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 #include <rdma/ib_smi.h>
36 
37 #include "qib.h"
38 #include "qib_mad.h"
39 
40 /*
41  * Convert the AETH RNR timeout code into the number of microseconds.
42  */
43 const u32 ib_qib_rnr_table[32] = {
44 	655360,	/* 00: 655.36 */
45 	10,	/* 01:    .01 */
46 	20,	/* 02     .02 */
47 	30,	/* 03:    .03 */
48 	40,	/* 04:    .04 */
49 	60,	/* 05:    .06 */
50 	80,	/* 06:    .08 */
51 	120,	/* 07:    .12 */
52 	160,	/* 08:    .16 */
53 	240,	/* 09:    .24 */
54 	320,	/* 0A:    .32 */
55 	480,	/* 0B:    .48 */
56 	640,	/* 0C:    .64 */
57 	960,	/* 0D:    .96 */
58 	1280,	/* 0E:   1.28 */
59 	1920,	/* 0F:   1.92 */
60 	2560,	/* 10:   2.56 */
61 	3840,	/* 11:   3.84 */
62 	5120,	/* 12:   5.12 */
63 	7680,	/* 13:   7.68 */
64 	10240,	/* 14:  10.24 */
65 	15360,	/* 15:  15.36 */
66 	20480,	/* 16:  20.48 */
67 	30720,	/* 17:  30.72 */
68 	40960,	/* 18:  40.96 */
69 	61440,	/* 19:  61.44 */
70 	81920,	/* 1A:  81.92 */
71 	122880,	/* 1B: 122.88 */
72 	163840,	/* 1C: 163.84 */
73 	245760,	/* 1D: 245.76 */
74 	327680,	/* 1E: 327.68 */
75 	491520	/* 1F: 491.52 */
76 };
77 
78 /*
79  * Validate a RWQE and fill in the SGE state.
80  * Return 1 if OK.
81  */
82 static int qib_init_sge(struct rvt_qp *qp, struct rvt_rwqe *wqe)
83 {
84 	int i, j, ret;
85 	struct ib_wc wc;
86 	struct rvt_lkey_table *rkt;
87 	struct rvt_pd *pd;
88 	struct rvt_sge_state *ss;
89 
90 	rkt = &to_idev(qp->ibqp.device)->rdi.lkey_table;
91 	pd = ibpd_to_rvtpd(qp->ibqp.srq ? qp->ibqp.srq->pd : qp->ibqp.pd);
92 	ss = &qp->r_sge;
93 	ss->sg_list = qp->r_sg_list;
94 	qp->r_len = 0;
95 	for (i = j = 0; i < wqe->num_sge; i++) {
96 		if (wqe->sg_list[i].length == 0)
97 			continue;
98 		/* Check LKEY */
99 		if (!rvt_lkey_ok(rkt, pd, j ? &ss->sg_list[j - 1] : &ss->sge,
100 				 &wqe->sg_list[i], IB_ACCESS_LOCAL_WRITE))
101 			goto bad_lkey;
102 		qp->r_len += wqe->sg_list[i].length;
103 		j++;
104 	}
105 	ss->num_sge = j;
106 	ss->total_len = qp->r_len;
107 	ret = 1;
108 	goto bail;
109 
110 bad_lkey:
111 	while (j) {
112 		struct rvt_sge *sge = --j ? &ss->sg_list[j - 1] : &ss->sge;
113 
114 		rvt_put_mr(sge->mr);
115 	}
116 	ss->num_sge = 0;
117 	memset(&wc, 0, sizeof(wc));
118 	wc.wr_id = wqe->wr_id;
119 	wc.status = IB_WC_LOC_PROT_ERR;
120 	wc.opcode = IB_WC_RECV;
121 	wc.qp = &qp->ibqp;
122 	/* Signal solicited completion event. */
123 	rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
124 	ret = 0;
125 bail:
126 	return ret;
127 }
128 
129 /**
130  * qib_get_rwqe - copy the next RWQE into the QP's RWQE
131  * @qp: the QP
132  * @wr_id_only: update qp->r_wr_id only, not qp->r_sge
133  *
134  * Return -1 if there is a local error, 0 if no RWQE is available,
135  * otherwise return 1.
