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/sizes.h>
28 #include <linux/slab.h>
29 #include <linux/log2.h>
30 #include <linux/cma.h>
31 #include <linux/highmem.h>
33 #include <linux/kmemleak.h>
34 #include <trace/events/cma.h>
38 struct cma cma_areas[MAX_CMA_AREAS];
39 unsigned cma_area_count;
40 static DEFINE_MUTEX(cma_mutex);
42 phys_addr_t cma_get_base(const struct cma *cma)
44 return PFN_PHYS(cma->base_pfn);
47 unsigned long cma_get_size(const struct cma *cma)
49 return cma->count << PAGE_SHIFT;
52 const char *cma_get_name(const struct cma *cma)
57 static unsigned long cma_bitmap_aligned_mask(const struct cma *cma,
58 unsigned int align_order)
60 if (align_order <= cma->order_per_bit)
62 return (1UL << (align_order - cma->order_per_bit)) - 1;
66 * Find the offset of the base PFN from the specified align_order.
67 * The value returned is represented in order_per_bits.
69 static unsigned long cma_bitmap_aligned_offset(const struct cma *cma,
70 unsigned int align_order)
72 return (cma->base_pfn & ((1UL << align_order) - 1))
73 >> cma->order_per_bit;
76 static unsigned long cma_bitmap_pages_to_bits(const struct cma *cma,
79 return ALIGN(pages, 1UL << cma->order_per_bit) >> cma->order_per_bit;
82 static void cma_clear_bitmap(struct cma *cma, unsigned long pfn,
85 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 spin_lock_irqsave(&cma->lock, flags);
92 bitmap_clear(cma->bitmap, bitmap_no, bitmap_count);
93 spin_unlock_irqrestore(&cma->lock, flags);
96 static void __init cma_activate_area(struct cma *cma)
98 unsigned long base_pfn = cma->base_pfn, pfn;
101 cma->bitmap = bitmap_zalloc(cma_bitmap_maxno(cma), GFP_KERNEL);
106 * alloc_contig_range() requires the pfn range specified to be in the
107 * same zone. Simplify by forcing the entire CMA resv range to be in the
110 WARN_ON_ONCE(!pfn_valid(base_pfn));
111 zone = page_zone(pfn_to_page(base_pfn));
112 for (pfn = base_pfn + 1; pfn < base_pfn + cma->count; pfn++) {
113 WARN_ON_ONCE(!pfn_valid(pfn));
114 if (page_zone(pfn_to_page(pfn)) != zone)
118 for (pfn = base_pfn; pfn < base_pfn + cma->count;
119 pfn += pageblock_nr_pages)
120 init_cma_reserved_pageblock(pfn_to_page(pfn));
122 spin_lock_init(&cma->lock);
124 #ifdef CONFIG_CMA_DEBUGFS
125 INIT_HLIST_HEAD(&cma->mem_head);
126 spin_lock_init(&cma->mem_head_lock);
132 bitmap_free(cma->bitmap);
134 /* Expose all pages to the buddy, they are useless for CMA. */
135 for (pfn = base_pfn; pfn < base_pfn + cma->count; pfn++)
136 free_reserved_page(pfn_to_page(pfn));
137 totalcma_pages -= cma->count;
139 pr_err("CMA area %s could not be activated\n", cma->name);
143 static int __init cma_init_reserved_areas(void)
147 for (i = 0; i < cma_area_count; i++)
148 cma_activate_area(&cma_areas[i]);
152 core_initcall(cma_init_reserved_areas);
155 * cma_init_reserved_mem() - create custom contiguous area from reserved memory
156 * @base: Base address of the reserved area
157 * @size: Size of the reserved area (in bytes),
158 * @order_per_bit: Order of pages represented by one bit on bitmap.
159 * @name: The name of the area. If this parameter is NULL, the name of
160 * the area will be set to "cmaN", where N is a running counter of
162 * @res_cma: Pointer to store the created cma region.
164 * This function creates custom contiguous area from already reserved memory.
