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 enum resp_states { 14 RESPST_NONE, 15 RESPST_GET_REQ, 16 RESPST_CHK_PSN, 17 RESPST_CHK_OP_SEQ, 18 RESPST_CHK_OP_VALID, 19 RESPST_CHK_RESOURCE, 20 RESPST_CHK_LENGTH, 21 RESPST_CHK_RKEY, 22 RESPST_EXECUTE, 23 RESPST_READ_REPLY, 24 RESPST_ATOMIC_REPLY, 25 RESPST_COMPLETE, 26 RESPST_ACKNOWLEDGE, 27 RESPST_CLEANUP, 28 RESPST_DUPLICATE_REQUEST, 29 RESPST_ERR_MALFORMED_WQE, 30 RESPST_ERR_UNSUPPORTED_OPCODE, 31 RESPST_ERR_MISALIGNED_ATOMIC, 32 RESPST_ERR_PSN_OUT_OF_SEQ, 33 RESPST_ERR_MISSING_OPCODE_FIRST, 34 RESPST_ERR_MISSING_OPCODE_LAST_C, 35 RESPST_ERR_MISSING_OPCODE_LAST_D1E, 36 RESPST_ERR_TOO_MANY_RDMA_ATM_REQ, 37 RESPST_ERR_RNR, 38 RESPST_ERR_RKEY_VIOLATION, 39 RESPST_ERR_INVALIDATE_RKEY, 40 RESPST_ERR_LENGTH, 41 RESPST_ERR_CQ_OVERFLOW, 42 RESPST_ERROR, 43 RESPST_RESET, 44 RESPST_DONE, 45 RESPST_EXIT, 46 }; 47 48 static char *resp_state_name[] = { 49 [RESPST_NONE] = "NONE", 50 [RESPST_GET_REQ] = "GET_REQ", 51 [RESPST_CHK_PSN] = "CHK_PSN", 52 [RESPST_CHK_OP_SEQ] = "CHK_OP_SEQ", 53 [RESPST_CHK_OP_VALID] = "CHK_OP_VALID", 54 [RESPST_CHK_RESOURCE] = "CHK_RESOURCE", 55 [RESPST_CHK_LENGTH] = "CHK_LENGTH", 56 [RESPST_CHK_RKEY] = "CHK_RKEY", 57 [RESPST_EXECUTE] = "EXECUTE", 58 [RESPST_READ_REPLY] = "READ_REPLY", 59 [RESPST_ATOMIC_REPLY] = "ATOMIC_REPLY", 60 [RESPST_COMPLETE] = "COMPLETE", 61 [RESPST_ACKNOWLEDGE] = "ACKNOWLEDGE", 62 [RESPST_CLEANUP] = "CLEANUP", 63 [RESPST_DUPLICATE_REQUEST] = "DUPLICATE_REQUEST", 64 [RESPST_ERR_MALFORMED_WQE] = "ERR_MALFORMED_WQE", 65 [RESPST_ERR_UNSUPPORTED_OPCODE] = "ERR_UNSUPPORTED_OPCODE", 66 [RESPST_ERR_MISALIGNED_ATOMIC] = "ERR_MISALIGNED_ATOMIC", 67 [RESPST_ERR_PSN_OUT_OF_SEQ] = "ERR_PSN_OUT_OF_SEQ", 68 [RESPST_ERR_MISSING_OPCODE_FIRST] = "ERR_MISSING_OPCODE_FIRST", 69 [RESPST_ERR_MISSING_OPCODE_LAST_C] = "ERR_MISSING_OPCODE_LAST_C", 70 [RESPST_ERR_MISSING_OPCODE_LAST_D1E] = "ERR_MISSING_OPCODE_LAST_D1E", 71 [RESPST_ERR_TOO_MANY_RDMA_ATM_REQ] = "ERR_TOO_MANY_RDMA_ATM_REQ", 72 [RESPST_ERR_RNR] = "ERR_RNR", 73 [RESPST_ERR_RKEY_VIOLATION] = "ERR_RKEY_VIOLATION", 74 [RESPST_ERR_INVALIDATE_RKEY] = "ERR_INVALIDATE_RKEY_VIOLATION", 75 [RESPST_ERR_LENGTH] = "ERR_LENGTH", 76 [RESPST_ERR_CQ_OVERFLOW] = "ERR_CQ_OVERFLOW", 77 [RESPST_ERROR] = "ERROR", 78 [RESPST_RESET] = "RESET", 79 [RESPST_DONE] = "DONE", 80 [RESPST_EXIT] = "EXIT", 81 }; 82 83 /* rxe_recv calls here to add a request packet to the input queue */ 84 void rxe_resp_queue_pkt(struct rxe_qp *qp, struct sk_buff *skb) 85 { 86 int must_sched; 87 struct rxe_pkt_info *pkt = SKB_TO_PKT(skb); 88 89 skb_queue_tail(&qp->req_pkts, skb); 90 91 must_sched = (pkt->opcode == IB_OPCODE_RC_RDMA_READ_REQUEST) || 92 (skb_queue_len(&qp->req_pkts) > 1); 93 94 rxe_run_task(&qp->resp.task, must_sched); 95 } 96 97 static inline enum resp_states get_req(struct rxe_qp *qp, 98 struct rxe_pkt_info **pkt_p) 99 { 100 struct sk_buff *skb; 101 102 if (qp->resp.state == QP_STATE_ERROR) { 103 while ((skb = skb_dequeue(&qp->req_pkts))) { 104 rxe_put(qp); 105 kfree_skb(skb); 106 ib_device_put(qp->ibqp.device); 107 } 108 109 /* go drain recv wr queue */ 110 return RESPST_CHK_RESOURCE; 111 } 112 113 skb = skb_peek(&qp->req_pkts); 114 if (!skb) 115 return RESPST_EXIT; 116 117 *pkt_p = SKB_TO_PKT(skb); 118 119 return (qp->resp.res) ? RESPST_READ_REPLY : RESPST_CHK_PSN; 120 } 121 122 static enum resp_states check_psn(struct rxe_qp *qp, 123 struct rxe_pkt_info *pkt) 124 { 125 int diff = psn_compare(pkt->psn, qp->resp.psn); 126 struct rxe_dev *rxe = to_rdev(qp->ibqp.device); 127 128 switch (qp_type(qp)) { 129 case IB_QPT_RC: 130 if (diff > 0) { 131 if (qp->resp.sent_psn_nak) 132 return RESPST_CLEANUP; 133 134 qp->resp.sent_psn_nak = 1; 135 rxe_counter_inc(rxe, RXE_CNT_OUT_OF_SEQ_REQ); 136 return RESPST_ERR_PSN_OUT_OF_SEQ; 137 138 } else if (diff < 0) { 139 rxe_counter_inc(rxe, RXE_CNT_DUP_REQ); 140 return RESPST_DUPLICATE_REQUEST; 141 } 142 143 if (qp->resp.sent_psn_nak) 144 qp->resp.sent_psn_nak = 0; 145 146 break; 147 148 case IB_QPT_UC: 149 if (qp->resp.drop_msg || diff != 0) { 150 if (pkt->mask & RXE_START_MASK) { 151 qp->resp.drop_msg = 0; 152 return RESPST_CHK_OP_SEQ; 153 } 154 155 qp->resp.drop_msg = 1; 156 return RESPST_CLEANUP; 157 } 158 break; 159 default: 160 break; 161 } 162 163 return RESPST_CHK_OP_SEQ; 164 } 165 166 static enum resp_states check_op_seq(struct rxe_qp *qp, 167 struct rxe_pkt_info *pkt) 168 { 169 switch (qp_type(qp)) { 170 case IB_QPT_RC: 171 switch (qp->resp.opcode) { 172 case IB_OPCODE_RC_SEND_FIRST: 173 case IB_OPCODE_RC_SEND_MIDDLE: 174 switch (pkt->opcode) { 175 case IB_OPCODE_RC_SEND_MIDDLE: 176 case IB_OPCODE_RC_SEND_LAST: 177 case IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE: 