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 void __pvrdma_destroy_qp(struct pvrdma_dev *dev,
56 				struct pvrdma_qp *qp);
57 
58 static inline void get_cqs(struct pvrdma_qp *qp, struct pvrdma_cq **send_cq,
59 			   struct pvrdma_cq **recv_cq)
60 {
61 	*send_cq = to_vcq(qp->ibqp.send_cq);
62 	*recv_cq = to_vcq(qp->ibqp.recv_cq);
63 }
64 
65 static void pvrdma_lock_cqs(struct pvrdma_cq *scq, struct pvrdma_cq *rcq,
66 			    unsigned long *scq_flags,
67 			    unsigned long *rcq_flags)
68 	__acquires(scq->cq_lock) __acquires(rcq->cq_lock)
69 {
70 	if (scq == rcq) {
71 		spin_lock_irqsave(&scq->cq_lock, *scq_flags);
72 		__acquire(rcq->cq_lock);
73 	} else if (scq->cq_handle < rcq->cq_handle) {
74 		spin_lock_irqsave(&scq->cq_lock, *scq_flags);
75 		spin_lock_irqsave_nested(&rcq->cq_lock, *rcq_flags,
76 					 SINGLE_DEPTH_NESTING);
77 	} else {
78 		spin_lock_irqsave(&rcq->cq_lock, *rcq_flags);
79 		spin_lock_irqsave_nested(&scq->cq_lock, *scq_flags,
80 					 SINGLE_DEPTH_NESTING);
81 	}
82 }
83 
84 static void pvrdma_unlock_cqs(struct pvrdma_cq *scq, struct pvrdma_cq *rcq,
85 			      unsigned long *scq_flags,
86 			      unsigned long *rcq_flags)
87 	__releases(scq->cq_lock) __releases(rcq->cq_lock)
88 {
89 	if (scq == rcq) {
90 		__release(rcq->cq_lock);
91 		spin_unlock_irqrestore(&scq->cq_lock, *scq_flags);
92 	} else if (scq->cq_handle < rcq->cq_handle) {
93 		spin_unlock_irqrestore(&rcq->cq_lock, *rcq_flags);
94 		spin_unlock_irqrestore(&scq->cq_lock, *scq_flags);
95 	} else {
96 		spin_unlock_irqrestore(&scq->cq_lock, *scq_flags);
97 		spin_unlock_irqrestore(&rcq->cq_lock, *rcq_flags);
98 	}
99 }
100 
101 static void pvrdma_reset_qp(struct pvrdma_qp *qp)
102 {
103 	struct pvrdma_cq *scq, *rcq;
104 	unsigned long scq_flags, rcq_flags;
105 
106 	/* Clean up cqes */
107 	get_cqs(qp, &scq, &rcq);
108 	pvrdma_lock_cqs(scq, rcq, &scq_flags, &rcq_flags);
109 
110 	_pvrdma_flush_cqe(qp, scq);
111 	if (scq != rcq)
112 		_pvrdma_flush_cqe(qp, rcq);
113 
114 	pvrdma_unlock_cqs(scq, rcq, &scq_flags, &rcq_flags);
115 
116 	/*
117 	 * Reset queuepair. The checks are because usermode queuepairs won't
118 	 * have kernel ringstates.
119 	 */
120 	if (qp->rq.ring) {
121 		atomic_set(&qp->rq.ring->cons_head, 0);
122 		atomic_set(&qp->rq.ring->prod_tail, 0);
123 	}
124 	if (qp->sq.ring) {
125 		atomic_set(&qp->sq.ring->cons_head, 0);
126 		atomic_set(&qp->sq.ring->prod_tail, 0);
127 	}
128 }
129 
130 static int pvrdma_set_rq_size(struct pvrdma_dev *dev,
131 			      struct ib_qp_cap *req_cap,
132 			      struct pvrdma_qp *qp)
133 {
134 	if (req_cap->max_recv_wr > dev->dsr->caps.max_qp_wr ||
135 	    req_cap->max_recv_sge > dev->dsr->caps.max_sge) {
136 		dev_warn(&dev->pdev->dev, "recv queue size invalid\n");
137 		return -EINVAL;
138 	}
139 
140 	qp->rq.wqe_cnt = roundup_pow_of_two(max(1U, req_cap->max_recv_wr));
141 	qp->rq.max_sg = roundup_pow_of_two(max(1U, req_cap->max_recv_sge));
142 
143 	/* Write back */
144 	req_cap->max_recv_wr = qp->rq.wqe_cnt;
145 	req_cap->max_recv_sge = qp->rq.max_sg;
146 
147 	qp->rq.wqe_size = roundup_pow_of_two(sizeof(struct pvrdma_rq_wqe_hdr) +
148 					     sizeof(struct pvrdma_sge) *
149 					     qp->rq.max_sg);
150 	qp->npages_recv = (qp->rq.wqe_cnt * qp->rq.wqe_size + PAGE_SIZE - 1) /
151 			  PAGE_SIZE;
152 
153 	return 0;
154 }
155 
156 static int pvrdma_set_sq_size(struct pvrdma_dev *dev, struct ib_qp_cap *req_cap,
157 			      struct pvrdma_qp *qp)
158 {
159 	if (req_cap->max_send_wr > dev->dsr->caps.max_qp_wr ||
160 	    req_cap->max_send_sge > dev->dsr->caps.max_sge) {
161 		dev_warn(&dev->pdev->dev, "send queue size invalid\n");
162 		return -EINVAL;
163 	}
164 
165 	qp->sq.wqe_cnt = roundup_pow_of_two(max(1U, req_cap->max_send_wr));
166 	qp->sq.max_sg = roundup_pow_of_two(max(1U, req_cap->max_send_sge));
167 
168 	/* Write back */
169 	req_cap->max_send_wr = qp->sq.wqe_cnt;
170 	req_cap->max_send_sge = qp->sq.max_sg;
171 
172 	qp->sq.wqe_size = roundup_pow_of_two(sizeof(struct pvrdma_sq_wqe_hdr) +
173 					     sizeof(struct pvrdma_sge) *
174 					     qp->sq.max_sg);
175 	/* Note: one extra page for the header. */
176 	qp->npages_send = PVRDMA_QP_NUM_HEADER_PAGES +
177 			  (qp->sq.wqe_cnt * qp->sq.wqe_size + PAGE_SIZE - 1) /
178 								PAGE_SIZE;
179 
180 	return 0;
181 }
182 
183 /**
184  * pvrdma_create_qp - create queue pair
185  * @pd: protection domain
186  * @init_attr: queue pair attributes
187  * @udata: user data
188  *
189  * @return: the ib_qp pointer on success, otherwise returns an errno.
