xref: /linux/drivers/infiniband/sw/rxe/rxe_qp.c (revision 37744feebc086908fd89760650f458ab19071750)
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 	qp->sq.max_wr		= init->cap.max_send_wr;
241 
242 	/* These caps are limited by rxe_qp_chk_cap() done by the caller */
243 	wqe_size = max_t(int, init->cap.max_send_sge * sizeof(struct ib_sge),
244 			 init->cap.max_inline_data);
245 	qp->sq.max_sge = init->cap.max_send_sge =
246 		wqe_size / sizeof(struct ib_sge);
247 	qp->sq.max_inline = init->cap.max_inline_data = wqe_size;
248 	wqe_size += sizeof(struct rxe_send_wqe);
249 
250 	qp->sq.queue = rxe_queue_init(rxe, &qp->sq.max_wr, wqe_size);
251 	if (!qp->sq.queue)
252 		return -ENOMEM;
253 
254 	err = do_mmap_info(rxe, uresp ? &uresp->sq_mi : NULL, udata,
255 			   qp->sq.queue->buf, qp->sq.queue->buf_size,
256 			   &qp->sq.queue->ip);
257 
258 	if (err) {
259 		vfree(qp->sq.queue->buf);
260 		kfree(qp->sq.queue);
261 		return err;
262 	}
263 
264 	qp->req.wqe_index	= producer_index(qp->sq.queue);
265 	qp->req.state		= QP_STATE_RESET;
266 	qp->req.opcode		= -1;
267 	qp->comp.opcode		= -1;
268 
269 	spin_lock_init(&qp->sq.sq_lock);
270 	skb_queue_head_init(&qp->req_pkts);
271 
272 	rxe_init_task(rxe, &qp->req.task, qp,
273 		      rxe_requester, "req");
274 	rxe_init_task(rxe, &qp->comp.task, qp,
275 		      rxe_completer, "comp");
276 
277 	qp->qp_timeout_jiffies = 0; /* Can't be set for UD/UC in modify_qp */
278 	if (init->qp_type == IB_QPT_RC) {
279 		timer_setup(&qp->rnr_nak_timer, rnr_nak_timer, 0);
280 		timer_setup(&qp->retrans_timer, retransmit_timer, 0);
281 	}
282 	return 0;
283 }
284 
285 static int rxe_qp_init_resp(struct rxe_dev *rxe, struct rxe_qp *qp,
286 			    struct ib_qp_init_attr *init,
287 			    struct ib_udata *udata,
288 			    struct rxe_create_qp_resp __user *uresp)
289 {
290 	int err;
291 	int wqe_size;
292 
293 	if (!qp->srq) {
294 		qp->rq.max_wr		= init->cap.max_recv_wr;
295 		qp->rq.max_sge		= init->cap.max_recv_sge;
296 
297 		wqe_size = rcv_wqe_size(qp->rq.max_sge);
298 
299 		pr_debug("qp#%d max_wr = %d, max_sge = %d, wqe_size = %d\n",
300 			 qp_num(qp), qp->rq.max_wr, qp->rq.max_sge, wqe_size);
301 
302 		qp->rq.queue = rxe_queue_init(rxe,
303 					      &qp->rq.max_wr,
304 					      wqe_size);
305 		if (!qp->rq.queue)
306 			return -ENOMEM;
307 
308 		err = do_mmap_info(rxe, uresp ? &uresp->rq_mi : NULL, udata,
309 				   qp->rq.queue->buf, qp->rq.queue->buf_size,
310 				   &qp->rq.queue->ip);
311 		if (err) {
312 			vfree(qp->rq.queue->buf);
313 			kfree(qp->rq.queue);
314 			return err;
315 		}
316 	}
317 
318 	spin_lock_init(&qp->rq.producer_lock);
