xref: /linux/drivers/infiniband/hw/hns/hns_roce_mr.c (revision e9f0878c4b2004ac19581274c1ae4c61ae3ca70e)
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/platform_device.h>
35 #include <linux/vmalloc.h>
36 #include <rdma/ib_umem.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(unsigned long ind)
42 {
43 	return (u32)(ind >> 24) | (ind << 8);
44 }
45 
46 unsigned long key_to_hw_index(u32 key)
47 {
48 	return (key << 24) | (key >> 8);
49 }
50 EXPORT_SYMBOL_GPL(key_to_hw_index);
51 
52 static int hns_roce_sw2hw_mpt(struct hns_roce_dev *hr_dev,
53 			      struct hns_roce_cmd_mailbox *mailbox,
54 			      unsigned long mpt_index)
55 {
56 	return hns_roce_cmd_mbox(hr_dev, mailbox->dma, 0, mpt_index, 0,
57 				 HNS_ROCE_CMD_SW2HW_MPT,
58 				 HNS_ROCE_CMD_TIMEOUT_MSECS);
59 }
60 
61 int hns_roce_hw2sw_mpt(struct hns_roce_dev *hr_dev,
62 			      struct hns_roce_cmd_mailbox *mailbox,
63 			      unsigned long mpt_index)
64 {
65 	return hns_roce_cmd_mbox(hr_dev, 0, mailbox ? mailbox->dma : 0,
66 				 mpt_index, !mailbox, HNS_ROCE_CMD_HW2SW_MPT,
67 				 HNS_ROCE_CMD_TIMEOUT_MSECS);
68 }
69 EXPORT_SYMBOL_GPL(hns_roce_hw2sw_mpt);
70 
71 static int hns_roce_buddy_alloc(struct hns_roce_buddy *buddy, int order,
72 				unsigned long *seg)
73 {
74 	int o;
75 	u32 m;
76 
77 	spin_lock(&buddy->lock);
78 
79 	for (o = order; o <= buddy->max_order; ++o) {
80 		if (buddy->num_free[o]) {
81 			m = 1 << (buddy->max_order - o);
82 			*seg = find_first_bit(buddy->bits[o], m);
83 			if (*seg < m)
84 				goto found;
85 		}
86 	}
87 	spin_unlock(&buddy->lock);
88 	return -1;
89 
90  found:
91 	clear_bit(*seg, buddy->bits[o]);
92 	--buddy->num_free[o];
93 
94 	while (o > order) {
95 		--o;
96 		*seg <<= 1;
97 		set_bit(*seg ^ 1, buddy->bits[o]);
98 		++buddy->num_free[o];
99 	}
100 
101 	spin_unlock(&buddy->lock);
102 
103 	*seg <<= order;
104 	return 0;
105 }
106 
107 static void hns_roce_buddy_free(struct hns_roce_buddy *buddy, unsigned long seg,
108 				int order)
109 {
110 	seg >>= order;
111 
112 	spin_lock(&buddy->lock);
113 
114 	while (test_bit(seg ^ 1, buddy->bits[order])) {
115 		clear_bit(seg ^ 1, buddy->bits[order]);
116 		--buddy->num_free[order];
117 		seg >>= 1;
118 		++order;
119 	}
120 
121 	set_bit(seg, buddy->bits[order]);
122 	++buddy->num_free[order];
123 
124 	spin_unlock(&buddy->lock);
125 }
126 
127 static int hns_roce_buddy_init(struct hns_roce_buddy *buddy, int max_order)
128 {
129 	int i, s;
130 
131 	buddy->max_order = max_order;
132 	spin_lock_init(&buddy->lock);
133 	buddy->bits = kcalloc(buddy->max_order + 1,
134 			      sizeof(*buddy->bits),
135 			      GFP_KERNEL);
136 	buddy->num_free = kcalloc(buddy->max_order + 1,
137 				  sizeof(*buddy->num_free),
138 				  GFP_KERNEL);
139 	if (!buddy->bits || !buddy->num_free)
140 		goto err_out;
141 
142 	for (i = 0; i <= buddy->max_order; ++i) {
143 		s = BITS_TO_LONGS(1 << (buddy->max_order - i));
144 		buddy->bits[i] = kcalloc(s, sizeof(long), GFP_KERNEL |
145 					 __GFP_NOWARN);
146 		if (!buddy->bits[i]) {
147 			buddy->bits[i] = vzalloc(array_size(s, sizeof(long)));
148 			if (!buddy->bits[i])
149 				goto err_out_free;
150 		}
151 	}
152 
153 	set_bit(0, buddy->bits[buddy->max_order]);
154 	buddy->num_free[buddy->max_order] = 1;
155 
156 	return 0;
157 
158 err_out_free:
159 	for (i = 0; i <= buddy->max_order; ++i)
160 		kvfree(buddy->bits[i]);
161 
162 err_out:
163 	kfree(buddy->bits);
164 	kfree(buddy->num_free);
165 	return -ENOMEM;
166 }
167 
168 static void hns_roce_buddy_cleanup(struct hns_roce_buddy *buddy)
169 {
170 	int i;
171 
172 	for (i = 0; i <= buddy->max_order; ++i)
173 		kvfree(buddy->bits[i]);
174 
175 	kfree(buddy->bits);
176 	kfree(buddy->num_free);
177 }
178 
179 static int hns_roce_alloc_mtt_range(struct hns_roce_dev *hr_dev, int order,
180 				    unsigned long *seg, u32 mtt_type)
181 {
182 	struct hns_roce_mr_table *mr_table = &hr_dev->mr_table;
183 	struct hns_roce_hem_table *table;
184 	struct hns_roce_buddy *buddy;
185 	int ret;
186 
