xref: /linux/drivers/infiniband/hw/vmw_pvrdma/pvrdma_cq.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
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  */
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  * @attrs: bundle
98  *
99  * @return: 0 on success
100  */
101 int pvrdma_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
102 		     struct uverbs_attr_bundle *attrs)
103 {
104 	struct ib_udata *udata = &attrs->driver_udata;
105 	struct ib_device *ibdev = ibcq->device;
106 	int entries = attr->cqe;
107 	struct pvrdma_dev *dev = to_vdev(ibdev);
108 	struct pvrdma_cq *cq = to_vcq(ibcq);
109 	int ret;
110 	int npages;
111 	unsigned long flags;
112 	union pvrdma_cmd_req req;
113 	union pvrdma_cmd_resp rsp;
114 	struct pvrdma_cmd_create_cq *cmd = &req.create_cq;
115 	struct pvrdma_cmd_create_cq_resp *resp = &rsp.create_cq_resp;
116 	struct pvrdma_create_cq_resp cq_resp = {};
117 	struct pvrdma_create_cq ucmd;
118 	struct pvrdma_ucontext *context = rdma_udata_to_drv_context(
119 		udata, struct pvrdma_ucontext, ibucontext);
120 
121 	BUILD_BUG_ON(sizeof(struct pvrdma_cqe) != 64);
122 
123 	if (attr->flags)
124 		return -EOPNOTSUPP;
125 
126 	entries = roundup_pow_of_two(entries);
127 	if (entries < 1 || entries > dev->dsr->caps.max_cqe)
128 		return -EINVAL;
129 
130 	if (!atomic_add_unless(&dev->num_cqs, 1, dev->dsr->caps.max_cq))
131 		return -ENOMEM;
132 
133 	cq->ibcq.cqe = entries;
134 	cq->is_kernel = !udata;
135 
136 	if (!cq->is_kernel) {
137 		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
138 			ret = -EFAULT;
139 			goto err_cq;
140 		}
141 
142 		cq->umem = ib_umem_get(ibdev, ucmd.buf_addr, ucmd.buf_size,
143 				       IB_ACCESS_LOCAL_WRITE);
144 		if (IS_ERR(cq->umem)) {
145 			ret = PTR_ERR(cq->umem);
146 			goto err_cq;
147 		}
148 
149 		npages = ib_umem_num_dma_blocks(cq->umem, PAGE_SIZE);
150 	} else {
151 		/* One extra page for shared ring state */
152 		npages = 1 + (entries * sizeof(struct pvrdma_cqe) +
153 			      PAGE_SIZE - 1) / PAGE_SIZE;
154 
155 		/* Skip header page. */
156 		cq->offset = PAGE_SIZE;
157 	}
158 
159 	if (npages < 0 || npages > PVRDMA_PAGE_DIR_MAX_PAGES) {
160 		dev_warn(&dev->pdev->dev,
161 			 "overflow pages in completion queue\n");
162 		ret = -EINVAL;
163 		goto err_umem;
164 	}
165 
166 	ret = pvrdma_page_dir_init(dev, &cq->pdir, npages, cq->is_kernel);
167 	if (ret) {
168 		dev_warn(&dev->pdev->dev,
169 			 "could not allocate page directory\n");
170 		goto err_umem;
171 	}
172 
173 	/* Ring state is always the first page. Set in library for user cq. */
174 	if (cq->is_kernel)
175 		cq->ring_state = cq->pdir.pages[0];
176 	else
177 		pvrdma_page_dir_insert_umem(&cq->pdir, cq->umem, 0);
178 
179 	refcount_set(&cq->refcnt, 1);
180 	init_completion(&cq->free);
181 	spin_lock_init(&cq->cq_lock);
182 
183 	memset(cmd, 0, sizeof(*cmd));
184 	cmd->hdr.cmd = PVRDMA_CMD_CREATE_CQ;
185 	cmd->nchunks = npages;
186 	cmd->ctx_handle = context ? context->ctx_handle : 0;
187 	cmd->cqe = entries;
188 	cmd->pdir_dma = cq->pdir.dir_dma;
189 	ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_CREATE_CQ_RESP);
