xref: /openbmc/linux/drivers/infiniband/hw/hfi1/rc.c (revision bc5aa3a0)
1 /*
2  * Copyright(c) 2015, 2016 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 "verbs_txreq.h"
55 #include "trace.h"
56 
57 /* cut down ridiculously long IB macro names */
58 #define OP(x) IB_OPCODE_RC_##x
59 
60 /**
61  * hfi1_add_retry_timer - add/start a retry timer
62  * @qp - the QP
63  *
64  * add a retry timer on the QP
65  */
66 static inline void hfi1_add_retry_timer(struct rvt_qp *qp)
67 {
68 	struct ib_qp *ibqp = &qp->ibqp;
69 	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
70 
71 	qp->s_flags |= RVT_S_TIMER;
72 	/* 4.096 usec. * (1 << qp->timeout) */
73 	qp->s_timer.expires = jiffies + qp->timeout_jiffies +
74 			      rdi->busy_jiffies;
75 	add_timer(&qp->s_timer);
76 }
77 
78 /**
79  * hfi1_add_rnr_timer - add/start an rnr timer
80  * @qp - the QP
81  * @to - timeout in usecs
82  *
83  * add an rnr timer on the QP
84  */
85 void hfi1_add_rnr_timer(struct rvt_qp *qp, u32 to)
86 {
87 	struct hfi1_qp_priv *priv = qp->priv;
88 
89 	qp->s_flags |= RVT_S_WAIT_RNR;
90 	qp->s_timer.expires = jiffies + usecs_to_jiffies(to);
91 	add_timer(&priv->s_rnr_timer);
92 }
93 
94 /**
95  * hfi1_mod_retry_timer - mod a retry timer
96  * @qp - the QP
97  *
98  * Modify a potentially already running retry
99  * timer
100  */
101 static inline void hfi1_mod_retry_timer(struct rvt_qp *qp)
102 {
103 	struct ib_qp *ibqp = &qp->ibqp;
104 	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
105 
106 	qp->s_flags |= RVT_S_TIMER;
107 	/* 4.096 usec. * (1 << qp->timeout) */
108 	mod_timer(&qp->s_timer, jiffies + qp->timeout_jiffies +
109 		  rdi->busy_jiffies);
110 }
111 
112 /**
113  * hfi1_stop_retry_timer - stop a retry timer
114  * @qp - the QP
115  *
116  * stop a retry timer and return if the timer
117  * had been pending.
118  */
119 static inline int hfi1_stop_retry_timer(struct rvt_qp *qp)
120 {
121 	int rval = 0;
122 
123 	/* Remove QP from retry */
124 	if (qp->s_flags & RVT_S_TIMER) {
125 		qp->s_flags &= ~RVT_S_TIMER;
126 		rval = del_timer(&qp->s_timer);
127 	}
128 	return rval;
129 }
130 
131 /**
132  * hfi1_stop_rc_timers - stop all timers
133  * @qp - the QP
134  *
135  * stop any pending timers
136  */
137 void hfi1_stop_rc_timers(struct rvt_qp *qp)
138 {
139 	struct hfi1_qp_priv *priv = qp->priv;
140 
141 	/* Remove QP from all timers */
142 	if (qp->s_flags & (RVT_S_TIMER | RVT_S_WAIT_RNR)) {
143 		qp->s_flags &= ~(RVT_S_TIMER | RVT_S_WAIT_RNR);
144 		del_timer(&qp->s_timer);
145 		del_timer(&priv->s_rnr_timer);
146 	}
147 }
148 
149 /**
150  * hfi1_stop_rnr_timer - stop an rnr timer
151  * @qp - the QP
152  *
153  * stop an rnr timer and return if the timer
154  * had been pending.
155  */
156 static inline int hfi1_stop_rnr_timer(struct rvt_qp *qp)
157 {
158 	int rval = 0;
159 	struct hfi1_qp_priv *priv = qp->priv;
160 
161 	/* Remove QP from rnr timer */
162 	if (qp->s_flags & RVT_S_WAIT_RNR) {
163 		qp->s_flags &= ~RVT_S_WAIT_RNR;
164 		rval = del_timer(&priv->s_rnr_timer);
165 	}
166 	return rval;
167 }
168 
169 /**
170  * hfi1_del_timers_sync - wait for any timeout routines to exit
171  * @qp - the QP
172  */
173 void hfi1_del_timers_sync(struct rvt_qp *qp)
174 {
175 	struct hfi1_qp_priv *priv = qp->priv;
176 
177 	del_timer_sync(&qp->s_timer);
178 	del_timer_sync(&priv->s_rnr_timer);
179 }
180 
181 /* only opcode mask for adaptive pio */
182 const u32 rc_only_opcode =
183 	BIT(OP(SEND_ONLY) & 0x1f) |
184 	BIT(OP(SEND_ONLY_WITH_IMMEDIATE & 0x1f)) |
185 	BIT(OP(RDMA_WRITE_ONLY & 0x1f)) |
186 	BIT(OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE & 0x1f)) |
187 	BIT(OP(RDMA_READ_REQUEST & 0x1f)) |
188 	BIT(OP(ACKNOWLEDGE & 0x1f)) |
189 	BIT(OP(ATOMIC_ACKNOWLEDGE & 0x1f)) |
190 	BIT(OP(COMPARE_SWAP & 0x1f)) |
191 	BIT(OP(FETCH_ADD & 0x1f));
192 
193 static u32 restart_sge(struct rvt_sge_state *ss, struct rvt_swqe *wqe,
194 		       u32 psn, u32 pmtu)
195 {
196 	u32 len;
197 
198 	len = delta_psn(psn, wqe->psn) * pmtu;
199 	ss->sge = wqe->sg_list[0];
200 	ss->sg_list = wqe->sg_list + 1;
201 	ss->num_sge = wqe->wr.num_sge;
202 	ss->total_len = wqe->length;
203 	hfi1_skip_sge(ss, len, 0);
204 	return wqe->length - len;
205 }
206 
207 /**
208  * make_rc_ack - construct a response packet (ACK, NAK, or RDMA read)
209  * @dev: the device for this QP
210  * @qp: a pointer to the QP
211  * @ohdr: a pointer to the IB header being constructed
212  * @ps: the xmit packet state
213  *
214  * Return 1 if constructed; otherwise, return 0.
215  * Note that we are in the responder's side of the QP context.
216  * Note the QP s_lock must be held.
217  */
218 static int make_rc_ack(struct hfi1_ibdev *dev, struct rvt_qp *qp,
219 		       struct hfi1_other_headers *ohdr,
220 		       struct hfi1_pkt_state *ps)
221 {
222 	struct rvt_ack_entry *e;
223 	u32 hwords;
224 	u32 len;
225 	u32 bth0;
226 	u32 bth2;
227 	int middle = 0;
228 	u32 pmtu = qp->pmtu;
229 	struct hfi1_qp_priv *priv = qp->priv;
230 
231 	/* Don't send an ACK if we aren't supposed to. */
232 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
233 		goto bail;
234 
235 	/* header size in 32-bit words LRH+BTH = (8+12)/4. */
236 	hwords = 5;
237 
238 	switch (qp->s_ack_state) {
239 	case OP(RDMA_READ_RESPONSE_LAST):
240 	case OP(RDMA_READ_RESPONSE_ONLY):
241 		e = &qp->s_ack_queue[qp->s_tail_ack_queue];
242 		if (e->rdma_sge.mr) {
243 			rvt_put_mr(e->rdma_sge.mr);
244 			e->rdma_sge.mr = NULL;
245 		}
246 		/* FALLTHROUGH */
247 	case OP(ATOMIC_ACKNOWLEDGE):
248 		/*
249 		 * We can increment the tail pointer now that the last
250 		 * response has been sent instead of only being
251 		 * constructed.
252 		 */
253 		if (++qp->s_tail_ack_queue > HFI1_MAX_RDMA_ATOMIC)
254 			qp->s_tail_ack_queue = 0;
255 		/* FALLTHROUGH */
256 	case OP(SEND_ONLY):
257 	case OP(ACKNOWLEDGE):
258 		/* Check for no next entry in the queue. */
259 		if (qp->r_head_ack_queue == qp->s_tail_ack_queue) {
260 			if (qp->s_flags & RVT_S_ACK_PENDING)
261 				goto normal;
262 			goto bail;
263 		}
264 
265 		e = &qp->s_ack_queue[qp->s_tail_ack_queue];
266 		if (e->opcode == OP(RDMA_READ_REQUEST)) {
267 			/*
268 			 * If a RDMA read response is being resent and
269 			 * we haven't seen the duplicate request yet,
270 			 * then stop sending the remaining responses the
271 			 * responder has seen until the requester re-sends it.
272 			 */
273 			len = e->rdma_sge.sge_length;
274 			if (len && !e->rdma_sge.mr) {
275 				qp->s_tail_ack_queue = qp->r_head_ack_queue;
276 				goto bail;
277 			}
278 			/* Copy SGE state in case we need to resend */
279 			ps->s_txreq->mr = e->rdma_sge.mr;
280 			if (ps->s_txreq->mr)
281 				rvt_get_mr(ps->s_txreq->mr);
282 			qp->s_ack_rdma_sge.sge = e->rdma_sge;
283 			qp->s_ack_rdma_sge.num_sge = 1;
284 			qp->s_cur_sge = &qp->s_ack_rdma_sge;
285 			if (len > pmtu) {
286 				len = pmtu;
287 				qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST);
288 			} else {
289 				qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY);
290 				e->sent = 1;
291 			}
292 			ohdr->u.aeth = hfi1_compute_aeth(qp);
293 			hwords++;
294 			qp->s_ack_rdma_psn = e->psn;
295 			bth2 = mask_psn(qp->s_ack_rdma_psn++);
296 		} else {
297 			/* COMPARE_SWAP or FETCH_ADD */
298 			qp->s_cur_sge = NULL;
299 			len = 0;
300 			qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE);
301 			ohdr->u.at.aeth = hfi1_compute_aeth(qp);
302 			ohdr->u.at.atomic_ack_eth[0] =
303 				cpu_to_be32(e->atomic_data >> 32);
304 			ohdr->u.at.atomic_ack_eth[1] =
305 				cpu_to_be32(e->atomic_data);
306 			hwords += sizeof(ohdr->u.at) / sizeof(u32);
307 			bth2 = mask_psn(e->psn);
308 			e->sent = 1;
309 		}
310 		bth0 = qp->s_ack_state << 24;
311 		break;
312 
313 	case OP(RDMA_READ_RESPONSE_FIRST):
314 		qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE);
315 		/* FALLTHROUGH */
316 	case OP(RDMA_READ_RESPONSE_MIDDLE):
317 		qp->s_cur_sge = &qp->s_ack_rdma_sge;
318 		ps->s_txreq->mr = qp->s_ack_rdma_sge.sge.mr;
319 		if (ps->s_txreq->mr)
320 			rvt_get_mr(ps->s_txreq->mr);
321 		len = qp->s_ack_rdma_sge.sge.sge_length;
322 		if (len > pmtu) {
323 			len = pmtu;
324 			middle = HFI1_CAP_IS_KSET(SDMA_AHG);
325 		} else {
326 			ohdr->u.aeth = hfi1_compute_aeth(qp);
327 			hwords++;
328 			qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
329 			e = &qp->s_ack_queue[qp->s_tail_ack_queue];
330 			e->sent = 1;
331 		}
332 		bth0 = qp->s_ack_state << 24;
333 		bth2 = mask_psn(qp->s_ack_rdma_psn++);
334 		break;
335 
336 	default:
337 normal:
338 		/*
339 		 * Send a regular ACK.
340 		 * Set the s_ack_state so we wait until after sending
341 		 * the ACK before setting s_ack_state to ACKNOWLEDGE
342 		 * (see above).
343 		 */
344 		qp->s_ack_state = OP(SEND_ONLY);
345 		qp->s_flags &= ~RVT_S_ACK_PENDING;
346 		qp->s_cur_sge = NULL;
347 		if (qp->s_nak_state)
348 			ohdr->u.aeth =
349 				cpu_to_be32((qp->r_msn & HFI1_MSN_MASK) |
350 					    (qp->s_nak_state <<
351 					     HFI1_AETH_CREDIT_SHIFT));
352 		else
353 			ohdr->u.aeth = hfi1_compute_aeth(qp);
354 		hwords++;
355 		len = 0;
356 		bth0 = OP(ACKNOWLEDGE) << 24;
357 		bth2 = mask_psn(qp->s_ack_psn);
358 	}
359 	qp->s_rdma_ack_cnt++;
360 	qp->s_hdrwords = hwords;
361 	ps->s_txreq->sde = priv->s_sde;
362 	qp->s_cur_size = len;
363 	hfi1_make_ruc_header(qp, ohdr, bth0, bth2, middle, ps);
364 	/* pbc */
365 	ps->s_txreq->hdr_dwords = qp->s_hdrwords + 2;
366 	return 1;
367 
368 bail:
369 	qp->s_ack_state = OP(ACKNOWLEDGE);
370 	/*
371 	 * Ensure s_rdma_ack_cnt changes are committed prior to resetting
372 	 * RVT_S_RESP_PENDING
373 	 */
374 	smp_wmb();
375 	qp->s_flags &= ~(RVT_S_RESP_PENDING
376 				| RVT_S_ACK_PENDING
377 				| RVT_S_AHG_VALID);
378 	return 0;
379 }
380 
381 /**
382  * hfi1_make_rc_req - construct a request packet (SEND, RDMA r/w, ATOMIC)
383  * @qp: a pointer to the QP
384  *
385  * Assumes s_lock is held.
