xref: /linux/drivers/infiniband/hw/hns/hns_roce_hem.c (revision 77ec462536a13d4b428a1eead725c4818a49f0b1)
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 "hns_roce_device.h"
36 #include "hns_roce_hem.h"
37 #include "hns_roce_common.h"
38 
39 #define DMA_ADDR_T_SHIFT		12
40 #define BT_BA_SHIFT			32
41 
42 #define HEM_INDEX_BUF			BIT(0)
43 #define HEM_INDEX_L0			BIT(1)
44 #define HEM_INDEX_L1			BIT(2)
45 struct hns_roce_hem_index {
46 	u64 buf;
47 	u64 l0;
48 	u64 l1;
49 	u32 inited; /* indicate which index is available */
50 };
51 
52 bool hns_roce_check_whether_mhop(struct hns_roce_dev *hr_dev, u32 type)
53 {
54 	int hop_num = 0;
55 
56 	switch (type) {
57 	case HEM_TYPE_QPC:
58 		hop_num = hr_dev->caps.qpc_hop_num;
59 		break;
60 	case HEM_TYPE_MTPT:
61 		hop_num = hr_dev->caps.mpt_hop_num;
62 		break;
63 	case HEM_TYPE_CQC:
64 		hop_num = hr_dev->caps.cqc_hop_num;
65 		break;
66 	case HEM_TYPE_SRQC:
67 		hop_num = hr_dev->caps.srqc_hop_num;
68 		break;
69 	case HEM_TYPE_SCCC:
70 		hop_num = hr_dev->caps.sccc_hop_num;
71 		break;
72 	case HEM_TYPE_QPC_TIMER:
73 		hop_num = hr_dev->caps.qpc_timer_hop_num;
74 		break;
75 	case HEM_TYPE_CQC_TIMER:
76 		hop_num = hr_dev->caps.cqc_timer_hop_num;
77 		break;
78 	case HEM_TYPE_GMV:
79 		hop_num = hr_dev->caps.gmv_hop_num;
80 		break;
81 	default:
82 		return false;
83 	}
84 
85 	return hop_num ? true : false;
86 }
87 
88 static bool hns_roce_check_hem_null(struct hns_roce_hem **hem, u64 hem_idx,
89 				    u32 bt_chunk_num, u64 hem_max_num)
90 {
91 	u64 start_idx = round_down(hem_idx, bt_chunk_num);
92 	u64 check_max_num = start_idx + bt_chunk_num;
93 	u64 i;
94 
95 	for (i = start_idx; (i < check_max_num) && (i < hem_max_num); i++)
96 		if (i != hem_idx && hem[i])
97 			return false;
98 
99 	return true;
100 }
101 
102 static bool hns_roce_check_bt_null(u64 **bt, u64 ba_idx, u32 bt_chunk_num)
103 {
104 	u64 start_idx = round_down(ba_idx, bt_chunk_num);
105 	int i;
106 
107 	for (i = 0; i < bt_chunk_num; i++)
108 		if (i != ba_idx && bt[start_idx + i])
109 			return false;
110 
111 	return true;
112 }
113 
114 static int hns_roce_get_bt_num(u32 table_type, u32 hop_num)
115 {
116 	if (check_whether_bt_num_3(table_type, hop_num))
117 		return 3;
118 	else if (check_whether_bt_num_2(table_type, hop_num))
119 		return 2;
120 	else if (check_whether_bt_num_1(table_type, hop_num))
121 		return 1;
122 	else
123 		return 0;
124 }
125 
126 static int get_hem_table_config(struct hns_roce_dev *hr_dev,
127 				struct hns_roce_hem_mhop *mhop,
128 				u32 type)
129 {
130 	struct device *dev = hr_dev->dev;
131 
132 	switch (type) {
133 	case HEM_TYPE_QPC:
134 		mhop->buf_chunk_size = 1 << (hr_dev->caps.qpc_buf_pg_sz
135 					     + PAGE_SHIFT);
136 		mhop->bt_chunk_size = 1 << (hr_dev->caps.qpc_ba_pg_sz
137 					     + PAGE_SHIFT);
138 		mhop->ba_l0_num = hr_dev->caps.qpc_bt_num;
139 		mhop->hop_num = hr_dev->caps.qpc_hop_num;
140 		break;
141 	case HEM_TYPE_MTPT:
142 		mhop->buf_chunk_size = 1 << (hr_dev->caps.mpt_buf_pg_sz
143 					     + PAGE_SHIFT);
144 		mhop->bt_chunk_size = 1 << (hr_dev->caps.mpt_ba_pg_sz
145 					     + PAGE_SHIFT);
146 		mhop->ba_l0_num = hr_dev->caps.mpt_bt_num;
147 		mhop->hop_num = hr_dev->caps.mpt_hop_num;
148 		break;
149 	case HEM_TYPE_CQC:
150 		mhop->buf_chunk_size = 1 << (hr_dev->caps.cqc_buf_pg_sz
151 					     + PAGE_SHIFT);
152 		mhop->bt_chunk_size = 1 << (hr_dev->caps.cqc_ba_pg_sz
153 					    + PAGE_SHIFT);
154 		mhop->ba_l0_num = hr_dev->caps.cqc_bt_num;
155 		mhop->hop_num = hr_dev->caps.cqc_hop_num;
156 		break;
157 	case HEM_TYPE_SCCC:
158 		mhop->buf_chunk_size = 1 << (hr_dev->caps.sccc_buf_pg_sz
159 					     + PAGE_SHIFT);
160 		mhop->bt_chunk_size = 1 << (hr_dev->caps.sccc_ba_pg_sz
161 					    + PAGE_SHIFT);
162 		mhop->ba_l0_num = hr_dev->caps.sccc_bt_num;
163 		mhop->hop_num = hr_dev->caps.sccc_hop_num;
164 		break;
165 	case HEM_TYPE_QPC_TIMER:
166 		mhop->buf_chunk_size = 1 << (hr_dev->caps.qpc_timer_buf_pg_sz
167 					     + PAGE_SHIFT);
168 		mhop->bt_chunk_size = 1 << (hr_dev->caps.qpc_timer_ba_pg_sz
169 					    + PAGE_SHIFT);
170 		mhop->ba_l0_num = hr_dev->caps.qpc_timer_bt_num;
171 		mhop->hop_num = hr_dev->caps.qpc_timer_hop_num;
172 		break;
173 	case HEM_TYPE_CQC_TIMER:
174 		mhop->buf_chunk_size = 1 << (hr_dev->caps.cqc_timer_buf_pg_sz
175 					     + PAGE_SHIFT);
176 		mhop->bt_chunk_size = 1 << (hr_dev->caps.cqc_timer_ba_pg_sz
177 					    + PAGE_SHIFT);
178 		mhop->ba_l0_num = hr_dev->caps.cqc_timer_bt_num;
179 		mhop->hop_num = hr_dev->caps.cqc_timer_hop_num;
180 		break;
181 	case HEM_TYPE_SRQC:
182 		mhop->buf_chunk_size = 1 << (hr_dev->caps.srqc_buf_pg_sz
183 					     + PAGE_SHIFT);
184 		mhop->bt_chunk_size = 1 << (hr_dev->caps.srqc_ba_pg_sz
185 					     + PAGE_SHIFT);
186 		mhop->ba_l0_num = hr_dev->caps.srqc_bt_num;
187 		mhop->hop_num = hr_dev->caps.srqc_hop_num;
188 		break;
189 	case HEM_TYPE_GMV:
190 		mhop->buf_chunk_size = 1 << (hr_dev->caps.gmv_buf_pg_sz +
191 					     PAGE_SHIFT);
192 		mhop->bt_chunk_size = 1 << (hr_dev->caps.gmv_ba_pg_sz +
193 					    PAGE_SHIFT);
194 		mhop->ba_l0_num = hr_dev->caps.gmv_bt_num;
195 		mhop->hop_num = hr_dev->caps.gmv_hop_num;
196 		break;
197 	default:
198 		dev_err(dev, "table %u not support multi-hop addressing!\n",
199 			type);
200 		return -EINVAL;
201 	}
202 
203 	return 0;
204 }
205 
206 int hns_roce_calc_hem_mhop(struct hns_roce_dev *hr_dev,
207 			   struct hns_roce_hem_table *table, unsigned long *obj,
208 			   struct hns_roce_hem_mhop *mhop)
209 {
210 	struct device *dev = hr_dev->dev;
211 	u32 chunk_ba_num;
212 	u32 chunk_size;
213 	u32 table_idx;
214 	u32 bt_num;
215 
216 	if (get_hem_table_config(hr_dev, mhop, table->type))
217 		return -EINVAL;
218 
219 	if (!obj)
220 		return 0;
221 
222 	/*
223 	 * QPC/MTPT/CQC/SRQC/SCCC alloc hem for buffer pages.
