xref: /linux/drivers/infiniband/hw/vmw_pvrdma/pvrdma_qp.c (revision ec8a42e7343234802b9054874fe01810880289ce)
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(-EOPNOTSUPP);
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 			dev_warn(&dev->pdev->dev, "invalid queuepair attrs\n");
237 			ret = -EINVAL;
238 			goto err_qp;
239 		}
240 		fallthrough;
241 	case IB_QPT_RC:
242 	case IB_QPT_UD:
243 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
244 		if (!qp) {
245 			ret = -ENOMEM;
246 			goto err_qp;
247 		}
248 
249 		spin_lock_init(&qp->sq.lock);
250 		spin_lock_init(&qp->rq.lock);
251 		mutex_init(&qp->mutex);
252 		refcount_set(&qp->refcnt, 1);
253 		init_completion(&qp->free);
254 
255 		qp->state = IB_QPS_RESET;
256 		qp->is_kernel = !udata;
257 
258 		if (!qp->is_kernel) {
259 			dev_dbg(&dev->pdev->dev,
260 				"create queuepair from user space\n");
261 
262 			if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
263 				ret = -EFAULT;
264 				goto err_qp;
265 			}
266 
267 			/* Userspace supports qpn and qp handles? */
268 			if (dev->dsr_version >= PVRDMA_QPHANDLE_VERSION &&
269 			    udata->outlen < sizeof(qp_resp)) {
270 				dev_warn(&dev->pdev->dev,
271 					 "create queuepair not supported\n");
272 				ret = -EOPNOTSUPP;
273 				goto err_qp;
274 			}
275 
276 			if (!is_srq) {
277 				/* set qp->sq.wqe_cnt, shift, buf_size.. */
278 				qp->rumem =
279 					ib_umem_get(pd->device, ucmd.rbuf_addr,
280 						    ucmd.rbuf_size, 0);
281 				if (IS_ERR(qp->rumem)) {
282 					ret = PTR_ERR(qp->rumem);
283 					goto err_qp;
284 				}
285 				qp->srq = NULL;
286 			} else {
287 				qp->rumem = NULL;
288 				qp->srq = to_vsrq(init_attr->srq);
289 			}
290 
291 			qp->sumem = ib_umem_get(pd->device, ucmd.sbuf_addr,
292 						ucmd.sbuf_size, 0);
293 			if (IS_ERR(qp->sumem)) {
294 				if (!is_srq)
295 					ib_umem_release(qp->rumem);
296 				ret = PTR_ERR(qp->sumem);
297 				goto err_qp;
298 			}
299 
300 			qp->npages_send =
301 				ib_umem_num_dma_blocks(qp->sumem, PAGE_SIZE);
302 			if (!is_srq)
303 				qp->npages_recv = ib_umem_num_dma_blocks(
304 					qp->rumem, PAGE_SIZE);
305 			else
306 				qp->npages_recv = 0;
307 			qp->npages = qp->npages_send + qp->npages_recv;
308 		} else {
309 			ret = pvrdma_set_sq_size(to_vdev(pd->device),
310 						 &init_attr->cap, qp);
311 			if (ret)
312 				goto err_qp;
313 
314 			ret = pvrdma_set_rq_size(to_vdev(pd->device),
315 						 &init_attr->cap, qp);
316 			if (ret)
317 				goto err_qp;
318 
319 			qp->npages = qp->npages_send + qp->npages_recv;
320 
321 			/* Skip header page. */
322 			qp->sq.offset = PVRDMA_QP_NUM_HEADER_PAGES * PAGE_SIZE;
323 
324 			/* Recv queue pages are after send pages. */
325 			qp->rq.offset = qp->npages_send * PAGE_SIZE;
326 		}
327 
328 		if (qp->npages < 0 || qp->npages > PVRDMA_PAGE_DIR_MAX_PAGES) {
329 			dev_warn(&dev->pdev->dev,
330 				 "overflow pages in queuepair\n");
331 			ret = -EINVAL;
332 			goto err_umem;
333 		}
334 
