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 		qib_put_mr(sge->mr);
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 r_lock held.
264  *
265  * The s_lock will be acquired around the qib_migrate_qp() call.
266  */
267 int qib_ruc_check_hdr(struct qib_ibport *ibp, struct qib_ib_header *hdr,
268 		      int has_grh, struct qib_qp *qp, u32 bth0)
269 {
270 	__be64 guid;
271 	unsigned long flags;
272 
273 	if (qp->s_mig_state == IB_MIG_ARMED && (bth0 & IB_BTH_MIG_REQ)) {
274 		if (!has_grh) {
275 			if (qp->alt_ah_attr.ah_flags & IB_AH_GRH)
276 				goto err;
277 		} else {
278 			if (!(qp->alt_ah_attr.ah_flags & IB_AH_GRH))
279 				goto err;
280 			guid = get_sguid(ibp, qp->alt_ah_attr.grh.sgid_index);
281 			if (!gid_ok(&hdr->u.l.grh.dgid, ibp->gid_prefix, guid))
282 				goto err;
283 			if (!gid_ok(&hdr->u.l.grh.sgid,
284 			    qp->alt_ah_attr.grh.dgid.global.subnet_prefix,
285 			    qp->alt_ah_attr.grh.dgid.global.interface_id))
286 				goto err;
287 		}
288 		if (!qib_pkey_ok((u16)bth0,
289 				 qib_get_pkey(ibp, qp->s_alt_pkey_index))) {
290 			qib_bad_pqkey(ibp, IB_NOTICE_TRAP_BAD_PKEY,
291 				      (u16)bth0,
292 				      (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF,
293 				      0, qp->ibqp.qp_num,
294 				      hdr->lrh[3], hdr->lrh[1]);
295 			goto err;
296 		}
297 		/* Validate the SLID. See Ch. 9.6.1.5 and 17.2.8 */
298 		if (be16_to_cpu(hdr->lrh[3]) != qp->alt_ah_attr.dlid ||
299 		    ppd_from_ibp(ibp)->port != qp->alt_ah_attr.port_num)
300 			goto err;
301 		spin_lock_irqsave(&qp->s_lock, flags);
302 		qib_migrate_qp(qp);
303 		spin_unlock_irqrestore(&qp->s_lock, flags);
304 	} else {
305 		if (!has_grh) {
306 			if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
307 				goto err;
308 		} else {
309 			if (!(qp->remote_ah_attr.ah_flags & IB_AH_GRH))
310 				goto err;
311 			guid = get_sguid(ibp,
312 					 qp->remote_ah_attr.grh.sgid_index);
313 			if (!gid_ok(&hdr->u.l.grh.dgid, ibp->gid_prefix, guid))
314 				goto err;
315 			if (!gid_ok(&hdr->u.l.grh.sgid,
316 			    qp->remote_ah_attr.grh.dgid.global.subnet_prefix,
317 			    qp->remote_ah_attr.grh.dgid.global.interface_id))
318 				goto err;
319 		}
320 		if (!qib_pkey_ok((u16)bth0,
321 				 qib_get_pkey(ibp, qp->s_pkey_index))) {
322 			qib_bad_pqkey(ibp, IB_NOTICE_TRAP_BAD_PKEY,
323 				      (u16)bth0,
324 				      (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF,
325 				      0, qp->ibqp.qp_num,
326 				      hdr->lrh[3], hdr->lrh[1]);
327 			goto err;
328 		}
329 		/* Validate the SLID. See Ch. 9.6.1.5 */
330 		if (be16_to_cpu(hdr->lrh[3]) != qp->remote_ah_attr.dlid ||
331 		    ppd_from_ibp(ibp)->port != qp->port_num)
332 			goto err;
333 		if (qp->s_mig_state == IB_MIG_REARM &&
334 		    !(bth0 & IB_BTH_MIG_REQ))
335 			qp->s_mig_state = IB_MIG_ARMED;
336 	}
337 
338 	return 0;
339 
340 err:
341 	return 1;
342 }
343 
344 /**
345  * qib_ruc_loopback - handle UC and RC lookback requests
346  * @sqp: the sending QP
347  *
348  * This is called from qib_do_send() to
349  * forward a WQE addressed to the same HCA.
