xref: /linux/drivers/infiniband/hw/hfi1/qp.c (revision 86919f9dd2dbb88bbbe4e381dc8e6b9e9a1b9eb5)
1 /*
2  * Copyright(c) 2015 - 2018 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 };
136 
137 static void flush_list_head(struct list_head *l)
138 {
139 	while (!list_empty(l)) {
140 		struct sdma_txreq *tx;
141 
142 		tx = list_first_entry(
143 			l,
144 			struct sdma_txreq,
145 			list);
146 		list_del_init(&tx->list);
147 		hfi1_put_txreq(
148 			container_of(tx, struct verbs_txreq, txreq));
149 	}
150 }
151 
152 static void flush_tx_list(struct rvt_qp *qp)
153 {
154 	struct hfi1_qp_priv *priv = qp->priv;
155 
156 	flush_list_head(&iowait_get_ib_work(&priv->s_iowait)->tx_head);
157 	flush_list_head(&iowait_get_tid_work(&priv->s_iowait)->tx_head);
158 }
159 
160 static void flush_iowait(struct rvt_qp *qp)
161 {
162 	struct hfi1_qp_priv *priv = qp->priv;
163 	unsigned long flags;
164 	seqlock_t *lock = priv->s_iowait.lock;
165 
166 	if (!lock)
167 		return;
168 	write_seqlock_irqsave(lock, flags);
169 	if (!list_empty(&priv->s_iowait.list)) {
170 		list_del_init(&priv->s_iowait.list);
171 		priv->s_iowait.lock = NULL;
172 		rvt_put_qp(qp);
173 	}
174 	write_sequnlock_irqrestore(lock, flags);
175 }
176 
177 static inline int opa_mtu_enum_to_int(int mtu)
178 {
179 	switch (mtu) {
180 	case OPA_MTU_8192:  return 8192;
181 	case OPA_MTU_10240: return 10240;
182 	default:            return -1;
183 	}
184 }
185 
186 /**
187  * This function is what we would push to the core layer if we wanted to be a
188  * "first class citizen".  Instead we hide this here and rely on Verbs ULPs
189  * to blindly pass the MTU enum value from the PathRecord to us.
190  */
191 static inline int verbs_mtu_enum_to_int(struct ib_device *dev, enum ib_mtu mtu)
192 {
193 	int val;
194 
195 	/* Constraining 10KB packets to 8KB packets */
196 	if (mtu == (enum ib_mtu)OPA_MTU_10240)
197 		mtu = OPA_MTU_8192;
198 	val = opa_mtu_enum_to_int((int)mtu);
199 	if (val > 0)
200 		return val;
201 	return ib_mtu_enum_to_int(mtu);
202 }
203 
204 int hfi1_check_modify_qp(struct rvt_qp *qp, struct ib_qp_attr *attr,
205 			 int attr_mask, struct ib_udata *udata)
206 {
207 	struct ib_qp *ibqp = &qp->ibqp;
208 	struct hfi1_ibdev *dev = to_idev(ibqp->device);
209 	struct hfi1_devdata *dd = dd_from_dev(dev);
210 	u8 sc;
211 
212 	if (attr_mask & IB_QP_AV) {
213 		sc = ah_to_sc(ibqp->device, &attr->ah_attr);
214 		if (sc == 0xf)
215 			return -EINVAL;
216 
217 		if (!qp_to_sdma_engine(qp, sc) &&
218 		    dd->flags & HFI1_HAS_SEND_DMA)
219 			return -EINVAL;
220 
221 		if (!qp_to_send_context(qp, sc))
222 			return -EINVAL;
223 	}
224 
225 	if (attr_mask & IB_QP_ALT_PATH) {
226 		sc = ah_to_sc(ibqp->device, &attr->alt_ah_attr);
227 		if (sc == 0xf)
228 			return -EINVAL;
229 
230 		if (!qp_to_sdma_engine(qp, sc) &&
231 		    dd->flags & HFI1_HAS_SEND_DMA)
232 			return -EINVAL;
233 
234 		if (!qp_to_send_context(qp, sc))
235 			return -EINVAL;
236 	}
237 
238 	return 0;
239 }
240 
241 /*
