1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * Contiguous Memory Allocator
4 *
5 * Copyright (c) 2010-2011 by Samsung Electronics.
6 * Copyright IBM Corporation, 2013
7 * Copyright LG Electronics Inc., 2014
8 * Written by:
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>
13 */
14
15 #define pr_fmt(fmt) "cma: " fmt
16
17 #define CREATE_TRACE_POINTS
18
19 #include <linux/memblock.h>
20 #include <linux/err.h>
21 #include <linux/mm.h>
22 #include <linux/sizes.h>
23 #include <linux/slab.h>
24 #include <linux/log2.h>
25 #include <linux/cma.h>
26 #include <linux/highmem.h>
27 #include <linux/io.h>
28 #include <linux/kmemleak.h>
29 #include <trace/events/cma.h>
30
31 #include "internal.h"
32 #include "cma.h"
33
34 struct cma cma_areas[MAX_CMA_AREAS];
35 unsigned cma_area_count;
36 static DEFINE_MUTEX(cma_mutex);
37
38 phys_addr_t cma_get_base(const struct cma *cma)
39 {
40 return PFN_PHYS(cma->base_pfn);
41 }
42
43 unsigned long cma_get_size(const struct cma *cma)
44 {
45 return cma->count << PAGE_SHIFT;
46 }
47
48 const char *cma_get_name(const struct cma *cma)
49 {
50 return cma->name;
51 }
52
53 static unsigned long cma_bitmap_aligned_mask(const struct cma *cma,
54 unsigned int align_order)
55 {
56 if (align_order <= cma->order_per_bit)
57 return 0;
58 return (1UL << (align_order - cma->order_per_bit)) - 1;
59 }
60
61 /*
62 * Find the offset of the base PFN from the specified align_order.
63 * The value returned is represented in order_per_bits.
64 */
65 static unsigned long cma_bitmap_aligned_offset(const struct cma *cma,
66 unsigned int align_order)
67 {
68 return (cma->base_pfn & ((1UL << align_order) - 1))
69 >> cma->order_per_bit;
70 }
71
72 static unsigned long cma_bitmap_pages_to_bits(const struct cma *cma,
73 unsigned long pages)
74 {
75 return ALIGN(pages, 1UL << cma->order_per_bit) >> cma->order_per_bit;
76 }
77
78 static void cma_clear_bitmap(struct cma *cma, unsigned long pfn,
79 unsigned long count)
80 {
81 unsigned long bitmap_no, bitmap_count;
82 unsigned long flags;
83
84 bitmap_no = (pfn - cma->base_pfn) >> cma->order_per_bit;
85 bitmap_count = cma_bitmap_pages_to_bits(cma, count);
86
87 spin_lock_irqsave(&cma->lock, flags);
88 bitmap_clear(cma->bitmap, bitmap_no, bitmap_count);
89 spin_unlock_irqrestore(&cma->lock, flags);
90 }
91
92 static void __init cma_activate_area(struct cma *cma)
93 {
94 unsigned long base_pfn = cma->base_pfn, pfn;
95 struct zone *zone;
96
97 cma->bitmap = bitmap_zalloc(cma_bitmap_maxno(cma), GFP_KERNEL);
98 if (!cma->bitmap)
99 goto out_error;
100
101 /*
102 * alloc_contig_range() requires the pfn range specified to be in the
103 * same zone. Simplify by forcing the entire CMA resv range to be in the
104 * same zone.
