xref: /linux/drivers/infiniband/sw/rxe/rxe_qp.c (revision 44f57d78)
1 /*
2  * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved.
3  * Copyright (c) 2015 System Fabric Works, Inc. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *	   Redistribution and use in source and binary forms, with or
12  *	   without modification, are permitted provided that the following
13  *	   conditions are met:
14  *
15  *		- Redistributions of source code must retain the above
16  *		  copyright notice, this list of conditions and the following
17  *		  disclaimer.
18  *
19  *		- Redistributions in binary form must reproduce the above
20  *		  copyright notice, this list of conditions and the following
21  *		  disclaimer in the documentation and/or other materials
22  *		  provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33 
34 #include <linux/skbuff.h>
35 #include <linux/delay.h>
36 #include <linux/sched.h>
37 #include <linux/vmalloc.h>
38 #include <rdma/uverbs_ioctl.h>
39 
40 #include "rxe.h"
41 #include "rxe_loc.h"
42 #include "rxe_queue.h"
43 #include "rxe_task.h"
44 
45 static int rxe_qp_chk_cap(struct rxe_dev *rxe, struct ib_qp_cap *cap,
46 			  int has_srq)
47 {
48 	if (cap->max_send_wr > rxe->attr.max_qp_wr) {
49 		pr_warn("invalid send wr = %d > %d\n",
50 			cap->max_send_wr, rxe->attr.max_qp_wr);
51 		goto err1;
52 	}
53 
54 	if (cap->max_send_sge > rxe->attr.max_send_sge) {
55 		pr_warn("invalid send sge = %d > %d\n",
56 			cap->max_send_sge, rxe->attr.max_send_sge);
57 		goto err1;
58 	}
59 
60 	if (!has_srq) {
61 		if (cap->max_recv_wr > rxe->attr.max_qp_wr) {
62 			pr_warn("invalid recv wr = %d > %d\n",
63 				cap->max_recv_wr, rxe->attr.max_qp_wr);
64 			goto err1;
65 		}
66 
67 		if (cap->max_recv_sge > rxe->attr.max_recv_sge) {
68 			pr_warn("invalid recv sge = %d > %d\n",
69 				cap->max_recv_sge, rxe->attr.max_recv_sge);
70 			goto err1;
71 		}
72 	}
73 
74 	if (cap->max_inline_data > rxe->max_inline_data) {
75 		pr_warn("invalid max inline data = %d > %d\n",
76 			cap->max_inline_data, rxe->max_inline_data);
77 		goto err1;
78 	}
79 
80 	return 0;
81 
82 err1:
83 	return -EINVAL;
84 }
85 
86 int rxe_qp_chk_init(struct rxe_dev *rxe, struct ib_qp_init_attr *init)
87 {
88 	struct ib_qp_cap *cap = &init->cap;
89 	struct rxe_port *port;
90 	int port_num = init->port_num;
91 
92 	if (!init->recv_cq || !init->send_cq) {
93 		pr_warn("missing cq\n");
94 		goto err1;
95 	}
96 
97 	if (rxe_qp_chk_cap(rxe, cap, !!init->srq))
98 		goto err1;
99 
100 	if (init->qp_type == IB_QPT_SMI || init->qp_type == IB_QPT_GSI) {
101 		if (!rdma_is_port_valid(&rxe->ib_dev, port_num)) {
102 			pr_warn("invalid port = %d\n", port_num);
103 			goto err1;
104 		}
105 
106 		port = &rxe->port;
107 
108 		if (init->qp_type == IB_QPT_SMI && port->qp_smi_index) {
109 			pr_warn("SMI QP exists for port %d\n", port_num);
110 			goto err1;
111 		}
112 
113 		if (init->qp_type == IB_QPT_GSI && port->qp_gsi_index) {
114 			pr_warn("GSI QP exists for port %d\n", port_num);
115 			goto err1;
116 		}
117 	}
118 