242  * qp_set_16b - Set the hdr_type based on whether the slid or the
243  * dlid in the connection is extended. Only applicable for RC and UC
244  * QPs. UD QPs determine this on the fly from the ah in the wqe
245  */
246 static inline void qp_set_16b(struct rvt_qp *qp)
247 {
248 	struct hfi1_pportdata *ppd;
249 	struct hfi1_ibport *ibp;
250 	struct hfi1_qp_priv *priv = qp->priv;
251 
252 	/* Update ah_attr to account for extended LIDs */
253 	hfi1_update_ah_attr(qp->ibqp.device, &qp->remote_ah_attr);
254 
255 	/* Create 32 bit LIDs */
256 	hfi1_make_opa_lid(&qp->remote_ah_attr);
257 
258 	if (!(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH))
259 		return;
260 
261 	ibp = to_iport(qp->ibqp.device, qp->port_num);
262 	ppd = ppd_from_ibp(ibp);
263 	priv->hdr_type = hfi1_get_hdr_type(ppd->lid, &qp->remote_ah_attr);
264 }
265 
266 void hfi1_modify_qp(struct rvt_qp *qp, struct ib_qp_attr *attr,
267 		    int attr_mask, struct ib_udata *udata)
268 {
269 	struct ib_qp *ibqp = &qp->ibqp;
270 	struct hfi1_qp_priv *priv = qp->priv;
271 
272 	if (attr_mask & IB_QP_AV) {
273 		priv->s_sc = ah_to_sc(ibqp->device, &qp->remote_ah_attr);
274 		priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
275 		priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
276 		qp_set_16b(qp);
277 	}
278 
279 	if (attr_mask & IB_QP_PATH_MIG_STATE &&
280 	    attr->path_mig_state == IB_MIG_MIGRATED &&
281 	    qp->s_mig_state == IB_MIG_ARMED) {
282 		qp->s_flags |= HFI1_S_AHG_CLEAR;
283 		priv->s_sc = ah_to_sc(ibqp->device, &qp->remote_ah_attr);
284 		priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
285 		priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
286 		qp_set_16b(qp);
287 	}
288 }
289 
290 /**
291  * hfi1_setup_wqe - set up the wqe
292  * @qp - The qp
293  * @wqe - The built wqe
294  * @call_send - Determine if the send should be posted or scheduled.
295  *
296  * Perform setup of the wqe.  This is called
297  * prior to inserting the wqe into the ring but after
298  * the wqe has been setup by RDMAVT. This function
299  * allows the driver the opportunity to perform
300  * validation and additional setup of the wqe.
301  *
302  * Returns 0 on success, -EINVAL on failure
303  *
304  */
305 int hfi1_setup_wqe(struct rvt_qp *qp, struct rvt_swqe *wqe, bool *call_send)
306 {
307 	struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
308 	struct rvt_ah *ah;
309 	struct hfi1_pportdata *ppd;
310 	struct hfi1_devdata *dd;
311 
312 	switch (qp->ibqp.qp_type) {
313 	case IB_QPT_RC:
314 	case IB_QPT_UC:
315 		if (wqe->length > 0x80000000U)
316 			return -EINVAL;
317 		if (wqe->length > qp->pmtu)
318 			*call_send = false;
319 		break;
320 	case IB_QPT_SMI:
321 		/*
322 		 * SM packets should exclusively use VL15 and their SL is
323 		 * ignored (IBTA v1.3, Section 3.5.8.2). Therefore, when ah
324 		 * is created, SL is 0 in most cases and as a result some
325 		 * fields (vl and pmtu) in ah may not be set correctly,
326 		 * depending on the SL2SC and SC2VL tables at the time.
