1 /*
2  * Copyright (c) 2012-2016 VMware, Inc.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of EITHER the GNU General Public License
6  * version 2 as published by the Free Software Foundation or the BSD
7  * 2-Clause License. This program is distributed in the hope that it
8  * will be useful, but WITHOUT ANY WARRANTY; WITHOUT EVEN THE IMPLIED
9  * WARRANTY OF MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE.
10  * See the GNU General Public License version 2 for more details at
11  * http://www.gnu.org/licenses/old-licenses/gpl-2.0.en.html.
12  *
13  * You should have received a copy of the GNU General Public License
14  * along with this program available in the file COPYING in the main
15  * directory of this source tree.
16  *
17  * The BSD 2-Clause License
18  *
19  *     Redistribution and use in source and binary forms, with or
20  *     without modification, are permitted provided that the following
21  *     conditions are met:
22  *
23  *      - Redistributions of source code must retain the above
24  *        copyright notice, this list of conditions and the following
25  *        disclaimer.
26  *
27  *      - Redistributions in binary form must reproduce the above
28  *        copyright notice, this list of conditions and the following
29  *        disclaimer in the documentation and/or other materials
30  *        provided with the distribution.
31  *
32  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
33  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
34  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
35  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
36  * COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
37  * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
38  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
39  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
40  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
41  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
42  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
43  * OF THE POSSIBILITY OF SUCH DAMAGE.
44  */
45 
46 #include <asm/page.h>
47 #include <linux/io.h>
48 #include <linux/wait.h>
49 #include <rdma/ib_addr.h>
50 #include <rdma/ib_smi.h>
51 #include <rdma/ib_user_verbs.h>
52 
53 #include "pvrdma.h"
54 
55 static inline void get_cqs(struct pvrdma_qp *qp, struct pvrdma_cq **send_cq,
56 			   struct pvrdma_cq **recv_cq)
57 {
58 	*send_cq = to_vcq(qp->ibqp.send_cq);
59 	*recv_cq = to_vcq(qp->ibqp.recv_cq);
60 }
61 
62 static void pvrdma_lock_cqs(struct pvrdma_cq *scq, struct pvrdma_cq *rcq,
63 			    unsigned long *scq_flags,
64 			    unsigned long *rcq_flags)
65 	__acquires(scq->cq_lock) __acquires(rcq->cq_lock)
66 {
67 	if (scq == rcq) {
68 		spin_lock_irqsave(&scq->cq_lock, *scq_flags);
69 		__acquire(rcq->cq_lock);
70 	} else if (scq->cq_handle < rcq->cq_handle) {
71 		spin_lock_irqsave(&scq->cq_lock, *scq_flags);
72 		spin_lock_irqsave_nested(&rcq->cq_lock, *rcq_flags,
73 					 SINGLE_DEPTH_NESTING);
74 	} else {
75 		spin_lock_irqsave(&rcq->cq_lock, *rcq_flags);
76 		spin_lock_irqsave_nested(&scq->cq_lock, *scq_flags,
77 					 SINGLE_DEPTH_NESTING);
78 	}
79 }
80 
81 static void pvrdma_unlock_cqs(struct pvrdma_cq *scq, struct pvrdma_cq *rcq,
82 			      unsigned long *scq_flags,
83 			      unsigned long *rcq_flags)
84 	__releases(scq->cq_lock) __releases(rcq->cq_lock)
85 {
86 	if (scq == rcq) {
87 		__release(rcq->cq_lock);
88 		spin_unlock_irqrestore(&scq->cq_lock, *scq_flags);
89 	} else if (scq->cq_handle < rcq->cq_handle) {
90 		spin_unlock_irqrestore(&rcq->cq_lock, *rcq_flags);
91 		spin_unlock_irqrestore(&scq->cq_lock, *scq_flags);
92 	} else {
93 		spin_unlock_irqrestore(&scq->cq_lock, *scq_flags);
94 		spin_unlock_irqrestore(&rcq->cq_lock, *rcq_flags);
95 	}
96 }
97 
98 static void pvrdma_reset_qp(struct pvrdma_qp *qp)
99 {
100 	struct pvrdma_cq *scq, *rcq;
101 	unsigned long scq_flags, rcq_flags;
102 
103 	/* Clean up cqes */
104 	get_cqs(qp, &scq, &rcq);
105 	pvrdma_lock_cqs(scq, rcq, &scq_flags, &rcq_flags);
106 
107 	_pvrdma_flush_cqe(qp, scq);
108 	if (scq != rcq)
109 		_pvrdma_flush_cqe(qp, rcq);
110 
111 	pvrdma_unlock_cqs(scq, rcq, &scq_flags, &rcq_flags);
112 
113 	/*
114 	 * Reset queuepair. The checks are because usermode queuepairs won't
115 	 * have kernel ringstates.
