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 			      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 = PVRDMA_QP_NUM_HEADER_PAGES +
174 			  (qp->sq.wqe_cnt * qp->sq.wqe_size + PAGE_SIZE - 1) /
175 								PAGE_SIZE;
176 
177 	return 0;
178 }
179 
180 /**
181  * pvrdma_create_qp - create queue pair
182  * @pd: protection domain
183  * @init_attr: queue pair attributes
184  * @udata: user data
185  *
186  * @return: the ib_qp pointer on success, otherwise returns an errno.
187  */
188 struct ib_qp *pvrdma_create_qp(struct ib_pd *pd,
189 			       struct ib_qp_init_attr *init_attr,
190 			       struct ib_udata *udata)
191 {
192 	struct pvrdma_qp *qp = NULL;
193 	struct pvrdma_dev *dev = to_vdev(pd->device);
194 	union pvrdma_cmd_req req;
195 	union pvrdma_cmd_resp rsp;
196 	struct pvrdma_cmd_create_qp *cmd = &req.create_qp;
197 	struct pvrdma_cmd_create_qp_resp *resp = &rsp.create_qp_resp;
198 	struct pvrdma_create_qp ucmd;
199 	unsigned long flags;
200 	int ret;
201 	bool is_srq = !!init_attr->srq;
202 
203 	if (init_attr->create_flags) {
204 		dev_warn(&dev->pdev->dev,
205 			 "invalid create queuepair flags %#x\n",
206 			 init_attr->create_flags);
207 		return ERR_PTR(-EINVAL);
208 	}
209 
210 	if (init_attr->qp_type != IB_QPT_RC &&
211 	    init_attr->qp_type != IB_QPT_UD &&
212 	    init_attr->qp_type != IB_QPT_GSI) {
213 		dev_warn(&dev->pdev->dev, "queuepair type %d not supported\n",
214 			 init_attr->qp_type);
215 		return ERR_PTR(-EINVAL);
216 	}
217 
218 	if (is_srq && !dev->dsr->caps.max_srq) {
219 		dev_warn(&dev->pdev->dev,
220 			 "SRQs not supported by device\n");
221 		return ERR_PTR(-EINVAL);
222 	}
223 
224 	if (!atomic_add_unless(&dev->num_qps, 1, dev->dsr->caps.max_qp))
225 		return ERR_PTR(-ENOMEM);
226 
227 	switch (init_attr->qp_type) {
228 	case IB_QPT_GSI:
229 		if (init_attr->port_num == 0 ||
230 		    init_attr->port_num > pd->device->phys_port_cnt ||
231 		    udata) {
232 			dev_warn(&dev->pdev->dev, "invalid queuepair attrs\n");
233 			ret = -EINVAL;
234 			goto err_qp;
235 		}
236 		/* fall through */
237 	case IB_QPT_RC:
238 	case IB_QPT_UD:
239 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
240 		if (!qp) {
241 			ret = -ENOMEM;
242 			goto err_qp;
243 		}
244 
245 		spin_lock_init(&qp->sq.lock);
246 		spin_lock_init(&qp->rq.lock);
247 		mutex_init(&qp->mutex);
248 		refcount_set(&qp->refcnt, 1);
249 		init_completion(&qp->free);
250 
251 		qp->state = IB_QPS_RESET;
252 		qp->is_kernel = !udata;
253 
254 		if (!qp->is_kernel) {
255 			dev_dbg(&dev->pdev->dev,
256 				"create queuepair from user space\n");
257 
258 			if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
259 				ret = -EFAULT;
260 				goto err_qp;
261 			}
262 
263 			if (!is_srq) {
264 				/* set qp->sq.wqe_cnt, shift, buf_size.. */
265 				qp->rumem = ib_umem_get(udata, ucmd.rbuf_addr,
266 							ucmd.rbuf_size, 0, 0);
267 				if (IS_ERR(qp->rumem)) {
268 					ret = PTR_ERR(qp->rumem);
269 					goto err_qp;
270 				}
271 				qp->srq = NULL;
272 			} else {
