xref: /freebsd/sys/dev/mlx5/mlx5_ib/mlx5_ib_srq.c (revision 5f4c09dd85bff675e0ca63c55ea3c517e0fddfcc)
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 
26 #include "opt_rss.h"
27 #include "opt_ratelimit.h"
28 
29 #include <linux/module.h>
30 #include <dev/mlx5/qp.h>
31 #include <dev/mlx5/srq.h>
32 #include <linux/slab.h>
33 #include <rdma/ib_umem.h>
34 #include <rdma/ib_user_verbs.h>
35 #include <rdma/uverbs_ioctl.h>
36 
37 #include <dev/mlx5/mlx5_ib/mlx5_ib.h>
38 
39 static void *get_wqe(struct mlx5_ib_srq *srq, int n)
40 {
41 	return mlx5_buf_offset(&srq->buf, n << srq->msrq.wqe_shift);
42 }
43 
44 static void mlx5_ib_srq_event(struct mlx5_core_srq *srq, int type)
45 {
46 	struct ib_event event;
47 	struct ib_srq *ibsrq = &to_mibsrq(srq)->ibsrq;
48 
49 	if (ibsrq->event_handler) {
50 		event.device      = ibsrq->device;
51 		event.element.srq = ibsrq;
52 		switch (type) {
53 		case MLX5_EVENT_TYPE_SRQ_RQ_LIMIT:
54 			event.event = IB_EVENT_SRQ_LIMIT_REACHED;
55 			break;
56 		case MLX5_EVENT_TYPE_SRQ_CATAS_ERROR:
57 			event.event = IB_EVENT_SRQ_ERR;
58 			break;
59 		default:
60 			pr_warn("mlx5_ib: Unexpected event type %d on SRQ %06x\n",
61 				type, srq->srqn);
62 			return;
63 		}
64 
65 		ibsrq->event_handler(&event, ibsrq->srq_context);
66 	}
67 }
68 
69 static int create_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq,
70 			   struct mlx5_srq_attr *in,
71 			   struct ib_udata *udata, int buf_size)
72 {
73 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
74 	struct mlx5_ib_create_srq ucmd = {};
75 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
76 		udata, struct mlx5_ib_ucontext, ibucontext);
77 	size_t ucmdlen;
78 	int err;
79 	int npages;
80 	int page_shift;
81 	int ncont;
82 	u32 offset;
83 	u32 uidx = MLX5_IB_DEFAULT_UIDX;
84 
85 	ucmdlen = min(udata->inlen, sizeof(ucmd));
86 
87 	if (ib_copy_from_udata(&ucmd, udata, ucmdlen)) {
88 		mlx5_ib_dbg(dev, "failed copy udata\n");
89 		return -EFAULT;
90 	}
91 
92 	if (ucmd.reserved0 || ucmd.reserved1)
93 		return -EINVAL;
94 
95 	if (udata->inlen > sizeof(ucmd) &&
96 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
97 				 udata->inlen - sizeof(ucmd)))
98 		return -EINVAL;
99 
100 	if (in->type != IB_SRQT_BASIC) {
101 		err = get_srq_user_index(ucontext, &ucmd, udata->inlen, &uidx);
102 		if (err)
103 			return err;
104 	}
105 
106 	srq->wq_sig = !!(ucmd.flags & MLX5_SRQ_FLAG_SIGNATURE);
107 
108 	srq->umem = ib_umem_get(&ucontext->ibucontext, ucmd.buf_addr, buf_size, 0, 0);
109 	if (IS_ERR(srq->umem)) {
110 		mlx5_ib_dbg(dev, "failed umem get, size %d\n", buf_size);
111 		err = PTR_ERR(srq->umem);
112 		return err;
113 	}
114 
115 	mlx5_ib_cont_pages(srq->umem, ucmd.buf_addr, 0, &npages,
116 			   &page_shift, &ncont, NULL);
117 	err = mlx5_ib_get_buf_offset(ucmd.buf_addr, page_shift,
118 				     &offset);
119 	if (err) {
120 		mlx5_ib_warn(dev, "bad offset\n");
121 		goto err_umem;
122 	}
123 
124 	in->pas = mlx5_vzalloc(sizeof(*in->pas) * ncont);
125 	if (!in->pas) {
126 		err = -ENOMEM;
127 		goto err_umem;
128 	}
129 
130 	mlx5_ib_populate_pas(dev, srq->umem, page_shift, in->pas, 0);
131 
132 	err = mlx5_ib_db_map_user(ucontext, ucmd.db_addr, &srq->db);
133 	if (err) {
