xref: /openbmc/linux/drivers/infiniband/hw/hfi1/ruc.c (revision 293d5b43)
1 /*
2  * Copyright(c) 2015, 2016 Intel Corporation.
3  *
4  * This file is provided under a dual BSD/GPLv2 license.  When using or
5  * redistributing this file, you may do so under either license.
6  *
7  * GPL LICENSE SUMMARY
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of version 2 of the GNU General Public License as
11  * published by the Free Software Foundation.
12  *
13  * This program is distributed in the hope that it will be useful, but
14  * WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
16  * General Public License for more details.
17  *
18  * BSD LICENSE
19  *
20  * Redistribution and use in source and binary forms, with or without
21  * modification, are permitted provided that the following conditions
22  * are met:
23  *
24  *  - Redistributions of source code must retain the above copyright
25  *    notice, this list of conditions and the following disclaimer.
26  *  - Redistributions in binary form must reproduce the above copyright
27  *    notice, this list of conditions and the following disclaimer in
28  *    the documentation and/or other materials provided with the
29  *    distribution.
30  *  - Neither the name of Intel Corporation nor the names of its
31  *    contributors may be used to endorse or promote products derived
32  *    from this software without specific prior written permission.
33  *
34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
45  *
46  */
47 
48 #include <linux/spinlock.h>
49 
50 #include "hfi.h"
51 #include "mad.h"
52 #include "qp.h"
53 #include "verbs_txreq.h"
54 #include "trace.h"
55 
56 /*
57  * Convert the AETH RNR timeout code into the number of microseconds.
58  */
59 const u32 ib_hfi1_rnr_table[32] = {
60 	655360,	/* 00: 655.36 */
61 	10,	/* 01:    .01 */
62 	20,	/* 02     .02 */
63 	30,	/* 03:    .03 */
64 	40,	/* 04:    .04 */
65 	60,	/* 05:    .06 */
66 	80,	/* 06:    .08 */
67 	120,	/* 07:    .12 */
68 	160,	/* 08:    .16 */
69 	240,	/* 09:    .24 */
70 	320,	/* 0A:    .32 */
71 	480,	/* 0B:    .48 */
72 	640,	/* 0C:    .64 */
73 	960,	/* 0D:    .96 */
74 	1280,	/* 0E:   1.28 */
75 	1920,	/* 0F:   1.92 */
76 	2560,	/* 10:   2.56 */
77 	3840,	/* 11:   3.84 */
78 	5120,	/* 12:   5.12 */
79 	7680,	/* 13:   7.68 */
80 	10240,	/* 14:  10.24 */
81 	15360,	/* 15:  15.36 */
82 	20480,	/* 16:  20.48 */
83 	30720,	/* 17:  30.72 */
84 	40960,	/* 18:  40.96 */
85 	61440,	/* 19:  61.44 */
86 	81920,	/* 1A:  81.92 */
87 	122880,	/* 1B: 122.88 */
88 	163840,	/* 1C: 163.84 */
89 	245760,	/* 1D: 245.76 */
90 	327680,	/* 1E: 327.68 */
91 	491520	/* 1F: 491.52 */
92 };
93 
94 /*
95  * Validate a RWQE and fill in the SGE state.
96  * Return 1 if OK.
97  */
98 static int init_sge(struct rvt_qp *qp, struct rvt_rwqe *wqe)
99 {
100 	int i, j, ret;
101 	struct ib_wc wc;
102 	struct rvt_lkey_table *rkt;
103 	struct rvt_pd *pd;
104 	struct rvt_sge_state *ss;
105 
106 	rkt = &to_idev(qp->ibqp.device)->rdi.lkey_table;
107 	pd = ibpd_to_rvtpd(qp->ibqp.srq ? qp->ibqp.srq->pd : qp->ibqp.pd);
108 	ss = &qp->r_sge;
109 	ss->sg_list = qp->r_sg_list;
110 	qp->r_len = 0;
111 	for (i = j = 0; i < wqe->num_sge; i++) {
112 		if (wqe->sg_list[i].length == 0)
113 			continue;
114 		/* Check LKEY */
115 		if (!rvt_lkey_ok(rkt, pd, j ? &ss->sg_list[j - 1] : &ss->sge,
116 				 &wqe->sg_list[i], IB_ACCESS_LOCAL_WRITE))
117 			goto bad_lkey;
118 		qp->r_len += wqe->sg_list[i].length;
119 		j++;
120 	}
121 	ss->num_sge = j;
122 	ss->total_len = qp->r_len;
123 	ret = 1;
124 	goto bail;
125 
126 bad_lkey:
127 	while (j) {
128 		struct rvt_sge *sge = --j ? &ss->sg_list[j - 1] : &ss->sge;
129 
130 		rvt_put_mr(sge->mr);
131 	}
132 	ss->num_sge = 0;
133 	memset(&wc, 0, sizeof(wc));
134 	wc.wr_id = wqe->wr_id;
135 	wc.status = IB_WC_LOC_PROT_ERR;
136 	wc.opcode = IB_WC_RECV;
137 	wc.qp = &qp->ibqp;
138 	/* Signal solicited completion event. */
139 	rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
140 	ret = 0;
141 bail:
142 	return ret;
143 }
144 
145 /**
146  * hfi1_rvt_get_rwqe - copy the next RWQE into the QP's RWQE
147  * @qp: the QP
148  * @wr_id_only: update qp->r_wr_id only, not qp->r_sge
149  *
150  * Return -1 if there is a local error, 0 if no RWQE is available,
151  * otherwise return 1.
