1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB 2 /* 3 * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved. 4 * Copyright (c) 2015 System Fabric Works, Inc. All rights reserved. 5 */ 6 7 #include <linux/skbuff.h> 8 9 #include "rxe.h" 10 #include "rxe_loc.h" 11 #include "rxe_queue.h" 12 13 static int next_opcode(struct rxe_qp *qp, struct rxe_send_wqe *wqe, 14 u32 opcode); 15 16 static inline void retry_first_write_send(struct rxe_qp *qp, 17 struct rxe_send_wqe *wqe, int npsn) 18 { 19 int i; 20 21 for (i = 0; i < npsn; i++) { 22 int to_send = (wqe->dma.resid > qp->mtu) ? 23 qp->mtu : wqe->dma.resid; 24 25 qp->req.opcode = next_opcode(qp, wqe, 26 wqe->wr.opcode); 27 28 if (wqe->wr.send_flags & IB_SEND_INLINE) { 29 wqe->dma.resid -= to_send; 30 wqe->dma.sge_offset += to_send; 31 } else { 32 advance_dma_data(&wqe->dma, to_send); 33 } 34 } 35 } 36 37 static void req_retry(struct rxe_qp *qp) 38 { 39 struct rxe_send_wqe *wqe; 40 unsigned int wqe_index; 41 unsigned int mask; 42 int npsn; 43 int first = 1; 44 struct rxe_queue *q = qp->sq.queue; 45 unsigned int cons; 46 unsigned int prod; 47 48 cons = queue_get_consumer(q, QUEUE_TYPE_FROM_CLIENT); 49 prod = queue_get_producer(q, QUEUE_TYPE_FROM_CLIENT); 50 51 qp->req.wqe_index = cons; 52 qp->req.psn = qp->comp.psn; 53 qp->req.opcode = -1; 54 55 for (wqe_index = cons; wqe_index != prod; 56 wqe_index = queue_next_index(q, wqe_index)) { 57 wqe = queue_addr_from_index(qp->sq.queue, wqe_index); 58 mask = wr_opcode_mask(wqe->wr.opcode, qp); 59 60 if (wqe->state == wqe_state_posted) 61 break; 62 63 if (wqe->state == wqe_state_done) 64 continue; 65 66 wqe->iova = (mask & WR_ATOMIC_MASK) ? 67 wqe->wr.wr.atomic.remote_addr : 68 (mask & WR_READ_OR_WRITE_MASK) ? 69 wqe->wr.wr.rdma.remote_addr : 70 0; 71 72 if (!first || (mask & WR_READ_MASK) == 0) { 73 wqe->dma.resid = wqe->dma.length; 74 wqe->dma.cur_sge = 0; 75 wqe->dma.sge_offset = 0; 76 } 77 78 if (first) { 79 first = 0; 80 81 if (mask & WR_WRITE_OR_SEND_MASK) { 82 npsn = (qp->comp.psn - wqe->first_psn) & 83 BTH_PSN_MASK; 84 retry_first_write_send(qp, wqe, npsn); 85 } 86 87 if (mask & WR_READ_MASK) { 88 npsn = (wqe->dma.length - wqe->dma.resid) / 89 qp->mtu; 90 wqe->iova += npsn * qp->mtu; 91 } 92 } 93 94 wqe->state = wqe_state_posted; 95 } 96 } 97 98 void rnr_nak_timer(struct timer_list *t) 99 { 100 struct rxe_qp *qp = from_timer(qp, t, rnr_nak_timer); 101 unsigned long flags; 102 103 rxe_dbg_qp(qp, "nak timer fired\n"); 104 105 spin_lock_irqsave(&qp->state_lock, flags); 106 if (qp->valid) { 107 /* request a send queue retry */ 108 qp->req.need_retry = 1; 109 qp->req.wait_for_rnr_timer = 0; 110 rxe_sched_task(&qp->send_task); 111 } 112 spin_unlock_irqrestore(&qp->state_lock, flags); 113 } 114 115 static void req_check_sq_drain_done(struct