273 				qp->rumem = NULL;
274 				qp->srq = to_vsrq(init_attr->srq);
275 			}
276 
277 			qp->sumem = ib_umem_get(udata, ucmd.sbuf_addr,
278 						ucmd.sbuf_size, 0, 0);
279 			if (IS_ERR(qp->sumem)) {
280 				if (!is_srq)
281 					ib_umem_release(qp->rumem);
282 				ret = PTR_ERR(qp->sumem);
283 				goto err_qp;
284 			}
285 
286 			qp->npages_send = ib_umem_page_count(qp->sumem);
287 			if (!is_srq)
288 				qp->npages_recv = ib_umem_page_count(qp->rumem);
289 			else
290 				qp->npages_recv = 0;
291 			qp->npages = qp->npages_send + qp->npages_recv;
292 		} else {
293 			ret = pvrdma_set_sq_size(to_vdev(pd->device),
294 						 &init_attr->cap, qp);
295 			if (ret)
296 				goto err_qp;
297 
298 			ret = pvrdma_set_rq_size(to_vdev(pd->device),
299 						 &init_attr->cap, qp);
300 			if (ret)
301 				goto err_qp;
302 
303 			qp->npages = qp->npages_send + qp->npages_recv;
304 
305 			/* Skip header page. */
306 			qp->sq.offset = PVRDMA_QP_NUM_HEADER_PAGES * PAGE_SIZE;
307 
308 			/* Recv queue pages are after send pages. */
309 			qp->rq.offset = qp->npages_send * PAGE_SIZE;
310 		}
311 
312 		if (qp->npages < 0 || qp->npages > PVRDMA_PAGE_DIR_MAX_PAGES) {
313 			dev_warn(&dev->pdev->dev,
314 				 "overflow pages in queuepair\n");
315 			ret = -EINVAL;
316 			goto err_umem;
317 		}
318 
319 		ret = pvrdma_page_dir_init(dev, &qp->pdir, qp->npages,
320 					   qp->is_kernel);
321 		if (ret) {
322 			dev_warn(&dev->pdev->dev,
323 				 "could not allocate page directory\n");
324 			goto err_umem;
325 		}
326 
327 		if (!qp->is_kernel) {
328 			pvrdma_page_dir_insert_umem(&qp->pdir, qp->sumem, 0);
329 			if (!is_srq)
330 				pvrdma_page_dir_insert_umem(&qp->pdir,
331 							    qp->rumem,
332 							    qp->npages_send);
333 		} else {
334 			/* Ring state is always the first page. */
335 			qp->sq.ring = qp->pdir.pages[0];
336 			qp->rq.ring = is_srq ? NULL : &qp->sq.ring[1];
337 		}
338 		break;
339 	default:
340 		ret = -EINVAL;
341 		goto err_qp;
342 	}
343 
344 	/* Not supported */
345 	init_attr->cap.max_inline_data = 0;
346 
347 	memset(cmd, 0, sizeof(*cmd));
348 	cmd->hdr.cmd = PVRDMA_CMD_CREATE_QP;
349 	cmd->pd_handle = to_vpd(pd)->pd_handle;
350 	cmd->send_cq_handle = to_vcq(init_attr->send_cq)->cq_handle;
351 	cmd->recv_cq_handle = to_vcq(init_attr->recv_cq)->cq_handle;
352 	if (is_srq)
353 		cmd->srq_handle = to_vsrq(init_attr->srq)->srq_handle;
354 	else
355 		cmd->srq_handle = 0;
356 	cmd->max_send_wr = init_attr->cap.max_send_wr;
357 	cmd->max_recv_wr = init_attr->cap.max_recv_wr;
358 	cmd->max_send_sge = init_attr->cap.max_send_sge;
359 	cmd->max_recv_sge = init_attr->cap.max_recv_sge;
360 	cmd->max_inline_data = init_attr->cap.max_inline_data;
361 	cmd->sq_sig_all = (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) ? 1 : 0;
362 	cmd->qp_type = ib_qp_type_to_pvrdma(init_attr->qp_type);
363 	cmd->is_srq = is_srq;
364 	cmd->lkey = 0;
365 	cmd->access_flags = IB_ACCESS_LOCAL_WRITE;
366 	cmd->total_chunks = qp->npages;
367 	cmd->send_chunks = qp->npages_send - PVRDMA_QP_NUM_HEADER_PAGES;
