2 * Contiguous Memory Allocator
4 * Copyright (c) 2010-2011 by Samsung Electronics.
5 * Copyright IBM Corporation, 2013
6 * Copyright LG Electronics Inc., 2014
8 * Marek Szyprowski <m.szyprowski@samsung.com>
9 * Michal Nazarewicz <mina86@mina86.com>
10 * Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
11 * Joonsoo Kim <iamjoonsoo.kim@lge.com>
13 * This program is free software; you can redistribute it and/or
14 * modify it under the terms of the GNU General Public License as
15 * published by the Free Software Foundation; either version 2 of the
16 * License or (at your optional) any later version of the license.
19 #define pr_fmt(fmt) "cma: " fmt
21 #ifdef CONFIG_CMA_DEBUG
26 #define CREATE_TRACE_POINTS
28 #include <linux/memblock.h>
29 #include <linux/err.h>
31 #include <linux/mutex.h>
32 #include <linux/sizes.h>
33 #include <linux/slab.h>
34 #include <linux/log2.h>
35 #include <linux/cma.h>
36 #include <linux/highmem.h>
38 #include <linux/kmemleak.h>
39 #include <trace/events/cma.h>
44 struct cma cma_areas[MAX_CMA_AREAS];
45 unsigned cma_area_count;
46 static DEFINE_MUTEX(cma_mutex);
48 phys_addr_t cma_get_base(const struct cma *cma)
50 return PFN_PHYS(cma->base_pfn);
53 unsigned long cma_get_size(const struct cma *cma)
55 return cma->count << PAGE_SHIFT;
58 const char *cma_get_name(const struct cma *cma)
60 return cma->name ? cma->name : "(undefined)";
63 static unsigned long cma_bitmap_aligned_mask(const struct cma *cma,
64 unsigned int align_order)
66 if (align_order <= cma->order_per_bit)
68 return (1UL << (align_order - cma->order_per_bit)) - 1;
72 * Find the offset of the base PFN from the specified align_order.
73 * The value returned is represented in order_per_bits.
75 static unsigned long cma_bitmap_aligned_offset(const struct cma *cma,
76 unsigned int align_order)
78 return (cma->base_pfn & ((1UL << align_order) - 1))
79 >> cma->order_per_bit;
82 static unsigned long cma_bitmap_pages_to_bits(const struct cma *cma,
85 return ALIGN(pages, 1UL << cma->order_per_bit) >> cma->order_per_bit;
88 static void cma_clear_bitmap(struct cma *cma, unsigned long pfn,
91 unsigned long bitmap_no, bitmap_count;
93 bitmap_no = (pfn - cma->base_pfn) >> cma->order_per_bit;
94 bitmap_count = cma_bitmap_pages_to_bits(cma, count);
96 mutex_lock(&cma->lock);
97 bitmap_clear(cma->bitmap, bitmap_no, bitmap_count);
98 mutex_unlock(&cma->lock);
101 static int __init cma_activate_area(struct cma *cma)
103 int bitmap_size = BITS_TO_LONGS(cma_bitmap_maxno(cma)) * sizeof(long);
104 unsigned long base_pfn = cma->base_pfn, pfn = base_pfn;
105 unsigned i = cma->count >> pageblock_order;
108 cma->bitmap = kzalloc(bitmap_size, GFP_KERNEL);
117 if (!pfn_valid(base_pfn))
120 zone = page_zone(pfn_to_page(base_pfn));
121 for (j = pageblock_nr_pages; j; --j, pfn++) {
126 * In init_cma_reserved_pageblock(), present_pages
127 * is adjusted with assumption that all pages in
128 * the pageblock come from a single zone.
130 if (page_zone(pfn_to_page(pfn)) != zone)
133 init_cma_reserved_pageblock(pfn_to_page(base_pfn));
136 mutex_init(&cma->lock);
138 #ifdef CONFIG_CMA_DEBUGFS
139 INIT_HLIST_HEAD(&cma->mem_head);
140 spin_lock_init(&cma->mem_head_lock);
146 pr_err("CMA area %s could not be activated\n", cma->name);
152 static int __init cma_init_reserved_areas(void)
161 for_each_online_pgdat(pgdat) {
162 unsigned long start_pfn = UINT_MAX, end_pfn = 0;
164 zone = &pgdat->node_zones[ZONE_MOVABLE];
167 * In this case, we cannot adjust the zone range
168 * since it is now maximum node span and we don't
169 * know original zone range.