119 	return 0;
120 
121 err1:
122 	return -EINVAL;
123 }
124 
125 static int alloc_rd_atomic_resources(struct rxe_qp *qp, unsigned int n)
126 {
127 	qp->resp.res_head = 0;
128 	qp->resp.res_tail = 0;
129 	qp->resp.resources = kcalloc(n, sizeof(struct resp_res), GFP_KERNEL);
130 
131 	if (!qp->resp.resources)
132 		return -ENOMEM;
133 
134 	return 0;
135 }
136 
137 static void free_rd_atomic_resources(struct rxe_qp *qp)
138 {
139 	if (qp->resp.resources) {
140 		int i;
141 
142 		for (i = 0; i < qp->attr.max_dest_rd_atomic; i++) {
143 			struct resp_res *res = &qp->resp.resources[i];
144 
145 			free_rd_atomic_resource(qp, res);
146 		}
147 		kfree(qp->resp.resources);
148 		qp->resp.resources = NULL;
149 	}
150 }
151 
152 void free_rd_atomic_resource(struct rxe_qp *qp, struct resp_res *res)
153 {
154 	if (res->type == RXE_ATOMIC_MASK) {
155 		rxe_drop_ref(qp);
156 		kfree_skb(res->atomic.skb);
157 	} else if (res->type == RXE_READ_MASK) {
158 		if (res->read.mr)
159 			rxe_drop_ref(res->read.mr);
160 	}
161 	res->type = 0;
162 }
163 
164 static void cleanup_rd_atomic_resources(struct rxe_qp *qp)
165 {
166 	int i;
167 	struct resp_res *res;
168 
169 	if (qp->resp.resources) {
170 		for (i = 0; i < qp->attr.max_dest_rd_atomic; i++) {
171 			res = &qp->resp.resources[i];
172 			free_rd_atomic_resource(qp, res);
173 		}
174 	}
175 }
176 
177 static void rxe_qp_init_misc(struct rxe_dev *rxe, struct rxe_qp *qp,
178 			     struct ib_qp_init_attr *init)
179 {
180 	struct rxe_port *port;
181 	u32 qpn;
182 
183 	qp->sq_sig_type		= init->sq_sig_type;
184 	qp->attr.path_mtu	= 1;
185 	qp->mtu			= ib_mtu_enum_to_int(qp->attr.path_mtu);
186 
187 	qpn			= qp->pelem.index;
188 	port			= &rxe->port;
189 
190 	switch (init->qp_type) {
191 	case IB_QPT_SMI:
192 		qp->ibqp.qp_num		= 0;
193 		port->qp_smi_index	= qpn;
194 		qp->attr.port_num	= init->port_num;
195 		break;
196 
197 	case IB_QPT_GSI:
198 		qp->ibqp.qp_num		= 1;
199 		port->qp_gsi_index	= qpn;
200 		qp->attr.port_num	= init->port_num;
201 		break;
202 
203 	default:
204 		qp->ibqp.qp_num		= qpn;
205 		break;
206 	}
207 
208 	INIT_LIST_HEAD(&qp->grp_list);
209 
210 	skb_queue_head_init(&qp->send_pkts);
211 
212 	spin_lock_init(&qp->grp_lock);
213 	spin_lock_init(&qp->state_lock);
214 
215 	atomic_set(&qp->ssn, 0);
216 	atomic_set(&qp->skb_out, 0);
217 }
218 
219 static int rxe_qp_init_req(struct rxe_dev *rxe, struct rxe_qp *qp,
220 			   struct ib_qp_init_attr *init, struct ib_udata *udata,
221 			   struct rxe_create_qp_resp __user *uresp)
222 {
223 	int err;
224 	int wqe_size;
225 
226 	err = sock_create_kern(&init_net, AF_INET, SOCK_DGRAM, 0, &qp->sk);
227 	if (err < 0)
228 		return err;
229 	qp->sk->sk->sk_user_data = qp;
230 
231 	/* pick a source UDP port number for this QP based on
232 	 * the source QPN. this spreads traffic for different QPs
233 	 * across different NIC RX queues (while using a single
234 	 * flow for a given QP to maintain packet order).
235 	 * the port number must be in the Dynamic Ports range
236 	 * (0xc000 - 0xffff).
