1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
3 * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved.
4 * Copyright (c) 2015 System Fabric Works, Inc. All rights reserved.
7 #include <linux/skbuff.h>
11 #include "rxe_queue.h"
13 static char *resp_state_name[] = {
14 [RESPST_NONE] = "NONE",
15 [RESPST_GET_REQ] = "GET_REQ",
16 [RESPST_CHK_PSN] = "CHK_PSN",
17 [RESPST_CHK_OP_SEQ] = "CHK_OP_SEQ",
18 [RESPST_CHK_OP_VALID] = "CHK_OP_VALID",
19 [RESPST_CHK_RESOURCE] = "CHK_RESOURCE",
20 [RESPST_CHK_LENGTH] = "CHK_LENGTH",
21 [RESPST_CHK_RKEY] = "CHK_RKEY",
22 [RESPST_EXECUTE] = "EXECUTE",
23 [RESPST_READ_REPLY] = "READ_REPLY",
24 [RESPST_ATOMIC_REPLY] = "ATOMIC_REPLY",
25 [RESPST_ATOMIC_WRITE_REPLY] = "ATOMIC_WRITE_REPLY",
26 [RESPST_PROCESS_FLUSH] = "PROCESS_FLUSH",
27 [RESPST_COMPLETE] = "COMPLETE",
28 [RESPST_ACKNOWLEDGE] = "ACKNOWLEDGE",
29 [RESPST_CLEANUP] = "CLEANUP",
30 [RESPST_DUPLICATE_REQUEST] = "DUPLICATE_REQUEST",
31 [RESPST_ERR_MALFORMED_WQE] = "ERR_MALFORMED_WQE",
32 [RESPST_ERR_UNSUPPORTED_OPCODE] = "ERR_UNSUPPORTED_OPCODE",
33 [RESPST_ERR_MISALIGNED_ATOMIC] = "ERR_MISALIGNED_ATOMIC",
34 [RESPST_ERR_PSN_OUT_OF_SEQ] = "ERR_PSN_OUT_OF_SEQ",
35 [RESPST_ERR_MISSING_OPCODE_FIRST] = "ERR_MISSING_OPCODE_FIRST",
36 [RESPST_ERR_MISSING_OPCODE_LAST_C] = "ERR_MISSING_OPCODE_LAST_C",
37 [RESPST_ERR_MISSING_OPCODE_LAST_D1E] = "ERR_MISSING_OPCODE_LAST_D1E",
38 [RESPST_ERR_TOO_MANY_RDMA_ATM_REQ] = "ERR_TOO_MANY_RDMA_ATM_REQ",
39 [RESPST_ERR_RNR] = "ERR_RNR",
40 [RESPST_ERR_RKEY_VIOLATION] = "ERR_RKEY_VIOLATION",
41 [RESPST_ERR_INVALIDATE_RKEY] = "ERR_INVALIDATE_RKEY_VIOLATION",
42 [RESPST_ERR_LENGTH] = "ERR_LENGTH",
43 [RESPST_ERR_CQ_OVERFLOW] = "ERR_CQ_OVERFLOW",
44 [RESPST_ERROR] = "ERROR",
45 [RESPST_DONE] = "DONE",
46 [RESPST_EXIT] = "EXIT",
49 /* rxe_recv calls here to add a request packet to the input queue */
50 void rxe_resp_queue_pkt(struct rxe_qp *qp, struct sk_buff *skb)
53 struct rxe_pkt_info *pkt = SKB_TO_PKT(skb);
55 skb_queue_tail(&qp->req_pkts, skb);
57 must_sched = (pkt->opcode == IB_OPCODE_RC_RDMA_READ_REQUEST) ||
58 (skb_queue_len(&qp->req_pkts) > 1);
61 rxe_sched_task(&qp->resp.task);
63 rxe_run_task(&qp->resp.task);
66 static inline enum resp_states get_req(struct rxe_qp *qp,
67 struct rxe_pkt_info **pkt_p)
71 skb = skb_peek(&qp->req_pkts);
75 *pkt_p = SKB_TO_PKT(skb);
77 return (qp->resp.res) ? RESPST_READ_REPLY : RESPST_CHK_PSN;
80 static enum resp_states check_psn(struct rxe_qp *qp,
81 struct rxe_pkt_info *pkt)
83 int diff = psn_compare(pkt->psn, qp->resp.psn);
84 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
86 switch (qp_type(qp)) {
89 if (qp->resp.sent_psn_nak)
90 return RESPST_CLEANUP;
92 qp->resp.sent_psn_nak = 1;
93 rxe_counter_inc(rxe, RXE_CNT_OUT_OF_SEQ_REQ);
94 return RESPST_ERR_PSN_OUT_OF_SEQ;
96 } else if (diff < 0) {
97 rxe_counter_inc(rxe, RXE_CNT_DUP_REQ);
98 return RESPST_DUPLICATE_REQUEST;
101 if (qp->resp.sent_psn_nak)
102 qp->resp.sent_psn_nak = 0;
107 if (qp->resp.drop_msg || diff != 0) {
108 if (pkt->mask & RXE_START_MASK) {
109 qp->resp.drop_msg = 0;
110 return RESPST_CHK_OP_SEQ;
113 qp->resp.drop_msg = 1;
114 return RESPST_CLEANUP;
121 return RESPST_CHK_OP_SEQ;
124 static enum resp_states check_op_seq(struct rxe_qp *qp,
125 struct rxe_pkt_info *pkt)
127 switch (qp_type(qp)) {
129 switch (qp->resp.opcode) {
130 case IB_OPCODE_RC_SEND_FIRST:
131 case IB_OPCODE_RC_SEND_MIDDLE:
132 switch (pkt->opcode) {
133 case IB_OPCODE_RC_SEND_MIDDLE:
134 case IB_OPCODE_RC_SEND_LAST:
135 case IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE:
136 case IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE:
137 return RESPST_CHK_OP_VALID;
139 return RESPST_ERR_MISSING_OPCODE_LAST_C;
142 case IB_OPCODE_RC_RDMA_WRITE_FIRST:
143 case IB_OPCODE_RC_RDMA_WRITE_MIDDLE:
144 switch (pkt->opcode) {
145 case IB_OPCODE_RC_RDMA_WRITE_MIDDLE:
146 case IB_OPCODE_RC_RDMA_WRITE_LAST:
147 case IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
148 return RESPST_CHK_OP_VALID;
150 return RESPST_ERR_MISSING_OPCODE_LAST_C;
154 switch (pkt->opcode) {
155 case IB_OPCODE_RC_SEND_MIDDLE:
156 case IB_OPCODE_RC_SEND_LAST:
157 case IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE:
158 case IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE:
159 case IB_OPCODE_RC_RDMA_WRITE_MIDDLE:
160 case IB_OPCODE_RC_RDMA_WRITE_LAST:
161 case IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
162 return RESPST_ERR_MISSING_OPCODE_FIRST;
164 return RESPST_CHK_OP_VALID;
