xref: /openbmc/linux/drivers/infiniband/hw/hfi1/qp.c (revision ecd25094)
1 /*
2  * Copyright(c) 2015 - 2019 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/err.h>
49 #include <linux/vmalloc.h>
50 #include <linux/hash.h>
51 #include <linux/module.h>
52 #include <linux/seq_file.h>
53 #include <rdma/rdma_vt.h>
54 #include <rdma/rdmavt_qp.h>
55 #include <rdma/ib_verbs.h>
56 
57 #include "hfi.h"
58 #include "qp.h"
59 #include "trace.h"
60 #include "verbs_txreq.h"
61 
62 unsigned int hfi1_qp_table_size = 256;
63 module_param_named(qp_table_size, hfi1_qp_table_size, uint, S_IRUGO);
64 MODULE_PARM_DESC(qp_table_size, "QP table size");
65 
66 static void flush_tx_list(struct rvt_qp *qp);
67 static int iowait_sleep(
68 	struct sdma_engine *sde,
69 	struct iowait_work *wait,
70 	struct sdma_txreq *stx,
71 	unsigned int seq,
72 	bool pkts_sent);
73 static void iowait_wakeup(struct iowait *wait, int reason);
74 static void iowait_sdma_drained(struct iowait *wait);
75 static void qp_pio_drain(struct rvt_qp *qp);
76 
77 const struct rvt_operation_params hfi1_post_parms[RVT_OPERATION_MAX] = {
78 [IB_WR_RDMA_WRITE] = {
79 	.length = sizeof(struct ib_rdma_wr),
80 	.qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
81 },
82 
83 [IB_WR_RDMA_READ] = {
84 	.length = sizeof(struct ib_rdma_wr),
85 	.qpt_support = BIT(IB_QPT_RC),
86 	.flags = RVT_OPERATION_ATOMIC,
87 },
88 
89 [IB_WR_ATOMIC_CMP_AND_SWP] = {
90 	.length = sizeof(struct ib_atomic_wr),
91 	.qpt_support = BIT(IB_QPT_RC),
92 	.flags = RVT_OPERATION_ATOMIC | RVT_OPERATION_ATOMIC_SGE,
93 },
94 
95 [IB_WR_ATOMIC_FETCH_AND_ADD] = {
96 	.length = sizeof(struct ib_atomic_wr),
97 	.qpt_support = BIT(IB_QPT_RC),
98 	.flags = RVT_OPERATION_ATOMIC | RVT_OPERATION_ATOMIC_SGE,
99 },
100 
101 [IB_WR_RDMA_WRITE_WITH_IMM] = {
102 	.length = sizeof(struct ib_rdma_wr),
103 	.qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
104 },
105 
106 [IB_WR_SEND] = {
107 	.length = sizeof(struct ib_send_wr),
108 	.qpt_support = BIT(IB_QPT_UD) | BIT(IB_QPT_SMI) | BIT(IB_QPT_GSI) |
109 		       BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
110 },
111 
112 [IB_WR_SEND_WITH_IMM] = {
113 	.length = sizeof(struct ib_send_wr),
114 	.qpt_support = BIT(IB_QPT_UD) | BIT(IB_QPT_SMI) | BIT(IB_QPT_GSI) |
115 		       BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
116 },
117 
118 [IB_WR_REG_MR] = {
119 	.length = sizeof(struct ib_reg_wr),
120 	.qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
121 	.flags = RVT_OPERATION_LOCAL,
122 },
123 
124 [IB_WR_LOCAL_INV] = {
125 	.length = sizeof(struct ib_send_wr),
126 	.qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
127 	.flags = RVT_OPERATION_LOCAL,
128 },
129 
130 [IB_WR_SEND_WITH_INV] = {
131 	.length = sizeof(struct ib_send_wr),
132 	.qpt_support = BIT(IB_QPT_RC),
133 },
134 
135 [IB_WR_OPFN] = {
136 	.length = sizeof(struct ib_atomic_wr),
137 	.qpt_support = BIT(IB_QPT_RC),
138 	.flags = RVT_OPERATION_USE_RESERVE,
139 },
140 
141 [IB_WR_TID_RDMA_WRITE] = {
142 	.length = sizeof(struct ib_rdma_wr),
143 	.qpt_support = BIT(IB_QPT_RC),
144 	.flags = RVT_OPERATION_IGN_RNR_CNT,
145 },
146 
147 };
148 
149 static void flush_list_head(struct list_head *l)
150 {
151 	while (!list_empty(l)) {
152 		struct sdma_txreq *tx;
153 
154 		tx = list_first_entry(
155 			l,
156 			struct sdma_txreq,
157 			list);
158 		list_del_init(&tx->list);
159 		hfi1_put_txreq(
160 			container_of(tx, struct verbs_txreq, txreq));
161 	}
162 }
163 
164 static void flush_tx_list(struct rvt_qp *qp)
165 {
166 	struct hfi1_qp_priv *priv = qp->priv;
167 
168 	flush_list_head(&iowait_get_ib_work(&priv->s_iowait)->tx_head);
169 	flush_list_head(&iowait_get_tid_work(&priv->s_iowait)->tx_head);
170 }
171 
172 static void flush_iowait(struct rvt_qp *qp)
173 {
174 	struct hfi1_qp_priv *priv = qp->priv;
175 	unsigned long flags;
176 	seqlock_t *lock = priv->s_iowait.lock;
177 
178 	if (!lock)
179 		return;
