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 #include <rdma/uverbs_ioctl.h>
53 
54 #include "pvrdma.h"
55 
56 /**
57  * pvrdma_req_notify_cq - request notification for a completion queue
58  * @ibcq: the completion queue
59  * @notify_flags: notification flags
60  *
61  * @return: 0 for success.
62  */
pvrdma_req_notify_cq(struct ib_cq * ibcq,enum ib_cq_notify_flags notify_flags)63 int pvrdma_req_notify_cq(struct ib_cq *ibcq,
64 			 enum ib_cq_notify_flags notify_flags)
65 {
66 	struct pvrdma_dev *dev = to_vdev(ibcq->device);
67 	struct pvrdma_cq *cq = to_vcq(ibcq);
68 	u32 val = cq->cq_handle;
69 	unsigned long flags;
70 	int has_data = 0;
71 
72 	val |= (notify_flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
73 		PVRDMA_UAR_CQ_ARM_SOL : PVRDMA_UAR_CQ_ARM;
74 
75 	spin_lock_irqsave(&cq->cq_lock, flags);
76 
77 	pvrdma_write_uar_cq(dev, val);
78 
79 	if (notify_flags & IB_CQ_REPORT_MISSED_EVENTS) {
80 		unsigned int head;
81 
82 		has_data = pvrdma_idx_ring_has_data(&cq->ring_state->rx,
83 						    cq->ibcq.cqe, &head);
84 		if (unlikely(has_data == PVRDMA_INVALID_IDX))
85 			dev_err(&dev->pdev->dev, "CQ ring state invalid\n");
86 	}
87 
88 	spin_unlock_irqrestore(&cq->cq_lock, flags);
89 
90 	return has_data;
91 }
92 
93 /**
94  * pvrdma_create_cq - create completion queue
95  * @ibcq: Allocated CQ
96  * @attr: completion queue attributes
97  * @udata: user data
98  *
99  * @return: 0 on success
100  */
pvrdma_create_cq(struct ib_cq * ibcq,const struct ib_cq_init_attr * attr,struct ib_udata * udata)101 int pvrdma_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
102 		     struct ib_udata *udata)
103 {
104 	struct ib_device *ibdev = ibcq->device;
105 	int entries = attr->cqe;
106 	struct pvrdma_dev *dev = to_vdev(ibdev);
107 	struct pvrdma_cq *cq = to_vcq(ibcq);
108 	int ret;
109 	int npages;
110 	unsigned long flags;
111 	union pvrdma_cmd_req req;
112 	union pvrdma_cmd_resp rsp;
113 	struct pvrdma_cmd_create_cq *cmd = &req.create_cq;
114 	struct pvrdma_cmd_create_cq_resp *resp = &rsp.create_cq_resp;
115 	struct pvrdma_create_cq_resp cq_resp = {};
116 	struct pvrdma_create_cq ucmd;
117 	struct pvrdma_ucontext *context = rdma_udata_to_drv_context(
118 		udata, struct pvrdma_ucontext, ibucontext);
119 
120 	BUILD_BUG_ON(sizeof(struct pvrdma_cqe) != 64);
121 
122 	if (attr->flags)
123 		return -EOPNOTSUPP;
124 
125 	entries = roundup_pow_of_two(entries);
126 	if (entries < 1 || entries > dev->dsr->caps.max_cqe)
127 		return -EINVAL;
128 
129 	if (!atomic_add_unless(&dev->num_cqs, 1, dev->dsr->caps.max_cq))
130 		return -ENOMEM;
131 
132 	cq->ibcq.cqe = entries;
133 	cq->is_kernel = !udata;
134 
135 	if (!cq->is_kernel) {
136 		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
137 			ret = -EFAULT;
138 			goto err_cq;
139 		}
140 
141 		cq->umem = ib_umem_get(ibdev, ucmd.buf_addr, ucmd.buf_size,
142 				       IB_ACCESS_LOCAL_WRITE);
143 		if (IS_ERR(cq->umem)) {
144 			ret = PTR_ERR(cq->umem);
145 			goto err_cq;
146 		}
147 
148 		npages = ib_umem_num_dma_blocks(cq->umem, PAGE_SIZE);
149 	} else {
150 		/* One extra page for shared ring state */
151 		npages = 1 + (entries * sizeof(struct pvrdma_cqe) +
152 			      PAGE_SIZE - 1) / PAGE_SIZE;
153 
154 		/* Skip header page. */
155 		cq->offset = PAGE_SIZE;
156 	}
157 
158 	if (npages < 0 || npages > PVRDMA_PAGE_DIR_MAX_PAGES) {
159 		dev_warn(&dev->pdev->dev,
160 			 "overflow pages in completion queue\n");
161 		ret = -EINVAL;
162 		goto err_umem;
163 	}
164 
165 	ret = pvrdma_page_dir_init(dev, &cq->pdir, npages, cq->is_kernel);
166 	if (ret) {
167 		dev_warn(&dev->pdev->dev,
