xref: /linux/drivers/infiniband/hw/qib/qib_rc.c (revision bd628c1bed7902ec1f24ba0fe70758949146abbe)
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 	ss->sge = wqe->sg_list[0];
49 	ss->sg_list = wqe->sg_list + 1;
50 	ss->num_sge = wqe->wr.num_sge;
51 	ss->total_len = wqe->length;
52 	rvt_skip_sge(ss, len, false);
53 	return wqe->length - len;
54 }
55 
56 /**
57  * qib_make_rc_ack - construct a response packet (ACK, NAK, or RDMA read)
58  * @dev: the device for this QP
59  * @qp: a pointer to the QP
60  * @ohdr: a pointer to the IB header being constructed
61  * @pmtu: the path MTU
62  *
63  * Return 1 if constructed; otherwise, return 0.
64  * Note that we are in the responder's side of the QP context.
65  * Note the QP s_lock must be held.
66  */
67 static int qib_make_rc_ack(struct qib_ibdev *dev, struct rvt_qp *qp,
68 			   struct ib_other_headers *ohdr, u32 pmtu)
69 {
70 	struct rvt_ack_entry *e;
71 	u32 hwords;
72 	u32 len;
73 	u32 bth0;
74 	u32 bth2;
75 
76 	/* Don't send an ACK if we aren't supposed to. */
77 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
78 		goto bail;
79 
80 	/* header size in 32-bit words LRH+BTH = (8+12)/4. */
81 	hwords = 5;
82 
83 	switch (qp->s_ack_state) {
84 	case OP(RDMA_READ_RESPONSE_LAST):
85 	case OP(RDMA_READ_RESPONSE_ONLY):
86 		e = &qp->s_ack_queue[qp->s_tail_ack_queue];
87 		if (e->rdma_sge.mr) {
88 			rvt_put_mr(e->rdma_sge.mr);
89 			e->rdma_sge.mr = NULL;
90 		}
91 		/* FALLTHROUGH */
92 	case OP(ATOMIC_ACKNOWLEDGE):
93 		/*
94 		 * We can increment the tail pointer now that the last
95 		 * response has been sent instead of only being
96 		 * constructed.
97 		 */
98 		if (++qp->s_tail_ack_queue > QIB_MAX_RDMA_ATOMIC)
99 			qp->s_tail_ack_queue = 0;
100 		/* FALLTHROUGH */
101 	case OP(SEND_ONLY):
102 	case OP(ACKNOWLEDGE):
103 		/* Check for no next entry in the queue. */
104 		if (qp->r_head_ack_queue == qp->s_tail_ack_queue) {
105 			if (qp->s_flags & RVT_S_ACK_PENDING)
106 				goto normal;
107 			goto bail;
108 		}
109 
110 		e = &qp->s_ack_queue[qp->s_tail_ack_queue];
111 		if (e->opcode == OP(RDMA_READ_REQUEST)) {
112 			/*
113 			 * If a RDMA read response is being resent and
114 			 * we haven't seen the duplicate request yet,
115 			 * then stop sending the remaining responses the
116 			 * responder has seen until the requester resends it.
117 			 */
118 			len = e->rdma_sge.sge_length;
119 			if (len && !e->rdma_sge.mr) {
120 				qp->s_tail_ack_queue = qp->r_head_ack_queue;
121 				goto bail;
122 			}
123 			/* Copy SGE state in case we need to resend */
124 			qp->s_rdma_mr = e->rdma_sge.mr;
125 			if (qp->s_rdma_mr)
126 				rvt_get_mr(qp->s_rdma_mr);
127 			qp->s_ack_rdma_sge.sge = e->rdma_sge;
128 			qp->s_ack_rdma_sge.num_sge = 1;
129 			qp->s_cur_sge = &qp->s_ack_rdma_sge;
130 			if (len > pmtu) {
131 				len = pmtu;
132 				qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST);
133 			} else {
134 				qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY);
135 				e->sent = 1;
136 			}
137 			ohdr->u.aeth = rvt_compute_aeth(qp);
138 			hwords++;
139 			qp->s_ack_rdma_psn = e->psn;
140 			bth2 = qp->s_ack_rdma_psn++ & QIB_PSN_MASK;
141 		} else {
142 			/* COMPARE_SWAP or FETCH_ADD */
143 			qp->s_cur_sge = NULL;
144 			len = 0;
145 			qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE);
146 			ohdr->u.at.aeth = rvt_compute_aeth(qp);
147 			ib_u64_put(e->atomic_data, &ohdr->u.at.atomic_ack_eth);
148 			hwords += sizeof(ohdr->u.at) / sizeof(u32);
149 			bth2 = e->psn & QIB_PSN_MASK;
150 			e->sent = 1;
151 		}
152 		bth0 = qp->s_ack_state << 24;
153 		break;
154 
155 	case OP(RDMA_READ_RESPONSE_FIRST):
156 		qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE);
157 		/* FALLTHROUGH */
158 	case OP(RDMA_READ_RESPONSE_MIDDLE):
159 		qp->s_cur_sge = &qp->s_ack_rdma_sge;
160 		qp->s_rdma_mr = qp->s_ack_rdma_sge.sge.mr;
161 		if (qp->s_rdma_mr)
162 			rvt_get_mr(qp->s_rdma_mr);
163 		len = qp->s_ack_rdma_sge.sge.sge_length;
164 		if (len > pmtu)
165 			len = pmtu;
166 		else {
167 			ohdr->u.aeth = rvt_compute_aeth(qp);
168 			hwords++;
169 			qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
170 			e = &qp->s_ack_queue[qp->s_tail_ack_queue];
171 			e->sent = 1;
172 		}
173 		bth0 = qp->s_ack_state << 24;
174 		bth2 = qp->s_ack_rdma_psn++ & QIB_PSN_MASK;
175 		break;
176 
177 	default:
178 normal:
179 		/*
180 		 * Send a regular ACK.
181 		 * Set the s_ack_state so we wait until after sending
182 		 * the ACK before setting s_ack_state to ACKNOWLEDGE
183 		 * (see above).
184 		 */
185 		qp->s_ack_state = OP(SEND_ONLY);
186 		qp->s_flags &= ~RVT_S_ACK_PENDING;
187 		qp->s_cur_sge = NULL;
188 		if (qp->s_nak_state)
189 			ohdr->u.aeth =
190 				cpu_to_be32((qp->r_msn & IB_MSN_MASK) |
191 					    (qp->s_nak_state <<
192 					     IB_AETH_CREDIT_SHIFT));
193 		else
194 			ohdr->u.aeth = rvt_compute_aeth(qp);
195 		hwords++;
196 		len = 0;
197 		bth0 = OP(ACKNOWLEDGE) << 24;
198 		bth2 = qp->s_ack_psn & QIB_PSN_MASK;
199 	}
200 	qp->s_rdma_ack_cnt++;
201 	qp->s_hdrwords = hwords;
202 	qp->s_cur_size = len;
203 	qib_make_ruc_header(qp, ohdr, bth0, bth2);
204 	return 1;
205 
206 bail:
207 	qp->s_ack_state = OP(ACKNOWLEDGE);
208 	qp->s_flags &= ~(RVT_S_RESP_PENDING | RVT_S_ACK_PENDING);
209 	return 0;
210 }
211 
212 /**
213  * qib_make_rc_req - construct a request packet (SEND, RDMA r/w, ATOMIC)
214  * @qp: a pointer to the QP
215  *
216  * Assumes the s_lock is held.
217  *
218  * Return 1 if constructed; otherwise, return 0.
219  */
220 int qib_make_rc_req(struct rvt_qp *qp, unsigned long *flags)
221 {
222 	struct qib_qp_priv *priv = qp->priv;
223 	struct qib_ibdev *dev = to_idev(qp->ibqp.device);
224 	struct ib_other_headers *ohdr;
225 	struct rvt_sge_state *ss;
226 	struct rvt_swqe *wqe;
227 	u32 hwords;
228 	u32 len;
229 	u32 bth0;
230 	u32 bth2;
231 	u32 pmtu = qp->pmtu;
232 	char newreq;
233 	int ret = 0;
234 	int delta;
235 
236 	ohdr = &priv->s_hdr->u.oth;
237 	if (rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH)
238 		ohdr = &priv->s_hdr->u.l.oth;
239 
240 	/* Sending responses has higher priority over sending requests. */
241 	if ((qp->s_flags & RVT_S_RESP_PENDING) &&
242 	    qib_make_rc_ack(dev, qp, ohdr, pmtu))
243 		goto done;
244 
245 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_SEND_OK)) {
246 		if (!(ib_rvt_state_ops[qp->state] & RVT_FLUSH_SEND))
247 			goto bail;
248 		/* We are in the error state, flush the work request. */
249 		if (qp->s_last == READ_ONCE(qp->s_head))
250 			goto bail;
251 		/* If DMAs are in progress, we can't flush immediately. */
252 		if (atomic_read(&priv->s_dma_busy)) {
253 			qp->s_flags |= RVT_S_WAIT_DMA;
254 			goto bail;
255 		}
256 		wqe = rvt_get_swqe_ptr(qp, qp->s_last);
257 		rvt_send_complete(qp, wqe, qp->s_last != qp->s_acked ?
258 			IB_WC_SUCCESS : IB_WC_WR_FLUSH_ERR);
259 		/* will get called again */
260 		goto done;
261 	}
262 
263 	if (qp->s_flags & (RVT_S_WAIT_RNR | RVT_S_WAIT_ACK))
264 		goto bail;
265 
266 	if (qib_cmp24(qp->s_psn, qp->s_sending_hpsn) <= 0) {
267 		if (qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) <= 0) {
268 			qp->s_flags |= RVT_S_WAIT_PSN;
269 			goto bail;
270 		}
271 		qp->s_sending_psn = qp->s_psn;
272 		qp->s_sending_hpsn = qp->s_psn - 1;
273 	}
274 
275 	/* header size in 32-bit words LRH+BTH = (8+12)/4. */
276 	hwords = 5;
277 	bth0 = 0;
278 
279 	/* Send a request. */
280 	wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
281 	switch (qp->s_state) {
282 	default:
283 		if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_NEXT_SEND_OK))
284 			goto bail;
285 		/*
286 		 * Resend an old request or start a new one.
287 		 *
288 		 * We keep track of the current SWQE so that
289 		 * we don't reset the "furthest progress" state
290 		 * if we need to back up.
291 		 */
292 		newreq = 0;
293 		if (qp->s_cur == qp->s_tail) {
294 			/* Check if send work queue is empty. */
295 			if (qp->s_tail == READ_ONCE(qp->s_head))
296 				goto bail;
297 			/*
298 			 * If a fence is requested, wait for previous
299 			 * RDMA read and atomic operations to finish.
300 			 */
301 			if ((wqe->wr.send_flags & IB_SEND_FENCE) &&
302 			    qp->s_num_rd_atomic) {
303 				qp->s_flags |= RVT_S_WAIT_FENCE;
304 				goto bail;
305 			}
306 			newreq = 1;
307 			qp->s_psn = wqe->psn;
308 		}
309 		/*
310 		 * Note that we have to be careful not to modify the
311 		 * original work request since we may need to resend
312 		 * it.
