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 #include "rxe_task.h" 13 14 enum comp_state { 15 COMPST_GET_ACK, 16 COMPST_GET_WQE, 17 COMPST_COMP_WQE, 18 COMPST_COMP_ACK, 19 COMPST_CHECK_PSN, 20 COMPST_CHECK_ACK, 21 COMPST_READ, 22 COMPST_ATOMIC, 23 COMPST_WRITE_SEND, 24 COMPST_UPDATE_COMP, 25 COMPST_ERROR_RETRY, 26 COMPST_RNR_RETRY, 27 COMPST_ERROR, 28 COMPST_EXIT, /* We have an issue, and we want to rerun the completer */ 29 COMPST_DONE, /* The completer finished successflly */ 30 }; 31 32 static char *comp_state_name[] = { 33 [COMPST_GET_ACK] = "GET ACK", 34 [COMPST_GET_WQE] = "GET WQE", 35 [COMPST_COMP_WQE] = "COMP WQE", 36 [COMPST_COMP_ACK] = "COMP ACK", 37 [COMPST_CHECK_PSN] = "CHECK PSN", 38 [COMPST_CHECK_ACK] = "CHECK ACK", 39 [COMPST_READ] = "READ", 40 [COMPST_ATOMIC] = "ATOMIC", 41 [COMPST_WRITE_SEND] = "WRITE/SEND", 42 [COMPST_UPDATE_COMP] = "UPDATE COMP", 43 [COMPST_ERROR_RETRY] = "ERROR RETRY", 44 [COMPST_RNR_RETRY] = "RNR RETRY", 45 [COMPST_ERROR] = "ERROR", 46 [COMPST_EXIT] = "EXIT", 47 [COMPST_DONE] = "DONE", 48 }; 49 50 static unsigned long rnrnak_usec[32] = { 51 [IB_RNR_TIMER_655_36] = 655360, 52 [IB_RNR_TIMER_000_01] = 10, 53 [IB_RNR_TIMER_000_02] = 20, 54 [IB_RNR_TIMER_000_03] = 30, 55 [IB_RNR_TIMER_000_04] = 40, 56 [IB_RNR_TIMER_000_06] = 60, 57 [IB_RNR_TIMER_000_08] = 80, 58 [IB_RNR_TIMER_000_12] = 120, 59 [IB_RNR_TIMER_000_16] = 160, 60 [IB_RNR_TIMER_000_24] = 240, 61 [IB_RNR_TIMER_000_32] = 320, 62 [IB_RNR_TIMER_000_48] = 480, 63 [IB_RNR_TIMER_000_64] = 640, 64 [IB_RNR_TIMER_000_96] = 960, 65 [IB_RNR_TIMER_001_28] = 1280, 66 [IB_RNR_TIMER_001_92] = 1920, 67 [IB_RNR_TIMER_002_56] = 2560, 68 [IB_RNR_TIMER_003_84] = 3840, 69 [IB_RNR_TIMER_005_12] = 5120, 70 [IB_RNR_TIMER_007_68] = 7680, 71 [IB_RNR_TIMER_010_24] = 10240, 72 [IB_RNR_TIMER_015_36] = 15360, 73 [IB_RNR_TIMER_020_48] = 20480, 74 [IB_RNR_TIMER_030_72] = 30720, 75 [IB_RNR_TIMER_040_96] = 40960, 76 [IB_RNR_TIMER_061_44] = 61410, 77 [IB_RNR_TIMER_081_92] = 81920, 78 [IB_RNR_TIMER_122_88] = 122880, 79 [IB_RNR_TIMER_163_84] = 163840, 80 [IB_RNR_TIMER_245_76] = 245760, 81 [IB_RNR_TIMER_327_68] = 327680, 82 [IB_RNR_TIMER_491_52] = 491520, 83 }; 84 85 static inline unsigned long rnrnak_jiffies(u8 timeout) 86 { 87 return max_t(unsigned long, 88 usecs_to_jiffies(rnrnak_usec[timeout]), 1); 89 } 90 91 static enum ib_wc_opcode wr_to_wc_opcode(enum ib_wr_opcode opcode) 92 { 93 switch (opcode) { 94 case IB_WR_RDMA_WRITE: return IB_WC_RDMA_WRITE; 95 case IB_WR_RDMA_WRITE_WITH_IMM: return IB_WC_RDMA_WRITE; 96 case IB_WR_SEND: return IB_WC_SEND; 97 case IB_WR_SEND_WITH_IMM: return IB_WC_SEND; 98 case IB_WR_RDMA_READ: return IB_WC_RDMA_READ; 99 case IB_WR_ATOMIC_CMP_AND_SWP: return IB_WC_COMP_SWAP; 100 case IB_WR_ATOMIC_FETCH_AND_ADD: return IB_WC_FETCH_ADD; 101 case IB_WR_LSO: return IB_WC_LSO; 102 case IB_WR_SEND_WITH_INV: return IB_WC_SEND; 103 