178 case IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE: 179 return RESPST_CHK_OP_VALID; 180 default: 181 return RESPST_ERR_MISSING_OPCODE_LAST_C; 182 } 183 184 case IB_OPCODE_RC_RDMA_WRITE_FIRST: 185 case IB_OPCODE_RC_RDMA_WRITE_MIDDLE: 186 switch (pkt->opcode) { 187 case IB_OPCODE_RC_RDMA_WRITE_MIDDLE: 188 case IB_OPCODE_RC_RDMA_WRITE_LAST: 189 case IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE: 190 return RESPST_CHK_OP_VALID; 191 default: 192 return RESPST_ERR_MISSING_OPCODE_LAST_C; 193 } 194 195 default: 196 switch (pkt->opcode) { 197 case IB_OPCODE_RC_SEND_MIDDLE: 198 case IB_OPCODE_RC_SEND_LAST: 199 case IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE: 200 case IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE: 201 case IB_OPCODE_RC_RDMA_WRITE_MIDDLE: 202 case IB_OPCODE_RC_RDMA_WRITE_LAST: 203 case IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE: 204 return RESPST_ERR_MISSING_OPCODE_FIRST; 205 default: 206 return RESPST_CHK_OP_VALID; 207 } 208 } 209 break; 210 211 case IB_QPT_UC: 212 switch (qp->resp.opcode) { 213 case IB_OPCODE_UC_SEND_FIRST: 214 case IB_OPCODE_UC_SEND_MIDDLE: 215 switch (pkt->opcode) { 216 case IB_OPCODE_UC_SEND_MIDDLE: 217 case IB_OPCODE_UC_SEND_LAST: 218 case IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE: 219 return RESPST_CHK_OP_VALID; 220 default: 221 return RESPST_ERR_MISSING_OPCODE_LAST_D1E; 222 } 223 224 case IB_OPCODE_UC_RDMA_WRITE_FIRST: 225 case IB_OPCODE_UC_RDMA_WRITE_MIDDLE: 226 switch (pkt->opcode) { 227 case IB_OPCODE_UC_RDMA_WRITE_MIDDLE: 228 case IB_OPCODE_UC_RDMA_WRITE_LAST: 229 case IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE: 230 return RESPST_CHK_OP_VALID; 231 default: 232 return RESPST_ERR_MISSING_OPCODE_LAST_D1E; 233 } 234 235 default: 236 switch (pkt->opcode) { 237 case IB_OPCODE_UC_SEND_MIDDLE: 238 case IB_OPCODE_UC_SEND_LAST: 239 case IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE: 240 case IB_OPCODE_UC_RDMA_WRITE_MIDDLE: 241 case IB_OPCODE_UC_RDMA_WRITE_LAST: 242 case IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE: 243 qp->resp.drop_msg = 1; 244 return RESPST_CLEANUP; 245 default: 246 return RESPST_CHK_OP_VALID; 247 } 248 } 249 break; 250 251 default: 252 return RESPST_CHK_OP_VALID; 253 } 254 } 255 256 static enum resp_states check_op_valid(struct rxe_qp *qp, 257 struct rxe_pkt_info *pkt) 258 { 259 switch (qp_type(qp)) { 260 case IB_QPT_RC: 261 if (((pkt->mask & RXE_READ_MASK) && 262 !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_READ)) || 263 ((pkt->mask & RXE_WRITE_MASK) && 264 !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_WRITE)) || 265 ((pkt->mask & RXE_ATOMIC_MASK) && 266 !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_ATOMIC))) { 267 return RESPST_ERR_UNSUPPORTED_OPCODE; 268 } 269 270 break; 271 272 case IB_QPT_UC: 273 if ((pkt->mask & RXE_WRITE_MASK) && 274 !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_WRITE)) { 275 qp->resp.drop_msg = 1; 276 return RESPST_CLEANUP; 277 } 278 279 break; 280 281 case IB_QPT_UD: 282 case IB_QPT_GSI: 283 break; 284 285 default: 286 WARN_ON_ONCE(1); 287 break; 288 } 289 290 return RESPST_CHK_RESOURCE; 291 } 292 293 static enum resp_states get_srq_wqe(struct rxe_qp *qp) 294 { 295 struct rxe_srq *srq = qp->srq; 296 struct rxe_queue *q = srq->rq.queue; 297 struct rxe_recv_wqe *wqe; 298 struct ib_event ev; 299 unsigned int count; 300 size_t size; 301 unsigned long flags; 302 303 if (srq->error) 304 return RESPST_ERR_RNR; 305 306 spin_lock_irqsave(&srq->rq.consumer_lock, flags); 307 308 wqe = queue_head(q, QUEUE_TYPE_FROM_CLIENT); 309 if (!wqe) { 310 spin_unlock_irqrestore(&srq->rq.consumer_lock, flags); 311 return RESPST_ERR_RNR; 312 } 313 314 /* don't trust user space data */ 315 if (unlikely(wqe->dma.num_sge > srq->rq.max_sge)) { 316 spin_unlock_irqrestore(&srq->rq.consumer_lock, flags); 317 pr_warn("%s: invalid num_sge in SRQ entry\n", __func__); 318 return RESPST_ERR_MALFORMED_WQE; 319 } 320 size = sizeof(*wqe) + wqe->dma.num_sge*sizeof(struct rxe_sge); 321 memcpy(&qp->resp.srq_wqe, wqe, size); 322 323 qp->resp.wqe = &qp->resp.srq_wqe.wqe; 324 queue_advance_consumer(q, QUEUE_TYPE_FROM_CLIENT); 325 count = queue_count(q, QUEUE_TYPE_FROM_CLIENT); 326 327 if (srq->limit && srq->ibsrq.event_handler && (count < srq->limit)) { 328 srq->limit = 0; 329 goto event; 330 } 331 332 spin_unlock_irqrestore(&srq->rq.consumer_lock, flags); 333 return RESPST_CHK_LENGTH; 334 335 event: 336 spin_unlock_irqrestore(&srq->rq.consumer_lock, flags); 337 ev.device = qp->ibqp.device; 338 ev.element.srq = qp->ibqp.srq; 339 ev.event = IB_EVENT_SRQ_LIMIT_REACHED; 340 srq->ibsrq.event_handler(&ev, srq->ibsrq.srq_context); 