1 /* 2 * Copyright 2008 Advanced Micro Devices, Inc. 3 * Copyright 2008 Red Hat Inc. 4 * Copyright 2009 Jerome Glisse. 5 * 6 * Permission is hereby granted, free of charge, to any person obtaining a 7 * copy of this software and associated documentation files (the "Software"), 8 * to deal in the Software without restriction, including without limitation 9 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 10 * and/or sell copies of the Software, and to permit persons to whom the 11 * Software is furnished to do so, subject to the following conditions: 12 * 13 * The above copyright notice and this permission notice shall be included in 14 * all copies or substantial portions of the Software. 15 * 16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 22 * OTHER DEALINGS IN THE SOFTWARE. 23 * 24 * Authors: Dave Airlie 25 * Alex Deucher 26 * Jerome Glisse 27 */ 28 #include <drm/drmP.h> 29 #include <drm/amdgpu_drm.h> 30 #ifdef CONFIG_X86 31 #include <asm/set_memory.h> 32 #endif 33 #include "amdgpu.h" 34 35 /* 36 * GART 37 * The GART (Graphics Aperture Remapping Table) is an aperture 38 * in the GPU's address space. System pages can be mapped into 39 * the aperture and look like contiguous pages from the GPU's 40 * perspective. A page table maps the pages in the aperture 41 * to the actual backing pages in system memory. 42 * 43 * Radeon GPUs support both an internal GART, as described above, 44 * and AGP. AGP works similarly, but the GART table is configured 45 * and maintained by the northbridge rather than the driver. 46 * Radeon hw has a separate AGP aperture that is programmed to 47 * point to the AGP aperture provided by the northbridge and the 48 * requests are passed through to the northbridge aperture. 49 * Both AGP and internal GART can be used at the same time, however 50 * that is not currently supported by the driver. 51 * 52 * This file handles the common internal GART management. 53 */ 54 55 /* 56 * Common GART table functions. 57 */ 58 59 /** 60 * amdgpu_dummy_page_init - init dummy page used by the driver 61 * 62 * @adev: amdgpu_device pointer 63 * 64 * Allocate the dummy page used by the driver (all asics). 65 * This dummy page is used by the driver as a filler for gart entries 66 * when pages are taken out of the GART 67 * Returns 0 on sucess, -ENOMEM on failure. 68 */ 69 static int amdgpu_gart_dummy_page_init(struct amdgpu_device *adev) 70 { 71 if (adev->dummy_page.page) 72 return 0; 73 adev->dummy_page.page = alloc_page(GFP_DMA32 | GFP_KERNEL | __GFP_ZERO); 74 if (adev->dummy_page.page == NULL) 75 return -ENOMEM; 76 adev->dummy_page.addr = pci_map_page(adev->pdev, adev->dummy_page.page, 77 0, PAGE_SIZE, PCI_DMA_BIDIRECTIONAL); 78 if (pci_dma_mapping_error(adev->pdev, adev->dummy_page.addr)) { 79 dev_err(&adev->pdev->dev, "Failed to DMA MAP the dummy page\n"); 80 __free_page(adev->dummy_page.page); 81 adev->dummy_page.page = NULL; 82 return -ENOMEM; 83 } 84 return 0; 85 } 86 87 /** 88 * amdgpu_dummy_page_fini - free dummy page used by the driver 89 * 90 * @adev: amdgpu_device pointer 91 * 92 * Frees the dummy page used by the driver (all asics). 93 */ 94 static void amdgpu_gart_dummy_page_fini(struct amdgpu_device *adev) 95 { 96 if (adev->dummy_page.page == NULL) 97 return; 98 pci_unmap_page(adev->pdev, adev->dummy_page.addr, 99 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL); 100 __free_page(adev->dummy_page.page); 101 adev->dummy_page.page = NULL; 102 } 103 104 /** 105 * amdgpu_gart_table_vram_alloc - allocate vram for gart page table 106 * 107 * @adev: amdgpu_device pointer 108 * 109 * Allocate video memory for GART page table 110 * (pcie r4xx, r5xx+). These asics require the 111 * gart table to be in video memory. 112 * Returns 0 for success, error for failure. 