xref: /linux/drivers/infiniband/sw/rxe/rxe_qp.c (revision 4b132aacb0768ac1e652cf517097ea6f237214b9) 
1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /*
3  * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved.
4  * Copyright (c) 2015 System Fabric Works, Inc. All rights reserved.
5  */
6 
7 #include <linux/skbuff.h>
8 #include <linux/delay.h>
9 #include <linux/sched.h>
10 #include <linux/vmalloc.h>
11 #include <rdma/uverbs_ioctl.h>
12 
13 #include "rxe.h"
14 #include "rxe_loc.h"
15 #include "rxe_queue.h"
16 #include "rxe_task.h"
17 
18 static int rxe_qp_chk_cap(struct rxe_dev *rxe, struct ib_qp_cap *cap,
19 			  int has_srq)
20 {
21 	if (cap->max_send_wr > rxe->attr.max_qp_wr) {
22 		rxe_dbg_dev(rxe, "invalid send wr = %u > %d\n",
23 			 cap->max_send_wr, rxe->attr.max_qp_wr);
24 		goto err1;
25 	}
26 
27 	if (cap->max_send_sge > rxe->attr.max_send_sge) {
28 		rxe_dbg_dev(rxe, "invalid send sge = %u > %d\n",
29 			 cap->max_send_sge, rxe->attr.max_send_sge);
30 		goto err1;
31 	}
32 
33 	if (!has_srq) {
34 		if (cap->max_recv_wr > rxe->attr.max_qp_wr) {
35 			rxe_dbg_dev(rxe, "invalid recv wr = %u > %d\n",
36 				 cap->max_recv_wr, rxe->attr.max_qp_wr);
37 			goto err1;
38 		}
39 
40 		if (cap->max_recv_sge > rxe->attr.max_recv_sge) {
41 			rxe_dbg_dev(rxe, "invalid recv sge = %u > %d\n",
42 				 cap->max_recv_sge, rxe->attr.max_recv_sge);
43 			goto err1;
44 		}
45 	}
46 
47 	if (cap->max_inline_data > rxe->max_inline_data) {
48 		rxe_dbg_dev(rxe, "invalid max inline data = %u > %d\n",
49 			 cap->max_inline_data, rxe->max_inline_data);
50 		goto err1;
51 	}
52 
53 	return 0;
54 
55 err1:
56 	return -EINVAL;
57 }
58 
59 int rxe_qp_chk_init(struct rxe_dev *rxe, struct ib_qp_init_attr *init)
60 {
61 	struct ib_qp_cap *cap = &init->cap;
62 	struct rxe_port *port;
63 	int port_num = init->port_num;
64 
65 	switch (init->qp_type) {
66 	case IB_QPT_GSI:
67 	case IB_QPT_RC:
68 	case IB_QPT_UC:
69 	case IB_QPT_UD:
70 		break;
71 	default:
72 		return -EOPNOTSUPP;
73 	}
74 
75 	if (!init->recv_cq || !init->send_cq) {
76 		rxe_dbg_dev(rxe, "missing cq\n");
77 		goto err1;
78 	}
79 
80 	if (rxe_qp_chk_cap(rxe, cap, !!init->srq))
81 		goto err1;
82 
83 	if (init->qp_type == IB_QPT_GSI) {
84 		if (!rdma_is_port_valid(&rxe->ib_dev, port_num)) {
85 			rxe_dbg_dev(rxe, "invalid port = %d\n", port_num);
86 			goto err1;
87 		}
88 
89 		port = &rxe->port;
90 
91 		if (init->qp_type == IB_QPT_GSI && port->qp_gsi_index) {
92 			rxe_dbg_dev(rxe, "GSI QP exists for port %d\n", port_num);
93 			goto err1;
94 		}
95 	}
96 
97 	return 0;
98 
99 err1:
100 	return -EINVAL;
101 }
102 
103 static int alloc_rd_atomic_resources(struct rxe_qp *qp, unsigned int n)
104 {
105 	qp->resp.res_head = 0;
106 	qp->resp.res_tail = 0;
107 	qp->resp.resources = kcalloc(n, sizeof(struct resp_res), GFP_KERNEL);
108 
109 	if (!qp->resp.resources)
110 		return -ENOMEM;
111 
112 	return 0;
113 }
114 
115 static void free_rd_atomic_resources(struct rxe_qp *qp)
116 {
117 	if (qp->resp.resources) {
118 		int i;
119 
120 		for (i = 0; i < qp->attr.max_dest_rd_atomic; i++) {
121 			struct resp_res *res = &qp->resp.resources[i];
122 
123 			free_rd_atomic_resource(res);
124 		}
125 		kfree(qp->resp.resources);
126 		qp->resp.resources = NULL;
127 	}
128 }
129 
130 void free_rd_atomic_resource(struct resp_res *res)
131 {
132 	res->type = 0;
133 }
134 
135 static void cleanup_rd_atomic_resources(struct rxe_qp *qp)
136 {
137 	int i;
