xref: /openbmc/linux/drivers/infiniband/hw/mlx5/srq.c (revision 7587eb18)
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 #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 			   int is_xrc)
80 {
81 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
82 	struct mlx5_ib_create_srq ucmd = {};
83 	size_t ucmdlen;
84 	void *xsrqc;
85 	int err;
86 	int npages;
87 	int page_shift;
88 	int ncont;
89 	u32 offset;
90 	u32 uidx = MLX5_IB_DEFAULT_UIDX;
91 
92 	ucmdlen = min(udata->inlen, 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 (ucmd.reserved0 || ucmd.reserved1)
100 		return -EINVAL;
101 
102 	if (udata->inlen > sizeof(ucmd) &&
103 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
104 				 udata->inlen - sizeof(ucmd)))
105 		return -EINVAL;
106 
107 	if (is_xrc) {
108 		err = get_srq_user_index(to_mucontext(pd->uobject->context),
109 					 &ucmd, udata->inlen, &uidx);
110 		if (err)
111 			return err;
112 	}
113 
114 	srq->wq_sig = !!(ucmd.flags & MLX5_SRQ_FLAG_SIGNATURE);
115 
116 	srq->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr, buf_size,
117 				0, 0);
118 	if (IS_ERR(srq->umem)) {
119 		mlx5_ib_dbg(dev, "failed umem get, size %d\n", buf_size);
120 		err = PTR_ERR(srq->umem);
121 		return err;
122 	}
123 
124 	mlx5_ib_cont_pages(srq->umem, ucmd.buf_addr, &npages,
125 			   &page_shift, &ncont, NULL);
126 	err = mlx5_ib_get_buf_offset(ucmd.buf_addr, page_shift,
127 				     &offset);
128 	if (err) {
129 		mlx5_ib_warn(dev, "bad offset\n");
130 		goto err_umem;
131 	}
132 
133 	*inlen = sizeof(**in) + sizeof(*(*in)->pas) * ncont;
134 	*in = mlx5_vzalloc(*inlen);
135 	if (!(*in)) {
136 		err = -ENOMEM;
137 		goto err_umem;
138 	}
139 
140 	mlx5_ib_populate_pas(dev, srq->umem, page_shift, (*in)->pas, 0);
141 
142 	err = mlx5_ib_db_map_user(to_mucontext(pd->uobject->context),
143 				  ucmd.db_addr, &srq->db);
144 	if (err) {
145 		mlx5_ib_dbg(dev, "map doorbell failed\n");
146 		goto err_in;
147 	}
148 
149 	(*in)->ctx.log_pg_sz = page_shift - MLX5_ADAPTER_PAGE_SHIFT;
150 	(*in)->ctx.pgoff_cqn = cpu_to_be32(offset << 26);
151 
152 	if ((MLX5_CAP_GEN(dev->mdev, cqe_version) == MLX5_CQE_VERSION_V1) &&
153 	     is_xrc){
154 		xsrqc = MLX5_ADDR_OF(create_xrc_srq_in, *in,
155 				     xrc_srq_context_entry);
156 		MLX5_SET(xrc_srqc, xsrqc, user_index, uidx);
157 	}
158 
159 	return 0;
160 
161 err_in:
162 	kvfree(*in);
163 
164 err_umem:
165 	ib_umem_release(srq->umem);
166 
167 	return err;
168 }
169 
170 static int create_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq,
171 			     struct mlx5_create_srq_mbox_in **in, int buf_size,
172 			     int *inlen, int is_xrc)
173 {
174 	int err;
175 	int i;
176 	struct mlx5_wqe_srq_next_seg *next;
177 	int page_shift;
178 	int npages;
179 	void *xsrqc;
180 
181 	err = mlx5_db_alloc(dev->mdev, &srq->db);
182 	if (err) {
183 		mlx5_ib_warn(dev, "alloc dbell rec failed\n");
184 		return err;
185 	}
186 
187 	if (mlx5_buf_alloc(dev->mdev, buf_size, &srq->buf)) {
188 		mlx5_ib_dbg(dev, "buf alloc failed\n");
189 		err = -ENOMEM;
190 		goto err_db;
191 	}
192 	page_shift = srq->buf.page_shift;
193 