136  *
137  * Can be called from interrupt level.
138  */
139 int qib_get_rwqe(struct rvt_qp *qp, int wr_id_only)
140 {
141 	unsigned long flags;
142 	struct rvt_rq *rq;
143 	struct rvt_rwq *wq;
144 	struct rvt_srq *srq;
145 	struct rvt_rwqe *wqe;
146 	void (*handler)(struct ib_event *, void *);
147 	u32 tail;
148 	int ret;
149 
150 	if (qp->ibqp.srq) {
151 		srq = ibsrq_to_rvtsrq(qp->ibqp.srq);
152 		handler = srq->ibsrq.event_handler;
153 		rq = &srq->rq;
154 	} else {
155 		srq = NULL;
156 		handler = NULL;
157 		rq = &qp->r_rq;
158 	}
159 
160 	spin_lock_irqsave(&rq->lock, flags);
161 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) {
162 		ret = 0;
163 		goto unlock;
164 	}
165 
166 	wq = rq->wq;
167 	tail = wq->tail;
168 	/* Validate tail before using it since it is user writable. */
169 	if (tail >= rq->size)
170 		tail = 0;
171 	if (unlikely(tail == wq->head)) {
172 		ret = 0;
173 		goto unlock;
174 	}
175 	/* Make sure entry is read after head index is read. */
176 	smp_rmb();
177 	wqe = rvt_get_rwqe_ptr(rq, tail);
178 	/*
179 	 * Even though we update the tail index in memory, the verbs
180 	 * consumer is not supposed to post more entries until a
181 	 * completion is generated.
182 	 */
183 	if (++tail >= rq->size)
184 		tail = 0;
185 	wq->tail = tail;
186 	if (!wr_id_only && !qib_init_sge(qp, wqe)) {
187 		ret = -1;
188 		goto unlock;
189 	}
190 	qp->r_wr_id = wqe->wr_id;
191 
192 	ret = 1;
193 	set_bit(RVT_R_WRID_VALID, &qp->r_aflags);
194 	if (handler) {
195 		u32 n;
196 
197 		/*
198 		 * Validate head pointer value and compute
199 		 * the number of remaining WQEs.
200 		 */
201 		n = wq->head;
202 		if (n >= rq->size)
203 			n = 0;
204 		if (n < tail)
205 			n += rq->size - tail;
206 		else
207 			n -= tail;
208 		if (n < srq->limit) {
209 			struct ib_event ev;
210 
211 			srq->limit = 0;
212 			spin_unlock_irqrestore(&rq->lock, flags);
213 			ev.device = qp->ibqp.device;
214 			ev.element.srq = qp->ibqp.srq;
215 			ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
216 			handler(&ev, srq->ibsrq.srq_context);
217 			goto bail;
218 		}
219 	}
220 unlock:
221 	spin_unlock_irqrestore(&rq->lock, flags);
222 bail:
223 	return ret;
224 }
225 
226 /*
227  * Switch to alternate path.
228  * The QP s_lock should be held and interrupts disabled.
229  */
230 void qib_migrate_qp(struct rvt_qp *qp)
231 {
232 	struct ib_event ev;
233 
234 	qp->s_mig_state = IB_MIG_MIGRATED;
235 	qp->remote_ah_attr = qp->alt_ah_attr;
236 	qp->port_num = qp->alt_ah_attr.port_num;
237 	qp->s_pkey_index = qp->s_alt_pkey_index;
238 
239 	ev.device = qp->ibqp.device;
240 	ev.element.qp = &qp->ibqp;
241 	ev.event = IB_EVENT_PATH_MIG;
242 	qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
243 }
244 
245 static __be64 get_sguid(struct qib_ibport *ibp, unsigned index)
246 {
247 	if (!index) {
248 		struct qib_pportdata *ppd = ppd_from_ibp(ibp);
249 
250 		return ppd->guid;
251 	}
252 	return ibp->guids[index - 1];
253 }
254 
255 static int gid_ok(union ib_gid *gid, __be64 gid_prefix, __be64 id)
256 {
257 	return (gid->global.interface_id == id &&
258 		(gid->global.subnet_prefix == gid_prefix ||
259 		 gid->global.subnet_prefix == IB_DEFAULT_GID_PREFIX));
260 }
261 
262 /*
263  *
264  * This should be called with the QP r_lock held.
