xref: /freebsd/sys/dev/mlx4/mlx4_core/mlx4_alloc.c (revision 3873b9a8b3a51731e5db92ab0c5c59797e416015)
1 /*
2  * Copyright (c) 2006, 2007 Cisco Systems, Inc.  All rights reserved.
3  * Copyright (c) 2007, 2008, 2014 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/module.h>
38 #include <linux/dma-mapping.h>
39 #include <linux/vmalloc.h>
40 
41 #include "mlx4.h"
42 
mlx4_bitmap_alloc(struct mlx4_bitmap * bitmap)43 u32 mlx4_bitmap_alloc(struct mlx4_bitmap *bitmap)
44 {
45 	u32 obj;
46 
47 	spin_lock(&bitmap->lock);
48 
49 	obj = find_next_zero_bit(bitmap->table, bitmap->max, bitmap->last);
50 	if (obj >= bitmap->max) {
51 		bitmap->top = (bitmap->top + bitmap->max + bitmap->reserved_top)
52 				& bitmap->mask;
53 		obj = find_first_zero_bit(bitmap->table, bitmap->max);
54 	}
55 
56 	if (obj < bitmap->max) {
57 		set_bit(obj, bitmap->table);
58 		bitmap->last = (obj + 1);
59 		if (bitmap->last == bitmap->max)
60 			bitmap->last = 0;
61 		obj |= bitmap->top;
62 	} else
63 		obj = -1;
64 
65 	if (obj != -1)
66 		--bitmap->avail;
67 
68 	spin_unlock(&bitmap->lock);
69 
70 	return obj;
71 }
72 
mlx4_bitmap_free(struct mlx4_bitmap * bitmap,u32 obj,int use_rr)73 void mlx4_bitmap_free(struct mlx4_bitmap *bitmap, u32 obj, int use_rr)
74 {
75 	mlx4_bitmap_free_range(bitmap, obj, 1, use_rr);
76 }
77 
find_aligned_range(unsigned long * bitmap,u32 start,u32 nbits,int len,int align,u32 skip_mask)78 static unsigned long find_aligned_range(unsigned long *bitmap,
79 					u32 start, u32 nbits,
80 					int len, int align, u32 skip_mask)
81 {
82 	unsigned long end, i;
83 
84 again:
85 	start = ALIGN(start, align);
86 
87 	while ((start < nbits) && (test_bit(start, bitmap) ||
88 				   (start & skip_mask)))
89 		start += align;
90 
91 	if (start >= nbits)
92 		return -1;
93 
94 	end = start+len;
95 	if (end > nbits)
96 		return -1;
97 
98 	for (i = start + 1; i < end; i++) {
99 		if (test_bit(i, bitmap) || ((u32)i & skip_mask)) {
100 			start = i + 1;
101 			goto again;
102 		}
103 	}
104 
105 	return start;
106 }
107 
mlx4_bitmap_alloc_range(struct mlx4_bitmap * bitmap,int cnt,int align,u32 skip_mask)108 u32 mlx4_bitmap_alloc_range(struct mlx4_bitmap *bitmap, int cnt,
109 			    int align, u32 skip_mask)
110 {
111 	u32 obj;
112 
113 	if (likely(cnt == 1 && align == 1 && !skip_mask))
114 		return mlx4_bitmap_alloc(bitmap);
115 
116 	spin_lock(&bitmap->lock);
117 
118 	obj = find_aligned_range(bitmap->table, bitmap->last,
119 				 bitmap->max, cnt, align, skip_mask);
120 	if (obj >= bitmap->max) {
121 		bitmap->top = (bitmap->top + bitmap->max + bitmap->reserved_top)
122 				& bitmap->mask;
123 		obj = find_aligned_range(bitmap->table, 0, bitmap->max,
124 					 cnt, align, skip_mask);
125 	}
126 
127 	if (obj < bitmap->max) {
128 		bitmap_set(bitmap->table, obj, cnt);
129 		if (obj == bitmap->last) {
130 			bitmap->last = (obj + cnt);
131 			if (bitmap->last >= bitmap->max)
132 				bitmap->last = 0;
133 		}
134 		obj |= bitmap->top;
135 	} else
136 		obj = -1;
137 
138 	if (obj != -1)
139 		bitmap->avail -= cnt;
140 
141 	spin_unlock(&bitmap->lock);
142 
143 	return obj;
144 }
145 
