xref: /openbmc/linux/drivers/infiniband/hw/qib/qib_rc.c (revision ecd25094)
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  *
211  * Assumes the s_lock is held.
212  *
213  * Return 1 if constructed; otherwise, return 0.
214  */
215 int qib_make_rc_req(struct rvt_qp *qp, unsigned long *flags)
216 {
217 	struct qib_qp_priv *priv = qp->priv;
218 	struct qib_ibdev *dev = to_idev(qp->ibqp.device);
219 	struct ib_other_headers *ohdr;
220 	struct rvt_sge_state *ss;
221 	struct rvt_swqe *wqe;
222 	u32 hwords;
223 	u32 len;
224 	u32 bth0;
225 	u32 bth2;
226 	u32 pmtu = qp->pmtu;
227 	char newreq;
228 	int ret = 0;
229 	int delta;
230 
231 	ohdr = &priv->s_hdr->u.oth;
232 	if (rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH)
233 		ohdr = &priv->s_hdr->u.l.oth;
234 
235 	/* Sending responses has higher priority over sending requests. */
236 	if ((qp->s_flags & RVT_S_RESP_PENDING) &&
237 	    qib_make_rc_ack(dev, qp, ohdr, pmtu))
238 		goto done;
239 
240 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_SEND_OK)) {
241 		if (!(ib_rvt_state_ops[qp->state] & RVT_FLUSH_SEND))
242 			goto bail;
243 		/* We are in the error state, flush the work request. */
244 		if (qp->s_last == READ_ONCE(qp->s_head))
245 			goto bail;
246 		/* If DMAs are in progress, we can't flush immediately. */
247 		if (atomic_read(&priv->s_dma_busy)) {
248 			qp->s_flags |= RVT_S_WAIT_DMA;
249 			goto bail;
250 		}
251 		wqe = rvt_get_swqe_ptr(qp, qp->s_last);
252 		rvt_send_complete(qp, wqe, qp->s_last != qp->s_acked ?
253 			IB_WC_SUCCESS : IB_WC_WR_FLUSH_ERR);
254 		/* will get called again */
255 		goto done;
256 	}
257 
258 	if (qp->s_flags & (RVT_S_WAIT_RNR | RVT_S_WAIT_ACK))
259 		goto bail;
260 
261 	if (qib_cmp24(qp->s_psn, qp->s_sending_hpsn) <= 0) {
262 		if (qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) <= 0) {
263 			qp->s_flags |= RVT_S_WAIT_PSN;
264 			goto bail;
265 		}
266 		qp->s_sending_psn = qp->s_psn;
267 		qp->s_sending_hpsn = qp->s_psn - 1;
268 	}
269 
270 	/* header size in 32-bit words LRH+BTH = (8+12)/4. */
271 	hwords = 5;
272 	bth0 = 0;
273 
274 	/* Send a request. */
275 	wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
276 	switch (qp->s_state) {
277 	default:
278 		if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_NEXT_SEND_OK))
279 			goto bail;
280 		/*
281 		 * Resend an old request or start a new one.
282 		 *
283 		 * We keep track of the current SWQE so that
284 		 * we don't reset the "furthest progress" state
285 		 * if we need to back up.
286 		 */
287 		newreq = 0;
288 		if (qp->s_cur == qp->s_tail) {
289 			/* Check if send work queue is empty. */
290 			if (qp->s_tail == READ_ONCE(qp->s_head))
291 				goto bail;
292 			/*
293 			 * If a fence is requested, wait for previous
294 			 * RDMA read and atomic operations to finish.
295 			 */
296 			if ((wqe->wr.send_flags & IB_SEND_FENCE) &&
297 			    qp->s_num_rd_atomic) {
298 				qp->s_flags |= RVT_S_WAIT_FENCE;
299 				goto bail;
300 			}
301 			newreq = 1;
302 			qp->s_psn = wqe->psn;
303 		}
304 		/*
305 		 * Note that we have to be careful not to modify the
306 		 * original work request since we may need to resend
307 		 * it.
308 		 */
309 		len = wqe->length;
310 		ss = &qp->s_sge;
311 		bth2 = qp->s_psn & QIB_PSN_MASK;
312 		switch (wqe->wr.opcode) {
313 		case IB_WR_SEND:
314 		case IB_WR_SEND_WITH_IMM:
315 			/* If no credit, return. */
316 			if (!rvt_rc_credit_avail(qp, wqe))
317 				goto bail;
318 			if (len > pmtu) {
319 				qp->s_state = OP(SEND_FIRST);
320 				len = pmtu;
321 				break;
322 			}
323 			if (wqe->wr.opcode == IB_WR_SEND)
324 				qp->s_state = OP(SEND_ONLY);
325 			else {
326 				qp->s_state = OP(SEND_ONLY_WITH_IMMEDIATE);
327 				/* Immediate data comes after the BTH */
328 				ohdr->u.imm_data = wqe->wr.ex.imm_data;
329 				hwords += 1;
330 			}
331 			if (wqe->wr.send_flags & IB_SEND_SOLICITED)
332 				bth0 |= IB_BTH_SOLICITED;
333 			bth2 |= IB_BTH_REQ_ACK;
334 			if (++qp->s_cur == qp->s_size)
335 				qp->s_cur = 0;
336 			break;
337 
338 		case IB_WR_RDMA_WRITE:
339 			if (newreq && !(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
340 				qp->s_lsn++;
341 			goto no_flow_control;
342 		case IB_WR_RDMA_WRITE_WITH_IMM:
343 			/* If no credit, return. */
344 			if (!rvt_rc_credit_avail(qp, wqe))
345 				goto bail;
346 no_flow_control:
347 			ohdr->u.rc.reth.vaddr =
348 				cpu_to_be64(wqe->rdma_wr.remote_addr);
349 			ohdr->u.rc.reth.rkey =
350 				cpu_to_be32(wqe->rdma_wr.rkey);
351 			ohdr->u.rc.reth.length = cpu_to_be32(len);
352 			hwords += sizeof(struct ib_reth) / sizeof(u32);
353 			if (len > pmtu) {
354 				qp->s_state = OP(RDMA_WRITE_FIRST);
355 				len = pmtu;
356 				break;
357 			}
358 			if (wqe->rdma_wr.wr.opcode == IB_WR_RDMA_WRITE)
359 				qp->s_state = OP(RDMA_WRITE_ONLY);
360 			else {
361 				qp->s_state = OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
362 				/* Immediate data comes after RETH */
363 				ohdr->u.rc.imm_data =
364 					wqe->rdma_wr.wr.ex.imm_data;
365 				hwords += 1;
366 				if (wqe->rdma_wr.wr.send_flags & IB_SEND_SOLICITED)
367 					bth0 |= IB_BTH_SOLICITED;
368 			}
369 			bth2 |= IB_BTH_REQ_ACK;
370 			if (++qp->s_cur == qp->s_size)
371 				qp->s_cur = 0;
372 			break;
373 
374 		case IB_WR_RDMA_READ:
375 			/*
376 			 * Don't allow more operations to be started
377 			 * than the QP limits allow.
378 			 */
379 			if (newreq) {
380 				if (qp->s_num_rd_atomic >=
381 				    qp->s_max_rd_atomic) {
382 					qp->s_flags |= RVT_S_WAIT_RDMAR;
383 					goto bail;
384 				}
385 				qp->s_num_rd_atomic++;
386 				if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
387 					qp->s_lsn++;
388 			}
389 
390 			ohdr->u.rc.reth.vaddr =
391 				cpu_to_be64(wqe->rdma_wr.remote_addr);
392 			ohdr->u.rc.reth.rkey =
393 				cpu_to_be32(wqe->rdma_wr.rkey);
394 			ohdr->u.rc.reth.length = cpu_to_be32(len);
395 			qp->s_state = OP(RDMA_READ_REQUEST);
396 			hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
397 			ss = NULL;
398 			len = 0;
399 			bth2 |= IB_BTH_REQ_ACK;
400 			if (++qp->s_cur == qp->s_size)
401 				qp->s_cur = 0;
402 			break;
403 
404 		case IB_WR_ATOMIC_CMP_AND_SWP:
405 		case IB_WR_ATOMIC_FETCH_AND_ADD:
406 			/*
407 			 * Don't allow more operations to be started
408 			 * than the QP limits allow.
409 			 */
410 			if (newreq) {
411 				if (qp->s_num_rd_atomic >=
412 				    qp->s_max_rd_atomic) {
413 					qp->s_flags |= RVT_S_WAIT_RDMAR;
414 					goto bail;
415 				}
416 				qp->s_num_rd_atomic++;
417 				if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
418 					qp->s_lsn++;
419 			}
420 			if (wqe->atomic_wr.wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
421 				qp->s_state = OP(COMPARE_SWAP);
422 				put_ib_ateth_swap(wqe->atomic_wr.swap,
423 						  &ohdr->u.atomic_eth);
424 				put_ib_ateth_compare(wqe->atomic_wr.compare_add,
425 						     &ohdr->u.atomic_eth);
426 			} else {
427 				qp->s_state = OP(FETCH_ADD);
428 				put_ib_ateth_swap(wqe->atomic_wr.compare_add,
429 						  &ohdr->u.atomic_eth);
430 				put_ib_ateth_compare(0, &ohdr->u.atomic_eth);
431 			}
432 			put_ib_ateth_vaddr(wqe->atomic_wr.remote_addr,
433 					   &ohdr->u.atomic_eth);
434 			ohdr->u.atomic_eth.rkey = cpu_to_be32(
435 				wqe->atomic_wr.rkey);
436 			hwords += sizeof(struct ib_atomic_eth) / sizeof(u32);
437 			ss = NULL;
438 			len = 0;
439 			bth2 |= IB_BTH_REQ_ACK;
440 			if (++qp->s_cur == qp->s_size)
441 				qp->s_cur = 0;
442 			break;
443 
444 		default:
445 			goto bail;
446 		}
447 		qp->s_sge.sge = wqe->sg_list[0];
448 		qp->s_sge.sg_list = wqe->sg_list + 1;
449 		qp->s_sge.num_sge = wqe->wr.num_sge;
450 		qp->s_sge.total_len = wqe->length;
451 		qp->s_len = wqe->length;
452 		if (newreq) {
453 			qp->s_tail++;
454 			if (qp->s_tail >= qp->s_size)
455 				qp->s_tail = 0;
456 		}
457 		if (wqe->wr.opcode == IB_WR_RDMA_READ)
458 			qp->s_psn = wqe->lpsn + 1;
459 		else
460 			qp->s_psn++;
461 		break;
462 
463 	case OP(RDMA_READ_RESPONSE_FIRST):
464 		/*
465 		 * qp->s_state is normally set to the opcode of the
466 		 * last packet constructed for new requests and therefore
467 		 * is never set to RDMA read response.