case IB_WR_RDMA_READ_WITH_INV: return IB_WC_RDMA_READ; 104 case IB_WR_LOCAL_INV: return IB_WC_LOCAL_INV; 105 case IB_WR_REG_MR: return IB_WC_REG_MR; 106 case IB_WR_BIND_MW: return IB_WC_BIND_MW; 107 108 default: 109 return 0xff; 110 } 111 } 112 113 void retransmit_timer(struct timer_list *t) 114 { 115 struct rxe_qp *qp = from_timer(qp, t, retrans_timer); 116 117 if (qp->valid) { 118 qp->comp.timeout = 1; 119 rxe_run_task(&qp->comp.task, 1); 120 } 121 } 122 123 void rxe_comp_queue_pkt(struct rxe_qp *qp, struct sk_buff *skb) 124 { 125 int must_sched; 126 127 skb_queue_tail(&qp->resp_pkts, skb); 128 129 must_sched = skb_queue_len(&qp->resp_pkts) > 1; 130 if (must_sched != 0) 131 rxe_counter_inc(SKB_TO_PKT(skb)->rxe, RXE_CNT_COMPLETER_SCHED); 132 133 rxe_run_task(&qp->comp.task, must_sched); 134 } 135 136 static inline enum comp_state get_wqe(struct rxe_qp *qp, 137 struct rxe_pkt_info *pkt, 138 struct rxe_send_wqe **wqe_p) 139 { 140 struct rxe_send_wqe *wqe; 141 142 /* we come here whether or not we found a response packet to see if 143 * there are any posted WQEs 144 */ 145 if (qp->is_user) 146 wqe = queue_head(qp->sq.queue, QUEUE_TYPE_FROM_USER); 147 else 148 wqe = queue_head(qp->sq.queue, QUEUE_TYPE_KERNEL); 149 *wqe_p = wqe; 150 151 /* no WQE or requester has not started it yet */ 152 if (!wqe || wqe->state == wqe_state_posted) 153 return pkt ? COMPST_DONE : COMPST_EXIT; 154 155 /* WQE does not require an ack */ 156 if (wqe->state == wqe_state_done) 157 return COMPST_COMP_WQE; 158 159 /* WQE caused an error */ 160 if (wqe->state == wqe_state_error) 161 return COMPST_ERROR; 162 163 /* we have a WQE, if we also have an ack check its PSN */ 164 return pkt ? COMPST_CHECK_PSN : COMPST_EXIT; 165 } 166 167 static inline void reset_retry_counters(struct rxe_qp *qp) 168 { 169 qp->comp.retry_cnt = qp->attr.retry_cnt; 170 qp->comp.rnr_retry = qp->attr.rnr_retry; 171 qp->comp.started_retry = 0; 172 } 173 174 static inline enum comp_state check_psn(struct rxe_qp *qp, 175 struct rxe_pkt_info *pkt, 176 struct rxe_send_wqe *wqe) 177 { 178 s32 diff; 179 180 /* check to see if response is past the oldest WQE. if it is, complete 181 * send/write or error read/atomic 182 */ 183 diff = psn_compare(pkt->psn, wqe->last_psn); 184 if (diff > 0) { 185 if (wqe->state == wqe_state_pending) { 186 if (wqe->mask & WR_ATOMIC_OR_READ_MASK) 187 return COMPST_ERROR_RETRY; 188 189 reset_retry_counters(qp); 190 return COMPST_COMP_WQE; 191 } else { 192 return COMPST_DONE; 193 } 194 } 195 196 /* compare response packet to expected response */ 197 diff = psn_compare(pkt->psn, qp->comp.psn); 198 if (diff < 0) { 199 /* response is most likely a retried packet if it matches an 200 * uncompleted WQE go complete it else ignore it 201 */ 202 if (pkt->psn == wqe->last_psn) 203 return COMPST_COMP_ACK; 204 else 205 return COMPST_DONE; 206 } else if ((diff > 0) && (wqe->mask & WR_ATOMIC_OR_READ_MASK)) { 207 return COMPST_DONE; 208 } else { 209 return COMPST_CHECK_ACK; 210 } 211 } 212 213 static inline enum comp_state check_ack(struct rxe_qp *qp, 214 struct rxe_pkt_info *pkt, 215 struct rxe_send_wqe *wqe) 216 { 217 unsigned int mask = pkt->mask; 218 u8 syn; 219 struct rxe_dev *rxe = to_rdev(qp->ibqp.device); 220 221 /* Check the sequence only */ 222 switch (qp->comp.opcode) { 223 case -1: 224 /* Will catch all *_ONLY cases. */ 225 if (!(mask & RXE_START_MASK)) 226 return COMPST_ERROR; 227 228 break; 229 230 case IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST: 231 case IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE: 232 if (pkt->opcode != IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE && 233 pkt->opcode != IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST) { 234 /* read retries of partial data may restart from 235 * read response first or response only. 236 */ 237 if ((pkt->psn == wqe->first_psn && 238 pkt->opcode == 239 IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST) || 240 (wqe->first_psn == wqe->last_psn && 241 pkt->opcode == 242 IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY)) 243 break; 244 245 return COMPST_ERROR; 246 } 247 break; 248 default: 249 WARN_ON_ONCE(1); 250 } 251 252 /* Check operation validity. */ 253 switch (pkt->opcode) { 254 case IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST: 255 case IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST: 256 case IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY: 257 syn = aeth_syn(pkt); 258 259 if ((syn & AETH_TYPE_MASK) != AETH_ACK) 260 return COMPST_ERROR; 261 262 fallthrough; 263 /* (IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE doesn't have an AETH) 264 */ 265 case IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE: 266 if (wqe->wr.opcode != IB_WR_RDMA_READ && 267 wqe->wr.opcode != IB_WR_RDMA_READ_WITH_INV) { 268 wqe->status = IB_WC_FATAL_ERR; 269 return COMPST_ERROR; 270 } 271 reset_retry_counters(qp); 272 return COMPST_READ; 273 274 case IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE: 275 syn = aeth_syn(pkt); 276 277 if ((syn & AETH_TYPE_MASK) != AETH_ACK) 278 return COMPST_ERROR; 279 280 if (wqe->wr.opcode != IB_WR_ATOMIC_CMP_AND_SWP && 281 wqe->wr.opcode != IB_WR_ATOMIC_FETCH_AND_ADD) 282 return COMPST_ERROR; 283 reset_retry_counters(qp); 284 return COMPST_ATOMIC; 285 286 case IB_OPCODE_RC_ACKNOWLEDGE: 287 syn = aeth_syn(pkt); 288 switch (syn & AETH_TYPE_MASK) { 289 case AETH_ACK: 290 reset_retry_counters(qp); 291 return COMPST_WRITE_SEND; 292 293 case AETH_RNR_NAK: 294 rxe_counter_inc(rxe, RXE_CNT_RCV_RNR); 295 return COMPST_RNR_RETRY; 296 297 case AETH_NAK: 298 switch (syn) { 299 case AETH_NAK_PSN_SEQ_ERROR: 300 /* a nak implicitly acks all packets with psns 301 * before 302 */ 303 if (psn_compare(pkt->psn, qp->comp.psn) > 