?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
WIP FPC-III support
[linux/fpc-iii.git] / drivers / infiniband / hw / hfi1 / rc.c
blob1bb5f57152d3525ebd1a30214e41039dcd99dac0
1 /*
2 * Copyright(c) 2015 - 2018 Intel Corporation.
3 *
4 * This file is provided under a dual BSD/GPLv2 license. When using or
5 * redistributing this file, you may do so under either license.
6 *
7 * GPL LICENSE SUMMARY
8 *
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of version 2 of the GNU General Public License as
11 * published by the Free Software Foundation.
12 *
13 * This program is distributed in the hope that it will be useful, but
14 * WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * General Public License for more details.
17 *
18 * BSD LICENSE
19 *
20 * Redistribution and use in source and binary forms, with or without
21 * modification, are permitted provided that the following conditions
22 * are met:
23 *
24 * - Redistributions of source code must retain the above copyright
25 * notice, this list of conditions and the following disclaimer.
26 * - Redistributions in binary form must reproduce the above copyright
27 * notice, this list of conditions and the following disclaimer in
28 * the documentation and/or other materials provided with the
29 * distribution.
30 * - Neither the name of Intel Corporation nor the names of its
31 * contributors may be used to endorse or promote products derived
32 * from this software without specific prior written permission.
33 *
34 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
45 *
46 */
47
48 #include <linux/io.h>
49 #include <rdma/rdma_vt.h>
50 #include <rdma/rdmavt_qp.h>
51
52 #include "hfi.h"
53 #include "qp.h"
54 #include "rc.h"
55 #include "verbs_txreq.h"
56 #include "trace.h"
57
58 struct rvt_ack_entry *find_prev_entry(struct rvt_qp *qp, u32 psn, u8 *prev,
59 u8 *prev_ack, bool *scheduled)
60 __must_hold(&qp->s_lock)
61 {
62 struct rvt_ack_entry *e = NULL;
63 u8 i, p;
64 bool s = true;
65
66 for (i = qp->r_head_ack_queue; ; i = p) {
67 if (i == qp->s_tail_ack_queue)
68 s = false;
69 if (i)
70 p = i - 1;
71 else
72 p = rvt_size_atomic(ib_to_rvt(qp->ibqp.device));
73 if (p == qp->r_head_ack_queue) {
74 e = NULL;
75 break;
76 }
77 e = &qp->s_ack_queue[p];
78 if (!e->opcode) {
79 e = NULL;
80 break;
81 }
82 if (cmp_psn(psn, e->psn) >= 0) {
83 if (p == qp->s_tail_ack_queue &&
84 cmp_psn(psn, e->lpsn) <= 0)
85 s = false;
86 break;
87 }
88 }
89 if (prev)
90 *prev = p;
91 if (prev_ack)
92 *prev_ack = i;
93 if (scheduled)
94 *scheduled = s;
95 return e;
96 }
97
98 /**
99 * make_rc_ack - construct a response packet (ACK, NAK, or RDMA read)
100 * @dev: the device for this QP
101 * @qp: a pointer to the QP
102 * @ohdr: a pointer to the IB header being constructed
103 * @ps: the xmit packet state
104 *
105 * Return 1 if constructed; otherwise, return 0.
106 * Note that we are in the responder's side of the QP context.
107 * Note the QP s_lock must be held.
108 */
109 static int make_rc_ack(struct hfi1_ibdev *dev, struct rvt_qp *qp,
110 struct ib_other_headers *ohdr,
111 struct hfi1_pkt_state *ps)
112 {
113 struct rvt_ack_entry *e;
114 u32 hwords, hdrlen;
115 u32 len = 0;
116 u32 bth0 = 0, bth2 = 0;
117 u32 bth1 = qp->remote_qpn | (HFI1_CAP_IS_KSET(OPFN) << IB_BTHE_E_SHIFT);
118 int middle = 0;
119 u32 pmtu = qp->pmtu;
120 struct hfi1_qp_priv *qpriv = qp->priv;
121 bool last_pkt;
122 u32 delta;
123 u8 next = qp->s_tail_ack_queue;
124 struct tid_rdma_request *req;
125
126 trace_hfi1_rsp_make_rc_ack(qp, 0);
127 lockdep_assert_held(&qp->s_lock);
128 /* Don't send an ACK if we aren't supposed to. */
129 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
130 goto bail;
131
132 if (qpriv->hdr_type == HFI1_PKT_TYPE_9B)
133 /* header size in 32-bit words LRH+BTH = (8+12)/4. */
134 hwords = 5;
135 else
136 /* header size in 32-bit words 16B LRH+BTH = (16+12)/4. */
137 hwords = 7;
138
139 switch (qp->s_ack_state) {
140 case OP(RDMA_READ_RESPONSE_LAST):
141 case OP(RDMA_READ_RESPONSE_ONLY):
142 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
143 release_rdma_sge_mr(e);
144 fallthrough;
145 case OP(ATOMIC_ACKNOWLEDGE):
146 /*
147 * We can increment the tail pointer now that the last
148 * response has been sent instead of only being
149 * constructed.
150 */
151 if (++next > rvt_size_atomic(&dev->rdi))
152 next = 0;
153 /*
154 * Only advance the s_acked_ack_queue pointer if there
155 * have been no TID RDMA requests.
156 */
157 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
158 if (e->opcode != TID_OP(WRITE_REQ) &&
159 qp->s_acked_ack_queue == qp->s_tail_ack_queue)
160 qp->s_acked_ack_queue = next;
161 qp->s_tail_ack_queue = next;
162 trace_hfi1_rsp_make_rc_ack(qp, e->psn);
163 fallthrough;
164 case OP(SEND_ONLY):
165 case OP(ACKNOWLEDGE):
166 /* Check for no next entry in the queue. */
167 if (qp->r_head_ack_queue == qp->s_tail_ack_queue) {
168 if (qp->s_flags & RVT_S_ACK_PENDING)
169 goto normal;
170 goto bail;
171 }
172
173 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
174 /* Check for tid write fence */
175 if ((qpriv->s_flags & HFI1_R_TID_WAIT_INTERLCK) ||
176 hfi1_tid_rdma_ack_interlock(qp, e)) {
177 iowait_set_flag(&qpriv->s_iowait, IOWAIT_PENDING_IB);
178 goto bail;
179 }
180 if (e->opcode == OP(RDMA_READ_REQUEST)) {
181 /*
182 * If a RDMA read response is being resent and
183 * we haven't seen the duplicate request yet,
184 * then stop sending the remaining responses the
185 * responder has seen until the requester re-sends it.
186 */
187 len = e->rdma_sge.sge_length;
188 if (len && !e->rdma_sge.mr) {
189 if (qp->s_acked_ack_queue ==
190 qp->s_tail_ack_queue)
191 qp->s_acked_ack_queue =
192 qp->r_head_ack_queue;
193 qp->s_tail_ack_queue = qp->r_head_ack_queue;
194 goto bail;
195 }
196 /* Copy SGE state in case we need to resend */
197 ps->s_txreq->mr = e->rdma_sge.mr;
198 if (ps->s_txreq->mr)
199 rvt_get_mr(ps->s_txreq->mr);
200 qp->s_ack_rdma_sge.sge = e->rdma_sge;
201 qp->s_ack_rdma_sge.num_sge = 1;
202 ps->s_txreq->ss = &qp->s_ack_rdma_sge;
203 if (len > pmtu) {
204 len = pmtu;
205 qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST);
206 } else {
207 qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY);
208 e->sent = 1;
209 }
210 ohdr->u.aeth = rvt_compute_aeth(qp);
211 hwords++;
212 qp->s_ack_rdma_psn = e->psn;
213 bth2 = mask_psn(qp->s_ack_rdma_psn++);
214 } else if (e->opcode == TID_OP(WRITE_REQ)) {
215 /*
216 * If a TID RDMA WRITE RESP is being resent, we have to
217 * wait for the actual request. All requests that are to
218 * be resent will have their state set to
219 * TID_REQUEST_RESEND. When the new request arrives, the
220 * state will be changed to TID_REQUEST_RESEND_ACTIVE.
221 */
222 req = ack_to_tid_req(e);
223 if (req->state == TID_REQUEST_RESEND ||
224 req->state == TID_REQUEST_INIT_RESEND)
225 goto bail;
226 qp->s_ack_state = TID_OP(WRITE_RESP);
227 qp->s_ack_rdma_psn = mask_psn(e->psn + req->cur_seg);
228 goto write_resp;
229 } else if (e->opcode == TID_OP(READ_REQ)) {
230 /*
231 * If a TID RDMA read response is being resent and
232 * we haven't seen the duplicate request yet,
233 * then stop sending the remaining responses the
234 * responder has seen until the requester re-sends it.
235 */
236 len = e->rdma_sge.sge_length;
237 if (len && !e->rdma_sge.mr) {
238 if (qp->s_acked_ack_queue ==
239 qp->s_tail_ack_queue)
240 qp->s_acked_ack_queue =
241 qp->r_head_ack_queue;
242 qp->s_tail_ack_queue = qp->r_head_ack_queue;
243 goto bail;
244 }
245 /* Copy SGE state in case we need to resend */
246 ps->s_txreq->mr = e->rdma_sge.mr;
247 if (ps->s_txreq->mr)
248 rvt_get_mr(ps->s_txreq->mr);
249 qp->s_ack_rdma_sge.sge = e->rdma_sge;
250 qp->s_ack_rdma_sge.num_sge = 1;
251 qp->s_ack_state = TID_OP(READ_RESP);
252 goto read_resp;
253 } else {
254 /* COMPARE_SWAP or FETCH_ADD */
255 ps->s_txreq->ss = NULL;
256 len = 0;
257 qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE);
258 ohdr->u.at.aeth = rvt_compute_aeth(qp);
259 ib_u64_put(e->atomic_data, &ohdr->u.at.atomic_ack_eth);
260 hwords += sizeof(ohdr->u.at) / sizeof(u32);
261 bth2 = mask_psn(e->psn);
262 e->sent = 1;
263 }
264 trace_hfi1_tid_write_rsp_make_rc_ack(qp);
265 bth0 = qp->s_ack_state << 24;
266 break;
267
268 case OP(RDMA_READ_RESPONSE_FIRST):
269 qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE);
270 fallthrough;
271 case OP(RDMA_READ_RESPONSE_MIDDLE):
272 ps->s_txreq->ss = &qp->s_ack_rdma_sge;
273 ps->s_txreq->mr = qp->s_ack_rdma_sge.sge.mr;
274 if (ps->s_txreq->mr)
275 rvt_get_mr(ps->s_txreq->mr);
276 len = qp->s_ack_rdma_sge.sge.sge_length;
277 if (len > pmtu) {
278 len = pmtu;
279 middle = HFI1_CAP_IS_KSET(SDMA_AHG);
280 } else {
281 ohdr->u.aeth = rvt_compute_aeth(qp);
282 hwords++;
283 qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
284 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
285 e->sent = 1;
286 }
287 bth0 = qp->s_ack_state << 24;
288 bth2 = mask_psn(qp->s_ack_rdma_psn++);
289 break;
290
291 case TID_OP(WRITE_RESP):
292 write_resp:
293 /*
294 * 1. Check if RVT_S_ACK_PENDING is set. If yes,
295 * goto normal.
296 * 2. Attempt to allocate TID resources.
297 * 3. Remove RVT_S_RESP_PENDING flags from s_flags
298 * 4. If resources not available:
299 * 4.1 Set RVT_S_WAIT_TID_SPACE
300 * 4.2 Queue QP on RCD TID queue
301 * 4.3 Put QP on iowait list.
302 * 4.4 Build IB RNR NAK with appropriate timeout value
303 * 4.5 Return indication progress made.
304 * 5. If resources are available:
305 * 5.1 Program HW flow CSRs
306 * 5.2 Build TID RDMA WRITE RESP packet
307 * 5.3 If more resources needed, do 2.1 - 2.3.
308 * 5.4 Wake up next QP on RCD TID queue.
309 * 5.5 Return indication progress made.
310 */
311
312 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
313 req = ack_to_tid_req(e);
314
315 /*
316 * Send scheduled RNR NAK's. RNR NAK's need to be sent at
317 * segment boundaries, not at request boundaries. Don't change
318 * s_ack_state because we are still in the middle of a request
319 */
320 if (qpriv->rnr_nak_state == TID_RNR_NAK_SEND &&
321 qp->s_tail_ack_queue == qpriv->r_tid_alloc &&
322 req->cur_seg == req->alloc_seg) {
323 qpriv->rnr_nak_state = TID_RNR_NAK_SENT;
324 goto normal_no_state;
325 }
326
327 bth2 = mask_psn(qp->s_ack_rdma_psn);
328 hdrlen = hfi1_build_tid_rdma_write_resp(qp, e, ohdr, &bth1,
329 bth2, &len,
330 &ps->s_txreq->ss);
331 if (!hdrlen)
332 return 0;
333
334 hwords += hdrlen;
335 bth0 = qp->s_ack_state << 24;
336 qp->s_ack_rdma_psn++;
337 trace_hfi1_tid_req_make_rc_ack_write(qp, 0, e->opcode, e->psn,
338 e->lpsn, req);
339 if (req->cur_seg != req->total_segs)
340 break;
341
342 e->sent = 1;
343 /* Do not free e->rdma_sge until all data are received */
344 qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE);
345 break;
346
347 case TID_OP(READ_RESP):
348 read_resp:
349 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
350 ps->s_txreq->ss = &qp->s_ack_rdma_sge;
351 delta = hfi1_build_tid_rdma_read_resp(qp, e, ohdr, &bth0,
352 &bth1, &bth2, &len,
353 &last_pkt);
354 if (delta == 0)
355 goto error_qp;
356 hwords += delta;
357 if (last_pkt) {
358 e->sent = 1;
359 /*
360 * Increment qp->s_tail_ack_queue through s_ack_state
361 * transition.
362 */
363 qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
364 }
365 break;
366 case TID_OP(READ_REQ):
367 goto bail;
368
369 default:
370 normal:
371 /*
372 * Send a regular ACK.
373 * Set the s_ack_state so we wait until after sending
374 * the ACK before setting s_ack_state to ACKNOWLEDGE
375 * (see above).
376 */
377 qp->s_ack_state = OP(SEND_ONLY);
378 normal_no_state:
379 if (qp->s_nak_state)
380 ohdr->u.aeth =
381 cpu_to_be32((qp->r_msn & IB_MSN_MASK) |
382 (qp->s_nak_state <<
383 IB_AETH_CREDIT_SHIFT));
384 else
385 ohdr->u.aeth = rvt_compute_aeth(qp);
386 hwords++;
387 len = 0;
388 bth0 = OP(ACKNOWLEDGE) << 24;
389 bth2 = mask_psn(qp->s_ack_psn);
390 qp->s_flags &= ~RVT_S_ACK_PENDING;
391 ps->s_txreq->txreq.flags |= SDMA_TXREQ_F_VIP;
392 ps->s_txreq->ss = NULL;
393 }
394 qp->s_rdma_ack_cnt++;
395 ps->s_txreq->sde = qpriv->s_sde;
396 ps->s_txreq->s_cur_size = len;
397 ps->s_txreq->hdr_dwords = hwords;
398 hfi1_make_ruc_header(qp, ohdr, bth0, bth1, bth2, middle, ps);
399 return 1;
400 error_qp:
401 spin_unlock_irqrestore(&qp->s_lock, ps->flags);
402 spin_lock_irqsave(&qp->r_lock, ps->flags);
403 spin_lock(&qp->s_lock);
404 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
405 spin_unlock(&qp->s_lock);
406 spin_unlock_irqrestore(&qp->r_lock, ps->flags);
407 spin_lock_irqsave(&qp->s_lock, ps->flags);
408 bail:
409 qp->s_ack_state = OP(ACKNOWLEDGE);
410 /*
411 * Ensure s_rdma_ack_cnt changes are committed prior to resetting
412 * RVT_S_RESP_PENDING
413 */
414 smp_wmb();
415 qp->s_flags &= ~(RVT_S_RESP_PENDING
416 | RVT_S_ACK_PENDING
417 | HFI1_S_AHG_VALID);
418 return 0;
419 }
420
421 /**
422 * hfi1_make_rc_req - construct a request packet (SEND, RDMA r/w, ATOMIC)
423 * @qp: a pointer to the QP
424 *
425 * Assumes s_lock is held.
426 *
427 * Return 1 if constructed; otherwise, return 0.
