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/spinlock.h>
35 
36 #include "qib.h"
37 #include "qib_mad.h"
38 
39 /*
40  * Convert the AETH RNR timeout code into the number of microseconds.
41  */
42 const u32 ib_qib_rnr_table[32] = {
43 	655360,	/* 00: 655.36 */
44 	10,	/* 01:    .01 */
45 	20,	/* 02     .02 */
46 	30,	/* 03:    .03 */
47 	40,	/* 04:    .04 */
48 	60,	/* 05:    .06 */
49 	80,	/* 06:    .08 */
50 	120,	/* 07:    .12 */
51 	160,	/* 08:    .16 */
52 	240,	/* 09:    .24 */
53 	320,	/* 0A:    .32 */
54 	480,	/* 0B:    .48 */
55 	640,	/* 0C:    .64 */
56 	960,	/* 0D:    .96 */
57 	1280,	/* 0E:   1.28 */
58 	1920,	/* 0F:   1.92 */
59 	2560,	/* 10:   2.56 */
60 	3840,	/* 11:   3.84 */
61 	5120,	/* 12:   5.12 */
62 	7680,	/* 13:   7.68 */
63 	10240,	/* 14:  10.24 */
64 	15360,	/* 15:  15.36 */
65 	20480,	/* 16:  20.48 */
66 	30720,	/* 17:  30.72 */
67 	40960,	/* 18:  40.96 */
68 	61440,	/* 19:  61.44 */
69 	81920,	/* 1A:  81.92 */
70 	122880,	/* 1B: 122.88 */
71 	163840,	/* 1C: 163.84 */
72 	245760,	/* 1D: 245.76 */
73 	327680,	/* 1E: 327.68 */
74 	491520	/* 1F: 491.52 */
75 };
76 
77 /*
78  * Validate a RWQE and fill in the SGE state.
79  * Return 1 if OK.
80  */
81 static int qib_init_sge(struct qib_qp *qp, struct qib_rwqe *wqe)
82 {
83 	int i, j, ret;
84 	struct ib_wc wc;
85 	struct qib_lkey_table *rkt;
86 	struct qib_pd *pd;
87 	struct qib_sge_state *ss;
88 
89 	rkt = &to_idev(qp->ibqp.device)->lk_table;
90 	pd = to_ipd(qp->ibqp.srq ? qp->ibqp.srq->pd : qp->ibqp.pd);
91 	ss = &qp->r_sge;
92 	ss->sg_list = qp->r_sg_list;
93 	qp->r_len = 0;
94 	for (i = j = 0; i < wqe->num_sge; i++) {
95 		if (wqe->sg_list[i].length == 0)
96 			continue;
97 		/* Check LKEY */
98 		if (!qib_lkey_ok(rkt, pd, j ? &ss->sg_list[j - 1] : &ss->sge,
99 				 &wqe->sg_list[i], IB_ACCESS_LOCAL_WRITE))
100 			goto bad_lkey;
101 		qp->r_len += wqe->sg_list[i].length;
102 		j++;
103 	}
104 	ss->num_sge = j;
105 	ss->total_len = qp->r_len;
106 	ret = 1;
107 	goto bail;
108 
109 bad_lkey:
110 	while (j) {
111 		struct qib_sge *sge = --j ? &ss->sg_list[j - 1] : &ss->sge;
112 
113 		atomic_dec(&sge->mr->refcount);
114 	}
115 	ss->num_sge = 0;
116 	memset(&wc, 0, sizeof(wc));
117 	wc.wr_id = wqe->wr_id;
118 	wc.status = IB_WC_LOC_PROT_ERR;
119 	wc.opcode = IB_WC_RECV;
120 	wc.qp = &qp->ibqp;
121 	/* Signal solicited completion event. */
122 	qib_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
123 	ret = 0;
124 bail:
125 	return ret;
126 }
127 
128 /**
129  * qib_get_rwqe - copy the next RWQE into the QP's RWQE
130  * @qp: the QP
131  * @wr_id_only: update qp->r_wr_id only, not qp->r_sge
132  *
133  * Return -1 if there is a local error, 0 if no RWQE is available,
134  * otherwise return 1.