350  * Note that although we are single threaded due to the tasklet, we still
351  * have to protect against post_send().  We don't have to worry about
352  * receive interrupts since this is a connected protocol and all packets
353  * will pass through here.
354  */
355 static void qib_ruc_loopback(struct qib_qp *sqp)
356 {
357 	struct qib_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num);
358 	struct qib_qp *qp;
359 	struct qib_swqe *wqe;
360 	struct qib_sge *sge;
361 	unsigned long flags;
362 	struct ib_wc wc;
363 	u64 sdata;
364 	atomic64_t *maddr;
365 	enum ib_wc_status send_status;
366 	int release;
367 	int ret;
368 
369 	/*
370 	 * Note that we check the responder QP state after
371 	 * checking the requester's state.
372 	 */
373 	qp = qib_lookup_qpn(ibp, sqp->remote_qpn);
374 
375 	spin_lock_irqsave(&sqp->s_lock, flags);
376 
377 	/* Return if we are already busy processing a work request. */
378 	if ((sqp->s_flags & (QIB_S_BUSY | QIB_S_ANY_WAIT)) ||
379 	    !(ib_qib_state_ops[sqp->state] & QIB_PROCESS_OR_FLUSH_SEND))
380 		goto unlock;
381 
382 	sqp->s_flags |= QIB_S_BUSY;
383 
384 again:
385 	if (sqp->s_last == sqp->s_head)
386 		goto clr_busy;
387 	wqe = get_swqe_ptr(sqp, sqp->s_last);
388 
389 	/* Return if it is not OK to start a new work reqeust. */
390 	if (!(ib_qib_state_ops[sqp->state] & QIB_PROCESS_NEXT_SEND_OK)) {
391 		if (!(ib_qib_state_ops[sqp->state] & QIB_FLUSH_SEND))
392 			goto clr_busy;
393 		/* We are in the error state, flush the work request. */
394 		send_status = IB_WC_WR_FLUSH_ERR;
395 		goto flush_send;
396 	}
397 
398 	/*
399 	 * We can rely on the entry not changing without the s_lock
400 	 * being held until we update s_last.
401 	 * We increment s_cur to indicate s_last is in progress.
402 	 */
403 	if (sqp->s_last == sqp->s_cur) {
404 		if (++sqp->s_cur >= sqp->s_size)
405 			sqp->s_cur = 0;
406 	}
407 	spin_unlock_irqrestore(&sqp->s_lock, flags);
408 
409 	if (!qp || !(ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK) ||
410 	    qp->ibqp.qp_type != sqp->ibqp.qp_type) {
411 		ibp->n_pkt_drops++;
412 		/*
413 		 * For RC, the requester would timeout and retry so
414 		 * shortcut the timeouts and just signal too many retries.
415 		 */
416 		if (sqp->ibqp.qp_type == IB_QPT_RC)
417 			send_status = IB_WC_RETRY_EXC_ERR;
418 		else
419 			send_status = IB_WC_SUCCESS;
420 		goto serr;
421 	}
422 
423 	memset(&wc, 0, sizeof wc);
424 	send_status = IB_WC_SUCCESS;
425 
426 	release = 1;
427 	sqp->s_sge.sge = wqe->sg_list[0];
428 	sqp->s_sge.sg_list = wqe->sg_list + 1;
429 	sqp->s_sge.num_sge = wqe->wr.num_sge;
430 	sqp->s_len = wqe->length;
431 	switch (wqe->wr.opcode) {
432 	case IB_WR_SEND_WITH_IMM:
433 		wc.wc_flags = IB_WC_WITH_IMM;
434 		wc.ex.imm_data = wqe->wr.ex.imm_data;
435 		/* FALLTHROUGH */
436 	case IB_WR_SEND:
437 		ret = qib_get_rwqe(qp, 0);
438 		if (ret < 0)
439 			goto op_err;
440 		if (!ret)
441 			goto rnr_nak;
442 		break;
443 
444 	case IB_WR_RDMA_WRITE_WITH_IMM:
445 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
446 			goto inv_err;
447 		wc.wc_flags = IB_WC_WITH_IMM;
448 		wc.ex.imm_data = wqe->wr.ex.imm_data;