468 		 * RDMA_READ_RESPONSE_FIRST is used by the ACK processing
469 		 * thread to indicate a SEND needs to be restarted from an
470 		 * earlier PSN without interferring with the sending thread.
471 		 * See qib_restart_rc().
472 		 */
473 		qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu);
474 		/* FALLTHROUGH */
475 	case OP(SEND_FIRST):
476 		qp->s_state = OP(SEND_MIDDLE);
477 		/* FALLTHROUGH */
478 	case OP(SEND_MIDDLE):
479 		bth2 = qp->s_psn++ & QIB_PSN_MASK;
480 		ss = &qp->s_sge;
481 		len = qp->s_len;
482 		if (len > pmtu) {
483 			len = pmtu;
484 			break;
485 		}
486 		if (wqe->wr.opcode == IB_WR_SEND)
487 			qp->s_state = OP(SEND_LAST);
488 		else {
489 			qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
490 			/* Immediate data comes after the BTH */
491 			ohdr->u.imm_data = wqe->wr.ex.imm_data;
492 			hwords += 1;
493 		}
494 		if (wqe->wr.send_flags & IB_SEND_SOLICITED)
495 			bth0 |= IB_BTH_SOLICITED;
496 		bth2 |= IB_BTH_REQ_ACK;
497 		qp->s_cur++;
498 		if (qp->s_cur >= qp->s_size)
499 			qp->s_cur = 0;
500 		break;
501 
502 	case OP(RDMA_READ_RESPONSE_LAST):
503 		/*
504 		 * qp->s_state is normally set to the opcode of the
505 		 * last packet constructed for new requests and therefore
506 		 * is never set to RDMA read response.
507 		 * RDMA_READ_RESPONSE_LAST is used by the ACK processing
508 		 * thread to indicate a RDMA write needs to be restarted from
509 		 * an earlier PSN without interferring with the sending thread.
510 		 * See qib_restart_rc().
511 		 */
512 		qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu);
513 		/* FALLTHROUGH */
514 	case OP(RDMA_WRITE_FIRST):
515 		qp->s_state = OP(RDMA_WRITE_MIDDLE);
516 		/* FALLTHROUGH */
517 	case OP(RDMA_WRITE_MIDDLE):
518 		bth2 = qp->s_psn++ & QIB_PSN_MASK;
519 		ss = &qp->s_sge;
520 		len = qp->s_len;
521 		if (len > pmtu) {
522 			len = pmtu;
523 			break;
524 		}
525 		if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
526 			qp->s_state = OP(RDMA_WRITE_LAST);
527 		else {
528 			qp->s_state = OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
529 			/* Immediate data comes after the BTH */
530 			ohdr->u.imm_data = wqe->wr.ex.imm_data;
531 			hwords += 1;
532 			if (wqe->wr.send_flags & IB_SEND_SOLICITED)
533 				bth0 |= IB_BTH_SOLICITED;
534 		}
535 		bth2 |= IB_BTH_REQ_ACK;
536 		qp->s_cur++;
537 		if (qp->s_cur >= qp->s_size)
538 			qp->s_cur = 0;
539 		break;
540 
541 	case OP(RDMA_READ_RESPONSE_MIDDLE):
542 		/*
543 		 * qp->s_state is normally set to the opcode of the
544 		 * last packet constructed for new requests and therefore
545 		 * is never set to RDMA read response.
546 		 * RDMA_READ_RESPONSE_MIDDLE is used by the ACK processing
547 		 * thread to indicate a RDMA read needs to be restarted from
548 		 * an earlier PSN without interferring with the sending thread.
549 		 * See qib_restart_rc().
550 		 */
551 		len = ((qp->s_psn - wqe->psn) & QIB_PSN_MASK) * pmtu;
552 		ohdr->u.rc.reth.vaddr =
553 			cpu_to_be64(wqe->rdma_wr.remote_addr + len);
554 		ohdr->u.rc.reth.rkey =
555 			cpu_to_be32(wqe->rdma_wr.rkey);
556 		ohdr->u.rc.reth.length = cpu_to_be32(wqe->length - len);
557 		qp->s_state = OP(RDMA_READ_REQUEST);
558 		hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
559 		bth2 = (qp->s_psn & QIB_PSN_MASK) | IB_BTH_REQ_ACK;
560 		qp->s_psn = wqe->lpsn + 1;
561 		ss = NULL;
562 		len = 0;
563 		qp->s_cur++;
564 		if (qp->s_cur == qp->s_size)
565 			qp->s_cur = 0;
566 		break;
567 	}
568 	qp->s_sending_hpsn = bth2;
569 	delta = (((int) bth2 - (int) wqe->psn) << 8) >> 8;
570 	if (delta && delta % QIB_PSN_CREDIT == 0)
571 		bth2 |= IB_BTH_REQ_ACK;
572 	if (qp->s_flags & RVT_S_SEND_ONE) {
573 		qp->s_flags &= ~RVT_S_SEND_ONE;
574 		qp->s_flags |= RVT_S_WAIT_ACK;
575 		bth2 |= IB_BTH_REQ_ACK;
576 	}
577 	qp->s_len -= len;
578 	qp->s_hdrwords = hwords;
579 	qp->s_cur_sge = ss;
580 	qp->s_cur_size = len;
581 	qib_make_ruc_header(qp, ohdr, bth0 | (qp->s_state << 24), bth2);
582 done:
583 	return 1;
584 bail:
585 	qp->s_flags &= ~RVT_S_BUSY;
586 	return ret;
587 }
588 
589 /**
590  * qib_send_rc_ack - Construct an ACK packet and send it
591  * @qp: a pointer to the QP
592  *
593  * This is called from qib_rc_rcv() and qib_kreceive().
594  * Note that RDMA reads and atomics are handled in the
595  * send side QP state and tasklet.
596  */
597 void qib_send_rc_ack(struct rvt_qp *qp)
598 {
599 	struct qib_devdata *dd = dd_from_ibdev(qp->ibqp.device);
600 	struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
601 	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
602 	u64 pbc;
603 	u16 lrh0;
604 	u32 bth0;
605 	u32 hwords;
606 	u32 pbufn;
607 	u32 __iomem *piobuf;
608 	struct ib_header hdr;
609 	struct ib_other_headers *ohdr;
610 	u32 control;
611 	unsigned long flags;
612 
613 	spin_lock_irqsave(&qp->s_lock, flags);
614 
615 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
616 		goto unlock;
617 
618 	/* Don't send ACK or NAK if a RDMA read or atomic is pending. */
619 	if ((qp->s_flags & RVT_S_RESP_PENDING) || qp->s_rdma_ack_cnt)
620 		goto queue_ack;
621 
622 	/* Construct the header with s_lock held so APM doesn't change it. */
623 	ohdr = &hdr.u.oth;
624 	lrh0 = QIB_LRH_BTH;
625 	/* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4. */
626 	hwords = 6;
627 	if (unlikely(rdma_ah_get_ah_flags(&qp->remote_ah_attr) &
628 		     IB_AH_GRH)) {
629 		hwords += qib_make_grh(ibp, &hdr.u.l.grh,
630 				       rdma_ah_read_grh(&qp->remote_ah_attr),
631 				       hwords, 0);
632 		ohdr = &hdr.u.l.oth;
633 		lrh0 = QIB_LRH_GRH;
634 	}
635 	/* read pkey_index w/o lock (its atomic) */
636 	bth0 = qib_get_pkey(ibp, qp->s_pkey_index) | (OP(ACKNOWLEDGE) << 24);
637 	if (qp->s_mig_state == IB_MIG_MIGRATED)
638 		bth0 |= IB_BTH_MIG_REQ;
639 	if (qp->r_nak_state)
640 		ohdr->u.aeth = cpu_to_be32((qp->r_msn & IB_MSN_MASK) |
641 					    (qp->r_nak_state <<
642 					     IB_AETH_CREDIT_SHIFT));
643 	else
644 		ohdr->u.aeth = rvt_compute_aeth(qp);
645 	lrh0 |= ibp->sl_to_vl[rdma_ah_get_sl(&qp->remote_ah_attr)] << 12 |
646 		rdma_ah_get_sl(&qp->remote_ah_attr) << 4;
647 	hdr.lrh[0] = cpu_to_be16(lrh0);
648 	hdr.lrh[1] = cpu_to_be16(rdma_ah_get_dlid(&qp->remote_ah_attr));
649 	hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC);
650 	hdr.lrh[3] = cpu_to_be16(ppd->lid |
651 				 rdma_ah_get_path_bits(&qp->remote_ah_attr));
652 	ohdr->bth[0] = cpu_to_be32(bth0);
653 	ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
654 	ohdr->bth[2] = cpu_to_be32(qp->r_ack_psn & QIB_PSN_MASK);
655 
656 	spin_unlock_irqrestore(&qp->s_lock, flags);
657 
658 	/* Don't try to send ACKs if the link isn't ACTIVE */
659 	if (!(ppd->lflags & QIBL_LINKACTIVE))
660 		goto done;
661 
662 	control = dd->f_setpbc_control(ppd, hwords + SIZE_OF_CRC,
663 				       qp->s_srate, lrh0 >> 12);
664 	/* length is + 1 for the control dword */
665 	pbc = ((u64) control << 32) | (hwords + 1);
666 
667 	piobuf = dd->f_getsendbuf(ppd, pbc, &pbufn);
668 	if (!piobuf) {
669 		/*
670 		 * We are out of PIO buffers at the moment.