265  *
266  * The s_lock will be acquired around the qib_migrate_qp() call.
267  */
268 int qib_ruc_check_hdr(struct qib_ibport *ibp, struct qib_ib_header *hdr,
269 		      int has_grh, struct rvt_qp *qp, u32 bth0)
270 {
271 	__be64 guid;
272 	unsigned long flags;
273 
274 	if (qp->s_mig_state == IB_MIG_ARMED && (bth0 & IB_BTH_MIG_REQ)) {
275 		if (!has_grh) {
276 			if (qp->alt_ah_attr.ah_flags & IB_AH_GRH)
277 				goto err;
278 		} else {
279 			if (!(qp->alt_ah_attr.ah_flags & IB_AH_GRH))
280 				goto err;
281 			guid = get_sguid(ibp, qp->alt_ah_attr.grh.sgid_index);
282 			if (!gid_ok(&hdr->u.l.grh.dgid,
283 				    ibp->rvp.gid_prefix, guid))
284 				goto err;
285 			if (!gid_ok(&hdr->u.l.grh.sgid,
286 			    qp->alt_ah_attr.grh.dgid.global.subnet_prefix,
287 			    qp->alt_ah_attr.grh.dgid.global.interface_id))
288 				goto err;
289 		}
290 		if (!qib_pkey_ok((u16)bth0,
291 				 qib_get_pkey(ibp, qp->s_alt_pkey_index))) {
292 			qib_bad_pqkey(ibp, IB_NOTICE_TRAP_BAD_PKEY,
293 				      (u16)bth0,
294 				      (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF,
295 				      0, qp->ibqp.qp_num,
296 				      hdr->lrh[3], hdr->lrh[1]);
297 			goto err;
298 		}
299 		/* Validate the SLID. See Ch. 9.6.1.5 and 17.2.8 */
300 		if (be16_to_cpu(hdr->lrh[3]) != qp->alt_ah_attr.dlid ||
301 		    ppd_from_ibp(ibp)->port != qp->alt_ah_attr.port_num)
302 			goto err;
303 		spin_lock_irqsave(&qp->s_lock, flags);
304 		qib_migrate_qp(qp);
305 		spin_unlock_irqrestore(&qp->s_lock, flags);
306 	} else {
307 		if (!has_grh) {
308 			if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
309 				goto err;
310 		} else {
311 			if (!(qp->remote_ah_attr.ah_flags & IB_AH_GRH))
312 				goto err;
313 			guid = get_sguid(ibp,
314 					 qp->remote_ah_attr.grh.sgid_index);
315 			if (!gid_ok(&hdr->u.l.grh.dgid,
316 				    ibp->rvp.gid_prefix, guid))
317 				goto err;
318 			if (!gid_ok(&hdr->u.l.grh.sgid,
319 			    qp->remote_ah_attr.grh.dgid.global.subnet_prefix,
320 			    qp->remote_ah_attr.grh.dgid.global.interface_id))
321 				goto err;
322 		}
323 		if (!qib_pkey_ok((u16)bth0,
324 				 qib_get_pkey(ibp, qp->s_pkey_index))) {
325 			qib_bad_pqkey(ibp, IB_NOTICE_TRAP_BAD_PKEY,
326 				      (u16)bth0,
327 				      (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF,
328 				      0, qp->ibqp.qp_num,
329 				      hdr->lrh[3], hdr->lrh[1]);
330 			goto err;
331 		}
332 		/* Validate the SLID. See Ch. 9.6.1.5 */
333 		if (be16_to_cpu(hdr->lrh[3]) != qp->remote_ah_attr.dlid ||
334 		    ppd_from_ibp(ibp)->port != qp->port_num)
335 			goto err;
336 		if (qp->s_mig_state == IB_MIG_REARM &&
337 		    !(bth0 & IB_BTH_MIG_REQ))
338 			qp->s_mig_state = IB_MIG_ARMED;
339 	}
340 
341 	return 0;
342 
343 err:
344 	return 1;
345 }
346 
347 /**
348  * qib_ruc_loopback - handle UC and RC lookback requests
349  * @sqp: the sending QP
350  *
351  * This is called from qib_do_send() to
352  * forward a WQE addressed to the same HCA.