152  *
153  * Can be called from interrupt level.
154  */
155 int hfi1_rvt_get_rwqe(struct rvt_qp *qp, int wr_id_only)
156 {
157 	unsigned long flags;
158 	struct rvt_rq *rq;
159 	struct rvt_rwq *wq;
160 	struct rvt_srq *srq;
161 	struct rvt_rwqe *wqe;
162 	void (*handler)(struct ib_event *, void *);
163 	u32 tail;
164 	int ret;
165 
166 	if (qp->ibqp.srq) {
167 		srq = ibsrq_to_rvtsrq(qp->ibqp.srq);
168 		handler = srq->ibsrq.event_handler;
169 		rq = &srq->rq;
170 	} else {
171 		srq = NULL;
172 		handler = NULL;
173 		rq = &qp->r_rq;
174 	}
175 
176 	spin_lock_irqsave(&rq->lock, flags);
177 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) {
178 		ret = 0;
179 		goto unlock;
180 	}
181 
182 	wq = rq->wq;
183 	tail = wq->tail;
184 	/* Validate tail before using it since it is user writable. */
185 	if (tail >= rq->size)
186 		tail = 0;
187 	if (unlikely(tail == wq->head)) {
188 		ret = 0;
189 		goto unlock;
190 	}
191 	/* Make sure entry is read after head index is read. */
192 	smp_rmb();
193 	wqe = rvt_get_rwqe_ptr(rq, tail);
194 	/*
195 	 * Even though we update the tail index in memory, the verbs
196 	 * consumer is not supposed to post more entries until a
197 	 * completion is generated.
198 	 */
199 	if (++tail >= rq->size)
200 		tail = 0;
201 	wq->tail = tail;
202 	if (!wr_id_only && !init_sge(qp, wqe)) {
203 		ret = -1;
204 		goto unlock;
205 	}
206 	qp->r_wr_id = wqe->wr_id;
207 
208 	ret = 1;
209 	set_bit(RVT_R_WRID_VALID, &qp->r_aflags);
210 	if (handler) {
211 		u32 n;
212 
213 		/*
214 		 * Validate head pointer value and compute
215 		 * the number of remaining WQEs.
216 		 */
217 		n = wq->head;
218 		if (n >= rq->size)
219 			n = 0;
220 		if (n < tail)
221 			n += rq->size - tail;
222 		else
223 			n -= tail;
224 		if (n < srq->limit) {
225 			struct ib_event ev;
226 
227 			srq->limit = 0;
228 			spin_unlock_irqrestore(&rq->lock, flags);
229 			ev.device = qp->ibqp.device;
230 			ev.element.srq = qp->ibqp.srq;
231 			ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
232 			handler(&ev, srq->ibsrq.srq_context);
233 			goto bail;
234 		}
235 	}
236 unlock:
237 	spin_unlock_irqrestore(&rq->lock, flags);
238 bail:
239 	return ret;
240 }
241 
242 static __be64 get_sguid(struct hfi1_ibport *ibp, unsigned index)
243 {
244 	if (!index) {
245 		struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
246 
247 		return cpu_to_be64(ppd->guid);
248 	}
249 	return ibp->guids[index - 1];
250 }
251 
252 static int gid_ok(union ib_gid *gid, __be64 gid_prefix, __be64 id)
253 {
254 	return (gid->global.interface_id == id &&
255 		(gid->global.subnet_prefix == gid_prefix ||
256 		 gid->global.subnet_prefix == IB_DEFAULT_GID_PREFIX));
257 }
258 
259 /*
260  *
261  * This should be called with the QP r_lock held.
262  *
263  * The s_lock will be acquired around the hfi1_migrate_qp() call.
