xref: /openbmc/linux/drivers/infiniband/hw/hfi1/qp.c (revision 78700c0a)
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/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 
56 #include "hfi.h"
57 #include "qp.h"
58 #include "trace.h"
59 #include "verbs_txreq.h"
60 
61 unsigned int hfi1_qp_table_size = 256;
62 module_param_named(qp_table_size, hfi1_qp_table_size, uint, S_IRUGO);
63 MODULE_PARM_DESC(qp_table_size, "QP table size");
64 
65 static void flush_tx_list(struct rvt_qp *qp);
66 static int iowait_sleep(
67 	struct sdma_engine *sde,
68 	struct iowait *wait,
69 	struct sdma_txreq *stx,
70 	unsigned seq);
71 static void iowait_wakeup(struct iowait *wait, int reason);
72 static void iowait_sdma_drained(struct iowait *wait);
73 static void qp_pio_drain(struct rvt_qp *qp);
74 
75 static inline unsigned mk_qpn(struct rvt_qpn_table *qpt,
76 			      struct rvt_qpn_map *map, unsigned off)
77 {
78 	return (map - qpt->map) * RVT_BITS_PER_PAGE + off;
79 }
80 
81 /*
82  * Convert the AETH credit code into the number of credits.
83  */
84 static const u16 credit_table[31] = {
85 	0,                      /* 0 */
86 	1,                      /* 1 */
87 	2,                      /* 2 */
88 	3,                      /* 3 */
89 	4,                      /* 4 */
90 	6,                      /* 5 */
91 	8,                      /* 6 */
92 	12,                     /* 7 */
93 	16,                     /* 8 */
94 	24,                     /* 9 */
95 	32,                     /* A */
96 	48,                     /* B */
97 	64,                     /* C */
98 	96,                     /* D */
99 	128,                    /* E */
100 	192,                    /* F */
101 	256,                    /* 10 */
102 	384,                    /* 11 */
103 	512,                    /* 12 */
104 	768,                    /* 13 */
105 	1024,                   /* 14 */
106 	1536,                   /* 15 */
107 	2048,                   /* 16 */
108 	3072,                   /* 17 */
109 	4096,                   /* 18 */
110 	6144,                   /* 19 */
111 	8192,                   /* 1A */
112 	12288,                  /* 1B */
113 	16384,                  /* 1C */
114 	24576,                  /* 1D */
115 	32768                   /* 1E */
116 };
117 
118 static void flush_tx_list(struct rvt_qp *qp)
119 {
120 	struct hfi1_qp_priv *priv = qp->priv;
121 
122 	while (!list_empty(&priv->s_iowait.tx_head)) {
123 		struct sdma_txreq *tx;
124 
125 		tx = list_first_entry(
126 			&priv->s_iowait.tx_head,
127 			struct sdma_txreq,
128 			list);
129 		list_del_init(&tx->list);
130 		hfi1_put_txreq(
131 			container_of(tx, struct verbs_txreq, txreq));
132 	}
133 }
134 
135 static void flush_iowait(struct rvt_qp *qp)
136 {
137 	struct hfi1_qp_priv *priv = qp->priv;
138 	struct hfi1_ibdev *dev = to_idev(qp->ibqp.device);
139 	unsigned long flags;
140 
141 	write_seqlock_irqsave(&dev->iowait_lock, flags);
142 	if (!list_empty(&priv->s_iowait.list)) {
143 		list_del_init(&priv->s_iowait.list);
144 		if (atomic_dec_and_test(&qp->refcount))
145 			wake_up(&qp->wait);
146 	}
147 	write_sequnlock_irqrestore(&dev->iowait_lock, flags);
148 }
149 
150 static inline int opa_mtu_enum_to_int(int mtu)
151 {
152 	switch (mtu) {
153 	case OPA_MTU_8192:  return 8192;
154 	case OPA_MTU_10240: return 10240;
155 	default:            return -1;
156 	}
157 }
158 
159 /**
160  * This function is what we would push to the core layer if we wanted to be a
161  * "first class citizen".  Instead we hide this here and rely on Verbs ULPs
162  * to blindly pass the MTU enum value from the PathRecord to us.