449 		ret = qib_get_rwqe(qp, 1);
450 		if (ret < 0)
451 			goto op_err;
452 		if (!ret)
453 			goto rnr_nak;
454 		/* FALLTHROUGH */
455 	case IB_WR_RDMA_WRITE:
456 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
457 			goto inv_err;
458 		if (wqe->length == 0)
459 			break;
460 		if (unlikely(!qib_rkey_ok(qp, &qp->r_sge.sge, wqe->length,
461 					  wqe->wr.wr.rdma.remote_addr,
462 					  wqe->wr.wr.rdma.rkey,
463 					  IB_ACCESS_REMOTE_WRITE)))
464 			goto acc_err;
465 		qp->r_sge.sg_list = NULL;
466 		qp->r_sge.num_sge = 1;
467 		qp->r_sge.total_len = wqe->length;
468 		break;
469 
470 	case IB_WR_RDMA_READ:
471 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
472 			goto inv_err;
473 		if (unlikely(!qib_rkey_ok(qp, &sqp->s_sge.sge, wqe->length,
474 					  wqe->wr.wr.rdma.remote_addr,
475 					  wqe->wr.wr.rdma.rkey,
476 					  IB_ACCESS_REMOTE_READ)))
477 			goto acc_err;
478 		release = 0;
479 		sqp->s_sge.sg_list = NULL;
480 		sqp->s_sge.num_sge = 1;
481 		qp->r_sge.sge = wqe->sg_list[0];
482 		qp->r_sge.sg_list = wqe->sg_list + 1;
483 		qp->r_sge.num_sge = wqe->wr.num_sge;
484 		qp->r_sge.total_len = wqe->length;
485 		break;
486 
487 	case IB_WR_ATOMIC_CMP_AND_SWP:
488 	case IB_WR_ATOMIC_FETCH_AND_ADD:
489 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
490 			goto inv_err;
491 		if (unlikely(!qib_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64),
492 					  wqe->wr.wr.atomic.remote_addr,
493 					  wqe->wr.wr.atomic.rkey,
494 					  IB_ACCESS_REMOTE_ATOMIC)))
495 			goto acc_err;
496 		/* Perform atomic OP and save result. */
497 		maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
498 		sdata = wqe->wr.wr.atomic.compare_add;
499 		*(u64 *) sqp->s_sge.sge.vaddr =
500 			(wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ?
501 			(u64) atomic64_add_return(sdata, maddr) - sdata :
502 			(u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
503 				      sdata, wqe->wr.wr.atomic.swap);
504 		qib_put_mr(qp->r_sge.sge.mr);
505 		qp->r_sge.num_sge = 0;
506 		goto send_comp;
507 
508 	default:
509 		send_status = IB_WC_LOC_QP_OP_ERR;
510 		goto serr;
511 	}
512 
513 	sge = &sqp->s_sge.sge;
514 	while (sqp->s_len) {
515 		u32 len = sqp->s_len;
516 
517 		if (len > sge->length)
518 			len = sge->length;
519 		if (len > sge->sge_length)
520 			len = sge->sge_length;
521 		BUG_ON(len == 0);
522 		qib_copy_sge(&qp->r_sge, sge->vaddr, len, release);
523 		sge->vaddr += len;
524 		sge->length -= len;
525 		sge->sge_length -= len;
526 		if (sge->sge_length == 0) {
527 			if (!release)
528 				qib_put_mr(sge->mr);
529 			if (--sqp->s_sge.num_sge)
530 				*sge = *sqp->s_sge.sg_list++;
531 		} else if (sge->length == 0 && sge->mr->lkey) {
532 			if (++sge->n >= QIB_SEGSZ) {
533 				if (++sge->m >= sge->mr->mapsz)
534 					break;
535 				sge->n = 0;
536 			}
537 			sge->vaddr =
538 				sge->mr->map[sge->m]->segs[sge->n].vaddr;
539 			sge->length =
540 				sge->mr->map[sge->m]->segs[sge->n].length;
541 		}
542 		sqp->s_len -= len;
543 	}
544 	if (release)
545 		qib_put_ss(&qp->r_sge);
546 
547 	if (!test_and_clear_bit(QIB_R_WRID_VALID, &qp->r_aflags))
548 		goto send_comp;
549 