138 	struct resp_res *res;
139 
140 	if (qp->resp.resources) {
141 		for (i = 0; i < qp->attr.max_dest_rd_atomic; i++) {
142 			res = &qp->resp.resources[i];
143 			free_rd_atomic_resource(res);
144 		}
145 	}
146 }
147 
148 static void rxe_qp_init_misc(struct rxe_dev *rxe, struct rxe_qp *qp,
149 			     struct ib_qp_init_attr *init)
150 {
151 	struct rxe_port *port;
152 	u32 qpn;
153 
154 	qp->sq_sig_type		= init->sq_sig_type;
155 	qp->attr.path_mtu	= 1;
156 	qp->mtu			= ib_mtu_enum_to_int(qp->attr.path_mtu);
157 
158 	qpn			= qp->elem.index;
159 	port			= &rxe->port;
160 
161 	switch (init->qp_type) {
162 	case IB_QPT_GSI:
163 		qp->ibqp.qp_num		= 1;
164 		port->qp_gsi_index	= qpn;
165 		qp->attr.port_num	= init->port_num;
166 		break;
167 
168 	default:
169 		qp->ibqp.qp_num		= qpn;
170 		break;
171 	}
172 
173 	spin_lock_init(&qp->state_lock);
174 
175 	spin_lock_init(&qp->sq.sq_lock);
176 	spin_lock_init(&qp->rq.producer_lock);
177 	spin_lock_init(&qp->rq.consumer_lock);
178 
179 	skb_queue_head_init(&qp->req_pkts);
180 	skb_queue_head_init(&qp->resp_pkts);
181 
182 	atomic_set(&qp->ssn, 0);
183 	atomic_set(&qp->skb_out, 0);
184 }
185 
186 static int rxe_init_sq(struct rxe_qp *qp, struct ib_qp_init_attr *init,
187 		       struct ib_udata *udata,
188 		       struct rxe_create_qp_resp __user *uresp)
189 {
190 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
191 	int wqe_size;
192 	int err;
193 
194 	qp->sq.max_wr = init->cap.max_send_wr;
195 	wqe_size = max_t(int, init->cap.max_send_sge * sizeof(struct ib_sge),
196 			 init->cap.max_inline_data);
197 	qp->sq.max_sge = wqe_size / sizeof(struct ib_sge);
198 	qp->sq.max_inline = wqe_size;
199 	wqe_size += sizeof(struct rxe_send_wqe);
200 
201 	qp->sq.queue = rxe_queue_init(rxe, &qp->sq.max_wr, wqe_size,
202 				      QUEUE_TYPE_FROM_CLIENT);
203 	if (!qp->sq.queue) {
204 		rxe_err_qp(qp, "Unable to allocate send queue\n");
205 		err = -ENOMEM;
206 		goto err_out;
207 	}
208 
209 	/* prepare info for caller to mmap send queue if user space qp */
210 	err = do_mmap_info(rxe, uresp ? &uresp->sq_mi : NULL, udata,
211 			   qp->sq.queue->buf, qp->sq.queue->buf_size,
212 			   &qp->sq.queue->ip);
213 	if (err) {
214 		rxe_err_qp(qp, "do_mmap_info failed, err = %d\n", err);
215 		goto err_free;
216 	}
217 
218 	/* return actual capabilities to caller which may be larger
219 	 * than requested
220 	 */
221 	init->cap.max_send_wr = qp->sq.max_wr;
222 	init->cap.max_send_sge = qp->sq.max_sge;
223 	init->cap.max_inline_data = qp->sq.max_inline;
224 
225 	return 0;
226 
227 err_free:
228 	vfree(qp->sq.queue->buf);
229 	kfree(qp->sq.queue);
230 	qp->sq.queue = NULL;
231 err_out:
232 	return err;
233 }
234 
235 static int rxe_qp_init_req(struct rxe_dev *rxe, struct rxe_qp *qp,
236 			   struct ib_qp_init_attr *init, struct ib_udata *udata,
237 			   struct rxe_create_qp_resp __user *uresp)
238 {
239 	int err;
240 
241 	/* if we don't finish qp create make sure queue is valid */
242 	skb_queue_head_init(&qp->req_pkts);
243 
244 	err = sock_create_kern(&init_net, AF_INET, SOCK_DGRAM, 0, &qp->sk);
245 	if (err < 0)
246 		return err;
247 	qp->sk->sk->sk_user_data = (void *)(uintptr_t)qp->elem.index;
248 
249 	/* pick a source UDP port number for this QP based on
250 	 * the source QPN. this spreads traffic for different QPs
251 	 * across different NIC RX queues (while using a single
252 	 * flow for a given QP to maintain packet order).
