xref: /openbmc/linux/drivers/infiniband/hw/hfi1/qp.c (revision e2c75e76)
1 /*
2  * Copyright(c) 2015 - 2017 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 *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 };
136 
137 static void flush_tx_list(struct rvt_qp *qp)
138 {
139 	struct hfi1_qp_priv *priv = qp->priv;
140 
141 	while (!list_empty(&priv->s_iowait.tx_head)) {
142 		struct sdma_txreq *tx;
143 
144 		tx = list_first_entry(
145 			&priv->s_iowait.tx_head,
146 			struct sdma_txreq,
147 			list);
148 		list_del_init(&tx->list);
149 		hfi1_put_txreq(
150 			container_of(tx, struct verbs_txreq, txreq));
151 	}
152 }
153 
154 static void flush_iowait(struct rvt_qp *qp)
155 {
156 	struct hfi1_qp_priv *priv = qp->priv;
157 	unsigned long flags;
158 	seqlock_t *lock = priv->s_iowait.lock;
159 
160 	if (!lock)
161 		return;
162 	write_seqlock_irqsave(lock, flags);
163 	if (!list_empty(&priv->s_iowait.list)) {
164 		list_del_init(&priv->s_iowait.list);
165 		priv->s_iowait.lock = NULL;
166 		rvt_put_qp(qp);
167 	}
168 	write_sequnlock_irqrestore(lock, flags);
169 }
170 
171 static inline int opa_mtu_enum_to_int(int mtu)
172 {
173 	switch (mtu) {
174 	case OPA_MTU_8192:  return 8192;
175 	case OPA_MTU_10240: return 10240;
176 	default:            return -1;
177 	}
178 }
179 
180 /**
181  * This function is what we would push to the core layer if we wanted to be a
182  * "first class citizen".  Instead we hide this here and rely on Verbs ULPs
183  * to blindly pass the MTU enum value from the PathRecord to us.
184  */
185 static inline int verbs_mtu_enum_to_int(struct ib_device *dev, enum ib_mtu mtu)
186 {
187 	int val;
188 
189 	/* Constraining 10KB packets to 8KB packets */
190 	if (mtu == (enum ib_mtu)OPA_MTU_10240)
191 		mtu = OPA_MTU_8192;
192 	val = opa_mtu_enum_to_int((int)mtu);
193 	if (val > 0)
194 		return val;
195 	return ib_mtu_enum_to_int(mtu);
196 }
197 
198 int hfi1_check_modify_qp(struct rvt_qp *qp, struct ib_qp_attr *attr,
199 			 int attr_mask, struct ib_udata *udata)
200 {
201 	struct ib_qp *ibqp = &qp->ibqp;
202 	struct hfi1_ibdev *dev = to_idev(ibqp->device);
203 	struct hfi1_devdata *dd = dd_from_dev(dev);
204 	u8 sc;
205 
206 	if (attr_mask & IB_QP_AV) {
207 		sc = ah_to_sc(ibqp->device, &attr->ah_attr);
208 		if (sc == 0xf)
209 			return -EINVAL;
210 
211 		if (!qp_to_sdma_engine(qp, sc) &&
212 		    dd->flags & HFI1_HAS_SEND_DMA)
213 			return -EINVAL;
214 
215 		if (!qp_to_send_context(qp, sc))
216 			return -EINVAL;
217 	}
218 
219 	if (attr_mask & IB_QP_ALT_PATH) {
220 		sc = ah_to_sc(ibqp->device, &attr->alt_ah_attr);
221 		if (sc == 0xf)
222 			return -EINVAL;
223 
224 		if (!qp_to_sdma_engine(qp, sc) &&
225 		    dd->flags & HFI1_HAS_SEND_DMA)
226 			return -EINVAL;
227 
228 		if (!qp_to_send_context(qp, sc))
229 			return -EINVAL;
230 	}
231 
232 	return 0;
233 }
234 
235 /*
236  * qp_set_16b - Set the hdr_type based on whether the slid or the
237  * dlid in the connection is extended. Only applicable for RC and UC
238  * QPs. UD QPs determine this on the fly from the ah in the wqe
239  */
240 static inline void qp_set_16b(struct rvt_qp *qp)
241 {
