Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 1 | // SPDX-License-Identifier: GPL-2.0 |
| 2 | /* |
| 3 | * Coherent per-device memory handling. |
| 4 | * Borrowed from i386 |
| 5 | */ |
| 6 | #include <linux/io.h> |
| 7 | #include <linux/slab.h> |
| 8 | #include <linux/kernel.h> |
| 9 | #include <linux/module.h> |
| 10 | #include <linux/dma-mapping.h> |
| 11 | |
| 12 | struct dma_coherent_mem { |
| 13 | void *virt_base; |
| 14 | dma_addr_t device_base; |
| 15 | unsigned long pfn_base; |
| 16 | int size; |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 17 | unsigned long *bitmap; |
| 18 | spinlock_t spinlock; |
| 19 | bool use_dev_dma_pfn_offset; |
| 20 | }; |
| 21 | |
| 22 | static struct dma_coherent_mem *dma_coherent_default_memory __ro_after_init; |
| 23 | |
| 24 | static inline struct dma_coherent_mem *dev_get_coherent_memory(struct device *dev) |
| 25 | { |
| 26 | if (dev && dev->dma_mem) |
| 27 | return dev->dma_mem; |
| 28 | return NULL; |
| 29 | } |
| 30 | |
| 31 | static inline dma_addr_t dma_get_device_base(struct device *dev, |
| 32 | struct dma_coherent_mem * mem) |
| 33 | { |
| 34 | if (mem->use_dev_dma_pfn_offset) |
| 35 | return (mem->pfn_base - dev->dma_pfn_offset) << PAGE_SHIFT; |
| 36 | else |
| 37 | return mem->device_base; |
| 38 | } |
| 39 | |
David Brazdil | 0f672f6 | 2019-12-10 10:32:29 +0000 | [diff] [blame^] | 40 | static int dma_init_coherent_memory(phys_addr_t phys_addr, |
| 41 | dma_addr_t device_addr, size_t size, |
| 42 | struct dma_coherent_mem **mem) |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 43 | { |
| 44 | struct dma_coherent_mem *dma_mem = NULL; |
David Brazdil | 0f672f6 | 2019-12-10 10:32:29 +0000 | [diff] [blame^] | 45 | void *mem_base = NULL; |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 46 | int pages = size >> PAGE_SHIFT; |
| 47 | int bitmap_size = BITS_TO_LONGS(pages) * sizeof(long); |
| 48 | int ret; |
| 49 | |
| 50 | if (!size) { |
| 51 | ret = -EINVAL; |
| 52 | goto out; |
| 53 | } |
| 54 | |
| 55 | mem_base = memremap(phys_addr, size, MEMREMAP_WC); |
| 56 | if (!mem_base) { |
| 57 | ret = -EINVAL; |
| 58 | goto out; |
| 59 | } |
| 60 | dma_mem = kzalloc(sizeof(struct dma_coherent_mem), GFP_KERNEL); |
| 61 | if (!dma_mem) { |
| 62 | ret = -ENOMEM; |
| 63 | goto out; |
| 64 | } |
| 65 | dma_mem->bitmap = kzalloc(bitmap_size, GFP_KERNEL); |
| 66 | if (!dma_mem->bitmap) { |
| 67 | ret = -ENOMEM; |
| 68 | goto out; |
| 69 | } |
| 70 | |
| 71 | dma_mem->virt_base = mem_base; |
| 72 | dma_mem->device_base = device_addr; |
| 73 | dma_mem->pfn_base = PFN_DOWN(phys_addr); |
| 74 | dma_mem->size = pages; |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 75 | spin_lock_init(&dma_mem->spinlock); |
| 76 | |
| 77 | *mem = dma_mem; |
| 78 | return 0; |
| 79 | |
| 80 | out: |
| 81 | kfree(dma_mem); |
| 82 | if (mem_base) |
| 83 | memunmap(mem_base); |
| 84 | return ret; |
| 85 | } |
| 86 | |
| 87 | static void dma_release_coherent_memory(struct dma_coherent_mem *mem) |
| 88 | { |
| 89 | if (!mem) |
| 90 | return; |
| 91 | |
| 92 | memunmap(mem->virt_base); |
| 93 | kfree(mem->bitmap); |
| 94 | kfree(mem); |
| 95 | } |
| 96 | |
| 97 | static int dma_assign_coherent_memory(struct device *dev, |
| 98 | struct dma_coherent_mem *mem) |
| 99 | { |
| 100 | if (!dev) |
| 101 | return -ENODEV; |
| 102 | |
| 103 | if (dev->dma_mem) |
| 104 | return -EBUSY; |
| 105 | |
| 106 | dev->dma_mem = mem; |
| 107 | return 0; |
| 108 | } |
| 109 | |
| 110 | int dma_declare_coherent_memory(struct device *dev, phys_addr_t phys_addr, |