253 	 * the port number must be in the Dynamic Ports range
254 	 * (0xc000 - 0xffff).
255 	 */
256 	qp->src_port = RXE_ROCE_V2_SPORT + (hash_32(qp_num(qp), 14) & 0x3fff);
257 
258 	err = rxe_init_sq(qp, init, udata, uresp);
259 	if (err)
260 		return err;
261 
262 	qp->req.wqe_index = queue_get_producer(qp->sq.queue,
263 					       QUEUE_TYPE_FROM_CLIENT);
264 
265 	qp->req.opcode		= -1;
266 	qp->comp.opcode		= -1;
267 
268 	rxe_init_task(&qp->send_task, qp, rxe_sender);
269 
270 	qp->qp_timeout_jiffies = 0; /* Can't be set for UD/UC in modify_qp */
271 	if (init->qp_type == IB_QPT_RC) {
272 		timer_setup(&qp->rnr_nak_timer, rnr_nak_timer, 0);
273 		timer_setup(&qp->retrans_timer, retransmit_timer, 0);
274 	}
275 	return 0;
276 }
277 
278 static int rxe_init_rq(struct rxe_qp *qp, struct ib_qp_init_attr *init,
279 		       struct ib_udata *udata,
280 		       struct rxe_create_qp_resp __user *uresp)
281 {
282 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
283 	int wqe_size;
284 	int err;
285 
286 	qp->rq.max_wr = init->cap.max_recv_wr;
287 	qp->rq.max_sge = init->cap.max_recv_sge;
288 	wqe_size = sizeof(struct rxe_recv_wqe) +
289 				qp->rq.max_sge*sizeof(struct ib_sge);
290 
291 	qp->rq.queue = rxe_queue_init(rxe, &qp->rq.max_wr, wqe_size,
292 				      QUEUE_TYPE_FROM_CLIENT);
293 	if (!qp->rq.queue) {
294 		rxe_err_qp(qp, "Unable to allocate recv queue\n");
295 		err = -ENOMEM;
296 		goto err_out;
297 	}
298 
299 	/* prepare info for caller to mmap recv queue if user space qp */
300 	err = do_mmap_info(rxe, uresp ? &uresp->rq_mi : NULL, udata,
301 			   qp->rq.queue->buf, qp->rq.queue->buf_size,
302 			   &qp->rq.queue->ip);
303 	if (err) {
304 		rxe_err_qp(qp, "do_mmap_info failed, err = %d\n", err);
305 		goto err_free;
306 	}
307 
308 	/* return actual capabilities to caller which may be larger
309 	 * than requested
310 	 */
311 	init->cap.max_recv_wr = qp->rq.max_wr;
312 
313 	return 0;
314 
315 err_free:
316 	vfree(qp->rq.queue->buf);
317 	kfree(qp->rq.queue);
318 	qp->rq.queue = NULL;
319 err_out:
320 	return err;
321 }
322 
323 static int rxe_qp_init_resp(struct rxe_dev *rxe, struct rxe_qp *qp,
324 			    struct ib_qp_init_attr *init,
325 			    struct ib_udata *udata,
326 			    struct rxe_create_qp_resp __user *uresp)
327 {
328 	int err;
329 
330 	/* if we don't finish qp create make sure queue is valid */
331 	skb_queue_head_init(&qp->resp_pkts);
332 
333 	if (!qp->srq) {
334 		err = rxe_init_rq(qp, init, udata, uresp);
335 		if (err)
336 			return err;
337 	}
338 
339 	rxe_init_task(&qp->recv_task, qp, rxe_receiver);
340 
341 	qp->resp.opcode		= OPCODE_NONE;
342 	qp->resp.msn		= 0;
343 
344 	return 0;
345 }
346 
347 /* called by the create qp verb */
348 int rxe_qp_from_init(struct rxe_dev *rxe, struct rxe_qp *qp, struct rxe_pd *pd,
349 		     struct ib_qp_init_attr *init,
350 		     struct rxe_create_qp_resp __user *uresp,
351 		     struct ib_pd *ibpd,
352 		     struct ib_udata *udata)
353 {