134 		mlx5_ib_dbg(dev, "map doorbell failed\n");
135 		goto err_in;
136 	}
137 
138 	in->log_page_size = page_shift - MLX5_ADAPTER_PAGE_SHIFT;
139 	in->page_offset = offset;
140 	in->uid = (in->type != IB_SRQT_XRC) ?  to_mpd(pd)->uid : 0;
141 	if (MLX5_CAP_GEN(dev->mdev, cqe_version) == MLX5_CQE_VERSION_V1 &&
142 	    in->type != IB_SRQT_BASIC)
143 		in->user_index = uidx;
144 
145 	return 0;
146 
147 err_in:
148 	kvfree(in->pas);
149 
150 err_umem:
151 	ib_umem_release(srq->umem);
152 
153 	return err;
154 }
155 
156 static int create_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq,
157 			     struct mlx5_srq_attr *in, int buf_size)
158 {
159 	int err;
160 	int i;
161 	struct mlx5_wqe_srq_next_seg *next;
162 
163 	err = mlx5_db_alloc(dev->mdev, &srq->db);
164 	if (err) {
165 		mlx5_ib_warn(dev, "alloc dbell rec failed\n");
166 		return err;
167 	}
168 
169 	if (mlx5_buf_alloc(dev->mdev, buf_size, 2 * PAGE_SIZE, &srq->buf)) {
170 		mlx5_ib_dbg(dev, "buf alloc failed\n");
171 		err = -ENOMEM;
172 		goto err_db;
173 	}
174 
175 	srq->head    = 0;
176 	srq->tail    = srq->msrq.max - 1;
177 	srq->wqe_ctr = 0;
178 
179 	for (i = 0; i < srq->msrq.max; i++) {
180 		next = get_wqe(srq, i);
181 		next->next_wqe_index =
182 			cpu_to_be16((i + 1) & (srq->msrq.max - 1));
183 	}
184 
185 	mlx5_ib_dbg(dev, "srq->buf.page_shift = %d\n", srq->buf.page_shift);
186 	in->pas = mlx5_vzalloc(sizeof(*in->pas) * srq->buf.npages);
187 	if (!in->pas) {
188 		err = -ENOMEM;
189 		goto err_buf;
190 	}
191 	mlx5_fill_page_array(&srq->buf, in->pas);
192 
193 	srq->wrid = kmalloc(srq->msrq.max * sizeof(u64), GFP_KERNEL);
194 	if (!srq->wrid) {
195 		err = -ENOMEM;
196 		goto err_in;
197 	}
198 	srq->wq_sig = 0;
199 
200 	in->log_page_size = srq->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT;
201 	if (MLX5_CAP_GEN(dev->mdev, cqe_version) == MLX5_CQE_VERSION_V1 &&
202 	    in->type != IB_SRQT_BASIC)
203 		in->user_index = MLX5_IB_DEFAULT_UIDX;
204 
205 	return 0;
206 
207 err_in:
208 	kvfree(in->pas);
209 
210 err_buf:
211 	mlx5_buf_free(dev->mdev, &srq->buf);
212 
213 err_db:
214 	mlx5_db_free(dev->mdev, &srq->db);
215 	return err;
216 }
217 
218 static void destroy_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq,
219 			     struct ib_udata *udata)
220 {
221 	mlx5_ib_db_unmap_user(
222 		rdma_udata_to_drv_context(
223 			udata,
224 			struct mlx5_ib_ucontext,
225 			ibucontext),
226 		&srq->db);
227 	ib_umem_release(srq->umem);
228 }
229 
230 
231 static void destroy_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq)
232 {
233 	kfree(srq->wrid);
234 	mlx5_buf_free(dev->mdev, &srq->buf);
235 	mlx5_db_free(dev->mdev, &srq->db);
236 }
237 
238 int mlx5_ib_create_srq(struct ib_srq *ib_srq,
239 		       struct ib_srq_init_attr *init_attr,
240 		       struct ib_udata *udata)
241 {
242 	struct mlx5_ib_dev *dev = to_mdev(ib_srq->device);
243 	struct mlx5_ib_srq *srq = to_msrq(ib_srq);
244 	size_t desc_size;
245 	size_t buf_size;
246 	int err;
247 	struct mlx5_srq_attr in = {};
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 -EINVAL;
256 	}
257 
258 	mutex_init(&srq->mutex);
259 	spin_lock_init(&srq->lock);
260 	srq->msrq.max    = roundup_pow_of_two(init_attr->attr.max_wr + 1);