386  *
387  * Return 1 if constructed; otherwise, return 0.
388  */
389 int hfi1_make_rc_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
390 {
391 	struct hfi1_qp_priv *priv = qp->priv;
392 	struct hfi1_ibdev *dev = to_idev(qp->ibqp.device);
393 	struct hfi1_other_headers *ohdr;
394 	struct rvt_sge_state *ss;
395 	struct rvt_swqe *wqe;
396 	/* header size in 32-bit words LRH+BTH = (8+12)/4. */
397 	u32 hwords = 5;
398 	u32 len;
399 	u32 bth0 = 0;
400 	u32 bth2;
401 	u32 pmtu = qp->pmtu;
402 	char newreq;
403 	int middle = 0;
404 	int delta;
405 
406 	ps->s_txreq = get_txreq(ps->dev, qp);
407 	if (IS_ERR(ps->s_txreq))
408 		goto bail_no_tx;
409 
410 	ohdr = &ps->s_txreq->phdr.hdr.u.oth;
411 	if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
412 		ohdr = &ps->s_txreq->phdr.hdr.u.l.oth;
413 
414 	/* Sending responses has higher priority over sending requests. */
415 	if ((qp->s_flags & RVT_S_RESP_PENDING) &&
416 	    make_rc_ack(dev, qp, ohdr, ps))
417 		return 1;
418 
419 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_SEND_OK)) {
420 		if (!(ib_rvt_state_ops[qp->state] & RVT_FLUSH_SEND))
421 			goto bail;
422 		/* We are in the error state, flush the work request. */
423 		smp_read_barrier_depends(); /* see post_one_send() */
424 		if (qp->s_last == ACCESS_ONCE(qp->s_head))
425 			goto bail;
426 		/* If DMAs are in progress, we can't flush immediately. */
427 		if (iowait_sdma_pending(&priv->s_iowait)) {
428 			qp->s_flags |= RVT_S_WAIT_DMA;
429 			goto bail;
430 		}
431 		clear_ahg(qp);
432 		wqe = rvt_get_swqe_ptr(qp, qp->s_last);
433 		hfi1_send_complete(qp, wqe, qp->s_last != qp->s_acked ?
434 			IB_WC_SUCCESS : IB_WC_WR_FLUSH_ERR);
435 		/* will get called again */
436 		goto done_free_tx;
437 	}
438 
439 	if (qp->s_flags & (RVT_S_WAIT_RNR | RVT_S_WAIT_ACK))
440 		goto bail;
441 
442 	if (cmp_psn(qp->s_psn, qp->s_sending_hpsn) <= 0) {
443 		if (cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) <= 0) {
444 			qp->s_flags |= RVT_S_WAIT_PSN;
445 			goto bail;
446 		}
447 		qp->s_sending_psn = qp->s_psn;
448 		qp->s_sending_hpsn = qp->s_psn - 1;
449 	}
450 
451 	/* Send a request. */
452 	wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
453 	switch (qp->s_state) {
454 	default:
455 		if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_NEXT_SEND_OK))
456 			goto bail;
457 		/*
458 		 * Resend an old request or start a new one.
459 		 *
460 		 * We keep track of the current SWQE so that
461 		 * we don't reset the "furthest progress" state
462 		 * if we need to back up.
463 		 */
464 		newreq = 0;
465 		if (qp->s_cur == qp->s_tail) {
466 			/* Check if send work queue is empty. */
467 			if (qp->s_tail == qp->s_head) {
468 				clear_ahg(qp);
469 				goto bail;
470 			}
471 			/*
472 			 * If a fence is requested, wait for previous
473 			 * RDMA read and atomic operations to finish.
474 			 */
475 			if ((wqe->wr.send_flags & IB_SEND_FENCE) &&
476 			    qp->s_num_rd_atomic) {
477 				qp->s_flags |= RVT_S_WAIT_FENCE;
478 				goto bail;
479 			}
480 			/*
481 			 * Local operations are processed immediately
482 			 * after all prior requests have completed
483 			 */
484 			if (wqe->wr.opcode == IB_WR_REG_MR ||
485 			    wqe->wr.opcode == IB_WR_LOCAL_INV) {
486 				int local_ops = 0;
487 				int err = 0;
488 
489 				if (qp->s_last != qp->s_cur)
490 					goto bail;
491 				if (++qp->s_cur == qp->s_size)
492 					qp->s_cur = 0;
493 				if (++qp->s_tail == qp->s_size)
494 					qp->s_tail = 0;
495 				if (!(wqe->wr.send_flags &
496 				      RVT_SEND_COMPLETION_ONLY)) {
497 					err = rvt_invalidate_rkey(
498 						qp,
499 						wqe->wr.ex.invalidate_rkey);
500 					local_ops = 1;
501 				}
502 				hfi1_send_complete(qp, wqe,
503 						   err ? IB_WC_LOC_PROT_ERR
504 						       : IB_WC_SUCCESS);
505 				if (local_ops)
506 					atomic_dec(&qp->local_ops_pending);
507 				qp->s_hdrwords = 0;
508 				goto done_free_tx;
509 			}
510 
511 			newreq = 1;
512 			qp->s_psn = wqe->psn;
513 		}
514 		/*
515 		 * Note that we have to be careful not to modify the
516 		 * original work request since we may need to resend
517 		 * it.
518 		 */
519 		len = wqe->length;
520 		ss = &qp->s_sge;
521 		bth2 = mask_psn(qp->s_psn);
522 		switch (wqe->wr.opcode) {
523 		case IB_WR_SEND:
524 		case IB_WR_SEND_WITH_IMM:
525 		case IB_WR_SEND_WITH_INV:
526 			/* If no credit, return. */
527 			if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT) &&
528 			    cmp_msn(wqe->ssn, qp->s_lsn + 1) > 0) {
529 				qp->s_flags |= RVT_S_WAIT_SSN_CREDIT;
530 				goto bail;
531 			}
532 			if (len > pmtu) {
533 				qp->s_state = OP(SEND_FIRST);
534 				len = pmtu;
535 				break;
536 			}
537 			if (wqe->wr.opcode == IB_WR_SEND) {
538 				qp->s_state = OP(SEND_ONLY);
539 			} else if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
540 				qp->s_state = OP(SEND_ONLY_WITH_IMMEDIATE);
541 				/* Immediate data comes after the BTH */
542 				ohdr->u.imm_data = wqe->wr.ex.imm_data;
543 				hwords += 1;
544 			} else {
545 				qp->s_state = OP(SEND_ONLY_WITH_INVALIDATE);
546 				/* Invalidate rkey comes after the BTH */
547 				ohdr->u.ieth = cpu_to_be32(
548 						wqe->wr.ex.invalidate_rkey);
549 				hwords += 1;
550 			}
551 			if (wqe->wr.send_flags & IB_SEND_SOLICITED)
552 				bth0 |= IB_BTH_SOLICITED;
553 			bth2 |= IB_BTH_REQ_ACK;
554 			if (++qp->s_cur == qp->s_size)
555 				qp->s_cur = 0;
556 			break;
557 
558 		case IB_WR_RDMA_WRITE:
559 			if (newreq && !(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
560 				qp->s_lsn++;
561 			/* FALLTHROUGH */
562 		case IB_WR_RDMA_WRITE_WITH_IMM:
563 			/* If no credit, return. */
564 			if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT) &&
565 			    cmp_msn(wqe->ssn, qp->s_lsn + 1) > 0) {
566 				qp->s_flags |= RVT_S_WAIT_SSN_CREDIT;
567 				goto bail;
568 			}
569 			ohdr->u.rc.reth.vaddr =
570 				cpu_to_be64(wqe->rdma_wr.remote_addr);
571 			ohdr->u.rc.reth.rkey =
572 				cpu_to_be32(wqe->rdma_wr.rkey);
573 			ohdr->u.rc.reth.length = cpu_to_be32(len);
574 			hwords += sizeof(struct ib_reth) / sizeof(u32);
575 			if (len > pmtu) {
576 				qp->s_state = OP(RDMA_WRITE_FIRST);
577 				len = pmtu;
578 				break;
579 			}
580 			if (wqe->wr.opcode == IB_WR_RDMA_WRITE) {
581 				qp->s_state = OP(RDMA_WRITE_ONLY);
582 			} else {
583 				qp->s_state =
584 					OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
585 				/* Immediate data comes after RETH */
586 				ohdr->u.rc.imm_data = wqe->wr.ex.imm_data;
587 				hwords += 1;
588 				if (wqe->wr.send_flags & IB_SEND_SOLICITED)
589 					bth0 |= IB_BTH_SOLICITED;
590 			}
591 			bth2 |= IB_BTH_REQ_ACK;
592 			if (++qp->s_cur == qp->s_size)
593 				qp->s_cur = 0;
594 			break;
595 
596 		case IB_WR_RDMA_READ:
597 			/*
598 			 * Don't allow more operations to be started
599 			 * than the QP limits allow.
600 			 */
601 			if (newreq) {
602 				if (qp->s_num_rd_atomic >=
603 				    qp->s_max_rd_atomic) {
604 					qp->s_flags |= RVT_S_WAIT_RDMAR;
605 					goto bail;
606 				}
607 				qp->s_num_rd_atomic++;
608 				if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
609 					qp->s_lsn++;
610 			}
611 			ohdr->u.rc.reth.vaddr =
612 				cpu_to_be64(wqe->rdma_wr.remote_addr);
613 			ohdr->u.rc.reth.rkey =
614 				cpu_to_be32(wqe->rdma_wr.rkey);
615 			ohdr->u.rc.reth.length = cpu_to_be32(len);
616 			qp->s_state = OP(RDMA_READ_REQUEST);
617 			hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
618 			ss = NULL;
619 			len = 0;
620 			bth2 |= IB_BTH_REQ_ACK;
621 			if (++qp->s_cur == qp->s_size)
622 				qp->s_cur = 0;
623 			break;
624 
625 		case IB_WR_ATOMIC_CMP_AND_SWP:
626 		case IB_WR_ATOMIC_FETCH_AND_ADD:
627 			/*
628 			 * Don't allow more operations to be started
629 			 * than the QP limits allow.
630 			 */
631 			if (newreq) {
632 				if (qp->s_num_rd_atomic >=
633 				    qp->s_max_rd_atomic) {
634 					qp->s_flags |= RVT_S_WAIT_RDMAR;
635 					goto bail;
636 				}
637 				qp->s_num_rd_atomic++;
638 				if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
639 					qp->s_lsn++;
640 			}
641 			if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
642 				qp->s_state = OP(COMPARE_SWAP);
643 				ohdr->u.atomic_eth.swap_data = cpu_to_be64(
644 					wqe->atomic_wr.swap);
645 				ohdr->u.atomic_eth.compare_data = cpu_to_be64(
646 					wqe->atomic_wr.compare_add);
647 			} else {
648 				qp->s_state = OP(FETCH_ADD);
649 				ohdr->u.atomic_eth.swap_data = cpu_to_be64(
650 					wqe->atomic_wr.compare_add);
651 				ohdr->u.atomic_eth.compare_data = 0;
652 			}
653 			ohdr->u.atomic_eth.vaddr[0] = cpu_to_be32(
654 				wqe->atomic_wr.remote_addr >> 32);
655 			ohdr->u.atomic_eth.vaddr[1] = cpu_to_be32(
656 				wqe->atomic_wr.remote_addr);
657 			ohdr->u.atomic_eth.rkey = cpu_to_be32(
658 				wqe->atomic_wr.rkey);
659 			hwords += sizeof(struct ib_atomic_eth) / sizeof(u32);
660 			ss = NULL;
661 			len = 0;
662 			bth2 |= IB_BTH_REQ_ACK;
663 			if (++qp->s_cur == qp->s_size)
664 				qp->s_cur = 0;
665 			break;
666 
667 		default:
668 			goto bail;
669 		}
670 		qp->s_sge.sge = wqe->sg_list[0];
671 		qp->s_sge.sg_list = wqe->sg_list + 1;
672 		qp->s_sge.num_sge = wqe->wr.num_sge;
673 		qp->s_sge.total_len = wqe->length;
674 		qp->s_len = wqe->length;
675 		if (newreq) {
676 			qp->s_tail++;
677 			if (qp->s_tail >= qp->s_size)
678 				qp->s_tail = 0;
679 		}
680 		if (wqe->wr.opcode == IB_WR_RDMA_READ)
681 			qp->s_psn = wqe->lpsn + 1;
682 		else
683 			qp->s_psn++;
684 		break;
685 
686 	case OP(RDMA_READ_RESPONSE_FIRST):
687 		/*
688 		 * qp->s_state is normally set to the opcode of the
689 		 * last packet constructed for new requests and therefore
690 		 * is never set to RDMA read response.
691 		 * RDMA_READ_RESPONSE_FIRST is used by the ACK processing
692 		 * thread to indicate a SEND needs to be restarted from an
693 		 * earlier PSN without interfering with the sending thread.
694 		 * See restart_rc().
