xref: /freebsd/sys/dev/mlx5/mlx5_ib/mlx5_ib_srq.c (revision 8ddb146abcdf061be9f2c0db7e391697dafad85c)
1 /*-
2  * Copyright (c) 2013-2020, Mellanox Technologies, Ltd.  All rights reserved.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. Redistributions in binary form must reproduce the above copyright
10  *    notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  *
13  * THIS SOFTWARE IS PROVIDED BY AUTHOR AND CONTRIBUTORS `AS IS' AND
14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16  * ARE DISCLAIMED.  IN NO EVENT SHALL AUTHOR OR CONTRIBUTORS BE LIABLE
17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23  * SUCH DAMAGE.
24  *
25  * $FreeBSD$
26  */
27 
28 #include "opt_rss.h"
29 #include "opt_ratelimit.h"
30 
31 #include <linux/module.h>
32 #include <dev/mlx5/qp.h>
33 #include <dev/mlx5/srq.h>
34 #include <linux/slab.h>
35 #include <rdma/ib_umem.h>
36 #include <rdma/ib_user_verbs.h>
37 #include <rdma/uverbs_ioctl.h>
38 
39 #include <dev/mlx5/mlx5_ib/mlx5_ib.h>
40 
41 static void *get_wqe(struct mlx5_ib_srq *srq, int n)
42 {
43 	return mlx5_buf_offset(&srq->buf, n << srq->msrq.wqe_shift);
44 }
45 
46 static void mlx5_ib_srq_event(struct mlx5_core_srq *srq, int 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 MLX5_EVENT_TYPE_SRQ_RQ_LIMIT:
56 			event.event = IB_EVENT_SRQ_LIMIT_REACHED;
57 			break;
58 		case MLX5_EVENT_TYPE_SRQ_CATAS_ERROR:
59 			event.event = IB_EVENT_SRQ_ERR;
60 			break;
61 		default:
62 			pr_warn("mlx5_ib: Unexpected event type %d on SRQ %06x\n",
63 				type, srq->srqn);
64 			return;
65 		}
66 
67 		ibsrq->event_handler(&event, ibsrq->srq_context);
68 	}
69 }
70 
71 static int create_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq,
72 			   struct mlx5_srq_attr *in,
73 			   struct ib_udata *udata, int buf_size)
74 {
75 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
76 	struct mlx5_ib_create_srq ucmd = {};
77 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
78 		udata, struct mlx5_ib_ucontext, ibucontext);
79 	size_t ucmdlen;
80 	int err;
81 	int npages;
82 	int page_shift;
83 	int ncont;
84 	u32 offset;
85 	u32 uidx = MLX5_IB_DEFAULT_UIDX;
86 
87 	ucmdlen = min(udata->inlen, sizeof(ucmd));
88 
89 	if (ib_copy_from_udata(&ucmd, udata, ucmdlen)) {
90 		mlx5_ib_dbg(dev, "failed copy udata\n");
91 		return -EFAULT;
92 	}
93 
94 	if (ucmd.reserved0 || ucmd.reserved1)
95 		return -EINVAL;
96 
97 	if (udata->inlen > sizeof(ucmd) &&
98 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
99 				 udata->inlen - sizeof(ucmd)))
100 		return -EINVAL;
101 
102 	if (in->type != IB_SRQT_BASIC) {
103 		err = get_srq_user_index(ucontext, &ucmd, udata->inlen, &uidx);
104 		if (err)
105 			return err;
106 	}
107 
108 	srq->wq_sig = !!(ucmd.flags & MLX5_SRQ_FLAG_SIGNATURE);
109 
110 	srq->umem = ib_umem_get(&ucontext->ibucontext, ucmd.buf_addr, buf_size, 0, 0);
111 	if (IS_ERR(srq->umem)) {
112 		mlx5_ib_dbg(dev, "failed umem get, size %d\n", buf_size);
113 		err = PTR_ERR(srq->umem);
114 		return err;
115 	}
116 
117 	mlx5_ib_cont_pages(srq->umem, ucmd.buf_addr, 0, &npages,
118 			   &page_shift, &ncont, NULL);
119 	err = mlx5_ib_get_buf_offset(ucmd.buf_addr, page_shift,
120 				     &offset);
121 	if (err) {
122 		mlx5_ib_warn(dev, "bad offset\n");
123 		goto err_umem;
124 	}
125 
126 	in->pas = mlx5_vzalloc(sizeof(*in->pas) * ncont);
127 	if (!in->pas) {
128 		err = -ENOMEM;
