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 <m.szyprowski@samsung.com> |
7 | * Michal Nazarewicz <mina86@mina86.com> |
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 | |
55 | struct cma *dma_contiguous_default_area; |
56 | |
57 | /* |
58 | * Default global CMA area size can be defined in kernel's .config. |
59 | * This is useful mainly for distro maintainers to create a kernel |
60 | * that works correctly for most supported systems. |
61 | * The size can be set in bytes or as a percentage of the total memory |
62 | * in the system. |
63 | * |
64 | * Users, who want to set the size of global CMA area for their system |
65 | * should use cma= kernel parameter. |
66 | */ |
67 | #define size_bytes ((phys_addr_t)CMA_SIZE_MBYTES * SZ_1M) |
68 | static phys_addr_t size_cmdline __initdata = -1; |
69 | static phys_addr_t base_cmdline __initdata; |
70 | static phys_addr_t limit_cmdline __initdata; |
71 | |
72 | static int __init early_cma(char *p) |
73 | { |
74 | if (!p) { |
75 | pr_err("Config string not provided\n"); |
76 | return -EINVAL; |
77 | } |
78 | |
79 | size_cmdline = memparse(ptr: p, retptr: &p); |
80 | if (*p != '@') |
81 | return 0; |
82 | base_cmdline = memparse(ptr: p + 1, retptr: &p); |
83 | if (*p != '-') { |
84 | limit_cmdline = base_cmdline + size_cmdline; |
85 | return 0; |
86 | } |
87 | limit_cmdline = memparse(ptr: p + 1, retptr: &p); |
88 | |
89 | return 0; |
90 | } |
91 | early_param("cma", early_cma); |
92 | |
93 | #ifdef CONFIG_DMA_NUMA_CMA |
94 | |
95 | static struct cma *dma_contiguous_numa_area[MAX_NUMNODES]; |
96 | static phys_addr_t numa_cma_size[MAX_NUMNODES] __initdata; |
97 | static struct cma *dma_contiguous_pernuma_area[MAX_NUMNODES]; |
98 | static phys_addr_t pernuma_size_bytes __initdata; |
99 | |
100 | static int __init early_numa_cma(char *p) |
101 | { |
102 | int nid, count = 0; |
103 | unsigned long tmp; |
104 | char *s = p; |
105 | |
106 | while (*s) { |
107 | if (sscanf(s, "%lu%n", &tmp, &count) != 1) |
108 | break; |
109 | |
110 | if (s[count] == ':') { |
111 | if (tmp >= MAX_NUMNODES) |
112 | break; |
113 | nid = array_index_nospec(tmp, MAX_NUMNODES); |
114 | |
115 | s += count + 1; |
116 | tmp = memparse(ptr: s, retptr: &s); |
117 | numa_cma_size[nid] = tmp; |
118 | |
119 | if (*s == ',') |
120 | s++; |
121 | else |
122 | break; |
123 | } else |
124 | break; |
125 | } |
126 | |
127 | return 0; |
128 | } |
129 | early_param("numa_cma", early_numa_cma); |
130 | |
131 | static int __init early_cma_pernuma(char *p) |
132 | { |
133 | pernuma_size_bytes = memparse(ptr: p, retptr: &p); |
134 | return 0; |
135 | } |
136 | early_param("cma_pernuma", early_cma_pernuma); |
137 | #endif |
138 | |
139 | #ifdef CONFIG_CMA_SIZE_PERCENTAGE |
140 | |
141 | static phys_addr_t __init __maybe_unused cma_early_percent_memory(void) |
142 | { |
143 | unsigned long total_pages = PHYS_PFN(memblock_phys_mem_size()); |
144 | |
145 | return (total_pages * CONFIG_CMA_SIZE_PERCENTAGE / 100) << PAGE_SHIFT; |
146 | } |
147 | |
148 | #else |
149 | |
150 | static inline __maybe_unused phys_addr_t cma_early_percent_memory(void) |
151 | { |
152 | return 0; |
153 | } |
154 | |
155 | #endif |
156 | |
157 | #ifdef CONFIG_DMA_NUMA_CMA |
158 | static void __init dma_numa_cma_reserve(void) |
159 | { |
160 | int nid; |