170 switch (qp->resp.opcode) {
171 case IB_OPCODE_UC_SEND_FIRST:
172 case IB_OPCODE_UC_SEND_MIDDLE:
173 switch (pkt->opcode) {
174 case IB_OPCODE_UC_SEND_MIDDLE:
175 case IB_OPCODE_UC_SEND_LAST:
176 case IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE:
177 return RESPST_CHK_OP_VALID;
179 return RESPST_ERR_MISSING_OPCODE_LAST_D1E;
182 case IB_OPCODE_UC_RDMA_WRITE_FIRST:
183 case IB_OPCODE_UC_RDMA_WRITE_MIDDLE:
184 switch (pkt->opcode) {
185 case IB_OPCODE_UC_RDMA_WRITE_MIDDLE:
186 case IB_OPCODE_UC_RDMA_WRITE_LAST:
187 case IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
188 return RESPST_CHK_OP_VALID;
190 return RESPST_ERR_MISSING_OPCODE_LAST_D1E;
194 switch (pkt->opcode) {
195 case IB_OPCODE_UC_SEND_MIDDLE:
196 case IB_OPCODE_UC_SEND_LAST:
197 case IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE:
198 case IB_OPCODE_UC_RDMA_WRITE_MIDDLE:
199 case IB_OPCODE_UC_RDMA_WRITE_LAST:
200 case IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
201 qp->resp.drop_msg = 1;
202 return RESPST_CLEANUP;
204 return RESPST_CHK_OP_VALID;
210 return RESPST_CHK_OP_VALID;
214 static bool check_qp_attr_access(struct rxe_qp *qp,
215 struct rxe_pkt_info *pkt)
217 if (((pkt->mask & RXE_READ_MASK) &&
218 !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_READ)) ||
219 ((pkt->mask & (RXE_WRITE_MASK | RXE_ATOMIC_WRITE_MASK)) &&
220 !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_WRITE)) ||
221 ((pkt->mask & RXE_ATOMIC_MASK) &&
222 !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
225 if (pkt->mask & RXE_FLUSH_MASK) {
226 u32 flush_type = feth_plt(pkt);
228 if ((flush_type & IB_FLUSH_GLOBAL &&
229 !(qp->attr.qp_access_flags & IB_ACCESS_FLUSH_GLOBAL)) ||
230 (flush_type & IB_FLUSH_PERSISTENT &&
231 !(qp->attr.qp_access_flags & IB_ACCESS_FLUSH_PERSISTENT)))
238 static enum resp_states check_op_valid(struct rxe_qp *qp,
239 struct rxe_pkt_info *pkt)
241 switch (qp_type(qp)) {
243 if (!check_qp_attr_access(qp, pkt))
244 return RESPST_ERR_UNSUPPORTED_OPCODE;
249 if ((pkt->mask & RXE_WRITE_MASK) &&
250 !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_WRITE)) {
251 qp->resp.drop_msg = 1;
252 return RESPST_CLEANUP;
266 return RESPST_CHK_RESOURCE;
269 static enum resp_states get_srq_wqe(struct rxe_qp *qp)
271 struct rxe_srq *srq = qp->srq;
272 struct rxe_queue *q = srq->rq.queue;
273 struct rxe_recv_wqe *wqe;
280 return RESPST_ERR_RNR;
282 spin_lock_irqsave(&srq->rq.consumer_lock, flags);
284 wqe = queue_head(q, QUEUE_TYPE_FROM_CLIENT);
286 spin_unlock_irqrestore(&srq->rq.consumer_lock, flags);
287 return RESPST_ERR_RNR;
290 /* don't trust user space data */
291 if (unlikely(wqe->dma.num_sge > srq->rq.max_sge)) {
292 spin_unlock_irqrestore(&srq->rq.consumer_lock, flags);
293 rxe_dbg_qp(qp, "invalid num_sge in SRQ entry\n");
294 return RESPST_ERR_MALFORMED_WQE;
296 size = sizeof(*wqe) + wqe->dma.num_sge*sizeof(struct rxe_sge);
297 memcpy(&qp->resp.srq_wqe, wqe, size);
299 qp->resp.wqe = &qp->resp.srq_wqe.wqe;
300 queue_advance_consumer(q, QUEUE_TYPE_FROM_CLIENT);
301 count = queue_count(q, QUEUE_TYPE_FROM_CLIENT);
303 if (srq->limit && srq->ibsrq.event_handler && (count < srq->limit)) {
308 spin_unlock_irqrestore(&srq->rq.consumer_lock, flags);
309 return RESPST_CHK_LENGTH;
312 spin_unlock_irqrestore(&srq->rq.consumer_lock, flags);
313 ev.device = qp->ibqp.device;
314 ev.element.srq = qp->ibqp.srq;
315 ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
316 srq->ibsrq.event_handler(&ev, srq->ibsrq.srq_context);
317 return RESPST_CHK_LENGTH;
320 static enum resp_states check_resource(struct rxe_qp *qp,
321 struct rxe_pkt_info *pkt)
323 struct rxe_srq *srq = qp->srq;
325 if (pkt->mask & (RXE_READ_OR_ATOMIC_MASK | RXE_ATOMIC_WRITE_MASK)) {
326 /* it is the requesters job to not send
327 * too many read/atomic ops, we just
328 * recycle the responder resource queue
330 if (likely(qp->attr.max_dest_rd_atomic > 0))
331 return RESPST_CHK_LENGTH;
333 return RESPST_ERR_TOO_MANY_RDMA_ATM_REQ;
336 if (pkt->mask & RXE_RWR_MASK) {
338 return get_srq_wqe(qp);
340 qp->resp.wqe = queue_head(qp->rq.queue,
341 QUEUE_TYPE_FROM_CLIENT);
342 return (qp->resp.wqe) ? RESPST_CHK_LENGTH : RESPST_ERR_RNR;
345 return RESPST_CHK_LENGTH;
348 static enum resp_states rxe_resp_check_length(struct rxe_qp *qp,
349 struct rxe_pkt_info *pkt)
353 * For UD QPs we only check if the packet will fit in the
354 * receive buffer later. For rmda operations additional
355 * length checks are performed in check_rkey.
