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 "user.h"
40
get_wqe(struct mlx4_ib_srq * srq,int n)41 static void *get_wqe(struct mlx4_ib_srq *srq, int n)
42 {
43 return mlx4_buf_offset(&srq->buf, n << srq->msrq.wqe_shift);
44 }
45
mlx4_ib_srq_event(struct mlx4_srq * srq,enum mlx4_event type)46 static void mlx4_ib_srq_event(struct mlx4_srq *srq, enum mlx4_event type)
47 {
48 struct ib_event event;
49 struct ib_srq *ibsrq = &to_mibsrq(srq)->ibsrq;
50
51 if (ibsrq->event_handler) {
52 event.device = ibsrq->device;
53 event.element.srq = ibsrq;
54 switch (type) {
55 case MLX4_EVENT_TYPE_SRQ_LIMIT:
56 event.event = IB_EVENT_SRQ_LIMIT_REACHED;
57 break;
58 case MLX4_EVENT_TYPE_SRQ_CATAS_ERROR:
59 event.event = IB_EVENT_SRQ_ERR;
60 break;
61 default:
62 printk(KERN_WARNING "mlx4_ib: Unexpected event type %d "
63 "on SRQ %06x\n", type, srq->srqn);
64 return;
65 }
66
67 ibsrq->event_handler(&event, ibsrq->srq_context);
68 }
69 }
70
mlx4_ib_create_srq(struct ib_pd * pd,struct ib_srq_init_attr * init_attr,struct ib_udata * udata)71 struct ib_srq *mlx4_ib_create_srq(struct ib_pd *pd,
72 struct ib_srq_init_attr *init_attr,
73 struct ib_udata *udata)
74 {
75 struct mlx4_ib_dev *dev = to_mdev(pd->device);
76 struct mlx4_ib_srq *srq;
77 struct mlx4_wqe_srq_next_seg *next;
78 struct mlx4_wqe_data_seg *scatter;
79 u32 cqn;
80 u16 xrcdn;
81 int desc_size;
82 int buf_size;
83 int err;
84 int i;
85
86 /* Sanity check SRQ size before proceeding */
87 if (init_attr->attr.max_wr >= dev->dev->caps.max_srq_wqes ||
88 init_attr->attr.max_sge > dev->dev->caps.max_srq_sge)
89 return ERR_PTR(-EINVAL);
90
91 srq = kmalloc(sizeof *srq, GFP_KERNEL);
92 if (!srq)
93 return ERR_PTR(-ENOMEM);
94
95 mutex_init(&srq->mutex);
96 spin_lock_init(&srq->lock);
97 srq->msrq.max = roundup_pow_of_two(init_attr->attr.max_wr + 1);
98 srq->msrq.max_gs = init_attr->attr.max_sge;
99
100 desc_size = max(32UL,
101 roundup_pow_of_two(sizeof (struct mlx4_wqe_srq_next_seg) +
102 srq->msrq.max_gs *
103 sizeof (struct mlx4_wqe_data_seg)));
104 srq->msrq.wqe_shift = ilog2(desc_size);
105
106 buf_size = srq->msrq.max * desc_size;
107
108 if (pd->uobject) {
109 struct mlx4_ib_create_srq ucmd;
110
111 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
112 err = -EFAULT;
113 goto err_srq;
114 }
115
116 srq->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
117 buf_size, 0, 0);
118 if (IS_ERR(srq->umem)) {
119 err = PTR_ERR(srq->umem);
120 goto err_srq;
121 }
122
123 err = mlx4_mtt_init(dev->dev, ib_umem_page_count(srq->umem),
124 ilog2(srq->umem->page_size), &srq->mtt);
125 if (err)
126 goto err_buf;
127
128 err = mlx4_ib_umem_write_mtt(dev, &srq->mtt, srq->umem);
129 if (err)
130 goto err_mtt;
131
132 err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
133 ucmd.db_addr, &srq->db);
134 if (err)
135 goto err_mtt;
136 } else {
137 err = mlx4_db_alloc(dev->dev, &srq->db, 0);
138 if (err)
139 goto err_srq;
140
141 *srq->db.db = 0;
142