0) { 304 rxe_counter_inc(rxe, 305 RXE_CNT_RCV_SEQ_ERR); 306 qp->comp.psn = pkt->psn; 307 if (qp->req.wait_psn) { 308 qp->req.wait_psn = 0; 309 rxe_run_task(&qp->req.task, 0); 310 } 311 } 312 return COMPST_ERROR_RETRY; 313 314 case AETH_NAK_INVALID_REQ: 315 wqe->status = IB_WC_REM_INV_REQ_ERR; 316 return COMPST_ERROR; 317 318 case AETH_NAK_REM_ACC_ERR: 319 wqe->status = IB_WC_REM_ACCESS_ERR; 320 return COMPST_ERROR; 321 322 case AETH_NAK_REM_OP_ERR: 323 wqe->status = IB_WC_REM_OP_ERR; 324 return COMPST_ERROR; 325 326 default: 327 pr_warn("unexpected nak %x\n", syn); 328 wqe->status = IB_WC_REM_OP_ERR; 329 return COMPST_ERROR; 330 } 331 332 default: 333 return COMPST_ERROR; 334 } 335 break; 336 337 default: 338 pr_warn("unexpected opcode\n"); 339 } 340 341 return COMPST_ERROR; 342 } 343 344 static inline enum comp_state do_read(struct rxe_qp *qp, 345 struct rxe_pkt_info *pkt, 346 struct rxe_send_wqe *wqe) 347 { 348 int ret; 349 350 ret = copy_data(qp->pd, IB_ACCESS_LOCAL_WRITE, 351 &wqe->dma, payload_addr(pkt), 352 payload_size(pkt), RXE_TO_MR_OBJ, NULL); 353 if (ret) { 354 wqe->status = IB_WC_LOC_PROT_ERR; 355 return COMPST_ERROR; 356 } 357 358 if (wqe->dma.resid == 0 && (pkt->mask & RXE_END_MASK)) 359 return COMPST_COMP_ACK; 360 361 return COMPST_UPDATE_COMP; 362 } 363 364 static inline enum comp_state do_atomic(struct rxe_qp *qp, 365 struct rxe_pkt_info *pkt, 366 struct rxe_send_wqe *wqe) 367 { 368 int ret; 369 370 u64 atomic_orig = atmack_orig(pkt); 371 372 ret = copy_data(qp->pd, IB_ACCESS_LOCAL_WRITE, 373 &wqe->dma, &atomic_orig, 374 sizeof(u64), RXE_TO_MR_OBJ, NULL); 375 if (ret) { 376 wqe->status = IB_WC_LOC_PROT_ERR; 377 return COMPST_ERROR; 378 } 379 380 return COMPST_COMP_ACK; 381 } 382 383 static void make_send_cqe(struct rxe_qp *qp, struct rxe_send_wqe *wqe, 384 struct rxe_cqe *cqe) 385 { 386 memset(cqe, 0, sizeof(*cqe)); 387 388 if (!qp->is_user) { 389 struct ib_wc *wc = &cqe->ibwc; 390 391 wc->wr_id = wqe->wr.wr_id; 392 wc->status = wqe->status; 393 wc->opcode = wr_to_wc_opcode(wqe->wr.opcode); 394 if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM || 395 wqe->wr.opcode == IB_WR_SEND_WITH_IMM) 396 wc->wc_flags = IB_WC_WITH_IMM; 397 wc->byte_len = wqe->dma.length; 398 wc->qp = &qp->ibqp; 399 } else { 400 struct ib_uverbs_wc *uwc = &cqe->uibwc; 401 402 uwc->wr_id = wqe->wr.wr_id; 403 uwc->status = wqe->status; 404 uwc->opcode = wr_to_wc_opcode(wqe->wr.opcode); 405 if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM || 406 wqe->wr.opcode == IB_WR_SEND_WITH_IMM) 407 uwc->wc_flags = IB_WC_WITH_IMM; 408 uwc->byte_len = wqe->dma.length; 409 uwc->qp_num = qp->ibqp.qp_num; 410 } 411 } 412 413 /* 414 * IBA Spec. Section 10.7.3.1 SIGNALED COMPLETIONS 415 * ---------8<---------8<------------- 416 * ...Note that if a completion error occurs, a Work Completion 417 * will always be generated, even if the signaling 418 * indicator requests an Unsignaled Completion. 