xref: /linux/drivers/infiniband/hw/hns/hns_roce_srq.c (revision d53b8e36925256097a08d7cb749198d85cbf9b2b)
1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /*
3  * Copyright (c) 2018 Hisilicon Limited.
4  */
5 
6 #include <linux/pci.h>
7 #include <rdma/ib_umem.h>
8 #include <rdma/uverbs_ioctl.h>
9 #include "hns_roce_device.h"
10 #include "hns_roce_cmd.h"
11 #include "hns_roce_hem.h"
12 
13 void hns_roce_srq_event(struct hns_roce_dev *hr_dev, u32 srqn, int event_type)
14 {
15 	struct hns_roce_srq_table *srq_table = &hr_dev->srq_table;
16 	struct hns_roce_srq *srq;
17 
18 	xa_lock(&srq_table->xa);
19 	srq = xa_load(&srq_table->xa, srqn & (hr_dev->caps.num_srqs - 1));
20 	if (srq)
21 		refcount_inc(&srq->refcount);
22 	xa_unlock(&srq_table->xa);
23 
24 	if (!srq) {
25 		dev_warn(hr_dev->dev, "Async event for bogus SRQ %08x\n", srqn);
26 		return;
27 	}
28 
29 	srq->event(srq, event_type);
30 
31 	if (refcount_dec_and_test(&srq->refcount))
32 		complete(&srq->free);
33 }
34 
35 static void hns_roce_ib_srq_event(struct hns_roce_srq *srq,
36 				  enum hns_roce_event event_type)
37 {
38 	struct hns_roce_dev *hr_dev = to_hr_dev(srq->ibsrq.device);
39 	struct ib_srq *ibsrq = &srq->ibsrq;
40 	struct ib_event event;
41 
42 	if (ibsrq->event_handler) {
43 		event.device      = ibsrq->device;
44 		event.element.srq = ibsrq;
45 		switch (event_type) {
46 		case HNS_ROCE_EVENT_TYPE_SRQ_LIMIT_REACH:
47 			event.event = IB_EVENT_SRQ_LIMIT_REACHED;
48 			break;
49 		case HNS_ROCE_EVENT_TYPE_SRQ_CATAS_ERROR:
50 			event.event = IB_EVENT_SRQ_ERR;
51 			break;
52 		default:
53 			dev_err(hr_dev->dev,
54 			   "hns_roce:Unexpected event type 0x%x on SRQ %06lx\n",
55 			   event_type, srq->srqn);
56 			return;
57 		}
58 
59 		ibsrq->event_handler(&event, ibsrq->srq_context);
60 	}
61 }
62 
63 static int alloc_srqn(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq)
64 {
65 	struct hns_roce_ida *srq_ida = &hr_dev->srq_table.srq_ida;
66 	int id;
67 
68 	id = ida_alloc_range(&srq_ida->ida, srq_ida->min, srq_ida->max,
69 			     GFP_KERNEL);
70 	if (id < 0) {
71 		ibdev_err(&hr_dev->ib_dev, "failed to alloc srq(%d).\n", id);
72 		return -ENOMEM;
73 	}
74 
75 	srq->srqn = id;
76 
77 	return 0;
78 }
79 
80 static void free_srqn(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq)
81 {
82 	ida_free(&hr_dev->srq_table.srq_ida.ida, (int)srq->srqn);
83 }
84 
85 static int hns_roce_create_srqc(struct hns_roce_dev *hr_dev,
86 				struct hns_roce_srq *srq)
87 {
88 	struct ib_device *ibdev = &hr_dev->ib_dev;
89 	struct hns_roce_cmd_mailbox *mailbox;
90 	int ret;
91 
92 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
93 	if (IS_ERR(mailbox)) {
94 		ibdev_err(ibdev, "failed to alloc mailbox for SRQC.\n");
95 		return PTR_ERR(mailbox);
96 	}
97 
98 	ret = hr_dev->hw->write_srqc(srq, mailbox->buf);
99 	if (ret) {
100 		ibdev_err(ibdev, "failed to write SRQC.\n");
101 		goto err_mbox;
102 	}
103 
104 	ret = hns_roce_create_hw_ctx(hr_dev, mailbox, HNS_ROCE_CMD_CREATE_SRQ,
105 				     srq->srqn);
106 	if (ret)
107 		ibdev_err(ibdev, "failed to config SRQC, ret = %d.\n", ret);
108 
109 err_mbox:
110 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
111 	return ret;
112 }
113 
