xref: /linux/drivers/infiniband/hw/hfi1/qp.c (revision 23c48a124b469cee2eb0c75e6d22d366d1caa118)
1 // SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause
2 /*
3  * Copyright(c) 2015 - 2020 Intel Corporation.
4  */
5 
6 #include <linux/err.h>
7 #include <linux/vmalloc.h>
8 #include <linux/hash.h>
9 #include <linux/module.h>
10 #include <linux/seq_file.h>
11 #include <rdma/rdma_vt.h>
12 #include <rdma/rdmavt_qp.h>
13 #include <rdma/ib_verbs.h>
14 
15 #include "hfi.h"
16 #include "qp.h"
17 #include "trace.h"
18 #include "verbs_txreq.h"
19 
20 unsigned int hfi1_qp_table_size = 256;
21 module_param_named(qp_table_size, hfi1_qp_table_size, uint, S_IRUGO);
22 MODULE_PARM_DESC(qp_table_size, "QP table size");
23 
24 static void flush_tx_list(struct rvt_qp *qp);
25 static int iowait_sleep(
26 	struct sdma_engine *sde,
27 	struct iowait_work *wait,
28 	struct sdma_txreq *stx,
29 	unsigned int seq,
30 	bool pkts_sent);
31 static void iowait_wakeup(struct iowait *wait, int reason);
32 static void iowait_sdma_drained(struct iowait *wait);
33 static void qp_pio_drain(struct rvt_qp *qp);
34 
35 const struct rvt_operation_params hfi1_post_parms[RVT_OPERATION_MAX] = {
36 [IB_WR_RDMA_WRITE] = {
37 	.length = sizeof(struct ib_rdma_wr),
38 	.qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
39 },
40 
41 [IB_WR_RDMA_READ] = {
42 	.length = sizeof(struct ib_rdma_wr),
43 	.qpt_support = BIT(IB_QPT_RC),
44 	.flags = RVT_OPERATION_ATOMIC,
45 },
46 
47 [IB_WR_ATOMIC_CMP_AND_SWP] = {
48 	.length = sizeof(struct ib_atomic_wr),
49 	.qpt_support = BIT(IB_QPT_RC),
50 	.flags = RVT_OPERATION_ATOMIC | RVT_OPERATION_ATOMIC_SGE,
51 },
52 
53 [IB_WR_ATOMIC_FETCH_AND_ADD] = {
54 	.length = sizeof(struct ib_atomic_wr),
55 	.qpt_support = BIT(IB_QPT_RC),
56 	.flags = RVT_OPERATION_ATOMIC | RVT_OPERATION_ATOMIC_SGE,
57 },
58 
59 [IB_WR_RDMA_WRITE_WITH_IMM] = {
60 	.length = sizeof(struct ib_rdma_wr),
61 	.qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
62 },
63 
64 [IB_WR_SEND] = {
65 	.length = sizeof(struct ib_send_wr),
66 	.qpt_support = BIT(IB_QPT_UD) | BIT(IB_QPT_SMI) | BIT(IB_QPT_GSI) |
67 		       BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
68 },
69 
70 [IB_WR_SEND_WITH_IMM] = {
71 	.length = sizeof(struct ib_send_wr),
72 	.qpt_support = BIT(IB_QPT_UD) | BIT(IB_QPT_SMI) | BIT(IB_QPT_GSI) |
73 		       BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
74 },
75 
76 [IB_WR_REG_MR] = {
77 	.length = sizeof(struct ib_reg_wr),
78 	.qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
79 	.flags = RVT_OPERATION_LOCAL,
80 },
81 
82 [IB_WR_LOCAL_INV] = {
83 	.length = sizeof(struct ib_send_wr),
84 	.qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
85 	.flags = RVT_OPERATION_LOCAL,
86 },
87 
88 [IB_WR_SEND_WITH_INV] = {
89 	.length = sizeof(struct ib_send_wr),
90 	.qpt_support = BIT(IB_QPT_RC),
91 },
92 
93 [IB_WR_OPFN] = {
94 	.length = sizeof(struct ib_atomic_wr),
95 	.qpt_support = BIT(IB_QPT_RC),
96 	.flags = RVT_OPERATION_USE_RESERVE,
97 },
98 
99 [IB_WR_TID_RDMA_WRITE] = {
100 	.length = sizeof(struct ib_rdma_wr),
101 	.qpt_support = BIT(IB_QPT_RC),
102 	.flags = RVT_OPERATION_IGN_RNR_CNT,
103 },
104 
105 };
106 
107 static void flush_list_head(struct list_head *l)
108 {
109 	while (!list_empty(l)) {
110 		struct sdma_txreq *tx;
111 
112 		tx = list_first_entry(
113 			l,
114 			struct sdma_txreq,
115 			list);
