2 * Copyright (c) 2006, 2007 Cisco Systems, Inc. All rights reserved.
3 * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
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:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
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.
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
34 #include <linux/errno.h>
35 #include <linux/slab.h>
37 #include <linux/export.h>
38 #include <linux/bitmap.h>
39 #include <linux/dma-mapping.h>
40 #include <linux/vmalloc.h>
44 u32 mlx4_bitmap_alloc(struct mlx4_bitmap *bitmap)
48 spin_lock(&bitmap->lock);
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)
54 obj = find_first_zero_bit(bitmap->table, bitmap->max);
57 if (obj < bitmap->max) {
58 set_bit(obj, bitmap->table);
59 bitmap->last = (obj + 1);
60 if (bitmap->last == bitmap->max)
69 spin_unlock(&bitmap->lock);
74 void mlx4_bitmap_free(struct mlx4_bitmap *bitmap, u32 obj, int use_rr)
76 mlx4_bitmap_free_range(bitmap, obj, 1, use_rr);
79 static unsigned long find_aligned_range(unsigned long *bitmap,
81 int len, int align, u32 skip_mask)
86 start = ALIGN(start, align);
88 while ((start < nbits) && (test_bit(start, bitmap) ||
99 for (i = start + 1; i < end; i++) {
100 if (test_bit(i, bitmap) || ((u32)i & skip_mask)) {
109 u32 mlx4_bitmap_alloc_range(struct mlx4_bitmap *bitmap, int cnt,
110 int align, u32 skip_mask)
114 if (likely(cnt == 1 && align == 1 && !skip_mask))
115 return mlx4_bitmap_alloc(bitmap);
117 spin_lock(&bitmap->lock);
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)
124 obj = find_aligned_range(bitmap->table, 0, bitmap->max,
125 cnt, align, skip_mask);
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)
140 bitmap->avail -= cnt;
142 spin_unlock(&bitmap->lock);
147 u32 mlx4_bitmap_avail(struct mlx4_bitmap *bitmap)
149 return bitmap->avail;
152 static u32 mlx4_bitmap_masked_value(struct mlx4_bitmap *bitmap, u32 obj)
154 return obj & (bitmap->max + bitmap->reserved_top - 1);
157 void mlx4_bitmap_free_range(struct mlx4_bitmap *bitmap, u32 obj, int cnt,
160 obj &= bitmap->max + bitmap->reserved_top - 1;
162 spin_lock(&bitmap->lock);
164 bitmap->last = min(bitmap->last, obj);
165 bitmap->top = (bitmap->top + bitmap->max + bitmap->reserved_top)
168 bitmap_clear(bitmap->table, obj, cnt);
169 bitmap->avail += cnt;
170 spin_unlock(&bitmap->lock);
173 int mlx4_bitmap_init(struct mlx4_bitmap *bitmap, u32 num, u32 mask,
174 u32 reserved_bot, u32 reserved_top)
176 /* num must be a power of 2 */
177 if (num != roundup_pow_of_two(num))
182 bitmap->max = num - reserved_top;
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);
192 bitmap_set(bitmap->table, 0, reserved_bot);
197 void mlx4_bitmap_cleanup(struct mlx4_bitmap *bitmap)
199 bitmap_free(bitmap->table);
202 struct mlx4_zone_allocator {
203 struct list_head entries;
204 struct list_head prios;
207 /* protect the zone_allocator from concurrent accesses */
209 enum mlx4_zone_alloc_flags flags;
212 struct mlx4_zone_entry {
213 struct list_head list;
214 struct list_head prio_list;
216 struct mlx4_zone_allocator *allocator;
217 struct mlx4_bitmap *bitmap;
