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/dma-mapping.h>
35 #include <net/addrconf.h>
36 #include "rxe.h"
37 #include "rxe_loc.h"
38 #include "rxe_queue.h"
39 #include "rxe_hw_counters.h"
40 
41 static int rxe_query_device(struct ib_device *dev,
42 			    struct ib_device_attr *attr,
43 			    struct ib_udata *uhw)
44 {
45 	struct rxe_dev *rxe = to_rdev(dev);
46 
47 	if (uhw->inlen || uhw->outlen)
48 		return -EINVAL;
49 
50 	*attr = rxe->attr;
51 	return 0;
52 }
53 
54 static int rxe_query_port(struct ib_device *dev,
55 			  u8 port_num, struct ib_port_attr *attr)
56 {
57 	struct rxe_dev *rxe = to_rdev(dev);
58 	struct rxe_port *port;
59 	int rc = -EINVAL;
60 
61 	if (unlikely(port_num != 1)) {
62 		pr_warn("invalid port_number %d\n", port_num);
63 		goto out;
64 	}
65 
66 	port = &rxe->port;
67 
68 	/* *attr being zeroed by the caller, avoid zeroing it here */
69 	*attr = port->attr;
70 
71 	mutex_lock(&rxe->usdev_lock);
72 	rc = ib_get_eth_speed(dev, port_num, &attr->active_speed,
73 			      &attr->active_width);
74 	mutex_unlock(&rxe->usdev_lock);
75 
76 out:
77 	return rc;
78 }
79 
80 static struct net_device *rxe_get_netdev(struct ib_device *device,
81 					 u8 port_num)
82 {
83 	struct rxe_dev *rxe = to_rdev(device);
84 
85 	if (rxe->ndev) {
86 		dev_hold(rxe->ndev);
87 		return rxe->ndev;
88 	}
89 
90 	return NULL;
91 }
92 
93 static int rxe_query_pkey(struct ib_device *device,
94 			  u8 port_num, u16 index, u16 *pkey)
95 {
96 	struct rxe_dev *rxe = to_rdev(device);
97 	struct rxe_port *port;
98 
99 	if (unlikely(port_num != 1)) {
100 		dev_warn(device->dev.parent, "invalid port_num = %d\n",
101 			 port_num);
102 		goto err1;
103 	}
104 
105 	port = &rxe->port;
106 
107 	if (unlikely(index >= port->attr.pkey_tbl_len)) {
108 		dev_warn(device->dev.parent, "invalid index = %d\n",
109 			 index);
110 		goto err1;
111 	}
112 
113 	*pkey = port->pkey_tbl[index];
114 	return 0;
115 
116 err1:
117 	return -EINVAL;
118 }
119 
120 static int rxe_modify_device(struct ib_device *dev,
121 			     int mask, struct ib_device_modify *attr)
122 {
123 	struct rxe_dev *rxe = to_rdev(dev);
124 
125 	if (mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID)
126 		rxe->attr.sys_image_guid = cpu_to_be64(attr->sys_image_guid);
127 
128 	if (mask & IB_DEVICE_MODIFY_NODE_DESC) {
129 		memcpy(rxe->ib_dev.node_desc,
130 		       attr->node_desc, sizeof(rxe->ib_dev.node_desc));
131 	}
132 
133 	return 0;
134 }
135 
136 static int rxe_modify_port(struct ib_device *dev,
137 			   u8 port_num, int mask, struct ib_port_modify *attr)
138 {
139 	struct rxe_dev *rxe = to_rdev(dev);
140 	struct rxe_port *port;
141 
142 	if (unlikely(port_num != 1)) {
143 		pr_warn("invalid port_num = %d\n", port_num);
144 		goto err1;
145 	}
146 
147 	port = &rxe->port;
148 
149 	port->attr.port_cap_flags |= attr->set_port_cap_mask;
150 	port->attr.port_cap_flags &= ~attr->clr_port_cap_mask;
151 
152 	if (mask & IB_PORT_RESET_QKEY_CNTR)
153 		port->attr.qkey_viol_cntr = 0;
154 
155 	return 0;
156 
157 err1:
158 	return -EINVAL;
159 }
160 
161 static enum rdma_link_layer rxe_get_link_layer(struct ib_device *dev,
162 					       u8 port_num)
163 {
164 	struct rxe_dev *rxe = to_rdev(dev);
165 
166 	return rxe_link_layer(rxe, port_num);
167 }
168 
169 static struct ib_ucontext *rxe_alloc_ucontext(struct ib_device *dev,
170 					      struct ib_udata *udata)
171 {
172 	struct rxe_dev *rxe = to_rdev(dev);
173 	struct rxe_ucontext *uc;
174 
175 	uc = rxe_alloc(&rxe->uc_pool);
176 	return uc ? &uc->ibuc : ERR_PTR(-ENOMEM);
177 }
178 
179 static int rxe_dealloc_ucontext(struct ib_ucontext *ibuc)
180 {
181 	struct rxe_ucontext *uc = to_ruc(ibuc);
182 
183 	rxe_drop_ref(uc);
184 	return 0;
185 }
186 
187 static int rxe_port_immutable(struct ib_device *dev, u8 port_num,
188 			      struct ib_port_immutable *immutable)
189 {
190 	int err;
191 	struct ib_port_attr attr;
192 
193 	immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
194 
195 	err = ib_query_port(dev, port_num, &attr);
196 	if (err)
197 		return err;
198 