224 	 * MTT/CQE alloc hem for bt pages.
225 	 */
226 	bt_num = hns_roce_get_bt_num(table->type, mhop->hop_num);
227 	chunk_ba_num = mhop->bt_chunk_size / BA_BYTE_LEN;
228 	chunk_size = table->type < HEM_TYPE_MTT ? mhop->buf_chunk_size :
229 			      mhop->bt_chunk_size;
230 	table_idx = (*obj & (table->num_obj - 1)) /
231 		     (chunk_size / table->obj_size);
232 	switch (bt_num) {
233 	case 3:
234 		mhop->l2_idx = table_idx & (chunk_ba_num - 1);
235 		mhop->l1_idx = table_idx / chunk_ba_num & (chunk_ba_num - 1);
236 		mhop->l0_idx = (table_idx / chunk_ba_num) / chunk_ba_num;
237 		break;
238 	case 2:
239 		mhop->l1_idx = table_idx & (chunk_ba_num - 1);
240 		mhop->l0_idx = table_idx / chunk_ba_num;
241 		break;
242 	case 1:
243 		mhop->l0_idx = table_idx;
244 		break;
245 	default:
246 		dev_err(dev, "table %u not support hop_num = %u!\n",
247 			table->type, mhop->hop_num);
248 		return -EINVAL;
249 	}
250 	if (mhop->l0_idx >= mhop->ba_l0_num)
251 		mhop->l0_idx %= mhop->ba_l0_num;
252 
253 	return 0;
254 }
255 
256 static struct hns_roce_hem *hns_roce_alloc_hem(struct hns_roce_dev *hr_dev,
257 					       int npages,
258 					       unsigned long hem_alloc_size,
259 					       gfp_t gfp_mask)
260 {
261 	struct hns_roce_hem_chunk *chunk = NULL;
262 	struct hns_roce_hem *hem;
263 	struct scatterlist *mem;
264 	int order;
265 	void *buf;
266 
267 	WARN_ON(gfp_mask & __GFP_HIGHMEM);
268 
269 	hem = kmalloc(sizeof(*hem),
270 		      gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
271 	if (!hem)
272 		return NULL;
273 
274 	hem->refcount = 0;
275 	INIT_LIST_HEAD(&hem->chunk_list);
276 
277 	order = get_order(hem_alloc_size);
278 
279 	while (npages > 0) {
280 		if (!chunk) {
281 			chunk = kmalloc(sizeof(*chunk),
282 				gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
283 			if (!chunk)
284 				goto fail;
285 
286 			sg_init_table(chunk->mem, HNS_ROCE_HEM_CHUNK_LEN);
287 			chunk->npages = 0;
288 			chunk->nsg = 0;
289 			memset(chunk->buf, 0, sizeof(chunk->buf));
290 			list_add_tail(&chunk->list, &hem->chunk_list);
291 		}
292 
293 		while (1 << order > npages)
294 			--order;
295 
296 		/*
297 		 * Alloc memory one time. If failed, don't alloc small block
298 		 * memory, directly return fail.
299 		 */
300 		mem = &chunk->mem[chunk->npages];
301 		buf = dma_alloc_coherent(hr_dev->dev, PAGE_SIZE << order,
302 				&sg_dma_address(mem), gfp_mask);
303 		if (!buf)
304 			goto fail;
305 
306 		chunk->buf[chunk->npages] = buf;
307 		sg_dma_len(mem) = PAGE_SIZE << order;
308 
309 		++chunk->npages;
310 		++chunk->nsg;
311 		npages -= 1 << order;
312 	}
313 
314 	return hem;
315 
316 fail:
317 	hns_roce_free_hem(hr_dev, hem);
318 	return NULL;
319 }
320 
321 void hns_roce_free_hem(struct hns_roce_dev *hr_dev, struct hns_roce_hem *hem)
322 {
323 	struct hns_roce_hem_chunk *chunk, *tmp;
324 	int i;
325 
326 	if (!hem)
327 		return;
328 
329 	list_for_each_entry_safe(chunk, tmp, &hem->chunk_list, list) {
330 		for (i = 0; i < chunk->npages; ++i)
331 			dma_free_coherent(hr_dev->dev,
332 				   sg_dma_len(&chunk->mem[i]),
333 				   chunk->buf[i],
334 				   sg_dma_address(&chunk->mem[i]));
335 		kfree(chunk);
336 	}
337 
338 	kfree(hem);
339 }
340 
341 static int hns_roce_set_hem(struct hns_roce_dev *hr_dev,
342 			    struct hns_roce_hem_table *table, unsigned long obj)
343 {
344 	spinlock_t *lock = &hr_dev->bt_cmd_lock;
345 	struct device *dev = hr_dev->dev;
346 	struct hns_roce_hem_iter iter;
347 	void __iomem *bt_cmd;
348 	__le32 bt_cmd_val[2];
349 	__le32 bt_cmd_h = 0;
350 	unsigned long flags;
351 	__le32 bt_cmd_l;
352 	int ret = 0;
353 	u64 bt_ba;
354 	long end;
355 
356 	/* Find the HEM(Hardware Entry Memory) entry */
357 	unsigned long i = (obj & (table->num_obj - 1)) /
358 			  (table->table_chunk_size / table->obj_size);
359 
360 	switch (table->type) {
361 	case HEM_TYPE_QPC:
362 	case HEM_TYPE_MTPT:
363 	case HEM_TYPE_CQC:
364 	case HEM_TYPE_SRQC:
365 		roce_set_field(bt_cmd_h, ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_M,
366 			ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_S, table->type);
367 		break;
368 	default:
369 		return ret;
370 	}
371 
372 	roce_set_field(bt_cmd_h, ROCEE_BT_CMD_H_ROCEE_BT_CMD_IN_MDF_M,
373 		       ROCEE_BT_CMD_H_ROCEE_BT_CMD_IN_MDF_S, obj);
374 	roce_set_bit(bt_cmd_h, ROCEE_BT_CMD_H_ROCEE_BT_CMD_S, 0);
375 	roce_set_bit(bt_cmd_h, ROCEE_BT_CMD_H_ROCEE_BT_CMD_HW_SYNS_S, 1);
376 
377 	/* Currently iter only a chunk */
378 	for (hns_roce_hem_first(table->hem[i], &iter);
379 	     !hns_roce_hem_last(&iter); hns_roce_hem_next(&iter)) {
380 		bt_ba = hns_roce_hem_addr(&iter) >> DMA_ADDR_T_SHIFT;
381 
382 		spin_lock_irqsave(lock, flags);