171 if (populated_zone(zone))
174 for (i = 0; i < cma_area_count; i++) {
175 if (pfn_to_nid(cma_areas[i].base_pfn) !=
179 start_pfn = min(start_pfn, cma_areas[i].base_pfn);
180 end_pfn = max(end_pfn, cma_areas[i].base_pfn +
187 zone->zone_start_pfn = start_pfn;
188 zone->spanned_pages = end_pfn - start_pfn;
191 for (i = 0; i < cma_area_count; i++) {
192 int ret = cma_activate_area(&cma_areas[i]);
199 * Reserved pages for ZONE_MOVABLE are now activated and
200 * this would change ZONE_MOVABLE's managed page counter and
201 * the other zones' present counter. We need to re-calculate
202 * various zone information that depends on this initialization.
204 build_all_zonelists(NULL);
205 for_each_populated_zone(zone) {
206 if (zone_idx(zone) == ZONE_MOVABLE) {
207 zone_pcp_reset(zone);
208 setup_zone_pageset(zone);
210 zone_pcp_update(zone);
212 set_zone_contiguous(zone);
216 * We need to re-init per zone wmark by calling
217 * init_per_zone_wmark_min() but doesn't call here because it is
218 * registered on core_initcall and it will be called later than us.
223 pure_initcall(cma_init_reserved_areas);
226 * cma_init_reserved_mem() - create custom contiguous area from reserved memory
227 * @base: Base address of the reserved area
228 * @size: Size of the reserved area (in bytes),
229 * @order_per_bit: Order of pages represented by one bit on bitmap.
230 * @name: The name of the area. If this parameter is NULL, the name of
231 * the area will be set to "cmaN", where N is a running counter of
233 * @res_cma: Pointer to store the created cma region.
235 * This function creates custom contiguous area from already reserved memory.
237 int __init cma_init_reserved_mem(phys_addr_t base, phys_addr_t size,
238 unsigned int order_per_bit,
240 struct cma **res_cma)
243 phys_addr_t alignment;
246 if (cma_area_count == ARRAY_SIZE(cma_areas)) {
247 pr_err("Not enough slots for CMA reserved regions!\n");
251 if (!size || !memblock_is_region_reserved(base, size))
254 /* ensure minimal alignment required by mm core */
255 alignment = PAGE_SIZE <<
256 max_t(unsigned long, MAX_ORDER - 1, pageblock_order);
258 /* alignment should be aligned with order_per_bit */
259 if (!IS_ALIGNED(alignment >> PAGE_SHIFT, 1 << order_per_bit))
262 if (ALIGN(base, alignment) != base || ALIGN(size, alignment) != size)
266 * Each reserved area must be initialised later, when more kernel
267 * subsystems (like slab allocator) are available.
269 cma = &cma_areas[cma_area_count];
273 cma->name = kasprintf(GFP_KERNEL, "cma%d\n", cma_area_count);
277 cma->base_pfn = PFN_DOWN(base);
278 cma->count = size >> PAGE_SHIFT;
279 cma->order_per_bit = order_per_bit;
282 totalcma_pages += (size / PAGE_SIZE);
288 * cma_declare_contiguous() - reserve custom contiguous area
289 * @base: Base address of the reserved area optional, use 0 for any
290 * @size: Size of the reserved area (in bytes),
291 * @limit: End address of the reserved memory (optional, 0 for any).
292 * @alignment: Alignment for the CMA area, should be power of 2 or zero
293 * @order_per_bit: Order of pages represented by one bit on bitmap.
294 * @fixed: hint about where to place the reserved area
295 * @name: The name of the area. See function cma_init_reserved_mem()
296 * @res_cma: Pointer to store the created cma region.