mlx4_bitmap_avail(struct mlx4_bitmap * bitmap)146 u32 mlx4_bitmap_avail(struct mlx4_bitmap *bitmap)
147 {
148 	return bitmap->avail;
149 }
150 
mlx4_bitmap_masked_value(struct mlx4_bitmap * bitmap,u32 obj)151 static u32 mlx4_bitmap_masked_value(struct mlx4_bitmap *bitmap, u32 obj)
152 {
153 	return obj & (bitmap->max + bitmap->reserved_top - 1);
154 }
155 
mlx4_bitmap_free_range(struct mlx4_bitmap * bitmap,u32 obj,int cnt,int use_rr)156 void mlx4_bitmap_free_range(struct mlx4_bitmap *bitmap, u32 obj, int cnt,
157 			    int use_rr)
158 {
159 	obj &= bitmap->max + bitmap->reserved_top - 1;
160 
161 	spin_lock(&bitmap->lock);
162 	if (!use_rr) {
163 		bitmap->last = min(bitmap->last, obj);
164 		bitmap->top = (bitmap->top + bitmap->max + bitmap->reserved_top)
165 				& bitmap->mask;
166 	}
167 	bitmap_clear(bitmap->table, obj, cnt);
168 	bitmap->avail += cnt;
169 	spin_unlock(&bitmap->lock);
170 }
171 
mlx4_bitmap_init(struct mlx4_bitmap * bitmap,u32 num,u32 mask,u32 reserved_bot,u32 reserved_top)172 int mlx4_bitmap_init(struct mlx4_bitmap *bitmap, u32 num, u32 mask,
173 		     u32 reserved_bot, u32 reserved_top)
174 {
175 	/* num must be a power of 2 */
176 	if (!is_power_of_2(num))
177 		return -EINVAL;
178 
179 	bitmap->last = 0;
180 	bitmap->top  = 0;
181 	bitmap->max  = num - reserved_top;
182 	bitmap->mask = mask;
183 	bitmap->reserved_top = reserved_top;
184 	bitmap->avail = num - reserved_top - reserved_bot;
185 	bitmap->effective_len = bitmap->avail;
186 	spin_lock_init(&bitmap->lock);
187 	bitmap->table = kzalloc(BITS_TO_LONGS(bitmap->max) *
188 				sizeof (long), 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 	kfree(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 int __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 	return 0;
319 }
320 
mlx4_zone_allocator_destroy(struct mlx4_zone_allocator * zone_alloc)321 void mlx4_zone_allocator_destroy(struct mlx4_zone_allocator *zone_alloc)
322 {
323 	struct mlx4_zone_entry *zone, *tmp;
324 
325 	spin_lock(&zone_alloc->lock);
326 
327 	list_for_each_entry_safe(zone, tmp, &zone_alloc->entries, list) {
328 		list_del(&zone->list);
329 		list_del(&zone->prio_list);
330 		kfree(zone);
331 	}
332 
333 	spin_unlock(&zone_alloc->lock);
334 	kfree(zone_alloc);
335 }
336 
337 /* Should be called under a lock */
__mlx4_alloc_from_zone(struct mlx4_zone_entry * zone,int count,int align,u32 skip_mask,u32 * puid)338 static u32 __mlx4_alloc_from_zone(struct mlx4_zone_entry *zone, int count,
339 				  int align, u32 skip_mask, u32 *puid)
340 {
341 	u32 uid = 0;
342 	u32 res;
343 	struct mlx4_zone_allocator *zone_alloc = zone->allocator;
344 	struct mlx4_zone_entry *curr_node;
345 
346 	res = mlx4_bitmap_alloc_range(zone->bitmap, count,
347 				      align, skip_mask);
348 
349 	if (res != (u32)-1) {
350 		res += zone->offset;
351 		uid = zone->uid;
352 		goto out;
353 	}
354 
355 	list_for_each_entry(curr_node, &zone_alloc->prios, prio_list) {
356 		if (unlikely(curr_node->priority == zone->priority))
357 			break;
358 	}
359 
360 	if (zone->flags & MLX4_ZONE_ALLOW_ALLOC_FROM_LOWER_PRIO) {
361 		struct mlx4_zone_entry *it = curr_node;