313 		 */
314 		len = wqe->length;
315 		ss = &qp->s_sge;
316 		bth2 = qp->s_psn & QIB_PSN_MASK;
317 		switch (wqe->wr.opcode) {
318 		case IB_WR_SEND:
319 		case IB_WR_SEND_WITH_IMM:
320 			/* If no credit, return. */
321 			if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT) &&
322 			    rvt_cmp_msn(wqe->ssn, qp->s_lsn + 1) > 0) {
323 				qp->s_flags |= RVT_S_WAIT_SSN_CREDIT;
324 				goto bail;
325 			}
326 			if (len > pmtu) {
327 				qp->s_state = OP(SEND_FIRST);
328 				len = pmtu;
329 				break;
330 			}
331 			if (wqe->wr.opcode == IB_WR_SEND)
332 				qp->s_state = OP(SEND_ONLY);
333 			else {
334 				qp->s_state = OP(SEND_ONLY_WITH_IMMEDIATE);
335 				/* Immediate data comes after the BTH */
336 				ohdr->u.imm_data = wqe->wr.ex.imm_data;
337 				hwords += 1;
338 			}
339 			if (wqe->wr.send_flags & IB_SEND_SOLICITED)
340 				bth0 |= IB_BTH_SOLICITED;
341 			bth2 |= IB_BTH_REQ_ACK;
342 			if (++qp->s_cur == qp->s_size)
343 				qp->s_cur = 0;
344 			break;
345 
346 		case IB_WR_RDMA_WRITE:
347 			if (newreq && !(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
348 				qp->s_lsn++;
349 			goto no_flow_control;
350 		case IB_WR_RDMA_WRITE_WITH_IMM:
351 			/* If no credit, return. */
352 			if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT) &&
353 			    rvt_cmp_msn(wqe->ssn, qp->s_lsn + 1) > 0) {
354 				qp->s_flags |= RVT_S_WAIT_SSN_CREDIT;
355 				goto bail;
356 			}
357 no_flow_control:
358 			ohdr->u.rc.reth.vaddr =
359 				cpu_to_be64(wqe->rdma_wr.remote_addr);
360 			ohdr->u.rc.reth.rkey =
361 				cpu_to_be32(wqe->rdma_wr.rkey);
362 			ohdr->u.rc.reth.length = cpu_to_be32(len);
363 			hwords += sizeof(struct ib_reth) / sizeof(u32);
364 			if (len > pmtu) {
365 				qp->s_state = OP(RDMA_WRITE_FIRST);
366 				len = pmtu;
367 				break;
368 			}
369 			if (wqe->rdma_wr.wr.opcode == IB_WR_RDMA_WRITE)
370 				qp->s_state = OP(RDMA_WRITE_ONLY);
371 			else {
372 				qp->s_state = OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
373 				/* Immediate data comes after RETH */
374 				ohdr->u.rc.imm_data =
375 					wqe->rdma_wr.wr.ex.imm_data;
376 				hwords += 1;
377 				if (wqe->rdma_wr.wr.send_flags & IB_SEND_SOLICITED)
378 					bth0 |= IB_BTH_SOLICITED;
379 			}
380 			bth2 |= IB_BTH_REQ_ACK;
381 			if (++qp->s_cur == qp->s_size)
382 				qp->s_cur = 0;
383 			break;
384 
385 		case IB_WR_RDMA_READ:
386 			/*
387 			 * Don't allow more operations to be started
388 			 * than the QP limits allow.
389 			 */
390 			if (newreq) {
391 				if (qp->s_num_rd_atomic >=
392 				    qp->s_max_rd_atomic) {
393 					qp->s_flags |= RVT_S_WAIT_RDMAR;
394 					goto bail;
395 				}
396 				qp->s_num_rd_atomic++;
397 				if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
398 					qp->s_lsn++;
399 			}
400 
401 			ohdr->u.rc.reth.vaddr =
402 				cpu_to_be64(wqe->rdma_wr.remote_addr);
403 			ohdr->u.rc.reth.rkey =
404 				cpu_to_be32(wqe->rdma_wr.rkey);
405 			ohdr->u.rc.reth.length = cpu_to_be32(len);
406 			qp->s_state = OP(RDMA_READ_REQUEST);
407 			hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
408 			ss = NULL;
409 			len = 0;
410 			bth2 |= IB_BTH_REQ_ACK;
411 			if (++qp->s_cur == qp->s_size)
412 				qp->s_cur = 0;
413 			break;
414 
415 		case IB_WR_ATOMIC_CMP_AND_SWP:
416 		case IB_WR_ATOMIC_FETCH_AND_ADD:
417 			/*
418 			 * Don't allow more operations to be started
419 			 * than the QP limits allow.
420 			 */
421 			if (newreq) {
422 				if (qp->s_num_rd_atomic >=
423 				    qp->s_max_rd_atomic) {
424 					qp->s_flags |= RVT_S_WAIT_RDMAR;
425 					goto bail;
426 				}
427 				qp->s_num_rd_atomic++;
428 				if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
429 					qp->s_lsn++;
430 			}
431 			if (wqe->atomic_wr.wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
432 				qp->s_state = OP(COMPARE_SWAP);
433 				put_ib_ateth_swap(wqe->atomic_wr.swap,
434 						  &ohdr->u.atomic_eth);
435 				put_ib_ateth_compare(wqe->atomic_wr.compare_add,
436 						     &ohdr->u.atomic_eth);
437 			} else {
438 				qp->s_state = OP(FETCH_ADD);
439 				put_ib_ateth_swap(wqe->atomic_wr.compare_add,
440 						  &ohdr->u.atomic_eth);
441 				put_ib_ateth_compare(0, &ohdr->u.atomic_eth);
442 			}
443 			put_ib_ateth_vaddr(wqe->atomic_wr.remote_addr,
444 					   &ohdr->u.atomic_eth);
445 			ohdr->u.atomic_eth.rkey = cpu_to_be32(
446 				wqe->atomic_wr.rkey);
447 			hwords += sizeof(struct ib_atomic_eth) / sizeof(u32);
448 			ss = NULL;
449 			len = 0;
450 			bth2 |= IB_BTH_REQ_ACK;
451 			if (++qp->s_cur == qp->s_size)
452 				qp->s_cur = 0;
453 			break;
454 
455 		default:
456 			goto bail;
457 		}
458 		qp->s_sge.sge = wqe->sg_list[0];
459 		qp->s_sge.sg_list = wqe->sg_list + 1;
460 		qp->s_sge.num_sge = wqe->wr.num_sge;
461 		qp->s_sge.total_len = wqe->length;
462 		qp->s_len = wqe->length;
463 		if (newreq) {
464 			qp->s_tail++;
465 			if (qp->s_tail >= qp->s_size)
466 				qp->s_tail = 0;
467 		}
468 		if (wqe->wr.opcode == IB_WR_RDMA_READ)
469 			qp->s_psn = wqe->lpsn + 1;
470 		else
471 			qp->s_psn++;
472 		break;
473 
474 	case OP(RDMA_READ_RESPONSE_FIRST):
475 		/*
476 		 * qp->s_state is normally set to the opcode of the
477 		 * last packet constructed for new requests and therefore
478 		 * is never set to RDMA read response.
479 		 * RDMA_READ_RESPONSE_FIRST is used by the ACK processing
480 		 * thread to indicate a SEND needs to be restarted from an
481 		 * earlier PSN without interferring with the sending thread.
482 		 * See qib_restart_rc().
483 		 */
484 		qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu);
485 		/* FALLTHROUGH */
486 	case OP(SEND_FIRST):
487 		qp->s_state = OP(SEND_MIDDLE);
488 		/* FALLTHROUGH */
489 	case OP(SEND_MIDDLE):
490 		bth2 = qp->s_psn++ & QIB_PSN_MASK;
491 		ss = &qp->s_sge;
492 		len = qp->s_len;
493 		if (len > pmtu) {
494 			len = pmtu;
495 			break;
496 		}
497 		if (wqe->wr.opcode == IB_WR_SEND)
498 			qp->s_state = OP(SEND_LAST);
499 		else {
500 			qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
501 			/* Immediate data comes after the BTH */
502 			ohdr->u.imm_data = wqe->wr.ex.imm_data;
503 			hwords += 1;
504 		}
505 		if (wqe->wr.send_flags & IB_SEND_SOLICITED)
506 			bth0 |= IB_BTH_SOLICITED;
507 		bth2 |= IB_BTH_REQ_ACK;
508 		qp->s_cur++;
509 		if (qp->s_cur >= qp->s_size)
510 			qp->s_cur = 0;
511 		break;
512 
513 	case OP(RDMA_READ_RESPONSE_LAST):
514 		/*
515 		 * qp->s_state is normally set to the opcode of the
516 		 * last packet constructed for new requests and therefore
517 		 * is never set to RDMA read response.
518 		 * RDMA_READ_RESPONSE_LAST is used by the ACK processing
519 		 * thread to indicate a RDMA write needs to be restarted from
520 		 * an earlier PSN without interferring with the sending thread.
521 		 * See qib_restart_rc().
522 		 */
523 		qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu);
524 		/* FALLTHROUGH */
525 	case OP(RDMA_WRITE_FIRST):
526 		qp->s_state = OP(RDMA_WRITE_MIDDLE);
527 		/* FALLTHROUGH */
528 	case OP(RDMA_WRITE_MIDDLE):
529 		bth2 = qp->s_psn++ & QIB_PSN_MASK;
530 		ss = &qp->s_sge;
531 		len = qp->s_len;
532 		if (len > pmtu) {
533 			len = pmtu;
534 			break;
535 		}
536 		if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
537 			qp->s_state = OP(RDMA_WRITE_LAST);
538 		else {
539 			qp->s_state = OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
540 			/* Immediate data comes after the BTH */
541 			ohdr->u.imm_data = wqe->wr.ex.imm_data;
542 			hwords += 1;
543 			if (wqe->wr.send_flags & IB_SEND_SOLICITED)
544 				bth0 |= IB_BTH_SOLICITED;
545 		}
546 		bth2 |= IB_BTH_REQ_ACK;
547 		qp->s_cur++;
548 		if (qp->s_cur >= qp->s_size)
549 			qp->s_cur = 0;
550 		break;
551 
552 	case OP(RDMA_READ_RESPONSE_MIDDLE):
553 		/*
554 		 * qp->s_state is normally set to the opcode of the
555 		 * last packet constructed for new requests and therefore
556 		 * is never set to RDMA read response.
557 		 * RDMA_READ_RESPONSE_MIDDLE is used by the ACK processing
558 		 * thread to indicate a RDMA read needs to be restarted from
559 		 * an earlier PSN without interferring with the sending thread.
560 		 * See qib_restart_rc().