335 		ret = pvrdma_page_dir_init(dev, &qp->pdir, qp->npages,
336 					   qp->is_kernel);
337 		if (ret) {
338 			dev_warn(&dev->pdev->dev,
339 				 "could not allocate page directory\n");
340 			goto err_umem;
341 		}
342 
343 		if (!qp->is_kernel) {
344 			pvrdma_page_dir_insert_umem(&qp->pdir, qp->sumem, 0);
345 			if (!is_srq)
346 				pvrdma_page_dir_insert_umem(&qp->pdir,
347 							    qp->rumem,
348 							    qp->npages_send);
349 		} else {
350 			/* Ring state is always the first page. */
351 			qp->sq.ring = qp->pdir.pages[0];
352 			qp->rq.ring = is_srq ? NULL : &qp->sq.ring[1];
353 		}
354 		break;
355 	default:
356 		ret = -EINVAL;
357 		goto err_qp;
358 	}
359 
360 	/* Not supported */
361 	init_attr->cap.max_inline_data = 0;
362 
363 	memset(cmd, 0, sizeof(*cmd));
364 	cmd->hdr.cmd = PVRDMA_CMD_CREATE_QP;
365 	cmd->pd_handle = to_vpd(pd)->pd_handle;
366 	cmd->send_cq_handle = to_vcq(init_attr->send_cq)->cq_handle;
367 	cmd->recv_cq_handle = to_vcq(init_attr->recv_cq)->cq_handle;
368 	if (is_srq)
369 		cmd->srq_handle = to_vsrq(init_attr->srq)->srq_handle;
370 	else
371 		cmd->srq_handle = 0;
372 	cmd->max_send_wr = init_attr->cap.max_send_wr;
373 	cmd->max_recv_wr = init_attr->cap.max_recv_wr;
374 	cmd->max_send_sge = init_attr->cap.max_send_sge;
375 	cmd->max_recv_sge = init_attr->cap.max_recv_sge;
376 	cmd->max_inline_data = init_attr->cap.max_inline_data;
377 	cmd->sq_sig_all = (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) ? 1 : 0;
378 	cmd->qp_type = ib_qp_type_to_pvrdma(init_attr->qp_type);
379 	cmd->is_srq = is_srq;
380 	cmd->lkey = 0;
381 	cmd->access_flags = IB_ACCESS_LOCAL_WRITE;
382 	cmd->total_chunks = qp->npages;
383 	cmd->send_chunks = qp->npages_send - PVRDMA_QP_NUM_HEADER_PAGES;
384 	cmd->pdir_dma = qp->pdir.dir_dma;
385 
386 	dev_dbg(&dev->pdev->dev, "create queuepair with %d, %d, %d, %d\n",
387 		cmd->max_send_wr, cmd->max_recv_wr, cmd->max_send_sge,
388 		cmd->max_recv_sge);
389 
390 	ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_CREATE_QP_RESP);
391 	if (ret < 0) {
392 		dev_warn(&dev->pdev->dev,
393 			 "could not create queuepair, error: %d\n", ret);
394 		goto err_pdir;
395 	}
396 
397 	/* max_send_wr/_recv_wr/_send_sge/_recv_sge/_inline_data */
398 	qp->port = init_attr->port_num;
399 
400 	if (dev->dsr_version >= PVRDMA_QPHANDLE_VERSION) {
401 		qp->ibqp.qp_num = resp_v2->qpn;
402 		qp->qp_handle = resp_v2->qp_handle;
403 	} else {
404 		qp->ibqp.qp_num = resp->qpn;
405 		qp->qp_handle = resp->qpn;
406 	}
407 
408 	spin_lock_irqsave(&dev->qp_tbl_lock, flags);
409 	dev->qp_tbl[qp->qp_handle % dev->dsr->caps.max_qp] = qp;
410 	spin_unlock_irqrestore(&dev->qp_tbl_lock, flags);
411 
412 	if (udata) {
413 		qp_resp.qpn = qp->ibqp.qp_num;
414 		qp_resp.qp_handle = qp->qp_handle;
415 
416 		if (ib_copy_to_udata(udata, &qp_resp,
417 				     min(udata->outlen, sizeof(qp_resp)))) {
418 			dev_warn(&dev->pdev->dev,
419 				 "failed to copy back udata\n");
420 			__pvrdma_destroy_qp(dev, qp);