357 if (pkt->mask & RXE_PAYLOAD_MASK && ((qp_type(qp) == IB_QPT_RC) ||
358 (qp_type(qp) == IB_QPT_UC))) {
359 unsigned int mtu = qp->mtu;
360 unsigned int payload = payload_size(pkt);
362 if ((pkt->mask & RXE_START_MASK) &&
363 (pkt->mask & RXE_END_MASK)) {
364 if (unlikely(payload > mtu)) {
365 rxe_dbg_qp(qp, "only packet too long");
366 return RESPST_ERR_LENGTH;
368 } else if ((pkt->mask & RXE_START_MASK) ||
369 (pkt->mask & RXE_MIDDLE_MASK)) {
370 if (unlikely(payload != mtu)) {
371 rxe_dbg_qp(qp, "first or middle packet not mtu");
372 return RESPST_ERR_LENGTH;
374 } else if (pkt->mask & RXE_END_MASK) {
375 if (unlikely((payload == 0) || (payload > mtu))) {
376 rxe_dbg_qp(qp, "last packet zero or too long");
377 return RESPST_ERR_LENGTH;
383 if (pkt->mask & RXE_RETH_MASK) {
384 if (reth_len(pkt) > (1U << 31)) {
385 rxe_dbg_qp(qp, "dma length too long");
386 return RESPST_ERR_LENGTH;
390 if (pkt->mask & RXE_RDMA_OP_MASK)
391 return RESPST_CHK_RKEY;
393 return RESPST_EXECUTE;
396 /* if the reth length field is zero we can assume nothing
397 * about the rkey value and should not validate or use it.
398 * Instead set qp->resp.rkey to 0 which is an invalid rkey
399 * value since the minimum index part is 1.
401 static void qp_resp_from_reth(struct rxe_qp *qp, struct rxe_pkt_info *pkt)
403 unsigned int length = reth_len(pkt);
405 qp->resp.va = reth_va(pkt);
407 qp->resp.resid = length;
408 qp->resp.length = length;
409 if (pkt->mask & RXE_READ_OR_WRITE_MASK && length == 0)
412 qp->resp.rkey = reth_rkey(pkt);
415 static void qp_resp_from_atmeth(struct rxe_qp *qp, struct rxe_pkt_info *pkt)
417 qp->resp.va = atmeth_va(pkt);
419 qp->resp.rkey = atmeth_rkey(pkt);
420 qp->resp.resid = sizeof(u64);
423 /* resolve the packet rkey to qp->resp.mr or set qp->resp.mr to NULL
424 * if an invalid rkey is received or the rdma length is zero. For middle
425 * or last packets use the stored value of mr.
427 static enum resp_states check_rkey(struct rxe_qp *qp,
428 struct rxe_pkt_info *pkt)
430 struct rxe_mr *mr = NULL;
431 struct rxe_mw *mw = NULL;
437 enum resp_states state;
440 /* parse RETH or ATMETH header for first/only packets
441 * for va, length, rkey, etc. or use current value for
442 * middle/last packets.
444 if (pkt->mask & (RXE_READ_OR_WRITE_MASK | RXE_ATOMIC_WRITE_MASK)) {
445 if (pkt->mask & RXE_RETH_MASK)
446 qp_resp_from_reth(qp, pkt);
448 access = (pkt->mask & RXE_READ_MASK) ? IB_ACCESS_REMOTE_READ
449 : IB_ACCESS_REMOTE_WRITE;
450 } else if (pkt->mask & RXE_FLUSH_MASK) {
451 u32 flush_type = feth_plt(pkt);
453 if (pkt->mask & RXE_RETH_MASK)
454 qp_resp_from_reth(qp, pkt);
456 if (flush_type & IB_FLUSH_GLOBAL)
457 access |= IB_ACCESS_FLUSH_GLOBAL;
458 if (flush_type & IB_FLUSH_PERSISTENT)
459 access |= IB_ACCESS_FLUSH_PERSISTENT;
460 } else if (pkt->mask & RXE_ATOMIC_MASK) {
461 qp_resp_from_atmeth(qp, pkt);
462 access = IB_ACCESS_REMOTE_ATOMIC;
464 /* shouldn't happen */
468 /* A zero-byte read or write op is not required to
469 * set an addr or rkey. See C9-88
471 if ((pkt->mask & RXE_READ_OR_WRITE_MASK) &&
472 (pkt->mask & RXE_RETH_MASK) && reth_len(pkt) == 0) {
474 return RESPST_EXECUTE;
478 rkey = qp->resp.rkey;
479 resid = qp->resp.resid;
480 pktlen = payload_size(pkt);
482 if (rkey_is_mw(rkey)) {
483 mw = rxe_lookup_mw(qp, access, rkey);
485 rxe_dbg_qp(qp, "no MW matches rkey %#x\n", rkey);
486 state = RESPST_ERR_RKEY_VIOLATION;
492 rxe_dbg_qp(qp, "MW doesn't have an MR\n");
493 state = RESPST_ERR_RKEY_VIOLATION;
497 if (mw->access & IB_ZERO_BASED)
498 qp->resp.offset = mw->addr;
504 mr = lookup_mr(qp->pd, access, rkey, RXE_LOOKUP_REMOTE);
506 rxe_dbg_qp(qp, "no MR matches rkey %#x\n", rkey);
507 state = RESPST_ERR_RKEY_VIOLATION;
512 if (pkt->mask & RXE_FLUSH_MASK) {
513 /* FLUSH MR may not set va or resid
514 * no need to check range since we will flush whole mr
516 if (feth_sel(pkt) == IB_FLUSH_MR)
517 goto skip_check_range;
520 if (mr_check_range(mr, va + qp->resp.offset, resid)) {
521 state = RESPST_ERR_RKEY_VIOLATION;
526 if (pkt->mask & (RXE_WRITE_MASK | RXE_ATOMIC_WRITE_MASK)) {
528 if (pktlen != mtu || bth_pad(pkt)) {
529 state = RESPST_ERR_LENGTH;
533 if (pktlen != resid) {
534 state = RESPST_ERR_LENGTH;
537 if ((bth_pad(pkt) != (0x3 & (-resid)))) {
538 /* This case may not be exactly that
539 * but nothing else fits.