428 */
429 int hfi1_make_rc_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
430 {
431 struct hfi1_qp_priv *priv = qp->priv;
432 struct hfi1_ibdev *dev = to_idev(qp->ibqp.device);
433 struct ib_other_headers *ohdr;
434 struct rvt_sge_state *ss = NULL;
435 struct rvt_swqe *wqe;
436 struct hfi1_swqe_priv *wpriv;
437 struct tid_rdma_request *req = NULL;
438 /* header size in 32-bit words LRH+BTH = (8+12)/4. */
439 u32 hwords = 5;
440 u32 len = 0;
441 u32 bth0 = 0, bth2 = 0;
442 u32 bth1 = qp->remote_qpn | (HFI1_CAP_IS_KSET(OPFN) << IB_BTHE_E_SHIFT);
443 u32 pmtu = qp->pmtu;
444 char newreq;
445 int middle = 0;
446 int delta;
447 struct tid_rdma_flow *flow = NULL;
448 struct tid_rdma_params *remote;
449
450 trace_hfi1_sender_make_rc_req(qp);
451 lockdep_assert_held(&qp->s_lock);
452 ps->s_txreq = get_txreq(ps->dev, qp);
453 if (!ps->s_txreq)
454 goto bail_no_tx;
455
456 if (priv->hdr_type == HFI1_PKT_TYPE_9B) {
457 /* header size in 32-bit words LRH+BTH = (8+12)/4. */
458 hwords = 5;
459 if (rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH)
460 ohdr = &ps->s_txreq->phdr.hdr.ibh.u.l.oth;
461 else
462 ohdr = &ps->s_txreq->phdr.hdr.ibh.u.oth;
463 } else {
464 /* header size in 32-bit words 16B LRH+BTH = (16+12)/4. */
465 hwords = 7;
466 if ((rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH) &&
467 (hfi1_check_mcast(rdma_ah_get_dlid(&qp->remote_ah_attr))))
468 ohdr = &ps->s_txreq->phdr.hdr.opah.u.l.oth;
469 else
470 ohdr = &ps->s_txreq->phdr.hdr.opah.u.oth;
471 }
472
473 /* Sending responses has higher priority over sending requests. */
474 if ((qp->s_flags & RVT_S_RESP_PENDING) &&
475 make_rc_ack(dev, qp, ohdr, ps))
476 return 1;
477
478 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_SEND_OK)) {
479 if (!(ib_rvt_state_ops[qp->state] & RVT_FLUSH_SEND))
480 goto bail;
481 /* We are in the error state, flush the work request. */
482 if (qp->s_last == READ_ONCE(qp->s_head))
483 goto bail;
484 /* If DMAs are in progress, we can't flush immediately. */
485 if (iowait_sdma_pending(&priv->s_iowait)) {
486 qp->s_flags |= RVT_S_WAIT_DMA;
487 goto bail;
488 }
489 clear_ahg(qp);
490 wqe = rvt_get_swqe_ptr(qp, qp->s_last);
491 hfi1_trdma_send_complete(qp, wqe, qp->s_last != qp->s_acked ?
492 IB_WC_SUCCESS : IB_WC_WR_FLUSH_ERR);
493 /* will get called again */
494 goto done_free_tx;
495 }
496
497 if (qp->s_flags & (RVT_S_WAIT_RNR | RVT_S_WAIT_ACK | HFI1_S_WAIT_HALT))
498 goto bail;
499
500 if (cmp_psn(qp->s_psn, qp->s_sending_hpsn) <= 0) {
501 if (cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) <= 0) {
502 qp->s_flags |= RVT_S_WAIT_PSN;
503 goto bail;
504 }
505 qp->s_sending_psn = qp->s_psn;
506 qp->s_sending_hpsn = qp->s_psn - 1;
507 }
508
509 /* Send a request. */
510 wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
511 check_s_state:
512 switch (qp->s_state) {
513 default:
514 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_NEXT_SEND_OK))
515 goto bail;
516 /*
517 * Resend an old request or start a new one.
518 *
519 * We keep track of the current SWQE so that
520 * we don't reset the "furthest progress" state
521 * if we need to back up.
522 */
523 newreq = 0;
524 if (qp->s_cur == qp->s_tail) {
525 /* Check if send work queue is empty. */
526 if (qp->s_tail == READ_ONCE(qp->s_head)) {
527 clear_ahg(qp);
528 goto bail;
529 }
530 /*
531 * If a fence is requested, wait for previous
532 * RDMA read and atomic operations to finish.
533 * However, there is no need to guard against
534 * TID RDMA READ after TID RDMA READ.
535 */
536 if ((wqe->wr.send_flags & IB_SEND_FENCE) &&
537 qp->s_num_rd_atomic &&
538 (wqe->wr.opcode != IB_WR_TID_RDMA_READ ||
539 priv->pending_tid_r_segs < qp->s_num_rd_atomic)) {
540 qp->s_flags |= RVT_S_WAIT_FENCE;
541 goto bail;
542 }
543 /*
544 * Local operations are processed immediately
545 * after all prior requests have completed
546 */
547 if (wqe->wr.opcode == IB_WR_REG_MR ||
548 wqe->wr.opcode == IB_WR_LOCAL_INV) {
549 int local_ops = 0;
550 int err = 0;
551
552 if (qp->s_last != qp->s_cur)
553 goto bail;
554 if (++qp->s_cur == qp->s_size)
555 qp->s_cur = 0;
556 if (++qp->s_tail == qp->s_size)
557 qp->s_tail = 0;
558 if (!(wqe->wr.send_flags &
559 RVT_SEND_COMPLETION_ONLY)) {
560 err = rvt_invalidate_rkey(
561 qp,
562 wqe->wr.ex.invalidate_rkey);
563 local_ops = 1;
564 }
565 rvt_send_complete(qp, wqe,
566 err ? IB_WC_LOC_PROT_ERR
567 : IB_WC_SUCCESS);
568 if (local_ops)
569 atomic_dec(&qp->local_ops_pending);
570 goto done_free_tx;
571 }
572
573 newreq = 1;
574 qp->s_psn = wqe->psn;
575 }
576 /*
577 * Note that we have to be careful not to modify the
578 * original work request since we may need to resend
579 * it.
580 */
581 len = wqe->length;
582 ss = &qp->s_sge;
583 bth2 = mask_psn(qp->s_psn);
584
585 /*
586 * Interlock between various IB requests and TID RDMA
587 * if necessary.
588 */
589 if ((priv->s_flags & HFI1_S_TID_WAIT_INTERLCK) ||
590 hfi1_tid_rdma_wqe_interlock(qp, wqe))
591 goto bail;
592
593 switch (wqe->wr.opcode) {
594 case IB_WR_SEND:
595 case IB_WR_SEND_WITH_IMM:
596 case IB_WR_SEND_WITH_INV:
597 /* If no credit, return. */
598 if (!rvt_rc_credit_avail(qp, wqe))
599 goto bail;
600 if (len > pmtu) {
601 qp->s_state = OP(SEND_FIRST);
602 len = pmtu;
603 break;
604 }
605 if (wqe->wr.opcode == IB_WR_SEND) {
606 qp->s_state = OP(SEND_ONLY);
607 } else if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
608 qp->s_state = OP(SEND_ONLY_WITH_IMMEDIATE);
609 /* Immediate data comes after the BTH */
610 ohdr->u.imm_data = wqe->wr.ex.imm_data;
611 hwords += 1;
612 } else {
613 qp->s_state = OP(SEND_ONLY_WITH_INVALIDATE);
614 /* Invalidate rkey comes after the BTH */
615 ohdr->u.ieth = cpu_to_be32(
616 wqe->wr.ex.invalidate_rkey);
617 hwords += 1;
618 }
619 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
620 bth0 |= IB_BTH_SOLICITED;
621 bth2 |= IB_BTH_REQ_ACK;
622 if (++qp->s_cur == qp->s_size)
623 qp->s_cur = 0;
624 break;
625
626 case IB_WR_RDMA_WRITE:
627 if (newreq && !(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
628 qp->s_lsn++;
629 goto no_flow_control;
630 case IB_WR_RDMA_WRITE_WITH_IMM:
631 /* If no credit, return. */
632 if (!rvt_rc_credit_avail(qp, wqe))
633 goto bail;
634 no_flow_control:
635 put_ib_reth_vaddr(
636 wqe->rdma_wr.remote_addr,
637 &ohdr->u.rc.reth);
638 ohdr->u.rc.reth.rkey =
639 cpu_to_be32(wqe->rdma_wr.rkey);
640 ohdr->u.rc.reth.length = cpu_to_be32(len);
641 hwords += sizeof(struct ib_reth) / sizeof(u32);
642 if (len > pmtu) {
643 qp->s_state = OP(RDMA_WRITE_FIRST);
644 len = pmtu;
645 break;
646 }
647 if (wqe->wr.opcode == IB_WR_RDMA_WRITE) {
648 qp->s_state = OP(RDMA_WRITE_ONLY);
649 } else {
650 qp->s_state =
651 OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
652 /* Immediate data comes after RETH */
653 ohdr->u.rc.imm_data = wqe->wr.ex.imm_data;
654 hwords += 1;
655 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
656 bth0 |= IB_BTH_SOLICITED;
657 }
658 bth2 |= IB_BTH_REQ_ACK;
659 if (++qp->s_cur == qp->s_size)
660 qp->s_cur = 0;
661 break;
662
663 case IB_WR_TID_RDMA_WRITE:
664 if (newreq) {
665 /*
666 * Limit the number of TID RDMA WRITE requests.
667 */
668 if (atomic_read(&priv->n_tid_requests) >=
669 HFI1_TID_RDMA_WRITE_CNT)
670 goto bail;
671
672 if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
673 qp->s_lsn++;
674 }
675
676 hwords += hfi1_build_tid_rdma_write_req(qp, wqe, ohdr,
677 &bth1, &bth2,
678 &len);
679 ss = NULL;
680 if (priv->s_tid_cur == HFI1_QP_WQE_INVALID) {
681 priv->s_tid_cur = qp->s_cur;
682 if (priv->s_tid_tail == HFI1_QP_WQE_INVALID) {
683 priv->s_tid_tail = qp->s_cur;
684 priv->s_state = TID_OP(WRITE_RESP);
685 }
686 } else if (priv->s_tid_cur == priv->s_tid_head) {
687 struct rvt_swqe *__w;
688 struct tid_rdma_request *__r;
689
690 __w = rvt_get_swqe_ptr(qp, priv->s_tid_cur);
691 __r = wqe_to_tid_req(__w);
692
693 /*
694 * The s_tid_cur pointer is advanced to s_cur if
695 * any of the following conditions about the WQE
696 * to which s_ti_cur currently points to are
697 * satisfied:
698 * 1. The request is not a TID RDMA WRITE
699 * request,
700 * 2. The request is in the INACTIVE or
701 * COMPLETE states (TID RDMA READ requests
702 * stay at INACTIVE and TID RDMA WRITE
703 * transition to COMPLETE when done),
704 * 3. The request is in the ACTIVE or SYNC
705 * state and the number of completed
706 * segments is equal to the total segment
707 * count.
708 * (If ACTIVE, the request is waiting for
709 * ACKs. If SYNC, the request has not
710 * received any responses because it's
711 * waiting on a sync point.)
712 */
713 if (__w->wr.opcode != IB_WR_TID_RDMA_WRITE ||
714 __r->state == TID_REQUEST_INACTIVE ||
715 __r->state == TID_REQUEST_COMPLETE ||
716 ((__r->state == TID_REQUEST_ACTIVE ||
717 __r->state == TID_REQUEST_SYNC) &&
718 __r->comp_seg == __r->total_segs)) {
719 if (priv->s_tid_tail ==
720 priv->s_tid_cur &&
721 priv->s_state ==
722 TID_OP(WRITE_DATA_LAST)) {
723 priv->s_tid_tail = qp->s_cur;
724 priv->s_state =
725 TID_OP(WRITE_RESP);
726 }
727 priv->s_tid_cur = qp->s_cur;
728 }
729 /*
730 * A corner case: when the last TID RDMA WRITE
731 * request was completed, s_tid_head,
732 * s_tid_cur, and s_tid_tail all point to the
733 * same location. Other requests are posted and
734 * s_cur wraps around to the same location,
735 * where a new TID RDMA WRITE is posted. In
736 * this case, none of the indices need to be
737 * updated. However, the priv->s_state should.
738 */
739 if (priv->s_tid_tail == qp->s_cur &&
740 priv->s_state == TID_OP(WRITE_DATA_LAST))
741 priv->s_state = TID_OP(WRITE_RESP);
742 }
743 req = wqe_to_tid_req(wqe);
744 if (newreq) {
745 priv->s_tid_head = qp->s_cur;
746 priv->pending_tid_w_resp += req->total_segs;
747 atomic_inc(&priv->n_tid_requests);
748 atomic_dec(&priv->n_requests);
749 } else {
750 req->state = TID_REQUEST_RESEND;
751 req->comp_seg = delta_psn(bth2, wqe->psn);
752 /*
753 * Pull back any segments since we are going
754 * to re-receive them.
755 */
756 req->setup_head = req->clear_tail;
757 priv->pending_tid_w_resp +=
758 delta_psn(wqe->lpsn, bth2) + 1;
759 }
760
761 trace_hfi1_tid_write_sender_make_req(qp, newreq);
762 trace_hfi1_tid_req_make_req_write(qp, newreq,
763 wqe->wr.opcode,
764 wqe->psn, wqe->lpsn,
765 req);
766 if (++qp->s_cur == qp->s_size)
767 qp->s_cur = 0;
768 break;
769
770 case IB_WR_RDMA_READ:
771 /*
772 * Don't allow more operations to be started
773 * than the QP limits allow.
774 */
775 if (qp->s_num_rd_atomic >=
776 qp->s_max_rd_atomic) {
777 qp->s_flags |= RVT_S_WAIT_RDMAR;
778 goto bail;
779 }
780 qp->s_num_rd_atomic++;
781 if (newreq && !(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
782 qp->s_lsn++;
783 put_ib_reth_vaddr(
784 wqe->rdma_wr.remote_addr,
785 &ohdr->u.rc.reth);
786 ohdr->u.rc.reth.rkey =
787 cpu_to_be32(wqe->rdma_wr.rkey);
788 ohdr->u.rc.reth.length = cpu_to_be32(len);
789 qp->s_state = OP(RDMA_READ_REQUEST);
790 hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
791 ss = NULL;
792 len = 0;
793 bth2 |= IB_BTH_REQ_ACK;
794 if (++qp->s_cur == qp->s_size)
795 qp->s_cur = 0;
796 break;
797
798 case IB_WR_TID_RDMA_READ:
799 trace_hfi1_tid_read_sender_make_req(qp, newreq);
800 wpriv = wqe->priv;
801 req = wqe_to_tid_req(wqe);
802 trace_hfi1_tid_req_make_req_read(qp, newreq,
803 wqe->wr.opcode,
804 wqe->psn, wqe->lpsn,
805 req);
806 delta = cmp_psn(qp->s_psn, wqe->psn);
807
808 /*
809 * Don't allow more operations to be started
810 * than the QP limits allow. We could get here under
811 * three conditions; (1) It's a new request; (2) We are
812 * sending the second or later segment of a request,
813 * but the qp->s_state is set to OP(RDMA_READ_REQUEST)
814 * when the last segment of a previous request is
815 * received just before this; (3) We are re-sending a
816 * request.
817 */
818 if (qp->s_num_rd_atomic >= qp->s_max_rd_atomic) {
819 qp->s_flags |= RVT_S_WAIT_RDMAR;
820 goto bail;
821 }
822 if (newreq) {
823 struct tid_rdma_flow *flow =
824 &req->flows[req->setup_head];
825
826 /*
827 * Set up s_sge as it is needed for TID
828 * allocation. However, if the pages have been
829 * walked and mapped, skip it. An earlier try
830 * has failed to allocate the TID entries.
831 */
832 if (!flow->npagesets) {
833 qp->s_sge.sge = wqe->sg_list[0];
834 qp->s_sge.sg_list = wqe->sg_list + 1;
835 qp->s_sge.num_sge = wqe->wr.num_sge;
836 qp->s_sge.total_len = wqe->length;
837 qp->s_len = wqe->length;
838 req->isge = 0;
839 req->clear_tail = req->setup_head;
840 req->flow_idx = req->setup_head;
841 req->state = TID_REQUEST_ACTIVE;
842 }
843 } else if (delta == 0) {
844 /* Re-send a request */
845 req->cur_seg = 0;
846 req->comp_seg = 0;
847 req->ack_pending = 0;
848 req->flow_idx = req->clear_tail;
849 req->state = TID_REQUEST_RESEND;
850 }
851 req->s_next_psn = qp->s_psn;
852 /* Read one segment at a time */
853 len = min_t(u32, req->seg_len,
854 wqe->length - req->seg_len * req->cur_seg);
855 delta = hfi1_build_tid_rdma_read_req(qp, wqe, ohdr,
856 &bth1, &bth2,
857 &len);
858 if (delta <= 0) {
859 /* Wait for TID space */
860 goto bail;
861 }
862 if (newreq && !(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
863 qp->s_lsn++;
864 hwords += delta;
865 ss = &wpriv->ss;
866 /* Check if this is the last segment */
867 if (req->cur_seg >= req->total_segs &&
868 ++qp->s_cur == qp->s_size)
869 qp->s_cur = 0;
870 break;
871
872 case IB_WR_ATOMIC_CMP_AND_SWP:
873 case IB_WR_ATOMIC_FETCH_AND_ADD:
874 /*
875 * Don't allow more operations to be started
876 * than the QP limits allow.