362 
363 		list_for_each_entry_continue_reverse(it, &zone_alloc->entries, list) {
364 			res = mlx4_bitmap_alloc_range(it->bitmap, count,
365 						      align, skip_mask);
366 			if (res != (u32)-1) {
367 				res += it->offset;
368 				uid = it->uid;
369 				goto out;
370 			}
371 		}
372 	}
373 
374 	if (zone->flags & MLX4_ZONE_ALLOW_ALLOC_FROM_EQ_PRIO) {
375 		struct mlx4_zone_entry *it = curr_node;
376 
377 		list_for_each_entry_from(it, &zone_alloc->entries, list) {
378 			if (unlikely(it == zone))
379 				continue;
380 
381 			if (unlikely(it->priority != curr_node->priority))
382 				break;
383 
384 			res = mlx4_bitmap_alloc_range(it->bitmap, count,
385 						      align, skip_mask);
386 			if (res != (u32)-1) {
387 				res += it->offset;
388 				uid = it->uid;
389 				goto out;
390 			}
391 		}
392 	}
393 
394 	if (zone->flags & MLX4_ZONE_FALLBACK_TO_HIGHER_PRIO) {
395 		if (list_is_last(&curr_node->prio_list, &zone_alloc->prios))
396 			goto out;
397 
398 		curr_node = list_first_entry(&curr_node->prio_list,
399 					     typeof(*curr_node),
400 					     prio_list);
401 
402 		list_for_each_entry_from(curr_node, &zone_alloc->entries, list) {
403 			res = mlx4_bitmap_alloc_range(curr_node->bitmap, count,
404 						      align, skip_mask);
405 			if (res != (u32)-1) {
406 				res += curr_node->offset;
407 				uid = curr_node->uid;
408 				goto out;
409 			}
410 		}
411 	}
412 
413 out:
414 	if (NULL != puid && res != (u32)-1)
415 		*puid = uid;
416 	return res;
417 }
418 
419 /* Should be called under a lock */
__mlx4_free_from_zone(struct mlx4_zone_entry * zone,u32 obj,u32 count)420 static void __mlx4_free_from_zone(struct mlx4_zone_entry *zone, u32 obj,
421 				  u32 count)
422 {
423 	mlx4_bitmap_free_range(zone->bitmap, obj - zone->offset, count, zone->use_rr);
424 }
425 
426 /* Should be called under a lock */
__mlx4_find_zone_by_uid(struct mlx4_zone_allocator * zones,u32 uid)427 static struct mlx4_zone_entry *__mlx4_find_zone_by_uid(
428 		struct mlx4_zone_allocator *zones, u32 uid)
429 {
430 	struct mlx4_zone_entry *zone;
431 
432 	list_for_each_entry(zone, &zones->entries, list) {
433 		if (zone->uid == uid)
434 			return zone;
435 	}
436 
437 	return NULL;
438 }
439 
mlx4_zone_get_bitmap(struct mlx4_zone_allocator * zones,u32 uid)440 struct mlx4_bitmap *mlx4_zone_get_bitmap(struct mlx4_zone_allocator *zones, u32 uid)
441 {
442 	struct mlx4_zone_entry *zone;
443 	struct mlx4_bitmap *bitmap;
444 
445 	spin_lock(&zones->lock);
446 
447 	zone = __mlx4_find_zone_by_uid(zones, uid);
448 
449 	bitmap = zone == NULL ? NULL : zone->bitmap;
450 
451 	spin_unlock(&zones->lock);
452 
453 	return bitmap;
454 }
455 
mlx4_zone_remove_one(struct mlx4_zone_allocator * zones,u32 uid)456 int mlx4_zone_remove_one(struct mlx4_zone_allocator *zones, u32 uid)
457 {
458 	struct mlx4_zone_entry *zone;
459 	int res;
460 
461 	spin_lock(&zones->lock);
462 
463 	zone = __mlx4_find_zone_by_uid(zones, uid);
464 
465 	if (NULL == zone) {
466 		res = -1;
467 		goto out;
468 	}
469 
470 	res = __mlx4_zone_remove_one_entry(zone);
471 
472 out:
473 	spin_unlock(&zones->lock);