353  * Note that although we are single threaded due to the tasklet, we still
354  * have to protect against post_send().  We don't have to worry about
355  * receive interrupts since this is a connected protocol and all packets
356  * will pass through here.
357  */
358 static void qib_ruc_loopback(struct rvt_qp *sqp)
359 {
360 	struct qib_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num);
361 	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
362 	struct qib_devdata *dd = ppd->dd;
363 	struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
364 	struct rvt_qp *qp;
365 	struct rvt_swqe *wqe;
366 	struct rvt_sge *sge;
367 	unsigned long flags;
368 	struct ib_wc wc;
369 	u64 sdata;
370 	atomic64_t *maddr;
371 	enum ib_wc_status send_status;
372 	int release;
373 	int ret;
374 
375 	rcu_read_lock();
376 	/*
377 	 * Note that we check the responder QP state after
378 	 * checking the requester's state.
379 	 */
380 	qp = rvt_lookup_qpn(rdi, &ibp->rvp, sqp->remote_qpn);
381 	if (!qp)
382 		goto done;
383 
384 	spin_lock_irqsave(&sqp->s_lock, flags);
385 
386 	/* Return if we are already busy processing a work request. */
387 	if ((sqp->s_flags & (RVT_S_BUSY | RVT_S_ANY_WAIT)) ||
388 	    !(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_OR_FLUSH_SEND))
389 		goto unlock;
390 
391 	sqp->s_flags |= RVT_S_BUSY;
392 
393 again:
394 	smp_read_barrier_depends(); /* see post_one_send() */
395 	if (sqp->s_last == ACCESS_ONCE(sqp->s_head))
396 		goto clr_busy;
397 	wqe = rvt_get_swqe_ptr(sqp, sqp->s_last);
398 
399 	/* Return if it is not OK to start a new work reqeust. */
400 	if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_NEXT_SEND_OK)) {
401 		if (!(ib_rvt_state_ops[sqp->state] & RVT_FLUSH_SEND))
402 			goto clr_busy;
403 		/* We are in the error state, flush the work request. */
404 		send_status = IB_WC_WR_FLUSH_ERR;
405 		goto flush_send;
406 	}
407 
408 	/*
409 	 * We can rely on the entry not changing without the s_lock
410 	 * being held until we update s_last.
411 	 * We increment s_cur to indicate s_last is in progress.
412 	 */
413 	if (sqp->s_last == sqp->s_cur) {
414 		if (++sqp->s_cur >= sqp->s_size)
415 			sqp->s_cur = 0;
416 	}
417 	spin_unlock_irqrestore(&sqp->s_lock, flags);
418 
419 	if (!qp || !(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) ||
420 	    qp->ibqp.qp_type != sqp->ibqp.qp_type) {
421 		ibp->rvp.n_pkt_drops++;
422 		/*
423 		 * For RC, the requester would timeout and retry so
424 		 * shortcut the timeouts and just signal too many retries.