368 	cmd->pdir_dma = qp->pdir.dir_dma;
369 
370 	dev_dbg(&dev->pdev->dev, "create queuepair with %d, %d, %d, %d\n",
371 		cmd->max_send_wr, cmd->max_recv_wr, cmd->max_send_sge,
372 		cmd->max_recv_sge);
373 
374 	ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_CREATE_QP_RESP);
375 	if (ret < 0) {
376 		dev_warn(&dev->pdev->dev,
377 			 "could not create queuepair, error: %d\n", ret);
378 		goto err_pdir;
379 	}
380 
381 	/* max_send_wr/_recv_wr/_send_sge/_recv_sge/_inline_data */
382 	qp->qp_handle = resp->qpn;
383 	qp->port = init_attr->port_num;
384 	qp->ibqp.qp_num = resp->qpn;
385 	spin_lock_irqsave(&dev->qp_tbl_lock, flags);
386 	dev->qp_tbl[qp->qp_handle % dev->dsr->caps.max_qp] = qp;
387 	spin_unlock_irqrestore(&dev->qp_tbl_lock, flags);
388 
389 	return &qp->ibqp;
390 
391 err_pdir:
392 	pvrdma_page_dir_cleanup(dev, &qp->pdir);
393 err_umem:
394 	ib_umem_release(qp->rumem);
395 	ib_umem_release(qp->sumem);
396 err_qp:
397 	kfree(qp);
398 	atomic_dec(&dev->num_qps);
399 
400 	return ERR_PTR(ret);
401 }
402 
403 static void pvrdma_free_qp(struct pvrdma_qp *qp)
404 {
405 	struct pvrdma_dev *dev = to_vdev(qp->ibqp.device);
406 	struct pvrdma_cq *scq;
407 	struct pvrdma_cq *rcq;
408 	unsigned long flags, scq_flags, rcq_flags;
409 
410 	/* In case cq is polling */
411 	get_cqs(qp, &scq, &rcq);
412 	pvrdma_lock_cqs(scq, rcq, &scq_flags, &rcq_flags);
413 
414 	_pvrdma_flush_cqe(qp, scq);
415 	if (scq != rcq)
416 		_pvrdma_flush_cqe(qp, rcq);
417 
418 	spin_lock_irqsave(&dev->qp_tbl_lock, flags);
419 	dev->qp_tbl[qp->qp_handle] = NULL;
420 	spin_unlock_irqrestore(&dev->qp_tbl_lock, flags);
421 
422 	pvrdma_unlock_cqs(scq, rcq, &scq_flags, &rcq_flags);
423 
424 	if (refcount_dec_and_test(&qp->refcnt))
425 		complete(&qp->free);
426 	wait_for_completion(&qp->free);
427 
428 	ib_umem_release(qp->rumem);
429 	ib_umem_release(qp->sumem);
430 
431 	pvrdma_page_dir_cleanup(dev, &qp->pdir);
432 
433 	kfree(qp);
434 
435 	atomic_dec(&dev->num_qps);
436 }
437 
438 /**
439  * pvrdma_destroy_qp - destroy a queue pair
440  * @qp: the queue pair to destroy
441  * @udata: user data or null for kernel object
442  *
443  * @return: 0 on success.
444  */
445 int pvrdma_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
446 {
447 	struct pvrdma_qp *vqp = to_vqp(qp);
448 	union pvrdma_cmd_req req;
449 	struct pvrdma_cmd_destroy_qp *cmd = &req.destroy_qp;
450 	int ret;
451 
452 	memset(cmd, 0, sizeof(*cmd));
453 	cmd->hdr.cmd = PVRDMA_CMD_DESTROY_QP;
454 	cmd->qp_handle = vqp->qp_handle;
455 
456 	ret = pvrdma_cmd_post(to_vdev(qp->device), &req, NULL, 0);
457 	if (ret < 0)
458 		dev_warn(&to_vdev(qp->device)->pdev->dev,
459 			 "destroy queuepair failed, error: %d\n", ret);
460 
461 	pvrdma_free_qp(vqp);
462 
463 	return 0;
464 }
465 
466 /**
467  * pvrdma_modify_qp - modify queue pair attributes
468  * @ibqp: the queue pair
469  * @attr: the new queue pair's attributes
470  * @attr_mask: attributes mask
471  * @udata: user data
472  *
473  * @returns 0 on success, otherwise returns an errno.
