kernel/dma/contiguous.c

0001 // SPDX-License-Identifier: GPL-2.0+
0002 /*
0003  * Contiguous Memory Allocator for DMA mapping framework
0004  * Copyright (c) 2010-2011 by Samsung Electronics.
0005  * Written by:
0006  *  Marek Szyprowski <m.szyprowski@samsung.com>
0007  *  Michal Nazarewicz <mina86@mina86.com>
0008  *
0009  * Contiguous Memory Allocator
0010  *
0011  *   The Contiguous Memory Allocator (CMA) makes it possible to
0012  *   allocate big contiguous chunks of memory after the system has
0013  *   booted.
0014  *
0015  * Why is it needed?
0016  *
0017  *   Various devices on embedded systems have no scatter-getter and/or
0018  *   IO map support and require contiguous blocks of memory to
0019  *   operate.  They include devices such as cameras, hardware video
0020  *   coders, etc.
0021  *
0022  *   Such devices often require big memory buffers (a full HD frame
0023  *   is, for instance, more than 2 mega pixels large, i.e. more than 6
0024  *   MB of memory), which makes mechanisms such as kmalloc() or
0025  *   alloc_page() ineffective.
0026  *
0027  *   At the same time, a solution where a big memory region is
0028  *   reserved for a device is suboptimal since often more memory is
0029  *   reserved then strictly required and, moreover, the memory is
0030  *   inaccessible to page system even if device drivers don't use it.
0031  *
0032  *   CMA tries to solve this issue by operating on memory regions
0033  *   where only movable pages can be allocated from.  This way, kernel
0034  *   can use the memory for pagecache and when device driver requests
0035  *   it, allocated pages can be migrated.
0036  */
0037
0038 #define pr_fmt(fmt) "cma: " fmt
0039
0040 #ifdef CONFIG_CMA_DEBUG
0041 #ifndef DEBUG
0042 #  define DEBUG
0043 #endif
0044 #endif
0045
0046 #include <asm/page.h>
0047
0048 #include <linux/memblock.h>
0049 #include <linux/err.h>
0050 #include <linux/sizes.h>
0051 #include <linux/dma-map-ops.h>
0052 #include <linux/cma.h>
0053
0054 #ifdef CONFIG_CMA_SIZE_MBYTES
0055 #define CMA_SIZE_MBYTES CONFIG_CMA_SIZE_MBYTES
0056 #else
0057 #define CMA_SIZE_MBYTES 0
0058 #endif
0059
0060 struct cma *dma_contiguous_default_area;
0061
0062 /*
0063  * Default global CMA area size can be defined in kernel's .config.
0064  * This is useful mainly for distro maintainers to create a kernel
0065  * that works correctly for most supported systems.
0066  * The size can be set in bytes or as a percentage of the total memory
0067  * in the system.
0068  *
0069  * Users, who want to set the size of global CMA area for their system
0070  * should use cma= kernel parameter.
0071  */
0072 static const phys_addr_t size_bytes __initconst =
0073     (phys_addr_t)CMA_SIZE_MBYTES * SZ_1M;
0074 static phys_addr_t  size_cmdline __initdata = -1;
0075 static phys_addr_t base_cmdline __initdata;
0076 static phys_addr_t limit_cmdline __initdata;
0077
0078 static int __init early_cma(char *p)
0079 {
0080     if (!p) {
0081         pr_err("Config string not provided\n");
0082         return -EINVAL;
0083     }
0084
0085     size_cmdline = memparse(p, &p);
0086     if (*p != '@')
0087         return 0;
0088     base_cmdline = memparse(p + 1, &p);
0089     if (*p != '-') {
0090         limit_cmdline = base_cmdline + size_cmdline;
0091         return 0;
0092     }
0093     limit_cmdline = memparse(p + 1, &p);
0094
0095     return 0;
0096 }
0097 early_param("cma", early_cma);
0098
0099 #ifdef CONFIG_DMA_PERNUMA_CMA
0100
0101 static struct cma *dma_contiguous_pernuma_area[MAX_NUMNODES];
0102 static phys_addr_t pernuma_size_bytes __initdata;
0103
0104 static int __init early_cma_pernuma(char *p)
0105 {
0106     pernuma_size_bytes = memparse(p, &p);
0107     return 0;
0108 }
0109 early_param("cma_pernuma", early_cma_pernuma);
0110 #endif
0111
0112 #ifdef CONFIG_CMA_SIZE_PERCENTAGE
0113
0114 static phys_addr_t __init __maybe_unused cma_early_percent_memory(void)
0115 {
