]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - drivers/gpu/drm/amd/amdgpu/amdgpu_gart.c
Merge tag 'perf-urgent-for-mingo-4.14-20170928' of git://git.kernel.org/pub/scm/linux...
[mirror_ubuntu-bionic-kernel.git] / drivers / gpu / drm / amd / amdgpu / amdgpu_gart.c
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_gart_table_ram_alloc - allocate system ram for gart page table
61 *
62 * @adev: amdgpu_device pointer
63 *
64 * Allocate system memory for GART page table
65 * (r1xx-r3xx, non-pcie r4xx, rs400). These asics require the
66 * gart table to be in system memory.
67 * Returns 0 for success, -ENOMEM for failure.
68 */
69 int amdgpu_gart_table_ram_alloc(struct amdgpu_device *adev)
70 {
71 void *ptr;
72
73 ptr = pci_alloc_consistent(adev->pdev, adev->gart.table_size,
74 &adev->gart.table_addr);
75 if (ptr == NULL) {
76 return -ENOMEM;
77 }
78 #ifdef CONFIG_X86
79 if (0) {
80 set_memory_uc((unsigned long)ptr,
81 adev->gart.table_size >> PAGE_SHIFT);
82 }
83 #endif
84 adev->gart.ptr = ptr;
85 memset((void *)adev->gart.ptr, 0, adev->gart.table_size);
86 return 0;
87 }
88
89 /**
90 * amdgpu_gart_table_ram_free - free system ram for gart page table
91 *
92 * @adev: amdgpu_device pointer
93 *
94 * Free system memory for GART page table
95 * (r1xx-r3xx, non-pcie r4xx, rs400). These asics require the
96 * gart table to be in system memory.
97 */
98 void amdgpu_gart_table_ram_free(struct amdgpu_device *adev)
99 {
100 if (adev->gart.ptr == NULL) {
101 return;
102 }
103 #ifdef CONFIG_X86
104 if (0) {
105 set_memory_wb((unsigned long)adev->gart.ptr,
106 adev->gart.table_size >> PAGE_SHIFT);
107 }
108 #endif
109 pci_free_consistent(adev->pdev, adev->gart.table_size,
110 (void *)adev->gart.ptr,
111 adev->gart.table_addr);
112 adev->gart.ptr = NULL;
113 adev->gart.table_addr = 0;
114 }
115
116 /**
117 * amdgpu_gart_table_vram_alloc - allocate vram for gart page table
118 *
119 * @adev: amdgpu_device pointer
120 *
121 * Allocate video memory for GART page table
122 * (pcie r4xx, r5xx+). These asics require the
123 * gart table to be in video memory.
124 * Returns 0 for success, error for failure.
125 */
126 int amdgpu_gart_table_vram_alloc(struct amdgpu_device *adev)
127 {
128 int r;
129
130 if (adev->gart.robj == NULL) {
131 r = amdgpu_bo_create(adev, adev->gart.table_size,
132 PAGE_SIZE, true, AMDGPU_GEM_DOMAIN_VRAM,
133 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
134 AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS,
135 NULL, NULL, 0, &adev->gart.robj);
136 if (r) {
137 return r;
138 }
139 }
140 return 0;
141 }
142
143 /**
144 * amdgpu_gart_table_vram_pin - pin gart page table in vram
145 *
146 * @adev: amdgpu_device pointer
147 *
148 * Pin the GART page table in vram so it will not be moved
149 * by the memory manager (pcie r4xx, r5xx+). These asics require the
150 * gart table to be in video memory.
151 * Returns 0 for success, error for failure.
152 */
153 int amdgpu_gart_table_vram_pin(struct amdgpu_device *adev)
154 {
155 uint64_t gpu_addr;
156 int r;
157
158 r = amdgpu_bo_reserve(adev->gart.robj, false);
159 if (unlikely(r != 0))
160 return r;
161 r = amdgpu_bo_pin(adev->gart.robj,
162 AMDGPU_GEM_DOMAIN_VRAM, &gpu_addr);
163 if (r) {
164 amdgpu_bo_unreserve(adev->gart.robj);
165 return r;
166 }
167 r = amdgpu_bo_kmap(adev->gart.robj, &adev->gart.ptr);
168 if (r)
169 amdgpu_bo_unpin(adev->gart.robj);
170 amdgpu_bo_unreserve(adev->gart.robj);
171 adev->gart.table_addr = gpu_addr;
172 return r;
173 }
174
175 /**
176 * amdgpu_gart_table_vram_unpin - unpin gart page table in vram
177 *
178 * @adev: amdgpu_device pointer
179 *
180 * Unpin the GART page table in vram (pcie r4xx, r5xx+).
181 * These asics require the gart table to be in video memory.