105 */
106 WARN_ON_ONCE(!pfn_valid(base_pfn));
107 zone = page_zone(pfn_to_page(base_pfn));
108 for (pfn = base_pfn + 1; pfn < base_pfn + cma->count; pfn++) {
109 WARN_ON_ONCE(!pfn_valid(pfn));
110 if (page_zone(pfn_to_page(pfn)) != zone)
111 goto not_in_zone;
112 }
113
114 for (pfn = base_pfn; pfn < base_pfn + cma->count;
115 pfn += pageblock_nr_pages)
116 init_cma_reserved_pageblock(pfn_to_page(pfn));
117
118 spin_lock_init(&cma->lock);
119
120 #ifdef CONFIG_CMA_DEBUGFS
121 INIT_HLIST_HEAD(&cma->mem_head);
122 spin_lock_init(&cma->mem_head_lock);
123 #endif
124
125 return;
126
127 not_in_zone:
128 bitmap_free(cma->bitmap);
129 out_error:
130 /* Expose all pages to the buddy, they are useless for CMA. */
131 if (!cma->reserve_pages_on_error) {
132 for (pfn = base_pfn; pfn < base_pfn + cma->count; pfn++)
133 free_reserved_page(pfn_to_page(pfn));
134 }
135 totalcma_pages -= cma->count;
136 cma->count = 0;
137 pr_err("CMA area %s could not be activated\n", cma->name);
138 return;
139 }
140
141 static int __init cma_init_reserved_areas(void)
142 {
143 int i;
144
145 for (i = 0; i < cma_area_count; i++)
146 cma_activate_area(&cma_areas[i]);
147
148 return 0;
149 }
150 core_initcall(cma_init_reserved_areas);
151
152 void __init cma_reserve_pages_on_error(struct cma *cma)
153 {
154 cma->reserve_pages_on_error = true;
155 }
156
157 /**
158 * cma_init_reserved_mem() - create custom contiguous area from reserved memory
159 * @base: Base address of the reserved area
160 * @size: Size of the reserved area (in bytes),
161 * @order_per_bit: Order of pages represented by one bit on bitmap.
162 * @name: The name of the area. If this parameter is NULL, the name of
163 * the area will be set to "cmaN", where N is a running counter of
164 * used areas.
165 * @res_cma: Pointer to store the created cma region.
166 *
167 * This function creates custom contiguous area from already reserved memory.
168 */
169 int __init cma_init_reserved_mem(phys_addr_t base, phys_addr_t size,
170 unsigned int order_per_bit,
171 const char *name,
172 struct cma **res_cma)
173 {
174 struct cma *cma;
175
176 /* Sanity checks */
177 if (cma_area_count == ARRAY_SIZE(cma_areas)) {
178 pr_err("Not enough slots for CMA reserved regions!\n");
179 return -ENOSPC;
180 }
181
182 if (!size || !memblock_is_region_reserved(base, size))
183 return -EINVAL;
184
185 /* ensure minimal alignment required by mm core */
186 if (!IS_ALIGNED(base | size, CMA_MIN_ALIGNMENT_BYTES))
187 return -EINVAL;
188
189 /*
190 * Each reserved area must be initialised later, when more kernel
191 * subsystems (like slab allocator) are available.
192 */
193 cma = &cma_areas[cma_area_count];
194
195 if (name)
196 snprintf(cma->name, CMA_MAX_NAME, name);
197 else
198 snprintf(cma->name, CMA_MAX_NAME, "cma%d\n", cma_area_count);
199
200 cma->base_pfn = PFN_DOWN(base);
201 cma->count = size >> PAGE_SHIFT;
202 cma->order_per_bit = order_per_bit;
203 *res_cma = cma;
204 cma_area_count++;
205 totalcma_pages += (size / PAGE_SIZE);
206
207 return 0;
208 }
209
210 /**
211 * cma_declare_contiguous_nid() - reserve custom contiguous area
212 * @base: Base address of the reserved area optional, use 0 for any
213 * @size: Size of the reserved area (in bytes),
214 * @limit: End address of the reserved memory (optional, 0 for any).
215 * @alignment: Alignment for the CMA area, should be power of 2 or zero
216 * @order_per_bit: Order of pages represented by one bit on bitmap.
217 * @fixed: hint about where to place the reserved area
218 * @name: The name of the area. See function cma_init_reserved_mem()
219 * @res_cma: Pointer to store the created cma region.
220 * @nid: nid of the free area to find, %NUMA_NO_NODE for any node
221 *
222 * This function reserves memory from early allocator. It should be
223 * called by arch specific code once the early allocator (memblock or bootmem)
224 * has been activated and all other subsystems have already allocated/reserved
225 * memory. This function allows to create custom reserved areas.