261 	srq->msrq.max_gs = init_attr->attr.max_sge;
262 
263 	desc_size = sizeof(struct mlx5_wqe_srq_next_seg) +
264 		    srq->msrq.max_gs * sizeof(struct mlx5_wqe_data_seg);
265 	if (desc_size == 0 || srq->msrq.max_gs > desc_size)
266 		return -EINVAL;
267 
268 	desc_size = roundup_pow_of_two(desc_size);
269 	desc_size = max_t(size_t, 32, desc_size);
270 	if (desc_size < sizeof(struct mlx5_wqe_srq_next_seg))
271 		return -EINVAL;
272 
273 	srq->msrq.max_avail_gather = (desc_size - sizeof(struct mlx5_wqe_srq_next_seg)) /
274 		sizeof(struct mlx5_wqe_data_seg);
275 	srq->msrq.wqe_shift = ilog2(desc_size);
276 	buf_size = srq->msrq.max * desc_size;
277 	if (buf_size < desc_size)
278 		return -EINVAL;
279 
280 	in.type = init_attr->srq_type;
281 
282 	if (udata)
283 		err = create_srq_user(ib_srq->pd, srq, &in, udata, buf_size);
284 	else
285 		err = create_srq_kernel(dev, srq, &in, buf_size);
286 
287 	if (err) {
288 		mlx5_ib_warn(dev, "create srq %s failed, err %d\n",
289 			     udata ? "user" : "kernel", err);
290 		return err;
291 	}
292 
293 	in.log_size = ilog2(srq->msrq.max);
294 	in.wqe_shift = srq->msrq.wqe_shift - 4;
295 	if (srq->wq_sig)
296 		in.flags |= MLX5_SRQ_FLAG_WQ_SIG;
297 
298 	if (init_attr->srq_type == IB_SRQT_XRC)
299 		in.xrcd = to_mxrcd(init_attr->ext.xrc.xrcd)->xrcdn;
300 	else
301 		in.xrcd = to_mxrcd(dev->devr.x0)->xrcdn;
302 
303 	if (ib_srq_has_cq(init_attr->srq_type))
304 		in.cqn = to_mcq(init_attr->ext.cq)->mcq.cqn;
305 	else
306 		in.cqn = to_mcq(dev->devr.c0)->mcq.cqn;
307 
308 	in.pd = to_mpd(ib_srq->pd)->pdn;
309 	in.db_record = srq->db.dma;
310 	err = mlx5_core_create_srq(dev->mdev, &srq->msrq, &in);
311 	kvfree(in.pas);
312 	if (err) {
313 		mlx5_ib_dbg(dev, "create SRQ failed, err %d\n", err);
314 		goto err_usr_kern_srq;
315 	}
316 
317 	mlx5_ib_dbg(dev, "create SRQ with srqn 0x%x\n", srq->msrq.srqn);
318 
319 	srq->msrq.event = mlx5_ib_srq_event;
320 	srq->ibsrq.ext.xrc.srq_num = srq->msrq.srqn;
321 
322 	if (udata)
323 		if (ib_copy_to_udata(udata, &srq->msrq.srqn, sizeof(__u32))) {
324 			mlx5_ib_dbg(dev, "copy to user failed\n");
325 			err = -EFAULT;
326 			goto err_core;
327 		}
328 
329 	init_attr->attr.max_wr = srq->msrq.max - 1;
330 
331 	return 0;
332 
333 err_core:
334 	mlx5_core_destroy_srq(dev->mdev, &srq->msrq);
335 
336 err_usr_kern_srq:
337 	if (udata)
338 		destroy_srq_user(ib_srq->pd, srq, udata);
339 	else
340 		destroy_srq_kernel(dev, srq);
341 
342 	return err;
343 }
344 
345 int mlx5_ib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
346 		       enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
347 {
348 	struct mlx5_ib_dev *dev = to_mdev(ibsrq->device);
349 	struct mlx5_ib_srq *srq = to_msrq(ibsrq);
350 	int ret;
351 
352 	/* We don't support resizing SRQs yet */
353 	if (attr_mask & IB_SRQ_MAX_WR)
354 		return -EINVAL;
355 
356 	if (attr_mask & IB_SRQ_LIMIT) {
357 		if (attr->srq_limit >= srq->msrq.max)
358 			return -EINVAL;
359 
360 		mutex_lock(&srq->mutex);
361 		ret = mlx5_core_arm_srq(dev->mdev, &srq->msrq, attr->srq_limit, 1);
362 		mutex_unlock(&srq->mutex);
363 
364 		if (ret)
365 			return ret;
366 	}
367 
368 	return 0;
369 }
370 
371 int mlx5_ib_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
372 {
