xref: /openbmc/linux/drivers/infiniband/hw/hns/hns_roce_mr.c (revision 4ed91d48259d9ddd378424d008f2e6559f7e78f8)
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 <rdma/ib_umem.h>
36 #include "hns_roce_device.h"
37 #include "hns_roce_cmd.h"
38 #include "hns_roce_hem.h"
39 
40 static u32 hw_index_to_key(unsigned long ind)
41 {
42 	return (u32)(ind >> 24) | (ind << 8);
43 }
44 
45 unsigned long key_to_hw_index(u32 key)
46 {
47 	return (key << 24) | (key >> 8);
48 }
49 
50 static int hns_roce_sw2hw_mpt(struct hns_roce_dev *hr_dev,
51 			      struct hns_roce_cmd_mailbox *mailbox,
52 			      unsigned long mpt_index)
53 {
54 	return hns_roce_cmd_mbox(hr_dev, mailbox->dma, 0, mpt_index, 0,
55 				 HNS_ROCE_CMD_SW2HW_MPT,
56 				 HNS_ROCE_CMD_TIMEOUT_MSECS);
57 }
58 
59 int hns_roce_hw2sw_mpt(struct hns_roce_dev *hr_dev,
60 			      struct hns_roce_cmd_mailbox *mailbox,
61 			      unsigned long mpt_index)
62 {
63 	return hns_roce_cmd_mbox(hr_dev, 0, mailbox ? mailbox->dma : 0,
64 				 mpt_index, !mailbox, HNS_ROCE_CMD_HW2SW_MPT,
65 				 HNS_ROCE_CMD_TIMEOUT_MSECS);
66 }
67 
68 static int hns_roce_buddy_alloc(struct hns_roce_buddy *buddy, int order,
69 				unsigned long *seg)
70 {
71 	int o;
72 	u32 m;
73 
74 	spin_lock(&buddy->lock);
75 
76 	for (o = order; o <= buddy->max_order; ++o) {
77 		if (buddy->num_free[o]) {
78 			m = 1 << (buddy->max_order - o);
79 			*seg = find_first_bit(buddy->bits[o], m);
80 			if (*seg < m)
81 				goto found;
82 		}
83 	}
84 	spin_unlock(&buddy->lock);
85 	return -1;
86 
87  found:
88 	clear_bit(*seg, buddy->bits[o]);
89 	--buddy->num_free[o];
90 
91 	while (o > order) {
92 		--o;
93 		*seg <<= 1;
94 		set_bit(*seg ^ 1, buddy->bits[o]);
95 		++buddy->num_free[o];
96 	}
97 
98 	spin_unlock(&buddy->lock);
99 
100 	*seg <<= order;
101 	return 0;
102 }
103 
104 static void hns_roce_buddy_free(struct hns_roce_buddy *buddy, unsigned long seg,
105 				int order)
106 {
107 	seg >>= order;
108 
109 	spin_lock(&buddy->lock);
110 
111 	while (test_bit(seg ^ 1, buddy->bits[order])) {
112 		clear_bit(seg ^ 1, buddy->bits[order]);
113 		--buddy->num_free[order];
114 		seg >>= 1;
115 		++order;
116 	}
117 
118 	set_bit(seg, buddy->bits[order]);
119 	++buddy->num_free[order];
120 
121 	spin_unlock(&buddy->lock);
122 }
123 
124 static int hns_roce_buddy_init(struct hns_roce_buddy *buddy, int max_order)
125 {
126 	int i, s;
127 
128 	buddy->max_order = max_order;
129 	spin_lock_init(&buddy->lock);
130 
131 	buddy->bits = kzalloc((buddy->max_order + 1) * sizeof(long *),
132 			       GFP_KERNEL);
133 	buddy->num_free = kzalloc((buddy->max_order + 1) * sizeof(int *),
134 				   GFP_KERNEL);
135 	if (!buddy->bits || !buddy->num_free)
136 		goto err_out;
137 
138 	for (i = 0; i <= buddy->max_order; ++i) {
139 		s = BITS_TO_LONGS(1 << (buddy->max_order - i));
140 		buddy->bits[i] = kcalloc(s, sizeof(long), GFP_KERNEL |
141 					 __GFP_NOWARN);
142 		if (!buddy->bits[i]) {
143 			buddy->bits[i] = vzalloc(s * sizeof(long));
144 			if (!buddy->bits[i])
145 				goto err_out_free;
146 		}
147 	}
148 
149 	set_bit(0, buddy->bits[buddy->max_order]);