116 		list_del_init(&tx->list);
117 		hfi1_put_txreq(
118 			container_of(tx, struct verbs_txreq, txreq));
119 	}
120 }
121 
122 static void flush_tx_list(struct rvt_qp *qp)
123 {
124 	struct hfi1_qp_priv *priv = qp->priv;
125 
126 	flush_list_head(&iowait_get_ib_work(&priv->s_iowait)->tx_head);
127 	flush_list_head(&iowait_get_tid_work(&priv->s_iowait)->tx_head);
128 }
129 
130 static void flush_iowait(struct rvt_qp *qp)
131 {
132 	struct hfi1_qp_priv *priv = qp->priv;
133 	unsigned long flags;
134 	seqlock_t *lock = priv->s_iowait.lock;
135 
136 	if (!lock)
137 		return;
138 	write_seqlock_irqsave(lock, flags);
139 	if (!list_empty(&priv->s_iowait.list)) {
140 		list_del_init(&priv->s_iowait.list);
141 		priv->s_iowait.lock = NULL;
142 		rvt_put_qp(qp);
143 	}
144 	write_sequnlock_irqrestore(lock, flags);
145 }
146 
147 /*
148  * This function is what we would push to the core layer if we wanted to be a
149  * "first class citizen".  Instead we hide this here and rely on Verbs ULPs
150  * to blindly pass the MTU enum value from the PathRecord to us.
151  */
152 static inline int verbs_mtu_enum_to_int(struct ib_device *dev, enum ib_mtu mtu)
153 {
154 	/* Constraining 10KB packets to 8KB packets */
155 	if (mtu == (enum ib_mtu)OPA_MTU_10240)
156 		mtu = (enum ib_mtu)OPA_MTU_8192;
157 	return opa_mtu_enum_to_int((enum opa_mtu)mtu);
158 }
159 
160 int hfi1_check_modify_qp(struct rvt_qp *qp, struct ib_qp_attr *attr,
161 			 int attr_mask, struct ib_udata *udata)
162 {
163 	struct ib_qp *ibqp = &qp->ibqp;
164 	struct hfi1_ibdev *dev = to_idev(ibqp->device);
165 	struct hfi1_devdata *dd = dd_from_dev(dev);
166 	u8 sc;
167 
168 	if (attr_mask & IB_QP_AV) {
169 		sc = ah_to_sc(ibqp->device, &attr->ah_attr);
170 		if (sc == 0xf)
171 			return -EINVAL;
172 
173 		if (!qp_to_sdma_engine(qp, sc) &&
174 		    dd->flags & HFI1_HAS_SEND_DMA)
175 			return -EINVAL;
176 
177 		if (!qp_to_send_context(qp, sc))
178 			return -EINVAL;
179 	}
180 
181 	if (attr_mask & IB_QP_ALT_PATH) {
182 		sc = ah_to_sc(ibqp->device, &attr->alt_ah_attr);
183 		if (sc == 0xf)
184 			return -EINVAL;
185 
186 		if (!qp_to_sdma_engine(qp, sc) &&
187 		    dd->flags & HFI1_HAS_SEND_DMA)
188 			return -EINVAL;
189 
190 		if (!qp_to_send_context(qp, sc))
191 			return -EINVAL;
192 	}
193 
194 	return 0;
195 }
196 
197 /*
198  * qp_set_16b - Set the hdr_type based on whether the slid or the
199  * dlid in the connection is extended. Only applicable for RC and UC
200  * QPs. UD QPs determine this on the fly from the ah in the wqe
201  */
202 static inline void qp_set_16b(struct rvt_qp *qp)
203 {
204 	struct hfi1_pportdata *ppd;
205 	struct hfi1_ibport *ibp;
206 	struct hfi1_qp_priv *priv = qp->priv;
207 
208 	/* Update ah_attr to account for extended LIDs */
209 	hfi1_update_ah_attr(qp->ibqp.device, &qp->remote_ah_attr);
210 
211 	/* Create 32 bit LIDs */
212 	hfi1_make_opa_lid(&qp->remote_ah_attr);
213 
214 	if (!(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH))
215 		return;
216 
217 	ibp = to_iport(qp->ibqp.device, qp->port_num);
218 	ppd = ppd_from_ibp(ibp);
219 	priv->hdr_type = hfi1_get_hdr_type(ppd->lid, &qp->remote_ah_attr);
220 }
221 
222 void hfi1_modify_qp(struct rvt_qp *qp, struct ib_qp_attr *attr,
223 		    int attr_mask, struct ib_udata *udata)
224 {
225 	struct ib_qp *ibqp = &qp->ibqp;
226 	struct hfi1_qp_priv *priv = qp->priv;
227 
228 	if (attr_mask & IB_QP_AV) {
