xref: /linux/drivers/infiniband/hw/hfi1/qp.c (revision 32786fdc9506aeba98278c1844d4bfb766863832)
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 #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 seq);
72 static void iowait_wakeup(struct iowait *wait, int reason);
73 static void iowait_sdma_drained(struct iowait *wait);
74 static void qp_pio_drain(struct rvt_qp *qp);
75 
76 static inline unsigned mk_qpn(struct rvt_qpn_table *qpt,
77 			      struct rvt_qpn_map *map, unsigned off)
78 {
79 	return (map - qpt->map) * RVT_BITS_PER_PAGE + off;
80 }
81 
82 /*
83  * Convert the AETH credit code into the number of credits.
84  */
85 static const u16 credit_table[31] = {
86 	0,                      /* 0 */
87 	1,                      /* 1 */
88 	2,                      /* 2 */
89 	3,                      /* 3 */
90 	4,                      /* 4 */
91 	6,                      /* 5 */
92 	8,                      /* 6 */
93 	12,                     /* 7 */
94 	16,                     /* 8 */
95 	24,                     /* 9 */
96 	32,                     /* A */
97 	48,                     /* B */
98 	64,                     /* C */
99 	96,                     /* D */
100 	128,                    /* E */
101 	192,                    /* F */
102 	256,                    /* 10 */
103 	384,                    /* 11 */
104 	512,                    /* 12 */
105 	768,                    /* 13 */
106 	1024,                   /* 14 */
107 	1536,                   /* 15 */
108 	2048,                   /* 16 */
109 	3072,                   /* 17 */
110 	4096,                   /* 18 */
111 	6144,                   /* 19 */
112 	8192,                   /* 1A */
113 	12288,                  /* 1B */
114 	16384,                  /* 1C */
115 	24576,                  /* 1D */
116 	32768                   /* 1E */
117 };
118 
119 const struct rvt_operation_params hfi1_post_parms[RVT_OPERATION_MAX] = {
120 [IB_WR_RDMA_WRITE] = {
121 	.length = sizeof(struct ib_rdma_wr),
122 	.qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
123 },
124 
125 [IB_WR_RDMA_READ] = {
126 	.length = sizeof(struct ib_rdma_wr),
127 	.qpt_support = BIT(IB_QPT_RC),
128 	.flags = RVT_OPERATION_ATOMIC,
129 },
130 
131 [IB_WR_ATOMIC_CMP_AND_SWP] = {
132 	.length = sizeof(struct ib_atomic_wr),
133 	.qpt_support = BIT(IB_QPT_RC),
134 	.flags = RVT_OPERATION_ATOMIC | RVT_OPERATION_ATOMIC_SGE,
135 },
136 
137 [IB_WR_ATOMIC_FETCH_AND_ADD] = {
138 	.length = sizeof(struct ib_atomic_wr),
139 	.qpt_support = BIT(IB_QPT_RC),
140 	.flags = RVT_OPERATION_ATOMIC | RVT_OPERATION_ATOMIC_SGE,
141 },
142 
143 [IB_WR_RDMA_WRITE_WITH_IMM] = {
144 	.length = sizeof(struct ib_rdma_wr),
145 	.qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
146 },
147 
148 [IB_WR_SEND] = {
149 	.length = sizeof(struct ib_send_wr),
150 	.qpt_support = BIT(IB_QPT_UD) | BIT(IB_QPT_SMI) | BIT(IB_QPT_GSI) |
151 		       BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
152 },
153 
154 [IB_WR_SEND_WITH_IMM] = {
155 	.length = sizeof(struct ib_send_wr),
156 	.qpt_support = BIT(IB_QPT_UD) | BIT(IB_QPT_SMI) | BIT(IB_QPT_GSI) |
157 		       BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
158 },
159 
160 [IB_WR_REG_MR] = {
161 	.length = sizeof(struct ib_reg_wr),
162 	.qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
163 	.flags = RVT_OPERATION_LOCAL,
164 },
165 
166 [IB_WR_LOCAL_INV] = {
167 	.length = sizeof(struct ib_send_wr),
168 	.qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
169 	.flags = RVT_OPERATION_LOCAL,
170 },
171 
172 [IB_WR_SEND_WITH_INV] = {
173 	.length = sizeof(struct ib_send_wr),
174 	.qpt_support = BIT(IB_QPT_RC),
175 },
176 
177 };
178 
179 static void flush_tx_list(struct rvt_qp *qp)
180 {
181 	struct hfi1_qp_priv *priv = qp->priv;
182 
183 	while (!list_empty(&priv->s_iowait.tx_head)) {
184 		struct sdma_txreq *tx;
185 
186 		tx = list_first_entry(
187 			&priv->s_iowait.tx_head,
188 			struct sdma_txreq,
189 			list);
190 		list_del_init(&tx->list);
191 		hfi1_put_txreq(
192 			container_of(tx, struct verbs_txreq, txreq));
193 	}
194 }
195 
196 static void flush_iowait(struct rvt_qp *qp)
197 {
198 	struct hfi1_qp_priv *priv = qp->priv;
199 	unsigned long flags;
200 	seqlock_t *lock = priv->s_iowait.lock;
201 
202 	if (!lock)
203 		return;
204 	write_seqlock_irqsave(lock, flags);
205 	if (!list_empty(&priv->s_iowait.list)) {
206 		list_del_init(&priv->s_iowait.list);
207 		priv->s_iowait.lock = NULL;
208 		rvt_put_qp(qp);
209 	}
210 	write_sequnlock_irqrestore(lock, flags);
211 }
212 
213 static inline int opa_mtu_enum_to_int(int mtu)
214 {
215 	switch (mtu) {
216 	case OPA_MTU_8192:  return 8192;
217 	case OPA_MTU_10240: return 10240;
218 	default:            return -1;
219 	}
220 }
221 
222 /**
223  * This function is what we would push to the core layer if we wanted to be a
224  * "first class citizen".  Instead we hide this here and rely on Verbs ULPs
225  * to blindly pass the MTU enum value from the PathRecord to us.