298 * This function reserves memory from early allocator. It should be
299 * called by arch specific code once the early allocator (memblock or bootmem)
300 * has been activated and all other subsystems have already allocated/reserved
301 * memory. This function allows to create custom reserved areas.
303 * If @fixed is true, reserve contiguous area at exactly @base. If false,
304 * reserve in range from @base to @limit.
306 int __init cma_declare_contiguous(phys_addr_t base,
307 phys_addr_t size, phys_addr_t limit,
308 phys_addr_t alignment, unsigned int order_per_bit,
309 bool fixed, const char *name, struct cma **res_cma)
311 phys_addr_t memblock_end = memblock_end_of_DRAM();
312 phys_addr_t highmem_start;
316 * We can't use __pa(high_memory) directly, since high_memory
317 * isn't a valid direct map VA, and DEBUG_VIRTUAL will (validly)
318 * complain. Find the boundary by adding one to the last valid
321 highmem_start = __pa(high_memory - 1) + 1;
322 pr_debug("%s(size %pa, base %pa, limit %pa alignment %pa)\n",
323 __func__, &size, &base, &limit, &alignment);
325 if (cma_area_count == ARRAY_SIZE(cma_areas)) {
326 pr_err("Not enough slots for CMA reserved regions!\n");
333 if (alignment && !is_power_of_2(alignment))
337 * Sanitise input arguments.
338 * Pages both ends in CMA area could be merged into adjacent unmovable
339 * migratetype page by page allocator's buddy algorithm. In the case,
340 * you couldn't get a contiguous memory, which is not what we want.
342 alignment = max(alignment, (phys_addr_t)PAGE_SIZE <<
343 max_t(unsigned long, MAX_ORDER - 1, pageblock_order));
344 base = ALIGN(base, alignment);
345 size = ALIGN(size, alignment);
346 limit &= ~(alignment - 1);
351 /* size should be aligned with order_per_bit */
352 if (!IS_ALIGNED(size >> PAGE_SHIFT, 1 << order_per_bit))
356 * If allocating at a fixed base the request region must not cross the
357 * low/high memory boundary.
359 if (fixed && base < highmem_start && base + size > highmem_start) {
361 pr_err("Region at %pa defined on low/high memory boundary (%pa)\n",
362 &base, &highmem_start);
367 * If the limit is unspecified or above the memblock end, its effective
368 * value will be the memblock end. Set it explicitly to simplify further
371 if (limit == 0 || limit > memblock_end)
372 limit = memblock_end;
376 if (memblock_is_region_reserved(base, size) ||
377 memblock_reserve(base, size) < 0) {
382 phys_addr_t addr = 0;
385 * All pages in the reserved area must come from the same zone.
386 * If the requested region crosses the low/high memory boundary,
387 * try allocating from high memory first and fall back to low
388 * memory in case of failure.
390 if (base < highmem_start && limit > highmem_start) {
391 addr = memblock_alloc_range(size, alignment,
392 highmem_start, limit,
394 limit = highmem_start;
398 addr = memblock_alloc_range(size, alignment, base,
408 * kmemleak scans/reads tracked objects for pointers to other
409 * objects but this address isn't mapped and accessible
411 kmemleak_ignore_phys(addr);
415 ret = cma_init_reserved_mem(base, size, order_per_bit, name, res_cma);
419 pr_info("Reserved %ld MiB at %pa\n", (unsigned long)size / SZ_1M,
424 pr_err("Failed to reserve %ld MiB\n", (unsigned long)size / SZ_1M);
428 #ifdef CONFIG_CMA_DEBUG
429 static void cma_debug_show_areas(struct cma *cma)
431 unsigned long next_zero_bit, next_set_bit;
432 unsigned long start = 0;
433 unsigned int nr_zero, nr_total = 0;
435 mutex_lock(&cma->lock);
436 pr_info("number of available pages: ");
438 next_zero_bit = find_next_zero_bit(cma->bitmap, cma->count, start);
439 if (next_zero_bit >= cma->count)
441 next_set_bit = find_next_bit(cma->bitmap, cma->count, next_zero_bit);
442 nr_zero = next_set_bit - next_zero_bit;
443 pr_cont("%s%u@%lu", nr_total ? "+" : "", nr_zero, next_zero_bit);
445 start = next_zero_bit + nr_zero;
447 pr_cont("=> %u free of %lu total pages\n", nr_total, cma->count);
448 mutex_unlock(&cma->lock);
451 static inline void cma_debug_show_areas(struct cma *cma) { }
455 * cma_alloc() - allocate pages from contiguous area
456 * @cma: Contiguous memory region for which the allocation is performed.