373 	struct mlx5_ib_dev *dev = to_mdev(ibsrq->device);
374 	struct mlx5_ib_srq *srq = to_msrq(ibsrq);
375 	int ret;
376 	struct mlx5_srq_attr *out;
377 
378 	out = kzalloc(sizeof(*out), GFP_KERNEL);
379 	if (!out)
380 		return -ENOMEM;
381 
382 	ret = mlx5_core_query_srq(dev->mdev, &srq->msrq, out);
383 	if (ret)
384 		goto out_box;
385 
386 	srq_attr->srq_limit = out->lwm;
387 	srq_attr->max_wr    = srq->msrq.max - 1;
388 	srq_attr->max_sge   = srq->msrq.max_gs;
389 
390 out_box:
391 	kfree(out);
392 	return ret;
393 }
394 
395 void mlx5_ib_destroy_srq(struct ib_srq *srq, struct ib_udata *udata)
396 {
397 	struct mlx5_ib_dev *dev = to_mdev(srq->device);
398 	struct mlx5_ib_srq *msrq = to_msrq(srq);
399 
400 	mlx5_core_destroy_srq(dev->mdev, &msrq->msrq);
401 
402 	if (srq->uobject) {
403 		mlx5_ib_db_unmap_user(
404 			rdma_udata_to_drv_context(
405 				udata,
406 				struct mlx5_ib_ucontext,
407 				ibucontext),
408 			&msrq->db);
409 		ib_umem_release(msrq->umem);
410 	} else {
411 		destroy_srq_kernel(dev, msrq);
412 	}
413 }
414 
415 void mlx5_ib_free_srq_wqe(struct mlx5_ib_srq *srq, int wqe_index)
416 {
417 	struct mlx5_wqe_srq_next_seg *next;
418 
419 	/* always called with interrupts disabled. */
420 	spin_lock(&srq->lock);
421 
422 	next = get_wqe(srq, srq->tail);
423 	next->next_wqe_index = cpu_to_be16(wqe_index);
424 	srq->tail = wqe_index;
425 
426 	spin_unlock(&srq->lock);
427 }
428 
429 int mlx5_ib_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
430 			  const struct ib_recv_wr **bad_wr)
431 {
432 	struct mlx5_ib_srq *srq = to_msrq(ibsrq);
433 	struct mlx5_wqe_srq_next_seg *next;
434 	struct mlx5_wqe_data_seg *scat;
435 	struct mlx5_ib_dev *dev = to_mdev(ibsrq->device);
436 	struct mlx5_core_dev *mdev = dev->mdev;
437 	unsigned long flags;
438 	int err = 0;
439 	int nreq;
440 	int i;
441 
442 	spin_lock_irqsave(&srq->lock, flags);
443 
444 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
445 		err = -EIO;
446 		*bad_wr = wr;
447 		goto out;
448 	}
449 
450 	for (nreq = 0; wr; nreq++, wr = wr->next) {
451 		if (unlikely(wr->num_sge > srq->msrq.max_gs)) {
452 			err = -EINVAL;
453 			*bad_wr = wr;
454 			break;
455 		}
456 
457 		if (unlikely(srq->head == srq->tail)) {
458 			err = -ENOMEM;
459 			*bad_wr = wr;
460 			break;
461 		}
462 
463 		srq->wrid[srq->head] = wr->wr_id;
464 
465 		next      = get_wqe(srq, srq->head);
466 		srq->head = be16_to_cpu(next->next_wqe_index);
467 		scat      = (struct mlx5_wqe_data_seg *)(next + 1);
468 
469 		for (i = 0; i < wr->num_sge; i++) {
470 			scat[i].byte_count = cpu_to_be32(wr->sg_list[i].length);
471 			scat[i].lkey       = cpu_to_be32(wr->sg_list[i].lkey);
472 			scat[i].addr       = cpu_to_be64(wr->sg_list[i].addr);
473 		}
474 
475 		if (i < srq->msrq.max_avail_gather) {
476 			scat[i].byte_count = 0;
477 			scat[i].lkey       = cpu_to_be32(MLX5_INVALID_LKEY);
478 			scat[i].addr       = 0;
479 		}
480 	}
481 
482 	if (likely(nreq)) {
483 		srq->wqe_ctr += nreq;
484 
485 		/* Make sure that descriptors are written before
486 		 * doorbell record.
487 		 */
488 		wmb();
489 
490 		*srq->db.db = cpu_to_be32(srq->wqe_ctr);
491 	}
492 out:
493 	spin_unlock_irqrestore(&srq->lock, flags);
494 
495 	return err;
496 }
497