421 			return ERR_PTR(-EINVAL);
422 		}
423 	}
424 
425 	return &qp->ibqp;
426 
427 err_pdir:
428 	pvrdma_page_dir_cleanup(dev, &qp->pdir);
429 err_umem:
430 	ib_umem_release(qp->rumem);
431 	ib_umem_release(qp->sumem);
432 err_qp:
433 	kfree(qp);
434 	atomic_dec(&dev->num_qps);
435 
436 	return ERR_PTR(ret);
437 }
438 
439 static void _pvrdma_free_qp(struct pvrdma_qp *qp)
440 {
441 	unsigned long flags;
442 	struct pvrdma_dev *dev = to_vdev(qp->ibqp.device);
443 
444 	spin_lock_irqsave(&dev->qp_tbl_lock, flags);
445 	dev->qp_tbl[qp->qp_handle] = NULL;
446 	spin_unlock_irqrestore(&dev->qp_tbl_lock, flags);
447 
448 	if (refcount_dec_and_test(&qp->refcnt))
449 		complete(&qp->free);
450 	wait_for_completion(&qp->free);
451 
452 	ib_umem_release(qp->rumem);
453 	ib_umem_release(qp->sumem);
454 
455 	pvrdma_page_dir_cleanup(dev, &qp->pdir);
456 
457 	kfree(qp);
458 
459 	atomic_dec(&dev->num_qps);
460 }
461 
462 static void pvrdma_free_qp(struct pvrdma_qp *qp)
463 {
464 	struct pvrdma_cq *scq;
465 	struct pvrdma_cq *rcq;
466 	unsigned long scq_flags, rcq_flags;
467 
468 	/* In case cq is polling */
469 	get_cqs(qp, &scq, &rcq);
470 	pvrdma_lock_cqs(scq, rcq, &scq_flags, &rcq_flags);
471 
472 	_pvrdma_flush_cqe(qp, scq);
473 	if (scq != rcq)
474 		_pvrdma_flush_cqe(qp, rcq);
475 
476 	/*
477 	 * We're now unlocking the CQs before clearing out the qp handle this
478 	 * should still be safe. We have destroyed the backend QP and flushed
479 	 * the CQEs so there should be no other completions for this QP.
480 	 */
481 	pvrdma_unlock_cqs(scq, rcq, &scq_flags, &rcq_flags);
482 
483 	_pvrdma_free_qp(qp);
484 }
485 
486 static inline void _pvrdma_destroy_qp_work(struct pvrdma_dev *dev,
487 					   u32 qp_handle)
488 {
489 	union pvrdma_cmd_req req;
490 	struct pvrdma_cmd_destroy_qp *cmd = &req.destroy_qp;
491 	int ret;
492 
493 	memset(cmd, 0, sizeof(*cmd));
494 	cmd->hdr.cmd = PVRDMA_CMD_DESTROY_QP;
495 	cmd->qp_handle = qp_handle;
496 
497 	ret = pvrdma_cmd_post(dev, &req, NULL, 0);
498 	if (ret < 0)
499 		dev_warn(&dev->pdev->dev,
500 			 "destroy queuepair failed, error: %d\n", ret);
501 }
502 
503 /**
504  * pvrdma_destroy_qp - destroy a queue pair
505  * @qp: the queue pair to destroy
506  * @udata: user data or null for kernel object
507  *
508  * @return: always 0.
509  */
510 int pvrdma_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
511 {
512 	struct pvrdma_qp *vqp = to_vqp(qp);
513 
514 	_pvrdma_destroy_qp_work(to_vdev(qp->device), vqp->qp_handle);
515 	pvrdma_free_qp(vqp);
516 
517 	return 0;
518 }
519 
520 static void __pvrdma_destroy_qp(struct pvrdma_dev *dev,
521 				struct pvrdma_qp *qp)
522 {
523 	_pvrdma_destroy_qp_work(dev, qp->qp_handle);
524 	_pvrdma_free_qp(qp);
525 }
526 
527 /**
528  * pvrdma_modify_qp - modify queue pair attributes
529  * @ibqp: the queue pair
530  * @attr: the new queue pair's attributes
531  * @attr_mask: attributes mask
532  * @udata: user data
533  *
534  * @returns 0 on success, otherwise returns an errno.