161 | |
162 | for_each_node(nid) { |
163 | int ret; |
164 | char name[CMA_MAX_NAME]; |
165 | struct cma **cma; |
166 | |
167 | if (!node_online(nid)) { |
168 | if (pernuma_size_bytes || numa_cma_size[nid]) |
169 | pr_warn("invalid node %d specified\n", nid); |
170 | continue; |
171 | } |
172 | |
173 | if (pernuma_size_bytes) { |
174 | |
175 | cma = &dma_contiguous_pernuma_area[nid]; |
176 | snprintf(buf: name, size: sizeof(name), fmt: "pernuma%d", nid); |
177 | ret = cma_declare_contiguous_nid(base: 0, size: pernuma_size_bytes, limit: 0, alignment: 0, |
178 | order_per_bit: 0, fixed: false, name, res_cma: cma, nid); |
179 | if (ret) |
180 | pr_warn("%s: reservation failed: err %d, node %d", __func__, |
181 | ret, nid); |
182 | } |
183 | |
184 | if (numa_cma_size[nid]) { |
185 | |
186 | cma = &dma_contiguous_numa_area[nid]; |
187 | snprintf(buf: name, size: sizeof(name), fmt: "numa%d", nid); |
188 | ret = cma_declare_contiguous_nid(base: 0, size: numa_cma_size[nid], limit: 0, alignment: 0, order_per_bit: 0, fixed: false, |
189 | name, res_cma: cma, nid); |
190 | if (ret) |
191 | pr_warn("%s: reservation failed: err %d, node %d", __func__, |
192 | ret, nid); |
193 | } |
194 | } |
195 | } |
196 | #else |
197 | static inline void __init dma_numa_cma_reserve(void) |
198 | { |
199 | } |
200 | #endif |
201 | |
202 | /** |
203 | * dma_contiguous_reserve() - reserve area(s) for contiguous memory handling |
204 | * @limit: End address of the reserved memory (optional, 0 for any). |
205 | * |
206 | * This function reserves memory from early allocator. It should be |
207 | * called by arch specific code once the early allocator (memblock or bootmem) |
208 | * has been activated and all other subsystems have already allocated/reserved |
209 | * memory. |
210 | */ |
211 | void __init dma_contiguous_reserve(phys_addr_t limit) |
212 | { |
213 | phys_addr_t selected_size = 0; |
214 | phys_addr_t selected_base = 0; |
215 | phys_addr_t selected_limit = limit; |
216 | bool fixed = false; |
217 | |
218 | dma_numa_cma_reserve(); |
219 | |
220 | pr_debug("%s(limit %08lx)\n", __func__, (unsigned long)limit); |
221 | |
222 | if (size_cmdline != -1) { |
223 | selected_size = size_cmdline; |
224 | selected_base = base_cmdline; |
225 | selected_limit = min_not_zero(limit_cmdline, limit); |
226 | if (base_cmdline + size_cmdline == limit_cmdline) |
227 | fixed = true; |
228 | } else { |
229 | #ifdef CONFIG_CMA_SIZE_SEL_MBYTES |
230 | selected_size = size_bytes; |
231 | #elif defined(CONFIG_CMA_SIZE_SEL_PERCENTAGE) |
232 | selected_size = cma_early_percent_memory(); |
233 | #elif defined(CONFIG_CMA_SIZE_SEL_MIN) |
234 | selected_size = min(size_bytes, cma_early_percent_memory()); |
235 | #elif defined(CONFIG_CMA_SIZE_SEL_MAX) |
236 | selected_size = max(size_bytes, cma_early_percent_memory()); |
237 | #endif |
238 | } |
239 | |
240 | if (selected_size && !dma_contiguous_default_area) { |
241 | pr_debug("%s: reserving %ld MiB for global area\n", __func__, |
242 | (unsigned long)selected_size / SZ_1M); |
243 | |
244 | dma_contiguous_reserve_area(size: selected_size, base: selected_base, |
245 | limit: selected_limit, |
246 | res_cma: &dma_contiguous_default_area, |
247 | fixed); |
248 | } |
249 | } |
250 | |
251 | void __weak |
