1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Contiguous Memory Allocator
5 * Copyright (c) 2010-2011 by Samsung Electronics.
6 * Copyright IBM Corporation, 2013
7 * Copyright LG Electronics Inc., 2014
9 * Marek Szyprowski <m.szyprowski@samsung.com>
10 * Michal Nazarewicz <mina86@mina86.com>
11 * Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
12 * Joonsoo Kim <iamjoonsoo.kim@lge.com>
15 #define pr_fmt(fmt) "cma: " fmt
17 #ifdef CONFIG_CMA_DEBUG
22 #define CREATE_TRACE_POINTS
24 #include <linux/memblock.h>
25 #include <linux/err.h>
27 #include <linux/mutex.h>
28 #include <linux/sizes.h>
29 #include <linux/slab.h>
30 #include <linux/log2.h>
31 #include <linux/cma.h>
32 #include <linux/highmem.h>
34 #include <linux/kmemleak.h>
35 #include <trace/events/cma.h>
39 struct cma cma_areas[MAX_CMA_AREAS];
40 unsigned cma_area_count;
41 static DEFINE_MUTEX(cma_mutex);
43 phys_addr_t cma_get_base(const struct cma *cma)
45 return PFN_PHYS(cma->base_pfn);
48 unsigned long cma_get_size(const struct cma *cma)
50 return cma->count << PAGE_SHIFT;
53 const char *cma_get_name(const struct cma *cma)
58 static unsigned long cma_bitmap_aligned_mask(const struct cma *cma,
59 unsigned int align_order)
61 if (align_order <= cma->order_per_bit)
63 return (1UL << (align_order - cma->order_per_bit)) - 1;
67 * Find the offset of the base PFN from the specified align_order.
68 * The value returned is represented in order_per_bits.
70 static unsigned long cma_bitmap_aligned_offset(const struct cma *cma,
71 unsigned int align_order)
73 return (cma->base_pfn & ((1UL << align_order) - 1))
74 >> cma->order_per_bit;
77 static unsigned long cma_bitmap_pages_to_bits(const struct cma *cma,
80 return ALIGN(pages, 1UL << cma->order_per_bit) >> cma->order_per_bit;
83 static void cma_clear_bitmap(struct cma *cma, unsigned long pfn,
86 unsigned long bitmap_no, bitmap_count;
88 bitmap_no = (pfn - cma->base_pfn) >> cma->order_per_bit;
89 bitmap_count = cma_bitmap_pages_to_bits(cma, count);
91 mutex_lock(&cma->lock);
92 bitmap_clear(cma->bitmap, bitmap_no, bitmap_count);
93 mutex_unlock(&cma->lock);
96 static int __init cma_activate_area(struct cma *cma)
98 unsigned long base_pfn = cma->base_pfn, pfn = base_pfn;
99 unsigned i = cma->count >> pageblock_order;
102 cma->bitmap = bitmap_zalloc(cma_bitmap_maxno(cma), GFP_KERNEL);
108 WARN_ON_ONCE(!pfn_valid(pfn));
109 zone = page_zone(pfn_to_page(pfn));
115 for (j = pageblock_nr_pages; j; --j, pfn++) {
116 WARN_ON_ONCE(!pfn_valid(pfn));
118 * alloc_contig_range requires the pfn range
119 * specified to be in the same zone. Make this
120 * simple by forcing the entire CMA resv range
121 * to be in the same zone.
123 if (page_zone(pfn_to_page(pfn)) != zone)
126 init_cma_reserved_pageblock(pfn_to_page(base_pfn));
129 mutex_init(&cma->lock);
131 #ifdef CONFIG_CMA_DEBUGFS
132 INIT_HLIST_HEAD(&cma->mem_head);
133 spin_lock_init(&cma->mem_head_lock);
139 pr_err("CMA area %s could not be activated\n", cma->name);
140 bitmap_free(cma->bitmap);
145 static int __init cma_init_reserved_areas(void)
149 for (i = 0; i < cma_area_count; i++) {
150 int ret = cma_activate_area(&cma_areas[i]);
158 core_initcall(cma_init_reserved_areas);
161 * cma_init_reserved_mem() - create custom contiguous area from reserved memory
162 * @base: Base address of the reserved area
163 * @size: Size of the reserved area (in bytes),
164 * @order_per_bit: Order of pages represented by one bit on bitmap.