135  *
136  * Can be called from interrupt level.
137  */
138 int qib_get_rwqe(struct qib_qp *qp, int wr_id_only)
139 {
140 	unsigned long flags;
141 	struct qib_rq *rq;
142 	struct qib_rwq *wq;
143 	struct qib_srq *srq;
144 	struct qib_rwqe *wqe;
145 	void (*handler)(struct ib_event *, void *);
146 	u32 tail;
147 	int ret;
148 
149 	if (qp->ibqp.srq) {
150 		srq = to_isrq(qp->ibqp.srq);
151 		handler = srq->ibsrq.event_handler;
152 		rq = &srq->rq;
153 	} else {
154 		srq = NULL;
155 		handler = NULL;
156 		rq = &qp->r_rq;
157 	}
158 
159 	spin_lock_irqsave(&rq->lock, flags);
160 	if (!(ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK)) {
161 		ret = 0;
162 		goto unlock;
163 	}
164 
165 	wq = rq->wq;
166 	tail = wq->tail;
167 	/* Validate tail before using it since it is user writable. */
168 	if (tail >= rq->size)
169 		tail = 0;
170 	if (unlikely(tail == wq->head)) {
171 		ret = 0;
172 		goto unlock;
173 	}
174 	/* Make sure entry is read after head index is read. */
175 	smp_rmb();
176 	wqe = get_rwqe_ptr(rq, tail);
177 	/*
178 	 * Even though we update the tail index in memory, the verbs
179 	 * consumer is not supposed to post more entries until a
180 	 * completion is generated.
181 	 */
182 	if (++tail >= rq->size)
183 		tail = 0;
184 	wq->tail = tail;
185 	if (!wr_id_only && !qib_init_sge(qp, wqe)) {
186 		ret = -1;
187 		goto unlock;
188 	}
189 	qp->r_wr_id = wqe->wr_id;
190 
191 	ret = 1;
192 	set_bit(QIB_R_WRID_VALID, &qp->r_aflags);
193 	if (handler) {
194 		u32 n;
195 
196 		/*
197 		 * Validate head pointer value and compute
198 		 * the number of remaining WQEs.
199 		 */
200 		n = wq->head;
201 		if (n >= rq->size)
202 			n = 0;
203 		if (n < tail)
204 			n += rq->size - tail;
205 		else
206 			n -= tail;
207 		if (n < srq->limit) {
208 			struct ib_event ev;
209 
210 			srq->limit = 0;
211 			spin_unlock_irqrestore(&rq->lock, flags);
212 			ev.device = qp->ibqp.device;
213 			ev.element.srq = qp->ibqp.srq;
214 			ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
215 			handler(&ev, srq->ibsrq.srq_context);
216 			goto bail;
217 		}
218 	}
219 unlock:
220 	spin_unlock_irqrestore(&rq->lock, flags);
221 bail:
222 	return ret;
223 }
224 
225 /*
226  * Switch to alternate path.
227  * The QP s_lock should be held and interrupts disabled.
228  */
229 void qib_migrate_qp(struct qib_qp *qp)
230 {
231 	struct ib_event ev;
232 
233 	qp->s_mig_state = IB_MIG_MIGRATED;
234 	qp->remote_ah_attr = qp->alt_ah_attr;
235 	qp->port_num = qp->alt_ah_attr.port_num;
236 	qp->s_pkey_index = qp->s_alt_pkey_index;
237 
238 	ev.device = qp->ibqp.device;
239 	ev.element.qp = &qp->ibqp;
240 	ev.event = IB_EVENT_PATH_MIG;
241 	qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
242 }
243 
244 static __be64 get_sguid(struct qib_ibport *ibp, unsigned index)
245 {
246 	if (!index) {
247 		struct qib_pportdata *ppd = ppd_from_ibp(ibp);
248 
249 		return ppd->guid;
250 	} else
251 		return ibp->guids[index - 1];
252 }
253 
254 static int gid_ok(union ib_gid *gid, __be64 gid_prefix, __be64 id)
255 {
256 	return (gid->global.interface_id == id &&
257 		(gid->global.subnet_prefix == gid_prefix ||
258 		 gid->global.subnet_prefix == IB_DEFAULT_GID_PREFIX));
259 }
260 
261 /*
262  *
263  * This should be called with the QP s_lock held.