252 | dma_contiguous_early_fixup(phys_addr_t base, unsigned long size) |
253 | { |
254 | } |
255 | |
256 | /** |
257 | * dma_contiguous_reserve_area() - reserve custom contiguous area |
258 | * @size: Size of the reserved area (in bytes), |
259 | * @base: Base address of the reserved area optional, use 0 for any |
260 | * @limit: End address of the reserved memory (optional, 0 for any). |
261 | * @res_cma: Pointer to store the created cma region. |
262 | * @fixed: hint about where to place the reserved area |
263 | * |
264 | * This function reserves memory from early allocator. It should be |
265 | * called by arch specific code once the early allocator (memblock or bootmem) |
266 | * has been activated and all other subsystems have already allocated/reserved |
267 | * memory. This function allows to create custom reserved areas for specific |
268 | * devices. |
269 | * |
270 | * If @fixed is true, reserve contiguous area at exactly @base. If false, |
271 | * reserve in range from @base to @limit. |
272 | */ |
273 | int __init dma_contiguous_reserve_area(phys_addr_t size, phys_addr_t base, |
274 | phys_addr_t limit, struct cma **res_cma, |
275 | bool fixed) |
276 | { |
277 | int ret; |
278 | |
279 | ret = cma_declare_contiguous(base, size, limit, alignment: 0, order_per_bit: 0, fixed, |
280 | name: "reserved", res_cma); |
281 | if (ret) |
282 | return ret; |
283 | |
284 | /* Architecture specific contiguous memory fixup. */ |
285 | dma_contiguous_early_fixup(base: cma_get_base(cma: *res_cma), |
286 | size: cma_get_size(cma: *res_cma)); |
287 | |
288 | return 0; |
289 | } |
290 | |
291 | /** |
292 | * dma_alloc_from_contiguous() - allocate pages from contiguous area |
293 | * @dev: Pointer to device for which the allocation is performed. |
294 | * @count: Requested number of pages. |
295 | * @align: Requested alignment of pages (in PAGE_SIZE order). |
296 | * @no_warn: Avoid printing message about failed allocation. |
297 | * |
298 | * This function allocates memory buffer for specified device. It uses |
299 | * device specific contiguous memory area if available or the default |
300 | * global one. Requires architecture specific dev_get_cma_area() helper |
301 | * function. |
302 | */ |
303 | struct page *dma_alloc_from_contiguous(struct device *dev, size_t count, |
304 | unsigned int align, bool no_warn) |
305 | { |
306 | if (align > CONFIG_CMA_ALIGNMENT) |
307 | align = CONFIG_CMA_ALIGNMENT; |
308 | |
309 | return cma_alloc(cma: dev_get_cma_area(dev), count, align, no_warn); |
310 | } |
311 | |
312 | /** |
313 | * dma_release_from_contiguous() - release allocated pages |
314 | * @dev: Pointer to device for which the pages were allocated. |
315 | * @pages: Allocated pages. |
316 | * @count: Number of allocated pages. |
317 | * |
318 | * This function releases memory allocated by dma_alloc_from_contiguous(). |
319 | * It returns false when provided pages do not belong to contiguous area and |
320 | * true otherwise. |
321 | */ |
322 | bool dma_release_from_contiguous(struct device *dev, struct page *pages, |
323 | int count) |
324 | { |
325 | return cma_release(cma: dev_get_cma_area(dev), pages, count); |
326 | } |
327 | |
328 | static struct page *cma_alloc_aligned(struct cma *cma, size_t size, gfp_t gfp) |
329 | { |
