1 /*
2  * Copyright (c) 2016 Hisilicon Limited.
3  * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33 
34 #include <linux/vmalloc.h>
35 #include <rdma/ib_umem.h>
36 #include <linux/math.h>
37 #include "hns_roce_device.h"
38 #include "hns_roce_cmd.h"
39 #include "hns_roce_hem.h"
40 
41 static u32 hw_index_to_key(int ind)
42 {
43 	return ((u32)ind >> 24) | ((u32)ind << 8);
44 }
45 
46 unsigned long key_to_hw_index(u32 key)
47 {
48 	return (key << 24) | (key >> 8);
49 }
50 
51 static int alloc_mr_key(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr)
52 {
53 	struct hns_roce_ida *mtpt_ida = &hr_dev->mr_table.mtpt_ida;
54 	struct ib_device *ibdev = &hr_dev->ib_dev;
55 	int err;
56 	int id;
57 
58 	/* Allocate a key for mr from mr_table */
59 	id = ida_alloc_range(&mtpt_ida->ida, mtpt_ida->min, mtpt_ida->max,
60 			     GFP_KERNEL);
61 	if (id < 0) {
62 		ibdev_err(ibdev, "failed to alloc id for MR key, id(%d)\n", id);
63 		return -ENOMEM;
64 	}
65 
66 	mr->key = hw_index_to_key(id); /* MR key */
67 
68 	err = hns_roce_table_get(hr_dev, &hr_dev->mr_table.mtpt_table,
69 				 (unsigned long)id);
70 	if (err) {
71 		ibdev_err(ibdev, "failed to alloc mtpt, ret = %d.\n", err);
72 		goto err_free_bitmap;
73 	}
74 
75 	return 0;
76 err_free_bitmap:
77 	ida_free(&mtpt_ida->ida, id);
78 	return err;
79 }
80 
81 static void free_mr_key(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr)
82 {
83 	unsigned long obj = key_to_hw_index(mr->key);
84 
85 	hns_roce_table_put(hr_dev, &hr_dev->mr_table.mtpt_table, obj);
86 	ida_free(&hr_dev->mr_table.mtpt_ida.ida, (int)obj);
87 }
88 
89 static int alloc_mr_pbl(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr,
90 			struct ib_udata *udata, u64 start)
91 {
92 	struct ib_device *ibdev = &hr_dev->ib_dev;
93 	bool is_fast = mr->type == MR_TYPE_FRMR;
94 	struct hns_roce_buf_attr buf_attr = {};
95 	int err;
96 
97 	mr->pbl_hop_num = is_fast ? 1 : hr_dev->caps.pbl_hop_num;
98 	buf_attr.page_shift = is_fast ? PAGE_SHIFT :
99 			      hr_dev->caps.pbl_buf_pg_sz + PAGE_SHIFT;
100 	buf_attr.region[0].size = mr->size;
101 	buf_attr.region[0].hopnum = mr->pbl_hop_num;
102 	buf_attr.region_count = 1;
103 	buf_attr.user_access = mr->access;
104 	/* fast MR's buffer is alloced before mapping, not at creation */
105 	buf_attr.mtt_only = is_fast;
106 
107 	err = hns_roce_mtr_create(hr_dev, &mr->pbl_mtr, &buf_attr,
108 				  hr_dev->caps.pbl_ba_pg_sz + PAGE_SHIFT,
109 				  udata, start);
110 	if (err)
111 		ibdev_err(ibdev, "failed to alloc pbl mtr, ret = %d.\n", err);
112 	else
113 		mr->npages = mr->pbl_mtr.hem_cfg.buf_pg_count;
114 
115 	return err;
116 }
117 
118 static void free_mr_pbl(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr)
119 {
120 	hns_roce_mtr_destroy(hr_dev, &mr->pbl_mtr);
121 }
122 
123 static void hns_roce_mr_free(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr)
124 {
125 	struct ib_device *ibdev = &hr_dev->ib_dev;
126 	int ret;
127 
128 	if (mr->enabled) {
129 		ret = hns_roce_destroy_hw_ctx(hr_dev, HNS_ROCE_CMD_DESTROY_MPT,
130 					      key_to_hw_index(mr->key) &
131 					      (hr_dev->caps.num_mtpts - 1));
132 		if (ret)
133 			ibdev_warn(ibdev, "failed to destroy mpt, ret = %d.\n",
134 				   ret);
135 	}
136 
137 	free_mr_pbl(hr_dev, mr);
138 	free_mr_key(hr_dev, mr);
139 }
140 
141 static int hns_roce_mr_enable(struct hns_roce_dev *hr_dev,
142 			      struct hns_roce_mr *mr)
143 {
144 	unsigned long mtpt_idx = key_to_hw_index(mr->key);
145 	struct hns_roce_cmd_mailbox *mailbox;
146 	struct device *dev = hr_dev->dev;
147 	int ret;
148 
149 	/* Allocate mailbox memory */
150 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