165 * @name: The name of the area. If this parameter is NULL, the name of
166 * the area will be set to "cmaN", where N is a running counter of
168 * @res_cma: Pointer to store the created cma region.
170 * This function creates custom contiguous area from already reserved memory.
172 int __init cma_init_reserved_mem(phys_addr_t base, phys_addr_t size,
173 unsigned int order_per_bit,
175 struct cma **res_cma)
178 phys_addr_t alignment;
181 if (cma_area_count == ARRAY_SIZE(cma_areas)) {
182 pr_err("Not enough slots for CMA reserved regions!\n");
186 if (!size || !memblock_is_region_reserved(base, size))
189 /* ensure minimal alignment required by mm core */
190 alignment = PAGE_SIZE <<
191 max_t(unsigned long, MAX_ORDER - 1, pageblock_order);
193 /* alignment should be aligned with order_per_bit */
194 if (!IS_ALIGNED(alignment >> PAGE_SHIFT, 1 << order_per_bit))
197 if (ALIGN(base, alignment) != base || ALIGN(size, alignment) != size)
201 * Each reserved area must be initialised later, when more kernel
202 * subsystems (like slab allocator) are available.
204 cma = &cma_areas[cma_area_count];
207 snprintf(cma->name, CMA_MAX_NAME, name);
209 snprintf(cma->name, CMA_MAX_NAME, "cma%d\n", cma_area_count);
211 cma->base_pfn = PFN_DOWN(base);
212 cma->count = size >> PAGE_SHIFT;
213 cma->order_per_bit = order_per_bit;
216 totalcma_pages += (size / PAGE_SIZE);
222 * cma_declare_contiguous_nid() - reserve custom contiguous area
223 * @base: Base address of the reserved area optional, use 0 for any
224 * @size: Size of the reserved area (in bytes),
225 * @limit: End address of the reserved memory (optional, 0 for any).
226 * @alignment: Alignment for the CMA area, should be power of 2 or zero
227 * @order_per_bit: Order of pages represented by one bit on bitmap.
228 * @fixed: hint about where to place the reserved area
229 * @name: The name of the area. See function cma_init_reserved_mem()
230 * @res_cma: Pointer to store the created cma region.
231 * @nid: nid of the free area to find, %NUMA_NO_NODE for any node
233 * This function reserves memory from early allocator. It should be
234 * called by arch specific code once the early allocator (memblock or bootmem)
235 * has been activated and all other subsystems have already allocated/reserved
236 * memory. This function allows to create custom reserved areas.
238 * If @fixed is true, reserve contiguous area at exactly @base. If false,
239 * reserve in range from @base to @limit.
241 int __init cma_declare_contiguous_nid(phys_addr_t base,
242 phys_addr_t size, phys_addr_t limit,
243 phys_addr_t alignment, unsigned int order_per_bit,
244 bool fixed, const char *name, struct cma **res_cma,
247 phys_addr_t memblock_end = memblock_end_of_DRAM();
248 phys_addr_t highmem_start;
252 * We can't use __pa(high_memory) directly, since high_memory
253 * isn't a valid direct map VA, and DEBUG_VIRTUAL will (validly)
254 * complain. Find the boundary by adding one to the last valid
257 highmem_start = __pa(high_memory - 1) + 1;
258 pr_debug("%s(size %pa, base %pa, limit %pa alignment %pa)\n",
259 __func__, &size, &base, &limit, &alignment);
261 if (cma_area_count == ARRAY_SIZE(cma_areas)) {
262 pr_err("Not enough slots for CMA reserved regions!\n");
269 if (alignment && !is_power_of_2(alignment))
273 * Sanitise input arguments.
