projects/linux·blob v7.2 · refs

588 lines · 15.9 KiB

1// SPDX-License-Identifier: GPL-2.0+
2/*
3 * Contiguous Memory Allocator for DMA mapping framework
4 * Copyright (c) 2010-2011 by Samsung Electronics.
5 * Written by:
6 * Marek Szyprowski <[email protected]>
7 * Michal Nazarewicz <[email protected]>
8 *
9 * Contiguous Memory Allocator
10 *
11 * The Contiguous Memory Allocator (CMA) makes it possible to
12 * allocate big contiguous chunks of memory after the system has
13 * booted.
14 *
15 * Why is it needed?
16 *
17 * Various devices on embedded systems have no scatter-getter and/or
18 * IO map support and require contiguous blocks of memory to
19 * operate. They include devices such as cameras, hardware video
20 * coders, etc.
21 *
22 * Such devices often require big memory buffers (a full HD frame
23 * is, for instance, more than 2 mega pixels large, i.e. more than 6
24 * MB of memory), which makes mechanisms such as kmalloc() or
25 * alloc_page() ineffective.
26 *
27 * At the same time, a solution where a big memory region is
28 * reserved for a device is suboptimal since often more memory is
29 * reserved then strictly required and, moreover, the memory is
30 * inaccessible to page system even if device drivers don't use it.
31 *
32 * CMA tries to solve this issue by operating on memory regions
33 * where only movable pages can be allocated from. This way, kernel
34 * can use the memory for pagecache and when device driver requests
35 * it, allocated pages can be migrated.
36 */
37
38#define pr_fmt(fmt) "cma: " fmt
39
40#include <asm/page.h>
41
42#include <linux/memblock.h>
43#include <linux/err.h>
44#include <linux/sizes.h>
45#include <linux/dma-map-ops.h>
46#include <linux/cma.h>
47#include <linux/nospec.h>
48
49#ifdef CONFIG_CMA_SIZE_MBYTES
50#define CMA_SIZE_MBYTES CONFIG_CMA_SIZE_MBYTES
51#else
52#define CMA_SIZE_MBYTES 0
53#endif
54
55static struct cma *dma_contiguous_areas[MAX_CMA_AREAS];
56static unsigned int dma_contiguous_areas_num;
57
58static int dma_contiguous_insert_area(struct cma *cma)
59{
60 if (dma_contiguous_areas_num >= ARRAY_SIZE(dma_contiguous_areas))
61 return -EINVAL;
62
63 dma_contiguous_areas[dma_contiguous_areas_num++] = cma;
64
65 return 0;
66}
67
68/**
69 * dma_contiguous_get_area_by_idx() - Get contiguous area at given index
70 * @idx: index of the area we query
71 *
72 * Queries for the contiguous area located at index @idx.
73 *
74 * Returns:
75 * A pointer to the requested contiguous area, or NULL otherwise.
76 */
77struct cma *dma_contiguous_get_area_by_idx(unsigned int idx)
78{
79 if (idx >= dma_contiguous_areas_num)
80 return NULL;
81
82 return dma_contiguous_areas[idx];
83}
84EXPORT_SYMBOL_GPL(dma_contiguous_get_area_by_idx);
85
86static struct cma *dma_contiguous_default_area;
87
88/*
89 * Default global CMA area size can be defined in kernel's .config.
90 * This is useful mainly for distro maintainers to create a kernel
91 * that works correctly for most supported systems.
92 * The size can be set in bytes or as a percentage of the total memory
93 * in the system.
94 *
95 * Users, who want to set the size of global CMA area for their system
96 * should use cma= kernel parameter.