151 	if (IS_ERR(mailbox))
152 		return PTR_ERR(mailbox);
153 
154 	if (mr->type != MR_TYPE_FRMR)
155 		ret = hr_dev->hw->write_mtpt(hr_dev, mailbox->buf, mr);
156 	else
157 		ret = hr_dev->hw->frmr_write_mtpt(hr_dev, mailbox->buf, mr);
158 	if (ret) {
159 		dev_err(dev, "failed to write mtpt, ret = %d.\n", ret);
160 		goto err_page;
161 	}
162 
163 	ret = hns_roce_create_hw_ctx(hr_dev, mailbox, HNS_ROCE_CMD_CREATE_MPT,
164 				     mtpt_idx & (hr_dev->caps.num_mtpts - 1));
165 	if (ret) {
166 		dev_err(dev, "failed to create mpt, ret = %d.\n", ret);
167 		goto err_page;
168 	}
169 
170 	mr->enabled = 1;
171 
172 err_page:
173 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
174 
175 	return ret;
176 }
177 
178 void hns_roce_init_mr_table(struct hns_roce_dev *hr_dev)
179 {
180 	struct hns_roce_ida *mtpt_ida = &hr_dev->mr_table.mtpt_ida;
181 
182 	ida_init(&mtpt_ida->ida);
183 	mtpt_ida->max = hr_dev->caps.num_mtpts - 1;
184 	mtpt_ida->min = hr_dev->caps.reserved_mrws;
185 }
186 
187 struct ib_mr *hns_roce_get_dma_mr(struct ib_pd *pd, int acc)
188 {
189 	struct hns_roce_dev *hr_dev = to_hr_dev(pd->device);
190 	struct hns_roce_mr *mr;
191 	int ret;
192 
193 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
194 	if (!mr)
195 		return  ERR_PTR(-ENOMEM);
196 
197 	mr->type = MR_TYPE_DMA;
198 	mr->pd = to_hr_pd(pd)->pdn;
199 	mr->access = acc;
200 
201 	/* Allocate memory region key */
202 	hns_roce_hem_list_init(&mr->pbl_mtr.hem_list);
203 	ret = alloc_mr_key(hr_dev, mr);
204 	if (ret)
205 		goto err_free;
206 
207 	ret = hns_roce_mr_enable(hr_dev, mr);
208 	if (ret)
209 		goto err_mr;
210 
211 	mr->ibmr.rkey = mr->ibmr.lkey = mr->key;
212 
213 	return &mr->ibmr;
214 err_mr:
215 	free_mr_key(hr_dev, mr);
216 
217 err_free:
218 	kfree(mr);
219 	return ERR_PTR(ret);
220 }
221 
222 struct ib_mr *hns_roce_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
223 				   u64 virt_addr, int access_flags,
224 				   struct ib_udata *udata)
225 {
226 	struct hns_roce_dev *hr_dev = to_hr_dev(pd->device);
227 	struct hns_roce_mr *mr;
228 	int ret;
229 
230 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
231 	if (!mr)
232 		return ERR_PTR(-ENOMEM);
233 
234 	mr->iova = virt_addr;
235 	mr->size = length;
236 	mr->pd = to_hr_pd(pd)->pdn;
237 	mr->access = access_flags;
238 	mr->type = MR_TYPE_MR;
239 
240 	ret = alloc_mr_key(hr_dev, mr);
241 	if (ret)
242 		goto err_alloc_mr;
243 
244 	ret = alloc_mr_pbl(hr_dev, mr, udata, start);
245 	if (ret)
246 		goto err_alloc_key;
247 
248 	ret = hns_roce_mr_enable(hr_dev, mr);
249 	if (ret)
250 		goto err_alloc_pbl;
251 
252 	mr->ibmr.rkey = mr->ibmr.lkey = mr->key;
253 
254 	return &mr->ibmr;
255 
256 err_alloc_pbl:
257 	free_mr_pbl(hr_dev, mr);
258 err_alloc_key:
259 	free_mr_key(hr_dev, mr);
260 err_alloc_mr:
261 	kfree(mr);
262 	return ERR_PTR(ret);
263 }
264 
265 struct ib_mr *hns_roce_rereg_user_mr(struct ib_mr *ibmr, int flags, u64 start,
266 				     u64 length, u64 virt_addr,
267 				     int mr_access_flags, struct ib_pd *pd,
268 				     struct ib_udata *udata)
269 {
270 	struct hns_roce_dev *hr_dev = to_hr_dev(ibmr->device);
271 	struct ib_device *ib_dev = &hr_dev->ib_dev;
272 	struct hns_roce_mr *mr = to_hr_mr(ibmr);
273 	struct hns_roce_cmd_mailbox *mailbox;
274 	unsigned long mtpt_idx;
275 	int ret;
276 
277 	if (!mr->enabled)
278 		return ERR_PTR(-EINVAL);
279 
280 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
281 	if (IS_ERR(mailbox))
282 		return ERR_CAST(mailbox);
283 
284 	mtpt_idx = key_to_hw_index(mr->key) & (hr_dev->caps.num_mtpts - 1);
285 
286 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, HNS_ROCE_CMD_QUERY_MPT,