0116     unsigned long total_pages = PHYS_PFN(memblock_phys_mem_size());
0117
0118     return (total_pages * CONFIG_CMA_SIZE_PERCENTAGE / 100) << PAGE_SHIFT;
0119 }
0120
0121 #else
0122
0123 static inline __maybe_unused phys_addr_t cma_early_percent_memory(void)
0124 {
0125     return 0;
0126 }
0127
0128 #endif
0129
0130 #ifdef CONFIG_DMA_PERNUMA_CMA
0131 void __init dma_pernuma_cma_reserve(void)
0132 {
0133     int nid;
0134
0135     if (!pernuma_size_bytes)
0136         return;
0137
0138     for_each_online_node(nid) {
0139         int ret;
0140         char name[CMA_MAX_NAME];
0141         struct cma **cma = &dma_contiguous_pernuma_area[nid];
0142
0143         snprintf(name, sizeof(name), "pernuma%d", nid);
0144         ret = cma_declare_contiguous_nid(0, pernuma_size_bytes, 0, 0,
0145                          0, false, name, cma, nid);
0146         if (ret) {
0147             pr_warn("%s: reservation failed: err %d, node %d", __func__,
0148                 ret, nid);
0149             continue;
0150         }
0151
0152         pr_debug("%s: reserved %llu MiB on node %d\n", __func__,
0153             (unsigned long long)pernuma_size_bytes / SZ_1M, nid);
0154     }
0155 }
0156 #endif
0157
0158 /**
0159  * dma_contiguous_reserve() - reserve area(s) for contiguous memory handling
0160  * @limit: End address of the reserved memory (optional, 0 for any).
0161  *
0162  * This function reserves memory from early allocator. It should be
0163  * called by arch specific code once the early allocator (memblock or bootmem)
0164  * has been activated and all other subsystems have already allocated/reserved
0165  * memory.
0166  */
0167 void __init dma_contiguous_reserve(phys_addr_t limit)
0168 {
0169     phys_addr_t selected_size = 0;
0170     phys_addr_t selected_base = 0;
0171     phys_addr_t selected_limit = limit;
0172     bool fixed = false;
0173
0174     pr_debug("%s(limit %08lx)\n", __func__, (unsigned long)limit);
0175
0176     if (size_cmdline != -1) {
0177         selected_size = size_cmdline;
0178         selected_base = base_cmdline;
0179         selected_limit = min_not_zero(limit_cmdline, limit);
0180         if (base_cmdline + size_cmdline == limit_cmdline)
0181             fixed = true;
0182     } else {
0183 #ifdef CONFIG_CMA_SIZE_SEL_MBYTES
0184         selected_size = size_bytes;
0185 #elif defined(CONFIG_CMA_SIZE_SEL_PERCENTAGE)
0186         selected_size = cma_early_percent_memory();
0187 #elif defined(CONFIG_CMA_SIZE_SEL_MIN)
0188         selected_size = min(size_bytes, cma_early_percent_memory());
0189 #elif defined(CONFIG_CMA_SIZE_SEL_MAX)
0190         selected_size = max(size_bytes, cma_early_percent_memory());
0191 #endif
0192     }
0193
0194     if (selected_size && !dma_contiguous_default_area) {
0195         pr_debug("%s: reserving %ld MiB for global area\n", __func__,
0196              (unsigned long)selected_size / SZ_1M);
0197
0198         dma_contiguous_reserve_area(selected_size, selected_base,
0199                         selected_limit,
0200                         &dma_contiguous_default_area,
0201                         fixed);
0202     }
0203 }
0204
0205 void __weak
0206 dma_contiguous_early_fixup(phys_addr_t base, unsigned long size)
0207 {
0208 }
0209
0210 /**
0211  * dma_contiguous_reserve_area() - reserve custom contiguous area
0212  * @size: Size of the reserved area (in bytes),
0213  * @base: Base address of the reserved area optional, use 0 for any
0214  * @limit: End address of the reserved memory (optional, 0 for any).
0215  * @res_cma: Pointer to store the created cma region.
0216  * @fixed: hint about where to place the reserved area
0217  *
0218  * This function reserves memory from early allocator. It should be
0219  * called by arch specific code once the early allocator (memblock or bootmem)
0220  * has been activated and all other subsystems have already allocated/reserved
0221  * memory. This function allows to create custom reserved areas for specific
0222  * devices.