264  */
265 int hfi1_ruc_check_hdr(struct hfi1_ibport *ibp, struct hfi1_ib_header *hdr,
266 		       int has_grh, struct rvt_qp *qp, u32 bth0)
267 {
268 	__be64 guid;
269 	unsigned long flags;
270 	u8 sc5 = ibp->sl_to_sc[qp->remote_ah_attr.sl];
271 
272 	if (qp->s_mig_state == IB_MIG_ARMED && (bth0 & IB_BTH_MIG_REQ)) {
273 		if (!has_grh) {
274 			if (qp->alt_ah_attr.ah_flags & IB_AH_GRH)
275 				goto err;
276 		} else {
277 			if (!(qp->alt_ah_attr.ah_flags & IB_AH_GRH))
278 				goto err;
279 			guid = get_sguid(ibp, qp->alt_ah_attr.grh.sgid_index);
280 			if (!gid_ok(&hdr->u.l.grh.dgid, ibp->rvp.gid_prefix,
281 				    guid))
282 				goto err;
283 			if (!gid_ok(
284 				&hdr->u.l.grh.sgid,
285 				qp->alt_ah_attr.grh.dgid.global.subnet_prefix,
286 				qp->alt_ah_attr.grh.dgid.global.interface_id))
287 				goto err;
288 		}
289 		if (unlikely(rcv_pkey_check(ppd_from_ibp(ibp), (u16)bth0,
290 					    sc5, be16_to_cpu(hdr->lrh[3])))) {
291 			hfi1_bad_pqkey(ibp, OPA_TRAP_BAD_P_KEY,
292 				       (u16)bth0,
293 				       (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF,
294 				       0, qp->ibqp.qp_num,
295 				       be16_to_cpu(hdr->lrh[3]),
296 				       be16_to_cpu(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 		hfi1_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, ibp->rvp.gid_prefix,
316 				    guid))
317 				goto err;
318 			if (!gid_ok(
319 			     &hdr->u.l.grh.sgid,
320 			     qp->remote_ah_attr.grh.dgid.global.subnet_prefix,
321 			     qp->remote_ah_attr.grh.dgid.global.interface_id))
322 				goto err;
323 		}
324 		if (unlikely(rcv_pkey_check(ppd_from_ibp(ibp), (u16)bth0,
325 					    sc5, be16_to_cpu(hdr->lrh[3])))) {
326 			hfi1_bad_pqkey(ibp, OPA_TRAP_BAD_P_KEY,
327 				       (u16)bth0,
328 				       (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF,
329 				       0, qp->ibqp.qp_num,
330 				       be16_to_cpu(hdr->lrh[3]),
331 				       be16_to_cpu(hdr->lrh[1]));
332 			goto err;
333 		}
334 		/* Validate the SLID. See Ch. 9.6.1.5 */
335 		if (be16_to_cpu(hdr->lrh[3]) != qp->remote_ah_attr.dlid ||
336 		    ppd_from_ibp(ibp)->port != qp->port_num)
337 			goto err;
338 		if (qp->s_mig_state == IB_MIG_REARM &&
339 		    !(bth0 & IB_BTH_MIG_REQ))
340 			qp->s_mig_state = IB_MIG_ARMED;
341 	}
342 
343 	return 0;
344 
345 err:
346 	return 1;
347 }
348 
349 /**
350  * ruc_loopback - handle UC and RC loopback requests
351  * @sqp: the sending QP
352  *
353  * This is called from hfi1_do_send() to
354  * forward a WQE addressed to the same HFI.
355  * Note that although we are single threaded due to the tasklet, we still
356  * have to protect against post_send().  We don't have to worry about
357  * receive interrupts since this is a connected protocol and all packets
358  * will pass through here.
359  */
360 static void ruc_loopback(struct rvt_qp *sqp)
361 {
362 	struct hfi1_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num);
363 	struct rvt_qp *qp;
364 	struct rvt_swqe *wqe;
365 	struct rvt_sge *sge;
366 	unsigned long flags;
367 	struct ib_wc wc;
368 	u64 sdata;
369 	atomic64_t *maddr;
370 	enum ib_wc_status send_status;
371 	int release;
372 	int ret;
373 	int copy_last = 0;
374 	u32 to;
375 	int local_ops = 0;
376 
377 	rcu_read_lock();
378 
379 	/*
380 	 * Note that we check the responder QP state after
381 	 * checking the requester's state.
382 	 */
383 	qp = rvt_lookup_qpn(ib_to_rvt(sqp->ibqp.device), &ibp->rvp,
384 			    sqp->remote_qpn);
385 
386 	spin_lock_irqsave(&sqp->s_lock, flags);
387 
388 	/* Return if we are already busy processing a work request. */
389 	if ((sqp->s_flags & (RVT_S_BUSY | RVT_S_ANY_WAIT)) ||
390 	    !(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_OR_FLUSH_SEND))
391 		goto unlock;
392 
393 	sqp->s_flags |= RVT_S_BUSY;
394 
395 again:
396 	smp_read_barrier_depends(); /* see post_one_send() */
397 	if (sqp->s_last == ACCESS_ONCE(sqp->s_head))
398 		goto clr_busy;
399 	wqe = rvt_get_swqe_ptr(sqp, sqp->s_last);
400 
401 	/* Return if it is not OK to start a new work request. */
402 	if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_NEXT_SEND_OK)) {
403 		if (!(ib_rvt_state_ops[sqp->state] & RVT_FLUSH_SEND))
404 			goto clr_busy;
405 		/* We are in the error state, flush the work request. */
406 		send_status = IB_WC_WR_FLUSH_ERR;
407 		goto flush_send;
408 	}
409 
410 	/*
411 	 * We can rely on the entry not changing without the s_lock
412 	 * being held until we update s_last.
413 	 * We increment s_cur to indicate s_last is in progress.