264  */
265 int qib_ruc_check_hdr(struct qib_ibport *ibp, struct qib_ib_header *hdr,
266 		      int has_grh, struct qib_qp *qp, u32 bth0)
267 {
268 	__be64 guid;
269 
270 	if (qp->s_mig_state == IB_MIG_ARMED && (bth0 & IB_BTH_MIG_REQ)) {
271 		if (!has_grh) {
272 			if (qp->alt_ah_attr.ah_flags & IB_AH_GRH)
273 				goto err;
274 		} else {
275 			if (!(qp->alt_ah_attr.ah_flags & IB_AH_GRH))
276 				goto err;
277 			guid = get_sguid(ibp, qp->alt_ah_attr.grh.sgid_index);
278 			if (!gid_ok(&hdr->u.l.grh.dgid, ibp->gid_prefix, guid))
279 				goto err;
280 			if (!gid_ok(&hdr->u.l.grh.sgid,
281 			    qp->alt_ah_attr.grh.dgid.global.subnet_prefix,
282 			    qp->alt_ah_attr.grh.dgid.global.interface_id))
283 				goto err;
284 		}
285 		if (!qib_pkey_ok((u16)bth0,
286 				 qib_get_pkey(ibp, qp->s_alt_pkey_index))) {
287 			qib_bad_pqkey(ibp, IB_NOTICE_TRAP_BAD_PKEY,
288 				      (u16)bth0,
289 				      (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF,
290 				      0, qp->ibqp.qp_num,
291 				      hdr->lrh[3], hdr->lrh[1]);
292 			goto err;
293 		}
294 		/* Validate the SLID. See Ch. 9.6.1.5 and 17.2.8 */
295 		if (be16_to_cpu(hdr->lrh[3]) != qp->alt_ah_attr.dlid ||
296 		    ppd_from_ibp(ibp)->port != qp->alt_ah_attr.port_num)
297 			goto err;
298 		qib_migrate_qp(qp);
299 	} else {
300 		if (!has_grh) {
301 			if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
302 				goto err;
303 		} else {
304 			if (!(qp->remote_ah_attr.ah_flags & IB_AH_GRH))
305 				goto err;
306 			guid = get_sguid(ibp,
307 					 qp->remote_ah_attr.grh.sgid_index);
308 			if (!gid_ok(&hdr->u.l.grh.dgid, ibp->gid_prefix, guid))
309 				goto err;
310 			if (!gid_ok(&hdr->u.l.grh.sgid,
311 			    qp->remote_ah_attr.grh.dgid.global.subnet_prefix,
312 			    qp->remote_ah_attr.grh.dgid.global.interface_id))
313 				goto err;
314 		}
315 		if (!qib_pkey_ok((u16)bth0,
316 				 qib_get_pkey(ibp, qp->s_pkey_index))) {
317 			qib_bad_pqkey(ibp, IB_NOTICE_TRAP_BAD_PKEY,
318 				      (u16)bth0,
319 				      (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF,
320 				      0, qp->ibqp.qp_num,
321 				      hdr->lrh[3], hdr->lrh[1]);
322 			goto err;
323 		}
324 		/* Validate the SLID. See Ch. 9.6.1.5 */
325 		if (be16_to_cpu(hdr->lrh[3]) != qp->remote_ah_attr.dlid ||
326 		    ppd_from_ibp(ibp)->port != qp->port_num)
327 			goto err;
328 		if (qp->s_mig_state == IB_MIG_REARM &&
329 		    !(bth0 & IB_BTH_MIG_REQ))
330 			qp->s_mig_state = IB_MIG_ARMED;
331 	}
332 
333 	return 0;
334 
335 err:
336 	return 1;
337 }
338 
339 /**
340  * qib_ruc_loopback - handle UC and RC lookback requests
341  * @sqp: the sending QP
342  *
343  * This is called from qib_do_send() to
344  * forward a WQE addressed to the same HCA.
