2 * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved.
3 * Copyright (c) 2015 System Fabric Works, Inc. All rights reserved.
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the
9 * OpenIB.org BSD license below:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
19 * - Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials
22 * provided with the distribution.
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34 #include <linux/skbuff.h>
38 #include "rxe_queue.h"
54 RESPST_DUPLICATE_REQUEST,
55 RESPST_ERR_MALFORMED_WQE,
56 RESPST_ERR_UNSUPPORTED_OPCODE,
57 RESPST_ERR_MISALIGNED_ATOMIC,
58 RESPST_ERR_PSN_OUT_OF_SEQ,
59 RESPST_ERR_MISSING_OPCODE_FIRST,
60 RESPST_ERR_MISSING_OPCODE_LAST_C,
61 RESPST_ERR_MISSING_OPCODE_LAST_D1E,
62 RESPST_ERR_TOO_MANY_RDMA_ATM_REQ,
64 RESPST_ERR_RKEY_VIOLATION,
66 RESPST_ERR_CQ_OVERFLOW,
73 static char *resp_state_name[] = {
74 [RESPST_NONE] = "NONE",
75 [RESPST_GET_REQ] = "GET_REQ",
76 [RESPST_CHK_PSN] = "CHK_PSN",
77 [RESPST_CHK_OP_SEQ] = "CHK_OP_SEQ",
78 [RESPST_CHK_OP_VALID] = "CHK_OP_VALID",
79 [RESPST_CHK_RESOURCE] = "CHK_RESOURCE",
80 [RESPST_CHK_LENGTH] = "CHK_LENGTH",
81 [RESPST_CHK_RKEY] = "CHK_RKEY",
82 [RESPST_EXECUTE] = "EXECUTE",
83 [RESPST_READ_REPLY] = "READ_REPLY",
84 [RESPST_COMPLETE] = "COMPLETE",
85 [RESPST_ACKNOWLEDGE] = "ACKNOWLEDGE",
86 [RESPST_CLEANUP] = "CLEANUP",
87 [RESPST_DUPLICATE_REQUEST] = "DUPLICATE_REQUEST",
88 [RESPST_ERR_MALFORMED_WQE] = "ERR_MALFORMED_WQE",
89 [RESPST_ERR_UNSUPPORTED_OPCODE] = "ERR_UNSUPPORTED_OPCODE",
90 [RESPST_ERR_MISALIGNED_ATOMIC] = "ERR_MISALIGNED_ATOMIC",
91 [RESPST_ERR_PSN_OUT_OF_SEQ] = "ERR_PSN_OUT_OF_SEQ",
92 [RESPST_ERR_MISSING_OPCODE_FIRST] = "ERR_MISSING_OPCODE_FIRST",
93 [RESPST_ERR_MISSING_OPCODE_LAST_C] = "ERR_MISSING_OPCODE_LAST_C",
94 [RESPST_ERR_MISSING_OPCODE_LAST_D1E] = "ERR_MISSING_OPCODE_LAST_D1E",
95 [RESPST_ERR_TOO_MANY_RDMA_ATM_REQ] = "ERR_TOO_MANY_RDMA_ATM_REQ",
96 [RESPST_ERR_RNR] = "ERR_RNR",
97 [RESPST_ERR_RKEY_VIOLATION] = "ERR_RKEY_VIOLATION",
98 [RESPST_ERR_LENGTH] = "ERR_LENGTH",
99 [RESPST_ERR_CQ_OVERFLOW] = "ERR_CQ_OVERFLOW",
100 [RESPST_ERROR] = "ERROR",
101 [RESPST_RESET] = "RESET",
102 [RESPST_DONE] = "DONE",
103 [RESPST_EXIT] = "EXIT",
106 /* rxe_recv calls here to add a request packet to the input queue */
107 void rxe_resp_queue_pkt(struct rxe_dev *rxe, struct rxe_qp *qp,
111 struct rxe_pkt_info *pkt = SKB_TO_PKT(skb);
113 skb_queue_tail(&qp->req_pkts, skb);
115 must_sched = (pkt->opcode == IB_OPCODE_RC_RDMA_READ_REQUEST) ||
116 (skb_queue_len(&qp->req_pkts) > 1);
118 rxe_run_task(&qp->resp.task, must_sched);
121 static inline enum resp_states get_req(struct rxe_qp *qp,
122 struct rxe_pkt_info **pkt_p)
126 if (qp->resp.state == QP_STATE_ERROR) {
127 skb = skb_dequeue(&qp->req_pkts);
129 /* drain request packet queue */
132 return RESPST_GET_REQ;
135 /* go drain recv wr queue */
136 return RESPST_CHK_RESOURCE;
139 skb = skb_peek(&qp->req_pkts);
143 *pkt_p = SKB_TO_PKT(skb);
145 return (qp->resp.res) ? RESPST_READ_REPLY : RESPST_CHK_PSN;
148 static enum resp_states check_psn(struct rxe_qp *qp,
149 struct rxe_pkt_info *pkt)
151 int diff = psn_compare(pkt->psn, qp->resp.psn);
152 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
154 switch (qp_type(qp)) {
157 if (qp->resp.sent_psn_nak)
158 return RESPST_CLEANUP;
160 qp->resp.sent_psn_nak = 1;
161 rxe_counter_inc(rxe, RXE_CNT_OUT_OF_SEQ_REQ);
162 return RESPST_ERR_PSN_OUT_OF_SEQ;
164 } else if (diff < 0) {