199 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
200 	immutable->gid_tbl_len = attr.gid_tbl_len;
201 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
202 
203 	return 0;
204 }
205 
206 static struct ib_pd *rxe_alloc_pd(struct ib_device *dev,
207 				  struct ib_ucontext *context,
208 				  struct ib_udata *udata)
209 {
210 	struct rxe_dev *rxe = to_rdev(dev);
211 	struct rxe_pd *pd;
212 
213 	pd = rxe_alloc(&rxe->pd_pool);
214 	return pd ? &pd->ibpd : ERR_PTR(-ENOMEM);
215 }
216 
217 static int rxe_dealloc_pd(struct ib_pd *ibpd)
218 {
219 	struct rxe_pd *pd = to_rpd(ibpd);
220 
221 	rxe_drop_ref(pd);
222 	return 0;
223 }
224 
225 static void rxe_init_av(struct rxe_dev *rxe, struct rdma_ah_attr *attr,
226 			struct rxe_av *av)
227 {
228 	rxe_av_from_attr(rdma_ah_get_port_num(attr), av, attr);
229 	rxe_av_fill_ip_info(av, attr);
230 }
231 
232 static struct ib_ah *rxe_create_ah(struct ib_pd *ibpd,
233 				   struct rdma_ah_attr *attr,
234 				   struct ib_udata *udata)
235 
236 {
237 	int err;
238 	struct rxe_dev *rxe = to_rdev(ibpd->device);
239 	struct rxe_pd *pd = to_rpd(ibpd);
240 	struct rxe_ah *ah;
241 
242 	err = rxe_av_chk_attr(rxe, attr);
243 	if (err)
244 		return ERR_PTR(err);
245 
246 	ah = rxe_alloc(&rxe->ah_pool);
247 	if (!ah)
248 		return ERR_PTR(-ENOMEM);
249 
250 	rxe_add_ref(pd);
251 	ah->pd = pd;
252 
253 	rxe_init_av(rxe, attr, &ah->av);
254 	return &ah->ibah;
255 }
256 
257 static int rxe_modify_ah(struct ib_ah *ibah, struct rdma_ah_attr *attr)
258 {
259 	int err;
260 	struct rxe_dev *rxe = to_rdev(ibah->device);
261 	struct rxe_ah *ah = to_rah(ibah);
262 
263 	err = rxe_av_chk_attr(rxe, attr);
264 	if (err)
265 		return err;
266 
267 	rxe_init_av(rxe, attr, &ah->av);
268 	return 0;
269 }
270 
271 static int rxe_query_ah(struct ib_ah *ibah, struct rdma_ah_attr *attr)
272 {
273 	struct rxe_ah *ah = to_rah(ibah);
274 
275 	memset(attr, 0, sizeof(*attr));
276 	attr->type = ibah->type;
277 	rxe_av_to_attr(&ah->av, attr);
278 	return 0;
279 }
280 
281 static int rxe_destroy_ah(struct ib_ah *ibah)
282 {
283 	struct rxe_ah *ah = to_rah(ibah);
284 
285 	rxe_drop_ref(ah->pd);
286 	rxe_drop_ref(ah);
287 	return 0;
288 }
289 
290 static int post_one_recv(struct rxe_rq *rq, const struct ib_recv_wr *ibwr)
291 {
292 	int err;
293 	int i;
294 	u32 length;
295 	struct rxe_recv_wqe *recv_wqe;
296 	int num_sge = ibwr->num_sge;
297 
298 	if (unlikely(queue_full(rq->queue))) {
299 		err = -ENOMEM;
300 		goto err1;
301 	}
302 
303 	if (unlikely(num_sge > rq->max_sge)) {
304 		err = -EINVAL;
305 		goto err1;
306 	}
307 
308 	length = 0;
309 	for (i = 0; i < num_sge; i++)
310 		length += ibwr->sg_list[i].length;
311 
312 	recv_wqe = producer_addr(rq->queue);
313 	recv_wqe->wr_id = ibwr->wr_id;
314 	recv_wqe->num_sge = num_sge;
315 
316 	memcpy(recv_wqe->dma.sge, ibwr->sg_list,
317 	       num_sge * sizeof(struct ib_sge));
318 
319 	recv_wqe->dma.length		= length;
320 	recv_wqe->dma.resid		= length;
321 	recv_wqe->dma.num_sge		= num_sge;
322 	recv_wqe->dma.cur_sge		= 0;
323 	recv_wqe->dma.sge_offset	= 0;
324 
325 	/* make sure all changes to the work queue are written before we
326 	 * update the producer pointer
327 	 */
328 	smp_wmb();
329 
330 	advance_producer(rq->queue);
331 	return 0;
332 
333 err1:
334 	return err;
335 }
336 
337 static struct ib_srq *rxe_create_srq(struct ib_pd *ibpd,
338 				     struct ib_srq_init_attr *init,
339 				     struct ib_udata *udata)
340 {
341 	int err;
342 	struct rxe_dev *rxe = to_rdev(ibpd->device);
343 	struct rxe_pd *pd = to_rpd(ibpd);
344 	struct rxe_srq *srq;
345 	struct ib_ucontext *context = udata ? ibpd->uobject->context : NULL;
346 	struct rxe_create_srq_resp __user *uresp = NULL;
347 
348 	if (udata) {
349 		if (udata->outlen < sizeof(*uresp))
350 			return ERR_PTR(-EINVAL);
351 		uresp = udata->outbuf;
352 	}
353 
354 	err = rxe_srq_chk_attr(rxe, NULL, &init->attr, IB_SRQ_INIT_MASK);
355 	if (err)
356 		goto err1;
357 
358 	srq = rxe_alloc(&rxe->srq_pool);
359 	if (!srq) {
360 		err = -ENOMEM;
361 		goto err1;
362 	}
363 