143 if (mlx4_buf_alloc(dev->dev, buf_size, PAGE_SIZE * 2, &srq->buf)) {
144 err = -ENOMEM;
145 goto err_db;
146 }
147
148 srq->head = 0;
149 srq->tail = srq->msrq.max - 1;
150 srq->wqe_ctr = 0;
151
152 for (i = 0; i < srq->msrq.max; ++i) {
153 next = get_wqe(srq, i);
154 next->next_wqe_index =
155 cpu_to_be16((i + 1) & (srq->msrq.max - 1));
156
157 for (scatter = (void *) (next + 1);
158 (void *) scatter < (void *) next + desc_size;
159 ++scatter)
160 scatter->lkey = cpu_to_be32(MLX4_INVALID_LKEY);
161 }
162
163 err = mlx4_mtt_init(dev->dev, srq->buf.npages, srq->buf.page_shift,
164 &srq->mtt);
165 if (err)
166 goto err_buf;
167
168 err = mlx4_buf_write_mtt(dev->dev, &srq->mtt, &srq->buf);
169 if (err)
170 goto err_mtt;
171
172 srq->wrid = kmalloc(srq->msrq.max * sizeof (u64), GFP_KERNEL);
173 if (!srq->wrid) {
174 err = -ENOMEM;
175 goto err_mtt;
176 }
177 }
178
179 cqn = (init_attr->srq_type == IB_SRQT_XRC) ?
180 to_mcq(init_attr->ext.xrc.cq)->mcq.cqn : 0;
181 xrcdn = (init_attr->srq_type == IB_SRQT_XRC) ?
182 to_mxrcd(init_attr->ext.xrc.xrcd)->xrcdn :
183 (u16) dev->dev->caps.reserved_xrcds;
184 err = mlx4_srq_alloc(dev->dev, to_mpd(pd)->pdn, cqn, xrcdn, &srq->mtt,
185 srq->db.dma, &srq->msrq);
186 if (err)
187 goto err_wrid;
188
189 srq->msrq.event = mlx4_ib_srq_event;
190 srq->ibsrq.ext.xrc.srq_num = srq->msrq.srqn;
191
192 if (pd->uobject)
193 if (ib_copy_to_udata(udata, &srq->msrq.srqn, sizeof (__u32))) {
194 err = -EFAULT;
195 goto err_wrid;
196 }
197
198 init_attr->attr.max_wr = srq->msrq.max - 1;
199
200 return &srq->ibsrq;
201
202 err_wrid:
203 if (pd->uobject)
204 mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &srq->db);
205 else
206 kfree(srq->wrid);
207
208 err_mtt:
209 mlx4_mtt_cleanup(dev->dev, &srq->mtt);
210
211 err_buf:
212 if (pd->uobject)
213 ib_umem_release(srq->umem);
214 else
215 mlx4_buf_free(dev->dev, buf_size, &srq->buf);
216
217 err_db:
218 if (!pd->uobject)
219 mlx4_db_free(dev->dev, &srq->db);
220
221 err_srq:
222 kfree(srq);
223
224 return ERR_PTR(err);
225 }
226
mlx4_ib_modify_srq(struct ib_srq * ibsrq,struct ib_srq_attr * attr,enum ib_srq_attr_mask attr_mask,struct ib_udata * udata)227 int mlx4_ib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
228 enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
229 {
230 struct mlx4_ib_dev *dev = to_mdev(ibsrq->device);
231 struct mlx4_ib_srq *srq = to_msrq(ibsrq);
232 int ret;
233
234 /* We don't support resizing SRQs (yet?) */
235 if (attr_mask & IB_SRQ_MAX_WR)
236 return -EINVAL;
237
238 if (attr_mask & IB_SRQ_LIMIT) {
239 if (attr->srq_limit >= srq->msrq.max)
240 return -EINVAL;
241
242 mutex_lock(&srq->mutex);
243 ret = mlx4_srq_arm(dev->dev, &srq->msrq, attr->srq_limit);
244 mutex_unlock(&srq->mutex);
245
246 if (ret)
247 return ret;
248 }
249
250 return 0;
251 }
252
mlx4_ib_query_srq(struct ib_srq * ibsrq,struct ib_srq_attr * srq_attr)253 int mlx4_ib_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
254 {
