xref: /linux/drivers/net/ethernet/mellanox/mlx4/alloc.c (revision 214fa1c4)
1 /*
2  * Copyright (c) 2006, 2007 Cisco Systems, Inc.  All rights reserved.
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/errno.h>
35 #include <linux/slab.h>
36 #include <linux/mm.h>
37 #include <linux/export.h>
38 #include <linux/bitmap.h>
39 #include <linux/dma-mapping.h>
40 #include <linux/vmalloc.h>
41 
42 #include "mlx4.h"
43 
mlx4_bitmap_alloc(struct mlx4_bitmap * bitmap)44 u32 mlx4_bitmap_alloc(struct mlx4_bitmap *bitmap)
45 {
46 	u32 obj;
47 
48 	spin_lock(&bitmap->lock);
49 
50 	obj = find_next_zero_bit(bitmap->table, bitmap->max, bitmap->last);
51 	if (obj >= bitmap->max) {
52 		bitmap->top = (bitmap->top + bitmap->max + bitmap->reserved_top)
53 				& bitmap->mask;
54 		obj = find_first_zero_bit(bitmap->table, bitmap->max);
55 	}
56 
57 	if (obj < bitmap->max) {
58 		set_bit(obj, bitmap->table);
59 		bitmap->last = (obj + 1);
60 		if (bitmap->last == bitmap->max)
61 			bitmap->last = 0;
62 		obj |= bitmap->top;
63 	} else
64 		obj = -1;
65 
66 	if (obj != -1)
67 		--bitmap->avail;
68 
69 	spin_unlock(&bitmap->lock);
70 
71 	return obj;
72 }
73 
mlx4_bitmap_free(struct mlx4_bitmap * bitmap,u32 obj,int use_rr)74 void mlx4_bitmap_free(struct mlx4_bitmap *bitmap, u32 obj, int use_rr)
75 {
76 	mlx4_bitmap_free_range(bitmap, obj, 1, use_rr);
77 }
78 
find_aligned_range(unsigned long * bitmap,u32 start,u32 nbits,int len,int align,u32 skip_mask)79 static unsigned long find_aligned_range(unsigned long *bitmap,
80 					u32 start, u32 nbits,
81 					int len, int align, u32 skip_mask)
82 {
83 	unsigned long end, i;
84 
85 again:
86 	start = ALIGN(start, align);
87 
88 	while ((start < nbits) && (test_bit(start, bitmap) ||
89 				   (start & skip_mask)))
90 		start += align;
91 
92 	if (start >= nbits)
93 		return -1;
94 
95 	end = start+len;
96 	if (end > nbits)
97 		return -1;
98 
99 	for (i = start + 1; i < end; i++) {
100 		if (test_bit(i, bitmap) || ((u32)i & skip_mask)) {
101 			start = i + 1;
102 			goto again;
103 		}
104 	}
105 
106 	return start;
107 }
108 
mlx4_bitmap_alloc_range(struct mlx4_bitmap * bitmap,int cnt,int align,u32 skip_mask)109 u32 mlx4_bitmap_alloc_range(struct mlx4_bitmap *bitmap, int cnt,
110 			    int align, u32 skip_mask)
111 {
112 	u32 obj;
113 
114 	if (likely(cnt == 1 && align == 1 && !skip_mask))
115 		return mlx4_bitmap_alloc(bitmap);
116 
117 	spin_lock(&bitmap->lock);
118 
119 	obj = find_aligned_range(bitmap->table, bitmap->last,
120 				 bitmap->max, cnt, align, skip_mask);
121 	if (obj >= bitmap->max) {
122 		bitmap->top = (bitmap->top + bitmap->max + bitmap->reserved_top)
123 				& bitmap->mask;
124 		obj = find_aligned_range(bitmap->table, 0, bitmap->max,
125 					 cnt, align, skip_mask);
126 	}
127 
128 	if (obj < bitmap->max) {