187 	if (mtt_type == MTT_TYPE_WQE) {
188 		buddy = &mr_table->mtt_buddy;
189 		table = &mr_table->mtt_table;
190 	} else {
191 		buddy = &mr_table->mtt_cqe_buddy;
192 		table = &mr_table->mtt_cqe_table;
193 	}
194 
195 	ret = hns_roce_buddy_alloc(buddy, order, seg);
196 	if (ret == -1)
197 		return -1;
198 
199 	if (hns_roce_table_get_range(hr_dev, table, *seg,
200 				     *seg + (1 << order) - 1)) {
201 		hns_roce_buddy_free(buddy, *seg, order);
202 		return -1;
203 	}
204 
205 	return 0;
206 }
207 
208 int hns_roce_mtt_init(struct hns_roce_dev *hr_dev, int npages, int page_shift,
209 		      struct hns_roce_mtt *mtt)
210 {
211 	int ret;
212 	int i;
213 
214 	/* Page num is zero, correspond to DMA memory register */
215 	if (!npages) {
216 		mtt->order = -1;
217 		mtt->page_shift = HNS_ROCE_HEM_PAGE_SHIFT;
218 		return 0;
219 	}
220 
221 	/* Note: if page_shift is zero, FAST memory register */
222 	mtt->page_shift = page_shift;
223 
224 	/* Compute MTT entry necessary */
225 	for (mtt->order = 0, i = HNS_ROCE_MTT_ENTRY_PER_SEG; i < npages;
226 	     i <<= 1)
227 		++mtt->order;
228 
229 	/* Allocate MTT entry */
230 	ret = hns_roce_alloc_mtt_range(hr_dev, mtt->order, &mtt->first_seg,
231 				       mtt->mtt_type);
232 	if (ret == -1)
233 		return -ENOMEM;
234 
235 	return 0;
236 }
237 
238 void hns_roce_mtt_cleanup(struct hns_roce_dev *hr_dev, struct hns_roce_mtt *mtt)
239 {
240 	struct hns_roce_mr_table *mr_table = &hr_dev->mr_table;
241 
242 	if (mtt->order < 0)
243 		return;
244 
245 	if (mtt->mtt_type == MTT_TYPE_WQE) {
246 		hns_roce_buddy_free(&mr_table->mtt_buddy, mtt->first_seg,
247 				    mtt->order);
248 		hns_roce_table_put_range(hr_dev, &mr_table->mtt_table,
249 					mtt->first_seg,
250 					mtt->first_seg + (1 << mtt->order) - 1);
251 	} else {
252 		hns_roce_buddy_free(&mr_table->mtt_cqe_buddy, mtt->first_seg,
253 				    mtt->order);
254 		hns_roce_table_put_range(hr_dev, &mr_table->mtt_cqe_table,
255 					mtt->first_seg,
256 					mtt->first_seg + (1 << mtt->order) - 1);
257 	}
258 }
259 EXPORT_SYMBOL_GPL(hns_roce_mtt_cleanup);
260 
261 static void hns_roce_loop_free(struct hns_roce_dev *hr_dev,
262 			       struct hns_roce_mr *mr, int err_loop_index,
263 			       int loop_i, int loop_j)
264 {
265 	struct device *dev = hr_dev->dev;
266 	u32 mhop_num;
267 	u32 pbl_bt_sz;
268 	u64 bt_idx;
269 	int i, j;
270 
271 	pbl_bt_sz = 1 << (hr_dev->caps.pbl_ba_pg_sz + PAGE_SHIFT);
272 	mhop_num = hr_dev->caps.pbl_hop_num;
273 
274 	i = loop_i;
275 	if (mhop_num == 3 && err_loop_index == 2) {
276 		for (; i >= 0; i--) {
277 			dma_free_coherent(dev, pbl_bt_sz, mr->pbl_bt_l1[i],
278 					  mr->pbl_l1_dma_addr[i]);
279 
280 			for (j = 0; j < pbl_bt_sz / 8; j++) {
281 				if (i == loop_i && j >= loop_j)
282 					break;
283 
284 				bt_idx = i * pbl_bt_sz / 8 + j;
285 				dma_free_coherent(dev, pbl_bt_sz,
286 						  mr->pbl_bt_l2[bt_idx],
287 						  mr->pbl_l2_dma_addr[bt_idx]);
288 			}
289 		}
290 	} else if (mhop_num == 3 && err_loop_index == 1) {
291 		for (i -= 1; i >= 0; i--) {
292 			dma_free_coherent(dev, pbl_bt_sz, mr->pbl_bt_l1[i],
293 					  mr->pbl_l1_dma_addr[i]);
294 
295 			for (j = 0; j < pbl_bt_sz / 8; j++) {
296 				bt_idx = i * pbl_bt_sz / 8 + j;
297 				dma_free_coherent(dev, pbl_bt_sz,
298 						  mr->pbl_bt_l2[bt_idx],
299 						  mr->pbl_l2_dma_addr[bt_idx]);
300 			}
301 		}
302 	} else if (mhop_num == 2 && err_loop_index == 1) {
303 		for (i -= 1; i >= 0; i--)
304 			dma_free_coherent(dev, pbl_bt_sz, mr->pbl_bt_l1[i],
305 					  mr->pbl_l1_dma_addr[i]);
306 	} else {
307 		dev_warn(dev, "not support: mhop_num=%d, err_loop_index=%d.",
308 			 mhop_num, err_loop_index);
309 		return;
310 	}
311 
312 	dma_free_coherent(dev, pbl_bt_sz, mr->pbl_bt_l0, mr->pbl_l0_dma_addr);
313 	mr->pbl_bt_l0 = NULL;
314 	mr->pbl_l0_dma_addr = 0;
315 }
316 
317 /* PBL multi hop addressing */
318 static int hns_roce_mhop_alloc(struct hns_roce_dev *hr_dev, int npages,
319 			       struct hns_roce_mr *mr)
320 {
321 	struct device *dev = hr_dev->dev;
