xref: /linux/drivers/infiniband/hw/mlx5/srq.c (revision 93d90ad708b8da6efc0e487b66111aa9db7f70c7)
1 /*
2  * Copyright (c) 2013, Mellanox Technologies inc.  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 #include "user.h"
42 
43 /* not supported currently */
44 static int srq_signature;
45 
46 static void *get_wqe(struct mlx5_ib_srq *srq, int n)
47 {
48 	return mlx5_buf_offset(&srq->buf, n << srq->msrq.wqe_shift);
49 }
50 
51 static void mlx5_ib_srq_event(struct mlx5_core_srq *srq, enum mlx5_event type)
52 {
53 	struct ib_event event;
54 	struct ib_srq *ibsrq = &to_mibsrq(srq)->ibsrq;
55 
56 	if (ibsrq->event_handler) {
57 		event.device      = ibsrq->device;
58 		event.element.srq = ibsrq;
59 		switch (type) {
60 		case MLX5_EVENT_TYPE_SRQ_RQ_LIMIT:
61 			event.event = IB_EVENT_SRQ_LIMIT_REACHED;
62 			break;
63 		case MLX5_EVENT_TYPE_SRQ_CATAS_ERROR:
64 			event.event = IB_EVENT_SRQ_ERR;
65 			break;
66 		default:
67 			pr_warn("mlx5_ib: Unexpected event type %d on SRQ %06x\n",
68 				type, srq->srqn);
69 			return;
70 		}
71 
72 		ibsrq->event_handler(&event, ibsrq->srq_context);
73 	}
74 }
75 
76 static int create_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq,
77 			   struct mlx5_create_srq_mbox_in **in,
78 			   struct ib_udata *udata, int buf_size, int *inlen)
79 {
80 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
81 	struct mlx5_ib_create_srq ucmd;
82 	size_t ucmdlen;
83 	int err;
84 	int npages;
85 	int page_shift;
86 	int ncont;
87 	u32 offset;
88 
89 	ucmdlen =
90 		(udata->inlen - sizeof(struct ib_uverbs_cmd_hdr) <
91 		 sizeof(ucmd)) ? (sizeof(ucmd) -
92 				  sizeof(ucmd.reserved)) : sizeof(ucmd);
93 
94 	if (ib_copy_from_udata(&ucmd, udata, ucmdlen)) {
95 		mlx5_ib_dbg(dev, "failed copy udata\n");
96 		return -EFAULT;
97 	}
98 
99 	if (ucmdlen == sizeof(ucmd) &&
100 	    ucmd.reserved != 0)
101 		return -EINVAL;
102 
103 	srq->wq_sig = !!(ucmd.flags & MLX5_SRQ_FLAG_SIGNATURE);
104 
105 	srq->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr, buf_size,
106 				0, 0);
107 	if (IS_ERR(srq->umem)) {
108 		mlx5_ib_dbg(dev, "failed umem get, size %d\n", buf_size);
109 		err = PTR_ERR(srq->umem);
110 		return err;
111 	}
112 
113 	mlx5_ib_cont_pages(srq->umem, ucmd.buf_addr, &npages,
114 			   &page_shift, &ncont, NULL);
115 	err = mlx5_ib_get_buf_offset(ucmd.buf_addr, page_shift,
116 				     &offset);
117 	if (err) {
118 		mlx5_ib_warn(dev, "bad offset\n");
119 		goto err_umem;
120 	}
121 
122 	*inlen = sizeof(**in) + sizeof(*(*in)->pas) * ncont;
123 	*in = mlx5_vzalloc(*inlen);
124 	if (!(*in)) {
125 		err = -ENOMEM;
126 		goto err_umem;
127 	}
128 
129 	mlx5_ib_populate_pas(dev, srq->umem, page_shift, (*in)->pas, 0);
130 
131 	err = mlx5_ib_db_map_user(to_mucontext(pd->uobject->context),
132 				  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)->ctx.log_pg_sz = page_shift - MLX5_ADAPTER_PAGE_SHIFT;
139 	(*in)->ctx.pgoff_cqn = cpu_to_be32(offset << 26);
140 
141 	return 0;
142 
143 err_in:
144 	kvfree(*in);
145 
146 err_umem:
147 	ib_umem_release(srq->umem);
148 
149 	return err;
150 }
151 