226  */
227 static inline int verbs_mtu_enum_to_int(struct ib_device *dev, enum ib_mtu mtu)
228 {
229 	int val;
230 
231 	/* Constraining 10KB packets to 8KB packets */
232 	if (mtu == (enum ib_mtu)OPA_MTU_10240)
233 		mtu = OPA_MTU_8192;
234 	val = opa_mtu_enum_to_int((int)mtu);
235 	if (val > 0)
236 		return val;
237 	return ib_mtu_enum_to_int(mtu);
238 }
239 
240 int hfi1_check_modify_qp(struct rvt_qp *qp, struct ib_qp_attr *attr,
241 			 int attr_mask, struct ib_udata *udata)
242 {
243 	struct ib_qp *ibqp = &qp->ibqp;
244 	struct hfi1_ibdev *dev = to_idev(ibqp->device);
245 	struct hfi1_devdata *dd = dd_from_dev(dev);
246 	u8 sc;
247 
248 	if (attr_mask & IB_QP_AV) {
249 		sc = ah_to_sc(ibqp->device, &attr->ah_attr);
250 		if (sc == 0xf)
251 			return -EINVAL;
252 
253 		if (!qp_to_sdma_engine(qp, sc) &&
254 		    dd->flags & HFI1_HAS_SEND_DMA)
255 			return -EINVAL;
256 
257 		if (!qp_to_send_context(qp, sc))
258 			return -EINVAL;
259 	}
260 
261 	if (attr_mask & IB_QP_ALT_PATH) {
262 		sc = ah_to_sc(ibqp->device, &attr->alt_ah_attr);
263 		if (sc == 0xf)
264 			return -EINVAL;
265 
266 		if (!qp_to_sdma_engine(qp, sc) &&
267 		    dd->flags & HFI1_HAS_SEND_DMA)
268 			return -EINVAL;
269 
270 		if (!qp_to_send_context(qp, sc))
271 			return -EINVAL;
272 	}
273 
274 	return 0;
275 }
276 
277 void hfi1_modify_qp(struct rvt_qp *qp, struct ib_qp_attr *attr,
278 		    int attr_mask, struct ib_udata *udata)
279 {
280 	struct ib_qp *ibqp = &qp->ibqp;
281 	struct hfi1_qp_priv *priv = qp->priv;
282 
283 	if (attr_mask & IB_QP_AV) {
284 		priv->s_sc = ah_to_sc(ibqp->device, &qp->remote_ah_attr);
285 		priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
286 		priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
287 	}
288 
289 	if (attr_mask & IB_QP_PATH_MIG_STATE &&
290 	    attr->path_mig_state == IB_MIG_MIGRATED &&
291 	    qp->s_mig_state == IB_MIG_ARMED) {
292 		qp->s_flags |= RVT_S_AHG_CLEAR;
293 		priv->s_sc = ah_to_sc(ibqp->device, &qp->remote_ah_attr);
294 		priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
295 		priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
296 	}
297 }
298 
299 /**
300  * hfi1_check_send_wqe - validate wqe
301  * @qp - The qp
302  * @wqe - The built wqe
303  *
304  * validate wqe.  This is called
305  * prior to inserting the wqe into
306  * the ring but after the wqe has been
307  * setup.