168 			 "could not allocate page directory\n");
169 		goto err_umem;
170 	}
171 
172 	/* Ring state is always the first page. Set in library for user cq. */
173 	if (cq->is_kernel)
174 		cq->ring_state = cq->pdir.pages[0];
175 	else
176 		pvrdma_page_dir_insert_umem(&cq->pdir, cq->umem, 0);
177 
178 	refcount_set(&cq->refcnt, 1);
179 	init_completion(&cq->free);
180 	spin_lock_init(&cq->cq_lock);
181 
182 	memset(cmd, 0, sizeof(*cmd));
183 	cmd->hdr.cmd = PVRDMA_CMD_CREATE_CQ;
184 	cmd->nchunks = npages;
185 	cmd->ctx_handle = context ? context->ctx_handle : 0;
186 	cmd->cqe = entries;
187 	cmd->pdir_dma = cq->pdir.dir_dma;
188 	ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_CREATE_CQ_RESP);
189 	if (ret < 0) {
190 		dev_warn(&dev->pdev->dev,
191 			 "could not create completion queue, error: %d\n", ret);
192 		goto err_page_dir;
193 	}
194 
195 	cq->ibcq.cqe = resp->cqe;
196 	cq->cq_handle = resp->cq_handle;
197 	cq_resp.cqn = resp->cq_handle;
198 	spin_lock_irqsave(&dev->cq_tbl_lock, flags);
199 	dev->cq_tbl[cq->cq_handle % dev->dsr->caps.max_cq] = cq;
200 	spin_unlock_irqrestore(&dev->cq_tbl_lock, flags);
201 
202 	if (!cq->is_kernel) {
203 		cq->uar = &context->uar;
204 
205 		/* Copy udata back. */
206 		if (ib_copy_to_udata(udata, &cq_resp, sizeof(cq_resp))) {
207 			dev_warn(&dev->pdev->dev,
208 				 "failed to copy back udata\n");
209 			pvrdma_destroy_cq(&cq->ibcq, udata);
210 			return -EINVAL;
211 		}
212 	}
213 
214 	return 0;
215 
216 err_page_dir:
217 	pvrdma_page_dir_cleanup(dev, &cq->pdir);
218 err_umem:
219 	ib_umem_release(cq->umem);
220 err_cq:
221 	atomic_dec(&dev->num_cqs);
222 	return ret;
223 }
224 
pvrdma_free_cq(struct pvrdma_dev * dev,struct pvrdma_cq * cq)225 static void pvrdma_free_cq(struct pvrdma_dev *dev, struct pvrdma_cq *cq)
226 {
227 	if (refcount_dec_and_test(&cq->refcnt))
228 		complete(&cq->free);
229 	wait_for_completion(&cq->free);
230 
231 	ib_umem_release(cq->umem);
232 
233 	pvrdma_page_dir_cleanup(dev, &cq->pdir);
234 }
235 
236 /**
237  * pvrdma_destroy_cq - destroy completion queue
238  * @cq: the completion queue to destroy.
239  * @udata: user data or null for kernel object
240  */
pvrdma_destroy_cq(struct ib_cq * cq,struct ib_udata * udata)241 int pvrdma_destroy_cq(struct ib_cq *cq, struct ib_udata *udata)
242 {
243 	struct pvrdma_cq *vcq = to_vcq(cq);
244 	union pvrdma_cmd_req req;
245 	struct pvrdma_cmd_destroy_cq *cmd = &req.destroy_cq;
246 	struct pvrdma_dev *dev = to_vdev(cq->device);
247 	unsigned long flags;
248 	int ret;
249 
250 	memset(cmd, 0, sizeof(*cmd));
251 	cmd->hdr.cmd = PVRDMA_CMD_DESTROY_CQ;
252 	cmd->cq_handle = vcq->cq_handle;
253 
254 	ret = pvrdma_cmd_post(dev, &req, NULL, 0);
255 	if (ret < 0)
256 		dev_warn(&dev->pdev->dev,
257 			 "could not destroy completion queue, error: %d\n",
258 			 ret);
259 
260 	/* free cq's resources */
261 	spin_lock_irqsave(&dev->cq_tbl_lock, flags);
262 	dev->cq_tbl[vcq->cq_handle] = NULL;
263 	spin_unlock_irqrestore(&dev->cq_tbl_lock, flags);
264 
265 	pvrdma_free_cq(dev, vcq);
266 	atomic_dec(&dev->num_cqs);
267 	return 0;
268 }
269 
get_cqe(struct pvrdma_cq * cq,int i)270 static inline struct pvrdma_cqe *get_cqe(struct pvrdma_cq *cq, int i)
271 {
272 	return (struct pvrdma_cqe *)pvrdma_page_dir_get_ptr(
273 					&cq->pdir,
274 					cq->offset +
275 					sizeof(struct pvrdma_cqe) * i);
276 }
277 
_pvrdma_flush_cqe(struct pvrdma_qp * qp,struct pvrdma_cq * cq)278 void _pvrdma_flush_cqe(struct pvrdma_qp *qp, struct pvrdma_cq *cq)
279 {
280 	unsigned int head;
281 	int has_data;
282 
283 	if (!cq->is_kernel)
284 		return;
285 
286 	/* Lock held */