242 	struct hfi1_pportdata *ppd;
243 	struct hfi1_ibport *ibp;
244 	struct hfi1_qp_priv *priv = qp->priv;
245 
246 	/* Update ah_attr to account for extended LIDs */
247 	hfi1_update_ah_attr(qp->ibqp.device, &qp->remote_ah_attr);
248 
249 	/* Create 32 bit LIDs */
250 	hfi1_make_opa_lid(&qp->remote_ah_attr);
251 
252 	if (!(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH))
253 		return;
254 
255 	ibp = to_iport(qp->ibqp.device, qp->port_num);
256 	ppd = ppd_from_ibp(ibp);
257 	priv->hdr_type = hfi1_get_hdr_type(ppd->lid, &qp->remote_ah_attr);
258 }
259 
260 void hfi1_modify_qp(struct rvt_qp *qp, struct ib_qp_attr *attr,
261 		    int attr_mask, struct ib_udata *udata)
262 {
263 	struct ib_qp *ibqp = &qp->ibqp;
264 	struct hfi1_qp_priv *priv = qp->priv;
265 
266 	if (attr_mask & IB_QP_AV) {
267 		priv->s_sc = ah_to_sc(ibqp->device, &qp->remote_ah_attr);
268 		priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
269 		priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
270 		qp_set_16b(qp);
271 	}
272 
273 	if (attr_mask & IB_QP_PATH_MIG_STATE &&
274 	    attr->path_mig_state == IB_MIG_MIGRATED &&
275 	    qp->s_mig_state == IB_MIG_ARMED) {
276 		qp->s_flags |= RVT_S_AHG_CLEAR;
277 		priv->s_sc = ah_to_sc(ibqp->device, &qp->remote_ah_attr);
278 		priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
279 		priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
280 		qp_set_16b(qp);
281 	}
282 }
283 
284 /**
285  * hfi1_check_send_wqe - validate wqe
286  * @qp - The qp
287  * @wqe - The built wqe
288  *
289  * validate wqe.  This is called
290  * prior to inserting the wqe into
291  * the ring but after the wqe has been
292  * setup.
293  *
294  * Returns 0 on success, -EINVAL on failure
295  *
296  */
297 int hfi1_check_send_wqe(struct rvt_qp *qp,
298 			struct rvt_swqe *wqe)
299 {
300 	struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
301 	struct rvt_ah *ah;
302 
303 	switch (qp->ibqp.qp_type) {
304 	case IB_QPT_RC:
305 	case IB_QPT_UC:
306 		if (wqe->length > 0x80000000U)
307 			return -EINVAL;
308 		break;
309 	case IB_QPT_SMI:
310 		ah = ibah_to_rvtah(wqe->ud_wr.ah);
311 		if (wqe->length > (1 << ah->log_pmtu))
312 			return -EINVAL;
313 		break;
314 	case IB_QPT_GSI:
315 	case IB_QPT_UD:
316 		ah = ibah_to_rvtah(wqe->ud_wr.ah);
317 		if (wqe->length > (1 << ah->log_pmtu))
318 			return -EINVAL;
319 		if (ibp->sl_to_sc[rdma_ah_get_sl(&ah->attr)] == 0xf)
320 			return -EINVAL;
321 	default:
322 		break;
323 	}
324 	return wqe->length <= piothreshold;
325 }
326 
327 /**
328  * _hfi1_schedule_send - schedule progress
329  * @qp: the QP
330  *
331  * This schedules qp progress w/o regard to the s_flags.
332  *
333  * It is only used in the post send, which doesn't hold
334  * the s_lock.
335  */
336 void _hfi1_schedule_send(struct rvt_qp *qp)
337 {
338 	struct hfi1_qp_priv *priv = qp->priv;
339 	struct hfi1_ibport *ibp =
340 		to_iport(qp->ibqp.device, qp->port_num);
341 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
342 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
343 
344 	iowait_schedule(&priv->s_iowait, ppd->hfi1_wq,
345 			priv->s_sde ?