322 	int mr_alloc_done = 0;
323 	int npages_allocated;
324 	int i = 0, j = 0;
325 	u32 pbl_bt_sz;
326 	u32 mhop_num;
327 	u64 pbl_last_bt_num;
328 	u64 pbl_bt_cnt = 0;
329 	u64 bt_idx;
330 	u64 size;
331 
332 	mhop_num = hr_dev->caps.pbl_hop_num;
333 	pbl_bt_sz = 1 << (hr_dev->caps.pbl_ba_pg_sz + PAGE_SHIFT);
334 	pbl_last_bt_num = (npages + pbl_bt_sz / 8 - 1) / (pbl_bt_sz / 8);
335 
336 	if (mhop_num == HNS_ROCE_HOP_NUM_0)
337 		return 0;
338 
339 	/* hop_num = 1 */
340 	if (mhop_num == 1) {
341 		if (npages > pbl_bt_sz / 8) {
342 			dev_err(dev, "npages %d is larger than buf_pg_sz!",
343 				npages);
344 			return -EINVAL;
345 		}
346 		mr->pbl_buf = dma_alloc_coherent(dev, npages * 8,
347 						 &(mr->pbl_dma_addr),
348 						 GFP_KERNEL);
349 		if (!mr->pbl_buf)
350 			return -ENOMEM;
351 
352 		mr->pbl_size = npages;
353 		mr->pbl_ba = mr->pbl_dma_addr;
354 		mr->pbl_hop_num = hr_dev->caps.pbl_hop_num;
355 		mr->pbl_ba_pg_sz = hr_dev->caps.pbl_ba_pg_sz;
356 		mr->pbl_buf_pg_sz = hr_dev->caps.pbl_buf_pg_sz;
357 		return 0;
358 	}
359 
360 	mr->pbl_l1_dma_addr = kcalloc(pbl_bt_sz / 8,
361 				      sizeof(*mr->pbl_l1_dma_addr),
362 				      GFP_KERNEL);
363 	if (!mr->pbl_l1_dma_addr)
364 		return -ENOMEM;
365 
366 	mr->pbl_bt_l1 = kcalloc(pbl_bt_sz / 8, sizeof(*mr->pbl_bt_l1),
367 				GFP_KERNEL);
368 	if (!mr->pbl_bt_l1)
369 		goto err_kcalloc_bt_l1;
370 
371 	if (mhop_num == 3) {
372 		mr->pbl_l2_dma_addr = kcalloc(pbl_last_bt_num,
373 					      sizeof(*mr->pbl_l2_dma_addr),
374 					      GFP_KERNEL);
375 		if (!mr->pbl_l2_dma_addr)
376 			goto err_kcalloc_l2_dma;
377 
378 		mr->pbl_bt_l2 = kcalloc(pbl_last_bt_num,
379 					sizeof(*mr->pbl_bt_l2),
380 					GFP_KERNEL);
381 		if (!mr->pbl_bt_l2)
382 			goto err_kcalloc_bt_l2;
383 	}
384 
385 	/* alloc L0 BT */
386 	mr->pbl_bt_l0 = dma_alloc_coherent(dev, pbl_bt_sz,
387 					   &(mr->pbl_l0_dma_addr),
388 					   GFP_KERNEL);
389 	if (!mr->pbl_bt_l0)
390 		goto err_dma_alloc_l0;
391 
392 	if (mhop_num == 2) {
393 		/* alloc L1 BT */
394 		for (i = 0; i < pbl_bt_sz / 8; i++) {
395 			if (pbl_bt_cnt + 1 < pbl_last_bt_num) {
396 				size = pbl_bt_sz;
397 			} else {
398 				npages_allocated = i * (pbl_bt_sz / 8);
399 				size = (npages - npages_allocated) * 8;
400 			}
401 			mr->pbl_bt_l1[i] = dma_alloc_coherent(dev, size,
402 						    &(mr->pbl_l1_dma_addr[i]),
403 						    GFP_KERNEL);
404 			if (!mr->pbl_bt_l1[i]) {
405 				hns_roce_loop_free(hr_dev, mr, 1, i, 0);
406 				goto err_dma_alloc_l0;
407 			}
408 
409 			*(mr->pbl_bt_l0 + i) = mr->pbl_l1_dma_addr[i];
410 
411 			pbl_bt_cnt++;
412 			if (pbl_bt_cnt >= pbl_last_bt_num)
413 				break;
414 		}
415 	} else if (mhop_num == 3) {
416 		/* alloc L1, L2 BT */
417 		for (i = 0; i < pbl_bt_sz / 8; i++) {
418 			mr->pbl_bt_l1[i] = dma_alloc_coherent(dev, pbl_bt_sz,
419 						    &(mr->pbl_l1_dma_addr[i]),
420 						    GFP_KERNEL);
421 			if (!mr->pbl_bt_l1[i]) {
422 				hns_roce_loop_free(hr_dev, mr, 1, i, 0);
423 				goto err_dma_alloc_l0;
424 			}
425 
426 			*(mr->pbl_bt_l0 + i) = mr->pbl_l1_dma_addr[i];
427 
428 			for (j = 0; j < pbl_bt_sz / 8; j++) {
429 				bt_idx = i * pbl_bt_sz / 8 + j;
430 
431 				if (pbl_bt_cnt + 1 < pbl_last_bt_num) {
432 					size = pbl_bt_sz;
433 				} else {
434 					npages_allocated = bt_idx *
435 							   (pbl_bt_sz / 8);
436 					size = (npages - npages_allocated) * 8;
437 				}
438 				mr->pbl_bt_l2[bt_idx] = dma_alloc_coherent(
439 					      dev, size,
440 					      &(mr->pbl_l2_dma_addr[bt_idx]),
441 					      GFP_KERNEL);
442 				if (!mr->pbl_bt_l2[bt_idx]) {
443 					hns_roce_loop_free(hr_dev, mr, 2, i, j);
444 					goto err_dma_alloc_l0;
445 				}
446 
447 				*(mr->pbl_bt_l1[i] + j) =
448 						mr->pbl_l2_dma_addr[bt_idx];
449 
450 				pbl_bt_cnt++;
451 				if (pbl_bt_cnt >= pbl_last_bt_num) {
452 					mr_alloc_done = 1;
453 					break;
454 				}
455 			}
456 
457 			if (mr_alloc_done)
458 				break;
459 		}
460 	}
461 
462 	mr->l0_chunk_last_num = i + 1;
463 	if (mhop_num == 3)