327 		 */
328 		ppd = ppd_from_ibp(ibp);
329 		dd = dd_from_ppd(ppd);
330 		if (wqe->length > dd->vld[15].mtu)
331 			return -EINVAL;
332 		break;
333 	case IB_QPT_GSI:
334 	case IB_QPT_UD:
335 		ah = ibah_to_rvtah(wqe->ud_wr.ah);
336 		if (wqe->length > (1 << ah->log_pmtu))
337 			return -EINVAL;
338 		if (ibp->sl_to_sc[rdma_ah_get_sl(&ah->attr)] == 0xf)
339 			return -EINVAL;
340 	default:
341 		break;
342 	}
343 
344 	/*
345 	 * System latency between send and schedule is large enough that
346 	 * forcing call_send to true for piothreshold packets is necessary.
347 	 */
348 	if (wqe->length <= piothreshold)
349 		*call_send = true;
350 	return 0;
351 }
352 
353 /**
354  * _hfi1_schedule_send - schedule progress
355  * @qp: the QP
356  *
357  * This schedules qp progress w/o regard to the s_flags.
358  *
359  * It is only used in the post send, which doesn't hold
360  * the s_lock.
361  */
362 bool _hfi1_schedule_send(struct rvt_qp *qp)
363 {
364 	struct hfi1_qp_priv *priv = qp->priv;
365 	struct hfi1_ibport *ibp =
366 		to_iport(qp->ibqp.device, qp->port_num);
367 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
368 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
369 
370 	return iowait_schedule(&priv->s_iowait, ppd->hfi1_wq,
371 			       priv->s_sde ?
372 			       priv->s_sde->cpu :
373 			       cpumask_first(cpumask_of_node(dd->node)));
374 }
375 
376 static void qp_pio_drain(struct rvt_qp *qp)
377 {
378 	struct hfi1_ibdev *dev;
379 	struct hfi1_qp_priv *priv = qp->priv;
380 
381 	if (!priv->s_sendcontext)
382 		return;
383 	dev = to_idev(qp->ibqp.device);
384 	while (iowait_pio_pending(&priv->s_iowait)) {
385 		write_seqlock_irq(&dev->iowait_lock);
386 		hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 1);
387 		write_sequnlock_irq(&dev->iowait_lock);
388 		iowait_pio_drain(&priv->s_iowait);
389 		write_seqlock_irq(&dev->iowait_lock);
390 		hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 0);
391 		write_sequnlock_irq(&dev->iowait_lock);
392 	}
393 }
394 
395 /**
396  * hfi1_schedule_send - schedule progress
397  * @qp: the QP
398  *
399  * This schedules qp progress and caller should hold
400  * the s_lock.
401  * @return true if the first leg is scheduled;
402  * false if the first leg is not scheduled.