419 * ---------8<---------8<------------- 420 */ 421 static void do_complete(struct rxe_qp *qp, struct rxe_send_wqe *wqe) 422 { 423 struct rxe_dev *rxe = to_rdev(qp->ibqp.device); 424 struct rxe_cqe cqe; 425 bool post; 426 427 /* do we need to post a completion */ 428 post = ((qp->sq_sig_type == IB_SIGNAL_ALL_WR) || 429 (wqe->wr.send_flags & IB_SEND_SIGNALED) || 430 wqe->status != IB_WC_SUCCESS); 431 432 if (post) 433 make_send_cqe(qp, wqe, &cqe); 434 435 if (qp->is_user) 436 advance_consumer(qp->sq.queue, QUEUE_TYPE_FROM_USER); 437 else 438 advance_consumer(qp->sq.queue, QUEUE_TYPE_KERNEL); 439 440 if (post) 441 rxe_cq_post(qp->scq, &cqe, 0); 442 443 if (wqe->wr.opcode == IB_WR_SEND || 444 wqe->wr.opcode == IB_WR_SEND_WITH_IMM || 445 wqe->wr.opcode == IB_WR_SEND_WITH_INV) 446 rxe_counter_inc(rxe, RXE_CNT_RDMA_SEND); 447 448 /* 449 * we completed something so let req run again 450 * if it is trying to fence 451 */ 452 if (qp->req.wait_fence) { 453 qp->req.wait_fence = 0; 454 rxe_run_task(&qp->req.task, 0); 455 } 456 } 457 458 static inline enum comp_state complete_ack(struct rxe_qp *qp, 459 struct rxe_pkt_info *pkt, 460 struct rxe_send_wqe *wqe) 461 { 462 unsigned long flags; 463 464 if (wqe->has_rd_atomic) { 465 wqe->has_rd_atomic = 0; 466 atomic_inc(&qp->req.rd_atomic); 467 if (qp->req.need_rd_atomic) { 468 qp->comp.timeout_retry = 0; 469 qp->req.need_rd_atomic = 0; 470 rxe_run_task(&qp->req.task, 0); 471 } 472 } 473 474 if (unlikely(qp->req.state == QP_STATE_DRAIN)) { 475 /* state_lock used by requester & completer */ 476 spin_lock_irqsave(&qp->state_lock, flags); 477 if ((qp->req.state == QP_STATE_DRAIN) && 478 (qp->comp.psn == qp->req.psn)) { 479 qp->req.state = QP_STATE_DRAINED; 480 spin_unlock_irqrestore(&qp->state_lock, flags); 481 482 if (qp->ibqp.event_handler) { 483 struct ib_event ev; 484 485 ev.device = qp->ibqp.device; 486 ev.element.qp = &qp->ibqp; 487 ev.event = IB_EVENT_SQ_DRAINED; 488 qp->ibqp.event_handler(&ev, 489 qp->ibqp.qp_context); 490 } 491 } else { 492 spin_unlock_irqrestore(&qp->state_lock, flags); 493 } 494 } 495 496 do_complete(qp, wqe); 497 498 if (psn_compare(pkt->psn, qp->comp.psn) >= 0) 499 return COMPST_UPDATE_COMP; 500 else 501 return COMPST_DONE; 502 } 503 504 static inline enum comp_state complete_wqe(struct rxe_qp *qp, 505 struct rxe_pkt_info *pkt, 506 struct rxe_send_wqe *wqe) 507 { 508 if (pkt && wqe->state == wqe_state_pending) { 509 if (psn_compare(wqe->last_psn, qp->comp.psn) >= 0) { 510 qp->comp.psn = (wqe->last_psn + 1) & BTH_PSN_MASK; 511 qp->comp.opcode = -1; 512 } 513 514 if (qp->req.wait_psn) { 515 qp->req.wait_psn = 0; 516 rxe_run_task(&qp->req.task, 1); 517 } 518 } 519 520 do_complete(qp, wqe); 521 522 return COMPST_GET_WQE; 523 } 524 525 static void rxe_drain_resp_pkts(struct rxe_qp *qp, bool notify) 526 { 527 struct