287 				mtpt_idx);
288 	if (ret)
289 		goto free_cmd_mbox;
290 
291 	ret = hns_roce_destroy_hw_ctx(hr_dev, HNS_ROCE_CMD_DESTROY_MPT,
292 				      mtpt_idx);
293 	if (ret)
294 		ibdev_warn(ib_dev, "failed to destroy MPT, ret = %d.\n", ret);
295 
296 	mr->enabled = 0;
297 	mr->iova = virt_addr;
298 	mr->size = length;
299 
300 	if (flags & IB_MR_REREG_PD)
301 		mr->pd = to_hr_pd(pd)->pdn;
302 
303 	if (flags & IB_MR_REREG_ACCESS)
304 		mr->access = mr_access_flags;
305 
306 	if (flags & IB_MR_REREG_TRANS) {
307 		free_mr_pbl(hr_dev, mr);
308 		ret = alloc_mr_pbl(hr_dev, mr, udata, start);
309 		if (ret) {
310 			ibdev_err(ib_dev, "failed to alloc mr PBL, ret = %d.\n",
311 				  ret);
312 			goto free_cmd_mbox;
313 		}
314 	}
315 
316 	ret = hr_dev->hw->rereg_write_mtpt(hr_dev, mr, flags, mailbox->buf);
317 	if (ret) {
318 		ibdev_err(ib_dev, "failed to write mtpt, ret = %d.\n", ret);
319 		goto free_cmd_mbox;
320 	}
321 
322 	ret = hns_roce_create_hw_ctx(hr_dev, mailbox, HNS_ROCE_CMD_CREATE_MPT,
323 				     mtpt_idx);
324 	if (ret) {
325 		ibdev_err(ib_dev, "failed to create MPT, ret = %d.\n", ret);
326 		goto free_cmd_mbox;
327 	}
328 
329 	mr->enabled = 1;
330 
331 free_cmd_mbox:
332 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
333 
334 	if (ret)
335 		return ERR_PTR(ret);
336 	return NULL;
337 }
338 
339 int hns_roce_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
340 {
341 	struct hns_roce_dev *hr_dev = to_hr_dev(ibmr->device);
342 	struct hns_roce_mr *mr = to_hr_mr(ibmr);
343 
344 	if (hr_dev->hw->dereg_mr)
345 		hr_dev->hw->dereg_mr(hr_dev);
346 
347 	hns_roce_mr_free(hr_dev, mr);
348 	kfree(mr);
349 
350 	return 0;
351 }
352 
353 struct ib_mr *hns_roce_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
354 				u32 max_num_sg)
355 {
356 	struct hns_roce_dev *hr_dev = to_hr_dev(pd->device);
357 	struct device *dev = hr_dev->dev;
358 	struct hns_roce_mr *mr;
359 	int ret;
360 
361 	if (mr_type != IB_MR_TYPE_MEM_REG)
362 		return ERR_PTR(-EINVAL);
363 
364 	if (max_num_sg > HNS_ROCE_FRMR_MAX_PA) {
365 		dev_err(dev, "max_num_sg larger than %d\n",
366 			HNS_ROCE_FRMR_MAX_PA);
367 		return ERR_PTR(-EINVAL);
368 	}
369 
370 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
371 	if (!mr)
372 		return ERR_PTR(-ENOMEM);
373 
374 	mr->type = MR_TYPE_FRMR;
375 	mr->pd = to_hr_pd(pd)->pdn;
376 	mr->size = max_num_sg * (1 << PAGE_SHIFT);
377 
378 	/* Allocate memory region key */
379 	ret = alloc_mr_key(hr_dev, mr);
380 	if (ret)
381 		goto err_free;
382 
383 	ret = alloc_mr_pbl(hr_dev, mr, NULL, 0);
384 	if (ret)
385 		goto err_key;
386 
387 	ret = hns_roce_mr_enable(hr_dev, mr);
388 	if (ret)
389 		goto err_pbl;
390 
391 	mr->ibmr.rkey = mr->ibmr.lkey = mr->key;
392 	mr->ibmr.length = mr->size;
393 
394 	return &mr->ibmr;
395 
396 err_pbl:
397 	free_mr_pbl(hr_dev, mr);
398 err_key:
399 	free_mr_key(hr_dev, mr);
400 err_free:
401 	kfree(mr);
402 	return ERR_PTR(ret);
403 }
404 
405 static int hns_roce_set_page(struct ib_mr *ibmr, u64 addr)
406 {
407 	struct hns_roce_mr *mr = to_hr_mr(ibmr);
408 
409 	if (likely(mr->npages < mr->pbl_mtr.hem_cfg.buf_pg_count)) {
410 		mr->page_list[mr->npages++] = addr;
411 		return 0;
412 	}
413 
414 	return -ENOBUFS;
415 }
416 
417 int hns_roce_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
418 		       unsigned int *sg_offset)
419 {
420 	struct hns_roce_dev *hr_dev = to_hr_dev(ibmr->device);
421 	struct ib_device *ibdev = &hr_dev->ib_dev;
422 	struct hns_roce_mr *mr = to_hr_mr(ibmr);
423 	struct hns_roce_mtr *mtr = &mr->pbl_mtr;
424 	int ret = 0;
425 
426 	mr->npages = 0;