180 	write_seqlock_irqsave(lock, flags);
181 	if (!list_empty(&priv->s_iowait.list)) {
182 		list_del_init(&priv->s_iowait.list);
183 		priv->s_iowait.lock = NULL;
184 		rvt_put_qp(qp);
185 	}
186 	write_sequnlock_irqrestore(lock, flags);
187 }
188 
189 static inline int opa_mtu_enum_to_int(int mtu)
190 {
191 	switch (mtu) {
192 	case OPA_MTU_8192:  return 8192;
193 	case OPA_MTU_10240: return 10240;
194 	default:            return -1;
195 	}
196 }
197 
198 /**
199  * This function is what we would push to the core layer if we wanted to be a
200  * "first class citizen".  Instead we hide this here and rely on Verbs ULPs
201  * to blindly pass the MTU enum value from the PathRecord to us.
202  */
203 static inline int verbs_mtu_enum_to_int(struct ib_device *dev, enum ib_mtu mtu)
204 {
205 	int val;
206 
207 	/* Constraining 10KB packets to 8KB packets */
208 	if (mtu == (enum ib_mtu)OPA_MTU_10240)
209 		mtu = OPA_MTU_8192;
210 	val = opa_mtu_enum_to_int((int)mtu);
211 	if (val > 0)
212 		return val;
213 	return ib_mtu_enum_to_int(mtu);
214 }
215 
216 int hfi1_check_modify_qp(struct rvt_qp *qp, struct ib_qp_attr *attr,
217 			 int attr_mask, struct ib_udata *udata)
218 {
219 	struct ib_qp *ibqp = &qp->ibqp;
220 	struct hfi1_ibdev *dev = to_idev(ibqp->device);
221 	struct hfi1_devdata *dd = dd_from_dev(dev);
222 	u8 sc;
223 
224 	if (attr_mask & IB_QP_AV) {
225 		sc = ah_to_sc(ibqp->device, &attr->ah_attr);
226 		if (sc == 0xf)
227 			return -EINVAL;
228 
229 		if (!qp_to_sdma_engine(qp, sc) &&
230 		    dd->flags & HFI1_HAS_SEND_DMA)
231 			return -EINVAL;
232 
233 		if (!qp_to_send_context(qp, sc))
234 			return -EINVAL;
235 	}
236 
237 	if (attr_mask & IB_QP_ALT_PATH) {
238 		sc = ah_to_sc(ibqp->device, &attr->alt_ah_attr);
239 		if (sc == 0xf)
240 			return -EINVAL;
241 
242 		if (!qp_to_sdma_engine(qp, sc) &&
243 		    dd->flags & HFI1_HAS_SEND_DMA)
244 			return -EINVAL;
245 
246 		if (!qp_to_send_context(qp, sc))
247 			return -EINVAL;
248 	}
249 
250 	return 0;
251 }
252 
253 /*
254  * qp_set_16b - Set the hdr_type based on whether the slid or the
255  * dlid in the connection is extended. Only applicable for RC and UC
256  * QPs. UD QPs determine this on the fly from the ah in the wqe
257  */
258 static inline void qp_set_16b(struct rvt_qp *qp)
259 {
260 	struct hfi1_pportdata *ppd;
261 	struct hfi1_ibport *ibp;
262 	struct hfi1_qp_priv *priv = qp->priv;
263 
264 	/* Update ah_attr to account for extended LIDs */
265 	hfi1_update_ah_attr(qp->ibqp.device, &qp->remote_ah_attr);
266 
267 	/* Create 32 bit LIDs */
268 	hfi1_make_opa_lid(&qp->remote_ah_attr);
269 
270 	if (!(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH))
271 		return;
272 
273 	ibp = to_iport(qp->ibqp.device, qp->port_num);
274 	ppd = ppd_from_ibp(ibp);
275 	priv->hdr_type = hfi1_get_hdr_type(ppd->lid, &qp->remote_ah_attr);
276 }
277 
278 void hfi1_modify_qp(struct rvt_qp *qp, struct ib_qp_attr *attr,
279 		    int attr_mask, struct ib_udata *udata)
280 {
281 	struct ib_qp *ibqp = &qp->ibqp;
282 	struct hfi1_qp_priv *priv = qp->priv;
283 
284 	if (attr_mask & IB_QP_AV) {
285 		priv->s_sc = ah_to_sc(ibqp->device, &qp->remote_ah_attr);
286 		priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
287 		priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
288 		qp_set_16b(qp);
289 	}
290 
291 	if (attr_mask & IB_QP_PATH_MIG_STATE &&
292 	    attr->path_mig_state == IB_MIG_MIGRATED &&
293 	    qp->s_mig_state == IB_MIG_ARMED) {
294 		qp->s_flags |= HFI1_S_AHG_CLEAR;
295 		priv->s_sc = ah_to_sc(ibqp->device, &qp->remote_ah_attr);
296 		priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
297 		priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
298 		qp_set_16b(qp);
299 	}
300 
301 	opfn_qp_init(qp, attr, attr_mask);
302 }
303 
304 /**
305  * hfi1_setup_wqe - set up the wqe
306  * @qp - The qp
307  * @wqe - The built wqe
308  * @call_send - Determine if the send should be posted or scheduled.