345  * Note that although we are single threaded due to the tasklet, we still
346  * have to protect against post_send().  We don't have to worry about
347  * receive interrupts since this is a connected protocol and all packets
348  * will pass through here.
349  */
350 static void qib_ruc_loopback(struct qib_qp *sqp)
351 {
352 	struct qib_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num);
353 	struct qib_qp *qp;
354 	struct qib_swqe *wqe;
355 	struct qib_sge *sge;
356 	unsigned long flags;
357 	struct ib_wc wc;
358 	u64 sdata;
359 	atomic64_t *maddr;
360 	enum ib_wc_status send_status;
361 	int release;
362 	int ret;
363 
364 	/*
365 	 * Note that we check the responder QP state after
366 	 * checking the requester's state.
367 	 */
368 	qp = qib_lookup_qpn(ibp, sqp->remote_qpn);
369 
370 	spin_lock_irqsave(&sqp->s_lock, flags);
371 
372 	/* Return if we are already busy processing a work request. */
373 	if ((sqp->s_flags & (QIB_S_BUSY | QIB_S_ANY_WAIT)) ||
374 	    !(ib_qib_state_ops[sqp->state] & QIB_PROCESS_OR_FLUSH_SEND))
375 		goto unlock;
376 
377 	sqp->s_flags |= QIB_S_BUSY;
378 
379 again:
380 	if (sqp->s_last == sqp->s_head)
381 		goto clr_busy;
382 	wqe = get_swqe_ptr(sqp, sqp->s_last);
383 
384 	/* Return if it is not OK to start a new work reqeust. */
385 	if (!(ib_qib_state_ops[sqp->state] & QIB_PROCESS_NEXT_SEND_OK)) {
386 		if (!(ib_qib_state_ops[sqp->state] & QIB_FLUSH_SEND))
387 			goto clr_busy;
388 		/* We are in the error state, flush the work request. */
389 		send_status = IB_WC_WR_FLUSH_ERR;
390 		goto flush_send;
391 	}
392 
393 	/*
394 	 * We can rely on the entry not changing without the s_lock
395 	 * being held until we update s_last.
396 	 * We increment s_cur to indicate s_last is in progress.
397 	 */
398 	if (sqp->s_last == sqp->s_cur) {
399 		if (++sqp->s_cur >= sqp->s_size)
400 			sqp->s_cur = 0;
401 	}
402 	spin_unlock_irqrestore(&sqp->s_lock, flags);
403 
404 	if (!qp || !(ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK) ||
405 	    qp->ibqp.qp_type != sqp->ibqp.qp_type) {
406 		ibp->n_pkt_drops++;
407 		/*
408 		 * For RC, the requester would timeout and retry so
409 		 * shortcut the timeouts and just signal too many retries.
410 		 */
411 		if (sqp->ibqp.qp_type == IB_QPT_RC)
412 			send_status = IB_WC_RETRY_EXC_ERR;
413 		else
414 			send_status = IB_WC_SUCCESS;
415 		goto serr;
416 	}
417 
418 	memset(&wc, 0, sizeof wc);
419 	send_status = IB_WC_SUCCESS;
420 
421 	release = 1;
422 	sqp->s_sge.sge = wqe->sg_list[0];
423 	sqp->s_sge.sg_list = wqe->sg_list + 1;
424 	sqp->s_sge.num_sge = wqe->wr.num_sge;
425 	sqp->s_len = wqe->length;
426 	switch (wqe->wr.opcode) {
427 	case IB_WR_SEND_WITH_IMM:
428 		wc.wc_flags = IB_WC_WITH_IMM;
429 		wc.ex.imm_data = wqe->wr.ex.imm_data;
430 		/* FALLTHROUGH */
431 	case IB_WR_SEND:
432 		ret = qib_get_rwqe(qp, 0);
433 		if (ret < 0)
434 			goto op_err;
435 		if (!ret)
436 			goto rnr_nak;
437 		break;
438 
439 	case IB_WR_RDMA_WRITE_WITH_IMM:
440 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
441 			goto inv_err;
442 		wc.wc_flags = IB_WC_WITH_IMM;
443 		wc.ex.imm_data = wqe->wr.ex.imm_data;
444 		ret = qib_get_rwqe(qp, 1);
445 		if (ret < 0)
446 			goto op_err;