427 	mr->page_list = kvcalloc(mr->pbl_mtr.hem_cfg.buf_pg_count,
428 				 sizeof(dma_addr_t), GFP_KERNEL);
429 	if (!mr->page_list)
430 		return ret;
431 
432 	ret = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, hns_roce_set_page);
433 	if (ret < 1) {
434 		ibdev_err(ibdev, "failed to store sg pages %u %u, cnt = %d.\n",
435 			  mr->npages, mr->pbl_mtr.hem_cfg.buf_pg_count, ret);
436 		goto err_page_list;
437 	}
438 
439 	mtr->hem_cfg.region[0].offset = 0;
440 	mtr->hem_cfg.region[0].count = mr->npages;
441 	mtr->hem_cfg.region[0].hopnum = mr->pbl_hop_num;
442 	mtr->hem_cfg.region_count = 1;
443 	ret = hns_roce_mtr_map(hr_dev, mtr, mr->page_list, mr->npages);
444 	if (ret) {
445 		ibdev_err(ibdev, "failed to map sg mtr, ret = %d.\n", ret);
446 		ret = 0;
447 	} else {
448 		mr->pbl_mtr.hem_cfg.buf_pg_shift = (u32)ilog2(ibmr->page_size);
449 		ret = mr->npages;
450 	}
451 
452 err_page_list:
453 	kvfree(mr->page_list);
454 	mr->page_list = NULL;
455 
456 	return ret;
457 }
458 
459 static void hns_roce_mw_free(struct hns_roce_dev *hr_dev,
460 			     struct hns_roce_mw *mw)
461 {
462 	struct device *dev = hr_dev->dev;
463 	int ret;
464 
465 	if (mw->enabled) {
466 		ret = hns_roce_destroy_hw_ctx(hr_dev, HNS_ROCE_CMD_DESTROY_MPT,
467 					      key_to_hw_index(mw->rkey) &
468 					      (hr_dev->caps.num_mtpts - 1));
469 		if (ret)
470 			dev_warn(dev, "MW DESTROY_MPT failed (%d)\n", ret);
471 
472 		hns_roce_table_put(hr_dev, &hr_dev->mr_table.mtpt_table,
473 				   key_to_hw_index(mw->rkey));
474 	}
475 
476 	ida_free(&hr_dev->mr_table.mtpt_ida.ida,
477 		 (int)key_to_hw_index(mw->rkey));
478 }
479 
480 static int hns_roce_mw_enable(struct hns_roce_dev *hr_dev,
481 			      struct hns_roce_mw *mw)
482 {
483 	struct hns_roce_mr_table *mr_table = &hr_dev->mr_table;
484 	struct hns_roce_cmd_mailbox *mailbox;
485 	struct device *dev = hr_dev->dev;
486 	unsigned long mtpt_idx = key_to_hw_index(mw->rkey);
487 	int ret;
488 
489 	/* prepare HEM entry memory */
490 	ret = hns_roce_table_get(hr_dev, &mr_table->mtpt_table, mtpt_idx);
491 	if (ret)
492 		return ret;
493 
494 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
495 	if (IS_ERR(mailbox)) {
496 		ret = PTR_ERR(mailbox);
497 		goto err_table;
498 	}
499 
500 	ret = hr_dev->hw->mw_write_mtpt(mailbox->buf, mw);
501 	if (ret) {
502 		dev_err(dev, "MW write mtpt fail!\n");
503 		goto err_page;
504 	}
505 
506 	ret = hns_roce_create_hw_ctx(hr_dev, mailbox, HNS_ROCE_CMD_CREATE_MPT,
507 				     mtpt_idx & (hr_dev->caps.num_mtpts - 1));
508 	if (ret) {
509 		dev_err(dev, "MW CREATE_MPT failed (%d)\n", ret);
510 		goto err_page;
511 	}
512 
513 	mw->enabled = 1;
514 
515 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
516 
517 	return 0;
518 
519 err_page:
520 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
521 
522 err_table:
523 	hns_roce_table_put(hr_dev, &mr_table->mtpt_table, mtpt_idx);
524 
525 	return ret;
526 }
527 
528 int hns_roce_alloc_mw(struct ib_mw *ibmw, struct ib_udata *udata)
529 {
530 	struct hns_roce_dev *hr_dev = to_hr_dev(ibmw->device);
531 	struct hns_roce_ida *mtpt_ida = &hr_dev->mr_table.mtpt_ida;
532 	struct ib_device *ibdev = &hr_dev->ib_dev;
533 	struct hns_roce_mw *mw = to_hr_mw(ibmw);
534 	int ret;
535 	int id;
536 
537 	/* Allocate a key for mw from mr_table */
538 	id = ida_alloc_range(&mtpt_ida->ida, mtpt_ida->min, mtpt_ida->max,
539 			     GFP_KERNEL);
540 	if (id < 0) {
541 		ibdev_err(ibdev, "failed to alloc id for MW key, id(%d)\n", id);
542 		return -ENOMEM;
543 	}
544 
545 	mw->rkey = hw_index_to_key(id);
546 
547 	ibmw->rkey = mw->rkey;
548 	mw->pdn = to_hr_pd(ibmw->pd)->pdn;