695 		 */
696 		qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu);
697 		/* FALLTHROUGH */
698 	case OP(SEND_FIRST):
699 		qp->s_state = OP(SEND_MIDDLE);
700 		/* FALLTHROUGH */
701 	case OP(SEND_MIDDLE):
702 		bth2 = mask_psn(qp->s_psn++);
703 		ss = &qp->s_sge;
704 		len = qp->s_len;
705 		if (len > pmtu) {
706 			len = pmtu;
707 			middle = HFI1_CAP_IS_KSET(SDMA_AHG);
708 			break;
709 		}
710 		if (wqe->wr.opcode == IB_WR_SEND) {
711 			qp->s_state = OP(SEND_LAST);
712 		} else if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
713 			qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
714 			/* Immediate data comes after the BTH */
715 			ohdr->u.imm_data = wqe->wr.ex.imm_data;
716 			hwords += 1;
717 		} else {
718 			qp->s_state = OP(SEND_LAST_WITH_INVALIDATE);
719 			/* invalidate data comes after the BTH */
720 			ohdr->u.ieth = cpu_to_be32(wqe->wr.ex.invalidate_rkey);
721 			hwords += 1;
722 		}
723 		if (wqe->wr.send_flags & IB_SEND_SOLICITED)
724 			bth0 |= IB_BTH_SOLICITED;
725 		bth2 |= IB_BTH_REQ_ACK;
726 		qp->s_cur++;
727 		if (qp->s_cur >= qp->s_size)
728 			qp->s_cur = 0;
729 		break;
730 
731 	case OP(RDMA_READ_RESPONSE_LAST):
732 		/*
733 		 * qp->s_state is normally set to the opcode of the
734 		 * last packet constructed for new requests and therefore
735 		 * is never set to RDMA read response.
736 		 * RDMA_READ_RESPONSE_LAST is used by the ACK processing
737 		 * thread to indicate a RDMA write needs to be restarted from
738 		 * an earlier PSN without interfering with the sending thread.
739 		 * See restart_rc().
740 		 */
741 		qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu);
742 		/* FALLTHROUGH */
743 	case OP(RDMA_WRITE_FIRST):
744 		qp->s_state = OP(RDMA_WRITE_MIDDLE);
745 		/* FALLTHROUGH */
746 	case OP(RDMA_WRITE_MIDDLE):
747 		bth2 = mask_psn(qp->s_psn++);
748 		ss = &qp->s_sge;
749 		len = qp->s_len;
750 		if (len > pmtu) {
751 			len = pmtu;
752 			middle = HFI1_CAP_IS_KSET(SDMA_AHG);
753 			break;
754 		}
755 		if (wqe->wr.opcode == IB_WR_RDMA_WRITE) {
756 			qp->s_state = OP(RDMA_WRITE_LAST);
757 		} else {
758 			qp->s_state = OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
759 			/* Immediate data comes after the BTH */
760 			ohdr->u.imm_data = wqe->wr.ex.imm_data;
761 			hwords += 1;
762 			if (wqe->wr.send_flags & IB_SEND_SOLICITED)
763 				bth0 |= IB_BTH_SOLICITED;
764 		}
765 		bth2 |= IB_BTH_REQ_ACK;
766 		qp->s_cur++;
767 		if (qp->s_cur >= qp->s_size)
768 			qp->s_cur = 0;
769 		break;
770 
771 	case OP(RDMA_READ_RESPONSE_MIDDLE):
772 		/*
773 		 * qp->s_state is normally set to the opcode of the
774 		 * last packet constructed for new requests and therefore
775 		 * is never set to RDMA read response.
776 		 * RDMA_READ_RESPONSE_MIDDLE is used by the ACK processing
777 		 * thread to indicate a RDMA read needs to be restarted from
778 		 * an earlier PSN without interfering with the sending thread.
779 		 * See restart_rc().
780 		 */
781 		len = (delta_psn(qp->s_psn, wqe->psn)) * pmtu;
782 		ohdr->u.rc.reth.vaddr =
783 			cpu_to_be64(wqe->rdma_wr.remote_addr + len);
784 		ohdr->u.rc.reth.rkey =
785 			cpu_to_be32(wqe->rdma_wr.rkey);
786 		ohdr->u.rc.reth.length = cpu_to_be32(wqe->length - len);
787 		qp->s_state = OP(RDMA_READ_REQUEST);
788 		hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
789 		bth2 = mask_psn(qp->s_psn) | IB_BTH_REQ_ACK;
790 		qp->s_psn = wqe->lpsn + 1;
791 		ss = NULL;
792 		len = 0;
793 		qp->s_cur++;
794 		if (qp->s_cur == qp->s_size)
795 			qp->s_cur = 0;
796 		break;
797 	}
798 	qp->s_sending_hpsn = bth2;
799 	delta = delta_psn(bth2, wqe->psn);
800 	if (delta && delta % HFI1_PSN_CREDIT == 0)
801 		bth2 |= IB_BTH_REQ_ACK;
802 	if (qp->s_flags & RVT_S_SEND_ONE) {
803 		qp->s_flags &= ~RVT_S_SEND_ONE;
804 		qp->s_flags |= RVT_S_WAIT_ACK;
805 		bth2 |= IB_BTH_REQ_ACK;
806 	}
807 	qp->s_len -= len;
808 	qp->s_hdrwords = hwords;
809 	ps->s_txreq->sde = priv->s_sde;
810 	qp->s_cur_sge = ss;
811 	qp->s_cur_size = len;
812 	hfi1_make_ruc_header(
813 		qp,
814 		ohdr,
815 		bth0 | (qp->s_state << 24),
816 		bth2,
817 		middle,
818 		ps);
819 	/* pbc */
820 	ps->s_txreq->hdr_dwords = qp->s_hdrwords + 2;
821 	return 1;
822 
823 done_free_tx:
824 	hfi1_put_txreq(ps->s_txreq);
825 	ps->s_txreq = NULL;
826 	return 1;
827 
828 bail:
829 	hfi1_put_txreq(ps->s_txreq);
830 
831 bail_no_tx:
832 	ps->s_txreq = NULL;
833 	qp->s_flags &= ~RVT_S_BUSY;
834 	qp->s_hdrwords = 0;
835 	return 0;
836 }
837 
838 /**
839  * hfi1_send_rc_ack - Construct an ACK packet and send it
840  * @qp: a pointer to the QP
841  *
842  * This is called from hfi1_rc_rcv() and handle_receive_interrupt().
843  * Note that RDMA reads and atomics are handled in the
844  * send side QP state and tasklet.
845  */
846 void hfi1_send_rc_ack(struct hfi1_ctxtdata *rcd, struct rvt_qp *qp,
847 		      int is_fecn)
848 {
849 	struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
850 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
851 	u64 pbc, pbc_flags = 0;
852 	u16 lrh0;
853 	u16 sc5;
854 	u32 bth0;
855 	u32 hwords;
856 	u32 vl, plen;
857 	struct send_context *sc;
858 	struct pio_buf *pbuf;
859 	struct hfi1_ib_header hdr;
860 	struct hfi1_other_headers *ohdr;
861 	unsigned long flags;
862 
863 	/* Don't send ACK or NAK if a RDMA read or atomic is pending. */
864 	if (qp->s_flags & RVT_S_RESP_PENDING)
865 		goto queue_ack;
866 
867 	/* Ensure s_rdma_ack_cnt changes are committed */
868 	smp_read_barrier_depends();
869 	if (qp->s_rdma_ack_cnt)
870 		goto queue_ack;
871 
872 	/* Construct the header */
873 	/* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4 */
874 	hwords = 6;
875 	if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
876 		hwords += hfi1_make_grh(ibp, &hdr.u.l.grh,
877 				       &qp->remote_ah_attr.grh, hwords, 0);
878 		ohdr = &hdr.u.l.oth;
879 		lrh0 = HFI1_LRH_GRH;
880 	} else {
881 		ohdr = &hdr.u.oth;
882 		lrh0 = HFI1_LRH_BTH;
883 	}
884 	/* read pkey_index w/o lock (its atomic) */
885 	bth0 = hfi1_get_pkey(ibp, qp->s_pkey_index) | (OP(ACKNOWLEDGE) << 24);
886 	if (qp->s_mig_state == IB_MIG_MIGRATED)
887 		bth0 |= IB_BTH_MIG_REQ;
888 	if (qp->r_nak_state)
889 		ohdr->u.aeth = cpu_to_be32((qp->r_msn & HFI1_MSN_MASK) |
890 					    (qp->r_nak_state <<
891 					     HFI1_AETH_CREDIT_SHIFT));
892 	else
893 		ohdr->u.aeth = hfi1_compute_aeth(qp);
894 	sc5 = ibp->sl_to_sc[qp->remote_ah_attr.sl];
895 	/* set PBC_DC_INFO bit (aka SC[4]) in pbc_flags */
896 	pbc_flags |= ((!!(sc5 & 0x10)) << PBC_DC_INFO_SHIFT);
897 	lrh0 |= (sc5 & 0xf) << 12 | (qp->remote_ah_attr.sl & 0xf) << 4;
898 	hdr.lrh[0] = cpu_to_be16(lrh0);
899 	hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
900 	hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC);
901 	hdr.lrh[3] = cpu_to_be16(ppd->lid | qp->remote_ah_attr.src_path_bits);
902 	ohdr->bth[0] = cpu_to_be32(bth0);
903 	ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
904 	ohdr->bth[1] |= cpu_to_be32((!!is_fecn) << HFI1_BECN_SHIFT);
905 	ohdr->bth[2] = cpu_to_be32(mask_psn(qp->r_ack_psn));
906 
907 	/* Don't try to send ACKs if the link isn't ACTIVE */
908 	if (driver_lstate(ppd) != IB_PORT_ACTIVE)
909 		return;
910 
911 	sc = rcd->sc;
912 	plen = 2 /* PBC */ + hwords;
913 	vl = sc_to_vlt(ppd->dd, sc5);
914 	pbc = create_pbc(ppd, pbc_flags, qp->srate_mbps, vl, plen);
915 
916 	pbuf = sc_buffer_alloc(sc, plen, NULL, NULL);
917 	if (!pbuf) {
918 		/*
919 		 * We have no room to send at the moment.  Pass
920 		 * responsibility for sending the ACK to the send tasklet
921 		 * so that when enough buffer space becomes available,
922 		 * the ACK is sent ahead of other outgoing packets.
923 		 */
924 		goto queue_ack;
925 	}
926 
927 	trace_ack_output_ibhdr(dd_from_ibdev(qp->ibqp.device), &hdr);
928 
929 	/* write the pbc and data */
930 	ppd->dd->pio_inline_send(ppd->dd, pbuf, pbc, &hdr, hwords);
931 
932 	return;
933 
934 queue_ack:
935 	this_cpu_inc(*ibp->rvp.rc_qacks);
936 	spin_lock_irqsave(&qp->s_lock, flags);
937 	qp->s_flags |= RVT_S_ACK_PENDING | RVT_S_RESP_PENDING;
938 	qp->s_nak_state = qp->r_nak_state;
939 	qp->s_ack_psn = qp->r_ack_psn;
940 	if (is_fecn)
941 		qp->s_flags |= RVT_S_ECN;
942 
943 	/* Schedule the send tasklet. */
944 	hfi1_schedule_send(qp);
945 	spin_unlock_irqrestore(&qp->s_lock, flags);
946 }
947 
948 /**
949  * reset_psn - reset the QP state to send starting from PSN
950  * @qp: the QP
951  * @psn: the packet sequence number to restart at
952  *
953  * This is called from hfi1_rc_rcv() to process an incoming RC ACK
954  * for the given QP.
955  * Called at interrupt level with the QP s_lock held.
956  */
957 static void reset_psn(struct rvt_qp *qp, u32 psn)
958 {
959 	u32 n = qp->s_acked;
960 	struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, n);
961 	u32 opcode;
962 
963 	qp->s_cur = n;
964 
965 	/*
966 	 * If we are starting the request from the beginning,
967 	 * let the normal send code handle initialization.
968 	 */
969 	if (cmp_psn(psn, wqe->psn) <= 0) {
970 		qp->s_state = OP(SEND_LAST);
971 		goto done;
972 	}
973 
974 	/* Find the work request opcode corresponding to the given PSN. */
975 	opcode = wqe->wr.opcode;
976 	for (;;) {
977 		int diff;
978 
979 		if (++n == qp->s_size)
980 			n = 0;
981 		if (n == qp->s_tail)
982 			break;
983 		wqe = rvt_get_swqe_ptr(qp, n);
984 		diff = cmp_psn(psn, wqe->psn);
985 		if (diff < 0)
986 			break;
987 		qp->s_cur = n;
988 		/*
989 		 * If we are starting the request from the beginning,
990 		 * let the normal send code handle initialization.
991 		 */
992 		if (diff == 0) {
993 			qp->s_state = OP(SEND_LAST);
994 			goto done;
995 		}
996 		opcode = wqe->wr.opcode;
997 	}
998 
999 	/*
1000 	 * Set the state to restart in the middle of a request.
1001 	 * Don't change the s_sge, s_cur_sge, or s_cur_size.
1002 	 * See hfi1_make_rc_req().
1003 	 */
1004 	switch (opcode) {
1005 	case IB_WR_SEND:
1006 	case IB_WR_SEND_WITH_IMM:
1007 		qp->s_state = OP(RDMA_READ_RESPONSE_FIRST);
1008 		break;
1009 
1010 	case IB_WR_RDMA_WRITE:
1011 	case IB_WR_RDMA_WRITE_WITH_IMM:
1012 		qp->s_state = OP(RDMA_READ_RESPONSE_LAST);
1013 		break;
1014 
1015 	case IB_WR_RDMA_READ:
1016 		qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE);
1017 		break;
1018 
1019 	default:
1020 		/*
1021 		 * This case shouldn't happen since its only
1022 		 * one PSN per req.