425 		 */
426 		if (sqp->ibqp.qp_type == IB_QPT_RC)
427 			send_status = IB_WC_RETRY_EXC_ERR;
428 		else
429 			send_status = IB_WC_SUCCESS;
430 		goto serr;
431 	}
432 
433 	memset(&wc, 0, sizeof(wc));
434 	send_status = IB_WC_SUCCESS;
435 
436 	release = 1;
437 	sqp->s_sge.sge = wqe->sg_list[0];
438 	sqp->s_sge.sg_list = wqe->sg_list + 1;
439 	sqp->s_sge.num_sge = wqe->wr.num_sge;
440 	sqp->s_len = wqe->length;
441 	switch (wqe->wr.opcode) {
442 	case IB_WR_SEND_WITH_IMM:
443 		wc.wc_flags = IB_WC_WITH_IMM;
444 		wc.ex.imm_data = wqe->wr.ex.imm_data;
445 		/* FALLTHROUGH */
446 	case IB_WR_SEND:
447 		ret = qib_get_rwqe(qp, 0);
448 		if (ret < 0)
449 			goto op_err;
450 		if (!ret)
451 			goto rnr_nak;
452 		break;
453 
454 	case IB_WR_RDMA_WRITE_WITH_IMM:
455 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
456 			goto inv_err;
457 		wc.wc_flags = IB_WC_WITH_IMM;
458 		wc.ex.imm_data = wqe->wr.ex.imm_data;
459 		ret = qib_get_rwqe(qp, 1);
460 		if (ret < 0)
461 			goto op_err;
462 		if (!ret)
463 			goto rnr_nak;
464 		/* FALLTHROUGH */
465 	case IB_WR_RDMA_WRITE:
466 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
467 			goto inv_err;
468 		if (wqe->length == 0)
469 			break;
470 		if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, wqe->length,
471 					  wqe->rdma_wr.remote_addr,
472 					  wqe->rdma_wr.rkey,
473 					  IB_ACCESS_REMOTE_WRITE)))
474 			goto acc_err;
475 		qp->r_sge.sg_list = NULL;
476 		qp->r_sge.num_sge = 1;
477 		qp->r_sge.total_len = wqe->length;
478 		break;
479 
480 	case IB_WR_RDMA_READ:
481 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
482 			goto inv_err;
483 		if (unlikely(!rvt_rkey_ok(qp, &sqp->s_sge.sge, wqe->length,
484 					  wqe->rdma_wr.remote_addr,
485 					  wqe->rdma_wr.rkey,
486 					  IB_ACCESS_REMOTE_READ)))
487 			goto acc_err;
488 		release = 0;
489 		sqp->s_sge.sg_list = NULL;
490 		sqp->s_sge.num_sge = 1;
491 		qp->r_sge.sge = wqe->sg_list[0];
492 		qp->r_sge.sg_list = wqe->sg_list + 1;
493 		qp->r_sge.num_sge = wqe->wr.num_sge;
494 		qp->r_sge.total_len = wqe->length;
495 		break;
496 
497 	case IB_WR_ATOMIC_CMP_AND_SWP:
498 	case IB_WR_ATOMIC_FETCH_AND_ADD:
499 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
500 			goto inv_err;
501 		if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64),
502 					  wqe->atomic_wr.remote_addr,
503 					  wqe->atomic_wr.rkey,
504 					  IB_ACCESS_REMOTE_ATOMIC)))
505 			goto acc_err;
506 		/* Perform atomic OP and save result. */
507 		maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
508 		sdata = wqe->atomic_wr.compare_add;
509 		*(u64 *) sqp->s_sge.sge.vaddr =
510 			(wqe->atomic_wr.wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ?
511 			(u64) atomic64_add_return(sdata, maddr) - sdata :
512 			(u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
513 				      sdata, wqe->atomic_wr.swap);
514 		rvt_put_mr(qp->r_sge.sge.mr);
515 		qp->r_sge.num_sge = 0;
516 		goto send_comp;
517 
518 	default:
519 		send_status = IB_WC_LOC_QP_OP_ERR;
520 		goto serr;
521 	}
522 
523 	sge = &sqp->s_sge.sge;
524 	while (sqp->s_len) {
525 		u32 len = sqp->s_len;
526 
527 		if (len > sge->length)
528 			len = sge->length;
529 		if (len > sge->sge_length)
530 			len = sge->sge_length;
531 		BUG_ON(len == 0);
532 		qib_copy_sge(&qp->r_sge, sge->vaddr, len, release);
533 		sge->vaddr += len;
534 		sge->length -= len;
535 		sge->sge_length -= len;
536 		if (sge->sge_length == 0) {
537 			if (!release)
538 				rvt_put_mr(sge->mr);
539 			if (--sqp->s_sge.num_sge)
540 				*sge = *sqp->s_sge.sg_list++;
