| 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 | |
| 29 | #include <linux/pci.h> |
| 30 | #include <linux/vmalloc.h> |
| 31 | |
| 32 | #include <drm/radeon_drm.h> |
| 33 | #ifdef CONFIG_X86 |
| 34 | #include <asm/set_memory.h> |
| 35 | #endif |
| 36 | #include "radeon.h" |
| 37 | |
| 38 | /* |
| 39 | * GART |
| 40 | * The GART (Graphics Aperture Remapping Table) is an aperture |
| 41 | * in the GPU's address space. System pages can be mapped into |
| 42 | * the aperture and look like contiguous pages from the GPU's |
| 43 | * perspective. A page table maps the pages in the aperture |
| 44 | * to the actual backing pages in system memory. |
| 45 | * |
| 46 | * Radeon GPUs support both an internal GART, as described above, |
| 47 | * and AGP. AGP works similarly, but the GART table is configured |
| 48 | * and maintained by the northbridge rather than the driver. |
| 49 | * Radeon hw has a separate AGP aperture that is programmed to |
| 50 | * point to the AGP aperture provided by the northbridge and the |
| 51 | * requests are passed through to the northbridge aperture. |
| 52 | * Both AGP and internal GART can be used at the same time, however |
| 53 | * that is not currently supported by the driver. |
| 54 | * |
| 55 | * This file handles the common internal GART management. |
| 56 | */ |
| 57 | |
| 58 | /* |
| 59 | * Common GART table functions. |
| 60 | */ |
| 61 | /** |
| 62 | * radeon_gart_table_ram_alloc - allocate system ram for gart page table |
| 63 | * |
| 64 | * @rdev: radeon_device pointer |
| 65 | * |
| 66 | * Allocate system memory for GART page table |
| 67 | * (r1xx-r3xx, non-pcie r4xx, rs400). These asics require the |
| 68 | * gart table to be in system memory. |
| 69 | * Returns 0 for success, -ENOMEM for failure. |
| 70 | */ |
| 71 | int radeon_gart_table_ram_alloc(struct radeon_device *rdev) |
| 72 | { |
| 73 | void *ptr; |
| 74 | |
| 75 | ptr = dma_alloc_coherent(dev: &rdev->pdev->dev, size: rdev->gart.table_size, |
| 76 | dma_handle: &rdev->gart.table_addr, GFP_KERNEL); |
| 77 | if (!ptr) |
| 78 | return -ENOMEM; |
| 79 | |
| 80 | #ifdef CONFIG_X86 |
| 81 | if (rdev->family == CHIP_RS400 || rdev->family == CHIP_RS480 || |
| 82 | rdev->family == CHIP_RS690 || rdev->family == CHIP_RS740) { |
| 83 | set_memory_uc(addr: (unsigned long)ptr, |
| 84 | numpages: rdev->gart.table_size >> PAGE_SHIFT); |
| 85 | } |
| 86 | #endif |
| 87 | rdev->gart.ptr = ptr; |
| 88 | return 0; |
| 89 | } |
| 90 | |
| 91 | /** |
| 92 | * radeon_gart_table_ram_free - free system ram for gart page table |
| 93 | * |
| 94 | * @rdev: radeon_device pointer |
| 95 | * |
| 96 | * Free system memory for GART page table |
| 97 | * (r1xx-r3xx, non-pcie r4xx, rs400). These asics require the |
| 98 | * gart table to be in system memory. |
| 99 | */ |
| 100 | void radeon_gart_table_ram_free(struct radeon_device *rdev) |
| 101 | { |
| 102 | if (!rdev->gart.ptr) |
| 103 | return; |
| 104 | |
| 105 | #ifdef CONFIG_X86 |
| 106 | if (rdev->family == CHIP_RS400 || rdev->family == CHIP_RS480 || |
| 107 | rdev->family == CHIP_RS690 || rdev->family == CHIP_RS740) { |
| 108 | set_memory_wb(addr: (unsigned long)rdev->gart.ptr, |
| 109 | numpages: rdev->gart.table_size >> PAGE_SHIFT); |