550 	if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
551 		wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
552 	else
553 		wc.opcode = IB_WC_RECV;
554 	wc.wr_id = qp->r_wr_id;
555 	wc.status = IB_WC_SUCCESS;
556 	wc.byte_len = wqe->length;
557 	wc.qp = &qp->ibqp;
558 	wc.src_qp = qp->remote_qpn;
559 	wc.slid = qp->remote_ah_attr.dlid;
560 	wc.sl = qp->remote_ah_attr.sl;
561 	wc.port_num = 1;
562 	/* Signal completion event if the solicited bit is set. */
563 	qib_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
564 		       wqe->wr.send_flags & IB_SEND_SOLICITED);
565 
566 send_comp:
567 	spin_lock_irqsave(&sqp->s_lock, flags);
568 	ibp->n_loop_pkts++;
569 flush_send:
570 	sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
571 	qib_send_complete(sqp, wqe, send_status);
572 	goto again;
573 
574 rnr_nak:
575 	/* Handle RNR NAK */
576 	if (qp->ibqp.qp_type == IB_QPT_UC)
577 		goto send_comp;
578 	ibp->n_rnr_naks++;
579 	/*
580 	 * Note: we don't need the s_lock held since the BUSY flag
581 	 * makes this single threaded.
582 	 */
583 	if (sqp->s_rnr_retry == 0) {
584 		send_status = IB_WC_RNR_RETRY_EXC_ERR;
585 		goto serr;
586 	}
587 	if (sqp->s_rnr_retry_cnt < 7)
588 		sqp->s_rnr_retry--;
589 	spin_lock_irqsave(&sqp->s_lock, flags);
590 	if (!(ib_qib_state_ops[sqp->state] & QIB_PROCESS_RECV_OK))
591 		goto clr_busy;
592 	sqp->s_flags |= QIB_S_WAIT_RNR;
593 	sqp->s_timer.function = qib_rc_rnr_retry;
594 	sqp->s_timer.expires = jiffies +
595 		usecs_to_jiffies(ib_qib_rnr_table[qp->r_min_rnr_timer]);
596 	add_timer(&sqp->s_timer);
597 	goto clr_busy;
598 
599 op_err:
600 	send_status = IB_WC_REM_OP_ERR;
601 	wc.status = IB_WC_LOC_QP_OP_ERR;
602 	goto err;
603 
604 inv_err:
605 	send_status = IB_WC_REM_INV_REQ_ERR;
606 	wc.status = IB_WC_LOC_QP_OP_ERR;
607 	goto err;
608 
609 acc_err:
610 	send_status = IB_WC_REM_ACCESS_ERR;
611 	wc.status = IB_WC_LOC_PROT_ERR;
612 err:
613 	/* responder goes to error state */
614 	qib_rc_error(qp, wc.status);
615 
616 serr:
617 	spin_lock_irqsave(&sqp->s_lock, flags);
618 	qib_send_complete(sqp, wqe, send_status);
619 	if (sqp->ibqp.qp_type == IB_QPT_RC) {
620 		int lastwqe = qib_error_qp(sqp, IB_WC_WR_FLUSH_ERR);
621 
622 		sqp->s_flags &= ~QIB_S_BUSY;
623 		spin_unlock_irqrestore(&sqp->s_lock, flags);
624 		if (lastwqe) {
625 			struct ib_event ev;
626 
627 			ev.device = sqp->ibqp.device;
628 			ev.element.qp = &sqp->ibqp;
629 			ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
630 			sqp->ibqp.event_handler(&ev, sqp->ibqp.qp_context);
631 		}
632 		goto done;
633 	}
634 clr_busy:
635 	sqp->s_flags &= ~QIB_S_BUSY;
636 unlock:
637 	spin_unlock_irqrestore(&sqp->s_lock, flags);
638 done:
639 	if (qp && atomic_dec_and_test(&qp->refcount))
640 		wake_up(&qp->wait);
641 }
642 
643 /**
644  * qib_make_grh - construct a GRH header
645  * @ibp: a pointer to the IB port
646  * @hdr: a pointer to the GRH header being constructed
647  * @grh: the global route address to send to
648  * @hwords: the number of 32 bit words of header being sent
649  * @nwords: the number of 32 bit words of data being sent
650  *
651  * Return the size of the header in 32 bit words.