308  *
309  * Returns 0 on success, -EINVAL on failure
310  *
311  */
312 int hfi1_check_send_wqe(struct rvt_qp *qp,
313 			struct rvt_swqe *wqe)
314 {
315 	struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
316 	struct rvt_ah *ah;
317 
318 	switch (qp->ibqp.qp_type) {
319 	case IB_QPT_RC:
320 	case IB_QPT_UC:
321 		if (wqe->length > 0x80000000U)
322 			return -EINVAL;
323 		break;
324 	case IB_QPT_SMI:
325 		ah = ibah_to_rvtah(wqe->ud_wr.ah);
326 		if (wqe->length > (1 << ah->log_pmtu))
327 			return -EINVAL;
328 		break;
329 	case IB_QPT_GSI:
330 	case IB_QPT_UD:
331 		ah = ibah_to_rvtah(wqe->ud_wr.ah);
332 		if (wqe->length > (1 << ah->log_pmtu))
333 			return -EINVAL;
334 		if (ibp->sl_to_sc[ah->attr.sl] == 0xf)
335 			return -EINVAL;
336 	default:
337 		break;
338 	}
339 	return wqe->length <= piothreshold;
340 }
341 
342 /**
343  * hfi1_compute_aeth - compute the AETH (syndrome + MSN)
344  * @qp: the queue pair to compute the AETH for
345  *
346  * Returns the AETH.
347  */
348 __be32 hfi1_compute_aeth(struct rvt_qp *qp)
349 {
350 	u32 aeth = qp->r_msn & HFI1_MSN_MASK;
351 
352 	if (qp->ibqp.srq) {
353 		/*
354 		 * Shared receive queues don't generate credits.
355 		 * Set the credit field to the invalid value.
356 		 */
357 		aeth |= HFI1_AETH_CREDIT_INVAL << HFI1_AETH_CREDIT_SHIFT;
358 	} else {
359 		u32 min, max, x;
360 		u32 credits;
361 		struct rvt_rwq *wq = qp->r_rq.wq;
362 		u32 head;
363 		u32 tail;
364 
365 		/* sanity check pointers before trusting them */
366 		head = wq->head;
367 		if (head >= qp->r_rq.size)
368 			head = 0;
369 		tail = wq->tail;
370 		if (tail >= qp->r_rq.size)
371 			tail = 0;
372 		/*
373 		 * Compute the number of credits available (RWQEs).
374 		 * There is a small chance that the pair of reads are
375 		 * not atomic, which is OK, since the fuzziness is
376 		 * resolved as further ACKs go out.
377 		 */
378 		credits = head - tail;
379 		if ((int)credits < 0)
380 			credits += qp->r_rq.size;
381 		/*
382 		 * Binary search the credit table to find the code to
383 		 * use.
384 		 */
385 		min = 0;
386 		max = 31;
387 		for (;;) {
388 			x = (min + max) / 2;
389 			if (credit_table[x] == credits)
390 				break;
391 			if (credit_table[x] > credits) {
392 				max = x;
393 			} else {
394 				if (min == x)
395 					break;
396 				min = x;
397 			}
398 		}
399 		aeth |= x << HFI1_AETH_CREDIT_SHIFT;
400 	}
401 	return cpu_to_be32(aeth);
402 }
403 
404 /**
405  * _hfi1_schedule_send - schedule progress
406  * @qp: the QP
407  *
408  * This schedules qp progress w/o regard to the s_flags.
409  *
410  * It is only used in the post send, which doesn't hold
411  * the s_lock.
412  */
413 void _hfi1_schedule_send(struct rvt_qp *qp)
414 {
415 	struct hfi1_qp_priv *priv = qp->priv;
416 	struct hfi1_ibport *ibp =
417 		to_iport(qp->ibqp.device, qp->port_num);
418 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
419 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
420 
421 	iowait_schedule(&priv->s_iowait, ppd->hfi1_wq,
422 			priv->s_sde ?
423 			priv->s_sde->cpu :
424 			cpumask_first(cpumask_of_node(dd->node)));
425 }
426 
427 static void qp_pio_drain(struct rvt_qp *qp)
428 {
429 	struct hfi1_ibdev *dev;
430 	struct hfi1_qp_priv *priv = qp->priv;
431 
432 	if (!priv->s_sendcontext)
433 		return;
434 	dev = to_idev(qp->ibqp.device);
435 	while (iowait_pio_pending(&priv->s_iowait)) {
436 		write_seqlock_irq(&dev->iowait_lock);
437 		hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 1);
438 		write_sequnlock_irq(&dev->iowait_lock);
439 		iowait_pio_drain(&priv->s_iowait);
440 		write_seqlock_irq(&dev->iowait_lock);
441 		hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 0);
442 		write_sequnlock_irq(&dev->iowait_lock);
443 	}
444 }
445 
446 /**
447  * hfi1_schedule_send - schedule progress
448  * @qp: the QP
449  *
450  * This schedules qp progress and caller should hold
451  * the s_lock.