226 *
227 * If @fixed is true, reserve contiguous area at exactly @base. If false,
228 * reserve in range from @base to @limit.
229 */
230 int __init cma_declare_contiguous_nid(phys_addr_t base,
231 phys_addr_t size, phys_addr_t limit,
232 phys_addr_t alignment, unsigned int order_per_bit,
233 bool fixed, const char *name, struct cma **res_cma,
234 int nid)
235 {
236 phys_addr_t memblock_end = memblock_end_of_DRAM();
237 phys_addr_t highmem_start;
238 int ret;
239
240 /*
241 * We can't use __pa(high_memory) directly, since high_memory
242 * isn't a valid direct map VA, and DEBUG_VIRTUAL will (validly)
243 * complain. Find the boundary by adding one to the last valid
244 * address.
245 */
246 highmem_start = __pa(high_memory - 1) + 1;
247 pr_debug("%s(size %pa, base %pa, limit %pa alignment %pa)\n",
248 __func__, &size, &base, &limit, &alignment);
249
250 if (cma_area_count == ARRAY_SIZE(cma_areas)) {
251 pr_err("Not enough slots for CMA reserved regions!\n");
252 return -ENOSPC;
253 }
254
255 if (!size)
256 return -EINVAL;
257
258 if (alignment && !is_power_of_2(alignment))
259 return -EINVAL;
260
261 if (!IS_ENABLED(CONFIG_NUMA))
262 nid = NUMA_NO_NODE;
263
264 /* Sanitise input arguments. */
265 alignment = max_t(phys_addr_t, alignment, CMA_MIN_ALIGNMENT_BYTES);
266 if (fixed && base & (alignment - 1)) {
267 ret = -EINVAL;
268 pr_err("Region at %pa must be aligned to %pa bytes\n",
269 &base, &alignment);
270 goto err;
271 }
272 base = ALIGN(base, alignment);
273 size = ALIGN(size, alignment);
274 limit &= ~(alignment - 1);
275
276 if (!base)
277 fixed = false;
278
279 /* size should be aligned with order_per_bit */
280 if (!IS_ALIGNED(size >> PAGE_SHIFT, 1 << order_per_bit))
281 return -EINVAL;
282
283 /*
284 * If allocating at a fixed base the request region must not cross the
285 * low/high memory boundary.
286 */
287 if (fixed && base < highmem_start && base + size > highmem_start) {
288 ret = -EINVAL;
289 pr_err("Region at %pa defined on low/high memory boundary (%pa)\n",
290 &base, &highmem_start);
291 goto err;
292 }
293
294 /*
295 * If the limit is unspecified or above the memblock end, its effective
296 * value will be the memblock end. Set it explicitly to simplify further
297 * checks.
298 */
299 if (limit == 0 || limit > memblock_end)
300 limit = memblock_end;
301
302 if (base + size > limit) {
303 ret = -EINVAL;
304 pr_err("Size (%pa) of region at %pa exceeds limit (%pa)\n",
305 &size, &base, &limit);
306 goto err;
307 }
308
309 /* Reserve memory */
310 if (fixed) {
311 if (memblock_is_region_reserved(base, size) ||
312 memblock_reserve(base, size) < 0) {
313 ret = -EBUSY;
314 goto err;
315 }
316 } else {
317 phys_addr_t addr = 0;
318
319 /*
320 * If there is enough memory, try a bottom-up allocation first.
321 * It will place the new cma area close to the start of the node
322 * and guarantee that the compaction is moving pages out of the
323 * cma area and not into it.
324 * Avoid using first 4GB to not interfere with constrained zones
325 * like DMA/DMA32.
326 */
327 #ifdef CONFIG_PHYS_ADDR_T_64BIT
328 if (!memblock_bottom_up() && memblock_end >= SZ_4G + size) {
329 memblock_set_bottom_up(true);
330 addr = memblock_alloc_range_nid(size, alignment, SZ_4G,
331 limit, nid, true);
332 memblock_set_bottom_up(false);
333 }
334 #endif
335
336 /*
337 * All pages in the reserved area must come from the same zone.