541 state = RESPST_ERR_LENGTH;
547 WARN_ON_ONCE(qp->resp.mr);
550 return RESPST_EXECUTE;
562 static enum resp_states send_data_in(struct rxe_qp *qp, void *data_addr,
567 err = copy_data(qp->pd, IB_ACCESS_LOCAL_WRITE, &qp->resp.wqe->dma,
568 data_addr, data_len, RXE_TO_MR_OBJ);
570 return (err == -ENOSPC) ? RESPST_ERR_LENGTH
571 : RESPST_ERR_MALFORMED_WQE;
576 static enum resp_states write_data_in(struct rxe_qp *qp,
577 struct rxe_pkt_info *pkt)
579 enum resp_states rc = RESPST_NONE;
581 int data_len = payload_size(pkt);
583 err = rxe_mr_copy(qp->resp.mr, qp->resp.va + qp->resp.offset,
584 payload_addr(pkt), data_len, RXE_TO_MR_OBJ);
586 rc = RESPST_ERR_RKEY_VIOLATION;
590 qp->resp.va += data_len;
591 qp->resp.resid -= data_len;
597 static struct resp_res *rxe_prepare_res(struct rxe_qp *qp,
598 struct rxe_pkt_info *pkt,
601 struct resp_res *res;
604 res = &qp->resp.resources[qp->resp.res_head];
605 rxe_advance_resp_resource(qp);
606 free_rd_atomic_resource(res);
613 res->read.va = qp->resp.va + qp->resp.offset;
614 res->read.va_org = qp->resp.va + qp->resp.offset;
615 res->read.resid = qp->resp.resid;
616 res->read.length = qp->resp.resid;
617 res->read.rkey = qp->resp.rkey;
619 pkts = max_t(u32, (reth_len(pkt) + qp->mtu - 1)/qp->mtu, 1);
620 res->first_psn = pkt->psn;
621 res->cur_psn = pkt->psn;
622 res->last_psn = (pkt->psn + pkts - 1) & BTH_PSN_MASK;
624 res->state = rdatm_res_state_new;
626 case RXE_ATOMIC_MASK:
627 case RXE_ATOMIC_WRITE_MASK:
628 res->first_psn = pkt->psn;
629 res->last_psn = pkt->psn;
630 res->cur_psn = pkt->psn;
633 res->flush.va = qp->resp.va + qp->resp.offset;
634 res->flush.length = qp->resp.length;
635 res->flush.type = feth_plt(pkt);
636 res->flush.level = feth_sel(pkt);
642 static enum resp_states process_flush(struct rxe_qp *qp,
643 struct rxe_pkt_info *pkt)
646 struct rxe_mr *mr = qp->resp.mr;
647 struct resp_res *res = qp->resp.res;
649 /* oA19-14, oA19-15 */
650 if (res && res->replay)
651 return RESPST_ACKNOWLEDGE;
653 res = rxe_prepare_res(qp, pkt, RXE_FLUSH_MASK);
657 if (res->flush.level == IB_FLUSH_RANGE) {
658 start = res->flush.va;
659 length = res->flush.length;
660 } else { /* level == IB_FLUSH_MR */
661 start = mr->ibmr.iova;
662 length = mr->ibmr.length;
665 if (res->flush.type & IB_FLUSH_PERSISTENT) {
666 if (rxe_flush_pmem_iova(mr, start, length))
667 return RESPST_ERR_RKEY_VIOLATION;
668 /* Make data persistent. */
670 } else if (res->flush.type & IB_FLUSH_GLOBAL) {
671 /* Make data global visibility. */
677 /* next expected psn, read handles this separately */
678 qp->resp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
679 qp->resp.ack_psn = qp->resp.psn;
681 qp->resp.opcode = pkt->opcode;
682 qp->resp.status = IB_WC_SUCCESS;
684 return RESPST_ACKNOWLEDGE;
687 static enum resp_states atomic_reply(struct rxe_qp *qp,
688 struct rxe_pkt_info *pkt)
690 struct rxe_mr *mr = qp->resp.mr;
691 struct resp_res *res = qp->resp.res;
695 res = rxe_prepare_res(qp, pkt, RXE_ATOMIC_MASK);
700 u64 iova = qp->resp.va + qp->resp.offset;
702 err = rxe_mr_do_atomic_op(mr, iova, pkt->opcode,
704 atmeth_swap_add(pkt),
705 &res->atomic.orig_val);
711 /* next expected psn, read handles this separately */
712 qp->resp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
713 qp->resp.ack_psn = qp->resp.psn;
715 qp->resp.opcode = pkt->opcode;
716 qp->resp.status = IB_WC_SUCCESS;
719 return RESPST_ACKNOWLEDGE;
722 static enum resp_states atomic_write_reply(struct rxe_qp *qp,
723 struct rxe_pkt_info *pkt)
725 struct resp_res *res = qp->resp.res;
732 res = rxe_prepare_res(qp, pkt, RXE_ATOMIC_WRITE_MASK);
737 return RESPST_ACKNOWLEDGE;
740 value = *(u64 *)payload_addr(pkt);
741 iova = qp->resp.va + qp->resp.offset;
743 err = rxe_mr_do_atomic_write(mr, iova, value);
750 /* next expected psn, read handles this separately */
751 qp->resp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
752 qp->resp.ack_psn = qp->resp.psn;
754 qp->resp.opcode = pkt->opcode;
755 qp->resp.status = IB_WC_SUCCESS;
757 return RESPST_ACKNOWLEDGE;
760 static struct sk_buff *prepare_ack_packet(struct rxe_qp *qp,
761 struct rxe_pkt_info *ack,
767 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
776 pad = (-payload) & 0x3;
777 paylen = rxe_opcode[opcode].length + payload + pad + RXE_ICRC_SIZE;
779 skb = rxe_init_packet(rxe, &qp->pri_av, paylen, ack);
784 ack->opcode = opcode;
785 ack->mask = rxe_opcode[opcode].mask;
786 ack->paylen = paylen;
789 bth_init(ack, opcode, 0, 0, pad, IB_DEFAULT_PKEY_FULL,
790 qp->attr.dest_qp_num, 0, psn);
792 if (ack->mask & RXE_AETH_MASK) {
793 aeth_set_syn(ack, syndrome);
794 aeth_set_msn(ack, qp->resp.msn);
797 if (ack->mask & RXE_ATMACK_MASK)
798 atmack_set_orig(ack, qp->resp.res->atomic.orig_val);
800 err = rxe_prepare(&qp->pri_av, ack, skb);
810 * rxe_recheck_mr - revalidate MR from rkey and get a reference
814 * This code allows the MR to be invalidated or deregistered or
815 * the MW if one was used to be invalidated or deallocated.