237 	 */
238 	qp->src_port = RXE_ROCE_V2_SPORT +
239 		(hash_32_generic(qp_num(qp), 14) & 0x3fff);
240 
241 	qp->sq.max_wr		= init->cap.max_send_wr;
242 	qp->sq.max_sge		= init->cap.max_send_sge;
243 	qp->sq.max_inline	= init->cap.max_inline_data;
244 
245 	wqe_size = max_t(int, sizeof(struct rxe_send_wqe) +
246 			 qp->sq.max_sge * sizeof(struct ib_sge),
247 			 sizeof(struct rxe_send_wqe) +
248 			 qp->sq.max_inline);
249 
250 	qp->sq.queue = rxe_queue_init(rxe,
251 				      &qp->sq.max_wr,
252 				      wqe_size);
253 	if (!qp->sq.queue)
254 		return -ENOMEM;
255 
256 	err = do_mmap_info(rxe, uresp ? &uresp->sq_mi : NULL, udata,
257 			   qp->sq.queue->buf, qp->sq.queue->buf_size,
258 			   &qp->sq.queue->ip);
259 
260 	if (err) {
261 		vfree(qp->sq.queue->buf);
262 		kfree(qp->sq.queue);
263 		return err;
264 	}
265 
266 	qp->req.wqe_index	= producer_index(qp->sq.queue);
267 	qp->req.state		= QP_STATE_RESET;
268 	qp->req.opcode		= -1;
269 	qp->comp.opcode		= -1;
270 
271 	spin_lock_init(&qp->sq.sq_lock);
272 	skb_queue_head_init(&qp->req_pkts);
273 
274 	rxe_init_task(rxe, &qp->req.task, qp,
275 		      rxe_requester, "req");
276 	rxe_init_task(rxe, &qp->comp.task, qp,
277 		      rxe_completer, "comp");
278 
279 	qp->qp_timeout_jiffies = 0; /* Can't be set for UD/UC in modify_qp */
280 	if (init->qp_type == IB_QPT_RC) {
281 		timer_setup(&qp->rnr_nak_timer, rnr_nak_timer, 0);
282 		timer_setup(&qp->retrans_timer, retransmit_timer, 0);
283 	}
284 	return 0;
285 }
286 
287 static int rxe_qp_init_resp(struct rxe_dev *rxe, struct rxe_qp *qp,
288 			    struct ib_qp_init_attr *init,
289 			    struct ib_udata *udata,
290 			    struct rxe_create_qp_resp __user *uresp)
291 {
292 	int err;
293 	int wqe_size;
294 
295 	if (!qp->srq) {
296 		qp->rq.max_wr		= init->cap.max_recv_wr;
297 		qp->rq.max_sge		= init->cap.max_recv_sge;
298 
299 		wqe_size = rcv_wqe_size(qp->rq.max_sge);
300 
301 		pr_debug("qp#%d max_wr = %d, max_sge = %d, wqe_size = %d\n",
302 			 qp_num(qp), qp->rq.max_wr, qp->rq.max_sge, wqe_size);
303 
304 		qp->rq.queue = rxe_queue_init(rxe,
305 					      &qp->rq.max_wr,
306 					      wqe_size);
307 		if (!qp->rq.queue)
308 			return -ENOMEM;
309 
310 		err = do_mmap_info(rxe, uresp ? &uresp->rq_mi : NULL, udata,
311 				   qp->rq.queue->buf, qp->rq.queue->buf_size,
312 				   &qp->rq.queue->ip);
313 		if (err) {
314 			vfree(qp->rq.queue->buf);
315 			kfree(qp->rq.queue);
316 			return err;
317 		}
318 	}
319 
320 	spin_lock_init(&qp->rq.producer_lock);
321 	spin_lock_init(&qp->rq.consumer_lock);
322 
323 	skb_queue_head_init(&qp->resp_pkts);