403  */
404 bool hfi1_schedule_send(struct rvt_qp *qp)
405 {
406 	lockdep_assert_held(&qp->s_lock);
407 	if (hfi1_send_ok(qp)) {
408 		_hfi1_schedule_send(qp);
409 		return true;
410 	}
411 	if (qp->s_flags & HFI1_S_ANY_WAIT_IO)
412 		iowait_set_flag(&((struct hfi1_qp_priv *)qp->priv)->s_iowait,
413 				IOWAIT_PENDING_IB);
414 	return false;
415 }
416 
417 static void hfi1_qp_schedule(struct rvt_qp *qp)
418 {
419 	struct hfi1_qp_priv *priv = qp->priv;
420 	bool ret;
421 
422 	if (iowait_flag_set(&priv->s_iowait, IOWAIT_PENDING_IB)) {
423 		ret = hfi1_schedule_send(qp);
424 		if (ret)
425 			iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_IB);
426 	}
427 }
428 
429 void hfi1_qp_wakeup(struct rvt_qp *qp, u32 flag)
430 {
431 	unsigned long flags;
432 
433 	spin_lock_irqsave(&qp->s_lock, flags);
434 	if (qp->s_flags & flag) {
435 		qp->s_flags &= ~flag;
436 		trace_hfi1_qpwakeup(qp, flag);
437 		hfi1_qp_schedule(qp);
438 	}
439 	spin_unlock_irqrestore(&qp->s_lock, flags);
440 	/* Notify hfi1_destroy_qp() if it is waiting. */
441 	rvt_put_qp(qp);
442 }
443 
444 void hfi1_qp_unbusy(struct rvt_qp *qp, struct iowait_work *wait)
445 {
446 	if (iowait_set_work_flag(wait) == IOWAIT_IB_SE)
447 		qp->s_flags &= ~RVT_S_BUSY;
448 }
449 
450 static int iowait_sleep(
451 	struct sdma_engine *sde,
452 	struct iowait_work *wait,
453 	struct sdma_txreq *stx,
454 	uint seq,
455 	bool pkts_sent)
456 {
457 	struct verbs_txreq *tx = container_of(stx, struct verbs_txreq, txreq);
458 	struct rvt_qp *qp;
459 	struct hfi1_qp_priv *priv;
460 	unsigned long flags;
461 	int ret = 0;
462 	struct hfi1_ibdev *dev;
463 
464 	qp = tx->qp;
465 	priv = qp->priv;
466 
467 	spin_lock_irqsave(&qp->s_lock, flags);
468 	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
469 		/*
470 		 * If we couldn't queue the DMA request, save the info
471 		 * and try again later rather than destroying the
472 		 * buffer and undoing the side effects of the copy.
473 		 */
474 		/* Make a common routine? */
475 		dev = &sde->dd->verbs_dev;
476 		list_add_tail(&stx->list, &wait->tx_head);
477 		write_seqlock(&dev->iowait_lock);
478 		if (sdma_progress(sde, seq, stx))
479 			goto eagain;
480 		if (list_empty(&priv->s_iowait.list)) {
481 			struct hfi1_ibport *ibp =
482 				to_iport(qp->ibqp.device, qp->port_num);
483 
484 			ibp->rvp.n_dmawait++;
485 			qp->s_flags |= RVT_S_WAIT_DMA_DESC;
486 			iowait_queue(pkts_sent, &priv->s_iowait,
487 				     &sde->dmawait);
488 			priv->s_iowait.lock = &dev->iowait_lock;
489 			trace_hfi1_qpsleep(qp, RVT_S_WAIT_DMA_DESC);
490 			rvt_get_qp(qp);
491 		}
492 		write_sequnlock(&dev->iowait_lock);
493 		hfi1_qp_unbusy(qp, wait);
494 		spin_unlock_irqrestore(&qp->s_lock, flags);
495 		ret = -EBUSY;
496 	} else {
497 		spin_unlock_irqrestore(&qp->s_lock, flags);
498 		hfi1_put_txreq(tx);
499 	}
500 	return ret;
501 eagain:
502 	write_sequnlock(&dev->iowait_lock);
503 	spin_unlock_irqrestore(&qp->s_lock, flags);
504 	list_del_init(&stx->list);
505 	return -EAGAIN;
506 }
507 
508 static void iowait_wakeup(struct iowait *wait, int reason)
509 {
510 	struct rvt_qp *qp = iowait_to_qp(wait);
511 
512 	WARN_ON(reason != SDMA_AVAIL_REASON);
513 	hfi1_qp_wakeup(qp, RVT_S_WAIT_DMA_DESC);
514 }
515 
516 static void iowait_sdma_drained(struct iowait *wait)
517 {
518 	struct rvt_qp *qp = iowait_to_qp(wait);
519 	unsigned long flags;
520 
521 	/*
522 	 * This happens when the send engine notes
523 	 * a QP in the error state and cannot
524 	 * do the flush work until that QP's
525 	 * sdma work has finished.