116 	 */
117 	if (qp->rq.ring) {
118 		atomic_set(&qp->rq.ring->cons_head, 0);
119 		atomic_set(&qp->rq.ring->prod_tail, 0);
120 	}
121 	if (qp->sq.ring) {
122 		atomic_set(&qp->sq.ring->cons_head, 0);
123 		atomic_set(&qp->sq.ring->prod_tail, 0);
124 	}
125 }
126 
127 static int pvrdma_set_rq_size(struct pvrdma_dev *dev,
128 			      struct ib_qp_cap *req_cap,
129 			      struct pvrdma_qp *qp)
130 {
131 	if (req_cap->max_recv_wr > dev->dsr->caps.max_qp_wr ||
132 	    req_cap->max_recv_sge > dev->dsr->caps.max_sge) {
133 		dev_warn(&dev->pdev->dev, "recv queue size invalid\n");
134 		return -EINVAL;
135 	}
136 
137 	qp->rq.wqe_cnt = roundup_pow_of_two(max(1U, req_cap->max_recv_wr));
138 	qp->rq.max_sg = roundup_pow_of_two(max(1U, req_cap->max_recv_sge));
139 
140 	/* Write back */
141 	req_cap->max_recv_wr = qp->rq.wqe_cnt;
142 	req_cap->max_recv_sge = qp->rq.max_sg;
143 
144 	qp->rq.wqe_size = roundup_pow_of_two(sizeof(struct pvrdma_rq_wqe_hdr) +
145 					     sizeof(struct pvrdma_sge) *
146 					     qp->rq.max_sg);
147 	qp->npages_recv = (qp->rq.wqe_cnt * qp->rq.wqe_size + PAGE_SIZE - 1) /
148 			  PAGE_SIZE;
149 
150 	return 0;
151 }
152 
153 static int pvrdma_set_sq_size(struct pvrdma_dev *dev, struct ib_qp_cap *req_cap,
154 			      enum ib_qp_type type, struct pvrdma_qp *qp)
155 {
156 	if (req_cap->max_send_wr > dev->dsr->caps.max_qp_wr ||
157 	    req_cap->max_send_sge > dev->dsr->caps.max_sge) {
158 		dev_warn(&dev->pdev->dev, "send queue size invalid\n");
159 		return -EINVAL;
160 	}
161 
162 	qp->sq.wqe_cnt = roundup_pow_of_two(max(1U, req_cap->max_send_wr));
163 	qp->sq.max_sg = roundup_pow_of_two(max(1U, req_cap->max_send_sge));
164 
165 	/* Write back */
166 	req_cap->max_send_wr = qp->sq.wqe_cnt;
167 	req_cap->max_send_sge = qp->sq.max_sg;
168 
169 	qp->sq.wqe_size = roundup_pow_of_two(sizeof(struct pvrdma_sq_wqe_hdr) +
170 					     sizeof(struct pvrdma_sge) *
171 					     qp->sq.max_sg);
172 	/* Note: one extra page for the header. */
173 	qp->npages_send = 1 + (qp->sq.wqe_cnt * qp->sq.wqe_size +
174 			       PAGE_SIZE - 1) / PAGE_SIZE;
175 
176 	return 0;
177 }
178 
179 /**
180  * pvrdma_create_qp - create queue pair
181  * @pd: protection domain
182  * @init_attr: queue pair attributes
183  * @udata: user data
184  *
185  * @return: the ib_qp pointer on success, otherwise returns an errno.
186  */
187 struct ib_qp *pvrdma_create_qp(struct ib_pd *pd,
188 			       struct ib_qp_init_attr *init_attr,
189 			       struct ib_udata *udata)
190 {
191 	struct pvrdma_qp *qp = NULL;
192 	struct pvrdma_dev *dev = to_vdev(pd->device);
193 	union pvrdma_cmd_req req;
194 	union pvrdma_cmd_resp rsp;
195 	struct pvrdma_cmd_create_qp *cmd = &req.create_qp;
196 	struct pvrdma_cmd_create_qp_resp *resp = &rsp.create_qp_resp;
197 	struct pvrdma_create_qp ucmd;
198 	unsigned long flags;
199 	int ret;
200 
201 	if (init_attr->create_flags) {
202 		dev_warn(&dev->pdev->dev,
203 			 "invalid create queuepair flags %#x\n",
204 			 init_attr->create_flags);
205 		return ERR_PTR(-EINVAL);
206 	}
207 
208 	if (init_attr->qp_type != IB_QPT_RC &&
209 	    init_attr->qp_type != IB_QPT_UD &&
210 	    init_attr->qp_type != IB_QPT_GSI) {
211 		dev_warn(&dev->pdev->dev, "queuepair type %d not supported\n",
212 			 init_attr->qp_type);
213 		return ERR_PTR(-EINVAL);
214 	}
215 
216 	if (!atomic_add_unless(&dev->num_qps, 1, dev->dsr->caps.max_qp))
217 		return ERR_PTR(-ENOMEM);
218 
219 	switch (init_attr->qp_type) {
220 	case IB_QPT_GSI:
221 		if (init_attr->port_num == 0 ||
222 		    init_attr->port_num > pd->device->phys_port_cnt ||
223 		    udata) {
224 			dev_warn(&dev->pdev->dev, "invalid queuepair attrs\n");
225 			ret = -EINVAL;
226 			goto err_qp;
227 		}
228 		/* fall through */
229 	case IB_QPT_RC:
230 	case IB_QPT_UD:
231 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
232 		if (!qp) {
233 			ret = -ENOMEM;
234 			goto err_qp;