324 
325 	rxe_init_task(rxe, &qp->resp.task, qp,
326 		      rxe_responder, "resp");
327 
328 	qp->resp.opcode		= OPCODE_NONE;
329 	qp->resp.msn		= 0;
330 	qp->resp.state		= QP_STATE_RESET;
331 
332 	return 0;
333 }
334 
335 /* called by the create qp verb */
336 int rxe_qp_from_init(struct rxe_dev *rxe, struct rxe_qp *qp, struct rxe_pd *pd,
337 		     struct ib_qp_init_attr *init,
338 		     struct rxe_create_qp_resp __user *uresp,
339 		     struct ib_pd *ibpd,
340 		     struct ib_udata *udata)
341 {
342 	int err;
343 	struct rxe_cq *rcq = to_rcq(init->recv_cq);
344 	struct rxe_cq *scq = to_rcq(init->send_cq);
345 	struct rxe_srq *srq = init->srq ? to_rsrq(init->srq) : NULL;
346 
347 	rxe_add_ref(pd);
348 	rxe_add_ref(rcq);
349 	rxe_add_ref(scq);
350 	if (srq)
351 		rxe_add_ref(srq);
352 
353 	qp->pd			= pd;
354 	qp->rcq			= rcq;
355 	qp->scq			= scq;
356 	qp->srq			= srq;
357 
358 	rxe_qp_init_misc(rxe, qp, init);
359 
360 	err = rxe_qp_init_req(rxe, qp, init, udata, uresp);
361 	if (err)
362 		goto err1;
363 
364 	err = rxe_qp_init_resp(rxe, qp, init, udata, uresp);
365 	if (err)
366 		goto err2;
367 
368 	qp->attr.qp_state = IB_QPS_RESET;
369 	qp->valid = 1;
370 
371 	return 0;
372 
373 err2:
374 	rxe_queue_cleanup(qp->sq.queue);
375 err1:
376 	if (srq)
377 		rxe_drop_ref(srq);
378 	rxe_drop_ref(scq);
379 	rxe_drop_ref(rcq);
380 	rxe_drop_ref(pd);
381 
382 	return err;
383 }
384 
385 /* called by the query qp verb */
386 int rxe_qp_to_init(struct rxe_qp *qp, struct ib_qp_init_attr *init)
387 {
388 	init->event_handler		= qp->ibqp.event_handler;
389 	init->qp_context		= qp->ibqp.qp_context;
390 	init->send_cq			= qp->ibqp.send_cq;
391 	init->recv_cq			= qp->ibqp.recv_cq;
392 	init->srq			= qp->ibqp.srq;
393 
394 	init->cap.max_send_wr		= qp->sq.max_wr;
395 	init->cap.max_send_sge		= qp->sq.max_sge;
396 	init->cap.max_inline_data	= qp->sq.max_inline;
397 
398 	if (!qp->srq) {
399 		init->cap.max_recv_wr		= qp->rq.max_wr;
400 		init->cap.max_recv_sge		= qp->rq.max_sge;
401 	}
402 
403 	init->sq_sig_type		= qp->sq_sig_type;
404 
405 	init->qp_type			= qp->ibqp.qp_type;
406 	init->port_num			= 1;
407 
408 	return 0;
409 }
410 
411 /* called by the modify qp verb, this routine checks all the parameters before
412  * making any changes
413  */
414 int rxe_qp_chk_attr(struct rxe_dev *rxe, struct rxe_qp *qp,
415 		    struct ib_qp_attr *attr, int mask)
416 {
417 	enum ib_qp_state cur_state = (mask & IB_QP_CUR_STATE) ?
418 					attr->cur_qp_state : qp->attr.qp_state;
419 	enum ib_qp_state new_state = (mask & IB_QP_STATE) ?