129 		goto err_umem;
130 	}
131 
132 	mlx5_ib_populate_pas(dev, srq->umem, page_shift, in->pas, 0);
133 
134 	err = mlx5_ib_db_map_user(ucontext, ucmd.db_addr, &srq->db);
135 	if (err) {
136 		mlx5_ib_dbg(dev, "map doorbell failed\n");
137 		goto err_in;
138 	}
139 
140 	in->log_page_size = page_shift - MLX5_ADAPTER_PAGE_SHIFT;
141 	in->page_offset = offset;
142 	in->uid = (in->type != IB_SRQT_XRC) ?  to_mpd(pd)->uid : 0;
143 	if (MLX5_CAP_GEN(dev->mdev, cqe_version) == MLX5_CQE_VERSION_V1 &&
144 	    in->type != IB_SRQT_BASIC)
145 		in->user_index = uidx;
146 
147 	return 0;
148 
149 err_in:
150 	kvfree(in->pas);
151 
152 err_umem:
153 	ib_umem_release(srq->umem);
154 
155 	return err;
156 }
157 
158 static int create_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq,
159 			     struct mlx5_srq_attr *in, int buf_size)
160 {
161 	int err;
162 	int i;
163 	struct mlx5_wqe_srq_next_seg *next;
164 
165 	err = mlx5_db_alloc(dev->mdev, &srq->db);
166 	if (err) {
167 		mlx5_ib_warn(dev, "alloc dbell rec failed\n");
168 		return err;
169 	}
170 
171 	if (mlx5_buf_alloc(dev->mdev, buf_size, 2 * PAGE_SIZE, &srq->buf)) {
172 		mlx5_ib_dbg(dev, "buf alloc failed\n");
173 		err = -ENOMEM;
174 		goto err_db;
175 	}
176 
177 	srq->head    = 0;
178 	srq->tail    = srq->msrq.max - 1;
179 	srq->wqe_ctr = 0;
180 
181 	for (i = 0; i < srq->msrq.max; i++) {
182 		next = get_wqe(srq, i);
183 		next->next_wqe_index =
184 			cpu_to_be16((i + 1) & (srq->msrq.max - 1));
185 	}
186 
187 	mlx5_ib_dbg(dev, "srq->buf.page_shift = %d\n", srq->buf.page_shift);
188 	in->pas = mlx5_vzalloc(sizeof(*in->pas) * srq->buf.npages);
189 	if (!in->pas) {
190 		err = -ENOMEM;
191 		goto err_buf;
192 	}
193 	mlx5_fill_page_array(&srq->buf, in->pas);
194 
195 	srq->wrid = kmalloc(srq->msrq.max * sizeof(u64), GFP_KERNEL);
196 	if (!srq->wrid) {
197 		err = -ENOMEM;
198 		goto err_in;
199 	}
200 	srq->wq_sig = 0;
201 
202 	in->log_page_size = srq->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT;
203 	if (MLX5_CAP_GEN(dev->mdev, cqe_version) == MLX5_CQE_VERSION_V1 &&
204 	    in->type != IB_SRQT_BASIC)
205 		in->user_index = MLX5_IB_DEFAULT_UIDX;
206 
207 	return 0;
208 
209 err_in:
210 	kvfree(in->pas);
211 
212 err_buf:
213 	mlx5_buf_free(dev->mdev, &srq->buf);
214 
215 err_db:
216 	mlx5_db_free(dev->mdev, &srq->db);
217 	return err;
218 }
219 
220 static void destroy_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq,
221 			     struct ib_udata *udata)
222 {
223 	mlx5_ib_db_unmap_user(
224 		rdma_udata_to_drv_context(
225 			udata,
226 			struct mlx5_ib_ucontext,
227 			ibucontext),
228 		&srq->db);
229 	ib_umem_release(srq->umem);
230 }
231 
232 
233 static void destroy_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq)
234 {
235 	kfree(srq->wrid);
236 	mlx5_buf_free(dev->mdev, &srq->buf);
237 	mlx5_db_free(dev->mdev, &srq->db);
238 }
239 
240 int mlx5_ib_create_srq(struct ib_srq *ib_srq,
241 		       struct ib_srq_init_attr *init_attr,
242 		       struct ib_udata *udata)
243 {
244 	struct mlx5_ib_dev *dev = to_mdev(ib_srq->device);
245 	struct mlx5_ib_srq *srq = to_msrq(ib_srq);
246 	size_t desc_size;
247 	size_t buf_size;
248 	int err;
249 	struct mlx5_srq_attr in = {};
250 	__u32 max_srq_wqes = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
251 
252 	/* Sanity check SRQ size before proceeding */
253 	if (init_attr->attr.max_wr >= max_srq_wqes) {