652  */
653 u32 qib_make_grh(struct qib_ibport *ibp, struct ib_grh *hdr,
654 		 struct ib_global_route *grh, u32 hwords, u32 nwords)
655 {
656 	hdr->version_tclass_flow =
657 		cpu_to_be32((IB_GRH_VERSION << IB_GRH_VERSION_SHIFT) |
658 			    (grh->traffic_class << IB_GRH_TCLASS_SHIFT) |
659 			    (grh->flow_label << IB_GRH_FLOW_SHIFT));
660 	hdr->paylen = cpu_to_be16((hwords - 2 + nwords + SIZE_OF_CRC) << 2);
661 	/* next_hdr is defined by C8-7 in ch. 8.4.1 */
662 	hdr->next_hdr = IB_GRH_NEXT_HDR;
663 	hdr->hop_limit = grh->hop_limit;
664 	/* The SGID is 32-bit aligned. */
665 	hdr->sgid.global.subnet_prefix = ibp->gid_prefix;
666 	hdr->sgid.global.interface_id = grh->sgid_index ?
667 		ibp->guids[grh->sgid_index - 1] : ppd_from_ibp(ibp)->guid;
668 	hdr->dgid = grh->dgid;
669 
670 	/* GRH header size in 32-bit words. */
671 	return sizeof(struct ib_grh) / sizeof(u32);
672 }
673 
674 void qib_make_ruc_header(struct qib_qp *qp, struct qib_other_headers *ohdr,
675 			 u32 bth0, u32 bth2)
676 {
677 	struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
678 	u16 lrh0;
679 	u32 nwords;
680 	u32 extra_bytes;
681 
682 	/* Construct the header. */
683 	extra_bytes = -qp->s_cur_size & 3;
684 	nwords = (qp->s_cur_size + extra_bytes) >> 2;
685 	lrh0 = QIB_LRH_BTH;
686 	if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
687 		qp->s_hdrwords += qib_make_grh(ibp, &qp->s_hdr->u.l.grh,
688 					       &qp->remote_ah_attr.grh,
689 					       qp->s_hdrwords, nwords);
690 		lrh0 = QIB_LRH_GRH;
691 	}
692 	lrh0 |= ibp->sl_to_vl[qp->remote_ah_attr.sl] << 12 |
693 		qp->remote_ah_attr.sl << 4;
694 	qp->s_hdr->lrh[0] = cpu_to_be16(lrh0);
695 	qp->s_hdr->lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
696 	qp->s_hdr->lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC);
697 	qp->s_hdr->lrh[3] = cpu_to_be16(ppd_from_ibp(ibp)->lid |
698 				       qp->remote_ah_attr.src_path_bits);
699 	bth0 |= qib_get_pkey(ibp, qp->s_pkey_index);
700 	bth0 |= extra_bytes << 20;
701 	if (qp->s_mig_state == IB_MIG_MIGRATED)
702 		bth0 |= IB_BTH_MIG_REQ;
703 	ohdr->bth[0] = cpu_to_be32(bth0);
704 	ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
705 	ohdr->bth[2] = cpu_to_be32(bth2);
706 }
707 
708 /**
709  * qib_do_send - perform a send on a QP
710  * @work: contains a pointer to the QP
711  *
712  * Process entries in the send work queue until credit or queue is
713  * exhausted.  Only allow one CPU to send a packet per QP (tasklet).
714  * Otherwise, two threads could send packets out of order.