165 rxe_counter_inc(rxe, RXE_CNT_DUP_REQ);
166 return RESPST_DUPLICATE_REQUEST;
169 if (qp->resp.sent_psn_nak)
170 qp->resp.sent_psn_nak = 0;
175 if (qp->resp.drop_msg || diff != 0) {
176 if (pkt->mask & RXE_START_MASK) {
177 qp->resp.drop_msg = 0;
178 return RESPST_CHK_OP_SEQ;
181 qp->resp.drop_msg = 1;
182 return RESPST_CLEANUP;
189 return RESPST_CHK_OP_SEQ;
192 static enum resp_states check_op_seq(struct rxe_qp *qp,
193 struct rxe_pkt_info *pkt)
195 switch (qp_type(qp)) {
197 switch (qp->resp.opcode) {
198 case IB_OPCODE_RC_SEND_FIRST:
199 case IB_OPCODE_RC_SEND_MIDDLE:
200 switch (pkt->opcode) {
201 case IB_OPCODE_RC_SEND_MIDDLE:
202 case IB_OPCODE_RC_SEND_LAST:
203 case IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE:
204 case IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE:
205 return RESPST_CHK_OP_VALID;
207 return RESPST_ERR_MISSING_OPCODE_LAST_C;
210 case IB_OPCODE_RC_RDMA_WRITE_FIRST:
211 case IB_OPCODE_RC_RDMA_WRITE_MIDDLE:
212 switch (pkt->opcode) {
213 case IB_OPCODE_RC_RDMA_WRITE_MIDDLE:
214 case IB_OPCODE_RC_RDMA_WRITE_LAST:
215 case IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
216 return RESPST_CHK_OP_VALID;
218 return RESPST_ERR_MISSING_OPCODE_LAST_C;
222 switch (pkt->opcode) {
223 case IB_OPCODE_RC_SEND_MIDDLE:
224 case IB_OPCODE_RC_SEND_LAST:
225 case IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE:
226 case IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE:
227 case IB_OPCODE_RC_RDMA_WRITE_MIDDLE:
228 case IB_OPCODE_RC_RDMA_WRITE_LAST:
229 case IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
230 return RESPST_ERR_MISSING_OPCODE_FIRST;
232 return RESPST_CHK_OP_VALID;
238 switch (qp->resp.opcode) {
239 case IB_OPCODE_UC_SEND_FIRST:
240 case IB_OPCODE_UC_SEND_MIDDLE:
241 switch (pkt->opcode) {
242 case IB_OPCODE_UC_SEND_MIDDLE:
243 case IB_OPCODE_UC_SEND_LAST:
244 case IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE:
245 return RESPST_CHK_OP_VALID;
247 return RESPST_ERR_MISSING_OPCODE_LAST_D1E;
250 case IB_OPCODE_UC_RDMA_WRITE_FIRST:
251 case IB_OPCODE_UC_RDMA_WRITE_MIDDLE:
252 switch (pkt->opcode) {
253 case IB_OPCODE_UC_RDMA_WRITE_MIDDLE:
254 case IB_OPCODE_UC_RDMA_WRITE_LAST:
255 case IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
256 return RESPST_CHK_OP_VALID;
258 return RESPST_ERR_MISSING_OPCODE_LAST_D1E;
262 switch (pkt->opcode) {
263 case IB_OPCODE_UC_SEND_MIDDLE:
264 case IB_OPCODE_UC_SEND_LAST:
265 case IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE:
266 case IB_OPCODE_UC_RDMA_WRITE_MIDDLE:
267 case IB_OPCODE_UC_RDMA_WRITE_LAST:
268 case IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
269 qp->resp.drop_msg = 1;
270 return RESPST_CLEANUP;
272 return RESPST_CHK_OP_VALID;
278 return RESPST_CHK_OP_VALID;
282 static enum resp_states check_op_valid(struct rxe_qp *qp,
283 struct rxe_pkt_info *pkt)
285 switch (qp_type(qp)) {
287 if (((pkt->mask & RXE_READ_MASK) &&
288 !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_READ)) ||
289 ((pkt->mask & RXE_WRITE_MASK) &&
290 !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_WRITE)) ||
291 ((pkt->mask & RXE_ATOMIC_MASK) &&
292 !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_ATOMIC))) {
293 return RESPST_ERR_UNSUPPORTED_OPCODE;
299 if ((pkt->mask & RXE_WRITE_MASK) &&
300 !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_WRITE)) {
301 qp->resp.drop_msg = 1;
302 return RESPST_CLEANUP;
317 return RESPST_CHK_RESOURCE;
320 static enum resp_states get_srq_wqe(struct rxe_qp *qp)