254 		mlx5_ib_dbg(dev, "max_wr %d, cap %d\n",
255 			    init_attr->attr.max_wr,
256 			    max_srq_wqes);
257 		return -EINVAL;
258 	}
259 
260 	mutex_init(&srq->mutex);
261 	spin_lock_init(&srq->lock);
262 	srq->msrq.max    = roundup_pow_of_two(init_attr->attr.max_wr + 1);
263 	srq->msrq.max_gs = init_attr->attr.max_sge;
264 
265 	desc_size = sizeof(struct mlx5_wqe_srq_next_seg) +
266 		    srq->msrq.max_gs * sizeof(struct mlx5_wqe_data_seg);
267 	if (desc_size == 0 || srq->msrq.max_gs > desc_size)
268 		return -EINVAL;
269 
270 	desc_size = roundup_pow_of_two(desc_size);
271 	desc_size = max_t(size_t, 32, desc_size);
272 	if (desc_size < sizeof(struct mlx5_wqe_srq_next_seg))
273 		return -EINVAL;
274 
275 	srq->msrq.max_avail_gather = (desc_size - sizeof(struct mlx5_wqe_srq_next_seg)) /
276 		sizeof(struct mlx5_wqe_data_seg);
277 	srq->msrq.wqe_shift = ilog2(desc_size);
278 	buf_size = srq->msrq.max * desc_size;
279 	if (buf_size < desc_size)
280 		return -EINVAL;
281 
282 	in.type = init_attr->srq_type;
283 
284 	if (udata)
285 		err = create_srq_user(ib_srq->pd, srq, &in, udata, buf_size);
286 	else
287 		err = create_srq_kernel(dev, srq, &in, buf_size);
288 
289 	if (err) {
290 		mlx5_ib_warn(dev, "create srq %s failed, err %d\n",
291 			     udata ? "user" : "kernel", err);
292 		return err;
293 	}
294 
295 	in.log_size = ilog2(srq->msrq.max);
296 	in.wqe_shift = srq->msrq.wqe_shift - 4;
297 	if (srq->wq_sig)
298 		in.flags |= MLX5_SRQ_FLAG_WQ_SIG;
299 
300 	if (init_attr->srq_type == IB_SRQT_XRC)
301 		in.xrcd = to_mxrcd(init_attr->ext.xrc.xrcd)->xrcdn;
302 	else
303 		in.xrcd = to_mxrcd(dev->devr.x0)->xrcdn;
304 
305 	if (ib_srq_has_cq(init_attr->srq_type))
306 		in.cqn = to_mcq(init_attr->ext.cq)->mcq.cqn;
307 	else
308 		in.cqn = to_mcq(dev->devr.c0)->mcq.cqn;
309 
310 	in.pd = to_mpd(ib_srq->pd)->pdn;
311 	in.db_record = srq->db.dma;
312 	err = mlx5_core_create_srq(dev->mdev, &srq->msrq, &in);
313 	kvfree(in.pas);
314 	if (err) {
315 		mlx5_ib_dbg(dev, "create SRQ failed, err %d\n", err);
316 		goto err_usr_kern_srq;
317 	}
318 
319 	mlx5_ib_dbg(dev, "create SRQ with srqn 0x%x\n", srq->msrq.srqn);
320 
321 	srq->msrq.event = mlx5_ib_srq_event;
322 	srq->ibsrq.ext.xrc.srq_num = srq->msrq.srqn;
323 
324 	if (udata)
325 		if (ib_copy_to_udata(udata, &srq->msrq.srqn, sizeof(__u32))) {
326 			mlx5_ib_dbg(dev, "copy to user failed\n");
327 			err = -EFAULT;
328 			goto err_core;
329 		}
330 
331 	init_attr->attr.max_wr = srq->msrq.max - 1;
332 
333 	return 0;
334 
335 err_core:
336 	mlx5_core_destroy_srq(dev->mdev, &srq->msrq);
337 
338 err_usr_kern_srq:
339 	if (udata)
340 		destroy_srq_user(ib_srq->pd, srq, udata);
341 	else
342 		destroy_srq_kernel(dev, srq);
343 
344 	return err;
345 }
346 
347 int mlx5_ib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
348 		       enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
349 {
350 	struct mlx5_ib_dev *dev = to_mdev(ibsrq->device);
351 	struct mlx5_ib_srq *srq = to_msrq(ibsrq);
352 	int ret;
353 
354 	/* We don't support resizing SRQs yet */
355 	if (attr_mask & IB_SRQ_MAX_WR)
356 		return -EINVAL;
357 
358 	if (attr_mask & IB_SRQ_LIMIT) {
359 		if (attr->srq_limit >= srq->msrq.max)
360 			return -EINVAL;
361 
362 		mutex_lock(&srq->mutex);
363 		ret = mlx5_core_arm_srq(dev->mdev, &srq->msrq, attr->srq_limit, 1);
364 		mutex_unlock(&srq->mutex);
365 