474  */
475 int pvrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
476 		     int attr_mask, struct ib_udata *udata)
477 {
478 	struct pvrdma_dev *dev = to_vdev(ibqp->device);
479 	struct pvrdma_qp *qp = to_vqp(ibqp);
480 	union pvrdma_cmd_req req;
481 	union pvrdma_cmd_resp rsp;
482 	struct pvrdma_cmd_modify_qp *cmd = &req.modify_qp;
483 	enum ib_qp_state cur_state, next_state;
484 	int ret;
485 
486 	/* Sanity checking. Should need lock here */
487 	mutex_lock(&qp->mutex);
488 	cur_state = (attr_mask & IB_QP_CUR_STATE) ? attr->cur_qp_state :
489 		qp->state;
490 	next_state = (attr_mask & IB_QP_STATE) ? attr->qp_state : cur_state;
491 
492 	if (!ib_modify_qp_is_ok(cur_state, next_state, ibqp->qp_type,
493 				attr_mask)) {
494 		ret = -EINVAL;
495 		goto out;
496 	}
497 
498 	if (attr_mask & IB_QP_PORT) {
499 		if (attr->port_num == 0 ||
500 		    attr->port_num > ibqp->device->phys_port_cnt) {
501 			ret = -EINVAL;
502 			goto out;
503 		}
504 	}
505 
506 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
507 		if (attr->min_rnr_timer > 31) {
508 			ret = -EINVAL;
509 			goto out;
510 		}
511 	}
512 
513 	if (attr_mask & IB_QP_PKEY_INDEX) {
514 		if (attr->pkey_index >= dev->dsr->caps.max_pkeys) {
515 			ret = -EINVAL;
516 			goto out;
517 		}
518 	}
519 
520 	if (attr_mask & IB_QP_QKEY)
521 		qp->qkey = attr->qkey;
522 
523 	if (cur_state == next_state && cur_state == IB_QPS_RESET) {
524 		ret = 0;
525 		goto out;
526 	}
527 
528 	qp->state = next_state;
529 	memset(cmd, 0, sizeof(*cmd));
530 	cmd->hdr.cmd = PVRDMA_CMD_MODIFY_QP;
531 	cmd->qp_handle = qp->qp_handle;
532 	cmd->attr_mask = ib_qp_attr_mask_to_pvrdma(attr_mask);
533 	cmd->attrs.qp_state = ib_qp_state_to_pvrdma(attr->qp_state);
534 	cmd->attrs.cur_qp_state =
535 		ib_qp_state_to_pvrdma(attr->cur_qp_state);
536 	cmd->attrs.path_mtu = ib_mtu_to_pvrdma(attr->path_mtu);
537 	cmd->attrs.path_mig_state =
538 		ib_mig_state_to_pvrdma(attr->path_mig_state);
539 	cmd->attrs.qkey = attr->qkey;
540 	cmd->attrs.rq_psn = attr->rq_psn;
541 	cmd->attrs.sq_psn = attr->sq_psn;
542 	cmd->attrs.dest_qp_num = attr->dest_qp_num;
543 	cmd->attrs.qp_access_flags =
544 		ib_access_flags_to_pvrdma(attr->qp_access_flags);
545 	cmd->attrs.pkey_index = attr->pkey_index;
546 	cmd->attrs.alt_pkey_index = attr->alt_pkey_index;
547 	cmd->attrs.en_sqd_async_notify = attr->en_sqd_async_notify;
548 	cmd->attrs.sq_draining = attr->sq_draining;
549 	cmd->attrs.max_rd_atomic = attr->max_rd_atomic;
550 	cmd->attrs.max_dest_rd_atomic = attr->max_dest_rd_atomic;
551 	cmd->attrs.min_rnr_timer = attr->min_rnr_timer;
552 	cmd->attrs.port_num = attr->port_num;
553 	cmd->attrs.timeout = attr->timeout;
554 	cmd->attrs.retry_cnt = attr->retry_cnt;
555 	cmd->attrs.rnr_retry = attr->rnr_retry;
556 	cmd->attrs.alt_port_num = attr->alt_port_num;
557 	cmd->attrs.alt_timeout = attr->alt_timeout;
558 	ib_qp_cap_to_pvrdma(&cmd->attrs.cap, &attr->cap);
559 	rdma_ah_attr_to_pvrdma(&cmd->attrs.ah_attr, &attr->ah_attr);
560 	rdma_ah_attr_to_pvrdma(&cmd->attrs.alt_ah_attr, &attr->alt_ah_attr);
561 
562 	ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_MODIFY_QP_RESP);