414 	 */
415 	if (sqp->s_last == sqp->s_cur) {
416 		if (++sqp->s_cur >= sqp->s_size)
417 			sqp->s_cur = 0;
418 	}
419 	spin_unlock_irqrestore(&sqp->s_lock, flags);
420 
421 	if (!qp || !(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) ||
422 	    qp->ibqp.qp_type != sqp->ibqp.qp_type) {
423 		ibp->rvp.n_pkt_drops++;
424 		/*
425 		 * For RC, the requester would timeout and retry so
426 		 * shortcut the timeouts and just signal too many retries.
427 		 */
428 		if (sqp->ibqp.qp_type == IB_QPT_RC)
429 			send_status = IB_WC_RETRY_EXC_ERR;
430 		else
431 			send_status = IB_WC_SUCCESS;
432 		goto serr;
433 	}
434 
435 	memset(&wc, 0, sizeof(wc));
436 	send_status = IB_WC_SUCCESS;
437 
438 	release = 1;
439 	sqp->s_sge.sge = wqe->sg_list[0];
440 	sqp->s_sge.sg_list = wqe->sg_list + 1;
441 	sqp->s_sge.num_sge = wqe->wr.num_sge;
442 	sqp->s_len = wqe->length;
443 	switch (wqe->wr.opcode) {
444 	case IB_WR_REG_MR:
445 		goto send_comp;
446 
447 	case IB_WR_LOCAL_INV:
448 		if (!(wqe->wr.send_flags & RVT_SEND_COMPLETION_ONLY)) {
449 			if (rvt_invalidate_rkey(sqp,
450 						wqe->wr.ex.invalidate_rkey))
451 				send_status = IB_WC_LOC_PROT_ERR;
452 			local_ops = 1;
453 		}
454 		goto send_comp;
455 
456 	case IB_WR_SEND_WITH_INV:
457 		if (!rvt_invalidate_rkey(qp, wqe->wr.ex.invalidate_rkey)) {
458 			wc.wc_flags = IB_WC_WITH_INVALIDATE;
459 			wc.ex.invalidate_rkey = wqe->wr.ex.invalidate_rkey;
460 		}
461 		goto send;
462 
463 	case IB_WR_SEND_WITH_IMM:
464 		wc.wc_flags = IB_WC_WITH_IMM;
465 		wc.ex.imm_data = wqe->wr.ex.imm_data;
466 		/* FALLTHROUGH */
467 	case IB_WR_SEND:
468 send:
469 		ret = hfi1_rvt_get_rwqe(qp, 0);
470 		if (ret < 0)
471 			goto op_err;
472 		if (!ret)
473 			goto rnr_nak;
474 		break;
475 
476 	case IB_WR_RDMA_WRITE_WITH_IMM:
477 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
478 			goto inv_err;
479 		wc.wc_flags = IB_WC_WITH_IMM;
480 		wc.ex.imm_data = wqe->wr.ex.imm_data;
481 		ret = hfi1_rvt_get_rwqe(qp, 1);
482 		if (ret < 0)
483 			goto op_err;
484 		if (!ret)
485 			goto rnr_nak;
486 		/* skip copy_last set and qp_access_flags recheck */
487 		goto do_write;
488 	case IB_WR_RDMA_WRITE:
489 		copy_last = ibpd_to_rvtpd(qp->ibqp.pd)->user;
490 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
491 			goto inv_err;
492 do_write:
493 		if (wqe->length == 0)
494 			break;
495 		if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, wqe->length,
496 					  wqe->rdma_wr.remote_addr,
497 					  wqe->rdma_wr.rkey,
498 					  IB_ACCESS_REMOTE_WRITE)))
499 			goto acc_err;
500 		qp->r_sge.sg_list = NULL;
501 		qp->r_sge.num_sge = 1;
502 		qp->r_sge.total_len = wqe->length;
503 		break;
504 
505 	case IB_WR_RDMA_READ:
506 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
507 			goto inv_err;
508 		if (unlikely(!rvt_rkey_ok(qp, &sqp->s_sge.sge, wqe->length,
509 					  wqe->rdma_wr.remote_addr,
510 					  wqe->rdma_wr.rkey,
511 					  IB_ACCESS_REMOTE_READ)))
512 			goto acc_err;
513 		release = 0;
514 		sqp->s_sge.sg_list = NULL;
515 		sqp->s_sge.num_sge = 1;
516 		qp->r_sge.sge = wqe->sg_list[0];
517 		qp->r_sge.sg_list = wqe->sg_list + 1;
518 		qp->r_sge.num_sge = wqe->wr.num_sge;
519 		qp->r_sge.total_len = wqe->length;
520 		break;
521 
522 	case IB_WR_ATOMIC_CMP_AND_SWP:
523 	case IB_WR_ATOMIC_FETCH_AND_ADD:
524 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
525 			goto inv_err;
526 		if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64),
527 					  wqe->atomic_wr.remote_addr,
528 					  wqe->atomic_wr.rkey,
529 					  IB_ACCESS_REMOTE_ATOMIC)))
530 			goto acc_err;
531 		/* Perform atomic OP and save result. */
532 		maddr = (atomic64_t *)qp->r_sge.sge.vaddr;
533 		sdata = wqe->atomic_wr.compare_add;
534 		*(u64 *)sqp->s_sge.sge.vaddr =
535 			(wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ?