309  *
310  * Perform setup of the wqe.  This is called
311  * prior to inserting the wqe into the ring but after
312  * the wqe has been setup by RDMAVT. This function
313  * allows the driver the opportunity to perform
314  * validation and additional setup of the wqe.
315  *
316  * Returns 0 on success, -EINVAL on failure
317  *
318  */
319 int hfi1_setup_wqe(struct rvt_qp *qp, struct rvt_swqe *wqe, bool *call_send)
320 {
321 	struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
322 	struct rvt_ah *ah;
323 	struct hfi1_pportdata *ppd;
324 	struct hfi1_devdata *dd;
325 
326 	switch (qp->ibqp.qp_type) {
327 	case IB_QPT_RC:
328 		hfi1_setup_tid_rdma_wqe(qp, wqe);
329 		/* fall through */
330 	case IB_QPT_UC:
331 		if (wqe->length > 0x80000000U)
332 			return -EINVAL;
333 		if (wqe->length > qp->pmtu)
334 			*call_send = false;
335 		break;
336 	case IB_QPT_SMI:
337 		/*
338 		 * SM packets should exclusively use VL15 and their SL is
339 		 * ignored (IBTA v1.3, Section 3.5.8.2). Therefore, when ah
340 		 * is created, SL is 0 in most cases and as a result some
341 		 * fields (vl and pmtu) in ah may not be set correctly,
342 		 * depending on the SL2SC and SC2VL tables at the time.
343 		 */
344 		ppd = ppd_from_ibp(ibp);
345 		dd = dd_from_ppd(ppd);
346 		if (wqe->length > dd->vld[15].mtu)
347 			return -EINVAL;
348 		break;
349 	case IB_QPT_GSI:
350 	case IB_QPT_UD:
351 		ah = rvt_get_swqe_ah(wqe);
352 		if (wqe->length > (1 << ah->log_pmtu))
353 			return -EINVAL;
354 		if (ibp->sl_to_sc[rdma_ah_get_sl(&ah->attr)] == 0xf)
355 			return -EINVAL;
356 	default:
357 		break;
358 	}
359 
360 	/*
361 	 * System latency between send and schedule is large enough that
362 	 * forcing call_send to true for piothreshold packets is necessary.
363 	 */
364 	if (wqe->length <= piothreshold)
365 		*call_send = true;
366 	return 0;
367 }
368 
369 /**
370  * _hfi1_schedule_send - schedule progress
371  * @qp: the QP
372  *
373  * This schedules qp progress w/o regard to the s_flags.
374  *
375  * It is only used in the post send, which doesn't hold
376  * the s_lock.
377  */
378 bool _hfi1_schedule_send(struct rvt_qp *qp)
379 {
380 	struct hfi1_qp_priv *priv = qp->priv;
381 	struct hfi1_ibport *ibp =
382 		to_iport(qp->ibqp.device, qp->port_num);
383 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
384 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
385 
386 	return iowait_schedule(&priv->s_iowait, ppd->hfi1_wq,
387 			       priv->s_sde ?
388 			       priv->s_sde->cpu :
389 			       cpumask_first(cpumask_of_node(dd->node)));
390 }
391 
392 static void qp_pio_drain(struct rvt_qp *qp)
393 {
394 	struct hfi1_qp_priv *priv = qp->priv;
395 
396 	if (!priv->s_sendcontext)
397 		return;
398 	while (iowait_pio_pending(&priv->s_iowait)) {
399 		write_seqlock_irq(&priv->s_sendcontext->waitlock);
400 		hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 1);
401 		write_sequnlock_irq(&priv->s_sendcontext->waitlock);
402 		iowait_pio_drain(&priv->s_iowait);
403 		write_seqlock_irq(&priv->s_sendcontext->waitlock);
404 		hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 0);
405 		write_sequnlock_irq(&priv->s_sendcontext->waitlock);
406 	}
407 }
408 
409 /**
410  * hfi1_schedule_send - schedule progress
411  * @qp: the QP
412  *
413  * This schedules qp progress and caller should hold
414  * the s_lock.