163  */
164 static inline int verbs_mtu_enum_to_int(struct ib_device *dev, enum ib_mtu mtu)
165 {
166 	int val;
167 
168 	/* Constraining 10KB packets to 8KB packets */
169 	if (mtu == (enum ib_mtu)OPA_MTU_10240)
170 		mtu = OPA_MTU_8192;
171 	val = opa_mtu_enum_to_int((int)mtu);
172 	if (val > 0)
173 		return val;
174 	return ib_mtu_enum_to_int(mtu);
175 }
176 
177 int hfi1_check_modify_qp(struct rvt_qp *qp, struct ib_qp_attr *attr,
178 			 int attr_mask, struct ib_udata *udata)
179 {
180 	struct ib_qp *ibqp = &qp->ibqp;
181 	struct hfi1_ibdev *dev = to_idev(ibqp->device);
182 	struct hfi1_devdata *dd = dd_from_dev(dev);
183 	u8 sc;
184 
185 	if (attr_mask & IB_QP_AV) {
186 		sc = ah_to_sc(ibqp->device, &attr->ah_attr);
187 		if (sc == 0xf)
188 			return -EINVAL;
189 
190 		if (!qp_to_sdma_engine(qp, sc) &&
191 		    dd->flags & HFI1_HAS_SEND_DMA)
192 			return -EINVAL;
193 
194 		if (!qp_to_send_context(qp, sc))
195 			return -EINVAL;
196 	}
197 
198 	if (attr_mask & IB_QP_ALT_PATH) {
199 		sc = ah_to_sc(ibqp->device, &attr->alt_ah_attr);
200 		if (sc == 0xf)
201 			return -EINVAL;
202 
203 		if (!qp_to_sdma_engine(qp, sc) &&
204 		    dd->flags & HFI1_HAS_SEND_DMA)
205 			return -EINVAL;
206 
207 		if (!qp_to_send_context(qp, sc))
208 			return -EINVAL;
209 	}
210 
211 	return 0;
212 }
213 
214 void hfi1_modify_qp(struct rvt_qp *qp, struct ib_qp_attr *attr,
215 		    int attr_mask, struct ib_udata *udata)
216 {
217 	struct ib_qp *ibqp = &qp->ibqp;
218 	struct hfi1_qp_priv *priv = qp->priv;
219 
220 	if (attr_mask & IB_QP_AV) {
221 		priv->s_sc = ah_to_sc(ibqp->device, &qp->remote_ah_attr);
222 		priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
223 		priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
224 	}
225 
226 	if (attr_mask & IB_QP_PATH_MIG_STATE &&
227 	    attr->path_mig_state == IB_MIG_MIGRATED &&
228 	    qp->s_mig_state == IB_MIG_ARMED) {
229 		qp->s_flags |= RVT_S_AHG_CLEAR;
230 		priv->s_sc = ah_to_sc(ibqp->device, &qp->remote_ah_attr);
231 		priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
232 		priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
233 	}
234 }
235 
236 /**
237  * hfi1_check_send_wqe - validate wqe
238  * @qp - The qp
239  * @wqe - The built wqe
240  *
241  * validate wqe.  This is called
242  * prior to inserting the wqe into
243  * the ring but after the wqe has been
244  * setup.
245  *
246  * Returns 0 on success, -EINVAL on failure
247  *
248  */
249 int hfi1_check_send_wqe(struct rvt_qp *qp,
250 			struct rvt_swqe *wqe)
251 {
252 	struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
253 	struct rvt_ah *ah;
254 
255 	switch (qp->ibqp.qp_type) {
256 	case IB_QPT_RC:
257 	case IB_QPT_UC:
258 		if (wqe->length > 0x80000000U)
259 			return -EINVAL;
260 		break;
261 	case IB_QPT_SMI:
262 		ah = ibah_to_rvtah(wqe->ud_wr.ah);
263 		if (wqe->length > (1 << ah->log_pmtu))
264 			return -EINVAL;
265 		break;
266 	case IB_QPT_GSI:
267 	case IB_QPT_UD:
268 		ah = ibah_to_rvtah(wqe->ud_wr.ah);
269 		if (wqe->length > (1 << ah->log_pmtu))
270 			return -EINVAL;
271 		if (ibp->sl_to_sc[ah->attr.sl] == 0xf)
272 			return -EINVAL;
273 	default:
274 		break;
275 	}
276 	return wqe->length <= piothreshold;
277 }
278 
279 /**
280  * hfi1_compute_aeth - compute the AETH (syndrome + MSN)
281  * @qp: the queue pair to compute the AETH for
282  *
283  * Returns the AETH.