549 	mw->pbl_hop_num = hr_dev->caps.pbl_hop_num;
550 	mw->pbl_ba_pg_sz = hr_dev->caps.pbl_ba_pg_sz;
551 	mw->pbl_buf_pg_sz = hr_dev->caps.pbl_buf_pg_sz;
552 
553 	ret = hns_roce_mw_enable(hr_dev, mw);
554 	if (ret)
555 		goto err_mw;
556 
557 	return 0;
558 
559 err_mw:
560 	hns_roce_mw_free(hr_dev, mw);
561 	return ret;
562 }
563 
564 int hns_roce_dealloc_mw(struct ib_mw *ibmw)
565 {
566 	struct hns_roce_dev *hr_dev = to_hr_dev(ibmw->device);
567 	struct hns_roce_mw *mw = to_hr_mw(ibmw);
568 
569 	hns_roce_mw_free(hr_dev, mw);
570 	return 0;
571 }
572 
573 static int mtr_map_region(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr,
574 			  struct hns_roce_buf_region *region, dma_addr_t *pages,
575 			  int max_count)
576 {
577 	int count, npage;
578 	int offset, end;
579 	__le64 *mtts;
580 	u64 addr;
581 	int i;
582 
583 	offset = region->offset;
584 	end = offset + region->count;
585 	npage = 0;
586 	while (offset < end && npage < max_count) {
587 		count = 0;
588 		mtts = hns_roce_hem_list_find_mtt(hr_dev, &mtr->hem_list,
589 						  offset, &count);
590 		if (!mtts)
591 			return -ENOBUFS;
592 
593 		for (i = 0; i < count && npage < max_count; i++) {
594 			addr = pages[npage];
595 
596 			mtts[i] = cpu_to_le64(addr);
597 			npage++;
598 		}
599 		offset += count;
600 	}
601 
602 	return npage;
603 }
604 
605 static inline bool mtr_has_mtt(struct hns_roce_buf_attr *attr)
606 {
607 	int i;
608 
609 	for (i = 0; i < attr->region_count; i++)
610 		if (attr->region[i].hopnum != HNS_ROCE_HOP_NUM_0 &&
611 		    attr->region[i].hopnum > 0)
612 			return true;
613 
614 	/* because the mtr only one root base address, when hopnum is 0 means
615 	 * root base address equals the first buffer address, thus all alloced
616 	 * memory must in a continuous space accessed by direct mode.
617 	 */
618 	return false;
619 }
620 
621 static inline size_t mtr_bufs_size(struct hns_roce_buf_attr *attr)
622 {
623 	size_t size = 0;
624 	int i;
625 
626 	for (i = 0; i < attr->region_count; i++)
627 		size += attr->region[i].size;
628 
629 	return size;
630 }
631 
632 /*
633  * check the given pages in continuous address space
634  * Returns 0 on success, or the error page num.
635  */
636 static inline int mtr_check_direct_pages(dma_addr_t *pages, int page_count,
637 					 unsigned int page_shift)
638 {
639 	size_t page_size = 1 << page_shift;
640 	int i;
641 
642 	for (i = 1; i < page_count; i++)
643 		if (pages[i] - pages[i - 1] != page_size)
644 			return i;
645 
646 	return 0;
647 }
648 
649 static void mtr_free_bufs(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr)
650 {
651 	/* release user buffers */
652 	if (mtr->umem) {
653 		ib_umem_release(mtr->umem);
654 		mtr->umem = NULL;
655 	}
656 
657 	/* release kernel buffers */
658 	if (mtr->kmem) {
659 		hns_roce_buf_free(hr_dev, mtr->kmem);
660 		mtr->kmem = NULL;
661 	}
662 }
663 
664 static int mtr_alloc_bufs(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr,
665 			  struct hns_roce_buf_attr *buf_attr,
666 			  struct ib_udata *udata, unsigned long user_addr)
667 {
668 	struct ib_device *ibdev = &hr_dev->ib_dev;
669 	size_t total_size;
670 
671 	total_size = mtr_bufs_size(buf_attr);
672 
673 	if (udata) {
674 		mtr->kmem = NULL;
675 		mtr->umem = ib_umem_get(ibdev, user_addr, total_size,
676 					buf_attr->user_access);
677 		if (IS_ERR_OR_NULL(mtr->umem)) {
678 			ibdev_err(ibdev, "failed to get umem, ret = %ld.\n",
679 				  PTR_ERR(mtr->umem));
680 			return -ENOMEM;
681 		}
682 	} else {
683 		mtr->umem = NULL;
684 		mtr->kmem = hns_roce_buf_alloc(hr_dev, total_size,
685 					       buf_attr->page_shift,
686 					       mtr->hem_cfg.is_direct ?