0223  *
0224  * If @fixed is true, reserve contiguous area at exactly @base.  If false,
0225  * reserve in range from @base to @limit.
0226  */
0227 int __init dma_contiguous_reserve_area(phys_addr_t size, phys_addr_t base,
0228                        phys_addr_t limit, struct cma **res_cma,
0229                        bool fixed)
0230 {
0231     int ret;
0232
0233     ret = cma_declare_contiguous(base, size, limit, 0, 0, fixed,
0234                     "reserved", res_cma);
0235     if (ret)
0236         return ret;
0237
0238     /* Architecture specific contiguous memory fixup. */
0239     dma_contiguous_early_fixup(cma_get_base(*res_cma),
0240                 cma_get_size(*res_cma));
0241
0242     return 0;
0243 }
0244
0245 /**
0246  * dma_alloc_from_contiguous() - allocate pages from contiguous area
0247  * @dev:   Pointer to device for which the allocation is performed.
0248  * @count: Requested number of pages.
0249  * @align: Requested alignment of pages (in PAGE_SIZE order).
0250  * @no_warn: Avoid printing message about failed allocation.
0251  *
0252  * This function allocates memory buffer for specified device. It uses
0253  * device specific contiguous memory area if available or the default
0254  * global one. Requires architecture specific dev_get_cma_area() helper
0255  * function.
0256  */
0257 struct page *dma_alloc_from_contiguous(struct device *dev, size_t count,
0258                        unsigned int align, bool no_warn)
0259 {
0260     if (align > CONFIG_CMA_ALIGNMENT)
0261         align = CONFIG_CMA_ALIGNMENT;
0262
0263     return cma_alloc(dev_get_cma_area(dev), count, align, no_warn);
0264 }
0265
0266 /**
0267  * dma_release_from_contiguous() - release allocated pages
0268  * @dev:   Pointer to device for which the pages were allocated.
0269  * @pages: Allocated pages.
0270  * @count: Number of allocated pages.
0271  *
0272  * This function releases memory allocated by dma_alloc_from_contiguous().
0273  * It returns false when provided pages do not belong to contiguous area and
0274  * true otherwise.
0275  */
0276 bool dma_release_from_contiguous(struct device *dev, struct page *pages,
0277                  int count)
0278 {
0279     return cma_release(dev_get_cma_area(dev), pages, count);
0280 }
0281
0282 static struct page *cma_alloc_aligned(struct cma *cma, size_t size, gfp_t gfp)
0283 {
0284     unsigned int align = min(get_order(size), CONFIG_CMA_ALIGNMENT);
0285
0286     return cma_alloc(cma, size >> PAGE_SHIFT, align, gfp & __GFP_NOWARN);
0287 }
0288
0289 /**
0290  * dma_alloc_contiguous() - allocate contiguous pages
0291  * @dev:   Pointer to device for which the allocation is performed.
0292  * @size:  Requested allocation size.
0293  * @gfp:   Allocation flags.
0294  *
0295  * tries to use device specific contiguous memory area if available, or it
0296  * tries to use per-numa cma, if the allocation fails, it will fallback to
0297  * try default global one.
0298  *
0299  * Note that it bypass one-page size of allocations from the per-numa and
0300  * global area as the addresses within one page are always contiguous, so
0301  * there is no need to waste CMA pages for that kind; it also helps reduce
0302  * fragmentations.
0303  */
0304 struct page *dma_alloc_contiguous(struct device *dev, size_t size, gfp_t gfp)
0305 {
0306 #ifdef CONFIG_DMA_PERNUMA_CMA
0307     int nid = dev_to_node(dev);
0308 #endif
0309
0310     /* CMA can be used only in the context which permits sleeping */
0311     if (!gfpflags_allow_blocking(gfp))
0312         return NULL;
0313     if (dev->cma_area)
0314         return cma_alloc_aligned(dev->cma_area, size, gfp);
0315     if (size <= PAGE_SIZE)
0316         return NULL;
0317
0318 #ifdef CONFIG_DMA_PERNUMA_CMA
0319     if (nid != NUMA_NO_NODE && !(gfp & (GFP_DMA | GFP_DMA32))) {
0320         struct cma *cma = dma_contiguous_pernuma_area[nid];
0321         struct page *page;
0322
0323         if (cma) {
0324             page = cma_alloc_aligned(cma, size, gfp);
0325             if (page)
0326                 return page;
0327         }
0328     }
0329 #endif
0330     if (!dma_contiguous_default_area)
0331         return NULL;
0332
0333     return cma_alloc_aligned(dma_contiguous_default_area, size, gfp);
0334 }
0335
0336 /**
0337  * dma_free_contiguous() - release allocated pages
0338  * @dev:   Pointer to device for which the pages were allocated.