354 	int err;
355 	struct rxe_cq *rcq = to_rcq(init->recv_cq);
356 	struct rxe_cq *scq = to_rcq(init->send_cq);
357 	struct rxe_srq *srq = init->srq ? to_rsrq(init->srq) : NULL;
358 	unsigned long flags;
359 
360 	rxe_get(pd);
361 	rxe_get(rcq);
362 	rxe_get(scq);
363 	if (srq)
364 		rxe_get(srq);
365 
366 	qp->pd = pd;
367 	qp->rcq = rcq;
368 	qp->scq = scq;
369 	qp->srq = srq;
370 
371 	atomic_inc(&rcq->num_wq);
372 	atomic_inc(&scq->num_wq);
373 
374 	rxe_qp_init_misc(rxe, qp, init);
375 
376 	err = rxe_qp_init_req(rxe, qp, init, udata, uresp);
377 	if (err)
378 		goto err1;
379 
380 	err = rxe_qp_init_resp(rxe, qp, init, udata, uresp);
381 	if (err)
382 		goto err2;
383 
384 	spin_lock_irqsave(&qp->state_lock, flags);
385 	qp->attr.qp_state = IB_QPS_RESET;
386 	qp->valid = 1;
387 	spin_unlock_irqrestore(&qp->state_lock, flags);
388 
389 	return 0;
390 
391 err2:
392 	rxe_queue_cleanup(qp->sq.queue);
393 	qp->sq.queue = NULL;
394 err1:
395 	atomic_dec(&rcq->num_wq);
396 	atomic_dec(&scq->num_wq);
397 
398 	qp->pd = NULL;
399 	qp->rcq = NULL;
400 	qp->scq = NULL;
401 	qp->srq = NULL;
402 
403 	if (srq)
404 		rxe_put(srq);
405 	rxe_put(scq);
406 	rxe_put(rcq);
407 	rxe_put(pd);
408 
409 	return err;
410 }
411 
412 /* called by the query qp verb */
413 int rxe_qp_to_init(struct rxe_qp *qp, struct ib_qp_init_attr *init)
414 {
415 	init->event_handler		= qp->ibqp.event_handler;
416 	init->qp_context		= qp->ibqp.qp_context;
417 	init->send_cq			= qp->ibqp.send_cq;
418 	init->recv_cq			= qp->ibqp.recv_cq;
419 	init->srq			= qp->ibqp.srq;
420 
421 	init->cap.max_send_wr		= qp->sq.max_wr;
422 	init->cap.max_send_sge		= qp->sq.max_sge;
423 	init->cap.max_inline_data	= qp->sq.max_inline;
424 
425 	if (!qp->srq) {
426 		init->cap.max_recv_wr		= qp->rq.max_wr;
427 		init->cap.max_recv_sge		= qp->rq.max_sge;
428 	}
429 
430 	init->sq_sig_type		= qp->sq_sig_type;
431 
432 	init->qp_type			= qp->ibqp.qp_type;
433 	init->port_num			= 1;
434 
435 	return 0;
436 }
437 
438 int rxe_qp_chk_attr(struct rxe_dev *rxe, struct rxe_qp *qp,
439 		    struct ib_qp_attr *attr, int mask)
440 {
441 	if (mask & IB_QP_PORT) {
442 		if (!rdma_is_port_valid(&rxe->ib_dev, attr->port_num)) {
443 			rxe_dbg_qp(qp, "invalid port %d\n", attr->port_num);
444 			goto err1;
445 		}
446 	}
447 
448 	if (mask & IB_QP_CAP && rxe_qp_chk_cap(rxe, &attr->cap, !!qp->srq))
449 		goto err1;
450 
451 	if (mask & IB_QP_ACCESS_FLAGS) {
452 		if (!(qp_type(qp) == IB_QPT_RC || qp_type(qp) == IB_QPT_UC))
453 			goto err1;
454 		if (attr->qp_access_flags & ~RXE_ACCESS_SUPPORTED_QP)
455 			goto err1;
456 	}
457 
458 	if (mask & IB_QP_AV && rxe_av_chk_attr(qp, &attr->ah_attr))
459 		goto err1;
460 
461 	if (mask & IB_QP_ALT_PATH) {
462 		if (rxe_av_chk_attr(qp, &attr->alt_ah_attr))
463 			goto err1;
464 		if (!rdma_is_port_valid(&rxe->ib_dev, attr->alt_port_num))  {
465 			rxe_dbg_qp(qp, "invalid alt port %d\n", attr->alt_port_num);
466 			goto err1;
467 		}