129 		bitmap_set(bitmap->table, obj, cnt);
130 		if (obj == bitmap->last) {
131 			bitmap->last = (obj + cnt);
132 			if (bitmap->last >= bitmap->max)
133 				bitmap->last = 0;
134 		}
135 		obj |= bitmap->top;
136 	} else
137 		obj = -1;
138 
139 	if (obj != -1)
140 		bitmap->avail -= cnt;
141 
142 	spin_unlock(&bitmap->lock);
143 
144 	return obj;
145 }
146 
mlx4_bitmap_avail(struct mlx4_bitmap * bitmap)147 u32 mlx4_bitmap_avail(struct mlx4_bitmap *bitmap)
148 {
149 	return bitmap->avail;
150 }
151 
mlx4_bitmap_masked_value(struct mlx4_bitmap * bitmap,u32 obj)152 static u32 mlx4_bitmap_masked_value(struct mlx4_bitmap *bitmap, u32 obj)
153 {
154 	return obj & (bitmap->max + bitmap->reserved_top - 1);
155 }
156 
mlx4_bitmap_free_range(struct mlx4_bitmap * bitmap,u32 obj,int cnt,int use_rr)157 void mlx4_bitmap_free_range(struct mlx4_bitmap *bitmap, u32 obj, int cnt,
158 			    int use_rr)
159 {
160 	obj &= bitmap->max + bitmap->reserved_top - 1;
161 
162 	spin_lock(&bitmap->lock);
163 	if (!use_rr) {
164 		bitmap->last = min(bitmap->last, obj);
165 		bitmap->top = (bitmap->top + bitmap->max + bitmap->reserved_top)
166 				& bitmap->mask;
167 	}
168 	bitmap_clear(bitmap->table, obj, cnt);
169 	bitmap->avail += cnt;
170 	spin_unlock(&bitmap->lock);
171 }
172 
mlx4_bitmap_init(struct mlx4_bitmap * bitmap,u32 num,u32 mask,u32 reserved_bot,u32 reserved_top)173 int mlx4_bitmap_init(struct mlx4_bitmap *bitmap, u32 num, u32 mask,
174 		     u32 reserved_bot, u32 reserved_top)
175 {
176 	/* num must be a power of 2 */
177 	if (num != roundup_pow_of_two(num))
178 		return -EINVAL;
179 
180 	bitmap->last = 0;
181 	bitmap->top  = 0;
182 	bitmap->max  = num - reserved_top;
183 	bitmap->mask = mask;
184 	bitmap->reserved_top = reserved_top;
185 	bitmap->avail = num - reserved_top - reserved_bot;
186 	bitmap->effective_len = bitmap->avail;
187 	spin_lock_init(&bitmap->lock);
188 	bitmap->table = bitmap_zalloc(bitmap->max, GFP_KERNEL);
189 	if (!bitmap->table)
190 		return -ENOMEM;
191 
192 	bitmap_set(bitmap->table, 0, reserved_bot);
193 
194 	return 0;
195 }
196 
mlx4_bitmap_cleanup(struct mlx4_bitmap * bitmap)197 void mlx4_bitmap_cleanup(struct mlx4_bitmap *bitmap)
198 {
199 	bitmap_free(bitmap->table);
200 }
201 
202 struct mlx4_zone_allocator {
203 	struct list_head		entries;
204 	struct list_head		prios;
205 	u32				last_uid;
206 	u32				mask;
207 	/* protect the zone_allocator from concurrent accesses */
208 	spinlock_t			lock;
209 	enum mlx4_zone_alloc_flags	flags;
210 };
211 
212 struct mlx4_zone_entry {
213 	struct list_head		list;
214 	struct list_head		prio_list;
215 	u32				uid;
216 	struct mlx4_zone_allocator	*allocator;
217 	struct mlx4_bitmap		*bitmap;
218 	int				use_rr;
219 	int				priority;
220 	int				offset;
221 	enum mlx4_zone_flags		flags;