561 		 */
562 		len = ((qp->s_psn - wqe->psn) & QIB_PSN_MASK) * pmtu;
563 		ohdr->u.rc.reth.vaddr =
564 			cpu_to_be64(wqe->rdma_wr.remote_addr + len);
565 		ohdr->u.rc.reth.rkey =
566 			cpu_to_be32(wqe->rdma_wr.rkey);
567 		ohdr->u.rc.reth.length = cpu_to_be32(wqe->length - len);
568 		qp->s_state = OP(RDMA_READ_REQUEST);
569 		hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
570 		bth2 = (qp->s_psn & QIB_PSN_MASK) | IB_BTH_REQ_ACK;
571 		qp->s_psn = wqe->lpsn + 1;
572 		ss = NULL;
573 		len = 0;
574 		qp->s_cur++;
575 		if (qp->s_cur == qp->s_size)
576 			qp->s_cur = 0;
577 		break;
578 	}
579 	qp->s_sending_hpsn = bth2;
580 	delta = (((int) bth2 - (int) wqe->psn) << 8) >> 8;
581 	if (delta && delta % QIB_PSN_CREDIT == 0)
582 		bth2 |= IB_BTH_REQ_ACK;
583 	if (qp->s_flags & RVT_S_SEND_ONE) {
584 		qp->s_flags &= ~RVT_S_SEND_ONE;
585 		qp->s_flags |= RVT_S_WAIT_ACK;
586 		bth2 |= IB_BTH_REQ_ACK;
587 	}
588 	qp->s_len -= len;
589 	qp->s_hdrwords = hwords;
590 	qp->s_cur_sge = ss;
591 	qp->s_cur_size = len;
592 	qib_make_ruc_header(qp, ohdr, bth0 | (qp->s_state << 24), bth2);
593 done:
594 	return 1;
595 bail:
596 	qp->s_flags &= ~RVT_S_BUSY;
597 	return ret;
598 }
599 
600 /**
601  * qib_send_rc_ack - Construct an ACK packet and send it
602  * @qp: a pointer to the QP
603  *
604  * This is called from qib_rc_rcv() and qib_kreceive().
605  * Note that RDMA reads and atomics are handled in the
606  * send side QP state and tasklet.
607  */
608 void qib_send_rc_ack(struct rvt_qp *qp)
609 {
610 	struct qib_devdata *dd = dd_from_ibdev(qp->ibqp.device);
611 	struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
612 	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
613 	u64 pbc;
614 	u16 lrh0;
615 	u32 bth0;
616 	u32 hwords;
617 	u32 pbufn;
618 	u32 __iomem *piobuf;
619 	struct ib_header hdr;
620 	struct ib_other_headers *ohdr;
621 	u32 control;
622 	unsigned long flags;
623 
624 	spin_lock_irqsave(&qp->s_lock, flags);
625 
626 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
627 		goto unlock;
628 
629 	/* Don't send ACK or NAK if a RDMA read or atomic is pending. */
630 	if ((qp->s_flags & RVT_S_RESP_PENDING) || qp->s_rdma_ack_cnt)
631 		goto queue_ack;
632 
633 	/* Construct the header with s_lock held so APM doesn't change it. */
634 	ohdr = &hdr.u.oth;
635 	lrh0 = QIB_LRH_BTH;
636 	/* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4. */
637 	hwords = 6;
638 	if (unlikely(rdma_ah_get_ah_flags(&qp->remote_ah_attr) &
639 		     IB_AH_GRH)) {
640 		hwords += qib_make_grh(ibp, &hdr.u.l.grh,
641 				       rdma_ah_read_grh(&qp->remote_ah_attr),
642 				       hwords, 0);
643 		ohdr = &hdr.u.l.oth;
644 		lrh0 = QIB_LRH_GRH;
645 	}
646 	/* read pkey_index w/o lock (its atomic) */
647 	bth0 = qib_get_pkey(ibp, qp->s_pkey_index) | (OP(ACKNOWLEDGE) << 24);
648 	if (qp->s_mig_state == IB_MIG_MIGRATED)
649 		bth0 |= IB_BTH_MIG_REQ;
650 	if (qp->r_nak_state)
651 		ohdr->u.aeth = cpu_to_be32((qp->r_msn & IB_MSN_MASK) |
652 					    (qp->r_nak_state <<
653 					     IB_AETH_CREDIT_SHIFT));
654 	else
655 		ohdr->u.aeth = rvt_compute_aeth(qp);
656 	lrh0 |= ibp->sl_to_vl[rdma_ah_get_sl(&qp->remote_ah_attr)] << 12 |
657 		rdma_ah_get_sl(&qp->remote_ah_attr) << 4;
658 	hdr.lrh[0] = cpu_to_be16(lrh0);
659 	hdr.lrh[1] = cpu_to_be16(rdma_ah_get_dlid(&qp->remote_ah_attr));
660 	hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC);
661 	hdr.lrh[3] = cpu_to_be16(ppd->lid |
662 				 rdma_ah_get_path_bits(&qp->remote_ah_attr));
663 	ohdr->bth[0] = cpu_to_be32(bth0);
664 	ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
665 	ohdr->bth[2] = cpu_to_be32(qp->r_ack_psn & QIB_PSN_MASK);
666 
667 	spin_unlock_irqrestore(&qp->s_lock, flags);
668 
669 	/* Don't try to send ACKs if the link isn't ACTIVE */
670 	if (!(ppd->lflags & QIBL_LINKACTIVE))
671 		goto done;
672 
673 	control = dd->f_setpbc_control(ppd, hwords + SIZE_OF_CRC,
674 				       qp->s_srate, lrh0 >> 12);
675 	/* length is + 1 for the control dword */
676 	pbc = ((u64) control << 32) | (hwords + 1);
677 
678 	piobuf = dd->f_getsendbuf(ppd, pbc, &pbufn);
679 	if (!piobuf) {
680 		/*
681 		 * We are out of PIO buffers at the moment.
682 		 * Pass responsibility for sending the ACK to the
683 		 * send tasklet so that when a PIO buffer becomes
684 		 * available, the ACK is sent ahead of other outgoing
685 		 * packets.
686 		 */
687 		spin_lock_irqsave(&qp->s_lock, flags);
688 		goto queue_ack;
689 	}
690 
691 	/*
692 	 * Write the pbc.
693 	 * We have to flush after the PBC for correctness
694 	 * on some cpus or WC buffer can be written out of order.
695 	 */
696 	writeq(pbc, piobuf);
697 
698 	if (dd->flags & QIB_PIO_FLUSH_WC) {
699 		u32 *hdrp = (u32 *) &hdr;
700 
701 		qib_flush_wc();
702 		qib_pio_copy(piobuf + 2, hdrp, hwords - 1);
703 		qib_flush_wc();
704 		__raw_writel(hdrp[hwords - 1], piobuf + hwords + 1);
705 	} else
706 		qib_pio_copy(piobuf + 2, (u32 *) &hdr, hwords);
707 
708 	if (dd->flags & QIB_USE_SPCL_TRIG) {
709 		u32 spcl_off = (pbufn >= dd->piobcnt2k) ? 2047 : 1023;
710 
711 		qib_flush_wc();
712 		__raw_writel(0xaebecede, piobuf + spcl_off);
713 	}
714 
715 	qib_flush_wc();
716 	qib_sendbuf_done(dd, pbufn);
717 
718 	this_cpu_inc(ibp->pmastats->n_unicast_xmit);
719 	goto done;
720 
721 queue_ack:
722 	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
723 		this_cpu_inc(*ibp->rvp.rc_qacks);
724 		qp->s_flags |= RVT_S_ACK_PENDING | RVT_S_RESP_PENDING;
725 		qp->s_nak_state = qp->r_nak_state;
726 		qp->s_ack_psn = qp->r_ack_psn;
727 
728 		/* Schedule the send tasklet. */
729 		qib_schedule_send(qp);
730 	}
731 unlock:
732 	spin_unlock_irqrestore(&qp->s_lock, flags);
733 done:
734 	return;
735 }
736 
737 /**
738  * reset_psn - reset the QP state to send starting from PSN
739  * @qp: the QP
740  * @psn: the packet sequence number to restart at
741  *
742  * This is called from qib_rc_rcv() to process an incoming RC ACK
743  * for the given QP.
744  * Called at interrupt level with the QP s_lock held.
745  */
746 static void reset_psn(struct rvt_qp *qp, u32 psn)
747 {
748 	u32 n = qp->s_acked;
749 	struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, n);
750 	u32 opcode;
751 
752 	qp->s_cur = n;
753 
754 	/*
755 	 * If we are starting the request from the beginning,
756 	 * let the normal send code handle initialization.
757 	 */
758 	if (qib_cmp24(psn, wqe->psn) <= 0) {
759 		qp->s_state = OP(SEND_LAST);
760 		goto done;
761 	}
762 
763 	/* Find the work request opcode corresponding to the given PSN. */
764 	opcode = wqe->wr.opcode;
765 	for (;;) {
766 		int diff;
767 
768 		if (++n == qp->s_size)
769 			n = 0;
770 		if (n == qp->s_tail)
771 			break;
772 		wqe = rvt_get_swqe_ptr(qp, n);
773 		diff = qib_cmp24(psn, wqe->psn);
774 		if (diff < 0)
775 			break;
776 		qp->s_cur = n;
777 		/*
778 		 * If we are starting the request from the beginning,
779 		 * let the normal send code handle initialization.
780 		 */
781 		if (diff == 0) {
782 			qp->s_state = OP(SEND_LAST);
783 			goto done;
784 		}
785 		opcode = wqe->wr.opcode;
786 	}
787 
788 	/*
789 	 * Set the state to restart in the middle of a request.
790 	 * Don't change the s_sge, s_cur_sge, or s_cur_size.
791 	 * See qib_make_rc_req().
792 	 */
793 	switch (opcode) {
794 	case IB_WR_SEND:
795 	case IB_WR_SEND_WITH_IMM:
796 		qp->s_state = OP(RDMA_READ_RESPONSE_FIRST);
797 		break;
798 
799 	case IB_WR_RDMA_WRITE:
800 	case IB_WR_RDMA_WRITE_WITH_IMM:
801 		qp->s_state = OP(RDMA_READ_RESPONSE_LAST);
802 		break;
803 
804 	case IB_WR_RDMA_READ:
805 		qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE);
806 		break;
807 
808 	default:
809 		/*
810 		 * This case shouldn't happen since its only
811 		 * one PSN per req.
812 		 */
813 		qp->s_state = OP(SEND_LAST);
814 	}
815 done:
816 	qp->s_psn = psn;
817 	/*
818 	 * Set RVT_S_WAIT_PSN as qib_rc_complete() may start the timer
819 	 * asynchronously before the send tasklet can get scheduled.
820 	 * Doing it in qib_make_rc_req() is too late.
821 	 */
822 	if ((qib_cmp24(qp->s_psn, qp->s_sending_hpsn) <= 0) &&
823 	    (qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) <= 0))
824 		qp->s_flags |= RVT_S_WAIT_PSN;
825 }
826 
827 /*
828  * Back up requester to resend the last un-ACKed request.