526 	 */
527 	spin_lock_irqsave(&qp->s_lock, flags);
528 	if (qp->s_flags & RVT_S_WAIT_DMA) {
529 		qp->s_flags &= ~RVT_S_WAIT_DMA;
530 		hfi1_schedule_send(qp);
531 	}
532 	spin_unlock_irqrestore(&qp->s_lock, flags);
533 }
534 
535 /**
536  * qp_to_sdma_engine - map a qp to a send engine
537  * @qp: the QP
538  * @sc5: the 5 bit sc
539  *
540  * Return:
541  * A send engine for the qp or NULL for SMI type qp.
542  */
543 struct sdma_engine *qp_to_sdma_engine(struct rvt_qp *qp, u8 sc5)
544 {
545 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
546 	struct sdma_engine *sde;
547 
548 	if (!(dd->flags & HFI1_HAS_SEND_DMA))
549 		return NULL;
550 	switch (qp->ibqp.qp_type) {
551 	case IB_QPT_SMI:
552 		return NULL;
553 	default:
554 		break;
555 	}
556 	sde = sdma_select_engine_sc(dd, qp->ibqp.qp_num >> dd->qos_shift, sc5);
557 	return sde;
558 }
559 
560 /*
561  * qp_to_send_context - map a qp to a send context
562  * @qp: the QP
563  * @sc5: the 5 bit sc
564  *
565  * Return:
566  * A send context for the qp
567  */
568 struct send_context *qp_to_send_context(struct rvt_qp *qp, u8 sc5)
569 {
570 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
571 
572 	switch (qp->ibqp.qp_type) {
573 	case IB_QPT_SMI:
574 		/* SMA packets to VL15 */
575 		return dd->vld[15].sc;
576 	default:
577 		break;
578 	}
579 
580 	return pio_select_send_context_sc(dd, qp->ibqp.qp_num >> dd->qos_shift,
581 					  sc5);
582 }
583 
584 static const char * const qp_type_str[] = {
585 	"SMI", "GSI", "RC", "UC", "UD",
586 };
587 
588 static int qp_idle(struct rvt_qp *qp)
589 {
590 	return
591 		qp->s_last == qp->s_acked &&
592 		qp->s_acked == qp->s_cur &&
593 		qp->s_cur == qp->s_tail &&
594 		qp->s_tail == qp->s_head;
595 }
596 
597 /**
598  * qp_iter_print - print the qp information to seq_file
599  * @s: the seq_file to emit the qp information on
600  * @iter: the iterator for the qp hash list
601  */
602 void qp_iter_print(struct seq_file *s, struct rvt_qp_iter *iter)
603 {
604 	struct rvt_swqe *wqe;
605 	struct rvt_qp *qp = iter->qp;
606 	struct hfi1_qp_priv *priv = qp->priv;
607 	struct sdma_engine *sde;
608 	struct send_context *send_context;
609 	struct rvt_ack_entry *e = NULL;
610 	struct rvt_srq *srq = qp->ibqp.srq ?