415  * @return true if the first leg is scheduled;
416  * false if the first leg is not scheduled.
417  */
418 bool hfi1_schedule_send(struct rvt_qp *qp)
419 {
420 	lockdep_assert_held(&qp->s_lock);
421 	if (hfi1_send_ok(qp)) {
422 		_hfi1_schedule_send(qp);
423 		return true;
424 	}
425 	if (qp->s_flags & HFI1_S_ANY_WAIT_IO)
426 		iowait_set_flag(&((struct hfi1_qp_priv *)qp->priv)->s_iowait,
427 				IOWAIT_PENDING_IB);
428 	return false;
429 }
430 
431 static void hfi1_qp_schedule(struct rvt_qp *qp)
432 {
433 	struct hfi1_qp_priv *priv = qp->priv;
434 	bool ret;
435 
436 	if (iowait_flag_set(&priv->s_iowait, IOWAIT_PENDING_IB)) {
437 		ret = hfi1_schedule_send(qp);
438 		if (ret)
439 			iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_IB);
440 	}
441 	if (iowait_flag_set(&priv->s_iowait, IOWAIT_PENDING_TID)) {
442 		ret = hfi1_schedule_tid_send(qp);
443 		if (ret)
444 			iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_TID);
445 	}
446 }
447 
448 void hfi1_qp_wakeup(struct rvt_qp *qp, u32 flag)
449 {
450 	unsigned long flags;
451 
452 	spin_lock_irqsave(&qp->s_lock, flags);
453 	if (qp->s_flags & flag) {
454 		qp->s_flags &= ~flag;
455 		trace_hfi1_qpwakeup(qp, flag);
456 		hfi1_qp_schedule(qp);
457 	}
458 	spin_unlock_irqrestore(&qp->s_lock, flags);
459 	/* Notify hfi1_destroy_qp() if it is waiting. */
460 	rvt_put_qp(qp);
461 }
462 
463 void hfi1_qp_unbusy(struct rvt_qp *qp, struct iowait_work *wait)
464 {
465 	struct hfi1_qp_priv *priv = qp->priv;
466 
467 	if (iowait_set_work_flag(wait) == IOWAIT_IB_SE) {
468 		qp->s_flags &= ~RVT_S_BUSY;
469 		/*
470 		 * If we are sending a first-leg packet from the second leg,
471 		 * we need to clear the busy flag from priv->s_flags to
472 		 * avoid a race condition when the qp wakes up before
473 		 * the call to hfi1_verbs_send() returns to the second
474 		 * leg. In that case, the second leg will terminate without
475 		 * being re-scheduled, resulting in failure to send TID RDMA
476 		 * WRITE DATA and TID RDMA ACK packets.
477 		 */
478 		if (priv->s_flags & HFI1_S_TID_BUSY_SET) {
479 			priv->s_flags &= ~(HFI1_S_TID_BUSY_SET |
480 					   RVT_S_BUSY);
481 			iowait_set_flag(&priv->s_iowait, IOWAIT_PENDING_TID);
482 		}
483 	} else {
484 		priv->s_flags &= ~RVT_S_BUSY;
485 	}
486 }
487 
488 static int iowait_sleep(
489 	struct sdma_engine *sde,
490 	struct iowait_work *wait,
491 	struct sdma_txreq *stx,
492 	uint seq,
493 	bool pkts_sent)
494 {
495 	struct verbs_txreq *tx = container_of(stx, struct verbs_txreq, txreq);
496 	struct rvt_qp *qp;
497 	struct hfi1_qp_priv *priv;
498 	unsigned long flags;
499 	int ret = 0;
500 
501 	qp = tx->qp;
502 	priv = qp->priv;
503 
504 	spin_lock_irqsave(&qp->s_lock, flags);
505 	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
506 		/*
507 		 * If we couldn't queue the DMA request, save the info
508 		 * and try again later rather than destroying the
509 		 * buffer and undoing the side effects of the copy.
