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 = kzalloc(BITS_TO_LONGS(bitmap->max) *
189 sizeof (long), GFP_KERNEL);
193 bitmap_set(bitmap->table, 0, reserved_bot);
198 void mlx4_bitmap_cleanup(struct mlx4_bitmap *bitmap)
200 kfree(bitmap->table);
203 struct mlx4_zone_allocator {
204 struct list_head entries;
205 struct list_head prios;
208 /* protect the zone_allocator from concurrent accesses */
210 enum mlx4_zone_alloc_flags flags;
213 struct mlx4_zone_entry {
214 struct list_head list;
215 struct list_head prio_list;
217 struct mlx4_zone_allocator *allocator;
218 struct mlx4_bitmap *bitmap;
222 enum mlx4_zone_flags flags;
225 struct mlx4_zone_allocator *mlx4_zone_allocator_create(enum mlx4_zone_alloc_flags flags)
227 struct mlx4_zone_allocator *zones = kmalloc(sizeof(*zones), GFP_KERNEL);
232 INIT_LIST_HEAD(&zones->entries);
233 INIT_LIST_HEAD(&zones->prios);
234 spin_lock_init(&zones->lock);
237 zones->flags = flags;
242 int mlx4_zone_add_one(struct mlx4_zone_allocator *zone_alloc,
243 struct mlx4_bitmap *bitmap,
249 u32 mask = mlx4_bitmap_masked_value(bitmap, (u32)-1);
250 struct mlx4_zone_entry *it;
251 struct mlx4_zone_entry *zone = kmalloc(sizeof(*zone), GFP_KERNEL);
257 zone->bitmap = bitmap;
258 zone->use_rr = (flags & MLX4_ZONE_USE_RR) ? MLX4_USE_RR : 0;
259 zone->priority = priority;
260 zone->offset = offset;
262 spin_lock(&zone_alloc->lock);
264 zone->uid = zone_alloc->last_uid++;
265 zone->allocator = zone_alloc;
267 if (zone_alloc->mask < mask)
268 zone_alloc->mask = mask;
270 list_for_each_entry(it, &zone_alloc->prios, prio_list)
271 if (it->priority >= priority)
274 if (&it->prio_list == &zone_alloc->prios || it->priority > priority)
275 list_add_tail(&zone->prio_list, &it->prio_list);
276 list_add_tail(&zone->list, &it->list);
278 spin_unlock(&zone_alloc->lock);
285 /* Should be called under a lock */
286 static int __mlx4_zone_remove_one_entry(struct mlx4_zone_entry *entry)
288 struct mlx4_zone_allocator *zone_alloc = entry->allocator;
290 if (!list_empty(&entry->prio_list)) {
291 /* Check if we need to add an alternative node to the prio list */
292 if (!list_is_last(&entry->list, &zone_alloc->entries)) {
293 struct mlx4_zone_entry *next = list_first_entry(&entry->list,
297 if (next->priority == entry->priority)
298 list_add_tail(&next->prio_list, &entry->prio_list);
301 list_del(&entry->prio_list);
304 list_del(&entry->list);
306 if (zone_alloc->flags & MLX4_ZONE_ALLOC_FLAGS_NO_OVERLAP) {
308 struct mlx4_zone_entry *it;
310 list_for_each_entry(it, &zone_alloc->prios, prio_list) {
311 u32 cur_mask = mlx4_bitmap_masked_value(it->bitmap, (u32)-1);
316 zone_alloc->mask = mask;
322 void mlx4_zone_allocator_destroy(struct mlx4_zone_allocator *zone_alloc)
324 struct mlx4_zone_entry *zone, *tmp;
326 spin_lock(&zone_alloc->lock);
328 list_for_each_entry_safe(zone, tmp, &zone_alloc->entries, list) {
329 list_del(&zone->list);
330 list_del(&zone->prio_list);
334 spin_unlock(&zone_alloc->lock);