877 */
878 if (qp->s_num_rd_atomic >=
879 qp->s_max_rd_atomic) {
880 qp->s_flags |= RVT_S_WAIT_RDMAR;
881 goto bail;
882 }
883 qp->s_num_rd_atomic++;
884 fallthrough;
885 case IB_WR_OPFN:
886 if (newreq && !(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
887 qp->s_lsn++;
888 if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
889 wqe->wr.opcode == IB_WR_OPFN) {
890 qp->s_state = OP(COMPARE_SWAP);
891 put_ib_ateth_swap(wqe->atomic_wr.swap,
892 &ohdr->u.atomic_eth);
893 put_ib_ateth_compare(wqe->atomic_wr.compare_add,
894 &ohdr->u.atomic_eth);
895 } else {
896 qp->s_state = OP(FETCH_ADD);
897 put_ib_ateth_swap(wqe->atomic_wr.compare_add,
898 &ohdr->u.atomic_eth);
899 put_ib_ateth_compare(0, &ohdr->u.atomic_eth);
900 }
901 put_ib_ateth_vaddr(wqe->atomic_wr.remote_addr,
902 &ohdr->u.atomic_eth);
903 ohdr->u.atomic_eth.rkey = cpu_to_be32(
904 wqe->atomic_wr.rkey);
905 hwords += sizeof(struct ib_atomic_eth) / sizeof(u32);
906 ss = NULL;
907 len = 0;
908 bth2 |= IB_BTH_REQ_ACK;
909 if (++qp->s_cur == qp->s_size)
910 qp->s_cur = 0;
911 break;
912
913 default:
914 goto bail;
915 }
916 if (wqe->wr.opcode != IB_WR_TID_RDMA_READ) {
917 qp->s_sge.sge = wqe->sg_list[0];
918 qp->s_sge.sg_list = wqe->sg_list + 1;
919 qp->s_sge.num_sge = wqe->wr.num_sge;
920 qp->s_sge.total_len = wqe->length;
921 qp->s_len = wqe->length;
922 }
923 if (newreq) {
924 qp->s_tail++;
925 if (qp->s_tail >= qp->s_size)
926 qp->s_tail = 0;
927 }
928 if (wqe->wr.opcode == IB_WR_RDMA_READ ||
929 wqe->wr.opcode == IB_WR_TID_RDMA_WRITE)
930 qp->s_psn = wqe->lpsn + 1;
931 else if (wqe->wr.opcode == IB_WR_TID_RDMA_READ)
932 qp->s_psn = req->s_next_psn;
933 else
934 qp->s_psn++;
935 break;
936
937 case OP(RDMA_READ_RESPONSE_FIRST):
938 /*
939 * qp->s_state is normally set to the opcode of the
940 * last packet constructed for new requests and therefore
941 * is never set to RDMA read response.
942 * RDMA_READ_RESPONSE_FIRST is used by the ACK processing
943 * thread to indicate a SEND needs to be restarted from an
944 * earlier PSN without interfering with the sending thread.
945 * See restart_rc().
946 */
947 qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu);
948 fallthrough;
949 case OP(SEND_FIRST):
950 qp->s_state = OP(SEND_MIDDLE);
951 fallthrough;
952 case OP(SEND_MIDDLE):
953 bth2 = mask_psn(qp->s_psn++);
954 ss = &qp->s_sge;
955 len = qp->s_len;
956 if (len > pmtu) {
957 len = pmtu;
958 middle = HFI1_CAP_IS_KSET(SDMA_AHG);
959 break;
960 }
961 if (wqe->wr.opcode == IB_WR_SEND) {
962 qp->s_state = OP(SEND_LAST);
963 } else if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
964 qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
965 /* Immediate data comes after the BTH */
966 ohdr->u.imm_data = wqe->wr.ex.imm_data;
967 hwords += 1;
968 } else {
969 qp->s_state = OP(SEND_LAST_WITH_INVALIDATE);
970 /* invalidate data comes after the BTH */
971 ohdr->u.ieth = cpu_to_be32(wqe->wr.ex.invalidate_rkey);
972 hwords += 1;
973 }
974 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
975 bth0 |= IB_BTH_SOLICITED;
976 bth2 |= IB_BTH_REQ_ACK;
977 qp->s_cur++;
978 if (qp->s_cur >= qp->s_size)
979 qp->s_cur = 0;
980 break;
981
982 case OP(RDMA_READ_RESPONSE_LAST):
983 /*
984 * qp->s_state is normally set to the opcode of the
985 * last packet constructed for new requests and therefore
986 * is never set to RDMA read response.
987 * RDMA_READ_RESPONSE_LAST is used by the ACK processing
988 * thread to indicate a RDMA write needs to be restarted from
989 * an earlier PSN without interfering with the sending thread.
990 * See restart_rc().
991 */
992 qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu);
993 fallthrough;
994 case OP(RDMA_WRITE_FIRST):
995 qp->s_state = OP(RDMA_WRITE_MIDDLE);
996 fallthrough;
997 case OP(RDMA_WRITE_MIDDLE):
998 bth2 = mask_psn(qp->s_psn++);
999 ss = &qp->s_sge;
1000 len = qp->s_len;
1001 if (len > pmtu) {
1002 len = pmtu;
1003 middle = HFI1_CAP_IS_KSET(SDMA_AHG);
1004 break;
1005 }
1006 if (wqe->wr.opcode == IB_WR_RDMA_WRITE) {
1007 qp->s_state = OP(RDMA_WRITE_LAST);
1008 } else {
1009 qp->s_state = OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
1010 /* Immediate data comes after the BTH */
1011 ohdr->u.imm_data = wqe->wr.ex.imm_data;
1012 hwords += 1;
1013 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
1014 bth0 |= IB_BTH_SOLICITED;
1015 }
1016 bth2 |= IB_BTH_REQ_ACK;
1017 qp->s_cur++;
1018 if (qp->s_cur >= qp->s_size)
1019 qp->s_cur = 0;
1020 break;
1021
1022 case OP(RDMA_READ_RESPONSE_MIDDLE):
1023 /*
1024 * qp->s_state is normally set to the opcode of the
1025 * last packet constructed for new requests and therefore
1026 * is never set to RDMA read response.
1027 * RDMA_READ_RESPONSE_MIDDLE is used by the ACK processing
1028 * thread to indicate a RDMA read needs to be restarted from
1029 * an earlier PSN without interfering with the sending thread.
1030 * See restart_rc().
1031 */
1032 len = (delta_psn(qp->s_psn, wqe->psn)) * pmtu;
1033 put_ib_reth_vaddr(
1034 wqe->rdma_wr.remote_addr + len,
1035 &ohdr->u.rc.reth);
1036 ohdr->u.rc.reth.rkey =
1037 cpu_to_be32(wqe->rdma_wr.rkey);
1038 ohdr->u.rc.reth.length = cpu_to_be32(wqe->length - len);
1039 qp->s_state = OP(RDMA_READ_REQUEST);
1040 hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
1041 bth2 = mask_psn(qp->s_psn) | IB_BTH_REQ_ACK;
1042 qp->s_psn = wqe->lpsn + 1;
1043 ss = NULL;
1044 len = 0;
1045 qp->s_cur++;
1046 if (qp->s_cur == qp->s_size)
1047 qp->s_cur = 0;
1048 break;
1049
1050 case TID_OP(WRITE_RESP):
1051 /*
1052 * This value for s_state is used for restarting a TID RDMA
1053 * WRITE request. See comment in OP(RDMA_READ_RESPONSE_MIDDLE
1054 * for more).
1055 */
1056 req = wqe_to_tid_req(wqe);
1057 req->state = TID_REQUEST_RESEND;
1058 rcu_read_lock();
1059 remote = rcu_dereference(priv->tid_rdma.remote);
1060 req->comp_seg = delta_psn(qp->s_psn, wqe->psn);
1061 len = wqe->length - (req->comp_seg * remote->max_len);
1062 rcu_read_unlock();
1063
1064 bth2 = mask_psn(qp->s_psn);
1065 hwords += hfi1_build_tid_rdma_write_req(qp, wqe, ohdr, &bth1,
1066 &bth2, &len);
1067 qp->s_psn = wqe->lpsn + 1;
1068 ss = NULL;
1069 qp->s_state = TID_OP(WRITE_REQ);
1070 priv->pending_tid_w_resp += delta_psn(wqe->lpsn, bth2) + 1;
1071 priv->s_tid_cur = qp->s_cur;
1072 if (++qp->s_cur == qp->s_size)
1073 qp->s_cur = 0;
1074 trace_hfi1_tid_req_make_req_write(qp, 0, wqe->wr.opcode,
1075 wqe->psn, wqe->lpsn, req);
1076 break;
1077
1078 case TID_OP(READ_RESP):
1079 if (wqe->wr.opcode != IB_WR_TID_RDMA_READ)
1080 goto bail;
1081 /* This is used to restart a TID read request */
1082 req = wqe_to_tid_req(wqe);
1083 wpriv = wqe->priv;
1084 /*
1085 * Back down. The field qp->s_psn has been set to the psn with
1086 * which the request should be restart. It's OK to use division
1087 * as this is on the retry path.
1088 */
1089 req->cur_seg = delta_psn(qp->s_psn, wqe->psn) / priv->pkts_ps;
1090
1091 /*
1092 * The following function need to be redefined to return the
1093 * status to make sure that we find the flow. At the same
1094 * time, we can use the req->state change to check if the
1095 * call succeeds or not.
1096 */
1097 req->state = TID_REQUEST_RESEND;
1098 hfi1_tid_rdma_restart_req(qp, wqe, &bth2);
1099 if (req->state != TID_REQUEST_ACTIVE) {
1100 /*
1101 * Failed to find the flow. Release all allocated tid
1102 * resources.
1103 */
1104 hfi1_kern_exp_rcv_clear_all(req);
1105 hfi1_kern_clear_hw_flow(priv->rcd, qp);
1106
1107 hfi1_trdma_send_complete(qp, wqe, IB_WC_LOC_QP_OP_ERR);
1108 goto bail;
1109 }
1110 req->state = TID_REQUEST_RESEND;
1111 len = min_t(u32, req->seg_len,
1112 wqe->length - req->seg_len * req->cur_seg);
1113 flow = &req->flows[req->flow_idx];
1114 len -= flow->sent;
1115 req->s_next_psn = flow->flow_state.ib_lpsn + 1;
1116 delta = hfi1_build_tid_rdma_read_packet(wqe, ohdr, &bth1,
1117 &bth2, &len);
1118 if (delta <= 0) {
1119 /* Wait for TID space */
1120 goto bail;
1121 }
1122 hwords += delta;
1123 ss = &wpriv->ss;
1124 /* Check if this is the last segment */
1125 if (req->cur_seg >= req->total_segs &&
1126 ++qp->s_cur == qp->s_size)
1127 qp->s_cur = 0;
1128 qp->s_psn = req->s_next_psn;
1129 trace_hfi1_tid_req_make_req_read(qp, 0, wqe->wr.opcode,
1130 wqe->psn, wqe->lpsn, req);
1131 break;
1132 case TID_OP(READ_REQ):
1133 req = wqe_to_tid_req(wqe);
1134 delta = cmp_psn(qp->s_psn, wqe->psn);
1135 /*
1136 * If the current WR is not TID RDMA READ, or this is the start
1137 * of a new request, we need to change the qp->s_state so that
1138 * the request can be set up properly.
1139 */
1140 if (wqe->wr.opcode != IB_WR_TID_RDMA_READ || delta == 0 ||
1141 qp->s_cur == qp->s_tail) {
1142 qp->s_state = OP(RDMA_READ_REQUEST);
1143 if (delta == 0 || qp->s_cur == qp->s_tail)
1144 goto check_s_state;
1145 else
1146 goto bail;
1147 }
1148
1149 /* Rate limiting */
1150 if (qp->s_num_rd_atomic >= qp->s_max_rd_atomic) {
1151 qp->s_flags |= RVT_S_WAIT_RDMAR;
1152 goto bail;
1153 }
1154
1155 wpriv = wqe->priv;
1156 /* Read one segment at a time */
1157 len = min_t(u32, req->seg_len,
1158 wqe->length - req->seg_len * req->cur_seg);
1159 delta = hfi1_build_tid_rdma_read_req(qp, wqe, ohdr, &bth1,
1160 &bth2, &len);
1161 if (delta <= 0) {
1162 /* Wait for TID space */
1163 goto bail;
1164 }
1165 hwords += delta;
1166 ss = &wpriv->ss;
1167 /* Check if this is the last segment */
1168 if (req->cur_seg >= req->total_segs &&
1169 ++qp->s_cur == qp->s_size)
1170 qp->s_cur = 0;
1171 qp->s_psn = req->s_next_psn;
1172 trace_hfi1_tid_req_make_req_read(qp, 0, wqe->wr.opcode,
1173 wqe->psn, wqe->lpsn, req);
1174 break;
1175 }
1176 qp->s_sending_hpsn = bth2;
1177 delta = delta_psn(bth2, wqe->psn);
1178 if (delta && delta % HFI1_PSN_CREDIT == 0 &&
1179 wqe->wr.opcode != IB_WR_TID_RDMA_WRITE)
1180 bth2 |= IB_BTH_REQ_ACK;
1181 if (qp->s_flags & RVT_S_SEND_ONE) {
1182 qp->s_flags &= ~RVT_S_SEND_ONE;
1183 qp->s_flags |= RVT_S_WAIT_ACK;
1184 bth2 |= IB_BTH_REQ_ACK;
1185 }
1186 qp->s_len -= len;
1187 ps->s_txreq->hdr_dwords = hwords;
1188 ps->s_txreq->sde = priv->s_sde;
1189 ps->s_txreq->ss = ss;
1190 ps->s_txreq->s_cur_size = len;
1191 hfi1_make_ruc_header(
1192 qp,
1193 ohdr,
1194 bth0 | (qp->s_state << 24),
1195 bth1,
1196 bth2,
1197 middle,
1198 ps);
1199 return 1;
1200
1201 done_free_tx:
1202 hfi1_put_txreq(ps->s_txreq);
1203 ps->s_txreq = NULL;
1204 return 1;
1205
1206 bail:
1207 hfi1_put_txreq(ps->s_txreq);
1208
1209 bail_no_tx:
1210 ps->s_txreq = NULL;
1211 qp->s_flags &= ~RVT_S_BUSY;
1212 /*
1213 * If we didn't get a txreq, the QP will be woken up later to try
1214 * again. Set the flags to indicate which work item to wake
1215 * up.