510 		 */
511 		/* Make a common routine? */
512 		list_add_tail(&stx->list, &wait->tx_head);
513 		write_seqlock(&sde->waitlock);
514 		if (sdma_progress(sde, seq, stx))
515 			goto eagain;
516 		if (list_empty(&priv->s_iowait.list)) {
517 			struct hfi1_ibport *ibp =
518 				to_iport(qp->ibqp.device, qp->port_num);
519 
520 			ibp->rvp.n_dmawait++;
521 			qp->s_flags |= RVT_S_WAIT_DMA_DESC;
522 			iowait_get_priority(&priv->s_iowait);
523 			iowait_queue(pkts_sent, &priv->s_iowait,
524 				     &sde->dmawait);
525 			priv->s_iowait.lock = &sde->waitlock;
526 			trace_hfi1_qpsleep(qp, RVT_S_WAIT_DMA_DESC);
527 			rvt_get_qp(qp);
528 		}
529 		write_sequnlock(&sde->waitlock);
530 		hfi1_qp_unbusy(qp, wait);
531 		spin_unlock_irqrestore(&qp->s_lock, flags);
532 		ret = -EBUSY;
533 	} else {
534 		spin_unlock_irqrestore(&qp->s_lock, flags);
535 		hfi1_put_txreq(tx);
536 	}
537 	return ret;
538 eagain:
539 	write_sequnlock(&sde->waitlock);
540 	spin_unlock_irqrestore(&qp->s_lock, flags);
541 	list_del_init(&stx->list);
542 	return -EAGAIN;
543 }
544 
545 static void iowait_wakeup(struct iowait *wait, int reason)
546 {
547 	struct rvt_qp *qp = iowait_to_qp(wait);
548 
549 	WARN_ON(reason != SDMA_AVAIL_REASON);
550 	hfi1_qp_wakeup(qp, RVT_S_WAIT_DMA_DESC);
551 }
552 
553 static void iowait_sdma_drained(struct iowait *wait)
554 {
555 	struct rvt_qp *qp = iowait_to_qp(wait);
556 	unsigned long flags;
557 
558 	/*
559 	 * This happens when the send engine notes
560 	 * a QP in the error state and cannot
561 	 * do the flush work until that QP's
562 	 * sdma work has finished.
563 	 */
564 	spin_lock_irqsave(&qp->s_lock, flags);
565 	if (qp->s_flags & RVT_S_WAIT_DMA) {
566 		qp->s_flags &= ~RVT_S_WAIT_DMA;
567 		hfi1_schedule_send(qp);
568 	}
569 	spin_unlock_irqrestore(&qp->s_lock, flags);
570 }
571 
572 static void hfi1_init_priority(struct iowait *w)
573 {
574 	struct rvt_qp *qp = iowait_to_qp(w);
575 	struct hfi1_qp_priv *priv = qp->priv;
576 
577 	if (qp->s_flags & RVT_S_ACK_PENDING)
578 		w->priority++;
579 	if (priv->s_flags & RVT_S_ACK_PENDING)
580 		w->priority++;
581 }
582 
583 /**
584  * qp_to_sdma_engine - map a qp to a send engine
585  * @qp: the QP
586  * @sc5: the 5 bit sc
587  *
588  * Return:
589  * A send engine for the qp or NULL for SMI type qp.
590  */
591 struct sdma_engine *qp_to_sdma_engine(struct rvt_qp *qp, u8 sc5)
592 {
593 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
594 	struct sdma_engine *sde;
595 
596 	if (!(dd->flags & HFI1_HAS_SEND_DMA))
597 		return NULL;
598 	switch (qp->ibqp.qp_type) {
599 	case IB_QPT_SMI:
600 		return NULL;
601 	default:
602 		break;
603 	}
604 	sde = sdma_select_engine_sc(dd, qp->ibqp.qp_num >> dd->qos_shift, sc5);
605 	return sde;
606 }
607 
608 /*
609  * qp_to_send_context - map a qp to a send context
610  * @qp: the QP
611  * @sc5: the 5 bit sc
612  *
613  * Return:
614  * A send context for the qp
615  */
616 struct send_context *qp_to_send_context(struct rvt_qp *qp, u8 sc5)
617 {
618 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
619 
620 	switch (qp->ibqp.qp_type) {
621 	case IB_QPT_SMI:
622 		/* SMA packets to VL15 */
623 		return dd->vld[15].sc;
624 	default:
625 		break;
626 	}
627 
628 	return pio_select_send_context_sc(dd, qp->ibqp.qp_num >> dd->qos_shift,
629 					  sc5);
630 }
631 
632 static const char * const qp_type_str[] = {
633 	"SMI", "GSI", "RC", "UC", "UD",
634 };
635 
636 static int qp_idle(struct rvt_qp *qp)
637 {
638 	return
639 		qp->s_last == qp->s_acked &&
640 		qp->s_acked == qp->s_cur &&
641 		qp->s_cur == qp->s_tail &&
642 		qp->s_tail == qp->s_head;
643 }
644 
645 /**
646  * qp_iter_print - print the qp information to seq_file
647  * @s: the seq_file to emit the qp information on
648  * @iter: the iterator for the qp hash list
649  */
650 void qp_iter_print(struct seq_file *s, struct rvt_qp_iter *iter)
651 {
652 	struct rvt_swqe *wqe;
653 	struct rvt_qp *qp = iter->qp;
654 	struct hfi1_qp_priv *priv = qp->priv;
655 	struct sdma_engine *sde;
656 	struct send_context *send_context;
657 	struct rvt_ack_entry *e = NULL;
658 	struct rvt_srq *srq = qp->ibqp.srq ?