611 		ibsrq_to_rvtsrq(qp->ibqp.srq) : NULL;
612 
613 	sde = qp_to_sdma_engine(qp, priv->s_sc);
614 	wqe = rvt_get_swqe_ptr(qp, qp->s_last);
615 	send_context = qp_to_send_context(qp, priv->s_sc);
616 	if (qp->s_ack_queue)
617 		e = &qp->s_ack_queue[qp->s_tail_ack_queue];
618 	seq_printf(s,
619 		   "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",
620 		   iter->n,
621 		   qp_idle(qp) ? "I" : "B",
622 		   qp->ibqp.qp_num,
623 		   atomic_read(&qp->refcount),
624 		   qp_type_str[qp->ibqp.qp_type],
625 		   qp->state,
626 		   wqe ? wqe->wr.opcode : 0,
627 		   qp->s_flags,
628 		   iowait_sdma_pending(&priv->s_iowait),
629 		   iowait_pio_pending(&priv->s_iowait),
630 		   !list_empty(&priv->s_iowait.list),
631 		   qp->timeout,
632 		   wqe ? wqe->ssn : 0,
633 		   qp->s_lsn,
634 		   qp->s_last_psn,
635 		   qp->s_psn, qp->s_next_psn,
636 		   qp->s_sending_psn, qp->s_sending_hpsn,
637 		   qp->r_psn,
638 		   qp->s_last, qp->s_acked, qp->s_cur,
639 		   qp->s_tail, qp->s_head, qp->s_size,
640 		   qp->s_avail,
641 		   /* ack_queue ring pointers, size */
642 		   qp->s_tail_ack_queue, qp->r_head_ack_queue,
643 		   rvt_max_atomic(&to_idev(qp->ibqp.device)->rdi),
644 		   /* remote QP info  */
645 		   qp->remote_qpn,
646 		   rdma_ah_get_dlid(&qp->remote_ah_attr),
647 		   rdma_ah_get_sl(&qp->remote_ah_attr),
648 		   qp->pmtu,
649 		   qp->s_retry,
650 		   qp->s_retry_cnt,
651 		   qp->s_rnr_retry_cnt,
652 		   qp->s_rnr_retry,
653 		   sde,
654 		   sde ? sde->this_idx : 0,
655 		   send_context,
656 		   send_context ? send_context->sw_index : 0,
657 		   ibcq_to_rvtcq(qp->ibqp.send_cq)->queue->head,
658 		   ibcq_to_rvtcq(qp->ibqp.send_cq)->queue->tail,
659 		   qp->pid,
660 		   qp->s_state,
661 		   qp->s_ack_state,
662 		   /* ack queue information */
663 		   e ? e->opcode : 0,
664 		   e ? e->psn : 0,
665 		   e ? e->lpsn : 0,
666 		   qp->r_min_rnr_timer,
667 		   srq ? "SRQ" : "RQ",
668 		   srq ? srq->rq.size : qp->r_rq.size
669 		);
670 }
671 
672 void *qp_priv_alloc(struct rvt_dev_info *rdi, struct rvt_qp *qp)
673 {
674 	struct hfi1_qp_priv *priv;
675 
676 	priv = kzalloc_node(sizeof(*priv), GFP_KERNEL, rdi->dparms.node);
677 	if (!priv)
678 		return ERR_PTR(-ENOMEM);
679 
680 	priv->owner = qp;
681 
682 	priv->s_ahg = kzalloc_node(sizeof(*priv->s_ahg), GFP_KERNEL,
683 				   rdi->dparms.node);
684 	if (!priv->s_ahg) {
685 		kfree(priv);
686 		return ERR_PTR(-ENOMEM);
687 	}
688 	iowait_init(
689 		&priv->s_iowait,
690 		1,
691 		_hfi1_do_send,
692 		NULL,
693 		iowait_sleep,
694 		iowait_wakeup,
695 		iowait_sdma_drained);
696 	return priv;
697 }
698 
699 void qp_priv_free(struct rvt_dev_info *rdi, struct rvt_qp *qp)
700 {
701 	struct hfi1_qp_priv *priv = qp->priv;
702 
703 	kfree(priv->s_ahg);
704 	kfree(priv);
705 }
706 
707 unsigned free_all_qps(struct rvt_dev_info *rdi)
708 {
709 	struct hfi1_ibdev *verbs_dev = container_of(rdi,