David Brazdil | 0f672f6 | 2019-12-10 10:32:29 +0000 | [diff] [blame^] | 111 | dma_addr_t device_addr, size_t size) |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 112 | { |
| 113 | struct dma_coherent_mem *mem; |
| 114 | int ret; |
| 115 | |
David Brazdil | 0f672f6 | 2019-12-10 10:32:29 +0000 | [diff] [blame^] | 116 | ret = dma_init_coherent_memory(phys_addr, device_addr, size, &mem); |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 117 | if (ret) |
| 118 | return ret; |
| 119 | |
| 120 | ret = dma_assign_coherent_memory(dev, mem); |
| 121 | if (ret) |
| 122 | dma_release_coherent_memory(mem); |
| 123 | return ret; |
| 124 | } |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 125 | |
| 126 | static void *__dma_alloc_from_coherent(struct dma_coherent_mem *mem, |
| 127 | ssize_t size, dma_addr_t *dma_handle) |
| 128 | { |
| 129 | int order = get_order(size); |
| 130 | unsigned long flags; |
| 131 | int pageno; |
| 132 | void *ret; |
| 133 | |
| 134 | spin_lock_irqsave(&mem->spinlock, flags); |
| 135 | |
| 136 | if (unlikely(size > (mem->size << PAGE_SHIFT))) |
| 137 | goto err; |
| 138 | |
| 139 | pageno = bitmap_find_free_region(mem->bitmap, mem->size, order); |
| 140 | if (unlikely(pageno < 0)) |
| 141 | goto err; |
| 142 | |
| 143 | /* |
| 144 | * Memory was found in the coherent area. |
| 145 | */ |
| 146 | *dma_handle = mem->device_base + (pageno << PAGE_SHIFT); |
| 147 | ret = mem->virt_base + (pageno << PAGE_SHIFT); |
| 148 | spin_unlock_irqrestore(&mem->spinlock, flags); |
| 149 | memset(ret, 0, size); |
| 150 | return ret; |
| 151 | err: |
| 152 | spin_unlock_irqrestore(&mem->spinlock, flags); |
| 153 | return NULL; |
| 154 | } |
| 155 | |
| 156 | /** |
| 157 | * dma_alloc_from_dev_coherent() - allocate memory from device coherent pool |
| 158 | * @dev: device from which we allocate memory |
| 159 | * @size: size of requested memory area |
| 160 | * @dma_handle: This will be filled with the correct dma handle |
| 161 | * @ret: This pointer will be filled with the virtual address |
| 162 | * to allocated area. |
| 163 | * |
| 164 | * This function should be only called from per-arch dma_alloc_coherent() |
| 165 | * to support allocation from per-device coherent memory pools. |
| 166 | * |
| 167 | * Returns 0 if dma_alloc_coherent should continue with allocating from |
| 168 | * generic memory areas, or !0 if dma_alloc_coherent should return @ret. |
| 169 | */ |
| 170 | int dma_alloc_from_dev_coherent(struct device *dev, ssize_t size, |
| 171 | dma_addr_t *dma_handle, void **ret) |
| 172 | { |
| 173 | struct dma_coherent_mem *mem = dev_get_coherent_memory(dev); |
| 174 | |
| 175 | if (!mem) |
| 176 | return 0; |
| 177 | |
| 178 | *ret = __dma_alloc_from_coherent(mem, size, dma_handle); |
David Brazdil | 0f672f6 | 2019-12-10 10:32:29 +0000 | [diff] [blame^] | 179 | return 1; |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 180 | } |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 181 | |
| 182 | void *dma_alloc_from_global_coherent(ssize_t size, dma_addr_t *dma_handle) |
| 183 | { |
| 184 | if (!dma_coherent_default_memory) |
| 185 | return NULL; |
| 186 | |
| 187 | return __dma_alloc_from_coherent(dma_coherent_default_memory, size, |
| 188 | dma_handle); |
| 189 | } |
| 190 | |
| 191 | static int __dma_release_from_coherent(struct dma_coherent_mem *mem, |
| 192 | int order, void *vaddr) |
| 193 | { |
| 194 | if (mem && vaddr >= mem->virt_base && vaddr < |
| 195 | (mem->virt_base + (mem->size << PAGE_SHIFT))) { |
| 196 | int page = (vaddr - mem->virt_base) >> PAGE_SHIFT; |
| 197 | unsigned long flags; |