150 	buddy->num_free[buddy->max_order] = 1;
151 
152 	return 0;
153 
154 err_out_free:
155 	for (i = 0; i <= buddy->max_order; ++i)
156 		kvfree(buddy->bits[i]);
157 
158 err_out:
159 	kfree(buddy->bits);
160 	kfree(buddy->num_free);
161 	return -ENOMEM;
162 }
163 
164 static void hns_roce_buddy_cleanup(struct hns_roce_buddy *buddy)
165 {
166 	int i;
167 
168 	for (i = 0; i <= buddy->max_order; ++i)
169 		kvfree(buddy->bits[i]);
170 
171 	kfree(buddy->bits);
172 	kfree(buddy->num_free);
173 }
174 
175 static int hns_roce_alloc_mtt_range(struct hns_roce_dev *hr_dev, int order,
176 				    unsigned long *seg)
177 {
178 	struct hns_roce_mr_table *mr_table = &hr_dev->mr_table;
179 	int ret = 0;
180 
181 	ret = hns_roce_buddy_alloc(&mr_table->mtt_buddy, order, seg);
182 	if (ret == -1)
183 		return -1;
184 
185 	if (hns_roce_table_get_range(hr_dev, &mr_table->mtt_table, *seg,
186 				     *seg + (1 << order) - 1)) {
187 		hns_roce_buddy_free(&mr_table->mtt_buddy, *seg, order);
188 		return -1;
189 	}
190 
191 	return 0;
192 }
193 
194 int hns_roce_mtt_init(struct hns_roce_dev *hr_dev, int npages, int page_shift,
195 		      struct hns_roce_mtt *mtt)
196 {
197 	int ret = 0;
198 	int i;
199 
200 	/* Page num is zero, correspond to DMA memory register */
201 	if (!npages) {
202 		mtt->order = -1;
203 		mtt->page_shift = HNS_ROCE_HEM_PAGE_SHIFT;
204 		return 0;
205 	}
206 
207 	/* Note: if page_shift is zero, FAST memory regsiter */
208 	mtt->page_shift = page_shift;
209 
210 	/* Compute MTT entry necessary */
211 	for (mtt->order = 0, i = HNS_ROCE_MTT_ENTRY_PER_SEG; i < npages;
212 	     i <<= 1)
213 		++mtt->order;
214 
215 	/* Allocate MTT entry */
216 	ret = hns_roce_alloc_mtt_range(hr_dev, mtt->order, &mtt->first_seg);
217 	if (ret == -1)
218 		return -ENOMEM;
219 
220 	return 0;
221 }
222 
223 void hns_roce_mtt_cleanup(struct hns_roce_dev *hr_dev, struct hns_roce_mtt *mtt)
224 {
225 	struct hns_roce_mr_table *mr_table = &hr_dev->mr_table;
226 
227 	if (mtt->order < 0)
228 		return;
229 
230 	hns_roce_buddy_free(&mr_table->mtt_buddy, mtt->first_seg, mtt->order);
231 	hns_roce_table_put_range(hr_dev, &mr_table->mtt_table, mtt->first_seg,
232 				 mtt->first_seg + (1 << mtt->order) - 1);
233 }
234 
235 static int hns_roce_mr_alloc(struct hns_roce_dev *hr_dev, u32 pd, u64 iova,
236 			     u64 size, u32 access, int npages,
237 			     struct hns_roce_mr *mr)
238 {
239 	unsigned long index = 0;
240 	int ret = 0;
241 	struct device *dev = &hr_dev->pdev->dev;
242 
243 	/* Allocate a key for mr from mr_table */
244 	ret = hns_roce_bitmap_alloc(&hr_dev->mr_table.mtpt_bitmap, &index);
245 	if (ret == -1)
246 		return -ENOMEM;
247 
248 	mr->iova = iova;			/* MR va starting addr */
249 	mr->size = size;			/* MR addr range */
250 	mr->pd = pd;				/* MR num */
251 	mr->access = access;			/* MR access permit */
252 	mr->enabled = 0;			/* MR active status */
253 	mr->key = hw_index_to_key(index);	/* MR key */
254 
255 	if (size == ~0ull) {
256 		mr->type = MR_TYPE_DMA;
257 		mr->pbl_buf = NULL;
258 		mr->pbl_dma_addr = 0;
259 	} else {
260 		mr->type = MR_TYPE_MR;