1023 		 */
1024 		qp->s_state = OP(SEND_LAST);
1025 	}
1026 done:
1027 	qp->s_psn = psn;
1028 	/*
1029 	 * Set RVT_S_WAIT_PSN as rc_complete() may start the timer
1030 	 * asynchronously before the send tasklet can get scheduled.
1031 	 * Doing it in hfi1_make_rc_req() is too late.
1032 	 */
1033 	if ((cmp_psn(qp->s_psn, qp->s_sending_hpsn) <= 0) &&
1034 	    (cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) <= 0))
1035 		qp->s_flags |= RVT_S_WAIT_PSN;
1036 	qp->s_flags &= ~RVT_S_AHG_VALID;
1037 }
1038 
1039 /*
1040  * Back up requester to resend the last un-ACKed request.
1041  * The QP r_lock and s_lock should be held and interrupts disabled.
1042  */
1043 static void restart_rc(struct rvt_qp *qp, u32 psn, int wait)
1044 {
1045 	struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1046 	struct hfi1_ibport *ibp;
1047 
1048 	if (qp->s_retry == 0) {
1049 		if (qp->s_mig_state == IB_MIG_ARMED) {
1050 			hfi1_migrate_qp(qp);
1051 			qp->s_retry = qp->s_retry_cnt;
1052 		} else if (qp->s_last == qp->s_acked) {
1053 			hfi1_send_complete(qp, wqe, IB_WC_RETRY_EXC_ERR);
1054 			rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1055 			return;
1056 		} else { /* need to handle delayed completion */
1057 			return;
1058 		}
1059 	} else {
1060 		qp->s_retry--;
1061 	}
1062 
1063 	ibp = to_iport(qp->ibqp.device, qp->port_num);
1064 	if (wqe->wr.opcode == IB_WR_RDMA_READ)
1065 		ibp->rvp.n_rc_resends++;
1066 	else
1067 		ibp->rvp.n_rc_resends += delta_psn(qp->s_psn, psn);
1068 
1069 	qp->s_flags &= ~(RVT_S_WAIT_FENCE | RVT_S_WAIT_RDMAR |
1070 			 RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_PSN |
1071 			 RVT_S_WAIT_ACK);
1072 	if (wait)
1073 		qp->s_flags |= RVT_S_SEND_ONE;
1074 	reset_psn(qp, psn);
1075 }
1076 
1077 /*
1078  * This is called from s_timer for missing responses.
1079  */
1080 void hfi1_rc_timeout(unsigned long arg)
1081 {
1082 	struct rvt_qp *qp = (struct rvt_qp *)arg;
1083 	struct hfi1_ibport *ibp;
1084 	unsigned long flags;
1085 
1086 	spin_lock_irqsave(&qp->r_lock, flags);
1087 	spin_lock(&qp->s_lock);
1088 	if (qp->s_flags & RVT_S_TIMER) {
1089 		ibp = to_iport(qp->ibqp.device, qp->port_num);
1090 		ibp->rvp.n_rc_timeouts++;
1091 		qp->s_flags &= ~RVT_S_TIMER;
1092 		del_timer(&qp->s_timer);
1093 		trace_hfi1_timeout(qp, qp->s_last_psn + 1);
1094 		restart_rc(qp, qp->s_last_psn + 1, 1);
1095 		hfi1_schedule_send(qp);
1096 	}
1097 	spin_unlock(&qp->s_lock);
1098 	spin_unlock_irqrestore(&qp->r_lock, flags);
1099 }
1100 
1101 /*
1102  * This is called from s_timer for RNR timeouts.
1103  */
1104 void hfi1_rc_rnr_retry(unsigned long arg)
1105 {
1106 	struct rvt_qp *qp = (struct rvt_qp *)arg;
1107 	unsigned long flags;
1108 
1109 	spin_lock_irqsave(&qp->s_lock, flags);
1110 	hfi1_stop_rnr_timer(qp);
1111 	hfi1_schedule_send(qp);
1112 	spin_unlock_irqrestore(&qp->s_lock, flags);
1113 }
1114 
1115 /*
1116  * Set qp->s_sending_psn to the next PSN after the given one.
1117  * This would be psn+1 except when RDMA reads are present.
1118  */
1119 static void reset_sending_psn(struct rvt_qp *qp, u32 psn)
1120 {
1121 	struct rvt_swqe *wqe;
1122 	u32 n = qp->s_last;
1123 
1124 	/* Find the work request corresponding to the given PSN. */
1125 	for (;;) {
1126 		wqe = rvt_get_swqe_ptr(qp, n);
1127 		if (cmp_psn(psn, wqe->lpsn) <= 0) {
1128 			if (wqe->wr.opcode == IB_WR_RDMA_READ)
1129 				qp->s_sending_psn = wqe->lpsn + 1;
1130 			else
1131 				qp->s_sending_psn = psn + 1;
1132 			break;
1133 		}
1134 		if (++n == qp->s_size)
1135 			n = 0;
1136 		if (n == qp->s_tail)
1137 			break;
1138 	}
1139 }
1140 
1141 /*
1142  * This should be called with the QP s_lock held and interrupts disabled.
1143  */
1144 void hfi1_rc_send_complete(struct rvt_qp *qp, struct hfi1_ib_header *hdr)
1145 {
1146 	struct hfi1_other_headers *ohdr;
1147 	struct rvt_swqe *wqe;
1148 	struct ib_wc wc;
1149 	unsigned i;
1150 	u32 opcode;
1151 	u32 psn;
1152 
1153 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_OR_FLUSH_SEND))
1154 		return;
1155 
1156 	/* Find out where the BTH is */
1157 	if ((be16_to_cpu(hdr->lrh[0]) & 3) == HFI1_LRH_BTH)
1158 		ohdr = &hdr->u.oth;
1159 	else
1160 		ohdr = &hdr->u.l.oth;
1161 
1162 	opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
1163 	if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
1164 	    opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
1165 		WARN_ON(!qp->s_rdma_ack_cnt);
1166 		qp->s_rdma_ack_cnt--;
1167 		return;
1168 	}
1169 
1170 	psn = be32_to_cpu(ohdr->bth[2]);
1171 	reset_sending_psn(qp, psn);
1172 
1173 	/*
1174 	 * Start timer after a packet requesting an ACK has been sent and
1175 	 * there are still requests that haven't been acked.
1176 	 */
1177 	if ((psn & IB_BTH_REQ_ACK) && qp->s_acked != qp->s_tail &&
1178 	    !(qp->s_flags &
1179 		(RVT_S_TIMER | RVT_S_WAIT_RNR | RVT_S_WAIT_PSN)) &&
1180 		(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
1181 		hfi1_add_retry_timer(qp);
1182 
1183 	while (qp->s_last != qp->s_acked) {
1184 		u32 s_last;
1185 
1186 		wqe = rvt_get_swqe_ptr(qp, qp->s_last);
1187 		if (cmp_psn(wqe->lpsn, qp->s_sending_psn) >= 0 &&
1188 		    cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) <= 0)
1189 			break;
1190 		s_last = qp->s_last;
1191 		if (++s_last >= qp->s_size)
1192 			s_last = 0;
1193 		qp->s_last = s_last;
1194 		/* see post_send() */
1195 		barrier();
1196 		for (i = 0; i < wqe->wr.num_sge; i++) {
1197 			struct rvt_sge *sge = &wqe->sg_list[i];
1198 
1199 			rvt_put_mr(sge->mr);
1200 		}
1201 		/* Post a send completion queue entry if requested. */
1202 		if (!(qp->s_flags & RVT_S_SIGNAL_REQ_WR) ||
1203 		    (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
1204 			memset(&wc, 0, sizeof(wc));
1205 			wc.wr_id = wqe->wr.wr_id;
1206 			wc.status = IB_WC_SUCCESS;
1207 			wc.opcode = ib_hfi1_wc_opcode[wqe->wr.opcode];
1208 			wc.byte_len = wqe->length;
1209 			wc.qp = &qp->ibqp;
1210 			rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.send_cq), &wc, 0);
1211 		}
1212 	}
1213 	/*
1214 	 * If we were waiting for sends to complete before re-sending,
1215 	 * and they are now complete, restart sending.
1216 	 */
1217 	trace_hfi1_sendcomplete(qp, psn);
1218 	if (qp->s_flags & RVT_S_WAIT_PSN &&
1219 	    cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) > 0) {
1220 		qp->s_flags &= ~RVT_S_WAIT_PSN;
1221 		qp->s_sending_psn = qp->s_psn;
1222 		qp->s_sending_hpsn = qp->s_psn - 1;
1223 		hfi1_schedule_send(qp);
1224 	}
1225 }
1226 
1227 static inline void update_last_psn(struct rvt_qp *qp, u32 psn)
1228 {
1229 	qp->s_last_psn = psn;
1230 }
1231 
1232 /*
1233  * Generate a SWQE completion.
1234  * This is similar to hfi1_send_complete but has to check to be sure
1235  * that the SGEs are not being referenced if the SWQE is being resent.
1236  */
1237 static struct rvt_swqe *do_rc_completion(struct rvt_qp *qp,
1238 					 struct rvt_swqe *wqe,
1239 					 struct hfi1_ibport *ibp)
1240 {
1241 	struct ib_wc wc;
1242 	unsigned i;
1243 
1244 	/*
1245 	 * Don't decrement refcount and don't generate a
1246 	 * completion if the SWQE is being resent until the send
1247 	 * is finished.
1248 	 */
1249 	if (cmp_psn(wqe->lpsn, qp->s_sending_psn) < 0 ||
1250 	    cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) > 0) {
1251 		u32 s_last;
1252 
1253 		for (i = 0; i < wqe->wr.num_sge; i++) {
1254 			struct rvt_sge *sge = &wqe->sg_list[i];
1255 
1256 			rvt_put_mr(sge->mr);
1257 		}
1258 		s_last = qp->s_last;
1259 		if (++s_last >= qp->s_size)
1260 			s_last = 0;
1261 		qp->s_last = s_last;
1262 		/* see post_send() */
1263 		barrier();
1264 		/* Post a send completion queue entry if requested. */
1265 		if (!(qp->s_flags & RVT_S_SIGNAL_REQ_WR) ||
1266 		    (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
1267 			memset(&wc, 0, sizeof(wc));
1268 			wc.wr_id = wqe->wr.wr_id;
1269 			wc.status = IB_WC_SUCCESS;
1270 			wc.opcode = ib_hfi1_wc_opcode[wqe->wr.opcode];
1271 			wc.byte_len = wqe->length;
1272 			wc.qp = &qp->ibqp;
1273 			rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.send_cq), &wc, 0);
1274 		}
1275 	} else {
1276 		struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
1277 
1278 		this_cpu_inc(*ibp->rvp.rc_delayed_comp);
1279 		/*
1280 		 * If send progress not running attempt to progress
1281 		 * SDMA queue.
1282 		 */
1283 		if (ppd->dd->flags & HFI1_HAS_SEND_DMA) {
1284 			struct sdma_engine *engine;
1285 			u8 sc5;
1286 
1287 			/* For now use sc to find engine */
1288 			sc5 = ibp->sl_to_sc[qp->remote_ah_attr.sl];
1289 			engine = qp_to_sdma_engine(qp, sc5);
1290 			sdma_engine_progress_schedule(engine);
1291 		}
1292 	}
1293 
1294 	qp->s_retry = qp->s_retry_cnt;
1295 	update_last_psn(qp, wqe->lpsn);
1296 
1297 	/*
1298 	 * If we are completing a request which is in the process of
1299 	 * being resent, we can stop re-sending it since we know the
1300 	 * responder has already seen it.
1301 	 */
1302 	if (qp->s_acked == qp->s_cur) {
1303 		if (++qp->s_cur >= qp->s_size)
1304 			qp->s_cur = 0;
1305 		qp->s_acked = qp->s_cur;
1306 		wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
1307 		if (qp->s_acked != qp->s_tail) {
1308 			qp->s_state = OP(SEND_LAST);
1309 			qp->s_psn = wqe->psn;
1310 		}
1311 	} else {
1312 		if (++qp->s_acked >= qp->s_size)
1313 			qp->s_acked = 0;
1314 		if (qp->state == IB_QPS_SQD && qp->s_acked == qp->s_cur)
1315 			qp->s_draining = 0;
1316 		wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1317 	}
1318 	return wqe;
1319 }
1320 
1321 /**
1322  * do_rc_ack - process an incoming RC ACK
1323  * @qp: the QP the ACK came in on
1324  * @psn: the packet sequence number of the ACK
1325  * @opcode: the opcode of the request that resulted in the ACK
1326  *
1327  * This is called from rc_rcv_resp() to process an incoming RC ACK
1328  * for the given QP.
1329  * May be called at interrupt level, with the QP s_lock held.
1330  * Returns 1 if OK, 0 if current operation should be aborted (NAK).
1331  */
1332 static int do_rc_ack(struct rvt_qp *qp, u32 aeth, u32 psn, int opcode,
1333 		     u64 val, struct hfi1_ctxtdata *rcd)
1334 {
1335 	struct hfi1_ibport *ibp;
1336 	enum ib_wc_status status;
1337 	struct rvt_swqe *wqe;
1338 	int ret = 0;
1339 	u32 ack_psn;
1340 	int diff;
1341 	unsigned long to;
1342 
1343 	/*
1344 	 * Note that NAKs implicitly ACK outstanding SEND and RDMA write
1345 	 * requests and implicitly NAK RDMA read and atomic requests issued
1346 	 * before the NAK'ed request.  The MSN won't include the NAK'ed
1347 	 * request but will include an ACK'ed request(s).