659 		ibsrq_to_rvtsrq(qp->ibqp.srq) : NULL;
660 
661 	sde = qp_to_sdma_engine(qp, priv->s_sc);
662 	wqe = rvt_get_swqe_ptr(qp, qp->s_last);
663 	send_context = qp_to_send_context(qp, priv->s_sc);
664 	if (qp->s_ack_queue)
665 		e = &qp->s_ack_queue[qp->s_tail_ack_queue];
666 	seq_printf(s,
667 		   "N %d %s QP %x R %u %s %u %u f=%x %u %u %u %u %u %u SPSN %x %x %x %x %x RPSN %x S(%u %u %u %u %u %u %u) R(%u %u %u) RQP %x LID %x SL %u MTU %u %u %u %u %u SDE %p,%u SC %p,%u SCQ %u %u PID %d OS %x %x E %x %x %x RNR %d %s %d\n",
668 		   iter->n,
669 		   qp_idle(qp) ? "I" : "B",
670 		   qp->ibqp.qp_num,
671 		   atomic_read(&qp->refcount),
672 		   qp_type_str[qp->ibqp.qp_type],
673 		   qp->state,
674 		   wqe ? wqe->wr.opcode : 0,
675 		   qp->s_flags,
676 		   iowait_sdma_pending(&priv->s_iowait),
677 		   iowait_pio_pending(&priv->s_iowait),
678 		   !list_empty(&priv->s_iowait.list),
679 		   qp->timeout,
680 		   wqe ? wqe->ssn : 0,
681 		   qp->s_lsn,
682 		   qp->s_last_psn,
683 		   qp->s_psn, qp->s_next_psn,
684 		   qp->s_sending_psn, qp->s_sending_hpsn,
685 		   qp->r_psn,
686 		   qp->s_last, qp->s_acked, qp->s_cur,
687 		   qp->s_tail, qp->s_head, qp->s_size,
688 		   qp->s_avail,
689 		   /* ack_queue ring pointers, size */
690 		   qp->s_tail_ack_queue, qp->r_head_ack_queue,
691 		   rvt_max_atomic(&to_idev(qp->ibqp.device)->rdi),
692 		   /* remote QP info  */
693 		   qp->remote_qpn,
694 		   rdma_ah_get_dlid(&qp->remote_ah_attr),
695 		   rdma_ah_get_sl(&qp->remote_ah_attr),
696 		   qp->pmtu,
697 		   qp->s_retry,
698 		   qp->s_retry_cnt,
699 		   qp->s_rnr_retry_cnt,
700 		   qp->s_rnr_retry,
701 		   sde,
702 		   sde ? sde->this_idx : 0,
703 		   send_context,
704 		   send_context ? send_context->sw_index : 0,
705 		   ib_cq_head(qp->ibqp.send_cq),
706 		   ib_cq_tail(qp->ibqp.send_cq),
707 		   qp->pid,
708 		   qp->s_state,
709 		   qp->s_ack_state,
710 		   /* ack queue information */
711 		   e ? e->opcode : 0,
712 		   e ? e->psn : 0,
713 		   e ? e->lpsn : 0,
714 		   qp->r_min_rnr_timer,
715 		   srq ? "SRQ" : "RQ",
716 		   srq ? srq->rq.size : qp->r_rq.size
717 		);
718 }
719 
720 void *qp_priv_alloc(struct rvt_dev_info *rdi, struct rvt_qp *qp)
721 {
722 	struct hfi1_qp_priv *priv;
723 
724 	priv = kzalloc_node(sizeof(*priv), GFP_KERNEL, rdi->dparms.node);
725 	if (!priv)
726 		return ERR_PTR(-ENOMEM);
727 
728 	priv->owner = qp;
729 
730 	priv->s_ahg = kzalloc_node(sizeof(*priv->s_ahg), GFP_KERNEL,
731 				   rdi->dparms.node);
732 	if (!priv->s_ahg) {
733 		kfree(priv);
734 		return ERR_PTR(-ENOMEM);
735 	}
736 	iowait_init(
737 		&priv->s_iowait,
738 		1,
739 		_hfi1_do_send,
740 		_hfi1_do_tid_send,
741 		iowait_sleep,
742 		iowait_wakeup,
743 		iowait_sdma_drained,
744 		hfi1_init_priority);
745 	/* Init to a value to start the running average correctly */
746 	priv->s_running_pkt_size = piothreshold / 2;
747 	return priv;
748 }
749 
750 void qp_priv_free(struct rvt_dev_info *rdi, struct rvt_qp *qp)
751 {
752 	struct hfi1_qp_priv *priv = qp->priv;
753 
754 	hfi1_qp_priv_tid_free(rdi, qp);
755 	kfree(priv->s_ahg);
756 	kfree(priv);
757 }
758 
759 unsigned free_all_qps(struct rvt_dev_info *rdi)
760 {
