xref: /openbmc/linux/drivers/infiniband/hw/qib/qib_rc.c (revision b737eecd)
1 /*
2  * Copyright (c) 2006, 2007, 2008, 2009 QLogic Corporation. All rights reserved.
3  * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33 
34 #include <linux/io.h>
35 
36 #include "qib.h"
37 
38 /* cut down ridiculously long IB macro names */
39 #define OP(x) IB_OPCODE_RC_##x
40 
41 
42 static u32 restart_sge(struct rvt_sge_state *ss, struct rvt_swqe *wqe,
43 		       u32 psn, u32 pmtu)
44 {
45 	u32 len;
46 
47 	len = ((psn - wqe->psn) & QIB_PSN_MASK) * pmtu;
48 	return rvt_restart_sge(ss, wqe, len);
49 }
50 
51 /**
52  * qib_make_rc_ack - construct a response packet (ACK, NAK, or RDMA read)
53  * @dev: the device for this QP
54  * @qp: a pointer to the QP
55  * @ohdr: a pointer to the IB header being constructed
56  * @pmtu: the path MTU
57  *
58  * Return 1 if constructed; otherwise, return 0.
59  * Note that we are in the responder's side of the QP context.
60  * Note the QP s_lock must be held.
61  */
62 static int qib_make_rc_ack(struct qib_ibdev *dev, struct rvt_qp *qp,
63 			   struct ib_other_headers *ohdr, u32 pmtu)
64 {
65 	struct rvt_ack_entry *e;
66 	u32 hwords;
67 	u32 len;
68 	u32 bth0;
69 	u32 bth2;
70 
71 	/* Don't send an ACK if we aren't supposed to. */
72 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
73 		goto bail;
74 
75 	/* header size in 32-bit words LRH+BTH = (8+12)/4. */
76 	hwords = 5;
77 
78 	switch (qp->s_ack_state) {
79 	case OP(RDMA_READ_RESPONSE_LAST):
80 	case OP(RDMA_READ_RESPONSE_ONLY):
81 		e = &qp->s_ack_queue[qp->s_tail_ack_queue];
82 		if (e->rdma_sge.mr) {
83 			rvt_put_mr(e->rdma_sge.mr);
84 			e->rdma_sge.mr = NULL;
85 		}
86 		fallthrough;
87 	case OP(ATOMIC_ACKNOWLEDGE):
88 		/*
89 		 * We can increment the tail pointer now that the last
90 		 * response has been sent instead of only being
91 		 * constructed.
92 		 */
93 		if (++qp->s_tail_ack_queue > QIB_MAX_RDMA_ATOMIC)
94 			qp->s_tail_ack_queue = 0;
95 		fallthrough;
96 	case OP(SEND_ONLY):
97 	case OP(ACKNOWLEDGE):
98 		/* Check for no next entry in the queue. */
99 		if (qp->r_head_ack_queue == qp->s_tail_ack_queue) {
100 			if (qp->s_flags & RVT_S_ACK_PENDING)
101 				goto normal;
102 			goto bail;
103 		}
104 
105 		e = &qp->s_ack_queue[qp->s_tail_ack_queue];
106 		if (e->opcode == OP(RDMA_READ_REQUEST)) {
107 			/*
108 			 * If a RDMA read response is being resent and
109 			 * we haven't seen the duplicate request yet,
110 			 * then stop sending the remaining responses the
111 			 * responder has seen until the requester resends it.
112 			 */
113 			len = e->rdma_sge.sge_length;
114 			if (len && !e->rdma_sge.mr) {
115 				qp->s_tail_ack_queue = qp->r_head_ack_queue;
116 				goto bail;
117 			}
118 			/* Copy SGE state in case we need to resend */
119 			qp->s_rdma_mr = e->rdma_sge.mr;
120 			if (qp->s_rdma_mr)
121 				rvt_get_mr(qp->s_rdma_mr);
122 			qp->s_ack_rdma_sge.sge = e->rdma_sge;
123 			qp->s_ack_rdma_sge.num_sge = 1;
124 			qp->s_cur_sge = &qp->s_ack_rdma_sge;
125 			if (len > pmtu) {
126 				len = pmtu;
127 				qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST);
128 			} else {
129 				qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY);
130 				e->sent = 1;
131 			}
132 			ohdr->u.aeth = rvt_compute_aeth(qp);
133 			hwords++;
134 			qp->s_ack_rdma_psn = e->psn;
135 			bth2 = qp->s_ack_rdma_psn++ & QIB_PSN_MASK;
136 		} else {
137 			/* COMPARE_SWAP or FETCH_ADD */
138 			qp->s_cur_sge = NULL;
139 			len = 0;
140 			qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE);
141 			ohdr->u.at.aeth = rvt_compute_aeth(qp);
142 			ib_u64_put(e->atomic_data, &ohdr->u.at.atomic_ack_eth);
143 			hwords += sizeof(ohdr->u.at) / sizeof(u32);
144 			bth2 = e->psn & QIB_PSN_MASK;
145 			e->sent = 1;
146 		}
147 		bth0 = qp->s_ack_state << 24;
148 		break;
149 
150 	case OP(RDMA_READ_RESPONSE_FIRST):
151 		qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE);
152 		fallthrough;
153 	case OP(RDMA_READ_RESPONSE_MIDDLE):
154 		qp->s_cur_sge = &qp->s_ack_rdma_sge;
155 		qp->s_rdma_mr = qp->s_ack_rdma_sge.sge.mr;
156 		if (qp->s_rdma_mr)
157 			rvt_get_mr(qp->s_rdma_mr);
158 		len = qp->s_ack_rdma_sge.sge.sge_length;
159 		if (len > pmtu)
160 			len = pmtu;
161 		else {
162 			ohdr->u.aeth = rvt_compute_aeth(qp);
163 			hwords++;
164 			qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
165 			e = &qp->s_ack_queue[qp->s_tail_ack_queue];
166 			e->sent = 1;
167 		}
168 		bth0 = qp->s_ack_state << 24;
169 		bth2 = qp->s_ack_rdma_psn++ & QIB_PSN_MASK;
170 		break;
171 
172 	default:
173 normal:
174 		/*
175 		 * Send a regular ACK.
176 		 * Set the s_ack_state so we wait until after sending
177 		 * the ACK before setting s_ack_state to ACKNOWLEDGE
178 		 * (see above).
179 		 */
180 		qp->s_ack_state = OP(SEND_ONLY);
181 		qp->s_flags &= ~RVT_S_ACK_PENDING;
182 		qp->s_cur_sge = NULL;
183 		if (qp->s_nak_state)
184 			ohdr->u.aeth =
185 				cpu_to_be32((qp->r_msn & IB_MSN_MASK) |
186 					    (qp->s_nak_state <<
187 					     IB_AETH_CREDIT_SHIFT));
188 		else
189 			ohdr->u.aeth = rvt_compute_aeth(qp);
190 		hwords++;
191 		len = 0;
192 		bth0 = OP(ACKNOWLEDGE) << 24;
193 		bth2 = qp->s_ack_psn & QIB_PSN_MASK;
194 	}
195 	qp->s_rdma_ack_cnt++;
196 	qp->s_hdrwords = hwords;
197 	qp->s_cur_size = len;
198 	qib_make_ruc_header(qp, ohdr, bth0, bth2);
199 	return 1;
200 
201 bail:
202 	qp->s_ack_state = OP(ACKNOWLEDGE);
203 	qp->s_flags &= ~(RVT_S_RESP_PENDING | RVT_S_ACK_PENDING);
204 	return 0;
205 }
206 
207 /**
208  * qib_make_rc_req - construct a request packet (SEND, RDMA r/w, ATOMIC)
209  * @qp: a pointer to the QP
210  * @flags: unused
211  *
212  * Assumes the s_lock is held.
213  *
214  * Return 1 if constructed; otherwise, return 0.
215  */
216 int qib_make_rc_req(struct rvt_qp *qp, unsigned long *flags)
217 {
218 	struct qib_qp_priv *priv = qp->priv;
219 	struct qib_ibdev *dev = to_idev(qp->ibqp.device);
220 	struct ib_other_headers *ohdr;
221 	struct rvt_sge_state *ss;
222 	struct rvt_swqe *wqe;
223 	u32 hwords;
224 	u32 len;
225 	u32 bth0;
226 	u32 bth2;
227 	u32 pmtu = qp->pmtu;
228 	char newreq;
229 	int ret = 0;
230 	int delta;
231 
232 	ohdr = &priv->s_hdr->u.oth;
233 	if (rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH)
234 		ohdr = &priv->s_hdr->u.l.oth;
235 
236 	/* Sending responses has higher priority over sending requests. */
237 	if ((qp->s_flags & RVT_S_RESP_PENDING) &&
238 	    qib_make_rc_ack(dev, qp, ohdr, pmtu))
239 		goto done;
240 
241 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_SEND_OK)) {
242 		if (!(ib_rvt_state_ops[qp->state] & RVT_FLUSH_SEND))
243 			goto bail;
244 		/* We are in the error state, flush the work request. */
245 		if (qp->s_last == READ_ONCE(qp->s_head))
246 			goto bail;
247 		/* If DMAs are in progress, we can't flush immediately. */
248 		if (atomic_read(&priv->s_dma_busy)) {
249 			qp->s_flags |= RVT_S_WAIT_DMA;
250 			goto bail;
251 		}
252 		wqe = rvt_get_swqe_ptr(qp, qp->s_last);
253 		rvt_send_complete(qp, wqe, qp->s_last != qp->s_acked ?
254 			IB_WC_SUCCESS : IB_WC_WR_FLUSH_ERR);
255 		/* will get called again */
256 		goto done;
257 	}
258 
259 	if (qp->s_flags & (RVT_S_WAIT_RNR | RVT_S_WAIT_ACK))
260 		goto bail;
261 
262 	if (qib_cmp24(qp->s_psn, qp->s_sending_hpsn) <= 0) {
263 		if (qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) <= 0) {
264 			qp->s_flags |= RVT_S_WAIT_PSN;
265 			goto bail;
266 		}
267 		qp->s_sending_psn = qp->s_psn;
268 		qp->s_sending_hpsn = qp->s_psn - 1;
269 	}
270 
271 	/* header size in 32-bit words LRH+BTH = (8+12)/4. */
272 	hwords = 5;
273 	bth0 = 0;
274 
275 	/* Send a request. */
276 	wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
277 	switch (qp->s_state) {
278 	default:
279 		if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_NEXT_SEND_OK))
280 			goto bail;
281 		/*
282 		 * Resend an old request or start a new one.
283 		 *
284 		 * We keep track of the current SWQE so that
285 		 * we don't reset the "furthest progress" state
286 		 * if we need to back up.
287 		 */
288 		newreq = 0;
289 		if (qp->s_cur == qp->s_tail) {
290 			/* Check if send work queue is empty. */
291 			if (qp->s_tail == READ_ONCE(qp->s_head))
292 				goto bail;
293 			/*
294 			 * If a fence is requested, wait for previous
295 			 * RDMA read and atomic operations to finish.