341 return RESPST_CHK_LENGTH; 342 } 343 344 static enum resp_states check_resource(struct rxe_qp *qp, 345 struct rxe_pkt_info *pkt) 346 { 347 struct rxe_srq *srq = qp->srq; 348 349 if (qp->resp.state == QP_STATE_ERROR) { 350 if (qp->resp.wqe) { 351 qp->resp.status = IB_WC_WR_FLUSH_ERR; 352 return RESPST_COMPLETE; 353 } else if (!srq) { 354 qp->resp.wqe = queue_head(qp->rq.queue, 355 QUEUE_TYPE_FROM_CLIENT); 356 if (qp->resp.wqe) { 357 qp->resp.status = IB_WC_WR_FLUSH_ERR; 358 return RESPST_COMPLETE; 359 } else { 360 return RESPST_EXIT; 361 } 362 } else { 363 return RESPST_EXIT; 364 } 365 } 366 367 if (pkt->mask & RXE_READ_OR_ATOMIC_MASK) { 368 /* it is the requesters job to not send 369 * too many read/atomic ops, we just 370 * recycle the responder resource queue 371 */ 372 if (likely(qp->attr.max_dest_rd_atomic > 0)) 373 return RESPST_CHK_LENGTH; 374 else 375 return RESPST_ERR_TOO_MANY_RDMA_ATM_REQ; 376 } 377 378 if (pkt->mask & RXE_RWR_MASK) { 379 if (srq) 380 return get_srq_wqe(qp); 381 382 qp->resp.wqe = queue_head(qp->rq.queue, 383 QUEUE_TYPE_FROM_CLIENT); 384 return (qp->resp.wqe) ? RESPST_CHK_LENGTH : RESPST_ERR_RNR; 385 } 386 387 return RESPST_CHK_LENGTH; 388 } 389 390 static enum resp_states check_length(struct rxe_qp *qp, 391 struct rxe_pkt_info *pkt) 392 { 393 switch (qp_type(qp)) { 394 case IB_QPT_RC: 395 return RESPST_CHK_RKEY; 396 397 case IB_QPT_UC: 398 return RESPST_CHK_RKEY; 399 400 default: 401 return RESPST_CHK_RKEY; 402 } 403 } 404 405 static enum resp_states check_rkey(struct rxe_qp *qp, 406 struct rxe_pkt_info *pkt) 407 { 408 struct rxe_mr *mr = NULL; 409 struct rxe_mw *mw = NULL; 410 u64 va; 411 u32 rkey; 412 u32 resid; 413 u32 pktlen; 414 int mtu = qp->mtu; 415 enum resp_states state; 416 int access; 417 418 if (pkt->mask & RXE_READ_OR_WRITE_MASK) { 419 if (pkt->mask & RXE_RETH_MASK) { 420 qp->resp.va = reth_va(pkt); 421 qp->resp.offset = 0; 422 qp->resp.rkey = reth_rkey(pkt); 423 qp->resp.resid = reth_len(pkt); 424 qp->resp.length = reth_len(pkt); 425 } 426 access = (pkt->mask & RXE_READ_MASK) ? IB_ACCESS_REMOTE_READ 427 : IB_ACCESS_REMOTE_WRITE; 428 } else if (pkt->mask & RXE_ATOMIC_MASK) { 429 qp->resp.va = atmeth_va(pkt); 430 qp->resp.offset = 0; 431 qp->resp.rkey = atmeth_rkey(pkt); 432 qp->resp.resid = sizeof(u64); 433 access = IB_ACCESS_REMOTE_ATOMIC; 434 } else { 435 return RESPST_EXECUTE; 436 } 437 438 /* A zero-byte op is not required to set an addr or rkey. */ 439 if ((pkt->mask & RXE_READ_OR_WRITE_MASK) && 440 (pkt->mask & RXE_RETH_MASK) && 441 reth_len(pkt) == 0) { 442 return RESPST_EXECUTE; 443 } 444 445 va = qp->resp.va; 446 rkey = qp->resp.rkey; 447 resid = qp->resp.resid; 448 pktlen = payload_size(pkt); 449 450 if (rkey_is_mw(rkey)) { 451 mw = rxe_lookup_mw(qp, access, rkey); 452 if (!mw) { 453 pr_debug("%s: no MW matches rkey %#x\n", 454 __func__, rkey); 455 state = RESPST_ERR_RKEY_VIOLATION; 456 goto err; 457 } 458 459 mr = mw->mr; 460 if (!mr) { 461 pr_err("%s: MW doesn't have an MR\n", __func__); 462 state = RESPST_ERR_RKEY_VIOLATION; 463 goto err; 464 } 465 466 if (mw->access & IB_ZERO_BASED) 467 qp->resp.offset = mw->addr; 468 469 rxe_put(mw); 470 rxe_get(mr); 471 } else { 472 mr = lookup_mr(qp->pd, access, rkey, RXE_LOOKUP_REMOTE); 473 if (!mr) { 474 pr_debug("%s: no MR matches rkey %#x\n", 475 __func__, rkey); 476 state = RESPST_ERR_RKEY_VIOLATION; 477 goto err; 478 } 479 } 480 481 if (mr_check_range(mr, va + qp->resp.offset, resid)) { 482 state = RESPST_ERR_RKEY_VIOLATION; 483 goto err; 484 } 485 486 if (pkt->mask & RXE_WRITE_MASK) { 487 if (resid > mtu) { 488 if (pktlen != mtu || bth_pad(pkt)) { 489 state = RESPST_ERR_LENGTH; 490 goto err; 491 } 492 } else { 493 if (pktlen != resid) { 494 state = RESPST_ERR_LENGTH; 495 goto err; 496 } 497 if ((bth_pad(pkt) != (0x3 & (-resid)))) { 498 /* This case may not be exactly that 499 * but nothing else fits. 500 */ 501 state = RESPST_ERR_LENGTH; 502 goto err; 503 } 504 } 505 } 506 507 WARN_ON_ONCE(qp->resp.mr); 508 509 qp->resp.mr = mr; 510 return RESPST_EXECUTE; 511 512 err: 513 if (mr) 514 rxe_put(mr); 515 if (mw) 516 rxe_put(mw); 517 518 return state; 519 } 520 521 static enum resp_states send_data_in(struct rxe_qp *qp, void *data_addr, 522 int data_len) 523 { 524 int err; 525 526 err = copy_data(qp->pd, IB_ACCESS_LOCAL_WRITE, &qp->resp.wqe->dma, 527 data_addr, data_len, RXE_TO_MR_OBJ); 528 if (unlikely(err)) 529 return (err == -ENOSPC) ? RESPST_ERR_LENGTH 530 : RESPST_ERR_MALFORMED_WQE; 531 532 return RESPST_NONE; 533 } 534 535 static enum resp_states write_data_in(struct rxe_qp *qp, 536 struct rxe_pkt_info *pkt) 537 { 538 enum resp_states rc = RESPST_NONE; 539 int err; 