114 static int alloc_srqc(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq)
115 {
116 	struct hns_roce_srq_table *srq_table = &hr_dev->srq_table;
117 	struct ib_device *ibdev = &hr_dev->ib_dev;
118 	int ret;
119 
120 	ret = hns_roce_table_get(hr_dev, &srq_table->table, srq->srqn);
121 	if (ret) {
122 		ibdev_err(ibdev, "failed to get SRQC table, ret = %d.\n", ret);
123 		return ret;
124 	}
125 
126 	ret = xa_err(xa_store_irq(&srq_table->xa, srq->srqn, srq, GFP_KERNEL));
127 	if (ret) {
128 		ibdev_err(ibdev, "failed to store SRQC, ret = %d.\n", ret);
129 		goto err_put;
130 	}
131 
132 	ret = hns_roce_create_srqc(hr_dev, srq);
133 	if (ret)
134 		goto err_xa;
135 
136 	return 0;
137 
138 err_xa:
139 	xa_erase_irq(&srq_table->xa, srq->srqn);
140 err_put:
141 	hns_roce_table_put(hr_dev, &srq_table->table, srq->srqn);
142 
143 	return ret;
144 }
145 
146 static void free_srqc(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq)
147 {
148 	struct hns_roce_srq_table *srq_table = &hr_dev->srq_table;
149 	int ret;
150 
151 	ret = hns_roce_destroy_hw_ctx(hr_dev, HNS_ROCE_CMD_DESTROY_SRQ,
152 				      srq->srqn);
153 	if (ret)
154 		dev_err(hr_dev->dev, "DESTROY_SRQ failed (%d) for SRQN %06lx\n",
155 			ret, srq->srqn);
156 
157 	xa_erase_irq(&srq_table->xa, srq->srqn);
158 
159 	if (refcount_dec_and_test(&srq->refcount))
160 		complete(&srq->free);
161 	wait_for_completion(&srq->free);
162 
163 	hns_roce_table_put(hr_dev, &srq_table->table, srq->srqn);
164 }
165 
166 static int alloc_srq_idx(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq,
167 			 struct ib_udata *udata, unsigned long addr)
168 {
169 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
170 	struct ib_device *ibdev = &hr_dev->ib_dev;
171 	struct hns_roce_buf_attr buf_attr = {};
172 	int ret;
173 
174 	srq->idx_que.entry_shift = ilog2(HNS_ROCE_IDX_QUE_ENTRY_SZ);
175 
176 	buf_attr.page_shift = hr_dev->caps.idx_buf_pg_sz + PAGE_SHIFT;
177 	buf_attr.region[0].size = to_hr_hem_entries_size(srq->wqe_cnt,
178 					srq->idx_que.entry_shift);
179 	buf_attr.region[0].hopnum = hr_dev->caps.idx_hop_num;
180 	buf_attr.region_count = 1;
181 
182 	ret = hns_roce_mtr_create(hr_dev, &idx_que->mtr, &buf_attr,
183 				  hr_dev->caps.idx_ba_pg_sz + PAGE_SHIFT,
184 				  udata, addr);
185 	if (ret) {
186 		ibdev_err(ibdev,
187 			  "failed to alloc SRQ idx mtr, ret = %d.\n", ret);
188 		return ret;
189 	}
190 
191 	if (!udata) {
192 		idx_que->bitmap = bitmap_zalloc(srq->wqe_cnt, GFP_KERNEL);
193 		if (!idx_que->bitmap) {
194 			ibdev_err(ibdev, "failed to alloc SRQ idx bitmap.\n");
195 			ret = -ENOMEM;
196 			goto err_idx_mtr;
197 		}
198 	}
199 
200 	idx_que->head = 0;
201 	idx_que->tail = 0;
202 
203 	return 0;
204 err_idx_mtr:
205 	hns_roce_mtr_destroy(hr_dev, &idx_que->mtr);
206 
207 	return ret;
208 }
209 
210 static void free_srq_idx(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq)
211 {
212 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
213 
214 	bitmap_free(idx_que->bitmap);
215 	idx_que->bitmap = NULL;
216 	hns_roce_mtr_destroy(hr_dev, &idx_que->mtr);
217 }
218 
219 static int alloc_srq_wqe_buf(struct hns_roce_dev *hr_dev,
220 			     struct hns_roce_srq *srq,