194 	srq->head    = 0;
195 	srq->tail    = srq->msrq.max - 1;
196 	srq->wqe_ctr = 0;
197 
198 	for (i = 0; i < srq->msrq.max; i++) {
199 		next = get_wqe(srq, i);
200 		next->next_wqe_index =
201 			cpu_to_be16((i + 1) & (srq->msrq.max - 1));
202 	}
203 
204 	npages = DIV_ROUND_UP(srq->buf.npages, 1 << (page_shift - PAGE_SHIFT));
205 	mlx5_ib_dbg(dev, "buf_size %d, page_shift %d, npages %d, calc npages %d\n",
206 		    buf_size, page_shift, srq->buf.npages, npages);
207 	*inlen = sizeof(**in) + sizeof(*(*in)->pas) * npages;
208 	*in = mlx5_vzalloc(*inlen);
209 	if (!*in) {
210 		err = -ENOMEM;
211 		goto err_buf;
212 	}
213 	mlx5_fill_page_array(&srq->buf, (*in)->pas);
214 
215 	srq->wrid = kmalloc(srq->msrq.max * sizeof(u64), GFP_KERNEL);
216 	if (!srq->wrid) {
217 		mlx5_ib_dbg(dev, "kmalloc failed %lu\n",
218 			    (unsigned long)(srq->msrq.max * sizeof(u64)));
219 		err = -ENOMEM;
220 		goto err_in;
221 	}
222 	srq->wq_sig = !!srq_signature;
223 
224 	(*in)->ctx.log_pg_sz = page_shift - MLX5_ADAPTER_PAGE_SHIFT;
225 
226 	if ((MLX5_CAP_GEN(dev->mdev, cqe_version) == MLX5_CQE_VERSION_V1) &&
227 	     is_xrc){
228 		xsrqc = MLX5_ADDR_OF(create_xrc_srq_in, *in,
229 				     xrc_srq_context_entry);
230 		/* 0xffffff means we ask to work with cqe version 0 */
231 		MLX5_SET(xrc_srqc, xsrqc, user_index, MLX5_IB_DEFAULT_UIDX);
232 	}
233 
234 	return 0;
235 
236 err_in:
237 	kvfree(*in);
238 
239 err_buf:
240 	mlx5_buf_free(dev->mdev, &srq->buf);
241 
242 err_db:
243 	mlx5_db_free(dev->mdev, &srq->db);
244 	return err;
245 }
246 
247 static void destroy_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq)
248 {
249 	mlx5_ib_db_unmap_user(to_mucontext(pd->uobject->context), &srq->db);
250 	ib_umem_release(srq->umem);
251 }
252 
253 
254 static void destroy_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq)
255 {
256 	kfree(srq->wrid);
257 	mlx5_buf_free(dev->mdev, &srq->buf);
258 	mlx5_db_free(dev->mdev, &srq->db);
259 }
260 
261 struct ib_srq *mlx5_ib_create_srq(struct ib_pd *pd,
262 				  struct ib_srq_init_attr *init_attr,
263 				  struct ib_udata *udata)
264 {
265 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
266 	struct mlx5_ib_srq *srq;
267 	int desc_size;
268 	int buf_size;
269 	int err;
270 	struct mlx5_create_srq_mbox_in *uninitialized_var(in);
271 	int uninitialized_var(inlen);
272 	int is_xrc;
273 	u32 flgs, xrcdn;
274 	__u32 max_srq_wqes = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
275 
276 	/* Sanity check SRQ size before proceeding */
277 	if (init_attr->attr.max_wr >= max_srq_wqes) {
278 		mlx5_ib_dbg(dev, "max_wr %d, cap %d\n",
279 			    init_attr->attr.max_wr,
280 			    max_srq_wqes);
281 		return ERR_PTR(-EINVAL);
282 	}
283 
284 	srq = kmalloc(sizeof(*srq), GFP_KERNEL);
285 	if (!srq)
286 		return ERR_PTR(-ENOMEM);
287 
288 	mutex_init(&srq->mutex);
289 	spin_lock_init(&srq->lock);
290 	srq->msrq.max    = roundup_pow_of_two(init_attr->attr.max_wr + 1);
291 	srq->msrq.max_gs = init_attr->attr.max_sge;
292 
293 	desc_size = sizeof(struct mlx5_wqe_srq_next_seg) +
294 		    srq->msrq.max_gs * sizeof(struct mlx5_wqe_data_seg);