221 enum mlx4_zone_flags flags;
224 struct mlx4_zone_allocator *mlx4_zone_allocator_create(enum mlx4_zone_alloc_flags flags)
226 struct mlx4_zone_allocator *zones = kmalloc(sizeof(*zones), GFP_KERNEL);
231 INIT_LIST_HEAD(&zones->entries);
232 INIT_LIST_HEAD(&zones->prios);
233 spin_lock_init(&zones->lock);
236 zones->flags = flags;
241 int mlx4_zone_add_one(struct mlx4_zone_allocator *zone_alloc,
242 struct mlx4_bitmap *bitmap,
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);
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;
261 spin_lock(&zone_alloc->lock);
263 zone->uid = zone_alloc->last_uid++;
264 zone->allocator = zone_alloc;
266 if (zone_alloc->mask < mask)
267 zone_alloc->mask = mask;
269 list_for_each_entry(it, &zone_alloc->prios, prio_list)
270 if (it->priority >= priority)
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);
277 spin_unlock(&zone_alloc->lock);
284 /* Should be called under a lock */
285 static void __mlx4_zone_remove_one_entry(struct mlx4_zone_entry *entry)
287 struct mlx4_zone_allocator *zone_alloc = entry->allocator;
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,
296 if (next->priority == entry->priority)
297 list_add_tail(&next->prio_list, &entry->prio_list);
300 list_del(&entry->prio_list);
303 list_del(&entry->list);
305 if (zone_alloc->flags & MLX4_ZONE_ALLOC_FLAGS_NO_OVERLAP) {
307 struct mlx4_zone_entry *it;
309 list_for_each_entry(it, &zone_alloc->prios, prio_list) {
310 u32 cur_mask = mlx4_bitmap_masked_value(it->bitmap, (u32)-1);
315 zone_alloc->mask = mask;
319 void mlx4_zone_allocator_destroy(struct mlx4_zone_allocator *zone_alloc)
321 struct mlx4_zone_entry *zone, *tmp;
323 spin_lock(&zone_alloc->lock);
325 list_for_each_entry_safe(zone, tmp, &zone_alloc->entries, list) {
326 list_del(&zone->list);
327 list_del(&zone->prio_list);
331 spin_unlock(&zone_alloc->lock);
335 /* Should be called under a lock */
336 static u32 __mlx4_alloc_from_zone(struct mlx4_zone_entry *zone, int count,
337 int align, u32 skip_mask, u32 *puid)
341 struct mlx4_zone_allocator *zone_alloc = zone->allocator;
342 struct mlx4_zone_entry *curr_node;
344 res = mlx4_bitmap_alloc_range(zone->bitmap, count,
347 if (res != (u32)-1) {
353 list_for_each_entry(curr_node, &zone_alloc->prios, prio_list) {
354 if (unlikely(curr_node->priority == zone->priority))
358 if (zone->flags & MLX4_ZONE_ALLOW_ALLOC_FROM_LOWER_PRIO) {
359 struct mlx4_zone_entry *it = curr_node;
361 list_for_each_entry_continue_reverse(it, &zone_alloc->entries, list) {
362 res = mlx4_bitmap_alloc_range(it->bitmap, count,
364 if (res != (u32)-1) {
372 if (zone->flags & MLX4_ZONE_ALLOW_ALLOC_FROM_EQ_PRIO) {
373 struct mlx4_zone_entry *it = curr_node;
375 list_for_each_entry_from(it, &zone_alloc->entries, list) {
376 if (unlikely(it == zone))
379 if (unlikely(it->priority != curr_node->priority))
382 res = mlx4_bitmap_alloc_range(it->bitmap, count,
384 if (res != (u32)-1) {
392 if (zone->flags & MLX4_ZONE_FALLBACK_TO_HIGHER_PRIO) {
393 if (list_is_last(&curr_node->prio_list, &zone_alloc->prios))
396 curr_node = list_first_entry(&curr_node->prio_list,