671 		 * Pass responsibility for sending the ACK to the
672 		 * send tasklet so that when a PIO buffer becomes
673 		 * available, the ACK is sent ahead of other outgoing
674 		 * packets.
675 		 */
676 		spin_lock_irqsave(&qp->s_lock, flags);
677 		goto queue_ack;
678 	}
679 
680 	/*
681 	 * Write the pbc.
682 	 * We have to flush after the PBC for correctness
683 	 * on some cpus or WC buffer can be written out of order.
684 	 */
685 	writeq(pbc, piobuf);
686 
687 	if (dd->flags & QIB_PIO_FLUSH_WC) {
688 		u32 *hdrp = (u32 *) &hdr;
689 
690 		qib_flush_wc();
691 		qib_pio_copy(piobuf + 2, hdrp, hwords - 1);
692 		qib_flush_wc();
693 		__raw_writel(hdrp[hwords - 1], piobuf + hwords + 1);
694 	} else
695 		qib_pio_copy(piobuf + 2, (u32 *) &hdr, hwords);
696 
697 	if (dd->flags & QIB_USE_SPCL_TRIG) {
698 		u32 spcl_off = (pbufn >= dd->piobcnt2k) ? 2047 : 1023;
699 
700 		qib_flush_wc();
701 		__raw_writel(0xaebecede, piobuf + spcl_off);
702 	}
703 
704 	qib_flush_wc();
705 	qib_sendbuf_done(dd, pbufn);
706 
707 	this_cpu_inc(ibp->pmastats->n_unicast_xmit);
708 	goto done;
709 
710 queue_ack:
711 	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
712 		this_cpu_inc(*ibp->rvp.rc_qacks);
713 		qp->s_flags |= RVT_S_ACK_PENDING | RVT_S_RESP_PENDING;
714 		qp->s_nak_state = qp->r_nak_state;
715 		qp->s_ack_psn = qp->r_ack_psn;
716 
717 		/* Schedule the send tasklet. */
718 		qib_schedule_send(qp);
719 	}
720 unlock:
721 	spin_unlock_irqrestore(&qp->s_lock, flags);
722 done:
723 	return;
724 }
725 
726 /**
727  * reset_psn - reset the QP state to send starting from PSN
728  * @qp: the QP
729  * @psn: the packet sequence number to restart at
730  *
731  * This is called from qib_rc_rcv() to process an incoming RC ACK
732  * for the given QP.
733  * Called at interrupt level with the QP s_lock held.
734  */
735 static void reset_psn(struct rvt_qp *qp, u32 psn)
736 {
737 	u32 n = qp->s_acked;
738 	struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, n);
739 	u32 opcode;
740 
741 	qp->s_cur = n;
742 
743 	/*
744 	 * If we are starting the request from the beginning,
745 	 * let the normal send code handle initialization.
746 	 */
747 	if (qib_cmp24(psn, wqe->psn) <= 0) {
748 		qp->s_state = OP(SEND_LAST);
749 		goto done;
750 	}
751 
752 	/* Find the work request opcode corresponding to the given PSN. */
753 	opcode = wqe->wr.opcode;
754 	for (;;) {
755 		int diff;
756 
757 		if (++n == qp->s_size)
758 			n = 0;
759 		if (n == qp->s_tail)
760 			break;
761 		wqe = rvt_get_swqe_ptr(qp, n);
762 		diff = qib_cmp24(psn, wqe->psn);
763 		if (diff < 0)
764 			break;
765 		qp->s_cur = n;
766 		/*
767 		 * If we are starting the request from the beginning,
768 		 * let the normal send code handle initialization.
769 		 */
770 		if (diff == 0) {
771 			qp->s_state = OP(SEND_LAST);
772 			goto done;
773 		}
774 		opcode = wqe->wr.opcode;
775 	}
776 
777 	/*
778 	 * Set the state to restart in the middle of a request.
779 	 * Don't change the s_sge, s_cur_sge, or s_cur_size.
780 	 * See qib_make_rc_req().
781 	 */
782 	switch (opcode) {
783 	case IB_WR_SEND:
784 	case IB_WR_SEND_WITH_IMM:
785 		qp->s_state = OP(RDMA_READ_RESPONSE_FIRST);
786 		break;
787 
788 	case IB_WR_RDMA_WRITE:
789 	case IB_WR_RDMA_WRITE_WITH_IMM:
790 		qp->s_state = OP(RDMA_READ_RESPONSE_LAST);
791 		break;
792 
793 	case IB_WR_RDMA_READ:
794 		qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE);
795 		break;
796 
797 	default:
798 		/*
799 		 * This case shouldn't happen since its only
800 		 * one PSN per req.
801 		 */
802 		qp->s_state = OP(SEND_LAST);
803 	}
804 done:
805 	qp->s_psn = psn;
806 	/*
807 	 * Set RVT_S_WAIT_PSN as qib_rc_complete() may start the timer
808 	 * asynchronously before the send tasklet can get scheduled.
809 	 * Doing it in qib_make_rc_req() is too late.
810 	 */
811 	if ((qib_cmp24(qp->s_psn, qp->s_sending_hpsn) <= 0) &&
812 	    (qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) <= 0))
813 		qp->s_flags |= RVT_S_WAIT_PSN;
814 }
815 
816 /*
817  * Back up requester to resend the last un-ACKed request.
818  * The QP r_lock and s_lock should be held and interrupts disabled.
819  */
820 void qib_restart_rc(struct rvt_qp *qp, u32 psn, int wait)
821 {
822 	struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
823 	struct qib_ibport *ibp;
824 
825 	if (qp->s_retry == 0) {
826 		if (qp->s_mig_state == IB_MIG_ARMED) {
827 			qib_migrate_qp(qp);
828 			qp->s_retry = qp->s_retry_cnt;
829 		} else if (qp->s_last == qp->s_acked) {
830 			rvt_send_complete(qp, wqe, IB_WC_RETRY_EXC_ERR);
831 			rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
832 			return;
833 		} else /* XXX need to handle delayed completion */
834 			return;
835 	} else
836 		qp->s_retry--;
837 
838 	ibp = to_iport(qp->ibqp.device, qp->port_num);
839 	if (wqe->wr.opcode == IB_WR_RDMA_READ)
840 		ibp->rvp.n_rc_resends++;
841 	else
842 		ibp->rvp.n_rc_resends += (qp->s_psn - psn) & QIB_PSN_MASK;
843 
844 	qp->s_flags &= ~(RVT_S_WAIT_FENCE | RVT_S_WAIT_RDMAR |
845 			 RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_PSN |
846 			 RVT_S_WAIT_ACK);
847 	if (wait)
848 		qp->s_flags |= RVT_S_SEND_ONE;
849 	reset_psn(qp, psn);
850 }
851 
852 /*
853  * Set qp->s_sending_psn to the next PSN after the given one.
854  * This would be psn+1 except when RDMA reads are present.
855  */
856 static void reset_sending_psn(struct rvt_qp *qp, u32 psn)
857 {
858 	struct rvt_swqe *wqe;
859 	u32 n = qp->s_last;
860 
861 	/* Find the work request corresponding to the given PSN. */
862 	for (;;) {
863 		wqe = rvt_get_swqe_ptr(qp, n);
864 		if (qib_cmp24(psn, wqe->lpsn) <= 0) {
865 			if (wqe->wr.opcode == IB_WR_RDMA_READ)
866 				qp->s_sending_psn = wqe->lpsn + 1;
867 			else
868 				qp->s_sending_psn = psn + 1;
869 			break;
870 		}
871 		if (++n == qp->s_size)
872 			n = 0;
873 		if (n == qp->s_tail)
874 			break;
875 	}
876 }
877 
878 /*
879  * This should be called with the QP s_lock held and interrupts disabled.
880  */
881 void qib_rc_send_complete(struct rvt_qp *qp, struct ib_header *hdr)
882 {
883 	struct ib_other_headers *ohdr;
884 	struct rvt_swqe *wqe;
885 	u32 opcode;
886 	u32 psn;
887 
888 	if (!(ib_rvt_state_ops[qp->state] & RVT_SEND_OR_FLUSH_OR_RECV_OK))
889 		return;
890 
891 	/* Find out where the BTH is */
892 	if ((be16_to_cpu(hdr->lrh[0]) & 3) == QIB_LRH_BTH)
893 		ohdr = &hdr->u.oth;
894 	else
895 		ohdr = &hdr->u.l.oth;
896 
897 	opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
898 	if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
899 	    opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
900 		WARN_ON(!qp->s_rdma_ack_cnt);
901 		qp->s_rdma_ack_cnt--;
902 		return;
903 	}
904 
905 	psn = be32_to_cpu(ohdr->bth[2]);
906 	reset_sending_psn(qp, psn);
907 
908 	/*
909 	 * Start timer after a packet requesting an ACK has been sent and
910 	 * there are still requests that haven't been acked.
911 	 */
912 	if ((psn & IB_BTH_REQ_ACK) && qp->s_acked != qp->s_tail &&
913 	    !(qp->s_flags & (RVT_S_TIMER | RVT_S_WAIT_RNR | RVT_S_WAIT_PSN)) &&
914 	    (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
915 		rvt_add_retry_timer(qp);
916 
917 	while (qp->s_last != qp->s_acked) {
918 		wqe = rvt_get_swqe_ptr(qp, qp->s_last);
919 		if (qib_cmp24(wqe->lpsn, qp->s_sending_psn) >= 0 &&
920 		    qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) <= 0)
921 			break;
922 		rvt_qp_complete_swqe(qp,
923 				     wqe,
924 				     ib_qib_wc_opcode[wqe->wr.opcode],
925 				     IB_WC_SUCCESS);
926 	}
927 	/*
928 	 * If we were waiting for sends to complete before resending,
929 	 * and they are now complete, restart sending.
