xref: /openbmc/linux/drivers/infiniband/hw/mlx4/srq.c (revision abcda807)
1 /*
2  * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
3  * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33 
34 #include <linux/mlx4/qp.h>
35 #include <linux/mlx4/srq.h>
36 #include <linux/slab.h>
37 
38 #include "mlx4_ib.h"
39 #include <rdma/mlx4-abi.h>
40 #include <rdma/uverbs_ioctl.h>
41 
42 static void *get_wqe(struct mlx4_ib_srq *srq, int n)
43 {
44 	return mlx4_buf_offset(&srq->buf, n << srq->msrq.wqe_shift);
45 }
46 
47 static void mlx4_ib_srq_event(struct mlx4_srq *srq, enum mlx4_event type)
48 {
49 	struct ib_event event;
50 	struct ib_srq *ibsrq = &to_mibsrq(srq)->ibsrq;
51 
52 	if (ibsrq->event_handler) {
53 		event.device      = ibsrq->device;
54 		event.element.srq = ibsrq;
55 		switch (type) {
56 		case MLX4_EVENT_TYPE_SRQ_LIMIT:
57 			event.event = IB_EVENT_SRQ_LIMIT_REACHED;
58 			break;
59 		case MLX4_EVENT_TYPE_SRQ_CATAS_ERROR:
60 			event.event = IB_EVENT_SRQ_ERR;
61 			break;
62 		default:
63 			pr_warn("Unexpected event type %d "
64 			       "on SRQ %06x\n", type, srq->srqn);
65 			return;
66 		}
67 
68 		ibsrq->event_handler(&event, ibsrq->srq_context);
69 	}
70 }
71 
72 int mlx4_ib_create_srq(struct ib_srq *ib_srq,
73 		       struct ib_srq_init_attr *init_attr,
74 		       struct ib_udata *udata)
75 {
76 	struct mlx4_ib_dev *dev = to_mdev(ib_srq->device);
77 	struct mlx4_ib_ucontext *ucontext = rdma_udata_to_drv_context(
78 		udata, struct mlx4_ib_ucontext, ibucontext);
79 	struct mlx4_ib_srq *srq = to_msrq(ib_srq);
80 	struct mlx4_wqe_srq_next_seg *next;
81 	struct mlx4_wqe_data_seg *scatter;
82 	u32 cqn;
83 	u16 xrcdn;
84 	int desc_size;
85 	int buf_size;
86 	int err;
87 	int i;
88 
89 	/* Sanity check SRQ size before proceeding */
90 	if (init_attr->attr.max_wr  >= dev->dev->caps.max_srq_wqes ||
91 	    init_attr->attr.max_sge >  dev->dev->caps.max_srq_sge)
92 		return -EINVAL;
93 
94 	mutex_init(&srq->mutex);
95 	spin_lock_init(&srq->lock);
96 	srq->msrq.max    = roundup_pow_of_two(init_attr->attr.max_wr + 1);
97 	srq->msrq.max_gs = init_attr->attr.max_sge;
98 
99 	desc_size = max(32UL,
100 			roundup_pow_of_two(sizeof (struct mlx4_wqe_srq_next_seg) +
101 					   srq->msrq.max_gs *
102 					   sizeof (struct mlx4_wqe_data_seg)));
103 	srq->msrq.wqe_shift = ilog2(desc_size);
104 
105 	buf_size = srq->msrq.max * desc_size;
106 
107 	if (udata) {
108 		struct mlx4_ib_create_srq ucmd;
109 
110 		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd)))
111 			return -EFAULT;
112 
113 		srq->umem =
114 			ib_umem_get(ib_srq->device, ucmd.buf_addr, buf_size, 0);
115 		if (IS_ERR(srq->umem))
116 			return PTR_ERR(srq->umem);
117 
118 		err = mlx4_mtt_init(
119 			dev->dev, ib_umem_num_dma_blocks(srq->umem, PAGE_SIZE),
120 			PAGE_SHIFT, &srq->mtt);
121 		if (err)
122 			goto err_buf;
123 
124 		err = mlx4_ib_umem_write_mtt(dev, &srq->mtt, srq->umem);
125 		if (err)
126 			goto err_mtt;
127 
128 		err = mlx4_ib_db_map_user(udata, ucmd.db_addr, &srq->db);
129 		if (err)
130 			goto err_mtt;
131 	} else {
132 		err = mlx4_db_alloc(dev->dev, &srq->db, 0);