447 		if (!ret)
448 			goto rnr_nak;
449 		/* FALLTHROUGH */
450 	case IB_WR_RDMA_WRITE:
451 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
452 			goto inv_err;
453 		if (wqe->length == 0)
454 			break;
455 		if (unlikely(!qib_rkey_ok(qp, &qp->r_sge.sge, wqe->length,
456 					  wqe->wr.wr.rdma.remote_addr,
457 					  wqe->wr.wr.rdma.rkey,
458 					  IB_ACCESS_REMOTE_WRITE)))
459 			goto acc_err;
460 		qp->r_sge.sg_list = NULL;
461 		qp->r_sge.num_sge = 1;
462 		qp->r_sge.total_len = wqe->length;
463 		break;
464 
465 	case IB_WR_RDMA_READ:
466 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
467 			goto inv_err;
468 		if (unlikely(!qib_rkey_ok(qp, &sqp->s_sge.sge, wqe->length,
469 					  wqe->wr.wr.rdma.remote_addr,
470 					  wqe->wr.wr.rdma.rkey,
471 					  IB_ACCESS_REMOTE_READ)))
472 			goto acc_err;
473 		release = 0;
474 		sqp->s_sge.sg_list = NULL;
475 		sqp->s_sge.num_sge = 1;
476 		qp->r_sge.sge = wqe->sg_list[0];
477 		qp->r_sge.sg_list = wqe->sg_list + 1;
478 		qp->r_sge.num_sge = wqe->wr.num_sge;
479 		qp->r_sge.total_len = wqe->length;
480 		break;
481 
482 	case IB_WR_ATOMIC_CMP_AND_SWP:
483 	case IB_WR_ATOMIC_FETCH_AND_ADD:
484 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
485 			goto inv_err;
486 		if (unlikely(!qib_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64),
487 					  wqe->wr.wr.atomic.remote_addr,
488 					  wqe->wr.wr.atomic.rkey,
489 					  IB_ACCESS_REMOTE_ATOMIC)))
490 			goto acc_err;
491 		/* Perform atomic OP and save result. */
492 		maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
493 		sdata = wqe->wr.wr.atomic.compare_add;
494 		*(u64 *) sqp->s_sge.sge.vaddr =
495 			(wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ?
496 			(u64) atomic64_add_return(sdata, maddr) - sdata :
497 			(u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
498 				      sdata, wqe->wr.wr.atomic.swap);
499 		atomic_dec(&qp->r_sge.sge.mr->refcount);
500 		qp->r_sge.num_sge = 0;
501 		goto send_comp;
502 
503 	default:
504 		send_status = IB_WC_LOC_QP_OP_ERR;
505 		goto serr;
506 	}
507 
508 	sge = &sqp->s_sge.sge;
509 	while (sqp->s_len) {
510 		u32 len = sqp->s_len;
511 
512 		if (len > sge->length)
513 			len = sge->length;
514 		if (len > sge->sge_length)
515 			len = sge->sge_length;
516 		BUG_ON(len == 0);
517 		qib_copy_sge(&qp->r_sge, sge->vaddr, len, release);
518 		sge->vaddr += len;
519 		sge->length -= len;
520 		sge->sge_length -= len;
521 		if (sge->sge_length == 0) {
522 			if (!release)
523 				atomic_dec(&sge->mr->refcount);
524 			if (--sqp->s_sge.num_sge)
525 				*sge = *sqp->s_sge.sg_list++;
526 		} else if (sge->length == 0 && sge->mr->lkey) {
527 			if (++sge->n >= QIB_SEGSZ) {
528 				if (++sge->m >= sge->mr->mapsz)
529 					break;
530 				sge->n = 0;
531 			}
532 			sge->vaddr =
533 				sge->mr->map[sge->m]->segs[sge->n].vaddr;
534 			sge->length =
535 				sge->mr->map[sge->m]->segs[sge->n].length;
536 		}
537 		sqp->s_len -= len;
538 	}
539 	if (release)
540 		while (qp->r_sge.num_sge) {
541 			atomic_dec(&qp->r_sge.sge.mr->refcount);
542 			if (--qp->r_sge.num_sge)
543 				qp->r_sge.sge = *qp->r_sge.sg_list++;
544 		}
545 
546 	if (!test_and_clear_bit(QIB_R_WRID_VALID, &qp->r_aflags))