97 */
98#define size_bytes ((phys_addr_t)CMA_SIZE_MBYTES * SZ_1M)
99static phys_addr_t size_cmdline __initdata = -1;
100static phys_addr_t base_cmdline __initdata;
101static phys_addr_t limit_cmdline __initdata;
102
103static int __init early_cma(char *p)
104{
105 if (!p) {
106 pr_err("Config string not provided\n");
107 return -EINVAL;
108 }
109
110 size_cmdline = memparse(p, &p);
111 if (*p != '@')
112 return 0;
113 base_cmdline = memparse(p + 1, &p);
114 if (*p != '-') {
115 limit_cmdline = base_cmdline + size_cmdline;
116 return 0;
117 }
118 limit_cmdline = memparse(p + 1, &p);
119
120 return 0;
121}
122early_param("cma", early_cma);
123
124struct cma *dev_get_cma_area(struct device *dev)
125{
126 if (dev && dev->cma_area)
127 return dev->cma_area;
128
129 return dma_contiguous_default_area;
130}
131EXPORT_SYMBOL_GPL(dev_get_cma_area);
132
133#ifdef CONFIG_DMA_NUMA_CMA
134
135static struct cma *dma_contiguous_numa_area[MAX_NUMNODES];
136static phys_addr_t numa_cma_size[MAX_NUMNODES] __initdata;
137static phys_addr_t pernuma_size_bytes __initdata;
138static bool numa_cma_configured __initdata;
139
140static int __init early_numa_cma(char *p)
141{
142 int nid, count = 0;
143 unsigned long tmp;
144 char *s = p;
145
146 while (*s) {
147 if (sscanf(s, "%lu%n", &tmp, &count) != 1)
148 break;
149
150 if (s[count] == ':') {
151 if (tmp >= MAX_NUMNODES)
152 break;
153 nid = array_index_nospec(tmp, MAX_NUMNODES);
154
155 s += count + 1;
156 tmp = memparse(s, &s);
157 numa_cma_size[nid] = tmp;
158
159 if (*s == ',')
160 s++;
161 else
162 break;
163 } else
164 break;
165 }
166
167 numa_cma_configured = true;
168 return 0;
169}
170early_param("numa_cma", early_numa_cma);
171
172static int __init early_cma_pernuma(char *p)
173{
174 pernuma_size_bytes = memparse(p, &p);
175 numa_cma_configured = true;
176 return 0;
177}
178early_param("cma_pernuma", early_cma_pernuma);
179#endif
180
181#ifdef CONFIG_CMA_SIZE_PERCENTAGE
182
183static phys_addr_t __init __maybe_unused cma_early_percent_memory(void)
184{
185 unsigned long total_pages = PHYS_PFN(memblock_phys_mem_size());
186
187 return (total_pages * CONFIG_CMA_SIZE_PERCENTAGE / 100) << PAGE_SHIFT;
188}
189
190#else
191
192static inline __maybe_unused phys_addr_t cma_early_percent_memory(void)
193{
194 return 0;
195}
196
197#endif
198
199#ifdef CONFIG_DMA_NUMA_CMA
200static void __init dma_numa_cma_reserve(void)
201{
202 int nid;
203
204 if (IS_ENABLED(CONFIG_CMA_SIZE_PERNUMA) &&
205 !numa_cma_configured && dma_contiguous_default_area &&
206 nr_online_nodes > 1)
207 pernuma_size_bytes = cma_get_size(dma_contiguous_default_area);
208
209 for_each_node(nid) {
210 int size, ret;
211 char name[CMA_MAX_NAME];
212 struct cma **cma;
213
214 if (!node_online(nid)) {
215 if (pernuma_size_bytes || numa_cma_size[nid])
216 pr_warn("invalid node %d specified\n", nid);
217 continue;
218 }
219
220 /* per-node numa setting has the priority */
221 size = numa_cma_size[nid] ?: pernuma_size_bytes;
222 if (!size)
223 continue;
224
225 cma = &dma_contiguous_numa_area[nid];
226 snprintf(name, sizeof(name), "numa%d", nid);
227 ret = cma_declare_contiguous_nid(0, size, 0, 0, 0, false, name, cma, nid);
228 if (ret)
229 pr_warn("%s: reservation failed: err %d, node %d", __func__,
230 ret, nid);
231 }
232}
233#else
234static inline void __init dma_numa_cma_reserve(void)
235{
236}
237#endif
238
239/**
240 * dma_contiguous_reserve() - reserve area(s) for contiguous memory handling
241 * @limit: End address of the reserved memory (optional, 0 for any).