296 			 */
297 			if ((wqe->wr.send_flags & IB_SEND_FENCE) &&
298 			    qp->s_num_rd_atomic) {
299 				qp->s_flags |= RVT_S_WAIT_FENCE;
300 				goto bail;
301 			}
302 			newreq = 1;
303 			qp->s_psn = wqe->psn;
304 		}
305 		/*
306 		 * Note that we have to be careful not to modify the
307 		 * original work request since we may need to resend
308 		 * it.
309 		 */
310 		len = wqe->length;
311 		ss = &qp->s_sge;
312 		bth2 = qp->s_psn & QIB_PSN_MASK;
313 		switch (wqe->wr.opcode) {
314 		case IB_WR_SEND:
315 		case IB_WR_SEND_WITH_IMM:
316 			/* If no credit, return. */
317 			if (!rvt_rc_credit_avail(qp, wqe))
318 				goto bail;
319 			if (len > pmtu) {
320 				qp->s_state = OP(SEND_FIRST);
321 				len = pmtu;
322 				break;
323 			}
324 			if (wqe->wr.opcode == IB_WR_SEND)
325 				qp->s_state = OP(SEND_ONLY);
326 			else {
327 				qp->s_state = OP(SEND_ONLY_WITH_IMMEDIATE);
328 				/* Immediate data comes after the BTH */
329 				ohdr->u.imm_data = wqe->wr.ex.imm_data;
330 				hwords += 1;
331 			}
332 			if (wqe->wr.send_flags & IB_SEND_SOLICITED)
333 				bth0 |= IB_BTH_SOLICITED;
334 			bth2 |= IB_BTH_REQ_ACK;
335 			if (++qp->s_cur == qp->s_size)
336 				qp->s_cur = 0;
337 			break;
338 
339 		case IB_WR_RDMA_WRITE:
340 			if (newreq && !(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
341 				qp->s_lsn++;
342 			goto no_flow_control;
343 		case IB_WR_RDMA_WRITE_WITH_IMM:
344 			/* If no credit, return. */
345 			if (!rvt_rc_credit_avail(qp, wqe))
346 				goto bail;
347 no_flow_control:
348 			ohdr->u.rc.reth.vaddr =
349 				cpu_to_be64(wqe->rdma_wr.remote_addr);
350 			ohdr->u.rc.reth.rkey =
351 				cpu_to_be32(wqe->rdma_wr.rkey);
352 			ohdr->u.rc.reth.length = cpu_to_be32(len);
353 			hwords += sizeof(struct ib_reth) / sizeof(u32);
354 			if (len > pmtu) {
355 				qp->s_state = OP(RDMA_WRITE_FIRST);
356 				len = pmtu;
357 				break;
358 			}
359 			if (wqe->rdma_wr.wr.opcode == IB_WR_RDMA_WRITE)
360 				qp->s_state = OP(RDMA_WRITE_ONLY);
361 			else {
362 				qp->s_state = OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
363 				/* Immediate data comes after RETH */
364 				ohdr->u.rc.imm_data =
365 					wqe->rdma_wr.wr.ex.imm_data;
366 				hwords += 1;
367 				if (wqe->rdma_wr.wr.send_flags & IB_SEND_SOLICITED)
368 					bth0 |= IB_BTH_SOLICITED;
369 			}
370 			bth2 |= IB_BTH_REQ_ACK;
371 			if (++qp->s_cur == qp->s_size)
372 				qp->s_cur = 0;
373 			break;
374 
375 		case IB_WR_RDMA_READ:
376 			/*
377 			 * Don't allow more operations to be started
378 			 * than the QP limits allow.
379 			 */
380 			if (newreq) {
381 				if (qp->s_num_rd_atomic >=
382 				    qp->s_max_rd_atomic) {
383 					qp->s_flags |= RVT_S_WAIT_RDMAR;
384 					goto bail;
385 				}
386 				qp->s_num_rd_atomic++;
387 				if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
388 					qp->s_lsn++;
389 			}
390 
391 			ohdr->u.rc.reth.vaddr =
392 				cpu_to_be64(wqe->rdma_wr.remote_addr);
393 			ohdr->u.rc.reth.rkey =
394 				cpu_to_be32(wqe->rdma_wr.rkey);
395 			ohdr->u.rc.reth.length = cpu_to_be32(len);
396 			qp->s_state = OP(RDMA_READ_REQUEST);
397 			hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
398 			ss = NULL;
399 			len = 0;
400 			bth2 |= IB_BTH_REQ_ACK;
401 			if (++qp->s_cur == qp->s_size)
402 				qp->s_cur = 0;
403 			break;
404 
405 		case IB_WR_ATOMIC_CMP_AND_SWP:
406 		case IB_WR_ATOMIC_FETCH_AND_ADD:
407 			/*
408 			 * Don't allow more operations to be started
409 			 * than the QP limits allow.
410 			 */
411 			if (newreq) {
412 				if (qp->s_num_rd_atomic >=
413 				    qp->s_max_rd_atomic) {
414 					qp->s_flags |= RVT_S_WAIT_RDMAR;
415 					goto bail;
416 				}
417 				qp->s_num_rd_atomic++;
418 				if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
419 					qp->s_lsn++;
420 			}
421 			if (wqe->atomic_wr.wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
422 				qp->s_state = OP(COMPARE_SWAP);
423 				put_ib_ateth_swap(wqe->atomic_wr.swap,
424 						  &ohdr->u.atomic_eth);
425 				put_ib_ateth_compare(wqe->atomic_wr.compare_add,
426 						     &ohdr->u.atomic_eth);
427 			} else {
428 				qp->s_state = OP(FETCH_ADD);
429 				put_ib_ateth_swap(wqe->atomic_wr.compare_add,
430 						  &ohdr->u.atomic_eth);
431 				put_ib_ateth_compare(0, &ohdr->u.atomic_eth);
432 			}
433 			put_ib_ateth_vaddr(wqe->atomic_wr.remote_addr,
434 					   &ohdr->u.atomic_eth);
435 			ohdr->u.atomic_eth.rkey = cpu_to_be32(
436 				wqe->atomic_wr.rkey);
437 			hwords += sizeof(struct ib_atomic_eth) / sizeof(u32);
438 			ss = NULL;
439 			len = 0;
440 			bth2 |= IB_BTH_REQ_ACK;
441 			if (++qp->s_cur == qp->s_size)
442 				qp->s_cur = 0;
443 			break;
444 
445 		default:
446 			goto bail;
447 		}
448 		qp->s_sge.sge = wqe->sg_list[0];
449 		qp->s_sge.sg_list = wqe->sg_list + 1;
450 		qp->s_sge.num_sge = wqe->wr.num_sge;
451 		qp->s_sge.total_len = wqe->length;
452 		qp->s_len = wqe->length;
453 		if (newreq) {
454 			qp->s_tail++;
455 			if (qp->s_tail >= qp->s_size)
456 				qp->s_tail = 0;
457 		}
458 		if (wqe->wr.opcode == IB_WR_RDMA_READ)
459 			qp->s_psn = wqe->lpsn + 1;
460 		else
461 			qp->s_psn++;
462 		break;
463 
464 	case OP(RDMA_READ_RESPONSE_FIRST):
465 		/*
466 		 * qp->s_state is normally set to the opcode of the
467 		 * last packet constructed for new requests and therefore
468 		 * is never set to RDMA read response.
469 		 * RDMA_READ_RESPONSE_FIRST is used by the ACK processing
470 		 * thread to indicate a SEND needs to be restarted from an
471 		 * earlier PSN without interferring with the sending thread.
472 		 * See qib_restart_rc().
473 		 */
474 		qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu);
475 		fallthrough;
476 	case OP(SEND_FIRST):
477 		qp->s_state = OP(SEND_MIDDLE);
478 		fallthrough;
479 	case OP(SEND_MIDDLE):
480 		bth2 = qp->s_psn++ & QIB_PSN_MASK;
481 		ss = &qp->s_sge;
482 		len = qp->s_len;
483 		if (len > pmtu) {
484 			len = pmtu;
485 			break;
486 		}
487 		if (wqe->wr.opcode == IB_WR_SEND)
488 			qp->s_state = OP(SEND_LAST);
489 		else {
490 			qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
491 			/* Immediate data comes after the BTH */
492 			ohdr->u.imm_data = wqe->wr.ex.imm_data;
493 			hwords += 1;
494 		}
495 		if (wqe->wr.send_flags & IB_SEND_SOLICITED)
496 			bth0 |= IB_BTH_SOLICITED;
497 		bth2 |= IB_BTH_REQ_ACK;
498 		qp->s_cur++;
499 		if (qp->s_cur >= qp->s_size)
500 			qp->s_cur = 0;
501 		break;
502 
503 	case OP(RDMA_READ_RESPONSE_LAST):
504 		/*
505 		 * qp->s_state is normally set to the opcode of the
506 		 * last packet constructed for new requests and therefore
507 		 * is never set to RDMA read response.
508 		 * RDMA_READ_RESPONSE_LAST is used by the ACK processing
509 		 * thread to indicate a RDMA write needs to be restarted from
510 		 * an earlier PSN without interferring with the sending thread.
511 		 * See qib_restart_rc().
512 		 */
513 		qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu);
514 		fallthrough;
515 	case OP(RDMA_WRITE_FIRST):
516 		qp->s_state = OP(RDMA_WRITE_MIDDLE);
517 		fallthrough;
518 	case OP(RDMA_WRITE_MIDDLE):
519 		bth2 = qp->s_psn++ & QIB_PSN_MASK;
520 		ss = &qp->s_sge;
521 		len = qp->s_len;
522 		if (len > pmtu) {
523 			len = pmtu;
524 			break;
525 		}
526 		if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
527 			qp->s_state = OP(RDMA_WRITE_LAST);
528 		else {
529 			qp->s_state = OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
530 			/* Immediate data comes after the BTH */
531 			ohdr->u.imm_data = wqe->wr.ex.imm_data;
532 			hwords += 1;
533 			if (wqe->wr.send_flags & IB_SEND_SOLICITED)
534 				bth0 |= IB_BTH_SOLICITED;
535 		}
536 		bth2 |= IB_BTH_REQ_ACK;
537 		qp->s_cur++;
538 		if (qp->s_cur >= qp->s_size)
539 			qp->s_cur = 0;
540 		break;
541 
542 	case OP(RDMA_READ_RESPONSE_MIDDLE):
543 		/*
544 		 * qp->s_state is normally set to the opcode of the
545 		 * last packet constructed for new requests and therefore
546 		 * is never set to RDMA read response.
547 		 * RDMA_READ_RESPONSE_MIDDLE is used by the ACK processing
548 		 * thread to indicate a RDMA read needs to be restarted from
549 		 * an earlier PSN without interferring with the sending thread.
550 		 * See qib_restart_rc().