222 };
223 
mlx4_zone_allocator_create(enum mlx4_zone_alloc_flags flags)224 struct mlx4_zone_allocator *mlx4_zone_allocator_create(enum mlx4_zone_alloc_flags flags)
225 {
226 	struct mlx4_zone_allocator *zones = kmalloc(sizeof(*zones), GFP_KERNEL);
227 
228 	if (NULL == zones)
229 		return NULL;
230 
231 	INIT_LIST_HEAD(&zones->entries);
232 	INIT_LIST_HEAD(&zones->prios);
233 	spin_lock_init(&zones->lock);
234 	zones->last_uid = 0;
235 	zones->mask = 0;
236 	zones->flags = flags;
237 
238 	return zones;
239 }
240 
mlx4_zone_add_one(struct mlx4_zone_allocator * zone_alloc,struct mlx4_bitmap * bitmap,u32 flags,int priority,int offset,u32 * puid)241 int mlx4_zone_add_one(struct mlx4_zone_allocator *zone_alloc,
242 		      struct mlx4_bitmap *bitmap,
243 		      u32 flags,
244 		      int priority,
245 		      int offset,
246 		      u32 *puid)
247 {
248 	u32 mask = mlx4_bitmap_masked_value(bitmap, (u32)-1);
249 	struct mlx4_zone_entry *it;
250 	struct mlx4_zone_entry *zone = kmalloc(sizeof(*zone), GFP_KERNEL);
251 
252 	if (NULL == zone)
253 		return -ENOMEM;
254 
255 	zone->flags = flags;
256 	zone->bitmap = bitmap;
257 	zone->use_rr = (flags & MLX4_ZONE_USE_RR) ? MLX4_USE_RR : 0;
258 	zone->priority = priority;
259 	zone->offset = offset;
260 
261 	spin_lock(&zone_alloc->lock);
262 
263 	zone->uid = zone_alloc->last_uid++;
264 	zone->allocator = zone_alloc;
265 
266 	if (zone_alloc->mask < mask)
267 		zone_alloc->mask = mask;
268 
269 	list_for_each_entry(it, &zone_alloc->prios, prio_list)
270 		if (it->priority >= priority)
271 			break;
272 
273 	if (&it->prio_list == &zone_alloc->prios || it->priority > priority)
274 		list_add_tail(&zone->prio_list, &it->prio_list);
275 	list_add_tail(&zone->list, &it->list);
276 
277 	spin_unlock(&zone_alloc->lock);
278 
279 	*puid = zone->uid;
280 
281 	return 0;
282 }
283 
284 /* Should be called under a lock */
__mlx4_zone_remove_one_entry(struct mlx4_zone_entry * entry)285 static void __mlx4_zone_remove_one_entry(struct mlx4_zone_entry *entry)
286 {
287 	struct mlx4_zone_allocator *zone_alloc = entry->allocator;
288 
289 	if (!list_empty(&entry->prio_list)) {
290 		/* Check if we need to add an alternative node to the prio list */
291 		if (!list_is_last(&entry->list, &zone_alloc->entries)) {
292 			struct mlx4_zone_entry *next = list_first_entry(&entry->list,
293 									typeof(*next),
294 									list);
295 
296 			if (next->priority == entry->priority)
297 				list_add_tail(&next->prio_list, &entry->prio_list);
298 		}
299 
300 		list_del(&entry->prio_list);
301 	}
302 
303 	list_del(&entry->list);
304 
305 	if (zone_alloc->flags & MLX4_ZONE_ALLOC_FLAGS_NO_OVERLAP) {
306 		u32 mask = 0;
307 		struct mlx4_zone_entry *it;
308 
309 		list_for_each_entry(it, &zone_alloc->prios, prio_list) {