255 struct mlx4_ib_dev *dev = to_mdev(ibsrq->device);
256 struct mlx4_ib_srq *srq = to_msrq(ibsrq);
257 int ret;
258 int limit_watermark;
259
260 ret = mlx4_srq_query(dev->dev, &srq->msrq, &limit_watermark);
261 if (ret)
262 return ret;
263
264 srq_attr->srq_limit = limit_watermark;
265 srq_attr->max_wr = srq->msrq.max - 1;
266 srq_attr->max_sge = srq->msrq.max_gs;
267
268 return 0;
269 }
270
mlx4_ib_destroy_srq(struct ib_srq * srq)271 int mlx4_ib_destroy_srq(struct ib_srq *srq)
272 {
273 struct mlx4_ib_dev *dev = to_mdev(srq->device);
274 struct mlx4_ib_srq *msrq = to_msrq(srq);
275
276 mlx4_srq_free(dev->dev, &msrq->msrq);
277 mlx4_mtt_cleanup(dev->dev, &msrq->mtt);
278
279 if (srq->uobject) {
280 mlx4_ib_db_unmap_user(to_mucontext(srq->uobject->context), &msrq->db);
281 ib_umem_release(msrq->umem);
282 } else {
283 kfree(msrq->wrid);
284 mlx4_buf_free(dev->dev, msrq->msrq.max << msrq->msrq.wqe_shift,
285 &msrq->buf);
286 mlx4_db_free(dev->dev, &msrq->db);
287 }
288
289 kfree(msrq);
290
291 return 0;
292 }
293
mlx4_ib_free_srq_wqe(struct mlx4_ib_srq * srq,int wqe_index)294 void mlx4_ib_free_srq_wqe(struct mlx4_ib_srq *srq, int wqe_index)
295 {
296 struct mlx4_wqe_srq_next_seg *next;
297
298 /* always called with interrupts disabled. */
299 spin_lock(&srq->lock);
300
301 next = get_wqe(srq, srq->tail);
302 next->next_wqe_index = cpu_to_be16(wqe_index);
303 srq->tail = wqe_index;
304
305 spin_unlock(&srq->lock);
306 }
307
mlx4_ib_post_srq_recv(struct ib_srq * ibsrq,struct ib_recv_wr * wr,struct ib_recv_wr ** bad_wr)308 int mlx4_ib_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
309 struct ib_recv_wr **bad_wr)
310 {
311 struct mlx4_ib_srq *srq = to_msrq(ibsrq);
312 struct mlx4_wqe_srq_next_seg *next;
313 struct mlx4_wqe_data_seg *scat;
314 unsigned long flags;
315 int err = 0;
316 int nreq;
317 int i;
318
319 spin_lock_irqsave(&srq->lock, flags);
320
321 for (nreq = 0; wr; ++nreq, wr = wr->next) {
322 if (unlikely(wr->num_sge > srq->msrq.max_gs)) {
323 err = -EINVAL;
324 *bad_wr = wr;
325 break;
326 }
327
328 if (unlikely(srq->head == srq->tail)) {
329 err = -ENOMEM;
330 *bad_wr = wr;
331 break;
332 }
333
334 srq->wrid[srq->head] = wr->wr_id;
335
336 next = get_wqe(srq, srq->head);
337 srq->head = be16_to_cpu(next->next_wqe_index);
338 scat = (struct mlx4_wqe_data_seg *) (next + 1);
339
340 for (i = 0; i < wr->num_sge; ++i) {
341 scat[i].byte_count = cpu_to_be32(wr->sg_list[i].length);
342 scat[i].lkey = cpu_to_be32(wr->sg_list[i].lkey);
343 scat[i].addr = cpu_to_be64(wr->sg_list[i].addr);
344 }
345
346 if (i < srq->msrq.max_gs) {
347 scat[i].byte_count = 0;
348 scat[i].lkey = cpu_to_be32(MLX4_INVALID_LKEY);
349 scat[i].addr = 0;
350 }
351 }
352
353 if (likely(nreq)) {
354 srq->wqe_ctr += nreq;
355
356 /*
357 * Make sure that descriptors are written before
358 * doorbell record.
359 */
360 wmb();
361
362 *srq->db.db = cpu_to_be32(srq->wqe_ctr);
363 }
364
365 spin_unlock_irqrestore(&srq->lock, flags);
366
367 return err;
368 }
369