551 		 */
552 		len = ((qp->s_psn - wqe->psn) & QIB_PSN_MASK) * pmtu;
553 		ohdr->u.rc.reth.vaddr =
554 			cpu_to_be64(wqe->rdma_wr.remote_addr + len);
555 		ohdr->u.rc.reth.rkey =
556 			cpu_to_be32(wqe->rdma_wr.rkey);
557 		ohdr->u.rc.reth.length = cpu_to_be32(wqe->length - len);
558 		qp->s_state = OP(RDMA_READ_REQUEST);
559 		hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
560 		bth2 = (qp->s_psn & QIB_PSN_MASK) | IB_BTH_REQ_ACK;
561 		qp->s_psn = wqe->lpsn + 1;
562 		ss = NULL;
563 		len = 0;
564 		qp->s_cur++;
565 		if (qp->s_cur == qp->s_size)
566 			qp->s_cur = 0;
567 		break;
568 	}
569 	qp->s_sending_hpsn = bth2;
570 	delta = (((int) bth2 - (int) wqe->psn) << 8) >> 8;
571 	if (delta && delta % QIB_PSN_CREDIT == 0)
572 		bth2 |= IB_BTH_REQ_ACK;
573 	if (qp->s_flags & RVT_S_SEND_ONE) {
574 		qp->s_flags &= ~RVT_S_SEND_ONE;
575 		qp->s_flags |= RVT_S_WAIT_ACK;
576 		bth2 |= IB_BTH_REQ_ACK;
577 	}
578 	qp->s_len -= len;
579 	qp->s_hdrwords = hwords;
580 	qp->s_cur_sge = ss;
581 	qp->s_cur_size = len;
582 	qib_make_ruc_header(qp, ohdr, bth0 | (qp->s_state << 24), bth2);
583 done:
584 	return 1;
585 bail:
586 	qp->s_flags &= ~RVT_S_BUSY;
587 	return ret;
588 }
589 
590 /**
591  * qib_send_rc_ack - Construct an ACK packet and send it
592  * @qp: a pointer to the QP
593  *
594  * This is called from qib_rc_rcv() and qib_kreceive().
595  * Note that RDMA reads and atomics are handled in the
596  * send side QP state and tasklet.
597  */
598 void qib_send_rc_ack(struct rvt_qp *qp)
599 {
600 	struct qib_devdata *dd = dd_from_ibdev(qp->ibqp.device);
601 	struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
602 	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
603 	u64 pbc;
604 	u16 lrh0;
605 	u32 bth0;
606 	u32 hwords;
607 	u32 pbufn;
608 	u32 __iomem *piobuf;
609 	struct ib_header hdr;
610 	struct ib_other_headers *ohdr;
611 	u32 control;
612 	unsigned long flags;
613 
614 	spin_lock_irqsave(&qp->s_lock, flags);
615 
616 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
617 		goto unlock;
618 
619 	/* Don't send ACK or NAK if a RDMA read or atomic is pending. */
620 	if ((qp->s_flags & RVT_S_RESP_PENDING) || qp->s_rdma_ack_cnt)
621 		goto queue_ack;
622 
623 	/* Construct the header with s_lock held so APM doesn't change it. */
624 	ohdr = &hdr.u.oth;
625 	lrh0 = QIB_LRH_BTH;
626 	/* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4. */
627 	hwords = 6;
628 	if (unlikely(rdma_ah_get_ah_flags(&qp->remote_ah_attr) &
629 		     IB_AH_GRH)) {
630 		hwords += qib_make_grh(ibp, &hdr.u.l.grh,
631 				       rdma_ah_read_grh(&qp->remote_ah_attr),
632 				       hwords, 0);
633 		ohdr = &hdr.u.l.oth;
634 		lrh0 = QIB_LRH_GRH;
635 	}
636 	/* read pkey_index w/o lock (its atomic) */
637 	bth0 = qib_get_pkey(ibp, qp->s_pkey_index) | (OP(ACKNOWLEDGE) << 24);
638 	if (qp->s_mig_state == IB_MIG_MIGRATED)
639 		bth0 |= IB_BTH_MIG_REQ;
640 	if (qp->r_nak_state)
641 		ohdr->u.aeth = cpu_to_be32((qp->r_msn & IB_MSN_MASK) |
642 					    (qp->r_nak_state <<
643 					     IB_AETH_CREDIT_SHIFT));
644 	else
645 		ohdr->u.aeth = rvt_compute_aeth(qp);
646 	lrh0 |= ibp->sl_to_vl[rdma_ah_get_sl(&qp->remote_ah_attr)] << 12 |
647 		rdma_ah_get_sl(&qp->remote_ah_attr) << 4;
648 	hdr.lrh[0] = cpu_to_be16(lrh0);
649 	hdr.lrh[1] = cpu_to_be16(rdma_ah_get_dlid(&qp->remote_ah_attr));
650 	hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC);
651 	hdr.lrh[3] = cpu_to_be16(ppd->lid |
652 				 rdma_ah_get_path_bits(&qp->remote_ah_attr));
653 	ohdr->bth[0] = cpu_to_be32(bth0);
654 	ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
655 	ohdr->bth[2] = cpu_to_be32(qp->r_ack_psn & QIB_PSN_MASK);
656 
657 	spin_unlock_irqrestore(&qp->s_lock, flags);
658 
659 	/* Don't try to send ACKs if the link isn't ACTIVE */
660 	if (!(ppd->lflags & QIBL_LINKACTIVE))
661 		goto done;
662 
663 	control = dd->f_setpbc_control(ppd, hwords + SIZE_OF_CRC,
664 				       qp->s_srate, lrh0 >> 12);
665 	/* length is + 1 for the control dword */
666 	pbc = ((u64) control << 32) | (hwords + 1);
667 
668 	piobuf = dd->f_getsendbuf(ppd, pbc, &pbufn);
669 	if (!piobuf) {
670 		/*
671 		 * We are out of PIO buffers at the moment.
672 		 * Pass responsibility for sending the ACK to the
673 		 * send tasklet so that when a PIO buffer becomes
674 		 * available, the ACK is sent ahead of other outgoing
675 		 * packets.
676 		 */
677 		spin_lock_irqsave(&qp->s_lock, flags);
678 		goto queue_ack;
679 	}
680 
681 	/*
682 	 * Write the pbc.
683 	 * We have to flush after the PBC for correctness
684 	 * on some cpus or WC buffer can be written out of order.
685 	 */
686 	writeq(pbc, piobuf);
687 
688 	if (dd->flags & QIB_PIO_FLUSH_WC) {
689 		u32 *hdrp = (u32 *) &hdr;
690 
691 		qib_flush_wc();
692 		qib_pio_copy(piobuf + 2, hdrp, hwords - 1);
693 		qib_flush_wc();
694 		__raw_writel(hdrp[hwords - 1], piobuf + hwords + 1);
695 	} else
696 		qib_pio_copy(piobuf + 2, (u32 *) &hdr, hwords);
697 
698 	if (dd->flags & QIB_USE_SPCL_TRIG) {
699 		u32 spcl_off = (pbufn >= dd->piobcnt2k) ? 2047 : 1023;
700 
701 		qib_flush_wc();
702 		__raw_writel(0xaebecede, piobuf + spcl_off);
703 	}
704 
705 	qib_flush_wc();
706 	qib_sendbuf_done(dd, pbufn);
707 
708 	this_cpu_inc(ibp->pmastats->n_unicast_xmit);
709 	goto done;
710 
711 queue_ack:
712 	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
713 		this_cpu_inc(*ibp->rvp.rc_qacks);
714 		qp->s_flags |= RVT_S_ACK_PENDING | RVT_S_RESP_PENDING;
715 		qp->s_nak_state = qp->r_nak_state;
716 		qp->s_ack_psn = qp->r_ack_psn;
717 
718 		/* Schedule the send tasklet. */
719 		qib_schedule_send(qp);
720 	}
721 unlock:
722 	spin_unlock_irqrestore(&qp->s_lock, flags);
723 done:
724 	return;
725 }
726 
727 /**
728  * reset_psn - reset the QP state to send starting from PSN
729  * @qp: the QP
730  * @psn: the packet sequence number to restart at
731  *
732  * This is called from qib_rc_rcv() to process an incoming RC ACK
733  * for the given QP.
734  * Called at interrupt level with the QP s_lock held.
735  */
736 static void reset_psn(struct rvt_qp *qp, u32 psn)
737 {
738 	u32 n = qp->s_acked;
739 	struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, n);
740 	u32 opcode;
741 
742 	qp->s_cur = n;
743 
744 	/*
745 	 * If we are starting the request from the beginning,
746 	 * let the normal send code handle initialization.
747 	 */
748 	if (qib_cmp24(psn, wqe->psn) <= 0) {
749 		qp->s_state = OP(SEND_LAST);
750 		goto done;
751 	}
752 
753 	/* Find the work request opcode corresponding to the given PSN. */
754 	opcode = wqe->wr.opcode;
755 	for (;;) {
756 		int diff;
757 
758 		if (++n == qp->s_size)
759 			n = 0;
760 		if (n == qp->s_tail)
761 			break;
762 		wqe = rvt_get_swqe_ptr(qp, n);
763 		diff = qib_cmp24(psn, wqe->psn);
764 		if (diff < 0)
765 			break;
766 		qp->s_cur = n;
767 		/*
768 		 * If we are starting the request from the beginning,
769 		 * let the normal send code handle initialization.
770 		 */
771 		if (diff == 0) {
772 			qp->s_state = OP(SEND_LAST);
773 			goto done;
774 		}
775 		opcode = wqe->wr.opcode;
776 	}
777 
778 	/*
779 	 * Set the state to restart in the middle of a request.
780 	 * Don't change the s_sge, s_cur_sge, or s_cur_size.
781 	 * See qib_make_rc_req().
782 	 */
783 	switch (opcode) {
784 	case IB_WR_SEND:
785 	case IB_WR_SEND_WITH_IMM:
786 		qp->s_state = OP(RDMA_READ_RESPONSE_FIRST);
787 		break;
788 
789 	case IB_WR_RDMA_WRITE:
790 	case IB_WR_RDMA_WRITE_WITH_IMM:
791 		qp->s_state = OP(RDMA_READ_RESPONSE_LAST);
792 		break;
793 
794 	case IB_WR_RDMA_READ:
795 		qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE);
796 		break;
797 
798 	default:
799 		/*
800 		 * This case shouldn't happen since its only
801 		 * one PSN per req.
802 		 */
803 		qp->s_state = OP(SEND_LAST);
804 	}
805 done:
806 	qp->s_psn = psn;
807 	/*
808 	 * Set RVT_S_WAIT_PSN as qib_rc_complete() may start the timer
809 	 * asynchronously before the send tasklet can get scheduled.
810 	 * Doing it in qib_make_rc_req() is too late.