452  */
453 void hfi1_schedule_send(struct rvt_qp *qp)
454 {
455 	lockdep_assert_held(&qp->s_lock);
456 	if (hfi1_send_ok(qp))
457 		_hfi1_schedule_send(qp);
458 }
459 
460 /**
461  * hfi1_get_credit - handle credit in aeth
462  * @qp: the qp
463  * @aeth: the Acknowledge Extended Transport Header
464  *
465  * The QP s_lock should be held.
466  */
467 void hfi1_get_credit(struct rvt_qp *qp, u32 aeth)
468 {
469 	u32 credit = (aeth >> HFI1_AETH_CREDIT_SHIFT) & HFI1_AETH_CREDIT_MASK;
470 
471 	lockdep_assert_held(&qp->s_lock);
472 	/*
473 	 * If the credit is invalid, we can send
474 	 * as many packets as we like.  Otherwise, we have to
475 	 * honor the credit field.
476 	 */
477 	if (credit == HFI1_AETH_CREDIT_INVAL) {
478 		if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT)) {
479 			qp->s_flags |= RVT_S_UNLIMITED_CREDIT;
480 			if (qp->s_flags & RVT_S_WAIT_SSN_CREDIT) {
481 				qp->s_flags &= ~RVT_S_WAIT_SSN_CREDIT;
482 				hfi1_schedule_send(qp);
483 			}
484 		}
485 	} else if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT)) {
486 		/* Compute new LSN (i.e., MSN + credit) */
487 		credit = (aeth + credit_table[credit]) & HFI1_MSN_MASK;
488 		if (cmp_msn(credit, qp->s_lsn) > 0) {
489 			qp->s_lsn = credit;
490 			if (qp->s_flags & RVT_S_WAIT_SSN_CREDIT) {
491 				qp->s_flags &= ~RVT_S_WAIT_SSN_CREDIT;
492 				hfi1_schedule_send(qp);
493 			}
494 		}
495 	}
496 }
497 
498 void hfi1_qp_wakeup(struct rvt_qp *qp, u32 flag)
499 {
500 	unsigned long flags;
501 
502 	spin_lock_irqsave(&qp->s_lock, flags);
503 	if (qp->s_flags & flag) {
504 		qp->s_flags &= ~flag;
505 		trace_hfi1_qpwakeup(qp, flag);
506 		hfi1_schedule_send(qp);
507 	}
508 	spin_unlock_irqrestore(&qp->s_lock, flags);
509 	/* Notify hfi1_destroy_qp() if it is waiting. */
510 	rvt_put_qp(qp);
511 }
512 
513 static int iowait_sleep(
514 	struct sdma_engine *sde,
515 	struct iowait *wait,
516 	struct sdma_txreq *stx,
517 	unsigned seq)
518 {
519 	struct verbs_txreq *tx = container_of(stx, struct verbs_txreq, txreq);
520 	struct rvt_qp *qp;
521 	struct hfi1_qp_priv *priv;
522 	unsigned long flags;
523 	int ret = 0;
524 	struct hfi1_ibdev *dev;
525 
526 	qp = tx->qp;
527 	priv = qp->priv;
528 
529 	spin_lock_irqsave(&qp->s_lock, flags);
530 	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
531 		/*
532 		 * If we couldn't queue the DMA request, save the info
533 		 * and try again later rather than destroying the
534 		 * buffer and undoing the side effects of the copy.