rxe_qp *qp) 116 { 117 struct rxe_queue *q; 118 unsigned int index; 119 unsigned int cons; 120 struct rxe_send_wqe *wqe; 121 unsigned long flags; 122 123 spin_lock_irqsave(&qp->state_lock, flags); 124 if (qp_state(qp) == IB_QPS_SQD) { 125 q = qp->sq.queue; 126 index = qp->req.wqe_index; 127 cons = queue_get_consumer(q, QUEUE_TYPE_FROM_CLIENT); 128 wqe = queue_addr_from_index(q, cons); 129 130 /* check to see if we are drained; 131 * state_lock used by requester and completer 132 */ 133 do { 134 if (!qp->attr.sq_draining) 135 /* comp just finished */ 136 break; 137 138 if (wqe && ((index != cons) || 139 (wqe->state != wqe_state_posted))) 140 /* comp not done yet */ 141 break; 142 143 qp->attr.sq_draining = 0; 144 spin_unlock_irqrestore(&qp->state_lock, flags); 145 146 if (qp->ibqp.event_handler) { 147 struct ib_event ev; 148 149 ev.device = qp->ibqp.device; 150 ev.element.qp = &qp->ibqp; 151 ev.event = IB_EVENT_SQ_DRAINED; 152 qp->ibqp.event_handler(&ev, 153 qp->ibqp.qp_context); 154 } 155 return; 156 } while (0); 157 } 158 spin_unlock_irqrestore(&qp->state_lock, flags); 159 } 160 161 static struct rxe_send_wqe *__req_next_wqe(struct rxe_qp *qp) 162 { 163 struct rxe_queue *q = qp->sq.queue; 164 unsigned int index = qp->req.wqe_index; 165 unsigned int prod; 166 167 prod = queue_get_producer(q, QUEUE_TYPE_FROM_CLIENT); 168 if (index == prod) 169 return NULL; 170 else 171 return queue_addr_from_index(q, index); 172 } 173 174 static struct rxe_send_wqe *req_next_wqe(struct rxe_qp *qp) 175 { 176 struct rxe_send_wqe *wqe; 177 unsigned long flags; 178 179 req_check_sq_drain_done(qp); 180 181 wqe = __req_next_wqe(qp); 182 if (wqe == NULL) 183 return NULL; 184 185 spin_lock_irqsave(&qp->state_lock, flags); 186 if (unlikely((qp_state(qp) == IB_QPS_SQD) && 187 (wqe->state != wqe_state_processing))) { 188 spin_unlock_irqrestore(&qp->state_lock, flags); 189 return NULL; 190 } 191 spin_unlock_irqrestore(&qp->state_lock, flags); 192 193 wqe->mask = wr_opcode_mask(wqe->wr.opcode, qp); 194 return wqe; 195 } 196 197 /** 198 * rxe_wqe_is_fenced - check if next wqe is fenced 199 * @qp: the queue pair 200 * @wqe: the next wqe 201 * 202 * Returns: 1 if wqe needs to wait 203 * 0 if wqe is ready to go 204 */ 205 static int rxe_wqe_is_fenced(struct rxe_qp *qp, struct rxe_send_wqe *wqe) 206 { 207 /* Local invalidate fence (LIF) see IBA 10.6.5.1 208 * Requires ALL previous operations on the send queue 209 * are complete. Make mandatory for the rxe driver. 210 */ 211 if (wqe->wr.opcode == IB_WR_LOCAL_INV) 212 return qp->req.wqe_index != queue_get_consumer(qp->sq.queue, 213 QUEUE_TYPE_FROM_CLIENT); 214 215 /* Fence see IBA 10.8.3.3 216 * Requires that all previous read and atomic operations 217 * are complete. 