1 /*
2 * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
3 *
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30 * SOFTWARE.
31 */
32
33 #include <linux/module.h>
34 #include <linux/mlx5/qp.h>
35 #include <linux/mlx5/srq.h>
36 #include <linux/slab.h>
37 #include <rdma/ib_umem.h>
38 #include <rdma/ib_user_verbs.h>
39
40 #include "mlx5_ib.h"
41
42 /* not supported currently */
43 static int srq_signature;
44
get_wqe(struct mlx5_ib_srq * srq,int n)45 static void *get_wqe(struct mlx5_ib_srq *srq, int n)
46 {
47 return mlx5_buf_offset(&srq->buf, n << srq->msrq.wqe_shift);
48 }
49
mlx5_ib_srq_event(struct mlx5_core_srq * srq,enum mlx5_event type)50 static void mlx5_ib_srq_event(struct mlx5_core_srq *srq, enum mlx5_event type)
51 {
52 struct ib_event event;
53 struct ib_srq *ibsrq = &to_mibsrq(srq)->ibsrq;
54
55 if (ibsrq->event_handler) {
56 event.device = ibsrq->device;
57 event.element.srq = ibsrq;
58 switch (type) {
59 case MLX5_EVENT_TYPE_SRQ_RQ_LIMIT:
60 event.event = IB_EVENT_SRQ_LIMIT_REACHED;
61 break;
62 case MLX5_EVENT_TYPE_SRQ_CATAS_ERROR:
63 event.event = IB_EVENT_SRQ_ERR;
64 break;
65 default:
66 pr_warn("mlx5_ib: Unexpected event type %d on SRQ %06x\n",
67 type, srq->srqn);
68 return;
69 }
70
71 ibsrq->event_handler(&event, ibsrq->srq_context);
72 }
73 }
74
create_srq_user(struct ib_pd * pd,struct mlx5_ib_srq * srq,struct mlx5_srq_attr * in,struct ib_udata * udata,int buf_size)75 static int create_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq,
76 struct mlx5_srq_attr *in,
77 struct ib_udata *udata, int buf_size)
78 {
79 struct mlx5_ib_dev *dev = to_mdev(pd->device);
80 struct mlx5_ib_create_srq ucmd = {};
81 size_t ucmdlen;
82 int err;
83 int npages;
84 int page_shift;
85 int ncont;
86 u32 offset;
87 u32 uidx = MLX5_IB_DEFAULT_UIDX;
88
89 ucmdlen = min(udata->inlen, sizeof(ucmd));
90
91 if (ib_copy_from_udata(&ucmd, udata, ucmdlen)) {
92 mlx5_ib_dbg(dev, "failed copy udata\n");
93 return -EFAULT;
94 }
95
96 if (ucmd.reserved0 || ucmd.reserved1)
97 return -EINVAL;
98
99 if (udata->inlen > sizeof(ucmd) &&
100 !ib_is_udata_cleared(udata, sizeof(ucmd),
101 udata->inlen - sizeof(ucmd)))
102 return -EINVAL;
103
104 if (in->type != IB_SRQT_BASIC) {
105 err = get_srq_user_index(to_mucontext(pd->uobject->context),
106 &ucmd, udata->inlen, &uidx);
107 if (err)
108 return err;
109 }
110
111 srq->wq_sig = !!(ucmd.flags & MLX5_SRQ_FLAG_SIGNATURE);
112
113 srq->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr, buf_size,
114 0, 0);
115 if (IS_ERR(srq->umem)) {
116 mlx5_ib_dbg(dev, "failed umem get, size %d\n", buf_size);
117 err = PTR_ERR(srq->umem);
118 return err;
119 }
120
121 mlx5_ib_cont_pages(srq->umem, ucmd.buf_addr, 0, &npages,
122 &page_shift, &ncont, NULL);