468 		if (attr->alt_timeout > 31) {
469 			rxe_dbg_qp(qp, "invalid alt timeout %d > 31\n",
470 				 attr->alt_timeout);
471 			goto err1;
472 		}
473 	}
474 
475 	if (mask & IB_QP_PATH_MTU) {
476 		struct rxe_port *port = &rxe->port;
477 
478 		enum ib_mtu max_mtu = port->attr.max_mtu;
479 		enum ib_mtu mtu = attr->path_mtu;
480 
481 		if (mtu > max_mtu) {
482 			rxe_dbg_qp(qp, "invalid mtu (%d) > (%d)\n",
483 				 ib_mtu_enum_to_int(mtu),
484 				 ib_mtu_enum_to_int(max_mtu));
485 			goto err1;
486 		}
487 	}
488 
489 	if (mask & IB_QP_MAX_QP_RD_ATOMIC) {
490 		if (attr->max_rd_atomic > rxe->attr.max_qp_rd_atom) {
491 			rxe_dbg_qp(qp, "invalid max_rd_atomic %d > %d\n",
492 				 attr->max_rd_atomic,
493 				 rxe->attr.max_qp_rd_atom);
494 			goto err1;
495 		}
496 	}
497 
498 	if (mask & IB_QP_TIMEOUT) {
499 		if (attr->timeout > 31) {
500 			rxe_dbg_qp(qp, "invalid timeout %d > 31\n",
501 					attr->timeout);
502 			goto err1;
503 		}
504 	}
505 
506 	return 0;
507 
508 err1:
509 	return -EINVAL;
510 }
511 
512 /* move the qp to the reset state */
513 static void rxe_qp_reset(struct rxe_qp *qp)
514 {
515 	/* stop tasks from running */
516 	rxe_disable_task(&qp->recv_task);
517 	rxe_disable_task(&qp->send_task);
518 
519 	/* drain work and packet queuesc */
520 	rxe_sender(qp);
521 	rxe_receiver(qp);
522 
523 	if (qp->rq.queue)
524 		rxe_queue_reset(qp->rq.queue);
525 	if (qp->sq.queue)
526 		rxe_queue_reset(qp->sq.queue);
527 
528 	/* cleanup attributes */
529 	atomic_set(&qp->ssn, 0);
530 	qp->req.opcode = -1;
531 	qp->req.need_retry = 0;
532 	qp->req.wait_for_rnr_timer = 0;
533 	qp->req.noack_pkts = 0;
534 	qp->resp.msn = 0;
535 	qp->resp.opcode = -1;
536 	qp->resp.drop_msg = 0;
537 	qp->resp.goto_error = 0;
538 	qp->resp.sent_psn_nak = 0;
539 
540 	if (qp->resp.mr) {
541 		rxe_put(qp->resp.mr);
542 		qp->resp.mr = NULL;
543 	}
544 
545 	cleanup_rd_atomic_resources(qp);
546 
547 	/* reenable tasks */
548 	rxe_enable_task(&qp->recv_task);
549 	rxe_enable_task(&qp->send_task);
550 }
551 
552 /* move the qp to the error state */
553 void rxe_qp_error(struct rxe_qp *qp)
554 {
555 	unsigned long flags;
556 
557 	spin_lock_irqsave(&qp->state_lock, flags);
558 	qp->attr.qp_state = IB_QPS_ERR;
559 
560 	/* drain work and packet queues */
561 	rxe_sched_task(&qp->recv_task);
562 	rxe_sched_task(&qp->send_task);
563 	spin_unlock_irqrestore(&qp->state_lock, flags);
564 }
565 
566 static void rxe_qp_sqd(struct rxe_qp *qp, struct ib_qp_attr *attr,
567 		       int mask)
568 {
569 	unsigned long flags;
570 
571 	spin_lock_irqsave(&qp->state_lock, flags);
572 	qp->attr.sq_draining = 1;
573 	rxe_sched_task(&qp->send_task);
574 	spin_unlock_irqrestore(&qp->state_lock, flags);
575 }
576 
577 /* caller should hold qp->state_lock */
578 static int __qp_chk_state(struct rxe_qp *qp, struct ib_qp_attr *attr,
579 			    int mask)
580 {
581 	enum ib_qp_state cur_state;
582 	enum ib_qp_state new_state;
583 
584 	cur_state = (mask & IB_QP_CUR_STATE) ?