474 	kfree(zone);
475 
476 	return res;
477 }
478 
479 /* Should be called under a lock */
__mlx4_find_zone_by_uid_unique(struct mlx4_zone_allocator * zones,u32 obj)480 static struct mlx4_zone_entry *__mlx4_find_zone_by_uid_unique(
481 		struct mlx4_zone_allocator *zones, u32 obj)
482 {
483 	struct mlx4_zone_entry *zone, *zone_candidate = NULL;
484 	u32 dist = (u32)-1;
485 
486 	/* Search for the smallest zone that this obj could be
487 	 * allocated from. This is done in order to handle
488 	 * situations when small bitmaps are allocated from bigger
489 	 * bitmaps (and the allocated space is marked as reserved in
490 	 * the bigger bitmap.
491 	 */
492 	list_for_each_entry(zone, &zones->entries, list) {
493 		if (obj >= zone->offset) {
494 			u32 mobj = (obj - zone->offset) & zones->mask;
495 
496 			if (mobj < zone->bitmap->max) {
497 				u32 curr_dist = zone->bitmap->effective_len;
498 
499 				if (curr_dist < dist) {
500 					dist = curr_dist;
501 					zone_candidate = zone;
502 				}
503 			}
504 		}
505 	}
506 
507 	return zone_candidate;
508 }
509 
mlx4_zone_alloc_entries(struct mlx4_zone_allocator * zones,u32 uid,int count,int align,u32 skip_mask,u32 * puid)510 u32 mlx4_zone_alloc_entries(struct mlx4_zone_allocator *zones, u32 uid, int count,
511 			    int align, u32 skip_mask, u32 *puid)
512 {
513 	struct mlx4_zone_entry *zone;
514 	int res = -1;
515 
516 	spin_lock(&zones->lock);
517 
518 	zone = __mlx4_find_zone_by_uid(zones, uid);
519 
520 	if (NULL == zone)
521 		goto out;
522 
523 	res = __mlx4_alloc_from_zone(zone, count, align, skip_mask, puid);
524 
525 out:
526 	spin_unlock(&zones->lock);
527 
528 	return res;
529 }
530 
mlx4_zone_free_entries(struct mlx4_zone_allocator * zones,u32 uid,u32 obj,u32 count)531 u32 mlx4_zone_free_entries(struct mlx4_zone_allocator *zones, u32 uid, u32 obj, u32 count)
532 {
533 	struct mlx4_zone_entry *zone;
534 	int res = 0;
535 
536 	spin_lock(&zones->lock);
537 
538 	zone = __mlx4_find_zone_by_uid(zones, uid);
539 
540 	if (NULL == zone) {
541 		res = -1;
542 		goto out;
543 	}
544 
545 	__mlx4_free_from_zone(zone, obj, count);
546 
547 out:
548 	spin_unlock(&zones->lock);
549 
550 	return res;
551 }
552 
mlx4_zone_free_entries_unique(struct mlx4_zone_allocator * zones,u32 obj,u32 count)553 u32 mlx4_zone_free_entries_unique(struct mlx4_zone_allocator *zones, u32 obj, u32 count)
554 {
555 	struct mlx4_zone_entry *zone;
556 	int res;
557 
558 	if (!(zones->flags & MLX4_ZONE_ALLOC_FLAGS_NO_OVERLAP))
559 		return -EFAULT;
560 
561 	spin_lock(&zones->lock);
562 
563 	zone = __mlx4_find_zone_by_uid_unique(zones, obj);
564 
565 	if (NULL == zone) {
566 		res = -1;
567 		goto out;
568 	}
569 
570 	__mlx4_free_from_zone(zone, obj, count);
571 	res = 0;
572 
573 out:
574 	spin_unlock(&zones->lock);
575 
576 	return res;
577 }
578 /*
579  * Handling for queue buffers -- we allocate a bunch of memory and
580  * register it in a memory region at HCA virtual address 0.  If the
581  * requested size is > max_direct, we split the allocation into
582  * multiple pages, so we don't require too much contiguous memory.