123 err = mlx5_ib_get_buf_offset(ucmd.buf_addr, page_shift,
124 &offset);
125 if (err) {
126 mlx5_ib_warn(dev, "bad offset\n");
127 goto err_umem;
128 }
129
130 in->pas = kvcalloc(ncont, sizeof(*in->pas), GFP_KERNEL);
131 if (!in->pas) {
132 err = -ENOMEM;
133 goto err_umem;
134 }
135
136 mlx5_ib_populate_pas(dev, srq->umem, page_shift, in->pas, 0);
137
138 err = mlx5_ib_db_map_user(to_mucontext(pd->uobject->context),
139 ucmd.db_addr, &srq->db);
140 if (err) {
141 mlx5_ib_dbg(dev, "map doorbell failed\n");
142 goto err_in;
143 }
144
145 in->log_page_size = page_shift - MLX5_ADAPTER_PAGE_SHIFT;
146 in->page_offset = offset;
147 if (MLX5_CAP_GEN(dev->mdev, cqe_version) == MLX5_CQE_VERSION_V1 &&
148 in->type != IB_SRQT_BASIC)
149 in->user_index = uidx;
150
151 return 0;
152
153 err_in:
154 kvfree(in->pas);
155
156 err_umem:
157 ib_umem_release(srq->umem);
158
159 return err;
160 }
161
create_srq_kernel(struct mlx5_ib_dev * dev,struct mlx5_ib_srq * srq,struct mlx5_srq_attr * in,int buf_size)162 static int create_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq,
163 struct mlx5_srq_attr *in, int buf_size)
164 {
165 int err;
166 int i;
167 struct mlx5_wqe_srq_next_seg *next;
168
169 err = mlx5_db_alloc(dev->mdev, &srq->db);
170 if (err) {
171 mlx5_ib_warn(dev, "alloc dbell rec failed\n");
172 return err;
173 }
174
175 if (mlx5_buf_alloc(dev->mdev, buf_size, &srq->buf)) {
176 mlx5_ib_dbg(dev, "buf alloc failed\n");
177 err = -ENOMEM;
178 goto err_db;
179 }
180
181 srq->head = 0;
182 srq->tail = srq->msrq.max - 1;
183 srq->wqe_ctr = 0;
184
185 for (i = 0; i < srq->msrq.max; i++) {
186 next = get_wqe(srq, i);
187 next->next_wqe_index =
188 cpu_to_be16((i + 1) & (srq->msrq.max - 1));
189 }
190
191 mlx5_ib_dbg(dev, "srq->buf.page_shift = %d\n", srq->buf.page_shift);
192 in->pas = kvcalloc(srq->buf.npages, sizeof(*in->pas), GFP_KERNEL);
193 if (!in->pas) {
194 err = -ENOMEM;
195 goto err_buf;
196 }
197 mlx5_fill_page_array(&srq->buf, in->pas);
198
199 srq->wrid = kvmalloc_array(srq->msrq.max, sizeof(u64), GFP_KERNEL);
200 if (!srq->wrid) {
201 err = -ENOMEM;
202 goto err_in;
203 }
204 srq->wq_sig = !!srq_signature;
205
206 in->log_page_size = srq->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT;
207 if (MLX5_CAP_GEN(dev->mdev, cqe_version) == MLX5_CQE_VERSION_V1 &&
208 in->type != IB_SRQT_BASIC)
209 in->user_index = MLX5_IB_DEFAULT_UIDX;
210
211 return 0;
212
213 err_in:
214 kvfree(in->pas);
215
216 err_buf:
217 mlx5_buf_free(dev->mdev, &srq->buf);
218
219 err_db:
220 mlx5_db_free(dev->mdev, &srq->db);
221 return err;
222 }
223
destroy_srq_user(struct ib_pd * pd,struct mlx5_ib_srq * srq)224 static void destroy_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq)
225 {