364 	rxe_add_index(srq);
365 	rxe_add_ref(pd);
366 	srq->pd = pd;
367 
368 	err = rxe_srq_from_init(rxe, srq, init, context, uresp);
369 	if (err)
370 		goto err2;
371 
372 	return &srq->ibsrq;
373 
374 err2:
375 	rxe_drop_ref(pd);
376 	rxe_drop_index(srq);
377 	rxe_drop_ref(srq);
378 err1:
379 	return ERR_PTR(err);
380 }
381 
382 static int rxe_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
383 			  enum ib_srq_attr_mask mask,
384 			  struct ib_udata *udata)
385 {
386 	int err;
387 	struct rxe_srq *srq = to_rsrq(ibsrq);
388 	struct rxe_dev *rxe = to_rdev(ibsrq->device);
389 	struct rxe_modify_srq_cmd ucmd = {};
390 
391 	if (udata) {
392 		if (udata->inlen < sizeof(ucmd))
393 			return -EINVAL;
394 
395 		err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
396 		if (err)
397 			return err;
398 	}
399 
400 	err = rxe_srq_chk_attr(rxe, srq, attr, mask);
401 	if (err)
402 		goto err1;
403 
404 	err = rxe_srq_from_attr(rxe, srq, attr, mask, &ucmd);
405 	if (err)
406 		goto err1;
407 
408 	return 0;
409 
410 err1:
411 	return err;
412 }
413 
414 static int rxe_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr)
415 {
416 	struct rxe_srq *srq = to_rsrq(ibsrq);
417 
418 	if (srq->error)
419 		return -EINVAL;
420 
421 	attr->max_wr = srq->rq.queue->buf->index_mask;
422 	attr->max_sge = srq->rq.max_sge;
423 	attr->srq_limit = srq->limit;
424 	return 0;
425 }
426 
427 static int rxe_destroy_srq(struct ib_srq *ibsrq)
428 {
429 	struct rxe_srq *srq = to_rsrq(ibsrq);
430 
431 	if (srq->rq.queue)
432 		rxe_queue_cleanup(srq->rq.queue);
433 
434 	rxe_drop_ref(srq->pd);
435 	rxe_drop_index(srq);
436 	rxe_drop_ref(srq);
437 
438 	return 0;
439 }
440 
441 static int rxe_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
442 			     const struct ib_recv_wr **bad_wr)
443 {
444 	int err = 0;
445 	unsigned long flags;
446 	struct rxe_srq *srq = to_rsrq(ibsrq);
447 
448 	spin_lock_irqsave(&srq->rq.producer_lock, flags);
449 
450 	while (wr) {
451 		err = post_one_recv(&srq->rq, wr);
452 		if (unlikely(err))
453 			break;
454 		wr = wr->next;
455 	}
456 
457 	spin_unlock_irqrestore(&srq->rq.producer_lock, flags);
458 
459 	if (err)
460 		*bad_wr = wr;
461 
462 	return err;
463 }
464 
465 static struct ib_qp *rxe_create_qp(struct ib_pd *ibpd,
466 				   struct ib_qp_init_attr *init,
467 				   struct ib_udata *udata)
468 {
469 	int err;
470 	struct rxe_dev *rxe = to_rdev(ibpd->device);
471 	struct rxe_pd *pd = to_rpd(ibpd);
472 	struct rxe_qp *qp;
473 	struct rxe_create_qp_resp __user *uresp = NULL;
474 
475 	if (udata) {
476 		if (udata->outlen < sizeof(*uresp))
477 			return ERR_PTR(-EINVAL);
478 		uresp = udata->outbuf;
479 	}
480 
481 	err = rxe_qp_chk_init(rxe, init);
482 	if (err)
483 		goto err1;
484 
485 	qp = rxe_alloc(&rxe->qp_pool);
486 	if (!qp) {
487 		err = -ENOMEM;
488 		goto err1;
489 	}
490 
491 	if (udata) {
492 		if (udata->inlen) {
493 			err = -EINVAL;
494 			goto err2;
495 		}
496 		qp->is_user = 1;
497 	}
498 
499 	rxe_add_index(qp);
500 
501 	err = rxe_qp_from_init(rxe, qp, pd, init, uresp, ibpd);
502 	if (err)
503 		goto err3;
504 
505 	return &qp->ibqp;
506 
507 err3:
508 	rxe_drop_index(qp);
509 err2:
510 	rxe_drop_ref(qp);
511 err1:
512 	return ERR_PTR(err);
513 }
514 
515 static int rxe_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
516 			 int mask, struct ib_udata *udata)
517 {
518 	int err;
519 	struct rxe_dev *rxe = to_rdev(ibqp->device);
520 	struct rxe_qp *qp = to_rqp(ibqp);
521 
522 	err = rxe_qp_chk_attr(rxe, qp, attr, mask);
523 	if (err)
524 		goto err1;
525 
526 	err = rxe_qp_from_attr(qp, attr, mask, udata);
527 	if (err)
528 		goto err1;
529 
530 	return 0;
531 
532 err1:
533 	return err;
534 }
535 
536 static int rxe_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
537 			int mask, struct ib_qp_init_attr *init)
538 {
539 	struct rxe_qp *qp = to_rqp(ibqp);
540 
541 	rxe_qp_to_init(qp, init);
542 	rxe_qp_to_attr(qp, attr, mask);
543 
544 	return 0;
545 }
546 
547 static int rxe_destroy_qp(struct ib_qp *ibqp)
548 {
549 	struct rxe_qp *qp = to_rqp(ibqp);