166 int __init cma_init_reserved_mem(phys_addr_t base, phys_addr_t size,
167 unsigned int order_per_bit,
169 struct cma **res_cma)
172 phys_addr_t alignment;
175 if (cma_area_count == ARRAY_SIZE(cma_areas)) {
176 pr_err("Not enough slots for CMA reserved regions!\n");
180 if (!size || !memblock_is_region_reserved(base, size))
183 /* ensure minimal alignment required by mm core */
184 alignment = PAGE_SIZE <<
185 max_t(unsigned long, MAX_ORDER - 1, pageblock_order);
187 /* alignment should be aligned with order_per_bit */
188 if (!IS_ALIGNED(alignment >> PAGE_SHIFT, 1 << order_per_bit))
191 if (ALIGN(base, alignment) != base || ALIGN(size, alignment) != size)
195 * Each reserved area must be initialised later, when more kernel
196 * subsystems (like slab allocator) are available.
198 cma = &cma_areas[cma_area_count];
201 snprintf(cma->name, CMA_MAX_NAME, name);
203 snprintf(cma->name, CMA_MAX_NAME, "cma%d\n", cma_area_count);
205 cma->base_pfn = PFN_DOWN(base);
206 cma->count = size >> PAGE_SHIFT;
207 cma->order_per_bit = order_per_bit;
210 totalcma_pages += (size / PAGE_SIZE);
216 * cma_declare_contiguous_nid() - reserve custom contiguous area
217 * @base: Base address of the reserved area optional, use 0 for any
218 * @size: Size of the reserved area (in bytes),
219 * @limit: End address of the reserved memory (optional, 0 for any).
220 * @alignment: Alignment for the CMA area, should be power of 2 or zero
221 * @order_per_bit: Order of pages represented by one bit on bitmap.
222 * @fixed: hint about where to place the reserved area
223 * @name: The name of the area. See function cma_init_reserved_mem()
224 * @res_cma: Pointer to store the created cma region.
225 * @nid: nid of the free area to find, %NUMA_NO_NODE for any node
227 * This function reserves memory from early allocator. It should be
228 * called by arch specific code once the early allocator (memblock or bootmem)
229 * has been activated and all other subsystems have already allocated/reserved
230 * memory. This function allows to create custom reserved areas.
232 * If @fixed is true, reserve contiguous area at exactly @base. If false,
233 * reserve in range from @base to @limit.
235 int __init cma_declare_contiguous_nid(phys_addr_t base,
236 phys_addr_t size, phys_addr_t limit,
237 phys_addr_t alignment, unsigned int order_per_bit,
238 bool fixed, const char *name, struct cma **res_cma,
241 phys_addr_t memblock_end = memblock_end_of_DRAM();
242 phys_addr_t highmem_start;
246 * We can't use __pa(high_memory) directly, since high_memory
247 * isn't a valid direct map VA, and DEBUG_VIRTUAL will (validly)
248 * complain. Find the boundary by adding one to the last valid
251 highmem_start = __pa(high_memory - 1) + 1;
252 pr_debug("%s(size %pa, base %pa, limit %pa alignment %pa)\n",
253 __func__, &size, &base, &limit, &alignment);
255 if (cma_area_count == ARRAY_SIZE(cma_areas)) {
256 pr_err("Not enough slots for CMA reserved regions!\n");
263 if (alignment && !is_power_of_2(alignment))
267 * Sanitise input arguments.
268 * Pages both ends in CMA area could be merged into adjacent unmovable
269 * migratetype page by page allocator's buddy algorithm. In the case,
270 * you couldn't get a contiguous memory, which is not what we want.
272 alignment = max(alignment, (phys_addr_t)PAGE_SIZE <<
273 max_t(unsigned long, MAX_ORDER - 1, pageblock_order));
274 if (fixed && base & (alignment - 1)) {
276 pr_err("Region at %pa must be aligned to %pa bytes\n",
280 base = ALIGN(base, alignment);
281 size = ALIGN(size, alignment);
282 limit &= ~(alignment - 1);
287 /* size should be aligned with order_per_bit */
288 if (!IS_ALIGNED(size >> PAGE_SHIFT, 1 << order_per_bit))
292 * If allocating at a fixed base the request region must not cross the
293 * low/high memory boundary.