930 	 */
931 	if (qp->s_flags & RVT_S_WAIT_PSN &&
932 	    qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) > 0) {
933 		qp->s_flags &= ~RVT_S_WAIT_PSN;
934 		qp->s_sending_psn = qp->s_psn;
935 		qp->s_sending_hpsn = qp->s_psn - 1;
936 		qib_schedule_send(qp);
937 	}
938 }
939 
940 static inline void update_last_psn(struct rvt_qp *qp, u32 psn)
941 {
942 	qp->s_last_psn = psn;
943 }
944 
945 /*
946  * Generate a SWQE completion.
947  * This is similar to qib_send_complete but has to check to be sure
948  * that the SGEs are not being referenced if the SWQE is being resent.
949  */
950 static struct rvt_swqe *do_rc_completion(struct rvt_qp *qp,
951 					 struct rvt_swqe *wqe,
952 					 struct qib_ibport *ibp)
953 {
954 	/*
955 	 * Don't decrement refcount and don't generate a
956 	 * completion if the SWQE is being resent until the send
957 	 * is finished.
958 	 */
959 	if (qib_cmp24(wqe->lpsn, qp->s_sending_psn) < 0 ||
960 	    qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) > 0)
961 		rvt_qp_complete_swqe(qp,
962 				     wqe,
963 				     ib_qib_wc_opcode[wqe->wr.opcode],
964 				     IB_WC_SUCCESS);
965 	else
966 		this_cpu_inc(*ibp->rvp.rc_delayed_comp);
967 
968 	qp->s_retry = qp->s_retry_cnt;
969 	update_last_psn(qp, wqe->lpsn);
970 
971 	/*
972 	 * If we are completing a request which is in the process of
973 	 * being resent, we can stop resending it since we know the
974 	 * responder has already seen it.
975 	 */
976 	if (qp->s_acked == qp->s_cur) {
977 		if (++qp->s_cur >= qp->s_size)
978 			qp->s_cur = 0;
979 		qp->s_acked = qp->s_cur;
980 		wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
981 		if (qp->s_acked != qp->s_tail) {
982 			qp->s_state = OP(SEND_LAST);
983 			qp->s_psn = wqe->psn;
984 		}
985 	} else {
986 		if (++qp->s_acked >= qp->s_size)
987 			qp->s_acked = 0;
988 		if (qp->state == IB_QPS_SQD && qp->s_acked == qp->s_cur)
989 			qp->s_draining = 0;
990 		wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
991 	}
992 	return wqe;
993 }
994 
995 /**
996  * do_rc_ack - process an incoming RC ACK
997  * @qp: the QP the ACK came in on
998  * @psn: the packet sequence number of the ACK
999  * @opcode: the opcode of the request that resulted in the ACK
1000  *
1001  * This is called from qib_rc_rcv_resp() to process an incoming RC ACK
1002  * for the given QP.
1003  * Called at interrupt level with the QP s_lock held.
1004  * Returns 1 if OK, 0 if current operation should be aborted (NAK).
1005  */
1006 static int do_rc_ack(struct rvt_qp *qp, u32 aeth, u32 psn, int opcode,
1007 		     u64 val, struct qib_ctxtdata *rcd)
1008 {
1009 	struct qib_ibport *ibp;
1010 	enum ib_wc_status status;
1011 	struct rvt_swqe *wqe;
1012 	int ret = 0;
1013 	u32 ack_psn;
1014 	int diff;
1015 
1016 	/*
1017 	 * Note that NAKs implicitly ACK outstanding SEND and RDMA write
1018 	 * requests and implicitly NAK RDMA read and atomic requests issued
1019 	 * before the NAK'ed request.  The MSN won't include the NAK'ed
1020 	 * request but will include an ACK'ed request(s).
1021 	 */
1022 	ack_psn = psn;
1023 	if (aeth >> IB_AETH_NAK_SHIFT)
1024 		ack_psn--;
1025 	wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1026 	ibp = to_iport(qp->ibqp.device, qp->port_num);
1027 
1028 	/*
1029 	 * The MSN might be for a later WQE than the PSN indicates so
1030 	 * only complete WQEs that the PSN finishes.
1031 	 */
1032 	while ((diff = qib_cmp24(ack_psn, wqe->lpsn)) >= 0) {
1033 		/*
1034 		 * RDMA_READ_RESPONSE_ONLY is a special case since
1035 		 * we want to generate completion events for everything
1036 		 * before the RDMA read, copy the data, then generate
1037 		 * the completion for the read.
1038 		 */
1039 		if (wqe->wr.opcode == IB_WR_RDMA_READ &&
1040 		    opcode == OP(RDMA_READ_RESPONSE_ONLY) &&
1041 		    diff == 0) {
1042 			ret = 1;
1043 			goto bail;
1044 		}
1045 		/*
1046 		 * If this request is a RDMA read or atomic, and the ACK is
1047 		 * for a later operation, this ACK NAKs the RDMA read or
1048 		 * atomic.  In other words, only a RDMA_READ_LAST or ONLY
1049 		 * can ACK a RDMA read and likewise for atomic ops.  Note
1050 		 * that the NAK case can only happen if relaxed ordering is
1051 		 * used and requests are sent after an RDMA read or atomic
1052 		 * is sent but before the response is received.
1053 		 */
1054 		if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
1055 		     (opcode != OP(RDMA_READ_RESPONSE_LAST) || diff != 0)) ||
1056 		    ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1057 		      wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
1058 		     (opcode != OP(ATOMIC_ACKNOWLEDGE) || diff != 0))) {
1059 			/* Retry this request. */
1060 			if (!(qp->r_flags & RVT_R_RDMAR_SEQ)) {
1061 				qp->r_flags |= RVT_R_RDMAR_SEQ;
1062 				qib_restart_rc(qp, qp->s_last_psn + 1, 0);
1063 				if (list_empty(&qp->rspwait)) {
1064 					qp->r_flags |= RVT_R_RSP_SEND;
1065 					rvt_get_qp(qp);
1066 					list_add_tail(&qp->rspwait,
1067 						      &rcd->qp_wait_list);
1068 				}
1069 			}
1070 			/*
1071 			 * No need to process the ACK/NAK since we are
1072 			 * restarting an earlier request.
1073 			 */
1074 			goto bail;
1075 		}
1076 		if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1077 		    wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
1078 			u64 *vaddr = wqe->sg_list[0].vaddr;
1079 			*vaddr = val;
1080 		}
1081 		if (qp->s_num_rd_atomic &&
1082 		    (wqe->wr.opcode == IB_WR_RDMA_READ ||
1083 		     wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1084 		     wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)) {
1085 			qp->s_num_rd_atomic--;
1086 			/* Restart sending task if fence is complete */
1087 			if ((qp->s_flags & RVT_S_WAIT_FENCE) &&
1088 			    !qp->s_num_rd_atomic) {
1089 				qp->s_flags &= ~(RVT_S_WAIT_FENCE |
1090 						 RVT_S_WAIT_ACK);
1091 				qib_schedule_send(qp);
1092 			} else if (qp->s_flags & RVT_S_WAIT_RDMAR) {
1093 				qp->s_flags &= ~(RVT_S_WAIT_RDMAR |
1094 						 RVT_S_WAIT_ACK);
1095 				qib_schedule_send(qp);
1096 			}
1097 		}
1098 		wqe = do_rc_completion(qp, wqe, ibp);
1099 		if (qp->s_acked == qp->s_tail)
1100 			break;
1101 	}
1102 
1103 	switch (aeth >> IB_AETH_NAK_SHIFT) {
1104 	case 0:         /* ACK */
1105 		this_cpu_inc(*ibp->rvp.rc_acks);
1106 		if (qp->s_acked != qp->s_tail) {
1107 			/*
1108 			 * We are expecting more ACKs so
1109 			 * reset the retransmit timer.
1110 			 */
1111 			rvt_mod_retry_timer(qp);
1112 			/*
1113 			 * We can stop resending the earlier packets and
1114 			 * continue with the next packet the receiver wants.
1115 			 */
1116 			if (qib_cmp24(qp->s_psn, psn) <= 0)
1117 				reset_psn(qp, psn + 1);
1118 		} else {
1119 			/* No more acks - kill all timers */
1120 			rvt_stop_rc_timers(qp);
1121 			if (qib_cmp24(qp->s_psn, psn) <= 0) {
1122 				qp->s_state = OP(SEND_LAST);
1123 				qp->s_psn = psn + 1;
1124 			}
1125 		}
1126 		if (qp->s_flags & RVT_S_WAIT_ACK) {
1127 			qp->s_flags &= ~RVT_S_WAIT_ACK;
1128 			qib_schedule_send(qp);
1129 		}
1130 		rvt_get_credit(qp, aeth);
1131 		qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1132 		qp->s_retry = qp->s_retry_cnt;
1133 		update_last_psn(qp, psn);
1134 		return 1;
1135 
1136 	case 1:         /* RNR NAK */
1137 		ibp->rvp.n_rnr_naks++;
1138 		if (qp->s_acked == qp->s_tail)
1139 			goto bail;
1140 		if (qp->s_flags & RVT_S_WAIT_RNR)
1141 			goto bail;
1142 		if (qp->s_rnr_retry == 0) {
1143 			status = IB_WC_RNR_RETRY_EXC_ERR;
1144 			goto class_b;
1145 		}
1146 		if (qp->s_rnr_retry_cnt < 7)
1147 			qp->s_rnr_retry--;
1148 
1149 		/* The last valid PSN is the previous PSN. */
1150 		update_last_psn(qp, psn - 1);
1151 
1152 		ibp->rvp.n_rc_resends += (qp->s_psn - psn) & QIB_PSN_MASK;
1153 
1154 		reset_psn(qp, psn);
1155 
1156 		qp->s_flags &= ~(RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_ACK);
1157 		rvt_stop_rc_timers(qp);
1158 		rvt_add_rnr_timer(qp, aeth);
1159 		return 0;
1160 
1161 	case 3:         /* NAK */
1162 		if (qp->s_acked == qp->s_tail)
1163 			goto bail;
1164 		/* The last valid PSN is the previous PSN. */
1165 		update_last_psn(qp, psn - 1);
1166 		switch ((aeth >> IB_AETH_CREDIT_SHIFT) &
1167 			IB_AETH_CREDIT_MASK) {
1168 		case 0: /* PSN sequence error */
1169 			ibp->rvp.n_seq_naks++;
1170 			/*
1171 			 * Back up to the responder's expected PSN.