400 list_for_each_entry_from(curr_node, &zone_alloc->entries, list) {
401 res = mlx4_bitmap_alloc_range(curr_node->bitmap, count,
403 if (res != (u32)-1) {
404 res += curr_node->offset;
405 uid = curr_node->uid;
412 if (NULL != puid && res != (u32)-1)
417 /* Should be called under a lock */
418 static void __mlx4_free_from_zone(struct mlx4_zone_entry *zone, u32 obj,
421 mlx4_bitmap_free_range(zone->bitmap, obj - zone->offset, count, zone->use_rr);
424 /* Should be called under a lock */
425 static struct mlx4_zone_entry *__mlx4_find_zone_by_uid(
426 struct mlx4_zone_allocator *zones, u32 uid)
428 struct mlx4_zone_entry *zone;
430 list_for_each_entry(zone, &zones->entries, list) {
431 if (zone->uid == uid)
438 struct mlx4_bitmap *mlx4_zone_get_bitmap(struct mlx4_zone_allocator *zones, u32 uid)
440 struct mlx4_zone_entry *zone;
441 struct mlx4_bitmap *bitmap;
443 spin_lock(&zones->lock);
445 zone = __mlx4_find_zone_by_uid(zones, uid);
447 bitmap = zone == NULL ? NULL : zone->bitmap;
449 spin_unlock(&zones->lock);
454 int mlx4_zone_remove_one(struct mlx4_zone_allocator *zones, u32 uid)
456 struct mlx4_zone_entry *zone;
459 spin_lock(&zones->lock);
461 zone = __mlx4_find_zone_by_uid(zones, uid);
468 __mlx4_zone_remove_one_entry(zone);
471 spin_unlock(&zones->lock);
477 /* Should be called under a lock */
478 static struct mlx4_zone_entry *__mlx4_find_zone_by_uid_unique(
479 struct mlx4_zone_allocator *zones, u32 obj)
481 struct mlx4_zone_entry *zone, *zone_candidate = NULL;
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
490 list_for_each_entry(zone, &zones->entries, list) {
491 if (obj >= zone->offset) {
492 u32 mobj = (obj - zone->offset) & zones->mask;
494 if (mobj < zone->bitmap->max) {
495 u32 curr_dist = zone->bitmap->effective_len;
497 if (curr_dist < dist) {
499 zone_candidate = zone;
505 return zone_candidate;
508 u32 mlx4_zone_alloc_entries(struct mlx4_zone_allocator *zones, u32 uid, int count,
509 int align, u32 skip_mask, u32 *puid)
511 struct mlx4_zone_entry *zone;
514 spin_lock(&zones->lock);
516 zone = __mlx4_find_zone_by_uid(zones, uid);
521 res = __mlx4_alloc_from_zone(zone, count, align, skip_mask, puid);
524 spin_unlock(&zones->lock);
529 u32 mlx4_zone_free_entries(struct mlx4_zone_allocator *zones, u32 uid, u32 obj, u32 count)
531 struct mlx4_zone_entry *zone;
534 spin_lock(&zones->lock);
536 zone = __mlx4_find_zone_by_uid(zones, uid);
543 __mlx4_free_from_zone(zone, obj, count);
546 spin_unlock(&zones->lock);
551 u32 mlx4_zone_free_entries_unique(struct mlx4_zone_allocator *zones, u32 obj, u32 count)
553 struct mlx4_zone_entry *zone;
556 if (!(zones->flags & MLX4_ZONE_ALLOC_FLAGS_NO_OVERLAP))
559 spin_lock(&zones->lock);
561 zone = __mlx4_find_zone_by_uid_unique(zones, obj);
568 __mlx4_free_from_zone(zone, obj, count);
572 spin_unlock(&zones->lock);
577 static int mlx4_buf_direct_alloc(struct mlx4_dev *dev, int size,
578 struct mlx4_buf *buf)
584 buf->page_shift = get_order(size) + PAGE_SHIFT;
586 dma_alloc_coherent(&dev->persist->pdev->dev, size, &t,
588 if (!buf->direct.buf)
593 while (t & ((1 << buf->page_shift) - 1)) {
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.