583  */
584 
mlx4_buf_alloc(struct mlx4_dev * dev,int size,int max_direct,struct mlx4_buf * buf,gfp_t gfp)585 int mlx4_buf_alloc(struct mlx4_dev *dev, int size, int max_direct,
586 		   struct mlx4_buf *buf, gfp_t gfp)
587 {
588 	dma_addr_t t;
589 
590 	if (size <= max_direct) {
591 		buf->nbufs        = 1;
592 		buf->npages       = 1;
593 		buf->page_shift   = get_order(size) + PAGE_SHIFT;
594 		buf->direct.buf   = dma_alloc_coherent(&dev->persist->pdev->dev,
595 						       size, &t, gfp);
596 		if (!buf->direct.buf)
597 			return -ENOMEM;
598 
599 		buf->direct.map = t;
600 
601 		while (t & ((1 << buf->page_shift) - 1)) {
602 			--buf->page_shift;
603 			buf->npages *= 2;
604 		}
605 
606 		memset(buf->direct.buf, 0, size);
607 	} else {
608 		int i;
609 
610 		buf->direct.buf  = NULL;
611 		buf->nbufs       = (size + PAGE_SIZE - 1) / PAGE_SIZE;
612 		buf->npages      = buf->nbufs;
613 		buf->page_shift  = PAGE_SHIFT;
614 		buf->page_list   = kcalloc(buf->nbufs, sizeof(*buf->page_list),
615 					   gfp);
616 		if (!buf->page_list)
617 			return -ENOMEM;
618 
619 		for (i = 0; i < buf->nbufs; ++i) {
620 			buf->page_list[i].buf =
621 				dma_alloc_coherent(&dev->persist->pdev->dev,
622 						   PAGE_SIZE,
623 						   &t, gfp);
624 			if (!buf->page_list[i].buf)
625 				goto err_free;
626 
627 			buf->page_list[i].map = t;
628 
629 			memset(buf->page_list[i].buf, 0, PAGE_SIZE);
630 		}
631 
632 		if (BITS_PER_LONG == 64) {
633 			struct page **pages;
634 			pages = kmalloc(sizeof *pages * buf->nbufs, gfp);
635 			if (!pages)
636 				goto err_free;
637 			for (i = 0; i < buf->nbufs; ++i)
638 				pages[i] = virt_to_page(buf->page_list[i].buf);
639 			buf->direct.buf = vmap(pages, buf->nbufs, VM_MAP, PAGE_KERNEL);
640 			kfree(pages);
641 			if (!buf->direct.buf)
642 				goto err_free;
643 		}
644 	}
645 
646 	return 0;
647 
648 err_free:
649 	mlx4_buf_free(dev, size, buf);
650 
651 	return -ENOMEM;
652 }
653 EXPORT_SYMBOL_GPL(mlx4_buf_alloc);
654 
mlx4_buf_free(struct mlx4_dev * dev,int size,struct mlx4_buf * buf)655 void mlx4_buf_free(struct mlx4_dev *dev, int size, struct mlx4_buf *buf)
656 {
657 	int i;
658 
659 	if (buf->nbufs == 1)
660 		dma_free_coherent(&dev->persist->pdev->dev, size,
661 				  buf->direct.buf,
662 				  buf->direct.map);
663 	else {
664 		if (BITS_PER_LONG == 64)
665 			vunmap(buf->direct.buf);
666 