1172 			 * Note that we might get a NAK in the middle of an
1173 			 * RDMA READ response which terminates the RDMA
1174 			 * READ.
1175 			 */
1176 			qib_restart_rc(qp, psn, 0);
1177 			qib_schedule_send(qp);
1178 			break;
1179 
1180 		case 1: /* Invalid Request */
1181 			status = IB_WC_REM_INV_REQ_ERR;
1182 			ibp->rvp.n_other_naks++;
1183 			goto class_b;
1184 
1185 		case 2: /* Remote Access Error */
1186 			status = IB_WC_REM_ACCESS_ERR;
1187 			ibp->rvp.n_other_naks++;
1188 			goto class_b;
1189 
1190 		case 3: /* Remote Operation Error */
1191 			status = IB_WC_REM_OP_ERR;
1192 			ibp->rvp.n_other_naks++;
1193 class_b:
1194 			if (qp->s_last == qp->s_acked) {
1195 				rvt_send_complete(qp, wqe, status);
1196 				rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1197 			}
1198 			break;
1199 
1200 		default:
1201 			/* Ignore other reserved NAK error codes */
1202 			goto reserved;
1203 		}
1204 		qp->s_retry = qp->s_retry_cnt;
1205 		qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1206 		goto bail;
1207 
1208 	default:                /* 2: reserved */
1209 reserved:
1210 		/* Ignore reserved NAK codes. */
1211 		goto bail;
1212 	}
1213 
1214 bail:
1215 	rvt_stop_rc_timers(qp);
1216 	return ret;
1217 }
1218 
1219 /*
1220  * We have seen an out of sequence RDMA read middle or last packet.
1221  * This ACKs SENDs and RDMA writes up to the first RDMA read or atomic SWQE.
1222  */
1223 static void rdma_seq_err(struct rvt_qp *qp, struct qib_ibport *ibp, u32 psn,
1224 			 struct qib_ctxtdata *rcd)
1225 {
1226 	struct rvt_swqe *wqe;
1227 
1228 	/* Remove QP from retry timer */
1229 	rvt_stop_rc_timers(qp);
1230 
1231 	wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1232 
1233 	while (qib_cmp24(psn, wqe->lpsn) > 0) {
1234 		if (wqe->wr.opcode == IB_WR_RDMA_READ ||
1235 		    wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1236 		    wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
1237 			break;
1238 		wqe = do_rc_completion(qp, wqe, ibp);
1239 	}
1240 
1241 	ibp->rvp.n_rdma_seq++;
1242 	qp->r_flags |= RVT_R_RDMAR_SEQ;
1243 	qib_restart_rc(qp, qp->s_last_psn + 1, 0);
1244 	if (list_empty(&qp->rspwait)) {
1245 		qp->r_flags |= RVT_R_RSP_SEND;
1246 		rvt_get_qp(qp);
1247 		list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
1248 	}
1249 }
1250 
1251 /**
1252  * qib_rc_rcv_resp - process an incoming RC response packet
1253  * @ibp: the port this packet came in on
1254  * @ohdr: the other headers for this packet
1255  * @data: the packet data
1256  * @tlen: the packet length
1257  * @qp: the QP for this packet
1258  * @opcode: the opcode for this packet
1259  * @psn: the packet sequence number for this packet
1260  * @hdrsize: the header length
1261  * @pmtu: the path MTU
1262  *
1263  * This is called from qib_rc_rcv() to process an incoming RC response
1264  * packet for the given QP.
1265  * Called at interrupt level.
1266  */
1267 static void qib_rc_rcv_resp(struct qib_ibport *ibp,
1268 			    struct ib_other_headers *ohdr,
1269 			    void *data, u32 tlen,
1270 			    struct rvt_qp *qp,
1271 			    u32 opcode,
1272 			    u32 psn, u32 hdrsize, u32 pmtu,
1273 			    struct qib_ctxtdata *rcd)
1274 {
1275 	struct rvt_swqe *wqe;
1276 	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1277 	enum ib_wc_status status;
1278 	unsigned long flags;
1279 	int diff;
1280 	u32 pad;
1281 	u32 aeth;
1282 	u64 val;
1283 
1284 	if (opcode != OP(RDMA_READ_RESPONSE_MIDDLE)) {
1285 		/*
1286 		 * If ACK'd PSN on SDMA busy list try to make progress to
1287 		 * reclaim SDMA credits.
1288 		 */
1289 		if ((qib_cmp24(psn, qp->s_sending_psn) >= 0) &&
1290 		    (qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) <= 0)) {
1291 
1292 			/*
1293 			 * If send tasklet not running attempt to progress
1294 			 * SDMA queue.
1295 			 */
1296 			if (!(qp->s_flags & RVT_S_BUSY)) {
1297 				/* Acquire SDMA Lock */
1298 				spin_lock_irqsave(&ppd->sdma_lock, flags);
1299 				/* Invoke sdma make progress */
1300 				qib_sdma_make_progress(ppd);
1301 				/* Release SDMA Lock */
1302 				spin_unlock_irqrestore(&ppd->sdma_lock, flags);
1303 			}
1304 		}
1305 	}
1306 
1307 	spin_lock_irqsave(&qp->s_lock, flags);
1308 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
1309 		goto ack_done;
1310 
1311 	/* Ignore invalid responses. */
1312 	if (qib_cmp24(psn, READ_ONCE(qp->s_next_psn)) >= 0)
1313 		goto ack_done;
1314 
1315 	/* Ignore duplicate responses. */
1316 	diff = qib_cmp24(psn, qp->s_last_psn);
1317 	if (unlikely(diff <= 0)) {
1318 		/* Update credits for "ghost" ACKs */
1319 		if (diff == 0 && opcode == OP(ACKNOWLEDGE)) {
1320 			aeth = be32_to_cpu(ohdr->u.aeth);
1321 			if ((aeth >> IB_AETH_NAK_SHIFT) == 0)
1322 				rvt_get_credit(qp, aeth);
1323 		}
1324 		goto ack_done;
1325 	}
1326 
1327 	/*
1328 	 * Skip everything other than the PSN we expect, if we are waiting
1329 	 * for a reply to a restarted RDMA read or atomic op.
1330 	 */
1331 	if (qp->r_flags & RVT_R_RDMAR_SEQ) {
1332 		if (qib_cmp24(psn, qp->s_last_psn + 1) != 0)
1333 			goto ack_done;
1334 		qp->r_flags &= ~RVT_R_RDMAR_SEQ;
1335 	}
1336 
1337 	if (unlikely(qp->s_acked == qp->s_tail))
1338 		goto ack_done;
1339 	wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1340 	status = IB_WC_SUCCESS;
1341 
1342 	switch (opcode) {
1343 	case OP(ACKNOWLEDGE):
1344 	case OP(ATOMIC_ACKNOWLEDGE):
1345 	case OP(RDMA_READ_RESPONSE_FIRST):
1346 		aeth = be32_to_cpu(ohdr->u.aeth);
1347 		if (opcode == OP(ATOMIC_ACKNOWLEDGE))
1348 			val = ib_u64_get(&ohdr->u.at.atomic_ack_eth);
1349 		else
1350 			val = 0;
1351 		if (!do_rc_ack(qp, aeth, psn, opcode, val, rcd) ||
1352 		    opcode != OP(RDMA_READ_RESPONSE_FIRST))
1353 			goto ack_done;
1354 		hdrsize += 4;
1355 		wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1356 		if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1357 			goto ack_op_err;
1358 		/*
1359 		 * If this is a response to a resent RDMA read, we
1360 		 * have to be careful to copy the data to the right
1361 		 * location.
1362 		 */
1363 		qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1364 						  wqe, psn, pmtu);
1365 		goto read_middle;
1366 
1367 	case OP(RDMA_READ_RESPONSE_MIDDLE):
1368 		/* no AETH, no ACK */
1369 		if (unlikely(qib_cmp24(psn, qp->s_last_psn + 1)))
1370 			goto ack_seq_err;
1371 		if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1372 			goto ack_op_err;
1373 read_middle:
1374 		if (unlikely(tlen != (hdrsize + pmtu + 4)))
1375 			goto ack_len_err;
1376 		if (unlikely(pmtu >= qp->s_rdma_read_len))
1377 			goto ack_len_err;
1378 
1379 		/*
1380 		 * We got a response so update the timeout.
1381 		 * 4.096 usec. * (1 << qp->timeout)
1382 		 */
1383 		rvt_mod_retry_timer(qp);
1384 		if (qp->s_flags & RVT_S_WAIT_ACK) {
1385 			qp->s_flags &= ~RVT_S_WAIT_ACK;
1386 			qib_schedule_send(qp);
1387 		}
1388 
1389 		if (opcode == OP(RDMA_READ_RESPONSE_MIDDLE))
1390 			qp->s_retry = qp->s_retry_cnt;
1391 
1392 		/*
1393 		 * Update the RDMA receive state but do the copy w/o
1394 		 * holding the locks and blocking interrupts.
1395 		 */
1396 		qp->s_rdma_read_len -= pmtu;
1397 		update_last_psn(qp, psn);
1398 		spin_unlock_irqrestore(&qp->s_lock, flags);
1399 		rvt_copy_sge(qp, &qp->s_rdma_read_sge,
1400 			     data, pmtu, false, false);
1401 		goto bail;
1402 
1403 	case OP(RDMA_READ_RESPONSE_ONLY):
1404 		aeth = be32_to_cpu(ohdr->u.aeth);
1405 		if (!do_rc_ack(qp, aeth, psn, opcode, 0, rcd))
1406 			goto ack_done;
1407 		/* Get the number of bytes the message was padded by. */
1408 		pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1409 		/*
1410 		 * Check that the data size is >= 0 && <= pmtu.
