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 | static const phys_addr_t size_bytes __initconst = |
68 | (phys_addr_t)CMA_SIZE_MBYTES * SZ_1M; |
69 | static phys_addr_t size_cmdline __initdata = -1; |
70 | static phys_addr_t base_cmdline __initdata; |
71 | static phys_addr_t limit_cmdline __initdata; |
72 | |
73 | static int __init early_cma(char *p) |
74 | { |
75 | if (!p) { |
76 | pr_err("Config string not provided\n"); |
77 | return -EINVAL; |
78 | } |
79 | |
80 | size_cmdline = memparse(ptr: p, retptr: &p); |
81 | if (*p != '@') |
82 | return 0; |
83 | base_cmdline = memparse(ptr: p + 1, retptr: &p); |
84 | if (*p != '-') { |
85 | limit_cmdline = base_cmdline + size_cmdline; |
86 | return 0; |
87 | } |
88 | limit_cmdline = memparse(ptr: p + 1, retptr: &p); |
89 | |
90 | return 0; |
91 | } |
92 | early_param("cma", early_cma); |
93 | |
94 | #ifdef CONFIG_DMA_NUMA_CMA |
95 | |
96 | static struct cma *dma_contiguous_numa_area[MAX_NUMNODES]; |
97 | static phys_addr_t numa_cma_size[MAX_NUMNODES] __initdata; |
98 | static struct cma *dma_contiguous_pernuma_area[MAX_NUMNODES]; |
99 | static phys_addr_t pernuma_size_bytes __initdata; |
100 | |
101 | static int __init early_numa_cma(char *p) |
102 | { |
103 | int nid, count = 0; |
104 | unsigned long tmp; |
105 | char *s = p; |
106 | |
107 | while (*s) { |
108 | if (sscanf(s, "%lu%n", &tmp, &count) != 1) |
109 | break; |
110 | |
111 | if (s[count] == ':') { |
112 | if (tmp >= MAX_NUMNODES) |
113 | break; |
114 | nid = array_index_nospec(tmp, MAX_NUMNODES); |
115 | |
116 | s += count + 1; |
117 | tmp = memparse(ptr: s, retptr: &s); |
118 | numa_cma_size[nid] = tmp; |
119 | |
120 | if (*s == ',') |
121 | s++; |
122 | else |
123 | break; |
124 | } else |
125 | break; |
126 | } |
127 | |
128 | return 0; |
129 | } |
130 | early_param("numa_cma", early_numa_cma); |
131 | |
132 | static int __init early_cma_pernuma(char *p) |
133 | { |
134 | pernuma_size_bytes = memparse(ptr: p, retptr: &p); |
135 | return 0; |
136 | } |
137 | early_param("cma_pernuma", early_cma_pernuma); |
138 | #endif |
139 | |
140 | #ifdef CONFIG_CMA_SIZE_PERCENTAGE |
141 | |
142 | static phys_addr_t __init __maybe_unused cma_early_percent_memory(void) |
143 | { |
144 | unsigned long total_pages = PHYS_PFN(memblock_phys_mem_size()); |
145 | |
146 | return (total_pages * CONFIG_CMA_SIZE_PERCENTAGE / 100) << PAGE_SHIFT; |
147 | } |
148 | |
149 | #else |
150 | |
151 | static inline __maybe_unused phys_addr_t cma_early_percent_memory(void) |
152 | { |
153 | return 0; |
154 | } |
155 | |
156 | #endif |
157 | |
158 | #ifdef CONFIG_DMA_NUMA_CMA |
159 | static void __init dma_numa_cma_reserve(void) |
160 | { |
161 | int nid; |
162 | |
163 | for_each_node(nid) { |
164 | int ret; |
165 | char name[CMA_MAX_NAME]; |
166 | struct cma **cma; |
167 | |
168 | if (!node_online(nid)) { |
169 | if (pernuma_size_bytes || numa_cma_size[nid]) |
170 | pr_warn("invalid node %d specified\n", nid); |
171 | continue; |
172 | } |
173 | |
174 | if (pernuma_size_bytes) { |
175 | |
176 | cma = &dma_contiguous_pernuma_area[nid]; |
177 | snprintf(buf: name, size: sizeof(name), fmt: "pernuma%d", nid); |
178 | ret = cma_declare_contiguous_nid(base: 0, size: pernuma_size_bytes, limit: 0, alignment: 0, |
179 | order_per_bit: 0, fixed: false, name, res_cma: cma, nid); |
180 | if (ret) |
181 | pr_warn("%s: reservation failed: err %d, node %d", __func__, |
182 | ret, nid); |
183 | } |
184 | |
185 | if (numa_cma_size[nid]) { |
186 | |
187 | cma = &dma_contiguous_numa_area[nid]; |
188 | snprintf(buf: name, size: sizeof(name), fmt: "numa%d", nid); |
189 | ret = cma_declare_contiguous_nid(base: 0, size: numa_cma_size[nid], limit: 0, alignment: 0, order_per_bit: 0, fixed: false, |
190 | name, res_cma: cma, nid); |
191 | if (ret) |
192 | pr_warn("%s: reservation failed: err %d, node %d", __func__, |