221 			     struct ib_udata *udata, unsigned long addr)
222 {
223 	struct ib_device *ibdev = &hr_dev->ib_dev;
224 	struct hns_roce_buf_attr buf_attr = {};
225 	int ret;
226 
227 	srq->wqe_shift = ilog2(roundup_pow_of_two(max(HNS_ROCE_SGE_SIZE,
228 						      HNS_ROCE_SGE_SIZE *
229 						      srq->max_gs)));
230 
231 	buf_attr.page_shift = hr_dev->caps.srqwqe_buf_pg_sz + PAGE_SHIFT;
232 	buf_attr.region[0].size = to_hr_hem_entries_size(srq->wqe_cnt,
233 							 srq->wqe_shift);
234 	buf_attr.region[0].hopnum = hr_dev->caps.srqwqe_hop_num;
235 	buf_attr.region_count = 1;
236 
237 	ret = hns_roce_mtr_create(hr_dev, &srq->buf_mtr, &buf_attr,
238 				  hr_dev->caps.srqwqe_ba_pg_sz + PAGE_SHIFT,
239 				  udata, addr);
240 	if (ret)
241 		ibdev_err(ibdev,
242 			  "failed to alloc SRQ buf mtr, ret = %d.\n", ret);
243 
244 	return ret;
245 }
246 
247 static void free_srq_wqe_buf(struct hns_roce_dev *hr_dev,
248 			     struct hns_roce_srq *srq)
249 {
250 	hns_roce_mtr_destroy(hr_dev, &srq->buf_mtr);
251 }
252 
253 static int alloc_srq_wrid(struct hns_roce_srq *srq)
254 {
255 	srq->wrid = kvmalloc_array(srq->wqe_cnt, sizeof(u64), GFP_KERNEL);
256 	if (!srq->wrid)
257 		return -ENOMEM;
258 
259 	return 0;
260 }
261 
262 static void free_srq_wrid(struct hns_roce_srq *srq)
263 {
264 	kvfree(srq->wrid);
265 	srq->wrid = NULL;
266 }
267 
268 static u32 proc_srq_sge(struct hns_roce_dev *dev, struct hns_roce_srq *hr_srq,
269 			bool user)
270 {
271 	u32 max_sge = dev->caps.max_srq_sges;
272 
273 	if (dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
274 		return max_sge;
275 
276 	/* Reserve SGEs only for HIP08 in kernel; The userspace driver will
277 	 * calculate number of max_sge with reserved SGEs when allocating wqe
278 	 * buf, so there is no need to do this again in kernel. But the number
279 	 * may exceed the capacity of SGEs recorded in the firmware, so the
280 	 * kernel driver should just adapt the value accordingly.
281 	 */
282 	if (user)
283 		max_sge = roundup_pow_of_two(max_sge + 1);
284 	else
285 		hr_srq->rsv_sge = 1;
286 
287 	return max_sge;
288 }
289 
290 static int set_srq_basic_param(struct hns_roce_srq *srq,
291 			       struct ib_srq_init_attr *init_attr,
292 			       struct ib_udata *udata)
293 {
294 	struct hns_roce_dev *hr_dev = to_hr_dev(srq->ibsrq.device);
295 	struct ib_srq_attr *attr = &init_attr->attr;
296 	u32 max_sge;
297 
298 	max_sge = proc_srq_sge(hr_dev, srq, !!udata);
299 	if (attr->max_wr > hr_dev->caps.max_srq_wrs ||
300 	    attr->max_sge > max_sge || !attr->max_sge) {
301 		ibdev_err(&hr_dev->ib_dev,
302 			  "invalid SRQ attr, depth = %u, sge = %u.\n",
303 			  attr->max_wr, attr->max_sge);
304 		return -EINVAL;
305 	}
306 
307 	attr->max_wr = max_t(u32, attr->max_wr, HNS_ROCE_MIN_SRQ_WQE_NUM);
308 	srq->wqe_cnt = roundup_pow_of_two(attr->max_wr);
309 	srq->max_gs = roundup_pow_of_two(attr->max_sge + srq->rsv_sge);
310 
311 	attr->max_wr = srq->wqe_cnt;
312 	attr->max_sge = srq->max_gs - srq->rsv_sge;
313 	attr->srq_limit = 0;
314 
315 	return 0;
316 }
317 
318 static void set_srq_ext_param(struct hns_roce_srq *srq,
319 			      struct ib_srq_init_attr *init_attr)
320 {
321 	srq->cqn = ib_srq_has_cq(init_attr->srq_type) ?