190 	if (ret < 0) {
191 		dev_warn(&dev->pdev->dev,
192 			 "could not create completion queue, error: %d\n", ret);
193 		goto err_page_dir;
194 	}
195 
196 	cq->ibcq.cqe = resp->cqe;
197 	cq->cq_handle = resp->cq_handle;
198 	cq_resp.cqn = resp->cq_handle;
199 	spin_lock_irqsave(&dev->cq_tbl_lock, flags);
200 	dev->cq_tbl[cq->cq_handle % dev->dsr->caps.max_cq] = cq;
201 	spin_unlock_irqrestore(&dev->cq_tbl_lock, flags);
202 
203 	if (!cq->is_kernel) {
204 		cq->uar = &context->uar;
205 
206 		/* Copy udata back. */
207 		if (ib_copy_to_udata(udata, &cq_resp, sizeof(cq_resp))) {
208 			dev_warn(&dev->pdev->dev,
209 				 "failed to copy back udata\n");
210 			pvrdma_destroy_cq(&cq->ibcq, udata);
211 			return -EINVAL;
212 		}
213 	}
214 
215 	return 0;
216 
217 err_page_dir:
218 	pvrdma_page_dir_cleanup(dev, &cq->pdir);
219 err_umem:
220 	ib_umem_release(cq->umem);
221 err_cq:
222 	atomic_dec(&dev->num_cqs);
223 	return ret;
224 }
225 
226 static void pvrdma_free_cq(struct pvrdma_dev *dev, struct pvrdma_cq *cq)
227 {
228 	if (refcount_dec_and_test(&cq->refcnt))
229 		complete(&cq->free);
230 	wait_for_completion(&cq->free);
231 
232 	ib_umem_release(cq->umem);
233 
234 	pvrdma_page_dir_cleanup(dev, &cq->pdir);
235 }
236 
237 /**
238  * pvrdma_destroy_cq - destroy completion queue
239  * @cq: the completion queue to destroy.
240  * @udata: user data or null for kernel object
241  */
242 int pvrdma_destroy_cq(struct ib_cq *cq, struct ib_udata *udata)
243 {
244 	struct pvrdma_cq *vcq = to_vcq(cq);
245 	union pvrdma_cmd_req req;
246 	struct pvrdma_cmd_destroy_cq *cmd = &req.destroy_cq;
247 	struct pvrdma_dev *dev = to_vdev(cq->device);
248 	unsigned long flags;
249 	int ret;
250 
251 	memset(cmd, 0, sizeof(*cmd));
252 	cmd->hdr.cmd = PVRDMA_CMD_DESTROY_CQ;
253 	cmd->cq_handle = vcq->cq_handle;
254 
255 	ret = pvrdma_cmd_post(dev, &req, NULL, 0);
256 	if (ret < 0)
257 		dev_warn(&dev->pdev->dev,
258 			 "could not destroy completion queue, error: %d\n",
259 			 ret);
260 
261 	/* free cq's resources */
262 	spin_lock_irqsave(&dev->cq_tbl_lock, flags);
263 	dev->cq_tbl[vcq->cq_handle] = NULL;
264 	spin_unlock_irqrestore(&dev->cq_tbl_lock, flags);
265 
266 	pvrdma_free_cq(dev, vcq);
267 	atomic_dec(&dev->num_cqs);
268 	return 0;
269 }
270 
271 static inline struct pvrdma_cqe *get_cqe(struct pvrdma_cq *cq, int i)
272 {
273 	return (struct pvrdma_cqe *)pvrdma_page_dir_get_ptr(
274 					&cq->pdir,
275 					cq->offset +
276 					sizeof(struct pvrdma_cqe) * i);
277 }
278 
279 void _pvrdma_flush_cqe(struct pvrdma_qp *qp, struct pvrdma_cq *cq)
280 {
281 	unsigned int head;
282 	int has_data;
283 
284 	if (!cq->is_kernel)
285 		return;
286 
287 	/* Lock held */
288 	has_data = pvrdma_idx_ring_has_data(&cq->ring_state->rx,
289 					    cq->ibcq.cqe, &head);
290 	if (unlikely(has_data > 0)) {
291 		int items;
292 		int curr;
293 		int tail = pvrdma_idx(&cq->ring_state->rx.prod_tail,
294 				      cq->ibcq.cqe);
295 		struct pvrdma_cqe *cqe;
296 		struct pvrdma_cqe *curr_cqe;
297 
298 		items = (tail > head) ? (tail - head) :