322 struct rxe_srq *srq = qp->srq;
323 struct rxe_queue *q = srq->rq.queue;
324 struct rxe_recv_wqe *wqe;
328 return RESPST_ERR_RNR;
330 spin_lock_bh(&srq->rq.consumer_lock);
334 spin_unlock_bh(&srq->rq.consumer_lock);
335 return RESPST_ERR_RNR;
338 /* note kernel and user space recv wqes have same size */
339 memcpy(&qp->resp.srq_wqe, wqe, sizeof(qp->resp.srq_wqe));
341 qp->resp.wqe = &qp->resp.srq_wqe.wqe;
344 if (srq->limit && srq->ibsrq.event_handler &&
345 (queue_count(q) < srq->limit)) {
350 spin_unlock_bh(&srq->rq.consumer_lock);
351 return RESPST_CHK_LENGTH;
354 spin_unlock_bh(&srq->rq.consumer_lock);
355 ev.device = qp->ibqp.device;
356 ev.element.srq = qp->ibqp.srq;
357 ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
358 srq->ibsrq.event_handler(&ev, srq->ibsrq.srq_context);
359 return RESPST_CHK_LENGTH;
362 static enum resp_states check_resource(struct rxe_qp *qp,
363 struct rxe_pkt_info *pkt)
365 struct rxe_srq *srq = qp->srq;
367 if (qp->resp.state == QP_STATE_ERROR) {
369 qp->resp.status = IB_WC_WR_FLUSH_ERR;
370 return RESPST_COMPLETE;
372 qp->resp.wqe = queue_head(qp->rq.queue);
374 qp->resp.status = IB_WC_WR_FLUSH_ERR;
375 return RESPST_COMPLETE;
384 if (pkt->mask & RXE_READ_OR_ATOMIC) {
385 /* it is the requesters job to not send
386 * too many read/atomic ops, we just
387 * recycle the responder resource queue
389 if (likely(qp->attr.max_dest_rd_atomic > 0))
390 return RESPST_CHK_LENGTH;
392 return RESPST_ERR_TOO_MANY_RDMA_ATM_REQ;
395 if (pkt->mask & RXE_RWR_MASK) {
397 return get_srq_wqe(qp);
399 qp->resp.wqe = queue_head(qp->rq.queue);
400 return (qp->resp.wqe) ? RESPST_CHK_LENGTH : RESPST_ERR_RNR;
403 return RESPST_CHK_LENGTH;
406 static enum resp_states check_length(struct rxe_qp *qp,
407 struct rxe_pkt_info *pkt)
409 switch (qp_type(qp)) {
411 return RESPST_CHK_RKEY;
414 return RESPST_CHK_RKEY;
417 return RESPST_CHK_RKEY;
421 static enum resp_states check_rkey(struct rxe_qp *qp,
422 struct rxe_pkt_info *pkt)
424 struct rxe_mem *mem = NULL;
430 enum resp_states state;
433 if (pkt->mask & (RXE_READ_MASK | RXE_WRITE_MASK)) {
434 if (pkt->mask & RXE_RETH_MASK) {
435 qp->resp.va = reth_va(pkt);
436 qp->resp.rkey = reth_rkey(pkt);
437 qp->resp.resid = reth_len(pkt);
438 qp->resp.length = reth_len(pkt);
440 access = (pkt->mask & RXE_READ_MASK) ? IB_ACCESS_REMOTE_READ
441 : IB_ACCESS_REMOTE_WRITE;
442 } else if (pkt->mask & RXE_ATOMIC_MASK) {
443 qp->resp.va = atmeth_va(pkt);
444 qp->resp.rkey = atmeth_rkey(pkt);
445 qp->resp.resid = sizeof(u64);
446 access = IB_ACCESS_REMOTE_ATOMIC;
448 return RESPST_EXECUTE;
451 /* A zero-byte op is not required to set an addr or rkey. */
452 if ((pkt->mask & (RXE_READ_MASK | RXE_WRITE_OR_SEND)) &&
453 (pkt->mask & RXE_RETH_MASK) &&
454 reth_len(pkt) == 0) {
455 return RESPST_EXECUTE;
459 rkey = qp->resp.rkey;
460 resid = qp->resp.resid;
461 pktlen = payload_size(pkt);
463 mem = lookup_mem(qp->pd, access, rkey, lookup_remote);
465 state = RESPST_ERR_RKEY_VIOLATION;
469 if (unlikely(mem->state == RXE_MEM_STATE_FREE)) {
470 state = RESPST_ERR_RKEY_VIOLATION;
474 if (mem_check_range(mem, va, resid)) {
475 state = RESPST_ERR_RKEY_VIOLATION;
479 if (pkt->mask & RXE_WRITE_MASK) {
481 if (pktlen != mtu || bth_pad(pkt)) {
482 state = RESPST_ERR_LENGTH;
486 if (pktlen != resid) {
487 state = RESPST_ERR_LENGTH;
490 if ((bth_pad(pkt) != (0x3 & (-resid)))) {
491 /* This case may not be exactly that
492 * but nothing else fits.