284  */
285 __be32 hfi1_compute_aeth(struct rvt_qp *qp)
286 {
287 	u32 aeth = qp->r_msn & HFI1_MSN_MASK;
288 
289 	if (qp->ibqp.srq) {
290 		/*
291 		 * Shared receive queues don't generate credits.
292 		 * Set the credit field to the invalid value.
293 		 */
294 		aeth |= HFI1_AETH_CREDIT_INVAL << HFI1_AETH_CREDIT_SHIFT;
295 	} else {
296 		u32 min, max, x;
297 		u32 credits;
298 		struct rvt_rwq *wq = qp->r_rq.wq;
299 		u32 head;
300 		u32 tail;
301 
302 		/* sanity check pointers before trusting them */
303 		head = wq->head;
304 		if (head >= qp->r_rq.size)
305 			head = 0;
306 		tail = wq->tail;
307 		if (tail >= qp->r_rq.size)
308 			tail = 0;
309 		/*
310 		 * Compute the number of credits available (RWQEs).
311 		 * There is a small chance that the pair of reads are
312 		 * not atomic, which is OK, since the fuzziness is
313 		 * resolved as further ACKs go out.
314 		 */
315 		credits = head - tail;
316 		if ((int)credits < 0)
317 			credits += qp->r_rq.size;
318 		/*
319 		 * Binary search the credit table to find the code to
320 		 * use.
321 		 */
322 		min = 0;
323 		max = 31;
324 		for (;;) {
325 			x = (min + max) / 2;
326 			if (credit_table[x] == credits)
327 				break;
328 			if (credit_table[x] > credits) {
329 				max = x;
330 			} else {
331 				if (min == x)
332 					break;
333 				min = x;
334 			}
335 		}
336 		aeth |= x << HFI1_AETH_CREDIT_SHIFT;
337 	}
338 	return cpu_to_be32(aeth);
339 }
340 
341 /**
342  * _hfi1_schedule_send - schedule progress
343  * @qp: the QP
344  *
345  * This schedules qp progress w/o regard to the s_flags.
346  *
347  * It is only used in the post send, which doesn't hold
348  * the s_lock.
349  */
350 void _hfi1_schedule_send(struct rvt_qp *qp)
351 {
352 	struct hfi1_qp_priv *priv = qp->priv;
353 	struct hfi1_ibport *ibp =
354 		to_iport(qp->ibqp.device, qp->port_num);
355 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
356 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
357 
358 	iowait_schedule(&priv->s_iowait, ppd->hfi1_wq,
359 			priv->s_sde ?
360 			priv->s_sde->cpu :
361 			cpumask_first(cpumask_of_node(dd->node)));
362 }
363 
364 static void qp_pio_drain(struct rvt_qp *qp)
365 {
366 	struct hfi1_ibdev *dev;
367 	struct hfi1_qp_priv *priv = qp->priv;
368 
369 	if (!priv->s_sendcontext)
370 		return;
371 	dev = to_idev(qp->ibqp.device);
372 	while (iowait_pio_pending(&priv->s_iowait)) {
373 		write_seqlock_irq(&dev->iowait_lock);
374 		hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 1);
375 		write_sequnlock_irq(&dev->iowait_lock);
376 		iowait_pio_drain(&priv->s_iowait);
377 		write_seqlock_irq(&dev->iowait_lock);
378 		hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 0);
379 		write_sequnlock_irq(&dev->iowait_lock);
380 	}
381 }
382 
383 /**
384  * hfi1_schedule_send - schedule progress
385  * @qp: the QP
386  *
387  * This schedules qp progress and caller should hold
388  * the s_lock.
389  */
390 void hfi1_schedule_send(struct rvt_qp *qp)
391 {
392 	if (hfi1_send_ok(qp))
393 		_hfi1_schedule_send(qp);
394 }
395 
396 /**
397  * hfi1_get_credit - flush the send work queue of a QP
398  * @qp: the qp who's send work queue to flush
399  * @aeth: the Acknowledge Extended Transport Header
400  *
401  * The QP s_lock should be held.
402  */
403 void hfi1_get_credit(struct rvt_qp *qp, u32 aeth)
404 {
405 	u32 credit = (aeth >> HFI1_AETH_CREDIT_SHIFT) & HFI1_AETH_CREDIT_MASK;
406 
407 	/*
408 	 * If the credit is invalid, we can send
409 	 * as many packets as we like.  Otherwise, we have to
410 	 * honor the credit field.