383 
384 		bt_cmd = hr_dev->reg_base + ROCEE_BT_CMD_H_REG;
385 
386 		end = HW_SYNC_TIMEOUT_MSECS;
387 		while (end > 0) {
388 			if (!(readl(bt_cmd) >> BT_CMD_SYNC_SHIFT))
389 				break;
390 
391 			mdelay(HW_SYNC_SLEEP_TIME_INTERVAL);
392 			end -= HW_SYNC_SLEEP_TIME_INTERVAL;
393 		}
394 
395 		if (end <= 0) {
396 			dev_err(dev, "Write bt_cmd err,hw_sync is not zero.\n");
397 			spin_unlock_irqrestore(lock, flags);
398 			return -EBUSY;
399 		}
400 
401 		bt_cmd_l = cpu_to_le32(bt_ba);
402 		roce_set_field(bt_cmd_h, ROCEE_BT_CMD_H_ROCEE_BT_CMD_BA_H_M,
403 			       ROCEE_BT_CMD_H_ROCEE_BT_CMD_BA_H_S,
404 			       bt_ba >> BT_BA_SHIFT);
405 
406 		bt_cmd_val[0] = bt_cmd_l;
407 		bt_cmd_val[1] = bt_cmd_h;
408 		hns_roce_write64_k(bt_cmd_val,
409 				   hr_dev->reg_base + ROCEE_BT_CMD_L_REG);
410 		spin_unlock_irqrestore(lock, flags);
411 	}
412 
413 	return ret;
414 }
415 
416 static int calc_hem_config(struct hns_roce_dev *hr_dev,
417 			   struct hns_roce_hem_table *table, unsigned long obj,
418 			   struct hns_roce_hem_mhop *mhop,
419 			   struct hns_roce_hem_index *index)
420 {
421 	struct ib_device *ibdev = &hr_dev->ib_dev;
422 	unsigned long mhop_obj = obj;
423 	u32 l0_idx, l1_idx, l2_idx;
424 	u32 chunk_ba_num;
425 	u32 bt_num;
426 	int ret;
427 
428 	ret = hns_roce_calc_hem_mhop(hr_dev, table, &mhop_obj, mhop);
429 	if (ret)
430 		return ret;
431 
432 	l0_idx = mhop->l0_idx;
433 	l1_idx = mhop->l1_idx;
434 	l2_idx = mhop->l2_idx;
435 	chunk_ba_num = mhop->bt_chunk_size / BA_BYTE_LEN;
436 	bt_num = hns_roce_get_bt_num(table->type, mhop->hop_num);
437 	switch (bt_num) {
438 	case 3:
439 		index->l1 = l0_idx * chunk_ba_num + l1_idx;
440 		index->l0 = l0_idx;
441 		index->buf = l0_idx * chunk_ba_num * chunk_ba_num +
442 			     l1_idx * chunk_ba_num + l2_idx;
443 		break;
444 	case 2:
445 		index->l0 = l0_idx;
446 		index->buf = l0_idx * chunk_ba_num + l1_idx;
447 		break;
448 	case 1:
449 		index->buf = l0_idx;
450 		break;
451 	default:
452 		ibdev_err(ibdev, "table %u not support mhop.hop_num = %u!\n",
453 			  table->type, mhop->hop_num);
454 		return -EINVAL;
455 	}
456 
457 	if (unlikely(index->buf >= table->num_hem)) {
458 		ibdev_err(ibdev, "table %u exceed hem limt idx %llu, max %lu!\n",
459 			  table->type, index->buf, table->num_hem);
460 		return -EINVAL;
461 	}
462 
463 	return 0;
464 }
465 
466 static void free_mhop_hem(struct hns_roce_dev *hr_dev,
467 			  struct hns_roce_hem_table *table,
468 			  struct hns_roce_hem_mhop *mhop,
469 			  struct hns_roce_hem_index *index)
470 {
471 	u32 bt_size = mhop->bt_chunk_size;
472 	struct device *dev = hr_dev->dev;
473 
474 	if (index->inited & HEM_INDEX_BUF) {
475 		hns_roce_free_hem(hr_dev, table->hem[index->buf]);
476 		table->hem[index->buf] = NULL;
477 	}
478 
479 	if (index->inited & HEM_INDEX_L1) {
480 		dma_free_coherent(dev, bt_size, table->bt_l1[index->l1],
481 				  table->bt_l1_dma_addr[index->l1]);
482 		table->bt_l1[index->l1] = NULL;
483 	}
484 
485 	if (index->inited & HEM_INDEX_L0) {
486 		dma_free_coherent(dev, bt_size, table->bt_l0[index->l0],
487 				  table->bt_l0_dma_addr[index->l0]);
488 		table->bt_l0[index->l0] = NULL;
489 	}
490 }
491 
492 static int alloc_mhop_hem(struct hns_roce_dev *hr_dev,
493 			  struct hns_roce_hem_table *table,
494 			  struct hns_roce_hem_mhop *mhop,
495 			  struct hns_roce_hem_index *index)
496 {
497 	u32 bt_size = mhop->bt_chunk_size;
498 	struct device *dev = hr_dev->dev;
499 	struct hns_roce_hem_iter iter;
500 	gfp_t flag;
501 	u64 bt_ba;
502 	u32 size;
503 	int ret;
504 
505 	/* alloc L1 BA's chunk */
506 	if ((check_whether_bt_num_3(table->type, mhop->hop_num) ||
507 	     check_whether_bt_num_2(table->type, mhop->hop_num)) &&
508 	     !table->bt_l0[index->l0]) {
509 		table->bt_l0[index->l0] = dma_alloc_coherent(dev, bt_size,
510 					    &table->bt_l0_dma_addr[index->l0],
511 					    GFP_KERNEL);
512 		if (!table->bt_l0[index->l0]) {
513 			ret = -ENOMEM;
514 			goto out;
515 		}
516 		index->inited |= HEM_INDEX_L0;
517 	}
518 
519 	/* alloc L2 BA's chunk */
520 	if (check_whether_bt_num_3(table->type, mhop->hop_num) &&
521 	    !table->bt_l1[index->l1])  {
522 		table->bt_l1[index->l1] = dma_alloc_coherent(dev, bt_size,
523 					    &table->bt_l1_dma_addr[index->l1],
524 					    GFP_KERNEL);
525 		if (!table->bt_l1[index->l1]) {
526 			ret = -ENOMEM;
527 			goto err_alloc_hem;
528 		}
529 		index->inited |= HEM_INDEX_L1;
530 		*(table->bt_l0[index->l0] + mhop->l1_idx) =
531 					       table->bt_l1_dma_addr[index->l1];
532 	}
533 
534 	/*
535 	 * alloc buffer space chunk for QPC/MTPT/CQC/SRQC/SCCC.
536 	 * alloc bt space chunk for MTT/CQE.