535  */
536 int pvrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
537 		     int attr_mask, struct ib_udata *udata)
538 {
539 	struct pvrdma_dev *dev = to_vdev(ibqp->device);
540 	struct pvrdma_qp *qp = to_vqp(ibqp);
541 	union pvrdma_cmd_req req;
542 	union pvrdma_cmd_resp rsp;
543 	struct pvrdma_cmd_modify_qp *cmd = &req.modify_qp;
544 	enum ib_qp_state cur_state, next_state;
545 	int ret;
546 
547 	if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
548 		return -EOPNOTSUPP;
549 
550 	/* Sanity checking. Should need lock here */
551 	mutex_lock(&qp->mutex);
552 	cur_state = (attr_mask & IB_QP_CUR_STATE) ? attr->cur_qp_state :
553 		qp->state;
554 	next_state = (attr_mask & IB_QP_STATE) ? attr->qp_state : cur_state;
555 
556 	if (!ib_modify_qp_is_ok(cur_state, next_state, ibqp->qp_type,
557 				attr_mask)) {
558 		ret = -EINVAL;
559 		goto out;
560 	}
561 
562 	if (attr_mask & IB_QP_PORT) {
563 		if (attr->port_num == 0 ||
564 		    attr->port_num > ibqp->device->phys_port_cnt) {
565 			ret = -EINVAL;
566 			goto out;
567 		}
568 	}
569 
570 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
571 		if (attr->min_rnr_timer > 31) {
572 			ret = -EINVAL;
573 			goto out;
574 		}
575 	}
576 
577 	if (attr_mask & IB_QP_PKEY_INDEX) {
578 		if (attr->pkey_index >= dev->dsr->caps.max_pkeys) {
579 			ret = -EINVAL;
580 			goto out;
581 		}
582 	}
583 
584 	if (attr_mask & IB_QP_QKEY)
585 		qp->qkey = attr->qkey;
586 
587 	if (cur_state == next_state && cur_state == IB_QPS_RESET) {
588 		ret = 0;
589 		goto out;
590 	}
591 
592 	qp->state = next_state;
593 	memset(cmd, 0, sizeof(*cmd));
594 	cmd->hdr.cmd = PVRDMA_CMD_MODIFY_QP;
595 	cmd->qp_handle = qp->qp_handle;
596 	cmd->attr_mask = ib_qp_attr_mask_to_pvrdma(attr_mask);
597 	cmd->attrs.qp_state = ib_qp_state_to_pvrdma(attr->qp_state);
598 	cmd->attrs.cur_qp_state =
599 		ib_qp_state_to_pvrdma(attr->cur_qp_state);
600 	cmd->attrs.path_mtu = ib_mtu_to_pvrdma(attr->path_mtu);
601 	cmd->attrs.path_mig_state =
602 		ib_mig_state_to_pvrdma(attr->path_mig_state);
603 	cmd->attrs.qkey = attr->qkey;
604 	cmd->attrs.rq_psn = attr->rq_psn;
605 	cmd->attrs.sq_psn = attr->sq_psn;
606 	cmd->attrs.dest_qp_num = attr->dest_qp_num;
607 	cmd->attrs.qp_access_flags =
608 		ib_access_flags_to_pvrdma(attr->qp_access_flags);
609 	cmd->attrs.pkey_index = attr->pkey_index;
610 	cmd->attrs.alt_pkey_index = attr->alt_pkey_index;
611 	cmd->attrs.en_sqd_async_notify = attr->en_sqd_async_notify;
612 	cmd->attrs.sq_draining = attr->sq_draining;
613 	cmd->attrs.max_rd_atomic = attr->max_rd_atomic;
614 	cmd->attrs.max_dest_rd_atomic = attr->max_dest_rd_atomic;
615 	cmd->attrs.min_rnr_timer = attr->min_rnr_timer;
616 	cmd->attrs.port_num = attr->port_num;
617 	cmd->attrs.timeout = attr->timeout;
618 	cmd->attrs.retry_cnt = attr->retry_cnt;
619 	cmd->attrs.rnr_retry = attr->rnr_retry;
620 	cmd->attrs.alt_port_num = attr->alt_port_num;
621 	cmd->attrs.alt_timeout = attr->alt_timeout;
622 	ib_qp_cap_to_pvrdma(&cmd->attrs.cap, &attr->cap);
623 	rdma_ah_attr_to_pvrdma(&cmd->attrs.ah_attr, &attr->ah_attr);
624 	rdma_ah_attr_to_pvrdma(&cmd->attrs.alt_ah_attr, &attr->alt_ah_attr);
625 