829  * The QP r_lock and s_lock should be held and interrupts disabled.
830  */
831 void qib_restart_rc(struct rvt_qp *qp, u32 psn, int wait)
832 {
833 	struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
834 	struct qib_ibport *ibp;
835 
836 	if (qp->s_retry == 0) {
837 		if (qp->s_mig_state == IB_MIG_ARMED) {
838 			qib_migrate_qp(qp);
839 			qp->s_retry = qp->s_retry_cnt;
840 		} else if (qp->s_last == qp->s_acked) {
841 			rvt_send_complete(qp, wqe, IB_WC_RETRY_EXC_ERR);
842 			rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
843 			return;
844 		} else /* XXX need to handle delayed completion */
845 			return;
846 	} else
847 		qp->s_retry--;
848 
849 	ibp = to_iport(qp->ibqp.device, qp->port_num);
850 	if (wqe->wr.opcode == IB_WR_RDMA_READ)
851 		ibp->rvp.n_rc_resends++;
852 	else
853 		ibp->rvp.n_rc_resends += (qp->s_psn - psn) & QIB_PSN_MASK;
854 
855 	qp->s_flags &= ~(RVT_S_WAIT_FENCE | RVT_S_WAIT_RDMAR |
856 			 RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_PSN |
857 			 RVT_S_WAIT_ACK);
858 	if (wait)
859 		qp->s_flags |= RVT_S_SEND_ONE;
860 	reset_psn(qp, psn);
861 }
862 
863 /*
864  * Set qp->s_sending_psn to the next PSN after the given one.
865  * This would be psn+1 except when RDMA reads are present.
866  */
867 static void reset_sending_psn(struct rvt_qp *qp, u32 psn)
868 {
869 	struct rvt_swqe *wqe;
870 	u32 n = qp->s_last;
871 
872 	/* Find the work request corresponding to the given PSN. */
873 	for (;;) {
874 		wqe = rvt_get_swqe_ptr(qp, n);
875 		if (qib_cmp24(psn, wqe->lpsn) <= 0) {
876 			if (wqe->wr.opcode == IB_WR_RDMA_READ)
877 				qp->s_sending_psn = wqe->lpsn + 1;
878 			else
879 				qp->s_sending_psn = psn + 1;
880 			break;
881 		}
882 		if (++n == qp->s_size)
883 			n = 0;
884 		if (n == qp->s_tail)
885 			break;
886 	}
887 }
888 
889 /*
890  * This should be called with the QP s_lock held and interrupts disabled.
891  */
892 void qib_rc_send_complete(struct rvt_qp *qp, struct ib_header *hdr)
893 {
894 	struct ib_other_headers *ohdr;
895 	struct rvt_swqe *wqe;
896 	u32 opcode;
897 	u32 psn;
898 
899 	if (!(ib_rvt_state_ops[qp->state] & RVT_SEND_OR_FLUSH_OR_RECV_OK))
900 		return;
901 
902 	/* Find out where the BTH is */
903 	if ((be16_to_cpu(hdr->lrh[0]) & 3) == QIB_LRH_BTH)
904 		ohdr = &hdr->u.oth;
905 	else
906 		ohdr = &hdr->u.l.oth;
907 
908 	opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
909 	if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
910 	    opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
911 		WARN_ON(!qp->s_rdma_ack_cnt);
912 		qp->s_rdma_ack_cnt--;
913 		return;
914 	}
915 
916 	psn = be32_to_cpu(ohdr->bth[2]);
917 	reset_sending_psn(qp, psn);
918 
919 	/*
920 	 * Start timer after a packet requesting an ACK has been sent and
921 	 * there are still requests that haven't been acked.
922 	 */
923 	if ((psn & IB_BTH_REQ_ACK) && qp->s_acked != qp->s_tail &&
924 	    !(qp->s_flags & (RVT_S_TIMER | RVT_S_WAIT_RNR | RVT_S_WAIT_PSN)) &&
925 	    (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
926 		rvt_add_retry_timer(qp);
927 
928 	while (qp->s_last != qp->s_acked) {
929 		u32 s_last;
930 
931 		wqe = rvt_get_swqe_ptr(qp, qp->s_last);
932 		if (qib_cmp24(wqe->lpsn, qp->s_sending_psn) >= 0 &&
933 		    qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) <= 0)
934 			break;
935 		s_last = qp->s_last;
936 		if (++s_last >= qp->s_size)
937 			s_last = 0;
938 		qp->s_last = s_last;
939 		/* see post_send() */
940 		barrier();
941 		rvt_put_swqe(wqe);
942 		rvt_qp_swqe_complete(qp,
943 				     wqe,
944 				     ib_qib_wc_opcode[wqe->wr.opcode],
945 				     IB_WC_SUCCESS);
946 	}
947 	/*
948 	 * If we were waiting for sends to complete before resending,
949 	 * and they are now complete, restart sending.
950 	 */
951 	if (qp->s_flags & RVT_S_WAIT_PSN &&
952 	    qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) > 0) {
953 		qp->s_flags &= ~RVT_S_WAIT_PSN;
954 		qp->s_sending_psn = qp->s_psn;
955 		qp->s_sending_hpsn = qp->s_psn - 1;
956 		qib_schedule_send(qp);
957 	}
958 }
959 
960 static inline void update_last_psn(struct rvt_qp *qp, u32 psn)
961 {
962 	qp->s_last_psn = psn;
963 }
964 
965 /*
966  * Generate a SWQE completion.
967  * This is similar to qib_send_complete but has to check to be sure
968  * that the SGEs are not being referenced if the SWQE is being resent.
969  */
970 static struct rvt_swqe *do_rc_completion(struct rvt_qp *qp,
971 					 struct rvt_swqe *wqe,
972 					 struct qib_ibport *ibp)
973 {
974 	/*
975 	 * Don't decrement refcount and don't generate a
976 	 * completion if the SWQE is being resent until the send
977 	 * is finished.
978 	 */
979 	if (qib_cmp24(wqe->lpsn, qp->s_sending_psn) < 0 ||
980 	    qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) > 0) {
981 		u32 s_last;
982 
983 		rvt_put_swqe(wqe);
984 		s_last = qp->s_last;
985 		if (++s_last >= qp->s_size)
986 			s_last = 0;
987 		qp->s_last = s_last;
988 		/* see post_send() */
989 		barrier();
990 		rvt_qp_swqe_complete(qp,
991 				     wqe,
992 				     ib_qib_wc_opcode[wqe->wr.opcode],
993 				     IB_WC_SUCCESS);
994 	} else
995 		this_cpu_inc(*ibp->rvp.rc_delayed_comp);
996 
997 	qp->s_retry = qp->s_retry_cnt;
998 	update_last_psn(qp, wqe->lpsn);
999 
1000 	/*
1001 	 * If we are completing a request which is in the process of
1002 	 * being resent, we can stop resending it since we know the
1003 	 * responder has already seen it.
1004 	 */
1005 	if (qp->s_acked == qp->s_cur) {
1006 		if (++qp->s_cur >= qp->s_size)
1007 			qp->s_cur = 0;
1008 		qp->s_acked = qp->s_cur;
1009 		wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
1010 		if (qp->s_acked != qp->s_tail) {
1011 			qp->s_state = OP(SEND_LAST);
1012 			qp->s_psn = wqe->psn;
1013 		}
1014 	} else {
1015 		if (++qp->s_acked >= qp->s_size)
1016 			qp->s_acked = 0;
1017 		if (qp->state == IB_QPS_SQD && qp->s_acked == qp->s_cur)
1018 			qp->s_draining = 0;
1019 		wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1020 	}
1021 	return wqe;
1022 }
1023 
1024 /**
1025  * do_rc_ack - process an incoming RC ACK
1026  * @qp: the QP the ACK came in on
1027  * @psn: the packet sequence number of the ACK
1028  * @opcode: the opcode of the request that resulted in the ACK
1029  *
1030  * This is called from qib_rc_rcv_resp() to process an incoming RC ACK
1031  * for the given QP.
1032  * Called at interrupt level with the QP s_lock held.
1033  * Returns 1 if OK, 0 if current operation should be aborted (NAK).
1034  */
1035 static int do_rc_ack(struct rvt_qp *qp, u32 aeth, u32 psn, int opcode,
1036 		     u64 val, struct qib_ctxtdata *rcd)
1037 {
1038 	struct qib_ibport *ibp;
1039 	enum ib_wc_status status;
1040 	struct rvt_swqe *wqe;
1041 	int ret = 0;
1042 	u32 ack_psn;
1043 	int diff;
1044 
1045 	/*
1046 	 * Note that NAKs implicitly ACK outstanding SEND and RDMA write
1047 	 * requests and implicitly NAK RDMA read and atomic requests issued
1048 	 * before the NAK'ed request.  The MSN won't include the NAK'ed
1049 	 * request but will include an ACK'ed request(s).
1050 	 */
1051 	ack_psn = psn;
1052 	if (aeth >> IB_AETH_NAK_SHIFT)
1053 		ack_psn--;
1054 	wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1055 	ibp = to_iport(qp->ibqp.device, qp->port_num);
1056 
1057 	/*
1058 	 * The MSN might be for a later WQE than the PSN indicates so
1059 	 * only complete WQEs that the PSN finishes.
1060 	 */
1061 	while ((diff = qib_cmp24(ack_psn, wqe->lpsn)) >= 0) {
1062 		/*
1063 		 * RDMA_READ_RESPONSE_ONLY is a special case since
1064 		 * we want to generate completion events for everything
1065 		 * before the RDMA read, copy the data, then generate
1066 		 * the completion for the read.
1067 		 */
1068 		if (wqe->wr.opcode == IB_WR_RDMA_READ &&
1069 		    opcode == OP(RDMA_READ_RESPONSE_ONLY) &&
1070 		    diff == 0) {
1071 			ret = 1;
1072 			goto bail;
1073 		}
1074 		/*
1075 		 * If this request is a RDMA read or atomic, and the ACK is
1076 		 * for a later operation, this ACK NAKs the RDMA read or
1077 		 * atomic.  In other words, only a RDMA_READ_LAST or ONLY
1078 		 * can ACK a RDMA read and likewise for atomic ops.  Note
1079 		 * that the NAK case can only happen if relaxed ordering is
1080 		 * used and requests are sent after an RDMA read or atomic
1081 		 * is sent but before the response is received.
1082 		 */
1083 		if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
1084 		     (opcode != OP(RDMA_READ_RESPONSE_LAST) || diff != 0)) ||
1085 		    ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1086 		      wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
1087 		     (opcode != OP(ATOMIC_ACKNOWLEDGE) || diff != 0))) {
1088 			/* Retry this request. */
1089 			if (!(qp->r_flags & RVT_R_RDMAR_SEQ)) {
1090 				qp->r_flags |= RVT_R_RDMAR_SEQ;
1091 				qib_restart_rc(qp, qp->s_last_psn + 1, 0);
1092 				if (list_empty(&qp->rspwait)) {
1093 					qp->r_flags |= RVT_R_RSP_SEND;
1094 					rvt_get_qp(qp);
1095 					list_add_tail(&qp->rspwait,
1096 						      &rcd->qp_wait_list);
1097 				}
1098 			}
1099 			/*
1100 			 * No need to process the ACK/NAK since we are
1101 			 * restarting an earlier request.