585 				attr->cur_qp_state : qp->attr.qp_state;
586 	new_state = (mask & IB_QP_STATE) ?
587 				attr->qp_state : cur_state;
588 
589 	if (!ib_modify_qp_is_ok(cur_state, new_state, qp_type(qp), mask))
590 		return -EINVAL;
591 
592 	if (mask & IB_QP_STATE && cur_state == IB_QPS_SQD) {
593 		if (qp->attr.sq_draining && new_state != IB_QPS_ERR)
594 			return -EINVAL;
595 	}
596 
597 	return 0;
598 }
599 
600 static const char *const qps2str[] = {
601 	[IB_QPS_RESET]	= "RESET",
602 	[IB_QPS_INIT]	= "INIT",
603 	[IB_QPS_RTR]	= "RTR",
604 	[IB_QPS_RTS]	= "RTS",
605 	[IB_QPS_SQD]	= "SQD",
606 	[IB_QPS_SQE]	= "SQE",
607 	[IB_QPS_ERR]	= "ERR",
608 };
609 
610 /* called by the modify qp verb */
611 int rxe_qp_from_attr(struct rxe_qp *qp, struct ib_qp_attr *attr, int mask,
612 		     struct ib_udata *udata)
613 {
614 	int err;
615 
616 	if (mask & IB_QP_CUR_STATE)
617 		qp->attr.cur_qp_state = attr->qp_state;
618 
619 	if (mask & IB_QP_STATE) {
620 		unsigned long flags;
621 
622 		spin_lock_irqsave(&qp->state_lock, flags);
623 		err = __qp_chk_state(qp, attr, mask);
624 		if (!err) {
625 			qp->attr.qp_state = attr->qp_state;
626 			rxe_dbg_qp(qp, "state -> %s\n",
627 					qps2str[attr->qp_state]);
628 		}
629 		spin_unlock_irqrestore(&qp->state_lock, flags);
630 
631 		if (err)
632 			return err;
633 
634 		switch (attr->qp_state) {
635 		case IB_QPS_RESET:
636 			rxe_qp_reset(qp);
637 			break;
638 		case IB_QPS_SQD:
639 			rxe_qp_sqd(qp, attr, mask);
640 			break;
641 		case IB_QPS_ERR:
642 			rxe_qp_error(qp);
643 			break;
644 		default:
645 			break;
646 		}
647 	}
648 
649 	if (mask & IB_QP_MAX_QP_RD_ATOMIC) {
650 		int max_rd_atomic = attr->max_rd_atomic ?
651 			roundup_pow_of_two(attr->max_rd_atomic) : 0;
652 
653 		qp->attr.max_rd_atomic = max_rd_atomic;
654 		atomic_set(&qp->req.rd_atomic, max_rd_atomic);
655 	}
656 
657 	if (mask & IB_QP_MAX_DEST_RD_ATOMIC) {
658 		int max_dest_rd_atomic = attr->max_dest_rd_atomic ?