457 * @count: Requested number of pages.
458 * @align: Requested alignment of pages (in PAGE_SIZE order).
459 * @gfp_mask: GFP mask to use during compaction
461 * This function allocates part of contiguous memory on specific
462 * contiguous memory area.
464 struct page *cma_alloc(struct cma *cma, size_t count, unsigned int align,
467 unsigned long mask, offset;
468 unsigned long pfn = -1;
469 unsigned long start = 0;
470 unsigned long bitmap_maxno, bitmap_no, bitmap_count;
471 struct page *page = NULL;
474 if (!cma || !cma->count)
477 pr_debug("%s(cma %p, count %zu, align %d)\n", __func__, (void *)cma,
483 mask = cma_bitmap_aligned_mask(cma, align);
484 offset = cma_bitmap_aligned_offset(cma, align);
485 bitmap_maxno = cma_bitmap_maxno(cma);
486 bitmap_count = cma_bitmap_pages_to_bits(cma, count);
488 if (bitmap_count > bitmap_maxno)
492 mutex_lock(&cma->lock);
493 bitmap_no = bitmap_find_next_zero_area_off(cma->bitmap,
494 bitmap_maxno, start, bitmap_count, mask,
496 if (bitmap_no >= bitmap_maxno) {
497 mutex_unlock(&cma->lock);
500 bitmap_set(cma->bitmap, bitmap_no, bitmap_count);
502 * It's safe to drop the lock here. We've marked this region for
503 * our exclusive use. If the migration fails we will take the
504 * lock again and unmark it.
506 mutex_unlock(&cma->lock);
508 pfn = cma->base_pfn + (bitmap_no << cma->order_per_bit);
509 mutex_lock(&cma_mutex);
510 ret = alloc_contig_range(pfn, pfn + count, MIGRATE_CMA,
512 mutex_unlock(&cma_mutex);
514 page = pfn_to_page(pfn);
518 cma_clear_bitmap(cma, pfn, count);
522 pr_debug("%s(): memory range at %p is busy, retrying\n",
523 __func__, pfn_to_page(pfn));
524 /* try again with a bit different memory target */
525 start = bitmap_no + mask + 1;
528 trace_cma_alloc(pfn, page, count, align);
530 if (ret && !(gfp_mask & __GFP_NOWARN)) {
531 pr_err("%s: alloc failed, req-size: %zu pages, ret: %d\n",
532 __func__, count, ret);
533 cma_debug_show_areas(cma);
536 pr_debug("%s(): returned %p\n", __func__, page);
541 * cma_release() - release allocated pages
542 * @cma: Contiguous memory region for which the allocation is performed.
543 * @pages: Allocated pages.
544 * @count: Number of allocated pages.
546 * This function releases memory allocated by alloc_cma().
547 * It returns false when provided pages do not belong to contiguous area and
550 bool cma_release(struct cma *cma, const struct page *pages, unsigned int count)
557 pr_debug("%s(page %p)\n", __func__, (void *)pages);
559 pfn = page_to_pfn(pages);
561 if (pfn < cma->base_pfn || pfn >= cma->base_pfn + cma->count)
564 VM_BUG_ON(pfn + count > cma->base_pfn + cma->count);
566 free_contig_range(pfn, count);
567 cma_clear_bitmap(cma, pfn, count);
568 trace_cma_release(pfn, pages, count);
573 int cma_for_each_area(int (*it)(struct cma *cma, void *data), void *data)
577 for (i = 0; i < cma_area_count; i++) {
578 int ret = it(&cma_areas[i], data);