547 		goto send_comp;
548 
549 	if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
550 		wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
551 	else
552 		wc.opcode = IB_WC_RECV;
553 	wc.wr_id = qp->r_wr_id;
554 	wc.status = IB_WC_SUCCESS;
555 	wc.byte_len = wqe->length;
556 	wc.qp = &qp->ibqp;
557 	wc.src_qp = qp->remote_qpn;
558 	wc.slid = qp->remote_ah_attr.dlid;
559 	wc.sl = qp->remote_ah_attr.sl;
560 	wc.port_num = 1;
561 	/* Signal completion event if the solicited bit is set. */
562 	qib_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
563 		       wqe->wr.send_flags & IB_SEND_SOLICITED);
564 
565 send_comp:
566 	spin_lock_irqsave(&sqp->s_lock, flags);
567 	ibp->n_loop_pkts++;
568 flush_send:
569 	sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
570 	qib_send_complete(sqp, wqe, send_status);
571 	goto again;
572 
573 rnr_nak:
574 	/* Handle RNR NAK */
575 	if (qp->ibqp.qp_type == IB_QPT_UC)
576 		goto send_comp;
577 	ibp->n_rnr_naks++;
578 	/*
579 	 * Note: we don't need the s_lock held since the BUSY flag
580 	 * makes this single threaded.
581 	 */
582 	if (sqp->s_rnr_retry == 0) {
583 		send_status = IB_WC_RNR_RETRY_EXC_ERR;
584 		goto serr;
585 	}
586 	if (sqp->s_rnr_retry_cnt < 7)
587 		sqp->s_rnr_retry--;
588 	spin_lock_irqsave(&sqp->s_lock, flags);
589 	if (!(ib_qib_state_ops[sqp->state] & QIB_PROCESS_RECV_OK))
590 		goto clr_busy;
591 	sqp->s_flags |= QIB_S_WAIT_RNR;
592 	sqp->s_timer.function = qib_rc_rnr_retry;
593 	sqp->s_timer.expires = jiffies +
594 		usecs_to_jiffies(ib_qib_rnr_table[qp->r_min_rnr_timer]);
595 	add_timer(&sqp->s_timer);
596 	goto clr_busy;
597 
598 op_err:
599 	send_status = IB_WC_REM_OP_ERR;
600 	wc.status = IB_WC_LOC_QP_OP_ERR;
601 	goto err;
602 
603 inv_err:
604 	send_status = IB_WC_REM_INV_REQ_ERR;
605 	wc.status = IB_WC_LOC_QP_OP_ERR;
606 	goto err;
607 
608 acc_err:
609 	send_status = IB_WC_REM_ACCESS_ERR;
610 	wc.status = IB_WC_LOC_PROT_ERR;
611 err:
612 	/* responder goes to error state */
613 	qib_rc_error(qp, wc.status);
614 
615 serr:
616 	spin_lock_irqsave(&sqp->s_lock, flags);
617 	qib_send_complete(sqp, wqe, send_status);
618 	if (sqp->ibqp.qp_type == IB_QPT_RC) {
619 		int lastwqe = qib_error_qp(sqp, IB_WC_WR_FLUSH_ERR);
620 
621 		sqp->s_flags &= ~QIB_S_BUSY;
622 		spin_unlock_irqrestore(&sqp->s_lock, flags);
623 		if (lastwqe) {
624 			struct ib_event ev;
625 
626 			ev.device = sqp->ibqp.device;
627 			ev.element.qp = &sqp->ibqp;
628 			ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
629 			sqp->ibqp.event_handler(&ev, sqp->ibqp.qp_context);
630 		}
631 		goto done;
632 	}
633 clr_busy:
634 	sqp->s_flags &= ~QIB_S_BUSY;
635 unlock:
636 	spin_unlock_irqrestore(&sqp->s_lock, flags);
637 done:
638 	if (qp && atomic_dec_and_test(&qp->refcount))
639 		wake_up(&qp->wait);
640 }
641 
642 /**
643  * qib_make_grh - construct a GRH header
644  * @ibp: a pointer to the IB port
645  * @hdr: a pointer to the GRH header being constructed
646  * @grh: the global route address to send to
647  * @hwords: the number of 32 bit words of header being sent
648  * @nwords: the number of 32 bit words of data being sent
649  *
650  * Return the size of the header in 32 bit words.