190  */
191 struct ib_qp *pvrdma_create_qp(struct ib_pd *pd,
192 			       struct ib_qp_init_attr *init_attr,
193 			       struct ib_udata *udata)
194 {
195 	struct pvrdma_qp *qp = NULL;
196 	struct pvrdma_dev *dev = to_vdev(pd->device);
197 	union pvrdma_cmd_req req;
198 	union pvrdma_cmd_resp rsp;
199 	struct pvrdma_cmd_create_qp *cmd = &req.create_qp;
200 	struct pvrdma_cmd_create_qp_resp *resp = &rsp.create_qp_resp;
201 	struct pvrdma_cmd_create_qp_resp_v2 *resp_v2 = &rsp.create_qp_resp_v2;
202 	struct pvrdma_create_qp ucmd;
203 	struct pvrdma_create_qp_resp qp_resp = {};
204 	unsigned long flags;
205 	int ret;
206 	bool is_srq = !!init_attr->srq;
207 
208 	if (init_attr->create_flags) {
209 		dev_warn(&dev->pdev->dev,
210 			 "invalid create queuepair flags %#x\n",
211 			 init_attr->create_flags);
212 		return ERR_PTR(-EINVAL);
213 	}
214 
215 	if (init_attr->qp_type != IB_QPT_RC &&
216 	    init_attr->qp_type != IB_QPT_UD &&
217 	    init_attr->qp_type != IB_QPT_GSI) {
218 		dev_warn(&dev->pdev->dev, "queuepair type %d not supported\n",
219 			 init_attr->qp_type);
220 		return ERR_PTR(-EOPNOTSUPP);
221 	}
222 
223 	if (is_srq && !dev->dsr->caps.max_srq) {
224 		dev_warn(&dev->pdev->dev,
225 			 "SRQs not supported by device\n");
226 		return ERR_PTR(-EINVAL);
227 	}
228 
229 	if (!atomic_add_unless(&dev->num_qps, 1, dev->dsr->caps.max_qp))
230 		return ERR_PTR(-ENOMEM);
231 
232 	switch (init_attr->qp_type) {
233 	case IB_QPT_GSI:
234 		if (init_attr->port_num == 0 ||
235 		    init_attr->port_num > pd->device->phys_port_cnt ||
236 		    udata) {
237 			dev_warn(&dev->pdev->dev, "invalid queuepair attrs\n");
238 			ret = -EINVAL;
239 			goto err_qp;
240 		}
241 		/* fall through */
242 	case IB_QPT_RC:
243 	case IB_QPT_UD:
244 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
245 		if (!qp) {
246 			ret = -ENOMEM;
247 			goto err_qp;
248 		}
249 
250 		spin_lock_init(&qp->sq.lock);
251 		spin_lock_init(&qp->rq.lock);
252 		mutex_init(&qp->mutex);
253 		refcount_set(&qp->refcnt, 1);
254 		init_completion(&qp->free);
255 
256 		qp->state = IB_QPS_RESET;
257 		qp->is_kernel = !udata;
258 
259 		if (!qp->is_kernel) {
260 			dev_dbg(&dev->pdev->dev,
261 				"create queuepair from user space\n");
262 
263 			if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
264 				ret = -EFAULT;
265 				goto err_qp;
266 			}
267 
268 			/* Userspace supports qpn and qp handles? */
269 			if (dev->dsr_version >= PVRDMA_QPHANDLE_VERSION &&
270 			    udata->outlen < sizeof(qp_resp)) {
271 				dev_warn(&dev->pdev->dev,
272 					 "create queuepair not supported\n");
273 				ret = -EOPNOTSUPP;
274 				goto err_qp;
275 			}
276 
277 			if (!is_srq) {
278 				/* set qp->sq.wqe_cnt, shift, buf_size.. */
279 				qp->rumem =
280 					ib_umem_get(pd->device, ucmd.rbuf_addr,
281 						    ucmd.rbuf_size, 0);
282 				if (IS_ERR(qp->rumem)) {
283 					ret = PTR_ERR(qp->rumem);
284 					goto err_qp;
285 				}
286 				qp->srq = NULL;
287 			} else {
288 				qp->rumem = NULL;
289 				qp->srq = to_vsrq(init_attr->srq);
290 			}
291 
292 			qp->sumem = ib_umem_get(pd->device, ucmd.sbuf_addr,
293 						ucmd.sbuf_size, 0);
294 			if (IS_ERR(qp->sumem)) {
295 				if (!is_srq)
296 					ib_umem_release(qp->rumem);
297 				ret = PTR_ERR(qp->sumem);
298 				goto err_qp;
299 			}
300 
301 			qp->npages_send = ib_umem_page_count(qp->sumem);
302 			if (!is_srq)
303 				qp->npages_recv = ib_umem_page_count(qp->rumem);