816 * It is assumed that the access permissions if originally good
817 * are OK and the mappings to be unchanged.
819 * TODO: If someone reregisters an MR to change its size or
820 * access permissions during the processing of an RDMA read
821 * we should kill the responder resource and complete the
822 * operation with an error.
824 * Return: mr on success else NULL
826 static struct rxe_mr *rxe_recheck_mr(struct rxe_qp *qp, u32 rkey)
828 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
832 if (rkey_is_mw(rkey)) {
833 mw = rxe_pool_get_index(&rxe->mw_pool, rkey >> 8);
838 if (mw->rkey != rkey || mw->state != RXE_MW_STATE_VALID ||
839 !mr || mr->state != RXE_MR_STATE_VALID) {
850 mr = rxe_pool_get_index(&rxe->mr_pool, rkey >> 8);
854 if (mr->rkey != rkey || mr->state != RXE_MR_STATE_VALID) {
862 /* RDMA read response. If res is not NULL, then we have a current RDMA request
863 * being processed or replayed.
865 static enum resp_states read_reply(struct rxe_qp *qp,
866 struct rxe_pkt_info *req_pkt)
868 struct rxe_pkt_info ack_pkt;
871 enum resp_states state;
875 struct resp_res *res = qp->resp.res;
879 res = rxe_prepare_res(qp, req_pkt, RXE_READ_MASK);
883 if (res->state == rdatm_res_state_new) {
884 if (!res->replay || qp->resp.length == 0) {
885 /* if length == 0 mr will be NULL (is ok)
886 * otherwise qp->resp.mr holds a ref on mr
887 * which we transfer to mr and drop below.
892 mr = rxe_recheck_mr(qp, res->read.rkey);
894 return RESPST_ERR_RKEY_VIOLATION;
897 if (res->read.resid <= mtu)
898 opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY;
900 opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST;
902 /* re-lookup mr from rkey on all later packets.
903 * length will be non-zero. This can fail if someone
904 * modifies or destroys the mr since the first packet.
906 mr = rxe_recheck_mr(qp, res->read.rkey);
908 return RESPST_ERR_RKEY_VIOLATION;
910 if (res->read.resid > mtu)
911 opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE;
913 opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST;
916 res->state = rdatm_res_state_next;
918 payload = min_t(int, res->read.resid, mtu);
920 skb = prepare_ack_packet(qp, &ack_pkt, opcode, payload,
921 res->cur_psn, AETH_ACK_UNLIMITED);
923 state = RESPST_ERR_RNR;
927 err = rxe_mr_copy(mr, res->read.va, payload_addr(&ack_pkt),
928 payload, RXE_FROM_MR_OBJ);
931 state = RESPST_ERR_RKEY_VIOLATION;
935 if (bth_pad(&ack_pkt)) {
936 u8 *pad = payload_addr(&ack_pkt) + payload;
938 memset(pad, 0, bth_pad(&ack_pkt));
941 /* rxe_xmit_packet always consumes the skb */
942 err = rxe_xmit_packet(qp, &ack_pkt, skb);
944 state = RESPST_ERR_RNR;
948 res->read.va += payload;
949 res->read.resid -= payload;
950 res->cur_psn = (res->cur_psn + 1) & BTH_PSN_MASK;
952 if (res->read.resid > 0) {
957 qp->resp.opcode = -1;
958 if (psn_compare(res->cur_psn, qp->resp.psn) >= 0)
959 qp->resp.psn = res->cur_psn;
960 state = RESPST_CLEANUP;
969 static int invalidate_rkey(struct rxe_qp *qp, u32 rkey)
971 if (rkey_is_mw(rkey))
972 return rxe_invalidate_mw(qp, rkey);
974 return rxe_invalidate_mr(qp, rkey);
977 /* Executes a new request. A retried request never reach that function (send
978 * and writes are discarded, and reads and atomics are retried elsewhere.
980 static enum resp_states execute(struct rxe_qp *qp, struct rxe_pkt_info *pkt)
982 enum resp_states err;
983 struct sk_buff *skb = PKT_TO_SKB(pkt);
984 union rdma_network_hdr hdr;
986 if (pkt->mask & RXE_SEND_MASK) {
987 if (qp_type(qp) == IB_QPT_UD ||
988 qp_type(qp) == IB_QPT_GSI) {
989 if (skb->protocol == htons(ETH_P_IP)) {
990 memset(&hdr.reserved, 0,
991 sizeof(hdr.reserved));
992 memcpy(&hdr.roce4grh, ip_hdr(skb),
993 sizeof(hdr.roce4grh));
994 err = send_data_in(qp, &hdr, sizeof(hdr));
996 err = send_data_in(qp, ipv6_hdr(skb),
1002 err = send_data_in(qp, payload_addr(pkt), payload_size(pkt));
1005 } else if (pkt->mask & RXE_WRITE_MASK) {
1006 err = write_data_in(qp, pkt);
1009 } else if (pkt->mask & RXE_READ_MASK) {
1010 /* For RDMA Read we can increment the msn now. See C9-148. */
1012 return RESPST_READ_REPLY;
1013 } else if (pkt->mask & RXE_ATOMIC_MASK) {
1014 return RESPST_ATOMIC_REPLY;
1015 } else if (pkt->mask & RXE_ATOMIC_WRITE_MASK) {
1016 return RESPST_ATOMIC_WRITE_REPLY;
1017 } else if (pkt->mask & RXE_FLUSH_MASK) {
1018 return RESPST_PROCESS_FLUSH;
1024 if (pkt->mask & RXE_IETH_MASK) {
1025 u32 rkey = ieth_rkey(pkt);
1027 err = invalidate_rkey(qp, rkey);
1029 return RESPST_ERR_INVALIDATE_RKEY;
1032 if (pkt->mask & RXE_END_MASK)
1033 /* We successfully processed this new request. */
1036 /* next expected psn, read handles this separately */
1037 qp->resp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
1038 qp->resp.ack_psn = qp->resp.psn;
1040 qp->resp.opcode = pkt->opcode;
1041 qp->resp.status = IB_WC_SUCCESS;
1043 if (pkt->mask & RXE_COMP_MASK)
1044 return RESPST_COMPLETE;
1045 else if (qp_type(qp) == IB_QPT_RC)
1046 return RESPST_ACKNOWLEDGE;
1048 return RESPST_CLEANUP;
1051 static enum resp_states do_complete(struct rxe_qp *qp,
1052 struct rxe_pkt_info *pkt)
1055 struct ib_wc *wc = &cqe.ibwc;
1056 struct ib_uverbs_wc *uwc = &cqe.uibwc;
1057 struct rxe_recv_wqe *wqe = qp->resp.wqe;
1058 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
1059 unsigned long flags;
1064 memset(&cqe, 0, sizeof(cqe));
1066 if (qp->rcq->is_user) {
1067 uwc->status = qp->resp.status;
1068 uwc->qp_num = qp->ibqp.qp_num;
1069 uwc->wr_id = wqe->wr_id;
1071 wc->status = qp->resp.status;
1073 wc->wr_id = wqe->wr_id;
1076 if (wc->status == IB_WC_SUCCESS) {
1077 rxe_counter_inc(rxe, RXE_CNT_RDMA_RECV);
1078 wc->opcode = (pkt->mask & RXE_IMMDT_MASK &&
1079 pkt->mask & RXE_WRITE_MASK) ?