1102 			 */
1103 			goto bail;
1104 		}
1105 		if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1106 		    wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
1107 			u64 *vaddr = wqe->sg_list[0].vaddr;
1108 			*vaddr = val;
1109 		}
1110 		if (qp->s_num_rd_atomic &&
1111 		    (wqe->wr.opcode == IB_WR_RDMA_READ ||
1112 		     wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1113 		     wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)) {
1114 			qp->s_num_rd_atomic--;
1115 			/* Restart sending task if fence is complete */
1116 			if ((qp->s_flags & RVT_S_WAIT_FENCE) &&
1117 			    !qp->s_num_rd_atomic) {
1118 				qp->s_flags &= ~(RVT_S_WAIT_FENCE |
1119 						 RVT_S_WAIT_ACK);
1120 				qib_schedule_send(qp);
1121 			} else if (qp->s_flags & RVT_S_WAIT_RDMAR) {
1122 				qp->s_flags &= ~(RVT_S_WAIT_RDMAR |
1123 						 RVT_S_WAIT_ACK);
1124 				qib_schedule_send(qp);
1125 			}
1126 		}
1127 		wqe = do_rc_completion(qp, wqe, ibp);
1128 		if (qp->s_acked == qp->s_tail)
1129 			break;
1130 	}
1131 
1132 	switch (aeth >> IB_AETH_NAK_SHIFT) {
1133 	case 0:         /* ACK */
1134 		this_cpu_inc(*ibp->rvp.rc_acks);
1135 		if (qp->s_acked != qp->s_tail) {
1136 			/*
1137 			 * We are expecting more ACKs so
1138 			 * reset the retransmit timer.
1139 			 */
1140 			rvt_mod_retry_timer(qp);
1141 			/*
1142 			 * We can stop resending the earlier packets and
1143 			 * continue with the next packet the receiver wants.
1144 			 */
1145 			if (qib_cmp24(qp->s_psn, psn) <= 0)
1146 				reset_psn(qp, psn + 1);
1147 		} else {
1148 			/* No more acks - kill all timers */
1149 			rvt_stop_rc_timers(qp);
1150 			if (qib_cmp24(qp->s_psn, psn) <= 0) {
1151 				qp->s_state = OP(SEND_LAST);
1152 				qp->s_psn = psn + 1;
1153 			}
1154 		}
1155 		if (qp->s_flags & RVT_S_WAIT_ACK) {
1156 			qp->s_flags &= ~RVT_S_WAIT_ACK;
1157 			qib_schedule_send(qp);
1158 		}
1159 		rvt_get_credit(qp, aeth);
1160 		qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1161 		qp->s_retry = qp->s_retry_cnt;
1162 		update_last_psn(qp, psn);
1163 		return 1;
1164 
1165 	case 1:         /* RNR NAK */
1166 		ibp->rvp.n_rnr_naks++;
1167 		if (qp->s_acked == qp->s_tail)
1168 			goto bail;
1169 		if (qp->s_flags & RVT_S_WAIT_RNR)
1170 			goto bail;
1171 		if (qp->s_rnr_retry == 0) {
1172 			status = IB_WC_RNR_RETRY_EXC_ERR;
1173 			goto class_b;
1174 		}
1175 		if (qp->s_rnr_retry_cnt < 7)
1176 			qp->s_rnr_retry--;
1177 
1178 		/* The last valid PSN is the previous PSN. */
1179 		update_last_psn(qp, psn - 1);
1180 
1181 		ibp->rvp.n_rc_resends += (qp->s_psn - psn) & QIB_PSN_MASK;
1182 
1183 		reset_psn(qp, psn);
1184 
1185 		qp->s_flags &= ~(RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_ACK);
1186 		rvt_stop_rc_timers(qp);
1187 		rvt_add_rnr_timer(qp, aeth);
1188 		return 0;
1189 
1190 	case 3:         /* NAK */
1191 		if (qp->s_acked == qp->s_tail)
1192 			goto bail;
1193 		/* The last valid PSN is the previous PSN. */
1194 		update_last_psn(qp, psn - 1);
1195 		switch ((aeth >> IB_AETH_CREDIT_SHIFT) &
1196 			IB_AETH_CREDIT_MASK) {
1197 		case 0: /* PSN sequence error */
1198 			ibp->rvp.n_seq_naks++;
1199 			/*
1200 			 * Back up to the responder's expected PSN.
1201 			 * Note that we might get a NAK in the middle of an
1202 			 * RDMA READ response which terminates the RDMA
1203 			 * READ.
1204 			 */
1205 			qib_restart_rc(qp, psn, 0);
1206 			qib_schedule_send(qp);
1207 			break;
1208 
1209 		case 1: /* Invalid Request */
1210 			status = IB_WC_REM_INV_REQ_ERR;
1211 			ibp->rvp.n_other_naks++;
1212 			goto class_b;
1213 
1214 		case 2: /* Remote Access Error */
1215 			status = IB_WC_REM_ACCESS_ERR;
1216 			ibp->rvp.n_other_naks++;
1217 			goto class_b;
1218 
1219 		case 3: /* Remote Operation Error */
1220 			status = IB_WC_REM_OP_ERR;
1221 			ibp->rvp.n_other_naks++;
1222 class_b:
1223 			if (qp->s_last == qp->s_acked) {
1224 				rvt_send_complete(qp, wqe, status);
1225 				rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1226 			}
1227 			break;
1228 
1229 		default:
1230 			/* Ignore other reserved NAK error codes */
1231 			goto reserved;
1232 		}
1233 		qp->s_retry = qp->s_retry_cnt;
1234 		qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1235 		goto bail;
1236 
1237 	default:                /* 2: reserved */
1238 reserved:
1239 		/* Ignore reserved NAK codes. */
1240 		goto bail;
1241 	}
1242 
1243 bail:
1244 	rvt_stop_rc_timers(qp);
1245 	return ret;
1246 }
1247 
1248 /*
1249  * We have seen an out of sequence RDMA read middle or last packet.
1250  * This ACKs SENDs and RDMA writes up to the first RDMA read or atomic SWQE.
1251  */
1252 static void rdma_seq_err(struct rvt_qp *qp, struct qib_ibport *ibp, u32 psn,
1253 			 struct qib_ctxtdata *rcd)
1254 {
1255 	struct rvt_swqe *wqe;
1256 
1257 	/* Remove QP from retry timer */
1258 	rvt_stop_rc_timers(qp);
1259 
1260 	wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1261 
1262 	while (qib_cmp24(psn, wqe->lpsn) > 0) {
1263 		if (wqe->wr.opcode == IB_WR_RDMA_READ ||
1264 		    wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1265 		    wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
1266 			break;
1267 		wqe = do_rc_completion(qp, wqe, ibp);
1268 	}
1269 
1270 	ibp->rvp.n_rdma_seq++;
1271 	qp->r_flags |= RVT_R_RDMAR_SEQ;
1272 	qib_restart_rc(qp, qp->s_last_psn + 1, 0);
1273 	if (list_empty(&qp->rspwait)) {
1274 		qp->r_flags |= RVT_R_RSP_SEND;
1275 		rvt_get_qp(qp);
1276 		list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
1277 	}
1278 }
1279 
1280 /**
1281  * qib_rc_rcv_resp - process an incoming RC response packet
1282  * @ibp: the port this packet came in on
1283  * @ohdr: the other headers for this packet
1284  * @data: the packet data
1285  * @tlen: the packet length
1286  * @qp: the QP for this packet
1287  * @opcode: the opcode for this packet
1288  * @psn: the packet sequence number for this packet
1289  * @hdrsize: the header length
1290  * @pmtu: the path MTU
1291  *
1292  * This is called from qib_rc_rcv() to process an incoming RC response
1293  * packet for the given QP.
1294  * Called at interrupt level.
1295  */
1296 static void qib_rc_rcv_resp(struct qib_ibport *ibp,
1297 			    struct ib_other_headers *ohdr,
1298 			    void *data, u32 tlen,
1299 			    struct rvt_qp *qp,
1300 			    u32 opcode,
1301 			    u32 psn, u32 hdrsize, u32 pmtu,
1302 			    struct qib_ctxtdata *rcd)
1303 {
1304 	struct rvt_swqe *wqe;
1305 	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1306 	enum ib_wc_status status;
1307 	unsigned long flags;
1308 	int diff;
1309 	u32 pad;
1310 	u32 aeth;
1311 	u64 val;
1312 
1313 	if (opcode != OP(RDMA_READ_RESPONSE_MIDDLE)) {
1314 		/*
1315 		 * If ACK'd PSN on SDMA busy list try to make progress to
1316 		 * reclaim SDMA credits.
1317 		 */
1318 		if ((qib_cmp24(psn, qp->s_sending_psn) >= 0) &&
1319 		    (qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) <= 0)) {
1320 
1321 			/*
1322 			 * If send tasklet not running attempt to progress
1323 			 * SDMA queue.
1324 			 */
1325 			if (!(qp->s_flags & RVT_S_BUSY)) {
1326 				/* Acquire SDMA Lock */
1327 				spin_lock_irqsave(&ppd->sdma_lock, flags);
1328 				/* Invoke sdma make progress */
1329 				qib_sdma_make_progress(ppd);
1330 				/* Release SDMA Lock */
1331 				spin_unlock_irqrestore(&ppd->sdma_lock, flags);
1332 			}
1333 		}
1334 	}
1335 
1336 	spin_lock_irqsave(&qp->s_lock, flags);
1337 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
1338 		goto ack_done;
1339 
1340 	/* Ignore invalid responses. */
1341 	if (qib_cmp24(psn, READ_ONCE(qp->s_next_psn)) >= 0)
1342 		goto ack_done;
1343 
1344 	/* Ignore duplicate responses. */
1345 	diff = qib_cmp24(psn, qp->s_last_psn);
1346 	if (unlikely(diff <= 0)) {
1347 		/* Update credits for "ghost" ACKs */
1348 		if (diff == 0 && opcode == OP(ACKNOWLEDGE)) {
1349 			aeth = be32_to_cpu(ohdr->u.aeth);
1350 			if ((aeth >> IB_AETH_NAK_SHIFT) == 0)
1351 				rvt_get_credit(qp, aeth);
1352 		}
1353 		goto ack_done;
1354 	}
1355 
1356 	/*
1357 	 * Skip everything other than the PSN we expect, if we are waiting
1358 	 * for a reply to a restarted RDMA read or atomic op.