464 		mr->l1_chunk_last_num = j + 1;
465 
466 	mr->pbl_size = npages;
467 	mr->pbl_ba = mr->pbl_l0_dma_addr;
468 	mr->pbl_hop_num = hr_dev->caps.pbl_hop_num;
469 	mr->pbl_ba_pg_sz = hr_dev->caps.pbl_ba_pg_sz;
470 	mr->pbl_buf_pg_sz = hr_dev->caps.pbl_buf_pg_sz;
471 
472 	return 0;
473 
474 err_dma_alloc_l0:
475 	kfree(mr->pbl_bt_l2);
476 	mr->pbl_bt_l2 = NULL;
477 
478 err_kcalloc_bt_l2:
479 	kfree(mr->pbl_l2_dma_addr);
480 	mr->pbl_l2_dma_addr = NULL;
481 
482 err_kcalloc_l2_dma:
483 	kfree(mr->pbl_bt_l1);
484 	mr->pbl_bt_l1 = NULL;
485 
486 err_kcalloc_bt_l1:
487 	kfree(mr->pbl_l1_dma_addr);
488 	mr->pbl_l1_dma_addr = NULL;
489 
490 	return -ENOMEM;
491 }
492 
493 static int hns_roce_mr_alloc(struct hns_roce_dev *hr_dev, u32 pd, u64 iova,
494 			     u64 size, u32 access, int npages,
495 			     struct hns_roce_mr *mr)
496 {
497 	struct device *dev = hr_dev->dev;
498 	unsigned long index = 0;
499 	int ret = 0;
500 
501 	/* Allocate a key for mr from mr_table */
502 	ret = hns_roce_bitmap_alloc(&hr_dev->mr_table.mtpt_bitmap, &index);
503 	if (ret == -1)
504 		return -ENOMEM;
505 
506 	mr->iova = iova;			/* MR va starting addr */
507 	mr->size = size;			/* MR addr range */
508 	mr->pd = pd;				/* MR num */
509 	mr->access = access;			/* MR access permit */
510 	mr->enabled = 0;			/* MR active status */
511 	mr->key = hw_index_to_key(index);	/* MR key */
512 
513 	if (size == ~0ull) {
514 		mr->type = MR_TYPE_DMA;
515 		mr->pbl_buf = NULL;
516 		mr->pbl_dma_addr = 0;
517 		/* PBL multi-hop addressing parameters */
518 		mr->pbl_bt_l2 = NULL;
519 		mr->pbl_bt_l1 = NULL;
520 		mr->pbl_bt_l0 = NULL;
521 		mr->pbl_l2_dma_addr = NULL;
522 		mr->pbl_l1_dma_addr = NULL;
523 		mr->pbl_l0_dma_addr = 0;
524 	} else {
525 		mr->type = MR_TYPE_MR;
526 		if (!hr_dev->caps.pbl_hop_num) {
527 			mr->pbl_buf = dma_alloc_coherent(dev, npages * 8,
528 							 &(mr->pbl_dma_addr),
529 							 GFP_KERNEL);
530 			if (!mr->pbl_buf)
531 				return -ENOMEM;
532 		} else {
533 			ret = hns_roce_mhop_alloc(hr_dev, npages, mr);
534 		}
535 	}
536 
537 	return ret;
538 }
539 
540 static void hns_roce_mhop_free(struct hns_roce_dev *hr_dev,
541 			       struct hns_roce_mr *mr)
542 {
543 	struct device *dev = hr_dev->dev;
544 	int npages_allocated;
545 	int npages;
546 	int i, j;
547 	u32 pbl_bt_sz;
548 	u32 mhop_num;
549 	u64 bt_idx;
550 
551 	npages = ib_umem_page_count(mr->umem);
552 	pbl_bt_sz = 1 << (hr_dev->caps.pbl_ba_pg_sz + PAGE_SHIFT);
553 	mhop_num = hr_dev->caps.pbl_hop_num;
554 
555 	if (mhop_num == HNS_ROCE_HOP_NUM_0)
556 		return;
557 
558 	/* hop_num = 1 */
559 	if (mhop_num == 1) {
560 		dma_free_coherent(dev, (unsigned int)(npages * 8),
561 				  mr->pbl_buf, mr->pbl_dma_addr);
562 		return;
563 	}
564 
565 	dma_free_coherent(dev, pbl_bt_sz, mr->pbl_bt_l0,
566 			  mr->pbl_l0_dma_addr);
567 
568 	if (mhop_num == 2) {
569 		for (i = 0; i < mr->l0_chunk_last_num; i++) {
570 			if (i == mr->l0_chunk_last_num - 1) {
571 				npages_allocated = i * (pbl_bt_sz / 8);
572 
573 				dma_free_coherent(dev,
574 					      (npages - npages_allocated) * 8,
575 					      mr->pbl_bt_l1[i],
576 					      mr->pbl_l1_dma_addr[i]);
577 
578 				break;
579 			}
580 
581 			dma_free_coherent(dev, pbl_bt_sz, mr->pbl_bt_l1[i],
582 					  mr->pbl_l1_dma_addr[i]);
583 		}
584 	} else if (mhop_num == 3) {
585 		for (i = 0; i < mr->l0_chunk_last_num; i++) {
586 			dma_free_coherent(dev, pbl_bt_sz, mr->pbl_bt_l1[i],
587 					  mr->pbl_l1_dma_addr[i]);
588 
589 			for (j = 0; j < pbl_bt_sz / 8; j++) {
590 				bt_idx = i * (pbl_bt_sz / 8) + j;
591 
592 				if ((i == mr->l0_chunk_last_num - 1)
593 				    && j == mr->l1_chunk_last_num - 1) {
594 					npages_allocated = bt_idx *
595 							   (pbl_bt_sz / 8);
596 
597 					dma_free_coherent(dev,
598 					      (npages - npages_allocated) * 8,
599 					      mr->pbl_bt_l2[bt_idx],
600 					      mr->pbl_l2_dma_addr[bt_idx]);
601 
602 					break;
603 				}
604 
605 				dma_free_coherent(dev, pbl_bt_sz,
606 						mr->pbl_bt_l2[bt_idx],
