xref: /openbmc/linux/kernel/dma/contiguous.c (revision 1edd0337)
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 #ifdef CONFIG_CMA_DEBUG
41 #ifndef DEBUG
42 #  define DEBUG
43 #endif
44 #endif
45 
46 #include <asm/page.h>
47 
48 #include <linux/memblock.h>
49 #include <linux/err.h>
50 #include <linux/sizes.h>
51 #include <linux/dma-map-ops.h>
52 #include <linux/cma.h>
53 
54 #ifdef CONFIG_CMA_SIZE_MBYTES
55 #define CMA_SIZE_MBYTES CONFIG_CMA_SIZE_MBYTES
56 #else
57 #define CMA_SIZE_MBYTES 0
58 #endif
59 
60 struct cma *dma_contiguous_default_area;
61 
62 /*
63  * Default global CMA area size can be defined in kernel's .config.
64  * This is useful mainly for distro maintainers to create a kernel
65  * that works correctly for most supported systems.
66  * The size can be set in bytes or as a percentage of the total memory
67  * in the system.
68  *
69  * Users, who want to set the size of global CMA area for their system
70  * should use cma= kernel parameter.
71  */
72 static const phys_addr_t size_bytes __initconst =
73 	(phys_addr_t)CMA_SIZE_MBYTES * SZ_1M;
74 static phys_addr_t  size_cmdline __initdata = -1;
75 static phys_addr_t base_cmdline __initdata;
76 static phys_addr_t limit_cmdline __initdata;
77 
78 static int __init early_cma(char *p)
79 {
80 	if (!p) {
81 		pr_err("Config string not provided\n");
82 		return -EINVAL;
83 	}
84 
85 	size_cmdline = memparse(p, &p);
86 	if (*p != '@')
87 		return 0;
88 	base_cmdline = memparse(p + 1, &p);
89 	if (*p != '-') {
90 		limit_cmdline = base_cmdline + size_cmdline;
91 		return 0;
92 	}
93 	limit_cmdline = memparse(p + 1, &p);
94 
95 	return 0;
96 }
97 early_param("cma", early_cma);
98 
99 #ifdef CONFIG_DMA_PERNUMA_CMA
100 
101 static struct cma *dma_contiguous_pernuma_area[MAX_NUMNODES];
102 static phys_addr_t pernuma_size_bytes __initdata;
103 
104 static int __init early_cma_pernuma(char *p)
105 {
106 	pernuma_size_bytes = memparse(p, &p);
107 	return 0;
108 }
109 early_param("cma_pernuma", early_cma_pernuma);
110 #endif
111 
112 #ifdef CONFIG_CMA_SIZE_PERCENTAGE
113 
114 static phys_addr_t __init __maybe_unused cma_early_percent_memory(void)
115 {
116 	unsigned long total_pages = PHYS_PFN(memblock_phys_mem_size());
117 
118 	return (total_pages * CONFIG_CMA_SIZE_PERCENTAGE / 100) << PAGE_SHIFT;
119 }
120 
121 #else
122 
123 static inline __maybe_unused phys_addr_t cma_early_percent_memory(void)
124 {
125 	return 0;
126 }
127 
128 #endif
129 
130 #ifdef CONFIG_DMA_PERNUMA_CMA
131 void __init dma_pernuma_cma_reserve(void)
132 {
133 	int nid;
134 
135 	if (!pernuma_size_bytes)
136 		return;
137 
138 	for_each_online_node(nid) {
139 		int ret;
140 		char name[CMA_MAX_NAME];
141 		struct cma **cma = &dma_contiguous_pernuma_area[nid];
142 
143 		snprintf(name, sizeof(name), "pernuma%d", nid);
144 		ret = cma_declare_contiguous_nid(0, pernuma_size_bytes, 0, 0,
145 						 0, false, name, cma, nid);
146 		if (ret) {
147 			pr_warn("%s: reservation failed: err %d, node %d", __func__,
148 				ret, nid);
149 			continue;
150 		}
151 
152 		pr_debug("%s: reserved %llu MiB on node %d\n", __func__,
153 			(unsigned long long)pernuma_size_bytes / SZ_1M, nid);
154 	}
155 }
156 #endif
157 
158 /**
159  * dma_contiguous_reserve() - reserve area(s) for contiguous memory handling
160  * @limit: End address of the reserved memory (optional, 0 for any).
