xref: /openbmc/linux/drivers/infiniband/hw/mlx5/srq.c (revision a06c488d)
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 {
80 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
81 	struct mlx5_ib_create_srq ucmd = {};
82 	size_t ucmdlen;
83 	void *xsrqc;
84 	int err;
85 	int npages;
86 	int page_shift;
87 	int ncont;
88 	u32 offset;
89 	u32 uidx = MLX5_IB_DEFAULT_UIDX;
90 	int drv_data = udata->inlen - sizeof(struct ib_uverbs_cmd_hdr);
91 
92 	if (drv_data < 0)
93 		return -EINVAL;
94 
95 	ucmdlen = (drv_data < sizeof(ucmd)) ?
96 		  drv_data : sizeof(ucmd);
97 
98 	if (ib_copy_from_udata(&ucmd, udata, ucmdlen)) {
99 		mlx5_ib_dbg(dev, "failed copy udata\n");
100 		return -EFAULT;
101 	}
102 
103 	if (ucmd.reserved0 || ucmd.reserved1)
104 		return -EINVAL;
105 
106 	if (drv_data > sizeof(ucmd) &&
107 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
108 				 drv_data - sizeof(ucmd)))
109 		return -EINVAL;
110 
111 	err = get_srq_user_index(to_mucontext(pd->uobject->context),
112 				 &ucmd, udata->inlen, &uidx);
113 	if (err)
114 		return err;
115 
116 	srq->wq_sig = !!(ucmd.flags & MLX5_SRQ_FLAG_SIGNATURE);
117 
118 	srq->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr, buf_size,
119 				0, 0);
120 	if (IS_ERR(srq->umem)) {
121 		mlx5_ib_dbg(dev, "failed umem get, size %d\n", buf_size);
122 		err = PTR_ERR(srq->umem);
123 		return err;
124 	}
125 
126 	mlx5_ib_cont_pages(srq->umem, ucmd.buf_addr, &npages,
127 			   &page_shift, &ncont, NULL);
128 	err = mlx5_ib_get_buf_offset(ucmd.buf_addr, page_shift,
129 				     &offset);
130 	if (err) {
131 		mlx5_ib_warn(dev, "bad offset\n");
132 		goto err_umem;
133 	}
134 
135 	*inlen = sizeof(**in) + sizeof(*(*in)->pas) * ncont;
136 	*in = mlx5_vzalloc(*inlen);
137 	if (!(*in)) {
138 		err = -ENOMEM;
139 		goto err_umem;
140 	}
141 
142 	mlx5_ib_populate_pas(dev, srq->umem, page_shift, (*in)->pas, 0);
143 
144 	err = mlx5_ib_db_map_user(to_mucontext(pd->uobject->context),
145 				  ucmd.db_addr, &srq->db);
146 	if (err) {
147 		mlx5_ib_dbg(dev, "map doorbell failed\n");
148 		goto err_in;
149 	}
150 
151 	(*in)->ctx.log_pg_sz = page_shift - MLX5_ADAPTER_PAGE_SHIFT;
152 	(*in)->ctx.pgoff_cqn = cpu_to_be32(offset << 26);
153 
154 	if (MLX5_CAP_GEN(dev->mdev, cqe_version) == MLX5_CQE_VERSION_V1) {
155 		xsrqc = MLX5_ADDR_OF(create_xrc_srq_in, *in,
156 				     xrc_srq_context_entry);
157 		MLX5_SET(xrc_srqc, xsrqc, user_index, uidx);
158 	}
159 
160 	return 0;
161 
162 err_in:
163 	kvfree(*in);
164 
165 err_umem:
166 	ib_umem_release(srq->umem);
167 
168 	return err;
169 }
170 
171 static int create_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq,
172 			     struct mlx5_create_srq_mbox_in **in, int buf_size,
173 			     int *inlen)
174 {
175 	int err;
176 	int i;
177 	struct mlx5_wqe_srq_next_seg *next;
178 	int page_shift;
179 	int npages;
180 	void *xsrqc;
181 
182 	err = mlx5_db_alloc(dev->mdev, &srq->db);
183 	if (err) {
184 		mlx5_ib_warn(dev, "alloc dbell rec failed\n");
185 		return err;
186 	}
187 
188 	if (mlx5_buf_alloc(dev->mdev, buf_size, &srq->buf)) {
189 		mlx5_ib_dbg(dev, "buf alloc failed\n");
190 		err = -ENOMEM;
191 		goto err_db;
192 	}
193 	page_shift = srq->buf.page_shift;
194 
195 	srq->head    = 0;
196 	srq->tail    = srq->msrq.max - 1;
197 	srq->wqe_ctr = 0;
198 
199 	for (i = 0; i < srq->msrq.max; i++) {