310 			u32 cur_mask = mlx4_bitmap_masked_value(it->bitmap, (u32)-1);
311 
312 			if (mask < cur_mask)
313 				mask = cur_mask;
314 		}
315 		zone_alloc->mask = mask;
316 	}
317 }
318 
mlx4_zone_allocator_destroy(struct mlx4_zone_allocator * zone_alloc)319 void mlx4_zone_allocator_destroy(struct mlx4_zone_allocator *zone_alloc)
320 {
321 	struct mlx4_zone_entry *zone, *tmp;
322 
323 	spin_lock(&zone_alloc->lock);
324 
325 	list_for_each_entry_safe(zone, tmp, &zone_alloc->entries, list) {
326 		list_del(&zone->list);
327 		list_del(&zone->prio_list);
328 		kfree(zone);
329 	}
330 
331 	spin_unlock(&zone_alloc->lock);
332 	kfree(zone_alloc);
333 }
334 
335 /* Should be called under a lock */
__mlx4_alloc_from_zone(struct mlx4_zone_entry * zone,int count,int align,u32 skip_mask,u32 * puid)336 static u32 __mlx4_alloc_from_zone(struct mlx4_zone_entry *zone, int count,
337 				  int align, u32 skip_mask, u32 *puid)
338 {
339 	u32 uid = 0;
340 	u32 res;
341 	struct mlx4_zone_allocator *zone_alloc = zone->allocator;
342 	struct mlx4_zone_entry *curr_node;
343 
344 	res = mlx4_bitmap_alloc_range(zone->bitmap, count,
345 				      align, skip_mask);
346 
347 	if (res != (u32)-1) {
348 		res += zone->offset;
349 		uid = zone->uid;
350 		goto out;
351 	}
352 
353 	list_for_each_entry(curr_node, &zone_alloc->prios, prio_list) {
354 		if (unlikely(curr_node->priority == zone->priority))
355 			break;
356 	}
357 
358 	if (zone->flags & MLX4_ZONE_ALLOW_ALLOC_FROM_LOWER_PRIO) {
359 		struct mlx4_zone_entry *it = curr_node;
360 
361 		list_for_each_entry_continue_reverse(it, &zone_alloc->entries, list) {
362 			res = mlx4_bitmap_alloc_range(it->bitmap, count,
363 						      align, skip_mask);
364 			if (res != (u32)-1) {
365 				res += it->offset;
366 				uid = it->uid;
367 				goto out;
368 			}
369 		}
370 	}
371 
372 	if (zone->flags & MLX4_ZONE_ALLOW_ALLOC_FROM_EQ_PRIO) {
373 		struct mlx4_zone_entry *it = curr_node;
374 
375 		list_for_each_entry_from(it, &zone_alloc->entries, list) {
376 			if (unlikely(it == zone))
377 				continue;
378 
379 			if (unlikely(it->priority != curr_node->priority))
380 				break;
381 
382 			res = mlx4_bitmap_alloc_range(it->bitmap, count,
383 						      align, skip_mask);
384 			if (res != (u32)-1) {
385 				res += it->offset;
386 				uid = it->uid;
387 				goto out;
388 			}
389 		}
390 	}
391 
392 	if (zone->flags & MLX4_ZONE_FALLBACK_TO_HIGHER_PRIO) {
393 		if (list_is_last(&curr_node->prio_list, &zone_alloc->prios))
394 			goto out;
395 
396 		curr_node = list_first_entry(&curr_node->prio_list,
397 					     typeof(*curr_node),
398 					     prio_list);
399 
400 		list_for_each_entry_from(curr_node, &zone_alloc->entries, list) {
401 			res = mlx4_bitmap_alloc_range(curr_node->bitmap, count,
402 						      align, skip_mask);