295 	desc_size = roundup_pow_of_two(desc_size);
296 	desc_size = max_t(int, 32, desc_size);
297 	srq->msrq.max_avail_gather = (desc_size - sizeof(struct mlx5_wqe_srq_next_seg)) /
298 		sizeof(struct mlx5_wqe_data_seg);
299 	srq->msrq.wqe_shift = ilog2(desc_size);
300 	buf_size = srq->msrq.max * desc_size;
301 	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",
302 		    desc_size, init_attr->attr.max_wr, srq->msrq.max, srq->msrq.max_gs,
303 		    srq->msrq.max_avail_gather);
304 
305 	is_xrc = (init_attr->srq_type == IB_SRQT_XRC);
306 
307 	if (pd->uobject)
308 		err = create_srq_user(pd, srq, &in, udata, buf_size, &inlen,
309 				      is_xrc);
310 	else
311 		err = create_srq_kernel(dev, srq, &in, buf_size, &inlen,
312 					is_xrc);
313 
314 	if (err) {
315 		mlx5_ib_warn(dev, "create srq %s failed, err %d\n",
316 			     pd->uobject ? "user" : "kernel", err);
317 		goto err_srq;
318 	}
319 
320 	in->ctx.state_log_sz = ilog2(srq->msrq.max);
321 	flgs = ((srq->msrq.wqe_shift - 4) | (is_xrc << 5) | (srq->wq_sig << 7)) << 24;
322 	xrcdn = 0;
323 	if (is_xrc) {
324 		xrcdn = to_mxrcd(init_attr->ext.xrc.xrcd)->xrcdn;
325 		in->ctx.pgoff_cqn |= cpu_to_be32(to_mcq(init_attr->ext.xrc.cq)->mcq.cqn);
326 	} else if (init_attr->srq_type == IB_SRQT_BASIC) {
327 		xrcdn = to_mxrcd(dev->devr.x0)->xrcdn;
328 		in->ctx.pgoff_cqn |= cpu_to_be32(to_mcq(dev->devr.c0)->mcq.cqn);
329 	}
330 
331 	in->ctx.flags_xrcd = cpu_to_be32((flgs & 0xFF000000) | (xrcdn & 0xFFFFFF));
332 
333 	in->ctx.pd = cpu_to_be32(to_mpd(pd)->pdn);
334 	in->ctx.db_record = cpu_to_be64(srq->db.dma);
335 	err = mlx5_core_create_srq(dev->mdev, &srq->msrq, in, inlen, is_xrc);
336 	kvfree(in);
337 	if (err) {
338 		mlx5_ib_dbg(dev, "create SRQ failed, err %d\n", err);
339 		goto err_usr_kern_srq;
340 	}
341 
342 	mlx5_ib_dbg(dev, "create SRQ with srqn 0x%x\n", srq->msrq.srqn);
343 
344 	srq->msrq.event = mlx5_ib_srq_event;
345 	srq->ibsrq.ext.xrc.srq_num = srq->msrq.srqn;
346 
347 	if (pd->uobject)
348 		if (ib_copy_to_udata(udata, &srq->msrq.srqn, sizeof(__u32))) {
349 			mlx5_ib_dbg(dev, "copy to user failed\n");
350 			err = -EFAULT;
351 			goto err_core;
352 		}
353 
354 	init_attr->attr.max_wr = srq->msrq.max - 1;
355 
356 	return &srq->ibsrq;
357 
358 err_core:
359 	mlx5_core_destroy_srq(dev->mdev, &srq->msrq);
360 
361 err_usr_kern_srq:
362 	if (pd->uobject)
363 		destroy_srq_user(pd, srq);
364 	else
365 		destroy_srq_kernel(dev, srq);
366 
367 err_srq:
368 	kfree(srq);
369 
370 	return ERR_PTR(err);
371 }
372 
373 int mlx5_ib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
374 		       enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
375 {
376 	struct mlx5_ib_dev *dev = to_mdev(ibsrq->device);
377 	struct mlx5_ib_srq *srq = to_msrq(ibsrq);
378 	int ret;
379 
380 	/* We don't support resizing SRQs yet */
381 	if (attr_mask & IB_SRQ_MAX_WR)
382 		return -EINVAL;
383 
384 	if (attr_mask & IB_SRQ_LIMIT) {
385 		if (attr->srq_limit >= srq->msrq.max)
386 			return -EINVAL;
387 
388 		mutex_lock(&srq->mutex);
389 		ret = mlx5_core_arm_srq(dev->mdev, &srq->msrq, attr->srq_limit, 1);
390 		mutex_unlock(&srq->mutex);