541 		} else if (sge->length == 0 && sge->mr->lkey) {
542 			if (++sge->n >= RVT_SEGSZ) {
543 				if (++sge->m >= sge->mr->mapsz)
544 					break;
545 				sge->n = 0;
546 			}
547 			sge->vaddr =
548 				sge->mr->map[sge->m]->segs[sge->n].vaddr;
549 			sge->length =
550 				sge->mr->map[sge->m]->segs[sge->n].length;
551 		}
552 		sqp->s_len -= len;
553 	}
554 	if (release)
555 		rvt_put_ss(&qp->r_sge);
556 
557 	if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
558 		goto send_comp;
559 
560 	if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
561 		wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
562 	else
563 		wc.opcode = IB_WC_RECV;
564 	wc.wr_id = qp->r_wr_id;
565 	wc.status = IB_WC_SUCCESS;
566 	wc.byte_len = wqe->length;
567 	wc.qp = &qp->ibqp;
568 	wc.src_qp = qp->remote_qpn;
569 	wc.slid = qp->remote_ah_attr.dlid;
570 	wc.sl = qp->remote_ah_attr.sl;
571 	wc.port_num = 1;
572 	/* Signal completion event if the solicited bit is set. */
573 	rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc,
574 		     wqe->wr.send_flags & IB_SEND_SOLICITED);
575 
576 send_comp:
577 	spin_lock_irqsave(&sqp->s_lock, flags);
578 	ibp->rvp.n_loop_pkts++;
579 flush_send:
580 	sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
581 	qib_send_complete(sqp, wqe, send_status);
582 	goto again;
583 
584 rnr_nak:
585 	/* Handle RNR NAK */
586 	if (qp->ibqp.qp_type == IB_QPT_UC)
587 		goto send_comp;
588 	ibp->rvp.n_rnr_naks++;
589 	/*
590 	 * Note: we don't need the s_lock held since the BUSY flag
591 	 * makes this single threaded.
592 	 */
593 	if (sqp->s_rnr_retry == 0) {
594 		send_status = IB_WC_RNR_RETRY_EXC_ERR;
595 		goto serr;
596 	}
597 	if (sqp->s_rnr_retry_cnt < 7)
598 		sqp->s_rnr_retry--;
599 	spin_lock_irqsave(&sqp->s_lock, flags);
600 	if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_RECV_OK))
601 		goto clr_busy;
602 	sqp->s_flags |= RVT_S_WAIT_RNR;
603 	sqp->s_timer.function = qib_rc_rnr_retry;
604 	sqp->s_timer.expires = jiffies +
605 		usecs_to_jiffies(ib_qib_rnr_table[qp->r_min_rnr_timer]);
606 	add_timer(&sqp->s_timer);
607 	goto clr_busy;
608 
609 op_err:
610 	send_status = IB_WC_REM_OP_ERR;
611 	wc.status = IB_WC_LOC_QP_OP_ERR;
612 	goto err;
613 
614 inv_err:
615 	send_status = IB_WC_REM_INV_REQ_ERR;
616 	wc.status = IB_WC_LOC_QP_OP_ERR;
617 	goto err;
618 
619 acc_err:
620 	send_status = IB_WC_REM_ACCESS_ERR;
621 	wc.status = IB_WC_LOC_PROT_ERR;
622 err:
623 	/* responder goes to error state */
624 	qib_rc_error(qp, wc.status);
625 
626 serr:
627 	spin_lock_irqsave(&sqp->s_lock, flags);
628 	qib_send_complete(sqp, wqe, send_status);
629 	if (sqp->ibqp.qp_type == IB_QPT_RC) {
630 		int lastwqe = rvt_error_qp(sqp, IB_WC_WR_FLUSH_ERR);
631 
632 		sqp->s_flags &= ~RVT_S_BUSY;
633 		spin_unlock_irqrestore(&sqp->s_lock, flags);
634 		if (lastwqe) {
635 			struct ib_event ev;
636 
637 			ev.device = sqp->ibqp.device;
638 			ev.element.qp = &sqp->ibqp;
639 			ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
640 			sqp->ibqp.event_handler(&ev, sqp->ibqp.qp_context);
641 		}
642 		goto done;
643 	}
644 clr_busy:
645 	sqp->s_flags &= ~RVT_S_BUSY;
646 unlock:
647 	spin_unlock_irqrestore(&sqp->s_lock, flags);
648 done:
649 	rcu_read_unlock();
650 }
651 
652 /**
653  * qib_make_grh - construct a GRH header
654  * @ibp: a pointer to the IB port
655  * @hdr: a pointer to the GRH header being constructed
656  * @grh: the global route address to send to
657  * @hwords: the number of 32 bit words of header being sent
658  * @nwords: the number of 32 bit words of data being sent
659  *
660  * Return the size of the header in 32 bit words.