133 		if (err)
134 			return err;
135 
136 		*srq->db.db = 0;
137 
138 		if (mlx4_buf_alloc(dev->dev, buf_size, PAGE_SIZE * 2,
139 				   &srq->buf)) {
140 			err = -ENOMEM;
141 			goto err_db;
142 		}
143 
144 		srq->head    = 0;
145 		srq->tail    = srq->msrq.max - 1;
146 		srq->wqe_ctr = 0;
147 
148 		for (i = 0; i < srq->msrq.max; ++i) {
149 			next = get_wqe(srq, i);
150 			next->next_wqe_index =
151 				cpu_to_be16((i + 1) & (srq->msrq.max - 1));
152 
153 			for (scatter = (void *) (next + 1);
154 			     (void *) scatter < (void *) next + desc_size;
155 			     ++scatter)
156 				scatter->lkey = cpu_to_be32(MLX4_INVALID_LKEY);
157 		}
158 
159 		err = mlx4_mtt_init(dev->dev, srq->buf.npages, srq->buf.page_shift,
160 				    &srq->mtt);
161 		if (err)
162 			goto err_buf;
163 
164 		err = mlx4_buf_write_mtt(dev->dev, &srq->mtt, &srq->buf);
165 		if (err)
166 			goto err_mtt;
167 
168 		srq->wrid = kvmalloc_array(srq->msrq.max,
169 					   sizeof(u64), GFP_KERNEL);
170 		if (!srq->wrid) {
171 			err = -ENOMEM;
172 			goto err_mtt;
173 		}
174 	}
175 
176 	cqn = ib_srq_has_cq(init_attr->srq_type) ?
177 		to_mcq(init_attr->ext.cq)->mcq.cqn : 0;
178 	xrcdn = (init_attr->srq_type == IB_SRQT_XRC) ?
179 		to_mxrcd(init_attr->ext.xrc.xrcd)->xrcdn :
180 		(u16) dev->dev->caps.reserved_xrcds;
181 	err = mlx4_srq_alloc(dev->dev, to_mpd(ib_srq->pd)->pdn, cqn, xrcdn,
182 			     &srq->mtt, srq->db.dma, &srq->msrq);
183 	if (err)
184 		goto err_wrid;
185 
186 	srq->msrq.event = mlx4_ib_srq_event;
187 	srq->ibsrq.ext.xrc.srq_num = srq->msrq.srqn;
188 
189 	if (udata)
190 		if (ib_copy_to_udata(udata, &srq->msrq.srqn, sizeof (__u32))) {
191 			err = -EFAULT;
192 			goto err_wrid;
193 		}
194 
195 	init_attr->attr.max_wr = srq->msrq.max - 1;
196 
197 	return 0;
198 
199 err_wrid:
200 	if (udata)
201 		mlx4_ib_db_unmap_user(ucontext, &srq->db);
202 	else
203 		kvfree(srq->wrid);
204 
205 err_mtt:
206 	mlx4_mtt_cleanup(dev->dev, &srq->mtt);
207 
208 err_buf:
209 	if (!srq->umem)
210 		mlx4_buf_free(dev->dev, buf_size, &srq->buf);
211 	ib_umem_release(srq->umem);
212 
213 err_db:
214 	if (!udata)
215 		mlx4_db_free(dev->dev, &srq->db);
216 
217 	return err;
218 }
219 
220 int mlx4_ib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
221 		       enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
222 {
223 	struct mlx4_ib_dev *dev = to_mdev(ibsrq->device);
224 	struct mlx4_ib_srq *srq = to_msrq(ibsrq);
225 	int ret;
226 
227 	/* We don't support resizing SRQs (yet?) */
228 	if (attr_mask & IB_SRQ_MAX_WR)
229 		return -EINVAL;
230 
231 	if (attr_mask & IB_SRQ_LIMIT) {
232 		if (attr->srq_limit >= srq->msrq.max)
233 			return -EINVAL;
234 
235 		mutex_lock(&srq->mutex);
236 		ret = mlx4_srq_arm(dev->dev, &srq->msrq, attr->srq_limit);
237 		mutex_unlock(&srq->mutex);
238 
239 		if (ret)
240 			return ret;
241 	}
242 
243 	return 0;
244 }
245 
246 int mlx4_ib_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
247 {
248 	struct mlx4_ib_dev *dev = to_mdev(ibsrq->device);
249 	struct mlx4_ib_srq *srq = to_msrq(ibsrq);
250 	int ret;
251 	int limit_watermark;
252 