1411 		 * Remember to account for the AETH header (4) and
1412 		 * ICRC (4).
1413 		 */
1414 		if (unlikely(tlen < (hdrsize + pad + 8)))
1415 			goto ack_len_err;
1416 		/*
1417 		 * If this is a response to a resent RDMA read, we
1418 		 * have to be careful to copy the data to the right
1419 		 * location.
1420 		 */
1421 		wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1422 		qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1423 						  wqe, psn, pmtu);
1424 		goto read_last;
1425 
1426 	case OP(RDMA_READ_RESPONSE_LAST):
1427 		/* ACKs READ req. */
1428 		if (unlikely(qib_cmp24(psn, qp->s_last_psn + 1)))
1429 			goto ack_seq_err;
1430 		if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1431 			goto ack_op_err;
1432 		/* Get the number of bytes the message was padded by. */
1433 		pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1434 		/*
1435 		 * Check that the data size is >= 1 && <= pmtu.
1436 		 * Remember to account for the AETH header (4) and
1437 		 * ICRC (4).
1438 		 */
1439 		if (unlikely(tlen <= (hdrsize + pad + 8)))
1440 			goto ack_len_err;
1441 read_last:
1442 		tlen -= hdrsize + pad + 8;
1443 		if (unlikely(tlen != qp->s_rdma_read_len))
1444 			goto ack_len_err;
1445 		aeth = be32_to_cpu(ohdr->u.aeth);
1446 		rvt_copy_sge(qp, &qp->s_rdma_read_sge,
1447 			     data, tlen, false, false);
1448 		WARN_ON(qp->s_rdma_read_sge.num_sge);
1449 		(void) do_rc_ack(qp, aeth, psn,
1450 				 OP(RDMA_READ_RESPONSE_LAST), 0, rcd);
1451 		goto ack_done;
1452 	}
1453 
1454 ack_op_err:
1455 	status = IB_WC_LOC_QP_OP_ERR;
1456 	goto ack_err;
1457 
1458 ack_seq_err:
1459 	rdma_seq_err(qp, ibp, psn, rcd);
1460 	goto ack_done;
1461 
1462 ack_len_err:
1463 	status = IB_WC_LOC_LEN_ERR;
1464 ack_err:
1465 	if (qp->s_last == qp->s_acked) {
1466 		rvt_send_complete(qp, wqe, status);
1467 		rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1468 	}
1469 ack_done:
1470 	spin_unlock_irqrestore(&qp->s_lock, flags);
1471 bail:
1472 	return;
1473 }
1474 
1475 /**
1476  * qib_rc_rcv_error - process an incoming duplicate or error RC packet
1477  * @ohdr: the other headers for this packet
1478  * @data: the packet data
1479  * @qp: the QP for this packet
1480  * @opcode: the opcode for this packet
1481  * @psn: the packet sequence number for this packet
1482  * @diff: the difference between the PSN and the expected PSN
1483  *
1484  * This is called from qib_rc_rcv() to process an unexpected
1485  * incoming RC packet for the given QP.
1486  * Called at interrupt level.
1487  * Return 1 if no more processing is needed; otherwise return 0 to
1488  * schedule a response to be sent.
1489  */
1490 static int qib_rc_rcv_error(struct ib_other_headers *ohdr,
1491 			    void *data,
1492 			    struct rvt_qp *qp,
1493 			    u32 opcode,
1494 			    u32 psn,
1495 			    int diff,
1496 			    struct qib_ctxtdata *rcd)
1497 {
1498 	struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
1499 	struct rvt_ack_entry *e;
1500 	unsigned long flags;
1501 	u8 i, prev;
1502 	int old_req;
1503 
1504 	if (diff > 0) {
1505 		/*
1506 		 * Packet sequence error.
1507 		 * A NAK will ACK earlier sends and RDMA writes.
1508 		 * Don't queue the NAK if we already sent one.
1509 		 */
1510 		if (!qp->r_nak_state) {
1511 			ibp->rvp.n_rc_seqnak++;
1512 			qp->r_nak_state = IB_NAK_PSN_ERROR;
1513 			/* Use the expected PSN. */
1514 			qp->r_ack_psn = qp->r_psn;
1515 			/*
1516 			 * Wait to send the sequence NAK until all packets
1517 			 * in the receive queue have been processed.
1518 			 * Otherwise, we end up propagating congestion.
1519 			 */
1520 			if (list_empty(&qp->rspwait)) {
1521 				qp->r_flags |= RVT_R_RSP_NAK;
1522 				rvt_get_qp(qp);
1523 				list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
1524 			}
1525 		}
1526 		goto done;
1527 	}
1528 
1529 	/*
1530 	 * Handle a duplicate request.  Don't re-execute SEND, RDMA
1531 	 * write or atomic op.  Don't NAK errors, just silently drop
1532 	 * the duplicate request.  Note that r_sge, r_len, and
1533 	 * r_rcv_len may be in use so don't modify them.
1534 	 *
1535 	 * We are supposed to ACK the earliest duplicate PSN but we
1536 	 * can coalesce an outstanding duplicate ACK.  We have to
1537 	 * send the earliest so that RDMA reads can be restarted at
1538 	 * the requester's expected PSN.
1539 	 *
1540 	 * First, find where this duplicate PSN falls within the
1541 	 * ACKs previously sent.
1542 	 * old_req is true if there is an older response that is scheduled
1543 	 * to be sent before sending this one.
1544 	 */
1545 	e = NULL;
1546 	old_req = 1;
1547 	ibp->rvp.n_rc_dupreq++;
1548 
1549 	spin_lock_irqsave(&qp->s_lock, flags);
1550 
1551 	for (i = qp->r_head_ack_queue; ; i = prev) {
1552 		if (i == qp->s_tail_ack_queue)
1553 			old_req = 0;
1554 		if (i)
1555 			prev = i - 1;
1556 		else
1557 			prev = QIB_MAX_RDMA_ATOMIC;
1558 		if (prev == qp->r_head_ack_queue) {
1559 			e = NULL;
1560 			break;
1561 		}
1562 		e = &qp->s_ack_queue[prev];
1563 		if (!e->opcode) {
1564 			e = NULL;
1565 			break;
1566 		}
1567 		if (qib_cmp24(psn, e->psn) >= 0) {
1568 			if (prev == qp->s_tail_ack_queue &&
1569 			    qib_cmp24(psn, e->lpsn) <= 0)
1570 				old_req = 0;
1571 			break;
1572 		}
1573 	}
1574 	switch (opcode) {
1575 	case OP(RDMA_READ_REQUEST): {
1576 		struct ib_reth *reth;
1577 		u32 offset;
1578 		u32 len;
1579 
1580 		/*
1581 		 * If we didn't find the RDMA read request in the ack queue,
1582 		 * we can ignore this request.
1583 		 */
1584 		if (!e || e->opcode != OP(RDMA_READ_REQUEST))
1585 			goto unlock_done;
1586 		/* RETH comes after BTH */
1587 		reth = &ohdr->u.rc.reth;
1588 		/*
1589 		 * Address range must be a subset of the original
1590 		 * request and start on pmtu boundaries.
1591 		 * We reuse the old ack_queue slot since the requester
1592 		 * should not back up and request an earlier PSN for the
1593 		 * same request.
1594 		 */
1595 		offset = ((psn - e->psn) & QIB_PSN_MASK) *
1596 			qp->pmtu;
1597 		len = be32_to_cpu(reth->length);
1598 		if (unlikely(offset + len != e->rdma_sge.sge_length))
1599 			goto unlock_done;
1600 		if (e->rdma_sge.mr) {
1601 			rvt_put_mr(e->rdma_sge.mr);
1602 			e->rdma_sge.mr = NULL;
1603 		}
1604 		if (len != 0) {
1605 			u32 rkey = be32_to_cpu(reth->rkey);
1606 			u64 vaddr = be64_to_cpu(reth->vaddr);
1607 			int ok;
1608 
1609 			ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr, rkey,
1610 					 IB_ACCESS_REMOTE_READ);
1611 			if (unlikely(!ok))
1612 				goto unlock_done;
1613 		} else {
1614 			e->rdma_sge.vaddr = NULL;
1615 			e->rdma_sge.length = 0;
1616 			e->rdma_sge.sge_length = 0;
1617 		}
1618 		e->psn = psn;
1619 		if (old_req)
1620 			goto unlock_done;
1621 		qp->s_tail_ack_queue = prev;
1622 		break;
1623 	}
1624 
1625 	case OP(COMPARE_SWAP):
1626 	case OP(FETCH_ADD): {
1627 		/*
1628 		 * If we didn't find the atomic request in the ack queue
1629 		 * or the send tasklet is already backed up to send an
1630 		 * earlier entry, we can ignore this request.
1631 		 */
1632 		if (!e || e->opcode != (u8) opcode || old_req)
1633 			goto unlock_done;
1634 		qp->s_tail_ack_queue = prev;
1635 		break;
1636 	}
1637 
1638 	default:
1639 		/*
1640 		 * Ignore this operation if it doesn't request an ACK
1641 		 * or an earlier RDMA read or atomic is going to be resent.
1642 		 */
1643 		if (!(psn & IB_BTH_REQ_ACK) || old_req)
1644 			goto unlock_done;
1645 		/*
1646 		 * Resend the most recent ACK if this request is
1647 		 * after all the previous RDMA reads and atomics.
1648 		 */
1649 		if (i == qp->r_head_ack_queue) {
1650 			spin_unlock_irqrestore(&qp->s_lock, flags);
1651 			qp->r_nak_state = 0;
1652 			qp->r_ack_psn = qp->r_psn - 1;
1653 			goto send_ack;
1654 		}
1655 		/*
1656 		 * Try to send a simple ACK to work around a Mellanox bug
1657 		 * which doesn't accept a RDMA read response or atomic
1658 		 * response as an ACK for earlier SENDs or RDMA writes.