537 	 */
538 	size = table->type < HEM_TYPE_MTT ? mhop->buf_chunk_size : bt_size;
539 	flag = (table->lowmem ? GFP_KERNEL : GFP_HIGHUSER) | __GFP_NOWARN;
540 	table->hem[index->buf] = hns_roce_alloc_hem(hr_dev, size >> PAGE_SHIFT,
541 						    size, flag);
542 	if (!table->hem[index->buf]) {
543 		ret = -ENOMEM;
544 		goto err_alloc_hem;
545 	}
546 
547 	index->inited |= HEM_INDEX_BUF;
548 	hns_roce_hem_first(table->hem[index->buf], &iter);
549 	bt_ba = hns_roce_hem_addr(&iter);
550 	if (table->type < HEM_TYPE_MTT) {
551 		if (mhop->hop_num == 2)
552 			*(table->bt_l1[index->l1] + mhop->l2_idx) = bt_ba;
553 		else if (mhop->hop_num == 1)
554 			*(table->bt_l0[index->l0] + mhop->l1_idx) = bt_ba;
555 	} else if (mhop->hop_num == 2) {
556 		*(table->bt_l0[index->l0] + mhop->l1_idx) = bt_ba;
557 	}
558 
559 	return 0;
560 err_alloc_hem:
561 	free_mhop_hem(hr_dev, table, mhop, index);
562 out:
563 	return ret;
564 }
565 
566 static int set_mhop_hem(struct hns_roce_dev *hr_dev,
567 			struct hns_roce_hem_table *table, unsigned long obj,
568 			struct hns_roce_hem_mhop *mhop,
569 			struct hns_roce_hem_index *index)
570 {
571 	struct ib_device *ibdev = &hr_dev->ib_dev;
572 	int step_idx;
573 	int ret = 0;
574 
575 	if (index->inited & HEM_INDEX_L0) {
576 		ret = hr_dev->hw->set_hem(hr_dev, table, obj, 0);
577 		if (ret) {
578 			ibdev_err(ibdev, "set HEM step 0 failed!\n");
579 			goto out;
580 		}
581 	}
582 
583 	if (index->inited & HEM_INDEX_L1) {
584 		ret = hr_dev->hw->set_hem(hr_dev, table, obj, 1);
585 		if (ret) {
586 			ibdev_err(ibdev, "set HEM step 1 failed!\n");
587 			goto out;
588 		}
589 	}
590 
591 	if (index->inited & HEM_INDEX_BUF) {
592 		if (mhop->hop_num == HNS_ROCE_HOP_NUM_0)
593 			step_idx = 0;
594 		else
595 			step_idx = mhop->hop_num;
596 		ret = hr_dev->hw->set_hem(hr_dev, table, obj, step_idx);
597 		if (ret)
598 			ibdev_err(ibdev, "set HEM step last failed!\n");
599 	}
600 out:
601 	return ret;
602 }
603 
604 static int hns_roce_table_mhop_get(struct hns_roce_dev *hr_dev,
605 				   struct hns_roce_hem_table *table,
606 				   unsigned long obj)
607 {
608 	struct ib_device *ibdev = &hr_dev->ib_dev;
609 	struct hns_roce_hem_index index = {};
610 	struct hns_roce_hem_mhop mhop = {};
611 	int ret;
612 
613 	ret = calc_hem_config(hr_dev, table, obj, &mhop, &index);
614 	if (ret) {
615 		ibdev_err(ibdev, "calc hem config failed!\n");
616 		return ret;
617 	}
618 
619 	mutex_lock(&table->mutex);
620 	if (table->hem[index.buf]) {
621 		++table->hem[index.buf]->refcount;
622 		goto out;
623 	}
624 
625 	ret = alloc_mhop_hem(hr_dev, table, &mhop, &index);
626 	if (ret) {
627 		ibdev_err(ibdev, "alloc mhop hem failed!\n");
628 		goto out;
629 	}
630 
631 	/* set HEM base address to hardware */
632 	if (table->type < HEM_TYPE_MTT) {
633 		ret = set_mhop_hem(hr_dev, table, obj, &mhop, &index);
634 		if (ret) {
635 			ibdev_err(ibdev, "set HEM address to HW failed!\n");
636 			goto err_alloc;
637 		}
638 	}
639 
640 	++table->hem[index.buf]->refcount;
641 	goto out;
642 
643 err_alloc:
644 	free_mhop_hem(hr_dev, table, &mhop, &index);
645 out:
646 	mutex_unlock(&table->mutex);
647 	return ret;
648 }
649 
650 int hns_roce_table_get(struct hns_roce_dev *hr_dev,
651 		       struct hns_roce_hem_table *table, unsigned long obj)
652 {
653 	struct device *dev = hr_dev->dev;
654 	unsigned long i;
655 	int ret = 0;
656 
657 	if (hns_roce_check_whether_mhop(hr_dev, table->type))
658 		return hns_roce_table_mhop_get(hr_dev, table, obj);
659 
660 	i = (obj & (table->num_obj - 1)) / (table->table_chunk_size /
661 	     table->obj_size);
662 
663 	mutex_lock(&table->mutex);
664 
665 	if (table->hem[i]) {
666 		++table->hem[i]->refcount;
667 		goto out;
668 	}
669 
670 	table->hem[i] = hns_roce_alloc_hem(hr_dev,
671 				       table->table_chunk_size >> PAGE_SHIFT,
672 				       table->table_chunk_size,
673 				       (table->lowmem ? GFP_KERNEL :
674 					GFP_HIGHUSER) | __GFP_NOWARN);
675 	if (!table->hem[i]) {
676 		ret = -ENOMEM;
677 		goto out;
678 	}
679 
680 	/* Set HEM base address(128K/page, pa) to Hardware */
681 	if (hns_roce_set_hem(hr_dev, table, obj)) {
682 		hns_roce_free_hem(hr_dev, table->hem[i]);
683 		table->hem[i] = NULL;
684 		ret = -ENODEV;
685 		dev_err(dev, "set HEM base address to HW failed.\n");
686 		goto out;
687 	}
688 
689 	++table->hem[i]->refcount;
690 out:
691 	mutex_unlock(&table->mutex);
692 	return ret;
693 }
694 
695 static void clear_mhop_hem(struct hns_roce_dev *hr_dev,
696 			   struct hns_roce_hem_table *table, unsigned long obj,
697 			   struct hns_roce_hem_mhop *mhop,
698 			   struct hns_roce_hem_index *index)
699 {
700 	struct ib_device *ibdev = &hr_dev->ib_dev;
701 	u32 hop_num = mhop->hop_num;
702 	u32 chunk_ba_num;
703 	int step_idx;
704 
705 	index->inited = HEM_INDEX_BUF;
706 	chunk_ba_num = mhop->bt_chunk_size / BA_BYTE_LEN;
707 	if (check_whether_bt_num_2(table->type, hop_num)) {
708 		if (hns_roce_check_hem_null(table->hem, index->buf,
709 					    chunk_ba_num, table->num_hem))
710 			index->inited |= HEM_INDEX_L0;
711 	} else if (check_whether_bt_num_3(table->type, hop_num)) {
712 		if (hns_roce_check_hem_null(table->hem, index->buf,
713 					    chunk_ba_num, table->num_hem)) {
714 			index->inited |= HEM_INDEX_L1;
715 			if (hns_roce_check_bt_null(table->bt_l1, index->l1,
716 						   chunk_ba_num))
717 				index->inited |= HEM_INDEX_L0;
718 		}
719 	}
720 
721 	if (table->type < HEM_TYPE_MTT) {
722 		if (hop_num == HNS_ROCE_HOP_NUM_0)
723 			step_idx = 0;
724 		else
725 			step_idx = hop_num;
726 
727 		if (hr_dev->hw->clear_hem(hr_dev, table, obj, step_idx))
728 			ibdev_warn(ibdev, "failed to clear hop%u HEM.\n", hop_num);
729 
730 		if (index->inited & HEM_INDEX_L1)
731 			if (hr_dev->hw->clear_hem(hr_dev, table, obj, 1))
732 				ibdev_warn(ibdev, "failed to clear HEM step 1.\n");
733 
734 		if (index->inited & HEM_INDEX_L0)
735 			if (hr_dev->hw->clear_hem(hr_dev, table, obj, 0))
736 				ibdev_warn(ibdev, "failed to clear HEM step 0.\n");
737 	}
738 }
739 
740 static void hns_roce_table_mhop_put(struct hns_roce_dev *hr_dev,
741 				    struct hns_roce_hem_table *table,
742 				    unsigned long obj,