346 			priv->s_sde->cpu :
347 			cpumask_first(cpumask_of_node(dd->node)));
348 }
349 
350 static void qp_pio_drain(struct rvt_qp *qp)
351 {
352 	struct hfi1_ibdev *dev;
353 	struct hfi1_qp_priv *priv = qp->priv;
354 
355 	if (!priv->s_sendcontext)
356 		return;
357 	dev = to_idev(qp->ibqp.device);
358 	while (iowait_pio_pending(&priv->s_iowait)) {
359 		write_seqlock_irq(&dev->iowait_lock);
360 		hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 1);
361 		write_sequnlock_irq(&dev->iowait_lock);
362 		iowait_pio_drain(&priv->s_iowait);
363 		write_seqlock_irq(&dev->iowait_lock);
364 		hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 0);
365 		write_sequnlock_irq(&dev->iowait_lock);
366 	}
367 }
368 
369 /**
370  * hfi1_schedule_send - schedule progress
371  * @qp: the QP
372  *
373  * This schedules qp progress and caller should hold
374  * the s_lock.
375  */
376 void hfi1_schedule_send(struct rvt_qp *qp)
377 {
378 	lockdep_assert_held(&qp->s_lock);
379 	if (hfi1_send_ok(qp))
380 		_hfi1_schedule_send(qp);
381 }
382 
383 void hfi1_qp_wakeup(struct rvt_qp *qp, u32 flag)
384 {
385 	unsigned long flags;
386 
387 	spin_lock_irqsave(&qp->s_lock, flags);
388 	if (qp->s_flags & flag) {
389 		qp->s_flags &= ~flag;
390 		trace_hfi1_qpwakeup(qp, flag);
391 		hfi1_schedule_send(qp);
392 	}
393 	spin_unlock_irqrestore(&qp->s_lock, flags);
394 	/* Notify hfi1_destroy_qp() if it is waiting. */
395 	rvt_put_qp(qp);
396 }
397 
398 static int iowait_sleep(
399 	struct sdma_engine *sde,
400 	struct iowait *wait,
401 	struct sdma_txreq *stx,
402 	uint seq,
403 	bool pkts_sent)
404 {
405 	struct verbs_txreq *tx = container_of(stx, struct verbs_txreq, txreq);
406 	struct rvt_qp *qp;
407 	struct hfi1_qp_priv *priv;
408 	unsigned long flags;
409 	int ret = 0;
410 	struct hfi1_ibdev *dev;
411 
412 	qp = tx->qp;
413 	priv = qp->priv;
414 
415 	spin_lock_irqsave(&qp->s_lock, flags);
416 	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
417 		/*
418 		 * If we couldn't queue the DMA request, save the info
419 		 * and try again later rather than destroying the
420 		 * buffer and undoing the side effects of the copy.
421 		 */
422 		/* Make a common routine? */
423 		dev = &sde->dd->verbs_dev;
424 		list_add_tail(&stx->list, &wait->tx_head);
425 		write_seqlock(&dev->iowait_lock);
426 		if (sdma_progress(sde, seq, stx))
427 			goto eagain;
428 		if (list_empty(&priv->s_iowait.list)) {
429 			struct hfi1_ibport *ibp =
430 				to_iport(qp->ibqp.device, qp->port_num);
431 
432 			ibp->rvp.n_dmawait++;
433 			qp->s_flags |= RVT_S_WAIT_DMA_DESC;
434 			iowait_queue(pkts_sent, &priv->s_iowait,
435 				     &sde->dmawait);
436 			priv->s_iowait.lock = &dev->iowait_lock;
437 			trace_hfi1_qpsleep(qp, RVT_S_WAIT_DMA_DESC);
438 			rvt_get_qp(qp);
439 		}
440 		write_sequnlock(&dev->iowait_lock);
441 		qp->s_flags &= ~RVT_S_BUSY;
442 		spin_unlock_irqrestore(&qp->s_lock, flags);
443 		ret = -EBUSY;
444 	} else {
445 		spin_unlock_irqrestore(&qp->s_lock, flags);
446 		hfi1_put_txreq(tx);
447 	}
448 	return ret;
449 eagain:
450 	write_sequnlock(&dev->iowait_lock);
451 	spin_unlock_irqrestore(&qp->s_lock, flags);
452 	list_del_init(&stx->list);
453 	return -EAGAIN;
454 }
455 
456 static void iowait_wakeup(struct iowait *wait, int reason)
457 {
458 	struct rvt_qp *qp = iowait_to_qp(wait);
459 
460 	WARN_ON(reason != SDMA_AVAIL_REASON);
461 	hfi1_qp_wakeup(qp, RVT_S_WAIT_DMA_DESC);
462 }
463 
464 static void iowait_sdma_drained(struct iowait *wait)
465 {
466 	struct rvt_qp *qp = iowait_to_qp(wait);
467 	unsigned long flags;
468 
469 	/*
470 	 * This happens when the send engine notes
471 	 * a QP in the error state and cannot
472 	 * do the flush work until that QP's
473 	 * sdma work has finished.