563 	if (ret < 0) {
564 		dev_warn(&dev->pdev->dev,
565 			 "could not modify queuepair, error: %d\n", ret);
566 	} else if (rsp.hdr.err > 0) {
567 		dev_warn(&dev->pdev->dev,
568 			 "cannot modify queuepair, error: %d\n", rsp.hdr.err);
569 		ret = -EINVAL;
570 	}
571 
572 	if (ret == 0 && next_state == IB_QPS_RESET)
573 		pvrdma_reset_qp(qp);
574 
575 out:
576 	mutex_unlock(&qp->mutex);
577 
578 	return ret;
579 }
580 
581 static inline void *get_sq_wqe(struct pvrdma_qp *qp, unsigned int n)
582 {
583 	return pvrdma_page_dir_get_ptr(&qp->pdir,
584 				       qp->sq.offset + n * qp->sq.wqe_size);
585 }
586 
587 static inline void *get_rq_wqe(struct pvrdma_qp *qp, unsigned int n)
588 {
589 	return pvrdma_page_dir_get_ptr(&qp->pdir,
590 				       qp->rq.offset + n * qp->rq.wqe_size);
591 }
592 
593 static int set_reg_seg(struct pvrdma_sq_wqe_hdr *wqe_hdr,
594 		       const struct ib_reg_wr *wr)
595 {
596 	struct pvrdma_user_mr *mr = to_vmr(wr->mr);
597 
598 	wqe_hdr->wr.fast_reg.iova_start = mr->ibmr.iova;
599 	wqe_hdr->wr.fast_reg.pl_pdir_dma = mr->pdir.dir_dma;
600 	wqe_hdr->wr.fast_reg.page_shift = mr->page_shift;
601 	wqe_hdr->wr.fast_reg.page_list_len = mr->npages;
602 	wqe_hdr->wr.fast_reg.length = mr->ibmr.length;
603 	wqe_hdr->wr.fast_reg.access_flags = wr->access;
604 	wqe_hdr->wr.fast_reg.rkey = wr->key;
605 
606 	return pvrdma_page_dir_insert_page_list(&mr->pdir, mr->pages,
607 						mr->npages);
608 }
609 
610 /**
611  * pvrdma_post_send - post send work request entries on a QP
612  * @ibqp: the QP
613  * @wr: work request list to post
614  * @bad_wr: the first bad WR returned
615  *
616  * @return: 0 on success, otherwise errno returned.
617  */
618 int pvrdma_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
619 		     const struct ib_send_wr **bad_wr)
620 {
621 	struct pvrdma_qp *qp = to_vqp(ibqp);
622 	struct pvrdma_dev *dev = to_vdev(ibqp->device);
623 	unsigned long flags;
624 	struct pvrdma_sq_wqe_hdr *wqe_hdr;
625 	struct pvrdma_sge *sge;
626 	int i, ret;
627 
628 	/*
629 	 * In states lower than RTS, we can fail immediately. In other states,
630 	 * just post and let the device figure it out.
631 	 */
632 	if (qp->state < IB_QPS_RTS) {
633 		*bad_wr = wr;
634 		return -EINVAL;
635 	}
636 
637 	spin_lock_irqsave(&qp->sq.lock, flags);
638 
639 	while (wr) {
640 		unsigned int tail = 0;
641 
642 		if (unlikely(!pvrdma_idx_ring_has_space(
643 				qp->sq.ring, qp->sq.wqe_cnt, &tail))) {
644 			dev_warn_ratelimited(&dev->pdev->dev,
645 					     "send queue is full\n");
646 			*bad_wr = wr;
647 			ret = -ENOMEM;
648 			goto out;
649 		}
650 
651 		if (unlikely(wr->num_sge > qp->sq.max_sg || wr->num_sge < 0)) {
652 			dev_warn_ratelimited(&dev->pdev->dev,
653 					     "send SGE overflow\n");
654 			*bad_wr = wr;
655 			ret = -EINVAL;
656 			goto out;
657 		}
658 
659 		if (unlikely(wr->opcode < 0)) {
660 			dev_warn_ratelimited(&dev->pdev->dev,
661 					     "invalid send opcode\n");
662 			*bad_wr = wr;
663 			ret = -EINVAL;
664 			goto out;
665 		}
666 
667 		/*
668 		 * Only support UD, RC.
669 		 * Need to check opcode table for thorough checking.