229 		priv->s_sc = ah_to_sc(ibqp->device, &qp->remote_ah_attr);
230 		priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
231 		priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
232 		qp_set_16b(qp);
233 	}
234 
235 	if (attr_mask & IB_QP_PATH_MIG_STATE &&
236 	    attr->path_mig_state == IB_MIG_MIGRATED &&
237 	    qp->s_mig_state == IB_MIG_ARMED) {
238 		qp->s_flags |= HFI1_S_AHG_CLEAR;
239 		priv->s_sc = ah_to_sc(ibqp->device, &qp->remote_ah_attr);
240 		priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
241 		priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
242 		qp_set_16b(qp);
243 	}
244 
245 	opfn_qp_init(qp, attr, attr_mask);
246 }
247 
248 /**
249  * hfi1_setup_wqe - set up the wqe
250  * @qp: The qp
251  * @wqe: The built wqe
252  * @call_send: Determine if the send should be posted or scheduled.
253  *
254  * Perform setup of the wqe.  This is called
255  * prior to inserting the wqe into the ring but after
256  * the wqe has been setup by RDMAVT. This function
257  * allows the driver the opportunity to perform
258  * validation and additional setup of the wqe.
259  *
260  * Returns 0 on success, -EINVAL on failure
261  *
262  */
263 int hfi1_setup_wqe(struct rvt_qp *qp, struct rvt_swqe *wqe, bool *call_send)
264 {
265 	struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
266 	struct rvt_ah *ah;
267 	struct hfi1_pportdata *ppd;
268 	struct hfi1_devdata *dd;
269 
270 	switch (qp->ibqp.qp_type) {
271 	case IB_QPT_RC:
272 		hfi1_setup_tid_rdma_wqe(qp, wqe);
273 		fallthrough;
274 	case IB_QPT_UC:
275 		if (wqe->length > 0x80000000U)
276 			return -EINVAL;
277 		if (wqe->length > qp->pmtu)
278 			*call_send = false;
279 		break;
280 	case IB_QPT_SMI:
281 		/*
282 		 * SM packets should exclusively use VL15 and their SL is
283 		 * ignored (IBTA v1.3, Section 3.5.8.2). Therefore, when ah
284 		 * is created, SL is 0 in most cases and as a result some
285 		 * fields (vl and pmtu) in ah may not be set correctly,
286 		 * depending on the SL2SC and SC2VL tables at the time.
287 		 */
288 		ppd = ppd_from_ibp(ibp);
289 		dd = dd_from_ppd(ppd);
290 		if (wqe->length > dd->vld[15].mtu)
291 			return -EINVAL;
292 		break;
293 	case IB_QPT_GSI:
294 	case IB_QPT_UD:
295 		ah = rvt_get_swqe_ah(wqe);
296 		if (wqe->length > (1 << ah->log_pmtu))
297 			return -EINVAL;
298 		if (ibp->sl_to_sc[rdma_ah_get_sl(&ah->attr)] == 0xf)
299 			return -EINVAL;
300 		break;
301 	default:
302 		break;
303 	}
304 
305 	/*
306 	 * System latency between send and schedule is large enough that
307 	 * forcing call_send to true for piothreshold packets is necessary.
308 	 */
309 	if (wqe->length <= piothreshold)
310 		*call_send = true;
311 	return 0;
312 }
313 
314 /**
315  * _hfi1_schedule_send - schedule progress
316  * @qp: the QP
317  *
318  * This schedules qp progress w/o regard to the s_flags.
319  *
320  * It is only used in the post send, which doesn't hold
321  * the s_lock.
322  */
323 bool _hfi1_schedule_send(struct rvt_qp *qp)
324 {
325 	struct hfi1_qp_priv *priv = qp->priv;
326 	struct hfi1_ibport *ibp =
327 		to_iport(qp->ibqp.device, qp->port_num);
328 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
329 	struct hfi1_devdata *dd = ppd->dd;
330 
331 	if (dd->flags & HFI1_SHUTDOWN)
332 		return true;
333 
334 	return iowait_schedule(&priv->s_iowait, ppd->hfi1_wq,
335 			       priv->s_sde ?