626 	ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_MODIFY_QP_RESP);
627 	if (ret < 0) {
628 		dev_warn(&dev->pdev->dev,
629 			 "could not modify queuepair, error: %d\n", ret);
630 	} else if (rsp.hdr.err > 0) {
631 		dev_warn(&dev->pdev->dev,
632 			 "cannot modify queuepair, error: %d\n", rsp.hdr.err);
633 		ret = -EINVAL;
634 	}
635 
636 	if (ret == 0 && next_state == IB_QPS_RESET)
637 		pvrdma_reset_qp(qp);
638 
639 out:
640 	mutex_unlock(&qp->mutex);
641 
642 	return ret;
643 }
644 
645 static inline void *get_sq_wqe(struct pvrdma_qp *qp, unsigned int n)
646 {
647 	return pvrdma_page_dir_get_ptr(&qp->pdir,
648 				       qp->sq.offset + n * qp->sq.wqe_size);
649 }
650 
651 static inline void *get_rq_wqe(struct pvrdma_qp *qp, unsigned int n)
652 {
653 	return pvrdma_page_dir_get_ptr(&qp->pdir,
654 				       qp->rq.offset + n * qp->rq.wqe_size);
655 }
656 
657 static int set_reg_seg(struct pvrdma_sq_wqe_hdr *wqe_hdr,
658 		       const struct ib_reg_wr *wr)
659 {
660 	struct pvrdma_user_mr *mr = to_vmr(wr->mr);
661 
662 	wqe_hdr->wr.fast_reg.iova_start = mr->ibmr.iova;
663 	wqe_hdr->wr.fast_reg.pl_pdir_dma = mr->pdir.dir_dma;
664 	wqe_hdr->wr.fast_reg.page_shift = mr->page_shift;
665 	wqe_hdr->wr.fast_reg.page_list_len = mr->npages;
666 	wqe_hdr->wr.fast_reg.length = mr->ibmr.length;
667 	wqe_hdr->wr.fast_reg.access_flags = wr->access;
668 	wqe_hdr->wr.fast_reg.rkey = wr->key;
669 
670 	return pvrdma_page_dir_insert_page_list(&mr->pdir, mr->pages,
671 						mr->npages);
672 }
673 
674 /**
675  * pvrdma_post_send - post send work request entries on a QP
676  * @ibqp: the QP
677  * @wr: work request list to post
678  * @bad_wr: the first bad WR returned
679  *
680  * @return: 0 on success, otherwise errno returned.
681  */
682 int pvrdma_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
683 		     const struct ib_send_wr **bad_wr)
684 {
685 	struct pvrdma_qp *qp = to_vqp(ibqp);
686 	struct pvrdma_dev *dev = to_vdev(ibqp->device);
687 	unsigned long flags;
688 	struct pvrdma_sq_wqe_hdr *wqe_hdr;
689 	struct pvrdma_sge *sge;
690 	int i, ret;
691 
692 	/*
693 	 * In states lower than RTS, we can fail immediately. In other states,
694 	 * just post and let the device figure it out.
695 	 */
696 	if (qp->state < IB_QPS_RTS) {
697 		*bad_wr = wr;
698 		return -EINVAL;
699 	}
700 
701 	spin_lock_irqsave(&qp->sq.lock, flags);
702 
703 	while (wr) {
704 		unsigned int tail = 0;
705 
706 		if (unlikely(!pvrdma_idx_ring_has_space(
707 				qp->sq.ring, qp->sq.wqe_cnt, &tail))) {
708 			dev_warn_ratelimited(&dev->pdev->dev,
709 					     "send queue is full\n");
710 			*bad_wr = wr;
711 			ret = -ENOMEM;
712 			goto out;
713 		}
714 
715 		if (unlikely(wr->num_sge > qp->sq.max_sg || wr->num_sge < 0)) {
716 			dev_warn_ratelimited(&dev->pdev->dev,
717 					     "send SGE overflow\n");
718 			*bad_wr = wr;
719 			ret = -EINVAL;
720 			goto out;
721 		}
722 
723 		if (unlikely(wr->opcode < 0)) {
724 			dev_warn_ratelimited(&dev->pdev->dev,
725 					     "invalid send opcode\n");
726 			*bad_wr = wr;
727 			ret = -EINVAL;
728 			goto out;
729 		}
730 
731 		/*
732 		 * Only support UD, RC.
733 		 * Need to check opcode table for thorough checking.