200 		next = get_wqe(srq, i);
201 		next->next_wqe_index =
202 			cpu_to_be16((i + 1) & (srq->msrq.max - 1));
203 	}
204 
205 	npages = DIV_ROUND_UP(srq->buf.npages, 1 << (page_shift - PAGE_SHIFT));
206 	mlx5_ib_dbg(dev, "buf_size %d, page_shift %d, npages %d, calc npages %d\n",
207 		    buf_size, page_shift, srq->buf.npages, npages);
208 	*inlen = sizeof(**in) + sizeof(*(*in)->pas) * npages;
209 	*in = mlx5_vzalloc(*inlen);
210 	if (!*in) {
211 		err = -ENOMEM;
212 		goto err_buf;
213 	}
214 	mlx5_fill_page_array(&srq->buf, (*in)->pas);
215 
216 	srq->wrid = kmalloc(srq->msrq.max * sizeof(u64), GFP_KERNEL);
217 	if (!srq->wrid) {
218 		mlx5_ib_dbg(dev, "kmalloc failed %lu\n",
219 			    (unsigned long)(srq->msrq.max * sizeof(u64)));
220 		err = -ENOMEM;
221 		goto err_in;
222 	}
223 	srq->wq_sig = !!srq_signature;
224 
225 	(*in)->ctx.log_pg_sz = page_shift - MLX5_ADAPTER_PAGE_SHIFT;
226 
227 	if (MLX5_CAP_GEN(dev->mdev, cqe_version) == MLX5_CQE_VERSION_V1) {
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 	if (pd->uobject)
306 		err = create_srq_user(pd, srq, &in, udata, buf_size, &inlen);
307 	else
308 		err = create_srq_kernel(dev, srq, &in, buf_size, &inlen);
309 
310 	if (err) {
311 		mlx5_ib_warn(dev, "create srq %s failed, err %d\n",
312 			     pd->uobject ? "user" : "kernel", err);
313 		goto err_srq;
314 	}
315 
316 	is_xrc = (init_attr->srq_type == IB_SRQT_XRC);
317 	in->ctx.state_log_sz = ilog2(srq->msrq.max);
318 	flgs = ((srq->msrq.wqe_shift - 4) | (is_xrc << 5) | (srq->wq_sig << 7)) << 24;
319 	xrcdn = 0;
320 	if (is_xrc) {
321 		xrcdn = to_mxrcd(init_attr->ext.xrc.xrcd)->xrcdn;
322 		in->ctx.pgoff_cqn |= cpu_to_be32(to_mcq(init_attr->ext.xrc.cq)->mcq.cqn);
323 	} else if (init_attr->srq_type == IB_SRQT_BASIC) {
324 		xrcdn = to_mxrcd(dev->devr.x0)->xrcdn;
325 		in->ctx.pgoff_cqn |= cpu_to_be32(to_mcq(dev->devr.c0)->mcq.cqn);
326 	}
327 
328 	in->ctx.flags_xrcd = cpu_to_be32((flgs & 0xFF000000) | (xrcdn & 0xFFFFFF));
329 
330 	in->ctx.pd = cpu_to_be32(to_mpd(pd)->pdn);
331 	in->ctx.db_record = cpu_to_be64(srq->db.dma);
332 	err = mlx5_core_create_srq(dev->mdev, &srq->msrq, in, inlen, is_xrc);
333 	kvfree(in);
334 	if (err) {
335 		mlx5_ib_dbg(dev, "create SRQ failed, err %d\n", err);
336 		goto err_usr_kern_srq;
337 	}
338 
339 	mlx5_ib_dbg(dev, "create SRQ with srqn 0x%x\n", srq->msrq.srqn);
340 
341 	srq->msrq.event = mlx5_ib_srq_event;
342 	srq->ibsrq.ext.xrc.srq_num = srq->msrq.srqn;
343 
344 	if (pd->uobject)
345 		if (ib_copy_to_udata(udata, &srq->msrq.srqn, sizeof(__u32))) {
346 			mlx5_ib_dbg(dev, "copy to user failed\n");
347 			err = -EFAULT;
348 			goto err_core;
349 		}
350 
351 	init_attr->attr.max_wr = srq->msrq.max - 1;
352 
353 	return &srq->ibsrq;
354 
355 err_core:
356 	mlx5_core_destroy_srq(dev->mdev, &srq->msrq);
357 
358 err_usr_kern_srq:
359 	if (pd->uobject)
360 		destroy_srq_user(pd, srq);
361 	else
362 		destroy_srq_kernel(dev, srq);
363 
364 err_srq:
365 	kfree(srq);
366 
367 	return ERR_PTR(err);
368 }
369 
370 int mlx5_ib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
371 		       enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
372 {
373 	struct mlx5_ib_dev *dev = to_mdev(ibsrq->device);
374 	struct mlx5_ib_srq *srq = to_msrq(ibsrq);
375 	int ret;
376 
377 	/* We don't support resizing SRQs yet */
378 	if (attr_mask & IB_SRQ_MAX_WR)
379 		return -EINVAL;
380 