743 				    int check_refcount)
744 {
745 	struct ib_device *ibdev = &hr_dev->ib_dev;
746 	struct hns_roce_hem_index index = {};
747 	struct hns_roce_hem_mhop mhop = {};
748 	int ret;
749 
750 	ret = calc_hem_config(hr_dev, table, obj, &mhop, &index);
751 	if (ret) {
752 		ibdev_err(ibdev, "calc hem config failed!\n");
753 		return;
754 	}
755 
756 	mutex_lock(&table->mutex);
757 	if (check_refcount && (--table->hem[index.buf]->refcount > 0)) {
758 		mutex_unlock(&table->mutex);
759 		return;
760 	}
761 
762 	clear_mhop_hem(hr_dev, table, obj, &mhop, &index);
763 	free_mhop_hem(hr_dev, table, &mhop, &index);
764 
765 	mutex_unlock(&table->mutex);
766 }
767 
768 void hns_roce_table_put(struct hns_roce_dev *hr_dev,
769 			struct hns_roce_hem_table *table, unsigned long obj)
770 {
771 	struct device *dev = hr_dev->dev;
772 	unsigned long i;
773 
774 	if (hns_roce_check_whether_mhop(hr_dev, table->type)) {
775 		hns_roce_table_mhop_put(hr_dev, table, obj, 1);
776 		return;
777 	}
778 
779 	i = (obj & (table->num_obj - 1)) /
780 	    (table->table_chunk_size / table->obj_size);
781 
782 	mutex_lock(&table->mutex);
783 
784 	if (--table->hem[i]->refcount == 0) {
785 		/* Clear HEM base address */
786 		if (hr_dev->hw->clear_hem(hr_dev, table, obj, 0))
787 			dev_warn(dev, "Clear HEM base address failed.\n");
788 
789 		hns_roce_free_hem(hr_dev, table->hem[i]);
790 		table->hem[i] = NULL;
791 	}
792 
793 	mutex_unlock(&table->mutex);
794 }
795 
796 void *hns_roce_table_find(struct hns_roce_dev *hr_dev,
797 			  struct hns_roce_hem_table *table,
798 			  unsigned long obj, dma_addr_t *dma_handle)
799 {
800 	struct hns_roce_hem_chunk *chunk;
801 	struct hns_roce_hem_mhop mhop;
802 	struct hns_roce_hem *hem;
803 	unsigned long mhop_obj = obj;
804 	unsigned long obj_per_chunk;
805 	unsigned long idx_offset;
806 	int offset, dma_offset;
807 	void *addr = NULL;
808 	u32 hem_idx = 0;
809 	int length;
810 	int i, j;
811 
812 	if (!table->lowmem)
813 		return NULL;
814 
815 	mutex_lock(&table->mutex);
816 
817 	if (!hns_roce_check_whether_mhop(hr_dev, table->type)) {
818 		obj_per_chunk = table->table_chunk_size / table->obj_size;
819 		hem = table->hem[(obj & (table->num_obj - 1)) / obj_per_chunk];
820 		idx_offset = (obj & (table->num_obj - 1)) % obj_per_chunk;
821 		dma_offset = offset = idx_offset * table->obj_size;
822 	} else {
823 		u32 seg_size = 64; /* 8 bytes per BA and 8 BA per segment */
824 
825 		if (hns_roce_calc_hem_mhop(hr_dev, table, &mhop_obj, &mhop))
826 			goto out;
827 		/* mtt mhop */
828 		i = mhop.l0_idx;
829 		j = mhop.l1_idx;
830 		if (mhop.hop_num == 2)
831 			hem_idx = i * (mhop.bt_chunk_size / BA_BYTE_LEN) + j;
832 		else if (mhop.hop_num == 1 ||
833 			 mhop.hop_num == HNS_ROCE_HOP_NUM_0)
834 			hem_idx = i;
835 
836 		hem = table->hem[hem_idx];
837 		dma_offset = offset = (obj & (table->num_obj - 1)) * seg_size %
838 				       mhop.bt_chunk_size;
839 		if (mhop.hop_num == 2)
840 			dma_offset = offset = 0;
841 	}
842 
843 	if (!hem)
844 		goto out;
845 
846 	list_for_each_entry(chunk, &hem->chunk_list, list) {
847 		for (i = 0; i < chunk->npages; ++i) {
848 			length = sg_dma_len(&chunk->mem[i]);
849 			if (dma_handle && dma_offset >= 0) {
850 				if (length > (u32)dma_offset)
851 					*dma_handle = sg_dma_address(
852 						&chunk->mem[i]) + dma_offset;
853 				dma_offset -= length;
854 			}
855 
856 			if (length > (u32)offset) {
857 				addr = chunk->buf[i] + offset;
858 				goto out;
859 			}
860 			offset -= length;
861 		}
862 	}
863 
864 out:
865 	mutex_unlock(&table->mutex);
866 	return addr;
867 }
868 
869 int hns_roce_init_hem_table(struct hns_roce_dev *hr_dev,
870 			    struct hns_roce_hem_table *table, u32 type,
871 			    unsigned long obj_size, unsigned long nobj,
872 			    int use_lowmem)
873 {
874 	unsigned long obj_per_chunk;
875 	unsigned long num_hem;
876 
877 	if (!hns_roce_check_whether_mhop(hr_dev, type)) {
878 		table->table_chunk_size = hr_dev->caps.chunk_sz;
879 		obj_per_chunk = table->table_chunk_size / obj_size;
880 		num_hem = (nobj + obj_per_chunk - 1) / obj_per_chunk;
881 
882 		table->hem = kcalloc(num_hem, sizeof(*table->hem), GFP_KERNEL);
883 		if (!table->hem)
884 			return -ENOMEM;
885 	} else {
886 		struct hns_roce_hem_mhop mhop = {};
887 		unsigned long buf_chunk_size;
888 		unsigned long bt_chunk_size;
889 		unsigned long bt_chunk_num;
890 		unsigned long num_bt_l0;
891 		u32 hop_num;
892 
893 		if (get_hem_table_config(hr_dev, &mhop, type))
894 			return -EINVAL;
895 
896 		buf_chunk_size = mhop.buf_chunk_size;
897 		bt_chunk_size = mhop.bt_chunk_size;
898 		num_bt_l0 = mhop.ba_l0_num;
899 		hop_num = mhop.hop_num;
900 
901 		obj_per_chunk = buf_chunk_size / obj_size;
902 		num_hem = (nobj + obj_per_chunk - 1) / obj_per_chunk;
903 		bt_chunk_num = bt_chunk_size / BA_BYTE_LEN;
904 		if (type >= HEM_TYPE_MTT)
905 			num_bt_l0 = bt_chunk_num;
906 
907 		table->hem = kcalloc(num_hem, sizeof(*table->hem),
908 					 GFP_KERNEL);
909 		if (!table->hem)
910 			goto err_kcalloc_hem_buf;
911 
912 		if (check_whether_bt_num_3(type, hop_num)) {
913 			unsigned long num_bt_l1;
914 
915 			num_bt_l1 = (num_hem + bt_chunk_num - 1) /
916 					     bt_chunk_num;
917 			table->bt_l1 = kcalloc(num_bt_l1,
918 					       sizeof(*table->bt_l1),
919 					       GFP_KERNEL);
920 			if (!table->bt_l1)
921 				goto err_kcalloc_bt_l1;
922 
923 			table->bt_l1_dma_addr = kcalloc(num_bt_l1,
924 						 sizeof(*table->bt_l1_dma_addr),
925 						 GFP_KERNEL);
926 
927 			if (!table->bt_l1_dma_addr)
928 				goto err_kcalloc_l1_dma;
929 		}
930 
931 		if (check_whether_bt_num_2(type, hop_num) ||
932 			check_whether_bt_num_3(type, hop_num)) {
933 			table->bt_l0 = kcalloc(num_bt_l0, sizeof(*table->bt_l0),
934 					       GFP_KERNEL);
935 			if (!table->bt_l0)
936 				goto err_kcalloc_bt_l0;
937 
938 			table->bt_l0_dma_addr = kcalloc(num_bt_l0,
939 						 sizeof(*table->bt_l0_dma_addr),
940 						 GFP_KERNEL);
941 			if (!table->bt_l0_dma_addr)
942 				goto err_kcalloc_l0_dma;
943 		}
944 	}
945 
946 	table->type = type;
947 	table->num_hem = num_hem;
948 	table->num_obj = nobj;
949 	table->obj_size = obj_size;