494 state = RESPST_ERR_LENGTH;
500 WARN_ON_ONCE(qp->resp.mr);
503 return RESPST_EXECUTE;
511 static enum resp_states send_data_in(struct rxe_qp *qp, void *data_addr,
515 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
517 err = copy_data(rxe, qp->pd, IB_ACCESS_LOCAL_WRITE, &qp->resp.wqe->dma,
518 data_addr, data_len, to_mem_obj, NULL);
520 return (err == -ENOSPC) ? RESPST_ERR_LENGTH
521 : RESPST_ERR_MALFORMED_WQE;
526 static enum resp_states write_data_in(struct rxe_qp *qp,
527 struct rxe_pkt_info *pkt)
529 enum resp_states rc = RESPST_NONE;
531 int data_len = payload_size(pkt);
533 err = rxe_mem_copy(qp->resp.mr, qp->resp.va, payload_addr(pkt),
534 data_len, to_mem_obj, NULL);
536 rc = RESPST_ERR_RKEY_VIOLATION;
540 qp->resp.va += data_len;
541 qp->resp.resid -= data_len;
547 /* Guarantee atomicity of atomic operations at the machine level. */
548 static DEFINE_SPINLOCK(atomic_ops_lock);
550 static enum resp_states process_atomic(struct rxe_qp *qp,
551 struct rxe_pkt_info *pkt)
553 u64 iova = atmeth_va(pkt);
555 enum resp_states ret;
556 struct rxe_mem *mr = qp->resp.mr;
558 if (mr->state != RXE_MEM_STATE_VALID) {
559 ret = RESPST_ERR_RKEY_VIOLATION;
563 vaddr = iova_to_vaddr(mr, iova, sizeof(u64));
565 /* check vaddr is 8 bytes aligned. */
566 if (!vaddr || (uintptr_t)vaddr & 7) {
567 ret = RESPST_ERR_MISALIGNED_ATOMIC;
571 spin_lock_bh(&atomic_ops_lock);
573 qp->resp.atomic_orig = *vaddr;
575 if (pkt->opcode == IB_OPCODE_RC_COMPARE_SWAP ||
576 pkt->opcode == IB_OPCODE_RD_COMPARE_SWAP) {
577 if (*vaddr == atmeth_comp(pkt))
578 *vaddr = atmeth_swap_add(pkt);
580 *vaddr += atmeth_swap_add(pkt);
583 spin_unlock_bh(&atomic_ops_lock);
590 static struct sk_buff *prepare_ack_packet(struct rxe_qp *qp,
591 struct rxe_pkt_info *pkt,
592 struct rxe_pkt_info *ack,
599 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
610 pad = (-payload) & 0x3;
611 paylen = rxe_opcode[opcode].length + payload + pad + RXE_ICRC_SIZE;
613 skb = rxe_init_packet(rxe, &qp->pri_av, paylen, ack);
618 ack->opcode = opcode;
619 ack->mask = rxe_opcode[opcode].mask;
620 ack->offset = pkt->offset;
621 ack->paylen = paylen;
623 /* fill in bth using the request packet headers */
624 memcpy(ack->hdr, pkt->hdr, pkt->offset + RXE_BTH_BYTES);
626 bth_set_opcode(ack, opcode);
627 bth_set_qpn(ack, qp->attr.dest_qp_num);
628 bth_set_pad(ack, pad);
630 bth_set_psn(ack, psn);
634 if (ack->mask & RXE_AETH_MASK) {
635 aeth_set_syn(ack, syndrome);
636 aeth_set_msn(ack, qp->resp.msn);
639 if (ack->mask & RXE_ATMACK_MASK)
640 atmack_set_orig(ack, qp->resp.atomic_orig);
642 err = rxe_prepare(rxe, ack, skb, &crc);
649 /* CRC computation will be continued by the caller */
652 p = payload_addr(ack) + payload + bth_pad(ack);
659 /* RDMA read response. If res is not NULL, then we have a current RDMA request
660 * being processed or replayed.