474 	 */
475 	spin_lock_irqsave(&qp->s_lock, flags);
476 	if (qp->s_flags & RVT_S_WAIT_DMA) {
477 		qp->s_flags &= ~RVT_S_WAIT_DMA;
478 		hfi1_schedule_send(qp);
479 	}
480 	spin_unlock_irqrestore(&qp->s_lock, flags);
481 }
482 
483 /**
484  *
485  * qp_to_sdma_engine - map a qp to a send engine
486  * @qp: the QP
487  * @sc5: the 5 bit sc
488  *
489  * Return:
490  * A send engine for the qp or NULL for SMI type qp.
491  */
492 struct sdma_engine *qp_to_sdma_engine(struct rvt_qp *qp, u8 sc5)
493 {
494 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
495 	struct sdma_engine *sde;
496 
497 	if (!(dd->flags & HFI1_HAS_SEND_DMA))
498 		return NULL;
499 	switch (qp->ibqp.qp_type) {
500 	case IB_QPT_SMI:
501 		return NULL;
502 	default:
503 		break;
504 	}
505 	sde = sdma_select_engine_sc(dd, qp->ibqp.qp_num >> dd->qos_shift, sc5);
506 	return sde;
507 }
508 
509 /*
510  * qp_to_send_context - map a qp to a send context
511  * @qp: the QP
512  * @sc5: the 5 bit sc
513  *
514  * Return:
515  * A send context for the qp
516  */
517 struct send_context *qp_to_send_context(struct rvt_qp *qp, u8 sc5)
518 {
519 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
520 
521 	switch (qp->ibqp.qp_type) {
522 	case IB_QPT_SMI:
523 		/* SMA packets to VL15 */
524 		return dd->vld[15].sc;
525 	default:
526 		break;
527 	}
528 
529 	return pio_select_send_context_sc(dd, qp->ibqp.qp_num >> dd->qos_shift,
530 					  sc5);
531 }
532 
533 static const char * const qp_type_str[] = {
534 	"SMI", "GSI", "RC", "UC", "UD",
535 };
536 
537 static int qp_idle(struct rvt_qp *qp)
538 {
539 	return
540 		qp->s_last == qp->s_acked &&
541 		qp->s_acked == qp->s_cur &&
542 		qp->s_cur == qp->s_tail &&
543 		qp->s_tail == qp->s_head;
544 }
545 
546 /**
547  * qp_iter_print - print the qp information to seq_file
548  * @s: the seq_file to emit the qp information on
549  * @iter: the iterator for the qp hash list
550  */
551 void qp_iter_print(struct seq_file *s, struct rvt_qp_iter *iter)
552 {
553 	struct rvt_swqe *wqe;
554 	struct rvt_qp *qp = iter->qp;
555 	struct hfi1_qp_priv *priv = qp->priv;
556 	struct sdma_engine *sde;
557 	struct send_context *send_context;
558 	struct rvt_ack_entry *e = NULL;
559 	struct rvt_srq *srq = qp->ibqp.srq ?