336 			       priv->s_sde->cpu :
337 			       cpumask_first(cpumask_of_node(dd->node)));
338 }
339 
340 static void qp_pio_drain(struct rvt_qp *qp)
341 {
342 	struct hfi1_qp_priv *priv = qp->priv;
343 
344 	if (!priv->s_sendcontext)
345 		return;
346 	while (iowait_pio_pending(&priv->s_iowait)) {
347 		write_seqlock_irq(&priv->s_sendcontext->waitlock);
348 		hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 1);
349 		write_sequnlock_irq(&priv->s_sendcontext->waitlock);
350 		iowait_pio_drain(&priv->s_iowait);
351 		write_seqlock_irq(&priv->s_sendcontext->waitlock);
352 		hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 0);
353 		write_sequnlock_irq(&priv->s_sendcontext->waitlock);
354 	}
355 }
356 
357 /**
358  * hfi1_schedule_send - schedule progress
359  * @qp: the QP
360  *
361  * This schedules qp progress and caller should hold
362  * the s_lock.
363  * @return true if the first leg is scheduled;
364  * false if the first leg is not scheduled.
365  */
366 bool hfi1_schedule_send(struct rvt_qp *qp)
367 {
368 	lockdep_assert_held(&qp->s_lock);
369 	if (hfi1_send_ok(qp)) {
370 		_hfi1_schedule_send(qp);
371 		return true;
372 	}
373 	if (qp->s_flags & HFI1_S_ANY_WAIT_IO)
374 		iowait_set_flag(&((struct hfi1_qp_priv *)qp->priv)->s_iowait,
375 				IOWAIT_PENDING_IB);
376 	return false;
377 }
378 
379 static void hfi1_qp_schedule(struct rvt_qp *qp)
380 {
381 	struct hfi1_qp_priv *priv = qp->priv;
382 	bool ret;
383 
384 	if (iowait_flag_set(&priv->s_iowait, IOWAIT_PENDING_IB)) {
385 		ret = hfi1_schedule_send(qp);
386 		if (ret)
387 			iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_IB);
388 	}
389 	if (iowait_flag_set(&priv->s_iowait, IOWAIT_PENDING_TID)) {
390 		ret = hfi1_schedule_tid_send(qp);
391 		if (ret)
392 			iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_TID);
393 	}
394 }
395 
396 void hfi1_qp_wakeup(struct rvt_qp *qp, u32 flag)
397 {
398 	unsigned long flags;
399 
400 	spin_lock_irqsave(&qp->s_lock, flags);
401 	if (qp->s_flags & flag) {
402 		qp->s_flags &= ~flag;
403 		trace_hfi1_qpwakeup(qp, flag);
404 		hfi1_qp_schedule(qp);
405 	}
406 	spin_unlock_irqrestore(&qp->s_lock, flags);
407 	/* Notify hfi1_destroy_qp() if it is waiting. */
408 	rvt_put_qp(qp);
409 }
410 
411 void hfi1_qp_unbusy(struct rvt_qp *qp, struct iowait_work *wait)
412 {
413 	struct hfi1_qp_priv *priv = qp->priv;
414 
415 	if (iowait_set_work_flag(wait) == IOWAIT_IB_SE) {
416 		qp->s_flags &= ~RVT_S_BUSY;
417 		/*
418 		 * If we are sending a first-leg packet from the second leg,
419 		 * we need to clear the busy flag from priv->s_flags to
420 		 * avoid a race condition when the qp wakes up before
421 		 * the call to hfi1_verbs_send() returns to the second
422 		 * leg. In that case, the second leg will terminate without
423 		 * being re-scheduled, resulting in failure to send TID RDMA
424 		 * WRITE DATA and TID RDMA ACK packets.
425 		 */
426 		if (priv->s_flags & HFI1_S_TID_BUSY_SET) {
427 			priv->s_flags &= ~(HFI1_S_TID_BUSY_SET |
428 					   RVT_S_BUSY);
429 			iowait_set_flag(&priv->s_iowait, IOWAIT_PENDING_TID);
430 		}
431 	} else {
432 		priv->s_flags &= ~RVT_S_BUSY;
433 	}
434 }
435 
436 static int iowait_sleep(
437 	struct sdma_engine *sde,
438 	struct iowait_work *wait,
439 	struct sdma_txreq *stx,
440 	uint seq,
441 	bool pkts_sent)
442 {
443 	struct verbs_txreq *tx = container_of(stx, struct verbs_txreq, txreq);
444 	struct rvt_qp *qp;
445 	struct hfi1_qp_priv *priv;
446 	unsigned long flags;
447 	int ret = 0;
448 
449 	qp = tx->qp;
450 	priv = qp->priv;
451 
452 	spin_lock_irqsave(&qp->s_lock, flags);
453 	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
454 		/*
455 		 * If we couldn't queue the DMA request, save the info
456 		 * and try again later rather than destroying the
457 		 * buffer and undoing the side effects of the copy.