811 	 */
812 	if ((qib_cmp24(qp->s_psn, qp->s_sending_hpsn) <= 0) &&
813 	    (qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) <= 0))
814 		qp->s_flags |= RVT_S_WAIT_PSN;
815 }
816 
817 /*
818  * Back up requester to resend the last un-ACKed request.
819  * The QP r_lock and s_lock should be held and interrupts disabled.
820  */
821 void qib_restart_rc(struct rvt_qp *qp, u32 psn, int wait)
822 {
823 	struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
824 	struct qib_ibport *ibp;
825 
826 	if (qp->s_retry == 0) {
827 		if (qp->s_mig_state == IB_MIG_ARMED) {
828 			qib_migrate_qp(qp);
829 			qp->s_retry = qp->s_retry_cnt;
830 		} else if (qp->s_last == qp->s_acked) {
831 			rvt_send_complete(qp, wqe, IB_WC_RETRY_EXC_ERR);
832 			rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
833 			return;
834 		} else /* XXX need to handle delayed completion */
835 			return;
836 	} else
837 		qp->s_retry--;
838 
839 	ibp = to_iport(qp->ibqp.device, qp->port_num);
840 	if (wqe->wr.opcode == IB_WR_RDMA_READ)
841 		ibp->rvp.n_rc_resends++;
842 	else
843 		ibp->rvp.n_rc_resends += (qp->s_psn - psn) & QIB_PSN_MASK;
844 
845 	qp->s_flags &= ~(RVT_S_WAIT_FENCE | RVT_S_WAIT_RDMAR |
846 			 RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_PSN |
847 			 RVT_S_WAIT_ACK);
848 	if (wait)
849 		qp->s_flags |= RVT_S_SEND_ONE;
850 	reset_psn(qp, psn);
851 }
852 
853 /*
854  * Set qp->s_sending_psn to the next PSN after the given one.
855  * This would be psn+1 except when RDMA reads are present.
856  */
857 static void reset_sending_psn(struct rvt_qp *qp, u32 psn)
858 {
859 	struct rvt_swqe *wqe;
860 	u32 n = qp->s_last;
861 
862 	/* Find the work request corresponding to the given PSN. */
863 	for (;;) {
864 		wqe = rvt_get_swqe_ptr(qp, n);
865 		if (qib_cmp24(psn, wqe->lpsn) <= 0) {
866 			if (wqe->wr.opcode == IB_WR_RDMA_READ)
867 				qp->s_sending_psn = wqe->lpsn + 1;
868 			else
869 				qp->s_sending_psn = psn + 1;
870 			break;
871 		}
872 		if (++n == qp->s_size)
873 			n = 0;
874 		if (n == qp->s_tail)
875 			break;
876 	}
877 }
878 
879 /*
880  * This should be called with the QP s_lock held and interrupts disabled.
881  */
882 void qib_rc_send_complete(struct rvt_qp *qp, struct ib_header *hdr)
883 {
884 	struct ib_other_headers *ohdr;
885 	struct rvt_swqe *wqe;
886 	u32 opcode;
887 	u32 psn;
888 
889 	if (!(ib_rvt_state_ops[qp->state] & RVT_SEND_OR_FLUSH_OR_RECV_OK))
890 		return;
891 
892 	/* Find out where the BTH is */
893 	if ((be16_to_cpu(hdr->lrh[0]) & 3) == QIB_LRH_BTH)
894 		ohdr = &hdr->u.oth;
895 	else
896 		ohdr = &hdr->u.l.oth;
897 
898 	opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
899 	if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
900 	    opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
901 		WARN_ON(!qp->s_rdma_ack_cnt);
902 		qp->s_rdma_ack_cnt--;
903 		return;
904 	}
905 
906 	psn = be32_to_cpu(ohdr->bth[2]);
907 	reset_sending_psn(qp, psn);
908 
909 	/*
910 	 * Start timer after a packet requesting an ACK has been sent and
911 	 * there are still requests that haven't been acked.
912 	 */
913 	if ((psn & IB_BTH_REQ_ACK) && qp->s_acked != qp->s_tail &&
914 	    !(qp->s_flags & (RVT_S_TIMER | RVT_S_WAIT_RNR | RVT_S_WAIT_PSN)) &&
915 	    (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
916 		rvt_add_retry_timer(qp);
917 
918 	while (qp->s_last != qp->s_acked) {
919 		wqe = rvt_get_swqe_ptr(qp, qp->s_last);
920 		if (qib_cmp24(wqe->lpsn, qp->s_sending_psn) >= 0 &&
921 		    qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) <= 0)
922 			break;
923 		rvt_qp_complete_swqe(qp,
924 				     wqe,
925 				     ib_qib_wc_opcode[wqe->wr.opcode],
926 				     IB_WC_SUCCESS);
927 	}
928 	/*
929 	 * If we were waiting for sends to complete before resending,
930 	 * and they are now complete, restart sending.
931 	 */
932 	if (qp->s_flags & RVT_S_WAIT_PSN &&
933 	    qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) > 0) {
934 		qp->s_flags &= ~RVT_S_WAIT_PSN;
935 		qp->s_sending_psn = qp->s_psn;
936 		qp->s_sending_hpsn = qp->s_psn - 1;
937 		qib_schedule_send(qp);
938 	}
939 }
940 
941 static inline void update_last_psn(struct rvt_qp *qp, u32 psn)
942 {
943 	qp->s_last_psn = psn;
944 }
945 
946 /*
947  * Generate a SWQE completion.
948  * This is similar to qib_send_complete but has to check to be sure
949  * that the SGEs are not being referenced if the SWQE is being resent.
950  */
951 static struct rvt_swqe *do_rc_completion(struct rvt_qp *qp,
952 					 struct rvt_swqe *wqe,
953 					 struct qib_ibport *ibp)
954 {
955 	/*
956 	 * Don't decrement refcount and don't generate a
957 	 * completion if the SWQE is being resent until the send
958 	 * is finished.
959 	 */
960 	if (qib_cmp24(wqe->lpsn, qp->s_sending_psn) < 0 ||
961 	    qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) > 0)
962 		rvt_qp_complete_swqe(qp,
963 				     wqe,
964 				     ib_qib_wc_opcode[wqe->wr.opcode],
965 				     IB_WC_SUCCESS);
966 	else
967 		this_cpu_inc(*ibp->rvp.rc_delayed_comp);
968 
969 	qp->s_retry = qp->s_retry_cnt;
970 	update_last_psn(qp, wqe->lpsn);
971 
972 	/*
973 	 * If we are completing a request which is in the process of
974 	 * being resent, we can stop resending it since we know the
975 	 * responder has already seen it.
976 	 */
977 	if (qp->s_acked == qp->s_cur) {
978 		if (++qp->s_cur >= qp->s_size)
979 			qp->s_cur = 0;
980 		qp->s_acked = qp->s_cur;
981 		wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
982 		if (qp->s_acked != qp->s_tail) {
983 			qp->s_state = OP(SEND_LAST);
984 			qp->s_psn = wqe->psn;
985 		}
986 	} else {
987 		if (++qp->s_acked >= qp->s_size)
988 			qp->s_acked = 0;
989 		if (qp->state == IB_QPS_SQD && qp->s_acked == qp->s_cur)
990 			qp->s_draining = 0;
991 		wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
992 	}
993 	return wqe;
994 }
995 
996 /*
997  * do_rc_ack - process an incoming RC ACK
998  * @qp: the QP the ACK came in on
999  * @psn: the packet sequence number of the ACK
1000  * @opcode: the opcode of the request that resulted in the ACK
1001  *
1002  * This is called from qib_rc_rcv_resp() to process an incoming RC ACK
1003  * for the given QP.
1004  * Called at interrupt level with the QP s_lock held.
1005  * Returns 1 if OK, 0 if current operation should be aborted (NAK).
1006  */
1007 static int do_rc_ack(struct rvt_qp *qp, u32 aeth, u32 psn, int opcode,
1008 		     u64 val, struct qib_ctxtdata *rcd)
1009 {
1010 	struct qib_ibport *ibp;
1011 	enum ib_wc_status status;
1012 	struct rvt_swqe *wqe;
1013 	int ret = 0;
1014 	u32 ack_psn;
1015 	int diff;
1016 
1017 	/*
1018 	 * Note that NAKs implicitly ACK outstanding SEND and RDMA write
1019 	 * requests and implicitly NAK RDMA read and atomic requests issued
1020 	 * before the NAK'ed request.  The MSN won't include the NAK'ed
1021 	 * request but will include an ACK'ed request(s).
1022 	 */
1023 	ack_psn = psn;
1024 	if (aeth >> IB_AETH_NAK_SHIFT)
1025 		ack_psn--;
1026 	wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1027 	ibp = to_iport(qp->ibqp.device, qp->port_num);
1028 
1029 	/*
1030 	 * The MSN might be for a later WQE than the PSN indicates so
1031 	 * only complete WQEs that the PSN finishes.
1032 	 */
1033 	while ((diff = qib_cmp24(ack_psn, wqe->lpsn)) >= 0) {
1034 		/*
1035 		 * RDMA_READ_RESPONSE_ONLY is a special case since
1036 		 * we want to generate completion events for everything
1037 		 * before the RDMA read, copy the data, then generate
1038 		 * the completion for the read.
1039 		 */
1040 		if (wqe->wr.opcode == IB_WR_RDMA_READ &&
1041 		    opcode == OP(RDMA_READ_RESPONSE_ONLY) &&
1042 		    diff == 0) {
1043 			ret = 1;
1044 			goto bail;
1045 		}
1046 		/*
1047 		 * If this request is a RDMA read or atomic, and the ACK is
1048 		 * for a later operation, this ACK NAKs the RDMA read or
1049 		 * atomic.  In other words, only a RDMA_READ_LAST or ONLY
1050 		 * can ACK a RDMA read and likewise for atomic ops.  Note
1051 		 * that the NAK case can only happen if relaxed ordering is
1052 		 * used and requests are sent after an RDMA read or atomic
1053 		 * is sent but before the response is received.
1054 		 */
1055 		if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
1056 		     (opcode != OP(RDMA_READ_RESPONSE_LAST) || diff != 0)) ||
1057 		    ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1058 		      wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
1059 		     (opcode != OP(ATOMIC_ACKNOWLEDGE) || diff != 0))) {
1060 			/* Retry this request. */
1061 			if (!(qp->r_flags & RVT_R_RDMAR_SEQ)) {
1062 				qp->r_flags |= RVT_R_RDMAR_SEQ;
1063 				qib_restart_rc(qp, qp->s_last_psn + 1, 0);
1064 				if (list_empty(&qp->rspwait)) {
1065 					qp->r_flags |= RVT_R_RSP_SEND;
1066 					rvt_get_qp(qp);
1067 					list_add_tail(&qp->rspwait,
1068 						      &rcd->qp_wait_list);
1069 				}
1070 			}
1071 			/*
1072 			 * No need to process the ACK/NAK since we are
1073 			 * restarting an earlier request.
