Commit | Line | Data |
---|---|---|
d38ceaf9 AD |
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> | |
ed3ba079 LA |
30 | #ifdef CONFIG_X86 |
31 | #include <asm/set_memory.h> | |
32 | #endif | |
d38ceaf9 AD |
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 | * amdgpu_gart_table_ram_alloc - allocate system ram for gart page table | |
60 | * | |
61 | * @adev: amdgpu_device pointer | |
62 | * | |
63 | * Allocate system memory for GART page table | |
64 | * (r1xx-r3xx, non-pcie r4xx, rs400). These asics require the | |
65 | * gart table to be in system memory. | |
66 | * Returns 0 for success, -ENOMEM for failure. | |
67 | */ | |
68 | int amdgpu_gart_table_ram_alloc(struct amdgpu_device *adev) | |
69 | { | |
70 | void *ptr; | |
71 | ||
72 | ptr = pci_alloc_consistent(adev->pdev, adev->gart.table_size, | |
73 | &adev->gart.table_addr); | |
74 | if (ptr == NULL) { | |
75 | return -ENOMEM; | |
76 | } | |
77 | #ifdef CONFIG_X86 | |
78 | if (0) { | |
79 | set_memory_uc((unsigned long)ptr, | |
80 | adev->gart.table_size >> PAGE_SHIFT); | |
81 | } | |
82 | #endif | |
83 | adev->gart.ptr = ptr; | |
84 | memset((void *)adev->gart.ptr, 0, adev->gart.table_size); | |
85 | return 0; | |
86 | } | |
87 | ||
88 | /** | |
89 | * amdgpu_gart_table_ram_free - free system ram for gart page table | |
90 | * | |
91 | * @adev: amdgpu_device pointer | |
92 | * | |
93 | * Free system memory for GART page table | |
94 | * (r1xx-r3xx, non-pcie r4xx, rs400). These asics require the | |
95 | * gart table to be in system memory. | |
96 | */ | |
97 | void amdgpu_gart_table_ram_free(struct amdgpu_device *adev) | |
98 | { | |
99 | if (adev->gart.ptr == NULL) { | |
100 | return; | |
101 | } | |
102 | #ifdef CONFIG_X86 | |
103 | if (0) { | |
104 | set_memory_wb((unsigned long)adev->gart.ptr, | |
105 | adev->gart.table_size >> PAGE_SHIFT); | |
106 | } | |
107 | #endif | |
108 | pci_free_consistent(adev->pdev, adev->gart.table_size, | |
109 | (void *)adev->gart.ptr, | |
110 | adev->gart.table_addr); | |
111 | adev->gart.ptr = NULL; | |
112 | adev->gart.table_addr = 0; | |
113 | } | |
114 | ||
115 | /** | |
116 | * amdgpu_gart_table_vram_alloc - allocate vram for gart page table | |
117 | * | |
118 | * @adev: amdgpu_device pointer | |
119 | * | |
120 | * Allocate video memory for GART page table | |
121 | * (pcie r4xx, r5xx+). These asics require the | |
122 | * gart table to be in video memory. | |
123 | * Returns 0 for success, error for failure. | |
124 | */ | |
125 | int amdgpu_gart_table_vram_alloc(struct amdgpu_device *adev) | |
126 | { | |
127 | int r; | |
128 | ||
129 | if (adev->gart.robj == NULL) { | |
130 | r = amdgpu_bo_create(adev, adev->gart.table_size, | |
857d913d | 131 | PAGE_SIZE, true, AMDGPU_GEM_DOMAIN_VRAM, |
03f48dd5 CK |
132 | AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED | |
133 | AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS, | |
72d7668b | 134 | NULL, NULL, &adev->gart.robj); |
d38ceaf9 AD |
135 | if (r) { |
136 | return r; | |
137 | } | |
138 | } | |
139 | return 0; | |
140 | } | |
141 | ||
142 | /** | |
143 | * amdgpu_gart_table_vram_pin - pin gart page table in vram | |
144 | * | |
145 | * @adev: amdgpu_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 amdgpu_gart_table_vram_pin(struct amdgpu_device *adev) | |
153 | { | |
154 | uint64_t gpu_addr; | |
155 | int r; | |
156 | ||
157 | r = amdgpu_bo_reserve(adev->gart.robj, false); | |
158 | if (unlikely(r != 0)) | |
159 | return r; | |
160 | r = amdgpu_bo_pin(adev->gart.robj, | |
161 | AMDGPU_GEM_DOMAIN_VRAM, &gpu_addr); | |
162 | if (r) { | |
163 | amdgpu_bo_unreserve(adev->gart.robj); | |
164 | return r; | |
165 | } | |