193 | ret, nid); |
194 | } |
195 | } |
196 | } |
197 | #else |
198 | static inline void __init dma_numa_cma_reserve(void) |
199 | { |
200 | } |
201 | #endif |
202 | |
203 | /** |
204 | * dma_contiguous_reserve() - reserve area(s) for contiguous memory handling |
205 | * @limit: End address of the reserved memory (optional, 0 for any). |
206 | * |
207 | * This function reserves memory from early allocator. It should be |
208 | * called by arch specific code once the early allocator (memblock or bootmem) |
209 | * has been activated and all other subsystems have already allocated/reserved |
210 | * memory. |
211 | */ |
212 | void __init dma_contiguous_reserve(phys_addr_t limit) |
213 | { |
214 | phys_addr_t selected_size = 0; |
215 | phys_addr_t selected_base = 0; |
216 | phys_addr_t selected_limit = limit; |
217 | bool fixed = false; |
218 | |
219 | dma_numa_cma_reserve(); |
220 | |
221 | pr_debug("%s(limit %08lx)\n", __func__, (unsigned long)limit); |
222 | |
223 | if (size_cmdline != -1) { |
224 | selected_size = size_cmdline; |
225 | selected_base = base_cmdline; |
226 | selected_limit = min_not_zero(limit_cmdline, limit); |
227 | if (base_cmdline + size_cmdline == limit_cmdline) |
228 | fixed = true; |
229 | } else { |
230 | #ifdef CONFIG_CMA_SIZE_SEL_MBYTES |
231 | selected_size = size_bytes; |
232 | #elif defined(CONFIG_CMA_SIZE_SEL_PERCENTAGE) |
233 | selected_size = cma_early_percent_memory(); |
234 | #elif defined(CONFIG_CMA_SIZE_SEL_MIN) |
235 | selected_size = min(size_bytes, cma_early_percent_memory()); |
236 | #elif defined(CONFIG_CMA_SIZE_SEL_MAX) |
237 | selected_size = max(size_bytes, cma_early_percent_memory()); |
238 | #endif |
239 | } |
240 | |
241 | if (selected_size && !dma_contiguous_default_area) { |
242 | pr_debug("%s: reserving %ld MiB for global area\n", __func__, |
243 | (unsigned long)selected_size / SZ_1M); |
244 | |
245 | dma_contiguous_reserve_area(size: selected_size, base: selected_base, |
246 | limit: selected_limit, |
247 | res_cma: &dma_contiguous_default_area, |
248 | fixed); |
249 | } |
250 | } |
251 | |
252 | void __weak |
253 | dma_contiguous_early_fixup(phys_addr_t base, unsigned long size) |
254 | { |
255 | } |
256 | |
257 | /** |
258 | * dma_contiguous_reserve_area() - reserve custom contiguous area |
259 | * @size: Size of the reserved area (in bytes), |
260 | * @base: Base address of the reserved area optional, use 0 for any |
261 | * @limit: End address of the reserved memory (optional, 0 for any). |
262 | * @res_cma: Pointer to store the created cma region. |
263 | * @fixed: hint about where to place the reserved area |
264 | * |
265 | * This function reserves memory from early allocator. It should be |
266 | * called by arch specific code once the early allocator (memblock or bootmem) |
267 | * has been activated and all other subsystems have already allocated/reserved |
268 | * memory. This function allows to create custom reserved areas for specific |
269 | * devices. |
270 | * |
271 | * If @fixed is true, reserve contiguous area at exactly @base. If false, |
272 | * reserve in range from @base to @limit. |
273 | */ |
274 | int __init dma_contiguous_reserve_area(phys_addr_t size, phys_addr_t base, |
275 | phys_addr_t limit, struct cma **res_cma, |
276 | bool fixed) |
277 | { |
278 | int ret; |
279 | |
280 | ret = cma_declare_contiguous(base, size, limit, alignment: 0, order_per_bit: 0, fixed, |
281 | name: "reserved", res_cma); |
282 | if (ret) |
283 | return ret; |
284 | |
285 | /* Architecture specific contiguous memory fixup. */ |
286 | dma_contiguous_early_fixup(base: cma_get_base(cma: *res_cma), |
287 | size: cma_get_size(cma: *res_cma)); |
288 | |
289 | return 0; |
290 | } |
291 | |
292 | /** |
293 | * dma_alloc_from_contiguous() - allocate pages from contiguous area |