687 					       HNS_ROCE_BUF_DIRECT : 0);
688 		if (IS_ERR(mtr->kmem)) {
689 			ibdev_err(ibdev, "failed to alloc kmem, ret = %ld.\n",
690 				  PTR_ERR(mtr->kmem));
691 			return PTR_ERR(mtr->kmem);
692 		}
693 	}
694 
695 	return 0;
696 }
697 
698 static int mtr_map_bufs(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr,
699 			int page_count, unsigned int page_shift)
700 {
701 	struct ib_device *ibdev = &hr_dev->ib_dev;
702 	dma_addr_t *pages;
703 	int npage;
704 	int ret;
705 
706 	/* alloc a tmp array to store buffer's dma address */
707 	pages = kvcalloc(page_count, sizeof(dma_addr_t), GFP_KERNEL);
708 	if (!pages)
709 		return -ENOMEM;
710 
711 	if (mtr->umem)
712 		npage = hns_roce_get_umem_bufs(hr_dev, pages, page_count,
713 					       mtr->umem, page_shift);
714 	else
715 		npage = hns_roce_get_kmem_bufs(hr_dev, pages, page_count,
716 					       mtr->kmem, page_shift);
717 
718 	if (npage != page_count) {
719 		ibdev_err(ibdev, "failed to get mtr page %d != %d.\n", npage,
720 			  page_count);
721 		ret = -ENOBUFS;
722 		goto err_alloc_list;
723 	}
724 
725 	if (mtr->hem_cfg.is_direct && npage > 1) {
726 		ret = mtr_check_direct_pages(pages, npage, page_shift);
727 		if (ret) {
728 			ibdev_err(ibdev, "failed to check %s page: %d / %d.\n",
729 				  mtr->umem ? "umtr" : "kmtr", ret, npage);
730 			ret = -ENOBUFS;
731 			goto err_alloc_list;
732 		}
733 	}
734 
735 	ret = hns_roce_mtr_map(hr_dev, mtr, pages, page_count);
736 	if (ret)
737 		ibdev_err(ibdev, "failed to map mtr pages, ret = %d.\n", ret);
738 
739 err_alloc_list:
740 	kvfree(pages);
741 
742 	return ret;
743 }
744 
745 int hns_roce_mtr_map(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr,
746 		     dma_addr_t *pages, unsigned int page_cnt)
747 {
748 	struct ib_device *ibdev = &hr_dev->ib_dev;
749 	struct hns_roce_buf_region *r;
750 	unsigned int i, mapped_cnt;
751 	int ret = 0;
752 
753 	/*
754 	 * Only use the first page address as root ba when hopnum is 0, this
755 	 * is because the addresses of all pages are consecutive in this case.
756 	 */
757 	if (mtr->hem_cfg.is_direct) {
758 		mtr->hem_cfg.root_ba = pages[0];
759 		return 0;
760 	}
761 
762 	for (i = 0, mapped_cnt = 0; i < mtr->hem_cfg.region_count &&
763 	     mapped_cnt < page_cnt; i++) {
764 		r = &mtr->hem_cfg.region[i];
765 		/* if hopnum is 0, no need to map pages in this region */
766 		if (!r->hopnum) {
767 			mapped_cnt += r->count;
768 			continue;
769 		}
770 
771 		if (r->offset + r->count > page_cnt) {
772 			ret = -EINVAL;
773 			ibdev_err(ibdev,
774 				  "failed to check mtr%u count %u + %u > %u.\n",
775 				  i, r->offset, r->count, page_cnt);
776 			return ret;
777 		}
778 
779 		ret = mtr_map_region(hr_dev, mtr, r, &pages[r->offset],
780 				     page_cnt - mapped_cnt);
781 		if (ret < 0) {
782 			ibdev_err(ibdev,
783 				  "failed to map mtr%u offset %u, ret = %d.\n",
784 				  i, r->offset, ret);
785 			return ret;
786 		}
787 		mapped_cnt += ret;
788 		ret = 0;
789 	}
790 
791 	if (mapped_cnt < page_cnt) {
792 		ret = -ENOBUFS;
793 		ibdev_err(ibdev, "failed to map mtr pages count: %u < %u.\n",
794 			  mapped_cnt, page_cnt);
795 	}
796 
797 	return ret;
798 }
799 
800 int hns_roce_mtr_find(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr,
801 		      u32 offset, u64 *mtt_buf, int mtt_max, u64 *base_addr)
802 {
803 	struct hns_roce_hem_cfg *cfg = &mtr->hem_cfg;
804 	int mtt_count, left;
805 	u32 start_index;
806 	int total = 0;
807 	__le64 *mtts;
808 	u32 npage;
809 	u64 addr;
810 
811 	if (!mtt_buf || mtt_max < 1)
812 		goto done;
813 
814 	/* no mtt memory in direct mode, so just return the buffer address */
815 	if (cfg->is_direct) {
816 		start_index = offset >> HNS_HW_PAGE_SHIFT;