662 static enum resp_states read_reply(struct rxe_qp *qp,
663 struct rxe_pkt_info *req_pkt)
665 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
666 struct rxe_pkt_info ack_pkt;
669 enum resp_states state;
673 struct resp_res *res = qp->resp.res;
678 /* This is the first time we process that request. Get a
681 res = &qp->resp.resources[qp->resp.res_head];
683 free_rd_atomic_resource(qp, res);
684 rxe_advance_resp_resource(qp);
686 res->type = RXE_READ_MASK;
689 res->read.va = qp->resp.va;
690 res->read.va_org = qp->resp.va;
692 res->first_psn = req_pkt->psn;
694 if (reth_len(req_pkt)) {
695 res->last_psn = (req_pkt->psn +
696 (reth_len(req_pkt) + mtu - 1) /
697 mtu - 1) & BTH_PSN_MASK;
699 res->last_psn = res->first_psn;
701 res->cur_psn = req_pkt->psn;
703 res->read.resid = qp->resp.resid;
704 res->read.length = qp->resp.resid;
705 res->read.rkey = qp->resp.rkey;
707 /* note res inherits the reference to mr from qp */
708 res->read.mr = qp->resp.mr;
712 res->state = rdatm_res_state_new;
715 if (res->state == rdatm_res_state_new) {
716 if (res->read.resid <= mtu)
717 opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY;
719 opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST;
721 if (res->read.resid > mtu)
722 opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE;
724 opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST;
727 res->state = rdatm_res_state_next;
729 payload = min_t(int, res->read.resid, mtu);
731 skb = prepare_ack_packet(qp, req_pkt, &ack_pkt, opcode, payload,
732 res->cur_psn, AETH_ACK_UNLIMITED, &icrc);
734 return RESPST_ERR_RNR;
736 err = rxe_mem_copy(res->read.mr, res->read.va, payload_addr(&ack_pkt),
737 payload, from_mem_obj, &icrc);
739 pr_err("Failed copying memory\n");
741 if (bth_pad(&ack_pkt)) {
742 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
743 u8 *pad = payload_addr(&ack_pkt) + payload;
745 memset(pad, 0, bth_pad(&ack_pkt));
746 icrc = rxe_crc32(rxe, icrc, pad, bth_pad(&ack_pkt));
748 p = payload_addr(&ack_pkt) + payload + bth_pad(&ack_pkt);
751 err = rxe_xmit_packet(rxe, qp, &ack_pkt, skb);
753 pr_err("Failed sending RDMA reply.\n");
754 return RESPST_ERR_RNR;
757 res->read.va += payload;
758 res->read.resid -= payload;
759 res->cur_psn = (res->cur_psn + 1) & BTH_PSN_MASK;
761 if (res->read.resid > 0) {
766 qp->resp.opcode = -1;
767 if (psn_compare(res->cur_psn, qp->resp.psn) >= 0)
768 qp->resp.psn = res->cur_psn;
769 state = RESPST_CLEANUP;
775 static void build_rdma_network_hdr(union rdma_network_hdr *hdr,
776 struct rxe_pkt_info *pkt)
778 struct sk_buff *skb = PKT_TO_SKB(pkt);
780 memset(hdr, 0, sizeof(*hdr));
781 if (skb->protocol == htons(ETH_P_IP))
782 memcpy(&hdr->roce4grh, ip_hdr(skb), sizeof(hdr->roce4grh));
783 else if (skb->protocol == htons(ETH_P_IPV6))
784 memcpy(&hdr->ibgrh, ipv6_hdr(skb), sizeof(hdr->ibgrh));
787 /* Executes a new request. A retried request never reach that function (send
788 * and writes are discarded, and reads and atomics are retried elsewhere.
790 static enum resp_states execute(struct rxe_qp *qp, struct rxe_pkt_info *pkt)
792 enum resp_states err;
794 if (pkt->mask & RXE_SEND_MASK) {
795 if (qp_type(qp) == IB_QPT_UD ||
796 qp_type(qp) == IB_QPT_SMI ||
797 qp_type(qp) == IB_QPT_GSI) {
798 union rdma_network_hdr hdr;
800 build_rdma_network_hdr(&hdr, pkt);
802 err = send_data_in(qp, &hdr, sizeof(hdr));
806 err = send_data_in(qp, payload_addr(pkt), payload_size(pkt));
809 } else if (pkt->mask & RXE_WRITE_MASK) {
810 err = write_data_in(qp, pkt);
813 } else if (pkt->mask & RXE_READ_MASK) {
814 /* For RDMA Read we can increment the msn now. See C9-148. */
816 return RESPST_READ_REPLY;
817 } else if (pkt->mask & RXE_ATOMIC_MASK) {
818 err = process_atomic(qp, pkt);
826 /* next expected psn, read handles this separately */
827 qp->resp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
828 qp->resp.ack_psn = qp->resp.psn;
830 qp->resp.opcode = pkt->opcode;
831 qp->resp.status = IB_WC_SUCCESS;
833 if (pkt->mask & RXE_COMP_MASK) {
834 /* We successfully processed this new request. */
836 return RESPST_COMPLETE;
837 } else if (qp_type(qp) == IB_QPT_RC)
838 return RESPST_ACKNOWLEDGE;
840 return RESPST_CLEANUP;
843 static enum resp_states do_complete(struct rxe_qp *qp,
844 struct rxe_pkt_info *pkt)
847 struct ib_wc *wc = &cqe.ibwc;
848 struct ib_uverbs_wc *uwc = &cqe.uibwc;
849 struct rxe_recv_wqe *wqe = qp->resp.wqe;
852 return RESPST_CLEANUP;
854 memset(&cqe, 0, sizeof(cqe));
856 if (qp->rcq->is_user) {
857 uwc->status = qp->resp.status;
858 uwc->qp_num = qp->ibqp.qp_num;
859 uwc->wr_id = wqe->wr_id;
861 wc->status = qp->resp.status;
863 wc->wr_id = wqe->wr_id;
866 if (wc->status == IB_WC_SUCCESS) {
867 wc->opcode = (pkt->mask & RXE_IMMDT_MASK &&
868 pkt->mask & RXE_WRITE_MASK) ?
869 IB_WC_RECV_RDMA_WITH_IMM : IB_WC_RECV;
871 wc->byte_len = (pkt->mask & RXE_IMMDT_MASK &&
872 pkt->mask & RXE_WRITE_MASK) ?