659 			roundup_pow_of_two(attr->max_dest_rd_atomic) : 0;
660 
661 		qp->attr.max_dest_rd_atomic = max_dest_rd_atomic;
662 
663 		free_rd_atomic_resources(qp);
664 
665 		err = alloc_rd_atomic_resources(qp, max_dest_rd_atomic);
666 		if (err)
667 			return err;
668 	}
669 
670 	if (mask & IB_QP_EN_SQD_ASYNC_NOTIFY)
671 		qp->attr.en_sqd_async_notify = attr->en_sqd_async_notify;
672 
673 	if (mask & IB_QP_ACCESS_FLAGS)
674 		qp->attr.qp_access_flags = attr->qp_access_flags;
675 
676 	if (mask & IB_QP_PKEY_INDEX)
677 		qp->attr.pkey_index = attr->pkey_index;
678 
679 	if (mask & IB_QP_PORT)
680 		qp->attr.port_num = attr->port_num;
681 
682 	if (mask & IB_QP_QKEY)
683 		qp->attr.qkey = attr->qkey;
684 
685 	if (mask & IB_QP_AV)
686 		rxe_init_av(&attr->ah_attr, &qp->pri_av);
687 
688 	if (mask & IB_QP_ALT_PATH) {
689 		rxe_init_av(&attr->alt_ah_attr, &qp->alt_av);
690 		qp->attr.alt_port_num = attr->alt_port_num;
691 		qp->attr.alt_pkey_index = attr->alt_pkey_index;
692 		qp->attr.alt_timeout = attr->alt_timeout;
693 	}
694 
695 	if (mask & IB_QP_PATH_MTU) {
696 		qp->attr.path_mtu = attr->path_mtu;
697 		qp->mtu = ib_mtu_enum_to_int(attr->path_mtu);
698 	}
699 
700 	if (mask & IB_QP_TIMEOUT) {
701 		qp->attr.timeout = attr->timeout;
702 		if (attr->timeout == 0) {
703 			qp->qp_timeout_jiffies = 0;
704 		} else {
705 			/* According to the spec, timeout = 4.096 * 2 ^ attr->timeout [us] */
706 			int j = nsecs_to_jiffies(4096ULL << attr->timeout);
707 
708 			qp->qp_timeout_jiffies = j ? j : 1;
709 		}
710 	}
711 
712 	if (mask & IB_QP_RETRY_CNT) {
713 		qp->attr.retry_cnt = attr->retry_cnt;
714 		qp->comp.retry_cnt = attr->retry_cnt;
715 		rxe_dbg_qp(qp, "set retry count = %d\n", attr->retry_cnt);
716 	}
717 
718 	if (mask & IB_QP_RNR_RETRY) {
719 		qp->attr.rnr_retry = attr->rnr_retry;
720 		qp->comp.rnr_retry = attr->rnr_retry;
721 		rxe_dbg_qp(qp, "set rnr retry count = %d\n", attr->rnr_retry);
722 	}
723 
724 	if (mask & IB_QP_RQ_PSN) {
725 		qp->attr.rq_psn = (attr->rq_psn & BTH_PSN_MASK);
726 		qp->resp.psn = qp->attr.rq_psn;
727 		rxe_dbg_qp(qp, "set resp psn = 0x%x\n", qp->resp.psn);
728 	}
729 
730 	if (mask & IB_QP_MIN_RNR_TIMER) {
731 		qp->attr.min_rnr_timer = attr->min_rnr_timer;
732 		rxe_dbg_qp(qp, "set min rnr timer = 0x%x\n",
733 			 attr->min_rnr_timer);
734 	}
735 
736 	if (mask & IB_QP_SQ_PSN) {
737 		qp->attr.sq_psn = (attr->sq_psn & BTH_PSN_MASK);
738 		qp->req.psn = qp->attr.sq_psn;
739 		qp->comp.psn = qp->attr.sq_psn;
740 		rxe_dbg_qp(qp, "set req psn = 0x%x\n", qp->req.psn);
741 	}
742 
743 	if (mask & IB_QP_PATH_MIG_STATE)
744 		qp->attr.path_mig_state = attr->path_mig_state;
745 
746 	if (mask & IB_QP_DEST_QPN)
747 		qp->attr.dest_qp_num = attr->dest_qp_num;
748 
749 	return 0;
750 }
751 
752 /* called by the query qp verb */
753 int rxe_qp_to_attr(struct rxe_qp *qp, struct ib_qp_attr *attr, int mask)
754 {
755 	unsigned long flags;
756 
757 	*attr = qp->attr;
758 
759 	attr->rq_psn				= qp->resp.psn;
760 	attr->sq_psn				= qp->req.psn;
761 
762 	attr->cap.max_send_wr			= qp->sq.max_wr;
763 	attr->cap.max_send_sge			= qp->sq.max_sge;
764 	attr->cap.max_inline_data		= qp->sq.max_inline;
765 
766 	if (!qp->srq) {
767 		attr->cap.max_recv_wr		= qp->rq.max_wr;
768 		attr->cap.max_recv_sge		= qp->rq.max_sge;
769 	}
770 
771 	rxe_av_to_attr(&qp->pri_av, &attr->ah_attr);
772 	rxe_av_to_attr(&qp->alt_av, &attr->alt_ah_attr);
773 
774 	/* Applications that get this state typically spin on it.