661  */
662 u32 qib_make_grh(struct qib_ibport *ibp, struct ib_grh *hdr,
663 		 struct ib_global_route *grh, u32 hwords, u32 nwords)
664 {
665 	hdr->version_tclass_flow =
666 		cpu_to_be32((IB_GRH_VERSION << IB_GRH_VERSION_SHIFT) |
667 			    (grh->traffic_class << IB_GRH_TCLASS_SHIFT) |
668 			    (grh->flow_label << IB_GRH_FLOW_SHIFT));
669 	hdr->paylen = cpu_to_be16((hwords - 2 + nwords + SIZE_OF_CRC) << 2);
670 	/* next_hdr is defined by C8-7 in ch. 8.4.1 */
671 	hdr->next_hdr = IB_GRH_NEXT_HDR;
672 	hdr->hop_limit = grh->hop_limit;
673 	/* The SGID is 32-bit aligned. */
674 	hdr->sgid.global.subnet_prefix = ibp->rvp.gid_prefix;
675 	hdr->sgid.global.interface_id = grh->sgid_index ?
676 		ibp->guids[grh->sgid_index - 1] : ppd_from_ibp(ibp)->guid;
677 	hdr->dgid = grh->dgid;
678 
679 	/* GRH header size in 32-bit words. */
680 	return sizeof(struct ib_grh) / sizeof(u32);
681 }
682 
683 void qib_make_ruc_header(struct rvt_qp *qp, struct qib_other_headers *ohdr,
684 			 u32 bth0, u32 bth2)
685 {
686 	struct qib_qp_priv *priv = qp->priv;
687 	struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
688 	u16 lrh0;
689 	u32 nwords;
690 	u32 extra_bytes;
691 
692 	/* Construct the header. */
693 	extra_bytes = -qp->s_cur_size & 3;
694 	nwords = (qp->s_cur_size + extra_bytes) >> 2;
695 	lrh0 = QIB_LRH_BTH;
696 	if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
697 		qp->s_hdrwords += qib_make_grh(ibp, &priv->s_hdr->u.l.grh,
698 					       &qp->remote_ah_attr.grh,
699 					       qp->s_hdrwords, nwords);
700 		lrh0 = QIB_LRH_GRH;
701 	}
702 	lrh0 |= ibp->sl_to_vl[qp->remote_ah_attr.sl] << 12 |
703 		qp->remote_ah_attr.sl << 4;
704 	priv->s_hdr->lrh[0] = cpu_to_be16(lrh0);
705 	priv->s_hdr->lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
706 	priv->s_hdr->lrh[2] =
707 			cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC);
708 	priv->s_hdr->lrh[3] = cpu_to_be16(ppd_from_ibp(ibp)->lid |
709 				       qp->remote_ah_attr.src_path_bits);
710 	bth0 |= qib_get_pkey(ibp, qp->s_pkey_index);
711 	bth0 |= extra_bytes << 20;
712 	if (qp->s_mig_state == IB_MIG_MIGRATED)
713 		bth0 |= IB_BTH_MIG_REQ;
714 	ohdr->bth[0] = cpu_to_be32(bth0);
715 	ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
716 	ohdr->bth[2] = cpu_to_be32(bth2);
717 	this_cpu_inc(ibp->pmastats->n_unicast_xmit);
718 }
719 
720 void _qib_do_send(struct work_struct *work)
721 {
722 	struct qib_qp_priv *priv = container_of(work, struct qib_qp_priv,
723 						s_work);
724 	struct rvt_qp *qp = priv->owner;
725 
726 	qib_do_send(qp);
727 }
728 
729 /**
730  * qib_do_send - perform a send on a QP
731  * @qp: pointer to the QP
732  *
733  * Process entries in the send work queue until credit or queue is
734  * exhausted.  Only allow one CPU to send a packet per QP (tasklet).
735  * Otherwise, two threads could send packets out of order.