299 			(cq->ibcq.cqe - head + tail);
300 		curr = --tail;
301 		while (items-- > 0) {
302 			if (curr < 0)
303 				curr = cq->ibcq.cqe - 1;
304 			if (tail < 0)
305 				tail = cq->ibcq.cqe - 1;
306 			curr_cqe = get_cqe(cq, curr);
307 			if ((curr_cqe->qp & 0xFFFF) != qp->qp_handle) {
308 				if (curr != tail) {
309 					cqe = get_cqe(cq, tail);
310 					*cqe = *curr_cqe;
311 				}
312 				tail--;
313 			} else {
314 				pvrdma_idx_ring_inc(
315 					&cq->ring_state->rx.cons_head,
316 					cq->ibcq.cqe);
317 			}
318 			curr--;
319 		}
320 	}
321 }
322 
323 static int pvrdma_poll_one(struct pvrdma_cq *cq, struct pvrdma_qp **cur_qp,
324 			   struct ib_wc *wc)
325 {
326 	struct pvrdma_dev *dev = to_vdev(cq->ibcq.device);
327 	int has_data;
328 	unsigned int head;
329 	bool tried = false;
330 	struct pvrdma_cqe *cqe;
331 
332 retry:
333 	has_data = pvrdma_idx_ring_has_data(&cq->ring_state->rx,
334 					    cq->ibcq.cqe, &head);
335 	if (has_data == 0) {
336 		if (tried)
337 			return -EAGAIN;
338 
339 		pvrdma_write_uar_cq(dev, cq->cq_handle | PVRDMA_UAR_CQ_POLL);
340 
341 		tried = true;
342 		goto retry;
343 	} else if (has_data == PVRDMA_INVALID_IDX) {
344 		dev_err(&dev->pdev->dev, "CQ ring state invalid\n");
345 		return -EAGAIN;
346 	}
347 
348 	cqe = get_cqe(cq, head);
349 
350 	/* Ensure cqe is valid. */
351 	rmb();
352 	if (dev->qp_tbl[cqe->qp & 0xffff])
353 		*cur_qp = (struct pvrdma_qp *)dev->qp_tbl[cqe->qp & 0xffff];
354 	else
355 		return -EAGAIN;
356 
357 	wc->opcode = pvrdma_wc_opcode_to_ib(cqe->opcode);
358 	wc->status = pvrdma_wc_status_to_ib(cqe->status);
359 	wc->wr_id = cqe->wr_id;
360 	wc->qp = &(*cur_qp)->ibqp;
361 	wc->byte_len = cqe->byte_len;
362 	wc->ex.imm_data = cqe->imm_data;
363 	wc->src_qp = cqe->src_qp;
364 	wc->wc_flags = pvrdma_wc_flags_to_ib(cqe->wc_flags);
365 	wc->pkey_index = cqe->pkey_index;
366 	wc->slid = cqe->slid;
367 	wc->sl = cqe->sl;
368 	wc->dlid_path_bits = cqe->dlid_path_bits;
369 	wc->port_num = cqe->port_num;
370 	wc->vendor_err = cqe->vendor_err;
371 	wc->network_hdr_type = pvrdma_network_type_to_ib(cqe->network_hdr_type);
372 
373 	/* Update shared ring state */
374 	pvrdma_idx_ring_inc(&cq->ring_state->rx.cons_head, cq->ibcq.cqe);
375 
376 	return 0;
377 }
378 
379 /**
380  * pvrdma_poll_cq - poll for work completion queue entries
381  * @ibcq: completion queue
382  * @num_entries: the maximum number of entries
383  * @wc: pointer to work completion array
384  *
385  * @return: number of polled completion entries
386  */
387 int pvrdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
388 {
389 	struct pvrdma_cq *cq = to_vcq(ibcq);
390 	struct pvrdma_qp *cur_qp = NULL;
391 	unsigned long flags;
392 	int npolled;
393 
394 	if (num_entries < 1 || wc == NULL)
395 		return 0;
396 
397 	spin_lock_irqsave(&cq->cq_lock, flags);
398 	for (npolled = 0; npolled < num_entries; ++npolled) {
399 		if (pvrdma_poll_one(cq, &cur_qp, wc + npolled))
400 			break;
401 	}
402 
403 	spin_unlock_irqrestore(&cq->cq_lock, flags);
404 
405 	/* Ensure we do not return errors from poll_cq */
406 	return npolled;
407 }
408