420 					attr->qp_state : cur_state;
421 
422 	if (!ib_modify_qp_is_ok(cur_state, new_state, qp_type(qp), mask)) {
423 		pr_warn("invalid mask or state for qp\n");
424 		goto err1;
425 	}
426 
427 	if (mask & IB_QP_STATE) {
428 		if (cur_state == IB_QPS_SQD) {
429 			if (qp->req.state == QP_STATE_DRAIN &&
430 			    new_state != IB_QPS_ERR)
431 				goto err1;
432 		}
433 	}
434 
435 	if (mask & IB_QP_PORT) {
436 		if (!rdma_is_port_valid(&rxe->ib_dev, attr->port_num)) {
437 			pr_warn("invalid port %d\n", attr->port_num);
438 			goto err1;
439 		}
440 	}
441 
442 	if (mask & IB_QP_CAP && rxe_qp_chk_cap(rxe, &attr->cap, !!qp->srq))
443 		goto err1;
444 
445 	if (mask & IB_QP_AV && rxe_av_chk_attr(rxe, &attr->ah_attr))
446 		goto err1;
447 
448 	if (mask & IB_QP_ALT_PATH) {
449 		if (rxe_av_chk_attr(rxe, &attr->alt_ah_attr))
450 			goto err1;
451 		if (!rdma_is_port_valid(&rxe->ib_dev, attr->alt_port_num))  {
452 			pr_warn("invalid alt port %d\n", attr->alt_port_num);
453 			goto err1;
454 		}
455 		if (attr->alt_timeout > 31) {
456 			pr_warn("invalid QP alt timeout %d > 31\n",
457 				attr->alt_timeout);
458 			goto err1;
459 		}
460 	}
461 
462 	if (mask & IB_QP_PATH_MTU) {
463 		struct rxe_port *port = &rxe->port;
464 
465 		enum ib_mtu max_mtu = port->attr.max_mtu;
466 		enum ib_mtu mtu = attr->path_mtu;
467 
468 		if (mtu > max_mtu) {
469 			pr_debug("invalid mtu (%d) > (%d)\n",
470 				 ib_mtu_enum_to_int(mtu),
471 				 ib_mtu_enum_to_int(max_mtu));
472 			goto err1;
473 		}
474 	}
475 
476 	if (mask & IB_QP_MAX_QP_RD_ATOMIC) {
477 		if (attr->max_rd_atomic > rxe->attr.max_qp_rd_atom) {
478 			pr_warn("invalid max_rd_atomic %d > %d\n",
479 				attr->max_rd_atomic,
480 				rxe->attr.max_qp_rd_atom);
481 			goto err1;
482 		}
483 	}
484 
485 	if (mask & IB_QP_TIMEOUT) {
486 		if (attr->timeout > 31) {
487 			pr_warn("invalid QP timeout %d > 31\n",
488 				attr->timeout);
489 			goto err1;
490 		}
491 	}
492 
493 	return 0;
494 
495 err1:
496 	return -EINVAL;
497 }
498 
499 /* move the qp to the reset state */
500 static void rxe_qp_reset(struct rxe_qp *qp)
501 {
502 	/* stop tasks from running */
503 	rxe_disable_task(&qp->resp.task);
504 
505 	/* stop request/comp */
506 	if (qp->sq.queue) {
507 		if (qp_type(qp) == IB_QPT_RC)
508 			rxe_disable_task(&qp->comp.task);
509 		rxe_disable_task(&qp->req.task);
510 	}
511 
512 	/* move qp to the reset state */
513 	qp->req.state = QP_STATE_RESET;
514 	qp->resp.state = QP_STATE_RESET;
515 
516 	/* let state machines reset themselves drain work and packet queues
517 	 * etc.