873 qp->resp.length : wqe->dma.length - wqe->dma.resid;
875 /* fields after byte_len are different between kernel and user
878 if (qp->rcq->is_user) {
879 uwc->wc_flags = IB_WC_GRH;
881 if (pkt->mask & RXE_IMMDT_MASK) {
882 uwc->wc_flags |= IB_WC_WITH_IMM;
884 (__u32 __force)immdt_imm(pkt);
887 if (pkt->mask & RXE_IETH_MASK) {
888 uwc->wc_flags |= IB_WC_WITH_INVALIDATE;
889 uwc->ex.invalidate_rkey = ieth_rkey(pkt);
892 uwc->qp_num = qp->ibqp.qp_num;
894 if (pkt->mask & RXE_DETH_MASK)
895 uwc->src_qp = deth_sqp(pkt);
897 uwc->port_num = qp->attr.port_num;
899 struct sk_buff *skb = PKT_TO_SKB(pkt);
901 wc->wc_flags = IB_WC_GRH | IB_WC_WITH_NETWORK_HDR_TYPE;
902 if (skb->protocol == htons(ETH_P_IP))
903 wc->network_hdr_type = RDMA_NETWORK_IPV4;
905 wc->network_hdr_type = RDMA_NETWORK_IPV6;
907 if (pkt->mask & RXE_IMMDT_MASK) {
908 wc->wc_flags |= IB_WC_WITH_IMM;
909 wc->ex.imm_data = immdt_imm(pkt);
912 if (pkt->mask & RXE_IETH_MASK) {
913 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
916 wc->wc_flags |= IB_WC_WITH_INVALIDATE;
917 wc->ex.invalidate_rkey = ieth_rkey(pkt);
919 rmr = rxe_pool_get_index(&rxe->mr_pool,
920 wc->ex.invalidate_rkey >> 8);
921 if (unlikely(!rmr)) {
922 pr_err("Bad rkey %#x invalidation\n",
923 wc->ex.invalidate_rkey);
926 rmr->state = RXE_MEM_STATE_FREE;
932 if (pkt->mask & RXE_DETH_MASK)
933 wc->src_qp = deth_sqp(pkt);
935 wc->port_num = qp->attr.port_num;
939 /* have copy for srq and reference for !srq */
941 advance_consumer(qp->rq.queue);
945 if (rxe_cq_post(qp->rcq, &cqe, pkt ? bth_se(pkt) : 1))
946 return RESPST_ERR_CQ_OVERFLOW;
948 if (qp->resp.state == QP_STATE_ERROR)
949 return RESPST_CHK_RESOURCE;
953 else if (qp_type(qp) == IB_QPT_RC)
954 return RESPST_ACKNOWLEDGE;
956 return RESPST_CLEANUP;
959 static int send_ack(struct rxe_qp *qp, struct rxe_pkt_info *pkt,
960 u8 syndrome, u32 psn)
963 struct rxe_pkt_info ack_pkt;
965 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
967 skb = prepare_ack_packet(qp, pkt, &ack_pkt, IB_OPCODE_RC_ACKNOWLEDGE,
968 0, psn, syndrome, NULL);
974 err = rxe_xmit_packet(rxe, qp, &ack_pkt, skb);
976 pr_err_ratelimited("Failed sending ack\n");
982 static int send_atomic_ack(struct rxe_qp *qp, struct rxe_pkt_info *pkt,
986 struct rxe_pkt_info ack_pkt;
988 struct sk_buff *skb_copy;
989 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
990 struct resp_res *res;
992 skb = prepare_ack_packet(qp, pkt, &ack_pkt,
993 IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE, 0, pkt->psn,
1000 skb_copy = skb_clone(skb, GFP_ATOMIC);
1002 rxe_add_ref(qp); /* for the new SKB */
1004 pr_warn("Could not clone atomic response\n");
1009 res = &qp->resp.resources[qp->resp.res_head];
1010 free_rd_atomic_resource(qp, res);
1011 rxe_advance_resp_resource(qp);
1013 memcpy(SKB_TO_PKT(skb), &ack_pkt, sizeof(ack_pkt));
1014 memset((unsigned char *)SKB_TO_PKT(skb) + sizeof(ack_pkt), 0,
1015 sizeof(skb->cb) - sizeof(ack_pkt));
1017 res->type = RXE_ATOMIC_MASK;
1018 res->atomic.skb = skb;
1019 res->first_psn = ack_pkt.psn;
1020 res->last_psn = ack_pkt.psn;
1021 res->cur_psn = ack_pkt.psn;
1023 rc = rxe_xmit_packet(rxe, qp, &ack_pkt, skb_copy);
1025 pr_err_ratelimited("Failed sending ack\n");
1027 kfree_skb(skb_copy);
1034 static enum resp_states acknowledge(struct rxe_qp *qp,
1035 struct rxe_pkt_info *pkt)
1037 if (qp_type(qp) != IB_QPT_RC)
1038 return RESPST_CLEANUP;
1040 if (qp->resp.aeth_syndrome != AETH_ACK_UNLIMITED)
1041 send_ack(qp, pkt, qp->resp.aeth_syndrome, pkt->psn);
1042 else if (pkt->mask & RXE_ATOMIC_MASK)