319 	spin_lock_init(&qp->rq.consumer_lock);
320 
321 	skb_queue_head_init(&qp->resp_pkts);
322 
323 	rxe_init_task(rxe, &qp->resp.task, qp,
324 		      rxe_responder, "resp");
325 
326 	qp->resp.opcode		= OPCODE_NONE;
327 	qp->resp.msn		= 0;
328 	qp->resp.state		= QP_STATE_RESET;
329 
330 	return 0;
331 }
332 
333 /* called by the create qp verb */
334 int rxe_qp_from_init(struct rxe_dev *rxe, struct rxe_qp *qp, struct rxe_pd *pd,
335 		     struct ib_qp_init_attr *init,
336 		     struct rxe_create_qp_resp __user *uresp,
337 		     struct ib_pd *ibpd,
338 		     struct ib_udata *udata)
339 {
340 	int err;
341 	struct rxe_cq *rcq = to_rcq(init->recv_cq);
342 	struct rxe_cq *scq = to_rcq(init->send_cq);
343 	struct rxe_srq *srq = init->srq ? to_rsrq(init->srq) : NULL;
344 
345 	rxe_add_ref(pd);
346 	rxe_add_ref(rcq);
347 	rxe_add_ref(scq);
348 	if (srq)
349 		rxe_add_ref(srq);
350 
351 	qp->pd			= pd;
352 	qp->rcq			= rcq;
353 	qp->scq			= scq;
354 	qp->srq			= srq;
355 
356 	rxe_qp_init_misc(rxe, qp, init);
357 
358 	err = rxe_qp_init_req(rxe, qp, init, udata, uresp);
359 	if (err)
360 		goto err1;
361 
362 	err = rxe_qp_init_resp(rxe, qp, init, udata, uresp);
363 	if (err)
364 		goto err2;
365 
366 	qp->attr.qp_state = IB_QPS_RESET;
367 	qp->valid = 1;
368 
369 	return 0;
370 
371 err2:
372 	rxe_queue_cleanup(qp->sq.queue);
373 err1:
374 	if (srq)
375 		rxe_drop_ref(srq);
376 	rxe_drop_ref(scq);
377 	rxe_drop_ref(rcq);
378 	rxe_drop_ref(pd);
379 
380 	return err;
381 }
382 
383 /* called by the query qp verb */
384 int rxe_qp_to_init(struct rxe_qp *qp, struct ib_qp_init_attr *init)
385 {
386 	init->event_handler		= qp->ibqp.event_handler;
387 	init->qp_context		= qp->ibqp.qp_context;
388 	init->send_cq			= qp->ibqp.send_cq;
389 	init->recv_cq			= qp->ibqp.recv_cq;
390 	init->srq			= qp->ibqp.srq;
391 
392 	init->cap.max_send_wr		= qp->sq.max_wr;
393 	init->cap.max_send_sge		= qp->sq.max_sge;
394 	init->cap.max_inline_data	= qp->sq.max_inline;
395 
396 	if (!qp->srq) {
397 		init->cap.max_recv_wr		= qp->rq.max_wr;
398 		init->cap.max_recv_sge		= qp->rq.max_sge;
399 	}
400 
401 	init->sq_sig_type		= qp->sq_sig_type;
402 
403 	init->qp_type			= qp->ibqp.qp_type;
404 	init->port_num			= 1;
405 
406 	return 0;
407 }
408 
409 /* called by the modify qp verb, this routine checks all the parameters before
410  * making any changes
411  */
412 int rxe_qp_chk_attr(struct rxe_dev *rxe, struct rxe_qp *qp,
413 		    struct ib_qp_attr *attr, int mask)
414 {
415 	enum ib_qp_state cur_state = (mask & IB_QP_CUR_STATE) ?
416 					attr->cur_qp_state : qp->attr.qp_state;
417 	enum ib_qp_state new_state = (mask & IB_QP_STATE) ?