710 						    struct hfi1_ibdev,
711 						    rdi);
712 	struct hfi1_devdata *dd = container_of(verbs_dev,
713 					       struct hfi1_devdata,
714 					       verbs_dev);
715 	int n;
716 	unsigned qp_inuse = 0;
717 
718 	for (n = 0; n < dd->num_pports; n++) {
719 		struct hfi1_ibport *ibp = &dd->pport[n].ibport_data;
720 
721 		rcu_read_lock();
722 		if (rcu_dereference(ibp->rvp.qp[0]))
723 			qp_inuse++;
724 		if (rcu_dereference(ibp->rvp.qp[1]))
725 			qp_inuse++;
726 		rcu_read_unlock();
727 	}
728 
729 	return qp_inuse;
730 }
731 
732 void flush_qp_waiters(struct rvt_qp *qp)
733 {
734 	lockdep_assert_held(&qp->s_lock);
735 	flush_iowait(qp);
736 }
737 
738 void stop_send_queue(struct rvt_qp *qp)
739 {
740 	struct hfi1_qp_priv *priv = qp->priv;
741 
742 	iowait_cancel_work(&priv->s_iowait);
743 }
744 
745 void quiesce_qp(struct rvt_qp *qp)
746 {
747 	struct hfi1_qp_priv *priv = qp->priv;
748 
749 	iowait_sdma_drain(&priv->s_iowait);
750 	qp_pio_drain(qp);
751 	flush_tx_list(qp);
752 }
753 
754 void notify_qp_reset(struct rvt_qp *qp)
755 {
756 	qp->r_adefered = 0;
757 	clear_ahg(qp);
758 }
759 
760 /*
761  * Switch to alternate path.
762  * The QP s_lock should be held and interrupts disabled.
763  */
764 void hfi1_migrate_qp(struct rvt_qp *qp)
765 {
766 	struct hfi1_qp_priv *priv = qp->priv;
767 	struct ib_event ev;
768 
769 	qp->s_mig_state = IB_MIG_MIGRATED;
770 	qp->remote_ah_attr = qp->alt_ah_attr;
771 	qp->port_num = rdma_ah_get_port_num(&qp->alt_ah_attr);
772 	qp->s_pkey_index = qp->s_alt_pkey_index;
773 	qp->s_flags |= HFI1_S_AHG_CLEAR;
774 	priv->s_sc = ah_to_sc(qp->ibqp.device, &qp->remote_ah_attr);
775 	priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
776 	qp_set_16b(qp);
777 
778 	ev.device = qp->ibqp.device;
779 	ev.element.qp = &qp->ibqp;
780 	ev.event = IB_EVENT_PATH_MIG;
781 	qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
782 }
783 
784 int mtu_to_path_mtu(u32 mtu)
785 {
786 	return mtu_to_enum(mtu, OPA_MTU_8192);
787 }
788 
789 u32 mtu_from_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp, u32 pmtu)
790 {
791 	u32 mtu;
792 	struct hfi1_ibdev *verbs_dev = container_of(rdi,
793 						    struct hfi1_ibdev,
794 						    rdi);
795 	struct hfi1_devdata *dd = container_of(verbs_dev,
796 					       struct hfi1_devdata,
797 					       verbs_dev);
798 	struct hfi1_ibport *ibp;
799 	u8 sc, vl;
800 
801 	ibp = &dd->pport[qp->port_num - 1].ibport_data;
802 	sc = ibp->sl_to_sc[rdma_ah_get_sl(&qp->remote_ah_attr)];
803 	vl = sc_to_vlt(dd, sc);
804 
805 	mtu = verbs_mtu_enum_to_int(qp->ibqp.device, pmtu);
806 	if (vl < PER_VL_SEND_CONTEXTS)
807 		mtu = min_t(u32, mtu, dd->vld[vl].mtu);
808 	return mtu;
809 }
810 
811 int get_pmtu_from_attr(struct rvt_dev_info *rdi, struct rvt_qp *qp,
812 		       struct ib_qp_attr *attr)
813 {
814 	int mtu, pidx = qp->port_num - 1;
815 	struct hfi1_ibdev *verbs_dev = container_of(rdi,