295 if (fixed && base < highmem_start && base + size > highmem_start) {
297 pr_err("Region at %pa defined on low/high memory boundary (%pa)\n",
298 &base, &highmem_start);
303 * If the limit is unspecified or above the memblock end, its effective
304 * value will be the memblock end. Set it explicitly to simplify further
307 if (limit == 0 || limit > memblock_end)
308 limit = memblock_end;
310 if (base + size > limit) {
312 pr_err("Size (%pa) of region at %pa exceeds limit (%pa)\n",
313 &size, &base, &limit);
319 if (memblock_is_region_reserved(base, size) ||
320 memblock_reserve(base, size) < 0) {
325 phys_addr_t addr = 0;
328 * All pages in the reserved area must come from the same zone.
329 * If the requested region crosses the low/high memory boundary,
330 * try allocating from high memory first and fall back to low
331 * memory in case of failure.
333 if (base < highmem_start && limit > highmem_start) {
334 addr = memblock_alloc_range_nid(size, alignment,
335 highmem_start, limit, nid, true);
336 limit = highmem_start;
340 * If there is enough memory, try a bottom-up allocation first.
341 * It will place the new cma area close to the start of the node
342 * and guarantee that the compaction is moving pages out of the
343 * cma area and not into it.
344 * Avoid using first 4GB to not interfere with constrained zones
347 #ifdef CONFIG_PHYS_ADDR_T_64BIT
348 if (!memblock_bottom_up() && memblock_end >= SZ_4G + size) {
349 memblock_set_bottom_up(true);
350 addr = memblock_alloc_range_nid(size, alignment, SZ_4G,
352 memblock_set_bottom_up(false);
357 addr = memblock_alloc_range_nid(size, alignment, base,
366 * kmemleak scans/reads tracked objects for pointers to other
367 * objects but this address isn't mapped and accessible
369 kmemleak_ignore_phys(addr);
373 ret = cma_init_reserved_mem(base, size, order_per_bit, name, res_cma);
377 pr_info("Reserved %ld MiB at %pa\n", (unsigned long)size / SZ_1M,
382 memblock_free(base, size);
384 pr_err("Failed to reserve %ld MiB\n", (unsigned long)size / SZ_1M);
388 #ifdef CONFIG_CMA_DEBUG
389 static void cma_debug_show_areas(struct cma *cma)
391 unsigned long next_zero_bit, next_set_bit, nr_zero;
392 unsigned long start = 0;
393 unsigned long nr_part, nr_total = 0;
394 unsigned long nbits = cma_bitmap_maxno(cma);
396 spin_lock_irq(&cma->lock);
397 pr_info("number of available pages: ");
399 next_zero_bit = find_next_zero_bit(cma->bitmap, nbits, start);
400 if (next_zero_bit >= nbits)
402 next_set_bit = find_next_bit(cma->bitmap, nbits, next_zero_bit);
403 nr_zero = next_set_bit - next_zero_bit;
404 nr_part = nr_zero << cma->order_per_bit;
405 pr_cont("%s%lu@%lu", nr_total ? "+" : "", nr_part,
408 start = next_zero_bit + nr_zero;
410 pr_cont("=> %lu free of %lu total pages\n", nr_total, cma->count);
411 spin_unlock_irq(&cma->lock);
414 static inline void cma_debug_show_areas(struct cma *cma) { }
418 * cma_alloc() - allocate pages from contiguous area
419 * @cma: Contiguous memory region for which the allocation is performed.
420 * @count: Requested number of pages.
421 * @align: Requested alignment of pages (in PAGE_SIZE order).
422 * @no_warn: Avoid printing message about failed allocation
424 * This function allocates part of contiguous memory on specific
425 * contiguous memory area.