152 static int create_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq,
153 			     struct mlx5_create_srq_mbox_in **in, int buf_size,
154 			     int *inlen)
155 {
156 	int err;
157 	int i;
158 	struct mlx5_wqe_srq_next_seg *next;
159 	int page_shift;
160 	int npages;
161 
162 	err = mlx5_db_alloc(dev->mdev, &srq->db);
163 	if (err) {
164 		mlx5_ib_warn(dev, "alloc dbell rec failed\n");
165 		return err;
166 	}
167 
168 	*srq->db.db = 0;
169 
170 	if (mlx5_buf_alloc(dev->mdev, buf_size, PAGE_SIZE * 2, &srq->buf)) {
171 		mlx5_ib_dbg(dev, "buf alloc failed\n");
172 		err = -ENOMEM;
173 		goto err_db;
174 	}
175 	page_shift = srq->buf.page_shift;
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 	npages = DIV_ROUND_UP(srq->buf.npages, 1 << (page_shift - PAGE_SHIFT));
188 	mlx5_ib_dbg(dev, "buf_size %d, page_shift %d, npages %d, calc npages %d\n",
189 		    buf_size, page_shift, srq->buf.npages, npages);
190 	*inlen = sizeof(**in) + sizeof(*(*in)->pas) * npages;
191 	*in = mlx5_vzalloc(*inlen);
192 	if (!*in) {
193 		err = -ENOMEM;
194 		goto err_buf;
195 	}
196 	mlx5_fill_page_array(&srq->buf, (*in)->pas);
197 
198 	srq->wrid = kmalloc(srq->msrq.max * sizeof(u64), GFP_KERNEL);
199 	if (!srq->wrid) {
200 		mlx5_ib_dbg(dev, "kmalloc failed %lu\n",
201 			    (unsigned long)(srq->msrq.max * sizeof(u64)));
202 		err = -ENOMEM;
203 		goto err_in;
204 	}
205 	srq->wq_sig = !!srq_signature;
206 
207 	(*in)->ctx.log_pg_sz = page_shift - MLX5_ADAPTER_PAGE_SHIFT;
208 
209 	return 0;
210 
211 err_in:
212 	kvfree(*in);
213 
214 err_buf:
215 	mlx5_buf_free(dev->mdev, &srq->buf);
216 
217 err_db:
218 	mlx5_db_free(dev->mdev, &srq->db);
219 	return err;
220 }
221 
222 static void destroy_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq)
223 {
224 	mlx5_ib_db_unmap_user(to_mucontext(pd->uobject->context), &srq->db);
225 	ib_umem_release(srq->umem);
226 }
227 
228 
229 static void destroy_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq)
230 {
231 	kfree(srq->wrid);
232 	mlx5_buf_free(dev->mdev, &srq->buf);
233 	mlx5_db_free(dev->mdev, &srq->db);
234 }
235 
236 struct ib_srq *mlx5_ib_create_srq(struct ib_pd *pd,
237 				  struct ib_srq_init_attr *init_attr,
238 				  struct ib_udata *udata)
239 {
240 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
241 	struct mlx5_general_caps *gen;
242 	struct mlx5_ib_srq *srq;
243 	int desc_size;
244 	int buf_size;
245 	int err;
246 	struct mlx5_create_srq_mbox_in *uninitialized_var(in);
247 	int uninitialized_var(inlen);
248 	int is_xrc;
249 	u32 flgs, xrcdn;
250 
251 	gen = &dev->mdev->caps.gen;
252 	/* Sanity check SRQ size before proceeding */
253 	if (init_attr->attr.max_wr >= gen->max_srq_wqes) {
254 		mlx5_ib_dbg(dev, "max_wr %d, cap %d\n",
255 			    init_attr->attr.max_wr,
256 			    gen->max_srq_wqes);
257 		return ERR_PTR(-EINVAL);
258 	}
259 
260 	srq = kmalloc(sizeof(*srq), GFP_KERNEL);
261 	if (!srq)
262 		return ERR_PTR(-ENOMEM);
263 
264 	mutex_init(&srq->mutex);
265 	spin_lock_init(&srq->lock);
266 	srq->msrq.max    = roundup_pow_of_two(init_attr->attr.max_wr + 1);
267 	srq->msrq.max_gs = init_attr->attr.max_sge;