606 int mlx4_buf_alloc(struct mlx4_dev *dev, int size, int max_direct,
607 struct mlx4_buf *buf)
609 if (size <= max_direct) {
610 return mlx4_buf_direct_alloc(dev, size, buf);
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),
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)
631 buf->page_list[i].map = t;
638 mlx4_buf_free(dev, size, buf);
642 EXPORT_SYMBOL_GPL(mlx4_buf_alloc);
644 void mlx4_buf_free(struct mlx4_dev *dev, int size, struct mlx4_buf *buf)
646 if (buf->nbufs == 1) {
647 dma_free_coherent(&dev->persist->pdev->dev, size,
648 buf->direct.buf, buf->direct.map);
652 for (i = 0; i < buf->nbufs; ++i)
653 if (buf->page_list[i].buf)
654 dma_free_coherent(&dev->persist->pdev->dev,
656 buf->page_list[i].buf,
657 buf->page_list[i].map);
658 kfree(buf->page_list);
661 EXPORT_SYMBOL_GPL(mlx4_buf_free);
663 static struct mlx4_db_pgdir *mlx4_alloc_db_pgdir(struct device *dma_device)
665 struct mlx4_db_pgdir *pgdir;
667 pgdir = kzalloc(sizeof(*pgdir), GFP_KERNEL);
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) {
684 static int mlx4_alloc_db_from_pgdir(struct mlx4_db_pgdir *pgdir,
685 struct mlx4_db *db, int order)
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)
699 clear_bit(i, pgdir->bits[o]);
704 set_bit(i ^ 1, pgdir->bits[order]);
708 db->db = pgdir->db_page + db->index;
709 db->dma = pgdir->db_dma + db->index * 4;
715 int mlx4_db_alloc(struct mlx4_dev *dev, struct mlx4_db *db, int order)
717 struct mlx4_priv *priv = mlx4_priv(dev);
718 struct mlx4_db_pgdir *pgdir;
721 mutex_lock(&priv->pgdir_mutex);
723 list_for_each_entry(pgdir, &priv->pgdir_list, list)
724 if (!mlx4_alloc_db_from_pgdir(pgdir, db, order))
727 pgdir = mlx4_alloc_db_pgdir(&dev->persist->pdev->dev);
733 list_add(&pgdir->list, &priv->pgdir_list);
735 /* This should never fail -- we just allocated an empty page: */
736 WARN_ON(mlx4_alloc_db_from_pgdir(pgdir, db, order));
739 mutex_unlock(&priv->pgdir_mutex);
743 EXPORT_SYMBOL_GPL(mlx4_db_alloc);
745 void mlx4_db_free(struct mlx4_dev *dev, struct mlx4_db *db)
747 struct mlx4_priv *priv = mlx4_priv(dev);
751 mutex_lock(&priv->pgdir_mutex);
756 if (db->order == 0 && test_bit(i ^ 1, db->u.pgdir->order0)) {
757 clear_bit(i ^ 1, db->u.pgdir->order0);
761 set_bit(i, db->u.pgdir->bits[o]);
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);
770 mutex_unlock(&priv->pgdir_mutex);
772 EXPORT_SYMBOL_GPL(mlx4_db_free);
774 int mlx4_alloc_hwq_res(struct mlx4_dev *dev, struct mlx4_hwq_resources *wqres,
779 err = mlx4_db_alloc(dev, &wqres->db, 1);
785 err = mlx4_buf_direct_alloc(dev, size, &wqres->buf);
789 err = mlx4_mtt_init(dev, wqres->buf.npages, wqres->buf.page_shift,
794 err = mlx4_buf_write_mtt(dev, &wqres->mtt, &wqres->buf);
801 mlx4_mtt_cleanup(dev, &wqres->mtt);
803 mlx4_buf_free(dev, size, &wqres->buf);
805 mlx4_db_free(dev, &wqres->db);
809 EXPORT_SYMBOL_GPL(mlx4_alloc_hwq_res);
811 void mlx4_free_hwq_res(struct mlx4_dev *dev, struct mlx4_hwq_resources *wqres,
814 mlx4_mtt_cleanup(dev, &wqres->mtt);
815 mlx4_buf_free(dev, size, &wqres->buf);
816 mlx4_db_free(dev, &wqres->db);
818 EXPORT_SYMBOL_GPL(mlx4_free_hwq_res);