458 		 */
459 		/* Make a common routine? */
460 		list_add_tail(&stx->list, &wait->tx_head);
461 		write_seqlock(&sde->waitlock);
462 		if (sdma_progress(sde, seq, stx))
463 			goto eagain;
464 		if (list_empty(&priv->s_iowait.list)) {
465 			struct hfi1_ibport *ibp =
466 				to_iport(qp->ibqp.device, qp->port_num);
467 
468 			ibp->rvp.n_dmawait++;
469 			qp->s_flags |= RVT_S_WAIT_DMA_DESC;
470 			iowait_get_priority(&priv->s_iowait);
471 			iowait_queue(pkts_sent, &priv->s_iowait,
472 				     &sde->dmawait);
473 			priv->s_iowait.lock = &sde->waitlock;
474 			trace_hfi1_qpsleep(qp, RVT_S_WAIT_DMA_DESC);
475 			rvt_get_qp(qp);
476 		}
477 		write_sequnlock(&sde->waitlock);
478 		hfi1_qp_unbusy(qp, wait);
479 		spin_unlock_irqrestore(&qp->s_lock, flags);
480 		ret = -EBUSY;
481 	} else {
482 		spin_unlock_irqrestore(&qp->s_lock, flags);
483 		hfi1_put_txreq(tx);
484 	}
485 	return ret;
486 eagain:
487 	write_sequnlock(&sde->waitlock);
488 	spin_unlock_irqrestore(&qp->s_lock, flags);
489 	list_del_init(&stx->list);
490 	return -EAGAIN;
491 }
492 
493 static void iowait_wakeup(struct iowait *wait, int reason)
494 {
495 	struct rvt_qp *qp = iowait_to_qp(wait);
496 
497 	WARN_ON(reason != SDMA_AVAIL_REASON);
498 	hfi1_qp_wakeup(qp, RVT_S_WAIT_DMA_DESC);
499 }
500 
501 static void iowait_sdma_drained(struct iowait *wait)
502 {
503 	struct rvt_qp *qp = iowait_to_qp(wait);
504 	unsigned long flags;
505 
506 	/*
507 	 * This happens when the send engine notes
508 	 * a QP in the error state and cannot
509 	 * do the flush work until that QP's
510 	 * sdma work has finished.
511 	 */
512 	spin_lock_irqsave(&qp->s_lock, flags);
513 	if (qp->s_flags & RVT_S_WAIT_DMA) {
514 		qp->s_flags &= ~RVT_S_WAIT_DMA;
515 		hfi1_schedule_send(qp);
516 	}
517 	spin_unlock_irqrestore(&qp->s_lock, flags);
518 }
519 
520 static void hfi1_init_priority(struct iowait *w)
521 {
522 	struct rvt_qp *qp = iowait_to_qp(w);
523 	struct hfi1_qp_priv *priv = qp->priv;
524 
525 	if (qp->s_flags & RVT_S_ACK_PENDING)
526 		w->priority++;
527 	if (priv->s_flags & RVT_S_ACK_PENDING)
528 		w->priority++;
529 }
530 
531 /**
532  * qp_to_sdma_engine - map a qp to a send engine
533  * @qp: the QP
534  * @sc5: the 5 bit sc
535  *
536  * Return:
537  * A send engine for the qp or NULL for SMI type qp.