330 | unsigned int align = min(get_order(size), CONFIG_CMA_ALIGNMENT); |
331 | |
332 | return cma_alloc(cma, count: size >> PAGE_SHIFT, align, no_warn: gfp & __GFP_NOWARN); |
333 | } |
334 | |
335 | /** |
336 | * dma_alloc_contiguous() - allocate contiguous pages |
337 | * @dev: Pointer to device for which the allocation is performed. |
338 | * @size: Requested allocation size. |
339 | * @gfp: Allocation flags. |
340 | * |
341 | * tries to use device specific contiguous memory area if available, or it |
342 | * tries to use per-numa cma, if the allocation fails, it will fallback to |
343 | * try default global one. |
344 | * |
345 | * Note that it bypass one-page size of allocations from the per-numa and |
346 | * global area as the addresses within one page are always contiguous, so |
347 | * there is no need to waste CMA pages for that kind; it also helps reduce |
348 | * fragmentations. |
349 | */ |
350 | struct page *dma_alloc_contiguous(struct device *dev, size_t size, gfp_t gfp) |
351 | { |
352 | #ifdef CONFIG_DMA_NUMA_CMA |
353 | int nid = dev_to_node(dev); |
354 | #endif |
355 | |
356 | /* CMA can be used only in the context which permits sleeping */ |
357 | if (!gfpflags_allow_blocking(gfp_flags: gfp)) |
358 | return NULL; |
359 | if (dev->cma_area) |
360 | return cma_alloc_aligned(cma: dev->cma_area, size, gfp); |
361 | if (size <= PAGE_SIZE) |
362 | return NULL; |
363 | |
364 | #ifdef CONFIG_DMA_NUMA_CMA |
365 | if (nid != NUMA_NO_NODE && !(gfp & (GFP_DMA | GFP_DMA32))) { |
366 | struct cma *cma = dma_contiguous_pernuma_area[nid]; |
367 | struct page *page; |
368 | |
369 | if (cma) { |
370 | page = cma_alloc_aligned(cma, size, gfp); |
371 | if (page) |
372 | return page; |
373 | } |
374 | |
375 | cma = dma_contiguous_numa_area[nid]; |
376 | if (cma) { |
377 | page = cma_alloc_aligned(cma, size, gfp); |
378 | if (page) |
379 | return page; |
380 | } |
381 | } |
382 | #endif |
383 | if (!dma_contiguous_default_area) |
384 | return NULL; |
385 | |
386 | return cma_alloc_aligned(cma: dma_contiguous_default_area, size, gfp); |
387 | } |
388 | |
389 | /** |
390 | * dma_free_contiguous() - release allocated pages |
391 | * @dev: Pointer to device for which the pages were allocated. |
392 | * @page: Pointer to the allocated pages. |
393 | * @size: Size of allocated pages. |
394 | * |
395 | * This function releases memory allocated by dma_alloc_contiguous(). As the |
396 | * cma_release returns false when provided pages do not belong to contiguous |
397 | * area and true otherwise, this function then does a fallback __free_pages() |
398 | * upon a false-return. |
399 | */ |
400 | void dma_free_contiguous(struct device *dev, struct page *page, size_t size) |
401 | { |
402 | unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT; |
403 | |
404 | /* if dev has its own cma, free page from there */ |
405 | if (dev->cma_area) { |
406 | if (cma_release(cma: dev->cma_area, pages: page, count)) |
407 | return; |
408 | } else { |
409 | /* |
410 | * otherwise, page is from either per-numa cma or default cma |
411 | */ |
412 | #ifdef CONFIG_DMA_NUMA_CMA |
413 | if (cma_release(cma: dma_contiguous_pernuma_area[page_to_nid(page)], |
414 | pages: page, count)) |
415 | return; |
416 | if (cma_release(cma: dma_contiguous_numa_area[page_to_nid(page)], |
417 | pages: page, count)) |