218 */ 219 return (wqe->wr.send_flags & IB_SEND_FENCE) && 220 atomic_read(&qp->req.rd_atomic) != qp->attr.max_rd_atomic; 221 } 222 223 static int next_opcode_rc(struct rxe_qp *qp, u32 opcode, int fits) 224 { 225 switch (opcode) { 226 case IB_WR_RDMA_WRITE: 227 if (qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_FIRST || 228 qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_MIDDLE) 229 return fits ? 230 IB_OPCODE_RC_RDMA_WRITE_LAST : 231 IB_OPCODE_RC_RDMA_WRITE_MIDDLE; 232 else 233 return fits ? 234 IB_OPCODE_RC_RDMA_WRITE_ONLY : 235 IB_OPCODE_RC_RDMA_WRITE_FIRST; 236 237 case IB_WR_RDMA_WRITE_WITH_IMM: 238 if (qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_FIRST || 239 qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_MIDDLE) 240 return fits ? 241 IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE : 242 IB_OPCODE_RC_RDMA_WRITE_MIDDLE; 243 else 244 return fits ? 245 IB_OPCODE_RC_RDMA_WRITE_ONLY_WITH_IMMEDIATE : 246 IB_OPCODE_RC_RDMA_WRITE_FIRST; 247 248 case IB_WR_SEND: 249 if (qp->req.opcode == IB_OPCODE_RC_SEND_FIRST || 250 qp->req.opcode == IB_OPCODE_RC_SEND_MIDDLE) 251 return fits ? 252 IB_OPCODE_RC_SEND_LAST : 253 IB_OPCODE_RC_SEND_MIDDLE; 254 else 255 return fits ? 256 IB_OPCODE_RC_SEND_ONLY : 257 IB_OPCODE_RC_SEND_FIRST; 258 259 case IB_WR_SEND_WITH_IMM: 260 if (qp->req.opcode == IB_OPCODE_RC_SEND_FIRST || 261 qp->req.opcode == IB_OPCODE_RC_SEND_MIDDLE) 262 return fits ? 263 IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE : 264 IB_OPCODE_RC_SEND_MIDDLE; 265 else 266 return fits ? 267 IB_OPCODE_RC_SEND_ONLY_WITH_IMMEDIATE : 268 IB_OPCODE_RC_SEND_FIRST; 269 270 case IB_WR_FLUSH: 271 return IB_OPCODE_RC_FLUSH; 272 273 case IB_WR_RDMA_READ: 274 return IB_OPCODE_RC_RDMA_READ_REQUEST; 275 276 case IB_WR_ATOMIC_CMP_AND_SWP: 277 return IB_OPCODE_RC_COMPARE_SWAP; 278 279 case IB_WR_ATOMIC_FETCH_AND_ADD: 280 return IB_OPCODE_RC_FETCH_ADD; 281 282 case IB_WR_SEND_WITH_INV: 283 if (qp->req.opcode == IB_OPCODE_RC_SEND_FIRST || 284 qp->req.opcode == IB_OPCODE_RC_SEND_MIDDLE) 285 return fits ? IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE : 286 IB_OPCODE_RC_SEND_MIDDLE; 287 else 288 return fits ? IB_OPCODE_RC_SEND_ONLY_WITH_INVALIDATE : 289 IB_OPCODE_RC_SEND_FIRST; 290 291 case IB_WR_ATOMIC_WRITE: 292 return IB_OPCODE_RC_ATOMIC_WRITE; 293 294 case IB_WR_REG_MR: 295 case IB_WR_LOCAL_INV: 296 return opcode; 297 } 298 299 return -EINVAL; 300 } 301 302 static int next_opcode_uc(struct rxe_qp *qp, u32 opcode, int fits) 303 { 304 switch (opcode) { 305 case IB_WR_RDMA_WRITE: 306 if (qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_FIRST || 307 qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_MIDDLE) 308 return fits ? 309 IB_OPCODE_UC_RDMA_WRITE_LAST : 310 IB_OPCODE_UC_RDMA_WRITE_MIDDLE; 311 else 312 return fits ? 