161  *
162  * This function reserves memory from early allocator. It should be
163  * called by arch specific code once the early allocator (memblock or bootmem)
164  * has been activated and all other subsystems have already allocated/reserved
165  * memory.
166  */
167 void __init dma_contiguous_reserve(phys_addr_t limit)
168 {
169 	phys_addr_t selected_size = 0;
170 	phys_addr_t selected_base = 0;
171 	phys_addr_t selected_limit = limit;
172 	bool fixed = false;
173 
174 	pr_debug("%s(limit %08lx)\n", __func__, (unsigned long)limit);
175 
176 	if (size_cmdline != -1) {
177 		selected_size = size_cmdline;
178 		selected_base = base_cmdline;
179 		selected_limit = min_not_zero(limit_cmdline, limit);
180 		if (base_cmdline + size_cmdline == limit_cmdline)
181 			fixed = true;
182 	} else {
183 #ifdef CONFIG_CMA_SIZE_SEL_MBYTES
184 		selected_size = size_bytes;
185 #elif defined(CONFIG_CMA_SIZE_SEL_PERCENTAGE)
186 		selected_size = cma_early_percent_memory();
187 #elif defined(CONFIG_CMA_SIZE_SEL_MIN)
188 		selected_size = min(size_bytes, cma_early_percent_memory());
189 #elif defined(CONFIG_CMA_SIZE_SEL_MAX)
190 		selected_size = max(size_bytes, cma_early_percent_memory());
191 #endif
192 	}
193 
194 	if (selected_size && !dma_contiguous_default_area) {
195 		pr_debug("%s: reserving %ld MiB for global area\n", __func__,
196 			 (unsigned long)selected_size / SZ_1M);
197 
198 		dma_contiguous_reserve_area(selected_size, selected_base,
199 					    selected_limit,
200 					    &dma_contiguous_default_area,
201 					    fixed);
202 	}
203 }
204 
205 void __weak
206 dma_contiguous_early_fixup(phys_addr_t base, unsigned long size)
207 {
208 }
209 
210 /**
211  * dma_contiguous_reserve_area() - reserve custom contiguous area
212  * @size: Size of the reserved area (in bytes),
213  * @base: Base address of the reserved area optional, use 0 for any
214  * @limit: End address of the reserved memory (optional, 0 for any).
215  * @res_cma: Pointer to store the created cma region.
216  * @fixed: hint about where to place the reserved area
217  *
218  * This function reserves memory from early allocator. It should be
219  * called by arch specific code once the early allocator (memblock or bootmem)
220  * has been activated and all other subsystems have already allocated/reserved
221  * memory. This function allows to create custom reserved areas for specific
222  * devices.
223  *
224  * If @fixed is true, reserve contiguous area at exactly @base.  If false,
225  * reserve in range from @base to @limit.
226  */
227 int __init dma_contiguous_reserve_area(phys_addr_t size, phys_addr_t base,
228 				       phys_addr_t limit, struct cma **res_cma,
229 				       bool fixed)
230 {
231 	int ret;
232 
233 	ret = cma_declare_contiguous(base, size, limit, 0, 0, fixed,
234 					"reserved", res_cma);
235 	if (ret)
236 		return ret;
237 
238 	/* Architecture specific contiguous memory fixup. */
239 	dma_contiguous_early_fixup(cma_get_base(*res_cma),
240 				cma_get_size(*res_cma));
241 
242 	return 0;
243 }
244 
245 /**
246  * dma_alloc_from_contiguous() - allocate pages from contiguous area
247  * @dev:   Pointer to device for which the allocation is performed.