538  */
539 struct sdma_engine *qp_to_sdma_engine(struct rvt_qp *qp, u8 sc5)
540 {
541 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
542 	struct sdma_engine *sde;
543 
544 	if (!(dd->flags & HFI1_HAS_SEND_DMA))
545 		return NULL;
546 	switch (qp->ibqp.qp_type) {
547 	case IB_QPT_SMI:
548 		return NULL;
549 	default:
550 		break;
551 	}
552 	sde = sdma_select_engine_sc(dd, qp->ibqp.qp_num >> dd->qos_shift, sc5);
553 	return sde;
554 }
555 
556 /**
557  * qp_to_send_context - map a qp to a send context
558  * @qp: the QP
559  * @sc5: the 5 bit sc
560  *
561  * Return:
562  * A send context for the qp
563  */
564 struct send_context *qp_to_send_context(struct rvt_qp *qp, u8 sc5)
565 {
566 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
567 
568 	switch (qp->ibqp.qp_type) {
569 	case IB_QPT_SMI:
570 		/* SMA packets to VL15 */
571 		return dd->vld[15].sc;
572 	default:
573 		break;
574 	}
575 
576 	return pio_select_send_context_sc(dd, qp->ibqp.qp_num >> dd->qos_shift,
577 					  sc5);
578 }
579 
580 static const char * const qp_type_str[] = {
581 	"SMI", "GSI", "RC", "UC", "UD",
582 };
583 
584 static int qp_idle(struct rvt_qp *qp)
585 {
586 	return
587 		qp->s_last == qp->s_acked &&
588 		qp->s_acked == qp->s_cur &&
589 		qp->s_cur == qp->s_tail &&
590 		qp->s_tail == qp->s_head;
591 }
592 
593 /**
594  * qp_iter_print - print the qp information to seq_file
595  * @s: the seq_file to emit the qp information on
596  * @iter: the iterator for the qp hash list
597  */
598 void qp_iter_print(struct seq_file *s, struct rvt_qp_iter *iter)
599 {
600 	struct rvt_swqe *wqe;
601 	struct rvt_qp *qp = iter->qp;
602 	struct hfi1_qp_priv *priv = qp->priv;
603 	struct sdma_engine *sde;
604 	struct send_context *send_context;
605 	struct rvt_ack_entry *e = NULL;
606 	struct rvt_srq *srq = qp->ibqp.srq ?
607 		ibsrq_to_rvtsrq(qp->ibqp.srq) : NULL;
608 
609 	sde = qp_to_sdma_engine(qp, priv->s_sc);
610 	wqe = rvt_get_swqe_ptr(qp, qp->s_last);
611 	send_context = qp_to_send_context(qp, priv->s_sc);
612 	if (qp->s_ack_queue)
613 		e = &qp->s_ack_queue[qp->s_tail_ack_queue];
614 	seq_printf(s,
615 		   "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",
616 		   iter->n,
617 		   qp_idle(qp) ? "I" : "B",
618 		   qp->ibqp.qp_num,
619 		   atomic_read(&qp->refcount),
620 		   qp_type_str[qp->ibqp.qp_type],
621 		   qp->state,
622 		   wqe ? wqe->wr.opcode : 0,
623 		   qp->s_flags,
624 		   iowait_sdma_pending(&priv->s_iowait),
625 		   iowait_pio_pending(&priv->s_iowait),
626 		   !list_empty(&priv->s_iowait.list),
627 		   qp->timeout,
628 		   wqe ? wqe->ssn : 0,
629 		   qp->s_lsn,
630 		   qp->s_last_psn,
631 		   qp->s_psn, qp->s_next_psn,
632 		   qp->s_sending_psn, qp->s_sending_hpsn,
633 		   qp->r_psn,
634 		   qp->s_last, qp->s_acked, qp->s_cur,
635 		   qp->s_tail, qp->s_head, qp->s_size,
636 		   qp->s_avail,
637 		   /* ack_queue ring pointers, size */
638 		   qp->s_tail_ack_queue, qp->r_head_ack_queue,
639 		   rvt_max_atomic(&to_idev(qp->ibqp.device)->rdi),
640 		   /* remote QP info  */
641 		   qp->remote_qpn,
642 		   rdma_ah_get_dlid(&qp->remote_ah_attr),
643 		   rdma_ah_get_sl(&qp->remote_ah_attr),
644 		   qp->pmtu,
645 		   qp->s_retry,
646 		   qp->s_retry_cnt,
647 		   qp->s_rnr_retry_cnt,
648 		   qp->s_rnr_retry,
649 		   sde,
650 		   sde ? sde->this_idx : 0,
651 		   send_context,
652 		   send_context ? send_context->sw_index : 0,
653 		   ib_cq_head(qp->ibqp.send_cq),
654 		   ib_cq_tail(qp->ibqp.send_cq),
655 		   qp->pid,