775 	 * Yield the processor
776 	 */
777 	spin_lock_irqsave(&qp->state_lock, flags);
778 	if (qp->attr.sq_draining) {
779 		spin_unlock_irqrestore(&qp->state_lock, flags);
780 		cond_resched();
781 	} else {
782 		spin_unlock_irqrestore(&qp->state_lock, flags);
783 	}
784 
785 	return 0;
786 }
787 
788 int rxe_qp_chk_destroy(struct rxe_qp *qp)
789 {
790 	/* See IBA o10-2.2.3
791 	 * An attempt to destroy a QP while attached to a mcast group
792 	 * will fail immediately.
793 	 */
794 	if (atomic_read(&qp->mcg_num)) {
795 		rxe_dbg_qp(qp, "Attempt to destroy while attached to multicast group\n");
796 		return -EBUSY;
797 	}
798 
799 	return 0;
800 }
801 
802 /* called when the last reference to the qp is dropped */
803 static void rxe_qp_do_cleanup(struct work_struct *work)
804 {
805 	struct rxe_qp *qp = container_of(work, typeof(*qp), cleanup_work.work);
806 	unsigned long flags;
807 
808 	spin_lock_irqsave(&qp->state_lock, flags);
809 	qp->valid = 0;
810 	spin_unlock_irqrestore(&qp->state_lock, flags);
811 	qp->qp_timeout_jiffies = 0;
812 
813 	if (qp_type(qp) == IB_QPT_RC) {
814 		del_timer_sync(&qp->retrans_timer);
815 		del_timer_sync(&qp->rnr_nak_timer);
816 	}
817 
818 	if (qp->recv_task.func)
819 		rxe_cleanup_task(&qp->recv_task);
820 
821 	if (qp->send_task.func)
822 		rxe_cleanup_task(&qp->send_task);
823 
824 	/* flush out any receive wr's or pending requests */
825 	rxe_sender(qp);
826 	rxe_receiver(qp);
827 
828 	if (qp->sq.queue)
829 		rxe_queue_cleanup(qp->sq.queue);
830 
831 	if (qp->srq)
832 		rxe_put(qp->srq);
833 
834 	if (qp->rq.queue)
835 		rxe_queue_cleanup(qp->rq.queue);
836 
837 	if (qp->scq) {
838 		atomic_dec(&qp->scq->num_wq);
839 		rxe_put(qp->scq);
840 	}
841 
842 	if (qp->rcq) {
843 		atomic_dec(&qp->rcq->num_wq);
844 		rxe_put(qp->rcq);
845 	}
846 
847 	if (qp->pd)
848 		rxe_put(qp->pd);
849 
850 	if (qp->resp.mr)
851 		rxe_put(qp->resp.mr);
852 
853 	free_rd_atomic_resources(qp);
854 
855 	if (qp->sk) {
856 		if (qp_type(qp) == IB_QPT_RC)
857 			sk_dst_reset(qp->sk->sk);
858 
859 		kernel_sock_shutdown(qp->sk, SHUT_RDWR);
860 		sock_release(qp->sk);
861 	}
862 }
863 
864 /* called when the last reference to the qp is dropped */
865 void rxe_qp_cleanup(struct rxe_pool_elem *elem)
866 {
867 	struct rxe_qp *qp = container_of(elem, typeof(*qp), elem);
868 
869 	execute_in_process_context(rxe_qp_do_cleanup, &qp->cleanup_work);
870 }
871