248  * @count: Requested number of pages.
249  * @align: Requested alignment of pages (in PAGE_SIZE order).
250  * @no_warn: Avoid printing message about failed allocation.
251  *
252  * This function allocates memory buffer for specified device. It uses
253  * device specific contiguous memory area if available or the default
254  * global one. Requires architecture specific dev_get_cma_area() helper
255  * function.
256  */
257 struct page *dma_alloc_from_contiguous(struct device *dev, size_t count,
258 				       unsigned int align, bool no_warn)
259 {
260 	if (align > CONFIG_CMA_ALIGNMENT)
261 		align = CONFIG_CMA_ALIGNMENT;
262 
263 	return cma_alloc(dev_get_cma_area(dev), count, align, no_warn);
264 }
265 
266 /**
267  * dma_release_from_contiguous() - release allocated pages
268  * @dev:   Pointer to device for which the pages were allocated.
269  * @pages: Allocated pages.
270  * @count: Number of allocated pages.
271  *
272  * This function releases memory allocated by dma_alloc_from_contiguous().
273  * It returns false when provided pages do not belong to contiguous area and
274  * true otherwise.
275  */
276 bool dma_release_from_contiguous(struct device *dev, struct page *pages,
277 				 int count)
278 {
279 	return cma_release(dev_get_cma_area(dev), pages, count);
280 }
281 
282 static struct page *cma_alloc_aligned(struct cma *cma, size_t size, gfp_t gfp)
283 {
284 	unsigned int align = min(get_order(size), CONFIG_CMA_ALIGNMENT);
285 
286 	return cma_alloc(cma, size >> PAGE_SHIFT, align, gfp & __GFP_NOWARN);
287 }
288 
289 /**
290  * dma_alloc_contiguous() - allocate contiguous pages
291  * @dev:   Pointer to device for which the allocation is performed.
292  * @size:  Requested allocation size.
293  * @gfp:   Allocation flags.
294  *
295  * tries to use device specific contiguous memory area if available, or it
296  * tries to use per-numa cma, if the allocation fails, it will fallback to
297  * try default global one.
298  *
299  * Note that it bypass one-page size of allocations from the per-numa and
300  * global area as the addresses within one page are always contiguous, so
301  * there is no need to waste CMA pages for that kind; it also helps reduce
302  * fragmentations.
303  */
304 struct page *dma_alloc_contiguous(struct device *dev, size_t size, gfp_t gfp)
305 {
306 #ifdef CONFIG_DMA_PERNUMA_CMA
307 	int nid = dev_to_node(dev);
308 #endif
309 
310 	/* CMA can be used only in the context which permits sleeping */
311 	if (!gfpflags_allow_blocking(gfp))
312 		return NULL;
313 	if (dev->cma_area)
314 		return cma_alloc_aligned(dev->cma_area, size, gfp);
315 	if (size <= PAGE_SIZE)
316 		return NULL;
317 
318 #ifdef CONFIG_DMA_PERNUMA_CMA
319 	if (nid != NUMA_NO_NODE && !(gfp & (GFP_DMA | GFP_DMA32))) {
320 		struct cma *cma = dma_contiguous_pernuma_area[nid];
321 		struct page *page;
322 
323 		if (cma) {
324 			page = cma_alloc_aligned(cma, size, gfp);
325 			if (page)
326 				return page;
327 		}
328 	}
329 #endif
330 	if (!dma_contiguous_default_area)
331 		return NULL;
332 
333 	return cma_alloc_aligned(dma_contiguous_default_area, size, gfp);
334 }
335 
336 /**
337  * dma_free_contiguous() - release allocated pages
338  * @dev:   Pointer to device for which the pages were allocated.
339  * @page:  Pointer to the allocated pages.