1216 */
1217 iowait_set_flag(&priv->s_iowait, IOWAIT_PENDING_IB);
1218 return 0;
1219 }
1220
1221 static inline void hfi1_make_bth_aeth(struct rvt_qp *qp,
1222 struct ib_other_headers *ohdr,
1223 u32 bth0, u32 bth1)
1224 {
1225 if (qp->r_nak_state)
1226 ohdr->u.aeth = cpu_to_be32((qp->r_msn & IB_MSN_MASK) |
1227 (qp->r_nak_state <<
1228 IB_AETH_CREDIT_SHIFT));
1229 else
1230 ohdr->u.aeth = rvt_compute_aeth(qp);
1231
1232 ohdr->bth[0] = cpu_to_be32(bth0);
1233 ohdr->bth[1] = cpu_to_be32(bth1 | qp->remote_qpn);
1234 ohdr->bth[2] = cpu_to_be32(mask_psn(qp->r_ack_psn));
1235 }
1236
1237 static inline void hfi1_queue_rc_ack(struct hfi1_packet *packet, bool is_fecn)
1238 {
1239 struct rvt_qp *qp = packet->qp;
1240 struct hfi1_ibport *ibp;
1241 unsigned long flags;
1242
1243 spin_lock_irqsave(&qp->s_lock, flags);
1244 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
1245 goto unlock;
1246 ibp = rcd_to_iport(packet->rcd);
1247 this_cpu_inc(*ibp->rvp.rc_qacks);
1248 qp->s_flags |= RVT_S_ACK_PENDING | RVT_S_RESP_PENDING;
1249 qp->s_nak_state = qp->r_nak_state;
1250 qp->s_ack_psn = qp->r_ack_psn;
1251 if (is_fecn)
1252 qp->s_flags |= RVT_S_ECN;
1253
1254 /* Schedule the send tasklet. */
1255 hfi1_schedule_send(qp);
1256 unlock:
1257 spin_unlock_irqrestore(&qp->s_lock, flags);
1258 }
1259
1260 static inline void hfi1_make_rc_ack_9B(struct hfi1_packet *packet,
1261 struct hfi1_opa_header *opa_hdr,
1262 u8 sc5, bool is_fecn,
1263 u64 *pbc_flags, u32 *hwords,
1264 u32 *nwords)
1265 {
1266 struct rvt_qp *qp = packet->qp;
1267 struct hfi1_ibport *ibp = rcd_to_iport(packet->rcd);
1268 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
1269 struct ib_header *hdr = &opa_hdr->ibh;
1270 struct ib_other_headers *ohdr;
1271 u16 lrh0 = HFI1_LRH_BTH;
1272 u16 pkey;
1273 u32 bth0, bth1;
1274
1275 opa_hdr->hdr_type = HFI1_PKT_TYPE_9B;
1276 ohdr = &hdr->u.oth;
1277 /* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4 */
1278 *hwords = 6;
1279
1280 if (unlikely(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH)) {
1281 *hwords += hfi1_make_grh(ibp, &hdr->u.l.grh,
1282 rdma_ah_read_grh(&qp->remote_ah_attr),
1283 *hwords - 2, SIZE_OF_CRC);
1284 ohdr = &hdr->u.l.oth;
1285 lrh0 = HFI1_LRH_GRH;
1286 }
1287 /* set PBC_DC_INFO bit (aka SC[4]) in pbc_flags */
1288 *pbc_flags |= ((!!(sc5 & 0x10)) << PBC_DC_INFO_SHIFT);
1289
1290 /* read pkey_index w/o lock (its atomic) */
1291 pkey = hfi1_get_pkey(ibp, qp->s_pkey_index);
1292
1293 lrh0 |= (sc5 & IB_SC_MASK) << IB_SC_SHIFT |
1294 (rdma_ah_get_sl(&qp->remote_ah_attr) & IB_SL_MASK) <<
1295 IB_SL_SHIFT;
1296
1297 hfi1_make_ib_hdr(hdr, lrh0, *hwords + SIZE_OF_CRC,
1298 opa_get_lid(rdma_ah_get_dlid(&qp->remote_ah_attr), 9B),
1299 ppd->lid | rdma_ah_get_path_bits(&qp->remote_ah_attr));
1300
1301 bth0 = pkey | (OP(ACKNOWLEDGE) << 24);
1302 if (qp->s_mig_state == IB_MIG_MIGRATED)
1303 bth0 |= IB_BTH_MIG_REQ;
1304 bth1 = (!!is_fecn) << IB_BECN_SHIFT;
1305 /*
1306 * Inline ACKs go out without the use of the Verbs send engine, so
1307 * we need to set the STL Verbs Extended bit here
1308 */
1309 bth1 |= HFI1_CAP_IS_KSET(OPFN) << IB_BTHE_E_SHIFT;
1310 hfi1_make_bth_aeth(qp, ohdr, bth0, bth1);
1311 }
1312
1313 static inline void hfi1_make_rc_ack_16B(struct hfi1_packet *packet,
1314 struct hfi1_opa_header *opa_hdr,
1315 u8 sc5, bool is_fecn,
1316 u64 *pbc_flags, u32 *hwords,
1317 u32 *nwords)
1318 {
1319 struct rvt_qp *qp = packet->qp;
1320 struct hfi1_ibport *ibp = rcd_to_iport(packet->rcd);
1321 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
1322 struct hfi1_16b_header *hdr = &opa_hdr->opah;
1323 struct ib_other_headers *ohdr;
1324 u32 bth0, bth1 = 0;
1325 u16 len, pkey;
1326 bool becn = is_fecn;
1327 u8 l4 = OPA_16B_L4_IB_LOCAL;
1328 u8 extra_bytes;
1329
1330 opa_hdr->hdr_type = HFI1_PKT_TYPE_16B;
1331 ohdr = &hdr->u.oth;
1332 /* header size in 32-bit words 16B LRH+BTH+AETH = (16+12+4)/4 */
1333 *hwords = 8;
1334 extra_bytes = hfi1_get_16b_padding(*hwords << 2, 0);
1335 *nwords = SIZE_OF_CRC + ((extra_bytes + SIZE_OF_LT) >> 2);
1336
1337 if (unlikely(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH) &&
1338 hfi1_check_mcast(rdma_ah_get_dlid(&qp->remote_ah_attr))) {
1339 *hwords += hfi1_make_grh(ibp, &hdr->u.l.grh,
1340 rdma_ah_read_grh(&qp->remote_ah_attr),
1341 *hwords - 4, *nwords);
1342 ohdr = &hdr->u.l.oth;
1343 l4 = OPA_16B_L4_IB_GLOBAL;
1344 }
1345 *pbc_flags |= PBC_PACKET_BYPASS | PBC_INSERT_BYPASS_ICRC;
1346
1347 /* read pkey_index w/o lock (its atomic) */
1348 pkey = hfi1_get_pkey(ibp, qp->s_pkey_index);
1349
1350 /* Convert dwords to flits */
1351 len = (*hwords + *nwords) >> 1;
1352
1353 hfi1_make_16b_hdr(hdr, ppd->lid |
1354 (rdma_ah_get_path_bits(&qp->remote_ah_attr) &
1355 ((1 << ppd->lmc) - 1)),
1356 opa_get_lid(rdma_ah_get_dlid(&qp->remote_ah_attr),
1357 16B), len, pkey, becn, 0, l4, sc5);
1358
1359 bth0 = pkey | (OP(ACKNOWLEDGE) << 24);
1360 bth0 |= extra_bytes << 20;
1361 if (qp->s_mig_state == IB_MIG_MIGRATED)
1362 bth1 = OPA_BTH_MIG_REQ;
1363 hfi1_make_bth_aeth(qp, ohdr, bth0, bth1);
1364 }
1365
1366 typedef void (*hfi1_make_rc_ack)(struct hfi1_packet *packet,
1367 struct hfi1_opa_header *opa_hdr,
1368 u8 sc5, bool is_fecn,
1369 u64 *pbc_flags, u32 *hwords,
1370 u32 *nwords);
1371
1372 /* We support only two types - 9B and 16B for now */
1373 static const hfi1_make_rc_ack hfi1_make_rc_ack_tbl[2] = {
1374 [HFI1_PKT_TYPE_9B] = &hfi1_make_rc_ack_9B,
1375 [HFI1_PKT_TYPE_16B] = &hfi1_make_rc_ack_16B
1376 };
1377
1378 /**
1379 * hfi1_send_rc_ack - Construct an ACK packet and send it
1380 * @qp: a pointer to the QP
1381 *
1382 * This is called from hfi1_rc_rcv() and handle_receive_interrupt().
1383 * Note that RDMA reads and atomics are handled in the
1384 * send side QP state and send engine.
1385 */
1386 void hfi1_send_rc_ack(struct hfi1_packet *packet, bool is_fecn)
1387 {
1388 struct hfi1_ctxtdata *rcd = packet->rcd;
1389 struct rvt_qp *qp = packet->qp;
1390 struct hfi1_ibport *ibp = rcd_to_iport(rcd);
1391 struct hfi1_qp_priv *priv = qp->priv;
1392 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
1393 u8 sc5 = ibp->sl_to_sc[rdma_ah_get_sl(&qp->remote_ah_attr)];
1394 u64 pbc, pbc_flags = 0;
1395 u32 hwords = 0;
1396 u32 nwords = 0;
1397 u32 plen;
1398 struct pio_buf *pbuf;
1399 struct hfi1_opa_header opa_hdr;
1400
1401 /* clear the defer count */
1402 qp->r_adefered = 0;
1403
1404 /* Don't send ACK or NAK if a RDMA read or atomic is pending. */
1405 if (qp->s_flags & RVT_S_RESP_PENDING) {
1406 hfi1_queue_rc_ack(packet, is_fecn);
1407 return;
1408 }
1409
1410 /* Ensure s_rdma_ack_cnt changes are committed */
1411 if (qp->s_rdma_ack_cnt) {
1412 hfi1_queue_rc_ack(packet, is_fecn);
1413 return;
1414 }
1415
1416 /* Don't try to send ACKs if the link isn't ACTIVE */
1417 if (driver_lstate(ppd) != IB_PORT_ACTIVE)
1418 return;
1419
1420 /* Make the appropriate header */
1421 hfi1_make_rc_ack_tbl[priv->hdr_type](packet, &opa_hdr, sc5, is_fecn,
1422 &pbc_flags, &hwords, &nwords);
1423
1424 plen = 2 /* PBC */ + hwords + nwords;
1425 pbc = create_pbc(ppd, pbc_flags, qp->srate_mbps,
1426 sc_to_vlt(ppd->dd, sc5), plen);
1427 pbuf = sc_buffer_alloc(rcd->sc, plen, NULL, NULL);
1428 if (IS_ERR_OR_NULL(pbuf)) {
1429 /*
1430 * We have no room to send at the moment. Pass
1431 * responsibility for sending the ACK to the send engine
1432 * so that when enough buffer space becomes available,
1433 * the ACK is sent ahead of other outgoing packets.
1434 */
1435 hfi1_queue_rc_ack(packet, is_fecn);
1436 return;
1437 }
1438 trace_ack_output_ibhdr(dd_from_ibdev(qp->ibqp.device),
1439 &opa_hdr, ib_is_sc5(sc5));
1440
1441 /* write the pbc and data */
1442 ppd->dd->pio_inline_send(ppd->dd, pbuf, pbc,
1443 (priv->hdr_type == HFI1_PKT_TYPE_9B ?
1444 (void *)&opa_hdr.ibh :
1445 (void *)&opa_hdr.opah), hwords);
1446 return;
1447 }
1448
1449 /**
1450 * update_num_rd_atomic - update the qp->s_num_rd_atomic
1451 * @qp: the QP
1452 * @psn: the packet sequence number to restart at
1453 * @wqe: the wqe
1454 *
1455 * This is called from reset_psn() to update qp->s_num_rd_atomic
1456 * for the current wqe.
1457 * Called at interrupt level with the QP s_lock held.
1458 */
1459 static void update_num_rd_atomic(struct rvt_qp *qp, u32 psn,
1460 struct rvt_swqe *wqe)
1461 {
1462 u32 opcode = wqe->wr.opcode;
1463
1464 if (opcode == IB_WR_RDMA_READ ||
1465 opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1466 opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
1467 qp->s_num_rd_atomic++;
1468 } else if (opcode == IB_WR_TID_RDMA_READ) {
1469 struct tid_rdma_request *req = wqe_to_tid_req(wqe);
1470 struct hfi1_qp_priv *priv = qp->priv;
1471
1472 if (cmp_psn(psn, wqe->lpsn) <= 0) {
1473 u32 cur_seg;
1474
1475 cur_seg = (psn - wqe->psn) / priv->pkts_ps;
1476 req->ack_pending = cur_seg - req->comp_seg;
1477 priv->pending_tid_r_segs += req->ack_pending;
1478 qp->s_num_rd_atomic += req->ack_pending;
1479 trace_hfi1_tid_req_update_num_rd_atomic(qp, 0,
1480 wqe->wr.opcode,
1481 wqe->psn,
1482 wqe->lpsn,
1483 req);
1484 } else {
1485 priv->pending_tid_r_segs += req->total_segs;
1486 qp->s_num_rd_atomic += req->total_segs;
1487 }
1488 }
1489 }
1490
1491 /**
1492 * reset_psn - reset the QP state to send starting from PSN
1493 * @qp: the QP
1494 * @psn: the packet sequence number to restart at
1495 *
1496 * This is called from hfi1_rc_rcv() to process an incoming RC ACK
1497 * for the given QP.
1498 * Called at interrupt level with the QP s_lock held.
1499 */
1500 static void reset_psn(struct rvt_qp *qp, u32 psn)
1501 {
1502 u32 n = qp->s_acked;
1503 struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, n);
1504 u32 opcode;
1505 struct hfi1_qp_priv *priv = qp->priv;
1506
1507 lockdep_assert_held(&qp->s_lock);
1508 qp->s_cur = n;
1509 priv->pending_tid_r_segs = 0;
1510 priv->pending_tid_w_resp = 0;
1511 qp->s_num_rd_atomic = 0;
1512
1513 /*
1514 * If we are starting the request from the beginning,
1515 * let the normal send code handle initialization.
1516 */
1517 if (cmp_psn(psn, wqe->psn) <= 0) {
1518 qp->s_state = OP(SEND_LAST);
1519 goto done;
1520 }
1521 update_num_rd_atomic(qp, psn, wqe);
1522
1523 /* Find the work request opcode corresponding to the given PSN. */
1524 for (;;) {
1525 int diff;
1526
1527 if (++n == qp->s_size)
1528 n = 0;
1529 if (n == qp->s_tail)
1530 break;
1531 wqe = rvt_get_swqe_ptr(qp, n);
1532 diff = cmp_psn(psn, wqe->psn);
1533 if (diff < 0) {
1534 /* Point wqe back to the previous one*/
1535 wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
1536 break;
1537 }
1538 qp->s_cur = n;
1539 /*
1540 * If we are starting the request from the beginning,
1541 * let the normal send code handle initialization.
1542 */
1543 if (diff == 0) {
1544 qp->s_state = OP(SEND_LAST);
1545 goto done;
1546 }
1547
1548 update_num_rd_atomic(qp, psn, wqe);
1549 }
1550 opcode = wqe->wr.opcode;
1551
1552 /*
1553 * Set the state to restart in the middle of a request.
1554 * Don't change the s_sge, s_cur_sge, or s_cur_size.
1555 * See hfi1_make_rc_req().
1556 */
1557 switch (opcode) {
1558 case IB_WR_SEND:
1559 case IB_WR_SEND_WITH_IMM:
1560 qp->s_state = OP(RDMA_READ_RESPONSE_FIRST);
1561 break;
1562
1563 case IB_WR_RDMA_WRITE:
1564 case IB_WR_RDMA_WRITE_WITH_IMM:
1565 qp->s_state = OP(RDMA_READ_RESPONSE_LAST);
1566 break;
1567
1568 case IB_WR_TID_RDMA_WRITE:
1569 qp->s_state = TID_OP(WRITE_RESP);
1570 break;
1571
1572 case IB_WR_RDMA_READ:
1573 qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE);
1574 break;
1575
1576 case IB_WR_TID_RDMA_READ:
1577 qp->s_state = TID_OP(READ_RESP);
1578 break;
1579
1580 default:
1581 /*
1582 * This case shouldn't happen since its only
1583 * one PSN per req.
1584 */
1585 qp->s_state = OP(SEND_LAST);
1586 }
1587 done:
1588 priv->s_flags &= ~HFI1_S_TID_WAIT_INTERLCK;
1589 qp->s_psn = psn;
1590 /*
1591 * Set RVT_S_WAIT_PSN as rc_complete() may start the timer
1592 * asynchronously before the send engine can get scheduled.
1593 * Doing it in hfi1_make_rc_req() is too late.
1594 */
1595 if ((cmp_psn(qp->s_psn, qp->s_sending_hpsn) <= 0) &&
1596 (cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) <= 0))
1597 qp->s_flags |= RVT_S_WAIT_PSN;
1598 qp->s_flags &= ~HFI1_S_AHG_VALID;
1599 trace_hfi1_sender_reset_psn(qp);
1600 }
1601
1602 /*
1603 * Back up requester to resend the last un-ACKed request.
1604 * The QP r_lock and s_lock should be held and interrupts disabled.
1605 */
1606 void hfi1_restart_rc(struct rvt_qp *qp, u32 psn, int wait)
1607 {
1608 struct hfi1_qp_priv *priv = qp->priv;
1609 struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1610 struct hfi1_ibport *ibp;
1611
1612 lockdep_assert_held(&qp->r_lock);
1613 lockdep_assert_held(&qp->s_lock);
1614 trace_hfi1_sender_restart_rc(qp);
1615 if (qp->s_retry == 0) {
1616 if (qp->s_mig_state == IB_MIG_ARMED) {
1617 hfi1_migrate_qp(qp);
1618 qp->s_retry = qp->s_retry_cnt;
1619 } else if (qp->s_last == qp->s_acked) {
1620 /*
1621 * We need special handling for the OPFN request WQEs as
1622 * they are not allowed to generate real user errors
1623 */
1624 if (wqe->wr.opcode == IB_WR_OPFN) {
1625 struct hfi1_ibport *ibp =
1626 to_iport(qp->ibqp.device, qp->port_num);
1627 /*
1628 * Call opfn_conn_reply() with capcode and
1629 * remaining data as 0 to close out the
1630 * current request
1631 */
1632 opfn_conn_reply(qp, priv->opfn.curr);
1633 wqe = do_rc_completion(qp, wqe, ibp);
1634 qp->s_flags &= ~RVT_S_WAIT_ACK;
1635 } else {
1636 trace_hfi1_tid_write_sender_restart_rc(qp, 0);
1637 if (wqe->wr.opcode == IB_WR_TID_RDMA_READ) {
1638 struct tid_rdma_request *req;
1639
1640 req = wqe_to_tid_req(wqe);
1641 hfi1_kern_exp_rcv_clear_all(req);
1642 hfi1_kern_clear_hw_flow(priv->rcd, qp);
1643 }
1644
1645 hfi1_trdma_send_complete(qp, wqe,
1646 IB_WC_RETRY_EXC_ERR);
1647 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1648 }
1649 return;
1650 } else { /* need to handle delayed completion */
1651 return;
1652 }
1653 } else {
1654 qp->s_retry--;
1655 }
1656
1657 ibp = to_iport(qp->ibqp.device, qp->port_num);
1658 if (wqe->wr.opcode == IB_WR_RDMA_READ ||
1659 wqe->wr.opcode == IB_WR_TID_RDMA_READ)
1660 ibp->rvp.n_rc_resends++;
1661 else
1662 ibp->rvp.n_rc_resends += delta_psn(qp->s_psn, psn);
1663
1664 qp->s_flags &= ~(RVT_S_WAIT_FENCE | RVT_S_WAIT_RDMAR |
1665 RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_PSN |
1666 RVT_S_WAIT_ACK | HFI1_S_WAIT_TID_RESP);
1667 if (wait)
1668 qp->s_flags |= RVT_S_SEND_ONE;
1669 reset_psn(qp, psn);
1670 }
1671
1672 /*
1673 * Set qp->s_sending_psn to the next PSN after the given one.