816 						    struct hfi1_ibdev,
817 						    rdi);
818 	struct hfi1_devdata *dd = container_of(verbs_dev,
819 					       struct hfi1_devdata,
820 					       verbs_dev);
821 	mtu = verbs_mtu_enum_to_int(qp->ibqp.device, attr->path_mtu);
822 	if (mtu == -1)
823 		return -1; /* values less than 0 are error */
824 
825 	if (mtu > dd->pport[pidx].ibmtu)
826 		return mtu_to_enum(dd->pport[pidx].ibmtu, IB_MTU_2048);
827 	else
828 		return attr->path_mtu;
829 }
830 
831 void notify_error_qp(struct rvt_qp *qp)
832 {
833 	struct hfi1_qp_priv *priv = qp->priv;
834 	seqlock_t *lock = priv->s_iowait.lock;
835 
836 	if (lock) {
837 		write_seqlock(lock);
838 		if (!list_empty(&priv->s_iowait.list) &&
839 		    !(qp->s_flags & RVT_S_BUSY)) {
840 			qp->s_flags &= ~RVT_S_ANY_WAIT_IO;
841 			list_del_init(&priv->s_iowait.list);
842 			priv->s_iowait.lock = NULL;
843 			rvt_put_qp(qp);
844 		}
845 		write_sequnlock(lock);
846 	}
847 
848 	if (!(qp->s_flags & RVT_S_BUSY)) {
849 		if (qp->s_rdma_mr) {
850 			rvt_put_mr(qp->s_rdma_mr);
851 			qp->s_rdma_mr = NULL;
852 		}
853 		flush_tx_list(qp);
854 	}
855 }
856 
857 /**
858  * hfi1_qp_iter_cb - callback for iterator
859  * @qp - the qp
860  * @v - the sl in low bits of v
861  *
862  * This is called from the iterator callback to work
863  * on an individual qp.
864  */
865 static void hfi1_qp_iter_cb(struct rvt_qp *qp, u64 v)
866 {
867 	int lastwqe;
868 	struct ib_event ev;
869 	struct hfi1_ibport *ibp =
870 		to_iport(qp->ibqp.device, qp->port_num);
871 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
872 	u8 sl = (u8)v;
873 
874 	if (qp->port_num != ppd->port ||
875 	    (qp->ibqp.qp_type != IB_QPT_UC &&
876 	     qp->ibqp.qp_type != IB_QPT_RC) ||
877 	    rdma_ah_get_sl(&qp->remote_ah_attr) != sl ||
878 	    !(ib_rvt_state_ops[qp->state] & RVT_POST_SEND_OK))
879 		return;
880 
881 	spin_lock_irq(&qp->r_lock);
882 	spin_lock(&qp->s_hlock);
883 	spin_lock(&qp->s_lock);
884 	lastwqe = rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
885 	spin_unlock(&qp->s_lock);
886 	spin_unlock(&qp->s_hlock);
887 	spin_unlock_irq(&qp->r_lock);
888 	if (lastwqe) {
889 		ev.device = qp->ibqp.device;
890 		ev.element.qp = &qp->ibqp;
891 		ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
892 		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
893 	}
894 }
895 
896 /**
897  * hfi1_error_port_qps - put a port's RC/UC qps into error state
898  * @ibp: the ibport.
899  * @sl: the service level.
900  *
901  * This function places all RC/UC qps with a given service level into error
902  * state. It is generally called to force upper lay apps to abandon stale qps
903  * after an sl->sc mapping change.
904  */
905 void hfi1_error_port_qps(struct hfi1_ibport *ibp, u8 sl)
906 {
907 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
908 	struct hfi1_ibdev *dev = &ppd->dd->verbs_dev;
909 
910 	rvt_qp_iter(&dev->rdi, sl, hfi1_qp_iter_cb);
911 }
912