226 mlx5_ib_db_unmap_user(to_mucontext(pd->uobject->context), &srq->db);
227 ib_umem_release(srq->umem);
228 }
229
230
destroy_srq_kernel(struct mlx5_ib_dev * dev,struct mlx5_ib_srq * srq)231 static void destroy_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq)
232 {
233 kvfree(srq->wrid);
234 mlx5_buf_free(dev->mdev, &srq->buf);
235 mlx5_db_free(dev->mdev, &srq->db);
236 }
237
mlx5_ib_create_srq(struct ib_pd * pd,struct ib_srq_init_attr * init_attr,struct ib_udata * udata)238 struct ib_srq *mlx5_ib_create_srq(struct ib_pd *pd,
239 struct ib_srq_init_attr *init_attr,
240 struct ib_udata *udata)
241 {
242 struct mlx5_ib_dev *dev = to_mdev(pd->device);
243 struct mlx5_ib_srq *srq;
244 size_t desc_size;
245 size_t buf_size;
246 int err;
247 struct mlx5_srq_attr in = {0};
248 __u32 max_srq_wqes = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
249
250 /* Sanity check SRQ size before proceeding */
251 if (init_attr->attr.max_wr >= max_srq_wqes) {
252 mlx5_ib_dbg(dev, "max_wr %d, cap %d\n",
253 init_attr->attr.max_wr,
254 max_srq_wqes);
255 return ERR_PTR(-EINVAL);
256 }
257
258 srq = kmalloc(sizeof(*srq), GFP_KERNEL);
259 if (!srq)
260 return ERR_PTR(-ENOMEM);
261
262 mutex_init(&srq->mutex);
263 spin_lock_init(&srq->lock);
264 srq->msrq.max = roundup_pow_of_two(init_attr->attr.max_wr + 1);
265 srq->msrq.max_gs = init_attr->attr.max_sge;
266
267 desc_size = sizeof(struct mlx5_wqe_srq_next_seg) +
268 srq->msrq.max_gs * sizeof(struct mlx5_wqe_data_seg);
269 if (desc_size == 0 || srq->msrq.max_gs > desc_size) {
270 err = -EINVAL;
271 goto err_srq;
272 }
273 desc_size = roundup_pow_of_two(desc_size);
274 desc_size = max_t(size_t, 32, desc_size);
275 if (desc_size < sizeof(struct mlx5_wqe_srq_next_seg)) {
276 err = -EINVAL;
277 goto err_srq;
278 }
279 srq->msrq.max_avail_gather = (desc_size - sizeof(struct mlx5_wqe_srq_next_seg)) /
280 sizeof(struct mlx5_wqe_data_seg);
281 srq->msrq.wqe_shift = ilog2(desc_size);
282 buf_size = srq->msrq.max * desc_size;
283 if (buf_size < desc_size) {
284 err = -EINVAL;
285 goto err_srq;
286 }
287 in.type = init_attr->srq_type;
288
289 if (pd->uobject)
290 err = create_srq_user(pd, srq, &in, udata, buf_size);
291 else
292 err = create_srq_kernel(dev, srq, &in, buf_size);
293
294 if (err) {
295 mlx5_ib_warn(dev, "create srq %s failed, err %d\n",
296 pd->uobject ? "user" : "kernel", err);
297 goto err_srq;
298 }
299
300 in.log_size = ilog2(srq->msrq.max);
301 in.wqe_shift = srq->msrq.wqe_shift - 4;
302 if (srq->wq_sig)
303 in.flags |= MLX5_SRQ_FLAG_WQ_SIG;
304
305 if (init_attr->srq_type == IB_SRQT_XRC)
306 in.xrcd = to_mxrcd(init_attr->ext.xrc.xrcd)->xrcdn;
307 else
308 in.xrcd = to_mxrcd(dev->devr.x0)->xrcdn;
309
310 if (init_attr->srq_type == IB_SRQT_TM) {
311 in.tm_log_list_size =