1080 IB_WC_RECV_RDMA_WITH_IMM : IB_WC_RECV;
1081 wc->byte_len = (pkt->mask & RXE_IMMDT_MASK &&
1082 pkt->mask & RXE_WRITE_MASK) ?
1083 qp->resp.length : wqe->dma.length - wqe->dma.resid;
1085 /* fields after byte_len are different between kernel and user
1088 if (qp->rcq->is_user) {
1089 uwc->wc_flags = IB_WC_GRH;
1091 if (pkt->mask & RXE_IMMDT_MASK) {
1092 uwc->wc_flags |= IB_WC_WITH_IMM;
1093 uwc->ex.imm_data = immdt_imm(pkt);
1096 if (pkt->mask & RXE_IETH_MASK) {
1097 uwc->wc_flags |= IB_WC_WITH_INVALIDATE;
1098 uwc->ex.invalidate_rkey = ieth_rkey(pkt);
1101 if (pkt->mask & RXE_DETH_MASK)
1102 uwc->src_qp = deth_sqp(pkt);
1104 uwc->port_num = qp->attr.port_num;
1106 struct sk_buff *skb = PKT_TO_SKB(pkt);
1108 wc->wc_flags = IB_WC_GRH | IB_WC_WITH_NETWORK_HDR_TYPE;
1109 if (skb->protocol == htons(ETH_P_IP))
1110 wc->network_hdr_type = RDMA_NETWORK_IPV4;
1112 wc->network_hdr_type = RDMA_NETWORK_IPV6;
1114 if (is_vlan_dev(skb->dev)) {
1115 wc->wc_flags |= IB_WC_WITH_VLAN;
1116 wc->vlan_id = vlan_dev_vlan_id(skb->dev);
1119 if (pkt->mask & RXE_IMMDT_MASK) {
1120 wc->wc_flags |= IB_WC_WITH_IMM;
1121 wc->ex.imm_data = immdt_imm(pkt);
1124 if (pkt->mask & RXE_IETH_MASK) {
1125 wc->wc_flags |= IB_WC_WITH_INVALIDATE;
1126 wc->ex.invalidate_rkey = ieth_rkey(pkt);
1129 if (pkt->mask & RXE_DETH_MASK)
1130 wc->src_qp = deth_sqp(pkt);
1132 wc->port_num = qp->attr.port_num;
1135 if (wc->status != IB_WC_WR_FLUSH_ERR)
1136 rxe_err_qp(qp, "non-flush error status = %d",
1140 /* have copy for srq and reference for !srq */
1142 queue_advance_consumer(qp->rq.queue, QUEUE_TYPE_FROM_CLIENT);
1144 qp->resp.wqe = NULL;
1146 if (rxe_cq_post(qp->rcq, &cqe, pkt ? bth_se(pkt) : 1))
1147 return RESPST_ERR_CQ_OVERFLOW;
1150 spin_lock_irqsave(&qp->state_lock, flags);
1151 if (unlikely(qp_state(qp) == IB_QPS_ERR)) {
1152 spin_unlock_irqrestore(&qp->state_lock, flags);
1153 return RESPST_CHK_RESOURCE;
1155 spin_unlock_irqrestore(&qp->state_lock, flags);
1159 if (qp_type(qp) == IB_QPT_RC)
1160 return RESPST_ACKNOWLEDGE;
1162 return RESPST_CLEANUP;
1166 static int send_common_ack(struct rxe_qp *qp, u8 syndrome, u32 psn,
1167 int opcode, const char *msg)
1170 struct rxe_pkt_info ack_pkt;
1171 struct sk_buff *skb;
1173 skb = prepare_ack_packet(qp, &ack_pkt, opcode, 0, psn, syndrome);
1177 err = rxe_xmit_packet(qp, &ack_pkt, skb);
1179 rxe_dbg_qp(qp, "Failed sending %s\n", msg);
1184 static int send_ack(struct rxe_qp *qp, u8 syndrome, u32 psn)
1186 return send_common_ack(qp, syndrome, psn,
1187 IB_OPCODE_RC_ACKNOWLEDGE, "ACK");
1190 static int send_atomic_ack(struct rxe_qp *qp, u8 syndrome, u32 psn)
1192 int ret = send_common_ack(qp, syndrome, psn,
1193 IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE, "ATOMIC ACK");
1195 /* have to clear this since it is used to trigger
1198 qp->resp.res = NULL;
1202 static int send_read_response_ack(struct rxe_qp *qp, u8 syndrome, u32 psn)
1204 int ret = send_common_ack(qp, syndrome, psn,
1205 IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY,
1206 "RDMA READ response of length zero ACK");
1208 /* have to clear this since it is used to trigger
1211 qp->resp.res = NULL;
1215 static enum resp_states acknowledge(struct rxe_qp *qp,
1216 struct rxe_pkt_info *pkt)
1218 if (qp_type(qp) != IB_QPT_RC)
1219 return RESPST_CLEANUP;
1221 if (qp->resp.aeth_syndrome != AETH_ACK_UNLIMITED)
1222 send_ack(qp, qp->resp.aeth_syndrome, pkt->psn);
1223 else if (pkt->mask & RXE_ATOMIC_MASK)
1224 send_atomic_ack(qp, AETH_ACK_UNLIMITED, pkt->psn);
1225 else if (pkt->mask & (RXE_FLUSH_MASK | RXE_ATOMIC_WRITE_MASK))
1226 send_read_response_ack(qp, AETH_ACK_UNLIMITED, pkt->psn);
1227 else if (bth_ack(pkt))
1228 send_ack(qp, AETH_ACK_UNLIMITED, pkt->psn);
1230 return RESPST_CLEANUP;
1233 static enum resp_states cleanup(struct rxe_qp *qp,
1234 struct rxe_pkt_info *pkt)
1236 struct sk_buff *skb;
1239 skb = skb_dequeue(&qp->req_pkts);
1242 ib_device_put(qp->ibqp.device);
1246 rxe_put(qp->resp.mr);
1253 static struct resp_res *find_resource(struct rxe_qp *qp, u32 psn)
1257 for (i = 0; i < qp->attr.max_dest_rd_atomic; i++) {
1258 struct resp_res *res = &qp->resp.resources[i];
1263 if (psn_compare(psn, res->first_psn) >= 0 &&
1264 psn_compare(psn, res->last_psn) <= 0) {