817 		for (mtt_count = 0; mtt_count < cfg->region_count &&
818 		     total < mtt_max; mtt_count++) {
819 			npage = cfg->region[mtt_count].offset;
820 			if (npage < start_index)
821 				continue;
822 
823 			addr = cfg->root_ba + (npage << HNS_HW_PAGE_SHIFT);
824 			mtt_buf[total] = addr;
825 
826 			total++;
827 		}
828 
829 		goto done;
830 	}
831 
832 	start_index = offset >> cfg->buf_pg_shift;
833 	left = mtt_max;
834 	while (left > 0) {
835 		mtt_count = 0;
836 		mtts = hns_roce_hem_list_find_mtt(hr_dev, &mtr->hem_list,
837 						  start_index + total,
838 						  &mtt_count);
839 		if (!mtts || !mtt_count)
840 			goto done;
841 
842 		npage = min(mtt_count, left);
843 		left -= npage;
844 		for (mtt_count = 0; mtt_count < npage; mtt_count++)
845 			mtt_buf[total++] = le64_to_cpu(mtts[mtt_count]);
846 	}
847 
848 done:
849 	if (base_addr)
850 		*base_addr = cfg->root_ba;
851 
852 	return total;
853 }
854 
855 static int mtr_init_buf_cfg(struct hns_roce_dev *hr_dev,
856 			    struct hns_roce_buf_attr *attr,
857 			    struct hns_roce_hem_cfg *cfg,
858 			    unsigned int *buf_page_shift, u64 unalinged_size)
859 {
860 	struct hns_roce_buf_region *r;
861 	u64 first_region_padding;
862 	int page_cnt, region_cnt;
863 	unsigned int page_shift;
864 	size_t buf_size;
865 
866 	/* If mtt is disabled, all pages must be within a continuous range */
867 	cfg->is_direct = !mtr_has_mtt(attr);
868 	buf_size = mtr_bufs_size(attr);
869 	if (cfg->is_direct) {
870 		/* When HEM buffer uses 0-level addressing, the page size is
871 		 * equal to the whole buffer size, and we split the buffer into
872 		 * small pages which is used to check whether the adjacent
873 		 * units are in the continuous space and its size is fixed to
874 		 * 4K based on hns ROCEE's requirement.
875 		 */
876 		page_shift = HNS_HW_PAGE_SHIFT;
877 
878 		/* The ROCEE requires the page size to be 4K * 2 ^ N. */
879 		cfg->buf_pg_count = 1;
880 		cfg->buf_pg_shift = HNS_HW_PAGE_SHIFT +
881 			order_base_2(DIV_ROUND_UP(buf_size, HNS_HW_PAGE_SIZE));
882 		first_region_padding = 0;
883 	} else {
884 		page_shift = attr->page_shift;
885 		cfg->buf_pg_count = DIV_ROUND_UP(buf_size + unalinged_size,
886 						 1 << page_shift);
887 		cfg->buf_pg_shift = page_shift;
888 		first_region_padding = unalinged_size;
889 	}
890 
891 	/* Convert buffer size to page index and page count for each region and
892 	 * the buffer's offset needs to be appended to the first region.
893 	 */
894 	for (page_cnt = 0, region_cnt = 0; region_cnt < attr->region_count &&
895 	     region_cnt < ARRAY_SIZE(cfg->region); region_cnt++) {
896 		r = &cfg->region[region_cnt];
897 		r->offset = page_cnt;
898 		buf_size = hr_hw_page_align(attr->region[region_cnt].size +
899 					    first_region_padding);
900 		r->count = DIV_ROUND_UP(buf_size, 1 << page_shift);
901 		first_region_padding = 0;
902 		page_cnt += r->count;
903 		r->hopnum = to_hr_hem_hopnum(attr->region[region_cnt].hopnum,
904 					     r->count);
905 	}
906 
907 	cfg->region_count = region_cnt;
908 	*buf_page_shift = page_shift;
909 
910 	return page_cnt;
911 }
912 
913 static u64 cal_pages_per_l1ba(unsigned int ba_per_bt, unsigned int hopnum)
914 {
915 	return int_pow(ba_per_bt, hopnum - 1);
916 }
917 
918 static unsigned int cal_best_bt_pg_sz(struct hns_roce_dev *hr_dev,
919 				      struct hns_roce_mtr *mtr,
920 				      unsigned int pg_shift)
921 {
922 	unsigned long cap = hr_dev->caps.page_size_cap;
923 	struct hns_roce_buf_region *re;
924 	unsigned int pgs_per_l1ba;
925 	unsigned int ba_per_bt;
926 	unsigned int ba_num;
927 	int i;
928 
929 	for_each_set_bit_from(pg_shift, &cap, sizeof(cap) * BITS_PER_BYTE) {
930 		if (!(BIT(pg_shift) & cap))