312 ilog2(init_attr->ext.tag_matching.max_num_tags) + 1;
313 if (in.tm_log_list_size >
314 MLX5_CAP_GEN(dev->mdev, log_tag_matching_list_sz)) {
315 mlx5_ib_dbg(dev, "TM SRQ max_num_tags exceeding limit\n");
316 err = -EINVAL;
317 goto err_usr_kern_srq;
318 }
319 in.flags |= MLX5_SRQ_FLAG_RNDV;
320 }
321
322 if (ib_srq_has_cq(init_attr->srq_type))
323 in.cqn = to_mcq(init_attr->ext.cq)->mcq.cqn;
324 else
325 in.cqn = to_mcq(dev->devr.c0)->mcq.cqn;
326
327 in.pd = to_mpd(pd)->pdn;
328 in.db_record = srq->db.dma;
329 err = mlx5_core_create_srq(dev->mdev, &srq->msrq, &in);
330 kvfree(in.pas);
331 if (err) {
332 mlx5_ib_dbg(dev, "create SRQ failed, err %d\n", err);
333 goto err_usr_kern_srq;
334 }
335
336 mlx5_ib_dbg(dev, "create SRQ with srqn 0x%x\n", srq->msrq.srqn);
337
338 srq->msrq.event = mlx5_ib_srq_event;
339 srq->ibsrq.ext.xrc.srq_num = srq->msrq.srqn;
340
341 if (pd->uobject)
342 if (ib_copy_to_udata(udata, &srq->msrq.srqn, sizeof(__u32))) {
343 mlx5_ib_dbg(dev, "copy to user failed\n");
344 err = -EFAULT;
345 goto err_core;
346 }
347
348 init_attr->attr.max_wr = srq->msrq.max - 1;
349
350 return &srq->ibsrq;
351
352 err_core:
353 mlx5_core_destroy_srq(dev->mdev, &srq->msrq);
354
355 err_usr_kern_srq:
356 if (pd->uobject)
357 destroy_srq_user(pd, srq);
358 else
359 destroy_srq_kernel(dev, srq);
360
361 err_srq:
362 kfree(srq);
363
364 return ERR_PTR(err);
365 }
366
mlx5_ib_modify_srq(struct ib_srq * ibsrq,struct ib_srq_attr * attr,enum ib_srq_attr_mask attr_mask,struct ib_udata * udata)367 int mlx5_ib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
368 enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
369 {
370 struct mlx5_ib_dev *dev = to_mdev(ibsrq->device);
371 struct mlx5_ib_srq *srq = to_msrq(ibsrq);
372 int ret;
373
374 /* We don't support resizing SRQs yet */
375 if (attr_mask & IB_SRQ_MAX_WR)
376 return -EINVAL;
377
378 if (attr_mask & IB_SRQ_LIMIT) {
379 if (attr->srq_limit >= srq->msrq.max)
380 return -EINVAL;
381
382 mutex_lock(&srq->mutex);
383 ret = mlx5_core_arm_srq(dev->mdev, &srq->msrq, attr->srq_limit, 1);
384 mutex_unlock(&srq->mutex);
385
386 if (ret)
387 return ret;
388 }
389
390 return 0;
391 }
392
mlx5_ib_query_srq(struct ib_srq * ibsrq,struct ib_srq_attr * srq_attr)393 int mlx5_ib_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
394 {
395 struct mlx5_ib_dev *dev = to_mdev(ibsrq->device);
396 struct mlx5_ib_srq *srq = to_msrq(ibsrq);
397 int ret;
398 struct mlx5_srq_attr *out;
399
400 out = kzalloc(sizeof(*out), GFP_KERNEL);
401 if (!out)
402 return -ENOMEM;
403
404 ret = mlx5_core_query_srq(dev->mdev, &srq->msrq, out);
405 if (ret)
406 goto out_box;
407
408 srq_attr->srq_limit = out->lwm;
409 srq_attr->max_wr = srq->msrq.max - 1;
410 srq_attr->max_sge = srq->msrq.max_gs;
