contiguous.c

  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-buf/heaps/cma.h>
 46#include <linux/dma-map-ops.h>
 47#include <linux/cma.h>
 48#include <linux/nospec.h>
 49
 50#ifdef CONFIG_CMA_SIZE_MBYTES
 51#define CMA_SIZE_MBYTES CONFIG_CMA_SIZE_MBYTES
 52#else
 53#define CMA_SIZE_MBYTES 0
 54#endif
 55
 56struct cma *dma_contiguous_default_area;
 57
 58/*
 59 * Default global CMA area size can be defined in kernel's .config.
 60 * This is useful mainly for distro maintainers to create a kernel
 61 * that works correctly for most supported systems.
 62 * The size can be set in bytes or as a percentage of the total memory
 63 * in the system.
 64 *
 65 * Users, who want to set the size of global CMA area for their system
 66 * should use cma= kernel parameter.
 67 */
 68#define size_bytes ((phys_addr_t)CMA_SIZE_MBYTES * SZ_1M)
 69static phys_addr_t  size_cmdline __initdata = -1;
 70static phys_addr_t base_cmdline __initdata;
 71static phys_addr_t limit_cmdline __initdata;
 72
 73static 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(p, &p);
 81	if (*p != '@')
 82		return 0;
 83	base_cmdline = memparse(p + 1, &p);
 84	if (*p != '-') {
 85		limit_cmdline = base_cmdline + size_cmdline;
 86		return 0;
 87	}
 88	limit_cmdline = memparse(p + 1, &p);
 89
 90	return 0;
 91}
 92early_param("cma", early_cma);
 93
 94/*
 95 * cma_skip_dt_default_reserved_mem - This is called from the
 96 * reserved_mem framework to detect if the default cma region is being
 97 * set by the "cma=" kernel parameter.
 98 */
 99bool __init cma_skip_dt_default_reserved_mem(void)
100{
101	return size_cmdline != -1;
102}
103
104#ifdef CONFIG_DMA_NUMA_CMA
105
106static struct cma *dma_contiguous_numa_area[MAX_NUMNODES];
107static phys_addr_t numa_cma_size[MAX_NUMNODES] __initdata;
108static struct cma *dma_contiguous_pernuma_area[MAX_NUMNODES];
109static phys_addr_t pernuma_size_bytes __initdata;
110
111static int __init early_numa_cma(char *p)
112{
113	int nid, count = 0;
114	unsigned long tmp;
115	char *s = p;
116
117	while (*s) {
118		if (sscanf(s, "%lu%n", &tmp, &count) != 1)
119			break;
120
121		if (s[count] == ':') {
122			if (tmp >= MAX_NUMNODES)
123				break;
124			nid = array_index_nospec(tmp, MAX_NUMNODES);
125
126			s += count + 1;
127			tmp = memparse(s, &s);
128			numa_cma_size[nid] = tmp;
129
130			if (*s == ',')
131				s++;
132			else
133				break;
134		} else
135			break;
136	}
137
138	return 0;
139}
140early_param("numa_cma", early_numa_cma);
141
142static int __init early_cma_pernuma(char *p)
143{
144	pernuma_size_bytes = memparse(p, &p);
145	return 0;
146}
147early_param("cma_pernuma", early_cma_pernuma);
148#endif
149
150#ifdef CONFIG_CMA_SIZE_PERCENTAGE
151
152static phys_addr_t __init __maybe_unused cma_early_percent_memory(void)
153{
154	unsigned long total_pages = PHYS_PFN(memblock_phys_mem_size());
155
156	return (total_pages * CONFIG_CMA_SIZE_PERCENTAGE / 100) << PAGE_SHIFT;
157}
158
159#else
160
161static inline __maybe_unused phys_addr_t cma_early_percent_memory(void)
162{
163	return 0;
164}
165
166#endif
167
168#ifdef CONFIG_DMA_NUMA_CMA
169static void __init dma_numa_cma_reserve(void)
170{
171	int nid;
172
173	for_each_node(nid) {
174		int ret;
175		char name[CMA_MAX_NAME];
176		struct cma **cma;
177
178		if (!node_online(nid)) {
179			if (pernuma_size_bytes || numa_cma_size[nid])
180				pr_warn("invalid node %d specified\n", nid);
181			continue;
182		}
183
184		if (pernuma_size_bytes) {
185
186			cma = &dma_contiguous_pernuma_area[nid];