560 		ibsrq_to_rvtsrq(qp->ibqp.srq) : NULL;
561 
562 	sde = qp_to_sdma_engine(qp, priv->s_sc);
563 	wqe = rvt_get_swqe_ptr(qp, qp->s_last);
564 	send_context = qp_to_send_context(qp, priv->s_sc);
565 	if (qp->s_ack_queue)
566 		e = &qp->s_ack_queue[qp->s_tail_ack_queue];
567 	seq_printf(s,
568 		   "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",
569 		   iter->n,
570 		   qp_idle(qp) ? "I" : "B",
571 		   qp->ibqp.qp_num,
572 		   atomic_read(&qp->refcount),
573 		   qp_type_str[qp->ibqp.qp_type],
574 		   qp->state,
575 		   wqe ? wqe->wr.opcode : 0,
576 		   qp->s_flags,
577 		   iowait_sdma_pending(&priv->s_iowait),
578 		   iowait_pio_pending(&priv->s_iowait),
579 		   !list_empty(&priv->s_iowait.list),
580 		   qp->timeout,
581 		   wqe ? wqe->ssn : 0,
582 		   qp->s_lsn,
583 		   qp->s_last_psn,
584 		   qp->s_psn, qp->s_next_psn,
585 		   qp->s_sending_psn, qp->s_sending_hpsn,
586 		   qp->r_psn,
587 		   qp->s_last, qp->s_acked, qp->s_cur,
588 		   qp->s_tail, qp->s_head, qp->s_size,
589 		   qp->s_avail,
590 		   /* ack_queue ring pointers, size */
591 		   qp->s_tail_ack_queue, qp->r_head_ack_queue,
592 		   rvt_max_atomic(&to_idev(qp->ibqp.device)->rdi),
593 		   /* remote QP info  */
594 		   qp->remote_qpn,
595 		   rdma_ah_get_dlid(&qp->remote_ah_attr),
596 		   rdma_ah_get_sl(&qp->remote_ah_attr),
597 		   qp->pmtu,
598 		   qp->s_retry,
599 		   qp->s_retry_cnt,
600 		   qp->s_rnr_retry_cnt,
601 		   qp->s_rnr_retry,
602 		   sde,
603 		   sde ? sde->this_idx : 0,
604 		   send_context,
605 		   send_context ? send_context->sw_index : 0,
606 		   ibcq_to_rvtcq(qp->ibqp.send_cq)->queue->head,
607 		   ibcq_to_rvtcq(qp->ibqp.send_cq)->queue->tail,
608 		   qp->pid,
609 		   qp->s_state,
610 		   qp->s_ack_state,
611 		   /* ack queue information */
612 		   e ? e->opcode : 0,
613 		   e ? e->psn : 0,
614 		   e ? e->lpsn : 0,
615 		   qp->r_min_rnr_timer,
616 		   srq ? "SRQ" : "RQ",
617 		   srq ? srq->rq.size : qp->r_rq.size
618 		);
619 }
620 
621 void *qp_priv_alloc(struct rvt_dev_info *rdi, struct rvt_qp *qp)
622 {
623 	struct hfi1_qp_priv *priv;
624 
625 	priv = kzalloc_node(sizeof(*priv), GFP_KERNEL, rdi->dparms.node);
626 	if (!priv)
627 		return ERR_PTR(-ENOMEM);
628 
629 	priv->owner = qp;
630 
631 	priv->s_ahg = kzalloc_node(sizeof(*priv->s_ahg), GFP_KERNEL,
632 				   rdi->dparms.node);
633 	if (!priv->s_ahg) {
634 		kfree(priv);
635 		return ERR_PTR(-ENOMEM);
636 	}
637 	iowait_init(
638 		&priv->s_iowait,
639 		1,
640 		_hfi1_do_send,
641 		iowait_sleep,
642 		iowait_wakeup,
643 		iowait_sdma_drained);
644 	return priv;
645 }
646 
647 void qp_priv_free(struct rvt_dev_info *rdi, struct rvt_qp *qp)
648 {
649 	struct hfi1_qp_priv *priv = qp->priv;
650 
651 	kfree(priv->s_ahg);
652 	kfree(priv);
653 }
654 
655 unsigned free_all_qps(struct rvt_dev_info *rdi)
656 {
657 	struct hfi1_ibdev *verbs_dev = container_of(rdi,