418 					attr->qp_state : cur_state;
419 
420 	if (!ib_modify_qp_is_ok(cur_state, new_state, qp_type(qp), mask)) {
421 		pr_warn("invalid mask or state for qp\n");
422 		goto err1;
423 	}
424 
425 	if (mask & IB_QP_STATE) {
426 		if (cur_state == IB_QPS_SQD) {
427 			if (qp->req.state == QP_STATE_DRAIN &&
428 			    new_state != IB_QPS_ERR)
429 				goto err1;
430 		}
431 	}
432 
433 	if (mask & IB_QP_PORT) {
434 		if (!rdma_is_port_valid(&rxe->ib_dev, attr->port_num)) {
435 			pr_warn("invalid port %d\n", attr->port_num);
436 			goto err1;
437 		}
438 	}
439 
440 	if (mask & IB_QP_CAP && rxe_qp_chk_cap(rxe, &attr->cap, !!qp->srq))
441 		goto err1;
442 
443 	if (mask & IB_QP_AV && rxe_av_chk_attr(rxe, &attr->ah_attr))
444 		goto err1;
445 
446 	if (mask & IB_QP_ALT_PATH) {
447 		if (rxe_av_chk_attr(rxe, &attr->alt_ah_attr))
448 			goto err1;
449 		if (!rdma_is_port_valid(&rxe->ib_dev, attr->alt_port_num))  {
450 			pr_warn("invalid alt port %d\n", attr->alt_port_num);
451 			goto err1;
452 		}
453 		if (attr->alt_timeout > 31) {
454 			pr_warn("invalid QP alt timeout %d > 31\n",
455 				attr->alt_timeout);
456 			goto err1;
457 		}
458 	}
459 
460 	if (mask & IB_QP_PATH_MTU) {
461 		struct rxe_port *port = &rxe->port;
462 
463 		enum ib_mtu max_mtu = port->attr.max_mtu;
464 		enum ib_mtu mtu = attr->path_mtu;
465 
466 		if (mtu > max_mtu) {
467 			pr_debug("invalid mtu (%d) > (%d)\n",
468 				 ib_mtu_enum_to_int(mtu),
469 				 ib_mtu_enum_to_int(max_mtu));
470 			goto err1;
471 		}
472 	}
473 
474 	if (mask & IB_QP_MAX_QP_RD_ATOMIC) {
475 		if (attr->max_rd_atomic > rxe->attr.max_qp_rd_atom) {
476 			pr_warn("invalid max_rd_atomic %d > %d\n",
477 				attr->max_rd_atomic,
478 				rxe->attr.max_qp_rd_atom);
479 			goto err1;
480 		}
481 	}
482 
483 	if (mask & IB_QP_TIMEOUT) {
484 		if (attr->timeout > 31) {
485 			pr_warn("invalid QP timeout %d > 31\n",
486 				attr->timeout);
487 			goto err1;
488 		}
489 	}
490 
491 	return 0;
492 
493 err1:
494 	return -EINVAL;
495 }
496 
497 /* move the qp to the reset state */
498 static void rxe_qp_reset(struct rxe_qp *qp)
499 {
500 	/* stop tasks from running */
501 	rxe_disable_task(&qp->resp.task);
502 
503 	/* stop request/comp */
504 	if (qp->sq.queue) {
505 		if (qp_type(qp) == IB_QPT_RC)
506 			rxe_disable_task(&qp->comp.task);
507 		rxe_disable_task(&qp->req.task);
508 	}
509 
510 	/* move qp to the reset state */
511 	qp->req.state = QP_STATE_RESET;
512 	qp->resp.state = QP_STATE_RESET;
513 
514 	/* let state machines reset themselves drain work and packet queues
515 	 * etc.