540 int data_len = payload_size(pkt); 541 542 err = rxe_mr_copy(qp->resp.mr, qp->resp.va + qp->resp.offset, 543 payload_addr(pkt), data_len, RXE_TO_MR_OBJ); 544 if (err) { 545 rc = RESPST_ERR_RKEY_VIOLATION; 546 goto out; 547 } 548 549 qp->resp.va += data_len; 550 qp->resp.resid -= data_len; 551 552 out: 553 return rc; 554 } 555 556 static struct resp_res *rxe_prepare_res(struct rxe_qp *qp, 557 struct rxe_pkt_info *pkt, 558 int type) 559 { 560 struct resp_res *res; 561 u32 pkts; 562 563 res = &qp->resp.resources[qp->resp.res_head]; 564 rxe_advance_resp_resource(qp); 565 free_rd_atomic_resource(res); 566 567 res->type = type; 568 res->replay = 0; 569 570 switch (type) { 571 case RXE_READ_MASK: 572 res->read.va = qp->resp.va + qp->resp.offset; 573 res->read.va_org = qp->resp.va + qp->resp.offset; 574 res->read.resid = qp->resp.resid; 575 res->read.length = qp->resp.resid; 576 res->read.rkey = qp->resp.rkey; 577 578 pkts = max_t(u32, (reth_len(pkt) + qp->mtu - 1)/qp->mtu, 1); 579 res->first_psn = pkt->psn; 580 res->cur_psn = pkt->psn; 581 res->last_psn = (pkt->psn + pkts - 1) & BTH_PSN_MASK; 582 583 res->state = rdatm_res_state_new; 584 break; 585 case RXE_ATOMIC_MASK: 586 res->first_psn = pkt->psn; 587 res->last_psn = pkt->psn; 588 res->cur_psn = pkt->psn; 589 break; 590 } 591 592 return res; 593 } 594 595 /* Guarantee atomicity of atomic operations at the machine level. */ 596 static DEFINE_SPINLOCK(atomic_ops_lock); 597 598 static enum resp_states atomic_reply(struct rxe_qp *qp, 599 struct rxe_pkt_info *pkt) 600 { 601 u64 *vaddr; 602 enum resp_states ret; 603 struct rxe_mr *mr = qp->resp.mr; 604 struct resp_res *res = qp->resp.res; 605 u64 value; 606 607 if (!res) { 608 res = rxe_prepare_res(qp, pkt, RXE_ATOMIC_MASK); 609 qp->resp.res = res; 610 } 611 612 if (!res->replay) { 613 if (mr->state != RXE_MR_STATE_VALID) { 614 ret = RESPST_ERR_RKEY_VIOLATION; 615 goto out; 616 } 617 618 vaddr = iova_to_vaddr(mr, qp->resp.va + qp->resp.offset, 619 sizeof(u64)); 620 621 /* check vaddr is 8 bytes aligned. */ 622 if (!vaddr || (uintptr_t)vaddr & 7) { 623 ret = RESPST_ERR_MISALIGNED_ATOMIC; 624 goto out; 625 } 626 627 spin_lock_bh(&atomic_ops_lock); 628 res->atomic.orig_val = value = *vaddr; 629 630 if (pkt->opcode == IB_OPCODE_RC_COMPARE_SWAP) { 631 if (value == atmeth_comp(pkt)) 632 value = atmeth_swap_add(pkt); 633 } else { 634 value += atmeth_swap_add(pkt); 635 } 636 637 *vaddr = value; 638 spin_unlock_bh(&atomic_ops_lock); 639 640 qp->resp.msn++; 641 642 /* next expected psn, read handles this separately */ 643 qp->resp.psn = (pkt->psn + 1) & BTH_PSN_MASK; 644 qp->resp.ack_psn = qp->resp.psn; 645 646 qp->resp.opcode = pkt->opcode; 647 qp->resp.status = IB_WC_SUCCESS; 648 } 649 650 ret = RESPST_ACKNOWLEDGE; 651 out: 652 return ret; 653 } 654 655 static struct sk_buff *prepare_ack_packet(struct rxe_qp *qp, 656 struct rxe_pkt_info *ack, 657 int opcode, 658 int payload, 659 u32 psn, 660 u8 syndrome) 661 { 662 struct rxe_dev *rxe = to_rdev(qp->ibqp.device); 663 struct sk_buff *skb; 664 int paylen; 665 int pad; 666 int err; 667 668 /* 669 * allocate packet 670 */ 671 pad = (-payload) & 0x3; 672 paylen = rxe_opcode[opcode].length + payload + pad + RXE_ICRC_SIZE; 673 674 skb = rxe_init_packet(rxe, &qp->pri_av, paylen, ack); 675 if (!skb) 676 return NULL; 677 678 ack->qp = qp; 679 ack->opcode = opcode; 680 ack->mask = rxe_opcode[opcode].mask; 681 ack->paylen = paylen; 682 ack->psn = psn; 683 684 bth_init(ack, opcode, 0, 0, pad, IB_DEFAULT_PKEY_FULL, 685 qp->attr.dest_qp_num, 0, psn); 686 687 if (ack->mask & RXE_AETH_MASK) { 688 aeth_set_syn(ack, syndrome); 689 aeth_set_msn(ack, qp->resp.msn); 690 } 691 692 if (ack->mask & RXE_ATMACK_MASK) 693 atmack_set_orig(ack, qp->resp.res->atomic.orig_val); 694 695 err = rxe_prepare(&qp->pri_av, ack, skb); 696 if (err) { 697 kfree_skb(skb); 698 return NULL; 699 } 700 701 return skb; 702 } 703 704 /** 705 * rxe_recheck_mr - revalidate MR from rkey and get a reference 706 * @qp: the qp 707 * @rkey: the rkey 708 * 709 * This code allows the MR to be invalidated or deregistered or 710 * the MW if one was used to be invalidated or deallocated. 711 * It is assumed that the access permissions if originally good 712 * are OK and the mappings to be unchanged. 713 * 714 * TODO: If someone reregisters an MR to change its size or 715 * access permissions during the processing of an RDMA read 716 * we should kill the responder resource and complete the 717 * operation with an error. 