304 			else
305 				qp->npages_recv = 0;
306 			qp->npages = qp->npages_send + qp->npages_recv;
307 		} else {
308 			ret = pvrdma_set_sq_size(to_vdev(pd->device),
309 						 &init_attr->cap, qp);
310 			if (ret)
311 				goto err_qp;
312 
313 			ret = pvrdma_set_rq_size(to_vdev(pd->device),
314 						 &init_attr->cap, qp);
315 			if (ret)
316 				goto err_qp;
317 
318 			qp->npages = qp->npages_send + qp->npages_recv;
319 
320 			/* Skip header page. */
321 			qp->sq.offset = PVRDMA_QP_NUM_HEADER_PAGES * PAGE_SIZE;
322 
323 			/* Recv queue pages are after send pages. */
324 			qp->rq.offset = qp->npages_send * PAGE_SIZE;
325 		}
326 
327 		if (qp->npages < 0 || qp->npages > PVRDMA_PAGE_DIR_MAX_PAGES) {
328 			dev_warn(&dev->pdev->dev,
329 				 "overflow pages in queuepair\n");
330 			ret = -EINVAL;
331 			goto err_umem;
332 		}
333 
334 		ret = pvrdma_page_dir_init(dev, &qp->pdir, qp->npages,
335 					   qp->is_kernel);
336 		if (ret) {
337 			dev_warn(&dev->pdev->dev,
338 				 "could not allocate page directory\n");
339 			goto err_umem;
340 		}
341 
342 		if (!qp->is_kernel) {
343 			pvrdma_page_dir_insert_umem(&qp->pdir, qp->sumem, 0);
344 			if (!is_srq)
345 				pvrdma_page_dir_insert_umem(&qp->pdir,
346 							    qp->rumem,
347 							    qp->npages_send);
348 		} else {
349 			/* Ring state is always the first page. */
350 			qp->sq.ring = qp->pdir.pages[0];
351 			qp->rq.ring = is_srq ? NULL : &qp->sq.ring[1];
352 		}
353 		break;
354 	default:
355 		ret = -EINVAL;
356 		goto err_qp;
357 	}
358 
359 	/* Not supported */
360 	init_attr->cap.max_inline_data = 0;
361 
362 	memset(cmd, 0, sizeof(*cmd));
363 	cmd->hdr.cmd = PVRDMA_CMD_CREATE_QP;
364 	cmd->pd_handle = to_vpd(pd)->pd_handle;
365 	cmd->send_cq_handle = to_vcq(init_attr->send_cq)->cq_handle;
366 	cmd->recv_cq_handle = to_vcq(init_attr->recv_cq)->cq_handle;
367 	if (is_srq)
368 		cmd->srq_handle = to_vsrq(init_attr->srq)->srq_handle;
369 	else
370 		cmd->srq_handle = 0;
371 	cmd->max_send_wr = init_attr->cap.max_send_wr;
372 	cmd->max_recv_wr = init_attr->cap.max_recv_wr;
373 	cmd->max_send_sge = init_attr->cap.max_send_sge;
374 	cmd->max_recv_sge = init_attr->cap.max_recv_sge;
375 	cmd->max_inline_data = init_attr->cap.max_inline_data;
376 	cmd->sq_sig_all = (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) ? 1 : 0;
377 	cmd->qp_type = ib_qp_type_to_pvrdma(init_attr->qp_type);
378 	cmd->is_srq = is_srq;
379 	cmd->lkey = 0;
380 	cmd->access_flags = IB_ACCESS_LOCAL_WRITE;
381 	cmd->total_chunks = qp->npages;
382 	cmd->send_chunks = qp->npages_send - PVRDMA_QP_NUM_HEADER_PAGES;
383 	cmd->pdir_dma = qp->pdir.dir_dma;
384 
385 	dev_dbg(&dev->pdev->dev, "create queuepair with %d, %d, %d, %d\n",
386 		cmd->max_send_wr, cmd->max_recv_wr, cmd->max_send_sge,
387 		cmd->max_recv_sge);
388 
389 	ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_CREATE_QP_RESP);
390 	if (ret < 0) {
391 		dev_warn(&dev->pdev->dev,
392 			 "could not create queuepair, error: %d\n", ret);
393 		goto err_pdir;
394 	}
395 
396 	/* max_send_wr/_recv_wr/_send_sge/_recv_sge/_inline_data */
397 	qp->port = init_attr->port_num;
398 
399 	if (dev->dsr_version >= PVRDMA_QPHANDLE_VERSION) {
400 		qp->ibqp.qp_num = resp_v2->qpn;
401 		qp->qp_handle = resp_v2->qp_handle;
402 	} else {
403 		qp->ibqp.qp_num = resp->qpn;
404 		qp->qp_handle = resp->qpn;
405 	}
406 
407 	spin_lock_irqsave(&dev->qp_tbl_lock, flags);