381 	if (attr_mask & IB_SRQ_LIMIT) {
382 		if (attr->srq_limit >= srq->msrq.max)
383 			return -EINVAL;
384 
385 		mutex_lock(&srq->mutex);
386 		ret = mlx5_core_arm_srq(dev->mdev, &srq->msrq, attr->srq_limit, 1);
387 		mutex_unlock(&srq->mutex);
388 
389 		if (ret)
390 			return ret;
391 	}
392 
393 	return 0;
394 }
395 
396 int mlx5_ib_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
397 {
398 	struct mlx5_ib_dev *dev = to_mdev(ibsrq->device);
399 	struct mlx5_ib_srq *srq = to_msrq(ibsrq);
400 	int ret;
401 	struct mlx5_query_srq_mbox_out *out;
402 
403 	out = kzalloc(sizeof(*out), GFP_KERNEL);
404 	if (!out)
405 		return -ENOMEM;
406 
407 	ret = mlx5_core_query_srq(dev->mdev, &srq->msrq, out);
408 	if (ret)
409 		goto out_box;
410 
411 	srq_attr->srq_limit = be16_to_cpu(out->ctx.lwm);
412 	srq_attr->max_wr    = srq->msrq.max - 1;
413 	srq_attr->max_sge   = srq->msrq.max_gs;
414 
415 out_box:
416 	kfree(out);
417 	return ret;
418 }
419 
420 int mlx5_ib_destroy_srq(struct ib_srq *srq)
421 {
422 	struct mlx5_ib_dev *dev = to_mdev(srq->device);
423 	struct mlx5_ib_srq *msrq = to_msrq(srq);
424 
425 	mlx5_core_destroy_srq(dev->mdev, &msrq->msrq);
426 
427 	if (srq->uobject) {
428 		mlx5_ib_db_unmap_user(to_mucontext(srq->uobject->context), &msrq->db);
429 		ib_umem_release(msrq->umem);
430 	} else {
431 		destroy_srq_kernel(dev, msrq);
432 	}
433 
434 	kfree(srq);
435 	return 0;
436 }
437 
438 void mlx5_ib_free_srq_wqe(struct mlx5_ib_srq *srq, int wqe_index)
439 {
440 	struct mlx5_wqe_srq_next_seg *next;
441 
442 	/* always called with interrupts disabled. */
443 	spin_lock(&srq->lock);
444 
445 	next = get_wqe(srq, srq->tail);
446 	next->next_wqe_index = cpu_to_be16(wqe_index);
447 	srq->tail = wqe_index;
448 
449 	spin_unlock(&srq->lock);
450 }
451 
452 int mlx5_ib_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
453 			  struct ib_recv_wr **bad_wr)
454 {
455 	struct mlx5_ib_srq *srq = to_msrq(ibsrq);
456 	struct mlx5_wqe_srq_next_seg *next;
457 	struct mlx5_wqe_data_seg *scat;
458 	unsigned long flags;
459 	int err = 0;
460 	int nreq;
461 	int i;
462 
463 	spin_lock_irqsave(&srq->lock, flags);
464 
465 	for (nreq = 0; wr; nreq++, wr = wr->next) {
466 		if (unlikely(wr->num_sge > srq->msrq.max_gs)) {
467 			err = -EINVAL;
468 			*bad_wr = wr;
469 			break;
470 		}
471 
472 		if (unlikely(srq->head == srq->tail)) {
473 			err = -ENOMEM;
474 			*bad_wr = wr;
475 			break;
476 		}
477 
478 		srq->wrid[srq->head] = wr->wr_id;
479 
480 		next      = get_wqe(srq, srq->head);
481 		srq->head = be16_to_cpu(next->next_wqe_index);
482 		scat      = (struct mlx5_wqe_data_seg *)(next + 1);
483 
484 		for (i = 0; i < wr->num_sge; i++) {
485 			scat[i].byte_count = cpu_to_be32(wr->sg_list[i].length);
486 			scat[i].lkey       = cpu_to_be32(wr->sg_list[i].lkey);
487 			scat[i].addr       = cpu_to_be64(wr->sg_list[i].addr);
488 		}
489 
490 		if (i < srq->msrq.max_avail_gather) {
491 			scat[i].byte_count = 0;
492 			scat[i].lkey       = cpu_to_be32(MLX5_INVALID_LKEY);
493 			scat[i].addr       = 0;
494 		}
495 	}
496 
497 	if (likely(nreq)) {
498 		srq->wqe_ctr += nreq;
499 
500 		/* Make sure that descriptors are written before
501 		 * doorbell record.
502 		 */
503 		wmb();
504 
505 		*srq->db.db = cpu_to_be32(srq->wqe_ctr);
506 	}
507 
508 	spin_unlock_irqrestore(&srq->lock, flags);
509 
510 	return err;
511 }
512