| 198 | |
| 199 | spin_lock_irqsave(&mem->spinlock, flags); |
| 200 | bitmap_release_region(mem->bitmap, page, order); |
| 201 | spin_unlock_irqrestore(&mem->spinlock, flags); |
| 202 | return 1; |
| 203 | } |
| 204 | return 0; |
| 205 | } |
| 206 | |
| 207 | /** |
| 208 | * dma_release_from_dev_coherent() - free memory to device coherent memory pool |
| 209 | * @dev: device from which the memory was allocated |
| 210 | * @order: the order of pages allocated |
| 211 | * @vaddr: virtual address of allocated pages |
| 212 | * |
| 213 | * This checks whether the memory was allocated from the per-device |
| 214 | * coherent memory pool and if so, releases that memory. |
| 215 | * |
| 216 | * Returns 1 if we correctly released the memory, or 0 if the caller should |
| 217 | * proceed with releasing memory from generic pools. |
| 218 | */ |
| 219 | int dma_release_from_dev_coherent(struct device *dev, int order, void *vaddr) |
| 220 | { |
| 221 | struct dma_coherent_mem *mem = dev_get_coherent_memory(dev); |
| 222 | |
| 223 | return __dma_release_from_coherent(mem, order, vaddr); |
| 224 | } |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 225 | |
| 226 | int dma_release_from_global_coherent(int order, void *vaddr) |
| 227 | { |
| 228 | if (!dma_coherent_default_memory) |
| 229 | return 0; |
| 230 | |
| 231 | return __dma_release_from_coherent(dma_coherent_default_memory, order, |
| 232 | vaddr); |
| 233 | } |
| 234 | |
| 235 | static int __dma_mmap_from_coherent(struct dma_coherent_mem *mem, |
| 236 | struct vm_area_struct *vma, void *vaddr, size_t size, int *ret) |
| 237 | { |
| 238 | if (mem && vaddr >= mem->virt_base && vaddr + size <= |
| 239 | (mem->virt_base + (mem->size << PAGE_SHIFT))) { |
| 240 | unsigned long off = vma->vm_pgoff; |
| 241 | int start = (vaddr - mem->virt_base) >> PAGE_SHIFT; |
| 242 | int user_count = vma_pages(vma); |
| 243 | int count = PAGE_ALIGN(size) >> PAGE_SHIFT; |
| 244 | |
| 245 | *ret = -ENXIO; |
| 246 | if (off < count && user_count <= count - off) { |
| 247 | unsigned long pfn = mem->pfn_base + start + off; |
| 248 | *ret = remap_pfn_range(vma, vma->vm_start, pfn, |
| 249 | user_count << PAGE_SHIFT, |
| 250 | vma->vm_page_prot); |
| 251 | } |
| 252 | return 1; |
| 253 | } |
| 254 | return 0; |
| 255 | } |
| 256 | |
| 257 | /** |
| 258 | * dma_mmap_from_dev_coherent() - mmap memory from the device coherent pool |
| 259 | * @dev: device from which the memory was allocated |
| 260 | * @vma: vm_area for the userspace memory |
| 261 | * @vaddr: cpu address returned by dma_alloc_from_dev_coherent |
| 262 | * @size: size of the memory buffer allocated |
| 263 | * @ret: result from remap_pfn_range() |
| 264 | * |
| 265 | * This checks whether the memory was allocated from the per-device |
| 266 | * coherent memory pool and if so, maps that memory to the provided vma. |
| 267 | * |
| 268 | * Returns 1 if @vaddr belongs to the device coherent pool and the caller |
| 269 | * should return @ret, or 0 if they should proceed with mapping memory from |
| 270 | * generic areas. |
| 271 | */ |
| 272 | int dma_mmap_from_dev_coherent(struct device *dev, struct vm_area_struct *vma, |
| 273 | void *vaddr, size_t size, int *ret) |
| 274 | { |
| 275 | struct dma_coherent_mem *mem = dev_get_coherent_memory(dev); |
| 276 | |
| 277 | return __dma_mmap_from_coherent(mem, vma, vaddr, size, ret); |
| 278 | } |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 279 | |
| 280 | int dma_mmap_from_global_coherent(struct vm_area_struct *vma, void *vaddr, |
| 281 | size_t size, int *ret) |
| 282 | { |