1348 	 */
1349 	ack_psn = psn;
1350 	if (aeth >> 29)
1351 		ack_psn--;
1352 	wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1353 	ibp = to_iport(qp->ibqp.device, qp->port_num);
1354 
1355 	/*
1356 	 * The MSN might be for a later WQE than the PSN indicates so
1357 	 * only complete WQEs that the PSN finishes.
1358 	 */
1359 	while ((diff = delta_psn(ack_psn, wqe->lpsn)) >= 0) {
1360 		/*
1361 		 * RDMA_READ_RESPONSE_ONLY is a special case since
1362 		 * we want to generate completion events for everything
1363 		 * before the RDMA read, copy the data, then generate
1364 		 * the completion for the read.
1365 		 */
1366 		if (wqe->wr.opcode == IB_WR_RDMA_READ &&
1367 		    opcode == OP(RDMA_READ_RESPONSE_ONLY) &&
1368 		    diff == 0) {
1369 			ret = 1;
1370 			goto bail_stop;
1371 		}
1372 		/*
1373 		 * If this request is a RDMA read or atomic, and the ACK is
1374 		 * for a later operation, this ACK NAKs the RDMA read or
1375 		 * atomic.  In other words, only a RDMA_READ_LAST or ONLY
1376 		 * can ACK a RDMA read and likewise for atomic ops.  Note
1377 		 * that the NAK case can only happen if relaxed ordering is
1378 		 * used and requests are sent after an RDMA read or atomic
1379 		 * is sent but before the response is received.
1380 		 */
1381 		if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
1382 		     (opcode != OP(RDMA_READ_RESPONSE_LAST) || diff != 0)) ||
1383 		    ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1384 		      wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
1385 		     (opcode != OP(ATOMIC_ACKNOWLEDGE) || diff != 0))) {
1386 			/* Retry this request. */
1387 			if (!(qp->r_flags & RVT_R_RDMAR_SEQ)) {
1388 				qp->r_flags |= RVT_R_RDMAR_SEQ;
1389 				restart_rc(qp, qp->s_last_psn + 1, 0);
1390 				if (list_empty(&qp->rspwait)) {
1391 					qp->r_flags |= RVT_R_RSP_SEND;
1392 					atomic_inc(&qp->refcount);
1393 					list_add_tail(&qp->rspwait,
1394 						      &rcd->qp_wait_list);
1395 				}
1396 			}
1397 			/*
1398 			 * No need to process the ACK/NAK since we are
1399 			 * restarting an earlier request.
1400 			 */
1401 			goto bail_stop;
1402 		}
1403 		if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1404 		    wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
1405 			u64 *vaddr = wqe->sg_list[0].vaddr;
1406 			*vaddr = val;
1407 		}
1408 		if (qp->s_num_rd_atomic &&
1409 		    (wqe->wr.opcode == IB_WR_RDMA_READ ||
1410 		     wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1411 		     wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)) {
1412 			qp->s_num_rd_atomic--;
1413 			/* Restart sending task if fence is complete */
1414 			if ((qp->s_flags & RVT_S_WAIT_FENCE) &&
1415 			    !qp->s_num_rd_atomic) {
1416 				qp->s_flags &= ~(RVT_S_WAIT_FENCE |
1417 						 RVT_S_WAIT_ACK);
1418 				hfi1_schedule_send(qp);
1419 			} else if (qp->s_flags & RVT_S_WAIT_RDMAR) {
1420 				qp->s_flags &= ~(RVT_S_WAIT_RDMAR |
1421 						 RVT_S_WAIT_ACK);
1422 				hfi1_schedule_send(qp);
1423 			}
1424 		}
1425 		wqe = do_rc_completion(qp, wqe, ibp);
1426 		if (qp->s_acked == qp->s_tail)
1427 			break;
1428 	}
1429 
1430 	switch (aeth >> 29) {
1431 	case 0:         /* ACK */
1432 		this_cpu_inc(*ibp->rvp.rc_acks);
1433 		if (qp->s_acked != qp->s_tail) {
1434 			/*
1435 			 * We are expecting more ACKs so
1436 			 * mod the retry timer.
1437 			 */
1438 			hfi1_mod_retry_timer(qp);
1439 			/*
1440 			 * We can stop re-sending the earlier packets and
1441 			 * continue with the next packet the receiver wants.
1442 			 */
1443 			if (cmp_psn(qp->s_psn, psn) <= 0)
1444 				reset_psn(qp, psn + 1);
1445 		} else {
1446 			/* No more acks - kill all timers */
1447 			hfi1_stop_rc_timers(qp);
1448 			if (cmp_psn(qp->s_psn, psn) <= 0) {
1449 				qp->s_state = OP(SEND_LAST);
1450 				qp->s_psn = psn + 1;
1451 			}
1452 		}
1453 		if (qp->s_flags & RVT_S_WAIT_ACK) {
1454 			qp->s_flags &= ~RVT_S_WAIT_ACK;
1455 			hfi1_schedule_send(qp);
1456 		}
1457 		hfi1_get_credit(qp, aeth);
1458 		qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1459 		qp->s_retry = qp->s_retry_cnt;
1460 		update_last_psn(qp, psn);
1461 		return 1;
1462 
1463 	case 1:         /* RNR NAK */
1464 		ibp->rvp.n_rnr_naks++;
1465 		if (qp->s_acked == qp->s_tail)
1466 			goto bail_stop;
1467 		if (qp->s_flags & RVT_S_WAIT_RNR)
1468 			goto bail_stop;
1469 		if (qp->s_rnr_retry == 0) {
1470 			status = IB_WC_RNR_RETRY_EXC_ERR;
1471 			goto class_b;
1472 		}
1473 		if (qp->s_rnr_retry_cnt < 7)
1474 			qp->s_rnr_retry--;
1475 
1476 		/* The last valid PSN is the previous PSN. */
1477 		update_last_psn(qp, psn - 1);
1478 
1479 		ibp->rvp.n_rc_resends += delta_psn(qp->s_psn, psn);
1480 
1481 		reset_psn(qp, psn);
1482 
1483 		qp->s_flags &= ~(RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_ACK);
1484 		hfi1_stop_rc_timers(qp);
1485 		to =
1486 			ib_hfi1_rnr_table[(aeth >> HFI1_AETH_CREDIT_SHIFT) &
1487 					   HFI1_AETH_CREDIT_MASK];
1488 		hfi1_add_rnr_timer(qp, to);
1489 		return 0;
1490 
1491 	case 3:         /* NAK */
1492 		if (qp->s_acked == qp->s_tail)
1493 			goto bail_stop;
1494 		/* The last valid PSN is the previous PSN. */
1495 		update_last_psn(qp, psn - 1);
1496 		switch ((aeth >> HFI1_AETH_CREDIT_SHIFT) &
1497 			HFI1_AETH_CREDIT_MASK) {
1498 		case 0: /* PSN sequence error */
1499 			ibp->rvp.n_seq_naks++;
1500 			/*
1501 			 * Back up to the responder's expected PSN.
1502 			 * Note that we might get a NAK in the middle of an
1503 			 * RDMA READ response which terminates the RDMA
1504 			 * READ.
1505 			 */
1506 			restart_rc(qp, psn, 0);
1507 			hfi1_schedule_send(qp);
1508 			break;
1509 
1510 		case 1: /* Invalid Request */
1511 			status = IB_WC_REM_INV_REQ_ERR;
1512 			ibp->rvp.n_other_naks++;
1513 			goto class_b;
1514 
1515 		case 2: /* Remote Access Error */
1516 			status = IB_WC_REM_ACCESS_ERR;
1517 			ibp->rvp.n_other_naks++;
1518 			goto class_b;
1519 
1520 		case 3: /* Remote Operation Error */
1521 			status = IB_WC_REM_OP_ERR;
1522 			ibp->rvp.n_other_naks++;
1523 class_b:
1524 			if (qp->s_last == qp->s_acked) {
1525 				hfi1_send_complete(qp, wqe, status);
1526 				rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1527 			}
1528 			break;
1529 
1530 		default:
1531 			/* Ignore other reserved NAK error codes */
1532 			goto reserved;
1533 		}
1534 		qp->s_retry = qp->s_retry_cnt;
1535 		qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1536 		goto bail_stop;
1537 
1538 	default:                /* 2: reserved */
1539 reserved:
1540 		/* Ignore reserved NAK codes. */
1541 		goto bail_stop;
1542 	}
1543 	/* cannot be reached  */
1544 bail_stop:
1545 	hfi1_stop_rc_timers(qp);
1546 	return ret;
1547 }
1548 
1549 /*
1550  * We have seen an out of sequence RDMA read middle or last packet.
1551  * This ACKs SENDs and RDMA writes up to the first RDMA read or atomic SWQE.
1552  */
1553 static void rdma_seq_err(struct rvt_qp *qp, struct hfi1_ibport *ibp, u32 psn,
1554 			 struct hfi1_ctxtdata *rcd)
1555 {
1556 	struct rvt_swqe *wqe;
1557 
1558 	/* Remove QP from retry timer */
1559 	hfi1_stop_rc_timers(qp);
1560 
1561 	wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1562 
1563 	while (cmp_psn(psn, wqe->lpsn) > 0) {
1564 		if (wqe->wr.opcode == IB_WR_RDMA_READ ||
1565 		    wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1566 		    wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
1567 			break;
1568 		wqe = do_rc_completion(qp, wqe, ibp);
1569 	}
1570 
1571 	ibp->rvp.n_rdma_seq++;
1572 	qp->r_flags |= RVT_R_RDMAR_SEQ;
1573 	restart_rc(qp, qp->s_last_psn + 1, 0);
1574 	if (list_empty(&qp->rspwait)) {
1575 		qp->r_flags |= RVT_R_RSP_SEND;
1576 		atomic_inc(&qp->refcount);
1577 		list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
1578 	}
1579 }
1580 
1581 /**
1582  * rc_rcv_resp - process an incoming RC response packet
1583  * @ibp: the port this packet came in on
1584  * @ohdr: the other headers for this packet
1585  * @data: the packet data
1586  * @tlen: the packet length
1587  * @qp: the QP for this packet
1588  * @opcode: the opcode for this packet
1589  * @psn: the packet sequence number for this packet
1590  * @hdrsize: the header length
1591  * @pmtu: the path MTU
1592  *
1593  * This is called from hfi1_rc_rcv() to process an incoming RC response
1594  * packet for the given QP.
1595  * Called at interrupt level.
1596  */
1597 static void rc_rcv_resp(struct hfi1_ibport *ibp,
1598 			struct hfi1_other_headers *ohdr,
1599 			void *data, u32 tlen, struct rvt_qp *qp,
1600 			u32 opcode, u32 psn, u32 hdrsize, u32 pmtu,
1601 			struct hfi1_ctxtdata *rcd)
1602 {
1603 	struct rvt_swqe *wqe;
1604 	enum ib_wc_status status;
1605 	unsigned long flags;
1606 	int diff;
1607 	u32 pad;
1608 	u32 aeth;
1609 	u64 val;
1610 
1611 	spin_lock_irqsave(&qp->s_lock, flags);
1612 
1613 	trace_hfi1_ack(qp, psn);
1614 
1615 	/* Ignore invalid responses. */
1616 	smp_read_barrier_depends(); /* see post_one_send */
1617 	if (cmp_psn(psn, ACCESS_ONCE(qp->s_next_psn)) >= 0)
1618 		goto ack_done;
1619 
1620 	/* Ignore duplicate responses. */
1621 	diff = cmp_psn(psn, qp->s_last_psn);
1622 	if (unlikely(diff <= 0)) {
1623 		/* Update credits for "ghost" ACKs */
1624 		if (diff == 0 && opcode == OP(ACKNOWLEDGE)) {
1625 			aeth = be32_to_cpu(ohdr->u.aeth);
1626 			if ((aeth >> 29) == 0)
1627 				hfi1_get_credit(qp, aeth);
1628 		}
1629 		goto ack_done;
1630 	}
1631 
1632 	/*
1633 	 * Skip everything other than the PSN we expect, if we are waiting
1634 	 * for a reply to a restarted RDMA read or atomic op.
1635 	 */
1636 	if (qp->r_flags & RVT_R_RDMAR_SEQ) {
1637 		if (cmp_psn(psn, qp->s_last_psn + 1) != 0)
1638 			goto ack_done;
1639 		qp->r_flags &= ~RVT_R_RDMAR_SEQ;
1640 	}
1641 
1642 	if (unlikely(qp->s_acked == qp->s_tail))
1643 		goto ack_done;
1644 	wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1645 	status = IB_WC_SUCCESS;
1646 
1647 	switch (opcode) {
1648 	case OP(ACKNOWLEDGE):
1649 	case OP(ATOMIC_ACKNOWLEDGE):
1650 	case OP(RDMA_READ_RESPONSE_FIRST):
1651 		aeth = be32_to_cpu(ohdr->u.aeth);
1652 		if (opcode == OP(ATOMIC_ACKNOWLEDGE)) {
1653 			__be32 *p = ohdr->u.at.atomic_ack_eth;
1654 
1655 			val = ((u64)be32_to_cpu(p[0]) << 32) |
1656 				be32_to_cpu(p[1]);
1657 		} else {
1658 			val = 0;
1659 		}
1660 		if (!do_rc_ack(qp, aeth, psn, opcode, val, rcd) ||
1661 		    opcode != OP(RDMA_READ_RESPONSE_FIRST))
1662 			goto ack_done;
1663 		wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1664 		if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1665 			goto ack_op_err;
1666 		/*
1667 		 * If this is a response to a resent RDMA read, we
1668 		 * have to be careful to copy the data to the right
1669 		 * location.