667 		for (i = 0; i < buf->nbufs; ++i)
668 			if (buf->page_list[i].buf)
669 				dma_free_coherent(&dev->persist->pdev->dev,
670 						  PAGE_SIZE,
671 						  buf->page_list[i].buf,
672 						  buf->page_list[i].map);
673 		kfree(buf->page_list);
674 	}
675 }
676 EXPORT_SYMBOL_GPL(mlx4_buf_free);
677 
mlx4_alloc_db_pgdir(struct device * dma_device,gfp_t gfp)678 static struct mlx4_db_pgdir *mlx4_alloc_db_pgdir(struct device *dma_device,
679 						 gfp_t gfp)
680 {
681 	struct mlx4_db_pgdir *pgdir;
682 
683 	pgdir = kzalloc(sizeof *pgdir, gfp);
684 	if (!pgdir)
685 		return NULL;
686 
687 	bitmap_fill(pgdir->order1, MLX4_DB_PER_PAGE / 2);
688 	pgdir->bits[0] = pgdir->order0;
689 	pgdir->bits[1] = pgdir->order1;
690 	pgdir->db_page = dma_alloc_coherent(dma_device, PAGE_SIZE,
691 					    &pgdir->db_dma, gfp);
692 	if (!pgdir->db_page) {
693 		kfree(pgdir);
694 		return NULL;
695 	}
696 
697 	return pgdir;
698 }
699 
mlx4_alloc_db_from_pgdir(struct mlx4_db_pgdir * pgdir,struct mlx4_db * db,int order)700 static int mlx4_alloc_db_from_pgdir(struct mlx4_db_pgdir *pgdir,
701 				    struct mlx4_db *db, int order)
702 {
703 	int o;
704 	int i;
705 
706 	for (o = order; o <= 1; ++o) {
707 		i = find_first_bit(pgdir->bits[o], MLX4_DB_PER_PAGE >> o);
708 		if (i < MLX4_DB_PER_PAGE >> o)
709 			goto found;
710 	}
711 
712 	return -ENOMEM;
713 
714 found:
715 	clear_bit(i, pgdir->bits[o]);
716 
717 	i <<= o;
718 
719 	if (o > order)
720 		set_bit(i ^ 1, pgdir->bits[order]);
721 
722 	db->u.pgdir = pgdir;
723 	db->index   = i;
724 	db->db      = pgdir->db_page + db->index;
725 	db->dma     = pgdir->db_dma  + db->index * 4;
726 	db->order   = order;
727 
728 	return 0;
729 }
730 
mlx4_db_alloc(struct mlx4_dev * dev,struct mlx4_db * db,int order,gfp_t gfp)731 int mlx4_db_alloc(struct mlx4_dev *dev, struct mlx4_db *db, int order, gfp_t gfp)
732 {
733 	struct mlx4_priv *priv = mlx4_priv(dev);
734 	struct mlx4_db_pgdir *pgdir;
735 	int ret = 0;
736 
737 	mutex_lock(&priv->pgdir_mutex);
738 
739 	list_for_each_entry(pgdir, &priv->pgdir_list, list)
740 		if (!mlx4_alloc_db_from_pgdir(pgdir, db, order))
741 			goto out;
742 
743 	pgdir = mlx4_alloc_db_pgdir(&dev->persist->pdev->dev, gfp);
744 	if (!pgdir) {
745 		ret = -ENOMEM;
746 		goto out;
747 	}
748 
749 	list_add(&pgdir->list, &priv->pgdir_list);
750 
751 	/* This should never fail -- we just allocated an empty page: */