313 IB_OPCODE_UC_RDMA_WRITE_ONLY : 314 IB_OPCODE_UC_RDMA_WRITE_FIRST; 315 316 case IB_WR_RDMA_WRITE_WITH_IMM: 317 if (qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_FIRST || 318 qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_MIDDLE) 319 return fits ? 320 IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE : 321 IB_OPCODE_UC_RDMA_WRITE_MIDDLE; 322 else 323 return fits ? 324 IB_OPCODE_UC_RDMA_WRITE_ONLY_WITH_IMMEDIATE : 325 IB_OPCODE_UC_RDMA_WRITE_FIRST; 326 327 case IB_WR_SEND: 328 if (qp->req.opcode == IB_OPCODE_UC_SEND_FIRST || 329 qp->req.opcode == IB_OPCODE_UC_SEND_MIDDLE) 330 return fits ? 331 IB_OPCODE_UC_SEND_LAST : 332 IB_OPCODE_UC_SEND_MIDDLE; 333 else 334 return fits ? 335 IB_OPCODE_UC_SEND_ONLY : 336 IB_OPCODE_UC_SEND_FIRST; 337 338 case IB_WR_SEND_WITH_IMM: 339 if (qp->req.opcode == IB_OPCODE_UC_SEND_FIRST || 340 qp->req.opcode == IB_OPCODE_UC_SEND_MIDDLE) 341 return fits ? 342 IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE : 343 IB_OPCODE_UC_SEND_MIDDLE; 344 else 345 return fits ? 346 IB_OPCODE_UC_SEND_ONLY_WITH_IMMEDIATE : 347 IB_OPCODE_UC_SEND_FIRST; 348 } 349 350 return -EINVAL; 351 } 352 353 static int next_opcode(struct rxe_qp *qp, struct rxe_send_wqe *wqe, 354 u32 opcode) 355 { 356 int fits = (wqe->dma.resid <= qp->mtu); 357 358 switch (qp_type(qp)) { 359 case IB_QPT_RC: 360 return next_opcode_rc(qp, opcode, fits); 361 362 case IB_QPT_UC: 363 return next_opcode_uc(qp, opcode, fits); 364 365 case IB_QPT_UD: 366 case IB_QPT_GSI: 367 switch (opcode) { 368 case IB_WR_SEND: 369 return IB_OPCODE_UD_SEND_ONLY; 370 371 case IB_WR_SEND_WITH_IMM: 372 return IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE; 373 } 374 break; 375 376 default: 377 break; 378 } 379 380 return -EINVAL; 381 } 382 383 static inline int check_init_depth(struct rxe_qp *qp, struct rxe_send_wqe *wqe) 384 { 385 int depth; 386 387 if (wqe->has_rd_atomic) 388 return 0; 389 390 qp->req.need_rd_atomic = 1; 391 depth = atomic_dec_return(&qp->req.rd_atomic); 392 393 if (depth >= 0) { 394 qp->req.need_rd_atomic = 0; 395 wqe->has_rd_atomic = 1; 396 return 0; 397 } 398 399 atomic_inc(&qp->req.rd_atomic); 400 return -EAGAIN; 401 } 402 403 static inline int get_mtu(struct rxe_qp *qp) 404 { 405 struct rxe_dev *rxe = to_rdev(qp->ibqp.device); 406 407 if ((qp_type(qp) == IB_QPT_RC) || (qp_type(qp) == IB_QPT_UC)) 408 return qp->mtu; 409 410 return rxe->port.mtu_cap; 411 } 412 413 static struct sk_buff *init_req_packet(struct rxe_qp *qp, 414 struct rxe_av *av, 415 struct rxe_send_wqe *wqe, 416 int opcode, u32 payload, 417 struct rxe_pkt_info *pkt) 418 { 419 struct rxe_dev *rxe = to_rdev(qp->ibqp.device); 420 struct sk_buff *skb; 421 struct rxe_send_wr *ibwr = &wqe->wr; 422 int pad = (-payload) & 0x3; 423 int paylen; 424 int solicited; 425 u32 qp_num; 