1074 			 */
1075 			goto bail;
1076 		}
1077 		if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1078 		    wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
1079 			u64 *vaddr = wqe->sg_list[0].vaddr;
1080 			*vaddr = val;
1081 		}
1082 		if (qp->s_num_rd_atomic &&
1083 		    (wqe->wr.opcode == IB_WR_RDMA_READ ||
1084 		     wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1085 		     wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)) {
1086 			qp->s_num_rd_atomic--;
1087 			/* Restart sending task if fence is complete */
1088 			if ((qp->s_flags & RVT_S_WAIT_FENCE) &&
1089 			    !qp->s_num_rd_atomic) {
1090 				qp->s_flags &= ~(RVT_S_WAIT_FENCE |
1091 						 RVT_S_WAIT_ACK);
1092 				qib_schedule_send(qp);
1093 			} else if (qp->s_flags & RVT_S_WAIT_RDMAR) {
1094 				qp->s_flags &= ~(RVT_S_WAIT_RDMAR |
1095 						 RVT_S_WAIT_ACK);
1096 				qib_schedule_send(qp);
1097 			}
1098 		}
1099 		wqe = do_rc_completion(qp, wqe, ibp);
1100 		if (qp->s_acked == qp->s_tail)
1101 			break;
1102 	}
1103 
1104 	switch (aeth >> IB_AETH_NAK_SHIFT) {
1105 	case 0:         /* ACK */
1106 		this_cpu_inc(*ibp->rvp.rc_acks);
1107 		if (qp->s_acked != qp->s_tail) {
1108 			/*
1109 			 * We are expecting more ACKs so
1110 			 * reset the retransmit timer.
1111 			 */
1112 			rvt_mod_retry_timer(qp);
1113 			/*
1114 			 * We can stop resending the earlier packets and
1115 			 * continue with the next packet the receiver wants.
1116 			 */
1117 			if (qib_cmp24(qp->s_psn, psn) <= 0)
1118 				reset_psn(qp, psn + 1);
1119 		} else {
1120 			/* No more acks - kill all timers */
1121 			rvt_stop_rc_timers(qp);
1122 			if (qib_cmp24(qp->s_psn, psn) <= 0) {
1123 				qp->s_state = OP(SEND_LAST);
1124 				qp->s_psn = psn + 1;
1125 			}
1126 		}
1127 		if (qp->s_flags & RVT_S_WAIT_ACK) {
1128 			qp->s_flags &= ~RVT_S_WAIT_ACK;
1129 			qib_schedule_send(qp);
1130 		}
1131 		rvt_get_credit(qp, aeth);
1132 		qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1133 		qp->s_retry = qp->s_retry_cnt;
1134 		update_last_psn(qp, psn);
1135 		return 1;
1136 
1137 	case 1:         /* RNR NAK */
1138 		ibp->rvp.n_rnr_naks++;
1139 		if (qp->s_acked == qp->s_tail)
1140 			goto bail;
1141 		if (qp->s_flags & RVT_S_WAIT_RNR)
1142 			goto bail;
1143 		if (qp->s_rnr_retry == 0) {
1144 			status = IB_WC_RNR_RETRY_EXC_ERR;
1145 			goto class_b;
1146 		}
1147 		if (qp->s_rnr_retry_cnt < 7)
1148 			qp->s_rnr_retry--;
1149 
1150 		/* The last valid PSN is the previous PSN. */
1151 		update_last_psn(qp, psn - 1);
1152 
1153 		ibp->rvp.n_rc_resends += (qp->s_psn - psn) & QIB_PSN_MASK;
1154 
1155 		reset_psn(qp, psn);
1156 
1157 		qp->s_flags &= ~(RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_ACK);
1158 		rvt_stop_rc_timers(qp);
1159 		rvt_add_rnr_timer(qp, aeth);
1160 		return 0;
1161 
1162 	case 3:         /* NAK */
1163 		if (qp->s_acked == qp->s_tail)
1164 			goto bail;
1165 		/* The last valid PSN is the previous PSN. */
1166 		update_last_psn(qp, psn - 1);
1167 		switch ((aeth >> IB_AETH_CREDIT_SHIFT) &
1168 			IB_AETH_CREDIT_MASK) {
1169 		case 0: /* PSN sequence error */
1170 			ibp->rvp.n_seq_naks++;
1171 			/*
1172 			 * Back up to the responder's expected PSN.
1173 			 * Note that we might get a NAK in the middle of an
1174 			 * RDMA READ response which terminates the RDMA
1175 			 * READ.
1176 			 */
1177 			qib_restart_rc(qp, psn, 0);
1178 			qib_schedule_send(qp);
1179 			break;
1180 
1181 		case 1: /* Invalid Request */
1182 			status = IB_WC_REM_INV_REQ_ERR;
1183 			ibp->rvp.n_other_naks++;
1184 			goto class_b;
1185 
1186 		case 2: /* Remote Access Error */
1187 			status = IB_WC_REM_ACCESS_ERR;
1188 			ibp->rvp.n_other_naks++;
1189 			goto class_b;
1190 
1191 		case 3: /* Remote Operation Error */
1192 			status = IB_WC_REM_OP_ERR;
1193 			ibp->rvp.n_other_naks++;
1194 class_b:
1195 			if (qp->s_last == qp->s_acked) {
1196 				rvt_send_complete(qp, wqe, status);
1197 				rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1198 			}
1199 			break;
1200 
1201 		default:
1202 			/* Ignore other reserved NAK error codes */
1203 			goto reserved;
1204 		}
1205 		qp->s_retry = qp->s_retry_cnt;
1206 		qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1207 		goto bail;
1208 
1209 	default:                /* 2: reserved */
1210 reserved:
1211 		/* Ignore reserved NAK codes. */
1212 		goto bail;
1213 	}
1214 
1215 bail:
1216 	rvt_stop_rc_timers(qp);
1217 	return ret;
1218 }
1219 
1220 /*
1221  * We have seen an out of sequence RDMA read middle or last packet.
1222  * This ACKs SENDs and RDMA writes up to the first RDMA read or atomic SWQE.
1223  */
1224 static void rdma_seq_err(struct rvt_qp *qp, struct qib_ibport *ibp, u32 psn,
1225 			 struct qib_ctxtdata *rcd)
1226 {
1227 	struct rvt_swqe *wqe;
1228 
1229 	/* Remove QP from retry timer */
1230 	rvt_stop_rc_timers(qp);
1231 
1232 	wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1233 
1234 	while (qib_cmp24(psn, wqe->lpsn) > 0) {
1235 		if (wqe->wr.opcode == IB_WR_RDMA_READ ||
1236 		    wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1237 		    wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
1238 			break;
1239 		wqe = do_rc_completion(qp, wqe, ibp);
1240 	}
1241 
1242 	ibp->rvp.n_rdma_seq++;
1243 	qp->r_flags |= RVT_R_RDMAR_SEQ;
1244 	qib_restart_rc(qp, qp->s_last_psn + 1, 0);
1245 	if (list_empty(&qp->rspwait)) {
1246 		qp->r_flags |= RVT_R_RSP_SEND;
1247 		rvt_get_qp(qp);
1248 		list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
1249 	}
1250 }
1251 
1252 /**
1253  * qib_rc_rcv_resp - process an incoming RC response packet
1254  * @ibp: the port this packet came in on
1255  * @ohdr: the other headers for this packet
1256  * @data: the packet data
1257  * @tlen: the packet length
1258  * @qp: the QP for this packet
1259  * @opcode: the opcode for this packet
1260  * @psn: the packet sequence number for this packet
1261  * @hdrsize: the header length
1262  * @pmtu: the path MTU
1263  * @rcd: the context pointer
1264  *
1265  * This is called from qib_rc_rcv() to process an incoming RC response
1266  * packet for the given QP.
1267  * Called at interrupt level.
1268  */
1269 static void qib_rc_rcv_resp(struct qib_ibport *ibp,
1270 			    struct ib_other_headers *ohdr,
1271 			    void *data, u32 tlen,
1272 			    struct rvt_qp *qp,
1273 			    u32 opcode,
1274 			    u32 psn, u32 hdrsize, u32 pmtu,
1275 			    struct qib_ctxtdata *rcd)
1276 {
1277 	struct rvt_swqe *wqe;
1278 	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1279 	enum ib_wc_status status;
1280 	unsigned long flags;
1281 	int diff;
1282 	u32 pad;
1283 	u32 aeth;
1284 	u64 val;
1285 
1286 	if (opcode != OP(RDMA_READ_RESPONSE_MIDDLE)) {
1287 		/*
1288 		 * If ACK'd PSN on SDMA busy list try to make progress to
1289 		 * reclaim SDMA credits.
1290 		 */
1291 		if ((qib_cmp24(psn, qp->s_sending_psn) >= 0) &&
1292 		    (qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) <= 0)) {
1293 
1294 			/*
1295 			 * If send tasklet not running attempt to progress
1296 			 * SDMA queue.
1297 			 */
1298 			if (!(qp->s_flags & RVT_S_BUSY)) {
1299 				/* Acquire SDMA Lock */
1300 				spin_lock_irqsave(&ppd->sdma_lock, flags);
1301 				/* Invoke sdma make progress */
1302 				qib_sdma_make_progress(ppd);
1303 				/* Release SDMA Lock */
1304 				spin_unlock_irqrestore(&ppd->sdma_lock, flags);
1305 			}
1306 		}
1307 	}
1308 
1309 	spin_lock_irqsave(&qp->s_lock, flags);
1310 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
1311 		goto ack_done;
1312 
1313 	/* Ignore invalid responses. */
1314 	if (qib_cmp24(psn, READ_ONCE(qp->s_next_psn)) >= 0)
1315 		goto ack_done;
1316 
1317 	/* Ignore duplicate responses. */
1318 	diff = qib_cmp24(psn, qp->s_last_psn);
1319 	if (unlikely(diff <= 0)) {
1320 		/* Update credits for "ghost" ACKs */
1321 		if (diff == 0 && opcode == OP(ACKNOWLEDGE)) {
1322 			aeth = be32_to_cpu(ohdr->u.aeth);
1323 			if ((aeth >> IB_AETH_NAK_SHIFT) == 0)
1324 				rvt_get_credit(qp, aeth);
1325 		}
1326 		goto ack_done;
1327 	}
1328 
1329 	/*
1330 	 * Skip everything other than the PSN we expect, if we are waiting
1331 	 * for a reply to a restarted RDMA read or atomic op.