| 110 | } |
| 111 | #endif |
| 112 | dma_free_coherent(dev: &rdev->pdev->dev, size: rdev->gart.table_size, |
| 113 | cpu_addr: (void *)rdev->gart.ptr, dma_handle: rdev->gart.table_addr); |
| 114 | rdev->gart.ptr = NULL; |
| 115 | rdev->gart.table_addr = 0; |
| 116 | } |
| 117 | |
| 118 | /** |
| 119 | * radeon_gart_table_vram_alloc - allocate vram for gart page table |
| 120 | * |
| 121 | * @rdev: radeon_device pointer |
| 122 | * |
| 123 | * Allocate video memory for GART page table |
| 124 | * (pcie r4xx, r5xx+). These asics require the |
| 125 | * gart table to be in video memory. |
| 126 | * Returns 0 for success, error for failure. |
| 127 | */ |
| 128 | int radeon_gart_table_vram_alloc(struct radeon_device *rdev) |
| 129 | { |
| 130 | int r; |
| 131 | |
| 132 | if (rdev->gart.robj == NULL) { |
| 133 | r = radeon_bo_create(rdev, size: rdev->gart.table_size, |
| 134 | PAGE_SIZE, kernel: true, RADEON_GEM_DOMAIN_VRAM, |
| 135 | flags: 0, NULL, NULL, bo_ptr: &rdev->gart.robj); |
| 136 | if (r) |
| 137 | return r; |
| 138 | } |
| 139 | return 0; |
| 140 | } |
| 141 | |
| 142 | /** |
| 143 | * radeon_gart_table_vram_pin - pin gart page table in vram |
| 144 | * |
| 145 | * @rdev: radeon_device pointer |
| 146 | * |
| 147 | * Pin the GART page table in vram so it will not be moved |
| 148 | * by the memory manager (pcie r4xx, r5xx+). These asics require the |
| 149 | * gart table to be in video memory. |
| 150 | * Returns 0 for success, error for failure. |
| 151 | */ |
| 152 | int radeon_gart_table_vram_pin(struct radeon_device *rdev) |
| 153 | { |
| 154 | uint64_t gpu_addr; |
| 155 | int r; |
| 156 | |
| 157 | r = radeon_bo_reserve(bo: rdev->gart.robj, no_intr: false); |
| 158 | if (unlikely(r != 0)) |
| 159 | return r; |
| 160 | r = radeon_bo_pin(bo: rdev->gart.robj, |
| 161 | RADEON_GEM_DOMAIN_VRAM, gpu_addr: &gpu_addr); |
| 162 | if (r) { |
| 163 | radeon_bo_unreserve(bo: rdev->gart.robj); |
| 164 | return r; |
| 165 | } |
| 166 | r = radeon_bo_kmap(bo: rdev->gart.robj, ptr: &rdev->gart.ptr); |
| 167 | if (r) |
| 168 | radeon_bo_unpin(bo: rdev->gart.robj); |
| 169 | radeon_bo_unreserve(bo: rdev->gart.robj); |
| 170 | rdev->gart.table_addr = gpu_addr; |
| 171 | |
| 172 | if (!r) { |
| 173 | int i; |
| 174 | |
| 175 | /* We might have dropped some GART table updates while it wasn't |
| 176 | * mapped, restore all entries |
| 177 | */ |
| 178 | for (i = 0; i < rdev->gart.num_gpu_pages; i++) |
| 179 | radeon_gart_set_page(rdev, i, rdev->gart.pages_entry[i]); |
| 180 | mb(); |
| 181 | radeon_gart_tlb_flush(rdev); |
| 182 | } |
| 183 | |
| 184 | return r; |
| 185 | } |
| 186 | |
| 187 | /** |
| 188 | * radeon_gart_table_vram_unpin - unpin gart page table in vram |
| 189 | * |
| 190 | * @rdev: radeon_device pointer |
| 191 | * |
| 192 | * Unpin the GART page table in vram (pcie r4xx, r5xx+). |
| 193 | * These asics require the gart table to be in video memory. |
| 194 | */ |
| 195 | void radeon_gart_table_vram_unpin(struct radeon_device *rdev) |
| 196 | { |
| 197 | int r; |
| 198 | |
| 199 | if (!rdev->gart.robj) |
| 200 | return; |
| 201 | |
| 202 | r = radeon_bo_reserve(bo: rdev->gart.robj, no_intr: false); |
| 203 | if (likely(r == 0)) { |