651  */
652 u32 qib_make_grh(struct qib_ibport *ibp, struct ib_grh *hdr,
653 		 struct ib_global_route *grh, u32 hwords, u32 nwords)
654 {
655 	hdr->version_tclass_flow =
656 		cpu_to_be32((IB_GRH_VERSION << IB_GRH_VERSION_SHIFT) |
657 			    (grh->traffic_class << IB_GRH_TCLASS_SHIFT) |
658 			    (grh->flow_label << IB_GRH_FLOW_SHIFT));
659 	hdr->paylen = cpu_to_be16((hwords - 2 + nwords + SIZE_OF_CRC) << 2);
660 	/* next_hdr is defined by C8-7 in ch. 8.4.1 */
661 	hdr->next_hdr = IB_GRH_NEXT_HDR;
662 	hdr->hop_limit = grh->hop_limit;
663 	/* The SGID is 32-bit aligned. */
664 	hdr->sgid.global.subnet_prefix = ibp->gid_prefix;
665 	hdr->sgid.global.interface_id = grh->sgid_index ?
666 		ibp->guids[grh->sgid_index - 1] : ppd_from_ibp(ibp)->guid;
667 	hdr->dgid = grh->dgid;
668 
669 	/* GRH header size in 32-bit words. */
670 	return sizeof(struct ib_grh) / sizeof(u32);
671 }
672 
673 void qib_make_ruc_header(struct qib_qp *qp, struct qib_other_headers *ohdr,
674 			 u32 bth0, u32 bth2)
675 {
676 	struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
677 	u16 lrh0;
678 	u32 nwords;
679 	u32 extra_bytes;
680 
681 	/* Construct the header. */
682 	extra_bytes = -qp->s_cur_size & 3;
683 	nwords = (qp->s_cur_size + extra_bytes) >> 2;
684 	lrh0 = QIB_LRH_BTH;
685 	if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
686 		qp->s_hdrwords += qib_make_grh(ibp, &qp->s_hdr.u.l.grh,
687 					       &qp->remote_ah_attr.grh,
688 					       qp->s_hdrwords, nwords);
689 		lrh0 = QIB_LRH_GRH;
690 	}
691 	lrh0 |= ibp->sl_to_vl[qp->remote_ah_attr.sl] << 12 |
692 		qp->remote_ah_attr.sl << 4;
693 	qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
694 	qp->s_hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
695 	qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC);
696 	qp->s_hdr.lrh[3] = cpu_to_be16(ppd_from_ibp(ibp)->lid |
697 				       qp->remote_ah_attr.src_path_bits);
698 	bth0 |= qib_get_pkey(ibp, qp->s_pkey_index);
699 	bth0 |= extra_bytes << 20;
700 	if (qp->s_mig_state == IB_MIG_MIGRATED)
701 		bth0 |= IB_BTH_MIG_REQ;
702 	ohdr->bth[0] = cpu_to_be32(bth0);
703 	ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
704 	ohdr->bth[2] = cpu_to_be32(bth2);
705 }
706 
707 /**
708  * qib_do_send - perform a send on a QP
709  * @work: contains a pointer to the QP
710  *
711  * Process entries in the send work queue until credit or queue is
712  * exhausted.  Only allow one CPU to send a packet per QP (tasklet).
713  * Otherwise, two threads could send packets out of order.