113 */ 114 int amdgpu_gart_table_vram_alloc(struct amdgpu_device *adev) 115 { 116 int r; 117 118 if (adev->gart.robj == NULL) { 119 r = amdgpu_bo_create(adev, adev->gart.table_size, 120 PAGE_SIZE, true, AMDGPU_GEM_DOMAIN_VRAM, 121 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED | 122 AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS, 123 NULL, NULL, 0, &adev->gart.robj); 124 if (r) { 125 return r; 126 } 127 } 128 return 0; 129 } 130 131 /** 132 * amdgpu_gart_table_vram_pin - pin gart page table in vram 133 * 134 * @adev: amdgpu_device pointer 135 * 136 * Pin the GART page table in vram so it will not be moved 137 * by the memory manager (pcie r4xx, r5xx+). These asics require the 138 * gart table to be in video memory. 139 * Returns 0 for success, error for failure. 140 */ 141 int amdgpu_gart_table_vram_pin(struct amdgpu_device *adev) 142 { 143 uint64_t gpu_addr; 144 int r; 145 146 r = amdgpu_bo_reserve(adev->gart.robj, false); 147 if (unlikely(r != 0)) 148 return r; 149 r = amdgpu_bo_pin(adev->gart.robj, 150 AMDGPU_GEM_DOMAIN_VRAM, &gpu_addr); 151 if (r) { 152 amdgpu_bo_unreserve(adev->gart.robj); 153 return r; 154 } 155 r = amdgpu_bo_kmap(adev->gart.robj, &adev->gart.ptr); 156 if (r) 157 amdgpu_bo_unpin(adev->gart.robj); 158 amdgpu_bo_unreserve(adev->gart.robj); 159 adev->gart.table_addr = gpu_addr; 160 return r; 161 } 162 163 /** 164 * amdgpu_gart_table_vram_unpin - unpin gart page table in vram 165 * 166 * @adev: amdgpu_device pointer 167 * 168 * Unpin the GART page table in vram (pcie r4xx, r5xx+). 169 * These asics require the gart table to be in video memory. 170 */ 171 void amdgpu_gart_table_vram_unpin(struct amdgpu_device *adev) 172 { 173 int r; 174 175 if (adev->gart.robj == NULL) { 176 return; 177 } 178 r = amdgpu_bo_reserve(adev->gart.robj, true); 179 if (likely(r == 0)) { 180 amdgpu_bo_kunmap(adev->gart.robj); 181 amdgpu_bo_unpin(adev->gart.robj); 182 amdgpu_bo_unreserve(adev->gart.robj); 183 adev->gart.ptr = NULL; 184 } 185 } 186 187 /** 188 * amdgpu_gart_table_vram_free - free gart page table vram 189 * 190 * @adev: amdgpu_device pointer 191 * 192 * Free the video memory used for the GART page table 193 * (pcie r4xx, r5xx+). These asics require the gart table to 194 * be in video memory. 195 */ 196 void amdgpu_gart_table_vram_free(struct amdgpu_device *adev) 197 { 198 if (adev->gart.robj == NULL) { 199 return; 200 } 201 amdgpu_bo_unref(&adev->gart.robj); 202 } 203 204 /* 205 * Common gart functions. 206 */ 207 /** 208 * amdgpu_gart_unbind - unbind pages from the gart page table 209 * 210 * @adev: amdgpu_device pointer 211 * @offset: offset into the GPU's gart aperture 212 * @pages: number of pages to unbind 213 * 214 * Unbinds the requested pages from the gart page table and 215 * replaces them with the dummy page (all asics). 216 * Returns 0 for success, -EINVAL for failure. 217 */ 218 int amdgpu_gart_unbind(struct amdgpu_device *adev, uint64_t offset, 219 int pages) 220 { 221 unsigned t; 222 unsigned p; 223 int i, j; 224 u64 page_base; 225 /* Starting from VEGA10, system bit must be 0 to mean invalid. */ 226 uint64_t flags = 0; 227 228 if (!adev->gart.ready) { 229 WARN(1, "trying to unbind memory from uninitialized GART !\n"); 230 return -EINVAL; 231 } 232 233 t = offset / AMDGPU_GPU_PAGE_SIZE; 234 p = t / (PAGE_SIZE / AMDGPU_GPU_PAGE_SIZE); 235 for (i = 0; i < pages; i++, p++) { 236 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS 237 adev->gart.pages[p] = NULL; 238 #endif 239 page_base = adev->dummy_page.addr; 240 if (!adev->gart.ptr) 241 continue; 242 243 for (j = 0; j < (PAGE_SIZE / AMDGPU_GPU_PAGE_SIZE); j++, t++) { 244 amdgpu_gmc_set_pte_pde(adev, adev->gart.ptr, 245 t, page_base, flags); 246 page_base += AMDGPU_GPU_PAGE_SIZE; 247 } 248 } 249 mb(); 250 amdgpu_asic_flush_hdp(adev); 251 amdgpu_gmc_flush_gpu_tlb(adev, 0); 252 return 0; 253 } 254 255 /** 256 * amdgpu_gart_map - map dma_addresses into GART entries 257 * 258 * @adev: amdgpu_device pointer 259 * @offset: offset into the GPU's gart aperture 260 * @pages: number of pages to bind 261 * @dma_addr: DMA addresses of pages 262 * 263 * Map the dma_addresses into GART entries (all asics). 