261 		mr->pbl_buf = dma_alloc_coherent(dev, npages * 8,
262 						 &(mr->pbl_dma_addr),
263 						 GFP_KERNEL);
264 		if (!mr->pbl_buf)
265 			return -ENOMEM;
266 	}
267 
268 	return 0;
269 }
270 
271 static void hns_roce_mr_free(struct hns_roce_dev *hr_dev,
272 			     struct hns_roce_mr *mr)
273 {
274 	struct device *dev = &hr_dev->pdev->dev;
275 	int npages = 0;
276 	int ret;
277 
278 	if (mr->enabled) {
279 		ret = hns_roce_hw2sw_mpt(hr_dev, NULL, key_to_hw_index(mr->key)
280 					 & (hr_dev->caps.num_mtpts - 1));
281 		if (ret)
282 			dev_warn(dev, "HW2SW_MPT failed (%d)\n", ret);
283 	}
284 
285 	if (mr->size != ~0ULL) {
286 		npages = ib_umem_page_count(mr->umem);
287 		dma_free_coherent(dev, (unsigned int)(npages * 8), mr->pbl_buf,
288 				  mr->pbl_dma_addr);
289 	}
290 
291 	hns_roce_bitmap_free(&hr_dev->mr_table.mtpt_bitmap,
292 			     key_to_hw_index(mr->key), BITMAP_NO_RR);
293 }
294 
295 static int hns_roce_mr_enable(struct hns_roce_dev *hr_dev,
296 			      struct hns_roce_mr *mr)
297 {
298 	int ret;
299 	unsigned long mtpt_idx = key_to_hw_index(mr->key);
300 	struct device *dev = &hr_dev->pdev->dev;
301 	struct hns_roce_cmd_mailbox *mailbox;
302 	struct hns_roce_mr_table *mr_table = &hr_dev->mr_table;
303 
304 	/* Prepare HEM entry memory */
305 	ret = hns_roce_table_get(hr_dev, &mr_table->mtpt_table, mtpt_idx);
306 	if (ret)
307 		return ret;
308 
309 	/* Allocate mailbox memory */
310 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
311 	if (IS_ERR(mailbox)) {
312 		ret = PTR_ERR(mailbox);
313 		goto err_table;
314 	}
315 
316 	ret = hr_dev->hw->write_mtpt(mailbox->buf, mr, mtpt_idx);
317 	if (ret) {
318 		dev_err(dev, "Write mtpt fail!\n");
319 		goto err_page;
320 	}
321 
322 	ret = hns_roce_sw2hw_mpt(hr_dev, mailbox,
323 				 mtpt_idx & (hr_dev->caps.num_mtpts - 1));
324 	if (ret) {
325 		dev_err(dev, "SW2HW_MPT failed (%d)\n", ret);
326 		goto err_page;
327 	}
328 
329 	mr->enabled = 1;
330 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
331 
332 	return 0;
333 
334 err_page:
335 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
336 
337 err_table:
338 	hns_roce_table_put(hr_dev, &mr_table->mtpt_table, mtpt_idx);
339 	return ret;
340 }
341 
342 static int hns_roce_write_mtt_chunk(struct hns_roce_dev *hr_dev,
343 				    struct hns_roce_mtt *mtt, u32 start_index,
344 				    u32 npages, u64 *page_list)
345 {
346 	u32 i = 0;
347 	__le64 *mtts = NULL;
348 	dma_addr_t dma_handle;
349 	u32 s = start_index * sizeof(u64);
350 
351 	/* All MTTs must fit in the same page */
352 	if (start_index / (PAGE_SIZE / sizeof(u64)) !=
353 		(start_index + npages - 1) / (PAGE_SIZE / sizeof(u64)))
354 		return -EINVAL;
355 
356 	if (start_index & (HNS_ROCE_MTT_ENTRY_PER_SEG - 1))
357 		return -EINVAL;
358 
359 	mtts = hns_roce_table_find(&hr_dev->mr_table.mtt_table,
360 				mtt->first_seg + s / hr_dev->caps.mtt_entry_sz,
361 				&dma_handle);
362 	if (!mtts)
363 		return -ENOMEM;
364 
365 	/* Save page addr, low 12 bits : 0 */
366 	for (i = 0; i < npages; ++i)
367 		mtts[i] = (cpu_to_le64(page_list[i])) >> PAGE_ADDR_SHIFT;
368 
369 	return 0;
370 }