714  */
715 void qib_do_send(struct work_struct *work)
716 {
717 	struct qib_qp *qp = container_of(work, struct qib_qp, s_work);
718 	struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
719 	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
720 	int (*make_req)(struct qib_qp *qp);
721 	unsigned long flags;
722 
723 	if ((qp->ibqp.qp_type == IB_QPT_RC ||
724 	     qp->ibqp.qp_type == IB_QPT_UC) &&
725 	    (qp->remote_ah_attr.dlid & ~((1 << ppd->lmc) - 1)) == ppd->lid) {
726 		qib_ruc_loopback(qp);
727 		return;
728 	}
729 
730 	if (qp->ibqp.qp_type == IB_QPT_RC)
731 		make_req = qib_make_rc_req;
732 	else if (qp->ibqp.qp_type == IB_QPT_UC)
733 		make_req = qib_make_uc_req;
734 	else
735 		make_req = qib_make_ud_req;
736 
737 	spin_lock_irqsave(&qp->s_lock, flags);
738 
739 	/* Return if we are already busy processing a work request. */
740 	if (!qib_send_ok(qp)) {
741 		spin_unlock_irqrestore(&qp->s_lock, flags);
742 		return;
743 	}
744 
745 	qp->s_flags |= QIB_S_BUSY;
746 
747 	spin_unlock_irqrestore(&qp->s_lock, flags);
748 
749 	do {
750 		/* Check for a constructed packet to be sent. */
751 		if (qp->s_hdrwords != 0) {
752 			/*
753 			 * If the packet cannot be sent now, return and
754 			 * the send tasklet will be woken up later.
755 			 */
756 			if (qib_verbs_send(qp, &qp->s_hdr, qp->s_hdrwords,
757 					   qp->s_cur_sge, qp->s_cur_size))
758 				break;
759 			/* Record that s_hdr is empty. */
760 			qp->s_hdrwords = 0;
761 		}
762 	} while (make_req(qp));
763 }
764 
765 /*
766  * This should be called with s_lock held.
767  */
768 void qib_send_complete(struct qib_qp *qp, struct qib_swqe *wqe,
769 		       enum ib_wc_status status)
770 {
771 	u32 old_last, last;
772 	unsigned i;
773 
774 	if (!(ib_qib_state_ops[qp->state] & QIB_PROCESS_OR_FLUSH_SEND))
775 		return;
776 
777 	for (i = 0; i < wqe->wr.num_sge; i++) {
778 		struct qib_sge *sge = &wqe->sg_list[i];
779 
780 		atomic_dec(&sge->mr->refcount);
781 	}
782 	if (qp->ibqp.qp_type == IB_QPT_UD ||
783 	    qp->ibqp.qp_type == IB_QPT_SMI ||
784 	    qp->ibqp.qp_type == IB_QPT_GSI)
785 		atomic_dec(&to_iah(wqe->wr.wr.ud.ah)->refcount);
786 
787 	/* See ch. 11.2.4.1 and 10.7.3.1 */
788 	if (!(qp->s_flags & QIB_S_SIGNAL_REQ_WR) ||
789 	    (wqe->wr.send_flags & IB_SEND_SIGNALED) ||
790 	    status != IB_WC_SUCCESS) {
791 		struct ib_wc wc;
792 
793 		memset(&wc, 0, sizeof wc);
794 		wc.wr_id = wqe->wr.wr_id;
795 		wc.status = status;
796 		wc.opcode = ib_qib_wc_opcode[wqe->wr.opcode];
797 		wc.qp = &qp->ibqp;
798 		if (status == IB_WC_SUCCESS)
799 			wc.byte_len = wqe->length;
800 		qib_cq_enter(to_icq(qp->ibqp.send_cq), &wc,
801 			     status != IB_WC_SUCCESS);
802 	}
803 
804 	last = qp->s_last;
805 	old_last = last;
806 	if (++last >= qp->s_size)
807 		last = 0;
808 	qp->s_last = last;
809 	if (qp->s_acked == old_last)
810 		qp->s_acked = last;
811 	if (qp->s_cur == old_last)
812 		qp->s_cur = last;
813 	if (qp->s_tail == old_last)
814 		qp->s_tail = last;
815 	if (qp->state == IB_QPS_SQD && last == qp->s_cur)
816 		qp->s_draining = 0;
817 }
818