656 		   qp->s_state,
657 		   qp->s_ack_state,
658 		   /* ack queue information */
659 		   e ? e->opcode : 0,
660 		   e ? e->psn : 0,
661 		   e ? e->lpsn : 0,
662 		   qp->r_min_rnr_timer,
663 		   srq ? "SRQ" : "RQ",
664 		   srq ? srq->rq.size : qp->r_rq.size
665 		);
666 }
667 
668 void *qp_priv_alloc(struct rvt_dev_info *rdi, struct rvt_qp *qp)
669 {
670 	struct hfi1_qp_priv *priv;
671 
672 	priv = kzalloc_node(sizeof(*priv), GFP_KERNEL, rdi->dparms.node);
673 	if (!priv)
674 		return ERR_PTR(-ENOMEM);
675 
676 	priv->owner = qp;
677 
678 	priv->s_ahg = kzalloc_node(sizeof(*priv->s_ahg), GFP_KERNEL,
679 				   rdi->dparms.node);
680 	if (!priv->s_ahg) {
681 		kfree(priv);
682 		return ERR_PTR(-ENOMEM);
683 	}
684 	iowait_init(
685 		&priv->s_iowait,
686 		1,
687 		_hfi1_do_send,
688 		_hfi1_do_tid_send,
689 		iowait_sleep,
690 		iowait_wakeup,
691 		iowait_sdma_drained,
692 		hfi1_init_priority);
693 	/* Init to a value to start the running average correctly */
694 	priv->s_running_pkt_size = piothreshold / 2;
695 	return priv;
696 }
697 
698 void qp_priv_free(struct rvt_dev_info *rdi, struct rvt_qp *qp)
699 {
700 	struct hfi1_qp_priv *priv = qp->priv;
701 
702 	hfi1_qp_priv_tid_free(rdi, qp);
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 	hfi1_tid_rdma_flush_wait(qp);
737 }
738 
739 void stop_send_queue(struct rvt_qp *qp)
740 {
741 	struct hfi1_qp_priv *priv = qp->priv;
742 
743 	iowait_cancel_work(&priv->s_iowait);
744 	if (cancel_work_sync(&priv->tid_rdma.trigger_work))
745 		rvt_put_qp(qp);
746 }
747 
748 void quiesce_qp(struct rvt_qp *qp)
749 {
750 	struct hfi1_qp_priv *priv = qp->priv;
751 
752 	hfi1_del_tid_reap_timer(qp);
753 	hfi1_del_tid_retry_timer(qp);
754 	iowait_sdma_drain(&priv->s_iowait);
755 	qp_pio_drain(qp);
756 	flush_tx_list(qp);
757 }
758 
759 void notify_qp_reset(struct rvt_qp *qp)
760 {
761 	hfi1_qp_kern_exp_rcv_clear_all(qp);
762 	qp->r_adefered = 0;
763 	clear_ahg(qp);
764 
765 	/* Clear any OPFN state */
766 	if (qp->ibqp.qp_type == IB_QPT_RC)
767 		opfn_conn_error(qp);
768 }
769 
770 /*
771  * Switch to alternate path.
772  * The QP s_lock should be held and interrupts disabled.
773  */
774 void hfi1_migrate_qp(struct rvt_qp *qp)
775 {
776 	struct hfi1_qp_priv *priv = qp->priv;
777 	struct ib_event ev;
778 
779 	qp->s_mig_state = IB_MIG_MIGRATED;
780 	qp->remote_ah_attr = qp->alt_ah_attr;
781 	qp->port_num = rdma_ah_get_port_num(&qp->alt_ah_attr);
782 	qp->s_pkey_index = qp->s_alt_pkey_index;
783 	qp->s_flags |= HFI1_S_AHG_CLEAR;
784 	priv->s_sc = ah_to_sc(qp->ibqp.device, &qp->remote_ah_attr);
785 	priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
786 	qp_set_16b(qp);
787 
788 	ev.device = qp->ibqp.device;
789 	ev.element.qp = &qp->ibqp;
790 	ev.event = IB_EVENT_PATH_MIG;
791 	qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
792 }
793 
794 int mtu_to_path_mtu(u32 mtu)
795 {
796 	return mtu_to_enum(mtu, OPA_MTU_8192);
797 }
798 
799 u32 mtu_from_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp, u32 pmtu)
800 {
801 	u32 mtu;
802 	struct hfi1_ibdev *verbs_dev = container_of(rdi,
803 						    struct hfi1_ibdev,
804 						    rdi);
805 	struct hfi1_devdata *dd = container_of(verbs_dev,
806 					       struct hfi1_devdata,
807 					       verbs_dev);
808 	struct hfi1_ibport *ibp;
809 	u8 sc, vl;
810 
811 	ibp = &dd->pport[qp->port_num - 1].ibport_data;
812 	sc = ibp->sl_to_sc[rdma_ah_get_sl(&qp->remote_ah_attr)];
813 	vl = sc_to_vlt(dd, sc);
814 
815 	mtu = verbs_mtu_enum_to_int(qp->ibqp.device, pmtu);