950 	table->lowmem = use_lowmem;
951 	mutex_init(&table->mutex);
952 
953 	return 0;
954 
955 err_kcalloc_l0_dma:
956 	kfree(table->bt_l0);
957 	table->bt_l0 = NULL;
958 
959 err_kcalloc_bt_l0:
960 	kfree(table->bt_l1_dma_addr);
961 	table->bt_l1_dma_addr = NULL;
962 
963 err_kcalloc_l1_dma:
964 	kfree(table->bt_l1);
965 	table->bt_l1 = NULL;
966 
967 err_kcalloc_bt_l1:
968 	kfree(table->hem);
969 	table->hem = NULL;
970 
971 err_kcalloc_hem_buf:
972 	return -ENOMEM;
973 }
974 
975 static void hns_roce_cleanup_mhop_hem_table(struct hns_roce_dev *hr_dev,
976 					    struct hns_roce_hem_table *table)
977 {
978 	struct hns_roce_hem_mhop mhop;
979 	u32 buf_chunk_size;
980 	u64 obj;
981 	int i;
982 
983 	if (hns_roce_calc_hem_mhop(hr_dev, table, NULL, &mhop))
984 		return;
985 	buf_chunk_size = table->type < HEM_TYPE_MTT ? mhop.buf_chunk_size :
986 					mhop.bt_chunk_size;
987 
988 	for (i = 0; i < table->num_hem; ++i) {
989 		obj = i * buf_chunk_size / table->obj_size;
990 		if (table->hem[i])
991 			hns_roce_table_mhop_put(hr_dev, table, obj, 0);
992 	}
993 
994 	kfree(table->hem);
995 	table->hem = NULL;
996 	kfree(table->bt_l1);
997 	table->bt_l1 = NULL;
998 	kfree(table->bt_l1_dma_addr);
999 	table->bt_l1_dma_addr = NULL;
1000 	kfree(table->bt_l0);
1001 	table->bt_l0 = NULL;
1002 	kfree(table->bt_l0_dma_addr);
1003 	table->bt_l0_dma_addr = NULL;
1004 }
1005 
1006 void hns_roce_cleanup_hem_table(struct hns_roce_dev *hr_dev,
1007 				struct hns_roce_hem_table *table)
1008 {
1009 	struct device *dev = hr_dev->dev;
1010 	unsigned long i;
1011 
1012 	if (hns_roce_check_whether_mhop(hr_dev, table->type)) {
1013 		hns_roce_cleanup_mhop_hem_table(hr_dev, table);
1014 		return;
1015 	}
1016 
1017 	for (i = 0; i < table->num_hem; ++i)
1018 		if (table->hem[i]) {
1019 			if (hr_dev->hw->clear_hem(hr_dev, table,
1020 			    i * table->table_chunk_size / table->obj_size, 0))
1021 				dev_err(dev, "Clear HEM base address failed.\n");
1022 
1023 			hns_roce_free_hem(hr_dev, table->hem[i]);
1024 		}
1025 
1026 	kfree(table->hem);
1027 }
1028 
1029 void hns_roce_cleanup_hem(struct hns_roce_dev *hr_dev)
1030 {
1031 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SRQ)
1032 		hns_roce_cleanup_hem_table(hr_dev,
1033 					   &hr_dev->srq_table.table);
1034 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->cq_table.table);
1035 	if (hr_dev->caps.qpc_timer_entry_sz)
1036 		hns_roce_cleanup_hem_table(hr_dev,
1037 					   &hr_dev->qpc_timer_table);
1038 	if (hr_dev->caps.cqc_timer_entry_sz)
1039 		hns_roce_cleanup_hem_table(hr_dev,
1040 					   &hr_dev->cqc_timer_table);
1041 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL)
1042 		hns_roce_cleanup_hem_table(hr_dev,
1043 					   &hr_dev->qp_table.sccc_table);
1044 	if (hr_dev->caps.trrl_entry_sz)
1045 		hns_roce_cleanup_hem_table(hr_dev,
1046 					   &hr_dev->qp_table.trrl_table);
1047 
1048 	if (hr_dev->caps.gmv_entry_sz)
1049 		hns_roce_cleanup_hem_table(hr_dev, &hr_dev->gmv_table);
1050 
1051 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->qp_table.irrl_table);
1052 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->qp_table.qp_table);
1053 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->mr_table.mtpt_table);
1054 }
1055 
1056 struct roce_hem_item {
1057 	struct list_head list; /* link all hems in the same bt level */
1058 	struct list_head sibling; /* link all hems in last hop for mtt */
1059 	void *addr;
1060 	dma_addr_t dma_addr;
1061 	size_t count; /* max ba numbers */
1062 	int start; /* start buf offset in this hem */
1063 	int end; /* end buf offset in this hem */
1064 };
1065 
1066 static struct roce_hem_item *hem_list_alloc_item(struct hns_roce_dev *hr_dev,
1067 						   int start, int end,
1068 						   int count, bool exist_bt,
1069 						   int bt_level)
1070 {
1071 	struct roce_hem_item *hem;
1072 
1073 	hem = kzalloc(sizeof(*hem), GFP_KERNEL);
1074 	if (!hem)
1075 		return NULL;
1076 
1077 	if (exist_bt) {
1078 		hem->addr = dma_alloc_coherent(hr_dev->dev, count * BA_BYTE_LEN,
1079 					       &hem->dma_addr, GFP_KERNEL);
1080 		if (!hem->addr) {
1081 			kfree(hem);
1082 			return NULL;
1083 		}
1084 	}
1085 
1086 	hem->count = count;
1087 	hem->start = start;
1088 	hem->end = end;
1089 	INIT_LIST_HEAD(&hem->list);
1090 	INIT_LIST_HEAD(&hem->sibling);
1091 
1092 	return hem;
1093 }
1094 
1095 static void hem_list_free_item(struct hns_roce_dev *hr_dev,
1096 			       struct roce_hem_item *hem, bool exist_bt)
1097 {
1098 	if (exist_bt)
1099 		dma_free_coherent(hr_dev->dev, hem->count * BA_BYTE_LEN,
1100 				  hem->addr, hem->dma_addr);
1101 	kfree(hem);
1102 }
1103 
1104 static void hem_list_free_all(struct hns_roce_dev *hr_dev,
1105 			      struct list_head *head, bool exist_bt)
1106 {
1107 	struct roce_hem_item *hem, *temp_hem;
1108 
1109 	list_for_each_entry_safe(hem, temp_hem, head, list) {
1110 		list_del(&hem->list);
1111 		hem_list_free_item(hr_dev, hem, exist_bt);
1112 	}
1113 }
1114 
1115 static void hem_list_link_bt(struct hns_roce_dev *hr_dev, void *base_addr,
1116 			     u64 table_addr)
1117 {
1118 	*(u64 *)(base_addr) = table_addr;
1119 }
1120 
1121 /* assign L0 table address to hem from root bt */
1122 static void hem_list_assign_bt(struct hns_roce_dev *hr_dev,
1123 			       struct roce_hem_item *hem, void *cpu_addr,
1124 			       u64 phy_addr)
1125 {
1126 	hem->addr = cpu_addr;
1127 	hem->dma_addr = (dma_addr_t)phy_addr;
1128 }
1129 
1130 static inline bool hem_list_page_is_in_range(struct roce_hem_item *hem,
1131 					     int offset)
1132 {
1133 	return (hem->start <= offset && offset <= hem->end);
1134 }
1135 
1136 static struct roce_hem_item *hem_list_search_item(struct list_head *ba_list,
1137 						    int page_offset)
1138 {
1139 	struct roce_hem_item *hem, *temp_hem;
1140 	struct roce_hem_item *found = NULL;
1141 
1142 	list_for_each_entry_safe(hem, temp_hem, ba_list, list) {
1143 		if (hem_list_page_is_in_range(hem, page_offset)) {
1144 			found = hem;
1145 			break;
1146 		}
1147 	}
1148 
1149 	return found;
1150 }
1151 
1152 static bool hem_list_is_bottom_bt(int hopnum, int bt_level)
1153 {
1154 	/*
1155 	 * hopnum    base address table levels
1156 	 * 0		L0(buf)