366 		if (ret)
367 			return ret;
368 	}
369 
370 	return 0;
371 }
372 
373 int mlx5_ib_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
374 {
375 	struct mlx5_ib_dev *dev = to_mdev(ibsrq->device);
376 	struct mlx5_ib_srq *srq = to_msrq(ibsrq);
377 	int ret;
378 	struct mlx5_srq_attr *out;
379 
380 	out = kzalloc(sizeof(*out), GFP_KERNEL);
381 	if (!out)
382 		return -ENOMEM;
383 
384 	ret = mlx5_core_query_srq(dev->mdev, &srq->msrq, out);
385 	if (ret)
386 		goto out_box;
387 
388 	srq_attr->srq_limit = out->lwm;
389 	srq_attr->max_wr    = srq->msrq.max - 1;
390 	srq_attr->max_sge   = srq->msrq.max_gs;
391 
392 out_box:
393 	kfree(out);
394 	return ret;
395 }
396 
397 void mlx5_ib_destroy_srq(struct ib_srq *srq, struct ib_udata *udata)
398 {
399 	struct mlx5_ib_dev *dev = to_mdev(srq->device);
400 	struct mlx5_ib_srq *msrq = to_msrq(srq);
401 
402 	mlx5_core_destroy_srq(dev->mdev, &msrq->msrq);
403 
404 	if (srq->uobject) {
405 		mlx5_ib_db_unmap_user(
406 			rdma_udata_to_drv_context(
407 				udata,
408 				struct mlx5_ib_ucontext,
409 				ibucontext),
410 			&msrq->db);
411 		ib_umem_release(msrq->umem);
412 	} else {
413 		destroy_srq_kernel(dev, msrq);
414 	}
415 }
416 
417 void mlx5_ib_free_srq_wqe(struct mlx5_ib_srq *srq, int wqe_index)
418 {
419 	struct mlx5_wqe_srq_next_seg *next;
420 
421 	/* always called with interrupts disabled. */
422 	spin_lock(&srq->lock);
423 
424 	next = get_wqe(srq, srq->tail);
425 	next->next_wqe_index = cpu_to_be16(wqe_index);
426 	srq->tail = wqe_index;
427 
428 	spin_unlock(&srq->lock);
429 }
430 
431 int mlx5_ib_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
432 			  const struct ib_recv_wr **bad_wr)
433 {
434 	struct mlx5_ib_srq *srq = to_msrq(ibsrq);
435 	struct mlx5_wqe_srq_next_seg *next;
436 	struct mlx5_wqe_data_seg *scat;
437 	struct mlx5_ib_dev *dev = to_mdev(ibsrq->device);
438 	struct mlx5_core_dev *mdev = dev->mdev;
439 	unsigned long flags;
440 	int err = 0;
441 	int nreq;
442 	int i;
443 
444 	spin_lock_irqsave(&srq->lock, flags);
445 
446 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
447 		err = -EIO;
448 		*bad_wr = wr;
449 		goto out;
450 	}
451 
452 	for (nreq = 0; wr; nreq++, wr = wr->next) {
453 		if (unlikely(wr->num_sge > srq->msrq.max_gs)) {
454 			err = -EINVAL;
455 			*bad_wr = wr;
456 			break;
457 		}
458 
459 		if (unlikely(srq->head == srq->tail)) {
460 			err = -ENOMEM;
461 			*bad_wr = wr;
462 			break;
463 		}
464 
465 		srq->wrid[srq->head] = wr->wr_id;
466 
467 		next      = get_wqe(srq, srq->head);
468 		srq->head = be16_to_cpu(next->next_wqe_index);
469 		scat      = (struct mlx5_wqe_data_seg *)(next + 1);
470 
471 		for (i = 0; i < wr->num_sge; i++) {
472 			scat[i].byte_count = cpu_to_be32(wr->sg_list[i].length);
473 			scat[i].lkey       = cpu_to_be32(wr->sg_list[i].lkey);
474 			scat[i].addr       = cpu_to_be64(wr->sg_list[i].addr);
475 		}
476 
477 		if (i < srq->msrq.max_avail_gather) {
478 			scat[i].byte_count = 0;
479 			scat[i].lkey       = cpu_to_be32(MLX5_INVALID_LKEY);
480 			scat[i].addr       = 0;
481 		}
482 	}
483 
484 	if (likely(nreq)) {
485 		srq->wqe_ctr += nreq;
486 
487 		/* Make sure that descriptors are written before
488 		 * doorbell record.
489 		 */
490 		wmb();
491 
492 		*srq->db.db = cpu_to_be32(srq->wqe_ctr);
493 	}
494 out:
495 	spin_unlock_irqrestore(&srq->lock, flags);
496 
497 	return err;
498 }
499