535 		 */
536 		/* Make a common routine? */
537 		dev = &sde->dd->verbs_dev;
538 		list_add_tail(&stx->list, &wait->tx_head);
539 		write_seqlock(&dev->iowait_lock);
540 		if (sdma_progress(sde, seq, stx))
541 			goto eagain;
542 		if (list_empty(&priv->s_iowait.list)) {
543 			struct hfi1_ibport *ibp =
544 				to_iport(qp->ibqp.device, qp->port_num);
545 
546 			ibp->rvp.n_dmawait++;
547 			qp->s_flags |= RVT_S_WAIT_DMA_DESC;
548 			list_add_tail(&priv->s_iowait.list, &sde->dmawait);
549 			priv->s_iowait.lock = &dev->iowait_lock;
550 			trace_hfi1_qpsleep(qp, RVT_S_WAIT_DMA_DESC);
551 			rvt_get_qp(qp);
552 		}
553 		write_sequnlock(&dev->iowait_lock);
554 		qp->s_flags &= ~RVT_S_BUSY;
555 		spin_unlock_irqrestore(&qp->s_lock, flags);
556 		ret = -EBUSY;
557 	} else {
558 		spin_unlock_irqrestore(&qp->s_lock, flags);
559 		hfi1_put_txreq(tx);
560 	}
561 	return ret;
562 eagain:
563 	write_sequnlock(&dev->iowait_lock);
564 	spin_unlock_irqrestore(&qp->s_lock, flags);
565 	list_del_init(&stx->list);
566 	return -EAGAIN;
567 }
568 
569 static void iowait_wakeup(struct iowait *wait, int reason)
570 {
571 	struct rvt_qp *qp = iowait_to_qp(wait);
572 
573 	WARN_ON(reason != SDMA_AVAIL_REASON);
574 	hfi1_qp_wakeup(qp, RVT_S_WAIT_DMA_DESC);
575 }
576 
577 static void iowait_sdma_drained(struct iowait *wait)
578 {
579 	struct rvt_qp *qp = iowait_to_qp(wait);
580 	unsigned long flags;
581 
582 	/*
583 	 * This happens when the send engine notes
584 	 * a QP in the error state and cannot
585 	 * do the flush work until that QP's
586 	 * sdma work has finished.
587 	 */
588 	spin_lock_irqsave(&qp->s_lock, flags);
589 	if (qp->s_flags & RVT_S_WAIT_DMA) {
590 		qp->s_flags &= ~RVT_S_WAIT_DMA;
591 		hfi1_schedule_send(qp);
592 	}
593 	spin_unlock_irqrestore(&qp->s_lock, flags);
594 }
595 
596 /**
597  *
598  * qp_to_sdma_engine - map a qp to a send engine
599  * @qp: the QP
600  * @sc5: the 5 bit sc
601  *
602  * Return:
603  * A send engine for the qp or NULL for SMI type qp.
604  */
605 struct sdma_engine *qp_to_sdma_engine(struct rvt_qp *qp, u8 sc5)
606 {
607 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
608 	struct sdma_engine *sde;
609 
610 	if (!(dd->flags & HFI1_HAS_SEND_DMA))
611 		return NULL;
612 	switch (qp->ibqp.qp_type) {
613 	case IB_QPT_SMI:
614 		return NULL;
615 	default:
616 		break;
617 	}
618 	sde = sdma_select_engine_sc(dd, qp->ibqp.qp_num >> dd->qos_shift, sc5);
619 	return sde;
620 }
621 
622 /*
623  * qp_to_send_context - map a qp to a send context
624  * @qp: the QP
625  * @sc5: the 5 bit sc
626  *
627  * Return:
628  * A send context for the qp
629  */
630 struct send_context *qp_to_send_context(struct rvt_qp *qp, u8 sc5)
631 {
632 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
633 
634 	switch (qp->ibqp.qp_type) {
635 	case IB_QPT_SMI:
636 		/* SMA packets to VL15 */
637 		return dd->vld[15].sc;
638 	default:
639 		break;
640 	}
641 
642 	return pio_select_send_context_sc(dd, qp->ibqp.qp_num >> dd->qos_shift,
643 					  sc5);
644 }
645 
646 struct qp_iter {
647 	struct hfi1_ibdev *dev;
648 	struct rvt_qp *qp;
649 	int specials;
650 	int n;
651 };
652 
653 struct qp_iter *qp_iter_init(struct hfi1_ibdev *dev)
654 {
655 	struct qp_iter *iter;
656 
657 	iter = kzalloc(sizeof(*iter), GFP_KERNEL);
658 	if (!iter)
659 		return NULL;
660 
661 	iter->dev = dev;
662 	iter->specials = dev->rdi.ibdev.phys_port_cnt * 2;
663 
664 	return iter;
665 }
666 
667 int qp_iter_next(struct qp_iter *iter)
668 {
669 	struct hfi1_ibdev *dev = iter->dev;
670 	int n = iter->n;
671 	int ret = 1;
672 	struct rvt_qp *pqp = iter->qp;
673 	struct rvt_qp *qp;
674 
675 	/*
676 	 * The approach is to consider the special qps
677 	 * as an additional table entries before the
678 	 * real hash table.  Since the qp code sets
679 	 * the qp->next hash link to NULL, this works just fine.