242 *
243 * This function reserves memory from early allocator. It should be
244 * called by arch specific code once the early allocator (memblock or bootmem)
245 * has been activated and all other subsystems have already allocated/reserved
246 * memory.
247 */
248void __init dma_contiguous_reserve(phys_addr_t limit)
249{
250 phys_addr_t selected_size = 0;
251 phys_addr_t selected_base = 0;
252 phys_addr_t selected_limit = limit;
253 bool fixed = false;
254
255 pr_debug("%s(limit %08lx)\n", __func__, (unsigned long)limit);
256
257 if (size_cmdline != -1) {
258 selected_size = size_cmdline;
259 selected_base = base_cmdline;
260
261 /* Hornor the user setup dma address limit */
262 selected_limit = limit_cmdline ?: limit;
263
264 if (base_cmdline + size_cmdline == limit_cmdline)
265 fixed = true;
266 } else {
267#ifdef CONFIG_CMA_SIZE_SEL_MBYTES
268 selected_size = size_bytes;
269#elif defined(CONFIG_CMA_SIZE_SEL_PERCENTAGE)
270 selected_size = cma_early_percent_memory();
271#elif defined(CONFIG_CMA_SIZE_SEL_MIN)
272 selected_size = min(size_bytes, cma_early_percent_memory());
273#elif defined(CONFIG_CMA_SIZE_SEL_MAX)
274 selected_size = max(size_bytes, cma_early_percent_memory());
275#endif
276 }
277
278 if (selected_size && !dma_contiguous_default_area) {
279 int ret;
280
281 pr_debug("%s: reserving %ld MiB for global area\n", __func__,
282 (unsigned long)selected_size / SZ_1M);
283
284 ret = dma_contiguous_reserve_area(selected_size, selected_base,
285 selected_limit,
286 &dma_contiguous_default_area,
287 fixed);
288 if (ret)
289 return;
290
291 /*
292 * We need to insert the new area in our list to avoid
293 * any inconsistencies between having the default area
294 * listed in the DT or not.
295 *
296 * The DT case is handled by rmem_cma_setup() and will
297 * always insert all its areas in our list. However, if
298 * it didn't run (because OF_RESERVED_MEM isn't set, or
299 * there's no DT region specified), then we don't have a
300 * default area yet, and no area in our list.
301 *
302 * This block creates the default area in such a case,
303 * but we also need to insert it in our list to avoid
304 * having a default area but an empty list.
305 */
306 ret = dma_contiguous_insert_area(dma_contiguous_default_area);
307 if (ret)
308 pr_warn("Couldn't queue default CMA region for heap creation.");
309 }
310
311 dma_numa_cma_reserve();
312}
313
314void __weak
315dma_contiguous_early_fixup(phys_addr_t base, unsigned long size)
316{
317}
318
319/**
320 * dma_contiguous_reserve_area() - reserve custom contiguous area
321 * @size: Size of the reserved area (in bytes),
322 * @base: Base address of the reserved area optional, use 0 for any
323 * @limit: End address of the reserved memory (optional, 0 for any).
324 * @res_cma: Pointer to store the created cma region.
325 * @fixed: hint about where to place the reserved area
326 *
327 * This function reserves memory from early allocator. It should be
328 * called by arch specific code once the early allocator (memblock or bootmem)
329 * has been activated and all other subsystems have already allocated/reserved
330 * memory. This function allows to create custom reserved areas for specific
331 * devices.
332 *
333 * If @fixed is true, reserve contiguous area at exactly @base. If false,
334 * reserve in range from @base to @limit.
335 */
336int __init dma_contiguous_reserve_area(phys_addr_t size, phys_addr_t base,
337 phys_addr_t limit, struct cma **res_cma,
338 bool fixed)
339{
340 int ret;
341
342 ret = cma_declare_contiguous(base, size, limit, 0, 0, fixed,
343 "reserved", res_cma);
344 if (ret)
345 return ret;
346
347 /* Architecture specific contiguous memory fixup. */
348 dma_contiguous_early_fixup(cma_get_base(*res_cma),
349 cma_get_size(*res_cma));
350
351 return 0;
352}
353
354/**
355 * dma_alloc_from_contiguous() - allocate pages from contiguous area
356 * @dev: Pointer to device for which the allocation is performed.