338 * If the requested region crosses the low/high memory boundary,
339 * try allocating from high memory first and fall back to low
340 * memory in case of failure.
341 */
342 if (!addr && base < highmem_start && limit > highmem_start) {
343 addr = memblock_alloc_range_nid(size, alignment,
344 highmem_start, limit, nid, true);
345 limit = highmem_start;
346 }
347
348 if (!addr) {
349 addr = memblock_alloc_range_nid(size, alignment, base,
350 limit, nid, true);
351 if (!addr) {
352 ret = -ENOMEM;
353 goto err;
354 }
355 }
356
357 /*
358 * kmemleak scans/reads tracked objects for pointers to other
359 * objects but this address isn't mapped and accessible
360 */
361 kmemleak_ignore_phys(addr);
362 base = addr;
363 }
364
365 ret = cma_init_reserved_mem(base, size, order_per_bit, name, res_cma);
366 if (ret)
367 goto free_mem;
368
369 pr_info("Reserved %ld MiB at %pa on node %d\n", (unsigned long)size / SZ_1M,
370 &base, nid);
371 return 0;
372
373 free_mem:
374 memblock_phys_free(base, size);
375 err:
376 pr_err("Failed to reserve %ld MiB on node %d\n", (unsigned long)size / SZ_1M,
377 nid);
378 return ret;
379 }
380
381 static void cma_debug_show_areas(struct cma *cma)
382 {
383 unsigned long next_zero_bit, next_set_bit, nr_zero;
384 unsigned long start = 0;
385 unsigned long nr_part, nr_total = 0;
386 unsigned long nbits = cma_bitmap_maxno(cma);
387
388 spin_lock_irq(&cma->lock);
389 pr_info("number of available pages: ");
390 for (;;) {
391 next_zero_bit = find_next_zero_bit(cma->bitmap, nbits, start);
392 if (next_zero_bit >= nbits)
393 break;
394 next_set_bit = find_next_bit(cma->bitmap, nbits, next_zero_bit);
395 nr_zero = next_set_bit - next_zero_bit;
396 nr_part = nr_zero << cma->order_per_bit;
397 pr_cont("%s%lu@%lu", nr_total ? "+" : "", nr_part,
398 next_zero_bit);
399 nr_total += nr_part;
400 start = next_zero_bit + nr_zero;
401 }
402 pr_cont("=> %lu free of %lu total pages\n", nr_total, cma->count);
403 spin_unlock_irq(&cma->lock);
404 }
405
406 /**
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
412 *
413 * This function allocates part of contiguous memory on specific
414 * contiguous memory area.
415 */
416 struct page *cma_alloc(struct cma *cma, unsigned long count,
417 unsigned int align, bool no_warn)
418 {
419 unsigned long mask, offset;
420 unsigned long pfn = -1;
421 unsigned long start = 0;
422 unsigned long bitmap_maxno, bitmap_no, bitmap_count;
423 unsigned long i;
424 struct page *page = NULL;
425 int ret = -ENOMEM;
426 const char *name = cma ? cma->name : NULL;
427
428 trace_cma_alloc_start(name, count, align);
429
430 if (!cma || !cma->count || !cma->bitmap)
431 return page;
432
433 pr_debug("%s(cma %p, name: %s, count %lu, align %d)\n", __func__,
434 (void *)cma, cma->name, count, align);
435
436 if (!count)
437 return page;
438
439 mask = cma_bitmap_aligned_mask(cma, align);
440 offset = cma_bitmap_aligned_offset(cma, align);
441 bitmap_maxno = cma_bitmap_maxno(cma);
442 bitmap_count = cma_bitmap_pages_to_bits(cma, count);
443
444 if (bitmap_count > bitmap_maxno)
445 return page;
446
447 for (;;) {
448 spin_lock_irq(&cma->lock);
449 bitmap_no = bitmap_find_next_zero_area_off(cma->bitmap,
450 bitmap_maxno, start, bitmap_count, mask,
451 offset);
452 if (bitmap_no >= bitmap_maxno) {
453 spin_unlock_irq(&cma->lock);
454 break;
455 }
456 bitmap_set(cma->bitmap, bitmap_no, bitmap_count);
457 /*
458 * It's safe to drop the lock here. We've marked this region for
459 * our exclusive use. If the migration fails we will take the
460 * lock again and unmark it.