268 
269 	desc_size = sizeof(struct mlx5_wqe_srq_next_seg) +
270 		    srq->msrq.max_gs * sizeof(struct mlx5_wqe_data_seg);
271 	desc_size = roundup_pow_of_two(desc_size);
272 	desc_size = max_t(int, 32, desc_size);
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 	mlx5_ib_dbg(dev, "desc_size 0x%x, req wr 0x%x, srq size 0x%x, max_gs 0x%x, max_avail_gather 0x%x\n",
278 		    desc_size, init_attr->attr.max_wr, srq->msrq.max, srq->msrq.max_gs,
279 		    srq->msrq.max_avail_gather);
280 
281 	if (pd->uobject)
282 		err = create_srq_user(pd, srq, &in, udata, buf_size, &inlen);
283 	else
284 		err = create_srq_kernel(dev, srq, &in, buf_size, &inlen);
285 
286 	if (err) {
287 		mlx5_ib_warn(dev, "create srq %s failed, err %d\n",
288 			     pd->uobject ? "user" : "kernel", err);
289 		goto err_srq;
290 	}
291 
292 	is_xrc = (init_attr->srq_type == IB_SRQT_XRC);
293 	in->ctx.state_log_sz = ilog2(srq->msrq.max);
294 	flgs = ((srq->msrq.wqe_shift - 4) | (is_xrc << 5) | (srq->wq_sig << 7)) << 24;
295 	xrcdn = 0;
296 	if (is_xrc) {
297 		xrcdn = to_mxrcd(init_attr->ext.xrc.xrcd)->xrcdn;
298 		in->ctx.pgoff_cqn |= cpu_to_be32(to_mcq(init_attr->ext.xrc.cq)->mcq.cqn);
299 	} else if (init_attr->srq_type == IB_SRQT_BASIC) {
300 		xrcdn = to_mxrcd(dev->devr.x0)->xrcdn;
301 		in->ctx.pgoff_cqn |= cpu_to_be32(to_mcq(dev->devr.c0)->mcq.cqn);
302 	}
303 
304 	in->ctx.flags_xrcd = cpu_to_be32((flgs & 0xFF000000) | (xrcdn & 0xFFFFFF));
305 
306 	in->ctx.pd = cpu_to_be32(to_mpd(pd)->pdn);
307 	in->ctx.db_record = cpu_to_be64(srq->db.dma);
308 	err = mlx5_core_create_srq(dev->mdev, &srq->msrq, in, inlen);
309 	kvfree(in);
310 	if (err) {
311 		mlx5_ib_dbg(dev, "create SRQ failed, err %d\n", err);
312 		goto err_usr_kern_srq;
313 	}
314 
315 	mlx5_ib_dbg(dev, "create SRQ with srqn 0x%x\n", srq->msrq.srqn);
316 
317 	srq->msrq.event = mlx5_ib_srq_event;
318 	srq->ibsrq.ext.xrc.srq_num = srq->msrq.srqn;
319 
320 	if (pd->uobject)
321 		if (ib_copy_to_udata(udata, &srq->msrq.srqn, sizeof(__u32))) {
322 			mlx5_ib_dbg(dev, "copy to user failed\n");
323 			err = -EFAULT;
324 			goto err_core;
325 		}
326 
327 	init_attr->attr.max_wr = srq->msrq.max - 1;
328 
329 	return &srq->ibsrq;
330 
331 err_core:
332 	mlx5_core_destroy_srq(dev->mdev, &srq->msrq);
333 
334 err_usr_kern_srq:
335 	if (pd->uobject)
336 		destroy_srq_user(pd, srq);
337 	else
338 		destroy_srq_kernel(dev, srq);
339 
340 err_srq:
341 	kfree(srq);
342 
343 	return ERR_PTR(err);
344 }
345 
346 int mlx5_ib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
347 		       enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
348 {
349 	struct mlx5_ib_dev *dev = to_mdev(ibsrq->device);
350 	struct mlx5_ib_srq *srq = to_msrq(ibsrq);
351 	int ret;
352 
353 	/* We don't support resizing SRQs yet */
354 	if (attr_mask & IB_SRQ_MAX_WR)
355 		return -EINVAL;
356 
357 	if (attr_mask & IB_SRQ_LIMIT) {
358 		if (attr->srq_limit >= srq->msrq.max)
359 			return -EINVAL;
360 
361 		mutex_lock(&srq->mutex);
362 		ret = mlx5_core_arm_srq(dev->mdev, &srq->msrq, attr->srq_limit, 1);
363 		mutex_unlock(&srq->mutex);
364 
365 		if (ret)