426 int ack_req = 0; 427 428 /* length from start of bth to end of icrc */ 429 paylen = rxe_opcode[opcode].length + payload + pad + RXE_ICRC_SIZE; 430 pkt->paylen = paylen; 431 432 /* init skb */ 433 skb = rxe_init_packet(rxe, av, paylen, pkt); 434 if (unlikely(!skb)) 435 return NULL; 436 437 /* init bth */ 438 solicited = (ibwr->send_flags & IB_SEND_SOLICITED) && 439 (pkt->mask & RXE_END_MASK) && 440 ((pkt->mask & (RXE_SEND_MASK)) || 441 (pkt->mask & (RXE_WRITE_MASK | RXE_IMMDT_MASK)) == 442 (RXE_WRITE_MASK | RXE_IMMDT_MASK)); 443 444 qp_num = (pkt->mask & RXE_DETH_MASK) ? ibwr->wr.ud.remote_qpn : 445 qp->attr.dest_qp_num; 446 447 if (qp_type(qp) != IB_QPT_UD && qp_type(qp) != IB_QPT_UC) 448 ack_req = ((pkt->mask & RXE_END_MASK) || 449 (qp->req.noack_pkts++ > RXE_MAX_PKT_PER_ACK)); 450 if (ack_req) 451 qp->req.noack_pkts = 0; 452 453 bth_init(pkt, pkt->opcode, solicited, 0, pad, IB_DEFAULT_PKEY_FULL, qp_num, 454 ack_req, pkt->psn); 455 456 /* init optional headers */ 457 if (pkt->mask & RXE_RETH_MASK) { 458 if (pkt->mask & RXE_FETH_MASK) 459 reth_set_rkey(pkt, ibwr->wr.flush.rkey); 460 else 461 reth_set_rkey(pkt, ibwr->wr.rdma.rkey); 462 reth_set_va(pkt, wqe->iova); 463 reth_set_len(pkt, wqe->dma.resid); 464 } 465 466 /* Fill Flush Extension Transport Header */ 467 if (pkt->mask & RXE_FETH_MASK) 468 feth_init(pkt, ibwr->wr.flush.type, ibwr->wr.flush.level); 469 470 if (pkt->mask & RXE_IMMDT_MASK) 471 immdt_set_imm(pkt, ibwr->ex.imm_data); 472 473 if (pkt->mask & RXE_IETH_MASK) 474 ieth_set_rkey(pkt, ibwr->ex.invalidate_rkey); 475 476 if (pkt->mask & RXE_ATMETH_MASK) { 477 atmeth_set_va(pkt, wqe->iova); 478 if (opcode == IB_OPCODE_RC_COMPARE_SWAP) { 479 atmeth_set_swap_add(pkt, ibwr->wr.atomic.swap); 480 atmeth_set_comp(pkt, ibwr->wr.atomic.compare_add); 481 } else { 482 atmeth_set_swap_add(pkt, ibwr->wr.atomic.compare_add); 483 } 484 atmeth_set_rkey(pkt, ibwr->wr.atomic.rkey); 485 } 486 487 if (pkt->mask & RXE_DETH_MASK) { 488 if (qp->ibqp.qp_num == 1) 489 deth_set_qkey(pkt, GSI_QKEY); 490 else 491 deth_set_qkey(pkt, ibwr->wr.ud.remote_qkey); 492 deth_set_sqp(pkt, qp->ibqp.qp_num); 493 } 494 495 return skb; 496 } 497 498 static int finish_packet(struct rxe_qp *qp, struct rxe_av *av, 499 struct rxe_send_wqe *wqe, struct rxe_pkt_info *pkt, 500 struct sk_buff *skb, u32 payload) 501 { 502 int err; 503 504 err = rxe_prepare(av, pkt, skb); 505 if (err) 506 return err; 507 508 if (pkt->mask & RXE_WRITE_OR_SEND_MASK) { 509 if (wqe->wr.send_flags & IB_SEND_INLINE) { 510 u8 *tmp = &wqe->dma.inline_data[wqe->dma.sge_offset]; 511 512 memcpy(payload_addr(pkt), tmp, payload); 513 514 wqe->dma.resid -= payload; 515 wqe->dma.sge_offset += payload; 516 } else { 517 err = copy_data(qp->pd, 0, &wqe->dma, 518 payload_addr(pkt), payload, 519 