1157 	 * 1		L0 -> buf
1158 	 * 2		L0 -> L1 -> buf
1159 	 * 3		L0 -> L1 -> L2 -> buf
1160 	 */
1161 	return bt_level >= (hopnum ? hopnum - 1 : hopnum);
1162 }
1163 
1164 /**
1165  * calc base address entries num
1166  * @hopnum: num of mutihop addressing
1167  * @bt_level: base address table level
1168  * @unit: ba entries per bt page
1169  */
1170 static u32 hem_list_calc_ba_range(int hopnum, int bt_level, int unit)
1171 {
1172 	u32 step;
1173 	int max;
1174 	int i;
1175 
1176 	if (hopnum <= bt_level)
1177 		return 0;
1178 	/*
1179 	 * hopnum  bt_level   range
1180 	 * 1	      0       unit
1181 	 * ------------
1182 	 * 2	      0       unit * unit
1183 	 * 2	      1       unit
1184 	 * ------------
1185 	 * 3	      0       unit * unit * unit
1186 	 * 3	      1       unit * unit
1187 	 * 3	      2       unit
1188 	 */
1189 	step = 1;
1190 	max = hopnum - bt_level;
1191 	for (i = 0; i < max; i++)
1192 		step = step * unit;
1193 
1194 	return step;
1195 }
1196 
1197 /**
1198  * calc the root ba entries which could cover all regions
1199  * @regions: buf region array
1200  * @region_cnt: array size of @regions
1201  * @unit: ba entries per bt page
1202  */
1203 int hns_roce_hem_list_calc_root_ba(const struct hns_roce_buf_region *regions,
1204 				   int region_cnt, int unit)
1205 {
1206 	struct hns_roce_buf_region *r;
1207 	int total = 0;
1208 	int step;
1209 	int i;
1210 
1211 	for (i = 0; i < region_cnt; i++) {
1212 		r = (struct hns_roce_buf_region *)&regions[i];
1213 		if (r->hopnum > 1) {
1214 			step = hem_list_calc_ba_range(r->hopnum, 1, unit);
1215 			if (step > 0)
1216 				total += (r->count + step - 1) / step;
1217 		} else {
1218 			total += r->count;
1219 		}
1220 	}
1221 
1222 	return total;
1223 }
1224 
1225 static int hem_list_alloc_mid_bt(struct hns_roce_dev *hr_dev,
1226 				 const struct hns_roce_buf_region *r, int unit,
1227 				 int offset, struct list_head *mid_bt,
1228 				 struct list_head *btm_bt)
1229 {
1230 	struct roce_hem_item *hem_ptrs[HNS_ROCE_MAX_BT_LEVEL] = { NULL };
1231 	struct list_head temp_list[HNS_ROCE_MAX_BT_LEVEL];
1232 	struct roce_hem_item *cur, *pre;
1233 	const int hopnum = r->hopnum;
1234 	int start_aligned;
1235 	int distance;
1236 	int ret = 0;
1237 	int max_ofs;
1238 	int level;
1239 	u32 step;
1240 	int end;
1241 
1242 	if (hopnum <= 1)
1243 		return 0;
1244 
1245 	if (hopnum > HNS_ROCE_MAX_BT_LEVEL) {
1246 		dev_err(hr_dev->dev, "invalid hopnum %d!\n", hopnum);
1247 		return -EINVAL;
1248 	}
1249 
1250 	if (offset < r->offset) {
1251 		dev_err(hr_dev->dev, "invalid offset %d, min %u!\n",
1252 			offset, r->offset);
1253 		return -EINVAL;
1254 	}
1255 
1256 	distance = offset - r->offset;
1257 	max_ofs = r->offset + r->count - 1;
1258 	for (level = 0; level < hopnum; level++)
1259 		INIT_LIST_HEAD(&temp_list[level]);
1260 
1261 	/* config L1 bt to last bt and link them to corresponding parent */
1262 	for (level = 1; level < hopnum; level++) {
1263 		cur = hem_list_search_item(&mid_bt[level], offset);
1264 		if (cur) {
1265 			hem_ptrs[level] = cur;
1266 			continue;
1267 		}
1268 
1269 		step = hem_list_calc_ba_range(hopnum, level, unit);
1270 		if (step < 1) {
1271 			ret = -EINVAL;
1272 			goto err_exit;
1273 		}
1274 
1275 		start_aligned = (distance / step) * step + r->offset;
1276 		end = min_t(int, start_aligned + step - 1, max_ofs);
1277 		cur = hem_list_alloc_item(hr_dev, start_aligned, end, unit,
1278 					  true, level);
1279 		if (!cur) {
1280 			ret = -ENOMEM;
1281 			goto err_exit;
1282 		}
1283 		hem_ptrs[level] = cur;
1284 		list_add(&cur->list, &temp_list[level]);
1285 		if (hem_list_is_bottom_bt(hopnum, level))
1286 			list_add(&cur->sibling, &temp_list[0]);
1287 
1288 		/* link bt to parent bt */
1289 		if (level > 1) {
1290 			pre = hem_ptrs[level - 1];
1291 			step = (cur->start - pre->start) / step * BA_BYTE_LEN;
1292 			hem_list_link_bt(hr_dev, pre->addr + step,
1293 					 cur->dma_addr);
1294 		}
1295 	}
1296 
1297 	list_splice(&temp_list[0], btm_bt);
1298 	for (level = 1; level < hopnum; level++)
1299 		list_splice(&temp_list[level], &mid_bt[level]);
1300 
1301 	return 0;
1302 
1303 err_exit:
1304 	for (level = 1; level < hopnum; level++)
1305 		hem_list_free_all(hr_dev, &temp_list[level], true);
1306 
1307 	return ret;
1308 }
1309 
1310 static int hem_list_alloc_root_bt(struct hns_roce_dev *hr_dev,
1311 				  struct hns_roce_hem_list *hem_list, int unit,
1312 				  const struct hns_roce_buf_region *regions,
1313 				  int region_cnt)
1314 {
1315 	struct list_head temp_list[HNS_ROCE_MAX_BT_REGION];
1316 	struct roce_hem_item *hem, *temp_hem, *root_hem;
1317 	const struct hns_roce_buf_region *r;
1318 	struct list_head temp_root;
1319 	struct list_head temp_btm;
1320 	void *cpu_base;
1321 	u64 phy_base;
1322 	int ret = 0;
1323 	int ba_num;
1324 	int offset;
1325 	int total;
1326 	int step;
1327 	int i;
1328 
1329 	r = &regions[0];
1330 	root_hem = hem_list_search_item(&hem_list->root_bt, r->offset);
1331 	if (root_hem)
1332 		return 0;
1333 
1334 	ba_num = hns_roce_hem_list_calc_root_ba(regions, region_cnt, unit);
1335 	if (ba_num < 1)
1336 		return -ENOMEM;
1337 
1338 	if (ba_num > unit)
1339 		return -ENOBUFS;
1340 
1341 	ba_num = min_t(int, ba_num, unit);
1342 	INIT_LIST_HEAD(&temp_root);
1343 	offset = r->offset;
1344 	/* indicate to last region */
1345 	r = &regions[region_cnt - 1];
1346 	root_hem = hem_list_alloc_item(hr_dev, offset, r->offset + r->count - 1,
1347 				       ba_num, true, 0);
1348 	if (!root_hem)
1349 		return -ENOMEM;
1350 	list_add(&root_hem->list, &temp_root);
1351 
1352 	hem_list->root_ba = root_hem->dma_addr;
1353 
1354 	INIT_LIST_HEAD(&temp_btm);
1355 	for (i = 0; i < region_cnt; i++)
1356 		INIT_LIST_HEAD(&temp_list[i]);
1357 
1358 	total = 0;
1359 	for (i = 0; i < region_cnt && total < ba_num; i++) {
1360 		r = &regions[i];
1361 		if (!r->count)
1362 			continue;
1363 
1364 		/* all regions's mid[x][0] shared the root_bt's trunk */
1365 		cpu_base = root_hem->addr + total * BA_BYTE_LEN;
1366 		phy_base = root_hem->dma_addr + total * BA_BYTE_LEN;
1367 
1368 		/* if hopnum is 0 or 1, cut a new fake hem from the root bt
1369 		 * which's address share to all regions.
