 |
xsk_buff = pool->xdp[--cq->rbpool->xdp_top]; |
| 1508 |
if (!xsk_buff) |
| 1509 |
return false; |
| 1510 |
|
| 1511 |
xsk_buff->data_end = xsk_buff->data + cqe->sg.seg_size; |
| 1512 |
act = bpf_prog_run_xdp(prog, xsk_buff); |
| 1513 |
goto handle_xdp_verdict; |
| 1514 |
} |
| 1515 |
|
| 1516 |
iova = cqe->sg.seg_addr - OTX2_HEAD_ROOM; |
| 1517 |
pa = otx2_iova_to_phys(pfvf->iommu_domain, iova); |
| 1518 |
page = virt_to_page(phys_to_virt(pa)); |
| 1519 |
|
| 1520 |
xdp_init_buff(&xdp, pfvf->rbsize, &cq->xdp_rxq); |
| 1521 |
|
| 1522 |
hard_start = (unsigned char *)phys_to_virt(pa); |
| 1523 |
xdp_prepare_buff(&xdp, hard_start, OTX2_HEAD_ROOM, |
| 1524 |
cqe->sg.seg_size, true); |
| 1525 |
|
| 1526 |
act = bpf_prog_run_xdp(prog, &xdp); |
| 1527 |
|
| 1528 |
handle_xdp_verdict: |
| 1529 |
switch (act) { |
| 1530 |
case XDP_PASS: |
| 1531 |
*metasize = xdp.data - xdp.data_meta; |
| 1532 |
break; |
| 1533 |
case XDP_TX: |
| 1534 |
qidx += pfvf->hw.tx_queues; |
| 1535 |
cq->pool_ptrs++; |
| 1536 |
xdpf = xdp_convert_buff_to_frame(&xdp); |
| 1537 |
return otx2_xdp_sq_append_pkt(pfvf, xdpf, |
| 1538 |
cqe->sg.seg_addr, |
| 1539 |
cqe->sg.seg_size, |
| 1540 |
qidx, OTX2_XDP_TX); |
| 1541 |
case XDP_REDIRECT: |
| 1542 |
cq->pool_ptrs++; |
| 1543 |
if (xsk_buff) { |
| 1544 |
err = xdp_do_redirect(pfvf->netdev, xsk_buff, prog); |
| 1545 |
if (!err) { |
| 1546 |
*need_xdp_flush = true; |
| 1547 |
return true; |
| 1548 |
} |
| 1549 |
return false; |
| 1550 |
} |
| 1551 |
|
| 1552 |
err = xdp_do_redirect(pfvf->netdev, &xdp, prog); |
| 1553 |
if (!err) { |
| 1554 |
*need_xdp_flush = true; |
| 1555 |
return true; |
| 1556 |
} |
| 1557 |
|
| 1558 |
otx2_dma_unmap_page(pfvf, iova, pfvf->rbsize, |
| 1559 |
DMA_FROM_DEVICE); |
| 1560 |
xdpf = xdp_convert_buff_to_frame(&xdp); |
| 1561 |
xdp_return_frame(xdpf); |
| 1562 |
break; |
| 1563 |
default: |
| 1564 |
bpf_warn_invalid_xdp_action(pfvf->netdev, prog, act); |
| 1565 |
fallthrough; |
| 1566 |
case XDP_ABORTED: |
| 1567 |
if (act == XDP_ABORTED) |
| 1568 |
trace_xdp_exception(pfvf->netdev, prog, act); |
| 1569 |
fallthrough; |
| 1570 |
case XDP_DROP: |
| 1571 |
cq->pool_ptrs++; |
| 1572 |
if (xsk_buff) { |
| 1573 |
xsk_buff_free(xsk_buff); |
| 1574 |
} else if (pp_page_to_nmdesc(page)->pp) { |
| 1575 |
page_pool_recycle_direct(pool->page_pool, page); |
| 1576 |
} else { |
| 1577 |
otx2_dma_unmap_page(pfvf, iova, pfvf->rbsize, |
| 1578 |
DMA_FROM_DEVICE); |
| 1579 |
put_page(page); |
| 1580 |
} |
| 1581 |
return true; |
| 1582 |
} |
| 1583 |
return false; |
| 1584 |
} |