461 */
462 spin_unlock_irq(&cma->lock);
463
464 pfn = cma->base_pfn + (bitmap_no << cma->order_per_bit);
465 mutex_lock(&cma_mutex);
466 ret = alloc_contig_range(pfn, pfn + count, MIGRATE_CMA,
467 GFP_KERNEL | (no_warn ? __GFP_NOWARN : 0));
468 mutex_unlock(&cma_mutex);
469 if (ret == 0) {
470 page = pfn_to_page(pfn);
471 break;
472 }
473
474 cma_clear_bitmap(cma, pfn, count);
475 if (ret != -EBUSY)
476 break;
477
478 pr_debug("%s(): memory range at pfn 0x%lx %p is busy, retrying\n",
479 __func__, pfn, pfn_to_page(pfn));
480
481 trace_cma_alloc_busy_retry(cma->name, pfn, pfn_to_page(pfn),
482 count, align);
483 /* try again with a bit different memory target */
484 start = bitmap_no + mask + 1;
485 }
486
487 /*
488 * CMA can allocate multiple page blocks, which results in different
489 * blocks being marked with different tags. Reset the tags to ignore
490 * those page blocks.
491 */
492 if (page) {
493 for (i = 0; i < count; i++)
494 page_kasan_tag_reset(nth_page(page, i));
495 }
496
497 if (ret && !no_warn) {
498 pr_err_ratelimited("%s: %s: alloc failed, req-size: %lu pages, ret: %d\n",
499 __func__, cma->name, count, ret);
500 cma_debug_show_areas(cma);
501 }
502
503 pr_debug("%s(): returned %p\n", __func__, page);
504 trace_cma_alloc_finish(name, pfn, page, count, align, ret);
505 if (page) {
506 count_vm_event(CMA_ALLOC_SUCCESS);
507 cma_sysfs_account_success_pages(cma, count);
508 } else {
509 count_vm_event(CMA_ALLOC_FAIL);
510 cma_sysfs_account_fail_pages(cma, count);
511 }
512
513 return page;
514 }
515
516 bool cma_pages_valid(struct cma *cma, const struct page *pages,
517 unsigned long count)
518 {
519 unsigned long pfn;
520
521 if (!cma || !pages)
522 return false;
523
524 pfn = page_to_pfn(pages);
525
526 if (pfn < cma->base_pfn || pfn >= cma->base_pfn + cma->count) {
527 pr_debug("%s(page %p, count %lu)\n", __func__,
528 (void *)pages, count);
529 return false;
530 }
531
532 return true;
533 }
534
535 /**
536 * cma_release() - release allocated pages
537 * @cma: Contiguous memory region for which the allocation is performed.
538 * @pages: Allocated pages.
539 * @count: Number of allocated pages.
540 *
541 * This function releases memory allocated by cma_alloc().
542 * It returns false when provided pages do not belong to contiguous area and
543 * true otherwise.
544 */
545 bool cma_release(struct cma *cma, const struct page *pages,
546 unsigned long count)
547 {
548 unsigned long pfn;
549
550 if (!cma_pages_valid(cma, pages, count))
551 return false;
552
553 pr_debug("%s(page %p, count %lu)\n", __func__, (void *)pages, count);
554
555 pfn = page_to_pfn(pages);
556
557 VM_BUG_ON(pfn + count > cma->base_pfn + cma->count);
558
559 free_contig_range(pfn, count);
560 cma_clear_bitmap(cma, pfn, count);
561 cma_sysfs_account_release_pages(cma, count);
562 trace_cma_release(cma->name, pfn, pages, count);
563
564 return true;
565 }
566
567 int cma_for_each_area(int (*it)(struct cma *cma, void *data), void *data)
568 {
569 int i;
570
571 for (i = 0; i < cma_area_count; i++) {
572 int ret = it(&cma_areas[i], data);
573
574 if (ret)
575 return ret;
576 }
577
578 return 0;
579 }
580
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.