411 	 */
412 	if (credit == HFI1_AETH_CREDIT_INVAL) {
413 		if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT)) {
414 			qp->s_flags |= RVT_S_UNLIMITED_CREDIT;
415 			if (qp->s_flags & RVT_S_WAIT_SSN_CREDIT) {
416 				qp->s_flags &= ~RVT_S_WAIT_SSN_CREDIT;
417 				hfi1_schedule_send(qp);
418 			}
419 		}
420 	} else if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT)) {
421 		/* Compute new LSN (i.e., MSN + credit) */
422 		credit = (aeth + credit_table[credit]) & HFI1_MSN_MASK;
423 		if (cmp_msn(credit, qp->s_lsn) > 0) {
424 			qp->s_lsn = credit;
425 			if (qp->s_flags & RVT_S_WAIT_SSN_CREDIT) {
426 				qp->s_flags &= ~RVT_S_WAIT_SSN_CREDIT;
427 				hfi1_schedule_send(qp);
428 			}
429 		}
430 	}
431 }
432 
433 void hfi1_qp_wakeup(struct rvt_qp *qp, u32 flag)
434 {
435 	unsigned long flags;
436 
437 	spin_lock_irqsave(&qp->s_lock, flags);
438 	if (qp->s_flags & flag) {
439 		qp->s_flags &= ~flag;
440 		trace_hfi1_qpwakeup(qp, flag);
441 		hfi1_schedule_send(qp);
442 	}
443 	spin_unlock_irqrestore(&qp->s_lock, flags);
444 	/* Notify hfi1_destroy_qp() if it is waiting. */
445 	if (atomic_dec_and_test(&qp->refcount))
446 		wake_up(&qp->wait);
447 }
448 
449 static int iowait_sleep(
450 	struct sdma_engine *sde,
451 	struct iowait *wait,
452 	struct sdma_txreq *stx,
453 	unsigned seq)
454 {
455 	struct verbs_txreq *tx = container_of(stx, struct verbs_txreq, txreq);
456 	struct rvt_qp *qp;
457 	struct hfi1_qp_priv *priv;
458 	unsigned long flags;
459 	int ret = 0;
460 	struct hfi1_ibdev *dev;
461 
462 	qp = tx->qp;
463 	priv = qp->priv;
464 
465 	spin_lock_irqsave(&qp->s_lock, flags);
466 	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
467 		/*
468 		 * If we couldn't queue the DMA request, save the info
469 		 * and try again later rather than destroying the
470 		 * buffer and undoing the side effects of the copy.
471 		 */
472 		/* Make a common routine? */
473 		dev = &sde->dd->verbs_dev;
474 		list_add_tail(&stx->list, &wait->tx_head);
475 		write_seqlock(&dev->iowait_lock);
476 		if (sdma_progress(sde, seq, stx))
477 			goto eagain;
478 		if (list_empty(&priv->s_iowait.list)) {
479 			struct hfi1_ibport *ibp =
480 				to_iport(qp->ibqp.device, qp->port_num);
481 
482 			ibp->rvp.n_dmawait++;
483 			qp->s_flags |= RVT_S_WAIT_DMA_DESC;
484 			list_add_tail(&priv->s_iowait.list, &sde->dmawait);
485 			trace_hfi1_qpsleep(qp, RVT_S_WAIT_DMA_DESC);
486 			atomic_inc(&qp->refcount);
487 		}
488 		write_sequnlock(&dev->iowait_lock);
489 		qp->s_flags &= ~RVT_S_BUSY;
490 		spin_unlock_irqrestore(&qp->s_lock, flags);
491 		ret = -EBUSY;
492 	} else {
493 		spin_unlock_irqrestore(&qp->s_lock, flags);
494 		hfi1_put_txreq(tx);
495 	}
496 	return ret;
497 eagain:
498 	write_sequnlock(&dev->iowait_lock);
499 	spin_unlock_irqrestore(&qp->s_lock, flags);
500 	list_del_init(&stx->list);
501 	return -EAGAIN;
502 }
503 
504 static void iowait_wakeup(struct iowait *wait, int reason)
505 {
506 	struct rvt_qp *qp = iowait_to_qp(wait);
507 
508 	WARN_ON(reason != SDMA_AVAIL_REASON);
509 	hfi1_qp_wakeup(qp, RVT_S_WAIT_DMA_DESC);
510 }
511 
512 static void iowait_sdma_drained(struct iowait *wait)
513 {
514 	struct rvt_qp *qp = iowait_to_qp(wait);
515 	unsigned long flags;
516 
517 	/*
518 	 * This happens when the send engine notes
519 	 * a QP in the error state and cannot
520 	 * do the flush work until that QP's
521 	 * sdma work has finished.