235 		}
236 
237 		spin_lock_init(&qp->sq.lock);
238 		spin_lock_init(&qp->rq.lock);
239 		mutex_init(&qp->mutex);
240 		atomic_set(&qp->refcnt, 1);
241 		init_waitqueue_head(&qp->wait);
242 
243 		qp->state = IB_QPS_RESET;
244 
245 		if (pd->uobject && udata) {
246 			dev_dbg(&dev->pdev->dev,
247 				"create queuepair from user space\n");
248 
249 			if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
250 				ret = -EFAULT;
251 				goto err_qp;
252 			}
253 
254 			/* set qp->sq.wqe_cnt, shift, buf_size.. */
255 			qp->rumem = ib_umem_get(pd->uobject->context,
256 						ucmd.rbuf_addr,
257 						ucmd.rbuf_size, 0, 0);
258 			if (IS_ERR(qp->rumem)) {
259 				ret = PTR_ERR(qp->rumem);
260 				goto err_qp;
261 			}
262 
263 			qp->sumem = ib_umem_get(pd->uobject->context,
264 						ucmd.sbuf_addr,
265 						ucmd.sbuf_size, 0, 0);
266 			if (IS_ERR(qp->sumem)) {
267 				ib_umem_release(qp->rumem);
268 				ret = PTR_ERR(qp->sumem);
269 				goto err_qp;
270 			}
271 
272 			qp->npages_send = ib_umem_page_count(qp->sumem);
273 			qp->npages_recv = ib_umem_page_count(qp->rumem);
274 			qp->npages = qp->npages_send + qp->npages_recv;
275 		} else {
276 			qp->is_kernel = true;
277 
278 			ret = pvrdma_set_sq_size(to_vdev(pd->device),
279 						 &init_attr->cap,
280 						 init_attr->qp_type, qp);
281 			if (ret)
282 				goto err_qp;
283 
284 			ret = pvrdma_set_rq_size(to_vdev(pd->device),
285 						 &init_attr->cap, qp);
286 			if (ret)
287 				goto err_qp;
288 
289 			qp->npages = qp->npages_send + qp->npages_recv;
290 
291 			/* Skip header page. */
292 			qp->sq.offset = PAGE_SIZE;
293 
294 			/* Recv queue pages are after send pages. */
295 			qp->rq.offset = qp->npages_send * PAGE_SIZE;
296 		}
297 
298 		if (qp->npages < 0 || qp->npages > PVRDMA_PAGE_DIR_MAX_PAGES) {
299 			dev_warn(&dev->pdev->dev,
300 				 "overflow pages in queuepair\n");
301 			ret = -EINVAL;
302 			goto err_umem;
303 		}
304 
305 		ret = pvrdma_page_dir_init(dev, &qp->pdir, qp->npages,
306 					   qp->is_kernel);
307 		if (ret) {
308 			dev_warn(&dev->pdev->dev,
309 				 "could not allocate page directory\n");
310 			goto err_umem;
311 		}
312 
313 		if (!qp->is_kernel) {
314 			pvrdma_page_dir_insert_umem(&qp->pdir, qp->sumem, 0);
315 			pvrdma_page_dir_insert_umem(&qp->pdir, qp->rumem,
316 						    qp->npages_send);
317 		} else {
318 			/* Ring state is always the first page. */
319 			qp->sq.ring = qp->pdir.pages[0];
320 			qp->rq.ring = &qp->sq.ring[1];
321 		}
322 		break;
323 	default:
324 		ret = -EINVAL;
325 		goto err_qp;
326 	}
327 
328 	/* Not supported */
329 	init_attr->cap.max_inline_data = 0;
330 
331 	memset(cmd, 0, sizeof(*cmd));
332 	cmd->hdr.cmd = PVRDMA_CMD_CREATE_QP;
333 	cmd->pd_handle = to_vpd(pd)->pd_handle;
334 	cmd->send_cq_handle = to_vcq(init_attr->send_cq)->cq_handle;
335 	cmd->recv_cq_handle = to_vcq(init_attr->recv_cq)->cq_handle;
336 	cmd->max_send_wr = init_attr->cap.max_send_wr;
337 	cmd->max_recv_wr = init_attr->cap.max_recv_wr;
338 	cmd->max_send_sge = init_attr->cap.max_send_sge;
339 	cmd->max_recv_sge = init_attr->cap.max_recv_sge;
340 	cmd->max_inline_data = init_attr->cap.max_inline_data;
341 	cmd->sq_sig_all = (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) ? 1 : 0;
342 	cmd->qp_type = ib_qp_type_to_pvrdma(init_attr->qp_type);
343 	cmd->access_flags = IB_ACCESS_LOCAL_WRITE;
344 	cmd->total_chunks = qp->npages;
345 	cmd->send_chunks = qp->npages_send - 1;
346 	cmd->pdir_dma = qp->pdir.dir_dma;
347 
348 	dev_dbg(&dev->pdev->dev, "create queuepair with %d, %d, %d, %d\n",
349 		cmd->max_send_wr, cmd->max_recv_wr, cmd->max_send_sge,
350 		cmd->max_recv_sge);
351 
352 	ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_CREATE_QP_RESP);
353 	if (ret < 0) {
354 		dev_warn(&dev->pdev->dev,
355 			 "could not create queuepair, error: %d\n", ret);
356 		goto err_pdir;
357 	}
358 