403 			if (res != (u32)-1) {
404 				res += curr_node->offset;
405 				uid = curr_node->uid;
406 				goto out;
407 			}
408 		}
409 	}
410 
411 out:
412 	if (NULL != puid && res != (u32)-1)
413 		*puid = uid;
414 	return res;
415 }
416 
417 /* Should be called under a lock */
__mlx4_free_from_zone(struct mlx4_zone_entry * zone,u32 obj,u32 count)418 static void __mlx4_free_from_zone(struct mlx4_zone_entry *zone, u32 obj,
419 				  u32 count)
420 {
421 	mlx4_bitmap_free_range(zone->bitmap, obj - zone->offset, count, zone->use_rr);
422 }
423 
424 /* Should be called under a lock */
__mlx4_find_zone_by_uid(struct mlx4_zone_allocator * zones,u32 uid)425 static struct mlx4_zone_entry *__mlx4_find_zone_by_uid(
426 		struct mlx4_zone_allocator *zones, u32 uid)
427 {
428 	struct mlx4_zone_entry *zone;
429 
430 	list_for_each_entry(zone, &zones->entries, list) {
431 		if (zone->uid == uid)
432 			return zone;
433 	}
434 
435 	return NULL;
436 }
437 
mlx4_zone_get_bitmap(struct mlx4_zone_allocator * zones,u32 uid)438 struct mlx4_bitmap *mlx4_zone_get_bitmap(struct mlx4_zone_allocator *zones, u32 uid)
439 {
440 	struct mlx4_zone_entry *zone;
441 	struct mlx4_bitmap *bitmap;
442 
443 	spin_lock(&zones->lock);
444 
445 	zone = __mlx4_find_zone_by_uid(zones, uid);
446 
447 	bitmap = zone == NULL ? NULL : zone->bitmap;
448 
449 	spin_unlock(&zones->lock);
450 
451 	return bitmap;
452 }
453 
mlx4_zone_remove_one(struct mlx4_zone_allocator * zones,u32 uid)454 int mlx4_zone_remove_one(struct mlx4_zone_allocator *zones, u32 uid)
455 {
456 	struct mlx4_zone_entry *zone;
457 	int res = 0;
458 
459 	spin_lock(&zones->lock);
460 
461 	zone = __mlx4_find_zone_by_uid(zones, uid);
462 
463 	if (NULL == zone) {
464 		res = -1;
465 		goto out;
466 	}
467 
468 	__mlx4_zone_remove_one_entry(zone);
469 
470 out:
471 	spin_unlock(&zones->lock);
472 	kfree(zone);
473 
474 	return res;
475 }
476 
477 /* Should be called under a lock */
__mlx4_find_zone_by_uid_unique(struct mlx4_zone_allocator * zones,u32 obj)478 static struct mlx4_zone_entry *__mlx4_find_zone_by_uid_unique(
479 		struct mlx4_zone_allocator *zones, u32 obj)
480 {
481 	struct mlx4_zone_entry *zone, *zone_candidate = NULL;
482 	u32 dist = (u32)-1;
483 
484 	/* Search for the smallest zone that this obj could be
485 	 * allocated from. This is done in order to handle
486 	 * situations when small bitmaps are allocated from bigger
487 	 * bitmaps (and the allocated space is marked as reserved in
488 	 * the bigger bitmap.
489 	 */
490 	list_for_each_entry(zone, &zones->entries, list) {
491 		if (obj >= zone->offset) {
492 			u32 mobj = (obj - zone->offset) & zones->mask;
493 
494 			if (mobj < zone->bitmap->max) {
495 				u32 curr_dist = zone->bitmap->effective_len;
496 
497 				if (curr_dist < dist) {
498 					dist = curr_dist;