550 
551 	rxe_qp_destroy(qp);
552 	rxe_drop_index(qp);
553 	rxe_drop_ref(qp);
554 	return 0;
555 }
556 
557 static int validate_send_wr(struct rxe_qp *qp, const struct ib_send_wr *ibwr,
558 			    unsigned int mask, unsigned int length)
559 {
560 	int num_sge = ibwr->num_sge;
561 	struct rxe_sq *sq = &qp->sq;
562 
563 	if (unlikely(num_sge > sq->max_sge))
564 		goto err1;
565 
566 	if (unlikely(mask & WR_ATOMIC_MASK)) {
567 		if (length < 8)
568 			goto err1;
569 
570 		if (atomic_wr(ibwr)->remote_addr & 0x7)
571 			goto err1;
572 	}
573 
574 	if (unlikely((ibwr->send_flags & IB_SEND_INLINE) &&
575 		     (length > sq->max_inline)))
576 		goto err1;
577 
578 	return 0;
579 
580 err1:
581 	return -EINVAL;
582 }
583 
584 static void init_send_wr(struct rxe_qp *qp, struct rxe_send_wr *wr,
585 			 const struct ib_send_wr *ibwr)
586 {
587 	wr->wr_id = ibwr->wr_id;
588 	wr->num_sge = ibwr->num_sge;
589 	wr->opcode = ibwr->opcode;
590 	wr->send_flags = ibwr->send_flags;
591 
592 	if (qp_type(qp) == IB_QPT_UD ||
593 	    qp_type(qp) == IB_QPT_SMI ||
594 	    qp_type(qp) == IB_QPT_GSI) {
595 		wr->wr.ud.remote_qpn = ud_wr(ibwr)->remote_qpn;
596 		wr->wr.ud.remote_qkey = ud_wr(ibwr)->remote_qkey;
597 		if (qp_type(qp) == IB_QPT_GSI)
598 			wr->wr.ud.pkey_index = ud_wr(ibwr)->pkey_index;
599 		if (wr->opcode == IB_WR_SEND_WITH_IMM)
600 			wr->ex.imm_data = ibwr->ex.imm_data;
601 	} else {
602 		switch (wr->opcode) {
603 		case IB_WR_RDMA_WRITE_WITH_IMM:
604 			wr->ex.imm_data = ibwr->ex.imm_data;
605 			/* fall through */
606 		case IB_WR_RDMA_READ:
607 		case IB_WR_RDMA_WRITE:
608 			wr->wr.rdma.remote_addr = rdma_wr(ibwr)->remote_addr;
609 			wr->wr.rdma.rkey	= rdma_wr(ibwr)->rkey;
610 			break;
611 		case IB_WR_SEND_WITH_IMM:
612 			wr->ex.imm_data = ibwr->ex.imm_data;
613 			break;
614 		case IB_WR_SEND_WITH_INV:
615 			wr->ex.invalidate_rkey = ibwr->ex.invalidate_rkey;
616 			break;
617 		case IB_WR_ATOMIC_CMP_AND_SWP:
618 		case IB_WR_ATOMIC_FETCH_AND_ADD:
619 			wr->wr.atomic.remote_addr =
620 				atomic_wr(ibwr)->remote_addr;
621 			wr->wr.atomic.compare_add =
622 				atomic_wr(ibwr)->compare_add;
623 			wr->wr.atomic.swap = atomic_wr(ibwr)->swap;
624 			wr->wr.atomic.rkey = atomic_wr(ibwr)->rkey;
625 			break;
626 		case IB_WR_LOCAL_INV:
627 			wr->ex.invalidate_rkey = ibwr->ex.invalidate_rkey;
628 		break;
629 		case IB_WR_REG_MR:
630 			wr->wr.reg.mr = reg_wr(ibwr)->mr;
631 			wr->wr.reg.key = reg_wr(ibwr)->key;
632 			wr->wr.reg.access = reg_wr(ibwr)->access;
633 		break;
634 		default:
635 			break;
636 		}
637 	}
638 }
639 
640 static int init_send_wqe(struct rxe_qp *qp, const struct ib_send_wr *ibwr,
641 			 unsigned int mask, unsigned int length,
642 			 struct rxe_send_wqe *wqe)
643 {
644 	int num_sge = ibwr->num_sge;
645 	struct ib_sge *sge;
646 	int i;
647 	u8 *p;
648 
649 	init_send_wr(qp, &wqe->wr, ibwr);
650 
651 	if (qp_type(qp) == IB_QPT_UD ||
652 	    qp_type(qp) == IB_QPT_SMI ||
653 	    qp_type(qp) == IB_QPT_GSI)
654 		memcpy(&wqe->av, &to_rah(ud_wr(ibwr)->ah)->av, sizeof(wqe->av));
655 
656 	if (unlikely(ibwr->send_flags & IB_SEND_INLINE)) {
657 		p = wqe->dma.inline_data;
658 
659 		sge = ibwr->sg_list;
660 		for (i = 0; i < num_sge; i++, sge++) {
661 			memcpy(p, (void *)(uintptr_t)sge->addr,
662 					sge->length);
663 
664 			p += sge->length;
665 		}
666 	} else if (mask & WR_REG_MASK) {
667 		wqe->mask = mask;
668 		wqe->state = wqe_state_posted;
669 		return 0;
670 	} else
671 		memcpy(wqe->dma.sge, ibwr->sg_list,
672 		       num_sge * sizeof(struct ib_sge));
673 
674 	wqe->iova = mask & WR_ATOMIC_MASK ? atomic_wr(ibwr)->remote_addr :
675 		mask & WR_READ_OR_WRITE_MASK ? rdma_wr(ibwr)->remote_addr : 0;
676 	wqe->mask		= mask;
677 	wqe->dma.length		= length;
678 	wqe->dma.resid		= length;
679 	wqe->dma.num_sge	= num_sge;
680 	wqe->dma.cur_sge	= 0;
681 	wqe->dma.sge_offset	= 0;
682 	wqe->state		= wqe_state_posted;
683 	wqe->ssn		= atomic_add_return(1, &qp->ssn);
684 
685 	return 0;
686 }
687 
688 static int post_one_send(struct rxe_qp *qp, const struct ib_send_wr *ibwr,