1674 * This would be psn+1 except when RDMA reads or TID RDMA ops
1675 * are present.
1676 */
1677 static void reset_sending_psn(struct rvt_qp *qp, u32 psn)
1678 {
1679 struct rvt_swqe *wqe;
1680 u32 n = qp->s_last;
1681
1682 lockdep_assert_held(&qp->s_lock);
1683 /* Find the work request corresponding to the given PSN. */
1684 for (;;) {
1685 wqe = rvt_get_swqe_ptr(qp, n);
1686 if (cmp_psn(psn, wqe->lpsn) <= 0) {
1687 if (wqe->wr.opcode == IB_WR_RDMA_READ ||
1688 wqe->wr.opcode == IB_WR_TID_RDMA_READ ||
1689 wqe->wr.opcode == IB_WR_TID_RDMA_WRITE)
1690 qp->s_sending_psn = wqe->lpsn + 1;
1691 else
1692 qp->s_sending_psn = psn + 1;
1693 break;
1694 }
1695 if (++n == qp->s_size)
1696 n = 0;
1697 if (n == qp->s_tail)
1698 break;
1699 }
1700 }
1701
1702 /**
1703 * hfi1_rc_verbs_aborted - handle abort status
1704 * @qp: the QP
1705 * @opah: the opa header
1706 *
1707 * This code modifies both ACK bit in BTH[2]
1708 * and the s_flags to go into send one mode.
1709 *
1710 * This serves to throttle the send engine to only
1711 * send a single packet in the likely case the
1712 * a link has gone down.
1713 */
1714 void hfi1_rc_verbs_aborted(struct rvt_qp *qp, struct hfi1_opa_header *opah)
1715 {
1716 struct ib_other_headers *ohdr = hfi1_get_rc_ohdr(opah);
1717 u8 opcode = ib_bth_get_opcode(ohdr);
1718 u32 psn;
1719
1720 /* ignore responses */
1721 if ((opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
1722 opcode <= OP(ATOMIC_ACKNOWLEDGE)) ||
1723 opcode == TID_OP(READ_RESP) ||
1724 opcode == TID_OP(WRITE_RESP))
1725 return;
1726
1727 psn = ib_bth_get_psn(ohdr) | IB_BTH_REQ_ACK;
1728 ohdr->bth[2] = cpu_to_be32(psn);
1729 qp->s_flags |= RVT_S_SEND_ONE;
1730 }
1731
1732 /*
1733 * This should be called with the QP s_lock held and interrupts disabled.
1734 */
1735 void hfi1_rc_send_complete(struct rvt_qp *qp, struct hfi1_opa_header *opah)
1736 {
1737 struct ib_other_headers *ohdr;
1738 struct hfi1_qp_priv *priv = qp->priv;
1739 struct rvt_swqe *wqe;
1740 u32 opcode, head, tail;
1741 u32 psn;
1742 struct tid_rdma_request *req;
1743
1744 lockdep_assert_held(&qp->s_lock);
1745 if (!(ib_rvt_state_ops[qp->state] & RVT_SEND_OR_FLUSH_OR_RECV_OK))
1746 return;
1747
1748 ohdr = hfi1_get_rc_ohdr(opah);
1749 opcode = ib_bth_get_opcode(ohdr);
1750 if ((opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
1751 opcode <= OP(ATOMIC_ACKNOWLEDGE)) ||
1752 opcode == TID_OP(READ_RESP) ||
1753 opcode == TID_OP(WRITE_RESP)) {
1754 WARN_ON(!qp->s_rdma_ack_cnt);
1755 qp->s_rdma_ack_cnt--;
1756 return;
1757 }
1758
1759 psn = ib_bth_get_psn(ohdr);
1760 /*
1761 * Don't attempt to reset the sending PSN for packets in the
1762 * KDETH PSN space since the PSN does not match anything.
1763 */
1764 if (opcode != TID_OP(WRITE_DATA) &&
1765 opcode != TID_OP(WRITE_DATA_LAST) &&
1766 opcode != TID_OP(ACK) && opcode != TID_OP(RESYNC))
1767 reset_sending_psn(qp, psn);
1768
1769 /* Handle TID RDMA WRITE packets differently */
1770 if (opcode >= TID_OP(WRITE_REQ) &&
1771 opcode <= TID_OP(WRITE_DATA_LAST)) {
1772 head = priv->s_tid_head;
1773 tail = priv->s_tid_cur;
1774 /*
1775 * s_tid_cur is set to s_tid_head in the case, where
1776 * a new TID RDMA request is being started and all
1777 * previous ones have been completed.
1778 * Therefore, we need to do a secondary check in order
1779 * to properly determine whether we should start the
1780 * RC timer.
1781 */
1782 wqe = rvt_get_swqe_ptr(qp, tail);
1783 req = wqe_to_tid_req(wqe);
1784 if (head == tail && req->comp_seg < req->total_segs) {
1785 if (tail == 0)
1786 tail = qp->s_size - 1;
1787 else
1788 tail -= 1;
1789 }
1790 } else {
1791 head = qp->s_tail;
1792 tail = qp->s_acked;
1793 }
1794
1795 /*
1796 * Start timer after a packet requesting an ACK has been sent and
1797 * there are still requests that haven't been acked.
1798 */
1799 if ((psn & IB_BTH_REQ_ACK) && tail != head &&
1800 opcode != TID_OP(WRITE_DATA) && opcode != TID_OP(WRITE_DATA_LAST) &&
1801 opcode != TID_OP(RESYNC) &&
1802 !(qp->s_flags &
1803 (RVT_S_TIMER | RVT_S_WAIT_RNR | RVT_S_WAIT_PSN)) &&
1804 (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) {
1805 if (opcode == TID_OP(READ_REQ))
1806 rvt_add_retry_timer_ext(qp, priv->timeout_shift);
1807 else
1808 rvt_add_retry_timer(qp);
1809 }
1810
1811 /* Start TID RDMA ACK timer */
1812 if ((opcode == TID_OP(WRITE_DATA) ||
1813 opcode == TID_OP(WRITE_DATA_LAST) ||
1814 opcode == TID_OP(RESYNC)) &&
1815 (psn & IB_BTH_REQ_ACK) &&
1816 !(priv->s_flags & HFI1_S_TID_RETRY_TIMER) &&
1817 (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) {
1818 /*
1819 * The TID RDMA ACK packet could be received before this
1820 * function is called. Therefore, add the timer only if TID
1821 * RDMA ACK packets are actually pending.
1822 */
1823 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1824 req = wqe_to_tid_req(wqe);
1825 if (wqe->wr.opcode == IB_WR_TID_RDMA_WRITE &&
1826 req->ack_seg < req->cur_seg)
1827 hfi1_add_tid_retry_timer(qp);
1828 }
1829
1830 while (qp->s_last != qp->s_acked) {
1831 wqe = rvt_get_swqe_ptr(qp, qp->s_last);
1832 if (cmp_psn(wqe->lpsn, qp->s_sending_psn) >= 0 &&
1833 cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) <= 0)
1834 break;
1835 trdma_clean_swqe(qp, wqe);
1836 trace_hfi1_qp_send_completion(qp, wqe, qp->s_last);
1837 rvt_qp_complete_swqe(qp,
1838 wqe,
1839 ib_hfi1_wc_opcode[wqe->wr.opcode],
1840 IB_WC_SUCCESS);
1841 }
1842 /*
1843 * If we were waiting for sends to complete before re-sending,
1844 * and they are now complete, restart sending.
1845 */
1846 trace_hfi1_sendcomplete(qp, psn);
1847 if (qp->s_flags & RVT_S_WAIT_PSN &&
1848 cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) > 0) {
1849 qp->s_flags &= ~RVT_S_WAIT_PSN;
1850 qp->s_sending_psn = qp->s_psn;
1851 qp->s_sending_hpsn = qp->s_psn - 1;
1852 hfi1_schedule_send(qp);
1853 }
1854 }
1855
1856 static inline void update_last_psn(struct rvt_qp *qp, u32 psn)
1857 {
1858 qp->s_last_psn = psn;
1859 }
1860
1861 /*
1862 * Generate a SWQE completion.
1863 * This is similar to hfi1_send_complete but has to check to be sure
1864 * that the SGEs are not being referenced if the SWQE is being resent.
1865 */
1866 struct rvt_swqe *do_rc_completion(struct rvt_qp *qp,
1867 struct rvt_swqe *wqe,
1868 struct hfi1_ibport *ibp)
1869 {
1870 struct hfi1_qp_priv *priv = qp->priv;
1871
1872 lockdep_assert_held(&qp->s_lock);
1873 /*
1874 * Don't decrement refcount and don't generate a
1875 * completion if the SWQE is being resent until the send
1876 * is finished.
1877 */
1878 trace_hfi1_rc_completion(qp, wqe->lpsn);
1879 if (cmp_psn(wqe->lpsn, qp->s_sending_psn) < 0 ||
1880 cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) > 0) {
1881 trdma_clean_swqe(qp, wqe);
1882 trace_hfi1_qp_send_completion(qp, wqe, qp->s_last);
1883 rvt_qp_complete_swqe(qp,
1884 wqe,
1885 ib_hfi1_wc_opcode[wqe->wr.opcode],
1886 IB_WC_SUCCESS);
1887 } else {
1888 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
1889
1890 this_cpu_inc(*ibp->rvp.rc_delayed_comp);
1891 /*
1892 * If send progress not running attempt to progress
1893 * SDMA queue.
1894 */
1895 if (ppd->dd->flags & HFI1_HAS_SEND_DMA) {
1896 struct sdma_engine *engine;
1897 u8 sl = rdma_ah_get_sl(&qp->remote_ah_attr);
1898 u8 sc5;
1899
1900 /* For now use sc to find engine */
1901 sc5 = ibp->sl_to_sc[sl];
1902 engine = qp_to_sdma_engine(qp, sc5);
1903 sdma_engine_progress_schedule(engine);
1904 }
1905 }
1906
1907 qp->s_retry = qp->s_retry_cnt;
1908 /*
1909 * Don't update the last PSN if the request being completed is
1910 * a TID RDMA WRITE request.
1911 * Completion of the TID RDMA WRITE requests are done by the
1912 * TID RDMA ACKs and as such could be for a request that has
1913 * already been ACKed as far as the IB state machine is
1914 * concerned.
1915 */
1916 if (wqe->wr.opcode != IB_WR_TID_RDMA_WRITE)
1917 update_last_psn(qp, wqe->lpsn);
1918
1919 /*
1920 * If we are completing a request which is in the process of
1921 * being resent, we can stop re-sending it since we know the
1922 * responder has already seen it.
1923 */
1924 if (qp->s_acked == qp->s_cur) {
1925 if (++qp->s_cur >= qp->s_size)
1926 qp->s_cur = 0;
1927 qp->s_acked = qp->s_cur;
1928 wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
1929 if (qp->s_acked != qp->s_tail) {
1930 qp->s_state = OP(SEND_LAST);
1931 qp->s_psn = wqe->psn;
1932 }
1933 } else {
1934 if (++qp->s_acked >= qp->s_size)
1935 qp->s_acked = 0;
1936 if (qp->state == IB_QPS_SQD && qp->s_acked == qp->s_cur)
1937 qp->s_draining = 0;
1938 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1939 }
1940 if (priv->s_flags & HFI1_S_TID_WAIT_INTERLCK) {
1941 priv->s_flags &= ~HFI1_S_TID_WAIT_INTERLCK;
1942 hfi1_schedule_send(qp);
1943 }
1944 return wqe;
1945 }
1946
1947 static void set_restart_qp(struct rvt_qp *qp, struct hfi1_ctxtdata *rcd)
1948 {
1949 /* Retry this request. */
1950 if (!(qp->r_flags & RVT_R_RDMAR_SEQ)) {
1951 qp->r_flags |= RVT_R_RDMAR_SEQ;
1952 hfi1_restart_rc(qp, qp->s_last_psn + 1, 0);
1953 if (list_empty(&qp->rspwait)) {
1954 qp->r_flags |= RVT_R_RSP_SEND;
1955 rvt_get_qp(qp);
1956 list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
1957 }
1958 }
1959 }
1960
1961 /**
1962 * update_qp_retry_state - Update qp retry state.
1963 * @qp: the QP
1964 * @psn: the packet sequence number of the TID RDMA WRITE RESP.
1965 * @spsn: The start psn for the given TID RDMA WRITE swqe.
1966 * @lpsn: The last psn for the given TID RDMA WRITE swqe.
1967 *
1968 * This function is called to update the qp retry state upon
1969 * receiving a TID WRITE RESP after the qp is scheduled to retry
1970 * a request.
1971 */
1972 static void update_qp_retry_state(struct rvt_qp *qp, u32 psn, u32 spsn,
1973 u32 lpsn)
1974 {
1975 struct hfi1_qp_priv *qpriv = qp->priv;
1976
1977 qp->s_psn = psn + 1;
1978 /*
1979 * If this is the first TID RDMA WRITE RESP packet for the current
1980 * request, change the s_state so that the retry will be processed
1981 * correctly. Similarly, if this is the last TID RDMA WRITE RESP
1982 * packet, change the s_state and advance the s_cur.
1983 */
1984 if (cmp_psn(psn, lpsn) >= 0) {
1985 qp->s_cur = qpriv->s_tid_cur + 1;
1986 if (qp->s_cur >= qp->s_size)
1987 qp->s_cur = 0;
1988 qp->s_state = TID_OP(WRITE_REQ);
1989 } else if (!cmp_psn(psn, spsn)) {
1990 qp->s_cur = qpriv->s_tid_cur;
1991 qp->s_state = TID_OP(WRITE_RESP);
1992 }
1993 }
1994
1995 /**
1996 * do_rc_ack - process an incoming RC ACK
1997 * @qp: the QP the ACK came in on
1998 * @psn: the packet sequence number of the ACK
1999 * @opcode: the opcode of the request that resulted in the ACK
2000 *
2001 * This is called from rc_rcv_resp() to process an incoming RC ACK
2002 * for the given QP.
2003 * May be called at interrupt level, with the QP s_lock held.
2004 * Returns 1 if OK, 0 if current operation should be aborted (NAK).
2005 */
2006 int do_rc_ack(struct rvt_qp *qp, u32 aeth, u32 psn, int opcode,
2007 u64 val, struct hfi1_ctxtdata *rcd)
2008 {
2009 struct hfi1_ibport *ibp;
2010 enum ib_wc_status status;
2011 struct hfi1_qp_priv *qpriv = qp->priv;
2012 struct rvt_swqe *wqe;
2013 int ret = 0;
2014 u32 ack_psn;
2015 int diff;
2016 struct rvt_dev_info *rdi;
2017
2018 lockdep_assert_held(&qp->s_lock);
2019 /*
2020 * Note that NAKs implicitly ACK outstanding SEND and RDMA write
2021 * requests and implicitly NAK RDMA read and atomic requests issued
2022 * before the NAK'ed request. The MSN won't include the NAK'ed
2023 * request but will include an ACK'ed request(s).
2024 */
2025 ack_psn = psn;
2026 if (aeth >> IB_AETH_NAK_SHIFT)
2027 ack_psn--;
2028 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
2029 ibp = rcd_to_iport(rcd);
2030
2031 /*
2032 * The MSN might be for a later WQE than the PSN indicates so
2033 * only complete WQEs that the PSN finishes.
2034 */
2035 while ((diff = delta_psn(ack_psn, wqe->lpsn)) >= 0) {
2036 /*
2037 * RDMA_READ_RESPONSE_ONLY is a special case since
2038 * we want to generate completion events for everything
2039 * before the RDMA read, copy the data, then generate
2040 * the completion for the read.