715  */
716 void qib_do_send(struct work_struct *work)
717 {
718 	struct qib_qp *qp = container_of(work, struct qib_qp, s_work);
719 	struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
720 	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
721 	int (*make_req)(struct qib_qp *qp);
722 	unsigned long flags;
723 
724 	if ((qp->ibqp.qp_type == IB_QPT_RC ||
725 	     qp->ibqp.qp_type == IB_QPT_UC) &&
726 	    (qp->remote_ah_attr.dlid & ~((1 << ppd->lmc) - 1)) == ppd->lid) {
727 		qib_ruc_loopback(qp);
728 		return;
729 	}
730 
731 	if (qp->ibqp.qp_type == IB_QPT_RC)
732 		make_req = qib_make_rc_req;
733 	else if (qp->ibqp.qp_type == IB_QPT_UC)
734 		make_req = qib_make_uc_req;
735 	else
736 		make_req = qib_make_ud_req;
737 
738 	spin_lock_irqsave(&qp->s_lock, flags);
739 
740 	/* Return if we are already busy processing a work request. */
741 	if (!qib_send_ok(qp)) {
742 		spin_unlock_irqrestore(&qp->s_lock, flags);
743 		return;
744 	}
745 
746 	qp->s_flags |= QIB_S_BUSY;
747 
748 	spin_unlock_irqrestore(&qp->s_lock, flags);
749 
750 	do {
751 		/* Check for a constructed packet to be sent. */
752 		if (qp->s_hdrwords != 0) {
753 			/*
754 			 * If the packet cannot be sent now, return and
755 			 * the send tasklet will be woken up later.
756 			 */
757 			if (qib_verbs_send(qp, qp->s_hdr, qp->s_hdrwords,
758 					   qp->s_cur_sge, qp->s_cur_size))
759 				break;
760 			/* Record that s_hdr is empty. */
761 			qp->s_hdrwords = 0;
762 		}
763 	} while (make_req(qp));
764 }
765 
766 /*
767  * This should be called with s_lock held.
768  */
769 void qib_send_complete(struct qib_qp *qp, struct qib_swqe *wqe,
770 		       enum ib_wc_status status)
771 {
772 	u32 old_last, last;
773 	unsigned i;
774 
775 	if (!(ib_qib_state_ops[qp->state] & QIB_PROCESS_OR_FLUSH_SEND))
776 		return;
777 
778 	for (i = 0; i < wqe->wr.num_sge; i++) {
779 		struct qib_sge *sge = &wqe->sg_list[i];
780 
781 		qib_put_mr(sge->mr);
782 	}
783 	if (qp->ibqp.qp_type == IB_QPT_UD ||
784 	    qp->ibqp.qp_type == IB_QPT_SMI ||
785 	    qp->ibqp.qp_type == IB_QPT_GSI)
786 		atomic_dec(&to_iah(wqe->wr.wr.ud.ah)->refcount);
787 
788 	/* See ch. 11.2.4.1 and 10.7.3.1 */
789 	if (!(qp->s_flags & QIB_S_SIGNAL_REQ_WR) ||
790 	    (wqe->wr.send_flags & IB_SEND_SIGNALED) ||
791 	    status != IB_WC_SUCCESS) {
792 		struct ib_wc wc;
793 
794 		memset(&wc, 0, sizeof wc);
795 		wc.wr_id = wqe->wr.wr_id;
796 		wc.status = status;
797 		wc.opcode = ib_qib_wc_opcode[wqe->wr.opcode];
798 		wc.qp = &qp->ibqp;
799 		if (status == IB_WC_SUCCESS)
800 			wc.byte_len = wqe->length;
801 		qib_cq_enter(to_icq(qp->ibqp.send_cq), &wc,
802 			     status != IB_WC_SUCCESS);
803 	}
804 
805 	last = qp->s_last;
806 	old_last = last;
807 	if (++last >= qp->s_size)
808 		last = 0;
809 	qp->s_last = last;
810 	if (qp->s_acked == old_last)
811 		qp->s_acked = last;
812 	if (qp->s_cur == old_last)
813 		qp->s_cur = last;
814 	if (qp->s_tail == old_last)
815 		qp->s_tail = last;
816 	if (qp->state == IB_QPS_SQD && last == qp->s_cur)
817 		qp->s_draining = 0;
818 }
819