761 	struct hfi1_ibdev *verbs_dev = container_of(rdi,
762 						    struct hfi1_ibdev,
763 						    rdi);
764 	struct hfi1_devdata *dd = container_of(verbs_dev,
765 					       struct hfi1_devdata,
766 					       verbs_dev);
767 	int n;
768 	unsigned qp_inuse = 0;
769 
770 	for (n = 0; n < dd->num_pports; n++) {
771 		struct hfi1_ibport *ibp = &dd->pport[n].ibport_data;
772 
773 		rcu_read_lock();
774 		if (rcu_dereference(ibp->rvp.qp[0]))
775 			qp_inuse++;
776 		if (rcu_dereference(ibp->rvp.qp[1]))
777 			qp_inuse++;
778 		rcu_read_unlock();
779 	}
780 
781 	return qp_inuse;
782 }
783 
784 void flush_qp_waiters(struct rvt_qp *qp)
785 {
786 	lockdep_assert_held(&qp->s_lock);
787 	flush_iowait(qp);
788 	hfi1_tid_rdma_flush_wait(qp);
789 }
790 
791 void stop_send_queue(struct rvt_qp *qp)
792 {
793 	struct hfi1_qp_priv *priv = qp->priv;
794 
795 	iowait_cancel_work(&priv->s_iowait);
796 	if (cancel_work_sync(&priv->tid_rdma.trigger_work))
797 		rvt_put_qp(qp);
798 }
799 
800 void quiesce_qp(struct rvt_qp *qp)
801 {
802 	struct hfi1_qp_priv *priv = qp->priv;
803 
804 	hfi1_del_tid_reap_timer(qp);
805 	hfi1_del_tid_retry_timer(qp);
806 	iowait_sdma_drain(&priv->s_iowait);
807 	qp_pio_drain(qp);
808 	flush_tx_list(qp);
809 }
810 
811 void notify_qp_reset(struct rvt_qp *qp)
812 {
813 	hfi1_qp_kern_exp_rcv_clear_all(qp);
814 	qp->r_adefered = 0;
815 	clear_ahg(qp);
816 
817 	/* Clear any OPFN state */
818 	if (qp->ibqp.qp_type == IB_QPT_RC)
819 		opfn_conn_error(qp);
820 }
821 
822 /*
823  * Switch to alternate path.
824  * The QP s_lock should be held and interrupts disabled.
825  */
826 void hfi1_migrate_qp(struct rvt_qp *qp)
827 {
828 	struct hfi1_qp_priv *priv = qp->priv;
829 	struct ib_event ev;
830 
831 	qp->s_mig_state = IB_MIG_MIGRATED;
832 	qp->remote_ah_attr = qp->alt_ah_attr;
833 	qp->port_num = rdma_ah_get_port_num(&qp->alt_ah_attr);
834 	qp->s_pkey_index = qp->s_alt_pkey_index;
835 	qp->s_flags |= HFI1_S_AHG_CLEAR;
836 	priv->s_sc = ah_to_sc(qp->ibqp.device, &qp->remote_ah_attr);
837 	priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
838 	qp_set_16b(qp);
839 
840 	ev.device = qp->ibqp.device;
841 	ev.element.qp = &qp->ibqp;
842 	ev.event = IB_EVENT_PATH_MIG;
843 	qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
844 }
845 
846 int mtu_to_path_mtu(u32 mtu)
847 {
848 	return mtu_to_enum(mtu, OPA_MTU_8192);
849 }
850 
851 u32 mtu_from_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp, u32 pmtu)
852 {
853 	u32 mtu;
854 	struct hfi1_ibdev *verbs_dev = container_of(rdi,
855 						    struct hfi1_ibdev,
856 						    rdi);
857 	struct hfi1_devdata *dd = container_of(verbs_dev,
858 					       struct hfi1_devdata,
859 					       verbs_dev);
860 	struct hfi1_ibport *ibp;
861 	u8 sc, vl;
862 
863 	ibp = &dd->pport[qp->port_num - 1].ibport_data;
864 	sc = ibp->sl_to_sc[rdma_ah_get_sl(&qp->remote_ah_attr)];
865 	vl = sc_to_vlt(dd, sc);
866 
867 	mtu = verbs_mtu_enum_to_int(qp->ibqp.device, pmtu);
868 	if (vl < PER_VL_SEND_CONTEXTS)
869 		mtu = min_t(u32, mtu, dd->vld[vl].mtu);
870 	return mtu;
871 }
872 
873 int get_pmtu_from_attr(struct rvt_dev_info *rdi, struct rvt_qp *qp,
874 		       struct ib_qp_attr *attr)
875 {
876 	int mtu, pidx = qp->port_num - 1;