2041 */
2042 if (wqe->wr.opcode == IB_WR_RDMA_READ &&
2043 opcode == OP(RDMA_READ_RESPONSE_ONLY) &&
2044 diff == 0) {
2045 ret = 1;
2046 goto bail_stop;
2047 }
2048 /*
2049 * If this request is a RDMA read or atomic, and the ACK is
2050 * for a later operation, this ACK NAKs the RDMA read or
2051 * atomic. In other words, only a RDMA_READ_LAST or ONLY
2052 * can ACK a RDMA read and likewise for atomic ops. Note
2053 * that the NAK case can only happen if relaxed ordering is
2054 * used and requests are sent after an RDMA read or atomic
2055 * is sent but before the response is received.
2056 */
2057 if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
2058 (opcode != OP(RDMA_READ_RESPONSE_LAST) || diff != 0)) ||
2059 (wqe->wr.opcode == IB_WR_TID_RDMA_READ &&
2060 (opcode != TID_OP(READ_RESP) || diff != 0)) ||
2061 ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
2062 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
2063 (opcode != OP(ATOMIC_ACKNOWLEDGE) || diff != 0)) ||
2064 (wqe->wr.opcode == IB_WR_TID_RDMA_WRITE &&
2065 (delta_psn(psn, qp->s_last_psn) != 1))) {
2066 set_restart_qp(qp, rcd);
2067 /*
2068 * No need to process the ACK/NAK since we are
2069 * restarting an earlier request.
2070 */
2071 goto bail_stop;
2072 }
2073 if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
2074 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
2075 u64 *vaddr = wqe->sg_list[0].vaddr;
2076 *vaddr = val;
2077 }
2078 if (wqe->wr.opcode == IB_WR_OPFN)
2079 opfn_conn_reply(qp, val);
2080
2081 if (qp->s_num_rd_atomic &&
2082 (wqe->wr.opcode == IB_WR_RDMA_READ ||
2083 wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
2084 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)) {
2085 qp->s_num_rd_atomic--;
2086 /* Restart sending task if fence is complete */
2087 if ((qp->s_flags & RVT_S_WAIT_FENCE) &&
2088 !qp->s_num_rd_atomic) {
2089 qp->s_flags &= ~(RVT_S_WAIT_FENCE |
2090 RVT_S_WAIT_ACK);
2091 hfi1_schedule_send(qp);
2092 } else if (qp->s_flags & RVT_S_WAIT_RDMAR) {
2093 qp->s_flags &= ~(RVT_S_WAIT_RDMAR |
2094 RVT_S_WAIT_ACK);
2095 hfi1_schedule_send(qp);
2096 }
2097 }
2098
2099 /*
2100 * TID RDMA WRITE requests will be completed by the TID RDMA
2101 * ACK packet handler (see tid_rdma.c).
2102 */
2103 if (wqe->wr.opcode == IB_WR_TID_RDMA_WRITE)
2104 break;
2105
2106 wqe = do_rc_completion(qp, wqe, ibp);
2107 if (qp->s_acked == qp->s_tail)
2108 break;
2109 }
2110
2111 trace_hfi1_rc_ack_do(qp, aeth, psn, wqe);
2112 trace_hfi1_sender_do_rc_ack(qp);
2113 switch (aeth >> IB_AETH_NAK_SHIFT) {
2114 case 0: /* ACK */
2115 this_cpu_inc(*ibp->rvp.rc_acks);
2116 if (wqe->wr.opcode == IB_WR_TID_RDMA_READ) {
2117 if (wqe_to_tid_req(wqe)->ack_pending)
2118 rvt_mod_retry_timer_ext(qp,
2119 qpriv->timeout_shift);
2120 else
2121 rvt_stop_rc_timers(qp);
2122 } else if (qp->s_acked != qp->s_tail) {
2123 struct rvt_swqe *__w = NULL;
2124
2125 if (qpriv->s_tid_cur != HFI1_QP_WQE_INVALID)
2126 __w = rvt_get_swqe_ptr(qp, qpriv->s_tid_cur);
2127
2128 /*
2129 * Stop timers if we've received all of the TID RDMA
2130 * WRITE * responses.
2131 */
2132 if (__w && __w->wr.opcode == IB_WR_TID_RDMA_WRITE &&
2133 opcode == TID_OP(WRITE_RESP)) {
2134 /*
2135 * Normally, the loop above would correctly
2136 * process all WQEs from s_acked onward and
2137 * either complete them or check for correct
2138 * PSN sequencing.
2139 * However, for TID RDMA, due to pipelining,
2140 * the response may not be for the request at
2141 * s_acked so the above look would just be
2142 * skipped. This does not allow for checking
2143 * the PSN sequencing. It has to be done
2144 * separately.
2145 */
2146 if (cmp_psn(psn, qp->s_last_psn + 1)) {
2147 set_restart_qp(qp, rcd);
2148 goto bail_stop;
2149 }
2150 /*
2151 * If the psn is being resent, stop the
2152 * resending.
2153 */
2154 if (qp->s_cur != qp->s_tail &&
2155 cmp_psn(qp->s_psn, psn) <= 0)
2156 update_qp_retry_state(qp, psn,
2157 __w->psn,
2158 __w->lpsn);
2159 else if (--qpriv->pending_tid_w_resp)
2160 rvt_mod_retry_timer(qp);
2161 else
2162 rvt_stop_rc_timers(qp);
2163 } else {
2164 /*
2165 * We are expecting more ACKs so
2166 * mod the retry timer.
2167 */
2168 rvt_mod_retry_timer(qp);
2169 /*
2170 * We can stop re-sending the earlier packets
2171 * and continue with the next packet the
2172 * receiver wants.
2173 */
2174 if (cmp_psn(qp->s_psn, psn) <= 0)
2175 reset_psn(qp, psn + 1);
2176 }
2177 } else {
2178 /* No more acks - kill all timers */
2179 rvt_stop_rc_timers(qp);
2180 if (cmp_psn(qp->s_psn, psn) <= 0) {
2181 qp->s_state = OP(SEND_LAST);
2182 qp->s_psn = psn + 1;
2183 }
2184 }
2185 if (qp->s_flags & RVT_S_WAIT_ACK) {
2186 qp->s_flags &= ~RVT_S_WAIT_ACK;
2187 hfi1_schedule_send(qp);
2188 }
2189 rvt_get_credit(qp, aeth);
2190 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
2191 qp->s_retry = qp->s_retry_cnt;
2192 /*
2193 * If the current request is a TID RDMA WRITE request and the
2194 * response is not a TID RDMA WRITE RESP packet, s_last_psn
2195 * can't be advanced.
2196 */
2197 if (wqe->wr.opcode == IB_WR_TID_RDMA_WRITE &&
2198 opcode != TID_OP(WRITE_RESP) &&
2199 cmp_psn(psn, wqe->psn) >= 0)
2200 return 1;
2201 update_last_psn(qp, psn);
2202 return 1;
2203
2204 case 1: /* RNR NAK */
2205 ibp->rvp.n_rnr_naks++;
2206 if (qp->s_acked == qp->s_tail)
2207 goto bail_stop;
2208 if (qp->s_flags & RVT_S_WAIT_RNR)
2209 goto bail_stop;
2210 rdi = ib_to_rvt(qp->ibqp.device);
2211 if (!(rdi->post_parms[wqe->wr.opcode].flags &
2212 RVT_OPERATION_IGN_RNR_CNT)) {
2213 if (qp->s_rnr_retry == 0) {
2214 status = IB_WC_RNR_RETRY_EXC_ERR;
2215 goto class_b;
2216 }
2217 if (qp->s_rnr_retry_cnt < 7 && qp->s_rnr_retry_cnt > 0)
2218 qp->s_rnr_retry--;
2219 }
2220
2221 /*
2222 * The last valid PSN is the previous PSN. For TID RDMA WRITE
2223 * request, s_last_psn should be incremented only when a TID
2224 * RDMA WRITE RESP is received to avoid skipping lost TID RDMA
2225 * WRITE RESP packets.
2226 */
2227 if (wqe->wr.opcode == IB_WR_TID_RDMA_WRITE) {
2228 reset_psn(qp, qp->s_last_psn + 1);
2229 } else {
2230 update_last_psn(qp, psn - 1);
2231 reset_psn(qp, psn);
2232 }
2233
2234 ibp->rvp.n_rc_resends += delta_psn(qp->s_psn, psn);
2235 qp->s_flags &= ~(RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_ACK);
2236 rvt_stop_rc_timers(qp);
2237 rvt_add_rnr_timer(qp, aeth);
2238 return 0;
2239
2240 case 3: /* NAK */
2241 if (qp->s_acked == qp->s_tail)
2242 goto bail_stop;
2243 /* The last valid PSN is the previous PSN. */
2244 update_last_psn(qp, psn - 1);
2245 switch ((aeth >> IB_AETH_CREDIT_SHIFT) &
2246 IB_AETH_CREDIT_MASK) {
2247 case 0: /* PSN sequence error */
2248 ibp->rvp.n_seq_naks++;
2249 /*
2250 * Back up to the responder's expected PSN.
2251 * Note that we might get a NAK in the middle of an
2252 * RDMA READ response which terminates the RDMA
2253 * READ.
2254 */
2255 hfi1_restart_rc(qp, psn, 0);
2256 hfi1_schedule_send(qp);
2257 break;
2258
2259 case 1: /* Invalid Request */
2260 status = IB_WC_REM_INV_REQ_ERR;
2261 ibp->rvp.n_other_naks++;
2262 goto class_b;
2263
2264 case 2: /* Remote Access Error */
2265 status = IB_WC_REM_ACCESS_ERR;
2266 ibp->rvp.n_other_naks++;
2267 goto class_b;
2268
2269 case 3: /* Remote Operation Error */
2270 status = IB_WC_REM_OP_ERR;
2271 ibp->rvp.n_other_naks++;
2272 class_b:
2273 if (qp->s_last == qp->s_acked) {
2274 if (wqe->wr.opcode == IB_WR_TID_RDMA_READ)
2275 hfi1_kern_read_tid_flow_free(qp);
2276
2277 hfi1_trdma_send_complete(qp, wqe, status);
2278 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
2279 }
2280 break;
2281
2282 default:
2283 /* Ignore other reserved NAK error codes */
2284 goto reserved;
2285 }
2286 qp->s_retry = qp->s_retry_cnt;
2287 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
2288 goto bail_stop;
2289
2290 default: /* 2: reserved */
2291 reserved:
2292 /* Ignore reserved NAK codes. */
2293 goto bail_stop;
2294 }
2295 /* cannot be reached */
2296 bail_stop:
2297 rvt_stop_rc_timers(qp);
2298 return ret;
2299 }
2300
2301 /*
2302 * We have seen an out of sequence RDMA read middle or last packet.
2303 * This ACKs SENDs and RDMA writes up to the first RDMA read or atomic SWQE.
2304 */
2305 static void rdma_seq_err(struct rvt_qp *qp, struct hfi1_ibport *ibp, u32 psn,
2306 struct hfi1_ctxtdata *rcd)
2307 {
2308 struct rvt_swqe *wqe;
2309
2310 lockdep_assert_held(&qp->s_lock);
2311 /* Remove QP from retry timer */
2312 rvt_stop_rc_timers(qp);
2313
2314 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
2315
2316 while (cmp_psn(psn, wqe->lpsn) > 0) {
2317 if (wqe->wr.opcode == IB_WR_RDMA_READ ||
2318 wqe->wr.opcode == IB_WR_TID_RDMA_READ ||
2319 wqe->wr.opcode == IB_WR_TID_RDMA_WRITE ||
2320 wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
2321 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
2322 break;
2323 wqe = do_rc_completion(qp, wqe, ibp);
2324 }
2325
2326 ibp->rvp.n_rdma_seq++;
2327 qp->r_flags |= RVT_R_RDMAR_SEQ;
2328 hfi1_restart_rc(qp, qp->s_last_psn + 1, 0);
2329 if (list_empty(&qp->rspwait)) {
2330 qp->r_flags |= RVT_R_RSP_SEND;
2331 rvt_get_qp(qp);
2332 list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
2333 }
2334 }
2335
2336 /**
2337 * rc_rcv_resp - process an incoming RC response packet
2338 * @packet: data packet information
2339 *
2340 * This is called from hfi1_rc_rcv() to process an incoming RC response
2341 * packet for the given QP.
2342 * Called at interrupt level.
2343 */
2344 static void rc_rcv_resp(struct hfi1_packet *packet)
2345 {
2346 struct hfi1_ctxtdata *rcd = packet->rcd;
2347 void *data = packet->payload;
2348 u32 tlen = packet->tlen;
2349 struct rvt_qp *qp = packet->qp;
2350 struct hfi1_ibport *ibp;
2351 struct ib_other_headers *ohdr = packet->ohdr;
2352 struct rvt_swqe *wqe;
2353 enum ib_wc_status status;
2354 unsigned long flags;
2355 int diff;
2356 u64 val;
2357 u32 aeth;
2358 u32 psn = ib_bth_get_psn(packet->ohdr);
2359 u32 pmtu = qp->pmtu;
2360 u16 hdrsize = packet->hlen;
2361 u8 opcode = packet->opcode;
2362 u8 pad = packet->pad;
2363 u8 extra_bytes = pad + packet->extra_byte + (SIZE_OF_CRC << 2);
2364
2365 spin_lock_irqsave(&qp->s_lock, flags);
2366 trace_hfi1_ack(qp, psn);
2367
2368 /* Ignore invalid responses. */
2369 if (cmp_psn(psn, READ_ONCE(qp->s_next_psn)) >= 0)
2370 goto ack_done;
2371
2372 /* Ignore duplicate responses. */
2373 diff = cmp_psn(psn, qp->s_last_psn);
2374 if (unlikely(diff <= 0)) {
2375 /* Update credits for "ghost" ACKs */
2376 if (diff == 0 && opcode == OP(ACKNOWLEDGE)) {
2377 aeth = be32_to_cpu(ohdr->u.aeth);
2378 if ((aeth >> IB_AETH_NAK_SHIFT) == 0)
2379 rvt_get_credit(qp, aeth);
2380 }
2381 goto ack_done;
2382 }
2383
2384 /*
2385 * Skip everything other than the PSN we expect, if we are waiting
2386 * for a reply to a restarted RDMA read or atomic op.
2387 */
2388 if (qp->r_flags & RVT_R_RDMAR_SEQ) {
2389 if (cmp_psn(psn, qp->s_last_psn + 1) != 0)
2390 goto ack_done;
2391 qp->r_flags &= ~RVT_R_RDMAR_SEQ;
2392 }
2393
2394 if (unlikely(qp->s_acked == qp->s_tail))
2395 goto ack_done;
2396 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
2397 status = IB_WC_SUCCESS;
2398
2399 switch (opcode) {
2400 case OP(ACKNOWLEDGE):
2401 case OP(ATOMIC_ACKNOWLEDGE):
2402 case OP(RDMA_READ_RESPONSE_FIRST):
2403 aeth = be32_to_cpu(ohdr->u.aeth);
2404 if (opcode == OP(ATOMIC_ACKNOWLEDGE))
2405 val = ib_u64_get(&ohdr->u.at.atomic_ack_eth);
2406 else
2407 val = 0;
2408 if (!do_rc_ack(qp, aeth, psn, opcode, val, rcd) ||
2409 opcode != OP(RDMA_READ_RESPONSE_FIRST))
2410 goto ack_done;
2411 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
2412 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
2413 goto ack_op_err;
2414 /*
2415 * If this is a response to a resent RDMA read, we
2416 * have to be careful to copy the data to the right
2417 * location.
2418 */
2419 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
2420 wqe, psn, pmtu);
2421 goto read_middle;
2422
2423 case OP(RDMA_READ_RESPONSE_MIDDLE):
2424 /* no AETH, no ACK */
2425 if (unlikely(cmp_psn(psn, qp->s_last_psn + 1)))
2426 goto ack_seq_err;
2427 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
2428 goto ack_op_err;
2429 read_middle:
2430 if (unlikely(tlen != (hdrsize + pmtu + extra_bytes)))
2431 goto ack_len_err;
2432 if (unlikely(pmtu >= qp->s_rdma_read_len))
2433 goto ack_len_err;
2434
2435 /*
2436 * We got a response so update the timeout.
2437 * 4.096 usec. * (1 << qp->timeout)
2438 */
2439 rvt_mod_retry_timer(qp);
2440 if (qp->s_flags & RVT_S_WAIT_ACK) {
2441 qp->s_flags &= ~RVT_S_WAIT_ACK;
2442 hfi1_schedule_send(qp);
2443 }
2444
2445 if (opcode == OP(RDMA_READ_RESPONSE_MIDDLE))
2446 qp->s_retry = qp->s_retry_cnt;
2447
2448 /*
2449 * Update the RDMA receive state but do the copy w/o
2450 * holding the locks and blocking interrupts.