408 	dev->qp_tbl[qp->qp_handle % dev->dsr->caps.max_qp] = qp;
409 	spin_unlock_irqrestore(&dev->qp_tbl_lock, flags);
410 
411 	if (udata) {
412 		qp_resp.qpn = qp->ibqp.qp_num;
413 		qp_resp.qp_handle = qp->qp_handle;
414 
415 		if (ib_copy_to_udata(udata, &qp_resp,
416 				     min(udata->outlen, sizeof(qp_resp)))) {
417 			dev_warn(&dev->pdev->dev,
418 				 "failed to copy back udata\n");
419 			__pvrdma_destroy_qp(dev, qp);
420 			return ERR_PTR(-EINVAL);
421 		}
422 	}
423 
424 	return &qp->ibqp;
425 
426 err_pdir:
427 	pvrdma_page_dir_cleanup(dev, &qp->pdir);
428 err_umem:
429 	ib_umem_release(qp->rumem);
430 	ib_umem_release(qp->sumem);
431 err_qp:
432 	kfree(qp);
433 	atomic_dec(&dev->num_qps);
434 
435 	return ERR_PTR(ret);
436 }
437 
438 static void _pvrdma_free_qp(struct pvrdma_qp *qp)
439 {
440 	unsigned long flags;
441 	struct pvrdma_dev *dev = to_vdev(qp->ibqp.device);
442 
443 	spin_lock_irqsave(&dev->qp_tbl_lock, flags);
444 	dev->qp_tbl[qp->qp_handle] = NULL;
445 	spin_unlock_irqrestore(&dev->qp_tbl_lock, flags);
446 
447 	if (refcount_dec_and_test(&qp->refcnt))
448 		complete(&qp->free);
449 	wait_for_completion(&qp->free);
450 
451 	ib_umem_release(qp->rumem);
452 	ib_umem_release(qp->sumem);
453 
454 	pvrdma_page_dir_cleanup(dev, &qp->pdir);
455 
456 	kfree(qp);
457 
458 	atomic_dec(&dev->num_qps);
459 }
460 
461 static void pvrdma_free_qp(struct pvrdma_qp *qp)
462 {
463 	struct pvrdma_cq *scq;
464 	struct pvrdma_cq *rcq;
465 	unsigned long scq_flags, rcq_flags;
466 
467 	/* In case cq is polling */
468 	get_cqs(qp, &scq, &rcq);
469 	pvrdma_lock_cqs(scq, rcq, &scq_flags, &rcq_flags);
470 
471 	_pvrdma_flush_cqe(qp, scq);
472 	if (scq != rcq)
473 		_pvrdma_flush_cqe(qp, rcq);
474 
475 	/*
476 	 * We're now unlocking the CQs before clearing out the qp handle this
477 	 * should still be safe. We have destroyed the backend QP and flushed
478 	 * the CQEs so there should be no other completions for this QP.
479 	 */
480 	pvrdma_unlock_cqs(scq, rcq, &scq_flags, &rcq_flags);
481 
482 	_pvrdma_free_qp(qp);
483 }
484 
485 static inline void _pvrdma_destroy_qp_work(struct pvrdma_dev *dev,
486 					   u32 qp_handle)
487 {
488 	union pvrdma_cmd_req req;
489 	struct pvrdma_cmd_destroy_qp *cmd = &req.destroy_qp;
490 	int ret;
491 
492 	memset(cmd, 0, sizeof(*cmd));
493 	cmd->hdr.cmd = PVRDMA_CMD_DESTROY_QP;
494 	cmd->qp_handle = qp_handle;
495 
496 	ret = pvrdma_cmd_post(dev, &req, NULL, 0);
497 	if (ret < 0)
498 		dev_warn(&dev->pdev->dev,
499 			 "destroy queuepair failed, error: %d\n", ret);
500 }
501 
502 /**
503  * pvrdma_destroy_qp - destroy a queue pair
504  * @qp: the queue pair to destroy
505  * @udata: user data or null for kernel object
506  *
507  * @return: always 0.
508  */
509 int pvrdma_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
510 {
511 	struct pvrdma_qp *vqp = to_vqp(qp);
512 
513 	_pvrdma_destroy_qp_work(to_vdev(qp->device), vqp->qp_handle);
514 	pvrdma_free_qp(vqp);
515 
516 	return 0;
517 }
518 
519 static void __pvrdma_destroy_qp(struct pvrdma_dev *dev,
520 				struct pvrdma_qp *qp)
521 {
522 	_pvrdma_destroy_qp_work(dev, qp->qp_handle);
523 	_pvrdma_free_qp(qp);
524 }
525 
526 /**
527  * pvrdma_modify_qp - modify queue pair attributes
528  * @ibqp: the queue pair
529  * @attr: the new queue pair's attributes
530  * @attr_mask: attributes mask
531  * @udata: user data
532  *
533  * @returns 0 on success, otherwise returns an errno.