338 /* Should be called under a lock */
339 static u32 __mlx4_alloc_from_zone(struct mlx4_zone_entry *zone, int count,
340 int align, u32 skip_mask, u32 *puid)
344 struct mlx4_zone_allocator *zone_alloc = zone->allocator;
345 struct mlx4_zone_entry *curr_node;
347 res = mlx4_bitmap_alloc_range(zone->bitmap, count,
350 if (res != (u32)-1) {
356 list_for_each_entry(curr_node, &zone_alloc->prios, prio_list) {
357 if (unlikely(curr_node->priority == zone->priority))
361 if (zone->flags & MLX4_ZONE_ALLOW_ALLOC_FROM_LOWER_PRIO) {
362 struct mlx4_zone_entry *it = curr_node;
364 list_for_each_entry_continue_reverse(it, &zone_alloc->entries, list) {
365 res = mlx4_bitmap_alloc_range(it->bitmap, count,
367 if (res != (u32)-1) {
375 if (zone->flags & MLX4_ZONE_ALLOW_ALLOC_FROM_EQ_PRIO) {
376 struct mlx4_zone_entry *it = curr_node;
378 list_for_each_entry_from(it, &zone_alloc->entries, list) {
379 if (unlikely(it == zone))
382 if (unlikely(it->priority != curr_node->priority))
385 res = mlx4_bitmap_alloc_range(it->bitmap, count,
387 if (res != (u32)-1) {
395 if (zone->flags & MLX4_ZONE_FALLBACK_TO_HIGHER_PRIO) {
396 if (list_is_last(&curr_node->prio_list, &zone_alloc->prios))
399 curr_node = list_first_entry(&curr_node->prio_list,
403 list_for_each_entry_from(curr_node, &zone_alloc->entries, list) {
404 res = mlx4_bitmap_alloc_range(curr_node->bitmap, count,
406 if (res != (u32)-1) {
407 res += curr_node->offset;
408 uid = curr_node->uid;
415 if (NULL != puid && res != (u32)-1)
420 /* Should be called under a lock */
421 static void __mlx4_free_from_zone(struct mlx4_zone_entry *zone, u32 obj,
424 mlx4_bitmap_free_range(zone->bitmap, obj - zone->offset, count, zone->use_rr);
427 /* Should be called under a lock */
428 static struct mlx4_zone_entry *__mlx4_find_zone_by_uid(
429 struct mlx4_zone_allocator *zones, u32 uid)
431 struct mlx4_zone_entry *zone;
433 list_for_each_entry(zone, &zones->entries, list) {
434 if (zone->uid == uid)
441 struct mlx4_bitmap *mlx4_zone_get_bitmap(struct mlx4_zone_allocator *zones, u32 uid)
443 struct mlx4_zone_entry *zone;
444 struct mlx4_bitmap *bitmap;
446 spin_lock(&zones->lock);
448 zone = __mlx4_find_zone_by_uid(zones, uid);
450 bitmap = zone == NULL ? NULL : zone->bitmap;
452 spin_unlock(&zones->lock);
457 int mlx4_zone_remove_one(struct mlx4_zone_allocator *zones, u32 uid)
459 struct mlx4_zone_entry *zone;
462 spin_lock(&zones->lock);
464 zone = __mlx4_find_zone_by_uid(zones, uid);
471 res = __mlx4_zone_remove_one_entry(zone);
474 spin_unlock(&zones->lock);
480 /* Should be called under a lock */
481 static struct mlx4_zone_entry *__mlx4_find_zone_by_uid_unique(
482 struct mlx4_zone_allocator *zones, u32 obj)
484 struct mlx4_zone_entry *zone, *zone_candidate = NULL;
487 /* Search for the smallest zone that this obj could be
488 * allocated from. This is done in order to handle
489 * situations when small bitmaps are allocated from bigger
490 * bitmaps (and the allocated space is marked as reserved in