187			snprintf(name, sizeof(name), "pernuma%d", nid);
188			ret = cma_declare_contiguous_nid(0, pernuma_size_bytes, 0, 0,
189							 0, false, name, cma, nid);
190			if (ret)
191				pr_warn("%s: reservation failed: err %d, node %d", __func__,
192					ret, nid);
193		}
194
195		if (numa_cma_size[nid]) {
196
197			cma = &dma_contiguous_numa_area[nid];
198			snprintf(name, sizeof(name), "numa%d", nid);
199			ret = cma_declare_contiguous_nid(0, numa_cma_size[nid], 0, 0, 0, false,
200							 name, cma, nid);
201			if (ret)
202				pr_warn("%s: reservation failed: err %d, node %d", __func__,
203					ret, nid);
204		}
205	}
206}
207#else
208static inline void __init dma_numa_cma_reserve(void)
209{
210}
211#endif
212
213/**
214 * dma_contiguous_reserve() - reserve area(s) for contiguous memory handling
215 * @limit: End address of the reserved memory (optional, 0 for any).
216 *
217 * This function reserves memory from early allocator. It should be
218 * called by arch specific code once the early allocator (memblock or bootmem)
219 * has been activated and all other subsystems have already allocated/reserved
220 * memory.
221 */
222void __init dma_contiguous_reserve(phys_addr_t limit)
223{
224	phys_addr_t selected_size = 0;
225	phys_addr_t selected_base = 0;
226	phys_addr_t selected_limit = limit;
227	bool fixed = false;
228
229	dma_numa_cma_reserve();
230
231	pr_debug("%s(limit %08lx)\n", __func__, (unsigned long)limit);
232
233	if (size_cmdline != -1) {
234		selected_size = size_cmdline;
235		selected_base = base_cmdline;
236
237		/* Hornor the user setup dma address limit */
238		selected_limit = limit_cmdline ?: limit;
239
240		if (base_cmdline + size_cmdline == limit_cmdline)
241			fixed = true;
242	} else {
243#ifdef CONFIG_CMA_SIZE_SEL_MBYTES
244		selected_size = size_bytes;
245#elif defined(CONFIG_CMA_SIZE_SEL_PERCENTAGE)
246		selected_size = cma_early_percent_memory();
247#elif defined(CONFIG_CMA_SIZE_SEL_MIN)
248		selected_size = min(size_bytes, cma_early_percent_memory());
249#elif defined(CONFIG_CMA_SIZE_SEL_MAX)
250		selected_size = max(size_bytes, cma_early_percent_memory());
251#endif
252	}
253
254	if (selected_size && !dma_contiguous_default_area) {
255		int ret;
256
257		pr_debug("%s: reserving %ld MiB for global area\n", __func__,
258			 (unsigned long)selected_size / SZ_1M);
259
260		ret = dma_contiguous_reserve_area(selected_size, selected_base,
261						  selected_limit,
262						  &dma_contiguous_default_area,
263						  fixed);
264		if (ret)
265			return;
266
267		ret = dma_heap_cma_register_heap(dma_contiguous_default_area);
268		if (ret)
269			pr_warn("Couldn't register default CMA heap.");
270	}
271}
272
273void __weak
274dma_contiguous_early_fixup(phys_addr_t base, unsigned long size)
275{
276}
277
278/**
279 * dma_contiguous_reserve_area() - reserve custom contiguous area
280 * @size: Size of the reserved area (in bytes),
281 * @base: Base address of the reserved area optional, use 0 for any
282 * @limit: End address of the reserved memory (optional, 0 for any).
283 * @res_cma: Pointer to store the created cma region.
284 * @fixed: hint about where to place the reserved area
285 *
286 * This function reserves memory from early allocator. It should be
287 * called by arch specific code once the early allocator (memblock or bootmem)
288 * has been activated and all other subsystems have already allocated/reserved
289 * memory. This function allows to create custom reserved areas for specific
290 * devices.