689 			 unsigned int mask, u32 length)
690 {
691 	int err;
692 	struct rxe_sq *sq = &qp->sq;
693 	struct rxe_send_wqe *send_wqe;
694 	unsigned long flags;
695 
696 	err = validate_send_wr(qp, ibwr, mask, length);
697 	if (err)
698 		return err;
699 
700 	spin_lock_irqsave(&qp->sq.sq_lock, flags);
701 
702 	if (unlikely(queue_full(sq->queue))) {
703 		err = -ENOMEM;
704 		goto err1;
705 	}
706 
707 	send_wqe = producer_addr(sq->queue);
708 
709 	err = init_send_wqe(qp, ibwr, mask, length, send_wqe);
710 	if (unlikely(err))
711 		goto err1;
712 
713 	/*
714 	 * make sure all changes to the work queue are
715 	 * written before we update the producer pointer
716 	 */
717 	smp_wmb();
718 
719 	advance_producer(sq->queue);
720 	spin_unlock_irqrestore(&qp->sq.sq_lock, flags);
721 
722 	return 0;
723 
724 err1:
725 	spin_unlock_irqrestore(&qp->sq.sq_lock, flags);
726 	return err;
727 }
728 
729 static int rxe_post_send_kernel(struct rxe_qp *qp, const struct ib_send_wr *wr,
730 				const struct ib_send_wr **bad_wr)
731 {
732 	int err = 0;
733 	unsigned int mask;
734 	unsigned int length = 0;
735 	int i;
736 
737 	while (wr) {
738 		mask = wr_opcode_mask(wr->opcode, qp);
739 		if (unlikely(!mask)) {
740 			err = -EINVAL;
741 			*bad_wr = wr;
742 			break;
743 		}
744 
745 		if (unlikely((wr->send_flags & IB_SEND_INLINE) &&
746 			     !(mask & WR_INLINE_MASK))) {
747 			err = -EINVAL;
748 			*bad_wr = wr;
749 			break;
750 		}
751 
752 		length = 0;
753 		for (i = 0; i < wr->num_sge; i++)
754 			length += wr->sg_list[i].length;
755 
756 		err = post_one_send(qp, wr, mask, length);
757 
758 		if (err) {
759 			*bad_wr = wr;
760 			break;
761 		}
762 		wr = wr->next;
763 	}
764 
765 	rxe_run_task(&qp->req.task, 1);
766 	if (unlikely(qp->req.state == QP_STATE_ERROR))
767 		rxe_run_task(&qp->comp.task, 1);
768 
769 	return err;
770 }
771 
772 static int rxe_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
773 			 const struct ib_send_wr **bad_wr)
774 {
775 	struct rxe_qp *qp = to_rqp(ibqp);
776 
777 	if (unlikely(!qp->valid)) {
778 		*bad_wr = wr;
779 		return -EINVAL;
780 	}
781 
782 	if (unlikely(qp->req.state < QP_STATE_READY)) {
783 		*bad_wr = wr;
784 		return -EINVAL;
785 	}
786 
787 	if (qp->is_user) {
788 		/* Utilize process context to do protocol processing */
789 		rxe_run_task(&qp->req.task, 0);
790 		return 0;
791 	} else
792 		return rxe_post_send_kernel(qp, wr, bad_wr);
793 }
794 
795 static int rxe_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
796 			 const struct ib_recv_wr **bad_wr)
797 {
798 	int err = 0;
799 	struct rxe_qp *qp = to_rqp(ibqp);
800 	struct rxe_rq *rq = &qp->rq;
801 	unsigned long flags;
802 
803 	if (unlikely((qp_state(qp) < IB_QPS_INIT) || !qp->valid)) {
804 		*bad_wr = wr;
805 		err = -EINVAL;
806 		goto err1;
807 	}
808 
809 	if (unlikely(qp->srq)) {
810 		*bad_wr = wr;
811 		err = -EINVAL;
812 		goto err1;
813 	}
814 
815 	spin_lock_irqsave(&rq->producer_lock, flags);
816 
817 	while (wr) {
818 		err = post_one_recv(rq, wr);
819 		if (unlikely(err)) {
820 			*bad_wr = wr;
821 			break;
822 		}
823 		wr = wr->next;
824 	}
825 
826 	spin_unlock_irqrestore(&rq->producer_lock, flags);
827 
828 	if (qp->resp.state == QP_STATE_ERROR)
829 		rxe_run_task(&qp->resp.task, 1);
830 
831 err1:
832 	return err;
833 }
834 
835 static struct ib_cq *rxe_create_cq(struct ib_device *dev,
836 				   const struct ib_cq_init_attr *attr,
837 				   struct ib_ucontext *context,
838 				   struct ib_udata *udata)
839 {
840 	int err;
841 	struct rxe_dev *rxe = to_rdev(dev);
842 	struct rxe_cq *cq;
843 	struct rxe_create_cq_resp __user *uresp = NULL;
844 
845 	if (udata) {
846 		if (udata->outlen < sizeof(*uresp))
847 			return ERR_PTR(-EINVAL);
848 		uresp = udata->outbuf;
849 	}
850 
851 	if (attr->flags)
852 		return ERR_PTR(-EINVAL);
853 
854 	err = rxe_cq_chk_attr(rxe, NULL, attr->cqe, attr->comp_vector);
855 	if (err)
856 		goto err1;
857 
858 	cq = rxe_alloc(&rxe->cq_pool);
859 	if (!cq) {
860 		err = -ENOMEM;
861 		goto err1;
862 	}
863 