2451 */
2452 qp->s_rdma_read_len -= pmtu;
2453 update_last_psn(qp, psn);
2454 spin_unlock_irqrestore(&qp->s_lock, flags);
2455 rvt_copy_sge(qp, &qp->s_rdma_read_sge,
2456 data, pmtu, false, false);
2457 goto bail;
2458
2459 case OP(RDMA_READ_RESPONSE_ONLY):
2460 aeth = be32_to_cpu(ohdr->u.aeth);
2461 if (!do_rc_ack(qp, aeth, psn, opcode, 0, rcd))
2462 goto ack_done;
2463 /*
2464 * Check that the data size is >= 0 && <= pmtu.
2465 * Remember to account for ICRC (4).
2466 */
2467 if (unlikely(tlen < (hdrsize + extra_bytes)))
2468 goto ack_len_err;
2469 /*
2470 * If this is a response to a resent RDMA read, we
2471 * have to be careful to copy the data to the right
2472 * location.
2473 */
2474 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
2475 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
2476 wqe, psn, pmtu);
2477 goto read_last;
2478
2479 case OP(RDMA_READ_RESPONSE_LAST):
2480 /* ACKs READ req. */
2481 if (unlikely(cmp_psn(psn, qp->s_last_psn + 1)))
2482 goto ack_seq_err;
2483 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
2484 goto ack_op_err;
2485 /*
2486 * Check that the data size is >= 1 && <= pmtu.
2487 * Remember to account for ICRC (4).
2488 */
2489 if (unlikely(tlen <= (hdrsize + extra_bytes)))
2490 goto ack_len_err;
2491 read_last:
2492 tlen -= hdrsize + extra_bytes;
2493 if (unlikely(tlen != qp->s_rdma_read_len))
2494 goto ack_len_err;
2495 aeth = be32_to_cpu(ohdr->u.aeth);
2496 rvt_copy_sge(qp, &qp->s_rdma_read_sge,
2497 data, tlen, false, false);
2498 WARN_ON(qp->s_rdma_read_sge.num_sge);
2499 (void)do_rc_ack(qp, aeth, psn,
2500 OP(RDMA_READ_RESPONSE_LAST), 0, rcd);
2501 goto ack_done;
2502 }
2503
2504 ack_op_err:
2505 status = IB_WC_LOC_QP_OP_ERR;
2506 goto ack_err;
2507
2508 ack_seq_err:
2509 ibp = rcd_to_iport(rcd);
2510 rdma_seq_err(qp, ibp, psn, rcd);
2511 goto ack_done;
2512
2513 ack_len_err:
2514 status = IB_WC_LOC_LEN_ERR;
2515 ack_err:
2516 if (qp->s_last == qp->s_acked) {
2517 rvt_send_complete(qp, wqe, status);
2518 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
2519 }
2520 ack_done:
2521 spin_unlock_irqrestore(&qp->s_lock, flags);
2522 bail:
2523 return;
2524 }
2525
2526 static inline void rc_cancel_ack(struct rvt_qp *qp)
2527 {
2528 qp->r_adefered = 0;
2529 if (list_empty(&qp->rspwait))
2530 return;
2531 list_del_init(&qp->rspwait);
2532 qp->r_flags &= ~RVT_R_RSP_NAK;
2533 rvt_put_qp(qp);
2534 }
2535
2536 /**
2537 * rc_rcv_error - process an incoming duplicate or error RC packet
2538 * @ohdr: the other headers for this packet
2539 * @data: the packet data
2540 * @qp: the QP for this packet
2541 * @opcode: the opcode for this packet
2542 * @psn: the packet sequence number for this packet
2543 * @diff: the difference between the PSN and the expected PSN
2544 *
2545 * This is called from hfi1_rc_rcv() to process an unexpected
2546 * incoming RC packet for the given QP.
2547 * Called at interrupt level.
2548 * Return 1 if no more processing is needed; otherwise return 0 to
2549 * schedule a response to be sent.
2550 */
2551 static noinline int rc_rcv_error(struct ib_other_headers *ohdr, void *data,
2552 struct rvt_qp *qp, u32 opcode, u32 psn,
2553 int diff, struct hfi1_ctxtdata *rcd)
2554 {
2555 struct hfi1_ibport *ibp = rcd_to_iport(rcd);
2556 struct rvt_ack_entry *e;
2557 unsigned long flags;
2558 u8 prev;
2559 u8 mra; /* most recent ACK */
2560 bool old_req;
2561
2562 trace_hfi1_rcv_error(qp, psn);
2563 if (diff > 0) {
2564 /*
2565 * Packet sequence error.
2566 * A NAK will ACK earlier sends and RDMA writes.
2567 * Don't queue the NAK if we already sent one.
2568 */
2569 if (!qp->r_nak_state) {
2570 ibp->rvp.n_rc_seqnak++;
2571 qp->r_nak_state = IB_NAK_PSN_ERROR;
2572 /* Use the expected PSN. */
2573 qp->r_ack_psn = qp->r_psn;
2574 /*
2575 * Wait to send the sequence NAK until all packets
2576 * in the receive queue have been processed.
2577 * Otherwise, we end up propagating congestion.
2578 */
2579 rc_defered_ack(rcd, qp);
2580 }
2581 goto done;
2582 }
2583
2584 /*
2585 * Handle a duplicate request. Don't re-execute SEND, RDMA
2586 * write or atomic op. Don't NAK errors, just silently drop
2587 * the duplicate request. Note that r_sge, r_len, and
2588 * r_rcv_len may be in use so don't modify them.
2589 *
2590 * We are supposed to ACK the earliest duplicate PSN but we
2591 * can coalesce an outstanding duplicate ACK. We have to
2592 * send the earliest so that RDMA reads can be restarted at
2593 * the requester's expected PSN.
2594 *
2595 * First, find where this duplicate PSN falls within the
2596 * ACKs previously sent.
2597 * old_req is true if there is an older response that is scheduled
2598 * to be sent before sending this one.
2599 */
2600 e = NULL;
2601 old_req = true;
2602 ibp->rvp.n_rc_dupreq++;
2603
2604 spin_lock_irqsave(&qp->s_lock, flags);
2605
2606 e = find_prev_entry(qp, psn, &prev, &mra, &old_req);
2607
2608 switch (opcode) {
2609 case OP(RDMA_READ_REQUEST): {
2610 struct ib_reth *reth;
2611 u32 offset;
2612 u32 len;
2613
2614 /*
2615 * If we didn't find the RDMA read request in the ack queue,
2616 * we can ignore this request.
2617 */
2618 if (!e || e->opcode != OP(RDMA_READ_REQUEST))
2619 goto unlock_done;
2620 /* RETH comes after BTH */
2621 reth = &ohdr->u.rc.reth;
2622 /*
2623 * Address range must be a subset of the original
2624 * request and start on pmtu boundaries.
2625 * We reuse the old ack_queue slot since the requester
2626 * should not back up and request an earlier PSN for the
2627 * same request.
2628 */
2629 offset = delta_psn(psn, e->psn) * qp->pmtu;
2630 len = be32_to_cpu(reth->length);
2631 if (unlikely(offset + len != e->rdma_sge.sge_length))
2632 goto unlock_done;
2633 release_rdma_sge_mr(e);
2634 if (len != 0) {
2635 u32 rkey = be32_to_cpu(reth->rkey);
2636 u64 vaddr = get_ib_reth_vaddr(reth);
2637 int ok;
2638
2639 ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr, rkey,
2640 IB_ACCESS_REMOTE_READ);
2641 if (unlikely(!ok))
2642 goto unlock_done;
2643 } else {
2644 e->rdma_sge.vaddr = NULL;
2645 e->rdma_sge.length = 0;
2646 e->rdma_sge.sge_length = 0;
2647 }
2648 e->psn = psn;
2649 if (old_req)
2650 goto unlock_done;
2651 if (qp->s_acked_ack_queue == qp->s_tail_ack_queue)
2652 qp->s_acked_ack_queue = prev;
2653 qp->s_tail_ack_queue = prev;
2654 break;
2655 }
2656
2657 case OP(COMPARE_SWAP):
2658 case OP(FETCH_ADD): {
2659 /*
2660 * If we didn't find the atomic request in the ack queue
2661 * or the send engine is already backed up to send an
2662 * earlier entry, we can ignore this request.
2663 */
2664 if (!e || e->opcode != (u8)opcode || old_req)
2665 goto unlock_done;
2666 if (qp->s_tail_ack_queue == qp->s_acked_ack_queue)
2667 qp->s_acked_ack_queue = prev;
2668 qp->s_tail_ack_queue = prev;
2669 break;
2670 }
2671
2672 default:
2673 /*
2674 * Ignore this operation if it doesn't request an ACK
2675 * or an earlier RDMA read or atomic is going to be resent.
2676 */
2677 if (!(psn & IB_BTH_REQ_ACK) || old_req)
2678 goto unlock_done;
2679 /*
2680 * Resend the most recent ACK if this request is
2681 * after all the previous RDMA reads and atomics.
2682 */
2683 if (mra == qp->r_head_ack_queue) {
2684 spin_unlock_irqrestore(&qp->s_lock, flags);
2685 qp->r_nak_state = 0;
2686 qp->r_ack_psn = qp->r_psn - 1;
2687 goto send_ack;
2688 }
2689
2690 /*
2691 * Resend the RDMA read or atomic op which
2692 * ACKs this duplicate request.
2693 */
2694 if (qp->s_tail_ack_queue == qp->s_acked_ack_queue)
2695 qp->s_acked_ack_queue = mra;
2696 qp->s_tail_ack_queue = mra;
2697 break;
2698 }
2699 qp->s_ack_state = OP(ACKNOWLEDGE);
2700 qp->s_flags |= RVT_S_RESP_PENDING;
2701 qp->r_nak_state = 0;
2702 hfi1_schedule_send(qp);
2703
2704 unlock_done:
2705 spin_unlock_irqrestore(&qp->s_lock, flags);
2706 done:
2707 return 1;
2708
2709 send_ack:
2710 return 0;
2711 }
2712
2713 static void log_cca_event(struct hfi1_pportdata *ppd, u8 sl, u32 rlid,
2714 u32 lqpn, u32 rqpn, u8 svc_type)
2715 {
2716 struct opa_hfi1_cong_log_event_internal *cc_event;
2717 unsigned long flags;
2718
2719 if (sl >= OPA_MAX_SLS)
2720 return;
2721
2722 spin_lock_irqsave(&ppd->cc_log_lock, flags);
2723
2724 ppd->threshold_cong_event_map[sl / 8] |= 1 << (sl % 8);
2725 ppd->threshold_event_counter++;
2726
2727 cc_event = &ppd->cc_events[ppd->cc_log_idx++];
2728 if (ppd->cc_log_idx == OPA_CONG_LOG_ELEMS)
2729 ppd->cc_log_idx = 0;
2730 cc_event->lqpn = lqpn & RVT_QPN_MASK;
2731 cc_event->rqpn = rqpn & RVT_QPN_MASK;
2732 cc_event->sl = sl;
2733 cc_event->svc_type = svc_type;
2734 cc_event->rlid = rlid;
2735 /* keep timestamp in units of 1.024 usec */
2736 cc_event->timestamp = ktime_get_ns() / 1024;
2737
2738 spin_unlock_irqrestore(&ppd->cc_log_lock, flags);
2739 }
2740
2741 void process_becn(struct hfi1_pportdata *ppd, u8 sl, u32 rlid, u32 lqpn,
2742 u32 rqpn, u8 svc_type)
2743 {
2744 struct cca_timer *cca_timer;
2745 u16 ccti, ccti_incr, ccti_timer, ccti_limit;
2746 u8 trigger_threshold;
2747 struct cc_state *cc_state;
2748 unsigned long flags;
2749
2750 if (sl >= OPA_MAX_SLS)
2751 return;
2752
2753 cc_state = get_cc_state(ppd);
2754
2755 if (!cc_state)
2756 return;
2757
2758 /*
2759 * 1) increase CCTI (for this SL)
2760 * 2) select IPG (i.e., call set_link_ipg())
2761 * 3) start timer
2762 */
2763 ccti_limit = cc_state->cct.ccti_limit;
2764 ccti_incr = cc_state->cong_setting.entries[sl].ccti_increase;
2765 ccti_timer = cc_state->cong_setting.entries[sl].ccti_timer;
2766 trigger_threshold =
2767 cc_state->cong_setting.entries[sl].trigger_threshold;
2768
2769 spin_lock_irqsave(&ppd->cca_timer_lock, flags);
2770
2771 cca_timer = &ppd->cca_timer[sl];
2772 if (cca_timer->ccti < ccti_limit) {
2773 if (cca_timer->ccti + ccti_incr <= ccti_limit)
2774 cca_timer->ccti += ccti_incr;
2775 else
2776 cca_timer->ccti = ccti_limit;
2777 set_link_ipg(ppd);
2778 }
2779
2780 ccti = cca_timer->ccti;
2781
2782 if (!hrtimer_active(&cca_timer->hrtimer)) {
2783 /* ccti_timer is in units of 1.024 usec */
2784 unsigned long nsec = 1024 * ccti_timer;
2785
2786 hrtimer_start(&cca_timer->hrtimer, ns_to_ktime(nsec),
2787 HRTIMER_MODE_REL_PINNED);
2788 }
2789
2790 spin_unlock_irqrestore(&ppd->cca_timer_lock, flags);
2791
2792 if ((trigger_threshold != 0) && (ccti >= trigger_threshold))
2793 log_cca_event(ppd, sl, rlid, lqpn, rqpn, svc_type);
2794 }
2795
2796 /**
2797 * hfi1_rc_rcv - process an incoming RC packet
2798 * @packet: data packet information
2799 *
2800 * This is called from qp_rcv() to process an incoming RC packet
2801 * for the given QP.
2802 * May be called at interrupt level.
2803 */
2804 void hfi1_rc_rcv(struct hfi1_packet *packet)
2805 {
2806 struct hfi1_ctxtdata *rcd = packet->rcd;
2807 void *data = packet->payload;
2808 u32 tlen = packet->tlen;
2809 struct rvt_qp *qp = packet->qp;
2810 struct hfi1_qp_priv *qpriv = qp->priv;
2811 struct hfi1_ibport *ibp = rcd_to_iport(rcd);
2812 struct ib_other_headers *ohdr = packet->ohdr;
2813 u32 opcode = packet->opcode;
2814 u32 hdrsize = packet->hlen;
2815 u32 psn = ib_bth_get_psn(packet->ohdr);
2816 u32 pad = packet->pad;
2817 struct ib_wc wc;
2818 u32 pmtu = qp->pmtu;
2819 int diff;
2820 struct ib_reth *reth;
2821 unsigned long flags;
2822 int ret;
2823 bool copy_last = false, fecn;
2824 u32 rkey;
2825 u8 extra_bytes = pad + packet->extra_byte + (SIZE_OF_CRC << 2);
2826
2827 lockdep_assert_held(&qp->r_lock);
2828
2829 if (hfi1_ruc_check_hdr(ibp, packet))
2830 return;
2831
2832 fecn = process_ecn(qp, packet);
2833 opfn_trigger_conn_request(qp, be32_to_cpu(ohdr->bth[1]));
2834
2835 /*
2836 * Process responses (ACKs) before anything else. Note that the
2837 * packet sequence number will be for something in the send work
2838 * queue rather than the expected receive packet sequence number.
2839 * In other words, this QP is the requester.
2840 */
2841 if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
2842 opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
2843 rc_rcv_resp(packet);
2844 return;
2845 }
2846
2847 /* Compute 24 bits worth of difference. */
2848 diff = delta_psn(psn, qp->r_psn);
2849 if (unlikely(diff)) {
2850 if (rc_rcv_error(ohdr, data, qp, opcode, psn, diff, rcd))
2851 return;
2852 goto send_ack;
2853 }
2854
2855 /* Check for opcode sequence errors. */
2856 switch (qp->r_state) {
2857 case OP(SEND_FIRST):
2858 case OP(SEND_MIDDLE):
2859 if (opcode == OP(SEND_MIDDLE) ||
2860 opcode == OP(SEND_LAST) ||
2861 opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
2862 opcode == OP(SEND_LAST_WITH_INVALIDATE))
2863 break;
2864 goto nack_inv;
2865
2866 case OP(RDMA_WRITE_FIRST):
2867 case OP(RDMA_WRITE_MIDDLE):
2868 if (opcode == OP(RDMA_WRITE_MIDDLE) ||
2869 opcode == OP(RDMA_WRITE_LAST) ||
2870 opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
2871 break;
2872 goto nack_inv;
2873
2874 default:
2875 if (opcode == OP(SEND_MIDDLE) ||
2876 opcode == OP(SEND_LAST) ||
2877 opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
2878 opcode == OP(SEND_LAST_WITH_INVALIDATE) ||
2879 opcode == OP(RDMA_WRITE_MIDDLE) ||
2880 opcode == OP(RDMA_WRITE_LAST) ||
2881 opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
2882 goto nack_inv;
2883 /*
2884 * Note that it is up to the requester to not send a new
2885 * RDMA read or atomic operation before receiving an ACK
2886 * for the previous operation.