536 			(u64)atomic64_add_return(sdata, maddr) - sdata :
537 			(u64)cmpxchg((u64 *)qp->r_sge.sge.vaddr,
538 				      sdata, wqe->atomic_wr.swap);
539 		rvt_put_mr(qp->r_sge.sge.mr);
540 		qp->r_sge.num_sge = 0;
541 		goto send_comp;
542 
543 	default:
544 		send_status = IB_WC_LOC_QP_OP_ERR;
545 		goto serr;
546 	}
547 
548 	sge = &sqp->s_sge.sge;
549 	while (sqp->s_len) {
550 		u32 len = sqp->s_len;
551 
552 		if (len > sge->length)
553 			len = sge->length;
554 		if (len > sge->sge_length)
555 			len = sge->sge_length;
556 		WARN_ON_ONCE(len == 0);
557 		hfi1_copy_sge(&qp->r_sge, sge->vaddr, len, release, copy_last);
558 		sge->vaddr += len;
559 		sge->length -= len;
560 		sge->sge_length -= len;
561 		if (sge->sge_length == 0) {
562 			if (!release)
563 				rvt_put_mr(sge->mr);
564 			if (--sqp->s_sge.num_sge)
565 				*sge = *sqp->s_sge.sg_list++;
566 		} else if (sge->length == 0 && sge->mr->lkey) {
567 			if (++sge->n >= RVT_SEGSZ) {
568 				if (++sge->m >= sge->mr->mapsz)
569 					break;
570 				sge->n = 0;
571 			}
572 			sge->vaddr =
573 				sge->mr->map[sge->m]->segs[sge->n].vaddr;
574 			sge->length =
575 				sge->mr->map[sge->m]->segs[sge->n].length;
576 		}
577 		sqp->s_len -= len;
578 	}
579 	if (release)
580 		rvt_put_ss(&qp->r_sge);
581 
582 	if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
583 		goto send_comp;
584 
585 	if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
586 		wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
587 	else
588 		wc.opcode = IB_WC_RECV;
589 	wc.wr_id = qp->r_wr_id;
590 	wc.status = IB_WC_SUCCESS;
591 	wc.byte_len = wqe->length;
592 	wc.qp = &qp->ibqp;
593 	wc.src_qp = qp->remote_qpn;
594 	wc.slid = qp->remote_ah_attr.dlid;
595 	wc.sl = qp->remote_ah_attr.sl;
596 	wc.port_num = 1;
597 	/* Signal completion event if the solicited bit is set. */
598 	rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc,
599 		     wqe->wr.send_flags & IB_SEND_SOLICITED);
600 
601 send_comp:
602 	spin_lock_irqsave(&sqp->s_lock, flags);
603 	ibp->rvp.n_loop_pkts++;
604 flush_send:
605 	sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
606 	hfi1_send_complete(sqp, wqe, send_status);
607 	if (local_ops) {
608 		atomic_dec(&sqp->local_ops_pending);
609 		local_ops = 0;
610 	}
611 	goto again;
612 
613 rnr_nak:
614 	/* Handle RNR NAK */
615 	if (qp->ibqp.qp_type == IB_QPT_UC)
616 		goto send_comp;
617 	ibp->rvp.n_rnr_naks++;
618 	/*
619 	 * Note: we don't need the s_lock held since the BUSY flag
620 	 * makes this single threaded.
621 	 */
622 	if (sqp->s_rnr_retry == 0) {
623 		send_status = IB_WC_RNR_RETRY_EXC_ERR;
624 		goto serr;
625 	}
626 	if (sqp->s_rnr_retry_cnt < 7)
627 		sqp->s_rnr_retry--;
628 	spin_lock_irqsave(&sqp->s_lock, flags);
629 	if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_RECV_OK))
630 		goto clr_busy;
631 	to = ib_hfi1_rnr_table[qp->r_min_rnr_timer];
632 	hfi1_add_rnr_timer(sqp, to);
633 	goto clr_busy;
634 
635 op_err:
636 	send_status = IB_WC_REM_OP_ERR;
637 	wc.status = IB_WC_LOC_QP_OP_ERR;
638 	goto err;
639 
640 inv_err:
641 	send_status = IB_WC_REM_INV_REQ_ERR;
642 	wc.status = IB_WC_LOC_QP_OP_ERR;
643 	goto err;
644 
645 acc_err:
646 	send_status = IB_WC_REM_ACCESS_ERR;
647 	wc.status = IB_WC_LOC_PROT_ERR;
648 err:
649 	/* responder goes to error state */
650 	hfi1_rc_error(qp, wc.status);
651 
652 serr:
653 	spin_lock_irqsave(&sqp->s_lock, flags);
654 	hfi1_send_complete(sqp, wqe, send_status);
655 	if (sqp->ibqp.qp_type == IB_QPT_RC) {
656 		int lastwqe = rvt_error_qp(sqp, IB_WC_WR_FLUSH_ERR);
657 
658 		sqp->s_flags &= ~RVT_S_BUSY;
659 		spin_unlock_irqrestore(&sqp->s_lock, flags);
660 		if (lastwqe) {
661 			struct ib_event ev;
662 
663 			ev.device = sqp->ibqp.device;
664 			ev.element.qp = &sqp->ibqp;
665 			ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
666 			sqp->ibqp.event_handler(&ev, sqp->ibqp.qp_context);
667 		}
668 		goto done;
669 	}
670 clr_busy:
671 	sqp->s_flags &= ~RVT_S_BUSY;
672 unlock:
673 	spin_unlock_irqrestore(&sqp->s_lock, flags);
674 done:
675 	rcu_read_unlock();
676 }
677 
678 /**
679  * hfi1_make_grh - construct a GRH header
680  * @ibp: a pointer to the IB port
681  * @hdr: a pointer to the GRH header being constructed
682  * @grh: the global route address to send to
683  * @hwords: the number of 32 bit words of header being sent
684  * @nwords: the number of 32 bit words of data being sent
685  *
686  * Return the size of the header in 32 bit words.