0339  * @page:  Pointer to the allocated pages.
0340  * @size:  Size of allocated pages.
0341  *
0342  * This function releases memory allocated by dma_alloc_contiguous(). As the
0343  * cma_release returns false when provided pages do not belong to contiguous
0344  * area and true otherwise, this function then does a fallback __free_pages()
0345  * upon a false-return.
0346  */
0347 void dma_free_contiguous(struct device *dev, struct page *page, size_t size)
0348 {
0349     unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT;
0350
0351     /* if dev has its own cma, free page from there */
0352     if (dev->cma_area) {
0353         if (cma_release(dev->cma_area, page, count))
0354             return;
0355     } else {
0356         /*
0357          * otherwise, page is from either per-numa cma or default cma
0358          */
0359 #ifdef CONFIG_DMA_PERNUMA_CMA
0360         if (cma_release(dma_contiguous_pernuma_area[page_to_nid(page)],
0361                     page, count))
0362             return;
0363 #endif
0364         if (cma_release(dma_contiguous_default_area, page, count))
0365             return;
0366     }
0367
0368     /* not in any cma, free from buddy */
0369     __free_pages(page, get_order(size));
0370 }
0371
0372 /*
0373  * Support for reserved memory regions defined in device tree
0374  */
0375 #ifdef CONFIG_OF_RESERVED_MEM
0376 #include <linux/of.h>
0377 #include <linux/of_fdt.h>
0378 #include <linux/of_reserved_mem.h>
0379
0380 #undef pr_fmt
0381 #define pr_fmt(fmt) fmt
0382
0383 static int rmem_cma_device_init(struct reserved_mem *rmem, struct device *dev)
0384 {
0385     dev->cma_area = rmem->priv;
0386     return 0;
0387 }
0388
0389 static void rmem_cma_device_release(struct reserved_mem *rmem,
0390                     struct device *dev)
0391 {
0392     dev->cma_area = NULL;
0393 }
0394
0395 static const struct reserved_mem_ops rmem_cma_ops = {
0396     .device_init    = rmem_cma_device_init,
0397     .device_release = rmem_cma_device_release,
0398 };
0399
0400 static int __init rmem_cma_setup(struct reserved_mem *rmem)
0401 {
0402     unsigned long node = rmem->fdt_node;
0403     bool default_cma = of_get_flat_dt_prop(node, "linux,cma-default", NULL);
0404     struct cma *cma;
0405     int err;
0406
0407     if (size_cmdline != -1 && default_cma) {
0408         pr_info("Reserved memory: bypass %s node, using cmdline CMA params instead\n",
0409             rmem->name);
0410         return -EBUSY;
0411     }
0412
0413     if (!of_get_flat_dt_prop(node, "reusable", NULL) ||
0414         of_get_flat_dt_prop(node, "no-map", NULL))
0415         return -EINVAL;
0416
0417     if (!IS_ALIGNED(rmem->base | rmem->size, CMA_MIN_ALIGNMENT_BYTES)) {
0418         pr_err("Reserved memory: incorrect alignment of CMA region\n");
0419         return -EINVAL;
0420     }
0421
0422     err = cma_init_reserved_mem(rmem->base, rmem->size, 0, rmem->name, &cma);
0423     if (err) {
0424         pr_err("Reserved memory: unable to setup CMA region\n");
0425         return err;
0426     }
0427     /* Architecture specific contiguous memory fixup. */
0428     dma_contiguous_early_fixup(rmem->base, rmem->size);
0429
0430     if (default_cma)
0431         dma_contiguous_default_area = cma;
0432
0433     rmem->ops = &rmem_cma_ops;
0434     rmem->priv = cma;
0435
0436     pr_info("Reserved memory: created CMA memory pool at %pa, size %ld MiB\n",
0437         &rmem->base, (unsigned long)rmem->size / SZ_1M);
0438
0439     return 0;
0440 }
0441 RESERVEDMEM_OF_DECLARE(cma, "shared-dma-pool", rmem_cma_setup);
0442 #endif