427 struct page *cma_alloc(struct cma *cma, unsigned long count,
428 unsigned int align, bool no_warn)
430 unsigned long mask, offset;
431 unsigned long pfn = -1;
432 unsigned long start = 0;
433 unsigned long bitmap_maxno, bitmap_no, bitmap_count;
435 struct page *page = NULL;
438 if (!cma || !cma->count || !cma->bitmap)
441 pr_debug("%s(cma %p, count %lu, align %d)\n", __func__, (void *)cma,
447 trace_cma_alloc_start(cma->name, count, align);
449 mask = cma_bitmap_aligned_mask(cma, align);
450 offset = cma_bitmap_aligned_offset(cma, align);
451 bitmap_maxno = cma_bitmap_maxno(cma);
452 bitmap_count = cma_bitmap_pages_to_bits(cma, count);
454 if (bitmap_count > bitmap_maxno)
458 spin_lock_irq(&cma->lock);
459 bitmap_no = bitmap_find_next_zero_area_off(cma->bitmap,
460 bitmap_maxno, start, bitmap_count, mask,
462 if (bitmap_no >= bitmap_maxno) {
463 spin_unlock_irq(&cma->lock);
466 bitmap_set(cma->bitmap, bitmap_no, bitmap_count);
468 * It's safe to drop the lock here. We've marked this region for
469 * our exclusive use. If the migration fails we will take the
470 * lock again and unmark it.
472 spin_unlock_irq(&cma->lock);
474 pfn = cma->base_pfn + (bitmap_no << cma->order_per_bit);
475 mutex_lock(&cma_mutex);
476 ret = alloc_contig_range(pfn, pfn + count, MIGRATE_CMA,
477 GFP_KERNEL | (no_warn ? __GFP_NOWARN : 0));
478 mutex_unlock(&cma_mutex);
480 page = pfn_to_page(pfn);
484 cma_clear_bitmap(cma, pfn, count);
488 pr_debug("%s(): memory range at %p is busy, retrying\n",
489 __func__, pfn_to_page(pfn));
491 trace_cma_alloc_busy_retry(cma->name, pfn, pfn_to_page(pfn),
493 /* try again with a bit different memory target */
494 start = bitmap_no + mask + 1;
497 trace_cma_alloc_finish(cma->name, pfn, page, count, align);
500 * CMA can allocate multiple page blocks, which results in different
501 * blocks being marked with different tags. Reset the tags to ignore
505 for (i = 0; i < count; i++)
506 page_kasan_tag_reset(page + i);
509 if (ret && !no_warn) {
510 pr_err_ratelimited("%s: %s: alloc failed, req-size: %lu pages, ret: %d\n",
511 __func__, cma->name, count, ret);
512 cma_debug_show_areas(cma);
515 pr_debug("%s(): returned %p\n", __func__, page);
518 count_vm_event(CMA_ALLOC_SUCCESS);
519 cma_sysfs_account_success_pages(cma, count);
521 count_vm_event(CMA_ALLOC_FAIL);
523 cma_sysfs_account_fail_pages(cma, count);
530 * cma_release() - release allocated pages
531 * @cma: Contiguous memory region for which the allocation is performed.
532 * @pages: Allocated pages.
533 * @count: Number of allocated pages.
535 * This function releases memory allocated by cma_alloc().
536 * It returns false when provided pages do not belong to contiguous area and
539 bool cma_release(struct cma *cma, const struct page *pages,
547 pr_debug("%s(page %p, count %lu)\n", __func__, (void *)pages, count);
549 pfn = page_to_pfn(pages);
551 if (pfn < cma->base_pfn || pfn >= cma->base_pfn + cma->count)
554 VM_BUG_ON(pfn + count > cma->base_pfn + cma->count);
556 free_contig_range(pfn, count);
557 cma_clear_bitmap(cma, pfn, count);
558 trace_cma_release(cma->name, pfn, pages, count);
563 int cma_for_each_area(int (*it)(struct cma *cma, void *data), void *data)
567 for (i = 0; i < cma_area_count; i++) {
568 int ret = it(&cma_areas[i], data);