607 						mr->pbl_l2_dma_addr[bt_idx]);
608 			}
609 		}
610 	}
611 
612 	kfree(mr->pbl_bt_l1);
613 	kfree(mr->pbl_l1_dma_addr);
614 	mr->pbl_bt_l1 = NULL;
615 	mr->pbl_l1_dma_addr = NULL;
616 	if (mhop_num == 3) {
617 		kfree(mr->pbl_bt_l2);
618 		kfree(mr->pbl_l2_dma_addr);
619 		mr->pbl_bt_l2 = NULL;
620 		mr->pbl_l2_dma_addr = NULL;
621 	}
622 }
623 
624 static void hns_roce_mr_free(struct hns_roce_dev *hr_dev,
625 			     struct hns_roce_mr *mr)
626 {
627 	struct device *dev = hr_dev->dev;
628 	int npages = 0;
629 	int ret;
630 
631 	if (mr->enabled) {
632 		ret = hns_roce_hw2sw_mpt(hr_dev, NULL, key_to_hw_index(mr->key)
633 					 & (hr_dev->caps.num_mtpts - 1));
634 		if (ret)
635 			dev_warn(dev, "HW2SW_MPT failed (%d)\n", ret);
636 	}
637 
638 	if (mr->size != ~0ULL) {
639 		npages = ib_umem_page_count(mr->umem);
640 
641 		if (!hr_dev->caps.pbl_hop_num)
642 			dma_free_coherent(dev, (unsigned int)(npages * 8),
643 					  mr->pbl_buf, mr->pbl_dma_addr);
644 		else
645 			hns_roce_mhop_free(hr_dev, mr);
646 	}
647 
648 	if (mr->enabled)
649 		hns_roce_table_put(hr_dev, &hr_dev->mr_table.mtpt_table,
650 				   key_to_hw_index(mr->key));
651 
652 	hns_roce_bitmap_free(&hr_dev->mr_table.mtpt_bitmap,
653 			     key_to_hw_index(mr->key), BITMAP_NO_RR);
654 }
655 
656 static int hns_roce_mr_enable(struct hns_roce_dev *hr_dev,
657 			      struct hns_roce_mr *mr)
658 {
659 	int ret;
660 	unsigned long mtpt_idx = key_to_hw_index(mr->key);
661 	struct device *dev = hr_dev->dev;
662 	struct hns_roce_cmd_mailbox *mailbox;
663 	struct hns_roce_mr_table *mr_table = &hr_dev->mr_table;
664 
665 	/* Prepare HEM entry memory */
666 	ret = hns_roce_table_get(hr_dev, &mr_table->mtpt_table, mtpt_idx);
667 	if (ret)
668 		return ret;
669 
670 	/* Allocate mailbox memory */
671 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
672 	if (IS_ERR(mailbox)) {
673 		ret = PTR_ERR(mailbox);
674 		goto err_table;
675 	}
676 
677 	ret = hr_dev->hw->write_mtpt(mailbox->buf, mr, mtpt_idx);
678 	if (ret) {
679 		dev_err(dev, "Write mtpt fail!\n");
680 		goto err_page;
681 	}
682 
683 	ret = hns_roce_sw2hw_mpt(hr_dev, mailbox,
684 				 mtpt_idx & (hr_dev->caps.num_mtpts - 1));
685 	if (ret) {
686 		dev_err(dev, "SW2HW_MPT failed (%d)\n", ret);
687 		goto err_page;
688 	}
689 
690 	mr->enabled = 1;
691 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
692 
693 	return 0;
694 
695 err_page:
696 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
697 
698 err_table:
699 	hns_roce_table_put(hr_dev, &mr_table->mtpt_table, mtpt_idx);
700 	return ret;
701 }
702 
703 static int hns_roce_write_mtt_chunk(struct hns_roce_dev *hr_dev,
704 				    struct hns_roce_mtt *mtt, u32 start_index,
705 				    u32 npages, u64 *page_list)
706 {
707 	struct hns_roce_hem_table *table;
708 	dma_addr_t dma_handle;
709 	__le64 *mtts;
710 	u32 s = start_index * sizeof(u64);
711 	u32 bt_page_size;
712 	u32 i;
713 
714 	if (mtt->mtt_type == MTT_TYPE_WQE)
715 		bt_page_size = 1 << (hr_dev->caps.mtt_ba_pg_sz + PAGE_SHIFT);
716 	else
717 		bt_page_size = 1 << (hr_dev->caps.cqe_ba_pg_sz + PAGE_SHIFT);
718 
719 	/* All MTTs must fit in the same page */
720 	if (start_index / (bt_page_size / sizeof(u64)) !=
721 		(start_index + npages - 1) / (bt_page_size / sizeof(u64)))
722 		return -EINVAL;
723 
724 	if (start_index & (HNS_ROCE_MTT_ENTRY_PER_SEG - 1))
725 		return -EINVAL;
726 
727 	if (mtt->mtt_type == MTT_TYPE_WQE)
728 		table = &hr_dev->mr_table.mtt_table;
729 	else
730 		table = &hr_dev->mr_table.mtt_cqe_table;
731 
732 	mtts = hns_roce_table_find(hr_dev, table,
733 				mtt->first_seg + s / hr_dev->caps.mtt_entry_sz,
734 				&dma_handle);
735 	if (!mtts)
736 		return -ENOMEM;
737 
738 	/* Save page addr, low 12 bits : 0 */
739 	for (i = 0; i < npages; ++i) {
740 		if (!hr_dev->caps.mtt_hop_num)
741 			mtts[i] = cpu_to_le64(page_list[i] >> PAGE_ADDR_SHIFT);
742 		else
743 			mtts[i] = cpu_to_le64(page_list[i]);
744 	}
745 
746 	return 0;
747 }
748 
749 static int hns_roce_write_mtt(struct hns_roce_dev *hr_dev,