359 	/* max_send_wr/_recv_wr/_send_sge/_recv_sge/_inline_data */
360 	qp->qp_handle = resp->qpn;
361 	qp->port = init_attr->port_num;
362 	qp->ibqp.qp_num = resp->qpn;
363 	spin_lock_irqsave(&dev->qp_tbl_lock, flags);
364 	dev->qp_tbl[qp->qp_handle % dev->dsr->caps.max_qp] = qp;
365 	spin_unlock_irqrestore(&dev->qp_tbl_lock, flags);
366 
367 	return &qp->ibqp;
368 
369 err_pdir:
370 	pvrdma_page_dir_cleanup(dev, &qp->pdir);
371 err_umem:
372 	if (pd->uobject && udata) {
373 		if (qp->rumem)
374 			ib_umem_release(qp->rumem);
375 		if (qp->sumem)
376 			ib_umem_release(qp->sumem);
377 	}
378 err_qp:
379 	kfree(qp);
380 	atomic_dec(&dev->num_qps);
381 
382 	return ERR_PTR(ret);
383 }
384 
385 static void pvrdma_free_qp(struct pvrdma_qp *qp)
386 {
387 	struct pvrdma_dev *dev = to_vdev(qp->ibqp.device);
388 	struct pvrdma_cq *scq;
389 	struct pvrdma_cq *rcq;
390 	unsigned long flags, scq_flags, rcq_flags;
391 
392 	/* In case cq is polling */
393 	get_cqs(qp, &scq, &rcq);
394 	pvrdma_lock_cqs(scq, rcq, &scq_flags, &rcq_flags);
395 
396 	_pvrdma_flush_cqe(qp, scq);
397 	if (scq != rcq)
398 		_pvrdma_flush_cqe(qp, rcq);
399 
400 	spin_lock_irqsave(&dev->qp_tbl_lock, flags);
401 	dev->qp_tbl[qp->qp_handle] = NULL;
402 	spin_unlock_irqrestore(&dev->qp_tbl_lock, flags);
403 
404 	pvrdma_unlock_cqs(scq, rcq, &scq_flags, &rcq_flags);
405 
406 	atomic_dec(&qp->refcnt);
407 	wait_event(qp->wait, !atomic_read(&qp->refcnt));
408 
409 	pvrdma_page_dir_cleanup(dev, &qp->pdir);
410 
411 	kfree(qp);
412 
413 	atomic_dec(&dev->num_qps);
414 }
415 
416 /**
417  * pvrdma_destroy_qp - destroy a queue pair
418  * @qp: the queue pair to destroy
419  *
420  * @return: 0 on success.
421  */
422 int pvrdma_destroy_qp(struct ib_qp *qp)
423 {
424 	struct pvrdma_qp *vqp = to_vqp(qp);
425 	union pvrdma_cmd_req req;
426 	struct pvrdma_cmd_destroy_qp *cmd = &req.destroy_qp;
427 	int ret;
428 
429 	memset(cmd, 0, sizeof(*cmd));
430 	cmd->hdr.cmd = PVRDMA_CMD_DESTROY_QP;
431 	cmd->qp_handle = vqp->qp_handle;
432 
433 	ret = pvrdma_cmd_post(to_vdev(qp->device), &req, NULL, 0);
434 	if (ret < 0)
435 		dev_warn(&to_vdev(qp->device)->pdev->dev,
436 			 "destroy queuepair failed, error: %d\n", ret);
437 
438 	pvrdma_free_qp(vqp);
439 
440 	return 0;
441 }
442 
443 /**
444  * pvrdma_modify_qp - modify queue pair attributes
445  * @ibqp: the queue pair
446  * @attr: the new queue pair's attributes
447  * @attr_mask: attributes mask
448  * @udata: user data
449  *
450  * @returns 0 on success, otherwise returns an errno.
451  */
452 int pvrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
453 		     int attr_mask, struct ib_udata *udata)
454 {
455 	struct pvrdma_dev *dev = to_vdev(ibqp->device);
456 	struct pvrdma_qp *qp = to_vqp(ibqp);
457 	union pvrdma_cmd_req req;
458 	union pvrdma_cmd_resp rsp;
459 	struct pvrdma_cmd_modify_qp *cmd = &req.modify_qp;
460 	int cur_state, next_state;
461 	int ret;
462 
463 	/* Sanity checking. Should need lock here */
464 	mutex_lock(&qp->mutex);
465 	cur_state = (attr_mask & IB_QP_CUR_STATE) ? attr->cur_qp_state :
466 		qp->state;
467 	next_state = (attr_mask & IB_QP_STATE) ? attr->qp_state : cur_state;
468 
469 	if (!ib_modify_qp_is_ok(cur_state, next_state, ibqp->qp_type,
470 				attr_mask, IB_LINK_LAYER_ETHERNET)) {
471 		ret = -EINVAL;
472 		goto out;
473 	}
474 
475 	if (attr_mask & IB_QP_PORT) {
476 		if (attr->port_num == 0 ||
477 		    attr->port_num > ibqp->device->phys_port_cnt) {
478 			ret = -EINVAL;
479 			goto out;
480 		}
481 	}
482 
483 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
484 		if (attr->min_rnr_timer > 31) {
485 			ret = -EINVAL;
486 			goto out;
487 		}
488 	}
489 
490 	if (attr_mask & IB_QP_PKEY_INDEX) {
491 		if (attr->pkey_index >= dev->dsr->caps.max_pkeys) {
492 			ret = -EINVAL;
493 			goto out;
494 		}
495 	}
496 
497 	if (attr_mask & IB_QP_QKEY)
498 		qp->qkey = attr->qkey;
499 