658 						    struct hfi1_ibdev,
659 						    rdi);
660 	struct hfi1_devdata *dd = container_of(verbs_dev,
661 					       struct hfi1_devdata,
662 					       verbs_dev);
663 	int n;
664 	unsigned qp_inuse = 0;
665 
666 	for (n = 0; n < dd->num_pports; n++) {
667 		struct hfi1_ibport *ibp = &dd->pport[n].ibport_data;
668 
669 		rcu_read_lock();
670 		if (rcu_dereference(ibp->rvp.qp[0]))
671 			qp_inuse++;
672 		if (rcu_dereference(ibp->rvp.qp[1]))
673 			qp_inuse++;
674 		rcu_read_unlock();
675 	}
676 
677 	return qp_inuse;
678 }
679 
680 void flush_qp_waiters(struct rvt_qp *qp)
681 {
682 	lockdep_assert_held(&qp->s_lock);
683 	flush_iowait(qp);
684 }
685 
686 void stop_send_queue(struct rvt_qp *qp)
687 {
688 	struct hfi1_qp_priv *priv = qp->priv;
689 
690 	cancel_work_sync(&priv->s_iowait.iowork);
691 }
692 
693 void quiesce_qp(struct rvt_qp *qp)
694 {
695 	struct hfi1_qp_priv *priv = qp->priv;
696 
697 	iowait_sdma_drain(&priv->s_iowait);
698 	qp_pio_drain(qp);
699 	flush_tx_list(qp);
700 }
701 
702 void notify_qp_reset(struct rvt_qp *qp)
703 {
704 	qp->r_adefered = 0;
705 	clear_ahg(qp);
706 }
707 
708 /*
709  * Switch to alternate path.
710  * The QP s_lock should be held and interrupts disabled.
711  */
712 void hfi1_migrate_qp(struct rvt_qp *qp)
713 {
714 	struct hfi1_qp_priv *priv = qp->priv;
715 	struct ib_event ev;
716 
717 	qp->s_mig_state = IB_MIG_MIGRATED;
718 	qp->remote_ah_attr = qp->alt_ah_attr;
719 	qp->port_num = rdma_ah_get_port_num(&qp->alt_ah_attr);
720 	qp->s_pkey_index = qp->s_alt_pkey_index;
721 	qp->s_flags |= RVT_S_AHG_CLEAR;
722 	priv->s_sc = ah_to_sc(qp->ibqp.device, &qp->remote_ah_attr);
723 	priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
724 	qp_set_16b(qp);
725 
726 	ev.device = qp->ibqp.device;
727 	ev.element.qp = &qp->ibqp;
728 	ev.event = IB_EVENT_PATH_MIG;
729 	qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
730 }
731 
732 int mtu_to_path_mtu(u32 mtu)
733 {
734 	return mtu_to_enum(mtu, OPA_MTU_8192);
735 }
736 
737 u32 mtu_from_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp, u32 pmtu)
738 {
739 	u32 mtu;
740 	struct hfi1_ibdev *verbs_dev = container_of(rdi,
741 						    struct hfi1_ibdev,
742 						    rdi);
743 	struct hfi1_devdata *dd = container_of(verbs_dev,
744 					       struct hfi1_devdata,
745 					       verbs_dev);
746 	struct hfi1_ibport *ibp;
747 	u8 sc, vl;
748 
749 	ibp = &dd->pport[qp->port_num - 1].ibport_data;
750 	sc = ibp->sl_to_sc[rdma_ah_get_sl(&qp->remote_ah_attr)];
751 	vl = sc_to_vlt(dd, sc);
752 
753 	mtu = verbs_mtu_enum_to_int(qp->ibqp.device, pmtu);
754 	if (vl < PER_VL_SEND_CONTEXTS)
755 		mtu = min_t(u32, mtu, dd->vld[vl].mtu);
756 	return mtu;
757 }
758 
759 int get_pmtu_from_attr(struct rvt_dev_info *rdi, struct rvt_qp *qp,
760 		       struct ib_qp_attr *attr)
761 {
762 	int mtu, pidx = qp->port_num - 1;
763 	struct hfi1_ibdev *verbs_dev = container_of(rdi,