718 * 719 * Return: mr on success else NULL 720 */ 721 static struct rxe_mr *rxe_recheck_mr(struct rxe_qp *qp, u32 rkey) 722 { 723 struct rxe_dev *rxe = to_rdev(qp->ibqp.device); 724 struct rxe_mr *mr; 725 struct rxe_mw *mw; 726 727 if (rkey_is_mw(rkey)) { 728 mw = rxe_pool_get_index(&rxe->mw_pool, rkey >> 8); 729 if (!mw) 730 return NULL; 731 732 mr = mw->mr; 733 if (mw->rkey != rkey || mw->state != RXE_MW_STATE_VALID || 734 !mr || mr->state != RXE_MR_STATE_VALID) { 735 rxe_put(mw); 736 return NULL; 737 } 738 739 rxe_get(mr); 740 rxe_put(mw); 741 742 return mr; 743 } 744 745 mr = rxe_pool_get_index(&rxe->mr_pool, rkey >> 8); 746 if (!mr) 747 return NULL; 748 749 if (mr->rkey != rkey || mr->state != RXE_MR_STATE_VALID) { 750 rxe_put(mr); 751 return NULL; 752 } 753 754 return mr; 755 } 756 757 /* RDMA read response. If res is not NULL, then we have a current RDMA request 758 * being processed or replayed. 759 */ 760 static enum resp_states read_reply(struct rxe_qp *qp, 761 struct rxe_pkt_info *req_pkt) 762 { 763 struct rxe_pkt_info ack_pkt; 764 struct sk_buff *skb; 765 int mtu = qp->mtu; 766 enum resp_states state; 767 int payload; 768 int opcode; 769 int err; 770 struct resp_res *res = qp->resp.res; 771 struct rxe_mr *mr; 772 773 if (!res) { 774 res = rxe_prepare_res(qp, req_pkt, RXE_READ_MASK); 775 qp->resp.res = res; 776 } 777 778 if (res->state == rdatm_res_state_new) { 779 if (!res->replay) { 780 mr = qp->resp.mr; 781 qp->resp.mr = NULL; 782 } else { 783 mr = rxe_recheck_mr(qp, res->read.rkey); 784 if (!mr) 785 return RESPST_ERR_RKEY_VIOLATION; 786 } 787 788 if (res->read.resid <= mtu) 789 opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY; 790 else 791 opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST; 792 } else { 793 mr = rxe_recheck_mr(qp, res->read.rkey); 794 if (!mr) 795 return RESPST_ERR_RKEY_VIOLATION; 796 797 if (res->read.resid > mtu) 798 opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE; 799 else 800 opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST; 801 } 802 803 res->state = rdatm_res_state_next; 804 805 payload = min_t(int, res->read.resid, mtu); 806 807 skb = prepare_ack_packet(qp, &ack_pkt, opcode, payload, 808 res->cur_psn, AETH_ACK_UNLIMITED); 809 if (!skb) 810 return RESPST_ERR_RNR; 811 812 rxe_mr_copy(mr, res->read.va, payload_addr(&ack_pkt), 813 payload, RXE_FROM_MR_OBJ); 814 if (mr) 815 rxe_put(mr); 816 817 if (bth_pad(&ack_pkt)) { 818 u8 *pad = payload_addr(&ack_pkt) + payload; 819 820 memset(pad, 0, bth_pad(&ack_pkt)); 821 } 822 823 err = rxe_xmit_packet(qp, &ack_pkt, skb); 824 if (err) 825 return RESPST_ERR_RNR; 826 827 res->read.va += payload; 828 res->read.resid -= payload; 829 res->cur_psn = (res->cur_psn + 1) & BTH_PSN_MASK; 830 831 if (res->read.resid > 0) { 832 state = RESPST_DONE; 833 } else { 834 qp->resp.res = NULL; 835 if (!res->replay) 836 qp->resp.opcode = -1; 837 if (psn_compare(res->cur_psn, qp->resp.psn) >= 0) 838 qp->resp.psn = res->cur_psn; 839 state = RESPST_CLEANUP; 840 } 841 842 return state; 843 } 844 845 static int invalidate_rkey(struct rxe_qp *qp, u32 rkey) 846 { 847 if (rkey_is_mw(rkey)) 848 return rxe_invalidate_mw(qp, rkey); 849 else 850 return rxe_invalidate_mr(qp, rkey); 851 } 852 853 /* Executes a new request. A retried request never reach that function (send 854 * and writes are discarded, and reads and atomics are retried elsewhere. 855 */ 856 static enum resp_states execute(struct rxe_qp *qp, struct rxe_pkt_info *pkt) 857 { 858 enum resp_states err; 859 struct sk_buff *skb = PKT_TO_SKB(pkt); 860 union rdma_network_hdr hdr; 861 862 if (pkt->mask & RXE_SEND_MASK) { 863 if (qp_type(qp) == IB_QPT_UD || 864 qp_type(qp) == IB_QPT_GSI) { 865 if (skb->protocol == htons(ETH_P_IP)) { 866 memset(&hdr.reserved, 0, 867 sizeof(hdr.reserved)); 868 memcpy(&hdr.roce4grh, ip_hdr(skb), 869 sizeof(hdr.roce4grh)); 870 err = send_data_in(qp, &hdr, sizeof(hdr)); 871 } else { 872 err = send_data_in(qp, ipv6_hdr(skb), 873 sizeof(hdr)); 874 } 875 if (err) 876 return err; 877 } 878 err = send_data_in(qp, payload_addr(pkt), payload_size(pkt)); 879 if (err) 880 return err; 881 } else if (pkt->mask & RXE_WRITE_MASK) { 882 err = write_data_in(qp, pkt); 883 if (err) 884 return err; 885 } else if (pkt->mask & RXE_READ_MASK) { 886 /* For RDMA Read we can increment the msn now. See C9-148. */ 887 qp->resp.msn++; 888 return RESPST_READ_REPLY; 889 } else if (pkt->mask & RXE_ATOMIC_MASK) { 890 return RESPST_ATOMIC_REPLY; 891 } else { 892 /* Unreachable */ 893 WARN_ON_ONCE(1); 894 } 895 896 if (pkt->mask & RXE_IETH_MASK) { 897 u32 rkey = ieth_rkey(pkt); 898 899 err = invalidate_rkey(qp, rkey); 900 if (err) 901 return RESPST_ERR_INVALIDATE_RKEY; 902 } 903 904 if (pkt->mask & RXE_END_MASK) 905 /* We successfully processed this new request. */ 906 qp->resp.msn++; 907 908 /* next expected psn, read handles this separately */ 909 qp->resp.psn = (pkt->psn + 1) & BTH_PSN_MASK; 910 qp->resp.ack_psn = qp->resp.psn; 911 912 qp->resp.opcode = pkt->opcode; 913 qp->resp.status = IB_WC_SUCCESS; 914 915 if (pkt->mask & RXE_COMP_MASK) 916 return RESPST_COMPLETE; 917 else if (qp_type(qp) == IB_QPT_RC) 918 return RESPST_ACKNOWLEDGE; 919 else 920 return RESPST_CLEANUP; 921 } 922 923 static enum resp_states do_complete(struct rxe_qp *qp, 924 struct rxe_pkt_info *pkt) 925 { 926 struct rxe_cqe cqe; 927 struct ib_wc *wc = &cqe.ibwc; 928 struct ib_uverbs_wc *uwc = &cqe.uibwc; 929 struct rxe_recv_wqe *wqe = qp->resp.wqe; 930 struct rxe_dev *rxe = to_rdev(qp->ibqp.device); 931 932 if (!wqe) 933 goto finish; 934 935 memset(&cqe, 0, sizeof(cqe)); 936 937 if (qp->rcq->is_user) { 938 uwc->status = qp->resp.status; 939 uwc->qp_num = qp->ibqp.qp_num; 940 uwc->wr_id = wqe->wr_id; 941 } else { 942 wc->status = qp->resp.status; 943 wc->qp = &qp->ibqp; 944 wc->wr_id = wqe->wr_id; 945 } 946 947 if (wc->status == IB_WC_SUCCESS) { 948 rxe_counter_inc(rxe, RXE_CNT_RDMA_RECV); 949 wc->opcode = (pkt->mask & RXE_IMMDT_MASK && 950 pkt->mask & RXE_WRITE_MASK) ? 951 IB_WC_RECV_RDMA_WITH_IMM : IB_WC_RECV; 952 wc->byte_len = (pkt->mask & RXE_IMMDT_MASK && 953 pkt->mask & RXE_WRITE_MASK) ? 954 qp->resp.length : wqe->dma.length - wqe->dma.resid; 955 956 /* fields after byte_len are different between kernel and user 957 * space 958 */ 959 if (qp->rcq->is_user) { 960 uwc->wc_flags = IB_WC_GRH; 961 962 if (pkt->mask & RXE_IMMDT_MASK) { 963 uwc->wc_flags |= IB_WC_WITH_IMM; 964 uwc->ex.imm_data = immdt_imm(pkt); 965 } 966 967 if (pkt->mask & RXE_IETH_MASK) { 968 uwc->wc_flags |= IB_WC_WITH_INVALIDATE; 969 uwc->ex.invalidate_rkey = ieth_rkey(pkt); 970 } 971 972 if (pkt->mask & RXE_DETH_MASK) 973 uwc->src_qp = deth_sqp(pkt); 974 975 uwc->port_num = qp->attr.port_num; 976 } else { 977 struct sk_buff *skb = PKT_TO_SKB(pkt); 978 979 wc->wc_flags = IB_WC_GRH | IB_WC_WITH_NETWORK_HDR_TYPE; 980 if (skb->protocol == htons(ETH_P_IP)) 981 wc->network_hdr_type = RDMA_NETWORK_IPV4; 982 else 983 wc->network_hdr_type = RDMA_NETWORK_IPV6; 984 985 if (is_vlan_dev(skb->dev)) { 986 wc->wc_flags |= IB_WC_WITH_VLAN; 987 wc->vlan_id = vlan_dev_vlan_id(skb->dev); 988 } 989 990 if (pkt->mask & RXE_IMMDT_MASK) { 991 wc->wc_flags |= IB_WC_WITH_IMM; 992 wc->ex.imm_data = immdt_imm(pkt); 993 } 994 995 if (pkt->mask & RXE_IETH_MASK) { 996 wc->wc_flags |= IB_WC_WITH_INVALIDATE; 997 wc->ex.invalidate_rkey = ieth_rkey(pkt); 998 } 999 1000 if (pkt->mask & RXE_DETH_MASK) 1001 wc->src_qp = deth_sqp(pkt); 1002 1003 wc->port_num = qp->attr.port_num; 1004 } 1005 } 1006 1007 /* have copy for srq and reference for !srq */ 1008 if (!qp->srq) 1009 queue_advance_consumer(qp->rq.queue, QUEUE_TYPE_FROM_CLIENT); 1010 1011 qp->resp.wqe = NULL; 1012 1013 if (rxe_cq_post(qp->rcq, &cqe, pkt ? bth_se(pkt) : 1)) 1014 return RESPST_ERR_CQ_OVERFLOW; 1015 1016 finish: 1017 if (unlikely(qp->resp.state == QP_STATE_ERROR)) 1018 return RESPST_CHK_RESOURCE; 1019 if (unlikely(!pkt)) 1020 return RESPST_DONE; 1021 if (qp_type(qp) == IB_QPT_RC) 1022 return RESPST_ACKNOWLEDGE; 1023 else 1024 return RESPST_CLEANUP; 1025 } 1026 1027 1028 static int send_common_ack(struct rxe_qp *qp, u8 syndrome, u32 psn, 1029 int opcode, const char *msg) 1030 { 1031 int err; 1032 struct rxe_pkt_info ack_pkt; 1033 struct sk_buff *skb; 1034 1035 skb = prepare_ack_packet(qp, &ack_pkt, opcode, 0, psn, syndrome); 1036 if (!skb) 1037 return -ENOMEM; 1038 1039 err = rxe_xmit_packet(qp, &ack_pkt, skb); 1040 if (err) 1041 pr_err_ratelimited("Failed sending %s\n", msg); 1042 1043 return err; 1044 } 1045 1046 static int send_ack(struct rxe_qp *qp, u8 syndrome, u32 psn) 1047 { 1048 return send_common_ack(qp, syndrome, psn, 1049 IB_OPCODE_RC_ACKNOWLEDGE, "ACK"); 1050 } 1051 1052 static int send_atomic_ack(struct rxe_qp *qp, u8 syndrome, u32 psn) 1053 { 1054 int ret = send_common_ack(qp, syndrome, psn, 1055 IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE, "ATOMIC ACK"); 1056 1057 /* have to clear this since it is used to trigger 1058 * long read replies 1059 */ 1060 qp->resp.res = NULL; 1061 return ret; 1062 } 1063 1064 static enum resp_states acknowledge(struct rxe_qp *qp, 1065 struct rxe_pkt_info *pkt) 1066 { 1067 if (qp_type(qp) != IB_QPT_RC) 1068 return RESPST_CLEANUP; 1069 1070 if (qp->resp.aeth_syndrome != AETH_ACK_UNLIMITED) 1071 send_ack(qp, qp->resp.aeth_syndrome, pkt->psn); 1072 else if (pkt->mask & RXE_ATOMIC_MASK) 1073 send_atomic_ack(qp, AETH_ACK_UNLIMITED, pkt->psn); 1074 else if (bth_ack(pkt)) 1075 send_ack(qp, AETH_ACK_UNLIMITED, pkt->psn); 1076 1077 return RESPST_CLEANUP; 1078 } 1079 1080 static enum resp_states cleanup(struct rxe_qp *qp, 1081 