750 			      struct hns_roce_mtt *mtt, u32 start_index,
751 			      u32 npages, u64 *page_list)
752 {
753 	int chunk;
754 	int ret;
755 	u32 bt_page_size;
756 
757 	if (mtt->order < 0)
758 		return -EINVAL;
759 
760 	if (mtt->mtt_type == MTT_TYPE_WQE)
761 		bt_page_size = 1 << (hr_dev->caps.mtt_ba_pg_sz + PAGE_SHIFT);
762 	else
763 		bt_page_size = 1 << (hr_dev->caps.cqe_ba_pg_sz + PAGE_SHIFT);
764 
765 	while (npages > 0) {
766 		chunk = min_t(int, bt_page_size / sizeof(u64), npages);
767 
768 		ret = hns_roce_write_mtt_chunk(hr_dev, mtt, start_index, chunk,
769 					       page_list);
770 		if (ret)
771 			return ret;
772 
773 		npages -= chunk;
774 		start_index += chunk;
775 		page_list += chunk;
776 	}
777 
778 	return 0;
779 }
780 
781 int hns_roce_buf_write_mtt(struct hns_roce_dev *hr_dev,
782 			   struct hns_roce_mtt *mtt, struct hns_roce_buf *buf)
783 {
784 	u64 *page_list;
785 	int ret;
786 	u32 i;
787 
788 	page_list = kmalloc_array(buf->npages, sizeof(*page_list), GFP_KERNEL);
789 	if (!page_list)
790 		return -ENOMEM;
791 
792 	for (i = 0; i < buf->npages; ++i) {
793 		if (buf->nbufs == 1)
794 			page_list[i] = buf->direct.map + (i << buf->page_shift);
795 		else
796 			page_list[i] = buf->page_list[i].map;
797 
798 	}
799 	ret = hns_roce_write_mtt(hr_dev, mtt, 0, buf->npages, page_list);
800 
801 	kfree(page_list);
802 
803 	return ret;
804 }
805 
806 int hns_roce_init_mr_table(struct hns_roce_dev *hr_dev)
807 {
808 	struct hns_roce_mr_table *mr_table = &hr_dev->mr_table;
809 	int ret;
810 
811 	ret = hns_roce_bitmap_init(&mr_table->mtpt_bitmap,
812 				   hr_dev->caps.num_mtpts,
813 				   hr_dev->caps.num_mtpts - 1,
814 				   hr_dev->caps.reserved_mrws, 0);
815 	if (ret)
816 		return ret;
817 
818 	ret = hns_roce_buddy_init(&mr_table->mtt_buddy,
819 				  ilog2(hr_dev->caps.num_mtt_segs));
820 	if (ret)
821 		goto err_buddy;
822 
823 	if (hns_roce_check_whether_mhop(hr_dev, HEM_TYPE_CQE)) {
824 		ret = hns_roce_buddy_init(&mr_table->mtt_cqe_buddy,
825 					  ilog2(hr_dev->caps.num_cqe_segs));
826 		if (ret)
827 			goto err_buddy_cqe;
828 	}
829 	return 0;
830 
831 err_buddy_cqe:
832 	hns_roce_buddy_cleanup(&mr_table->mtt_buddy);
833 
834 err_buddy:
835 	hns_roce_bitmap_cleanup(&mr_table->mtpt_bitmap);
836 	return ret;
837 }
838 
839 void hns_roce_cleanup_mr_table(struct hns_roce_dev *hr_dev)
840 {
841 	struct hns_roce_mr_table *mr_table = &hr_dev->mr_table;
842 
843 	hns_roce_buddy_cleanup(&mr_table->mtt_buddy);
844 	if (hns_roce_check_whether_mhop(hr_dev, HEM_TYPE_CQE))
845 		hns_roce_buddy_cleanup(&mr_table->mtt_cqe_buddy);
846 	hns_roce_bitmap_cleanup(&mr_table->mtpt_bitmap);
847 }
848 
849 struct ib_mr *hns_roce_get_dma_mr(struct ib_pd *pd, int acc)
850 {
851 	struct hns_roce_mr *mr;
852 	int ret;
853 
854 	mr = kmalloc(sizeof(*mr), GFP_KERNEL);
855 	if (mr == NULL)
856 		return  ERR_PTR(-ENOMEM);
857 
858 	/* Allocate memory region key */
859 	ret = hns_roce_mr_alloc(to_hr_dev(pd->device), to_hr_pd(pd)->pdn, 0,
860 				~0ULL, acc, 0, mr);
861 	if (ret)
862 		goto err_free;
863 
864 	ret = hns_roce_mr_enable(to_hr_dev(pd->device), mr);
865 	if (ret)
866 		goto err_mr;
867 
868 	mr->ibmr.rkey = mr->ibmr.lkey = mr->key;
869 	mr->umem = NULL;
870 
871 	return &mr->ibmr;
872 
873 err_mr:
874 	hns_roce_mr_free(to_hr_dev(pd->device), mr);
875 
876 err_free:
877 	kfree(mr);
878 	return ERR_PTR(ret);
879 }
880 
881 int hns_roce_ib_umem_write_mtt(struct hns_roce_dev *hr_dev,
882 			       struct hns_roce_mtt *mtt, struct ib_umem *umem)
883 {
884 	struct device *dev = hr_dev->dev;
885 	struct scatterlist *sg;
886 	unsigned int order;
887 	int i, k, entry;
888 	int npage = 0;
889 	int ret = 0;
890 	int len;
891 	u64 page_addr;
892 	u64 *pages;
893 	u32 bt_page_size;
894 	u32 n;
895 
896 	order = mtt->mtt_type == MTT_TYPE_WQE ? hr_dev->caps.mtt_ba_pg_sz :
897 		hr_dev->caps.cqe_ba_pg_sz;
898 	bt_page_size = 1 << (order + PAGE_SHIFT);
899 
900 	pages = (u64 *) __get_free_pages(GFP_KERNEL, order);
901 	if (!pages)