534  */
535 int pvrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
536 		     int attr_mask, struct ib_udata *udata)
537 {
538 	struct pvrdma_dev *dev = to_vdev(ibqp->device);
539 	struct pvrdma_qp *qp = to_vqp(ibqp);
540 	union pvrdma_cmd_req req;
541 	union pvrdma_cmd_resp rsp;
542 	struct pvrdma_cmd_modify_qp *cmd = &req.modify_qp;
543 	enum ib_qp_state cur_state, next_state;
544 	int ret;
545 
546 	/* Sanity checking. Should need lock here */
547 	mutex_lock(&qp->mutex);
548 	cur_state = (attr_mask & IB_QP_CUR_STATE) ? attr->cur_qp_state :
549 		qp->state;
550 	next_state = (attr_mask & IB_QP_STATE) ? attr->qp_state : cur_state;
551 
552 	if (!ib_modify_qp_is_ok(cur_state, next_state, ibqp->qp_type,
553 				attr_mask)) {
554 		ret = -EINVAL;
555 		goto out;
556 	}
557 
558 	if (attr_mask & IB_QP_PORT) {
559 		if (attr->port_num == 0 ||
560 		    attr->port_num > ibqp->device->phys_port_cnt) {
561 			ret = -EINVAL;
562 			goto out;
563 		}
564 	}
565 
566 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
567 		if (attr->min_rnr_timer > 31) {
568 			ret = -EINVAL;
569 			goto out;
570 		}
571 	}
572 
573 	if (attr_mask & IB_QP_PKEY_INDEX) {
574 		if (attr->pkey_index >= dev->dsr->caps.max_pkeys) {
575 			ret = -EINVAL;
576 			goto out;
577 		}
578 	}
579 
580 	if (attr_mask & IB_QP_QKEY)
581 		qp->qkey = attr->qkey;
582 
583 	if (cur_state == next_state && cur_state == IB_QPS_RESET) {
584 		ret = 0;
585 		goto out;
586 	}
587 
588 	qp->state = next_state;
589 	memset(cmd, 0, sizeof(*cmd));
590 	cmd->hdr.cmd = PVRDMA_CMD_MODIFY_QP;
591 	cmd->qp_handle = qp->qp_handle;
592 	cmd->attr_mask = ib_qp_attr_mask_to_pvrdma(attr_mask);
593 	cmd->attrs.qp_state = ib_qp_state_to_pvrdma(attr->qp_state);
594 	cmd->attrs.cur_qp_state =
595 		ib_qp_state_to_pvrdma(attr->cur_qp_state);
596 	cmd->attrs.path_mtu = ib_mtu_to_pvrdma(attr->path_mtu);
597 	cmd->attrs.path_mig_state =
598 		ib_mig_state_to_pvrdma(attr->path_mig_state);
599 	cmd->attrs.qkey = attr->qkey;
600 	cmd->attrs.rq_psn = attr->rq_psn;
601 	cmd->attrs.sq_psn = attr->sq_psn;
602 	cmd->attrs.dest_qp_num = attr->dest_qp_num;
603 	cmd->attrs.qp_access_flags =
604 		ib_access_flags_to_pvrdma(attr->qp_access_flags);
605 	cmd->attrs.pkey_index = attr->pkey_index;
606 	cmd->attrs.alt_pkey_index = attr->alt_pkey_index;
607 	cmd->attrs.en_sqd_async_notify = attr->en_sqd_async_notify;
608 	cmd->attrs.sq_draining = attr->sq_draining;
609 	cmd->attrs.max_rd_atomic = attr->max_rd_atomic;
610 	cmd->attrs.max_dest_rd_atomic = attr->max_dest_rd_atomic;
611 	cmd->attrs.min_rnr_timer = attr->min_rnr_timer;
612 	cmd->attrs.port_num = attr->port_num;
613 	cmd->attrs.timeout = attr->timeout;
614 	cmd->attrs.retry_cnt = attr->retry_cnt;
615 	cmd->attrs.rnr_retry = attr->rnr_retry;
616 	cmd->attrs.alt_port_num = attr->alt_port_num;
617 	cmd->attrs.alt_timeout = attr->alt_timeout;
618 	ib_qp_cap_to_pvrdma(&cmd->attrs.cap, &attr->cap);
619 	rdma_ah_attr_to_pvrdma(&cmd->attrs.ah_attr, &attr->ah_attr);
620 	rdma_ah_attr_to_pvrdma(&cmd->attrs.alt_ah_attr, &attr->alt_ah_attr);
621 
622 	ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_MODIFY_QP_RESP);
623 	if (ret < 0) {
624 		dev_warn(&dev->pdev->dev,
625 			 "could not modify queuepair, error: %d\n", ret);
626 	} else if (rsp.hdr.err > 0) {
627 		dev_warn(&dev->pdev->dev,
628 			 "cannot modify queuepair, error: %d\n", rsp.hdr.err);
629 		ret = -EINVAL;
630 	}
631 
632 	if (ret == 0 && next_state == IB_QPS_RESET)
633 		pvrdma_reset_qp(qp);
634 
635 out:
636 	mutex_unlock(&qp->mutex);
637 
638 	return ret;
639 }
640 
641 static inline void *get_sq_wqe(struct pvrdma_qp *qp, unsigned int n)
642 {
643 	return pvrdma_page_dir_get_ptr(&qp->pdir,
644 				       qp->sq.offset + n * qp->sq.wqe_size);
645 }
646 
647 static inline void *get_rq_wqe(struct pvrdma_qp *qp, unsigned int n)
648 {
649 	return pvrdma_page_dir_get_ptr(&qp->pdir,
650 				       qp->rq.offset + n * qp->rq.wqe_size);
651 }
652 
653 static int set_reg_seg(struct pvrdma_sq_wqe_hdr *wqe_hdr,
654 		       const struct ib_reg_wr *wr)
655 {
656 	struct pvrdma_user_mr *mr = to_vmr(wr->mr);
657 
658 	wqe_hdr->wr.fast_reg.iova_start = mr->ibmr.iova;
659 	wqe_hdr->wr.fast_reg.pl_pdir_dma = mr->pdir.dir_dma;
660 	wqe_hdr->wr.fast_reg.page_shift = mr->page_shift;
661 	wqe_hdr->wr.fast_reg.page_list_len = mr->npages;
662 	wqe_hdr->wr.fast_reg.length = mr->ibmr.length;
663 	wqe_hdr->wr.fast_reg.access_flags = wr->access;
664 	wqe_hdr->wr.fast_reg.rkey = wr->key;
665 
666 	return pvrdma_page_dir_insert_page_list(&mr->pdir, mr->pages,
667 						mr->npages);
668 }
669 
670 /**
671  * pvrdma_post_send - post send work request entries on a QP
672  * @ibqp: the QP
673  * @wr: work request list to post
674  * @bad_wr: the first bad WR returned
675  *
676  * @return: 0 on success, otherwise errno returned.