516 	 */
517 	__rxe_do_task(&qp->resp.task);
518 
519 	if (qp->sq.queue) {
520 		__rxe_do_task(&qp->comp.task);
521 		__rxe_do_task(&qp->req.task);
522 		rxe_queue_reset(qp->sq.queue);
523 	}
524 
525 	/* cleanup attributes */
526 	atomic_set(&qp->ssn, 0);
527 	qp->req.opcode = -1;
528 	qp->req.need_retry = 0;
529 	qp->req.noack_pkts = 0;
530 	qp->resp.msn = 0;
531 	qp->resp.opcode = -1;
532 	qp->resp.drop_msg = 0;
533 	qp->resp.goto_error = 0;
534 	qp->resp.sent_psn_nak = 0;
535 
536 	if (qp->resp.mr) {
537 		rxe_drop_ref(qp->resp.mr);
538 		qp->resp.mr = NULL;
539 	}
540 
541 	cleanup_rd_atomic_resources(qp);
542 
543 	/* reenable tasks */
544 	rxe_enable_task(&qp->resp.task);
545 
546 	if (qp->sq.queue) {
547 		if (qp_type(qp) == IB_QPT_RC)
548 			rxe_enable_task(&qp->comp.task);
549 
550 		rxe_enable_task(&qp->req.task);
551 	}
552 }
553 
554 /* drain the send queue */
555 static void rxe_qp_drain(struct rxe_qp *qp)
556 {
557 	if (qp->sq.queue) {
558 		if (qp->req.state != QP_STATE_DRAINED) {
559 			qp->req.state = QP_STATE_DRAIN;
560 			if (qp_type(qp) == IB_QPT_RC)
561 				rxe_run_task(&qp->comp.task, 1);
562 			else
563 				__rxe_do_task(&qp->comp.task);
564 			rxe_run_task(&qp->req.task, 1);
565 		}
566 	}
567 }
568 
569 /* move the qp to the error state */
570 void rxe_qp_error(struct rxe_qp *qp)
571 {
572 	qp->req.state = QP_STATE_ERROR;
573 	qp->resp.state = QP_STATE_ERROR;
574 	qp->attr.qp_state = IB_QPS_ERR;
575 
576 	/* drain work and packet queues */
577 	rxe_run_task(&qp->resp.task, 1);
578 
579 	if (qp_type(qp) == IB_QPT_RC)
580 		rxe_run_task(&qp->comp.task, 1);
581 	else
582 		__rxe_do_task(&qp->comp.task);
583 	rxe_run_task(&qp->req.task, 1);
584 }
585 
586 /* called by the modify qp verb */
587 int rxe_qp_from_attr(struct rxe_qp *qp, struct ib_qp_attr *attr, int mask,
588 		     struct ib_udata *udata)
589 {
590 	int err;
591 
592 	if (mask & IB_QP_MAX_QP_RD_ATOMIC) {
593 		int max_rd_atomic = attr->max_rd_atomic ?
594 			roundup_pow_of_two(attr->max_rd_atomic) : 0;
595 
596 		qp->attr.max_rd_atomic = max_rd_atomic;
597 		atomic_set(&qp->req.rd_atomic, max_rd_atomic);
598 	}
599 
600 	if (mask & IB_QP_MAX_DEST_RD_ATOMIC) {
601 		int max_dest_rd_atomic = attr->max_dest_rd_atomic ?
602 			roundup_pow_of_two(attr->max_dest_rd_atomic) : 0;
603 
604 		qp->attr.max_dest_rd_atomic = max_dest_rd_atomic;
605 
606 		free_rd_atomic_resources(qp);
607 
608 		err = alloc_rd_atomic_resources(qp, max_dest_rd_atomic);
609 		if (err)
610 			return err;
611 	}
612 
613 	if (mask & IB_QP_CUR_STATE)
614 		qp->attr.cur_qp_state = attr->qp_state;
615 
616 	if (mask & IB_QP_EN_SQD_ASYNC_NOTIFY)
617 		qp->attr.en_sqd_async_notify = attr->en_sqd_async_notify;
618 
619 	if (mask & IB_QP_ACCESS_FLAGS)
620 		qp->attr.qp_access_flags = attr->qp_access_flags;
621 
622 	if (mask & IB_QP_PKEY_INDEX)
623 		qp->attr.pkey_index = attr->pkey_index;
624 
625 	if (mask & IB_QP_PORT)
626 		qp->attr.port_num = attr->port_num;
627 
628 	if (mask & IB_QP_QKEY)
