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 56 struct 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) 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(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 } 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(s, &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(p, &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(name, sizeof(name), "pernuma%d", nid); 178 ret = cma_declare_contiguous_nid(0, pernuma_size_bytes, 0, 0, 179 0, false, name, 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(name, sizeof(name), "numa%d", nid); 189 ret = cma_declare_contiguous_nid(0, numa_cma_size[nid], 0, 0, 0, false, 190 name, 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; 226 227 /* Hornor the user setup dma address limit */ 228 selected_limit = limit_cmdline ?: limit; 229 230 if (base_cmdline + size_cmdline == limit_cmdline) 231 fixed = true; 232 } else { 233 #ifdef CONFIG_CMA_SIZE_SEL_MBYTES 234 selected_size = size_bytes; 235 #elif defined(CONFIG_CMA_SIZE_SEL_PERCENTAGE) 236 selected_size = cma_early_percent_memory(); 237 #elif defined(CONFIG_CMA_SIZE_SEL_MIN) 238 selected_size = min(size_bytes, cma_early_percent_memory()); 239 #elif defined(CONFIG_CMA_SIZE_SEL_MAX) 240 selected_size = max(size_bytes, cma_early_percent_memory()); 241 #endif 242 } 243 244 if (selected_size && !dma_contiguous_default_area) { 245 int ret; 246 247 pr_debug("%s: reserving %ld MiB for global area\n", __func__, 248 (unsigned long)selected_size / SZ_1M); 249 250 dma_contiguous_reserve_area(selected_size, selected_base, 251 selected_limit, 252 &dma_contiguous_default_area, 253 fixed); 254 255 ret = dma_heap_cma_register_heap(dma_contiguous_default_area); 256 if (ret) 257 pr_warn("Couldn't register default CMA heap."); 258 } 259 } 260 261 void __weak 262 dma_contiguous_early_fixup(phys_addr_t base, unsigned long size) 263 { 264 } 265 266 /** 267 * dma_contiguous_reserve_area() - reserve custom contiguous area 268 * @size: Size of the reserved area (in bytes), 269 * @base: Base address of the reserved area optional, use 0 for any 270 * @limit: End address of the reserved memory (optional, 0 for any). 271 * @res_cma: Pointer to store the created cma region. 272 * @fixed: hint about where to place the reserved area 273 * 274 * This function reserves memory from early allocator. It should be 275 * called by arch specific code once the early allocator (memblock or bootmem) 276 * has been activated and all other subsystems have already allocated/reserved 277 * memory. This function allows to create custom reserved areas for specific 278 * devices. 279 * 280 * If @fixed is true, reserve contiguous area at exactly @base. If false, 281 * reserve in range from @base to @limit. 282 */ 283 int __init dma_contiguous_reserve_area(phys_addr_t size, phys_addr_t base, 284 phys_addr_t limit, struct cma **res_cma, 285 bool fixed) 286 { 287 int ret; 288 289 ret = cma_declare_contiguous(base, size, limit, 0, 0, fixed, 290 "reserved", res_cma); 291 if (ret) 292 return ret; 293 294 /* Architecture specific contiguous memory fixup. */ 295 dma_contiguous_early_fixup(cma_get_base(*res_cma), 296 cma_get_size(*res_cma)); 297 298 return 0; 299 } 300 301 /** 302 * dma_alloc_from_contiguous() - allocate pages from contiguous area 303 * @dev: Pointer to device for which the allocation is performed. 304 * @count: Requested number of pages. 305 * @align: Requested alignment of pages (in PAGE_SIZE order). 306 * @no_warn: Avoid printing message about failed allocation. 307 * 308 * This function allocates memory buffer for specified device. It uses 309 * device specific contiguous memory area if available or the default 310 * global one. Requires architecture specific dev_get_cma_area() helper 311 * function. 312 */ 313 struct page *dma_alloc_from_contiguous(struct device *dev, size_t count, 314 unsigned int align, bool no_warn) 315 { 316 if (align > CONFIG_CMA_ALIGNMENT) 317 align = CONFIG_CMA_ALIGNMENT; 318 319 return cma_alloc(dev_get_cma_area(dev), count, align, no_warn); 320 } 321 322 /** 323 * dma_release_from_contiguous() - release allocated pages 324 * @dev: Pointer to device for which the pages were allocated. 325 * @pages: Allocated pages. 326 * @count: Number of allocated pages. 327 * 328 * This function releases memory allocated by dma_alloc_from_contiguous(). 329 * It returns false when provided pages do not belong to contiguous area and 330 * true otherwise. 331 */ 332 bool dma_release_from_contiguous(struct device *dev, struct page *pages, 333 int count) 334 { 335 return cma_release(dev_get_cma_area(dev), pages, count); 336 } 337 338 static struct page *cma_alloc_aligned(struct cma *cma, size_t size, gfp_t gfp) 339 { 340 unsigned int align = min(get_order(size), CONFIG_CMA_ALIGNMENT); 341 342 return cma_alloc(cma, size >> PAGE_SHIFT, align, gfp & __GFP_NOWARN); 343 } 344 345 /** 346 * dma_alloc_contiguous() - allocate contiguous pages 347 * @dev: Pointer to device for which the allocation is performed. 