687  */
688 u32 hfi1_make_grh(struct hfi1_ibport *ibp, struct ib_grh *hdr,
689 		  struct ib_global_route *grh, u32 hwords, u32 nwords)
690 {
691 	hdr->version_tclass_flow =
692 		cpu_to_be32((IB_GRH_VERSION << IB_GRH_VERSION_SHIFT) |
693 			    (grh->traffic_class << IB_GRH_TCLASS_SHIFT) |
694 			    (grh->flow_label << IB_GRH_FLOW_SHIFT));
695 	hdr->paylen = cpu_to_be16((hwords - 2 + nwords + SIZE_OF_CRC) << 2);
696 	/* next_hdr is defined by C8-7 in ch. 8.4.1 */
697 	hdr->next_hdr = IB_GRH_NEXT_HDR;
698 	hdr->hop_limit = grh->hop_limit;
699 	/* The SGID is 32-bit aligned. */
700 	hdr->sgid.global.subnet_prefix = ibp->rvp.gid_prefix;
701 	hdr->sgid.global.interface_id =
702 		grh->sgid_index && grh->sgid_index < ARRAY_SIZE(ibp->guids) ?
703 		ibp->guids[grh->sgid_index - 1] :
704 			cpu_to_be64(ppd_from_ibp(ibp)->guid);
705 	hdr->dgid = grh->dgid;
706 
707 	/* GRH header size in 32-bit words. */
708 	return sizeof(struct ib_grh) / sizeof(u32);
709 }
710 
711 #define BTH2_OFFSET (offsetof(struct hfi1_sdma_header, hdr.u.oth.bth[2]) / 4)
712 
713 /**
714  * build_ahg - create ahg in s_ahg
715  * @qp: a pointer to QP
716  * @npsn: the next PSN for the request/response
717  *
718  * This routine handles the AHG by allocating an ahg entry and causing the
719  * copy of the first middle.
720  *
721  * Subsequent middles use the copied entry, editing the
722  * PSN with 1 or 2 edits.
723  */
724 static inline void build_ahg(struct rvt_qp *qp, u32 npsn)
725 {
726 	struct hfi1_qp_priv *priv = qp->priv;
727 
728 	if (unlikely(qp->s_flags & RVT_S_AHG_CLEAR))
729 		clear_ahg(qp);
730 	if (!(qp->s_flags & RVT_S_AHG_VALID)) {
731 		/* first middle that needs copy  */
732 		if (qp->s_ahgidx < 0)
733 			qp->s_ahgidx = sdma_ahg_alloc(priv->s_sde);
734 		if (qp->s_ahgidx >= 0) {
735 			qp->s_ahgpsn = npsn;
736 			priv->s_ahg->tx_flags |= SDMA_TXREQ_F_AHG_COPY;
737 			/* save to protect a change in another thread */
738 			priv->s_ahg->ahgidx = qp->s_ahgidx;
739 			qp->s_flags |= RVT_S_AHG_VALID;
740 		}
741 	} else {
742 		/* subsequent middle after valid */
743 		if (qp->s_ahgidx >= 0) {
744 			priv->s_ahg->tx_flags |= SDMA_TXREQ_F_USE_AHG;
745 			priv->s_ahg->ahgidx = qp->s_ahgidx;
746 			priv->s_ahg->ahgcount++;
747 			priv->s_ahg->ahgdesc[0] =
748 				sdma_build_ahg_descriptor(
749 					(__force u16)cpu_to_be16((u16)npsn),
750 					BTH2_OFFSET,
751 					16,
752 					16);
753 			if ((npsn & 0xffff0000) !=
754 					(qp->s_ahgpsn & 0xffff0000)) {
755 				priv->s_ahg->ahgcount++;
756 				priv->s_ahg->ahgdesc[1] =
757 					sdma_build_ahg_descriptor(
758 						(__force u16)cpu_to_be16(
759 							(u16)(npsn >> 16)),
760 						BTH2_OFFSET,
761 						0,
762 						16);
763 			}
764 		}
765 	}
766 }
767 
768 void hfi1_make_ruc_header(struct rvt_qp *qp, struct hfi1_other_headers *ohdr,
769 			  u32 bth0, u32 bth2, int middle,
770 			  struct hfi1_pkt_state *ps)
771 {
772 	struct hfi1_qp_priv *priv = qp->priv;
773 	struct hfi1_ibport *ibp = ps->ibp;
774 	u16 lrh0;
775 	u32 nwords;
776 	u32 extra_bytes;
777 	u32 bth1;
778 
779 	/* Construct the header. */
780 	extra_bytes = -qp->s_cur_size & 3;
781 	nwords = (qp->s_cur_size + extra_bytes) >> 2;
782 	lrh0 = HFI1_LRH_BTH;
783 	if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
784 		qp->s_hdrwords += hfi1_make_grh(ibp,
785 						&ps->s_txreq->phdr.hdr.u.l.grh,
786 						&qp->remote_ah_attr.grh,
787 						qp->s_hdrwords, nwords);
788 		lrh0 = HFI1_LRH_GRH;
789 		middle = 0;
790 	}
791 	lrh0 |= (priv->s_sc & 0xf) << 12 | (qp->remote_ah_attr.sl & 0xf) << 4;
792 	/*
793 	 * reset s_ahg/AHG fields
794 	 *
795 	 * This insures that the ahgentry/ahgcount
796 	 * are at a non-AHG default to protect
797 	 * build_verbs_tx_desc() from using
798 	 * an include ahgidx.