629 		qp->attr.qkey = attr->qkey;
630 
631 	if (mask & IB_QP_AV) {
632 		rxe_init_av(&attr->ah_attr, &qp->pri_av);
633 	}
634 
635 	if (mask & IB_QP_ALT_PATH) {
636 		rxe_init_av(&attr->alt_ah_attr, &qp->alt_av);
637 		qp->attr.alt_port_num = attr->alt_port_num;
638 		qp->attr.alt_pkey_index = attr->alt_pkey_index;
639 		qp->attr.alt_timeout = attr->alt_timeout;
640 	}
641 
642 	if (mask & IB_QP_PATH_MTU) {
643 		qp->attr.path_mtu = attr->path_mtu;
644 		qp->mtu = ib_mtu_enum_to_int(attr->path_mtu);
645 	}
646 
647 	if (mask & IB_QP_TIMEOUT) {
648 		qp->attr.timeout = attr->timeout;
649 		if (attr->timeout == 0) {
650 			qp->qp_timeout_jiffies = 0;
651 		} else {
652 			/* According to the spec, timeout = 4.096 * 2 ^ attr->timeout [us] */
653 			int j = nsecs_to_jiffies(4096ULL << attr->timeout);
654 
655 			qp->qp_timeout_jiffies = j ? j : 1;
656 		}
657 	}
658 
659 	if (mask & IB_QP_RETRY_CNT) {
660 		qp->attr.retry_cnt = attr->retry_cnt;
661 		qp->comp.retry_cnt = attr->retry_cnt;
662 		pr_debug("qp#%d set retry count = %d\n", qp_num(qp),
663 			 attr->retry_cnt);
664 	}
665 
666 	if (mask & IB_QP_RNR_RETRY) {
667 		qp->attr.rnr_retry = attr->rnr_retry;
668 		qp->comp.rnr_retry = attr->rnr_retry;
669 		pr_debug("qp#%d set rnr retry count = %d\n", qp_num(qp),
670 			 attr->rnr_retry);
671 	}
672 
673 	if (mask & IB_QP_RQ_PSN) {
674 		qp->attr.rq_psn = (attr->rq_psn & BTH_PSN_MASK);
675 		qp->resp.psn = qp->attr.rq_psn;
676 		pr_debug("qp#%d set resp psn = 0x%x\n", qp_num(qp),
677 			 qp->resp.psn);
678 	}
679 
680 	if (mask & IB_QP_MIN_RNR_TIMER) {
681 		qp->attr.min_rnr_timer = attr->min_rnr_timer;
682 		pr_debug("qp#%d set min rnr timer = 0x%x\n", qp_num(qp),
683 			 attr->min_rnr_timer);
684 	}
685 
686 	if (mask & IB_QP_SQ_PSN) {
687 		qp->attr.sq_psn = (attr->sq_psn & BTH_PSN_MASK);
688 		qp->req.psn = qp->attr.sq_psn;
689 		qp->comp.psn = qp->attr.sq_psn;
690 		pr_debug("qp#%d set req psn = 0x%x\n", qp_num(qp), qp->req.psn);
691 	}
692 
693 	if (mask & IB_QP_PATH_MIG_STATE)
694 		qp->attr.path_mig_state = attr->path_mig_state;
695 
696 	if (mask & IB_QP_DEST_QPN)
697 		qp->attr.dest_qp_num = attr->dest_qp_num;
698 
699 	if (mask & IB_QP_STATE) {
700 		qp->attr.qp_state = attr->qp_state;
701 
702 		switch (attr->qp_state) {
703 		case IB_QPS_RESET:
704 			pr_debug("qp#%d state -> RESET\n", qp_num(qp));
705 			rxe_qp_reset(qp);
706 			break;
707 
708 		case IB_QPS_INIT:
709 			pr_debug("qp#%d state -> INIT\n", qp_num(qp));
710 			qp->req.state = QP_STATE_INIT;
711 			qp->resp.state = QP_STATE_INIT;
712 			break;
713 
714 		case IB_QPS_RTR:
715 			pr_debug("qp#%d state -> RTR\n", qp_num(qp));
716 			qp->resp.state = QP_STATE_READY;
717 			break;
718 
719 		case IB_QPS_RTS:
720 			pr_debug("qp#%d state -> RTS\n", qp_num(qp));
721 			qp->req.state = QP_STATE_READY;
722 			break;
723 
724 		case IB_QPS_SQD:
725 			pr_debug("qp#%d state -> SQD\n", qp_num(qp));
726 			rxe_qp_drain(qp);
727 			break;
728 
729 		case IB_QPS_SQE:
730 			pr_warn("qp#%d state -> SQE !!?\n", qp_num(qp));
731 			/* Not possible from modify_qp. */
732 			break;
733 
734 		case IB_QPS_ERR:
735 			pr_debug("qp#%d state -> ERR\n", qp_num(qp));
736 			rxe_qp_error(qp);
737 			break;
738 		}
739 	}
740 
741 	return 0;
742 }
743 
744 /* called by the query qp verb */
745 int rxe_qp_to_attr(struct rxe_qp *qp, struct ib_qp_attr *attr, int mask)
746 {
747 	*attr = qp->attr;
748 
749 	attr->rq_psn				= qp->resp.psn;
750 	attr->sq_psn				= qp->req.psn;
751 
752 	attr->cap.max_send_wr			= qp->sq.max_wr;
753 	attr->cap.max_send_sge			= qp->sq.max_sge;
754 	attr->cap.max_inline_data		= qp->sq.max_inline;
755 
756 	if (!qp->srq) {
757 		attr->cap.max_recv_wr		= qp->rq.max_wr;
758 		attr->cap.max_recv_sge		= qp->rq.max_sge;
759 	}
760 
761 	rxe_av_to_attr(&qp->pri_av, &attr->ah_attr);
762 	rxe_av_to_attr(&qp->alt_av, &attr->alt_ah_attr);
763 
764 	if (qp->req.state == QP_STATE_DRAIN) {
765 		attr->sq_draining = 1;
766 		/* applications that get this state
767 		 * typically spin on it. yield the
768 		 * processor
769 		 */
770 		cond_resched();
771 	} else {
772 		attr->sq_draining = 0;
773 	}
774 
775 	pr_debug("attr->sq_draining = %d\n", attr->sq_draining);
776 
777 	return 0;
778 }
779 
780 /* called by the destroy qp verb */
781 void rxe_qp_destroy(struct rxe_qp *qp)
782 {
783 	qp->valid = 0;
784 	qp->qp_timeout_jiffies = 0;
785 	rxe_cleanup_task(&qp->resp.task);
786 
787 	if (qp_type(qp) == IB_QPT_RC) {
788 		del_timer_sync(&qp->retrans_timer);
789 		del_timer_sync(&qp->rnr_nak_timer);
790 	}
791 
792 	rxe_cleanup_task(&qp->req.task);
793 	rxe_cleanup_task(&qp->comp.task);
794 
795 	/* flush out any receive wr's or pending requests */
796 	__rxe_do_task(&qp->req.task);
797 	if (qp->sq.queue) {
798 		__rxe_do_task(&qp->comp.task);
799 		__rxe_do_task(&qp->req.task);
800 	}
801 }
802 
803 /* called when the last reference to the qp is dropped */
804 static void rxe_qp_do_cleanup(struct work_struct *work)
805 {
806 	struct rxe_qp *qp = container_of(work, typeof(*qp), cleanup_work.work);
807 
808 	rxe_drop_all_mcast_groups(qp);
809 
810 	if (qp->sq.queue)
811 		rxe_queue_cleanup(qp->sq.queue);
812 
813 	if (qp->srq)
814 		rxe_drop_ref(qp->srq);
815 
816 	if (qp->rq.queue)
817 		rxe_queue_cleanup(qp->rq.queue);
818 
819 	if (qp->scq)
820 		rxe_drop_ref(qp->scq);
821 	if (qp->rcq)
822 		rxe_drop_ref(qp->rcq);
823 	if (qp->pd)
824 		rxe_drop_ref(qp->pd);
825 
826 	if (qp->resp.mr) {
827 		rxe_drop_ref(qp->resp.mr);
828 		qp->resp.mr = NULL;
829 	}
830 
831 	if (qp_type(qp) == IB_QPT_RC)
832 		sk_dst_reset(qp->sk->sk);
833 
834 	free_rd_atomic_resources(qp);
835 
836 	kernel_sock_shutdown(qp->sk, SHUT_RDWR);
837 	sock_release(qp->sk);
838 }
839 
840 /* called when the last reference to the qp is dropped */
841 void rxe_qp_cleanup(struct rxe_pool_entry *arg)
842 {
843 	struct rxe_qp *qp = container_of(arg, typeof(*qp), pelem);
844 
845 	execute_in_process_context(rxe_qp_do_cleanup, &qp->cleanup_work);
846 }
847