291 *
292 * If @fixed is true, reserve contiguous area at exactly @base.  If false,
293 * reserve in range from @base to @limit.
294 */
295int __init dma_contiguous_reserve_area(phys_addr_t size, phys_addr_t base,
296				       phys_addr_t limit, struct cma **res_cma,
297				       bool fixed)
298{
299	int ret;
300
301	ret = cma_declare_contiguous(base, size, limit, 0, 0, fixed,
302					"reserved", res_cma);
303	if (ret)
304		return ret;
305
306	/* Architecture specific contiguous memory fixup. */
307	dma_contiguous_early_fixup(cma_get_base(*res_cma),
308				cma_get_size(*res_cma));
309
310	return 0;
311}
312
313/**
314 * dma_alloc_from_contiguous() - allocate pages from contiguous area
315 * @dev:   Pointer to device for which the allocation is performed.
316 * @count: Requested number of pages.
317 * @align: Requested alignment of pages (in PAGE_SIZE order).
318 * @no_warn: Avoid printing message about failed allocation.
319 *
320 * This function allocates memory buffer for specified device. It uses
321 * device specific contiguous memory area if available or the default
322 * global one. Requires architecture specific dev_get_cma_area() helper
323 * function.
324 */
325struct page *dma_alloc_from_contiguous(struct device *dev, size_t count,
326				       unsigned int align, bool no_warn)
327{
328	if (align > CONFIG_CMA_ALIGNMENT)
329		align = CONFIG_CMA_ALIGNMENT;
330
331	return cma_alloc(dev_get_cma_area(dev), count, align, no_warn);
332}
333
334/**
335 * dma_release_from_contiguous() - release allocated pages
336 * @dev:   Pointer to device for which the pages were allocated.
337 * @pages: Allocated pages.
338 * @count: Number of allocated pages.
339 *
340 * This function releases memory allocated by dma_alloc_from_contiguous().
341 * It returns false when provided pages do not belong to contiguous area and
342 * true otherwise.
343 */
344bool dma_release_from_contiguous(struct device *dev, struct page *pages,
345				 int count)
346{
347	return cma_release(dev_get_cma_area(dev), pages, count);
348}
349
350static struct page *cma_alloc_aligned(struct cma *cma, size_t size, gfp_t gfp)
351{
352	unsigned int align = min(get_order(size), CONFIG_CMA_ALIGNMENT);
353
354	return cma_alloc(cma, size >> PAGE_SHIFT, align, gfp & __GFP_NOWARN);
355}
356
357/**
358 * dma_alloc_contiguous() - allocate contiguous pages
359 * @dev:   Pointer to device for which the allocation is performed.
360 * @size:  Requested allocation size.
361 * @gfp:   Allocation flags.
362 *
363 * tries to use device specific contiguous memory area if available, or it
364 * tries to use per-numa cma, if the allocation fails, it will fallback to
365 * try default global one.
366 *
367 * Note that it bypass one-page size of allocations from the per-numa and
368 * global area as the addresses within one page are always contiguous, so
369 * there is no need to waste CMA pages for that kind; it also helps reduce
370 * fragmentations.
371 */
372struct page *dma_alloc_contiguous(struct device *dev, size_t size, gfp_t gfp)
373{
374#ifdef CONFIG_DMA_NUMA_CMA
375	int nid = dev_to_node(dev);
376#endif
377
378	/* CMA can be used only in the context which permits sleeping */
379	if (!gfpflags_allow_blocking(gfp))
380		return NULL;
381	if (dev->cma_area)
382		return cma_alloc_aligned(dev->cma_area, size, gfp);
383	if (size <= PAGE_SIZE)
384		return NULL;
385
386#ifdef CONFIG_DMA_NUMA_CMA
387	if (nid != NUMA_NO_NODE && !(gfp & (GFP_DMA | GFP_DMA32))) {
388		struct cma *cma = dma_contiguous_pernuma_area[nid];
389		struct page *page;
390
391		if (cma) {
392			page = cma_alloc_aligned(cma, size, gfp);
393			if (page)
394				return page;
395		}
396
397		cma = dma_contiguous_numa_area[nid];
398		if (cma) {
399			page = cma_alloc_aligned(cma, size, gfp);
400			if (page)
401				return page;
402		}
403	}
404#endif
405	if (!dma_contiguous_default_area)
406		return NULL;
407
408	return cma_alloc_aligned(dma_contiguous_default_area, size, gfp);
409}
410
411/**
412 * dma_free_contiguous() - release allocated pages
413 * @dev:   Pointer to device for which the pages were allocated.