1272 static enum resp_states duplicate_request(struct rxe_qp *qp,
1273 struct rxe_pkt_info *pkt)
1275 enum resp_states rc;
1276 u32 prev_psn = (qp->resp.ack_psn - 1) & BTH_PSN_MASK;
1278 if (pkt->mask & RXE_SEND_MASK ||
1279 pkt->mask & RXE_WRITE_MASK) {
1280 /* SEND. Ack again and cleanup. C9-105. */
1281 send_ack(qp, AETH_ACK_UNLIMITED, prev_psn);
1282 return RESPST_CLEANUP;
1283 } else if (pkt->mask & RXE_FLUSH_MASK) {
1284 struct resp_res *res;
1286 /* Find the operation in our list of responder resources. */
1287 res = find_resource(qp, pkt->psn);
1290 res->cur_psn = pkt->psn;
1292 rc = RESPST_PROCESS_FLUSH;
1296 /* Resource not found. Class D error. Drop the request. */
1297 rc = RESPST_CLEANUP;
1299 } else if (pkt->mask & RXE_READ_MASK) {
1300 struct resp_res *res;
1302 res = find_resource(qp, pkt->psn);
1304 /* Resource not found. Class D error. Drop the
1307 rc = RESPST_CLEANUP;
1310 /* Ensure this new request is the same as the previous
1311 * one or a subset of it.
1313 u64 iova = reth_va(pkt);
1314 u32 resid = reth_len(pkt);
1316 if (iova < res->read.va_org ||
1317 resid > res->read.length ||
1318 (iova + resid) > (res->read.va_org +
1319 res->read.length)) {
1320 rc = RESPST_CLEANUP;
1324 if (reth_rkey(pkt) != res->read.rkey) {
1325 rc = RESPST_CLEANUP;
1329 res->cur_psn = pkt->psn;
1330 res->state = (pkt->psn == res->first_psn) ?
1331 rdatm_res_state_new :
1332 rdatm_res_state_replay;
1335 /* Reset the resource, except length. */
1336 res->read.va_org = iova;
1337 res->read.va = iova;
1338 res->read.resid = resid;
1340 /* Replay the RDMA read reply. */
1342 rc = RESPST_READ_REPLY;
1346 struct resp_res *res;
1348 /* Find the operation in our list of responder resources. */
1349 res = find_resource(qp, pkt->psn);
1352 res->cur_psn = pkt->psn;
1354 rc = pkt->mask & RXE_ATOMIC_MASK ?
1355 RESPST_ATOMIC_REPLY :
1356 RESPST_ATOMIC_WRITE_REPLY;
1360 /* Resource not found. Class D error. Drop the request. */
1361 rc = RESPST_CLEANUP;
1368 /* Process a class A or C. Both are treated the same in this implementation. */
1369 static void do_class_ac_error(struct rxe_qp *qp, u8 syndrome,
1370 enum ib_wc_status status)
1372 qp->resp.aeth_syndrome = syndrome;
1373 qp->resp.status = status;
1375 /* indicate that we should go through the ERROR state */
1376 qp->resp.goto_error = 1;
1379 static enum resp_states do_class_d1e_error(struct rxe_qp *qp)
1384 qp->resp.drop_msg = 1;
1386 qp->resp.status = IB_WC_REM_INV_REQ_ERR;
1387 return RESPST_COMPLETE;
1389 return RESPST_CLEANUP;
1392 /* Class D1. This packet may be the start of a
1393 * new message and could be valid. The previous
1394 * message is invalid and ignored. reset the
1395 * recv wr to its original state
1398 qp->resp.wqe->dma.resid = qp->resp.wqe->dma.length;
1399 qp->resp.wqe->dma.cur_sge = 0;
1400 qp->resp.wqe->dma.sge_offset = 0;
1401 qp->resp.opcode = -1;
1405 rxe_put(qp->resp.mr);
1409 return RESPST_CLEANUP;
1413 /* drain incoming request packet queue */
1414 static void drain_req_pkts(struct rxe_qp *qp)
1416 struct sk_buff *skb;
1418 while ((skb = skb_dequeue(&qp->req_pkts))) {
1421 ib_device_put(qp->ibqp.device);
1425 /* complete receive wqe with flush error */
1426 static int flush_recv_wqe(struct rxe_qp *qp, struct rxe_recv_wqe *wqe)
1428 struct rxe_cqe cqe = {};
1429 struct ib_wc *wc = &cqe.ibwc;
1430 struct ib_uverbs_wc *uwc = &cqe.uibwc;
1433 if (qp->rcq->is_user) {
1434 uwc->wr_id = wqe->wr_id;
1435 uwc->status = IB_WC_WR_FLUSH_ERR;
1436 uwc->qp_num = qp_num(qp);
1438 wc->wr_id = wqe->wr_id;
1439 wc->status = IB_WC_WR_FLUSH_ERR;
1443 err = rxe_cq_post(qp->rcq, &cqe, 0);
1445 rxe_dbg_cq(qp->rcq, "post cq failed err = %d", err);
1450 /* drain and optionally complete the recive queue
1451 * if unable to complete a wqe stop completing and
1452 * just flush the remaining wqes
1454 static void flush_recv_queue(struct rxe_qp *qp, bool notify)
1456 struct rxe_queue *q = qp->rq.queue;
1457 struct rxe_recv_wqe *wqe;
1461 if (notify && qp->ibqp.event_handler) {
1464 ev.device = qp->ibqp.device;
1465 ev.element.qp = &qp->ibqp;
1466 ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
1467 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
1472 /* recv queue not created. nothing to do. */