902 		return -ENOMEM;
903 
904 	i = n = 0;
905 
906 	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
907 		len = sg_dma_len(sg) >> PAGE_SHIFT;
908 		for (k = 0; k < len; ++k) {
909 			page_addr =
910 				sg_dma_address(sg) + (k << umem->page_shift);
911 			if (!(npage % (1 << (mtt->page_shift - PAGE_SHIFT)))) {
912 				if (page_addr & ((1 << mtt->page_shift) - 1)) {
913 					dev_err(dev, "page_addr 0x%llx is not page_shift %d alignment!\n",
914 						page_addr, mtt->page_shift);
915 					ret = -EINVAL;
916 					goto out;
917 				}
918 				pages[i++] = page_addr;
919 			}
920 			npage++;
921 			if (i == bt_page_size / sizeof(u64)) {
922 				ret = hns_roce_write_mtt(hr_dev, mtt, n, i,
923 							 pages);
924 				if (ret)
925 					goto out;
926 				n += i;
927 				i = 0;
928 			}
929 		}
930 	}
931 
932 	if (i)
933 		ret = hns_roce_write_mtt(hr_dev, mtt, n, i, pages);
934 
935 out:
936 	free_pages((unsigned long) pages, order);
937 	return ret;
938 }
939 
940 static int hns_roce_ib_umem_write_mr(struct hns_roce_dev *hr_dev,
941 				     struct hns_roce_mr *mr,
942 				     struct ib_umem *umem)
943 {
944 	struct scatterlist *sg;
945 	int i = 0, j = 0, k;
946 	int entry;
947 	int len;
948 	u64 page_addr;
949 	u32 pbl_bt_sz;
950 
951 	if (hr_dev->caps.pbl_hop_num == HNS_ROCE_HOP_NUM_0)
952 		return 0;
953 
954 	pbl_bt_sz = 1 << (hr_dev->caps.pbl_ba_pg_sz + PAGE_SHIFT);
955 	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
956 		len = sg_dma_len(sg) >> PAGE_SHIFT;
957 		for (k = 0; k < len; ++k) {
958 			page_addr = sg_dma_address(sg) +
959 				    (k << umem->page_shift);
960 
961 			if (!hr_dev->caps.pbl_hop_num) {
962 				mr->pbl_buf[i++] = page_addr >> 12;
963 			} else if (hr_dev->caps.pbl_hop_num == 1) {
964 				mr->pbl_buf[i++] = page_addr;
965 			} else {
966 				if (hr_dev->caps.pbl_hop_num == 2)
967 					mr->pbl_bt_l1[i][j] = page_addr;
968 				else if (hr_dev->caps.pbl_hop_num == 3)
969 					mr->pbl_bt_l2[i][j] = page_addr;
970 
971 				j++;
972 				if (j >= (pbl_bt_sz / 8)) {
973 					i++;
974 					j = 0;
975 				}
976 			}
977 		}
978 	}
979 
980 	/* Memory barrier */
981 	mb();
982 
983 	return 0;
984 }
985 
986 struct ib_mr *hns_roce_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
987 				   u64 virt_addr, int access_flags,
988 				   struct ib_udata *udata)
989 {
990 	struct hns_roce_dev *hr_dev = to_hr_dev(pd->device);
991 	struct device *dev = hr_dev->dev;
992 	struct hns_roce_mr *mr;
993 	int bt_size;
994 	int ret;
995 	int n;
996 	int i;
997 
998 	mr = kmalloc(sizeof(*mr), GFP_KERNEL);
999 	if (!mr)
1000 		return ERR_PTR(-ENOMEM);
1001 
1002 	mr->umem = ib_umem_get(pd->uobject->context, start, length,
1003 			       access_flags, 0);
1004 	if (IS_ERR(mr->umem)) {
1005 		ret = PTR_ERR(mr->umem);
1006 		goto err_free;
1007 	}
1008 
1009 	n = ib_umem_page_count(mr->umem);
1010 
1011 	if (!hr_dev->caps.pbl_hop_num) {
1012 		if (n > HNS_ROCE_MAX_MTPT_PBL_NUM) {
1013 			dev_err(dev,
1014 			     " MR len %lld err. MR is limited to 4G at most!\n",
1015 			     length);
1016 			ret = -EINVAL;
1017 			goto err_umem;
1018 		}
1019 	} else {
1020 		int pbl_size = 1;
1021 
1022 		bt_size = (1 << (hr_dev->caps.pbl_ba_pg_sz + PAGE_SHIFT)) / 8;
1023 		for (i = 0; i < hr_dev->caps.pbl_hop_num; i++)
1024 			pbl_size *= bt_size;
1025 		if (n > pbl_size) {
1026 			dev_err(dev,
1027 			    " MR len %lld err. MR page num is limited to %d!\n",
1028 			    length, pbl_size);
1029 			ret = -EINVAL;
1030 			goto err_umem;
1031 		}
1032 	}
1033 
1034 	ret = hns_roce_mr_alloc(hr_dev, to_hr_pd(pd)->pdn, virt_addr, length,
1035 				access_flags, n, mr);
1036 	if (ret)
1037 		goto err_umem;
1038 
1039 	ret = hns_roce_ib_umem_write_mr(hr_dev, mr, mr->umem);
1040 	if (ret)
1041 		goto err_mr;
1042 
1043 	ret = hns_roce_mr_enable(hr_dev, mr);
1044 	if (ret)
1045 		goto err_mr;
1046 
1047 	mr->ibmr.rkey = mr->ibmr.lkey = mr->key;
1048 
1049 	return &mr->ibmr;
1050 
1051 err_mr:
1052 	hns_roce_mr_free(hr_dev, mr);
1053 
1054 err_umem:
1055 	ib_umem_release(mr->umem);
1056 
1057 err_free:
1058 	kfree(mr);
1059 	return ERR_PTR(ret);
1060 }
1061 
1062 int hns_roce_rereg_user_mr(struct ib_mr *ibmr, int flags, u64 start, u64 length,
1063 			   u64 virt_addr, int mr_access_flags, struct ib_pd *pd,
1064 			   struct ib_udata *udata)
1065 {
1066 	struct hns_roce_dev *hr_dev = to_hr_dev(ibmr->device);
1067 	struct hns_roce_mr *mr = to_hr_mr(ibmr);
1068 	struct hns_roce_cmd_mailbox *mailbox;
1069 	struct device *dev = hr_dev->dev;
1070 	unsigned long mtpt_idx;
1071 	u32 pdn = 0;
1072 	int npages;
1073 	int ret;
1074 
1075 	if (!mr->enabled)
1076 		return -EINVAL;
1077 
1078 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
1079 	if (IS_ERR(mailbox))
1080 		return PTR_ERR(mailbox);
1081 
1082 	mtpt_idx = key_to_hw_index(mr->key) & (hr_dev->caps.num_mtpts - 1);
1083 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, mtpt_idx, 0,
1084 				HNS_ROCE_CMD_QUERY_MPT,
1085 				HNS_ROCE_CMD_TIMEOUT_MSECS);
1086 	if (ret)
1087 		goto free_cmd_mbox;
1088 
1089 	ret = hns_roce_hw2sw_mpt(hr_dev, NULL, mtpt_idx);
1090 	if (ret)
1091 		dev_warn(dev, "HW2SW_MPT failed (%d)\n", ret);
1092 
1093 	mr->enabled = 0;
1094 
1095 	if (flags & IB_MR_REREG_PD)
1096 		pdn = to_hr_pd(pd)->pdn;
1097 
1098 	if (flags & IB_MR_REREG_TRANS) {
1099 		if (mr->size != ~0ULL) {
1100 			npages = ib_umem_page_count(mr->umem);
1101 
1102 			if (hr_dev->caps.pbl_hop_num)
1103 				hns_roce_mhop_free(hr_dev, mr);
1104 			else
1105 				dma_free_coherent(dev, npages * 8, mr->pbl_buf,
1106 						  mr->pbl_dma_addr);
1107 		}
1108 		ib_umem_release(mr->umem);
1109 
1110 		mr->umem = ib_umem_get(ibmr->uobject->context, start, length,
1111 				       mr_access_flags, 0);
1112 		if (IS_ERR(mr->umem)) {
1113 			ret = PTR_ERR(mr->umem);
1114 			mr->umem = NULL;
1115 			goto free_cmd_mbox;
1116 		}
1117 		npages = ib_umem_page_count(mr->umem);
1118 
1119 		if (hr_dev->caps.pbl_hop_num) {
1120 			ret = hns_roce_mhop_alloc(hr_dev, npages, mr);
1121 			if (ret)
1122 				goto release_umem;
1123 		} else {
1124 			mr->pbl_buf = dma_alloc_coherent(dev, npages * 8,
1125 							 &(mr->pbl_dma_addr),
1126 							 GFP_KERNEL);
1127 			if (!mr->pbl_buf) {
1128 				ret = -ENOMEM;
1129 				goto release_umem;
1130 			}
1131 		}
1132 	}
1133 
1134 	ret = hr_dev->hw->rereg_write_mtpt(hr_dev, mr, flags, pdn,
1135 					   mr_access_flags, virt_addr,
1136 					   length, mailbox->buf);
1137 	if (ret) {
1138 		if (flags & IB_MR_REREG_TRANS)
1139 			goto release_umem;
1140 		else
1141 			goto free_cmd_mbox;
1142 	}
1143 
1144 	if (flags & IB_MR_REREG_TRANS) {
1145 		ret = hns_roce_ib_umem_write_mr(hr_dev, mr, mr->umem);
1146 		if (ret) {
1147 			if (mr->size != ~0ULL) {
1148 				npages = ib_umem_page_count(mr->umem);
1149 
1150 				if (hr_dev->caps.pbl_hop_num)
1151 					hns_roce_mhop_free(hr_dev, mr);
1152 				else
1153 					dma_free_coherent(dev, npages * 8,
1154 							  mr->pbl_buf,
1155 							  mr->pbl_dma_addr);
1156 			}
1157 
1158 			goto release_umem;
1159 		}
1160 	}
1161 
1162 	ret = hns_roce_sw2hw_mpt(hr_dev, mailbox, mtpt_idx);
1163 	if (ret) {
1164 		dev_err(dev, "SW2HW_MPT failed (%d)\n", ret);
1165 		goto release_umem;
1166 	}
1167 
1168 	mr->enabled = 1;
1169 	if (flags & IB_MR_REREG_ACCESS)
1170 		mr->access = mr_access_flags;
1171 
1172 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
1173 
1174 	return 0;
1175 
1176 release_umem:
1177 	ib_umem_release(mr->umem);
1178 
1179 free_cmd_mbox:
1180 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
1181 
1182 	return ret;
1183 }
1184 
1185 int hns_roce_dereg_mr(struct ib_mr *ibmr)
1186 {
1187 	struct hns_roce_dev *hr_dev = to_hr_dev(ibmr->device);
1188 	struct hns_roce_mr *mr = to_hr_mr(ibmr);
1189 	int ret = 0;
1190 
1191 	if (hr_dev->hw->dereg_mr) {
1192 		ret = hr_dev->hw->dereg_mr(hr_dev, mr);
1193 	} else {
1194 		hns_roce_mr_free(hr_dev, mr);
1195 
1196 		if (mr->umem)
1197 			ib_umem_release(mr->umem);
1198 
1199 		kfree(mr);
1200 	}
1201 
1202 	return ret;
1203 }
1204