371 
372 static int hns_roce_write_mtt(struct hns_roce_dev *hr_dev,
373 			      struct hns_roce_mtt *mtt, u32 start_index,
374 			      u32 npages, u64 *page_list)
375 {
376 	int chunk;
377 	int ret;
378 
379 	if (mtt->order < 0)
380 		return -EINVAL;
381 
382 	while (npages > 0) {
383 		chunk = min_t(int, PAGE_SIZE / sizeof(u64), npages);
384 
385 		ret = hns_roce_write_mtt_chunk(hr_dev, mtt, start_index, chunk,
386 					       page_list);
387 		if (ret)
388 			return ret;
389 
390 		npages -= chunk;
391 		start_index += chunk;
392 		page_list += chunk;
393 	}
394 
395 	return 0;
396 }
397 
398 int hns_roce_buf_write_mtt(struct hns_roce_dev *hr_dev,
399 			   struct hns_roce_mtt *mtt, struct hns_roce_buf *buf)
400 {
401 	u32 i = 0;
402 	int ret = 0;
403 	u64 *page_list = NULL;
404 
405 	page_list = kmalloc_array(buf->npages, sizeof(*page_list), GFP_KERNEL);
406 	if (!page_list)
407 		return -ENOMEM;
408 
409 	for (i = 0; i < buf->npages; ++i) {
410 		if (buf->nbufs == 1)
411 			page_list[i] = buf->direct.map + (i << buf->page_shift);
412 		else
413 			page_list[i] = buf->page_list[i].map;
414 
415 	}
416 	ret = hns_roce_write_mtt(hr_dev, mtt, 0, buf->npages, page_list);
417 
418 	kfree(page_list);
419 
420 	return ret;
421 }
422 
423 int hns_roce_init_mr_table(struct hns_roce_dev *hr_dev)
424 {
425 	struct hns_roce_mr_table *mr_table = &hr_dev->mr_table;
426 	int ret = 0;
427 
428 	ret = hns_roce_bitmap_init(&mr_table->mtpt_bitmap,
429 				   hr_dev->caps.num_mtpts,
430 				   hr_dev->caps.num_mtpts - 1,
431 				   hr_dev->caps.reserved_mrws, 0);
432 	if (ret)
433 		return ret;
434 
435 	ret = hns_roce_buddy_init(&mr_table->mtt_buddy,
436 				  ilog2(hr_dev->caps.num_mtt_segs));
437 	if (ret)
438 		goto err_buddy;
439 
440 	return 0;
441 
442 err_buddy:
443 	hns_roce_bitmap_cleanup(&mr_table->mtpt_bitmap);
444 	return ret;
445 }
446 
447 void hns_roce_cleanup_mr_table(struct hns_roce_dev *hr_dev)
448 {
449 	struct hns_roce_mr_table *mr_table = &hr_dev->mr_table;
450 
451 	hns_roce_buddy_cleanup(&mr_table->mtt_buddy);
452 	hns_roce_bitmap_cleanup(&mr_table->mtpt_bitmap);
453 }
454 
455 struct ib_mr *hns_roce_get_dma_mr(struct ib_pd *pd, int acc)
456 {
457 	int ret = 0;
458 	struct hns_roce_mr *mr = NULL;
459 
460 	mr = kmalloc(sizeof(*mr), GFP_KERNEL);
461 	if (mr == NULL)
462 		return  ERR_PTR(-ENOMEM);
463 
464 	/* Allocate memory region key */
465 	ret = hns_roce_mr_alloc(to_hr_dev(pd->device), to_hr_pd(pd)->pdn, 0,
466 				~0ULL, acc, 0, mr);
467 	if (ret)
468 		goto err_free;
469 
470 	ret = hns_roce_mr_enable(to_hr_dev(pd->device), mr);
471 	if (ret)
472 		goto err_mr;
473 
474 	mr->ibmr.rkey = mr->ibmr.lkey = mr->key;
475 	mr->umem = NULL;
476 
477 	return &mr->ibmr;
478 
479 err_mr:
480 	hns_roce_mr_free(to_hr_dev(pd->device), mr);
481 
482 err_free:
483 	kfree(mr);
484 	return ERR_PTR(ret);
485 }
486 
487 int hns_roce_ib_umem_write_mtt(struct hns_roce_dev *hr_dev,
488 			       struct hns_roce_mtt *mtt, struct ib_umem *umem)
489 {
490 	struct scatterlist *sg;
491 	int i, k, entry;
492 	int ret = 0;
493 	u64 *pages;
494 	u32 n;
495 	int len;
496 