677  */
678 int pvrdma_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
679 		     const struct ib_send_wr **bad_wr)
680 {
681 	struct pvrdma_qp *qp = to_vqp(ibqp);
682 	struct pvrdma_dev *dev = to_vdev(ibqp->device);
683 	unsigned long flags;
684 	struct pvrdma_sq_wqe_hdr *wqe_hdr;
685 	struct pvrdma_sge *sge;
686 	int i, ret;
687 
688 	/*
689 	 * In states lower than RTS, we can fail immediately. In other states,
690 	 * just post and let the device figure it out.
691 	 */
692 	if (qp->state < IB_QPS_RTS) {
693 		*bad_wr = wr;
694 		return -EINVAL;
695 	}
696 
697 	spin_lock_irqsave(&qp->sq.lock, flags);
698 
699 	while (wr) {
700 		unsigned int tail = 0;
701 
702 		if (unlikely(!pvrdma_idx_ring_has_space(
703 				qp->sq.ring, qp->sq.wqe_cnt, &tail))) {
704 			dev_warn_ratelimited(&dev->pdev->dev,
705 					     "send queue is full\n");
706 			*bad_wr = wr;
707 			ret = -ENOMEM;
708 			goto out;
709 		}
710 
711 		if (unlikely(wr->num_sge > qp->sq.max_sg || wr->num_sge < 0)) {
712 			dev_warn_ratelimited(&dev->pdev->dev,
713 					     "send SGE overflow\n");
714 			*bad_wr = wr;
715 			ret = -EINVAL;
716 			goto out;
717 		}
718 
719 		if (unlikely(wr->opcode < 0)) {
720 			dev_warn_ratelimited(&dev->pdev->dev,
721 					     "invalid send opcode\n");
722 			*bad_wr = wr;
723 			ret = -EINVAL;
724 			goto out;
725 		}
726 
727 		/*
728 		 * Only support UD, RC.
729 		 * Need to check opcode table for thorough checking.
730 		 * opcode		_UD	_UC	_RC
731 		 * _SEND		x	x	x
732 		 * _SEND_WITH_IMM	x	x	x
733 		 * _RDMA_WRITE			x	x
734 		 * _RDMA_WRITE_WITH_IMM		x	x
735 		 * _LOCAL_INV			x	x
736 		 * _SEND_WITH_INV		x	x
737 		 * _RDMA_READ				x
738 		 * _ATOMIC_CMP_AND_SWP			x
739 		 * _ATOMIC_FETCH_AND_ADD		x
740 		 * _MASK_ATOMIC_CMP_AND_SWP		x
741 		 * _MASK_ATOMIC_FETCH_AND_ADD		x
742 		 * _REG_MR				x
743 		 *
744 		 */
745 		if (qp->ibqp.qp_type != IB_QPT_UD &&
746 		    qp->ibqp.qp_type != IB_QPT_RC &&
747 			wr->opcode != IB_WR_SEND) {
748 			dev_warn_ratelimited(&dev->pdev->dev,
749 					     "unsupported queuepair type\n");
750 			*bad_wr = wr;
751 			ret = -EINVAL;
752 			goto out;
753 		} else if (qp->ibqp.qp_type == IB_QPT_UD ||
754 			   qp->ibqp.qp_type == IB_QPT_GSI) {
755 			if (wr->opcode != IB_WR_SEND &&
756 			    wr->opcode != IB_WR_SEND_WITH_IMM) {
757 				dev_warn_ratelimited(&dev->pdev->dev,
758 						     "invalid send opcode\n");
759 				*bad_wr = wr;
760 				ret = -EINVAL;
761 				goto out;
762 			}
763 		}
764 
765 		wqe_hdr = (struct pvrdma_sq_wqe_hdr *)get_sq_wqe(qp, tail);
766 		memset(wqe_hdr, 0, sizeof(*wqe_hdr));
767 		wqe_hdr->wr_id = wr->wr_id;
768 		wqe_hdr->num_sge = wr->num_sge;
769 		wqe_hdr->opcode = ib_wr_opcode_to_pvrdma(wr->opcode);
770 		wqe_hdr->send_flags = ib_send_flags_to_pvrdma(wr->send_flags);
771 		if (wr->opcode == IB_WR_SEND_WITH_IMM ||
772 		    wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
773 			wqe_hdr->ex.imm_data = wr->ex.imm_data;
774 
775 		if (unlikely(wqe_hdr->opcode == PVRDMA_WR_ERROR)) {
776 			*bad_wr = wr;
777 			ret = -EINVAL;
778 			goto out;
779 		}
780 
781 		switch (qp->ibqp.qp_type) {
782 		case IB_QPT_GSI:
783 		case IB_QPT_UD:
784 			if (unlikely(!ud_wr(wr)->ah)) {
785 				dev_warn_ratelimited(&dev->pdev->dev,
786 						     "invalid address handle\n");
787 				*bad_wr = wr;
788 				ret = -EINVAL;
789 				goto out;
790 			}
791 
792 			/*
793 			 * Use qkey from qp context if high order bit set,
794 			 * otherwise from work request.
795 			 */
796 			wqe_hdr->wr.ud.remote_qpn = ud_wr(wr)->remote_qpn;
797 			wqe_hdr->wr.ud.remote_qkey =
798 				ud_wr(wr)->remote_qkey & 0x80000000 ?