1043 send_atomic_ack(qp, pkt, AETH_ACK_UNLIMITED);
1044 else if (bth_ack(pkt))
1045 send_ack(qp, pkt, AETH_ACK_UNLIMITED, pkt->psn);
1047 return RESPST_CLEANUP;
1050 static enum resp_states cleanup(struct rxe_qp *qp,
1051 struct rxe_pkt_info *pkt)
1053 struct sk_buff *skb;
1056 skb = skb_dequeue(&qp->req_pkts);
1062 rxe_drop_ref(qp->resp.mr);
1069 static struct resp_res *find_resource(struct rxe_qp *qp, u32 psn)
1073 for (i = 0; i < qp->attr.max_dest_rd_atomic; i++) {
1074 struct resp_res *res = &qp->resp.resources[i];
1079 if (psn_compare(psn, res->first_psn) >= 0 &&
1080 psn_compare(psn, res->last_psn) <= 0) {
1088 static enum resp_states duplicate_request(struct rxe_qp *qp,
1089 struct rxe_pkt_info *pkt)
1091 enum resp_states rc;
1092 u32 prev_psn = (qp->resp.ack_psn - 1) & BTH_PSN_MASK;
1094 if (pkt->mask & RXE_SEND_MASK ||
1095 pkt->mask & RXE_WRITE_MASK) {
1096 /* SEND. Ack again and cleanup. C9-105. */
1098 send_ack(qp, pkt, AETH_ACK_UNLIMITED, prev_psn);
1099 rc = RESPST_CLEANUP;
1101 } else if (pkt->mask & RXE_READ_MASK) {
1102 struct resp_res *res;
1104 res = find_resource(qp, pkt->psn);
1106 /* Resource not found. Class D error. Drop the
1109 rc = RESPST_CLEANUP;
1112 /* Ensure this new request is the same as the previous
1113 * one or a subset of it.
1115 u64 iova = reth_va(pkt);
1116 u32 resid = reth_len(pkt);
1118 if (iova < res->read.va_org ||
1119 resid > res->read.length ||
1120 (iova + resid) > (res->read.va_org +
1121 res->read.length)) {
1122 rc = RESPST_CLEANUP;
1126 if (reth_rkey(pkt) != res->read.rkey) {
1127 rc = RESPST_CLEANUP;
1131 res->cur_psn = pkt->psn;
1132 res->state = (pkt->psn == res->first_psn) ?
1133 rdatm_res_state_new :
1134 rdatm_res_state_replay;
1137 /* Reset the resource, except length. */
1138 res->read.va_org = iova;
1139 res->read.va = iova;
1140 res->read.resid = resid;
1142 /* Replay the RDMA read reply. */
1144 rc = RESPST_READ_REPLY;
1148 struct resp_res *res;
1150 /* Find the operation in our list of responder resources. */
1151 res = find_resource(qp, pkt->psn);
1153 struct sk_buff *skb_copy;
1155 skb_copy = skb_clone(res->atomic.skb, GFP_ATOMIC);
1157 rxe_add_ref(qp); /* for the new SKB */
1159 pr_warn("Couldn't clone atomic resp\n");
1160 rc = RESPST_CLEANUP;
1164 /* Resend the result. */
1165 rc = rxe_xmit_packet(to_rdev(qp->ibqp.device), qp,
1168 pr_err("Failed resending result. This flow is not handled - skb ignored\n");
1170 rc = RESPST_CLEANUP;
1175 /* Resource not found. Class D error. Drop the request. */
1176 rc = RESPST_CLEANUP;
1183 /* Process a class A or C. Both are treated the same in this implementation. */
1184 static void do_class_ac_error(struct rxe_qp *qp, u8 syndrome,
1185 enum ib_wc_status status)
1187 qp->resp.aeth_syndrome = syndrome;
1188 qp->resp.status = status;
1190 /* indicate that we should go through the ERROR state */
1191 qp->resp.goto_error = 1;
1194 static enum resp_states do_class_d1e_error(struct rxe_qp *qp)
1199 qp->resp.drop_msg = 1;
1201 qp->resp.status = IB_WC_REM_INV_REQ_ERR;
1202 return RESPST_COMPLETE;
1204 return RESPST_CLEANUP;
1207 /* Class D1. This packet may be the start of a
1208 * new message and could be valid. The previous
1209 * message is invalid and ignored. reset the
1210 * recv wr to its original state
1213 qp->resp.wqe->dma.resid = qp->resp.wqe->dma.length;
1214 qp->resp.wqe->dma.cur_sge = 0;
1215 qp->resp.wqe->dma.sge_offset = 0;
1216 qp->resp.opcode = -1;
1220 rxe_drop_ref(qp->resp.mr);
1224 return RESPST_CLEANUP;
1228 static void rxe_drain_req_pkts(struct rxe_qp *qp, bool notify)