499 					zone_candidate = zone;
500 				}
501 			}
502 		}
503 	}
504 
505 	return zone_candidate;
506 }
507 
mlx4_zone_alloc_entries(struct mlx4_zone_allocator * zones,u32 uid,int count,int align,u32 skip_mask,u32 * puid)508 u32 mlx4_zone_alloc_entries(struct mlx4_zone_allocator *zones, u32 uid, int count,
509 			    int align, u32 skip_mask, u32 *puid)
510 {
511 	struct mlx4_zone_entry *zone;
512 	int res = -1;
513 
514 	spin_lock(&zones->lock);
515 
516 	zone = __mlx4_find_zone_by_uid(zones, uid);
517 
518 	if (NULL == zone)
519 		goto out;
520 
521 	res = __mlx4_alloc_from_zone(zone, count, align, skip_mask, puid);
522 
523 out:
524 	spin_unlock(&zones->lock);
525 
526 	return res;
527 }
528 
mlx4_zone_free_entries(struct mlx4_zone_allocator * zones,u32 uid,u32 obj,u32 count)529 u32 mlx4_zone_free_entries(struct mlx4_zone_allocator *zones, u32 uid, u32 obj, u32 count)
530 {
531 	struct mlx4_zone_entry *zone;
532 	int res = 0;
533 
534 	spin_lock(&zones->lock);
535 
536 	zone = __mlx4_find_zone_by_uid(zones, uid);
537 
538 	if (NULL == zone) {
539 		res = -1;
540 		goto out;
541 	}
542 
543 	__mlx4_free_from_zone(zone, obj, count);
544 
545 out:
546 	spin_unlock(&zones->lock);
547 
548 	return res;
549 }
550 
mlx4_zone_free_entries_unique(struct mlx4_zone_allocator * zones,u32 obj,u32 count)551 u32 mlx4_zone_free_entries_unique(struct mlx4_zone_allocator *zones, u32 obj, u32 count)
552 {
553 	struct mlx4_zone_entry *zone;
554 	int res;
555 
556 	if (!(zones->flags & MLX4_ZONE_ALLOC_FLAGS_NO_OVERLAP))
557 		return -EFAULT;
558 
559 	spin_lock(&zones->lock);
560 
561 	zone = __mlx4_find_zone_by_uid_unique(zones, obj);
562 
563 	if (NULL == zone) {
564 		res = -1;
565 		goto out;
566 	}
567 
568 	__mlx4_free_from_zone(zone, obj, count);
569 	res = 0;
570 
571 out:
572 	spin_unlock(&zones->lock);
573 
574 	return res;
575 }
576 
mlx4_buf_direct_alloc(struct mlx4_dev * dev,int size,struct mlx4_buf * buf)577 static int mlx4_buf_direct_alloc(struct mlx4_dev *dev, int size,
578 				 struct mlx4_buf *buf)
579 {
580 	dma_addr_t t;
581 
582 	buf->nbufs        = 1;
583 	buf->npages       = 1;
584 	buf->page_shift   = get_order(size) + PAGE_SHIFT;
585 	buf->direct.buf   =
586 		dma_alloc_coherent(&dev->persist->pdev->dev, size, &t,
587 				   GFP_KERNEL);
588 	if (!buf->direct.buf)
589 		return -ENOMEM;
590 
591 	buf->direct.map = t;
592 
593 	while (t & ((1 << buf->page_shift) - 1)) {
594 		--buf->page_shift;
595 		buf->npages *= 2;
596 	}
597 
598 	return 0;
599 }
600 
601 /* Handling for queue buffers -- we allocate a bunch of memory and
602  * register it in a memory region at HCA virtual address 0. If the
603  *  requested size is > max_direct, we split the allocation into
604  *  multiple pages, so we don't require too much contiguous memory.