322 		   to_hr_cq(init_attr->ext.cq)->cqn : 0;
323 
324 	srq->xrcdn = (init_attr->srq_type == IB_SRQT_XRC) ?
325 		     to_hr_xrcd(init_attr->ext.xrc.xrcd)->xrcdn : 0;
326 }
327 
328 static int set_srq_param(struct hns_roce_srq *srq,
329 			 struct ib_srq_init_attr *init_attr,
330 			 struct ib_udata *udata)
331 {
332 	int ret;
333 
334 	ret = set_srq_basic_param(srq, init_attr, udata);
335 	if (ret)
336 		return ret;
337 
338 	set_srq_ext_param(srq, init_attr);
339 
340 	return 0;
341 }
342 
343 static int alloc_srq_buf(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq,
344 			 struct ib_udata *udata)
345 {
346 	struct hns_roce_ib_create_srq ucmd = {};
347 	int ret;
348 
349 	if (udata) {
350 		ret = ib_copy_from_udata(&ucmd, udata,
351 					 min(udata->inlen, sizeof(ucmd)));
352 		if (ret) {
353 			ibdev_err(&hr_dev->ib_dev,
354 				  "failed to copy SRQ udata, ret = %d.\n",
355 				  ret);
356 			return ret;
357 		}
358 	}
359 
360 	ret = alloc_srq_idx(hr_dev, srq, udata, ucmd.que_addr);
361 	if (ret)
362 		return ret;
363 
364 	ret = alloc_srq_wqe_buf(hr_dev, srq, udata, ucmd.buf_addr);
365 	if (ret)
366 		goto err_idx;
367 
368 	if (!udata) {
369 		ret = alloc_srq_wrid(srq);
370 		if (ret)
371 			goto err_wqe_buf;
372 	}
373 
374 	return 0;
375 
376 err_wqe_buf:
377 	free_srq_wqe_buf(hr_dev, srq);
378 err_idx:
379 	free_srq_idx(hr_dev, srq);
380 
381 	return ret;
382 }
383 
384 static void free_srq_buf(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq)
385 {
386 	free_srq_wrid(srq);
387 	free_srq_wqe_buf(hr_dev, srq);
388 	free_srq_idx(hr_dev, srq);
389 }
390 
391 static int get_srq_ucmd(struct hns_roce_srq *srq, struct ib_udata *udata,
392 			struct hns_roce_ib_create_srq *ucmd)
393 {
394 	struct ib_device *ibdev = srq->ibsrq.device;
395 	int ret;
396 
397 	ret = ib_copy_from_udata(ucmd, udata, min(udata->inlen, sizeof(*ucmd)));
398 	if (ret) {
399 		ibdev_err(ibdev, "failed to copy SRQ udata, ret = %d.\n", ret);
400 		return ret;
401 	}
402 
403 	return 0;
404 }
405 
406 static void free_srq_db(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq,
407 			struct ib_udata *udata)
408 {
409 	struct hns_roce_ucontext *uctx;
410 
411 	if (!(srq->cap_flags & HNS_ROCE_SRQ_CAP_RECORD_DB))
412 		return;
413 
414 	srq->cap_flags &= ~HNS_ROCE_SRQ_CAP_RECORD_DB;
415 	if (udata) {
416 		uctx = rdma_udata_to_drv_context(udata,
417 						 struct hns_roce_ucontext,
418 						 ibucontext);
419 		hns_roce_db_unmap_user(uctx, &srq->rdb);
420 	} else {
421 		hns_roce_free_db(hr_dev, &srq->rdb);
422 	}
423 }
424 
425 static int alloc_srq_db(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq,
426 			struct ib_udata *udata,
427 			struct hns_roce_ib_create_srq_resp *resp)
428 {
429 	struct hns_roce_ib_create_srq ucmd = {};
430 	struct hns_roce_ucontext *uctx;
431 	int ret;
432 
433 	if (udata) {
434 		ret = get_srq_ucmd(srq, udata, &ucmd);
435 		if (ret)
436 			return ret;
437 
438 		if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SRQ_RECORD_DB) &&