340  * @size:  Size of allocated pages.
341  *
342  * This function releases memory allocated by dma_alloc_contiguous(). As the
343  * cma_release returns false when provided pages do not belong to contiguous
344  * area and true otherwise, this function then does a fallback __free_pages()
345  * upon a false-return.
346  */
347 void dma_free_contiguous(struct device *dev, struct page *page, size_t size)
348 {
349 	unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT;
350 
351 	/* if dev has its own cma, free page from there */
352 	if (dev->cma_area) {
353 		if (cma_release(dev->cma_area, page, count))
354 			return;
355 	} else {
356 		/*
357 		 * otherwise, page is from either per-numa cma or default cma
358 		 */
359 #ifdef CONFIG_DMA_PERNUMA_CMA
360 		if (cma_release(dma_contiguous_pernuma_area[page_to_nid(page)],
361 					page, count))
362 			return;
363 #endif
364 		if (cma_release(dma_contiguous_default_area, page, count))
365 			return;
366 	}
367 
368 	/* not in any cma, free from buddy */
369 	__free_pages(page, get_order(size));
370 }
371 
372 /*
373  * Support for reserved memory regions defined in device tree
374  */
375 #ifdef CONFIG_OF_RESERVED_MEM
376 #include <linux/of.h>
377 #include <linux/of_fdt.h>
378 #include <linux/of_reserved_mem.h>
379 
380 #undef pr_fmt
381 #define pr_fmt(fmt) fmt
382 
383 static int rmem_cma_device_init(struct reserved_mem *rmem, struct device *dev)
384 {
385 	dev->cma_area = rmem->priv;
386 	return 0;
387 }
388 
389 static void rmem_cma_device_release(struct reserved_mem *rmem,
390 				    struct device *dev)
391 {
392 	dev->cma_area = NULL;
393 }
394 
395 static const struct reserved_mem_ops rmem_cma_ops = {
396 	.device_init	= rmem_cma_device_init,
397 	.device_release = rmem_cma_device_release,
398 };
399 
400 static int __init rmem_cma_setup(struct reserved_mem *rmem)
401 {
402 	unsigned long node = rmem->fdt_node;
403 	bool default_cma = of_get_flat_dt_prop(node, "linux,cma-default", NULL);
404 	struct cma *cma;
405 	int err;
406 
407 	if (size_cmdline != -1 && default_cma) {
408 		pr_info("Reserved memory: bypass %s node, using cmdline CMA params instead\n",
409 			rmem->name);
410 		return -EBUSY;
411 	}
412 
413 	if (!of_get_flat_dt_prop(node, "reusable", NULL) ||
414 	    of_get_flat_dt_prop(node, "no-map", NULL))
415 		return -EINVAL;
416 
417 	if (!IS_ALIGNED(rmem->base | rmem->size, CMA_MIN_ALIGNMENT_BYTES)) {
418 		pr_err("Reserved memory: incorrect alignment of CMA region\n");
419 		return -EINVAL;
420 	}
421 
422 	err = cma_init_reserved_mem(rmem->base, rmem->size, 0, rmem->name, &cma);
423 	if (err) {
424 		pr_err("Reserved memory: unable to setup CMA region\n");
425 		return err;
426 	}
427 	/* Architecture specific contiguous memory fixup. */
428 	dma_contiguous_early_fixup(rmem->base, rmem->size);
429 
430 	if (default_cma)
431 		dma_contiguous_default_area = cma;
432 
433 	rmem->ops = &rmem_cma_ops;
434 	rmem->priv = cma;
435 
436 	pr_info("Reserved memory: created CMA memory pool at %pa, size %ld MiB\n",
437 		&rmem->base, (unsigned long)rmem->size / SZ_1M);
438 
439 	return 0;
440 }
441 RESERVEDMEM_OF_DECLARE(cma, "shared-dma-pool", rmem_cma_setup);
442 #endif
443