294 | * @dev: Pointer to device for which the allocation is performed. |
295 | * @count: Requested number of pages. |
296 | * @align: Requested alignment of pages (in PAGE_SIZE order). |
297 | * @no_warn: Avoid printing message about failed allocation. |
298 | * |
299 | * This function allocates memory buffer for specified device. It uses |
300 | * device specific contiguous memory area if available or the default |
301 | * global one. Requires architecture specific dev_get_cma_area() helper |
302 | * function. |
303 | */ |
304 | struct page *dma_alloc_from_contiguous(struct device *dev, size_t count, |
305 | unsigned int align, bool no_warn) |
306 | { |
307 | if (align > CONFIG_CMA_ALIGNMENT) |
308 | align = CONFIG_CMA_ALIGNMENT; |
309 | |
310 | return cma_alloc(cma: dev_get_cma_area(dev), count, align, no_warn); |
311 | } |
312 | |
313 | /** |
314 | * dma_release_from_contiguous() - release allocated pages |
315 | * @dev: Pointer to device for which the pages were allocated. |
316 | * @pages: Allocated pages. |
317 | * @count: Number of allocated pages. |
318 | * |
319 | * This function releases memory allocated by dma_alloc_from_contiguous(). |
320 | * It returns false when provided pages do not belong to contiguous area and |
321 | * true otherwise. |
322 | */ |
323 | bool dma_release_from_contiguous(struct device *dev, struct page *pages, |
324 | int count) |
325 | { |
326 | return cma_release(cma: dev_get_cma_area(dev), pages, count); |
327 | } |
328 | |
329 | static struct page *cma_alloc_aligned(struct cma *cma, size_t size, gfp_t gfp) |
330 | { |
331 | unsigned int align = min(get_order(size), CONFIG_CMA_ALIGNMENT); |
332 | |
333 | return cma_alloc(cma, count: size >> PAGE_SHIFT, align, no_warn: gfp & __GFP_NOWARN); |
334 | } |
335 | |
336 | /** |
337 | * dma_alloc_contiguous() - allocate contiguous pages |
338 | * @dev: Pointer to device for which the allocation is performed. |
339 | * @size: Requested allocation size. |
340 | * @gfp: Allocation flags. |
341 | * |
342 | * tries to use device specific contiguous memory area if available, or it |
343 | * tries to use per-numa cma, if the allocation fails, it will fallback to |
344 | * try default global one. |
345 | * |
346 | * Note that it bypass one-page size of allocations from the per-numa and |
347 | * global area as the addresses within one page are always contiguous, so |
348 | * there is no need to waste CMA pages for that kind; it also helps reduce |
349 | * fragmentations. |
350 | */ |
351 | struct page *dma_alloc_contiguous(struct device *dev, size_t size, gfp_t gfp) |
352 | { |
353 | #ifdef CONFIG_DMA_NUMA_CMA |
354 | int nid = dev_to_node(dev); |
355 | #endif |
356 | |
357 | /* CMA can be used only in the context which permits sleeping */ |
358 | if (!gfpflags_allow_blocking(gfp_flags: gfp)) |
359 | return NULL; |
360 | if (dev->cma_area) |
361 | return cma_alloc_aligned(cma: dev->cma_area, size, gfp); |
362 | if (size <= PAGE_SIZE) |
363 | return NULL; |
364 | |
365 | #ifdef CONFIG_DMA_NUMA_CMA |
366 | if (nid != NUMA_NO_NODE && !(gfp & (GFP_DMA | GFP_DMA32))) { |
367 | struct cma *cma = dma_contiguous_pernuma_area[nid]; |
368 | struct page *page; |
369 | |
370 | if (cma) { |
371 | page = cma_alloc_aligned(cma, size, gfp); |
372 | if (page) |
373 | return page; |
374 | } |
375 | |
376 | cma = dma_contiguous_numa_area[nid]; |
377 | if (cma) { |
378 | page = cma_alloc_aligned(cma, size, gfp); |
379 | if (page) |
380 | return page; |
381 | } |
382 | } |
383 | #endif |
384 | if (!dma_contiguous_default_area) |
385 | return NULL; |
386 | |
387 | return cma_alloc_aligned(cma: dma_contiguous_default_area, size, gfp); |
388 | } |
389 | |
390 | /** |
391 | * dma_free_contiguous() - release allocated pages |
392 | * @dev: Pointer to device for which the pages were allocated. |
393 | * @page: Pointer to the allocated pages. |