522 	 */
523 	spin_lock_irqsave(&qp->s_lock, flags);
524 	if (qp->s_flags & RVT_S_WAIT_DMA) {
525 		qp->s_flags &= ~RVT_S_WAIT_DMA;
526 		hfi1_schedule_send(qp);
527 	}
528 	spin_unlock_irqrestore(&qp->s_lock, flags);
529 }
530 
531 /**
532  *
533  * qp_to_sdma_engine - map a qp to a send engine
534  * @qp: the QP
535  * @sc5: the 5 bit sc
536  *
537  * Return:
538  * A send engine for the qp or NULL for SMI type qp.
539  */
540 struct sdma_engine *qp_to_sdma_engine(struct rvt_qp *qp, u8 sc5)
541 {
542 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
543 	struct sdma_engine *sde;
544 
545 	if (!(dd->flags & HFI1_HAS_SEND_DMA))
546 		return NULL;
547 	switch (qp->ibqp.qp_type) {
548 	case IB_QPT_SMI:
549 		return NULL;
550 	default:
551 		break;
552 	}
553 	sde = sdma_select_engine_sc(dd, qp->ibqp.qp_num >> dd->qos_shift, sc5);
554 	return sde;
555 }
556 
557 /*
558  * qp_to_send_context - map a qp to a send context
559  * @qp: the QP
560  * @sc5: the 5 bit sc
561  *
562  * Return:
563  * A send context for the qp
564  */
565 struct send_context *qp_to_send_context(struct rvt_qp *qp, u8 sc5)
566 {
567 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
568 
569 	switch (qp->ibqp.qp_type) {
570 	case IB_QPT_SMI:
571 		/* SMA packets to VL15 */
572 		return dd->vld[15].sc;
573 	default:
574 		break;
575 	}
576 
577 	return pio_select_send_context_sc(dd, qp->ibqp.qp_num >> dd->qos_shift,
578 					  sc5);
579 }
580 
581 struct qp_iter {
582 	struct hfi1_ibdev *dev;
583 	struct rvt_qp *qp;
584 	int specials;
585 	int n;
586 };
587 
588 struct qp_iter *qp_iter_init(struct hfi1_ibdev *dev)
589 {
590 	struct qp_iter *iter;
591 
592 	iter = kzalloc(sizeof(*iter), GFP_KERNEL);
593 	if (!iter)
594 		return NULL;
595 
596 	iter->dev = dev;
597 	iter->specials = dev->rdi.ibdev.phys_port_cnt * 2;
598 	if (qp_iter_next(iter)) {
599 		kfree(iter);
600 		return NULL;
601 	}
602 
603 	return iter;
604 }
605 
606 int qp_iter_next(struct qp_iter *iter)
607 {
608 	struct hfi1_ibdev *dev = iter->dev;
609 	int n = iter->n;
610 	int ret = 1;
611 	struct rvt_qp *pqp = iter->qp;
612 	struct rvt_qp *qp;
613 
614 	/*
615 	 * The approach is to consider the special qps
616 	 * as an additional table entries before the
617 	 * real hash table.  Since the qp code sets
618 	 * the qp->next hash link to NULL, this works just fine.