605  */
mlx4_buf_alloc(struct mlx4_dev * dev,int size,int max_direct,struct mlx4_buf * buf)606 int mlx4_buf_alloc(struct mlx4_dev *dev, int size, int max_direct,
607 		   struct mlx4_buf *buf)
608 {
609 	if (size <= max_direct) {
610 		return mlx4_buf_direct_alloc(dev, size, buf);
611 	} else {
612 		dma_addr_t t;
613 		int i;
614 
615 		buf->direct.buf = NULL;
616 		buf->nbufs      = DIV_ROUND_UP(size, PAGE_SIZE);
617 		buf->npages	= buf->nbufs;
618 		buf->page_shift  = PAGE_SHIFT;
619 		buf->page_list   = kcalloc(buf->nbufs, sizeof(*buf->page_list),
620 					   GFP_KERNEL);
621 		if (!buf->page_list)
622 			return -ENOMEM;
623 
624 		for (i = 0; i < buf->nbufs; ++i) {
625 			buf->page_list[i].buf =
626 				dma_alloc_coherent(&dev->persist->pdev->dev,
627 						   PAGE_SIZE, &t, GFP_KERNEL);
628 			if (!buf->page_list[i].buf)
629 				goto err_free;
630 
631 			buf->page_list[i].map = t;
632 		}
633 	}
634 
635 	return 0;
636 
637 err_free:
638 	mlx4_buf_free(dev, size, buf);
639 
640 	return -ENOMEM;
641 }
642 EXPORT_SYMBOL_GPL(mlx4_buf_alloc);
643 
mlx4_buf_free(struct mlx4_dev * dev,int size,struct mlx4_buf * buf)644 void mlx4_buf_free(struct mlx4_dev *dev, int size, struct mlx4_buf *buf)
645 {
646 	if (buf->nbufs == 1) {
647 		dma_free_coherent(&dev->persist->pdev->dev, size,
648 				  buf->direct.buf, buf->direct.map);
649 	} else {
650 		int i;
651 
652 		for (i = 0; i < buf->nbufs; ++i)
653 			if (buf->page_list[i].buf)
654 				dma_free_coherent(&dev->persist->pdev->dev,
655 						  PAGE_SIZE,
656 						  buf->page_list[i].buf,
657 						  buf->page_list[i].map);
658 		kfree(buf->page_list);
659 	}
660 }
661 EXPORT_SYMBOL_GPL(mlx4_buf_free);
662 
mlx4_alloc_db_pgdir(struct device * dma_device)663 static struct mlx4_db_pgdir *mlx4_alloc_db_pgdir(struct device *dma_device)
664 {
665 	struct mlx4_db_pgdir *pgdir;
666 
667 	pgdir = kzalloc(sizeof(*pgdir), GFP_KERNEL);
668 	if (!pgdir)
669 		return NULL;
670 
671 	bitmap_fill(pgdir->order1, MLX4_DB_PER_PAGE / 2);
672 	pgdir->bits[0] = pgdir->order0;
673 	pgdir->bits[1] = pgdir->order1;
674 	pgdir->db_page = dma_alloc_coherent(dma_device, PAGE_SIZE,
675 					    &pgdir->db_dma, GFP_KERNEL);
676 	if (!pgdir->db_page) {
677 		kfree(pgdir);
678 		return NULL;
679 	}
680 
681 	return pgdir;
682 }
683 
mlx4_alloc_db_from_pgdir(struct mlx4_db_pgdir * pgdir,struct mlx4_db * db,int order)684 static int mlx4_alloc_db_from_pgdir(struct mlx4_db_pgdir *pgdir,
685 				    struct mlx4_db *db, int order)
686 {
687 	int o;
688 	int i;
689 
690 	for (o = order; o <= 1; ++o) {
691 		i = find_first_bit(pgdir->bits[o], MLX4_DB_PER_PAGE >> o);
692 		if (i < MLX4_DB_PER_PAGE >> o)
693 			goto found;
694 	}
695 
696 	return -ENOMEM;
697 
698 found:
699 	clear_bit(i, pgdir->bits[o]);
700 
701 	i <<= o;
702 
703 	if (o > order)
704 		set_bit(i ^ 1, pgdir->bits[order]);
705 
706 	db->u.pgdir = pgdir;
707 	db->index   = i;
708 	db->db      = pgdir->db_page + db->index;
709 	db->dma     = pgdir->db_dma  + db->index * 4;