500 	if (cur_state == next_state && cur_state == IB_QPS_RESET) {
501 		ret = 0;
502 		goto out;
503 	}
504 
505 	qp->state = next_state;
506 	memset(cmd, 0, sizeof(*cmd));
507 	cmd->hdr.cmd = PVRDMA_CMD_MODIFY_QP;
508 	cmd->qp_handle = qp->qp_handle;
509 	cmd->attr_mask = ib_qp_attr_mask_to_pvrdma(attr_mask);
510 	cmd->attrs.qp_state = ib_qp_state_to_pvrdma(attr->qp_state);
511 	cmd->attrs.cur_qp_state =
512 		ib_qp_state_to_pvrdma(attr->cur_qp_state);
513 	cmd->attrs.path_mtu = ib_mtu_to_pvrdma(attr->path_mtu);
514 	cmd->attrs.path_mig_state =
515 		ib_mig_state_to_pvrdma(attr->path_mig_state);
516 	cmd->attrs.qkey = attr->qkey;
517 	cmd->attrs.rq_psn = attr->rq_psn;
518 	cmd->attrs.sq_psn = attr->sq_psn;
519 	cmd->attrs.dest_qp_num = attr->dest_qp_num;
520 	cmd->attrs.qp_access_flags =
521 		ib_access_flags_to_pvrdma(attr->qp_access_flags);
522 	cmd->attrs.pkey_index = attr->pkey_index;
523 	cmd->attrs.alt_pkey_index = attr->alt_pkey_index;
524 	cmd->attrs.en_sqd_async_notify = attr->en_sqd_async_notify;
525 	cmd->attrs.sq_draining = attr->sq_draining;
526 	cmd->attrs.max_rd_atomic = attr->max_rd_atomic;
527 	cmd->attrs.max_dest_rd_atomic = attr->max_dest_rd_atomic;
528 	cmd->attrs.min_rnr_timer = attr->min_rnr_timer;
529 	cmd->attrs.port_num = attr->port_num;
530 	cmd->attrs.timeout = attr->timeout;
531 	cmd->attrs.retry_cnt = attr->retry_cnt;
532 	cmd->attrs.rnr_retry = attr->rnr_retry;
533 	cmd->attrs.alt_port_num = attr->alt_port_num;
534 	cmd->attrs.alt_timeout = attr->alt_timeout;
535 	ib_qp_cap_to_pvrdma(&cmd->attrs.cap, &attr->cap);
536 	ib_ah_attr_to_pvrdma(&cmd->attrs.ah_attr, &attr->ah_attr);
537 	ib_ah_attr_to_pvrdma(&cmd->attrs.alt_ah_attr, &attr->alt_ah_attr);
538 
539 	ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_MODIFY_QP_RESP);
540 	if (ret < 0) {
541 		dev_warn(&dev->pdev->dev,
542 			 "could not modify queuepair, error: %d\n", ret);
543 	} else if (rsp.hdr.err > 0) {
544 		dev_warn(&dev->pdev->dev,
545 			 "cannot modify queuepair, error: %d\n", rsp.hdr.err);
546 		ret = -EINVAL;
547 	}
548 
549 	if (ret == 0 && next_state == IB_QPS_RESET)
550 		pvrdma_reset_qp(qp);
551 
552 out:
553 	mutex_unlock(&qp->mutex);
554 
555 	return ret;
556 }
557 
558 static inline void *get_sq_wqe(struct pvrdma_qp *qp, int n)
559 {
560 	return pvrdma_page_dir_get_ptr(&qp->pdir,
561 				       qp->sq.offset + n * qp->sq.wqe_size);
562 }
563 
564 static inline void *get_rq_wqe(struct pvrdma_qp *qp, int n)
565 {
566 	return pvrdma_page_dir_get_ptr(&qp->pdir,
567 				       qp->rq.offset + n * qp->rq.wqe_size);
568 }
569 
570 static int set_reg_seg(struct pvrdma_sq_wqe_hdr *wqe_hdr, struct ib_reg_wr *wr)
571 {
572 	struct pvrdma_user_mr *mr = to_vmr(wr->mr);
573 
574 	wqe_hdr->wr.fast_reg.iova_start = mr->ibmr.iova;
575 	wqe_hdr->wr.fast_reg.pl_pdir_dma = mr->pdir.dir_dma;
576 	wqe_hdr->wr.fast_reg.page_shift = mr->page_shift;
577 	wqe_hdr->wr.fast_reg.page_list_len = mr->npages;
578 	wqe_hdr->wr.fast_reg.length = mr->ibmr.length;
579 	wqe_hdr->wr.fast_reg.access_flags = wr->access;
580 	wqe_hdr->wr.fast_reg.rkey = wr->key;
581 
582 	return pvrdma_page_dir_insert_page_list(&mr->pdir, mr->pages,
583 						mr->npages);
584 }
585 
586 /**
587  * pvrdma_post_send - post send work request entries on a QP
588  * @ibqp: the QP
589  * @wr: work request list to post
590  * @bad_wr: the first bad WR returned
591  *
592  * @return: 0 on success, otherwise errno returned.
593  */
594 int pvrdma_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
595 		     struct ib_send_wr **bad_wr)
596 {
597 	struct pvrdma_qp *qp = to_vqp(ibqp);
598 	struct pvrdma_dev *dev = to_vdev(ibqp->device);
599 	unsigned long flags;
600 	struct pvrdma_sq_wqe_hdr *wqe_hdr;
601 	struct pvrdma_sge *sge;
602 	int i, index;
603 	int nreq;
604 	int ret;
605 
606 	/*
607 	 * In states lower than RTS, we can fail immediately. In other states,
608 	 * just post and let the device figure it out.