734 		 * opcode		_UD	_UC	_RC
735 		 * _SEND		x	x	x
736 		 * _SEND_WITH_IMM	x	x	x
737 		 * _RDMA_WRITE			x	x
738 		 * _RDMA_WRITE_WITH_IMM		x	x
739 		 * _LOCAL_INV			x	x
740 		 * _SEND_WITH_INV		x	x
741 		 * _RDMA_READ				x
742 		 * _ATOMIC_CMP_AND_SWP			x
743 		 * _ATOMIC_FETCH_AND_ADD		x
744 		 * _MASK_ATOMIC_CMP_AND_SWP		x
745 		 * _MASK_ATOMIC_FETCH_AND_ADD		x
746 		 * _REG_MR				x
747 		 *
748 		 */
749 		if (qp->ibqp.qp_type != IB_QPT_UD &&
750 		    qp->ibqp.qp_type != IB_QPT_RC &&
751 			wr->opcode != IB_WR_SEND) {
752 			dev_warn_ratelimited(&dev->pdev->dev,
753 					     "unsupported queuepair type\n");
754 			*bad_wr = wr;
755 			ret = -EINVAL;
756 			goto out;
757 		} else if (qp->ibqp.qp_type == IB_QPT_UD ||
758 			   qp->ibqp.qp_type == IB_QPT_GSI) {
759 			if (wr->opcode != IB_WR_SEND &&
760 			    wr->opcode != IB_WR_SEND_WITH_IMM) {
761 				dev_warn_ratelimited(&dev->pdev->dev,
762 						     "invalid send opcode\n");
763 				*bad_wr = wr;
764 				ret = -EINVAL;
765 				goto out;
766 			}
767 		}
768 
769 		wqe_hdr = (struct pvrdma_sq_wqe_hdr *)get_sq_wqe(qp, tail);
770 		memset(wqe_hdr, 0, sizeof(*wqe_hdr));
771 		wqe_hdr->wr_id = wr->wr_id;
772 		wqe_hdr->num_sge = wr->num_sge;
773 		wqe_hdr->opcode = ib_wr_opcode_to_pvrdma(wr->opcode);
774 		wqe_hdr->send_flags = ib_send_flags_to_pvrdma(wr->send_flags);
775 		if (wr->opcode == IB_WR_SEND_WITH_IMM ||
776 		    wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
777 			wqe_hdr->ex.imm_data = wr->ex.imm_data;
778 
779 		if (unlikely(wqe_hdr->opcode == PVRDMA_WR_ERROR)) {
780 			*bad_wr = wr;
781 			ret = -EINVAL;
782 			goto out;
783 		}
784 
785 		switch (qp->ibqp.qp_type) {
786 		case IB_QPT_GSI:
787 		case IB_QPT_UD:
788 			if (unlikely(!ud_wr(wr)->ah)) {
789 				dev_warn_ratelimited(&dev->pdev->dev,
790 						     "invalid address handle\n");
791 				*bad_wr = wr;
792 				ret = -EINVAL;
793 				goto out;
794 			}
795 
796 			/*
797 			 * Use qkey from qp context if high order bit set,
798 			 * otherwise from work request.
799 			 */
800 			wqe_hdr->wr.ud.remote_qpn = ud_wr(wr)->remote_qpn;
801 			wqe_hdr->wr.ud.remote_qkey =
802 				ud_wr(wr)->remote_qkey & 0x80000000 ?
803 				qp->qkey : ud_wr(wr)->remote_qkey;
804 			wqe_hdr->wr.ud.av = to_vah(ud_wr(wr)->ah)->av;
805 
806 			break;
807 		case IB_QPT_RC:
808 			switch (wr->opcode) {
809 			case IB_WR_RDMA_READ:
810 			case IB_WR_RDMA_WRITE:
811 			case IB_WR_RDMA_WRITE_WITH_IMM:
812 				wqe_hdr->wr.rdma.remote_addr =
813 					rdma_wr(wr)->remote_addr;
814 				wqe_hdr->wr.rdma.rkey = rdma_wr(wr)->rkey;
815 				break;
816 			case IB_WR_LOCAL_INV:
817 			case IB_WR_SEND_WITH_INV:
818 				wqe_hdr->ex.invalidate_rkey =
819 					wr->ex.invalidate_rkey;
820 				break;
821 			case IB_WR_ATOMIC_CMP_AND_SWP:
822 			case IB_WR_ATOMIC_FETCH_AND_ADD:
823 				wqe_hdr->wr.atomic.remote_addr =
824 					atomic_wr(wr)->remote_addr;
825 				wqe_hdr->wr.atomic.rkey = atomic_wr(wr)->rkey;
826 				wqe_hdr->wr.atomic.compare_add =
827 					atomic_wr(wr)->compare_add;
828 				if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP)
829 					wqe_hdr->wr.atomic.swap =
830 						atomic_wr(wr)->swap;
831 				break;
832 			case IB_WR_REG_MR:
833 				ret = set_reg_seg(wqe_hdr, reg_wr(wr));
834 				if (ret < 0) {
835 					dev_warn_ratelimited(&dev->pdev->dev,
836 							     "Failed to set fast register work request\n");
837 					*bad_wr = wr;
838 					goto out;
839 				}
840 				break;
841 			default:
842 				break;
843 			}
844 
845 			break;
846 		default:
847 			dev_warn_ratelimited(&dev->pdev->dev,
848 					     "invalid queuepair type\n");
849 			ret = -EINVAL;
850 			*bad_wr = wr;
851 			goto out;
852 		}
853 
854 		sge = (struct pvrdma_sge *)(wqe_hdr + 1);
855 		for (i = 0; i < wr->num_sge; i++) {
856 			/* Need to check wqe_size 0 or max size */
857 			sge->addr = wr->sg_list[i].addr;
858 			sge->length = wr->sg_list[i].length;
859 			sge->lkey = wr->sg_list[i].lkey;
860 			sge++;
861 		}
862 
863 		/* Make sure wqe is written before index update */
864 		smp_wmb();
865 
866 		/* Update shared sq ring */
867 		pvrdma_idx_ring_inc(&qp->sq.ring->prod_tail,
868 				    qp->sq.wqe_cnt);
869 
870 		wr = wr->next;
871 	}
872 
873 	ret = 0;
874 
875 out:
876 	spin_unlock_irqrestore(&qp->sq.lock, flags);
877 
878 	if (!ret)
879 		pvrdma_write_uar_qp(dev, PVRDMA_UAR_QP_SEND | qp->qp_handle);
880 
881 	return ret;
882 }
883 
884 /**
885  * pvrdma_post_receive - post receive work request entries on a QP
886  * @ibqp: the QP
887  * @wr: the work request list to post
888  * @bad_wr: the first bad WR returned
889  *
890  * @return: 0 on success, otherwise errno returned.