1476 while ((wqe = queue_head(q, q->type))) {
1478 err = flush_recv_wqe(qp, wqe);
1482 queue_advance_consumer(q, q->type);
1485 qp->resp.wqe = NULL;
1488 int rxe_responder(struct rxe_qp *qp)
1490 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
1491 enum resp_states state;
1492 struct rxe_pkt_info *pkt = NULL;
1494 unsigned long flags;
1496 spin_lock_irqsave(&qp->state_lock, flags);
1497 if (!qp->valid || qp_state(qp) == IB_QPS_ERR ||
1498 qp_state(qp) == IB_QPS_RESET) {
1499 bool notify = qp->valid && (qp_state(qp) == IB_QPS_ERR);
1502 flush_recv_queue(qp, notify);
1503 spin_unlock_irqrestore(&qp->state_lock, flags);
1506 spin_unlock_irqrestore(&qp->state_lock, flags);
1508 qp->resp.aeth_syndrome = AETH_ACK_UNLIMITED;
1510 state = RESPST_GET_REQ;
1513 rxe_dbg_qp(qp, "state = %s\n", resp_state_name[state]);
1515 case RESPST_GET_REQ:
1516 state = get_req(qp, &pkt);
1518 case RESPST_CHK_PSN:
1519 state = check_psn(qp, pkt);
1521 case RESPST_CHK_OP_SEQ:
1522 state = check_op_seq(qp, pkt);
1524 case RESPST_CHK_OP_VALID:
1525 state = check_op_valid(qp, pkt);
1527 case RESPST_CHK_RESOURCE:
1528 state = check_resource(qp, pkt);
1530 case RESPST_CHK_LENGTH:
1531 state = rxe_resp_check_length(qp, pkt);
1533 case RESPST_CHK_RKEY:
1534 state = check_rkey(qp, pkt);
1536 case RESPST_EXECUTE:
1537 state = execute(qp, pkt);
1539 case RESPST_COMPLETE:
1540 state = do_complete(qp, pkt);
1542 case RESPST_READ_REPLY:
1543 state = read_reply(qp, pkt);
1545 case RESPST_ATOMIC_REPLY:
1546 state = atomic_reply(qp, pkt);
1548 case RESPST_ATOMIC_WRITE_REPLY:
1549 state = atomic_write_reply(qp, pkt);
1551 case RESPST_PROCESS_FLUSH:
1552 state = process_flush(qp, pkt);
1554 case RESPST_ACKNOWLEDGE:
1555 state = acknowledge(qp, pkt);
1557 case RESPST_CLEANUP:
1558 state = cleanup(qp, pkt);
1560 case RESPST_DUPLICATE_REQUEST:
1561 state = duplicate_request(qp, pkt);
1563 case RESPST_ERR_PSN_OUT_OF_SEQ:
1564 /* RC only - Class B. Drop packet. */
1565 send_ack(qp, AETH_NAK_PSN_SEQ_ERROR, qp->resp.psn);
1566 state = RESPST_CLEANUP;
1569 case RESPST_ERR_TOO_MANY_RDMA_ATM_REQ:
1570 case RESPST_ERR_MISSING_OPCODE_FIRST:
1571 case RESPST_ERR_MISSING_OPCODE_LAST_C:
1572 case RESPST_ERR_UNSUPPORTED_OPCODE:
1573 case RESPST_ERR_MISALIGNED_ATOMIC:
1574 /* RC Only - Class C. */
1575 do_class_ac_error(qp, AETH_NAK_INVALID_REQ,
1576 IB_WC_REM_INV_REQ_ERR);
1577 state = RESPST_COMPLETE;
1580 case RESPST_ERR_MISSING_OPCODE_LAST_D1E:
1581 state = do_class_d1e_error(qp);
1583 case RESPST_ERR_RNR:
1584 if (qp_type(qp) == IB_QPT_RC) {
1585 rxe_counter_inc(rxe, RXE_CNT_SND_RNR);
1587 send_ack(qp, AETH_RNR_NAK |
1589 qp->attr.min_rnr_timer),
1592 /* UD/UC - class D */
1593 qp->resp.drop_msg = 1;
1595 state = RESPST_CLEANUP;
1598 case RESPST_ERR_RKEY_VIOLATION:
1599 if (qp_type(qp) == IB_QPT_RC) {
1601 do_class_ac_error(qp, AETH_NAK_REM_ACC_ERR,
1602 IB_WC_REM_ACCESS_ERR);
1603 state = RESPST_COMPLETE;
1605 qp->resp.drop_msg = 1;
1607 /* UC/SRQ Class D */
1608 qp->resp.status = IB_WC_REM_ACCESS_ERR;
1609 state = RESPST_COMPLETE;
1611 /* UC/non-SRQ Class E. */
1612 state = RESPST_CLEANUP;
1617 case RESPST_ERR_INVALIDATE_RKEY:
1619 qp->resp.goto_error = 1;
1620 qp->resp.status = IB_WC_REM_INV_REQ_ERR;
1621 state = RESPST_COMPLETE;
1624 case RESPST_ERR_LENGTH:
1625 if (qp_type(qp) == IB_QPT_RC) {
1627 do_class_ac_error(qp, AETH_NAK_INVALID_REQ,
1628 IB_WC_REM_INV_REQ_ERR);
1629 state = RESPST_COMPLETE;
1630 } else if (qp->srq) {
1631 /* UC/UD - class E */
1632 qp->resp.status = IB_WC_REM_INV_REQ_ERR;
1633 state = RESPST_COMPLETE;
1635 /* UC/UD - class D */
1636 qp->resp.drop_msg = 1;
1637 state = RESPST_CLEANUP;
1641 case RESPST_ERR_MALFORMED_WQE:
1643 do_class_ac_error(qp, AETH_NAK_REM_OP_ERR,
1644 IB_WC_LOC_QP_OP_ERR);
1645 state = RESPST_COMPLETE;
1648 case RESPST_ERR_CQ_OVERFLOW:
1650 state = RESPST_ERROR;
1654 if (qp->resp.goto_error) {
1655 state = RESPST_ERROR;
1662 if (qp->resp.goto_error) {
1663 state = RESPST_ERROR;
1670 qp->resp.goto_error = 0;
1671 rxe_dbg_qp(qp, "moved to error state\n");
1680 /* A non-zero return value will cause rxe_do_task to
1681 * exit its loop and end the work item. A zero return
1682 * will continue looping and return to rxe_responder