680 	 *
681 	 * iter->specials is 2 * # ports
682 	 *
683 	 * n = 0..iter->specials is the special qp indices
684 	 *
685 	 * n = iter->specials..dev->rdi.qp_dev->qp_table_size+iter->specials are
686 	 * the potential hash bucket entries
687 	 *
688 	 */
689 	for (; n <  dev->rdi.qp_dev->qp_table_size + iter->specials; n++) {
690 		if (pqp) {
691 			qp = rcu_dereference(pqp->next);
692 		} else {
693 			if (n < iter->specials) {
694 				struct hfi1_pportdata *ppd;
695 				struct hfi1_ibport *ibp;
696 				int pidx;
697 
698 				pidx = n % dev->rdi.ibdev.phys_port_cnt;
699 				ppd = &dd_from_dev(dev)->pport[pidx];
700 				ibp = &ppd->ibport_data;
701 
702 				if (!(n & 1))
703 					qp = rcu_dereference(ibp->rvp.qp[0]);
704 				else
705 					qp = rcu_dereference(ibp->rvp.qp[1]);
706 			} else {
707 				qp = rcu_dereference(
708 					dev->rdi.qp_dev->qp_table[
709 						(n - iter->specials)]);
710 			}
711 		}
712 		pqp = qp;
713 		if (qp) {
714 			iter->qp = qp;
715 			iter->n = n;
716 			return 0;
717 		}
718 	}
719 	return ret;
720 }
721 
722 static const char * const qp_type_str[] = {
723 	"SMI", "GSI", "RC", "UC", "UD",
724 };
725 
726 static int qp_idle(struct rvt_qp *qp)
727 {
728 	return
729 		qp->s_last == qp->s_acked &&
730 		qp->s_acked == qp->s_cur &&
731 		qp->s_cur == qp->s_tail &&
732 		qp->s_tail == qp->s_head;
733 }
734 
735 void qp_iter_print(struct seq_file *s, struct qp_iter *iter)
736 {
737 	struct rvt_swqe *wqe;
738 	struct rvt_qp *qp = iter->qp;
739 	struct hfi1_qp_priv *priv = qp->priv;
740 	struct sdma_engine *sde;
741 	struct send_context *send_context;
742 
743 	sde = qp_to_sdma_engine(qp, priv->s_sc);
744 	wqe = rvt_get_swqe_ptr(qp, qp->s_last);
745 	send_context = qp_to_send_context(qp, priv->s_sc);
746 	seq_printf(s,
747 		   "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",
748 		   iter->n,
749 		   qp_idle(qp) ? "I" : "B",
750 		   qp->ibqp.qp_num,
751 		   atomic_read(&qp->refcount),
752 		   qp_type_str[qp->ibqp.qp_type],
753 		   qp->state,
754 		   wqe ? wqe->wr.opcode : 0,
755 		   qp->s_hdrwords,
756 		   qp->s_flags,
757 		   iowait_sdma_pending(&priv->s_iowait),
758 		   iowait_pio_pending(&priv->s_iowait),
759 		   !list_empty(&priv->s_iowait.list),
760 		   qp->timeout,
761 		   wqe ? wqe->ssn : 0,
762 		   qp->s_lsn,
763 		   qp->s_last_psn,
764 		   qp->s_psn, qp->s_next_psn,
765 		   qp->s_sending_psn, qp->s_sending_hpsn,
766 		   qp->s_last, qp->s_acked, qp->s_cur,
767 		   qp->s_tail, qp->s_head, qp->s_size,
768 		   qp->s_avail,
769 		   qp->remote_qpn,
770 		   qp->remote_ah_attr.dlid,
771 		   qp->remote_ah_attr.sl,
772 		   qp->pmtu,
773 		   qp->s_retry,
774 		   qp->s_retry_cnt,
775 		   qp->s_rnr_retry_cnt,
776 		   sde,
777 		   sde ? sde->this_idx : 0,
778 		   send_context,
779 		   send_context ? send_context->sw_index : 0,
780 		   ibcq_to_rvtcq(qp->ibqp.send_cq)->queue->head,
781 		   ibcq_to_rvtcq(qp->ibqp.send_cq)->queue->tail,
782 		   qp->pid);
783 }
784 
785 void qp_comm_est(struct rvt_qp *qp)
786 {
787 	qp->r_flags |= RVT_R_COMM_EST;
788 	if (qp->ibqp.event_handler) {
789 		struct ib_event ev;
790 
791 		ev.device = qp->ibqp.device;
792 		ev.element.qp = &qp->ibqp;
793 		ev.event = IB_EVENT_COMM_EST;
794 		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
795 	}
796 }
797 
798 void *qp_priv_alloc(struct rvt_dev_info *rdi, struct rvt_qp *qp,
799 		    gfp_t gfp)