799 				qp->qkey : ud_wr(wr)->remote_qkey;
800 			wqe_hdr->wr.ud.av = to_vah(ud_wr(wr)->ah)->av;
801 
802 			break;
803 		case IB_QPT_RC:
804 			switch (wr->opcode) {
805 			case IB_WR_RDMA_READ:
806 			case IB_WR_RDMA_WRITE:
807 			case IB_WR_RDMA_WRITE_WITH_IMM:
808 				wqe_hdr->wr.rdma.remote_addr =
809 					rdma_wr(wr)->remote_addr;
810 				wqe_hdr->wr.rdma.rkey = rdma_wr(wr)->rkey;
811 				break;
812 			case IB_WR_LOCAL_INV:
813 			case IB_WR_SEND_WITH_INV:
814 				wqe_hdr->ex.invalidate_rkey =
815 					wr->ex.invalidate_rkey;
816 				break;
817 			case IB_WR_ATOMIC_CMP_AND_SWP:
818 			case IB_WR_ATOMIC_FETCH_AND_ADD:
819 				wqe_hdr->wr.atomic.remote_addr =
820 					atomic_wr(wr)->remote_addr;
821 				wqe_hdr->wr.atomic.rkey = atomic_wr(wr)->rkey;
822 				wqe_hdr->wr.atomic.compare_add =
823 					atomic_wr(wr)->compare_add;
824 				if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP)
825 					wqe_hdr->wr.atomic.swap =
826 						atomic_wr(wr)->swap;
827 				break;
828 			case IB_WR_REG_MR:
829 				ret = set_reg_seg(wqe_hdr, reg_wr(wr));
830 				if (ret < 0) {
831 					dev_warn_ratelimited(&dev->pdev->dev,
832 							     "Failed to set fast register work request\n");
833 					*bad_wr = wr;
834 					goto out;
835 				}
836 				break;
837 			default:
838 				break;
839 			}
840 
841 			break;
842 		default:
843 			dev_warn_ratelimited(&dev->pdev->dev,
844 					     "invalid queuepair type\n");
845 			ret = -EINVAL;
846 			*bad_wr = wr;
847 			goto out;
848 		}
849 
850 		sge = (struct pvrdma_sge *)(wqe_hdr + 1);
851 		for (i = 0; i < wr->num_sge; i++) {
852 			/* Need to check wqe_size 0 or max size */
853 			sge->addr = wr->sg_list[i].addr;
854 			sge->length = wr->sg_list[i].length;
855 			sge->lkey = wr->sg_list[i].lkey;
856 			sge++;
857 		}
858 
859 		/* Make sure wqe is written before index update */
860 		smp_wmb();
861 
862 		/* Update shared sq ring */
863 		pvrdma_idx_ring_inc(&qp->sq.ring->prod_tail,
864 				    qp->sq.wqe_cnt);
865 
866 		wr = wr->next;
867 	}
868 
869 	ret = 0;
870 
871 out:
872 	spin_unlock_irqrestore(&qp->sq.lock, flags);
873 
874 	if (!ret)
875 		pvrdma_write_uar_qp(dev, PVRDMA_UAR_QP_SEND | qp->qp_handle);
876 
877 	return ret;
878 }
879 
880 /**
881  * pvrdma_post_receive - post receive work request entries on a QP
882  * @ibqp: the QP
883  * @wr: the work request list to post
884  * @bad_wr: the first bad WR returned
885  *
886  * @return: 0 on success, otherwise errno returned.
887  */
888 int pvrdma_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
889 		     const struct ib_recv_wr **bad_wr)
890 {
891 	struct pvrdma_dev *dev = to_vdev(ibqp->device);
892 	unsigned long flags;
893 	struct pvrdma_qp *qp = to_vqp(ibqp);
894 	struct pvrdma_rq_wqe_hdr *wqe_hdr;
895 	struct pvrdma_sge *sge;
896 	int ret = 0;
897 	int i;
898 
899 	/*
900 	 * In the RESET state, we can fail immediately. For other states,
901 	 * just post and let the device figure it out.
902 	 */
903 	if (qp->state == IB_QPS_RESET) {
904 		*bad_wr = wr;
905 		return -EINVAL;
906 	}
907 
908 	if (qp->srq) {
909 		dev_warn(&dev->pdev->dev, "QP associated with SRQ\n");
910 		*bad_wr = wr;
911 		return -EINVAL;
912 	}
913 
914 	spin_lock_irqsave(&qp->rq.lock, flags);
915 
916 	while (wr) {
917 		unsigned int tail = 0;
918 
919 		if (unlikely(wr->num_sge > qp->rq.max_sg ||
920 			     wr->num_sge < 0)) {
921 			ret = -EINVAL;
922 			*bad_wr = wr;
923 			dev_warn_ratelimited(&dev->pdev->dev,
924 					     "recv SGE overflow\n");
925 			goto out;
926 		}
927 
928 		if (unlikely(!pvrdma_idx_ring_has_space(
929 				qp->rq.ring, qp->rq.wqe_cnt, &tail))) {
930 			ret = -ENOMEM;
931 			*bad_wr = wr;
932 			dev_warn_ratelimited(&dev->pdev->dev,
933 					     "recv queue full\n");
934 			goto out;
935 		}
936 
937 		wqe_hdr = (struct pvrdma_rq_wqe_hdr *)get_rq_wqe(qp, tail);
938 		wqe_hdr->wr_id = wr->wr_id;
939 		wqe_hdr->num_sge = wr->num_sge;
940 		wqe_hdr->total_len = 0;
941 
942 		sge = (struct pvrdma_sge *)(wqe_hdr + 1);
943 		for (i = 0; i < wr->num_sge; i++) {
944 			sge->addr = wr->sg_list[i].addr;
945 			sge->length = wr->sg_list[i].length;
946 			sge->lkey = wr->sg_list[i].lkey;
947 			sge++;
948 		}
949 
950 		/* Make sure wqe is written before index update */
951 		smp_wmb();
952 
953 		/* Update shared rq ring */
954 		pvrdma_idx_ring_inc(&qp->rq.ring->prod_tail,
955 				    qp->rq.wqe_cnt);
956 
957 		wr = wr->next;
958 	}
959 
960 	spin_unlock_irqrestore(&qp->rq.lock, flags);
961 
962 	pvrdma_write_uar_qp(dev, PVRDMA_UAR_QP_RECV | qp->qp_handle);
963 
964 	return ret;
965 
966 out:
967 	spin_unlock_irqrestore(&qp->rq.lock, flags);
968 
969 	return ret;
970 }
971 
972 /**
973  * pvrdma_query_qp - query a queue pair's attributes
974  * @ibqp: the queue pair to query
975  * @attr: the queue pair's attributes
976  * @attr_mask: attributes mask
977  * @init_attr: initial queue pair attributes
978  *
979  * @returns 0 on success, otherwise returns an errno.