670 		 * opcode		_UD	_UC	_RC
671 		 * _SEND		x	x	x
672 		 * _SEND_WITH_IMM	x	x	x
673 		 * _RDMA_WRITE			x	x
674 		 * _RDMA_WRITE_WITH_IMM		x	x
675 		 * _LOCAL_INV			x	x
676 		 * _SEND_WITH_INV		x	x
677 		 * _RDMA_READ				x
678 		 * _ATOMIC_CMP_AND_SWP			x
679 		 * _ATOMIC_FETCH_AND_ADD		x
680 		 * _MASK_ATOMIC_CMP_AND_SWP		x
681 		 * _MASK_ATOMIC_FETCH_AND_ADD		x
682 		 * _REG_MR				x
683 		 *
684 		 */
685 		if (qp->ibqp.qp_type != IB_QPT_UD &&
686 		    qp->ibqp.qp_type != IB_QPT_RC &&
687 			wr->opcode != IB_WR_SEND) {
688 			dev_warn_ratelimited(&dev->pdev->dev,
689 					     "unsupported queuepair type\n");
690 			*bad_wr = wr;
691 			ret = -EINVAL;
692 			goto out;
693 		} else if (qp->ibqp.qp_type == IB_QPT_UD ||
694 			   qp->ibqp.qp_type == IB_QPT_GSI) {
695 			if (wr->opcode != IB_WR_SEND &&
696 			    wr->opcode != IB_WR_SEND_WITH_IMM) {
697 				dev_warn_ratelimited(&dev->pdev->dev,
698 						     "invalid send opcode\n");
699 				*bad_wr = wr;
700 				ret = -EINVAL;
701 				goto out;
702 			}
703 		}
704 
705 		wqe_hdr = (struct pvrdma_sq_wqe_hdr *)get_sq_wqe(qp, tail);
706 		memset(wqe_hdr, 0, sizeof(*wqe_hdr));
707 		wqe_hdr->wr_id = wr->wr_id;
708 		wqe_hdr->num_sge = wr->num_sge;
709 		wqe_hdr->opcode = ib_wr_opcode_to_pvrdma(wr->opcode);
710 		wqe_hdr->send_flags = ib_send_flags_to_pvrdma(wr->send_flags);
711 		if (wr->opcode == IB_WR_SEND_WITH_IMM ||
712 		    wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
713 			wqe_hdr->ex.imm_data = wr->ex.imm_data;
714 
715 		if (unlikely(wqe_hdr->opcode == PVRDMA_WR_ERROR)) {
716 			*bad_wr = wr;
717 			ret = -EINVAL;
718 			goto out;
719 		}
720 
721 		switch (qp->ibqp.qp_type) {
722 		case IB_QPT_GSI:
723 		case IB_QPT_UD:
724 			if (unlikely(!ud_wr(wr)->ah)) {
725 				dev_warn_ratelimited(&dev->pdev->dev,
726 						     "invalid address handle\n");
727 				*bad_wr = wr;
728 				ret = -EINVAL;
729 				goto out;
730 			}
731 
732 			/*
733 			 * Use qkey from qp context if high order bit set,
734 			 * otherwise from work request.
735 			 */
736 			wqe_hdr->wr.ud.remote_qpn = ud_wr(wr)->remote_qpn;
737 			wqe_hdr->wr.ud.remote_qkey =
738 				ud_wr(wr)->remote_qkey & 0x80000000 ?
739 				qp->qkey : ud_wr(wr)->remote_qkey;
740 			wqe_hdr->wr.ud.av = to_vah(ud_wr(wr)->ah)->av;
741 
742 			break;
743 		case IB_QPT_RC:
744 			switch (wr->opcode) {
745 			case IB_WR_RDMA_READ:
746 			case IB_WR_RDMA_WRITE:
747 			case IB_WR_RDMA_WRITE_WITH_IMM:
748 				wqe_hdr->wr.rdma.remote_addr =
749 					rdma_wr(wr)->remote_addr;
750 				wqe_hdr->wr.rdma.rkey = rdma_wr(wr)->rkey;
751 				break;
752 			case IB_WR_LOCAL_INV:
753 			case IB_WR_SEND_WITH_INV:
754 				wqe_hdr->ex.invalidate_rkey =
755 					wr->ex.invalidate_rkey;
756 				break;
757 			case IB_WR_ATOMIC_CMP_AND_SWP:
758 			case IB_WR_ATOMIC_FETCH_AND_ADD:
759 				wqe_hdr->wr.atomic.remote_addr =
760 					atomic_wr(wr)->remote_addr;
761 				wqe_hdr->wr.atomic.rkey = atomic_wr(wr)->rkey;
762 				wqe_hdr->wr.atomic.compare_add =
763 					atomic_wr(wr)->compare_add;
764 				if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP)
765 					wqe_hdr->wr.atomic.swap =
766 						atomic_wr(wr)->swap;
767 				break;
768 			case IB_WR_REG_MR:
769 				ret = set_reg_seg(wqe_hdr, reg_wr(wr));
770 				if (ret < 0) {
771 					dev_warn_ratelimited(&dev->pdev->dev,
772 							     "Failed to set fast register work request\n");
773 					*bad_wr = wr;
774 					goto out;
775 				}
776 				break;
777 			default:
778 				break;
779 			}
780 
781 			break;
782 		default:
783 			dev_warn_ratelimited(&dev->pdev->dev,
784 					     "invalid queuepair type\n");
785 			ret = -EINVAL;
786 			*bad_wr = wr;
787 			goto out;
788 		}
789 
790 		sge = (struct pvrdma_sge *)(wqe_hdr + 1);
791 		for (i = 0; i < wr->num_sge; i++) {
792 			/* Need to check wqe_size 0 or max size */
793 			sge->addr = wr->sg_list[i].addr;
794 			sge->length = wr->sg_list[i].length;
795 			sge->lkey = wr->sg_list[i].lkey;
796 			sge++;
797 		}
798 
799 		/* Make sure wqe is written before index update */
800 		smp_wmb();
801 
802 		/* Update shared sq ring */
803 		pvrdma_idx_ring_inc(&qp->sq.ring->prod_tail,
804 				    qp->sq.wqe_cnt);
805 
806 		wr = wr->next;
807 	}
808 
809 	ret = 0;
810 
811 out:
812 	spin_unlock_irqrestore(&qp->sq.lock, flags);
813 
814 	if (!ret)
815 		pvrdma_write_uar_qp(dev, PVRDMA_UAR_QP_SEND | qp->qp_handle);
816 
817 	return ret;
818 }
819 
820 /**
821  * pvrdma_post_receive - post receive work request entries on a QP
822  * @ibqp: the QP
823  * @wr: the work request list to post
824  * @bad_wr: the first bad WR returned
825  *
826  * @return: 0 on success, otherwise errno returned.