418 | return; |
419 | #endif |
420 | if (cma_release(cma: dma_contiguous_default_area, pages: page, count)) |
421 | return; |
422 | } |
423 | |
424 | /* not in any cma, free from buddy */ |
425 | __free_pages(page, order: get_order(size)); |
426 | } |
427 | |
428 | /* |
429 | * Support for reserved memory regions defined in device tree |
430 | */ |
431 | #ifdef CONFIG_OF_RESERVED_MEM |
432 | #include <linux/of.h> |
433 | #include <linux/of_fdt.h> |
434 | #include <linux/of_reserved_mem.h> |
435 | |
436 | #undef pr_fmt |
437 | #define pr_fmt(fmt) fmt |
438 | |
439 | static int rmem_cma_device_init(struct reserved_mem *rmem, struct device *dev) |
440 | { |
441 | dev->cma_area = rmem->priv; |
442 | return 0; |
443 | } |
444 | |
445 | static void rmem_cma_device_release(struct reserved_mem *rmem, |
446 | struct device *dev) |
447 | { |
448 | dev->cma_area = NULL; |
449 | } |
450 | |
451 | static const struct reserved_mem_ops rmem_cma_ops = { |
452 | .device_init = rmem_cma_device_init, |
453 | .device_release = rmem_cma_device_release, |
454 | }; |
455 | |
456 | static int __init rmem_cma_setup(struct reserved_mem *rmem) |
457 | { |
458 | unsigned long node = rmem->fdt_node; |
459 | bool default_cma = of_get_flat_dt_prop(node, name: "linux,cma-default", NULL); |
460 | struct cma *cma; |
461 | int err; |
462 | |
463 | if (size_cmdline != -1 && default_cma) { |
464 | pr_info("Reserved memory: bypass %s node, using cmdline CMA params instead\n", |
465 | rmem->name); |
466 | return -EBUSY; |
467 | } |
468 | |
469 | if (!of_get_flat_dt_prop(node, name: "reusable", NULL) || |
470 | of_get_flat_dt_prop(node, name: "no-map", NULL)) |
471 | return -EINVAL; |
472 | |
473 | if (!IS_ALIGNED(rmem->base | rmem->size, CMA_MIN_ALIGNMENT_BYTES)) { |
474 | pr_err("Reserved memory: incorrect alignment of CMA region\n"); |
475 | return -EINVAL; |
476 | } |
477 | |
478 | err = cma_init_reserved_mem(base: rmem->base, size: rmem->size, order_per_bit: 0, name: rmem->name, res_cma: &cma); |
479 | if (err) { |
480 | pr_err("Reserved memory: unable to setup CMA region\n"); |
481 | return err; |
482 | } |
483 | /* Architecture specific contiguous memory fixup. */ |
484 | dma_contiguous_early_fixup(base: rmem->base, size: rmem->size); |
485 | |
486 | if (default_cma) |
487 | dma_contiguous_default_area = cma; |
488 | |
489 | rmem->ops = &rmem_cma_ops; |
490 | rmem->priv = cma; |
491 | |
492 | pr_info("Reserved memory: created CMA memory pool at %pa, size %ld MiB\n", |
493 | &rmem->base, (unsigned long)rmem->size / SZ_1M); |
494 | |
495 | return 0; |
496 | } |
497 | RESERVEDMEM_OF_DECLARE(cma, "shared-dma-pool", rmem_cma_setup); |
498 | #endif |
499 |
Definitions
- dma_contiguous_default_area
- size_cmdline
- base_cmdline
- limit_cmdline
- early_cma
- dma_contiguous_numa_area
- numa_cma_size
- dma_contiguous_pernuma_area
- pernuma_size_bytes
- early_numa_cma
- early_cma_pernuma
- cma_early_percent_memory
- dma_numa_cma_reserve
- dma_contiguous_reserve
- dma_contiguous_early_fixup
- dma_contiguous_reserve_area
- dma_alloc_from_contiguous
- dma_release_from_contiguous
- cma_alloc_aligned
- dma_alloc_contiguous
- dma_free_contiguous
- rmem_cma_device_init
- rmem_cma_device_release
- rmem_cma_ops
Improve your Profiling and Debugging skills
Find out more