1370 		 */
1371 		if (hem_list_is_bottom_bt(r->hopnum, 0)) {
1372 			hem = hem_list_alloc_item(hr_dev, r->offset,
1373 						  r->offset + r->count - 1,
1374 						  r->count, false, 0);
1375 			if (!hem) {
1376 				ret = -ENOMEM;
1377 				goto err_exit;
1378 			}
1379 			hem_list_assign_bt(hr_dev, hem, cpu_base, phy_base);
1380 			list_add(&hem->list, &temp_list[i]);
1381 			list_add(&hem->sibling, &temp_btm);
1382 			total += r->count;
1383 		} else {
1384 			step = hem_list_calc_ba_range(r->hopnum, 1, unit);
1385 			if (step < 1) {
1386 				ret = -EINVAL;
1387 				goto err_exit;
1388 			}
1389 			/* if exist mid bt, link L1 to L0 */
1390 			list_for_each_entry_safe(hem, temp_hem,
1391 					  &hem_list->mid_bt[i][1], list) {
1392 				offset = (hem->start - r->offset) / step *
1393 					  BA_BYTE_LEN;
1394 				hem_list_link_bt(hr_dev, cpu_base + offset,
1395 						 hem->dma_addr);
1396 				total++;
1397 			}
1398 		}
1399 	}
1400 
1401 	list_splice(&temp_btm, &hem_list->btm_bt);
1402 	list_splice(&temp_root, &hem_list->root_bt);
1403 	for (i = 0; i < region_cnt; i++)
1404 		list_splice(&temp_list[i], &hem_list->mid_bt[i][0]);
1405 
1406 	return 0;
1407 
1408 err_exit:
1409 	for (i = 0; i < region_cnt; i++)
1410 		hem_list_free_all(hr_dev, &temp_list[i], false);
1411 
1412 	hem_list_free_all(hr_dev, &temp_root, true);
1413 
1414 	return ret;
1415 }
1416 
1417 /* construct the base address table and link them by address hop config */
1418 int hns_roce_hem_list_request(struct hns_roce_dev *hr_dev,
1419 			      struct hns_roce_hem_list *hem_list,
1420 			      const struct hns_roce_buf_region *regions,
1421 			      int region_cnt, unsigned int bt_pg_shift)
1422 {
1423 	const struct hns_roce_buf_region *r;
1424 	int ofs, end;
1425 	int unit;
1426 	int ret;
1427 	int i;
1428 
1429 	if (region_cnt > HNS_ROCE_MAX_BT_REGION) {
1430 		dev_err(hr_dev->dev, "invalid region region_cnt %d!\n",
1431 			region_cnt);
1432 		return -EINVAL;
1433 	}
1434 
1435 	unit = (1 << bt_pg_shift) / BA_BYTE_LEN;
1436 	for (i = 0; i < region_cnt; i++) {
1437 		r = &regions[i];
1438 		if (!r->count)
1439 			continue;
1440 
1441 		end = r->offset + r->count;
1442 		for (ofs = r->offset; ofs < end; ofs += unit) {
1443 			ret = hem_list_alloc_mid_bt(hr_dev, r, unit, ofs,
1444 						    hem_list->mid_bt[i],
1445 						    &hem_list->btm_bt);
1446 			if (ret) {
1447 				dev_err(hr_dev->dev,
1448 					"alloc hem trunk fail ret=%d!\n", ret);
1449 				goto err_alloc;
1450 			}
1451 		}
1452 	}
1453 
1454 	ret = hem_list_alloc_root_bt(hr_dev, hem_list, unit, regions,
1455 				     region_cnt);
1456 	if (ret)
1457 		dev_err(hr_dev->dev, "alloc hem root fail ret=%d!\n", ret);
1458 	else
1459 		return 0;
1460 
1461 err_alloc:
1462 	hns_roce_hem_list_release(hr_dev, hem_list);
1463 
1464 	return ret;
1465 }
1466 
1467 void hns_roce_hem_list_release(struct hns_roce_dev *hr_dev,
1468 			       struct hns_roce_hem_list *hem_list)
1469 {
1470 	int i, j;
1471 
1472 	for (i = 0; i < HNS_ROCE_MAX_BT_REGION; i++)
1473 		for (j = 0; j < HNS_ROCE_MAX_BT_LEVEL; j++)
1474 			hem_list_free_all(hr_dev, &hem_list->mid_bt[i][j],
1475 					  j != 0);
1476 
1477 	hem_list_free_all(hr_dev, &hem_list->root_bt, true);
1478 	INIT_LIST_HEAD(&hem_list->btm_bt);
1479 	hem_list->root_ba = 0;
1480 }
1481 
1482 void hns_roce_hem_list_init(struct hns_roce_hem_list *hem_list)
1483 {
1484 	int i, j;
1485 
1486 	INIT_LIST_HEAD(&hem_list->root_bt);
1487 	INIT_LIST_HEAD(&hem_list->btm_bt);
1488 	for (i = 0; i < HNS_ROCE_MAX_BT_REGION; i++)
1489 		for (j = 0; j < HNS_ROCE_MAX_BT_LEVEL; j++)
1490 			INIT_LIST_HEAD(&hem_list->mid_bt[i][j]);
1491 }
1492 
1493 void *hns_roce_hem_list_find_mtt(struct hns_roce_dev *hr_dev,
1494 				 struct hns_roce_hem_list *hem_list,
1495 				 int offset, int *mtt_cnt, u64 *phy_addr)
1496 {
1497 	struct list_head *head = &hem_list->btm_bt;
1498 	struct roce_hem_item *hem, *temp_hem;
1499 	void *cpu_base = NULL;
1500 	u64 phy_base = 0;
1501 	int nr = 0;
1502 
1503 	list_for_each_entry_safe(hem, temp_hem, head, sibling) {
1504 		if (hem_list_page_is_in_range(hem, offset)) {
1505 			nr = offset - hem->start;
1506 			cpu_base = hem->addr + nr * BA_BYTE_LEN;
1507 			phy_base = hem->dma_addr + nr * BA_BYTE_LEN;
1508 			nr = hem->end + 1 - offset;
1509 			break;
1510 		}
1511 	}
1512 
1513 	if (mtt_cnt)
1514 		*mtt_cnt = nr;
1515 
1516 	if (phy_addr)
1517 		*phy_addr = phy_base;
1518 
1519 	return cpu_base;
1520 }
1521