619 	 *
620 	 * iter->specials is 2 * # ports
621 	 *
622 	 * n = 0..iter->specials is the special qp indices
623 	 *
624 	 * n = iter->specials..dev->rdi.qp_dev->qp_table_size+iter->specials are
625 	 * the potential hash bucket entries
626 	 *
627 	 */
628 	for (; n <  dev->rdi.qp_dev->qp_table_size + iter->specials; n++) {
629 		if (pqp) {
630 			qp = rcu_dereference(pqp->next);
631 		} else {
632 			if (n < iter->specials) {
633 				struct hfi1_pportdata *ppd;
634 				struct hfi1_ibport *ibp;
635 				int pidx;
636 
637 				pidx = n % dev->rdi.ibdev.phys_port_cnt;
638 				ppd = &dd_from_dev(dev)->pport[pidx];
639 				ibp = &ppd->ibport_data;
640 
641 				if (!(n & 1))
642 					qp = rcu_dereference(ibp->rvp.qp[0]);
643 				else
644 					qp = rcu_dereference(ibp->rvp.qp[1]);
645 			} else {
646 				qp = rcu_dereference(
647 					dev->rdi.qp_dev->qp_table[
648 						(n - iter->specials)]);
649 			}
650 		}
651 		pqp = qp;
652 		if (qp) {
653 			iter->qp = qp;
654 			iter->n = n;
655 			return 0;
656 		}
657 	}
658 	return ret;
659 }
660 
661 static const char * const qp_type_str[] = {
662 	"SMI", "GSI", "RC", "UC", "UD",
663 };
664 
665 static int qp_idle(struct rvt_qp *qp)
666 {
667 	return
668 		qp->s_last == qp->s_acked &&
669 		qp->s_acked == qp->s_cur &&
670 		qp->s_cur == qp->s_tail &&
671 		qp->s_tail == qp->s_head;
672 }
673 
674 void qp_iter_print(struct seq_file *s, struct qp_iter *iter)
675 {
676 	struct rvt_swqe *wqe;
677 	struct rvt_qp *qp = iter->qp;
678 	struct hfi1_qp_priv *priv = qp->priv;
679 	struct sdma_engine *sde;
680 	struct send_context *send_context;
681 
682 	sde = qp_to_sdma_engine(qp, priv->s_sc);
683 	wqe = rvt_get_swqe_ptr(qp, qp->s_last);
684 	send_context = qp_to_send_context(qp, priv->s_sc);
685 	seq_printf(s,
686 		   "N %d %s QP %x R %u %s %u %u %u f=%x %u %u %u %u %u %u PSN %x %x %x %x %x (%u %u %u %u %u %u %u) RQP %x LID %x SL %u MTU %u %u %u %u SDE %p,%u SC %p,%u SCQ %u %u PID %d\n",
687 		   iter->n,
688 		   qp_idle(qp) ? "I" : "B",
689 		   qp->ibqp.qp_num,
690 		   atomic_read(&qp->refcount),
691 		   qp_type_str[qp->ibqp.qp_type],
692 		   qp->state,
693 		   wqe ? wqe->wr.opcode : 0,
694 		   qp->s_hdrwords,
695 		   qp->s_flags,
696 		   iowait_sdma_pending(&priv->s_iowait),
697 		   iowait_pio_pending(&priv->s_iowait),
698 		   !list_empty(&priv->s_iowait.list),
699 		   qp->timeout,
700 		   wqe ? wqe->ssn : 0,
701 		   qp->s_lsn,
702 		   qp->s_last_psn,
703 		   qp->s_psn, qp->s_next_psn,
704 		   qp->s_sending_psn, qp->s_sending_hpsn,
705 		   qp->s_last, qp->s_acked, qp->s_cur,
706 		   qp->s_tail, qp->s_head, qp->s_size,
707 		   qp->s_avail,
708 		   qp->remote_qpn,
709 		   qp->remote_ah_attr.dlid,
710 		   qp->remote_ah_attr.sl,
711 		   qp->pmtu,
712 		   qp->s_retry,
713 		   qp->s_retry_cnt,
714 		   qp->s_rnr_retry_cnt,
715 		   sde,
716 		   sde ? sde->this_idx : 0,
717 		   send_context,
718 		   send_context ? send_context->sw_index : 0,
719 		   ibcq_to_rvtcq(qp->ibqp.send_cq)->queue->head,
720 		   ibcq_to_rvtcq(qp->ibqp.send_cq)->queue->tail,
721 		   qp->pid);
722 }
723 
724 void qp_comm_est(struct rvt_qp *qp)
725 {
726 	qp->r_flags |= RVT_R_COMM_EST;
727 	if (qp->ibqp.event_handler) {
728 		struct ib_event ev;
729 
730 		ev.device = qp->ibqp.device;
731 		ev.element.qp = &qp->ibqp;
732 		ev.event = IB_EVENT_COMM_EST;