264 * Returns 0 for success, -EINVAL for failure. 265 */ 266 int amdgpu_gart_map(struct amdgpu_device *adev, uint64_t offset, 267 int pages, dma_addr_t *dma_addr, uint64_t flags, 268 void *dst) 269 { 270 uint64_t page_base; 271 unsigned i, j, t; 272 273 if (!adev->gart.ready) { 274 WARN(1, "trying to bind memory to uninitialized GART !\n"); 275 return -EINVAL; 276 } 277 278 t = offset / AMDGPU_GPU_PAGE_SIZE; 279 280 for (i = 0; i < pages; i++) { 281 page_base = dma_addr[i]; 282 for (j = 0; j < (PAGE_SIZE / AMDGPU_GPU_PAGE_SIZE); j++, t++) { 283 amdgpu_gmc_set_pte_pde(adev, dst, t, page_base, flags); 284 page_base += AMDGPU_GPU_PAGE_SIZE; 285 } 286 } 287 return 0; 288 } 289 290 /** 291 * amdgpu_gart_bind - bind pages into the gart page table 292 * 293 * @adev: amdgpu_device pointer 294 * @offset: offset into the GPU's gart aperture 295 * @pages: number of pages to bind 296 * @pagelist: pages to bind 297 * @dma_addr: DMA addresses of pages 298 * 299 * Binds the requested pages to the gart page table 300 * (all asics). 301 * Returns 0 for success, -EINVAL for failure. 302 */ 303 int amdgpu_gart_bind(struct amdgpu_device *adev, uint64_t offset, 304 int pages, struct page **pagelist, dma_addr_t *dma_addr, 305 uint64_t flags) 306 { 307 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS 308 unsigned i,t,p; 309 #endif 310 int r; 311 312 if (!adev->gart.ready) { 313 WARN(1, "trying to bind memory to uninitialized GART !\n"); 314 return -EINVAL; 315 } 316 317 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS 318 t = offset / AMDGPU_GPU_PAGE_SIZE; 319 p = t / (PAGE_SIZE / AMDGPU_GPU_PAGE_SIZE); 320 for (i = 0; i < pages; i++, p++) 321 adev->gart.pages[p] = pagelist[i]; 322 #endif 323 324 if (!adev->gart.ptr) 325 return 0; 326 327 r = amdgpu_gart_map(adev, offset, pages, dma_addr, flags, 328 adev->gart.ptr); 329 if (r) 330 return r; 331 332 mb(); 333 amdgpu_asic_flush_hdp(adev); 334 amdgpu_gmc_flush_gpu_tlb(adev, 0); 335 return 0; 336 } 337 338 /** 339 * amdgpu_gart_init - init the driver info for managing the gart 340 * 341 * @adev: amdgpu_device pointer 342 * 343 * Allocate the dummy page and init the gart driver info (all asics). 344 * Returns 0 for success, error for failure. 345 */ 346 int amdgpu_gart_init(struct amdgpu_device *adev) 347 { 348 int r; 349 350 if (adev->dummy_page.page) 351 return 0; 352 353 /* We need PAGE_SIZE >= AMDGPU_GPU_PAGE_SIZE */ 354 if (PAGE_SIZE < AMDGPU_GPU_PAGE_SIZE) { 355 DRM_ERROR("Page size is smaller than GPU page size!\n"); 356 return -EINVAL; 357 } 358 r = amdgpu_gart_dummy_page_init(adev); 359 if (r) 360 return r; 361 /* Compute table size */ 362 adev->gart.num_cpu_pages = adev->gmc.gart_size / PAGE_SIZE; 363 adev->gart.num_gpu_pages = adev->gmc.gart_size / AMDGPU_GPU_PAGE_SIZE; 364 DRM_INFO("GART: num cpu pages %u, num gpu pages %u\n", 365 adev->gart.num_cpu_pages, adev->gart.num_gpu_pages); 366 367 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS 368 /* Allocate pages table */ 369 adev->gart.pages = vzalloc(sizeof(void *) * adev->gart.num_cpu_pages); 370 if (adev->gart.pages == NULL) 371 return -ENOMEM; 372 #endif 373 374 return 0; 375 } 376 377 /** 378 * amdgpu_gart_fini - tear down the driver info for managing the gart 379 * 380 * @adev: amdgpu_device pointer 381 * 382 * Tear down the gart driver info and free the dummy page (all asics). 383 */ 384 void amdgpu_gart_fini(struct amdgpu_device *adev) 385 { 386 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS 387 vfree(adev->gart.pages); 388 adev->gart.pages = NULL; 389 #endif 390 amdgpu_gart_dummy_page_fini(adev); 391 } 392