518 	 */
519 	__rxe_do_task(&qp->resp.task);
520 
521 	if (qp->sq.queue) {
522 		__rxe_do_task(&qp->comp.task);
523 		__rxe_do_task(&qp->req.task);
524 		rxe_queue_reset(qp->sq.queue);
525 	}
526 
527 	/* cleanup attributes */
528 	atomic_set(&qp->ssn, 0);
529 	qp->req.opcode = -1;
530 	qp->req.need_retry = 0;
531 	qp->req.noack_pkts = 0;
532 	qp->resp.msn = 0;
533 	qp->resp.opcode = -1;
534 	qp->resp.drop_msg = 0;
535 	qp->resp.goto_error = 0;
536 	qp->resp.sent_psn_nak = 0;
537 
538 	if (qp->resp.mr) {
539 		rxe_drop_ref(qp->resp.mr);
540 		qp->resp.mr = NULL;
541 	}
542 
543 	cleanup_rd_atomic_resources(qp);
544 
545 	/* reenable tasks */
546 	rxe_enable_task(&qp->resp.task);
547 
548 	if (qp->sq.queue) {
549 		if (qp_type(qp) == IB_QPT_RC)
550 			rxe_enable_task(&qp->comp.task);
551 
552 		rxe_enable_task(&qp->req.task);
553 	}
554 }
555 
556 /* drain the send queue */
557 static void rxe_qp_drain(struct rxe_qp *qp)
558 {
559 	if (qp->sq.queue) {
560 		if (qp->req.state != QP_STATE_DRAINED) {
561 			qp->req.state = QP_STATE_DRAIN;
562 			if (qp_type(qp) == IB_QPT_RC)
563 				rxe_run_task(&qp->comp.task, 1);
564 			else
565 				__rxe_do_task(&qp->comp.task);
566 			rxe_run_task(&qp->req.task, 1);
567 		}
568 	}
569 }
570 
571 /* move the qp to the error state */
572 void rxe_qp_error(struct rxe_qp *qp)
573 {
574 	qp->req.state = QP_STATE_ERROR;
575 	qp->resp.state = QP_STATE_ERROR;
576 	qp->attr.qp_state = IB_QPS_ERR;
577 
578 	/* drain work and packet queues */
579 	rxe_run_task(&qp->resp.task, 1);
580 
581 	if (qp_type(qp) == IB_QPT_RC)
582 		rxe_run_task(&qp->comp.task, 1);
583 	else
584 		__rxe_do_task(&qp->comp.task);
585 	rxe_run_task(&qp->req.task, 1);
586 }
587 
588 /* called by the modify qp verb */
589 int rxe_qp_from_attr(struct rxe_qp *qp, struct ib_qp_attr *attr, int mask,
590 		     struct ib_udata *udata)
591 {
592 	int err;
593 
594 	if (mask & IB_QP_MAX_QP_RD_ATOMIC) {
595 		int max_rd_atomic = __roundup_pow_of_two(attr->max_rd_atomic);
596 
597 		qp->attr.max_rd_atomic = max_rd_atomic;
598 		atomic_set(&qp->req.rd_atomic, max_rd_atomic);
599 	}
600 
601 	if (mask & IB_QP_MAX_DEST_RD_ATOMIC) {
602 		int max_dest_rd_atomic =
603 			__roundup_pow_of_two(attr->max_dest_rd_atomic);
604 
605 		qp->attr.max_dest_rd_atomic = max_dest_rd_atomic;
606 
607 		free_rd_atomic_resources(qp);
608 
609 		err = alloc_rd_atomic_resources(qp, max_dest_rd_atomic);
610 		if (err)
611 			return err;
612 	}
613 
614 	if (mask & IB_QP_CUR_STATE)
615 		qp->attr.cur_qp_state = attr->qp_state;
616 
617 	if (mask & IB_QP_EN_SQD_ASYNC_NOTIFY)
618 		qp->attr.en_sqd_async_notify = attr->en_sqd_async_notify;
619 
620 	if (mask & IB_QP_ACCESS_FLAGS)
621 		qp->attr.qp_access_flags = attr->qp_access_flags;
622 
623 	if (mask & IB_QP_PKEY_INDEX)
624 		qp->attr.pkey_index = attr->pkey_index;
625 
626 	if (mask & IB_QP_PORT)
627 		qp->attr.port_num = attr->port_num;
628 
629 	if (mask & IB_QP_QKEY)
630 		qp->attr.qkey = attr->qkey;
631 