864 	err = rxe_cq_from_init(rxe, cq, attr->cqe, attr->comp_vector,
865 			       context, uresp);
866 	if (err)
867 		goto err2;
868 
869 	return &cq->ibcq;
870 
871 err2:
872 	rxe_drop_ref(cq);
873 err1:
874 	return ERR_PTR(err);
875 }
876 
877 static int rxe_destroy_cq(struct ib_cq *ibcq)
878 {
879 	struct rxe_cq *cq = to_rcq(ibcq);
880 
881 	rxe_cq_disable(cq);
882 
883 	rxe_drop_ref(cq);
884 	return 0;
885 }
886 
887 static int rxe_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata)
888 {
889 	int err;
890 	struct rxe_cq *cq = to_rcq(ibcq);
891 	struct rxe_dev *rxe = to_rdev(ibcq->device);
892 	struct rxe_resize_cq_resp __user *uresp = NULL;
893 
894 	if (udata) {
895 		if (udata->outlen < sizeof(*uresp))
896 			return -EINVAL;
897 		uresp = udata->outbuf;
898 	}
899 
900 	err = rxe_cq_chk_attr(rxe, cq, cqe, 0);
901 	if (err)
902 		goto err1;
903 
904 	err = rxe_cq_resize_queue(cq, cqe, uresp);
905 	if (err)
906 		goto err1;
907 
908 	return 0;
909 
910 err1:
911 	return err;
912 }
913 
914 static int rxe_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
915 {
916 	int i;
917 	struct rxe_cq *cq = to_rcq(ibcq);
918 	struct rxe_cqe *cqe;
919 	unsigned long flags;
920 
921 	spin_lock_irqsave(&cq->cq_lock, flags);
922 	for (i = 0; i < num_entries; i++) {
923 		cqe = queue_head(cq->queue);
924 		if (!cqe)
925 			break;
926 
927 		memcpy(wc++, &cqe->ibwc, sizeof(*wc));
928 		advance_consumer(cq->queue);
929 	}
930 	spin_unlock_irqrestore(&cq->cq_lock, flags);
931 
932 	return i;
933 }
934 
935 static int rxe_peek_cq(struct ib_cq *ibcq, int wc_cnt)
936 {
937 	struct rxe_cq *cq = to_rcq(ibcq);
938 	int count = queue_count(cq->queue);
939 
940 	return (count > wc_cnt) ? wc_cnt : count;
941 }
942 
943 static int rxe_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
944 {
945 	struct rxe_cq *cq = to_rcq(ibcq);
946 	unsigned long irq_flags;
947 	int ret = 0;
948 
949 	spin_lock_irqsave(&cq->cq_lock, irq_flags);
950 	if (cq->notify != IB_CQ_NEXT_COMP)
951 		cq->notify = flags & IB_CQ_SOLICITED_MASK;
952 
953 	if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && !queue_empty(cq->queue))
954 		ret = 1;
955 
956 	spin_unlock_irqrestore(&cq->cq_lock, irq_flags);
957 
958 	return ret;
959 }
960 
961 static struct ib_mr *rxe_get_dma_mr(struct ib_pd *ibpd, int access)
962 {
963 	struct rxe_dev *rxe = to_rdev(ibpd->device);
964 	struct rxe_pd *pd = to_rpd(ibpd);
965 	struct rxe_mem *mr;
966 	int err;
967 
968 	mr = rxe_alloc(&rxe->mr_pool);
969 	if (!mr) {
970 		err = -ENOMEM;
971 		goto err1;
972 	}
973 
974 	rxe_add_index(mr);
975 
976 	rxe_add_ref(pd);
977 
978 	err = rxe_mem_init_dma(pd, access, mr);
979 	if (err)
980 		goto err2;
981 
982 	return &mr->ibmr;
983 
984 err2:
985 	rxe_drop_ref(pd);
986 	rxe_drop_index(mr);
987 	rxe_drop_ref(mr);
988 err1:
989 	return ERR_PTR(err);
990 }
991 
992 static struct ib_mr *rxe_reg_user_mr(struct ib_pd *ibpd,
993 				     u64 start,
994 				     u64 length,
995 				     u64 iova,
996 				     int access, struct ib_udata *udata)
997 {
998 	int err;
999 	struct rxe_dev *rxe = to_rdev(ibpd->device);
1000 	struct rxe_pd *pd = to_rpd(ibpd);
1001 	struct rxe_mem *mr;
1002 
1003 	mr = rxe_alloc(&rxe->mr_pool);
1004 	if (!mr) {
1005 		err = -ENOMEM;
1006 		goto err2;
1007 	}
1008 
1009 	rxe_add_index(mr);
1010 
1011 	rxe_add_ref(pd);
1012 
1013 	err = rxe_mem_init_user(pd, start, length, iova,
1014 				access, udata, mr);
1015 	if (err)
1016 		goto err3;
1017 
1018 	return &mr->ibmr;
1019 
1020 err3:
1021 	rxe_drop_ref(pd);
1022 	rxe_drop_index(mr);
1023 	rxe_drop_ref(mr);
1024 err2:
1025 	return ERR_PTR(err);
1026 }
1027 
1028 static int rxe_dereg_mr(struct ib_mr *ibmr)
1029 {
1030 	struct rxe_mem *mr = to_rmr(ibmr);
1031 
1032 	mr->state = RXE_MEM_STATE_ZOMBIE;
1033 	rxe_drop_ref(mr->pd);
1034 	rxe_drop_index(mr);
1035 	rxe_drop_ref(mr);
1036 	return 0;
1037 }
1038 
1039 static struct ib_mr *rxe_alloc_mr(struct ib_pd *ibpd,
1040 				  enum ib_mr_type mr_type,
1041 				  u32 max_num_sg)
1042 {
1043 	struct rxe_dev *rxe = to_rdev(ibpd->device);
1044 	struct rxe_pd *pd = to_rpd(ibpd);