sk_buff *skb; 528 struct rxe_send_wqe *wqe; 529 struct rxe_queue *q = qp->sq.queue; 530 531 while ((skb = skb_dequeue(&qp->resp_pkts))) { 532 rxe_drop_ref(qp); 533 kfree_skb(skb); 534 ib_device_put(qp->ibqp.device); 535 } 536 537 while ((wqe = queue_head(q, q->type))) { 538 if (notify) { 539 wqe->status = IB_WC_WR_FLUSH_ERR; 540 do_complete(qp, wqe); 541 } else { 542 advance_consumer(q, q->type); 543 } 544 } 545 } 546 547 static void free_pkt(struct rxe_pkt_info *pkt) 548 { 549 struct sk_buff *skb = PKT_TO_SKB(pkt); 550 struct rxe_qp *qp = pkt->qp; 551 struct ib_device *dev = qp->ibqp.device; 552 553 kfree_skb(skb); 554 rxe_drop_ref(qp); 555 ib_device_put(dev); 556 } 557 558 int rxe_completer(void *arg) 559 { 560 struct rxe_qp *qp = (struct rxe_qp *)arg; 561 struct rxe_dev *rxe = to_rdev(qp->ibqp.device); 562 struct rxe_send_wqe *wqe = NULL; 563 struct sk_buff *skb = NULL; 564 struct rxe_pkt_info *pkt = NULL; 565 enum comp_state state; 566 int ret = 0; 567 568 rxe_add_ref(qp); 569 570 if (!qp->valid || qp->req.state == QP_STATE_ERROR || 571 qp->req.state == QP_STATE_RESET) { 572 rxe_drain_resp_pkts(qp, qp->valid && 573 qp->req.state == QP_STATE_ERROR); 574 ret = -EAGAIN; 575 goto done; 576 } 577 578 if (qp->comp.timeout) { 579 qp->comp.timeout_retry = 1; 580 qp->comp.timeout = 0; 581 } else { 582 qp->comp.timeout_retry = 0; 583 } 584 585 if (qp->req.need_retry) { 586 ret = -EAGAIN; 587 goto done; 588 } 589 590 state = COMPST_GET_ACK; 591 592 while (1) { 593 pr_debug("qp#%d state = %s\n", qp_num(qp), 594 comp_state_name[state]); 595 switch (state) { 596 case COMPST_GET_ACK: 597 skb = skb_dequeue(&qp->resp_pkts); 598 if (skb) { 599 pkt = SKB_TO_PKT(skb); 600 qp->comp.timeout_retry = 0; 601 } 602 state = COMPST_GET_WQE; 603 break; 604 605 case COMPST_GET_WQE: 606 state = get_wqe(qp, pkt, &wqe); 607 break; 608 609 case COMPST_CHECK_PSN: 610 state = check_psn(qp, pkt, wqe); 611 break; 612 613 case COMPST_CHECK_ACK: 614 state = check_ack(qp, pkt, wqe); 615 break; 616 617 case COMPST_READ: 618 state = do_read(qp, pkt, wqe); 619 break; 620 621 case COMPST_ATOMIC: 622 state = do_atomic(qp, pkt, wqe); 623 break; 624 625 case COMPST_WRITE_SEND: 626 if (wqe->state == wqe_state_pending && 627 wqe->last_psn == pkt->psn) 628 state = COMPST_COMP_ACK; 629 else 630 state = COMPST_UPDATE_COMP; 631 break; 632 633 case COMPST_COMP_ACK: 634 state = complete_ack(qp, pkt, wqe); 635 break; 636 637 case COMPST_COMP_WQE: 638 state = complete_wqe(qp, pkt, wqe); 639 break; 640 641 case COMPST_UPDATE_COMP: 642 if (pkt->mask & RXE_END_MASK) 643 qp->comp.opcode = -1; 644 else 645 qp->comp.opcode = pkt->opcode; 646 647 if (psn_compare(pkt->psn, qp->comp.psn) >= 0) 648 qp->comp.psn = (pkt->psn + 1) & BTH_PSN_MASK; 649 650 if (qp->req.wait_psn) { 651 qp->req.wait_psn = 0; 652 rxe_run_task(&qp->req.task, 1); 653 } 654 655 state = COMPST_DONE; 656 break; 