1659 		 */
1660 		if (!(qp->s_flags & RVT_S_RESP_PENDING)) {
1661 			spin_unlock_irqrestore(&qp->s_lock, flags);
1662 			qp->r_nak_state = 0;
1663 			qp->r_ack_psn = qp->s_ack_queue[i].psn - 1;
1664 			goto send_ack;
1665 		}
1666 		/*
1667 		 * Resend the RDMA read or atomic op which
1668 		 * ACKs this duplicate request.
1669 		 */
1670 		qp->s_tail_ack_queue = i;
1671 		break;
1672 	}
1673 	qp->s_ack_state = OP(ACKNOWLEDGE);
1674 	qp->s_flags |= RVT_S_RESP_PENDING;
1675 	qp->r_nak_state = 0;
1676 	qib_schedule_send(qp);
1677 
1678 unlock_done:
1679 	spin_unlock_irqrestore(&qp->s_lock, flags);
1680 done:
1681 	return 1;
1682 
1683 send_ack:
1684 	return 0;
1685 }
1686 
1687 static inline void qib_update_ack_queue(struct rvt_qp *qp, unsigned n)
1688 {
1689 	unsigned next;
1690 
1691 	next = n + 1;
1692 	if (next > QIB_MAX_RDMA_ATOMIC)
1693 		next = 0;
1694 	qp->s_tail_ack_queue = next;
1695 	qp->s_ack_state = OP(ACKNOWLEDGE);
1696 }
1697 
1698 /**
1699  * qib_rc_rcv - process an incoming RC packet
1700  * @rcd: the context pointer
1701  * @hdr: the header of this packet
1702  * @has_grh: true if the header has a GRH
1703  * @data: the packet data
1704  * @tlen: the packet length
1705  * @qp: the QP for this packet
1706  *
1707  * This is called from qib_qp_rcv() to process an incoming RC packet
1708  * for the given QP.
1709  * Called at interrupt level.
1710  */
1711 void qib_rc_rcv(struct qib_ctxtdata *rcd, struct ib_header *hdr,
1712 		int has_grh, void *data, u32 tlen, struct rvt_qp *qp)
1713 {
1714 	struct qib_ibport *ibp = &rcd->ppd->ibport_data;
1715 	struct ib_other_headers *ohdr;
1716 	u32 opcode;
1717 	u32 hdrsize;
1718 	u32 psn;
1719 	u32 pad;
1720 	struct ib_wc wc;
1721 	u32 pmtu = qp->pmtu;
1722 	int diff;
1723 	struct ib_reth *reth;
1724 	unsigned long flags;
1725 	int ret;
1726 
1727 	/* Check for GRH */
1728 	if (!has_grh) {
1729 		ohdr = &hdr->u.oth;
1730 		hdrsize = 8 + 12;       /* LRH + BTH */
1731 	} else {
1732 		ohdr = &hdr->u.l.oth;
1733 		hdrsize = 8 + 40 + 12;  /* LRH + GRH + BTH */
1734 	}
1735 
1736 	opcode = be32_to_cpu(ohdr->bth[0]);
1737 	if (qib_ruc_check_hdr(ibp, hdr, has_grh, qp, opcode))
1738 		return;
1739 
1740 	psn = be32_to_cpu(ohdr->bth[2]);
1741 	opcode >>= 24;
1742 
1743 	/*
1744 	 * Process responses (ACKs) before anything else.  Note that the
1745 	 * packet sequence number will be for something in the send work
1746 	 * queue rather than the expected receive packet sequence number.
1747 	 * In other words, this QP is the requester.
1748 	 */
1749 	if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
1750 	    opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
1751 		qib_rc_rcv_resp(ibp, ohdr, data, tlen, qp, opcode, psn,
1752 				hdrsize, pmtu, rcd);
1753 		return;
1754 	}
1755 
1756 	/* Compute 24 bits worth of difference. */
1757 	diff = qib_cmp24(psn, qp->r_psn);
1758 	if (unlikely(diff)) {
1759 		if (qib_rc_rcv_error(ohdr, data, qp, opcode, psn, diff, rcd))
1760 			return;
1761 		goto send_ack;
1762 	}
1763 
1764 	/* Check for opcode sequence errors. */
1765 	switch (qp->r_state) {
1766 	case OP(SEND_FIRST):
1767 	case OP(SEND_MIDDLE):
1768 		if (opcode == OP(SEND_MIDDLE) ||
1769 		    opcode == OP(SEND_LAST) ||
1770 		    opcode == OP(SEND_LAST_WITH_IMMEDIATE))
1771 			break;
1772 		goto nack_inv;
1773 
1774 	case OP(RDMA_WRITE_FIRST):
1775 	case OP(RDMA_WRITE_MIDDLE):
1776 		if (opcode == OP(RDMA_WRITE_MIDDLE) ||
1777 		    opcode == OP(RDMA_WRITE_LAST) ||
1778 		    opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1779 			break;
1780 		goto nack_inv;
1781 
1782 	default:
1783 		if (opcode == OP(SEND_MIDDLE) ||
1784 		    opcode == OP(SEND_LAST) ||
1785 		    opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
1786 		    opcode == OP(RDMA_WRITE_MIDDLE) ||
1787 		    opcode == OP(RDMA_WRITE_LAST) ||
1788 		    opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1789 			goto nack_inv;
1790 		/*
1791 		 * Note that it is up to the requester to not send a new
1792 		 * RDMA read or atomic operation before receiving an ACK
1793 		 * for the previous operation.
1794 		 */
1795 		break;
1796 	}
1797 
1798 	if (qp->state == IB_QPS_RTR && !(qp->r_flags & RVT_R_COMM_EST))
1799 		rvt_comm_est(qp);
1800 
1801 	/* OK, process the packet. */
1802 	switch (opcode) {
1803 	case OP(SEND_FIRST):
1804 		ret = rvt_get_rwqe(qp, false);
1805 		if (ret < 0)
1806 			goto nack_op_err;
1807 		if (!ret)
1808 			goto rnr_nak;
1809 		qp->r_rcv_len = 0;
1810 		/* FALLTHROUGH */
1811 	case OP(SEND_MIDDLE):
1812 	case OP(RDMA_WRITE_MIDDLE):
1813 send_middle:
1814 		/* Check for invalid length PMTU or posted rwqe len. */
1815 		if (unlikely(tlen != (hdrsize + pmtu + 4)))
1816 			goto nack_inv;
1817 		qp->r_rcv_len += pmtu;
1818 		if (unlikely(qp->r_rcv_len > qp->r_len))
1819 			goto nack_inv;
1820 		rvt_copy_sge(qp, &qp->r_sge, data, pmtu, true, false);
1821 		break;
1822 
1823 	case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
1824 		/* consume RWQE */
1825 		ret = rvt_get_rwqe(qp, true);
1826 		if (ret < 0)
1827 			goto nack_op_err;
1828 		if (!ret)
1829 			goto rnr_nak;
1830 		goto send_last_imm;
1831 
1832 	case OP(SEND_ONLY):
1833 	case OP(SEND_ONLY_WITH_IMMEDIATE):
1834 		ret = rvt_get_rwqe(qp, false);
1835 		if (ret < 0)
1836 			goto nack_op_err;
1837 		if (!ret)
1838 			goto rnr_nak;
1839 		qp->r_rcv_len = 0;
1840 		if (opcode == OP(SEND_ONLY))
1841 			goto no_immediate_data;
1842 		/* fall through -- for SEND_ONLY_WITH_IMMEDIATE */
1843 	case OP(SEND_LAST_WITH_IMMEDIATE):
1844 send_last_imm:
1845 		wc.ex.imm_data = ohdr->u.imm_data;
1846 		hdrsize += 4;
1847 		wc.wc_flags = IB_WC_WITH_IMM;
1848 		goto send_last;
1849 	case OP(SEND_LAST):
1850 	case OP(RDMA_WRITE_LAST):
1851 no_immediate_data:
1852 		wc.wc_flags = 0;
1853 		wc.ex.imm_data = 0;
1854 send_last:
1855 		/* Get the number of bytes the message was padded by. */
1856 		pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1857 		/* Check for invalid length. */
1858 		/* XXX LAST len should be >= 1 */
1859 		if (unlikely(tlen < (hdrsize + pad + 4)))
1860 			goto nack_inv;
1861 		/* Don't count the CRC. */
1862 		tlen -= (hdrsize + pad + 4);
1863 		wc.byte_len = tlen + qp->r_rcv_len;
1864 		if (unlikely(wc.byte_len > qp->r_len))
1865 			goto nack_inv;
1866 		rvt_copy_sge(qp, &qp->r_sge, data, tlen, true, false);
1867 		rvt_put_ss(&qp->r_sge);
1868 		qp->r_msn++;
1869 		if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
1870 			break;
1871 		wc.wr_id = qp->r_wr_id;
1872 		wc.status = IB_WC_SUCCESS;
1873 		if (opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE) ||
1874 		    opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE))
1875 			wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
1876 		else
1877 			wc.opcode = IB_WC_RECV;
1878 		wc.qp = &qp->ibqp;
1879 		wc.src_qp = qp->remote_qpn;
1880 		wc.slid = rdma_ah_get_dlid(&qp->remote_ah_attr);
1881 		wc.sl = rdma_ah_get_sl(&qp->remote_ah_attr);
1882 		/* zero fields that are N/A */
1883 		wc.vendor_err = 0;
1884 		wc.pkey_index = 0;
1885 		wc.dlid_path_bits = 0;
1886 		wc.port_num = 0;
1887 		/* Signal completion event if the solicited bit is set. */
1888 		rvt_recv_cq(qp, &wc, ib_bth_is_solicited(ohdr));
1889 		break;
1890 
1891 	case OP(RDMA_WRITE_FIRST):
1892 	case OP(RDMA_WRITE_ONLY):