274 * Pages both ends in CMA area could be merged into adjacent unmovable
275 * migratetype page by page allocator's buddy algorithm. In the case,
276 * you couldn't get a contiguous memory, which is not what we want.
278 alignment = max(alignment, (phys_addr_t)PAGE_SIZE <<
279 max_t(unsigned long, MAX_ORDER - 1, pageblock_order));
280 if (fixed && base & (alignment - 1)) {
282 pr_err("Region at %pa must be aligned to %pa bytes\n",
286 base = ALIGN(base, alignment);
287 size = ALIGN(size, alignment);
288 limit &= ~(alignment - 1);
293 /* size should be aligned with order_per_bit */
294 if (!IS_ALIGNED(size >> PAGE_SHIFT, 1 << order_per_bit))
298 * If allocating at a fixed base the request region must not cross the
299 * low/high memory boundary.
301 if (fixed && base < highmem_start && base + size > highmem_start) {
303 pr_err("Region at %pa defined on low/high memory boundary (%pa)\n",
304 &base, &highmem_start);
309 * If the limit is unspecified or above the memblock end, its effective
310 * value will be the memblock end. Set it explicitly to simplify further
313 if (limit == 0 || limit > memblock_end)
314 limit = memblock_end;
316 if (base + size > limit) {
318 pr_err("Size (%pa) of region at %pa exceeds limit (%pa)\n",
319 &size, &base, &limit);
325 if (memblock_is_region_reserved(base, size) ||
326 memblock_reserve(base, size) < 0) {
331 phys_addr_t addr = 0;
334 * All pages in the reserved area must come from the same zone.
335 * If the requested region crosses the low/high memory boundary,
336 * try allocating from high memory first and fall back to low
337 * memory in case of failure.
339 if (base < highmem_start && limit > highmem_start) {
340 addr = memblock_alloc_range_nid(size, alignment,
341 highmem_start, limit, nid, true);
342 limit = highmem_start;
346 addr = memblock_alloc_range_nid(size, alignment, base,
355 * kmemleak scans/reads tracked objects for pointers to other
356 * objects but this address isn't mapped and accessible
358 kmemleak_ignore_phys(addr);
362 ret = cma_init_reserved_mem(base, size, order_per_bit, name, res_cma);
366 pr_info("Reserved %ld MiB at %pa\n", (unsigned long)size / SZ_1M,
371 memblock_free(base, size);
373 pr_err("Failed to reserve %ld MiB\n", (unsigned long)size / SZ_1M);
377 #ifdef CONFIG_CMA_DEBUG
378 static void cma_debug_show_areas(struct cma *cma)
380 unsigned long next_zero_bit, next_set_bit, nr_zero;
381 unsigned long start = 0;
382 unsigned long nr_part, nr_total = 0;
383 unsigned long nbits = cma_bitmap_maxno(cma);
385 mutex_lock(&cma->lock);
386 pr_info("number of available pages: ");
388 next_zero_bit = find_next_zero_bit(cma->bitmap, nbits, start);
389 if (next_zero_bit >= nbits)
391 next_set_bit = find_next_bit(cma->bitmap, nbits, next_zero_bit);
392 nr_zero = next_set_bit - next_zero_bit;
393 nr_part = nr_zero << cma->order_per_bit;
394 pr_cont("%s%lu@%lu", nr_total ? "+" : "", nr_part,
397 start = next_zero_bit + nr_zero;
399 pr_cont("=> %lu free of %lu total pages\n", nr_total, cma->count);
400 mutex_unlock(&cma->lock);
403 static inline void cma_debug_show_areas(struct cma *cma) { }
407 * cma_alloc() - allocate pages from contiguous area
408 * @cma: Contiguous memory region for which the allocation is performed.