710 	db->order   = order;
711 
712 	return 0;
713 }
714 
mlx4_db_alloc(struct mlx4_dev * dev,struct mlx4_db * db,int order)715 int mlx4_db_alloc(struct mlx4_dev *dev, struct mlx4_db *db, int order)
716 {
717 	struct mlx4_priv *priv = mlx4_priv(dev);
718 	struct mlx4_db_pgdir *pgdir;
719 	int ret = 0;
720 
721 	mutex_lock(&priv->pgdir_mutex);
722 
723 	list_for_each_entry(pgdir, &priv->pgdir_list, list)
724 		if (!mlx4_alloc_db_from_pgdir(pgdir, db, order))
725 			goto out;
726 
727 	pgdir = mlx4_alloc_db_pgdir(&dev->persist->pdev->dev);
728 	if (!pgdir) {
729 		ret = -ENOMEM;
730 		goto out;
731 	}
732 
733 	list_add(&pgdir->list, &priv->pgdir_list);
734 
735 	/* This should never fail -- we just allocated an empty page: */
736 	WARN_ON(mlx4_alloc_db_from_pgdir(pgdir, db, order));
737 
738 out:
739 	mutex_unlock(&priv->pgdir_mutex);
740 
741 	return ret;
742 }
743 EXPORT_SYMBOL_GPL(mlx4_db_alloc);
744 
mlx4_db_free(struct mlx4_dev * dev,struct mlx4_db * db)745 void mlx4_db_free(struct mlx4_dev *dev, struct mlx4_db *db)
746 {
747 	struct mlx4_priv *priv = mlx4_priv(dev);
748 	int o;
749 	int i;
750 
751 	mutex_lock(&priv->pgdir_mutex);
752 
753 	o = db->order;
754 	i = db->index;
755 
756 	if (db->order == 0 && test_bit(i ^ 1, db->u.pgdir->order0)) {
757 		clear_bit(i ^ 1, db->u.pgdir->order0);
758 		++o;
759 	}
760 	i >>= o;
761 	set_bit(i, db->u.pgdir->bits[o]);
762 
763 	if (bitmap_full(db->u.pgdir->order1, MLX4_DB_PER_PAGE / 2)) {
764 		dma_free_coherent(&dev->persist->pdev->dev, PAGE_SIZE,
765 				  db->u.pgdir->db_page, db->u.pgdir->db_dma);
766 		list_del(&db->u.pgdir->list);
767 		kfree(db->u.pgdir);
768 	}
769 
770 	mutex_unlock(&priv->pgdir_mutex);
771 }
772 EXPORT_SYMBOL_GPL(mlx4_db_free);
773 
mlx4_alloc_hwq_res(struct mlx4_dev * dev,struct mlx4_hwq_resources * wqres,int size)774 int mlx4_alloc_hwq_res(struct mlx4_dev *dev, struct mlx4_hwq_resources *wqres,
775 		       int size)
776 {
777 	int err;
778 
779 	err = mlx4_db_alloc(dev, &wqres->db, 1);
780 	if (err)
781 		return err;
782 
783 	*wqres->db.db = 0;
784 
785 	err = mlx4_buf_direct_alloc(dev, size, &wqres->buf);
786 	if (err)
787 		goto err_db;
788 
789 	err = mlx4_mtt_init(dev, wqres->buf.npages, wqres->buf.page_shift,
790 			    &wqres->mtt);
791 	if (err)
792 		goto err_buf;
793 
794 	err = mlx4_buf_write_mtt(dev, &wqres->mtt, &wqres->buf);
795 	if (err)
796 		goto err_mtt;
797 
798 	return 0;
799 
800 err_mtt:
801 	mlx4_mtt_cleanup(dev, &wqres->mtt);
802 err_buf:
803 	mlx4_buf_free(dev, size, &wqres->buf);
804 err_db:
805 	mlx4_db_free(dev, &wqres->db);
806 
807 	return err;
808 }
809 EXPORT_SYMBOL_GPL(mlx4_alloc_hwq_res);
810 
mlx4_free_hwq_res(struct mlx4_dev * dev,struct mlx4_hwq_resources * wqres,int size)811 void mlx4_free_hwq_res(struct mlx4_dev *dev, struct mlx4_hwq_resources *wqres,
812 		       int size)
813 {
814 	mlx4_mtt_cleanup(dev, &wqres->mtt);
815 	mlx4_buf_free(dev, size, &wqres->buf);
816 	mlx4_db_free(dev, &wqres->db);
817 }
818 EXPORT_SYMBOL_GPL(mlx4_free_hwq_res);
819