493 list_for_each_entry(zone, &zones->entries, list) {
494 if (obj >= zone->offset) {
495 u32 mobj = (obj - zone->offset) & zones->mask;
497 if (mobj < zone->bitmap->max) {
498 u32 curr_dist = zone->bitmap->effective_len;
500 if (curr_dist < dist) {
502 zone_candidate = zone;
508 return zone_candidate;
511 u32 mlx4_zone_alloc_entries(struct mlx4_zone_allocator *zones, u32 uid, int count,
512 int align, u32 skip_mask, u32 *puid)
514 struct mlx4_zone_entry *zone;
517 spin_lock(&zones->lock);
519 zone = __mlx4_find_zone_by_uid(zones, uid);
524 res = __mlx4_alloc_from_zone(zone, count, align, skip_mask, puid);
527 spin_unlock(&zones->lock);
532 u32 mlx4_zone_free_entries(struct mlx4_zone_allocator *zones, u32 uid, u32 obj, u32 count)
534 struct mlx4_zone_entry *zone;
537 spin_lock(&zones->lock);
539 zone = __mlx4_find_zone_by_uid(zones, uid);
546 __mlx4_free_from_zone(zone, obj, count);
549 spin_unlock(&zones->lock);
554 u32 mlx4_zone_free_entries_unique(struct mlx4_zone_allocator *zones, u32 obj, u32 count)
556 struct mlx4_zone_entry *zone;
559 if (!(zones->flags & MLX4_ZONE_ALLOC_FLAGS_NO_OVERLAP))
562 spin_lock(&zones->lock);
564 zone = __mlx4_find_zone_by_uid_unique(zones, obj);
571 __mlx4_free_from_zone(zone, obj, count);
575 spin_unlock(&zones->lock);
580 * Handling for queue buffers -- we allocate a bunch of memory and
581 * register it in a memory region at HCA virtual address 0. If the
582 * requested size is > max_direct, we split the allocation into
583 * multiple pages, so we don't require too much contiguous memory.
586 int mlx4_buf_alloc(struct mlx4_dev *dev, int size, int max_direct,
587 struct mlx4_buf *buf, gfp_t gfp)
591 if (size <= max_direct) {
594 buf->page_shift = get_order(size) + PAGE_SHIFT;
595 buf->direct.buf = dma_alloc_coherent(&dev->persist->pdev->dev,
597 if (!buf->direct.buf)
602 while (t & ((1 << buf->page_shift) - 1)) {
607 memset(buf->direct.buf, 0, size);
611 buf->direct.buf = NULL;
612 buf->nbufs = (size + PAGE_SIZE - 1) / PAGE_SIZE;
613 buf->npages = buf->nbufs;
614 buf->page_shift = PAGE_SHIFT;
615 buf->page_list = kcalloc(buf->nbufs, sizeof(*buf->page_list),
620 for (i = 0; i < buf->nbufs; ++i) {
621 buf->page_list[i].buf =
622 dma_alloc_coherent(&dev->persist->pdev->dev,
625 if (!buf->page_list[i].buf)
628 buf->page_list[i].map = t;
630 memset(buf->page_list[i].buf, 0, PAGE_SIZE);
633 if (BITS_PER_LONG == 64) {
635 pages = kmalloc(sizeof *pages * buf->nbufs, gfp);
638 for (i = 0; i < buf->nbufs; ++i)
639 pages[i] = virt_to_page(buf->page_list[i].buf);
640 buf->direct.buf = vmap(pages, buf->nbufs, VM_MAP, PAGE_KERNEL);
642 if (!buf->direct.buf)
650 mlx4_buf_free(dev, size, buf);
654 EXPORT_SYMBOL_GPL(mlx4_buf_alloc);
656 void mlx4_buf_free(struct mlx4_dev *dev, int size, struct mlx4_buf *buf)
661 dma_free_coherent(&dev->persist->pdev->dev, size,
665 if (BITS_PER_LONG == 64)
666 vunmap(buf->direct.buf);
668 for (i = 0; i < buf->nbufs; ++i)
669 if (buf->page_list[i].buf)