287 	has_data = pvrdma_idx_ring_has_data(&cq->ring_state->rx,
288 					    cq->ibcq.cqe, &head);
289 	if (unlikely(has_data > 0)) {
290 		int items;
291 		int curr;
292 		int tail = pvrdma_idx(&cq->ring_state->rx.prod_tail,
293 				      cq->ibcq.cqe);
294 		struct pvrdma_cqe *cqe;
295 		struct pvrdma_cqe *curr_cqe;
296 
297 		items = (tail > head) ? (tail - head) :
298 			(cq->ibcq.cqe - head + tail);
299 		curr = --tail;
300 		while (items-- > 0) {
301 			if (curr < 0)
302 				curr = cq->ibcq.cqe - 1;
303 			if (tail < 0)
304 				tail = cq->ibcq.cqe - 1;
305 			curr_cqe = get_cqe(cq, curr);
306 			if ((curr_cqe->qp & 0xFFFF) != qp->qp_handle) {
307 				if (curr != tail) {
308 					cqe = get_cqe(cq, tail);
309 					*cqe = *curr_cqe;
310 				}
311 				tail--;
312 			} else {
313 				pvrdma_idx_ring_inc(
314 					&cq->ring_state->rx.cons_head,
315 					cq->ibcq.cqe);
316 			}
317 			curr--;
318 		}
319 	}
320 }
321 
pvrdma_poll_one(struct pvrdma_cq * cq,struct pvrdma_qp ** cur_qp,struct ib_wc * wc)322 static int pvrdma_poll_one(struct pvrdma_cq *cq, struct pvrdma_qp **cur_qp,
323 			   struct ib_wc *wc)
324 {
325 	struct pvrdma_dev *dev = to_vdev(cq->ibcq.device);
326 	int has_data;
327 	unsigned int head;
328 	bool tried = false;
329 	struct pvrdma_cqe *cqe;
330 
331 retry:
332 	has_data = pvrdma_idx_ring_has_data(&cq->ring_state->rx,
333 					    cq->ibcq.cqe, &head);
334 	if (has_data == 0) {
335 		if (tried)
336 			return -EAGAIN;
337 
338 		pvrdma_write_uar_cq(dev, cq->cq_handle | PVRDMA_UAR_CQ_POLL);
339 
340 		tried = true;
341 		goto retry;
342 	} else if (has_data == PVRDMA_INVALID_IDX) {
343 		dev_err(&dev->pdev->dev, "CQ ring state invalid\n");
344 		return -EAGAIN;
345 	}
346 
347 	cqe = get_cqe(cq, head);
348 
349 	/* Ensure cqe is valid. */
350 	rmb();
351 	if (dev->qp_tbl[cqe->qp & 0xffff])
352 		*cur_qp = (struct pvrdma_qp *)dev->qp_tbl[cqe->qp & 0xffff];
353 	else
354 		return -EAGAIN;
355 
356 	wc->opcode = pvrdma_wc_opcode_to_ib(cqe->opcode);
357 	wc->status = pvrdma_wc_status_to_ib(cqe->status);
358 	wc->wr_id = cqe->wr_id;
359 	wc->qp = &(*cur_qp)->ibqp;
360 	wc->byte_len = cqe->byte_len;
361 	wc->ex.imm_data = cqe->imm_data;
362 	wc->src_qp = cqe->src_qp;
363 	wc->wc_flags = pvrdma_wc_flags_to_ib(cqe->wc_flags);
364 	wc->pkey_index = cqe->pkey_index;
365 	wc->slid = cqe->slid;
366 	wc->sl = cqe->sl;
367 	wc->dlid_path_bits = cqe->dlid_path_bits;
368 	wc->port_num = cqe->port_num;
369 	wc->vendor_err = cqe->vendor_err;
370 	wc->network_hdr_type = pvrdma_network_type_to_ib(cqe->network_hdr_type);
371 
372 	/* Update shared ring state */
373 	pvrdma_idx_ring_inc(&cq->ring_state->rx.cons_head, cq->ibcq.cqe);
374 
375 	return 0;
376 }
377 
378 /**
379  * pvrdma_poll_cq - poll for work completion queue entries
380  * @ibcq: completion queue
381  * @num_entries: the maximum number of entries
382  * @wc: pointer to work completion array
383  *
384  * @return: number of polled completion entries
385  */
pvrdma_poll_cq(struct ib_cq * ibcq,int num_entries,struct ib_wc * wc)386 int pvrdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
387 {
388 	struct pvrdma_cq *cq = to_vcq(ibcq);
389 	struct pvrdma_qp *cur_qp = NULL;
390 	unsigned long flags;
391 	int npolled;
392 
393 	if (num_entries < 1 || wc == NULL)
394 		return 0;
395 
396 	spin_lock_irqsave(&cq->cq_lock, flags);
397 	for (npolled = 0; npolled < num_entries; ++npolled) {
398 		if (pvrdma_poll_one(cq, &cur_qp, wc + npolled))
399 			break;
400 	}
401 
402 	spin_unlock_irqrestore(&cq->cq_lock, flags);
403 
404 	/* Ensure we do not return errors from poll_cq */
405 	return npolled;
406 }
407