1670 		 */
1671 		qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1672 						  wqe, psn, pmtu);
1673 		goto read_middle;
1674 
1675 	case OP(RDMA_READ_RESPONSE_MIDDLE):
1676 		/* no AETH, no ACK */
1677 		if (unlikely(cmp_psn(psn, qp->s_last_psn + 1)))
1678 			goto ack_seq_err;
1679 		if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1680 			goto ack_op_err;
1681 read_middle:
1682 		if (unlikely(tlen != (hdrsize + pmtu + 4)))
1683 			goto ack_len_err;
1684 		if (unlikely(pmtu >= qp->s_rdma_read_len))
1685 			goto ack_len_err;
1686 
1687 		/*
1688 		 * We got a response so update the timeout.
1689 		 * 4.096 usec. * (1 << qp->timeout)
1690 		 */
1691 		qp->s_flags |= RVT_S_TIMER;
1692 		mod_timer(&qp->s_timer, jiffies + qp->timeout_jiffies);
1693 		if (qp->s_flags & RVT_S_WAIT_ACK) {
1694 			qp->s_flags &= ~RVT_S_WAIT_ACK;
1695 			hfi1_schedule_send(qp);
1696 		}
1697 
1698 		if (opcode == OP(RDMA_READ_RESPONSE_MIDDLE))
1699 			qp->s_retry = qp->s_retry_cnt;
1700 
1701 		/*
1702 		 * Update the RDMA receive state but do the copy w/o
1703 		 * holding the locks and blocking interrupts.
1704 		 */
1705 		qp->s_rdma_read_len -= pmtu;
1706 		update_last_psn(qp, psn);
1707 		spin_unlock_irqrestore(&qp->s_lock, flags);
1708 		hfi1_copy_sge(&qp->s_rdma_read_sge, data, pmtu, 0, 0);
1709 		goto bail;
1710 
1711 	case OP(RDMA_READ_RESPONSE_ONLY):
1712 		aeth = be32_to_cpu(ohdr->u.aeth);
1713 		if (!do_rc_ack(qp, aeth, psn, opcode, 0, rcd))
1714 			goto ack_done;
1715 		/* Get the number of bytes the message was padded by. */
1716 		pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1717 		/*
1718 		 * Check that the data size is >= 0 && <= pmtu.
1719 		 * Remember to account for ICRC (4).
1720 		 */
1721 		if (unlikely(tlen < (hdrsize + pad + 4)))
1722 			goto ack_len_err;
1723 		/*
1724 		 * If this is a response to a resent RDMA read, we
1725 		 * have to be careful to copy the data to the right
1726 		 * location.
1727 		 */
1728 		wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1729 		qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1730 						  wqe, psn, pmtu);
1731 		goto read_last;
1732 
1733 	case OP(RDMA_READ_RESPONSE_LAST):
1734 		/* ACKs READ req. */
1735 		if (unlikely(cmp_psn(psn, qp->s_last_psn + 1)))
1736 			goto ack_seq_err;
1737 		if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1738 			goto ack_op_err;
1739 		/* Get the number of bytes the message was padded by. */
1740 		pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1741 		/*
1742 		 * Check that the data size is >= 1 && <= pmtu.
1743 		 * Remember to account for ICRC (4).
1744 		 */
1745 		if (unlikely(tlen <= (hdrsize + pad + 4)))
1746 			goto ack_len_err;
1747 read_last:
1748 		tlen -= hdrsize + pad + 4;
1749 		if (unlikely(tlen != qp->s_rdma_read_len))
1750 			goto ack_len_err;
1751 		aeth = be32_to_cpu(ohdr->u.aeth);
1752 		hfi1_copy_sge(&qp->s_rdma_read_sge, data, tlen, 0, 0);
1753 		WARN_ON(qp->s_rdma_read_sge.num_sge);
1754 		(void)do_rc_ack(qp, aeth, psn,
1755 				 OP(RDMA_READ_RESPONSE_LAST), 0, rcd);
1756 		goto ack_done;
1757 	}
1758 
1759 ack_op_err:
1760 	status = IB_WC_LOC_QP_OP_ERR;
1761 	goto ack_err;
1762 
1763 ack_seq_err:
1764 	rdma_seq_err(qp, ibp, psn, rcd);
1765 	goto ack_done;
1766 
1767 ack_len_err:
1768 	status = IB_WC_LOC_LEN_ERR;
1769 ack_err:
1770 	if (qp->s_last == qp->s_acked) {
1771 		hfi1_send_complete(qp, wqe, status);
1772 		rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1773 	}
1774 ack_done:
1775 	spin_unlock_irqrestore(&qp->s_lock, flags);
1776 bail:
1777 	return;
1778 }
1779 
1780 static inline void rc_defered_ack(struct hfi1_ctxtdata *rcd,
1781 				  struct rvt_qp *qp)
1782 {
1783 	if (list_empty(&qp->rspwait)) {
1784 		qp->r_flags |= RVT_R_RSP_NAK;
1785 		atomic_inc(&qp->refcount);
1786 		list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
1787 	}
1788 }
1789 
1790 static inline void rc_cancel_ack(struct rvt_qp *qp)
1791 {
1792 	struct hfi1_qp_priv *priv = qp->priv;
1793 
1794 	priv->r_adefered = 0;
1795 	if (list_empty(&qp->rspwait))
1796 		return;
1797 	list_del_init(&qp->rspwait);
1798 	qp->r_flags &= ~RVT_R_RSP_NAK;
1799 	if (atomic_dec_and_test(&qp->refcount))
1800 		wake_up(&qp->wait);
1801 }
1802 
1803 /**
1804  * rc_rcv_error - process an incoming duplicate or error RC packet
1805  * @ohdr: the other headers for this packet
1806  * @data: the packet data
1807  * @qp: the QP for this packet
1808  * @opcode: the opcode for this packet
1809  * @psn: the packet sequence number for this packet
1810  * @diff: the difference between the PSN and the expected PSN
1811  *
1812  * This is called from hfi1_rc_rcv() to process an unexpected
1813  * incoming RC packet for the given QP.
1814  * Called at interrupt level.
1815  * Return 1 if no more processing is needed; otherwise return 0 to
1816  * schedule a response to be sent.
1817  */
1818 static noinline int rc_rcv_error(struct hfi1_other_headers *ohdr, void *data,
1819 				 struct rvt_qp *qp, u32 opcode, u32 psn,
1820 				 int diff, struct hfi1_ctxtdata *rcd)
1821 {
1822 	struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
1823 	struct rvt_ack_entry *e;
1824 	unsigned long flags;
1825 	u8 i, prev;
1826 	int old_req;
1827 
1828 	trace_hfi1_rcv_error(qp, psn);
1829 	if (diff > 0) {
1830 		/*
1831 		 * Packet sequence error.
1832 		 * A NAK will ACK earlier sends and RDMA writes.
1833 		 * Don't queue the NAK if we already sent one.
1834 		 */
1835 		if (!qp->r_nak_state) {
1836 			ibp->rvp.n_rc_seqnak++;
1837 			qp->r_nak_state = IB_NAK_PSN_ERROR;
1838 			/* Use the expected PSN. */
1839 			qp->r_ack_psn = qp->r_psn;
1840 			/*
1841 			 * Wait to send the sequence NAK until all packets
1842 			 * in the receive queue have been processed.
1843 			 * Otherwise, we end up propagating congestion.
1844 			 */
1845 			rc_defered_ack(rcd, qp);
1846 		}
1847 		goto done;
1848 	}
1849 
1850 	/*
1851 	 * Handle a duplicate request.  Don't re-execute SEND, RDMA
1852 	 * write or atomic op.  Don't NAK errors, just silently drop
1853 	 * the duplicate request.  Note that r_sge, r_len, and
1854 	 * r_rcv_len may be in use so don't modify them.
1855 	 *
1856 	 * We are supposed to ACK the earliest duplicate PSN but we
1857 	 * can coalesce an outstanding duplicate ACK.  We have to
1858 	 * send the earliest so that RDMA reads can be restarted at
1859 	 * the requester's expected PSN.
1860 	 *
1861 	 * First, find where this duplicate PSN falls within the
1862 	 * ACKs previously sent.
1863 	 * old_req is true if there is an older response that is scheduled
1864 	 * to be sent before sending this one.
1865 	 */
1866 	e = NULL;
1867 	old_req = 1;
1868 	ibp->rvp.n_rc_dupreq++;
1869 
1870 	spin_lock_irqsave(&qp->s_lock, flags);
1871 
1872 	for (i = qp->r_head_ack_queue; ; i = prev) {
1873 		if (i == qp->s_tail_ack_queue)
1874 			old_req = 0;
1875 		if (i)
1876 			prev = i - 1;
1877 		else
1878 			prev = HFI1_MAX_RDMA_ATOMIC;
1879 		if (prev == qp->r_head_ack_queue) {
1880 			e = NULL;
1881 			break;
1882 		}
1883 		e = &qp->s_ack_queue[prev];
1884 		if (!e->opcode) {
1885 			e = NULL;
1886 			break;
1887 		}
1888 		if (cmp_psn(psn, e->psn) >= 0) {
1889 			if (prev == qp->s_tail_ack_queue &&
1890 			    cmp_psn(psn, e->lpsn) <= 0)
1891 				old_req = 0;
1892 			break;
1893 		}
1894 	}
1895 	switch (opcode) {
1896 	case OP(RDMA_READ_REQUEST): {
1897 		struct ib_reth *reth;
1898 		u32 offset;
1899 		u32 len;
1900 
1901 		/*
1902 		 * If we didn't find the RDMA read request in the ack queue,
1903 		 * we can ignore this request.
1904 		 */
1905 		if (!e || e->opcode != OP(RDMA_READ_REQUEST))
1906 			goto unlock_done;
1907 		/* RETH comes after BTH */
1908 		reth = &ohdr->u.rc.reth;
1909 		/*
1910 		 * Address range must be a subset of the original
1911 		 * request and start on pmtu boundaries.
1912 		 * We reuse the old ack_queue slot since the requester
1913 		 * should not back up and request an earlier PSN for the
1914 		 * same request.
1915 		 */
1916 		offset = delta_psn(psn, e->psn) * qp->pmtu;
1917 		len = be32_to_cpu(reth->length);
1918 		if (unlikely(offset + len != e->rdma_sge.sge_length))
1919 			goto unlock_done;
1920 		if (e->rdma_sge.mr) {
1921 			rvt_put_mr(e->rdma_sge.mr);
1922 			e->rdma_sge.mr = NULL;
1923 		}
1924 		if (len != 0) {
1925 			u32 rkey = be32_to_cpu(reth->rkey);
1926 			u64 vaddr = be64_to_cpu(reth->vaddr);
1927 			int ok;
1928 
1929 			ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr, rkey,
1930 					 IB_ACCESS_REMOTE_READ);
1931 			if (unlikely(!ok))
1932 				goto unlock_done;
1933 		} else {
1934 			e->rdma_sge.vaddr = NULL;
1935 			e->rdma_sge.length = 0;
1936 			e->rdma_sge.sge_length = 0;
1937 		}
1938 		e->psn = psn;
1939 		if (old_req)
1940 			goto unlock_done;
1941 		qp->s_tail_ack_queue = prev;
1942 		break;
1943 	}
1944 
1945 	case OP(COMPARE_SWAP):
1946 	case OP(FETCH_ADD): {
1947 		/*
1948 		 * If we didn't find the atomic request in the ack queue
1949 		 * or the send tasklet is already backed up to send an
1950 		 * earlier entry, we can ignore this request.
1951 		 */
1952 		if (!e || e->opcode != (u8)opcode || old_req)
1953 			goto unlock_done;
1954 		qp->s_tail_ack_queue = prev;
1955 		break;
1956 	}
1957 
1958 	default:
1959 		/*
1960 		 * Ignore this operation if it doesn't request an ACK
1961 		 * or an earlier RDMA read or atomic is going to be resent.
1962 		 */
1963 		if (!(psn & IB_BTH_REQ_ACK) || old_req)
1964 			goto unlock_done;
1965 		/*
1966 		 * Resend the most recent ACK if this request is
1967 		 * after all the previous RDMA reads and atomics.
1968 		 */
1969 		if (i == qp->r_head_ack_queue) {
1970 			spin_unlock_irqrestore(&qp->s_lock, flags);
1971 			qp->r_nak_state = 0;
1972 			qp->r_ack_psn = qp->r_psn - 1;
1973 			goto send_ack;
1974 		}
1975 
1976 		/*
1977 		 * Resend the RDMA read or atomic op which
1978 		 * ACKs this duplicate request.