RXE_FROM_MR_OBJ); 520 if (err) 521 return err; 522 } 523 if (bth_pad(pkt)) { 524 u8 *pad = payload_addr(pkt) + payload; 525 526 memset(pad, 0, bth_pad(pkt)); 527 } 528 } else if (pkt->mask & RXE_FLUSH_MASK) { 529 /* oA19-2: shall have no payload. */ 530 wqe->dma.resid = 0; 531 } 532 533 if (pkt->mask & RXE_ATOMIC_WRITE_MASK) { 534 memcpy(payload_addr(pkt), wqe->dma.atomic_wr, payload); 535 wqe->dma.resid -= payload; 536 } 537 538 return 0; 539 } 540 541 static void update_wqe_state(struct rxe_qp *qp, 542 struct rxe_send_wqe *wqe, 543 struct rxe_pkt_info *pkt) 544 { 545 if (pkt->mask & RXE_END_MASK) { 546 if (qp_type(qp) == IB_QPT_RC) 547 wqe->state = wqe_state_pending; 548 else 549 wqe->state = wqe_state_done; 550 } else { 551 wqe->state = wqe_state_processing; 552 } 553 } 554 555 static void update_wqe_psn(struct rxe_qp *qp, 556 struct rxe_send_wqe *wqe, 557 struct rxe_pkt_info *pkt, 558 u32 payload) 559 { 560 /* number of packets left to send including current one */ 561 int num_pkt = (wqe->dma.resid + payload + qp->mtu - 1) / qp->mtu; 562 563 /* handle zero length packet case */ 564 if (num_pkt == 0) 565 num_pkt = 1; 566 567 if (pkt->mask & RXE_START_MASK) { 568 wqe->first_psn = qp->req.psn; 569 wqe->last_psn = (qp->req.psn + num_pkt - 1) & BTH_PSN_MASK; 570 } 571 572 if (pkt->mask & RXE_READ_MASK) 573 qp->req.psn = (wqe->first_psn + num_pkt) & BTH_PSN_MASK; 574 else 575 qp->req.psn = (qp->req.psn + 1) & BTH_PSN_MASK; 576 } 577 578 static void update_state(struct rxe_qp *qp, struct rxe_pkt_info *pkt) 579 { 580 qp->req.opcode = pkt->opcode; 581 582 if (pkt->mask & RXE_END_MASK) 583 qp->req.wqe_index = queue_next_index(qp->sq.queue, 584 qp->req.wqe_index); 585 586 qp->need_req_skb = 0; 587 588 if (qp->qp_timeout_jiffies && !timer_pending(&qp->retrans_timer)) 589 mod_timer(&qp->retrans_timer, 590 jiffies + qp->qp_timeout_jiffies); 591 } 592 593 static int rxe_do_local_ops(struct rxe_qp *qp, struct rxe_send_wqe *wqe) 594 { 595 u8 opcode = wqe->wr.opcode; 596 u32 rkey; 597 int ret; 598 599 switch (opcode) { 600 case IB_WR_LOCAL_INV: 601 rkey = wqe->wr.ex.invalidate_rkey; 602 if (rkey_is_mw(rkey)) 603 ret = rxe_invalidate_mw(qp, rkey); 604 else 605 ret = rxe_invalidate_mr(qp, rkey); 606 607 if (unlikely(ret)) { 608 wqe->status = IB_WC_LOC_QP_OP_ERR; 609 return ret; 610 } 611 break; 612 case IB_WR_REG_MR: 613 ret = rxe_reg_fast_mr(qp, wqe); 614 if (unlikely(ret)) { 615 wqe->status = IB_WC_LOC_QP_OP_ERR; 616 return ret; 617 } 618 break; 619 case IB_WR_BIND_MW: 620 ret = rxe_bind_mw(qp, wqe); 621 if (unlikely(ret)) { 622 wqe->status = IB_WC_MW_BIND_ERR; 623 return ret; 624 } 625 break; 626 default: 627 rxe_dbg_qp(qp, "Unexpected send wqe opcode %d\n", opcode); 628 wqe->status = IB_WC_LOC_QP_OP_ERR; 629 return -EINVAL; 630 } 631 632 wqe->state = wqe_state_done; 633 wqe->status = IB_WC_SUCCESS; 634 qp->req.wqe_index = queue_next_index(qp->sq.queue, qp->req.wqe_index); 635 636 return 0; 637 } 638 639 int rxe_requester(struct rxe_qp *qp) 640 { 641 struct rxe_dev *rxe = to_rdev(qp->ibqp.device); 642 struct rxe_pkt_info pkt; 643 struct sk_buff *skb; 644 struct rxe_send_wqe *wqe; 645 enum rxe_hdr_mask mask; 646 u32 payload; 647 int mtu; 648 int opcode; 649 int err; 650 int ret; 651 struct rxe_queue *q = qp->sq.queue; 652 struct rxe_ah *ah; 653 struct rxe_av *av; 654 unsigned long flags; 655 656 spin_lock_irqsave(&qp->state_lock, flags); 657 if (unlikely(!qp->valid)) { 658 spin_unlock_irqrestore(&qp->state_lock, flags); 659 goto exit; 660 } 661 662 if (unlikely(qp_state(qp) == IB_QPS_ERR)) { 663 wqe = __req_next_wqe(qp); 664 spin_unlock_irqrestore(&qp->state_lock, flags); 665 if (wqe) { 666 wqe->status = IB_WC_WR_FLUSH_ERR; 667 goto err; 668 } else { 669 goto exit; 670 } 671 } 672 673 if (unlikely(qp_state(qp) == IB_QPS_RESET)) { 674 qp->req.wqe_index = queue_get_consumer(q, 675 QUEUE_TYPE_FROM_CLIENT); 676 qp->req.opcode = -1; 677 qp->req.need_rd_atomic = 0; 678 qp->req.wait_psn = 0; 679 qp->req.need_retry = 0; 680 qp->req.wait_for_rnr_timer = 0; 681 spin_unlock_irqrestore(&qp->state_lock, flags); 682 goto exit; 683 } 684 spin_unlock_irqrestore(&qp->state_lock, flags); 685 686 /* we come here if the retransmit timer has fired 687 * or if the rnr timer has fired. If the retransmit 688 * timer fires while we are processing an RNR NAK wait 689 * until the rnr timer has fired before starting the 690 * retry flow 691 */ 692 if (unlikely(qp->req.need_retry && !qp->req.wait_for_rnr_timer)) { 693 req_retry(qp); 694 qp->req.need_retry = 0; 695 } 696 697 wqe = req_next_wqe(qp); 698 if (unlikely(!wqe)) 699 goto exit; 700 701 if (rxe_wqe_is_fenced(qp, wqe)) { 702 qp->req.wait_fence = 1; 703 goto exit; 704 } 705 706 if (wqe->mask & WR_LOCAL_OP_MASK) { 707 err = rxe_do_local_ops(qp, wqe); 708 if (unlikely(err)) 709 goto err; 710 else 711 goto done; 712 } 713 714 if (unlikely(qp_type(qp) == IB_QPT_RC && 715 psn_compare(qp->req.psn, (qp->comp.psn + 716 RXE_MAX_UNACKED_PSNS)) > 0)) { 717 qp->req.wait_psn = 1; 718 goto exit; 719 } 720 721 /* Limit the number of inflight SKBs per QP */ 722 if (unlikely(atomic_read(&qp->skb_out) > 723 RXE_INFLIGHT_SKBS_PER_QP_HIGH)) { 724 qp->need_req_skb = 1; 725 goto exit; 726 } 727 728 opcode = next_opcode(qp, wqe, wqe->wr.opcode); 729 if (unlikely(opcode < 0)) { 730 wqe->status = IB_WC_LOC_QP_OP_ERR; 731 goto err; 732 } 733 734 mask = rxe_opcode[opcode].mask; 735 if (unlikely(mask & (RXE_READ_OR_ATOMIC_MASK | 736 RXE_ATOMIC_WRITE_MASK))) { 737 if (check_init_depth(qp, wqe)) 738 goto exit; 739 } 740 741 mtu = get_mtu(qp); 742 payload = (mask & (RXE_WRITE_OR_SEND_MASK | RXE_ATOMIC_WRITE_MASK)) ? 