609 	 */
610 	if (qp->state < IB_QPS_RTS) {
611 		*bad_wr = wr;
612 		return -EINVAL;
613 	}
614 
615 	spin_lock_irqsave(&qp->sq.lock, flags);
616 
617 	index = pvrdma_idx(&qp->sq.ring->prod_tail, qp->sq.wqe_cnt);
618 	for (nreq = 0; wr; nreq++, wr = wr->next) {
619 		unsigned int tail;
620 
621 		if (unlikely(!pvrdma_idx_ring_has_space(
622 				qp->sq.ring, qp->sq.wqe_cnt, &tail))) {
623 			dev_warn_ratelimited(&dev->pdev->dev,
624 					     "send queue is full\n");
625 			*bad_wr = wr;
626 			ret = -ENOMEM;
627 			goto out;
628 		}
629 
630 		if (unlikely(wr->num_sge > qp->sq.max_sg || wr->num_sge < 0)) {
631 			dev_warn_ratelimited(&dev->pdev->dev,
632 					     "send SGE overflow\n");
633 			*bad_wr = wr;
634 			ret = -EINVAL;
635 			goto out;
636 		}
637 
638 		if (unlikely(wr->opcode < 0)) {
639 			dev_warn_ratelimited(&dev->pdev->dev,
640 					     "invalid send opcode\n");
641 			*bad_wr = wr;
642 			ret = -EINVAL;
643 			goto out;
644 		}
645 
646 		/*
647 		 * Only support UD, RC.
648 		 * Need to check opcode table for thorough checking.
649 		 * opcode		_UD	_UC	_RC
650 		 * _SEND		x	x	x
651 		 * _SEND_WITH_IMM	x	x	x
652 		 * _RDMA_WRITE			x	x
653 		 * _RDMA_WRITE_WITH_IMM		x	x
654 		 * _LOCAL_INV			x	x
655 		 * _SEND_WITH_INV		x	x
656 		 * _RDMA_READ				x
657 		 * _ATOMIC_CMP_AND_SWP			x
658 		 * _ATOMIC_FETCH_AND_ADD		x
659 		 * _MASK_ATOMIC_CMP_AND_SWP		x
660 		 * _MASK_ATOMIC_FETCH_AND_ADD		x
661 		 * _REG_MR				x
662 		 *
663 		 */
664 		if (qp->ibqp.qp_type != IB_QPT_UD &&
665 		    qp->ibqp.qp_type != IB_QPT_RC &&
666 			wr->opcode != IB_WR_SEND) {
667 			dev_warn_ratelimited(&dev->pdev->dev,
668 					     "unsupported queuepair type\n");
669 			*bad_wr = wr;
670 			ret = -EINVAL;
671 			goto out;
672 		} else if (qp->ibqp.qp_type == IB_QPT_UD ||
673 			   qp->ibqp.qp_type == IB_QPT_GSI) {
674 			if (wr->opcode != IB_WR_SEND &&
675 			    wr->opcode != IB_WR_SEND_WITH_IMM) {
676 				dev_warn_ratelimited(&dev->pdev->dev,
677 						     "invalid send opcode\n");
678 				*bad_wr = wr;
679 				ret = -EINVAL;
680 				goto out;
681 			}
682 		}
683 
684 		wqe_hdr = (struct pvrdma_sq_wqe_hdr *)get_sq_wqe(qp, index);
685 		memset(wqe_hdr, 0, sizeof(*wqe_hdr));
686 		wqe_hdr->wr_id = wr->wr_id;
687 		wqe_hdr->num_sge = wr->num_sge;
688 		wqe_hdr->opcode = ib_wr_opcode_to_pvrdma(wr->opcode);
689 		wqe_hdr->send_flags = ib_send_flags_to_pvrdma(wr->send_flags);
690 		if (wr->opcode == IB_WR_SEND_WITH_IMM ||
691 		    wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
692 			wqe_hdr->ex.imm_data = wr->ex.imm_data;
693 
694 		switch (qp->ibqp.qp_type) {
695 		case IB_QPT_GSI:
696 		case IB_QPT_UD:
697 			if (unlikely(!ud_wr(wr)->ah)) {
698 				dev_warn_ratelimited(&dev->pdev->dev,
699 						     "invalid address handle\n");
700 				*bad_wr = wr;
701 				ret = -EINVAL;
702 				goto out;
703 			}
704 
705 			/*
706 			 * Use qkey from qp context if high order bit set,
707 			 * otherwise from work request.
708 			 */
709 			wqe_hdr->wr.ud.remote_qpn = ud_wr(wr)->remote_qpn;
710 			wqe_hdr->wr.ud.remote_qkey =
711 				ud_wr(wr)->remote_qkey & 0x80000000 ?