827  */
828 int pvrdma_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
829 		     const struct ib_recv_wr **bad_wr)
830 {
831 	struct pvrdma_dev *dev = to_vdev(ibqp->device);
832 	unsigned long flags;
833 	struct pvrdma_qp *qp = to_vqp(ibqp);
834 	struct pvrdma_rq_wqe_hdr *wqe_hdr;
835 	struct pvrdma_sge *sge;
836 	int ret = 0;
837 	int i;
838 
839 	/*
840 	 * In the RESET state, we can fail immediately. For other states,
841 	 * just post and let the device figure it out.
842 	 */
843 	if (qp->state == IB_QPS_RESET) {
844 		*bad_wr = wr;
845 		return -EINVAL;
846 	}
847 
848 	if (qp->srq) {
849 		dev_warn(&dev->pdev->dev, "QP associated with SRQ\n");
850 		*bad_wr = wr;
851 		return -EINVAL;
852 	}
853 
854 	spin_lock_irqsave(&qp->rq.lock, flags);
855 
856 	while (wr) {
857 		unsigned int tail = 0;
858 
859 		if (unlikely(wr->num_sge > qp->rq.max_sg ||
860 			     wr->num_sge < 0)) {
861 			ret = -EINVAL;
862 			*bad_wr = wr;
863 			dev_warn_ratelimited(&dev->pdev->dev,
864 					     "recv SGE overflow\n");
865 			goto out;
866 		}
867 
868 		if (unlikely(!pvrdma_idx_ring_has_space(
869 				qp->rq.ring, qp->rq.wqe_cnt, &tail))) {
870 			ret = -ENOMEM;
871 			*bad_wr = wr;
872 			dev_warn_ratelimited(&dev->pdev->dev,
873 					     "recv queue full\n");
874 			goto out;
875 		}
876 
877 		wqe_hdr = (struct pvrdma_rq_wqe_hdr *)get_rq_wqe(qp, tail);
878 		wqe_hdr->wr_id = wr->wr_id;
879 		wqe_hdr->num_sge = wr->num_sge;
880 		wqe_hdr->total_len = 0;
881 
882 		sge = (struct pvrdma_sge *)(wqe_hdr + 1);
883 		for (i = 0; i < wr->num_sge; i++) {
884 			sge->addr = wr->sg_list[i].addr;
885 			sge->length = wr->sg_list[i].length;
886 			sge->lkey = wr->sg_list[i].lkey;
887 			sge++;
888 		}
889 
890 		/* Make sure wqe is written before index update */
891 		smp_wmb();
892 
893 		/* Update shared rq ring */
894 		pvrdma_idx_ring_inc(&qp->rq.ring->prod_tail,
895 				    qp->rq.wqe_cnt);
896 
897 		wr = wr->next;
898 	}
899 
900 	spin_unlock_irqrestore(&qp->rq.lock, flags);
901 
902 	pvrdma_write_uar_qp(dev, PVRDMA_UAR_QP_RECV | qp->qp_handle);
903 
904 	return ret;
905 
906 out:
907 	spin_unlock_irqrestore(&qp->rq.lock, flags);
908 
909 	return ret;
910 }
911 
912 /**
913  * pvrdma_query_qp - query a queue pair's attributes
914  * @ibqp: the queue pair to query
915  * @attr: the queue pair's attributes
916  * @attr_mask: attributes mask
917  * @init_attr: initial queue pair attributes
918  *
919  * @returns 0 on success, otherwise returns an errno.
