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