| 204 | radeon_bo_kunmap(bo: rdev->gart.robj); |
| 205 | radeon_bo_unpin(bo: rdev->gart.robj); |
| 206 | radeon_bo_unreserve(bo: rdev->gart.robj); |
| 207 | rdev->gart.ptr = NULL; |
| 208 | } |
| 209 | } |
| 210 | |
| 211 | /** |
| 212 | * radeon_gart_table_vram_free - free gart page table vram |
| 213 | * |
| 214 | * @rdev: radeon_device pointer |
| 215 | * |
| 216 | * Free the video memory used for the GART page table |
| 217 | * (pcie r4xx, r5xx+). These asics require the gart table to |
| 218 | * be in video memory. |
| 219 | */ |
| 220 | void radeon_gart_table_vram_free(struct radeon_device *rdev) |
| 221 | { |
| 222 | if (!rdev->gart.robj) |
| 223 | return; |
| 224 | |
| 225 | radeon_bo_unref(bo: &rdev->gart.robj); |
| 226 | } |
| 227 | |
| 228 | /* |
| 229 | * Common gart functions. |
| 230 | */ |
| 231 | /** |
| 232 | * radeon_gart_unbind - unbind pages from the gart page table |
| 233 | * |
| 234 | * @rdev: radeon_device pointer |
| 235 | * @offset: offset into the GPU's gart aperture |
| 236 | * @pages: number of pages to unbind |
| 237 | * |
| 238 | * Unbinds the requested pages from the gart page table and |
| 239 | * replaces them with the dummy page (all asics). |
| 240 | */ |
| 241 | void radeon_gart_unbind(struct radeon_device *rdev, unsigned int offset, |
| 242 | int pages) |
| 243 | { |
| 244 | unsigned int t, p; |
| 245 | int i, j; |
| 246 | |
| 247 | if (!rdev->gart.ready) { |
| 248 | WARN(1, "trying to unbind memory from uninitialized GART !\n" ); |
| 249 | return; |
| 250 | } |
| 251 | t = offset / RADEON_GPU_PAGE_SIZE; |
| 252 | p = t / (PAGE_SIZE / RADEON_GPU_PAGE_SIZE); |
| 253 | for (i = 0; i < pages; i++, p++) { |
| 254 | if (rdev->gart.pages[p]) { |
| 255 | rdev->gart.pages[p] = NULL; |
| 256 | for (j = 0; j < (PAGE_SIZE / RADEON_GPU_PAGE_SIZE); j++, t++) { |
| 257 | rdev->gart.pages_entry[t] = rdev->dummy_page.entry; |
| 258 | if (rdev->gart.ptr) { |
| 259 | radeon_gart_set_page(rdev, t, |
| 260 | rdev->dummy_page.entry); |
| 261 | } |
| 262 | } |
| 263 | } |
| 264 | } |
| 265 | if (rdev->gart.ptr) { |
| 266 | mb(); |
| 267 | radeon_gart_tlb_flush(rdev); |
| 268 | } |
| 269 | } |
| 270 | |
| 271 | /** |
| 272 | * radeon_gart_bind - bind pages into the gart page table |
| 273 | * |
| 274 | * @rdev: radeon_device pointer |
| 275 | * @offset: offset into the GPU's gart aperture |
| 276 | * @pages: number of pages to bind |
| 277 | * @pagelist: pages to bind |
| 278 | * @dma_addr: DMA addresses of pages |
| 279 | * @flags: RADEON_GART_PAGE_* flags |
| 280 | * |
| 281 | * Binds the requested pages to the gart page table |
| 282 | * (all asics). |
| 283 | * Returns 0 for success, -EINVAL for failure. |
| 284 | */ |
| 285 | int radeon_gart_bind(struct radeon_device *rdev, unsigned int offset, |
| 286 | int pages, struct page **pagelist, dma_addr_t *dma_addr, |
| 287 | uint32_t flags) |
| 288 | { |
| 289 | unsigned int t, p; |
| 290 | uint64_t page_base, page_entry; |
| 291 | int i, j; |
| 292 | |
| 293 | if (!rdev->gart.ready) { |
| 294 | WARN(1, "trying to bind memory to uninitialized GART !\n" ); |
| 295 | return -EINVAL; |
| 296 | } |
| 297 | t = offset / RADEON_GPU_PAGE_SIZE; |
| 298 | p = t / (PAGE_SIZE / RADEON_GPU_PAGE_SIZE); |