980  */
981 int pvrdma_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
982 		    int attr_mask, struct ib_qp_init_attr *init_attr)
983 {
984 	struct pvrdma_dev *dev = to_vdev(ibqp->device);
985 	struct pvrdma_qp *qp = to_vqp(ibqp);
986 	union pvrdma_cmd_req req;
987 	union pvrdma_cmd_resp rsp;
988 	struct pvrdma_cmd_query_qp *cmd = &req.query_qp;
989 	struct pvrdma_cmd_query_qp_resp *resp = &rsp.query_qp_resp;
990 	int ret = 0;
991 
992 	mutex_lock(&qp->mutex);
993 
994 	if (qp->state == IB_QPS_RESET) {
995 		attr->qp_state = IB_QPS_RESET;
996 		goto out;
997 	}
998 
999 	memset(cmd, 0, sizeof(*cmd));
1000 	cmd->hdr.cmd = PVRDMA_CMD_QUERY_QP;
1001 	cmd->qp_handle = qp->qp_handle;
1002 	cmd->attr_mask = ib_qp_attr_mask_to_pvrdma(attr_mask);
1003 
1004 	ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_QUERY_QP_RESP);
1005 	if (ret < 0) {
1006 		dev_warn(&dev->pdev->dev,
1007 			 "could not query queuepair, error: %d\n", ret);
1008 		goto out;
1009 	}
1010 
1011 	attr->qp_state = pvrdma_qp_state_to_ib(resp->attrs.qp_state);
1012 	attr->cur_qp_state =
1013 		pvrdma_qp_state_to_ib(resp->attrs.cur_qp_state);
1014 	attr->path_mtu = pvrdma_mtu_to_ib(resp->attrs.path_mtu);
1015 	attr->path_mig_state =
1016 		pvrdma_mig_state_to_ib(resp->attrs.path_mig_state);
1017 	attr->qkey = resp->attrs.qkey;
1018 	attr->rq_psn = resp->attrs.rq_psn;
1019 	attr->sq_psn = resp->attrs.sq_psn;
1020 	attr->dest_qp_num = resp->attrs.dest_qp_num;
1021 	attr->qp_access_flags =
1022 		pvrdma_access_flags_to_ib(resp->attrs.qp_access_flags);
1023 	attr->pkey_index = resp->attrs.pkey_index;
1024 	attr->alt_pkey_index = resp->attrs.alt_pkey_index;
1025 	attr->en_sqd_async_notify = resp->attrs.en_sqd_async_notify;
1026 	attr->sq_draining = resp->attrs.sq_draining;
1027 	attr->max_rd_atomic = resp->attrs.max_rd_atomic;
1028 	attr->max_dest_rd_atomic = resp->attrs.max_dest_rd_atomic;
1029 	attr->min_rnr_timer = resp->attrs.min_rnr_timer;
1030 	attr->port_num = resp->attrs.port_num;
1031 	attr->timeout = resp->attrs.timeout;
1032 	attr->retry_cnt = resp->attrs.retry_cnt;
1033 	attr->rnr_retry = resp->attrs.rnr_retry;
1034 	attr->alt_port_num = resp->attrs.alt_port_num;
1035 	attr->alt_timeout = resp->attrs.alt_timeout;
1036 	pvrdma_qp_cap_to_ib(&attr->cap, &resp->attrs.cap);
1037 	pvrdma_ah_attr_to_rdma(&attr->ah_attr, &resp->attrs.ah_attr);
1038 	pvrdma_ah_attr_to_rdma(&attr->alt_ah_attr, &resp->attrs.alt_ah_attr);
1039 
1040 	qp->state = attr->qp_state;
1041 
1042 	ret = 0;
1043 
1044 out:
1045 	attr->cur_qp_state = attr->qp_state;
1046 
1047 	init_attr->event_handler = qp->ibqp.event_handler;
1048 	init_attr->qp_context = qp->ibqp.qp_context;
1049 	init_attr->send_cq = qp->ibqp.send_cq;
1050 	init_attr->recv_cq = qp->ibqp.recv_cq;
1051 	init_attr->srq = qp->ibqp.srq;
1052 	init_attr->xrcd = NULL;
1053 	init_attr->cap = attr->cap;
1054 	init_attr->sq_sig_type = 0;
1055 	init_attr->qp_type = qp->ibqp.qp_type;
1056 	init_attr->create_flags = 0;
1057 	init_attr->port_num = qp->port;
1058 
1059 	mutex_unlock(&qp->mutex);
1060 	return ret;
1061 }
1062