764 						    struct hfi1_ibdev,
765 						    rdi);
766 	struct hfi1_devdata *dd = container_of(verbs_dev,
767 					       struct hfi1_devdata,
768 					       verbs_dev);
769 	mtu = verbs_mtu_enum_to_int(qp->ibqp.device, attr->path_mtu);
770 	if (mtu == -1)
771 		return -1; /* values less than 0 are error */
772 
773 	if (mtu > dd->pport[pidx].ibmtu)
774 		return mtu_to_enum(dd->pport[pidx].ibmtu, IB_MTU_2048);
775 	else
776 		return attr->path_mtu;
777 }
778 
779 void notify_error_qp(struct rvt_qp *qp)
780 {
781 	struct hfi1_qp_priv *priv = qp->priv;
782 	seqlock_t *lock = priv->s_iowait.lock;
783 
784 	if (lock) {
785 		write_seqlock(lock);
786 		if (!list_empty(&priv->s_iowait.list) &&
787 		    !(qp->s_flags & RVT_S_BUSY)) {
788 			qp->s_flags &= ~RVT_S_ANY_WAIT_IO;
789 			list_del_init(&priv->s_iowait.list);
790 			priv->s_iowait.lock = NULL;
791 			rvt_put_qp(qp);
792 		}
793 		write_sequnlock(lock);
794 	}
795 
796 	if (!(qp->s_flags & RVT_S_BUSY)) {
797 		if (qp->s_rdma_mr) {
798 			rvt_put_mr(qp->s_rdma_mr);
799 			qp->s_rdma_mr = NULL;
800 		}
801 		flush_tx_list(qp);
802 	}
803 }
804 
805 /**
806  * hfi1_qp_iter_cb - callback for iterator
807  * @qp - the qp
808  * @v - the sl in low bits of v
809  *
810  * This is called from the iterator callback to work
811  * on an individual qp.
812  */
813 static void hfi1_qp_iter_cb(struct rvt_qp *qp, u64 v)
814 {
815 	int lastwqe;
816 	struct ib_event ev;
817 	struct hfi1_ibport *ibp =
818 		to_iport(qp->ibqp.device, qp->port_num);
819 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
820 	u8 sl = (u8)v;
821 
822 	if (qp->port_num != ppd->port ||
823 	    (qp->ibqp.qp_type != IB_QPT_UC &&
824 	     qp->ibqp.qp_type != IB_QPT_RC) ||
825 	    rdma_ah_get_sl(&qp->remote_ah_attr) != sl ||
826 	    !(ib_rvt_state_ops[qp->state] & RVT_POST_SEND_OK))
827 		return;
828 
829 	spin_lock_irq(&qp->r_lock);
830 	spin_lock(&qp->s_hlock);
831 	spin_lock(&qp->s_lock);
832 	lastwqe = rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
833 	spin_unlock(&qp->s_lock);
834 	spin_unlock(&qp->s_hlock);
835 	spin_unlock_irq(&qp->r_lock);
836 	if (lastwqe) {
837 		ev.device = qp->ibqp.device;
838 		ev.element.qp = &qp->ibqp;
839 		ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
840 		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
841 	}
842 }
843 
844 /**
845  * hfi1_error_port_qps - put a port's RC/UC qps into error state
846  * @ibp: the ibport.
847  * @sl: the service level.
848  *
849  * This function places all RC/UC qps with a given service level into error
850  * state. It is generally called to force upper lay apps to abandon stale qps
851  * after an sl->sc mapping change.
852  */
853 void hfi1_error_port_qps(struct hfi1_ibport *ibp, u8 sl)
854 {
855 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
856 	struct hfi1_ibdev *dev = &ppd->dd->verbs_dev;
857 
858 	rvt_qp_iter(&dev->rdi, sl, hfi1_qp_iter_cb);
859 }
860