1332 	 */
1333 	if (qp->r_flags & RVT_R_RDMAR_SEQ) {
1334 		if (qib_cmp24(psn, qp->s_last_psn + 1) != 0)
1335 			goto ack_done;
1336 		qp->r_flags &= ~RVT_R_RDMAR_SEQ;
1337 	}
1338 
1339 	if (unlikely(qp->s_acked == qp->s_tail))
1340 		goto ack_done;
1341 	wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1342 	status = IB_WC_SUCCESS;
1343 
1344 	switch (opcode) {
1345 	case OP(ACKNOWLEDGE):
1346 	case OP(ATOMIC_ACKNOWLEDGE):
1347 	case OP(RDMA_READ_RESPONSE_FIRST):
1348 		aeth = be32_to_cpu(ohdr->u.aeth);
1349 		if (opcode == OP(ATOMIC_ACKNOWLEDGE))
1350 			val = ib_u64_get(&ohdr->u.at.atomic_ack_eth);
1351 		else
1352 			val = 0;
1353 		if (!do_rc_ack(qp, aeth, psn, opcode, val, rcd) ||
1354 		    opcode != OP(RDMA_READ_RESPONSE_FIRST))
1355 			goto ack_done;
1356 		hdrsize += 4;
1357 		wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1358 		if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1359 			goto ack_op_err;
1360 		/*
1361 		 * If this is a response to a resent RDMA read, we
1362 		 * have to be careful to copy the data to the right
1363 		 * location.
1364 		 */
1365 		qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1366 						  wqe, psn, pmtu);
1367 		goto read_middle;
1368 
1369 	case OP(RDMA_READ_RESPONSE_MIDDLE):
1370 		/* no AETH, no ACK */
1371 		if (unlikely(qib_cmp24(psn, qp->s_last_psn + 1)))
1372 			goto ack_seq_err;
1373 		if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1374 			goto ack_op_err;
1375 read_middle:
1376 		if (unlikely(tlen != (hdrsize + pmtu + 4)))
1377 			goto ack_len_err;
1378 		if (unlikely(pmtu >= qp->s_rdma_read_len))
1379 			goto ack_len_err;
1380 
1381 		/*
1382 		 * We got a response so update the timeout.
1383 		 * 4.096 usec. * (1 << qp->timeout)
1384 		 */
1385 		rvt_mod_retry_timer(qp);
1386 		if (qp->s_flags & RVT_S_WAIT_ACK) {
1387 			qp->s_flags &= ~RVT_S_WAIT_ACK;
1388 			qib_schedule_send(qp);
1389 		}
1390 
1391 		if (opcode == OP(RDMA_READ_RESPONSE_MIDDLE))
1392 			qp->s_retry = qp->s_retry_cnt;
1393 
1394 		/*
1395 		 * Update the RDMA receive state but do the copy w/o
1396 		 * holding the locks and blocking interrupts.
1397 		 */
1398 		qp->s_rdma_read_len -= pmtu;
1399 		update_last_psn(qp, psn);
1400 		spin_unlock_irqrestore(&qp->s_lock, flags);
1401 		rvt_copy_sge(qp, &qp->s_rdma_read_sge,
1402 			     data, pmtu, false, false);
1403 		goto bail;
1404 
1405 	case OP(RDMA_READ_RESPONSE_ONLY):
1406 		aeth = be32_to_cpu(ohdr->u.aeth);
1407 		if (!do_rc_ack(qp, aeth, psn, opcode, 0, rcd))
1408 			goto ack_done;
1409 		/* Get the number of bytes the message was padded by. */
1410 		pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1411 		/*
1412 		 * Check that the data size is >= 0 && <= pmtu.
1413 		 * Remember to account for the AETH header (4) and
1414 		 * ICRC (4).
1415 		 */
1416 		if (unlikely(tlen < (hdrsize + pad + 8)))
1417 			goto ack_len_err;
1418 		/*
1419 		 * If this is a response to a resent RDMA read, we
1420 		 * have to be careful to copy the data to the right
1421 		 * location.
1422 		 */
1423 		wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1424 		qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1425 						  wqe, psn, pmtu);
1426 		goto read_last;
1427 
1428 	case OP(RDMA_READ_RESPONSE_LAST):
1429 		/* ACKs READ req. */
1430 		if (unlikely(qib_cmp24(psn, qp->s_last_psn + 1)))
1431 			goto ack_seq_err;
1432 		if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1433 			goto ack_op_err;
1434 		/* Get the number of bytes the message was padded by. */
1435 		pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1436 		/*
1437 		 * Check that the data size is >= 1 && <= pmtu.
1438 		 * Remember to account for the AETH header (4) and
1439 		 * ICRC (4).
1440 		 */
1441 		if (unlikely(tlen <= (hdrsize + pad + 8)))
1442 			goto ack_len_err;
1443 read_last:
1444 		tlen -= hdrsize + pad + 8;
1445 		if (unlikely(tlen != qp->s_rdma_read_len))
1446 			goto ack_len_err;
1447 		aeth = be32_to_cpu(ohdr->u.aeth);
1448 		rvt_copy_sge(qp, &qp->s_rdma_read_sge,
1449 			     data, tlen, false, false);
1450 		WARN_ON(qp->s_rdma_read_sge.num_sge);
1451 		(void) do_rc_ack(qp, aeth, psn,
1452 				 OP(RDMA_READ_RESPONSE_LAST), 0, rcd);
1453 		goto ack_done;
1454 	}
1455 
1456 ack_op_err:
1457 	status = IB_WC_LOC_QP_OP_ERR;
1458 	goto ack_err;
1459 
1460 ack_seq_err:
1461 	rdma_seq_err(qp, ibp, psn, rcd);
1462 	goto ack_done;
1463 
1464 ack_len_err:
1465 	status = IB_WC_LOC_LEN_ERR;
1466 ack_err:
1467 	if (qp->s_last == qp->s_acked) {
1468 		rvt_send_complete(qp, wqe, status);
1469 		rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1470 	}
1471 ack_done:
1472 	spin_unlock_irqrestore(&qp->s_lock, flags);
1473 bail:
1474 	return;
1475 }
1476 
1477 /**
1478  * qib_rc_rcv_error - process an incoming duplicate or error RC packet
1479  * @ohdr: the other headers for this packet
1480  * @data: the packet data
1481  * @qp: the QP for this packet
1482  * @opcode: the opcode for this packet
1483  * @psn: the packet sequence number for this packet
1484  * @diff: the difference between the PSN and the expected PSN
1485  * @rcd: the context pointer
1486  *
1487  * This is called from qib_rc_rcv() to process an unexpected
1488  * incoming RC packet for the given QP.
1489  * Called at interrupt level.
1490  * Return 1 if no more processing is needed; otherwise return 0 to
1491  * schedule a response to be sent.
1492  */
1493 static int qib_rc_rcv_error(struct ib_other_headers *ohdr,
1494 			    void *data,
1495 			    struct rvt_qp *qp,
1496 			    u32 opcode,
1497 			    u32 psn,
1498 			    int diff,
1499 			    struct qib_ctxtdata *rcd)
1500 {
1501 	struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
1502 	struct rvt_ack_entry *e;
1503 	unsigned long flags;
1504 	u8 i, prev;
1505 	int old_req;
1506 
1507 	if (diff > 0) {
1508 		/*
1509 		 * Packet sequence error.
1510 		 * A NAK will ACK earlier sends and RDMA writes.
1511 		 * Don't queue the NAK if we already sent one.
1512 		 */
1513 		if (!qp->r_nak_state) {
1514 			ibp->rvp.n_rc_seqnak++;
1515 			qp->r_nak_state = IB_NAK_PSN_ERROR;
1516 			/* Use the expected PSN. */
1517 			qp->r_ack_psn = qp->r_psn;
1518 			/*
1519 			 * Wait to send the sequence NAK until all packets
1520 			 * in the receive queue have been processed.
1521 			 * Otherwise, we end up propagating congestion.
1522 			 */
1523 			if (list_empty(&qp->rspwait)) {
1524 				qp->r_flags |= RVT_R_RSP_NAK;
1525 				rvt_get_qp(qp);
1526 				list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
1527 			}
1528 		}
1529 		goto done;
1530 	}
1531 
1532 	/*
1533 	 * Handle a duplicate request.  Don't re-execute SEND, RDMA
1534 	 * write or atomic op.  Don't NAK errors, just silently drop
1535 	 * the duplicate request.  Note that r_sge, r_len, and
1536 	 * r_rcv_len may be in use so don't modify them.
1537 	 *
1538 	 * We are supposed to ACK the earliest duplicate PSN but we
1539 	 * can coalesce an outstanding duplicate ACK.  We have to
1540 	 * send the earliest so that RDMA reads can be restarted at
1541 	 * the requester's expected PSN.
1542 	 *
1543 	 * First, find where this duplicate PSN falls within the
1544 	 * ACKs previously sent.
1545 	 * old_req is true if there is an older response that is scheduled
1546 	 * to be sent before sending this one.
1547 	 */
1548 	e = NULL;
1549 	old_req = 1;
1550 	ibp->rvp.n_rc_dupreq++;
1551 
1552 	spin_lock_irqsave(&qp->s_lock, flags);
1553 
1554 	for (i = qp->r_head_ack_queue; ; i = prev) {
1555 		if (i == qp->s_tail_ack_queue)
1556 			old_req = 0;
1557 		if (i)
1558 			prev = i - 1;
1559 		else
1560 			prev = QIB_MAX_RDMA_ATOMIC;
1561 		if (prev == qp->r_head_ack_queue) {
1562 			e = NULL;
1563 			break;
1564 		}
1565 		e = &qp->s_ack_queue[prev];
1566 		if (!e->opcode) {
1567 			e = NULL;
1568 			break;
1569 		}
1570 		if (qib_cmp24(psn, e->psn) >= 0) {
1571 			if (prev == qp->s_tail_ack_queue &&
1572 			    qib_cmp24(psn, e->lpsn) <= 0)
1573 				old_req = 0;
1574 			break;
1575 		}
1576 	}
1577 	switch (opcode) {
1578 	case OP(RDMA_READ_REQUEST): {
1579 		struct ib_reth *reth;
1580 		u32 offset;
1581 		u32 len;
1582 
1583 		/*
1584 		 * If we didn't find the RDMA read request in the ack queue,
1585 		 * we can ignore this request.
1586 		 */
1587 		if (!e || e->opcode != OP(RDMA_READ_REQUEST))
1588 			goto unlock_done;
1589 		/* RETH comes after BTH */
1590 		reth = &ohdr->u.rc.reth;
1591 		/*
1592 		 * Address range must be a subset of the original
1593 		 * request and start on pmtu boundaries.
1594 		 * We reuse the old ack_queue slot since the requester
1595 		 * should not back up and request an earlier PSN for the
1596 		 * same request.