497 	pages = (u64 *) __get_free_page(GFP_KERNEL);
498 	if (!pages)
499 		return -ENOMEM;
500 
501 	i = n = 0;
502 
503 	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
504 		len = sg_dma_len(sg) >> mtt->page_shift;
505 		for (k = 0; k < len; ++k) {
506 			pages[i++] = sg_dma_address(sg) + umem->page_size * k;
507 			if (i == PAGE_SIZE / sizeof(u64)) {
508 				ret = hns_roce_write_mtt(hr_dev, mtt, n, i,
509 							 pages);
510 				if (ret)
511 					goto out;
512 				n += i;
513 				i = 0;
514 			}
515 		}
516 	}
517 
518 	if (i)
519 		ret = hns_roce_write_mtt(hr_dev, mtt, n, i, pages);
520 
521 out:
522 	free_page((unsigned long) pages);
523 	return ret;
524 }
525 
526 static int hns_roce_ib_umem_write_mr(struct hns_roce_mr *mr,
527 				     struct ib_umem *umem)
528 {
529 	int i = 0;
530 	int entry;
531 	struct scatterlist *sg;
532 
533 	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
534 		mr->pbl_buf[i] = ((u64)sg_dma_address(sg)) >> 12;
535 		i++;
536 	}
537 
538 	/* Memory barrier */
539 	mb();
540 
541 	return 0;
542 }
543 
544 struct ib_mr *hns_roce_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
545 				   u64 virt_addr, int access_flags,
546 				   struct ib_udata *udata)
547 {
548 	struct hns_roce_dev *hr_dev = to_hr_dev(pd->device);
549 	struct device *dev = &hr_dev->pdev->dev;
550 	struct hns_roce_mr *mr = NULL;
551 	int ret = 0;
552 	int n = 0;
553 
554 	mr = kmalloc(sizeof(*mr), GFP_KERNEL);
555 	if (!mr)
556 		return ERR_PTR(-ENOMEM);
557 
558 	mr->umem = ib_umem_get(pd->uobject->context, start, length,
559 			       access_flags, 0);
560 	if (IS_ERR(mr->umem)) {
561 		ret = PTR_ERR(mr->umem);
562 		goto err_free;
563 	}
564 
565 	n = ib_umem_page_count(mr->umem);
566 	if (mr->umem->page_size != HNS_ROCE_HEM_PAGE_SIZE) {
567 		dev_err(dev, "Just support 4K page size but is 0x%x now!\n",
568 			mr->umem->page_size);
569 		ret = -EINVAL;
570 		goto err_umem;
571 	}
572 
573 	if (n > HNS_ROCE_MAX_MTPT_PBL_NUM) {
574 		dev_err(dev, " MR len %lld err. MR is limited to 4G at most!\n",
575 			length);
576 		ret = -EINVAL;
577 		goto err_umem;
578 	}
579 
580 	ret = hns_roce_mr_alloc(hr_dev, to_hr_pd(pd)->pdn, virt_addr, length,
581 				access_flags, n, mr);
582 	if (ret)
583 		goto err_umem;
584 
585 	ret = hns_roce_ib_umem_write_mr(mr, mr->umem);
586 	if (ret)
587 		goto err_mr;
588 
589 	ret = hns_roce_mr_enable(hr_dev, mr);
590 	if (ret)
591 		goto err_mr;
592 
593 	mr->ibmr.rkey = mr->ibmr.lkey = mr->key;
594 
595 	return &mr->ibmr;
596 
597 err_mr:
598 	hns_roce_mr_free(hr_dev, mr);
599 
600 err_umem:
601 	ib_umem_release(mr->umem);
602 
603 err_free:
604 	kfree(mr);
605 	return ERR_PTR(ret);
606 }
607 
608 int hns_roce_dereg_mr(struct ib_mr *ibmr)
609 {
610 	struct hns_roce_dev *hr_dev = to_hr_dev(ibmr->device);
611 	struct hns_roce_mr *mr = to_hr_mr(ibmr);
612 	int ret = 0;
613 
614 	if (hr_dev->hw->dereg_mr) {
615 		ret = hr_dev->hw->dereg_mr(hr_dev, mr);
616 	} else {
617 		hns_roce_mr_free(hr_dev, mr);
618 
619 		if (mr->umem)
620 			ib_umem_release(mr->umem);
621 
622 		kfree(mr);
623 	}
624 
625 	return ret;
626 }
627