799 	 *
800 	 * build_ahg() will modify as appropriate
801 	 * to use the AHG feature.
802 	 */
803 	priv->s_ahg->tx_flags = 0;
804 	priv->s_ahg->ahgcount = 0;
805 	priv->s_ahg->ahgidx = 0;
806 	if (qp->s_mig_state == IB_MIG_MIGRATED)
807 		bth0 |= IB_BTH_MIG_REQ;
808 	else
809 		middle = 0;
810 	if (middle)
811 		build_ahg(qp, bth2);
812 	else
813 		qp->s_flags &= ~RVT_S_AHG_VALID;
814 	ps->s_txreq->phdr.hdr.lrh[0] = cpu_to_be16(lrh0);
815 	ps->s_txreq->phdr.hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
816 	ps->s_txreq->phdr.hdr.lrh[2] =
817 		cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC);
818 	ps->s_txreq->phdr.hdr.lrh[3] = cpu_to_be16(ppd_from_ibp(ibp)->lid |
819 				       qp->remote_ah_attr.src_path_bits);
820 	bth0 |= hfi1_get_pkey(ibp, qp->s_pkey_index);
821 	bth0 |= extra_bytes << 20;
822 	ohdr->bth[0] = cpu_to_be32(bth0);
823 	bth1 = qp->remote_qpn;
824 	if (qp->s_flags & RVT_S_ECN) {
825 		qp->s_flags &= ~RVT_S_ECN;
826 		/* we recently received a FECN, so return a BECN */
827 		bth1 |= (HFI1_BECN_MASK << HFI1_BECN_SHIFT);
828 	}
829 	ohdr->bth[1] = cpu_to_be32(bth1);
830 	ohdr->bth[2] = cpu_to_be32(bth2);
831 }
832 
833 /* when sending, force a reschedule every one of these periods */
834 #define SEND_RESCHED_TIMEOUT (5 * HZ)  /* 5s in jiffies */
835 
836 void _hfi1_do_send(struct work_struct *work)
837 {
838 	struct iowait *wait = container_of(work, struct iowait, iowork);
839 	struct rvt_qp *qp = iowait_to_qp(wait);
840 
841 	hfi1_do_send(qp);
842 }
843 
844 /**
845  * hfi1_do_send - perform a send on a QP
846  * @work: contains a pointer to the QP
847  *
848  * Process entries in the send work queue until credit or queue is
849  * exhausted.  Only allow one CPU to send a packet per QP (tasklet).
850  * Otherwise, two threads could send packets out of order.