800 {
801 	struct hfi1_qp_priv *priv;
802 
803 	priv = kzalloc_node(sizeof(*priv), gfp, rdi->dparms.node);
804 	if (!priv)
805 		return ERR_PTR(-ENOMEM);
806 
807 	priv->owner = qp;
808 
809 	priv->s_ahg = kzalloc_node(sizeof(*priv->s_ahg), gfp,
810 				   rdi->dparms.node);
811 	if (!priv->s_ahg) {
812 		kfree(priv);
813 		return ERR_PTR(-ENOMEM);
814 	}
815 	iowait_init(
816 		&priv->s_iowait,
817 		1,
818 		_hfi1_do_send,
819 		iowait_sleep,
820 		iowait_wakeup,
821 		iowait_sdma_drained);
822 	setup_timer(&priv->s_rnr_timer, hfi1_rc_rnr_retry, (unsigned long)qp);
823 	qp->s_timer.function = hfi1_rc_timeout;
824 	return priv;
825 }
826 
827 void qp_priv_free(struct rvt_dev_info *rdi, struct rvt_qp *qp)
828 {
829 	struct hfi1_qp_priv *priv = qp->priv;
830 
831 	kfree(priv->s_ahg);
832 	kfree(priv);
833 }
834 
835 unsigned free_all_qps(struct rvt_dev_info *rdi)
836 {
837 	struct hfi1_ibdev *verbs_dev = container_of(rdi,
838 						    struct hfi1_ibdev,
839 						    rdi);
840 	struct hfi1_devdata *dd = container_of(verbs_dev,
841 					       struct hfi1_devdata,
842 					       verbs_dev);
843 	int n;
844 	unsigned qp_inuse = 0;
845 
846 	for (n = 0; n < dd->num_pports; n++) {
847 		struct hfi1_ibport *ibp = &dd->pport[n].ibport_data;
848 
849 		rcu_read_lock();
850 		if (rcu_dereference(ibp->rvp.qp[0]))
851 			qp_inuse++;
852 		if (rcu_dereference(ibp->rvp.qp[1]))
853 			qp_inuse++;
854 		rcu_read_unlock();
855 	}
856 
857 	return qp_inuse;
858 }
859 
860 void flush_qp_waiters(struct rvt_qp *qp)
861 {
862 	lockdep_assert_held(&qp->s_lock);
863 	flush_iowait(qp);
864 	hfi1_stop_rc_timers(qp);
865 }
866 
867 void stop_send_queue(struct rvt_qp *qp)
868 {
869 	struct hfi1_qp_priv *priv = qp->priv;
870 
871 	cancel_work_sync(&priv->s_iowait.iowork);
872 	hfi1_del_timers_sync(qp);
873 }
874 
875 void quiesce_qp(struct rvt_qp *qp)
876 {
877 	struct hfi1_qp_priv *priv = qp->priv;
878 
879 	iowait_sdma_drain(&priv->s_iowait);
880 	qp_pio_drain(qp);
881 	flush_tx_list(qp);
882 }
883 
884 void notify_qp_reset(struct rvt_qp *qp)
885 {
886 	struct hfi1_qp_priv *priv = qp->priv;
887 
888 	priv->r_adefered = 0;
889 	clear_ahg(qp);
890 }
891 
892 /*
893  * Switch to alternate path.
894  * The QP s_lock should be held and interrupts disabled.
895  */
896 void hfi1_migrate_qp(struct rvt_qp *qp)
897 {
898 	struct hfi1_qp_priv *priv = qp->priv;
899 	struct ib_event ev;
900 
901 	qp->s_mig_state = IB_MIG_MIGRATED;
902 	qp->remote_ah_attr = qp->alt_ah_attr;
903 	qp->port_num = qp->alt_ah_attr.port_num;
904 	qp->s_pkey_index = qp->s_alt_pkey_index;
905 	qp->s_flags |= RVT_S_AHG_CLEAR;
906 	priv->s_sc = ah_to_sc(qp->ibqp.device, &qp->remote_ah_attr);
907 	priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
908 
909 	ev.device = qp->ibqp.device;
910 	ev.element.qp = &qp->ibqp;
911 	ev.event = IB_EVENT_PATH_MIG;
912 	qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
913 }
914 
915 int mtu_to_path_mtu(u32 mtu)
916 {
917 	return mtu_to_enum(mtu, OPA_MTU_8192);
918 }
919 
920 u32 mtu_from_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp, u32 pmtu)
921 {
922 	u32 mtu;
923 	struct hfi1_ibdev *verbs_dev = container_of(rdi,
924 						    struct hfi1_ibdev,
925 						    rdi);
926 	struct hfi1_devdata *dd = container_of(verbs_dev,
927 					       struct hfi1_devdata,
928 					       verbs_dev);
929 	struct hfi1_ibport *ibp;
930 	u8 sc, vl;