743 wqe->dma.resid : 0; 744 if (payload > mtu) { 745 if (qp_type(qp) == IB_QPT_UD) { 746 /* C10-93.1.1: If the total sum of all the buffer lengths specified for a 747 * UD message exceeds the MTU of the port as returned by QueryHCA, the CI 748 * shall not emit any packets for this message. Further, the CI shall not 749 * generate an error due to this condition. 750 */ 751 752 /* fake a successful UD send */ 753 wqe->first_psn = qp->req.psn; 754 wqe->last_psn = qp->req.psn; 755 qp->req.psn = (qp->req.psn + 1) & BTH_PSN_MASK; 756 qp->req.opcode = IB_OPCODE_UD_SEND_ONLY; 757 qp->req.wqe_index = queue_next_index(qp->sq.queue, 758 qp->req.wqe_index); 759 wqe->state = wqe_state_done; 760 wqe->status = IB_WC_SUCCESS; 761 goto done; 762 } 763 payload = mtu; 764 } 765 766 pkt.rxe = rxe; 767 pkt.opcode = opcode; 768 pkt.qp = qp; 769 pkt.psn = qp->req.psn; 770 pkt.mask = rxe_opcode[opcode].mask; 771 pkt.wqe = wqe; 772 773 av = rxe_get_av(&pkt, &ah); 774 if (unlikely(!av)) { 775 rxe_dbg_qp(qp, "Failed no address vector\n"); 776 wqe->status = IB_WC_LOC_QP_OP_ERR; 777 goto err; 778 } 779 780 skb = init_req_packet(qp, av, wqe, opcode, payload, &pkt); 781 if (unlikely(!skb)) { 782 rxe_dbg_qp(qp, "Failed allocating skb\n"); 783 wqe->status = IB_WC_LOC_QP_OP_ERR; 784 if (ah) 785 rxe_put(ah); 786 goto err; 787 } 788 789 err = finish_packet(qp, av, wqe, &pkt, skb, payload); 790 if (unlikely(err)) { 791 rxe_dbg_qp(qp, "Error during finish packet\n"); 792 if (err == -EFAULT) 793 wqe->status = IB_WC_LOC_PROT_ERR; 794 else 795 wqe->status = IB_WC_LOC_QP_OP_ERR; 796 kfree_skb(skb); 797 if (ah) 798 rxe_put(ah); 799 goto err; 800 } 801 802 if (ah) 803 rxe_put(ah); 804 805 err = rxe_xmit_packet(qp, &pkt, skb); 806 if (err) { 807 wqe->status = IB_WC_LOC_QP_OP_ERR; 808 goto err; 809 } 810 811 update_wqe_state(qp, wqe, &pkt); 812 update_wqe_psn(qp, wqe, &pkt, payload); 813 update_state(qp, &pkt); 814 815 /* A non-zero return value will cause rxe_do_task to 816 * exit its loop and end the work item. A zero return 817 * will continue looping and return to rxe_requester 818 */ 819 done: 820 ret = 0; 821 goto out; 822 err: 823 /* update wqe_index for each wqe completion */ 824 qp->req.wqe_index = queue_next_index(qp->sq.queue, qp->req.wqe_index); 825 wqe->state = wqe_state_error; 826 rxe_qp_error(qp); 827 exit: 828 ret = -EAGAIN; 829 out: 830 return ret; 831 } 832 833 int rxe_sender(struct rxe_qp *qp) 834 { 835 int req_ret; 836 int comp_ret; 837 838 /* process the send queue */ 839 req_ret = rxe_requester(qp); 840 841 /* process the response queue */ 842 comp_ret = rxe_completer(qp); 843 844 /* exit the task loop if both requester and completer 845 * are ready 846 */ 847 return (req_ret && comp_ret) ? -EAGAIN : 0; 848 } 849