851  */
852 void hfi1_do_send(struct rvt_qp *qp)
853 {
854 	struct hfi1_pkt_state ps;
855 	struct hfi1_qp_priv *priv = qp->priv;
856 	int (*make_req)(struct rvt_qp *qp, struct hfi1_pkt_state *ps);
857 	unsigned long timeout;
858 	unsigned long timeout_int;
859 	int cpu;
860 
861 	ps.dev = to_idev(qp->ibqp.device);
862 	ps.ibp = to_iport(qp->ibqp.device, qp->port_num);
863 	ps.ppd = ppd_from_ibp(ps.ibp);
864 
865 	switch (qp->ibqp.qp_type) {
866 	case IB_QPT_RC:
867 		if (!loopback && ((qp->remote_ah_attr.dlid & ~((1 << ps.ppd->lmc
868 								) - 1)) ==
869 				 ps.ppd->lid)) {
870 			ruc_loopback(qp);
871 			return;
872 		}
873 		make_req = hfi1_make_rc_req;
874 		timeout_int = (qp->timeout_jiffies);
875 		break;
876 	case IB_QPT_UC:
877 		if (!loopback && ((qp->remote_ah_attr.dlid & ~((1 << ps.ppd->lmc
878 								) - 1)) ==
879 				 ps.ppd->lid)) {
880 			ruc_loopback(qp);
881 			return;
882 		}
883 		make_req = hfi1_make_uc_req;
884 		timeout_int = SEND_RESCHED_TIMEOUT;
885 		break;
886 	default:
887 		make_req = hfi1_make_ud_req;
888 		timeout_int = SEND_RESCHED_TIMEOUT;
889 	}
890 
891 	spin_lock_irqsave(&qp->s_lock, ps.flags);
892 
893 	/* Return if we are already busy processing a work request. */
894 	if (!hfi1_send_ok(qp)) {
895 		spin_unlock_irqrestore(&qp->s_lock, ps.flags);
896 		return;
897 	}
898 
899 	qp->s_flags |= RVT_S_BUSY;
900 
901 	timeout = jiffies + (timeout_int) / 8;
902 	cpu = priv->s_sde ? priv->s_sde->cpu :
903 			cpumask_first(cpumask_of_node(ps.ppd->dd->node));
904 	/* insure a pre-built packet is handled  */
905 	ps.s_txreq = get_waiting_verbs_txreq(qp);
906 	do {
907 		/* Check for a constructed packet to be sent. */
908 		if (qp->s_hdrwords != 0) {
909 			spin_unlock_irqrestore(&qp->s_lock, ps.flags);
910 			/*
911 			 * If the packet cannot be sent now, return and
912 			 * the send tasklet will be woken up later.
913 			 */
914 			if (hfi1_verbs_send(qp, &ps))
915 				return;
916 			/* Record that s_ahg is empty. */
917 			qp->s_hdrwords = 0;
918 			/* allow other tasks to run */
919 			if (unlikely(time_after(jiffies, timeout))) {
920 				if (workqueue_congested(cpu,
921 							ps.ppd->hfi1_wq)) {
922 					spin_lock_irqsave(
923 						&qp->s_lock,
924 						ps.flags);
925 					qp->s_flags &= ~RVT_S_BUSY;
926 					hfi1_schedule_send(qp);
927 					spin_unlock_irqrestore(
928 						&qp->s_lock,
929 						ps.flags);
930 					this_cpu_inc(
931 						*ps.ppd->dd->send_schedule);
932 					return;
933 				}
934 				if (!irqs_disabled()) {
935 					cond_resched();
936 					this_cpu_inc(
937 					   *ps.ppd->dd->send_schedule);
938 				}
939 				timeout = jiffies + (timeout_int) / 8;
940 			}
941 			spin_lock_irqsave(&qp->s_lock, ps.flags);
942 		}
943 	} while (make_req(qp, &ps));
944 
945 	spin_unlock_irqrestore(&qp->s_lock, ps.flags);
946 }
947 
948 /*
949  * This should be called with s_lock held.
950  */
951 void hfi1_send_complete(struct rvt_qp *qp, struct rvt_swqe *wqe,
952 			enum ib_wc_status status)
953 {
954 	u32 old_last, last;
955 	unsigned i;
956 
957 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_OR_FLUSH_SEND))
958 		return;
959 
960 	last = qp->s_last;
961 	old_last = last;
962 	if (++last >= qp->s_size)
963 		last = 0;
964 	qp->s_last = last;
965 	/* See post_send() */
966 	barrier();
967 	for (i = 0; i < wqe->wr.num_sge; i++) {
968 		struct rvt_sge *sge = &wqe->sg_list[i];
969 
970 		rvt_put_mr(sge->mr);
971 	}
972 	if (qp->ibqp.qp_type == IB_QPT_UD ||
973 	    qp->ibqp.qp_type == IB_QPT_SMI ||
974 	    qp->ibqp.qp_type == IB_QPT_GSI)
975 		atomic_dec(&ibah_to_rvtah(wqe->ud_wr.ah)->refcount);
976 
977 	/* See ch. 11.2.4.1 and 10.7.3.1 */
978 	if (!(qp->s_flags & RVT_S_SIGNAL_REQ_WR) ||
979 	    (wqe->wr.send_flags & IB_SEND_SIGNALED) ||
980 	    status != IB_WC_SUCCESS) {
981 		struct ib_wc wc;
982 
983 		memset(&wc, 0, sizeof(wc));
984 		wc.wr_id = wqe->wr.wr_id;
985 		wc.status = status;
986 		wc.opcode = ib_hfi1_wc_opcode[wqe->wr.opcode];
987 		wc.qp = &qp->ibqp;
988 		if (status == IB_WC_SUCCESS)
989 			wc.byte_len = wqe->length;
990 		rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.send_cq), &wc,
991 			     status != IB_WC_SUCCESS);
992 	}
993 
994 	if (qp->s_acked == old_last)
995 		qp->s_acked = last;
996 	if (qp->s_cur == old_last)
997 		qp->s_cur = last;
998 	if (qp->s_tail == old_last)
999 		qp->s_tail = last;
1000 	if (qp->state == IB_QPS_SQD && last == qp->s_cur)
1001 		qp->s_draining = 0;
1002 }
1003