877 	struct hfi1_ibdev *verbs_dev = container_of(rdi,
878 						    struct hfi1_ibdev,
879 						    rdi);
880 	struct hfi1_devdata *dd = container_of(verbs_dev,
881 					       struct hfi1_devdata,
882 					       verbs_dev);
883 	mtu = verbs_mtu_enum_to_int(qp->ibqp.device, attr->path_mtu);
884 	if (mtu == -1)
885 		return -1; /* values less than 0 are error */
886 
887 	if (mtu > dd->pport[pidx].ibmtu)
888 		return mtu_to_enum(dd->pport[pidx].ibmtu, IB_MTU_2048);
889 	else
890 		return attr->path_mtu;
891 }
892 
893 void notify_error_qp(struct rvt_qp *qp)
894 {
895 	struct hfi1_qp_priv *priv = qp->priv;
896 	seqlock_t *lock = priv->s_iowait.lock;
897 
898 	if (lock) {
899 		write_seqlock(lock);
900 		if (!list_empty(&priv->s_iowait.list) &&
901 		    !(qp->s_flags & RVT_S_BUSY) &&
902 		    !(priv->s_flags & RVT_S_BUSY)) {
903 			qp->s_flags &= ~HFI1_S_ANY_WAIT_IO;
904 			iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_IB);
905 			iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_TID);
906 			list_del_init(&priv->s_iowait.list);
907 			priv->s_iowait.lock = NULL;
908 			rvt_put_qp(qp);
909 		}
910 		write_sequnlock(lock);
911 	}
912 
913 	if (!(qp->s_flags & RVT_S_BUSY) && !(priv->s_flags & RVT_S_BUSY)) {
914 		qp->s_hdrwords = 0;
915 		if (qp->s_rdma_mr) {
916 			rvt_put_mr(qp->s_rdma_mr);
917 			qp->s_rdma_mr = NULL;
918 		}
919 		flush_tx_list(qp);
920 	}
921 }
922 
923 /**
924  * hfi1_qp_iter_cb - callback for iterator
925  * @qp - the qp
926  * @v - the sl in low bits of v
927  *
928  * This is called from the iterator callback to work
929  * on an individual qp.
930  */
931 static void hfi1_qp_iter_cb(struct rvt_qp *qp, u64 v)
932 {
933 	int lastwqe;
934 	struct ib_event ev;
935 	struct hfi1_ibport *ibp =
936 		to_iport(qp->ibqp.device, qp->port_num);
937 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
938 	u8 sl = (u8)v;
939 
940 	if (qp->port_num != ppd->port ||
941 	    (qp->ibqp.qp_type != IB_QPT_UC &&
942 	     qp->ibqp.qp_type != IB_QPT_RC) ||
943 	    rdma_ah_get_sl(&qp->remote_ah_attr) != sl ||
944 	    !(ib_rvt_state_ops[qp->state] & RVT_POST_SEND_OK))
945 		return;
946 
947 	spin_lock_irq(&qp->r_lock);
948 	spin_lock(&qp->s_hlock);
949 	spin_lock(&qp->s_lock);
950 	lastwqe = rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
951 	spin_unlock(&qp->s_lock);
952 	spin_unlock(&qp->s_hlock);
953 	spin_unlock_irq(&qp->r_lock);
954 	if (lastwqe) {
955 		ev.device = qp->ibqp.device;
956 		ev.element.qp = &qp->ibqp;
957 		ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
958 		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
959 	}
960 }
961 
962 /**
963  * hfi1_error_port_qps - put a port's RC/UC qps into error state
964  * @ibp: the ibport.
965  * @sl: the service level.
966  *
967  * This function places all RC/UC qps with a given service level into error
968  * state. It is generally called to force upper lay apps to abandon stale qps
969  * after an sl->sc mapping change.
970  */
971 void hfi1_error_port_qps(struct hfi1_ibport *ibp, u8 sl)
972 {
973 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
974 	struct hfi1_ibdev *dev = &ppd->dd->verbs_dev;
975 
976 	rvt_qp_iter(&dev->rdi, sl, hfi1_qp_iter_cb);
977 }
978