657 658 case COMPST_DONE: 659 goto done; 660 661 case COMPST_EXIT: 662 if (qp->comp.timeout_retry && wqe) { 663 state = COMPST_ERROR_RETRY; 664 break; 665 } 666 667 /* re reset the timeout counter if 668 * (1) QP is type RC 669 * (2) the QP is alive 670 * (3) there is a packet sent by the requester that 671 * might be acked (we still might get spurious 672 * timeouts but try to keep them as few as possible) 673 * (4) the timeout parameter is set 674 */ 675 if ((qp_type(qp) == IB_QPT_RC) && 676 (qp->req.state == QP_STATE_READY) && 677 (psn_compare(qp->req.psn, qp->comp.psn) > 0) && 678 qp->qp_timeout_jiffies) 679 mod_timer(&qp->retrans_timer, 680 jiffies + qp->qp_timeout_jiffies); 681 ret = -EAGAIN; 682 goto done; 683 684 case COMPST_ERROR_RETRY: 685 /* we come here if the retry timer fired and we did 686 * not receive a response packet. try to retry the send 687 * queue if that makes sense and the limits have not 688 * been exceeded. remember that some timeouts are 689 * spurious since we do not reset the timer but kick 690 * it down the road or let it expire 691 */ 692 693 /* there is nothing to retry in this case */ 694 if (!wqe || (wqe->state == wqe_state_posted)) { 695 ret = -EAGAIN; 696 goto done; 697 } 698 699 /* if we've started a retry, don't start another 700 * retry sequence, unless this is a timeout. 701 */ 702 if (qp->comp.started_retry && 703 !qp->comp.timeout_retry) 704 goto done; 705 706 if (qp->comp.retry_cnt > 0) { 707 if (qp->comp.retry_cnt != 7) 708 qp->comp.retry_cnt--; 709 710 /* no point in retrying if we have already 711 * seen the last ack that the requester could 712 * have caused 713 */ 714 if (psn_compare(qp->req.psn, 715 qp->comp.psn) > 0) { 716 /* tell the requester to retry the 717 * send queue next time around 718 */ 719 rxe_counter_inc(rxe, 720 RXE_CNT_COMP_RETRY); 721 qp->req.need_retry = 1; 722 qp->comp.started_retry = 1; 723 rxe_run_task(&qp->req.task, 0); 724 } 725 goto done; 726 727 } else { 728 rxe_counter_inc(rxe, RXE_CNT_RETRY_EXCEEDED); 729 wqe->status = IB_WC_RETRY_EXC_ERR; 730 state = COMPST_ERROR; 731 } 732 break; 733 734 case COMPST_RNR_RETRY: 735 if (qp->comp.rnr_retry > 0) { 736 if (qp->comp.rnr_retry != 7) 737 qp->comp.rnr_retry--; 738 739 qp->req.need_retry = 1; 740 pr_debug("qp#%d set rnr nak timer\n", 741 qp_num(qp)); 742 mod_timer(&qp->rnr_nak_timer, 743 jiffies + rnrnak_jiffies(aeth_syn(pkt) 744 & ~AETH_TYPE_MASK)); 745 ret = -EAGAIN; 746 goto done; 747 } else { 748 rxe_counter_inc(rxe, 749 RXE_CNT_RNR_RETRY_EXCEEDED); 750 wqe->status = IB_WC_RNR_RETRY_EXC_ERR; 751 state = COMPST_ERROR; 752 } 753 break; 754 755 case COMPST_ERROR: 756 WARN_ON_ONCE(wqe->status == IB_WC_SUCCESS); 757 do_complete(qp, wqe); 758 rxe_qp_error(qp); 759 ret = -EAGAIN; 760 goto done; 761 } 762 } 763 764 done: 765 if (pkt) 766 free_pkt(pkt); 767 rxe_drop_ref(qp); 768 769 return ret; 770 } 771