409 * @count: Requested number of pages.
410 * @align: Requested alignment of pages (in PAGE_SIZE order).
411 * @no_warn: Avoid printing message about failed allocation
413 * This function allocates part of contiguous memory on specific
414 * contiguous memory area.
416 struct page *cma_alloc(struct cma *cma, size_t count, unsigned int align,
419 unsigned long mask, offset;
420 unsigned long pfn = -1;
421 unsigned long start = 0;
422 unsigned long bitmap_maxno, bitmap_no, bitmap_count;
424 struct page *page = NULL;
427 if (!cma || !cma->count || !cma->bitmap)
430 pr_debug("%s(cma %p, count %zu, align %d)\n", __func__, (void *)cma,
436 mask = cma_bitmap_aligned_mask(cma, align);
437 offset = cma_bitmap_aligned_offset(cma, align);
438 bitmap_maxno = cma_bitmap_maxno(cma);
439 bitmap_count = cma_bitmap_pages_to_bits(cma, count);
441 if (bitmap_count > bitmap_maxno)
445 mutex_lock(&cma->lock);
446 bitmap_no = bitmap_find_next_zero_area_off(cma->bitmap,
447 bitmap_maxno, start, bitmap_count, mask,
449 if (bitmap_no >= bitmap_maxno) {
450 mutex_unlock(&cma->lock);
453 bitmap_set(cma->bitmap, bitmap_no, bitmap_count);
455 * It's safe to drop the lock here. We've marked this region for
456 * our exclusive use. If the migration fails we will take the
457 * lock again and unmark it.
459 mutex_unlock(&cma->lock);
461 pfn = cma->base_pfn + (bitmap_no << cma->order_per_bit);
462 mutex_lock(&cma_mutex);
463 ret = alloc_contig_range(pfn, pfn + count, MIGRATE_CMA,
464 GFP_KERNEL | (no_warn ? __GFP_NOWARN : 0));
465 mutex_unlock(&cma_mutex);
467 page = pfn_to_page(pfn);
471 cma_clear_bitmap(cma, pfn, count);
475 pr_debug("%s(): memory range at %p is busy, retrying\n",
476 __func__, pfn_to_page(pfn));
477 /* try again with a bit different memory target */
478 start = bitmap_no + mask + 1;
481 trace_cma_alloc(pfn, page, count, align);
484 * CMA can allocate multiple page blocks, which results in different
485 * blocks being marked with different tags. Reset the tags to ignore
489 for (i = 0; i < count; i++)
490 page_kasan_tag_reset(page + i);
493 if (ret && !no_warn) {
494 pr_err("%s: alloc failed, req-size: %zu pages, ret: %d\n",
495 __func__, count, ret);
496 cma_debug_show_areas(cma);
499 pr_debug("%s(): returned %p\n", __func__, page);
504 * cma_release() - release allocated pages
505 * @cma: Contiguous memory region for which the allocation is performed.
506 * @pages: Allocated pages.
507 * @count: Number of allocated pages.
509 * This function releases memory allocated by cma_alloc().
510 * It returns false when provided pages do not belong to contiguous area and
513 bool cma_release(struct cma *cma, const struct page *pages, unsigned int count)
520 pr_debug("%s(page %p)\n", __func__, (void *)pages);
522 pfn = page_to_pfn(pages);
524 if (pfn < cma->base_pfn || pfn >= cma->base_pfn + cma->count)
527 VM_BUG_ON(pfn + count > cma->base_pfn + cma->count);
529 free_contig_range(pfn, count);
530 cma_clear_bitmap(cma, pfn, count);
531 trace_cma_release(pfn, pages, count);
536 int cma_for_each_area(int (*it)(struct cma *cma, void *data), void *data)
540 for (i = 0; i < cma_area_count; i++) {
541 int ret = it(&cma_areas[i], data);