1979 		 */
1980 		qp->s_tail_ack_queue = i;
1981 		break;
1982 	}
1983 	qp->s_ack_state = OP(ACKNOWLEDGE);
1984 	qp->s_flags |= RVT_S_RESP_PENDING;
1985 	qp->r_nak_state = 0;
1986 	hfi1_schedule_send(qp);
1987 
1988 unlock_done:
1989 	spin_unlock_irqrestore(&qp->s_lock, flags);
1990 done:
1991 	return 1;
1992 
1993 send_ack:
1994 	return 0;
1995 }
1996 
1997 void hfi1_rc_error(struct rvt_qp *qp, enum ib_wc_status err)
1998 {
1999 	unsigned long flags;
2000 	int lastwqe;
2001 
2002 	spin_lock_irqsave(&qp->s_lock, flags);
2003 	lastwqe = rvt_error_qp(qp, err);
2004 	spin_unlock_irqrestore(&qp->s_lock, flags);
2005 
2006 	if (lastwqe) {
2007 		struct ib_event ev;
2008 
2009 		ev.device = qp->ibqp.device;
2010 		ev.element.qp = &qp->ibqp;
2011 		ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
2012 		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
2013 	}
2014 }
2015 
2016 static inline void update_ack_queue(struct rvt_qp *qp, unsigned n)
2017 {
2018 	unsigned next;
2019 
2020 	next = n + 1;
2021 	if (next > HFI1_MAX_RDMA_ATOMIC)
2022 		next = 0;
2023 	qp->s_tail_ack_queue = next;
2024 	qp->s_ack_state = OP(ACKNOWLEDGE);
2025 }
2026 
2027 static void log_cca_event(struct hfi1_pportdata *ppd, u8 sl, u32 rlid,
2028 			  u32 lqpn, u32 rqpn, u8 svc_type)
2029 {
2030 	struct opa_hfi1_cong_log_event_internal *cc_event;
2031 	unsigned long flags;
2032 
2033 	if (sl >= OPA_MAX_SLS)
2034 		return;
2035 
2036 	spin_lock_irqsave(&ppd->cc_log_lock, flags);
2037 
2038 	ppd->threshold_cong_event_map[sl / 8] |= 1 << (sl % 8);
2039 	ppd->threshold_event_counter++;
2040 
2041 	cc_event = &ppd->cc_events[ppd->cc_log_idx++];
2042 	if (ppd->cc_log_idx == OPA_CONG_LOG_ELEMS)
2043 		ppd->cc_log_idx = 0;
2044 	cc_event->lqpn = lqpn & RVT_QPN_MASK;
2045 	cc_event->rqpn = rqpn & RVT_QPN_MASK;
2046 	cc_event->sl = sl;
2047 	cc_event->svc_type = svc_type;
2048 	cc_event->rlid = rlid;
2049 	/* keep timestamp in units of 1.024 usec */
2050 	cc_event->timestamp = ktime_to_ns(ktime_get()) / 1024;
2051 
2052 	spin_unlock_irqrestore(&ppd->cc_log_lock, flags);
2053 }
2054 
2055 void process_becn(struct hfi1_pportdata *ppd, u8 sl, u16 rlid, u32 lqpn,
2056 		  u32 rqpn, u8 svc_type)
2057 {
2058 	struct cca_timer *cca_timer;
2059 	u16 ccti, ccti_incr, ccti_timer, ccti_limit;
2060 	u8 trigger_threshold;
2061 	struct cc_state *cc_state;
2062 	unsigned long flags;
2063 
2064 	if (sl >= OPA_MAX_SLS)
2065 		return;
2066 
2067 	cc_state = get_cc_state(ppd);
2068 
2069 	if (!cc_state)
2070 		return;
2071 
2072 	/*
2073 	 * 1) increase CCTI (for this SL)
2074 	 * 2) select IPG (i.e., call set_link_ipg())
2075 	 * 3) start timer
2076 	 */
2077 	ccti_limit = cc_state->cct.ccti_limit;
2078 	ccti_incr = cc_state->cong_setting.entries[sl].ccti_increase;
2079 	ccti_timer = cc_state->cong_setting.entries[sl].ccti_timer;
2080 	trigger_threshold =
2081 		cc_state->cong_setting.entries[sl].trigger_threshold;
2082 
2083 	spin_lock_irqsave(&ppd->cca_timer_lock, flags);
2084 
2085 	cca_timer = &ppd->cca_timer[sl];
2086 	if (cca_timer->ccti < ccti_limit) {
2087 		if (cca_timer->ccti + ccti_incr <= ccti_limit)
2088 			cca_timer->ccti += ccti_incr;
2089 		else
2090 			cca_timer->ccti = ccti_limit;
2091 		set_link_ipg(ppd);
2092 	}
2093 
2094 	ccti = cca_timer->ccti;
2095 
2096 	if (!hrtimer_active(&cca_timer->hrtimer)) {
2097 		/* ccti_timer is in units of 1.024 usec */
2098 		unsigned long nsec = 1024 * ccti_timer;
2099 
2100 		hrtimer_start(&cca_timer->hrtimer, ns_to_ktime(nsec),
2101 			      HRTIMER_MODE_REL);
2102 	}
2103 
2104 	spin_unlock_irqrestore(&ppd->cca_timer_lock, flags);
2105 
2106 	if ((trigger_threshold != 0) && (ccti >= trigger_threshold))
2107 		log_cca_event(ppd, sl, rlid, lqpn, rqpn, svc_type);
2108 }
2109 
2110 /**
2111  * hfi1_rc_rcv - process an incoming RC packet
2112  * @rcd: the context pointer
2113  * @hdr: the header of this packet
2114  * @rcv_flags: flags relevant to rcv processing
2115  * @data: the packet data
2116  * @tlen: the packet length
2117  * @qp: the QP for this packet
2118  *
2119  * This is called from qp_rcv() to process an incoming RC packet
2120  * for the given QP.
2121  * May be called at interrupt level.
2122  */
2123 void hfi1_rc_rcv(struct hfi1_packet *packet)
2124 {
2125 	struct hfi1_ctxtdata *rcd = packet->rcd;
2126 	struct hfi1_ib_header *hdr = packet->hdr;
2127 	u32 rcv_flags = packet->rcv_flags;
2128 	void *data = packet->ebuf;
2129 	u32 tlen = packet->tlen;
2130 	struct rvt_qp *qp = packet->qp;
2131 	struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
2132 	struct hfi1_other_headers *ohdr = packet->ohdr;
2133 	u32 bth0, opcode;
2134 	u32 hdrsize = packet->hlen;
2135 	u32 psn;
2136 	u32 pad;
2137 	struct ib_wc wc;
2138 	u32 pmtu = qp->pmtu;
2139 	int diff;
2140 	struct ib_reth *reth;
2141 	unsigned long flags;
2142 	int ret, is_fecn = 0;
2143 	int copy_last = 0;
2144 	u32 rkey;
2145 
2146 	bth0 = be32_to_cpu(ohdr->bth[0]);
2147 	if (hfi1_ruc_check_hdr(ibp, hdr, rcv_flags & HFI1_HAS_GRH, qp, bth0))
2148 		return;
2149 
2150 	is_fecn = process_ecn(qp, packet, false);
2151 
2152 	psn = be32_to_cpu(ohdr->bth[2]);
2153 	opcode = (bth0 >> 24) & 0xff;
2154 
2155 	/*
2156 	 * Process responses (ACKs) before anything else.  Note that the
2157 	 * packet sequence number will be for something in the send work
2158 	 * queue rather than the expected receive packet sequence number.
2159 	 * In other words, this QP is the requester.
2160 	 */
2161 	if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
2162 	    opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
2163 		rc_rcv_resp(ibp, ohdr, data, tlen, qp, opcode, psn,
2164 			    hdrsize, pmtu, rcd);
2165 		if (is_fecn)
2166 			goto send_ack;
2167 		return;
2168 	}
2169 
2170 	/* Compute 24 bits worth of difference. */
2171 	diff = delta_psn(psn, qp->r_psn);
2172 	if (unlikely(diff)) {
2173 		if (rc_rcv_error(ohdr, data, qp, opcode, psn, diff, rcd))
2174 			return;
2175 		goto send_ack;
2176 	}
2177 
2178 	/* Check for opcode sequence errors. */
2179 	switch (qp->r_state) {
2180 	case OP(SEND_FIRST):
2181 	case OP(SEND_MIDDLE):
2182 		if (opcode == OP(SEND_MIDDLE) ||
2183 		    opcode == OP(SEND_LAST) ||
2184 		    opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
2185 		    opcode == OP(SEND_LAST_WITH_INVALIDATE))
2186 			break;
2187 		goto nack_inv;
2188 
2189 	case OP(RDMA_WRITE_FIRST):
2190 	case OP(RDMA_WRITE_MIDDLE):
2191 		if (opcode == OP(RDMA_WRITE_MIDDLE) ||
2192 		    opcode == OP(RDMA_WRITE_LAST) ||
2193 		    opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
2194 			break;
2195 		goto nack_inv;
2196 
2197 	default:
2198 		if (opcode == OP(SEND_MIDDLE) ||
2199 		    opcode == OP(SEND_LAST) ||
2200 		    opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
2201 		    opcode == OP(SEND_LAST_WITH_INVALIDATE) ||
2202 		    opcode == OP(RDMA_WRITE_MIDDLE) ||
2203 		    opcode == OP(RDMA_WRITE_LAST) ||
2204 		    opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
2205 			goto nack_inv;
2206 		/*
2207 		 * Note that it is up to the requester to not send a new
2208 		 * RDMA read or atomic operation before receiving an ACK
2209 		 * for the previous operation.
2210 		 */
2211 		break;
2212 	}
2213 
2214 	if (qp->state == IB_QPS_RTR && !(qp->r_flags & RVT_R_COMM_EST))
2215 		qp_comm_est(qp);
2216 
2217 	/* OK, process the packet. */
2218 	switch (opcode) {
2219 	case OP(SEND_FIRST):
2220 		ret = hfi1_rvt_get_rwqe(qp, 0);
2221 		if (ret < 0)
2222 			goto nack_op_err;
2223 		if (!ret)
2224 			goto rnr_nak;
2225 		qp->r_rcv_len = 0;
2226 		/* FALLTHROUGH */
2227 	case OP(SEND_MIDDLE):
2228 	case OP(RDMA_WRITE_MIDDLE):
2229 send_middle:
2230 		/* Check for invalid length PMTU or posted rwqe len. */
2231 		if (unlikely(tlen != (hdrsize + pmtu + 4)))
2232 			goto nack_inv;
2233 		qp->r_rcv_len += pmtu;
2234 		if (unlikely(qp->r_rcv_len > qp->r_len))
2235 			goto nack_inv;
2236 		hfi1_copy_sge(&qp->r_sge, data, pmtu, 1, 0);
2237 		break;
2238 
2239 	case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
2240 		/* consume RWQE */
2241 		ret = hfi1_rvt_get_rwqe(qp, 1);
2242 		if (ret < 0)
2243 			goto nack_op_err;
2244 		if (!ret)
2245 			goto rnr_nak;
2246 		goto send_last_imm;
2247 
2248 	case OP(SEND_ONLY):
2249 	case OP(SEND_ONLY_WITH_IMMEDIATE):
2250 	case OP(SEND_ONLY_WITH_INVALIDATE):
2251 		ret = hfi1_rvt_get_rwqe(qp, 0);
2252 		if (ret < 0)
2253 			goto nack_op_err;
2254 		if (!ret)
2255 			goto rnr_nak;
2256 		qp->r_rcv_len = 0;
2257 		if (opcode == OP(SEND_ONLY))
2258 			goto no_immediate_data;
2259 		if (opcode == OP(SEND_ONLY_WITH_INVALIDATE))
2260 			goto send_last_inv;
2261 		/* FALLTHROUGH for SEND_ONLY_WITH_IMMEDIATE */
2262 	case OP(SEND_LAST_WITH_IMMEDIATE):
2263 send_last_imm:
2264 		wc.ex.imm_data = ohdr->u.imm_data;
2265 		wc.wc_flags = IB_WC_WITH_IMM;
2266 		goto send_last;
2267 	case OP(SEND_LAST_WITH_INVALIDATE):
2268 send_last_inv:
2269 		rkey = be32_to_cpu(ohdr->u.ieth);
2270 		if (rvt_invalidate_rkey(qp, rkey))
2271 			goto no_immediate_data;
2272 		wc.ex.invalidate_rkey = rkey;
2273 		wc.wc_flags = IB_WC_WITH_INVALIDATE;
2274 		goto send_last;
2275 	case OP(RDMA_WRITE_LAST):
2276 		copy_last = ibpd_to_rvtpd(qp->ibqp.pd)->user;
2277 		/* fall through */
2278 	case OP(SEND_LAST):
2279 no_immediate_data:
2280 		wc.wc_flags = 0;
2281 		wc.ex.imm_data = 0;
2282 send_last:
2283 		/* Get the number of bytes the message was padded by. */
2284 		pad = (bth0 >> 20) & 3;
2285 		/* Check for invalid length. */
2286 		/* LAST len should be >= 1 */
2287 		if (unlikely(tlen < (hdrsize + pad + 4)))
2288 			goto nack_inv;
2289 		/* Don't count the CRC. */
2290 		tlen -= (hdrsize + pad + 4);
2291 		wc.byte_len = tlen + qp->r_rcv_len;
2292 		if (unlikely(wc.byte_len > qp->r_len))
2293 			goto nack_inv;
2294 		hfi1_copy_sge(&qp->r_sge, data, tlen, 1, copy_last);
2295 		rvt_put_ss(&qp->r_sge);
2296 		qp->r_msn++;
2297 		if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
2298 			break;
2299 		wc.wr_id = qp->r_wr_id;
2300 		wc.status = IB_WC_SUCCESS;
2301 		if (opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE) ||
2302 		    opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE))
2303 			wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
2304 		else
2305 			wc.opcode = IB_WC_RECV;
2306 		wc.qp = &qp->ibqp;
2307 		wc.src_qp = qp->remote_qpn;
2308 		wc.slid = qp->remote_ah_attr.dlid;
2309 		/*
2310 		 * It seems that IB mandates the presence of an SL in a
2311 		 * work completion only for the UD transport (see section
2312 		 * 11.4.2 of IBTA Vol. 1).