357 * @count: Requested number of pages.
358 * @align: Requested alignment of pages (in PAGE_SIZE order).
359 * @no_warn: Avoid printing message about failed allocation.
360 *
361 * This function allocates memory buffer for specified device. It uses
362 * device specific contiguous memory area if available or the default
363 * global one. Requires architecture specific dev_get_cma_area() helper
364 * function.
365 */
366struct page *dma_alloc_from_contiguous(struct device *dev, size_t count,
367 unsigned int align, bool no_warn)
368{
369 if (align > CONFIG_CMA_ALIGNMENT)
370 align = CONFIG_CMA_ALIGNMENT;
371
372 return cma_alloc(dev_get_cma_area(dev), count, align, no_warn);
373}
374
375/**
376 * dma_release_from_contiguous() - release allocated pages
377 * @dev: Pointer to device for which the pages were allocated.
378 * @pages: Allocated pages.
379 * @count: Number of allocated pages.
380 *
381 * This function releases memory allocated by dma_alloc_from_contiguous().
382 * It returns false when provided pages do not belong to contiguous area and
383 * true otherwise.
384 */
385bool dma_release_from_contiguous(struct device *dev, struct page *pages,
386 int count)
387{
388 return cma_release(dev_get_cma_area(dev), pages, count);
389}
390
391static struct page *cma_alloc_aligned(struct cma *cma, size_t size, gfp_t gfp)
392{
393 unsigned int align = min(get_order(size), CONFIG_CMA_ALIGNMENT);
394
395 return cma_alloc(cma, size >> PAGE_SHIFT, align, gfp & __GFP_NOWARN);
396}
397
398/**
399 * dma_alloc_contiguous() - allocate contiguous pages
400 * @dev: Pointer to device for which the allocation is performed.
401 * @size: Requested allocation size.
402 * @gfp: Allocation flags.
403 *
404 * tries to use device specific contiguous memory area if available, or it
405 * tries to use per-numa cma, if the allocation fails, it will fallback to
406 * try default global one.
407 *
408 * Note that it bypass one-page size of allocations from the per-numa and
409 * global area as the addresses within one page are always contiguous, so
410 * there is no need to waste CMA pages for that kind; it also helps reduce
411 * fragmentations.
412 */
413struct page *dma_alloc_contiguous(struct device *dev, size_t size, gfp_t gfp)
414{
415#ifdef CONFIG_DMA_NUMA_CMA
416 int nid = dev_to_node(dev);
417#endif
418
419 /* CMA can be used only in the context which permits sleeping */
420 if (!gfpflags_allow_blocking(gfp))
421 return NULL;
422 if (dev->cma_area)
423 return cma_alloc_aligned(dev->cma_area, size, gfp);
424 if (size <= PAGE_SIZE)
425 return NULL;
426
427#ifdef CONFIG_DMA_NUMA_CMA
428 if (nid != NUMA_NO_NODE && !(gfp & (GFP_DMA | GFP_DMA32))) {
429 struct cma *cma = dma_contiguous_numa_area[nid];
430 struct page *page;
431 if (cma) {
432 page = cma_alloc_aligned(cma, size, gfp);
433 if (page)
434 return page;
435 }
436 }
437#endif
438 if (!dma_contiguous_default_area)
439 return NULL;
440
441 return cma_alloc_aligned(dma_contiguous_default_area, size, gfp);
442}
443
444/**
445 * dma_free_contiguous() - release allocated pages
446 * @dev: Pointer to device for which the pages were allocated.
447 * @page: Pointer to the allocated pages.
448 * @size: Size of allocated pages.
449 *
450 * This function releases memory allocated by dma_alloc_contiguous(). As the
451 * cma_release returns false when provided pages do not belong to contiguous
452 * area and true otherwise, this function then does a fallback __free_pages()
453 * upon a false-return.