166 | r = amdgpu_bo_kmap(adev->gart.robj, &adev->gart.ptr); | |
167 | if (r) | |
168 | amdgpu_bo_unpin(adev->gart.robj); | |
169 | amdgpu_bo_unreserve(adev->gart.robj); | |
170 | adev->gart.table_addr = gpu_addr; | |
171 | return r; | |
172 | } | |
173 | ||
174 | /** | |
175 | * amdgpu_gart_table_vram_unpin - unpin gart page table in vram | |
176 | * | |
177 | * @adev: amdgpu_device pointer | |
178 | * | |
179 | * Unpin the GART page table in vram (pcie r4xx, r5xx+). | |
180 | * These asics require the gart table to be in video memory. | |
181 | */ | |
182 | void amdgpu_gart_table_vram_unpin(struct amdgpu_device *adev) | |
183 | { | |
184 | int r; | |
185 | ||
186 | if (adev->gart.robj == NULL) { | |
187 | return; | |
188 | } | |
c81a1a74 | 189 | r = amdgpu_bo_reserve(adev->gart.robj, true); |
d38ceaf9 AD |
190 | if (likely(r == 0)) { |
191 | amdgpu_bo_kunmap(adev->gart.robj); | |
192 | amdgpu_bo_unpin(adev->gart.robj); | |
193 | amdgpu_bo_unreserve(adev->gart.robj); | |
194 | adev->gart.ptr = NULL; | |
195 | } | |
196 | } | |
197 | ||
198 | /** | |
199 | * amdgpu_gart_table_vram_free - free gart page table vram | |
200 | * | |
201 | * @adev: amdgpu_device pointer | |
202 | * | |
203 | * Free the video memory used for the GART page table | |
204 | * (pcie r4xx, r5xx+). These asics require the gart table to | |
205 | * be in video memory. | |
206 | */ | |
207 | void amdgpu_gart_table_vram_free(struct amdgpu_device *adev) | |
208 | { | |
209 | if (adev->gart.robj == NULL) { | |
210 | return; | |
211 | } | |
212 | amdgpu_bo_unref(&adev->gart.robj); | |
213 | } | |
214 | ||
215 | /* | |
216 | * Common gart functions. | |
217 | */ | |
218 | /** | |
219 | * amdgpu_gart_unbind - unbind pages from the gart page table | |
220 | * | |
221 | * @adev: amdgpu_device pointer | |
222 | * @offset: offset into the GPU's gart aperture | |
223 | * @pages: number of pages to unbind | |
224 | * | |
225 | * Unbinds the requested pages from the gart page table and | |
226 | * replaces them with the dummy page (all asics). | |
227 | */ | |
cab0b8d5 | 228 | void amdgpu_gart_unbind(struct amdgpu_device *adev, uint64_t offset, |
d38ceaf9 AD |
229 | int pages) |
230 | { | |
231 | unsigned t; | |
232 | unsigned p; | |
233 | int i, j; | |
234 | u64 page_base; | |
a0676f60 AD |
235 | /* Starting from VEGA10, system bit must be 0 to mean invalid. */ |
236 | uint64_t flags = 0; | |
d38ceaf9 AD |
237 | |
238 | if (!adev->gart.ready) { | |
239 | WARN(1, "trying to unbind memory from uninitialized GART !\n"); | |
240 | return; | |
241 | } | |
242 | ||
243 | t = offset / AMDGPU_GPU_PAGE_SIZE; | |
244 | p = t / (PAGE_SIZE / AMDGPU_GPU_PAGE_SIZE); | |
245 | for (i = 0; i < pages; i++, p++) { | |
186294f9 | 246 | #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS |
a1d29476 CK |
247 | adev->gart.pages[p] = NULL; |
248 | #endif | |
249 | page_base = adev->dummy_page.addr; | |
250 | if (!adev->gart.ptr) | |
251 | continue; | |
d38ceaf9 | 252 | |
a1d29476 CK |
253 | for (j = 0; j < (PAGE_SIZE / AMDGPU_GPU_PAGE_SIZE); j++, t++) { |
254 | amdgpu_gart_set_pte_pde(adev, adev->gart.ptr, | |
255 | t, page_base, flags); | |
256 | page_base += AMDGPU_GPU_PAGE_SIZE; | |
d38ceaf9 AD |
257 | } |
258 | } | |
259 | mb(); | |
260 | amdgpu_gart_flush_gpu_tlb(adev, 0); | |
261 | } | |
262 | ||
263 | /** | |
264 | * amdgpu_gart_bind - bind pages into the gart page table | |
265 | * | |
266 | * @adev: amdgpu_device pointer | |
267 | * @offset: offset into the GPU's gart aperture | |
268 | * @pages: number of pages to bind | |
269 | * @pagelist: pages to bind | |
270 | * @dma_addr: DMA addresses of pages | |
271 | * | |
272 | * Binds the requested pages to the gart page table | |