752 	WARN_ON(mlx4_alloc_db_from_pgdir(pgdir, db, order));
753 
754 out:
755 	mutex_unlock(&priv->pgdir_mutex);
756 
757 	return ret;
758 }
759 EXPORT_SYMBOL_GPL(mlx4_db_alloc);
760 
mlx4_db_free(struct mlx4_dev * dev,struct mlx4_db * db)761 void mlx4_db_free(struct mlx4_dev *dev, struct mlx4_db *db)
762 {
763 	struct mlx4_priv *priv = mlx4_priv(dev);
764 	int o;
765 	int i;
766 
767 	mutex_lock(&priv->pgdir_mutex);
768 
769 	o = db->order;
770 	i = db->index;
771 
772 	if (db->order == 0 && test_bit(i ^ 1, db->u.pgdir->order0)) {
773 		clear_bit(i ^ 1, db->u.pgdir->order0);
774 		++o;
775 	}
776 	i >>= o;
777 	set_bit(i, db->u.pgdir->bits[o]);
778 
779 	if (bitmap_full(db->u.pgdir->order1, MLX4_DB_PER_PAGE / 2)) {
780 		dma_free_coherent(&dev->persist->pdev->dev, PAGE_SIZE,
781 				  db->u.pgdir->db_page, db->u.pgdir->db_dma);
782 		list_del(&db->u.pgdir->list);
783 		kfree(db->u.pgdir);
784 	}
785 
786 	mutex_unlock(&priv->pgdir_mutex);
787 }
788 EXPORT_SYMBOL_GPL(mlx4_db_free);
789 
mlx4_alloc_hwq_res(struct mlx4_dev * dev,struct mlx4_hwq_resources * wqres,int size,int max_direct)790 int mlx4_alloc_hwq_res(struct mlx4_dev *dev, struct mlx4_hwq_resources *wqres,
791 		       int size, int max_direct)
792 {
793 	int err;
794 
795 	err = mlx4_db_alloc(dev, &wqres->db, 1, GFP_KERNEL);
796 	if (err)
797 		return err;
798 
799 	*wqres->db.db = 0;
800 
801 	err = mlx4_buf_alloc(dev, size, max_direct, &wqres->buf, GFP_KERNEL);
802 	if (err)
803 		goto err_db;
804 
805 	err = mlx4_mtt_init(dev, wqres->buf.npages, wqres->buf.page_shift,
806 			    &wqres->mtt);
807 	if (err)
808 		goto err_buf;
809 
810 	err = mlx4_buf_write_mtt(dev, &wqres->mtt, &wqres->buf, GFP_KERNEL);
811 	if (err)
812 		goto err_mtt;
813 
814 	return 0;
815 
816 err_mtt:
817 	mlx4_mtt_cleanup(dev, &wqres->mtt);
818 err_buf:
819 	mlx4_buf_free(dev, size, &wqres->buf);
820 err_db:
821 	mlx4_db_free(dev, &wqres->db);
822 
823 	return err;
824 }
825 EXPORT_SYMBOL_GPL(mlx4_alloc_hwq_res);
826 
mlx4_free_hwq_res(struct mlx4_dev * dev,struct mlx4_hwq_resources * wqres,int size)827 void mlx4_free_hwq_res(struct mlx4_dev *dev, struct mlx4_hwq_resources *wqres,
828 		       int size)
829 {
830 	mlx4_mtt_cleanup(dev, &wqres->mtt);
831 	mlx4_buf_free(dev, size, &wqres->buf);
832 	mlx4_db_free(dev, &wqres->db);
833 }
834 EXPORT_SYMBOL_GPL(mlx4_free_hwq_res);
835