1230 struct sk_buff *skb;
1232 while ((skb = skb_dequeue(&qp->req_pkts))) {
1240 while (!qp->srq && qp->rq.queue && queue_head(qp->rq.queue))
1241 advance_consumer(qp->rq.queue);
1244 int rxe_responder(void *arg)
1246 struct rxe_qp *qp = (struct rxe_qp *)arg;
1247 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
1248 enum resp_states state;
1249 struct rxe_pkt_info *pkt = NULL;
1254 qp->resp.aeth_syndrome = AETH_ACK_UNLIMITED;
1261 switch (qp->resp.state) {
1262 case QP_STATE_RESET:
1263 state = RESPST_RESET;
1267 state = RESPST_GET_REQ;
1272 pr_debug("qp#%d state = %s\n", qp_num(qp),
1273 resp_state_name[state]);
1275 case RESPST_GET_REQ:
1276 state = get_req(qp, &pkt);
1278 case RESPST_CHK_PSN:
1279 state = check_psn(qp, pkt);
1281 case RESPST_CHK_OP_SEQ:
1282 state = check_op_seq(qp, pkt);
1284 case RESPST_CHK_OP_VALID:
1285 state = check_op_valid(qp, pkt);
1287 case RESPST_CHK_RESOURCE:
1288 state = check_resource(qp, pkt);
1290 case RESPST_CHK_LENGTH:
1291 state = check_length(qp, pkt);
1293 case RESPST_CHK_RKEY:
1294 state = check_rkey(qp, pkt);
1296 case RESPST_EXECUTE:
1297 state = execute(qp, pkt);
1299 case RESPST_COMPLETE:
1300 state = do_complete(qp, pkt);
1302 case RESPST_READ_REPLY:
1303 state = read_reply(qp, pkt);
1305 case RESPST_ACKNOWLEDGE:
1306 state = acknowledge(qp, pkt);
1308 case RESPST_CLEANUP:
1309 state = cleanup(qp, pkt);
1311 case RESPST_DUPLICATE_REQUEST:
1312 state = duplicate_request(qp, pkt);
1314 case RESPST_ERR_PSN_OUT_OF_SEQ:
1315 /* RC only - Class B. Drop packet. */
1316 send_ack(qp, pkt, AETH_NAK_PSN_SEQ_ERROR, qp->resp.psn);
1317 state = RESPST_CLEANUP;
1320 case RESPST_ERR_TOO_MANY_RDMA_ATM_REQ:
1321 case RESPST_ERR_MISSING_OPCODE_FIRST:
1322 case RESPST_ERR_MISSING_OPCODE_LAST_C:
1323 case RESPST_ERR_UNSUPPORTED_OPCODE:
1324 case RESPST_ERR_MISALIGNED_ATOMIC:
1325 /* RC Only - Class C. */
1326 do_class_ac_error(qp, AETH_NAK_INVALID_REQ,
1327 IB_WC_REM_INV_REQ_ERR);
1328 state = RESPST_COMPLETE;
1331 case RESPST_ERR_MISSING_OPCODE_LAST_D1E:
1332 state = do_class_d1e_error(qp);
1334 case RESPST_ERR_RNR:
1335 if (qp_type(qp) == IB_QPT_RC) {
1336 rxe_counter_inc(rxe, RXE_CNT_SND_RNR);
1338 send_ack(qp, pkt, AETH_RNR_NAK |
1340 qp->attr.min_rnr_timer),
1343 /* UD/UC - class D */
1344 qp->resp.drop_msg = 1;
1346 state = RESPST_CLEANUP;
1349 case RESPST_ERR_RKEY_VIOLATION:
1350 if (qp_type(qp) == IB_QPT_RC) {
1352 do_class_ac_error(qp, AETH_NAK_REM_ACC_ERR,
1353 IB_WC_REM_ACCESS_ERR);
1354 state = RESPST_COMPLETE;
1356 qp->resp.drop_msg = 1;
1358 /* UC/SRQ Class D */
1359 qp->resp.status = IB_WC_REM_ACCESS_ERR;
1360 state = RESPST_COMPLETE;
1362 /* UC/non-SRQ Class E. */
1363 state = RESPST_CLEANUP;
1368 case RESPST_ERR_LENGTH:
1369 if (qp_type(qp) == IB_QPT_RC) {
1371 do_class_ac_error(qp, AETH_NAK_INVALID_REQ,
1372 IB_WC_REM_INV_REQ_ERR);
1373 state = RESPST_COMPLETE;
1374 } else if (qp->srq) {
1375 /* UC/UD - class E */
1376 qp->resp.status = IB_WC_REM_INV_REQ_ERR;
1377 state = RESPST_COMPLETE;
1379 /* UC/UD - class D */
1380 qp->resp.drop_msg = 1;
1381 state = RESPST_CLEANUP;
1385 case RESPST_ERR_MALFORMED_WQE:
1387 do_class_ac_error(qp, AETH_NAK_REM_OP_ERR,
1388 IB_WC_LOC_QP_OP_ERR);
1389 state = RESPST_COMPLETE;
1392 case RESPST_ERR_CQ_OVERFLOW:
1394 state = RESPST_ERROR;
1398 if (qp->resp.goto_error) {
1399 state = RESPST_ERROR;
1406 if (qp->resp.goto_error) {
1407 state = RESPST_ERROR;
1414 rxe_drain_req_pkts(qp, false);
1415 qp->resp.wqe = NULL;
1419 qp->resp.goto_error = 0;
1420 pr_warn("qp#%d moved to error state\n", qp_num(qp));