273 | * (all asics). | |
274 | * Returns 0 for success, -EINVAL for failure. | |
275 | */ | |
cab0b8d5 | 276 | int amdgpu_gart_bind(struct amdgpu_device *adev, uint64_t offset, |
d38ceaf9 | 277 | int pages, struct page **pagelist, dma_addr_t *dma_addr, |
6b777607 | 278 | uint64_t flags) |
d38ceaf9 AD |
279 | { |
280 | unsigned t; | |
281 | unsigned p; | |
282 | uint64_t page_base; | |
283 | int i, j; | |
284 | ||
285 | if (!adev->gart.ready) { | |
286 | WARN(1, "trying to bind memory to uninitialized GART !\n"); | |
287 | return -EINVAL; | |
288 | } | |
289 | ||
290 | t = offset / AMDGPU_GPU_PAGE_SIZE; | |
291 | p = t / (PAGE_SIZE / AMDGPU_GPU_PAGE_SIZE); | |
292 | ||
293 | for (i = 0; i < pages; i++, p++) { | |
186294f9 | 294 | #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS |
d38ceaf9 | 295 | adev->gart.pages[p] = pagelist[i]; |
a1d29476 | 296 | #endif |
d38ceaf9 | 297 | if (adev->gart.ptr) { |
43251981 | 298 | page_base = dma_addr[i]; |
d38ceaf9 AD |
299 | for (j = 0; j < (PAGE_SIZE / AMDGPU_GPU_PAGE_SIZE); j++, t++) { |
300 | amdgpu_gart_set_pte_pde(adev, adev->gart.ptr, t, page_base, flags); | |
301 | page_base += AMDGPU_GPU_PAGE_SIZE; | |
302 | } | |
303 | } | |
304 | } | |
305 | mb(); | |
306 | amdgpu_gart_flush_gpu_tlb(adev, 0); | |
307 | return 0; | |
308 | } | |
309 | ||
310 | /** | |
311 | * amdgpu_gart_init - init the driver info for managing the gart | |
312 | * | |
313 | * @adev: amdgpu_device pointer | |
314 | * | |
315 | * Allocate the dummy page and init the gart driver info (all asics). | |
316 | * Returns 0 for success, error for failure. | |
317 | */ | |
318 | int amdgpu_gart_init(struct amdgpu_device *adev) | |
319 | { | |
43251981 | 320 | int r; |
d38ceaf9 | 321 | |
a1d29476 | 322 | if (adev->dummy_page.page) |
d38ceaf9 | 323 | return 0; |
a1d29476 | 324 | |
d38ceaf9 AD |
325 | /* We need PAGE_SIZE >= AMDGPU_GPU_PAGE_SIZE */ |
326 | if (PAGE_SIZE < AMDGPU_GPU_PAGE_SIZE) { | |
327 | DRM_ERROR("Page size is smaller than GPU page size!\n"); | |
328 | return -EINVAL; | |
329 | } | |
330 | r = amdgpu_dummy_page_init(adev); | |
331 | if (r) | |
332 | return r; | |
333 | /* Compute table size */ | |
334 | adev->gart.num_cpu_pages = adev->mc.gtt_size / PAGE_SIZE; | |
335 | adev->gart.num_gpu_pages = adev->mc.gtt_size / AMDGPU_GPU_PAGE_SIZE; | |
336 | DRM_INFO("GART: num cpu pages %u, num gpu pages %u\n", | |
337 | adev->gart.num_cpu_pages, adev->gart.num_gpu_pages); | |
a1d29476 | 338 | |
186294f9 | 339 | #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS |
d38ceaf9 AD |
340 | /* Allocate pages table */ |
341 | adev->gart.pages = vzalloc(sizeof(void *) * adev->gart.num_cpu_pages); | |
342 | if (adev->gart.pages == NULL) { | |
343 | amdgpu_gart_fini(adev); | |
344 | return -ENOMEM; | |
345 | } | |
a1d29476 CK |
346 | #endif |
347 | ||
d38ceaf9 AD |
348 | return 0; |
349 | } | |
350 | ||
351 | /** | |
352 | * amdgpu_gart_fini - tear down the driver info for managing the gart | |
353 | * | |
354 | * @adev: amdgpu_device pointer | |
355 | * | |
356 | * Tear down the gart driver info and free the dummy page (all asics). | |
357 | */ | |
358 | void amdgpu_gart_fini(struct amdgpu_device *adev) | |
359 | { | |
a1d29476 | 360 | if (adev->gart.ready) { |
d38ceaf9 AD |
361 | /* unbind pages */ |
362 | amdgpu_gart_unbind(adev, 0, adev->gart.num_cpu_pages); | |
363 | } | |
364 | adev->gart.ready = false; | |
186294f9 | 365 | #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS |
d38ceaf9 | 366 | vfree(adev->gart.pages); |
d38ceaf9 | 367 | adev->gart.pages = NULL; |
a1d29476 | 368 | #endif |
d38ceaf9 AD |
369 | amdgpu_dummy_page_fini(adev); |
370 | } |