366 			return ret;
367 	}
368 
369 	return 0;
370 }
371 
372 int mlx5_ib_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
373 {
374 	struct mlx5_ib_dev *dev = to_mdev(ibsrq->device);
375 	struct mlx5_ib_srq *srq = to_msrq(ibsrq);
376 	int ret;
377 	struct mlx5_query_srq_mbox_out *out;
378 
379 	out = kzalloc(sizeof(*out), GFP_KERNEL);
380 	if (!out)
381 		return -ENOMEM;
382 
383 	ret = mlx5_core_query_srq(dev->mdev, &srq->msrq, out);
384 	if (ret)
385 		goto out_box;
386 
387 	srq_attr->srq_limit = be16_to_cpu(out->ctx.lwm);
388 	srq_attr->max_wr    = srq->msrq.max - 1;
389 	srq_attr->max_sge   = srq->msrq.max_gs;
390 
391 out_box:
392 	kfree(out);
393 	return ret;
394 }
395 
396 int mlx5_ib_destroy_srq(struct ib_srq *srq)
397 {
398 	struct mlx5_ib_dev *dev = to_mdev(srq->device);
399 	struct mlx5_ib_srq *msrq = to_msrq(srq);
400 
401 	mlx5_core_destroy_srq(dev->mdev, &msrq->msrq);
402 
403 	if (srq->uobject) {
404 		mlx5_ib_db_unmap_user(to_mucontext(srq->uobject->context), &msrq->db);
405 		ib_umem_release(msrq->umem);
406 	} else {
407 		destroy_srq_kernel(dev, msrq);
408 	}
409 
410 	kfree(srq);
411 	return 0;
412 }
413 
414 void mlx5_ib_free_srq_wqe(struct mlx5_ib_srq *srq, int wqe_index)
415 {
416 	struct mlx5_wqe_srq_next_seg *next;
417 
418 	/* always called with interrupts disabled. */
419 	spin_lock(&srq->lock);
420 
421 	next = get_wqe(srq, srq->tail);
422 	next->next_wqe_index = cpu_to_be16(wqe_index);
423 	srq->tail = wqe_index;
424 
425 	spin_unlock(&srq->lock);
426 }
427 
428 int mlx5_ib_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
429 			  struct ib_recv_wr **bad_wr)
430 {
431 	struct mlx5_ib_srq *srq = to_msrq(ibsrq);
432 	struct mlx5_wqe_srq_next_seg *next;
433 	struct mlx5_wqe_data_seg *scat;
434 	unsigned long flags;
435 	int err = 0;
436 	int nreq;
437 	int i;
438 
439 	spin_lock_irqsave(&srq->lock, flags);
440 
441 	for (nreq = 0; wr; nreq++, wr = wr->next) {
442 		if (unlikely(wr->num_sge > srq->msrq.max_gs)) {
443 			err = -EINVAL;
444 			*bad_wr = wr;
445 			break;
446 		}
447 
448 		if (unlikely(srq->head == srq->tail)) {
449 			err = -ENOMEM;
450 			*bad_wr = wr;
451 			break;
452 		}
453 
454 		srq->wrid[srq->head] = wr->wr_id;
455 
456 		next      = get_wqe(srq, srq->head);
457 		srq->head = be16_to_cpu(next->next_wqe_index);
458 		scat      = (struct mlx5_wqe_data_seg *)(next + 1);
459 
460 		for (i = 0; i < wr->num_sge; i++) {
461 			scat[i].byte_count = cpu_to_be32(wr->sg_list[i].length);
462 			scat[i].lkey       = cpu_to_be32(wr->sg_list[i].lkey);
463 			scat[i].addr       = cpu_to_be64(wr->sg_list[i].addr);
464 		}
465 
466 		if (i < srq->msrq.max_avail_gather) {
467 			scat[i].byte_count = 0;
468 			scat[i].lkey       = cpu_to_be32(MLX5_INVALID_LKEY);
469 			scat[i].addr       = 0;
470 		}
471 	}
472 
473 	if (likely(nreq)) {
474 		srq->wqe_ctr += nreq;
475 
476 		/* Make sure that descriptors are written before
477 		 * doorbell record.
478 		 */
479 		wmb();
480 
481 		*srq->db.db = cpu_to_be32(srq->wqe_ctr);
482 	}
483 
484 	spin_unlock_irqrestore(&srq->lock, flags);
485 
486 	return err;
487 }
488