| 283 | if (!dma_coherent_default_memory) |
| 284 | return 0; |
| 285 | |
| 286 | return __dma_mmap_from_coherent(dma_coherent_default_memory, vma, |
| 287 | vaddr, size, ret); |
| 288 | } |
| 289 | |
| 290 | /* |
| 291 | * Support for reserved memory regions defined in device tree |
| 292 | */ |
| 293 | #ifdef CONFIG_OF_RESERVED_MEM |
| 294 | #include <linux/of.h> |
| 295 | #include <linux/of_fdt.h> |
| 296 | #include <linux/of_reserved_mem.h> |
| 297 | |
| 298 | static struct reserved_mem *dma_reserved_default_memory __initdata; |
| 299 | |
| 300 | static int rmem_dma_device_init(struct reserved_mem *rmem, struct device *dev) |
| 301 | { |
| 302 | struct dma_coherent_mem *mem = rmem->priv; |
| 303 | int ret; |
| 304 | |
| 305 | if (!mem) { |
| 306 | ret = dma_init_coherent_memory(rmem->base, rmem->base, |
David Brazdil | 0f672f6 | 2019-12-10 10:32:29 +0000 | [diff] [blame^] | 307 | rmem->size, &mem); |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 308 | if (ret) { |
| 309 | pr_err("Reserved memory: failed to init DMA memory pool at %pa, size %ld MiB\n", |
| 310 | &rmem->base, (unsigned long)rmem->size / SZ_1M); |
| 311 | return ret; |
| 312 | } |
| 313 | } |
| 314 | mem->use_dev_dma_pfn_offset = true; |
| 315 | rmem->priv = mem; |
| 316 | dma_assign_coherent_memory(dev, mem); |
| 317 | return 0; |
| 318 | } |
| 319 | |
| 320 | static void rmem_dma_device_release(struct reserved_mem *rmem, |
| 321 | struct device *dev) |
| 322 | { |
| 323 | if (dev) |
| 324 | dev->dma_mem = NULL; |
| 325 | } |
| 326 | |
| 327 | static const struct reserved_mem_ops rmem_dma_ops = { |
| 328 | .device_init = rmem_dma_device_init, |
| 329 | .device_release = rmem_dma_device_release, |
| 330 | }; |
| 331 | |
| 332 | static int __init rmem_dma_setup(struct reserved_mem *rmem) |
| 333 | { |
| 334 | unsigned long node = rmem->fdt_node; |
| 335 | |
| 336 | if (of_get_flat_dt_prop(node, "reusable", NULL)) |
| 337 | return -EINVAL; |
| 338 | |
| 339 | #ifdef CONFIG_ARM |
| 340 | if (!of_get_flat_dt_prop(node, "no-map", NULL)) { |
| 341 | pr_err("Reserved memory: regions without no-map are not yet supported\n"); |
| 342 | return -EINVAL; |
| 343 | } |
| 344 | |
| 345 | if (of_get_flat_dt_prop(node, "linux,dma-default", NULL)) { |
| 346 | WARN(dma_reserved_default_memory, |
| 347 | "Reserved memory: region for default DMA coherent area is redefined\n"); |
| 348 | dma_reserved_default_memory = rmem; |
| 349 | } |
| 350 | #endif |
| 351 | |
| 352 | rmem->ops = &rmem_dma_ops; |
| 353 | pr_info("Reserved memory: created DMA memory pool at %pa, size %ld MiB\n", |
| 354 | &rmem->base, (unsigned long)rmem->size / SZ_1M); |
| 355 | return 0; |
| 356 | } |
| 357 | |
| 358 | static int __init dma_init_reserved_memory(void) |
| 359 | { |
| 360 | const struct reserved_mem_ops *ops; |
| 361 | int ret; |
| 362 | |
| 363 | if (!dma_reserved_default_memory) |
| 364 | return -ENOMEM; |
| 365 | |
| 366 | ops = dma_reserved_default_memory->ops; |
| 367 | |
| 368 | /* |
| 369 | * We rely on rmem_dma_device_init() does not propagate error of |
| 370 | * dma_assign_coherent_memory() for "NULL" device. |
| 371 | */ |
| 372 | ret = ops->device_init(dma_reserved_default_memory, NULL); |
| 373 | |
| 374 | if (!ret) { |
| 375 | dma_coherent_default_memory = dma_reserved_default_memory->priv; |
| 376 | pr_info("DMA: default coherent area is set\n"); |
| 377 | } |
| 378 | |
| 379 | return ret; |
| 380 | } |
| 381 | |
| 382 | core_initcall(dma_init_reserved_memory); |
| 383 | |
| 384 | RESERVEDMEM_OF_DECLARE(dma, "shared-dma-pool", rmem_dma_setup); |
| 385 | #endif |