391 
392 		if (ret)
393 			return ret;
394 	}
395 
396 	return 0;
397 }
398 
399 int mlx5_ib_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
400 {
401 	struct mlx5_ib_dev *dev = to_mdev(ibsrq->device);
402 	struct mlx5_ib_srq *srq = to_msrq(ibsrq);
403 	int ret;
404 	struct mlx5_query_srq_mbox_out *out;
405 
406 	out = kzalloc(sizeof(*out), GFP_KERNEL);
407 	if (!out)
408 		return -ENOMEM;
409 
410 	ret = mlx5_core_query_srq(dev->mdev, &srq->msrq, out);
411 	if (ret)
412 		goto out_box;
413 
414 	srq_attr->srq_limit = be16_to_cpu(out->ctx.lwm);
415 	srq_attr->max_wr    = srq->msrq.max - 1;
416 	srq_attr->max_sge   = srq->msrq.max_gs;
417 
418 out_box:
419 	kfree(out);
420 	return ret;
421 }
422 
423 int mlx5_ib_destroy_srq(struct ib_srq *srq)
424 {
425 	struct mlx5_ib_dev *dev = to_mdev(srq->device);
426 	struct mlx5_ib_srq *msrq = to_msrq(srq);
427 
428 	mlx5_core_destroy_srq(dev->mdev, &msrq->msrq);
429 
430 	if (srq->uobject) {
431 		mlx5_ib_db_unmap_user(to_mucontext(srq->uobject->context), &msrq->db);
432 		ib_umem_release(msrq->umem);
433 	} else {
434 		destroy_srq_kernel(dev, msrq);
435 	}
436 
437 	kfree(srq);
438 	return 0;
439 }
440 
441 void mlx5_ib_free_srq_wqe(struct mlx5_ib_srq *srq, int wqe_index)
442 {
443 	struct mlx5_wqe_srq_next_seg *next;
444 
445 	/* always called with interrupts disabled. */
446 	spin_lock(&srq->lock);
447 
448 	next = get_wqe(srq, srq->tail);
449 	next->next_wqe_index = cpu_to_be16(wqe_index);
450 	srq->tail = wqe_index;
451 
452 	spin_unlock(&srq->lock);
453 }
454 
455 int mlx5_ib_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
456 			  struct ib_recv_wr **bad_wr)
457 {
458 	struct mlx5_ib_srq *srq = to_msrq(ibsrq);
459 	struct mlx5_wqe_srq_next_seg *next;
460 	struct mlx5_wqe_data_seg *scat;
461 	unsigned long flags;
462 	int err = 0;
463 	int nreq;
464 	int i;
465 
466 	spin_lock_irqsave(&srq->lock, flags);
467 
468 	for (nreq = 0; wr; nreq++, wr = wr->next) {
469 		if (unlikely(wr->num_sge > srq->msrq.max_gs)) {
470 			err = -EINVAL;
471 			*bad_wr = wr;
472 			break;
473 		}
474 
475 		if (unlikely(srq->head == srq->tail)) {
476 			err = -ENOMEM;
477 			*bad_wr = wr;
478 			break;
479 		}
480 
481 		srq->wrid[srq->head] = wr->wr_id;
482 
483 		next      = get_wqe(srq, srq->head);
484 		srq->head = be16_to_cpu(next->next_wqe_index);
485 		scat      = (struct mlx5_wqe_data_seg *)(next + 1);
486 
487 		for (i = 0; i < wr->num_sge; i++) {
488 			scat[i].byte_count = cpu_to_be32(wr->sg_list[i].length);
489 			scat[i].lkey       = cpu_to_be32(wr->sg_list[i].lkey);
490 			scat[i].addr       = cpu_to_be64(wr->sg_list[i].addr);
491 		}
492 
493 		if (i < srq->msrq.max_avail_gather) {
494 			scat[i].byte_count = 0;
495 			scat[i].lkey       = cpu_to_be32(MLX5_INVALID_LKEY);
496 			scat[i].addr       = 0;
497 		}
498 	}
499 
500 	if (likely(nreq)) {
501 		srq->wqe_ctr += nreq;
502 
503 		/* Make sure that descriptors are written before
504 		 * doorbell record.
505 		 */
506 		wmb();
507 
508 		*srq->db.db = cpu_to_be32(srq->wqe_ctr);
509 	}
510 
511 	spin_unlock_irqrestore(&srq->lock, flags);
512 
513 	return err;
514 }
515