712 				qp->qkey : ud_wr(wr)->remote_qkey;
713 			wqe_hdr->wr.ud.av = to_vah(ud_wr(wr)->ah)->av;
714 
715 			break;
716 		case IB_QPT_RC:
717 			switch (wr->opcode) {
718 			case IB_WR_RDMA_READ:
719 			case IB_WR_RDMA_WRITE:
720 			case IB_WR_RDMA_WRITE_WITH_IMM:
721 				wqe_hdr->wr.rdma.remote_addr =
722 					rdma_wr(wr)->remote_addr;
723 				wqe_hdr->wr.rdma.rkey = rdma_wr(wr)->rkey;
724 				break;
725 			case IB_WR_LOCAL_INV:
726 			case IB_WR_SEND_WITH_INV:
727 				wqe_hdr->ex.invalidate_rkey =
728 					wr->ex.invalidate_rkey;
729 				break;
730 			case IB_WR_ATOMIC_CMP_AND_SWP:
731 			case IB_WR_ATOMIC_FETCH_AND_ADD:
732 				wqe_hdr->wr.atomic.remote_addr =
733 					atomic_wr(wr)->remote_addr;
734 				wqe_hdr->wr.atomic.rkey = atomic_wr(wr)->rkey;
735 				wqe_hdr->wr.atomic.compare_add =
736 					atomic_wr(wr)->compare_add;
737 				if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP)
738 					wqe_hdr->wr.atomic.swap =
739 						atomic_wr(wr)->swap;
740 				break;
741 			case IB_WR_REG_MR:
742 				ret = set_reg_seg(wqe_hdr, reg_wr(wr));
743 				if (ret < 0) {
744 					dev_warn_ratelimited(&dev->pdev->dev,
745 							     "Failed to set fast register work request\n");
746 					*bad_wr = wr;
747 					goto out;
748 				}
749 				break;
750 			default:
751 				break;
752 			}
753 
754 			break;
755 		default:
756 			dev_warn_ratelimited(&dev->pdev->dev,
757 					     "invalid queuepair type\n");
758 			ret = -EINVAL;
759 			*bad_wr = wr;
760 			goto out;
761 		}
762 
763 		sge = (struct pvrdma_sge *)(wqe_hdr + 1);
764 		for (i = 0; i < wr->num_sge; i++) {
765 			/* Need to check wqe_size 0 or max size */
766 			sge->addr = wr->sg_list[i].addr;
767 			sge->length = wr->sg_list[i].length;
768 			sge->lkey = wr->sg_list[i].lkey;
769 			sge++;
770 		}
771 
772 		/* Make sure wqe is written before index update */
773 		smp_wmb();
774 
775 		index++;
776 		if (unlikely(index >= qp->sq.wqe_cnt))
777 			index = 0;
778 		/* Update shared sq ring */
779 		pvrdma_idx_ring_inc(&qp->sq.ring->prod_tail,
780 				    qp->sq.wqe_cnt);
781 	}
782 
783 	ret = 0;
784 
785 out:
786 	spin_unlock_irqrestore(&qp->sq.lock, flags);
787 
788 	if (!ret)
789 		pvrdma_write_uar_qp(dev, PVRDMA_UAR_QP_SEND | qp->qp_handle);
790 
791 	return ret;
792 }
793 
794 /**
795  * pvrdma_post_receive - post receive work request entries on a QP
796  * @ibqp: the QP
797  * @wr: the work request list to post
798  * @bad_wr: the first bad WR returned
799  *
800  * @return: 0 on success, otherwise errno returned.
801  */
802 int pvrdma_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
803 		     struct ib_recv_wr **bad_wr)
804 {
805 	struct pvrdma_dev *dev = to_vdev(ibqp->device);
806 	unsigned long flags;
807 	struct pvrdma_qp *qp = to_vqp(ibqp);
808 	struct pvrdma_rq_wqe_hdr *wqe_hdr;
809 	struct pvrdma_sge *sge;
810 	int index, nreq;
811 	int ret = 0;
812 	int i;
813 
814 	/*
815 	 * In the RESET state, we can fail immediately. For other states,
816 	 * just post and let the device figure it out.
817 	 */
818 	if (qp->state == IB_QPS_RESET) {
819 		*bad_wr = wr;
820 		return -EINVAL;
821 	}
822 
823 	spin_lock_irqsave(&qp->rq.lock, flags);
824 
825 	index = pvrdma_idx(&qp->rq.ring->prod_tail, qp->rq.wqe_cnt);
826 	for (nreq = 0; wr; nreq++, wr = wr->next) {
827 		unsigned int tail;
828 
829 		if (unlikely(wr->num_sge > qp->rq.max_sg ||
830 			     wr->num_sge < 0)) {
831 			ret = -EINVAL;
832 			*bad_wr = wr;
833 			dev_warn_ratelimited(&dev->pdev->dev,
834 					     "recv SGE overflow\n");
835 			goto out;
836 		}
837 
838 		if (unlikely(!pvrdma_idx_ring_has_space(
839 				qp->rq.ring, qp->rq.wqe_cnt, &tail))) {
840 			ret = -ENOMEM;
841 			*bad_wr = wr;
842 			dev_warn_ratelimited(&dev->pdev->dev,
843 					     "recv queue full\n");
844 			goto out;
845 		}
846 
847 		wqe_hdr = (struct pvrdma_rq_wqe_hdr *)get_rq_wqe(qp, index);
848 		wqe_hdr->wr_id = wr->wr_id;
849 		wqe_hdr->num_sge = wr->num_sge;
850 		wqe_hdr->total_len = 0;
851 
852 		sge = (struct pvrdma_sge *)(wqe_hdr + 1);
853 		for (i = 0; i < wr->num_sge; i++) {
854 			sge->addr = wr->sg_list[i].addr;
855 			sge->length = wr->sg_list[i].length;
856 			sge->lkey = wr->sg_list[i].lkey;
857 			sge++;
858 		}
859 
860 		/* Make sure wqe is written before index update */
861 		smp_wmb();
862 
863 		index++;
864 		if (unlikely(index >= qp->rq.wqe_cnt))
865 			index = 0;
866 		/* Update shared rq ring */
867 		pvrdma_idx_ring_inc(&qp->rq.ring->prod_tail,
868 				    qp->rq.wqe_cnt);
869 	}
870 
871 	spin_unlock_irqrestore(&qp->rq.lock, flags);
872 
873 	pvrdma_write_uar_qp(dev, PVRDMA_UAR_QP_RECV | qp->qp_handle);
874 
875 	return ret;
876 
877 out:
878 	spin_unlock_irqrestore(&qp->rq.lock, flags);
879 
880 	return ret;
881 }
882 
883 /**
884  * pvrdma_query_qp - query a queue pair's attributes
885  * @ibqp: the queue pair to query
886  * @attr: the queue pair's attributes
887  * @attr_mask: attributes mask
888  * @init_attr: initial queue pair attributes
889  *
890  * @returns 0 on success, otherwise returns an errno.