1893 	case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
1894 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
1895 			goto nack_inv;
1896 		/* consume RWQE */
1897 		reth = &ohdr->u.rc.reth;
1898 		hdrsize += sizeof(*reth);
1899 		qp->r_len = be32_to_cpu(reth->length);
1900 		qp->r_rcv_len = 0;
1901 		qp->r_sge.sg_list = NULL;
1902 		if (qp->r_len != 0) {
1903 			u32 rkey = be32_to_cpu(reth->rkey);
1904 			u64 vaddr = be64_to_cpu(reth->vaddr);
1905 			int ok;
1906 
1907 			/* Check rkey & NAK */
1908 			ok = rvt_rkey_ok(qp, &qp->r_sge.sge, qp->r_len, vaddr,
1909 					 rkey, IB_ACCESS_REMOTE_WRITE);
1910 			if (unlikely(!ok))
1911 				goto nack_acc;
1912 			qp->r_sge.num_sge = 1;
1913 		} else {
1914 			qp->r_sge.num_sge = 0;
1915 			qp->r_sge.sge.mr = NULL;
1916 			qp->r_sge.sge.vaddr = NULL;
1917 			qp->r_sge.sge.length = 0;
1918 			qp->r_sge.sge.sge_length = 0;
1919 		}
1920 		if (opcode == OP(RDMA_WRITE_FIRST))
1921 			goto send_middle;
1922 		else if (opcode == OP(RDMA_WRITE_ONLY))
1923 			goto no_immediate_data;
1924 		ret = rvt_get_rwqe(qp, true);
1925 		if (ret < 0)
1926 			goto nack_op_err;
1927 		if (!ret) {
1928 			rvt_put_ss(&qp->r_sge);
1929 			goto rnr_nak;
1930 		}
1931 		wc.ex.imm_data = ohdr->u.rc.imm_data;
1932 		hdrsize += 4;
1933 		wc.wc_flags = IB_WC_WITH_IMM;
1934 		goto send_last;
1935 
1936 	case OP(RDMA_READ_REQUEST): {
1937 		struct rvt_ack_entry *e;
1938 		u32 len;
1939 		u8 next;
1940 
1941 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
1942 			goto nack_inv;
1943 		next = qp->r_head_ack_queue + 1;
1944 		/* s_ack_queue is size QIB_MAX_RDMA_ATOMIC+1 so use > not >= */
1945 		if (next > QIB_MAX_RDMA_ATOMIC)
1946 			next = 0;
1947 		spin_lock_irqsave(&qp->s_lock, flags);
1948 		if (unlikely(next == qp->s_tail_ack_queue)) {
1949 			if (!qp->s_ack_queue[next].sent)
1950 				goto nack_inv_unlck;
1951 			qib_update_ack_queue(qp, next);
1952 		}
1953 		e = &qp->s_ack_queue[qp->r_head_ack_queue];
1954 		if (e->opcode == OP(RDMA_READ_REQUEST) && e->rdma_sge.mr) {
1955 			rvt_put_mr(e->rdma_sge.mr);
1956 			e->rdma_sge.mr = NULL;
1957 		}
1958 		reth = &ohdr->u.rc.reth;
1959 		len = be32_to_cpu(reth->length);
1960 		if (len) {
1961 			u32 rkey = be32_to_cpu(reth->rkey);
1962 			u64 vaddr = be64_to_cpu(reth->vaddr);
1963 			int ok;
1964 
1965 			/* Check rkey & NAK */
1966 			ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr,
1967 					 rkey, IB_ACCESS_REMOTE_READ);
1968 			if (unlikely(!ok))
1969 				goto nack_acc_unlck;
1970 			/*
1971 			 * Update the next expected PSN.  We add 1 later
1972 			 * below, so only add the remainder here.
1973 			 */
1974 			qp->r_psn += rvt_div_mtu(qp, len - 1);
1975 		} else {
1976 			e->rdma_sge.mr = NULL;
1977 			e->rdma_sge.vaddr = NULL;
1978 			e->rdma_sge.length = 0;
1979 			e->rdma_sge.sge_length = 0;
1980 		}
1981 		e->opcode = opcode;
1982 		e->sent = 0;
1983 		e->psn = psn;
1984 		e->lpsn = qp->r_psn;
1985 		/*
1986 		 * We need to increment the MSN here instead of when we
1987 		 * finish sending the result since a duplicate request would
1988 		 * increment it more than once.
1989 		 */
1990 		qp->r_msn++;
1991 		qp->r_psn++;
1992 		qp->r_state = opcode;
1993 		qp->r_nak_state = 0;
1994 		qp->r_head_ack_queue = next;
1995 
1996 		/* Schedule the send tasklet. */
1997 		qp->s_flags |= RVT_S_RESP_PENDING;
1998 		qib_schedule_send(qp);
1999 
2000 		goto sunlock;
2001 	}
2002 
2003 	case OP(COMPARE_SWAP):
2004 	case OP(FETCH_ADD): {
2005 		struct ib_atomic_eth *ateth;
2006 		struct rvt_ack_entry *e;
2007 		u64 vaddr;
2008 		atomic64_t *maddr;
2009 		u64 sdata;
2010 		u32 rkey;
2011 		u8 next;
2012 
2013 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
2014 			goto nack_inv;
2015 		next = qp->r_head_ack_queue + 1;
2016 		if (next > QIB_MAX_RDMA_ATOMIC)
2017 			next = 0;
2018 		spin_lock_irqsave(&qp->s_lock, flags);
2019 		if (unlikely(next == qp->s_tail_ack_queue)) {
2020 			if (!qp->s_ack_queue[next].sent)
2021 				goto nack_inv_unlck;
2022 			qib_update_ack_queue(qp, next);
2023 		}
2024 		e = &qp->s_ack_queue[qp->r_head_ack_queue];
2025 		if (e->opcode == OP(RDMA_READ_REQUEST) && e->rdma_sge.mr) {
2026 			rvt_put_mr(e->rdma_sge.mr);
2027 			e->rdma_sge.mr = NULL;
2028 		}
2029 		ateth = &ohdr->u.atomic_eth;
2030 		vaddr = get_ib_ateth_vaddr(ateth);
2031 		if (unlikely(vaddr & (sizeof(u64) - 1)))
2032 			goto nack_inv_unlck;
2033 		rkey = be32_to_cpu(ateth->rkey);
2034 		/* Check rkey & NAK */
2035 		if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64),
2036 					  vaddr, rkey,
2037 					  IB_ACCESS_REMOTE_ATOMIC)))
2038 			goto nack_acc_unlck;
2039 		/* Perform atomic OP and save result. */
2040 		maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
2041 		sdata = get_ib_ateth_swap(ateth);
2042 		e->atomic_data = (opcode == OP(FETCH_ADD)) ?
2043 			(u64) atomic64_add_return(sdata, maddr) - sdata :
2044 			(u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
2045 				      get_ib_ateth_compare(ateth),
2046 				      sdata);
2047 		rvt_put_mr(qp->r_sge.sge.mr);
2048 		qp->r_sge.num_sge = 0;
2049 		e->opcode = opcode;
2050 		e->sent = 0;
2051 		e->psn = psn;
2052 		e->lpsn = psn;
2053 		qp->r_msn++;
2054 		qp->r_psn++;
2055 		qp->r_state = opcode;
2056 		qp->r_nak_state = 0;
2057 		qp->r_head_ack_queue = next;
2058 
2059 		/* Schedule the send tasklet. */
2060 		qp->s_flags |= RVT_S_RESP_PENDING;
2061 		qib_schedule_send(qp);
2062 
2063 		goto sunlock;
2064 	}
2065 
2066 	default:
2067 		/* NAK unknown opcodes. */
2068 		goto nack_inv;
2069 	}
2070 	qp->r_psn++;
2071 	qp->r_state = opcode;
2072 	qp->r_ack_psn = psn;
2073 	qp->r_nak_state = 0;
2074 	/* Send an ACK if requested or required. */
2075 	if (psn & (1 << 31))
2076 		goto send_ack;
2077 	return;
2078 
2079 rnr_nak:
2080 	qp->r_nak_state = IB_RNR_NAK | qp->r_min_rnr_timer;
2081 	qp->r_ack_psn = qp->r_psn;
2082 	/* Queue RNR NAK for later */
2083 	if (list_empty(&qp->rspwait)) {
2084 		qp->r_flags |= RVT_R_RSP_NAK;
2085 		rvt_get_qp(qp);
2086 		list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
2087 	}
2088 	return;
2089 
2090 nack_op_err:
2091 	rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
2092 	qp->r_nak_state = IB_NAK_REMOTE_OPERATIONAL_ERROR;
2093 	qp->r_ack_psn = qp->r_psn;
2094 	/* Queue NAK for later */
2095 	if (list_empty(&qp->rspwait)) {
2096 		qp->r_flags |= RVT_R_RSP_NAK;
2097 		rvt_get_qp(qp);
2098 		list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
2099 	}
2100 	return;
2101 
2102 nack_inv_unlck:
2103 	spin_unlock_irqrestore(&qp->s_lock, flags);
2104 nack_inv:
2105 	rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
2106 	qp->r_nak_state = IB_NAK_INVALID_REQUEST;
2107 	qp->r_ack_psn = qp->r_psn;
2108 	/* Queue NAK for later */
2109 	if (list_empty(&qp->rspwait)) {
2110 		qp->r_flags |= RVT_R_RSP_NAK;
2111 		rvt_get_qp(qp);
2112 		list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
2113 	}
2114 	return;
2115 
2116 nack_acc_unlck:
2117 	spin_unlock_irqrestore(&qp->s_lock, flags);
2118 nack_acc:
2119 	rvt_rc_error(qp, IB_WC_LOC_PROT_ERR);
2120 	qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
2121 	qp->r_ack_psn = qp->r_psn;
2122 send_ack:
2123 	qib_send_rc_ack(qp);
2124 	return;
2125 
2126 sunlock:
2127 	spin_unlock_irqrestore(&qp->s_lock, flags);
2128 }
2129