2313 		 *
2314 		 * However, the way the SL is chosen below is consistent
2315 		 * with the way that IB/qib works and is trying avoid
2316 		 * introducing incompatibilities.
2317 		 *
2318 		 * See also OPA Vol. 1, section 9.7.6, and table 9-17.
2319 		 */
2320 		wc.sl = qp->remote_ah_attr.sl;
2321 		/* zero fields that are N/A */
2322 		wc.vendor_err = 0;
2323 		wc.pkey_index = 0;
2324 		wc.dlid_path_bits = 0;
2325 		wc.port_num = 0;
2326 		/* Signal completion event if the solicited bit is set. */
2327 		rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc,
2328 			     (bth0 & IB_BTH_SOLICITED) != 0);
2329 		break;
2330 
2331 	case OP(RDMA_WRITE_ONLY):
2332 		copy_last = 1;
2333 		/* fall through */
2334 	case OP(RDMA_WRITE_FIRST):
2335 	case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
2336 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
2337 			goto nack_inv;
2338 		/* consume RWQE */
2339 		reth = &ohdr->u.rc.reth;
2340 		qp->r_len = be32_to_cpu(reth->length);
2341 		qp->r_rcv_len = 0;
2342 		qp->r_sge.sg_list = NULL;
2343 		if (qp->r_len != 0) {
2344 			u32 rkey = be32_to_cpu(reth->rkey);
2345 			u64 vaddr = be64_to_cpu(reth->vaddr);
2346 			int ok;
2347 
2348 			/* Check rkey & NAK */
2349 			ok = rvt_rkey_ok(qp, &qp->r_sge.sge, qp->r_len, vaddr,
2350 					 rkey, IB_ACCESS_REMOTE_WRITE);
2351 			if (unlikely(!ok))
2352 				goto nack_acc;
2353 			qp->r_sge.num_sge = 1;
2354 		} else {
2355 			qp->r_sge.num_sge = 0;
2356 			qp->r_sge.sge.mr = NULL;
2357 			qp->r_sge.sge.vaddr = NULL;
2358 			qp->r_sge.sge.length = 0;
2359 			qp->r_sge.sge.sge_length = 0;
2360 		}
2361 		if (opcode == OP(RDMA_WRITE_FIRST))
2362 			goto send_middle;
2363 		else if (opcode == OP(RDMA_WRITE_ONLY))
2364 			goto no_immediate_data;
2365 		ret = hfi1_rvt_get_rwqe(qp, 1);
2366 		if (ret < 0)
2367 			goto nack_op_err;
2368 		if (!ret)
2369 			goto rnr_nak;
2370 		wc.ex.imm_data = ohdr->u.rc.imm_data;
2371 		wc.wc_flags = IB_WC_WITH_IMM;
2372 		goto send_last;
2373 
2374 	case OP(RDMA_READ_REQUEST): {
2375 		struct rvt_ack_entry *e;
2376 		u32 len;
2377 		u8 next;
2378 
2379 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
2380 			goto nack_inv;
2381 		next = qp->r_head_ack_queue + 1;
2382 		/* s_ack_queue is size HFI1_MAX_RDMA_ATOMIC+1 so use > not >= */
2383 		if (next > HFI1_MAX_RDMA_ATOMIC)
2384 			next = 0;
2385 		spin_lock_irqsave(&qp->s_lock, flags);
2386 		if (unlikely(next == qp->s_tail_ack_queue)) {
2387 			if (!qp->s_ack_queue[next].sent)
2388 				goto nack_inv_unlck;
2389 			update_ack_queue(qp, next);
2390 		}
2391 		e = &qp->s_ack_queue[qp->r_head_ack_queue];
2392 		if (e->opcode == OP(RDMA_READ_REQUEST) && e->rdma_sge.mr) {
2393 			rvt_put_mr(e->rdma_sge.mr);
2394 			e->rdma_sge.mr = NULL;
2395 		}
2396 		reth = &ohdr->u.rc.reth;
2397 		len = be32_to_cpu(reth->length);
2398 		if (len) {
2399 			u32 rkey = be32_to_cpu(reth->rkey);
2400 			u64 vaddr = be64_to_cpu(reth->vaddr);
2401 			int ok;
2402 
2403 			/* Check rkey & NAK */
2404 			ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr,
2405 					 rkey, IB_ACCESS_REMOTE_READ);
2406 			if (unlikely(!ok))
2407 				goto nack_acc_unlck;
2408 			/*
2409 			 * Update the next expected PSN.  We add 1 later
2410 			 * below, so only add the remainder here.
2411 			 */
2412 			if (len > pmtu)
2413 				qp->r_psn += (len - 1) / pmtu;
2414 		} else {
2415 			e->rdma_sge.mr = NULL;
2416 			e->rdma_sge.vaddr = NULL;
2417 			e->rdma_sge.length = 0;
2418 			e->rdma_sge.sge_length = 0;
2419 		}
2420 		e->opcode = opcode;
2421 		e->sent = 0;
2422 		e->psn = psn;
2423 		e->lpsn = qp->r_psn;
2424 		/*
2425 		 * We need to increment the MSN here instead of when we
2426 		 * finish sending the result since a duplicate request would
2427 		 * increment it more than once.
2428 		 */
2429 		qp->r_msn++;
2430 		qp->r_psn++;
2431 		qp->r_state = opcode;
2432 		qp->r_nak_state = 0;
2433 		qp->r_head_ack_queue = next;
2434 
2435 		/* Schedule the send tasklet. */
2436 		qp->s_flags |= RVT_S_RESP_PENDING;
2437 		hfi1_schedule_send(qp);
2438 
2439 		spin_unlock_irqrestore(&qp->s_lock, flags);
2440 		if (is_fecn)
2441 			goto send_ack;
2442 		return;
2443 	}
2444 
2445 	case OP(COMPARE_SWAP):
2446 	case OP(FETCH_ADD): {
2447 		struct ib_atomic_eth *ateth;
2448 		struct rvt_ack_entry *e;
2449 		u64 vaddr;
2450 		atomic64_t *maddr;
2451 		u64 sdata;
2452 		u32 rkey;
2453 		u8 next;
2454 
2455 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
2456 			goto nack_inv;
2457 		next = qp->r_head_ack_queue + 1;
2458 		if (next > HFI1_MAX_RDMA_ATOMIC)
2459 			next = 0;
2460 		spin_lock_irqsave(&qp->s_lock, flags);
2461 		if (unlikely(next == qp->s_tail_ack_queue)) {
2462 			if (!qp->s_ack_queue[next].sent)
2463 				goto nack_inv_unlck;
2464 			update_ack_queue(qp, next);
2465 		}
2466 		e = &qp->s_ack_queue[qp->r_head_ack_queue];
2467 		if (e->opcode == OP(RDMA_READ_REQUEST) && e->rdma_sge.mr) {
2468 			rvt_put_mr(e->rdma_sge.mr);
2469 			e->rdma_sge.mr = NULL;
2470 		}
2471 		ateth = &ohdr->u.atomic_eth;
2472 		vaddr = ((u64)be32_to_cpu(ateth->vaddr[0]) << 32) |
2473 			be32_to_cpu(ateth->vaddr[1]);
2474 		if (unlikely(vaddr & (sizeof(u64) - 1)))
2475 			goto nack_inv_unlck;
2476 		rkey = be32_to_cpu(ateth->rkey);
2477 		/* Check rkey & NAK */
2478 		if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64),
2479 					  vaddr, rkey,
2480 					  IB_ACCESS_REMOTE_ATOMIC)))
2481 			goto nack_acc_unlck;
2482 		/* Perform atomic OP and save result. */
2483 		maddr = (atomic64_t *)qp->r_sge.sge.vaddr;
2484 		sdata = be64_to_cpu(ateth->swap_data);
2485 		e->atomic_data = (opcode == OP(FETCH_ADD)) ?
2486 			(u64)atomic64_add_return(sdata, maddr) - sdata :
2487 			(u64)cmpxchg((u64 *)qp->r_sge.sge.vaddr,
2488 				      be64_to_cpu(ateth->compare_data),
2489 				      sdata);
2490 		rvt_put_mr(qp->r_sge.sge.mr);
2491 		qp->r_sge.num_sge = 0;
2492 		e->opcode = opcode;
2493 		e->sent = 0;
2494 		e->psn = psn;
2495 		e->lpsn = psn;
2496 		qp->r_msn++;
2497 		qp->r_psn++;
2498 		qp->r_state = opcode;
2499 		qp->r_nak_state = 0;
2500 		qp->r_head_ack_queue = next;
2501 
2502 		/* Schedule the send tasklet. */
2503 		qp->s_flags |= RVT_S_RESP_PENDING;
2504 		hfi1_schedule_send(qp);
2505 
2506 		spin_unlock_irqrestore(&qp->s_lock, flags);
2507 		if (is_fecn)
2508 			goto send_ack;
2509 		return;
2510 	}
2511 
2512 	default:
2513 		/* NAK unknown opcodes. */
2514 		goto nack_inv;
2515 	}
2516 	qp->r_psn++;
2517 	qp->r_state = opcode;
2518 	qp->r_ack_psn = psn;
2519 	qp->r_nak_state = 0;
2520 	/* Send an ACK if requested or required. */
2521 	if (psn & IB_BTH_REQ_ACK) {
2522 		struct hfi1_qp_priv *priv = qp->priv;
2523 
2524 		if (packet->numpkt == 0) {
2525 			rc_cancel_ack(qp);
2526 			goto send_ack;
2527 		}
2528 		if (priv->r_adefered >= HFI1_PSN_CREDIT) {
2529 			rc_cancel_ack(qp);
2530 			goto send_ack;
2531 		}
2532 		if (unlikely(is_fecn)) {
2533 			rc_cancel_ack(qp);
2534 			goto send_ack;
2535 		}
2536 		priv->r_adefered++;
2537 		rc_defered_ack(rcd, qp);
2538 	}
2539 	return;
2540 
2541 rnr_nak:
2542 	qp->r_nak_state = qp->r_min_rnr_timer | IB_RNR_NAK;
2543 	qp->r_ack_psn = qp->r_psn;
2544 	/* Queue RNR NAK for later */
2545 	rc_defered_ack(rcd, qp);
2546 	return;
2547 
2548 nack_op_err:
2549 	hfi1_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
2550 	qp->r_nak_state = IB_NAK_REMOTE_OPERATIONAL_ERROR;
2551 	qp->r_ack_psn = qp->r_psn;
2552 	/* Queue NAK for later */
2553 	rc_defered_ack(rcd, qp);
2554 	return;
2555 
2556 nack_inv_unlck:
2557 	spin_unlock_irqrestore(&qp->s_lock, flags);
2558 nack_inv:
2559 	hfi1_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
2560 	qp->r_nak_state = IB_NAK_INVALID_REQUEST;
2561 	qp->r_ack_psn = qp->r_psn;
2562 	/* Queue NAK for later */
2563 	rc_defered_ack(rcd, qp);
2564 	return;
2565 
2566 nack_acc_unlck:
2567 	spin_unlock_irqrestore(&qp->s_lock, flags);
2568 nack_acc:
2569 	hfi1_rc_error(qp, IB_WC_LOC_PROT_ERR);
2570 	qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
2571 	qp->r_ack_psn = qp->r_psn;
2572 send_ack:
2573 	hfi1_send_rc_ack(rcd, qp, is_fecn);
2574 }
2575 
2576 void hfi1_rc_hdrerr(
2577 	struct hfi1_ctxtdata *rcd,
2578 	struct hfi1_ib_header *hdr,
2579 	u32 rcv_flags,
2580 	struct rvt_qp *qp)
2581 {
2582 	int has_grh = rcv_flags & HFI1_HAS_GRH;
2583 	struct hfi1_other_headers *ohdr;
2584 	struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
2585 	int diff;
2586 	u32 opcode;
2587 	u32 psn, bth0;
2588 
2589 	/* Check for GRH */
2590 	ohdr = &hdr->u.oth;
2591 	if (has_grh)
2592 		ohdr = &hdr->u.l.oth;
2593 
2594 	bth0 = be32_to_cpu(ohdr->bth[0]);
2595 	if (hfi1_ruc_check_hdr(ibp, hdr, has_grh, qp, bth0))
2596 		return;
2597 
2598 	psn = be32_to_cpu(ohdr->bth[2]);
2599 	opcode = (bth0 >> 24) & 0xff;
2600 
2601 	/* Only deal with RDMA Writes for now */
2602 	if (opcode < IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST) {
2603 		diff = delta_psn(psn, qp->r_psn);
2604 		if (!qp->r_nak_state && diff >= 0) {
2605 			ibp->rvp.n_rc_seqnak++;
2606 			qp->r_nak_state = IB_NAK_PSN_ERROR;
2607 			/* Use the expected PSN. */
2608 			qp->r_ack_psn = qp->r_psn;
2609 			/*
2610 			 * Wait to send the sequence
2611 			 * NAK until all packets
2612 			 * in the receive queue have
2613 			 * been processed.
2614 			 * Otherwise, we end up
2615 			 * propagating congestion.
2616 			 */
2617 			rc_defered_ack(rcd, qp);
2618 		} /* Out of sequence NAK */
2619 	} /* QP Request NAKs */
2620 }
2621