736  */
737 void qib_do_send(struct rvt_qp *qp)
738 {
739 	struct qib_qp_priv *priv = qp->priv;
740 	struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
741 	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
742 	int (*make_req)(struct rvt_qp *qp, unsigned long *flags);
743 	unsigned long flags;
744 
745 	if ((qp->ibqp.qp_type == IB_QPT_RC ||
746 	     qp->ibqp.qp_type == IB_QPT_UC) &&
747 	    (qp->remote_ah_attr.dlid & ~((1 << ppd->lmc) - 1)) == ppd->lid) {
748 		qib_ruc_loopback(qp);
749 		return;
750 	}
751 
752 	if (qp->ibqp.qp_type == IB_QPT_RC)
753 		make_req = qib_make_rc_req;
754 	else if (qp->ibqp.qp_type == IB_QPT_UC)
755 		make_req = qib_make_uc_req;
756 	else
757 		make_req = qib_make_ud_req;
758 
759 	spin_lock_irqsave(&qp->s_lock, flags);
760 
761 	/* Return if we are already busy processing a work request. */
762 	if (!qib_send_ok(qp)) {
763 		spin_unlock_irqrestore(&qp->s_lock, flags);
764 		return;
765 	}
766 
767 	qp->s_flags |= RVT_S_BUSY;
768 
769 	do {
770 		/* Check for a constructed packet to be sent. */
771 		if (qp->s_hdrwords != 0) {
772 			spin_unlock_irqrestore(&qp->s_lock, flags);
773 			/*
774 			 * If the packet cannot be sent now, return and
775 			 * the send tasklet will be woken up later.
776 			 */
777 			if (qib_verbs_send(qp, priv->s_hdr, qp->s_hdrwords,
778 					   qp->s_cur_sge, qp->s_cur_size))
779 				return;
780 			/* Record that s_hdr is empty. */
781 			qp->s_hdrwords = 0;
782 			spin_lock_irqsave(&qp->s_lock, flags);
783 		}
784 	} while (make_req(qp, &flags));
785 
786 	spin_unlock_irqrestore(&qp->s_lock, flags);
787 }
788 
789 /*
790  * This should be called with s_lock held.
791  */
792 void qib_send_complete(struct rvt_qp *qp, struct rvt_swqe *wqe,
793 		       enum ib_wc_status status)
794 {
795 	u32 old_last, last;
796 	unsigned i;
797 
798 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_OR_FLUSH_SEND))
799 		return;
800 
801 	last = qp->s_last;
802 	old_last = last;
803 	if (++last >= qp->s_size)
804 		last = 0;
805 	qp->s_last = last;
806 	/* See post_send() */
807 	barrier();
808 	for (i = 0; i < wqe->wr.num_sge; i++) {
809 		struct rvt_sge *sge = &wqe->sg_list[i];
810 
811 		rvt_put_mr(sge->mr);
812 	}
813 	if (qp->ibqp.qp_type == IB_QPT_UD ||
814 	    qp->ibqp.qp_type == IB_QPT_SMI ||
815 	    qp->ibqp.qp_type == IB_QPT_GSI)
816 		atomic_dec(&ibah_to_rvtah(wqe->ud_wr.ah)->refcount);
817 
818 	/* See ch. 11.2.4.1 and 10.7.3.1 */
819 	if (!(qp->s_flags & RVT_S_SIGNAL_REQ_WR) ||
820 	    (wqe->wr.send_flags & IB_SEND_SIGNALED) ||
821 	    status != IB_WC_SUCCESS) {
822 		struct ib_wc wc;
823 
824 		memset(&wc, 0, sizeof(wc));
825 		wc.wr_id = wqe->wr.wr_id;
826 		wc.status = status;
827 		wc.opcode = ib_qib_wc_opcode[wqe->wr.opcode];
828 		wc.qp = &qp->ibqp;
829 		if (status == IB_WC_SUCCESS)
830 			wc.byte_len = wqe->length;
831 		rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.send_cq), &wc,
832 			     status != IB_WC_SUCCESS);
833 	}
834 
835 	if (qp->s_acked == old_last)
836 		qp->s_acked = last;
837 	if (qp->s_cur == old_last)
838 		qp->s_cur = last;
839 	if (qp->s_tail == old_last)
840 		qp->s_tail = last;
841 	if (qp->state == IB_QPS_SQD && last == qp->s_cur)
842 		qp->s_draining = 0;
843 }
844