920  */
921 int pvrdma_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
922 		    int attr_mask, struct ib_qp_init_attr *init_attr)
923 {
924 	struct pvrdma_dev *dev = to_vdev(ibqp->device);
925 	struct pvrdma_qp *qp = to_vqp(ibqp);
926 	union pvrdma_cmd_req req;
927 	union pvrdma_cmd_resp rsp;
928 	struct pvrdma_cmd_query_qp *cmd = &req.query_qp;
929 	struct pvrdma_cmd_query_qp_resp *resp = &rsp.query_qp_resp;
930 	int ret = 0;
931 
932 	mutex_lock(&qp->mutex);
933 
934 	if (qp->state == IB_QPS_RESET) {
935 		attr->qp_state = IB_QPS_RESET;
936 		goto out;
937 	}
938 
939 	memset(cmd, 0, sizeof(*cmd));
940 	cmd->hdr.cmd = PVRDMA_CMD_QUERY_QP;
941 	cmd->qp_handle = qp->qp_handle;
942 	cmd->attr_mask = ib_qp_attr_mask_to_pvrdma(attr_mask);
943 
944 	ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_QUERY_QP_RESP);
945 	if (ret < 0) {
946 		dev_warn(&dev->pdev->dev,
947 			 "could not query queuepair, error: %d\n", ret);
948 		goto out;
949 	}
950 
951 	attr->qp_state = pvrdma_qp_state_to_ib(resp->attrs.qp_state);
952 	attr->cur_qp_state =
953 		pvrdma_qp_state_to_ib(resp->attrs.cur_qp_state);
954 	attr->path_mtu = pvrdma_mtu_to_ib(resp->attrs.path_mtu);
955 	attr->path_mig_state =
956 		pvrdma_mig_state_to_ib(resp->attrs.path_mig_state);
957 	attr->qkey = resp->attrs.qkey;
958 	attr->rq_psn = resp->attrs.rq_psn;
959 	attr->sq_psn = resp->attrs.sq_psn;
960 	attr->dest_qp_num = resp->attrs.dest_qp_num;
961 	attr->qp_access_flags =
962 		pvrdma_access_flags_to_ib(resp->attrs.qp_access_flags);
963 	attr->pkey_index = resp->attrs.pkey_index;
964 	attr->alt_pkey_index = resp->attrs.alt_pkey_index;
965 	attr->en_sqd_async_notify = resp->attrs.en_sqd_async_notify;
966 	attr->sq_draining = resp->attrs.sq_draining;
967 	attr->max_rd_atomic = resp->attrs.max_rd_atomic;
968 	attr->max_dest_rd_atomic = resp->attrs.max_dest_rd_atomic;
969 	attr->min_rnr_timer = resp->attrs.min_rnr_timer;
970 	attr->port_num = resp->attrs.port_num;
971 	attr->timeout = resp->attrs.timeout;
972 	attr->retry_cnt = resp->attrs.retry_cnt;
973 	attr->rnr_retry = resp->attrs.rnr_retry;
974 	attr->alt_port_num = resp->attrs.alt_port_num;
975 	attr->alt_timeout = resp->attrs.alt_timeout;
976 	pvrdma_qp_cap_to_ib(&attr->cap, &resp->attrs.cap);
977 	pvrdma_ah_attr_to_rdma(&attr->ah_attr, &resp->attrs.ah_attr);
978 	pvrdma_ah_attr_to_rdma(&attr->alt_ah_attr, &resp->attrs.alt_ah_attr);
979 
980 	qp->state = attr->qp_state;
981 
982 	ret = 0;
983 
984 out:
985 	attr->cur_qp_state = attr->qp_state;
986 
987 	init_attr->event_handler = qp->ibqp.event_handler;
988 	init_attr->qp_context = qp->ibqp.qp_context;
989 	init_attr->send_cq = qp->ibqp.send_cq;
990 	init_attr->recv_cq = qp->ibqp.recv_cq;
991 	init_attr->srq = qp->ibqp.srq;
992 	init_attr->xrcd = NULL;
993 	init_attr->cap = attr->cap;
994 	init_attr->sq_sig_type = 0;
995 	init_attr->qp_type = qp->ibqp.qp_type;
996 	init_attr->create_flags = 0;
997 	init_attr->port_num = qp->port;
998 
999 	mutex_unlock(&qp->mutex);
1000 	return ret;
1001 }
1002