struct rxe_pkt_info *pkt) 1082 { 1083 struct sk_buff *skb; 1084 1085 if (pkt) { 1086 skb = skb_dequeue(&qp->req_pkts); 1087 rxe_put(qp); 1088 kfree_skb(skb); 1089 ib_device_put(qp->ibqp.device); 1090 } 1091 1092 if (qp->resp.mr) { 1093 rxe_put(qp->resp.mr); 1094 qp->resp.mr = NULL; 1095 } 1096 1097 return RESPST_DONE; 1098 } 1099 1100 static struct resp_res *find_resource(struct rxe_qp *qp, u32 psn) 1101 { 1102 int i; 1103 1104 for (i = 0; i < qp->attr.max_dest_rd_atomic; i++) { 1105 struct resp_res *res = &qp->resp.resources[i]; 1106 1107 if (res->type == 0) 1108 continue; 1109 1110 if (psn_compare(psn, res->first_psn) >= 0 && 1111 psn_compare(psn, res->last_psn) <= 0) { 1112 return res; 1113 } 1114 } 1115 1116 return NULL; 1117 } 1118 1119 static enum resp_states duplicate_request(struct rxe_qp *qp, 1120 struct rxe_pkt_info *pkt) 1121 { 1122 enum resp_states rc; 1123 u32 prev_psn = (qp->resp.ack_psn - 1) & BTH_PSN_MASK; 1124 1125 if (pkt->mask & RXE_SEND_MASK || 1126 pkt->mask & RXE_WRITE_MASK) { 1127 /* SEND. Ack again and cleanup. C9-105. */ 1128 send_ack(qp, AETH_ACK_UNLIMITED, prev_psn); 1129 return RESPST_CLEANUP; 1130 } else if (pkt->mask & RXE_READ_MASK) { 1131 struct resp_res *res; 1132 1133 res = find_resource(qp, pkt->psn); 1134 if (!res) { 1135 /* Resource not found. Class D error. Drop the 1136 * request. 1137 */ 1138 rc = RESPST_CLEANUP; 1139 goto out; 1140 } else { 1141 /* Ensure this new request is the same as the previous 1142 * one or a subset of it. 1143 */ 1144 u64 iova = reth_va(pkt); 1145 u32 resid = reth_len(pkt); 1146 1147 if (iova < res->read.va_org || 1148 resid > res->read.length || 1149 (iova + resid) > (res->read.va_org + 1150 res->read.length)) { 1151 rc = RESPST_CLEANUP; 1152 goto out; 1153 } 1154 1155 if (reth_rkey(pkt) != res->read.rkey) { 1156 rc = RESPST_CLEANUP; 1157 goto out; 1158 } 1159 1160 res->cur_psn = pkt->psn; 1161 res->state = (pkt->psn == res->first_psn) ? 1162 rdatm_res_state_new : 1163 rdatm_res_state_replay; 1164 res->replay = 1; 1165 1166 /* Reset the resource, except length. */ 1167 res->read.va_org = iova; 1168 res->read.va = iova; 1169 res->read.resid = resid; 1170 1171 /* Replay the RDMA read reply. */ 1172 qp->resp.res = res; 1173 rc = RESPST_READ_REPLY; 1174 goto out; 1175 } 1176 } else { 1177 struct resp_res *res; 1178 1179 /* Find the operation in our list of responder resources. */ 1180 res = find_resource(qp, pkt->psn); 1181 if (res) { 1182 res->replay = 1; 1183 res->cur_psn = pkt->psn; 1184 qp->resp.res = res; 1185 rc = RESPST_ATOMIC_REPLY; 1186 goto out; 1187 } 1188 1189 /* Resource not found. Class D error. Drop the request. */ 1190 rc = RESPST_CLEANUP; 1191 goto out; 1192 } 1193 out: 1194 return rc; 1195 } 1196 1197 /* Process a class A or C. Both are treated the same in this implementation. */ 1198 static void do_class_ac_error(struct rxe_qp *qp, u8 syndrome, 1199 enum ib_wc_status status) 1200 { 1201 qp->resp.aeth_syndrome = syndrome; 1202 qp->resp.status = status; 1203 1204 /* indicate that we should go through the ERROR state */ 1205 qp->resp.goto_error = 1; 1206 } 1207 1208 static enum resp_states do_class_d1e_error(struct rxe_qp *qp) 1209 { 1210 /* UC */ 1211 if (qp->srq) { 1212 /* Class E */ 1213 qp->resp.drop_msg = 1; 1214 if (qp->resp.wqe) { 1215 qp->resp.status = IB_WC_REM_INV_REQ_ERR; 1216 return RESPST_COMPLETE; 1217 } else { 1218 return RESPST_CLEANUP; 1219 } 1220 } else { 1221 /* Class D1. This packet may be the start of a 1222 * new message and could be valid. The previous 1223 * message is invalid and ignored. reset the 1224 * recv wr to its original state 1225 */ 1226 if (qp->resp.wqe) { 1227 qp->resp.wqe->dma.resid = qp->resp.wqe->dma.length; 1228 qp->resp.wqe->dma.cur_sge = 0; 1229 qp->resp.wqe->dma.sge_offset = 0; 1230 qp->resp.opcode = -1; 1231 } 1232 1233 if (qp->resp.mr) { 1234 rxe_put(qp->resp.mr); 1235 qp->resp.mr = NULL; 1236 } 1237 1238 return RESPST_CLEANUP; 1239 } 1240 } 1241 1242 static void rxe_drain_req_pkts(struct rxe_qp *qp, bool notify) 1243 { 1244 struct sk_buff *skb; 1245 struct rxe_queue *q = qp->rq.queue; 1246 1247 while ((skb = skb_dequeue(&qp->req_pkts))) { 1248 rxe_put(qp); 1249 kfree_skb(skb); 1250 ib_device_put(qp->ibqp.device); 1251 } 1252 1253 if (notify) 1254 return; 1255 1256 while (!qp->srq && q && queue_head(q, q->type)) 1257 queue_advance_consumer(q, q->type); 1258 } 1259 1260 int rxe_responder(void *arg) 1261 { 1262 struct rxe_qp *qp = (struct rxe_qp *)arg; 1263 struct rxe_dev *rxe = to_rdev(qp->ibqp.device); 1264 enum resp_states state; 1265 struct rxe_pkt_info *pkt = NULL; 1266 int ret; 1267 1268 if (!rxe_get(qp)) 1269 return -EAGAIN; 1270 1271 qp->resp.aeth_syndrome = AETH_ACK_UNLIMITED; 1272 1273 if (!qp->valid) 1274 goto exit; 1275 1276 switch (qp->resp.state) { 1277 case QP_STATE_RESET: 1278 state = RESPST_RESET; 1279 break; 1280 1281 default: 1282 state = RESPST_GET_REQ; 1283 break; 1284 } 1285 1286 while (1) { 1287 pr_debug("qp#%d state = %s\n", qp_num(qp), 1288 resp_state_name[state]); 1289 switch (state) { 1290 case RESPST_GET_REQ: 1291 state = get_req(qp, &pkt); 1292 break; 1293 case RESPST_CHK_PSN: 1294 state = check_psn(qp, pkt); 1295 break; 1296 case RESPST_CHK_OP_SEQ: 1297 state = check_op_seq(qp, pkt); 1298 break; 1299 case RESPST_CHK_OP_VALID: 1300 state = check_op_valid(qp, pkt); 1301 break; 1302 case RESPST_CHK_RESOURCE: 1303 state = check_resource(qp, pkt); 1304 break; 1305 case RESPST_CHK_LENGTH: 1306 state = check_length(qp, pkt); 1307 break; 1308 case RESPST_CHK_RKEY: 1309 state = check_rkey(qp, pkt); 1310 break; 1311 case RESPST_EXECUTE: 1312 state = execute(qp, pkt); 1313 break; 1314 case RESPST_COMPLETE: 1315 state = do_complete(qp, pkt); 1316 break; 1317 case RESPST_READ_REPLY: 1318 state = read_reply(qp, pkt); 1319 break; 1320 case RESPST_ATOMIC_REPLY: 1321 state = atomic_reply(qp, pkt); 1322 break; 1323 case RESPST_ACKNOWLEDGE: 1324 state = acknowledge(qp, pkt); 1325 break; 1326 case RESPST_CLEANUP: 1327 state = cleanup(qp, pkt); 1328 break; 1329 case RESPST_DUPLICATE_REQUEST: 1330 state = duplicate_request(qp, pkt); 1331 break; 1332 case RESPST_ERR_PSN_OUT_OF_SEQ: 1333 /* RC only - Class B. Drop packet. */ 1334 send_ack(qp, AETH_NAK_PSN_SEQ_ERROR, qp->resp.psn); 1335 state = RESPST_CLEANUP; 1336 break; 1337 1338 case RESPST_ERR_TOO_MANY_RDMA_ATM_REQ: 1339 case RESPST_ERR_MISSING_OPCODE_FIRST: 1340 case RESPST_ERR_MISSING_OPCODE_LAST_C: 1341 case RESPST_ERR_UNSUPPORTED_OPCODE: 1342 case RESPST_ERR_MISALIGNED_ATOMIC: 1343 /* RC Only - Class C. */ 1344 do_class_ac_error(qp, AETH_NAK_INVALID_REQ, 1345 IB_WC_REM_INV_REQ_ERR); 1346 state = RESPST_COMPLETE; 1347 break; 1348 1349 case RESPST_ERR_MISSING_OPCODE_LAST_D1E: 1350 state = do_class_d1e_error(qp); 1351 break; 1352 case RESPST_ERR_RNR: 1353 if (qp_type(qp) == IB_QPT_RC) { 1354 rxe_counter_inc(rxe, RXE_CNT_SND_RNR); 1355 /* RC - class B */ 1356 send_ack(qp, AETH_RNR_NAK | 1357 (~AETH_TYPE_MASK & 1358 qp->attr.min_rnr_timer), 1359 pkt->psn); 1360 } else { 1361 /* UD/UC - class D */ 1362 qp->resp.drop_msg = 1; 1363 } 1364 state = RESPST_CLEANUP; 1365 break; 1366 1367 case RESPST_ERR_RKEY_VIOLATION: 1368 if (qp_type(qp) == IB_QPT_RC) { 1369 /* Class C */ 1370 do_class_ac_error(qp, AETH_NAK_REM_ACC_ERR, 1371 IB_WC_REM_ACCESS_ERR); 1372 state = RESPST_COMPLETE; 1373 } else { 1374 qp->resp.drop_msg = 1; 1375 if (qp->srq) { 1376 /* UC/SRQ Class D */ 1377 qp->resp.status = IB_WC_REM_ACCESS_ERR; 1378 state = RESPST_COMPLETE; 1379 } else { 1380 /* UC/non-SRQ Class E. */ 1381 state = RESPST_CLEANUP; 1382 } 1383 } 1384 break; 1385 1386 case RESPST_ERR_INVALIDATE_RKEY: 1387 /* RC - Class J. */ 1388 qp->resp.goto_error = 1; 1389 qp->resp.status = IB_WC_REM_INV_REQ_ERR; 1390 state = RESPST_COMPLETE; 1391 break; 1392 1393 case RESPST_ERR_LENGTH: 1394 if (qp_type(qp) == IB_QPT_RC) { 1395 /* Class C */ 1396 do_class_ac_error(qp, AETH_NAK_INVALID_REQ, 1397 IB_WC_REM_INV_REQ_ERR); 1398 state = RESPST_COMPLETE; 1399 } else if (qp->srq) { 1400 /* UC/UD - class E */ 1401 qp->resp.status = IB_WC_REM_INV_REQ_ERR; 1402 state = RESPST_COMPLETE; 1403 } else { 1404 /* UC/UD - class D */ 1405 qp->resp.drop_msg = 1; 1406 state = RESPST_CLEANUP; 1407 } 1408 break; 1409 1410 case RESPST_ERR_MALFORMED_WQE: 1411 /* All, Class A. */ 1412 do_class_ac_error(qp, AETH_NAK_REM_OP_ERR, 1413 IB_WC_LOC_QP_OP_ERR); 1414 state = RESPST_COMPLETE; 1415 break; 1416 1417 case RESPST_ERR_CQ_OVERFLOW: 1418 /* All - Class G */ 1419 state = RESPST_ERROR; 1420 break; 1421 1422 case RESPST_DONE: 1423 if (qp->resp.goto_error) { 1424 state = RESPST_ERROR; 1425 break; 1426 } 1427 1428 goto done; 1429 1430 case RESPST_EXIT: 1431 if (qp->resp.goto_error) { 1432 state = RESPST_ERROR; 1433 break; 1434 } 1435 1436 goto exit; 1437 1438 case RESPST_RESET: 1439 rxe_drain_req_pkts(qp, false); 1440 qp->resp.wqe = NULL; 1441 goto exit; 1442 1443 case RESPST_ERROR: 1444 qp->resp.goto_error = 0; 1445 pr_debug("qp#%d moved to error state\n", qp_num(qp)); 1446 rxe_qp_error(qp); 1447 goto exit; 1448 1449 default: 1450 WARN_ON_ONCE(1); 1451 } 1452 } 1453 1454 /* A non-zero return value will cause rxe_do_task to 1455 * exit its loop and end the tasklet. A zero return 1456 * will continue looping and return to rxe_responder 1457 */ 1458 done: 1459 ret = 0; 1460 goto out; 1461 exit: 1462 ret = -EAGAIN; 1463 out: 1464 rxe_put(qp); 1465 return ret; 1466 } 1467