411
412 out_box:
413 kfree(out);
414 return ret;
415 }
416
mlx5_ib_destroy_srq(struct ib_srq * srq)417 int mlx5_ib_destroy_srq(struct ib_srq *srq)
418 {
419 struct mlx5_ib_dev *dev = to_mdev(srq->device);
420 struct mlx5_ib_srq *msrq = to_msrq(srq);
421
422 mlx5_core_destroy_srq(dev->mdev, &msrq->msrq);
423
424 if (srq->uobject) {
425 mlx5_ib_db_unmap_user(to_mucontext(srq->uobject->context), &msrq->db);
426 ib_umem_release(msrq->umem);
427 } else {
428 destroy_srq_kernel(dev, msrq);
429 }
430
431 kfree(srq);
432 return 0;
433 }
434
mlx5_ib_free_srq_wqe(struct mlx5_ib_srq * srq,int wqe_index)435 void mlx5_ib_free_srq_wqe(struct mlx5_ib_srq *srq, int wqe_index)
436 {
437 struct mlx5_wqe_srq_next_seg *next;
438
439 /* always called with interrupts disabled. */
440 spin_lock(&srq->lock);
441
442 next = get_wqe(srq, srq->tail);
443 next->next_wqe_index = cpu_to_be16(wqe_index);
444 srq->tail = wqe_index;
445
446 spin_unlock(&srq->lock);
447 }
448
mlx5_ib_post_srq_recv(struct ib_srq * ibsrq,const struct ib_recv_wr * wr,const struct ib_recv_wr ** bad_wr)449 int mlx5_ib_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
450 const struct ib_recv_wr **bad_wr)
451 {
452 struct mlx5_ib_srq *srq = to_msrq(ibsrq);
453 struct mlx5_wqe_srq_next_seg *next;
454 struct mlx5_wqe_data_seg *scat;
455 struct mlx5_ib_dev *dev = to_mdev(ibsrq->device);
456 struct mlx5_core_dev *mdev = dev->mdev;
457 unsigned long flags;
458 int err = 0;
459 int nreq;
460 int i;
461
462 spin_lock_irqsave(&srq->lock, flags);
463
464 if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
465 err = -EIO;
466 *bad_wr = wr;
467 goto out;
468 }
469
470 for (nreq = 0; wr; nreq++, wr = wr->next) {
471 if (unlikely(wr->num_sge > srq->msrq.max_gs)) {
472 err = -EINVAL;
473 *bad_wr = wr;
474 break;
475 }
476
477 if (unlikely(srq->head == srq->tail)) {
478 err = -ENOMEM;
479 *bad_wr = wr;
480 break;
481 }
482
483 srq->wrid[srq->head] = wr->wr_id;
484
485 next = get_wqe(srq, srq->head);
486 srq->head = be16_to_cpu(next->next_wqe_index);
487 scat = (struct mlx5_wqe_data_seg *)(next + 1);
488
489 for (i = 0; i < wr->num_sge; i++) {
490 scat[i].byte_count = cpu_to_be32(wr->sg_list[i].length);
491 scat[i].lkey = cpu_to_be32(wr->sg_list[i].lkey);
492 scat[i].addr = cpu_to_be64(wr->sg_list[i].addr);
493 }
494
495 if (i < srq->msrq.max_avail_gather) {
496 scat[i].byte_count = 0;
497 scat[i].lkey = cpu_to_be32(MLX5_INVALID_LKEY);
498 scat[i].addr = 0;
499 }
500 }
501
502 if (likely(nreq)) {
503 srq->wqe_ctr += nreq;
504
505 /* Make sure that descriptors are written before
506 * doorbell record.
507 */
508 wmb();
509
510 *srq->db.db = cpu_to_be32(srq->wqe_ctr);
511 }
512 out:
513 spin_unlock_irqrestore(&srq->lock, flags);
514
515 return err;
516 }
517