931 
932 	ibp = &dd->pport[qp->port_num - 1].ibport_data;
933 	sc = ibp->sl_to_sc[qp->remote_ah_attr.sl];
934 	vl = sc_to_vlt(dd, sc);
935 
936 	mtu = verbs_mtu_enum_to_int(qp->ibqp.device, pmtu);
937 	if (vl < PER_VL_SEND_CONTEXTS)
938 		mtu = min_t(u32, mtu, dd->vld[vl].mtu);
939 	return mtu;
940 }
941 
942 int get_pmtu_from_attr(struct rvt_dev_info *rdi, struct rvt_qp *qp,
943 		       struct ib_qp_attr *attr)
944 {
945 	int mtu, pidx = qp->port_num - 1;
946 	struct hfi1_ibdev *verbs_dev = container_of(rdi,
947 						    struct hfi1_ibdev,
948 						    rdi);
949 	struct hfi1_devdata *dd = container_of(verbs_dev,
950 					       struct hfi1_devdata,
951 					       verbs_dev);
952 	mtu = verbs_mtu_enum_to_int(qp->ibqp.device, attr->path_mtu);
953 	if (mtu == -1)
954 		return -1; /* values less than 0 are error */
955 
956 	if (mtu > dd->pport[pidx].ibmtu)
957 		return mtu_to_enum(dd->pport[pidx].ibmtu, IB_MTU_2048);
958 	else
959 		return attr->path_mtu;
960 }
961 
962 void notify_error_qp(struct rvt_qp *qp)
963 {
964 	struct hfi1_ibdev *dev = to_idev(qp->ibqp.device);
965 	struct hfi1_qp_priv *priv = qp->priv;
966 
967 	write_seqlock(&dev->iowait_lock);
968 	if (!list_empty(&priv->s_iowait.list) && !(qp->s_flags & RVT_S_BUSY)) {
969 		qp->s_flags &= ~RVT_S_ANY_WAIT_IO;
970 		list_del_init(&priv->s_iowait.list);
971 		priv->s_iowait.lock = NULL;
972 		rvt_put_qp(qp);
973 	}
974 	write_sequnlock(&dev->iowait_lock);
975 
976 	if (!(qp->s_flags & RVT_S_BUSY)) {
977 		qp->s_hdrwords = 0;
978 		if (qp->s_rdma_mr) {
979 			rvt_put_mr(qp->s_rdma_mr);
980 			qp->s_rdma_mr = NULL;
981 		}
982 		flush_tx_list(qp);
983 	}
984 }
985 
986 /**
987  * hfi1_error_port_qps - put a port's RC/UC qps into error state
988  * @ibp: the ibport.
989  * @sl: the service level.
990  *
991  * This function places all RC/UC qps with a given service level into error
992  * state. It is generally called to force upper lay apps to abandon stale qps
993  * after an sl->sc mapping change.
994  */
995 void hfi1_error_port_qps(struct hfi1_ibport *ibp, u8 sl)
996 {
997 	struct rvt_qp *qp = NULL;
998 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
999 	struct hfi1_ibdev *dev = &ppd->dd->verbs_dev;
1000 	int n;
1001 	int lastwqe;
1002 	struct ib_event ev;
1003 
1004 	rcu_read_lock();
1005 
1006 	/* Deal only with RC/UC qps that use the given SL. */
1007 	for (n = 0; n < dev->rdi.qp_dev->qp_table_size; n++) {
1008 		for (qp = rcu_dereference(dev->rdi.qp_dev->qp_table[n]); qp;
1009 			qp = rcu_dereference(qp->next)) {
1010 			if (qp->port_num == ppd->port &&
1011 			    (qp->ibqp.qp_type == IB_QPT_UC ||
1012 			     qp->ibqp.qp_type == IB_QPT_RC) &&
1013 			    qp->remote_ah_attr.sl == sl &&
1014 			    (ib_rvt_state_ops[qp->state] &
1015 			     RVT_POST_SEND_OK)) {
1016 				spin_lock_irq(&qp->r_lock);
1017 				spin_lock(&qp->s_hlock);
1018 				spin_lock(&qp->s_lock);
1019 				lastwqe = rvt_error_qp(qp,
1020 						       IB_WC_WR_FLUSH_ERR);
1021 				spin_unlock(&qp->s_lock);
1022 				spin_unlock(&qp->s_hlock);
1023 				spin_unlock_irq(&qp->r_lock);
1024 				if (lastwqe) {
1025 					ev.device = qp->ibqp.device;
1026 					ev.element.qp = &qp->ibqp;
1027 					ev.event =
1028 						IB_EVENT_QP_LAST_WQE_REACHED;
1029 					qp->ibqp.event_handler(&ev,
1030 						qp->ibqp.qp_context);
1031 				}
1032 			}
1033 		}
1034 	}
1035 
1036 	rcu_read_unlock();
1037 }
1038