1045 	struct rxe_mem *mr;
1046 	int err;
1047 
1048 	if (mr_type != IB_MR_TYPE_MEM_REG)
1049 		return ERR_PTR(-EINVAL);
1050 
1051 	mr = rxe_alloc(&rxe->mr_pool);
1052 	if (!mr) {
1053 		err = -ENOMEM;
1054 		goto err1;
1055 	}
1056 
1057 	rxe_add_index(mr);
1058 
1059 	rxe_add_ref(pd);
1060 
1061 	err = rxe_mem_init_fast(pd, max_num_sg, mr);
1062 	if (err)
1063 		goto err2;
1064 
1065 	return &mr->ibmr;
1066 
1067 err2:
1068 	rxe_drop_ref(pd);
1069 	rxe_drop_index(mr);
1070 	rxe_drop_ref(mr);
1071 err1:
1072 	return ERR_PTR(err);
1073 }
1074 
1075 static int rxe_set_page(struct ib_mr *ibmr, u64 addr)
1076 {
1077 	struct rxe_mem *mr = to_rmr(ibmr);
1078 	struct rxe_map *map;
1079 	struct rxe_phys_buf *buf;
1080 
1081 	if (unlikely(mr->nbuf == mr->num_buf))
1082 		return -ENOMEM;
1083 
1084 	map = mr->map[mr->nbuf / RXE_BUF_PER_MAP];
1085 	buf = &map->buf[mr->nbuf % RXE_BUF_PER_MAP];
1086 
1087 	buf->addr = addr;
1088 	buf->size = ibmr->page_size;
1089 	mr->nbuf++;
1090 
1091 	return 0;
1092 }
1093 
1094 static int rxe_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg,
1095 			 int sg_nents, unsigned int *sg_offset)
1096 {
1097 	struct rxe_mem *mr = to_rmr(ibmr);
1098 	int n;
1099 
1100 	mr->nbuf = 0;
1101 
1102 	n = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, rxe_set_page);
1103 
1104 	mr->va = ibmr->iova;
1105 	mr->iova = ibmr->iova;
1106 	mr->length = ibmr->length;
1107 	mr->page_shift = ilog2(ibmr->page_size);
1108 	mr->page_mask = ibmr->page_size - 1;
1109 	mr->offset = mr->iova & mr->page_mask;
1110 
1111 	return n;
1112 }
1113 
1114 static int rxe_attach_mcast(struct ib_qp *ibqp, union ib_gid *mgid, u16 mlid)
1115 {
1116 	int err;
1117 	struct rxe_dev *rxe = to_rdev(ibqp->device);
1118 	struct rxe_qp *qp = to_rqp(ibqp);
1119 	struct rxe_mc_grp *grp;
1120 
1121 	/* takes a ref on grp if successful */
1122 	err = rxe_mcast_get_grp(rxe, mgid, &grp);
1123 	if (err)
1124 		return err;
1125 
1126 	err = rxe_mcast_add_grp_elem(rxe, qp, grp);
1127 
1128 	rxe_drop_ref(grp);
1129 	return err;
1130 }
1131 
1132 static int rxe_detach_mcast(struct ib_qp *ibqp, union ib_gid *mgid, u16 mlid)
1133 {
1134 	struct rxe_dev *rxe = to_rdev(ibqp->device);
1135 	struct rxe_qp *qp = to_rqp(ibqp);
1136 
1137 	return rxe_mcast_drop_grp_elem(rxe, qp, mgid);
1138 }
1139 
1140 static ssize_t parent_show(struct device *device,
1141 			   struct device_attribute *attr, char *buf)
1142 {
1143 	struct rxe_dev *rxe = container_of(device, struct rxe_dev,
1144 					   ib_dev.dev);
1145 
1146 	return snprintf(buf, 16, "%s\n", rxe_parent_name(rxe, 1));
1147 }
1148 
1149 static DEVICE_ATTR_RO(parent);
1150 
1151 static struct device_attribute *rxe_dev_attributes[] = {
1152 	&dev_attr_parent,
1153 };
1154 
1155 int rxe_register_device(struct rxe_dev *rxe)
1156 {
1157 	int err;
1158 	int i;
1159 	struct ib_device *dev = &rxe->ib_dev;
1160 	struct crypto_shash *tfm;
1161 
1162 	strlcpy(dev->name, "rxe%d", IB_DEVICE_NAME_MAX);
1163 	strlcpy(dev->node_desc, "rxe", sizeof(dev->node_desc));
1164 
1165 	dev->owner = THIS_MODULE;
1166 	dev->node_type = RDMA_NODE_IB_CA;
1167 	dev->phys_port_cnt = 1;
1168 	dev->num_comp_vectors = num_possible_cpus();
1169 	dev->dev.parent = rxe_dma_device(rxe);
1170 	dev->local_dma_lkey = 0;
1171 	addrconf_addr_eui48((unsigned char *)&dev->node_guid,
1172 			    rxe->ndev->dev_addr);
1173 	dev->dev.dma_ops = &dma_virt_ops;
1174 	dma_coerce_mask_and_coherent(&dev->dev,
1175 				     dma_get_required_mask(&dev->dev));
1176 
1177 	dev->uverbs_abi_ver = RXE_UVERBS_ABI_VERSION;
1178 	dev->uverbs_cmd_mask = BIT_ULL(IB_USER_VERBS_CMD_GET_CONTEXT)
1179 	    | BIT_ULL(IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)
1180 	    | BIT_ULL(IB_USER_VERBS_CMD_QUERY_DEVICE)
1181 	    | BIT_ULL(IB_USER_VERBS_CMD_QUERY_PORT)
1182 	    | BIT_ULL(IB_USER_VERBS_CMD_ALLOC_PD)
1183 	    | BIT_ULL(IB_USER_VERBS_CMD_DEALLOC_PD)
1184 	    | BIT_ULL(IB_USER_VERBS_CMD_CREATE_SRQ)
1185 	    | BIT_ULL(IB_USER_VERBS_CMD_MODIFY_SRQ)
1186 	    | BIT_ULL(IB_USER_VERBS_CMD_QUERY_SRQ)