2887 */
2888 break;
2889 }
2890
2891 if (qp->state == IB_QPS_RTR && !(qp->r_flags & RVT_R_COMM_EST))
2892 rvt_comm_est(qp);
2893
2894 /* OK, process the packet. */
2895 switch (opcode) {
2896 case OP(SEND_FIRST):
2897 ret = rvt_get_rwqe(qp, false);
2898 if (ret < 0)
2899 goto nack_op_err;
2900 if (!ret)
2901 goto rnr_nak;
2902 qp->r_rcv_len = 0;
2903 fallthrough;
2904 case OP(SEND_MIDDLE):
2905 case OP(RDMA_WRITE_MIDDLE):
2906 send_middle:
2907 /* Check for invalid length PMTU or posted rwqe len. */
2908 /*
2909 * There will be no padding for 9B packet but 16B packets
2910 * will come in with some padding since we always add
2911 * CRC and LT bytes which will need to be flit aligned
2912 */
2913 if (unlikely(tlen != (hdrsize + pmtu + extra_bytes)))
2914 goto nack_inv;
2915 qp->r_rcv_len += pmtu;
2916 if (unlikely(qp->r_rcv_len > qp->r_len))
2917 goto nack_inv;
2918 rvt_copy_sge(qp, &qp->r_sge, data, pmtu, true, false);
2919 break;
2920
2921 case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
2922 /* consume RWQE */
2923 ret = rvt_get_rwqe(qp, true);
2924 if (ret < 0)
2925 goto nack_op_err;
2926 if (!ret)
2927 goto rnr_nak;
2928 goto send_last_imm;
2929
2930 case OP(SEND_ONLY):
2931 case OP(SEND_ONLY_WITH_IMMEDIATE):
2932 case OP(SEND_ONLY_WITH_INVALIDATE):
2933 ret = rvt_get_rwqe(qp, false);
2934 if (ret < 0)
2935 goto nack_op_err;
2936 if (!ret)
2937 goto rnr_nak;
2938 qp->r_rcv_len = 0;
2939 if (opcode == OP(SEND_ONLY))
2940 goto no_immediate_data;
2941 if (opcode == OP(SEND_ONLY_WITH_INVALIDATE))
2942 goto send_last_inv;
2943 fallthrough; /* for SEND_ONLY_WITH_IMMEDIATE */
2944 case OP(SEND_LAST_WITH_IMMEDIATE):
2945 send_last_imm:
2946 wc.ex.imm_data = ohdr->u.imm_data;
2947 wc.wc_flags = IB_WC_WITH_IMM;
2948 goto send_last;
2949 case OP(SEND_LAST_WITH_INVALIDATE):
2950 send_last_inv:
2951 rkey = be32_to_cpu(ohdr->u.ieth);
2952 if (rvt_invalidate_rkey(qp, rkey))
2953 goto no_immediate_data;
2954 wc.ex.invalidate_rkey = rkey;
2955 wc.wc_flags = IB_WC_WITH_INVALIDATE;
2956 goto send_last;
2957 case OP(RDMA_WRITE_LAST):
2958 copy_last = rvt_is_user_qp(qp);
2959 fallthrough;
2960 case OP(SEND_LAST):
2961 no_immediate_data:
2962 wc.wc_flags = 0;
2963 wc.ex.imm_data = 0;
2964 send_last:
2965 /* Check for invalid length. */
2966 /* LAST len should be >= 1 */
2967 if (unlikely(tlen < (hdrsize + extra_bytes)))
2968 goto nack_inv;
2969 /* Don't count the CRC(and padding and LT byte for 16B). */
2970 tlen -= (hdrsize + extra_bytes);
2971 wc.byte_len = tlen + qp->r_rcv_len;
2972 if (unlikely(wc.byte_len > qp->r_len))
2973 goto nack_inv;
2974 rvt_copy_sge(qp, &qp->r_sge, data, tlen, true, copy_last);
2975 rvt_put_ss(&qp->r_sge);
2976 qp->r_msn++;
2977 if (!__test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
2978 break;
2979 wc.wr_id = qp->r_wr_id;
2980 wc.status = IB_WC_SUCCESS;
2981 if (opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE) ||
2982 opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE))
2983 wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
2984 else
2985 wc.opcode = IB_WC_RECV;
2986 wc.qp = &qp->ibqp;
2987 wc.src_qp = qp->remote_qpn;
2988 wc.slid = rdma_ah_get_dlid(&qp->remote_ah_attr) & U16_MAX;
2989 /*
2990 * It seems that IB mandates the presence of an SL in a
2991 * work completion only for the UD transport (see section
2992 * 11.4.2 of IBTA Vol. 1).
2993 *
2994 * However, the way the SL is chosen below is consistent
2995 * with the way that IB/qib works and is trying avoid
2996 * introducing incompatibilities.
2997 *
2998 * See also OPA Vol. 1, section 9.7.6, and table 9-17.
2999 */
3000 wc.sl = rdma_ah_get_sl(&qp->remote_ah_attr);
3001 /* zero fields that are N/A */
3002 wc.vendor_err = 0;
3003 wc.pkey_index = 0;
3004 wc.dlid_path_bits = 0;
3005 wc.port_num = 0;
3006 /* Signal completion event if the solicited bit is set. */
3007 rvt_recv_cq(qp, &wc, ib_bth_is_solicited(ohdr));
3008 break;
3009
3010 case OP(RDMA_WRITE_ONLY):
3011 copy_last = rvt_is_user_qp(qp);
3012 fallthrough;
3013 case OP(RDMA_WRITE_FIRST):
3014 case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
3015 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
3016 goto nack_inv;
3017 /* consume RWQE */
3018 reth = &ohdr->u.rc.reth;
3019 qp->r_len = be32_to_cpu(reth->length);
3020 qp->r_rcv_len = 0;
3021 qp->r_sge.sg_list = NULL;
3022 if (qp->r_len != 0) {
3023 u32 rkey = be32_to_cpu(reth->rkey);
3024 u64 vaddr = get_ib_reth_vaddr(reth);
3025 int ok;
3026
3027 /* Check rkey & NAK */
3028 ok = rvt_rkey_ok(qp, &qp->r_sge.sge, qp->r_len, vaddr,
3029 rkey, IB_ACCESS_REMOTE_WRITE);
3030 if (unlikely(!ok))
3031 goto nack_acc;
3032 qp->r_sge.num_sge = 1;
3033 } else {
3034 qp->r_sge.num_sge = 0;
3035 qp->r_sge.sge.mr = NULL;
3036 qp->r_sge.sge.vaddr = NULL;
3037 qp->r_sge.sge.length = 0;
3038 qp->r_sge.sge.sge_length = 0;
3039 }
3040 if (opcode == OP(RDMA_WRITE_FIRST))
3041 goto send_middle;
3042 else if (opcode == OP(RDMA_WRITE_ONLY))
3043 goto no_immediate_data;
3044 ret = rvt_get_rwqe(qp, true);
3045 if (ret < 0)
3046 goto nack_op_err;
3047 if (!ret) {
3048 /* peer will send again */
3049 rvt_put_ss(&qp->r_sge);
3050 goto rnr_nak;
3051 }
3052 wc.ex.imm_data = ohdr->u.rc.imm_data;
3053 wc.wc_flags = IB_WC_WITH_IMM;
3054 goto send_last;
3055
3056 case OP(RDMA_READ_REQUEST): {
3057 struct rvt_ack_entry *e;
3058 u32 len;
3059 u8 next;
3060
3061 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
3062 goto nack_inv;
3063 next = qp->r_head_ack_queue + 1;
3064 /* s_ack_queue is size rvt_size_atomic()+1 so use > not >= */
3065 if (next > rvt_size_atomic(ib_to_rvt(qp->ibqp.device)))
3066 next = 0;
3067 spin_lock_irqsave(&qp->s_lock, flags);
3068 if (unlikely(next == qp->s_acked_ack_queue)) {
3069 if (!qp->s_ack_queue[next].sent)
3070 goto nack_inv_unlck;
3071 update_ack_queue(qp, next);
3072 }
3073 e = &qp->s_ack_queue[qp->r_head_ack_queue];
3074 release_rdma_sge_mr(e);
3075 reth = &ohdr->u.rc.reth;
3076 len = be32_to_cpu(reth->length);
3077 if (len) {
3078 u32 rkey = be32_to_cpu(reth->rkey);
3079 u64 vaddr = get_ib_reth_vaddr(reth);
3080 int ok;
3081
3082 /* Check rkey & NAK */
3083 ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr,
3084 rkey, IB_ACCESS_REMOTE_READ);
3085 if (unlikely(!ok))
3086 goto nack_acc_unlck;
3087 /*
3088 * Update the next expected PSN. We add 1 later
3089 * below, so only add the remainder here.
3090 */
3091 qp->r_psn += rvt_div_mtu(qp, len - 1);
3092 } else {
3093 e->rdma_sge.mr = NULL;
3094 e->rdma_sge.vaddr = NULL;
3095 e->rdma_sge.length = 0;
3096 e->rdma_sge.sge_length = 0;
3097 }
3098 e->opcode = opcode;
3099 e->sent = 0;
3100 e->psn = psn;
3101 e->lpsn = qp->r_psn;
3102 /*
3103 * We need to increment the MSN here instead of when we
3104 * finish sending the result since a duplicate request would
3105 * increment it more than once.
3106 */
3107 qp->r_msn++;
3108 qp->r_psn++;
3109 qp->r_state = opcode;
3110 qp->r_nak_state = 0;
3111 qp->r_head_ack_queue = next;
3112 qpriv->r_tid_alloc = qp->r_head_ack_queue;
3113
3114 /* Schedule the send engine. */
3115 qp->s_flags |= RVT_S_RESP_PENDING;
3116 if (fecn)
3117 qp->s_flags |= RVT_S_ECN;
3118 hfi1_schedule_send(qp);
3119
3120 spin_unlock_irqrestore(&qp->s_lock, flags);
3121 return;
3122 }
3123
3124 case OP(COMPARE_SWAP):
3125 case OP(FETCH_ADD): {
3126 struct ib_atomic_eth *ateth = &ohdr->u.atomic_eth;
3127 u64 vaddr = get_ib_ateth_vaddr(ateth);
3128 bool opfn = opcode == OP(COMPARE_SWAP) &&
3129 vaddr == HFI1_VERBS_E_ATOMIC_VADDR;
3130 struct rvt_ack_entry *e;
3131 atomic64_t *maddr;
3132 u64 sdata;
3133 u32 rkey;
3134 u8 next;
3135
3136 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC) &&
3137 !opfn))
3138 goto nack_inv;
3139 next = qp->r_head_ack_queue + 1;
3140 if (next > rvt_size_atomic(ib_to_rvt(qp->ibqp.device)))
3141 next = 0;
3142 spin_lock_irqsave(&qp->s_lock, flags);
3143 if (unlikely(next == qp->s_acked_ack_queue)) {
3144 if (!qp->s_ack_queue[next].sent)
3145 goto nack_inv_unlck;
3146 update_ack_queue(qp, next);
3147 }
3148 e = &qp->s_ack_queue[qp->r_head_ack_queue];
3149 release_rdma_sge_mr(e);
3150 /* Process OPFN special virtual address */
3151 if (opfn) {
3152 opfn_conn_response(qp, e, ateth);
3153 goto ack;
3154 }
3155 if (unlikely(vaddr & (sizeof(u64) - 1)))
3156 goto nack_inv_unlck;
3157 rkey = be32_to_cpu(ateth->rkey);
3158 /* Check rkey & NAK */
3159 if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64),
3160 vaddr, rkey,
3161 IB_ACCESS_REMOTE_ATOMIC)))
3162 goto nack_acc_unlck;
3163 /* Perform atomic OP and save result. */
3164 maddr = (atomic64_t *)qp->r_sge.sge.vaddr;
3165 sdata = get_ib_ateth_swap(ateth);
3166 e->atomic_data = (opcode == OP(FETCH_ADD)) ?
3167 (u64)atomic64_add_return(sdata, maddr) - sdata :
3168 (u64)cmpxchg((u64 *)qp->r_sge.sge.vaddr,
3169 get_ib_ateth_compare(ateth),
3170 sdata);
3171 rvt_put_mr(qp->r_sge.sge.mr);
3172 qp->r_sge.num_sge = 0;
3173 ack:
3174 e->opcode = opcode;
3175 e->sent = 0;
3176 e->psn = psn;
3177 e->lpsn = psn;
3178 qp->r_msn++;
3179 qp->r_psn++;
3180 qp->r_state = opcode;
3181 qp->r_nak_state = 0;
3182 qp->r_head_ack_queue = next;
3183 qpriv->r_tid_alloc = qp->r_head_ack_queue;
3184
3185 /* Schedule the send engine. */
3186 qp->s_flags |= RVT_S_RESP_PENDING;
3187 if (fecn)
3188 qp->s_flags |= RVT_S_ECN;
3189 hfi1_schedule_send(qp);
3190
3191 spin_unlock_irqrestore(&qp->s_lock, flags);
3192 return;
3193 }
3194
3195 default:
3196 /* NAK unknown opcodes. */
3197 goto nack_inv;
3198 }
3199 qp->r_psn++;
3200 qp->r_state = opcode;
3201 qp->r_ack_psn = psn;
3202 qp->r_nak_state = 0;
3203 /* Send an ACK if requested or required. */
3204 if (psn & IB_BTH_REQ_ACK || fecn) {
3205 if (packet->numpkt == 0 || fecn ||
3206 qp->r_adefered >= HFI1_PSN_CREDIT) {
3207 rc_cancel_ack(qp);
3208 goto send_ack;
3209 }
3210 qp->r_adefered++;
3211 rc_defered_ack(rcd, qp);
3212 }
3213 return;
3214
3215 rnr_nak:
3216 qp->r_nak_state = qp->r_min_rnr_timer | IB_RNR_NAK;
3217 qp->r_ack_psn = qp->r_psn;
3218 /* Queue RNR NAK for later */
3219 rc_defered_ack(rcd, qp);
3220 return;
3221
3222 nack_op_err:
3223 rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
3224 qp->r_nak_state = IB_NAK_REMOTE_OPERATIONAL_ERROR;
3225 qp->r_ack_psn = qp->r_psn;
3226 /* Queue NAK for later */
3227 rc_defered_ack(rcd, qp);
3228 return;
3229
3230 nack_inv_unlck:
3231 spin_unlock_irqrestore(&qp->s_lock, flags);
3232 nack_inv:
3233 rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
3234 qp->r_nak_state = IB_NAK_INVALID_REQUEST;
3235 qp->r_ack_psn = qp->r_psn;
3236 /* Queue NAK for later */
3237 rc_defered_ack(rcd, qp);
3238 return;
3239
3240 nack_acc_unlck:
3241 spin_unlock_irqrestore(&qp->s_lock, flags);
3242 nack_acc:
3243 rvt_rc_error(qp, IB_WC_LOC_PROT_ERR);
3244 qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
3245 qp->r_ack_psn = qp->r_psn;
3246 send_ack:
3247 hfi1_send_rc_ack(packet, fecn);
3248 }
3249
3250 void hfi1_rc_hdrerr(
3251 struct hfi1_ctxtdata *rcd,
3252 struct hfi1_packet *packet,
3253 struct rvt_qp *qp)
3254 {
3255 struct hfi1_ibport *ibp = rcd_to_iport(rcd);
3256 int diff;
3257 u32 opcode;
3258 u32 psn;
3259
3260 if (hfi1_ruc_check_hdr(ibp, packet))
3261 return;
3262
3263 psn = ib_bth_get_psn(packet->ohdr);
3264 opcode = ib_bth_get_opcode(packet->ohdr);
3265
3266 /* Only deal with RDMA Writes for now */
3267 if (opcode < IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST) {
3268 diff = delta_psn(psn, qp->r_psn);
3269 if (!qp->r_nak_state && diff >= 0) {
3270 ibp->rvp.n_rc_seqnak++;
3271 qp->r_nak_state = IB_NAK_PSN_ERROR;
3272 /* Use the expected PSN. */
3273 qp->r_ack_psn = qp->r_psn;
3274 /*
3275 * Wait to send the sequence
3276 * NAK until all packets
3277 * in the receive queue have
3278 * been processed.
3279 * Otherwise, we end up
3280 * propagating congestion.
3281 */
3282 rc_defered_ack(rcd, qp);
3283 } /* Out of sequence NAK */
3284 } /* QP Request NAKs */
3285 }
Linux with added FPC-III support