1359 	 */
1360 	if (qp->r_flags & RVT_R_RDMAR_SEQ) {
1361 		if (qib_cmp24(psn, qp->s_last_psn + 1) != 0)
1362 			goto ack_done;
1363 		qp->r_flags &= ~RVT_R_RDMAR_SEQ;
1364 	}
1365 
1366 	if (unlikely(qp->s_acked == qp->s_tail))
1367 		goto ack_done;
1368 	wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1369 	status = IB_WC_SUCCESS;
1370 
1371 	switch (opcode) {
1372 	case OP(ACKNOWLEDGE):
1373 	case OP(ATOMIC_ACKNOWLEDGE):
1374 	case OP(RDMA_READ_RESPONSE_FIRST):
1375 		aeth = be32_to_cpu(ohdr->u.aeth);
1376 		if (opcode == OP(ATOMIC_ACKNOWLEDGE))
1377 			val = ib_u64_get(&ohdr->u.at.atomic_ack_eth);
1378 		else
1379 			val = 0;
1380 		if (!do_rc_ack(qp, aeth, psn, opcode, val, rcd) ||
1381 		    opcode != OP(RDMA_READ_RESPONSE_FIRST))
1382 			goto ack_done;
1383 		hdrsize += 4;
1384 		wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1385 		if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1386 			goto ack_op_err;
1387 		/*
1388 		 * If this is a response to a resent RDMA read, we
1389 		 * have to be careful to copy the data to the right
1390 		 * location.
1391 		 */
1392 		qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1393 						  wqe, psn, pmtu);
1394 		goto read_middle;
1395 
1396 	case OP(RDMA_READ_RESPONSE_MIDDLE):
1397 		/* no AETH, no ACK */
1398 		if (unlikely(qib_cmp24(psn, qp->s_last_psn + 1)))
1399 			goto ack_seq_err;
1400 		if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1401 			goto ack_op_err;
1402 read_middle:
1403 		if (unlikely(tlen != (hdrsize + pmtu + 4)))
1404 			goto ack_len_err;
1405 		if (unlikely(pmtu >= qp->s_rdma_read_len))
1406 			goto ack_len_err;
1407 
1408 		/*
1409 		 * We got a response so update the timeout.
1410 		 * 4.096 usec. * (1 << qp->timeout)
1411 		 */
1412 		rvt_mod_retry_timer(qp);
1413 		if (qp->s_flags & RVT_S_WAIT_ACK) {
1414 			qp->s_flags &= ~RVT_S_WAIT_ACK;
1415 			qib_schedule_send(qp);
1416 		}
1417 
1418 		if (opcode == OP(RDMA_READ_RESPONSE_MIDDLE))
1419 			qp->s_retry = qp->s_retry_cnt;
1420 
1421 		/*
1422 		 * Update the RDMA receive state but do the copy w/o
1423 		 * holding the locks and blocking interrupts.
1424 		 */
1425 		qp->s_rdma_read_len -= pmtu;
1426 		update_last_psn(qp, psn);
1427 		spin_unlock_irqrestore(&qp->s_lock, flags);
1428 		rvt_copy_sge(qp, &qp->s_rdma_read_sge,
1429 			     data, pmtu, false, false);
1430 		goto bail;
1431 
1432 	case OP(RDMA_READ_RESPONSE_ONLY):
1433 		aeth = be32_to_cpu(ohdr->u.aeth);
1434 		if (!do_rc_ack(qp, aeth, psn, opcode, 0, rcd))
1435 			goto ack_done;
1436 		/* Get the number of bytes the message was padded by. */
1437 		pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1438 		/*
1439 		 * Check that the data size is >= 0 && <= pmtu.
1440 		 * Remember to account for the AETH header (4) and
1441 		 * ICRC (4).
1442 		 */
1443 		if (unlikely(tlen < (hdrsize + pad + 8)))
1444 			goto ack_len_err;
1445 		/*
1446 		 * If this is a response to a resent RDMA read, we
1447 		 * have to be careful to copy the data to the right
1448 		 * location.
1449 		 */
1450 		wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1451 		qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1452 						  wqe, psn, pmtu);
1453 		goto read_last;
1454 
1455 	case OP(RDMA_READ_RESPONSE_LAST):
1456 		/* ACKs READ req. */
1457 		if (unlikely(qib_cmp24(psn, qp->s_last_psn + 1)))
1458 			goto ack_seq_err;
1459 		if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1460 			goto ack_op_err;
1461 		/* Get the number of bytes the message was padded by. */
1462 		pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1463 		/*
1464 		 * Check that the data size is >= 1 && <= pmtu.
1465 		 * Remember to account for the AETH header (4) and
1466 		 * ICRC (4).
1467 		 */
1468 		if (unlikely(tlen <= (hdrsize + pad + 8)))
1469 			goto ack_len_err;
1470 read_last:
1471 		tlen -= hdrsize + pad + 8;
1472 		if (unlikely(tlen != qp->s_rdma_read_len))
1473 			goto ack_len_err;
1474 		aeth = be32_to_cpu(ohdr->u.aeth);
1475 		rvt_copy_sge(qp, &qp->s_rdma_read_sge,
1476 			     data, tlen, false, false);
1477 		WARN_ON(qp->s_rdma_read_sge.num_sge);
1478 		(void) do_rc_ack(qp, aeth, psn,
1479 				 OP(RDMA_READ_RESPONSE_LAST), 0, rcd);
1480 		goto ack_done;
1481 	}
1482 
1483 ack_op_err:
1484 	status = IB_WC_LOC_QP_OP_ERR;
1485 	goto ack_err;
1486 
1487 ack_seq_err:
1488 	rdma_seq_err(qp, ibp, psn, rcd);
1489 	goto ack_done;
1490 
1491 ack_len_err:
1492 	status = IB_WC_LOC_LEN_ERR;
1493 ack_err:
1494 	if (qp->s_last == qp->s_acked) {
1495 		rvt_send_complete(qp, wqe, status);
1496 		rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1497 	}
1498 ack_done:
1499 	spin_unlock_irqrestore(&qp->s_lock, flags);
1500 bail:
1501 	return;
1502 }
1503 
1504 /**
1505  * qib_rc_rcv_error - process an incoming duplicate or error RC packet
1506  * @ohdr: the other headers for this packet
1507  * @data: the packet data
1508  * @qp: the QP for this packet
1509  * @opcode: the opcode for this packet
1510  * @psn: the packet sequence number for this packet
1511  * @diff: the difference between the PSN and the expected PSN
1512  *
1513  * This is called from qib_rc_rcv() to process an unexpected
1514  * incoming RC packet for the given QP.
1515  * Called at interrupt level.
1516  * Return 1 if no more processing is needed; otherwise return 0 to
1517  * schedule a response to be sent.
1518  */
1519 static int qib_rc_rcv_error(struct ib_other_headers *ohdr,
1520 			    void *data,
1521 			    struct rvt_qp *qp,
1522 			    u32 opcode,
1523 			    u32 psn,
1524 			    int diff,
1525 			    struct qib_ctxtdata *rcd)
1526 {
1527 	struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
1528 	struct rvt_ack_entry *e;
1529 	unsigned long flags;
1530 	u8 i, prev;
1531 	int old_req;
1532 
1533 	if (diff > 0) {
1534 		/*
1535 		 * Packet sequence error.
1536 		 * A NAK will ACK earlier sends and RDMA writes.
1537 		 * Don't queue the NAK if we already sent one.
1538 		 */
1539 		if (!qp->r_nak_state) {
1540 			ibp->rvp.n_rc_seqnak++;
1541 			qp->r_nak_state = IB_NAK_PSN_ERROR;
1542 			/* Use the expected PSN. */
1543 			qp->r_ack_psn = qp->r_psn;
1544 			/*
1545 			 * Wait to send the sequence NAK until all packets
1546 			 * in the receive queue have been processed.
1547 			 * Otherwise, we end up propagating congestion.
1548 			 */
1549 			if (list_empty(&qp->rspwait)) {
1550 				qp->r_flags |= RVT_R_RSP_NAK;
1551 				rvt_get_qp(qp);
1552 				list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
1553 			}
1554 		}
1555 		goto done;
1556 	}
1557 
1558 	/*
1559 	 * Handle a duplicate request.  Don't re-execute SEND, RDMA
1560 	 * write or atomic op.  Don't NAK errors, just silently drop
1561 	 * the duplicate request.  Note that r_sge, r_len, and
1562 	 * r_rcv_len may be in use so don't modify them.
1563 	 *
1564 	 * We are supposed to ACK the earliest duplicate PSN but we
1565 	 * can coalesce an outstanding duplicate ACK.  We have to
1566 	 * send the earliest so that RDMA reads can be restarted at
1567 	 * the requester's expected PSN.
1568 	 *
1569 	 * First, find where this duplicate PSN falls within the
1570 	 * ACKs previously sent.
1571 	 * old_req is true if there is an older response that is scheduled
1572 	 * to be sent before sending this one.
1573 	 */
1574 	e = NULL;
1575 	old_req = 1;
1576 	ibp->rvp.n_rc_dupreq++;
1577 
1578 	spin_lock_irqsave(&qp->s_lock, flags);
1579 
1580 	for (i = qp->r_head_ack_queue; ; i = prev) {
1581 		if (i == qp->s_tail_ack_queue)
1582 			old_req = 0;
1583 		if (i)
1584 			prev = i - 1;
1585 		else
1586 			prev = QIB_MAX_RDMA_ATOMIC;
1587 		if (prev == qp->r_head_ack_queue) {
1588 			e = NULL;
1589 			break;
1590 		}
1591 		e = &qp->s_ack_queue[prev];
1592 		if (!e->opcode) {
1593 			e = NULL;
1594 			break;
1595 		}
1596 		if (qib_cmp24(psn, e->psn) >= 0) {
1597 			if (prev == qp->s_tail_ack_queue &&
1598 			    qib_cmp24(psn, e->lpsn) <= 0)
1599 				old_req = 0;
1600 			break;
1601 		}
1602 	}
1603 	switch (opcode) {
1604 	case OP(RDMA_READ_REQUEST): {
1605 		struct ib_reth *reth;
1606 		u32 offset;
1607 		u32 len;
1608 
1609 		/*
1610 		 * If we didn't find the RDMA read request in the ack queue,
1611 		 * we can ignore this request.
1612 		 */
1613 		if (!e || e->opcode != OP(RDMA_READ_REQUEST))
1614 			goto unlock_done;
1615 		/* RETH comes after BTH */
1616 		reth = &ohdr->u.rc.reth;
1617 		/*
1618 		 * Address range must be a subset of the original
1619 		 * request and start on pmtu boundaries.
1620 		 * We reuse the old ack_queue slot since the requester
1621 		 * should not back up and request an earlier PSN for the
1622 		 * same request.