733 		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
734 	}
735 }
736 
737 void *qp_priv_alloc(struct rvt_dev_info *rdi, struct rvt_qp *qp,
738 		    gfp_t gfp)
739 {
740 	struct hfi1_qp_priv *priv;
741 
742 	priv = kzalloc_node(sizeof(*priv), gfp, rdi->dparms.node);
743 	if (!priv)
744 		return ERR_PTR(-ENOMEM);
745 
746 	priv->owner = qp;
747 
748 	priv->s_hdr = kzalloc_node(sizeof(*priv->s_hdr), gfp, rdi->dparms.node);
749 	if (!priv->s_hdr) {
750 		kfree(priv);
751 		return ERR_PTR(-ENOMEM);
752 	}
753 	setup_timer(&priv->s_rnr_timer, hfi1_rc_rnr_retry, (unsigned long)qp);
754 	qp->s_timer.function = hfi1_rc_timeout;
755 	return priv;
756 }
757 
758 void qp_priv_free(struct rvt_dev_info *rdi, struct rvt_qp *qp)
759 {
760 	struct hfi1_qp_priv *priv = qp->priv;
761 
762 	kfree(priv->s_hdr);
763 	kfree(priv);
764 }
765 
766 unsigned free_all_qps(struct rvt_dev_info *rdi)
767 {
768 	struct hfi1_ibdev *verbs_dev = container_of(rdi,
769 						    struct hfi1_ibdev,
770 						    rdi);
771 	struct hfi1_devdata *dd = container_of(verbs_dev,
772 					       struct hfi1_devdata,
773 					       verbs_dev);
774 	int n;
775 	unsigned qp_inuse = 0;
776 
777 	for (n = 0; n < dd->num_pports; n++) {
778 		struct hfi1_ibport *ibp = &dd->pport[n].ibport_data;
779 
780 		rcu_read_lock();
781 		if (rcu_dereference(ibp->rvp.qp[0]))
782 			qp_inuse++;
783 		if (rcu_dereference(ibp->rvp.qp[1]))
784 			qp_inuse++;
785 		rcu_read_unlock();
786 	}
787 
788 	return qp_inuse;
789 }
790 
791 void flush_qp_waiters(struct rvt_qp *qp)
792 {
793 	flush_iowait(qp);
794 	hfi1_stop_rc_timers(qp);
795 }
796 
797 void stop_send_queue(struct rvt_qp *qp)
798 {
799 	struct hfi1_qp_priv *priv = qp->priv;
800 
801 	cancel_work_sync(&priv->s_iowait.iowork);
802 	hfi1_del_timers_sync(qp);
803 }
804 
805 void quiesce_qp(struct rvt_qp *qp)
806 {
807 	struct hfi1_qp_priv *priv = qp->priv;
808 
809 	iowait_sdma_drain(&priv->s_iowait);
810 	qp_pio_drain(qp);
811 	flush_tx_list(qp);
812 }
813 
814 void notify_qp_reset(struct rvt_qp *qp)
815 {
816 	struct hfi1_qp_priv *priv = qp->priv;
817 
818 	iowait_init(
819 		&priv->s_iowait,
820 		1,
821 		_hfi1_do_send,
822 		iowait_sleep,
823 		iowait_wakeup,
824 		iowait_sdma_drained);
825 	priv->r_adefered = 0;
826 	clear_ahg(qp);
827 }
828 
829 /*
830  * Switch to alternate path.
831  * The QP s_lock should be held and interrupts disabled.
832  */
833 void hfi1_migrate_qp(struct rvt_qp *qp)
834 {
835 	struct hfi1_qp_priv *priv = qp->priv;
836 	struct ib_event ev;
837 
838 	qp->s_mig_state = IB_MIG_MIGRATED;
839 	qp->remote_ah_attr = qp->alt_ah_attr;
840 	qp->port_num = qp->alt_ah_attr.port_num;
841 	qp->s_pkey_index = qp->s_alt_pkey_index;
842 	qp->s_flags |= RVT_S_AHG_CLEAR;
843 	priv->s_sc = ah_to_sc(qp->ibqp.device, &qp->remote_ah_attr);
844 	priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
845 
846 	ev.device = qp->ibqp.device;
847 	ev.element.qp = &qp->ibqp;
848 	ev.event = IB_EVENT_PATH_MIG;
849 	qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
850 }
851 
852 int mtu_to_path_mtu(u32 mtu)
853 {
854 	return mtu_to_enum(mtu, OPA_MTU_8192);
855 }
856 
857 u32 mtu_from_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp, u32 pmtu)
858 {
859 	u32 mtu;
860 	struct hfi1_ibdev *verbs_dev = container_of(rdi,
861 						    struct hfi1_ibdev,
862 						    rdi);