348 * @size: Requested allocation size. 349 * @gfp: Allocation flags. 350 * 351 * tries to use device specific contiguous memory area if available, or it 352 * tries to use per-numa cma, if the allocation fails, it will fallback to 353 * try default global one. 354 * 355 * Note that it bypass one-page size of allocations from the per-numa and 356 * global area as the addresses within one page are always contiguous, so 357 * there is no need to waste CMA pages for that kind; it also helps reduce 358 * fragmentations. 359 */ 360 struct page *dma_alloc_contiguous(struct device *dev, size_t size, gfp_t gfp) 361 { 362 #ifdef CONFIG_DMA_NUMA_CMA 363 int nid = dev_to_node(dev); 364 #endif 365 366 /* CMA can be used only in the context which permits sleeping */ 367 if (!gfpflags_allow_blocking(gfp)) 368 return NULL; 369 if (dev->cma_area) 370 return cma_alloc_aligned(dev->cma_area, size, gfp); 371 if (size <= PAGE_SIZE) 372 return NULL; 373 374 #ifdef CONFIG_DMA_NUMA_CMA 375 if (nid != NUMA_NO_NODE && !(gfp & (GFP_DMA | GFP_DMA32))) { 376 struct cma *cma = dma_contiguous_pernuma_area[nid]; 377 struct page *page; 378 379 if (cma) { 380 page = cma_alloc_aligned(cma, size, gfp); 381 if (page) 382 return page; 383 } 384 385 cma = dma_contiguous_numa_area[nid]; 386 if (cma) { 387 page = cma_alloc_aligned(cma, size, gfp); 388 if (page) 389 return page; 390 } 391 } 392 #endif 393 if (!dma_contiguous_default_area) 394 return NULL; 395 396 return cma_alloc_aligned(dma_contiguous_default_area, size, gfp); 397 } 398 399 /** 400 * dma_free_contiguous() - release allocated pages 401 * @dev: Pointer to device for which the pages were allocated. 402 * @page: Pointer to the allocated pages. 403 * @size: Size of allocated pages. 404 * 405 * This function releases memory allocated by dma_alloc_contiguous(). As the 406 * cma_release returns false when provided pages do not belong to contiguous 407 * area and true otherwise, this function then does a fallback __free_pages() 408 * upon a false-return. 409 */ 410 void dma_free_contiguous(struct device *dev, struct page *page, size_t size) 411 { 412 unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT; 413 414 /* if dev has its own cma, free page from there */ 415 if (dev->cma_area) { 416 if (cma_release(dev->cma_area, page, count)) 417 return; 418 } else { 419 /* 420 * otherwise, page is from either per-numa cma or default cma 421 */ 422 #ifdef CONFIG_DMA_NUMA_CMA 423 if (cma_release(dma_contiguous_pernuma_area[page_to_nid(page)], 424 page, count)) 425 return; 426 if (cma_release(dma_contiguous_numa_area[page_to_nid(page)], 427 page, count)) 428 return; 429 #endif 430 if (cma_release(dma_contiguous_default_area, page, count)) 431 return; 432 } 433 434 /* not in any cma, free from buddy */ 435 __free_pages(page, get_order(size)); 436 } 437 438 /* 439 * Support for reserved memory regions defined in device tree 440 */ 441 #ifdef CONFIG_OF_RESERVED_MEM 442 #include <linux/of.h> 443 #include <linux/of_fdt.h> 444 #include <linux/of_reserved_mem.h> 445 446 #undef pr_fmt 447 #define pr_fmt(fmt) fmt 448 449 static int rmem_cma_device_init(struct reserved_mem *rmem, struct device *dev) 450 { 451 dev->cma_area = rmem->priv; 452 return 0; 453 } 454 455 static void rmem_cma_device_release(struct reserved_mem *rmem, 456 struct device *dev) 457 { 458 dev->cma_area = NULL; 459 } 460 461 static const struct reserved_mem_ops rmem_cma_ops = { 462 .device_init = rmem_cma_device_init, 463 .device_release = rmem_cma_device_release, 464 }; 465 466 static int __init rmem_cma_setup(struct reserved_mem *rmem) 467 { 468 unsigned long node = rmem->fdt_node; 469 bool default_cma = of_get_flat_dt_prop(node, "linux,cma-default", NULL); 470 struct cma *cma; 471 int err; 472 473 if (size_cmdline != -1 && default_cma) { 474 pr_info("Reserved memory: bypass %s node, using cmdline CMA params instead\n", 475 rmem->name); 476 return -EBUSY; 477 } 478 479 if (!of_get_flat_dt_prop(node, "reusable", NULL) || 480 of_get_flat_dt_prop(node, "no-map", NULL)) 481 return -EINVAL; 482 483 if (!IS_ALIGNED(rmem->base | rmem->size, CMA_MIN_ALIGNMENT_BYTES)) { 484 pr_err("Reserved memory: incorrect alignment of CMA region\n"); 485 return -EINVAL; 486 } 487 488 err = cma_init_reserved_mem(rmem->base, rmem->size, 0, rmem->name, &cma); 489 if (err) { 490 pr_err("Reserved memory: unable to setup CMA region\n"); 491 return err; 492 } 493 494 if (default_cma) 495 dma_contiguous_default_area = cma; 496 497 rmem->ops = &rmem_cma_ops; 498 rmem->priv = cma; 499 500 pr_info("Reserved memory: created CMA memory pool at %pa, size %ld MiB\n", 501 &rmem->base, (unsigned long)rmem->size / SZ_1M); 502 503 err = dma_heap_cma_register_heap(cma); 504 if (err) 505 pr_warn("Couldn't register CMA heap."); 506 507 return 0; 508 } 509 RESERVEDMEM_OF_DECLARE(cma, "shared-dma-pool", rmem_cma_setup); 510 #endif 511