891  */
892 int pvrdma_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
893 		     const struct ib_recv_wr **bad_wr)
894 {
895 	struct pvrdma_dev *dev = to_vdev(ibqp->device);
896 	unsigned long flags;
897 	struct pvrdma_qp *qp = to_vqp(ibqp);
898 	struct pvrdma_rq_wqe_hdr *wqe_hdr;
899 	struct pvrdma_sge *sge;
900 	int ret = 0;
901 	int i;
902 
903 	/*
904 	 * In the RESET state, we can fail immediately. For other states,
905 	 * just post and let the device figure it out.
906 	 */
907 	if (qp->state == IB_QPS_RESET) {
908 		*bad_wr = wr;
909 		return -EINVAL;
910 	}
911 
912 	if (qp->srq) {
913 		dev_warn(&dev->pdev->dev, "QP associated with SRQ\n");
914 		*bad_wr = wr;
915 		return -EINVAL;
916 	}
917 
918 	spin_lock_irqsave(&qp->rq.lock, flags);
919 
920 	while (wr) {
921 		unsigned int tail = 0;
922 
923 		if (unlikely(wr->num_sge > qp->rq.max_sg ||
924 			     wr->num_sge < 0)) {
925 			ret = -EINVAL;
926 			*bad_wr = wr;
927 			dev_warn_ratelimited(&dev->pdev->dev,
928 					     "recv SGE overflow\n");
929 			goto out;
930 		}
931 
932 		if (unlikely(!pvrdma_idx_ring_has_space(
933 				qp->rq.ring, qp->rq.wqe_cnt, &tail))) {
934 			ret = -ENOMEM;
935 			*bad_wr = wr;
936 			dev_warn_ratelimited(&dev->pdev->dev,
937 					     "recv queue full\n");
938 			goto out;
939 		}
940 
941 		wqe_hdr = (struct pvrdma_rq_wqe_hdr *)get_rq_wqe(qp, tail);
942 		wqe_hdr->wr_id = wr->wr_id;
943 		wqe_hdr->num_sge = wr->num_sge;
944 		wqe_hdr->total_len = 0;
945 
946 		sge = (struct pvrdma_sge *)(wqe_hdr + 1);
947 		for (i = 0; i < wr->num_sge; i++) {
948 			sge->addr = wr->sg_list[i].addr;
949 			sge->length = wr->sg_list[i].length;
950 			sge->lkey = wr->sg_list[i].lkey;
951 			sge++;
952 		}
953 
954 		/* Make sure wqe is written before index update */
955 		smp_wmb();
956 
957 		/* Update shared rq ring */
958 		pvrdma_idx_ring_inc(&qp->rq.ring->prod_tail,
959 				    qp->rq.wqe_cnt);
960 
961 		wr = wr->next;
962 	}
963 
964 	spin_unlock_irqrestore(&qp->rq.lock, flags);
965 
966 	pvrdma_write_uar_qp(dev, PVRDMA_UAR_QP_RECV | qp->qp_handle);
967 
968 	return ret;
969 
970 out:
971 	spin_unlock_irqrestore(&qp->rq.lock, flags);
972 
973 	return ret;
974 }
975 
976 /**
977  * pvrdma_query_qp - query a queue pair's attributes
978  * @ibqp: the queue pair to query
979  * @attr: the queue pair's attributes
980  * @attr_mask: attributes mask
981  * @init_attr: initial queue pair attributes
982  *
983  * @returns 0 on success, otherwise returns an errno.