816 	if (vl < PER_VL_SEND_CONTEXTS)
817 		mtu = min_t(u32, mtu, dd->vld[vl].mtu);
818 	return mtu;
819 }
820 
821 int get_pmtu_from_attr(struct rvt_dev_info *rdi, struct rvt_qp *qp,
822 		       struct ib_qp_attr *attr)
823 {
824 	int mtu, pidx = qp->port_num - 1;
825 	struct hfi1_ibdev *verbs_dev = container_of(rdi,
826 						    struct hfi1_ibdev,
827 						    rdi);
828 	struct hfi1_devdata *dd = container_of(verbs_dev,
829 					       struct hfi1_devdata,
830 					       verbs_dev);
831 	mtu = verbs_mtu_enum_to_int(qp->ibqp.device, attr->path_mtu);
832 	if (mtu == -1)
833 		return -1; /* values less than 0 are error */
834 
835 	if (mtu > dd->pport[pidx].ibmtu)
836 		return mtu_to_enum(dd->pport[pidx].ibmtu, IB_MTU_2048);
837 	else
838 		return attr->path_mtu;
839 }
840 
841 void notify_error_qp(struct rvt_qp *qp)
842 {
843 	struct hfi1_qp_priv *priv = qp->priv;
844 	seqlock_t *lock = priv->s_iowait.lock;
845 
846 	if (lock) {
847 		write_seqlock(lock);
848 		if (!list_empty(&priv->s_iowait.list) &&
849 		    !(qp->s_flags & RVT_S_BUSY) &&
850 		    !(priv->s_flags & RVT_S_BUSY)) {
851 			qp->s_flags &= ~HFI1_S_ANY_WAIT_IO;
852 			iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_IB);
853 			iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_TID);
854 			list_del_init(&priv->s_iowait.list);
855 			priv->s_iowait.lock = NULL;
856 			rvt_put_qp(qp);
857 		}
858 		write_sequnlock(lock);
859 	}
860 
861 	if (!(qp->s_flags & RVT_S_BUSY) && !(priv->s_flags & RVT_S_BUSY)) {
862 		qp->s_hdrwords = 0;
863 		if (qp->s_rdma_mr) {
864 			rvt_put_mr(qp->s_rdma_mr);
865 			qp->s_rdma_mr = NULL;
866 		}
867 		flush_tx_list(qp);
868 	}
869 }
870 
871 /**
872  * hfi1_qp_iter_cb - callback for iterator
873  * @qp: the qp
874  * @v: the sl in low bits of v
875  *
876  * This is called from the iterator callback to work
877  * on an individual qp.
878  */
879 static void hfi1_qp_iter_cb(struct rvt_qp *qp, u64 v)
880 {
881 	int lastwqe;
882 	struct ib_event ev;
883 	struct hfi1_ibport *ibp =
884 		to_iport(qp->ibqp.device, qp->port_num);
885 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
886 	u8 sl = (u8)v;
887 
888 	if (qp->port_num != ppd->port ||
889 	    (qp->ibqp.qp_type != IB_QPT_UC &&
890 	     qp->ibqp.qp_type != IB_QPT_RC) ||
891 	    rdma_ah_get_sl(&qp->remote_ah_attr) != sl ||
892 	    !(ib_rvt_state_ops[qp->state] & RVT_POST_SEND_OK))
893 		return;
894 
895 	spin_lock_irq(&qp->r_lock);
896 	spin_lock(&qp->s_hlock);
897 	spin_lock(&qp->s_lock);
898 	lastwqe = rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
899 	spin_unlock(&qp->s_lock);
900 	spin_unlock(&qp->s_hlock);
901 	spin_unlock_irq(&qp->r_lock);
902 	if (lastwqe) {
903 		ev.device = qp->ibqp.device;
904 		ev.element.qp = &qp->ibqp;
905 		ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
906 		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
907 	}
908 }
909 
910 /**
911  * hfi1_error_port_qps - put a port's RC/UC qps into error state
912  * @ibp: the ibport.
913  * @sl: the service level.
914  *
915  * This function places all RC/UC qps with a given service level into error
916  * state. It is generally called to force upper lay apps to abandon stale qps
917  * after an sl->sc mapping change.
918  */
919 void hfi1_error_port_qps(struct hfi1_ibport *ibp, u8 sl)
920 {
921 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
922 	struct hfi1_ibdev *dev = &ppd->dd->verbs_dev;
923 
924 	rvt_qp_iter(&dev->rdi, sl, hfi1_qp_iter_cb);
925 }
926