1187 	    | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_SRQ)
1188 	    | BIT_ULL(IB_USER_VERBS_CMD_POST_SRQ_RECV)
1189 	    | BIT_ULL(IB_USER_VERBS_CMD_CREATE_QP)
1190 	    | BIT_ULL(IB_USER_VERBS_CMD_MODIFY_QP)
1191 	    | BIT_ULL(IB_USER_VERBS_CMD_QUERY_QP)
1192 	    | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_QP)
1193 	    | BIT_ULL(IB_USER_VERBS_CMD_POST_SEND)
1194 	    | BIT_ULL(IB_USER_VERBS_CMD_POST_RECV)
1195 	    | BIT_ULL(IB_USER_VERBS_CMD_CREATE_CQ)
1196 	    | BIT_ULL(IB_USER_VERBS_CMD_RESIZE_CQ)
1197 	    | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_CQ)
1198 	    | BIT_ULL(IB_USER_VERBS_CMD_POLL_CQ)
1199 	    | BIT_ULL(IB_USER_VERBS_CMD_PEEK_CQ)
1200 	    | BIT_ULL(IB_USER_VERBS_CMD_REQ_NOTIFY_CQ)
1201 	    | BIT_ULL(IB_USER_VERBS_CMD_REG_MR)
1202 	    | BIT_ULL(IB_USER_VERBS_CMD_DEREG_MR)
1203 	    | BIT_ULL(IB_USER_VERBS_CMD_CREATE_AH)
1204 	    | BIT_ULL(IB_USER_VERBS_CMD_MODIFY_AH)
1205 	    | BIT_ULL(IB_USER_VERBS_CMD_QUERY_AH)
1206 	    | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_AH)
1207 	    | BIT_ULL(IB_USER_VERBS_CMD_ATTACH_MCAST)
1208 	    | BIT_ULL(IB_USER_VERBS_CMD_DETACH_MCAST)
1209 	    ;
1210 
1211 	dev->query_device = rxe_query_device;
1212 	dev->modify_device = rxe_modify_device;
1213 	dev->query_port = rxe_query_port;
1214 	dev->modify_port = rxe_modify_port;
1215 	dev->get_link_layer = rxe_get_link_layer;
1216 	dev->get_netdev = rxe_get_netdev;
1217 	dev->query_pkey = rxe_query_pkey;
1218 	dev->alloc_ucontext = rxe_alloc_ucontext;
1219 	dev->dealloc_ucontext = rxe_dealloc_ucontext;
1220 	dev->mmap = rxe_mmap;
1221 	dev->get_port_immutable = rxe_port_immutable;
1222 	dev->alloc_pd = rxe_alloc_pd;
1223 	dev->dealloc_pd = rxe_dealloc_pd;
1224 	dev->create_ah = rxe_create_ah;
1225 	dev->modify_ah = rxe_modify_ah;
1226 	dev->query_ah = rxe_query_ah;
1227 	dev->destroy_ah = rxe_destroy_ah;
1228 	dev->create_srq = rxe_create_srq;
1229 	dev->modify_srq = rxe_modify_srq;
1230 	dev->query_srq = rxe_query_srq;
1231 	dev->destroy_srq = rxe_destroy_srq;
1232 	dev->post_srq_recv = rxe_post_srq_recv;
1233 	dev->create_qp = rxe_create_qp;
1234 	dev->modify_qp = rxe_modify_qp;
1235 	dev->query_qp = rxe_query_qp;
1236 	dev->destroy_qp = rxe_destroy_qp;
1237 	dev->post_send = rxe_post_send;
1238 	dev->post_recv = rxe_post_recv;
1239 	dev->create_cq = rxe_create_cq;
1240 	dev->destroy_cq = rxe_destroy_cq;
1241 	dev->resize_cq = rxe_resize_cq;
1242 	dev->poll_cq = rxe_poll_cq;
1243 	dev->peek_cq = rxe_peek_cq;
1244 	dev->req_notify_cq = rxe_req_notify_cq;
1245 	dev->get_dma_mr = rxe_get_dma_mr;
1246 	dev->reg_user_mr = rxe_reg_user_mr;
1247 	dev->dereg_mr = rxe_dereg_mr;
1248 	dev->alloc_mr = rxe_alloc_mr;
1249 	dev->map_mr_sg = rxe_map_mr_sg;
1250 	dev->attach_mcast = rxe_attach_mcast;
1251 	dev->detach_mcast = rxe_detach_mcast;
1252 	dev->get_hw_stats = rxe_ib_get_hw_stats;
1253 	dev->alloc_hw_stats = rxe_ib_alloc_hw_stats;
1254 
1255 	tfm = crypto_alloc_shash("crc32", 0, 0);
1256 	if (IS_ERR(tfm)) {
1257 		pr_err("failed to allocate crc algorithm err:%ld\n",
1258 		       PTR_ERR(tfm));
1259 		return PTR_ERR(tfm);
1260 	}
1261 	rxe->tfm = tfm;
1262 
1263 	dev->driver_id = RDMA_DRIVER_RXE;
1264 	err = ib_register_device(dev, NULL);
1265 	if (err) {
1266 		pr_warn("%s failed with error %d\n", __func__, err);
1267 		goto err1;
1268 	}
1269 
1270 	for (i = 0; i < ARRAY_SIZE(rxe_dev_attributes); ++i) {
1271 		err = device_create_file(&dev->dev, rxe_dev_attributes[i]);
1272 		if (err) {
1273 			pr_warn("%s failed with error %d for attr number %d\n",
1274 				__func__, err, i);
1275 			goto err2;
1276 		}
1277 	}
1278 
1279 	return 0;
1280 
1281 err2:
1282 	ib_unregister_device(dev);
1283 err1:
1284 	crypto_free_shash(rxe->tfm);
1285 
1286 	return err;
1287 }
1288 
1289 int rxe_unregister_device(struct rxe_dev *rxe)
1290 {
1291 	int i;
1292 	struct ib_device *dev = &rxe->ib_dev;
1293 
1294 	for (i = 0; i < ARRAY_SIZE(rxe_dev_attributes); ++i)
1295 		device_remove_file(&dev->dev, rxe_dev_attributes[i]);
1296 
1297 	ib_unregister_device(dev);
1298 
1299 	return 0;
1300 }
1301