1623 		 */
1624 		offset = ((psn - e->psn) & QIB_PSN_MASK) *
1625 			qp->pmtu;
1626 		len = be32_to_cpu(reth->length);
1627 		if (unlikely(offset + len != e->rdma_sge.sge_length))
1628 			goto unlock_done;
1629 		if (e->rdma_sge.mr) {
1630 			rvt_put_mr(e->rdma_sge.mr);
1631 			e->rdma_sge.mr = NULL;
1632 		}
1633 		if (len != 0) {
1634 			u32 rkey = be32_to_cpu(reth->rkey);
1635 			u64 vaddr = be64_to_cpu(reth->vaddr);
1636 			int ok;
1637 
1638 			ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr, rkey,
1639 					 IB_ACCESS_REMOTE_READ);
1640 			if (unlikely(!ok))
1641 				goto unlock_done;
1642 		} else {
1643 			e->rdma_sge.vaddr = NULL;
1644 			e->rdma_sge.length = 0;
1645 			e->rdma_sge.sge_length = 0;
1646 		}
1647 		e->psn = psn;
1648 		if (old_req)
1649 			goto unlock_done;
1650 		qp->s_tail_ack_queue = prev;
1651 		break;
1652 	}
1653 
1654 	case OP(COMPARE_SWAP):
1655 	case OP(FETCH_ADD): {
1656 		/*
1657 		 * If we didn't find the atomic request in the ack queue
1658 		 * or the send tasklet is already backed up to send an
1659 		 * earlier entry, we can ignore this request.
1660 		 */
1661 		if (!e || e->opcode != (u8) opcode || old_req)
1662 			goto unlock_done;
1663 		qp->s_tail_ack_queue = prev;
1664 		break;
1665 	}
1666 
1667 	default:
1668 		/*
1669 		 * Ignore this operation if it doesn't request an ACK
1670 		 * or an earlier RDMA read or atomic is going to be resent.
1671 		 */
1672 		if (!(psn & IB_BTH_REQ_ACK) || old_req)
1673 			goto unlock_done;
1674 		/*
1675 		 * Resend the most recent ACK if this request is
1676 		 * after all the previous RDMA reads and atomics.
1677 		 */
1678 		if (i == qp->r_head_ack_queue) {
1679 			spin_unlock_irqrestore(&qp->s_lock, flags);
1680 			qp->r_nak_state = 0;
1681 			qp->r_ack_psn = qp->r_psn - 1;
1682 			goto send_ack;
1683 		}
1684 		/*
1685 		 * Try to send a simple ACK to work around a Mellanox bug
1686 		 * which doesn't accept a RDMA read response or atomic
1687 		 * response as an ACK for earlier SENDs or RDMA writes.
1688 		 */
1689 		if (!(qp->s_flags & RVT_S_RESP_PENDING)) {
1690 			spin_unlock_irqrestore(&qp->s_lock, flags);
1691 			qp->r_nak_state = 0;
1692 			qp->r_ack_psn = qp->s_ack_queue[i].psn - 1;
1693 			goto send_ack;
1694 		}
1695 		/*
1696 		 * Resend the RDMA read or atomic op which
1697 		 * ACKs this duplicate request.
1698 		 */
1699 		qp->s_tail_ack_queue = i;
1700 		break;
1701 	}
1702 	qp->s_ack_state = OP(ACKNOWLEDGE);
1703 	qp->s_flags |= RVT_S_RESP_PENDING;
1704 	qp->r_nak_state = 0;
1705 	qib_schedule_send(qp);
1706 
1707 unlock_done:
1708 	spin_unlock_irqrestore(&qp->s_lock, flags);
1709 done:
1710 	return 1;
1711 
1712 send_ack:
1713 	return 0;
1714 }
1715 
1716 static inline void qib_update_ack_queue(struct rvt_qp *qp, unsigned n)
1717 {
1718 	unsigned next;
1719 
1720 	next = n + 1;
1721 	if (next > QIB_MAX_RDMA_ATOMIC)
1722 		next = 0;
1723 	qp->s_tail_ack_queue = next;
1724 	qp->s_ack_state = OP(ACKNOWLEDGE);
1725 }
1726 
1727 /**
1728  * qib_rc_rcv - process an incoming RC packet
1729  * @rcd: the context pointer
1730  * @hdr: the header of this packet
1731  * @has_grh: true if the header has a GRH
1732  * @data: the packet data
1733  * @tlen: the packet length
1734  * @qp: the QP for this packet
1735  *
1736  * This is called from qib_qp_rcv() to process an incoming RC packet
1737  * for the given QP.
1738  * Called at interrupt level.
1739  */
1740 void qib_rc_rcv(struct qib_ctxtdata *rcd, struct ib_header *hdr,
1741 		int has_grh, void *data, u32 tlen, struct rvt_qp *qp)
1742 {
1743 	struct qib_ibport *ibp = &rcd->ppd->ibport_data;
1744 	struct ib_other_headers *ohdr;
1745 	u32 opcode;
1746 	u32 hdrsize;
1747 	u32 psn;
1748 	u32 pad;
1749 	struct ib_wc wc;
1750 	u32 pmtu = qp->pmtu;
1751 	int diff;
1752 	struct ib_reth *reth;
1753 	unsigned long flags;
1754 	int ret;
1755 
1756 	/* Check for GRH */
1757 	if (!has_grh) {
1758 		ohdr = &hdr->u.oth;
1759 		hdrsize = 8 + 12;       /* LRH + BTH */
1760 	} else {
1761 		ohdr = &hdr->u.l.oth;
1762 		hdrsize = 8 + 40 + 12;  /* LRH + GRH + BTH */
1763 	}
1764 
1765 	opcode = be32_to_cpu(ohdr->bth[0]);
1766 	if (qib_ruc_check_hdr(ibp, hdr, has_grh, qp, opcode))
1767 		return;
1768 
1769 	psn = be32_to_cpu(ohdr->bth[2]);
1770 	opcode >>= 24;
1771 
1772 	/*
1773 	 * Process responses (ACKs) before anything else.  Note that the
1774 	 * packet sequence number will be for something in the send work
1775 	 * queue rather than the expected receive packet sequence number.
1776 	 * In other words, this QP is the requester.
1777 	 */
1778 	if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
1779 	    opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
1780 		qib_rc_rcv_resp(ibp, ohdr, data, tlen, qp, opcode, psn,
1781 				hdrsize, pmtu, rcd);
1782 		return;
1783 	}
1784 
1785 	/* Compute 24 bits worth of difference. */
1786 	diff = qib_cmp24(psn, qp->r_psn);
1787 	if (unlikely(diff)) {
1788 		if (qib_rc_rcv_error(ohdr, data, qp, opcode, psn, diff, rcd))
1789 			return;
1790 		goto send_ack;
1791 	}
1792 
1793 	/* Check for opcode sequence errors. */
1794 	switch (qp->r_state) {
1795 	case OP(SEND_FIRST):
1796 	case OP(SEND_MIDDLE):
1797 		if (opcode == OP(SEND_MIDDLE) ||
1798 		    opcode == OP(SEND_LAST) ||
1799 		    opcode == OP(SEND_LAST_WITH_IMMEDIATE))
1800 			break;
1801 		goto nack_inv;
1802 
1803 	case OP(RDMA_WRITE_FIRST):
1804 	case OP(RDMA_WRITE_MIDDLE):
1805 		if (opcode == OP(RDMA_WRITE_MIDDLE) ||
1806 		    opcode == OP(RDMA_WRITE_LAST) ||
1807 		    opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1808 			break;
1809 		goto nack_inv;
1810 
1811 	default:
1812 		if (opcode == OP(SEND_MIDDLE) ||
1813 		    opcode == OP(SEND_LAST) ||
1814 		    opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
1815 		    opcode == OP(RDMA_WRITE_MIDDLE) ||
1816 		    opcode == OP(RDMA_WRITE_LAST) ||
1817 		    opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1818 			goto nack_inv;
1819 		/*
1820 		 * Note that it is up to the requester to not send a new
1821 		 * RDMA read or atomic operation before receiving an ACK
1822 		 * for the previous operation.
1823 		 */
1824 		break;
1825 	}
1826 
1827 	if (qp->state == IB_QPS_RTR && !(qp->r_flags & RVT_R_COMM_EST))
1828 		rvt_comm_est(qp);
1829 
1830 	/* OK, process the packet. */
1831 	switch (opcode) {
1832 	case OP(SEND_FIRST):
1833 		ret = rvt_get_rwqe(qp, false);
1834 		if (ret < 0)
1835 			goto nack_op_err;
1836 		if (!ret)
1837 			goto rnr_nak;
1838 		qp->r_rcv_len = 0;
1839 		/* FALLTHROUGH */
1840 	case OP(SEND_MIDDLE):
1841 	case OP(RDMA_WRITE_MIDDLE):
1842 send_middle:
1843 		/* Check for invalid length PMTU or posted rwqe len. */
1844 		if (unlikely(tlen != (hdrsize + pmtu + 4)))
1845 			goto nack_inv;
1846 		qp->r_rcv_len += pmtu;
1847 		if (unlikely(qp->r_rcv_len > qp->r_len))
1848 			goto nack_inv;
1849 		rvt_copy_sge(qp, &qp->r_sge, data, pmtu, true, false);
1850 		break;
1851 
1852 	case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
1853 		/* consume RWQE */
1854 		ret = rvt_get_rwqe(qp, true);
1855 		if (ret < 0)
1856 			goto nack_op_err;
1857 		if (!ret)
1858 			goto rnr_nak;
1859 		goto send_last_imm;
1860 
1861 	case OP(SEND_ONLY):
1862 	case OP(SEND_ONLY_WITH_IMMEDIATE):
1863 		ret = rvt_get_rwqe(qp, false);
1864 		if (ret < 0)
1865 			goto nack_op_err;
1866 		if (!ret)
1867 			goto rnr_nak;
1868 		qp->r_rcv_len = 0;
1869 		if (opcode == OP(SEND_ONLY))
1870 			goto no_immediate_data;
1871 		/* fall through -- for SEND_ONLY_WITH_IMMEDIATE */
1872 	case OP(SEND_LAST_WITH_IMMEDIATE):
1873 send_last_imm:
1874 		wc.ex.imm_data = ohdr->u.imm_data;
1875 		hdrsize += 4;
1876 		wc.wc_flags = IB_WC_WITH_IMM;
1877 		goto send_last;
1878 	case OP(SEND_LAST):
1879 	case OP(RDMA_WRITE_LAST):
1880 no_immediate_data:
1881 		wc.wc_flags = 0;
1882 		wc.ex.imm_data = 0;
1883 send_last:
1884 		/* Get the number of bytes the message was padded by. */
1885 		pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1886 		/* Check for invalid length. */
1887 		/* XXX LAST len should be >= 1 */
1888 		if (unlikely(tlen < (hdrsize + pad + 4)))
1889 			goto nack_inv;
1890 		/* Don't count the CRC. */
1891 		tlen -= (hdrsize + pad + 4);
1892 		wc.byte_len = tlen + qp->r_rcv_len;
1893 		if (unlikely(wc.byte_len > qp->r_len))
1894 			goto nack_inv;
1895 		rvt_copy_sge(qp, &qp->r_sge, data, tlen, true, false);