454 */
455void dma_free_contiguous(struct device *dev, struct page *page, size_t size)
456{
457 unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT;
458
459 /* if dev has its own cma, free page from there */
460 if (dev->cma_area) {
461 if (cma_release(dev->cma_area, page, count))
462 return;
463 } else {
464 /*
465 * otherwise, page is from either per-numa cma or default cma
466 */
467#ifdef CONFIG_DMA_NUMA_CMA
468 if (cma_release(dma_contiguous_numa_area[page_to_nid(page)],
469 page, count))
470 return;
471#endif
472 if (cma_release(dma_contiguous_default_area, page, count))
473 return;
474 }
475
476 /* not in any cma, free from buddy */
477 __free_pages(page, get_order(size));
478}
479
480/*
481 * Support for reserved memory regions defined in device tree
482 */
483#ifdef CONFIG_OF_RESERVED_MEM
484#include <linux/of.h>
485#include <linux/of_fdt.h>
486#include <linux/of_reserved_mem.h>
487
488#undef pr_fmt
489#define pr_fmt(fmt) fmt
490
491static int rmem_cma_device_init(struct reserved_mem *rmem, struct device *dev)
492{
493 dev->cma_area = rmem->priv;
494 return 0;
495}
496
497static void rmem_cma_device_release(struct reserved_mem *rmem,
498 struct device *dev)
499{
500 dev->cma_area = NULL;
501}
502
503static int __init __rmem_cma_verify_node(unsigned long node)
504{
505 if (!of_get_flat_dt_prop(node, "reusable", NULL) ||
506 of_get_flat_dt_prop(node, "no-map", NULL))
507 return -ENODEV;
508
509 if (size_cmdline != -1 &&
510 of_get_flat_dt_prop(node, "linux,cma-default", NULL)) {
511 pr_err("Skipping dt linux,cma-default node in favor for \"cma=\" kernel param.\n");
512 return -EBUSY;
513 }
514 return 0;
515}
516
517static int __init rmem_cma_validate(unsigned long node, phys_addr_t *align)
518{
519 int ret = __rmem_cma_verify_node(node);
520
521 if (ret)
522 return ret;
523
524 if (align)
525 *align = max_t(phys_addr_t, *align, CMA_MIN_ALIGNMENT_BYTES);
526
527 return 0;
528}
529
530static int __init rmem_cma_fixup(unsigned long node, phys_addr_t base,
531 phys_addr_t size)
532{
533 int ret = __rmem_cma_verify_node(node);
534
535 if (ret)
536 return ret;
537
538 /* Architecture specific contiguous memory fixup. */
539 dma_contiguous_early_fixup(base, size);
540 return 0;
541}
542
543static int __init rmem_cma_setup(unsigned long node, struct reserved_mem *rmem)
544{
545 bool default_cma = of_get_flat_dt_prop(node, "linux,cma-default", NULL);
546 struct cma *cma;
547 int ret;
548
549 ret = __rmem_cma_verify_node(node);
550 if (ret)
551 return ret;
552
553 if (!IS_ALIGNED(rmem->base | rmem->size, CMA_MIN_ALIGNMENT_BYTES)) {
554 pr_err("Reserved memory: incorrect alignment of CMA region\n");
555 return -EINVAL;
556 }
557
558 ret = cma_init_reserved_mem(rmem->base, rmem->size, 0, rmem->name, &cma);
559 if (ret) {
560 pr_err("Reserved memory: unable to setup CMA region\n");
561 return ret;
562 }
563
564 if (default_cma)
565 dma_contiguous_default_area = cma;
566
567 rmem->priv = cma;
568
569 pr_info("Reserved memory: created CMA memory pool at %pa, size %ld MiB\n",
570 &rmem->base, (unsigned long)rmem->size / SZ_1M);
571
572 ret = dma_contiguous_insert_area(cma);
573 if (ret)
574 pr_warn("Couldn't store CMA reserved area.");
575
576 return 0;
577}
578
579static const struct reserved_mem_ops rmem_cma_ops = {
580 .node_validate = rmem_cma_validate,
581 .node_fixup = rmem_cma_fixup,
582 .node_init = rmem_cma_setup,
583 .device_init = rmem_cma_device_init,
584 .device_release = rmem_cma_device_release,
585};
586
587RESERVEDMEM_OF_DECLARE(cma, "shared-dma-pool", &rmem_cma_ops);
588#endif