1597 		 */
1598 		offset = ((psn - e->psn) & QIB_PSN_MASK) *
1599 			qp->pmtu;
1600 		len = be32_to_cpu(reth->length);
1601 		if (unlikely(offset + len != e->rdma_sge.sge_length))
1602 			goto unlock_done;
1603 		if (e->rdma_sge.mr) {
1604 			rvt_put_mr(e->rdma_sge.mr);
1605 			e->rdma_sge.mr = NULL;
1606 		}
1607 		if (len != 0) {
1608 			u32 rkey = be32_to_cpu(reth->rkey);
1609 			u64 vaddr = be64_to_cpu(reth->vaddr);
1610 			int ok;
1611 
1612 			ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr, rkey,
1613 					 IB_ACCESS_REMOTE_READ);
1614 			if (unlikely(!ok))
1615 				goto unlock_done;
1616 		} else {
1617 			e->rdma_sge.vaddr = NULL;
1618 			e->rdma_sge.length = 0;
1619 			e->rdma_sge.sge_length = 0;
1620 		}
1621 		e->psn = psn;
1622 		if (old_req)
1623 			goto unlock_done;
1624 		qp->s_tail_ack_queue = prev;
1625 		break;
1626 	}
1627 
1628 	case OP(COMPARE_SWAP):
1629 	case OP(FETCH_ADD): {
1630 		/*
1631 		 * If we didn't find the atomic request in the ack queue
1632 		 * or the send tasklet is already backed up to send an
1633 		 * earlier entry, we can ignore this request.
1634 		 */
1635 		if (!e || e->opcode != (u8) opcode || old_req)
1636 			goto unlock_done;
1637 		qp->s_tail_ack_queue = prev;
1638 		break;
1639 	}
1640 
1641 	default:
1642 		/*
1643 		 * Ignore this operation if it doesn't request an ACK
1644 		 * or an earlier RDMA read or atomic is going to be resent.
1645 		 */
1646 		if (!(psn & IB_BTH_REQ_ACK) || old_req)
1647 			goto unlock_done;
1648 		/*
1649 		 * Resend the most recent ACK if this request is
1650 		 * after all the previous RDMA reads and atomics.
1651 		 */
1652 		if (i == qp->r_head_ack_queue) {
1653 			spin_unlock_irqrestore(&qp->s_lock, flags);
1654 			qp->r_nak_state = 0;
1655 			qp->r_ack_psn = qp->r_psn - 1;
1656 			goto send_ack;
1657 		}
1658 		/*
1659 		 * Try to send a simple ACK to work around a Mellanox bug
1660 		 * which doesn't accept a RDMA read response or atomic
1661 		 * response as an ACK for earlier SENDs or RDMA writes.
1662 		 */
1663 		if (!(qp->s_flags & RVT_S_RESP_PENDING)) {
1664 			spin_unlock_irqrestore(&qp->s_lock, flags);
1665 			qp->r_nak_state = 0;
1666 			qp->r_ack_psn = qp->s_ack_queue[i].psn - 1;
1667 			goto send_ack;
1668 		}
1669 		/*
1670 		 * Resend the RDMA read or atomic op which
1671 		 * ACKs this duplicate request.
1672 		 */
1673 		qp->s_tail_ack_queue = i;
1674 		break;
1675 	}
1676 	qp->s_ack_state = OP(ACKNOWLEDGE);
1677 	qp->s_flags |= RVT_S_RESP_PENDING;
1678 	qp->r_nak_state = 0;
1679 	qib_schedule_send(qp);
1680 
1681 unlock_done:
1682 	spin_unlock_irqrestore(&qp->s_lock, flags);
1683 done:
1684 	return 1;
1685 
1686 send_ack:
1687 	return 0;
1688 }
1689 
1690 static inline void qib_update_ack_queue(struct rvt_qp *qp, unsigned n)
1691 {
1692 	unsigned next;
1693 
1694 	next = n + 1;
1695 	if (next > QIB_MAX_RDMA_ATOMIC)
1696 		next = 0;
1697 	qp->s_tail_ack_queue = next;
1698 	qp->s_ack_state = OP(ACKNOWLEDGE);
1699 }
1700 
1701 /**
1702  * qib_rc_rcv - process an incoming RC packet
1703  * @rcd: the context pointer
1704  * @hdr: the header of this packet
1705  * @has_grh: true if the header has a GRH
1706  * @data: the packet data
1707  * @tlen: the packet length
1708  * @qp: the QP for this packet
1709  *
1710  * This is called from qib_qp_rcv() to process an incoming RC packet
1711  * for the given QP.
1712  * Called at interrupt level.
1713  */
1714 void qib_rc_rcv(struct qib_ctxtdata *rcd, struct ib_header *hdr,
1715 		int has_grh, void *data, u32 tlen, struct rvt_qp *qp)
1716 {
1717 	struct qib_ibport *ibp = &rcd->ppd->ibport_data;
1718 	struct ib_other_headers *ohdr;
1719 	u32 opcode;
1720 	u32 hdrsize;
1721 	u32 psn;
1722 	u32 pad;
1723 	struct ib_wc wc;
1724 	u32 pmtu = qp->pmtu;
1725 	int diff;
1726 	struct ib_reth *reth;
1727 	unsigned long flags;
1728 	int ret;
1729 
1730 	/* Check for GRH */
1731 	if (!has_grh) {
1732 		ohdr = &hdr->u.oth;
1733 		hdrsize = 8 + 12;       /* LRH + BTH */
1734 	} else {
1735 		ohdr = &hdr->u.l.oth;
1736 		hdrsize = 8 + 40 + 12;  /* LRH + GRH + BTH */
1737 	}
1738 
1739 	opcode = be32_to_cpu(ohdr->bth[0]);
1740 	if (qib_ruc_check_hdr(ibp, hdr, has_grh, qp, opcode))
1741 		return;
1742 
1743 	psn = be32_to_cpu(ohdr->bth[2]);
1744 	opcode >>= 24;
1745 
1746 	/*
1747 	 * Process responses (ACKs) before anything else.  Note that the
1748 	 * packet sequence number will be for something in the send work
1749 	 * queue rather than the expected receive packet sequence number.
1750 	 * In other words, this QP is the requester.
1751 	 */
1752 	if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
1753 	    opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
1754 		qib_rc_rcv_resp(ibp, ohdr, data, tlen, qp, opcode, psn,
1755 				hdrsize, pmtu, rcd);
1756 		return;
1757 	}
1758 
1759 	/* Compute 24 bits worth of difference. */
1760 	diff = qib_cmp24(psn, qp->r_psn);
1761 	if (unlikely(diff)) {
1762 		if (qib_rc_rcv_error(ohdr, data, qp, opcode, psn, diff, rcd))
1763 			return;
1764 		goto send_ack;
1765 	}
1766 
1767 	/* Check for opcode sequence errors. */
1768 	switch (qp->r_state) {
1769 	case OP(SEND_FIRST):
1770 	case OP(SEND_MIDDLE):
1771 		if (opcode == OP(SEND_MIDDLE) ||
1772 		    opcode == OP(SEND_LAST) ||
1773 		    opcode == OP(SEND_LAST_WITH_IMMEDIATE))
1774 			break;
1775 		goto nack_inv;
1776 
1777 	case OP(RDMA_WRITE_FIRST):
1778 	case OP(RDMA_WRITE_MIDDLE):
1779 		if (opcode == OP(RDMA_WRITE_MIDDLE) ||
1780 		    opcode == OP(RDMA_WRITE_LAST) ||
1781 		    opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1782 			break;
1783 		goto nack_inv;
1784 
1785 	default:
1786 		if (opcode == OP(SEND_MIDDLE) ||
1787 		    opcode == OP(SEND_LAST) ||
1788 		    opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
1789 		    opcode == OP(RDMA_WRITE_MIDDLE) ||
1790 		    opcode == OP(RDMA_WRITE_LAST) ||
1791 		    opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1792 			goto nack_inv;
1793 		/*
1794 		 * Note that it is up to the requester to not send a new
1795 		 * RDMA read or atomic operation before receiving an ACK
1796 		 * for the previous operation.
1797 		 */
1798 		break;
1799 	}
1800 
1801 	if (qp->state == IB_QPS_RTR && !(qp->r_flags & RVT_R_COMM_EST))
1802 		rvt_comm_est(qp);
1803 
1804 	/* OK, process the packet. */
1805 	switch (opcode) {
1806 	case OP(SEND_FIRST):
1807 		ret = rvt_get_rwqe(qp, false);
1808 		if (ret < 0)
1809 			goto nack_op_err;
1810 		if (!ret)
1811 			goto rnr_nak;
1812 		qp->r_rcv_len = 0;
1813 		fallthrough;
1814 	case OP(SEND_MIDDLE):
1815 	case OP(RDMA_WRITE_MIDDLE):
1816 send_middle:
1817 		/* Check for invalid length PMTU or posted rwqe len. */
1818 		if (unlikely(tlen != (hdrsize + pmtu + 4)))
1819 			goto nack_inv;
1820 		qp->r_rcv_len += pmtu;
1821 		if (unlikely(qp->r_rcv_len > qp->r_len))
1822 			goto nack_inv;
1823 		rvt_copy_sge(qp, &qp->r_sge, data, pmtu, true, false);
1824 		break;
1825 
1826 	case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
1827 		/* consume RWQE */
1828 		ret = rvt_get_rwqe(qp, true);
1829 		if (ret < 0)
1830 			goto nack_op_err;
1831 		if (!ret)
1832 			goto rnr_nak;
1833 		goto send_last_imm;
1834 
1835 	case OP(SEND_ONLY):
1836 	case OP(SEND_ONLY_WITH_IMMEDIATE):
1837 		ret = rvt_get_rwqe(qp, false);
1838 		if (ret < 0)
1839 			goto nack_op_err;
1840 		if (!ret)
1841 			goto rnr_nak;
1842 		qp->r_rcv_len = 0;
1843 		if (opcode == OP(SEND_ONLY))
1844 			goto no_immediate_data;
1845 		fallthrough;	/* for SEND_ONLY_WITH_IMMEDIATE */
1846 	case OP(SEND_LAST_WITH_IMMEDIATE):
1847 send_last_imm:
1848 		wc.ex.imm_data = ohdr->u.imm_data;
1849 		hdrsize += 4;
1850 		wc.wc_flags = IB_WC_WITH_IMM;
1851 		goto send_last;
1852 	case OP(SEND_LAST):
1853 	case OP(RDMA_WRITE_LAST):
1854 no_immediate_data:
1855 		wc.wc_flags = 0;
1856 		wc.ex.imm_data = 0;
1857 send_last:
1858 		/* Get the number of bytes the message was padded by. */
1859 		pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1860 		/* Check for invalid length. */
1861 		/* XXX LAST len should be >= 1 */
1862 		if (unlikely(tlen < (hdrsize + pad + 4)))
1863 			goto nack_inv;
1864 		/* Don't count the CRC. */
1865 		tlen -= (hdrsize + pad + 4);
1866 		wc.byte_len = tlen + qp->r_rcv_len;
1867 		if (unlikely(wc.byte_len > qp->r_len))
1868 			goto nack_inv;