891  */
892 int pvrdma_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
893 		    int attr_mask, struct ib_qp_init_attr *init_attr)
894 {
895 	struct pvrdma_dev *dev = to_vdev(ibqp->device);
896 	struct pvrdma_qp *qp = to_vqp(ibqp);
897 	union pvrdma_cmd_req req;
898 	union pvrdma_cmd_resp rsp;
899 	struct pvrdma_cmd_query_qp *cmd = &req.query_qp;
900 	struct pvrdma_cmd_query_qp_resp *resp = &rsp.query_qp_resp;
901 	int ret = 0;
902 
903 	mutex_lock(&qp->mutex);
904 
905 	if (qp->state == IB_QPS_RESET) {
906 		attr->qp_state = IB_QPS_RESET;
907 		goto out;
908 	}
909 
910 	memset(cmd, 0, sizeof(*cmd));
911 	cmd->hdr.cmd = PVRDMA_CMD_QUERY_QP;
912 	cmd->qp_handle = qp->qp_handle;
913 	cmd->attr_mask = ib_qp_attr_mask_to_pvrdma(attr_mask);
914 
915 	ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_QUERY_QP_RESP);
916 	if (ret < 0) {
917 		dev_warn(&dev->pdev->dev,
918 			 "could not query queuepair, error: %d\n", ret);
919 		goto out;
920 	}
921 
922 	attr->qp_state = pvrdma_qp_state_to_ib(resp->attrs.qp_state);
923 	attr->cur_qp_state =
924 		pvrdma_qp_state_to_ib(resp->attrs.cur_qp_state);
925 	attr->path_mtu = pvrdma_mtu_to_ib(resp->attrs.path_mtu);
926 	attr->path_mig_state =
927 		pvrdma_mig_state_to_ib(resp->attrs.path_mig_state);
928 	attr->qkey = resp->attrs.qkey;
929 	attr->rq_psn = resp->attrs.rq_psn;
930 	attr->sq_psn = resp->attrs.sq_psn;
931 	attr->dest_qp_num = resp->attrs.dest_qp_num;
932 	attr->qp_access_flags =
933 		pvrdma_access_flags_to_ib(resp->attrs.qp_access_flags);
934 	attr->pkey_index = resp->attrs.pkey_index;
935 	attr->alt_pkey_index = resp->attrs.alt_pkey_index;
936 	attr->en_sqd_async_notify = resp->attrs.en_sqd_async_notify;
937 	attr->sq_draining = resp->attrs.sq_draining;
938 	attr->max_rd_atomic = resp->attrs.max_rd_atomic;
939 	attr->max_dest_rd_atomic = resp->attrs.max_dest_rd_atomic;
940 	attr->min_rnr_timer = resp->attrs.min_rnr_timer;
941 	attr->port_num = resp->attrs.port_num;
942 	attr->timeout = resp->attrs.timeout;
943 	attr->retry_cnt = resp->attrs.retry_cnt;
944 	attr->rnr_retry = resp->attrs.rnr_retry;
945 	attr->alt_port_num = resp->attrs.alt_port_num;
946 	attr->alt_timeout = resp->attrs.alt_timeout;
947 	pvrdma_qp_cap_to_ib(&attr->cap, &resp->attrs.cap);
948 	pvrdma_ah_attr_to_ib(&attr->ah_attr, &resp->attrs.ah_attr);
949 	pvrdma_ah_attr_to_ib(&attr->alt_ah_attr, &resp->attrs.alt_ah_attr);
950 
951 	qp->state = attr->qp_state;
952 
953 	ret = 0;
954 
955 out:
956 	attr->cur_qp_state = attr->qp_state;
957 
958 	init_attr->event_handler = qp->ibqp.event_handler;
959 	init_attr->qp_context = qp->ibqp.qp_context;
960 	init_attr->send_cq = qp->ibqp.send_cq;
961 	init_attr->recv_cq = qp->ibqp.recv_cq;
962 	init_attr->srq = qp->ibqp.srq;
963 	init_attr->xrcd = NULL;
964 	init_attr->cap = attr->cap;
965 	init_attr->sq_sig_type = 0;
966 	init_attr->qp_type = qp->ibqp.qp_type;
967 	init_attr->create_flags = 0;
968 	init_attr->port_num = qp->port;
969 
970 	mutex_unlock(&qp->mutex);
971 	return ret;
972 }
973