670 dma_free_coherent(&dev->persist->pdev->dev,
672 buf->page_list[i].buf,
673 buf->page_list[i].map);
674 kfree(buf->page_list);
677 EXPORT_SYMBOL_GPL(mlx4_buf_free);
679 static struct mlx4_db_pgdir *mlx4_alloc_db_pgdir(struct device *dma_device,
682 struct mlx4_db_pgdir *pgdir;
684 pgdir = kzalloc(sizeof *pgdir, gfp);
688 bitmap_fill(pgdir->order1, MLX4_DB_PER_PAGE / 2);
689 pgdir->bits[0] = pgdir->order0;
690 pgdir->bits[1] = pgdir->order1;
691 pgdir->db_page = dma_alloc_coherent(dma_device, PAGE_SIZE,
692 &pgdir->db_dma, gfp);
693 if (!pgdir->db_page) {
701 static int mlx4_alloc_db_from_pgdir(struct mlx4_db_pgdir *pgdir,
702 struct mlx4_db *db, int order)
707 for (o = order; o <= 1; ++o) {
708 i = find_first_bit(pgdir->bits[o], MLX4_DB_PER_PAGE >> o);
709 if (i < MLX4_DB_PER_PAGE >> o)
716 clear_bit(i, pgdir->bits[o]);
721 set_bit(i ^ 1, pgdir->bits[order]);
725 db->db = pgdir->db_page + db->index;
726 db->dma = pgdir->db_dma + db->index * 4;
732 int mlx4_db_alloc(struct mlx4_dev *dev, struct mlx4_db *db, int order, gfp_t gfp)
734 struct mlx4_priv *priv = mlx4_priv(dev);
735 struct mlx4_db_pgdir *pgdir;
738 mutex_lock(&priv->pgdir_mutex);
740 list_for_each_entry(pgdir, &priv->pgdir_list, list)
741 if (!mlx4_alloc_db_from_pgdir(pgdir, db, order))
744 pgdir = mlx4_alloc_db_pgdir(&dev->persist->pdev->dev, gfp);
750 list_add(&pgdir->list, &priv->pgdir_list);
752 /* This should never fail -- we just allocated an empty page: */
753 WARN_ON(mlx4_alloc_db_from_pgdir(pgdir, db, order));
756 mutex_unlock(&priv->pgdir_mutex);
760 EXPORT_SYMBOL_GPL(mlx4_db_alloc);
762 void mlx4_db_free(struct mlx4_dev *dev, struct mlx4_db *db)
764 struct mlx4_priv *priv = mlx4_priv(dev);
768 mutex_lock(&priv->pgdir_mutex);
773 if (db->order == 0 && test_bit(i ^ 1, db->u.pgdir->order0)) {
774 clear_bit(i ^ 1, db->u.pgdir->order0);
778 set_bit(i, db->u.pgdir->bits[o]);
780 if (bitmap_full(db->u.pgdir->order1, MLX4_DB_PER_PAGE / 2)) {
781 dma_free_coherent(&dev->persist->pdev->dev, PAGE_SIZE,
782 db->u.pgdir->db_page, db->u.pgdir->db_dma);
783 list_del(&db->u.pgdir->list);
787 mutex_unlock(&priv->pgdir_mutex);
789 EXPORT_SYMBOL_GPL(mlx4_db_free);
791 int mlx4_alloc_hwq_res(struct mlx4_dev *dev, struct mlx4_hwq_resources *wqres,
792 int size, int max_direct)
796 err = mlx4_db_alloc(dev, &wqres->db, 1, GFP_KERNEL);
802 err = mlx4_buf_alloc(dev, size, max_direct, &wqres->buf, GFP_KERNEL);
806 err = mlx4_mtt_init(dev, wqres->buf.npages, wqres->buf.page_shift,
811 err = mlx4_buf_write_mtt(dev, &wqres->mtt, &wqres->buf, GFP_KERNEL);
818 mlx4_mtt_cleanup(dev, &wqres->mtt);
820 mlx4_buf_free(dev, size, &wqres->buf);
822 mlx4_db_free(dev, &wqres->db);
826 EXPORT_SYMBOL_GPL(mlx4_alloc_hwq_res);
828 void mlx4_free_hwq_res(struct mlx4_dev *dev, struct mlx4_hwq_resources *wqres,
831 mlx4_mtt_cleanup(dev, &wqres->mtt);
832 mlx4_buf_free(dev, size, &wqres->buf);
833 mlx4_db_free(dev, &wqres->db);
835 EXPORT_SYMBOL_GPL(mlx4_free_hwq_res);