863 	struct hfi1_devdata *dd = container_of(verbs_dev,
864 					       struct hfi1_devdata,
865 					       verbs_dev);
866 	struct hfi1_ibport *ibp;
867 	u8 sc, vl;
868 
869 	ibp = &dd->pport[qp->port_num - 1].ibport_data;
870 	sc = ibp->sl_to_sc[qp->remote_ah_attr.sl];
871 	vl = sc_to_vlt(dd, sc);
872 
873 	mtu = verbs_mtu_enum_to_int(qp->ibqp.device, pmtu);
874 	if (vl < PER_VL_SEND_CONTEXTS)
875 		mtu = min_t(u32, mtu, dd->vld[vl].mtu);
876 	return mtu;
877 }
878 
879 int get_pmtu_from_attr(struct rvt_dev_info *rdi, struct rvt_qp *qp,
880 		       struct ib_qp_attr *attr)
881 {
882 	int mtu, pidx = qp->port_num - 1;
883 	struct hfi1_ibdev *verbs_dev = container_of(rdi,
884 						    struct hfi1_ibdev,
885 						    rdi);
886 	struct hfi1_devdata *dd = container_of(verbs_dev,
887 					       struct hfi1_devdata,
888 					       verbs_dev);
889 	mtu = verbs_mtu_enum_to_int(qp->ibqp.device, attr->path_mtu);
890 	if (mtu == -1)
891 		return -1; /* values less than 0 are error */
892 
893 	if (mtu > dd->pport[pidx].ibmtu)
894 		return mtu_to_enum(dd->pport[pidx].ibmtu, IB_MTU_2048);
895 	else
896 		return attr->path_mtu;
897 }
898 
899 void notify_error_qp(struct rvt_qp *qp)
900 {
901 	struct hfi1_ibdev *dev = to_idev(qp->ibqp.device);
902 	struct hfi1_qp_priv *priv = qp->priv;
903 
904 	write_seqlock(&dev->iowait_lock);
905 	if (!list_empty(&priv->s_iowait.list) && !(qp->s_flags & RVT_S_BUSY)) {
906 		qp->s_flags &= ~RVT_S_ANY_WAIT_IO;
907 		list_del_init(&priv->s_iowait.list);
908 		if (atomic_dec_and_test(&qp->refcount))
909 			wake_up(&qp->wait);
910 	}
911 	write_sequnlock(&dev->iowait_lock);
912 
913 	if (!(qp->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_error_port_qps - put a port's RC/UC qps into error state
925  * @ibp: the ibport.
926  * @sl: the service level.
927  *
928  * This function places all RC/UC qps with a given service level into error
929  * state. It is generally called to force upper lay apps to abandon stale qps
930  * after an sl->sc mapping change.
931  */
932 void hfi1_error_port_qps(struct hfi1_ibport *ibp, u8 sl)
933 {
934 	struct rvt_qp *qp = NULL;
935 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
936 	struct hfi1_ibdev *dev = &ppd->dd->verbs_dev;
937 	int n;
938 	int lastwqe;
939 	struct ib_event ev;
940 
941 	rcu_read_lock();
942 
943 	/* Deal only with RC/UC qps that use the given SL. */
944 	for (n = 0; n < dev->rdi.qp_dev->qp_table_size; n++) {
945 		for (qp = rcu_dereference(dev->rdi.qp_dev->qp_table[n]); qp;
946 			qp = rcu_dereference(qp->next)) {
947 			if (qp->port_num == ppd->port &&
948 			    (qp->ibqp.qp_type == IB_QPT_UC ||
949 			     qp->ibqp.qp_type == IB_QPT_RC) &&
950 			    qp->remote_ah_attr.sl == sl &&
951 			    (ib_rvt_state_ops[qp->state] &
952 			     RVT_POST_SEND_OK)) {
953 				spin_lock_irq(&qp->r_lock);
954 				spin_lock(&qp->s_hlock);
955 				spin_lock(&qp->s_lock);
956 				lastwqe = rvt_error_qp(qp,
957 						       IB_WC_WR_FLUSH_ERR);
958 				spin_unlock(&qp->s_lock);
959 				spin_unlock(&qp->s_hlock);
960 				spin_unlock_irq(&qp->r_lock);
961 				if (lastwqe) {
962 					ev.device = qp->ibqp.device;
963 					ev.element.qp = &qp->ibqp;
964 					ev.event =
965 						IB_EVENT_QP_LAST_WQE_REACHED;
966 					qp->ibqp.event_handler(&ev,
967 						qp->ibqp.qp_context);
968 				}
969 			}
970 		}
971 	}
972 
973 	rcu_read_unlock();
974 }
975