1869 		rvt_copy_sge(qp, &qp->r_sge, data, tlen, true, false);
1870 		rvt_put_ss(&qp->r_sge);
1871 		qp->r_msn++;
1872 		if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
1873 			break;
1874 		wc.wr_id = qp->r_wr_id;
1875 		wc.status = IB_WC_SUCCESS;
1876 		if (opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE) ||
1877 		    opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE))
1878 			wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
1879 		else
1880 			wc.opcode = IB_WC_RECV;
1881 		wc.qp = &qp->ibqp;
1882 		wc.src_qp = qp->remote_qpn;
1883 		wc.slid = rdma_ah_get_dlid(&qp->remote_ah_attr);
1884 		wc.sl = rdma_ah_get_sl(&qp->remote_ah_attr);
1885 		/* zero fields that are N/A */
1886 		wc.vendor_err = 0;
1887 		wc.pkey_index = 0;
1888 		wc.dlid_path_bits = 0;
1889 		wc.port_num = 0;
1890 		/* Signal completion event if the solicited bit is set. */
1891 		rvt_recv_cq(qp, &wc, ib_bth_is_solicited(ohdr));
1892 		break;
1893 
1894 	case OP(RDMA_WRITE_FIRST):
1895 	case OP(RDMA_WRITE_ONLY):
1896 	case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
1897 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
1898 			goto nack_inv;
1899 		/* consume RWQE */
1900 		reth = &ohdr->u.rc.reth;
1901 		hdrsize += sizeof(*reth);
1902 		qp->r_len = be32_to_cpu(reth->length);
1903 		qp->r_rcv_len = 0;
1904 		qp->r_sge.sg_list = NULL;
1905 		if (qp->r_len != 0) {
1906 			u32 rkey = be32_to_cpu(reth->rkey);
1907 			u64 vaddr = be64_to_cpu(reth->vaddr);
1908 			int ok;
1909 
1910 			/* Check rkey & NAK */
1911 			ok = rvt_rkey_ok(qp, &qp->r_sge.sge, qp->r_len, vaddr,
1912 					 rkey, IB_ACCESS_REMOTE_WRITE);
1913 			if (unlikely(!ok))
1914 				goto nack_acc;
1915 			qp->r_sge.num_sge = 1;
1916 		} else {
1917 			qp->r_sge.num_sge = 0;
1918 			qp->r_sge.sge.mr = NULL;
1919 			qp->r_sge.sge.vaddr = NULL;
1920 			qp->r_sge.sge.length = 0;
1921 			qp->r_sge.sge.sge_length = 0;
1922 		}
1923 		if (opcode == OP(RDMA_WRITE_FIRST))
1924 			goto send_middle;
1925 		else if (opcode == OP(RDMA_WRITE_ONLY))
1926 			goto no_immediate_data;
1927 		ret = rvt_get_rwqe(qp, true);
1928 		if (ret < 0)
1929 			goto nack_op_err;
1930 		if (!ret) {
1931 			rvt_put_ss(&qp->r_sge);
1932 			goto rnr_nak;
1933 		}
1934 		wc.ex.imm_data = ohdr->u.rc.imm_data;
1935 		hdrsize += 4;
1936 		wc.wc_flags = IB_WC_WITH_IMM;
1937 		goto send_last;
1938 
1939 	case OP(RDMA_READ_REQUEST): {
1940 		struct rvt_ack_entry *e;
1941 		u32 len;
1942 		u8 next;
1943 
1944 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
1945 			goto nack_inv;
1946 		next = qp->r_head_ack_queue + 1;
1947 		/* s_ack_queue is size QIB_MAX_RDMA_ATOMIC+1 so use > not >= */
1948 		if (next > QIB_MAX_RDMA_ATOMIC)
1949 			next = 0;
1950 		spin_lock_irqsave(&qp->s_lock, flags);
1951 		if (unlikely(next == qp->s_tail_ack_queue)) {
1952 			if (!qp->s_ack_queue[next].sent)
1953 				goto nack_inv_unlck;
1954 			qib_update_ack_queue(qp, next);
1955 		}
1956 		e = &qp->s_ack_queue[qp->r_head_ack_queue];
1957 		if (e->opcode == OP(RDMA_READ_REQUEST) && e->rdma_sge.mr) {
1958 			rvt_put_mr(e->rdma_sge.mr);
1959 			e->rdma_sge.mr = NULL;
1960 		}
1961 		reth = &ohdr->u.rc.reth;
1962 		len = be32_to_cpu(reth->length);
1963 		if (len) {
1964 			u32 rkey = be32_to_cpu(reth->rkey);
1965 			u64 vaddr = be64_to_cpu(reth->vaddr);
1966 			int ok;
1967 
1968 			/* Check rkey & NAK */
1969 			ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr,
1970 					 rkey, IB_ACCESS_REMOTE_READ);
1971 			if (unlikely(!ok))
1972 				goto nack_acc_unlck;
1973 			/*
1974 			 * Update the next expected PSN.  We add 1 later
1975 			 * below, so only add the remainder here.
1976 			 */
1977 			qp->r_psn += rvt_div_mtu(qp, len - 1);
1978 		} else {
1979 			e->rdma_sge.mr = NULL;
1980 			e->rdma_sge.vaddr = NULL;
1981 			e->rdma_sge.length = 0;
1982 			e->rdma_sge.sge_length = 0;
1983 		}
1984 		e->opcode = opcode;
1985 		e->sent = 0;
1986 		e->psn = psn;
1987 		e->lpsn = qp->r_psn;
1988 		/*
1989 		 * We need to increment the MSN here instead of when we
1990 		 * finish sending the result since a duplicate request would
1991 		 * increment it more than once.
1992 		 */
1993 		qp->r_msn++;
1994 		qp->r_psn++;
1995 		qp->r_state = opcode;
1996 		qp->r_nak_state = 0;
1997 		qp->r_head_ack_queue = next;
1998 
1999 		/* Schedule the send tasklet. */
2000 		qp->s_flags |= RVT_S_RESP_PENDING;
2001 		qib_schedule_send(qp);
2002 
2003 		goto sunlock;
2004 	}
2005 
2006 	case OP(COMPARE_SWAP):
2007 	case OP(FETCH_ADD): {
2008 		struct ib_atomic_eth *ateth;
2009 		struct rvt_ack_entry *e;
2010 		u64 vaddr;
2011 		atomic64_t *maddr;
2012 		u64 sdata;
2013 		u32 rkey;
2014 		u8 next;
2015 
2016 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
2017 			goto nack_inv;
2018 		next = qp->r_head_ack_queue + 1;
2019 		if (next > QIB_MAX_RDMA_ATOMIC)
2020 			next = 0;
2021 		spin_lock_irqsave(&qp->s_lock, flags);
2022 		if (unlikely(next == qp->s_tail_ack_queue)) {
2023 			if (!qp->s_ack_queue[next].sent)
2024 				goto nack_inv_unlck;
2025 			qib_update_ack_queue(qp, next);
2026 		}
2027 		e = &qp->s_ack_queue[qp->r_head_ack_queue];
2028 		if (e->opcode == OP(RDMA_READ_REQUEST) && e->rdma_sge.mr) {
2029 			rvt_put_mr(e->rdma_sge.mr);
2030 			e->rdma_sge.mr = NULL;
2031 		}
2032 		ateth = &ohdr->u.atomic_eth;
2033 		vaddr = get_ib_ateth_vaddr(ateth);
2034 		if (unlikely(vaddr & (sizeof(u64) - 1)))
2035 			goto nack_inv_unlck;
2036 		rkey = be32_to_cpu(ateth->rkey);
2037 		/* Check rkey & NAK */
2038 		if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64),
2039 					  vaddr, rkey,
2040 					  IB_ACCESS_REMOTE_ATOMIC)))
2041 			goto nack_acc_unlck;
2042 		/* Perform atomic OP and save result. */
2043 		maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
2044 		sdata = get_ib_ateth_swap(ateth);
2045 		e->atomic_data = (opcode == OP(FETCH_ADD)) ?
2046 			(u64) atomic64_add_return(sdata, maddr) - sdata :
2047 			(u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
2048 				      get_ib_ateth_compare(ateth),
2049 				      sdata);
2050 		rvt_put_mr(qp->r_sge.sge.mr);
2051 		qp->r_sge.num_sge = 0;
2052 		e->opcode = opcode;
2053 		e->sent = 0;
2054 		e->psn = psn;
2055 		e->lpsn = psn;
2056 		qp->r_msn++;
2057 		qp->r_psn++;
2058 		qp->r_state = opcode;
2059 		qp->r_nak_state = 0;
2060 		qp->r_head_ack_queue = next;
2061 
2062 		/* Schedule the send tasklet. */
2063 		qp->s_flags |= RVT_S_RESP_PENDING;
2064 		qib_schedule_send(qp);
2065 
2066 		goto sunlock;
2067 	}
2068 
2069 	default:
2070 		/* NAK unknown opcodes. */
2071 		goto nack_inv;
2072 	}
2073 	qp->r_psn++;
2074 	qp->r_state = opcode;
2075 	qp->r_ack_psn = psn;
2076 	qp->r_nak_state = 0;
2077 	/* Send an ACK if requested or required. */
2078 	if (psn & (1 << 31))
2079 		goto send_ack;
2080 	return;
2081 
2082 rnr_nak:
2083 	qp->r_nak_state = IB_RNR_NAK | qp->r_min_rnr_timer;
2084 	qp->r_ack_psn = qp->r_psn;
2085 	/* Queue RNR NAK for later */
2086 	if (list_empty(&qp->rspwait)) {
2087 		qp->r_flags |= RVT_R_RSP_NAK;
2088 		rvt_get_qp(qp);
2089 		list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
2090 	}
2091 	return;
2092 
2093 nack_op_err:
2094 	rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
2095 	qp->r_nak_state = IB_NAK_REMOTE_OPERATIONAL_ERROR;
2096 	qp->r_ack_psn = qp->r_psn;
2097 	/* Queue NAK for later */
2098 	if (list_empty(&qp->rspwait)) {
2099 		qp->r_flags |= RVT_R_RSP_NAK;
2100 		rvt_get_qp(qp);
2101 		list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
2102 	}
2103 	return;
2104 
2105 nack_inv_unlck:
2106 	spin_unlock_irqrestore(&qp->s_lock, flags);
2107 nack_inv:
2108 	rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
2109 	qp->r_nak_state = IB_NAK_INVALID_REQUEST;
2110 	qp->r_ack_psn = qp->r_psn;
2111 	/* Queue NAK for later */
2112 	if (list_empty(&qp->rspwait)) {
2113 		qp->r_flags |= RVT_R_RSP_NAK;
2114 		rvt_get_qp(qp);
2115 		list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
2116 	}
2117 	return;
2118 
2119 nack_acc_unlck:
2120 	spin_unlock_irqrestore(&qp->s_lock, flags);
2121 nack_acc:
2122 	rvt_rc_error(qp, IB_WC_LOC_PROT_ERR);
2123 	qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
2124 	qp->r_ack_psn = qp->r_psn;
2125 send_ack:
2126 	qib_send_rc_ack(qp);
2127 	return;
2128 
2129 sunlock:
2130 	spin_unlock_irqrestore(&qp->s_lock, flags);
2131 }
2132