| 299 | |
| 300 | for (i = 0; i < pages; i++, p++) { |
| 301 | rdev->gart.pages[p] = pagelist ? pagelist[i] : |
| 302 | rdev->dummy_page.page; |
| 303 | page_base = dma_addr[i]; |
| 304 | for (j = 0; j < (PAGE_SIZE / RADEON_GPU_PAGE_SIZE); j++, t++) { |
| 305 | page_entry = radeon_gart_get_page_entry(page_base, flags); |
| 306 | rdev->gart.pages_entry[t] = page_entry; |
| 307 | if (rdev->gart.ptr) |
| 308 | radeon_gart_set_page(rdev, t, page_entry); |
| 309 | |
| 310 | page_base += RADEON_GPU_PAGE_SIZE; |
| 311 | } |
| 312 | } |
| 313 | if (rdev->gart.ptr) { |
| 314 | mb(); |
| 315 | radeon_gart_tlb_flush(rdev); |
| 316 | } |
| 317 | return 0; |
| 318 | } |
| 319 | |
| 320 | /** |
| 321 | * radeon_gart_init - init the driver info for managing the gart |
| 322 | * |
| 323 | * @rdev: radeon_device pointer |
| 324 | * |
| 325 | * Allocate the dummy page and init the gart driver info (all asics). |
| 326 | * Returns 0 for success, error for failure. |
| 327 | */ |
| 328 | int radeon_gart_init(struct radeon_device *rdev) |
| 329 | { |
| 330 | int r, i; |
| 331 | |
| 332 | if (rdev->gart.pages) |
| 333 | return 0; |
| 334 | |
| 335 | /* We need PAGE_SIZE >= RADEON_GPU_PAGE_SIZE */ |
| 336 | if (PAGE_SIZE < RADEON_GPU_PAGE_SIZE) { |
| 337 | DRM_ERROR("Page size is smaller than GPU page size!\n" ); |
| 338 | return -EINVAL; |
| 339 | } |
| 340 | r = radeon_dummy_page_init(rdev); |
| 341 | if (r) |
| 342 | return r; |
| 343 | /* Compute table size */ |
| 344 | rdev->gart.num_cpu_pages = rdev->mc.gtt_size / PAGE_SIZE; |
| 345 | rdev->gart.num_gpu_pages = rdev->mc.gtt_size / RADEON_GPU_PAGE_SIZE; |
| 346 | DRM_INFO("GART: num cpu pages %u, num gpu pages %u\n" , |
| 347 | rdev->gart.num_cpu_pages, rdev->gart.num_gpu_pages); |
| 348 | /* Allocate pages table */ |
| 349 | rdev->gart.pages = vcalloc(rdev->gart.num_cpu_pages, |
| 350 | sizeof(void *)); |
| 351 | if (rdev->gart.pages == NULL) { |
| 352 | radeon_gart_fini(rdev); |
| 353 | return -ENOMEM; |
| 354 | } |
| 355 | rdev->gart.pages_entry = vmalloc_array(rdev->gart.num_gpu_pages, |
| 356 | sizeof(uint64_t)); |
| 357 | if (rdev->gart.pages_entry == NULL) { |
| 358 | radeon_gart_fini(rdev); |
| 359 | return -ENOMEM; |
| 360 | } |
| 361 | /* set GART entry to point to the dummy page by default */ |
| 362 | for (i = 0; i < rdev->gart.num_gpu_pages; i++) |
| 363 | rdev->gart.pages_entry[i] = rdev->dummy_page.entry; |
| 364 | return 0; |
| 365 | } |
| 366 | |
| 367 | /** |
| 368 | * radeon_gart_fini - tear down the driver info for managing the gart |
| 369 | * |
| 370 | * @rdev: radeon_device pointer |
| 371 | * |
| 372 | * Tear down the gart driver info and free the dummy page (all asics). |
| 373 | */ |
| 374 | void radeon_gart_fini(struct radeon_device *rdev) |
| 375 | { |
| 376 | if (rdev->gart.ready) { |
| 377 | /* unbind pages */ |
| 378 | radeon_gart_unbind(rdev, offset: 0, pages: rdev->gart.num_cpu_pages); |
| 379 | } |
| 380 | rdev->gart.ready = false; |
| 381 | vfree(addr: rdev->gart.pages); |
| 382 | vfree(addr: rdev->gart.pages_entry); |
| 383 | rdev->gart.pages = NULL; |
| 384 | rdev->gart.pages_entry = NULL; |
| 385 | |
| 386 | radeon_dummy_page_fini(rdev); |
| 387 | } |
| 388 | |