632 	if (mask & IB_QP_AV) {
633 		rxe_init_av(&attr->ah_attr, &qp->pri_av);
634 	}
635 
636 	if (mask & IB_QP_ALT_PATH) {
637 		rxe_init_av(&attr->alt_ah_attr, &qp->alt_av);
638 		qp->attr.alt_port_num = attr->alt_port_num;
639 		qp->attr.alt_pkey_index = attr->alt_pkey_index;
640 		qp->attr.alt_timeout = attr->alt_timeout;
641 	}
642 
643 	if (mask & IB_QP_PATH_MTU) {
644 		qp->attr.path_mtu = attr->path_mtu;
645 		qp->mtu = ib_mtu_enum_to_int(attr->path_mtu);
646 	}
647 
648 	if (mask & IB_QP_TIMEOUT) {
649 		qp->attr.timeout = attr->timeout;
650 		if (attr->timeout == 0) {
651 			qp->qp_timeout_jiffies = 0;
652 		} else {
653 			/* According to the spec, timeout = 4.096 * 2 ^ attr->timeout [us] */
654 			int j = nsecs_to_jiffies(4096ULL << attr->timeout);
655 
656 			qp->qp_timeout_jiffies = j ? j : 1;
657 		}
658 	}
659 
660 	if (mask & IB_QP_RETRY_CNT) {
661 		qp->attr.retry_cnt = attr->retry_cnt;
662 		qp->comp.retry_cnt = attr->retry_cnt;
663 		pr_debug("qp#%d set retry count = %d\n", qp_num(qp),
664 			 attr->retry_cnt);
665 	}
666 
667 	if (mask & IB_QP_RNR_RETRY) {
668 		qp->attr.rnr_retry = attr->rnr_retry;
669 		qp->comp.rnr_retry = attr->rnr_retry;
670 		pr_debug("qp#%d set rnr retry count = %d\n", qp_num(qp),
671 			 attr->rnr_retry);
672 	}
673 
674 	if (mask & IB_QP_RQ_PSN) {
675 		qp->attr.rq_psn = (attr->rq_psn & BTH_PSN_MASK);
676 		qp->resp.psn = qp->attr.rq_psn;
677 		pr_debug("qp#%d set resp psn = 0x%x\n", qp_num(qp),
678 			 qp->resp.psn);
679 	}
680 
681 	if (mask & IB_QP_MIN_RNR_TIMER) {
682 		qp->attr.min_rnr_timer = attr->min_rnr_timer;
683 		pr_debug("qp#%d set min rnr timer = 0x%x\n", qp_num(qp),
684 			 attr->min_rnr_timer);
685 	}
686 
687 	if (mask & IB_QP_SQ_PSN) {
688 		qp->attr.sq_psn = (attr->sq_psn & BTH_PSN_MASK);
689 		qp->req.psn = qp->attr.sq_psn;
690 		qp->comp.psn = qp->attr.sq_psn;
691 		pr_debug("qp#%d set req psn = 0x%x\n", qp_num(qp), qp->req.psn);
692 	}
693 
694 	if (mask & IB_QP_PATH_MIG_STATE)
695 		qp->attr.path_mig_state = attr->path_mig_state;
696 
697 	if (mask & IB_QP_DEST_QPN)
698 		qp->attr.dest_qp_num = attr->dest_qp_num;
699 
700 	if (mask & IB_QP_STATE) {
701 		qp->attr.qp_state = attr->qp_state;
702 
703 		switch (attr->qp_state) {
704 		case IB_QPS_RESET:
705 			pr_debug("qp#%d state -> RESET\n", qp_num(qp));
706 			rxe_qp_reset(qp);
707 			break;
708 
709 		case IB_QPS_INIT:
710 			pr_debug("qp#%d state -> INIT\n", qp_num(qp));
711 			qp->req.state = QP_STATE_INIT;
712 			qp->resp.state = QP_STATE_INIT;
713 			break;
714 
715 		case IB_QPS_RTR:
716 			pr_debug("qp#%d state -> RTR\n", qp_num(qp));
717 			qp->resp.state = QP_STATE_READY;
718 			break;
719 
720 		case IB_QPS_RTS:
721 			pr_debug("qp#%d state -> RTS\n", qp_num(qp));
722 			qp->req.state = QP_STATE_READY;
723 			break;
724 
725 		case IB_QPS_SQD:
726 			pr_debug("qp#%d state -> SQD\n", qp_num(qp));
727 			rxe_qp_drain(qp);
728 			break;
729 
730 		case IB_QPS_SQE:
731 			pr_warn("qp#%d state -> SQE !!?\n", qp_num(qp));
732 			/* Not possible from modify_qp. */
733 			break;
734 
735 		case IB_QPS_ERR:
736 			pr_debug("qp#%d state -> ERR\n", qp_num(qp));
737 			rxe_qp_error(qp);
738 			break;
739 		}
740 	}
741 
742 	return 0;
743 }
744 
745 /* called by the query qp verb */
746 int rxe_qp_to_attr(struct rxe_qp *qp, struct ib_qp_attr *attr, int mask)
747 {
748 	*attr = qp->attr;
749 
750 	attr->rq_psn				= qp->resp.psn;
751 	attr->sq_psn				= qp->req.psn;
752 
753 	attr->cap.max_send_wr			= qp->sq.max_wr;
754 	attr->cap.max_send_sge			= qp->sq.max_sge;
755 	attr->cap.max_inline_data		= qp->sq.max_inline;
756 
757 	if (!qp->srq) {
758 		attr->cap.max_recv_wr		= qp->rq.max_wr;
759 		attr->cap.max_recv_sge		= qp->rq.max_sge;
760 	}
761 
762 	rxe_av_to_attr(&qp->pri_av, &attr->ah_attr);
763 	rxe_av_to_attr(&qp->alt_av, &attr->alt_ah_attr);
764 
765 	if (qp->req.state == QP_STATE_DRAIN) {
766 		attr->sq_draining = 1;
767 		/* applications that get this state
768 		 * typically spin on it. yield the
769 		 * processor
770 		 */
771 		cond_resched();
772 	} else {
773 		attr->sq_draining = 0;
774 	}
775 
776 	pr_debug("attr->sq_draining = %d\n", attr->sq_draining);
777 
778 	return 0;
779 }
780 
781 /* called by the destroy qp verb */
782 void rxe_qp_destroy(struct rxe_qp *qp)
783 {
784 	qp->valid = 0;
785 	qp->qp_timeout_jiffies = 0;
786 	rxe_cleanup_task(&qp->resp.task);
787 
788 	if (qp_type(qp) == IB_QPT_RC) {
789 		del_timer_sync(&qp->retrans_timer);
790 		del_timer_sync(&qp->rnr_nak_timer);
791 	}
792 
793 	rxe_cleanup_task(&qp->req.task);
794 	rxe_cleanup_task(&qp->comp.task);
795 
796 	/* flush out any receive wr's or pending requests */
797 	__rxe_do_task(&qp->req.task);
798 	if (qp->sq.queue) {
799 		__rxe_do_task(&qp->comp.task);
800 		__rxe_do_task(&qp->req.task);
801 	}
802 }
803 
804 /* called when the last reference to the qp is dropped */
805 static void rxe_qp_do_cleanup(struct work_struct *work)
806 {
807 	struct rxe_qp *qp = container_of(work, typeof(*qp), cleanup_work.work);
808 
809 	rxe_drop_all_mcast_groups(qp);
810 
811 	if (qp->sq.queue)
812 		rxe_queue_cleanup(qp->sq.queue);
813 
814 	if (qp->srq)
815 		rxe_drop_ref(qp->srq);
816 
817 	if (qp->rq.queue)
818 		rxe_queue_cleanup(qp->rq.queue);
819 
820 	if (qp->scq)
821 		rxe_drop_ref(qp->scq);
822 	if (qp->rcq)
823 		rxe_drop_ref(qp->rcq);
824 	if (qp->pd)
825 		rxe_drop_ref(qp->pd);
826 
827 	if (qp->resp.mr) {
828 		rxe_drop_ref(qp->resp.mr);
829 		qp->resp.mr = NULL;
830 	}
831 
832 	if (qp_type(qp) == IB_QPT_RC)
833 		sk_dst_reset(qp->sk->sk);
834 
835 	free_rd_atomic_resources(qp);
836 
837 	kernel_sock_shutdown(qp->sk, SHUT_RDWR);
838 	sock_release(qp->sk);
839 }
840 
841 /* called when the last reference to the qp is dropped */
842 void rxe_qp_cleanup(struct rxe_pool_entry *arg)
843 {
844 	struct rxe_qp *qp = container_of(arg, typeof(*qp), pelem);
845 
846 	execute_in_process_context(rxe_qp_do_cleanup, &qp->cleanup_work);
847 }
848