1896 		rvt_put_ss(&qp->r_sge);
1897 		qp->r_msn++;
1898 		if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
1899 			break;
1900 		wc.wr_id = qp->r_wr_id;
1901 		wc.status = IB_WC_SUCCESS;
1902 		if (opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE) ||
1903 		    opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE))
1904 			wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
1905 		else
1906 			wc.opcode = IB_WC_RECV;
1907 		wc.qp = &qp->ibqp;
1908 		wc.src_qp = qp->remote_qpn;
1909 		wc.slid = rdma_ah_get_dlid(&qp->remote_ah_attr);
1910 		wc.sl = rdma_ah_get_sl(&qp->remote_ah_attr);
1911 		/* zero fields that are N/A */
1912 		wc.vendor_err = 0;
1913 		wc.pkey_index = 0;
1914 		wc.dlid_path_bits = 0;
1915 		wc.port_num = 0;
1916 		/* Signal completion event if the solicited bit is set. */
1917 		rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc,
1918 			     ib_bth_is_solicited(ohdr));
1919 		break;
1920 
1921 	case OP(RDMA_WRITE_FIRST):
1922 	case OP(RDMA_WRITE_ONLY):
1923 	case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
1924 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
1925 			goto nack_inv;
1926 		/* consume RWQE */
1927 		reth = &ohdr->u.rc.reth;
1928 		hdrsize += sizeof(*reth);
1929 		qp->r_len = be32_to_cpu(reth->length);
1930 		qp->r_rcv_len = 0;
1931 		qp->r_sge.sg_list = NULL;
1932 		if (qp->r_len != 0) {
1933 			u32 rkey = be32_to_cpu(reth->rkey);
1934 			u64 vaddr = be64_to_cpu(reth->vaddr);
1935 			int ok;
1936 
1937 			/* Check rkey & NAK */
1938 			ok = rvt_rkey_ok(qp, &qp->r_sge.sge, qp->r_len, vaddr,
1939 					 rkey, IB_ACCESS_REMOTE_WRITE);
1940 			if (unlikely(!ok))
1941 				goto nack_acc;
1942 			qp->r_sge.num_sge = 1;
1943 		} else {
1944 			qp->r_sge.num_sge = 0;
1945 			qp->r_sge.sge.mr = NULL;
1946 			qp->r_sge.sge.vaddr = NULL;
1947 			qp->r_sge.sge.length = 0;
1948 			qp->r_sge.sge.sge_length = 0;
1949 		}
1950 		if (opcode == OP(RDMA_WRITE_FIRST))
1951 			goto send_middle;
1952 		else if (opcode == OP(RDMA_WRITE_ONLY))
1953 			goto no_immediate_data;
1954 		ret = rvt_get_rwqe(qp, true);
1955 		if (ret < 0)
1956 			goto nack_op_err;
1957 		if (!ret) {
1958 			rvt_put_ss(&qp->r_sge);
1959 			goto rnr_nak;
1960 		}
1961 		wc.ex.imm_data = ohdr->u.rc.imm_data;
1962 		hdrsize += 4;
1963 		wc.wc_flags = IB_WC_WITH_IMM;
1964 		goto send_last;
1965 
1966 	case OP(RDMA_READ_REQUEST): {
1967 		struct rvt_ack_entry *e;
1968 		u32 len;
1969 		u8 next;
1970 
1971 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
1972 			goto nack_inv;
1973 		next = qp->r_head_ack_queue + 1;
1974 		/* s_ack_queue is size QIB_MAX_RDMA_ATOMIC+1 so use > not >= */
1975 		if (next > QIB_MAX_RDMA_ATOMIC)
1976 			next = 0;
1977 		spin_lock_irqsave(&qp->s_lock, flags);
1978 		if (unlikely(next == qp->s_tail_ack_queue)) {
1979 			if (!qp->s_ack_queue[next].sent)
1980 				goto nack_inv_unlck;
1981 			qib_update_ack_queue(qp, next);
1982 		}
1983 		e = &qp->s_ack_queue[qp->r_head_ack_queue];
1984 		if (e->opcode == OP(RDMA_READ_REQUEST) && e->rdma_sge.mr) {
1985 			rvt_put_mr(e->rdma_sge.mr);
1986 			e->rdma_sge.mr = NULL;
1987 		}
1988 		reth = &ohdr->u.rc.reth;
1989 		len = be32_to_cpu(reth->length);
1990 		if (len) {
1991 			u32 rkey = be32_to_cpu(reth->rkey);
1992 			u64 vaddr = be64_to_cpu(reth->vaddr);
1993 			int ok;
1994 
1995 			/* Check rkey & NAK */
1996 			ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr,
1997 					 rkey, IB_ACCESS_REMOTE_READ);
1998 			if (unlikely(!ok))
1999 				goto nack_acc_unlck;
2000 			/*
2001 			 * Update the next expected PSN.  We add 1 later
2002 			 * below, so only add the remainder here.
2003 			 */
2004 			qp->r_psn += rvt_div_mtu(qp, len - 1);
2005 		} else {
2006 			e->rdma_sge.mr = NULL;
2007 			e->rdma_sge.vaddr = NULL;
2008 			e->rdma_sge.length = 0;
2009 			e->rdma_sge.sge_length = 0;
2010 		}
2011 		e->opcode = opcode;
2012 		e->sent = 0;
2013 		e->psn = psn;
2014 		e->lpsn = qp->r_psn;
2015 		/*
2016 		 * We need to increment the MSN here instead of when we
2017 		 * finish sending the result since a duplicate request would
2018 		 * increment it more than once.
2019 		 */
2020 		qp->r_msn++;
2021 		qp->r_psn++;
2022 		qp->r_state = opcode;
2023 		qp->r_nak_state = 0;
2024 		qp->r_head_ack_queue = next;
2025 
2026 		/* Schedule the send tasklet. */
2027 		qp->s_flags |= RVT_S_RESP_PENDING;
2028 		qib_schedule_send(qp);
2029 
2030 		goto sunlock;
2031 	}
2032 
2033 	case OP(COMPARE_SWAP):
2034 	case OP(FETCH_ADD): {
2035 		struct ib_atomic_eth *ateth;
2036 		struct rvt_ack_entry *e;
2037 		u64 vaddr;
2038 		atomic64_t *maddr;
2039 		u64 sdata;
2040 		u32 rkey;
2041 		u8 next;
2042 
2043 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
2044 			goto nack_inv;
2045 		next = qp->r_head_ack_queue + 1;
2046 		if (next > QIB_MAX_RDMA_ATOMIC)
2047 			next = 0;
2048 		spin_lock_irqsave(&qp->s_lock, flags);
2049 		if (unlikely(next == qp->s_tail_ack_queue)) {
2050 			if (!qp->s_ack_queue[next].sent)
2051 				goto nack_inv_unlck;
2052 			qib_update_ack_queue(qp, next);
2053 		}
2054 		e = &qp->s_ack_queue[qp->r_head_ack_queue];
2055 		if (e->opcode == OP(RDMA_READ_REQUEST) && e->rdma_sge.mr) {
2056 			rvt_put_mr(e->rdma_sge.mr);
2057 			e->rdma_sge.mr = NULL;
2058 		}
2059 		ateth = &ohdr->u.atomic_eth;
2060 		vaddr = get_ib_ateth_vaddr(ateth);
2061 		if (unlikely(vaddr & (sizeof(u64) - 1)))
2062 			goto nack_inv_unlck;
2063 		rkey = be32_to_cpu(ateth->rkey);
2064 		/* Check rkey & NAK */
2065 		if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64),
2066 					  vaddr, rkey,
2067 					  IB_ACCESS_REMOTE_ATOMIC)))
2068 			goto nack_acc_unlck;
2069 		/* Perform atomic OP and save result. */
2070 		maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
2071 		sdata = get_ib_ateth_swap(ateth);
2072 		e->atomic_data = (opcode == OP(FETCH_ADD)) ?
2073 			(u64) atomic64_add_return(sdata, maddr) - sdata :
2074 			(u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
2075 				      get_ib_ateth_compare(ateth),
2076 				      sdata);
2077 		rvt_put_mr(qp->r_sge.sge.mr);
2078 		qp->r_sge.num_sge = 0;
2079 		e->opcode = opcode;
2080 		e->sent = 0;
2081 		e->psn = psn;
2082 		e->lpsn = psn;
2083 		qp->r_msn++;
2084 		qp->r_psn++;
2085 		qp->r_state = opcode;
2086 		qp->r_nak_state = 0;
2087 		qp->r_head_ack_queue = next;
2088 
2089 		/* Schedule the send tasklet. */
2090 		qp->s_flags |= RVT_S_RESP_PENDING;
2091 		qib_schedule_send(qp);
2092 
2093 		goto sunlock;
2094 	}
2095 
2096 	default:
2097 		/* NAK unknown opcodes. */
2098 		goto nack_inv;
2099 	}
2100 	qp->r_psn++;
2101 	qp->r_state = opcode;
2102 	qp->r_ack_psn = psn;
2103 	qp->r_nak_state = 0;
2104 	/* Send an ACK if requested or required. */
2105 	if (psn & (1 << 31))
2106 		goto send_ack;
2107 	return;
2108 
2109 rnr_nak:
2110 	qp->r_nak_state = IB_RNR_NAK | qp->r_min_rnr_timer;
2111 	qp->r_ack_psn = qp->r_psn;
2112 	/* Queue RNR NAK for later */
2113 	if (list_empty(&qp->rspwait)) {
2114 		qp->r_flags |= RVT_R_RSP_NAK;
2115 		rvt_get_qp(qp);
2116 		list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
2117 	}
2118 	return;
2119 
2120 nack_op_err:
2121 	rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
2122 	qp->r_nak_state = IB_NAK_REMOTE_OPERATIONAL_ERROR;
2123 	qp->r_ack_psn = qp->r_psn;
2124 	/* Queue NAK for later */
2125 	if (list_empty(&qp->rspwait)) {
2126 		qp->r_flags |= RVT_R_RSP_NAK;
2127 		rvt_get_qp(qp);
2128 		list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
2129 	}
2130 	return;
2131 
2132 nack_inv_unlck:
2133 	spin_unlock_irqrestore(&qp->s_lock, flags);
2134 nack_inv:
2135 	rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
2136 	qp->r_nak_state = IB_NAK_INVALID_REQUEST;
2137 	qp->r_ack_psn = qp->r_psn;
2138 	/* Queue NAK for later */
2139 	if (list_empty(&qp->rspwait)) {
2140 		qp->r_flags |= RVT_R_RSP_NAK;
2141 		rvt_get_qp(qp);
2142 		list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
2143 	}
2144 	return;
2145 
2146 nack_acc_unlck:
2147 	spin_unlock_irqrestore(&qp->s_lock, flags);
2148 nack_acc:
2149 	rvt_rc_error(qp, IB_WC_LOC_PROT_ERR);
2150 	qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
2151 	qp->r_ack_psn = qp->r_psn;
2152 send_ack:
2153 	qib_send_rc_ack(qp);
2154 	return;
2155 
2156 sunlock:
2157 	spin_unlock_irqrestore(&qp->s_lock, flags);
2158 }
2159