253 	ret = mlx4_srq_query(dev->dev, &srq->msrq, &limit_watermark);
254 	if (ret)
255 		return ret;
256 
257 	srq_attr->srq_limit = limit_watermark;
258 	srq_attr->max_wr    = srq->msrq.max - 1;
259 	srq_attr->max_sge   = srq->msrq.max_gs;
260 
261 	return 0;
262 }
263 
264 int mlx4_ib_destroy_srq(struct ib_srq *srq, struct ib_udata *udata)
265 {
266 	struct mlx4_ib_dev *dev = to_mdev(srq->device);
267 	struct mlx4_ib_srq *msrq = to_msrq(srq);
268 
269 	mlx4_srq_free(dev->dev, &msrq->msrq);
270 	mlx4_mtt_cleanup(dev->dev, &msrq->mtt);
271 
272 	if (udata) {
273 		mlx4_ib_db_unmap_user(
274 			rdma_udata_to_drv_context(
275 				udata,
276 				struct mlx4_ib_ucontext,
277 				ibucontext),
278 			&msrq->db);
279 	} else {
280 		kvfree(msrq->wrid);
281 		mlx4_buf_free(dev->dev, msrq->msrq.max << msrq->msrq.wqe_shift,
282 			      &msrq->buf);
283 		mlx4_db_free(dev->dev, &msrq->db);
284 	}
285 	ib_umem_release(msrq->umem);
286 	return 0;
287 }
288 
289 void mlx4_ib_free_srq_wqe(struct mlx4_ib_srq *srq, int wqe_index)
290 {
291 	struct mlx4_wqe_srq_next_seg *next;
292 
293 	/* always called with interrupts disabled. */
294 	spin_lock(&srq->lock);
295 
296 	next = get_wqe(srq, srq->tail);
297 	next->next_wqe_index = cpu_to_be16(wqe_index);
298 	srq->tail = wqe_index;
299 
300 	spin_unlock(&srq->lock);
301 }
302 
303 int mlx4_ib_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
304 			  const struct ib_recv_wr **bad_wr)
305 {
306 	struct mlx4_ib_srq *srq = to_msrq(ibsrq);
307 	struct mlx4_wqe_srq_next_seg *next;
308 	struct mlx4_wqe_data_seg *scat;
309 	unsigned long flags;
310 	int err = 0;
311 	int nreq;
312 	int i;
313 	struct mlx4_ib_dev *mdev = to_mdev(ibsrq->device);
314 
315 	spin_lock_irqsave(&srq->lock, flags);
316 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
317 		err = -EIO;
318 		*bad_wr = wr;
319 		nreq = 0;
320 		goto out;
321 	}
322 
323 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
324 		if (unlikely(wr->num_sge > srq->msrq.max_gs)) {
325 			err = -EINVAL;
326 			*bad_wr = wr;
327 			break;
328 		}
329 
330 		if (unlikely(srq->head == srq->tail)) {
331 			err = -ENOMEM;
332 			*bad_wr = wr;
333 			break;
334 		}
335 
336 		srq->wrid[srq->head] = wr->wr_id;
337 
338 		next      = get_wqe(srq, srq->head);
339 		srq->head = be16_to_cpu(next->next_wqe_index);
340 		scat      = (struct mlx4_wqe_data_seg *) (next + 1);
341 
342 		for (i = 0; i < wr->num_sge; ++i) {
343 			scat[i].byte_count = cpu_to_be32(wr->sg_list[i].length);
344 			scat[i].lkey       = cpu_to_be32(wr->sg_list[i].lkey);
345 			scat[i].addr       = cpu_to_be64(wr->sg_list[i].addr);
346 		}
347 
348 		if (i < srq->msrq.max_gs) {
349 			scat[i].byte_count = 0;
350 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
351 			scat[i].addr       = 0;
352 		}
353 	}
354 
355 	if (likely(nreq)) {
356 		srq->wqe_ctr += nreq;
357 
358 		/*
359 		 * Make sure that descriptors are written before
360 		 * doorbell record.
361 		 */
362 		wmb();
363 
364 		*srq->db.db = cpu_to_be32(srq->wqe_ctr);
365 	}
366 out:
367 
368 	spin_unlock_irqrestore(&srq->lock, flags);
369 
370 	return err;
371 }
372