931 			continue;
932 
933 		ba_per_bt = BIT(pg_shift) / BA_BYTE_LEN;
934 		ba_num = 0;
935 		for (i = 0; i < mtr->hem_cfg.region_count; i++) {
936 			re = &mtr->hem_cfg.region[i];
937 			if (re->hopnum == 0)
938 				continue;
939 
940 			pgs_per_l1ba = cal_pages_per_l1ba(ba_per_bt, re->hopnum);
941 			ba_num += DIV_ROUND_UP(re->count, pgs_per_l1ba);
942 		}
943 
944 		if (ba_num <= ba_per_bt)
945 			return pg_shift;
946 	}
947 
948 	return 0;
949 }
950 
951 static int mtr_alloc_mtt(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr,
952 			 unsigned int ba_page_shift)
953 {
954 	struct hns_roce_hem_cfg *cfg = &mtr->hem_cfg;
955 	int ret;
956 
957 	hns_roce_hem_list_init(&mtr->hem_list);
958 	if (!cfg->is_direct) {
959 		ba_page_shift = cal_best_bt_pg_sz(hr_dev, mtr, ba_page_shift);
960 		if (!ba_page_shift)
961 			return -ERANGE;
962 
963 		ret = hns_roce_hem_list_request(hr_dev, &mtr->hem_list,
964 						cfg->region, cfg->region_count,
965 						ba_page_shift);
966 		if (ret)
967 			return ret;
968 		cfg->root_ba = mtr->hem_list.root_ba;
969 		cfg->ba_pg_shift = ba_page_shift;
970 	} else {
971 		cfg->ba_pg_shift = cfg->buf_pg_shift;
972 	}
973 
974 	return 0;
975 }
976 
977 static void mtr_free_mtt(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr)
978 {
979 	hns_roce_hem_list_release(hr_dev, &mtr->hem_list);
980 }
981 
982 /**
983  * hns_roce_mtr_create - Create hns memory translate region.
984  *
985  * @hr_dev: RoCE device struct pointer
986  * @mtr: memory translate region
987  * @buf_attr: buffer attribute for creating mtr
988  * @ba_page_shift: page shift for multi-hop base address table
989  * @udata: user space context, if it's NULL, means kernel space
990  * @user_addr: userspace virtual address to start at
991  */
992 int hns_roce_mtr_create(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr,
993 			struct hns_roce_buf_attr *buf_attr,
994 			unsigned int ba_page_shift, struct ib_udata *udata,
995 			unsigned long user_addr)
996 {
997 	struct ib_device *ibdev = &hr_dev->ib_dev;
998 	unsigned int buf_page_shift = 0;
999 	int buf_page_cnt;
1000 	int ret;
1001 
1002 	buf_page_cnt = mtr_init_buf_cfg(hr_dev, buf_attr, &mtr->hem_cfg,
1003 					&buf_page_shift,
1004 					udata ? user_addr & ~PAGE_MASK : 0);
1005 	if (buf_page_cnt < 1 || buf_page_shift < HNS_HW_PAGE_SHIFT) {
1006 		ibdev_err(ibdev, "failed to init mtr cfg, count %d shift %u.\n",
1007 			  buf_page_cnt, buf_page_shift);
1008 		return -EINVAL;
1009 	}
1010 
1011 	ret = mtr_alloc_mtt(hr_dev, mtr, ba_page_shift);
1012 	if (ret) {
1013 		ibdev_err(ibdev, "failed to alloc mtr mtt, ret = %d.\n", ret);
1014 		return ret;
1015 	}
1016 
1017 	/* The caller has its own buffer list and invokes the hns_roce_mtr_map()
1018 	 * to finish the MTT configuration.
1019 	 */
1020 	if (buf_attr->mtt_only) {
1021 		mtr->umem = NULL;
1022 		mtr->kmem = NULL;
1023 		return 0;
1024 	}
1025 
1026 	ret = mtr_alloc_bufs(hr_dev, mtr, buf_attr, udata, user_addr);
1027 	if (ret) {
1028 		ibdev_err(ibdev, "failed to alloc mtr bufs, ret = %d.\n", ret);
1029 		goto err_alloc_mtt;
1030 	}
1031 
1032 	/* Write buffer's dma address to MTT */
1033 	ret = mtr_map_bufs(hr_dev, mtr, buf_page_cnt, buf_page_shift);
1034 	if (ret)
1035 		ibdev_err(ibdev, "failed to map mtr bufs, ret = %d.\n", ret);
1036 	else
1037 		return 0;
1038 
1039 	mtr_free_bufs(hr_dev, mtr);
1040 err_alloc_mtt:
1041 	mtr_free_mtt(hr_dev, mtr);
1042 	return ret;
1043 }
1044 
1045 void hns_roce_mtr_destroy(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr)
1046 {
1047 	/* release multi-hop addressing resource */
1048 	hns_roce_hem_list_release(hr_dev, &mtr->hem_list);
1049 
1050 	/* free buffers */
1051 	mtr_free_bufs(hr_dev, mtr);
1052 }
1053