984  */
985 int pvrdma_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
986 		    int attr_mask, struct ib_qp_init_attr *init_attr)
987 {
988 	struct pvrdma_dev *dev = to_vdev(ibqp->device);
989 	struct pvrdma_qp *qp = to_vqp(ibqp);
990 	union pvrdma_cmd_req req;
991 	union pvrdma_cmd_resp rsp;
992 	struct pvrdma_cmd_query_qp *cmd = &req.query_qp;
993 	struct pvrdma_cmd_query_qp_resp *resp = &rsp.query_qp_resp;
994 	int ret = 0;
995 
996 	mutex_lock(&qp->mutex);
997 
998 	if (qp->state == IB_QPS_RESET) {
999 		attr->qp_state = IB_QPS_RESET;
1000 		goto out;
1001 	}
1002 
1003 	memset(cmd, 0, sizeof(*cmd));
1004 	cmd->hdr.cmd = PVRDMA_CMD_QUERY_QP;
1005 	cmd->qp_handle = qp->qp_handle;
1006 	cmd->attr_mask = ib_qp_attr_mask_to_pvrdma(attr_mask);
1007 
1008 	ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_QUERY_QP_RESP);
1009 	if (ret < 0) {
1010 		dev_warn(&dev->pdev->dev,
1011 			 "could not query queuepair, error: %d\n", ret);
1012 		goto out;
1013 	}
1014 
1015 	attr->qp_state = pvrdma_qp_state_to_ib(resp->attrs.qp_state);
1016 	attr->cur_qp_state =
1017 		pvrdma_qp_state_to_ib(resp->attrs.cur_qp_state);
1018 	attr->path_mtu = pvrdma_mtu_to_ib(resp->attrs.path_mtu);
1019 	attr->path_mig_state =
1020 		pvrdma_mig_state_to_ib(resp->attrs.path_mig_state);
1021 	attr->qkey = resp->attrs.qkey;
1022 	attr->rq_psn = resp->attrs.rq_psn;
1023 	attr->sq_psn = resp->attrs.sq_psn;
1024 	attr->dest_qp_num = resp->attrs.dest_qp_num;
1025 	attr->qp_access_flags =
1026 		pvrdma_access_flags_to_ib(resp->attrs.qp_access_flags);
1027 	attr->pkey_index = resp->attrs.pkey_index;
1028 	attr->alt_pkey_index = resp->attrs.alt_pkey_index;
1029 	attr->en_sqd_async_notify = resp->attrs.en_sqd_async_notify;
1030 	attr->sq_draining = resp->attrs.sq_draining;
1031 	attr->max_rd_atomic = resp->attrs.max_rd_atomic;
1032 	attr->max_dest_rd_atomic = resp->attrs.max_dest_rd_atomic;
1033 	attr->min_rnr_timer = resp->attrs.min_rnr_timer;
1034 	attr->port_num = resp->attrs.port_num;
1035 	attr->timeout = resp->attrs.timeout;
1036 	attr->retry_cnt = resp->attrs.retry_cnt;
1037 	attr->rnr_retry = resp->attrs.rnr_retry;
1038 	attr->alt_port_num = resp->attrs.alt_port_num;
1039 	attr->alt_timeout = resp->attrs.alt_timeout;
1040 	pvrdma_qp_cap_to_ib(&attr->cap, &resp->attrs.cap);
1041 	pvrdma_ah_attr_to_rdma(&attr->ah_attr, &resp->attrs.ah_attr);
1042 	pvrdma_ah_attr_to_rdma(&attr->alt_ah_attr, &resp->attrs.alt_ah_attr);
1043 
1044 	qp->state = attr->qp_state;
1045 
1046 	ret = 0;
1047 
1048 out:
1049 	attr->cur_qp_state = attr->qp_state;
1050 
1051 	init_attr->event_handler = qp->ibqp.event_handler;
1052 	init_attr->qp_context = qp->ibqp.qp_context;
1053 	init_attr->send_cq = qp->ibqp.send_cq;
1054 	init_attr->recv_cq = qp->ibqp.recv_cq;
1055 	init_attr->srq = qp->ibqp.srq;
1056 	init_attr->xrcd = NULL;
1057 	init_attr->cap = attr->cap;
1058 	init_attr->sq_sig_type = 0;
1059 	init_attr->qp_type = qp->ibqp.qp_type;
1060 	init_attr->create_flags = 0;
1061 	init_attr->port_num = qp->port;
1062 
1063 	mutex_unlock(&qp->mutex);
1064 	return ret;
1065 }
1066