439 		    (ucmd.req_cap_flags & HNS_ROCE_SRQ_CAP_RECORD_DB)) {
440 			uctx = rdma_udata_to_drv_context(udata,
441 					struct hns_roce_ucontext, ibucontext);
442 			ret = hns_roce_db_map_user(uctx, ucmd.db_addr,
443 						   &srq->rdb);
444 			if (ret)
445 				return ret;
446 
447 			srq->cap_flags |= HNS_ROCE_RSP_SRQ_CAP_RECORD_DB;
448 		}
449 	} else {
450 		if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SRQ_RECORD_DB) {
451 			ret = hns_roce_alloc_db(hr_dev, &srq->rdb, 1);
452 			if (ret)
453 				return ret;
454 
455 			*srq->rdb.db_record = 0;
456 			srq->cap_flags |= HNS_ROCE_RSP_SRQ_CAP_RECORD_DB;
457 		}
458 		srq->db_reg = hr_dev->reg_base + SRQ_DB_REG;
459 	}
460 
461 	return 0;
462 }
463 
464 int hns_roce_create_srq(struct ib_srq *ib_srq,
465 			struct ib_srq_init_attr *init_attr,
466 			struct ib_udata *udata)
467 {
468 	struct hns_roce_dev *hr_dev = to_hr_dev(ib_srq->device);
469 	struct hns_roce_ib_create_srq_resp resp = {};
470 	struct hns_roce_srq *srq = to_hr_srq(ib_srq);
471 	int ret;
472 
473 	mutex_init(&srq->mutex);
474 	spin_lock_init(&srq->lock);
475 
476 	ret = set_srq_param(srq, init_attr, udata);
477 	if (ret)
478 		goto err_out;
479 
480 	ret = alloc_srq_buf(hr_dev, srq, udata);
481 	if (ret)
482 		goto err_out;
483 
484 	ret = alloc_srq_db(hr_dev, srq, udata, &resp);
485 	if (ret)
486 		goto err_srq_buf;
487 
488 	ret = alloc_srqn(hr_dev, srq);
489 	if (ret)
490 		goto err_srq_db;
491 
492 	ret = alloc_srqc(hr_dev, srq);
493 	if (ret)
494 		goto err_srqn;
495 
496 	if (udata) {
497 		resp.cap_flags = srq->cap_flags;
498 		resp.srqn = srq->srqn;
499 		if (ib_copy_to_udata(udata, &resp,
500 				     min(udata->outlen, sizeof(resp)))) {
501 			ret = -EFAULT;
502 			goto err_srqc;
503 		}
504 	}
505 
506 	srq->event = hns_roce_ib_srq_event;
507 	refcount_set(&srq->refcount, 1);
508 	init_completion(&srq->free);
509 
510 	return 0;
511 
512 err_srqc:
513 	free_srqc(hr_dev, srq);
514 err_srqn:
515 	free_srqn(hr_dev, srq);
516 err_srq_db:
517 	free_srq_db(hr_dev, srq, udata);
518 err_srq_buf:
519 	free_srq_buf(hr_dev, srq);
520 err_out:
521 	mutex_destroy(&srq->mutex);
522 	atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_SRQ_CREATE_ERR_CNT]);
523 
524 	return ret;
525 }
526 
527 int hns_roce_destroy_srq(struct ib_srq *ibsrq, struct ib_udata *udata)
528 {
529 	struct hns_roce_dev *hr_dev = to_hr_dev(ibsrq->device);
530 	struct hns_roce_srq *srq = to_hr_srq(ibsrq);
531 
532 	free_srqc(hr_dev, srq);
533 	free_srqn(hr_dev, srq);
534 	free_srq_db(hr_dev, srq, udata);
535 	free_srq_buf(hr_dev, srq);
536 	mutex_destroy(&srq->mutex);
537 	return 0;
538 }
539 
540 void hns_roce_init_srq_table(struct hns_roce_dev *hr_dev)
541 {
542 	struct hns_roce_srq_table *srq_table = &hr_dev->srq_table;
543 	struct hns_roce_ida *srq_ida = &srq_table->srq_ida;
544 
545 	xa_init(&srq_table->xa);
546 
547 	ida_init(&srq_ida->ida);
548 	srq_ida->max = hr_dev->caps.num_srqs - 1;
549 	srq_ida->min = hr_dev->caps.reserved_srqs;
550 }
551