394 | * @size: Size of allocated pages. |
395 | * |
396 | * This function releases memory allocated by dma_alloc_contiguous(). As the |
397 | * cma_release returns false when provided pages do not belong to contiguous |
398 | * area and true otherwise, this function then does a fallback __free_pages() |
399 | * upon a false-return. |
400 | */ |
401 | void dma_free_contiguous(struct device *dev, struct page *page, size_t size) |
402 | { |
403 | unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT; |
404 | |
405 | /* if dev has its own cma, free page from there */ |
406 | if (dev->cma_area) { |
407 | if (cma_release(cma: dev->cma_area, pages: page, count)) |
408 | return; |
409 | } else { |
410 | /* |
411 | * otherwise, page is from either per-numa cma or default cma |
412 | */ |
413 | #ifdef CONFIG_DMA_NUMA_CMA |
414 | if (cma_release(cma: dma_contiguous_pernuma_area[page_to_nid(page)], |
415 | pages: page, count)) |
416 | return; |
417 | if (cma_release(cma: dma_contiguous_numa_area[page_to_nid(page)], |
418 | pages: page, count)) |
419 | return; |
420 | #endif |
421 | if (cma_release(cma: dma_contiguous_default_area, pages: page, count)) |
422 | return; |
423 | } |
424 | |
425 | /* not in any cma, free from buddy */ |
426 | __free_pages(page, order: get_order(size)); |
427 | } |
428 | |
429 | /* |
430 | * Support for reserved memory regions defined in device tree |
431 | */ |
432 | #ifdef CONFIG_OF_RESERVED_MEM |
433 | #include <linux/of.h> |
434 | #include <linux/of_fdt.h> |
435 | #include <linux/of_reserved_mem.h> |
436 | |
437 | #undef pr_fmt |
438 | #define pr_fmt(fmt) fmt |
439 | |
440 | static int rmem_cma_device_init(struct reserved_mem *rmem, struct device *dev) |
441 | { |
442 | dev->cma_area = rmem->priv; |
443 | return 0; |
444 | } |
445 | |
446 | static void rmem_cma_device_release(struct reserved_mem *rmem, |
447 | struct device *dev) |
448 | { |
449 | dev->cma_area = NULL; |
450 | } |
451 | |
452 | static const struct reserved_mem_ops rmem_cma_ops = { |
453 | .device_init = rmem_cma_device_init, |
454 | .device_release = rmem_cma_device_release, |
455 | }; |
456 | |
457 | static int __init rmem_cma_setup(struct reserved_mem *rmem) |
458 | { |
459 | unsigned long node = rmem->fdt_node; |
460 | bool default_cma = of_get_flat_dt_prop(node, name: "linux,cma-default", NULL); |
461 | struct cma *cma; |
462 | int err; |
463 | |
464 | if (size_cmdline != -1 && default_cma) { |
465 | pr_info("Reserved memory: bypass %s node, using cmdline CMA params instead\n", |
466 | rmem->name); |
467 | return -EBUSY; |
468 | } |
469 | |
470 | if (!of_get_flat_dt_prop(node, name: "reusable", NULL) || |
471 | of_get_flat_dt_prop(node, name: "no-map", NULL)) |
472 | return -EINVAL; |
473 | |
474 | if (!IS_ALIGNED(rmem->base | rmem->size, CMA_MIN_ALIGNMENT_BYTES)) { |
475 | pr_err("Reserved memory: incorrect alignment of CMA region\n"); |
476 | return -EINVAL; |
477 | } |
478 | |
479 | err = cma_init_reserved_mem(base: rmem->base, size: rmem->size, order_per_bit: 0, name: rmem->name, res_cma: &cma); |
480 | if (err) { |
481 | pr_err("Reserved memory: unable to setup CMA region\n"); |
482 | return err; |
483 | } |
484 | /* Architecture specific contiguous memory fixup. */ |
485 | dma_contiguous_early_fixup(base: rmem->base, size: rmem->size); |
486 | |
487 | if (default_cma) |
488 | dma_contiguous_default_area = cma; |
489 | |
490 | rmem->ops = &rmem_cma_ops; |
491 | rmem->priv = cma; |
492 | |
493 | pr_info("Reserved memory: created CMA memory pool at %pa, size %ld MiB\n", |
494 | &rmem->base, (unsigned long)rmem->size / SZ_1M); |
495 | |
496 | return 0; |
497 | } |
498 | RESERVEDMEM_OF_DECLARE(cma, "shared-dma-pool", rmem_cma_setup); |
499 | #endif |
500 |
Definitions
- dma_contiguous_default_area
- size_bytes
- 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