414 * @page:  Pointer to the allocated pages.
415 * @size:  Size of allocated pages.
416 *
417 * This function releases memory allocated by dma_alloc_contiguous(). As the
418 * cma_release returns false when provided pages do not belong to contiguous
419 * area and true otherwise, this function then does a fallback __free_pages()
420 * upon a false-return.
421 */
422void dma_free_contiguous(struct device *dev, struct page *page, size_t size)
423{
424	unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT;
425
426	/* if dev has its own cma, free page from there */
427	if (dev->cma_area) {
428		if (cma_release(dev->cma_area, page, count))
429			return;
430	} else {
431		/*
432		 * otherwise, page is from either per-numa cma or default cma
433		 */
434#ifdef CONFIG_DMA_NUMA_CMA
435		if (cma_release(dma_contiguous_pernuma_area[page_to_nid(page)],
436					page, count))
437			return;
438		if (cma_release(dma_contiguous_numa_area[page_to_nid(page)],
439					page, count))
440			return;
441#endif
442		if (cma_release(dma_contiguous_default_area, page, count))
443			return;
444	}
445
446	/* not in any cma, free from buddy */
447	__free_pages(page, get_order(size));
448}
449
450/*
451 * Support for reserved memory regions defined in device tree
452 */
453#ifdef CONFIG_OF_RESERVED_MEM
454#include <linux/of.h>
455#include <linux/of_fdt.h>
456#include <linux/of_reserved_mem.h>
457
458#undef pr_fmt
459#define pr_fmt(fmt) fmt
460
461static int rmem_cma_device_init(struct reserved_mem *rmem, struct device *dev)
462{
463	dev->cma_area = rmem->priv;
464	return 0;
465}
466
467static void rmem_cma_device_release(struct reserved_mem *rmem,
468				    struct device *dev)
469{
470	dev->cma_area = NULL;
471}
472
473static const struct reserved_mem_ops rmem_cma_ops = {
474	.device_init	= rmem_cma_device_init,
475	.device_release = rmem_cma_device_release,
476};
477
478static int __init rmem_cma_setup(struct reserved_mem *rmem)
479{
480	unsigned long node = rmem->fdt_node;
481	bool default_cma = of_get_flat_dt_prop(node, "linux,cma-default", NULL);
482	struct cma *cma;
483	int err;
484
485	if (!of_get_flat_dt_prop(node, "reusable", NULL) ||
486	    of_get_flat_dt_prop(node, "no-map", NULL))
487		return -EINVAL;
488
489	if (!IS_ALIGNED(rmem->base | rmem->size, CMA_MIN_ALIGNMENT_BYTES)) {
490		pr_err("Reserved memory: incorrect alignment of CMA region\n");
491		return -EINVAL;
492	}
493
494	err = cma_init_reserved_mem(rmem->base, rmem->size, 0, rmem->name, &cma);
495	if (err) {
496		pr_err("Reserved memory: unable to setup CMA region\n");
497		return err;
498	}
499
500	if (default_cma)
501		dma_contiguous_default_area = cma;
502
503	rmem->ops = &rmem_cma_ops;
504	rmem->priv = cma;
505
506	pr_info("Reserved memory: created CMA memory pool at %pa, size %ld MiB\n",
507		&rmem->base, (unsigned long)rmem->size / SZ_1M);
508
509	err = dma_heap_cma_register_heap(cma);
510	if (err)
511		pr_warn("Couldn't register CMA heap.");
512
513	return 0;
514}
515RESERVEDMEM_OF_DECLARE(cma, "shared-dma-pool", rmem_cma_setup);
516#endif