182 */
183 void amdgpu_gart_table_vram_unpin(struct amdgpu_device *adev)
184 {
185 int r;
186
187 if (adev->gart.robj == NULL) {
188 return;
189 }
190 r = amdgpu_bo_reserve(adev->gart.robj, true);
191 if (likely(r == 0)) {
192 amdgpu_bo_kunmap(adev->gart.robj);
193 amdgpu_bo_unpin(adev->gart.robj);
194 amdgpu_bo_unreserve(adev->gart.robj);
195 adev->gart.ptr = NULL;
196 }
197 }
198
199 /**
200 * amdgpu_gart_table_vram_free - free gart page table vram
201 *
202 * @adev: amdgpu_device pointer
203 *
204 * Free the video memory used for the GART page table
205 * (pcie r4xx, r5xx+). These asics require the gart table to
206 * be in video memory.
207 */
208 void amdgpu_gart_table_vram_free(struct amdgpu_device *adev)
209 {
210 if (adev->gart.robj == NULL) {
211 return;
212 }
213 amdgpu_bo_unref(&adev->gart.robj);
214 }
215
216 /*
217 * Common gart functions.
218 */
219 /**
220 * amdgpu_gart_unbind - unbind pages from the gart page table
221 *
222 * @adev: amdgpu_device pointer
223 * @offset: offset into the GPU's gart aperture
224 * @pages: number of pages to unbind
225 *
226 * Unbinds the requested pages from the gart page table and
227 * replaces them with the dummy page (all asics).
228 * Returns 0 for success, -EINVAL for failure.
229 */
230 int amdgpu_gart_unbind(struct amdgpu_device *adev, uint64_t offset,
231 int pages)
232 {
233 unsigned t;
234 unsigned p;
235 int i, j;
236 u64 page_base;
237 /* Starting from VEGA10, system bit must be 0 to mean invalid. */
238 uint64_t flags = 0;
239
240 if (!adev->gart.ready) {
241 WARN(1, "trying to unbind memory from uninitialized GART !\n");
242 return -EINVAL;
243 }
244
245 t = offset / AMDGPU_GPU_PAGE_SIZE;
246 p = t / (PAGE_SIZE / AMDGPU_GPU_PAGE_SIZE);
247 for (i = 0; i < pages; i++, p++) {
248 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
249 adev->gart.pages[p] = NULL;
250 #endif
251 page_base = adev->dummy_page.addr;
252 if (!adev->gart.ptr)
253 continue;
254
255 for (j = 0; j < (PAGE_SIZE / AMDGPU_GPU_PAGE_SIZE); j++, t++) {
256 amdgpu_gart_set_pte_pde(adev, adev->gart.ptr,
257 t, page_base, flags);
258 page_base += AMDGPU_GPU_PAGE_SIZE;
259 }
260 }
261 mb();
262 amdgpu_gart_flush_gpu_tlb(adev, 0);
263 return 0;
264 }
265
266 /**
267 * amdgpu_gart_map - map dma_addresses into GART entries
268 *
269 * @adev: amdgpu_device pointer
270 * @offset: offset into the GPU's gart aperture
271 * @pages: number of pages to bind
272 * @dma_addr: DMA addresses of pages
273 *
274 * Map the dma_addresses into GART entries (all asics).
275 * Returns 0 for success, -EINVAL for failure.
276 */
277 int amdgpu_gart_map(struct amdgpu_device *adev, uint64_t offset,
278 int pages, dma_addr_t *dma_addr, uint64_t flags,
279 void *dst)
280 {
281 uint64_t page_base;
282 unsigned i, j, t;
283
284 if (!adev->gart.ready) {
285 WARN(1, "trying to bind memory to uninitialized GART !\n");
286 return -EINVAL;
287 }
288
289 t = offset / AMDGPU_GPU_PAGE_SIZE;
290
291 for (i = 0; i < pages; i++) {
292 page_base = dma_addr[i];
293 for (j = 0; j < (PAGE_SIZE / AMDGPU_GPU_PAGE_SIZE); j++, t++) {
294 amdgpu_gart_set_pte_pde(adev, dst, t, page_base, flags);
295 page_base += AMDGPU_GPU_PAGE_SIZE;
296 }
297 }
298 return 0;
299 }
300
301 /**
302 * amdgpu_gart_bind - bind pages into the gart page table
303 *
304 * @adev: amdgpu_device pointer
305 * @offset: offset into the GPU's gart aperture
306 * @pages: number of pages to bind
307 * @pagelist: pages to bind
308 * @dma_addr: DMA addresses of pages
309 *
310 * Binds the requested pages to the gart page table
311 * (all asics).
312 * Returns 0 for success, -EINVAL for failure.
313 */
314 int amdgpu_gart_bind(struct amdgpu_device *adev, uint64_t offset,
315 int pages, struct page **pagelist, dma_addr_t *dma_addr,
316 uint64_t flags)
317 {
318 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
319 unsigned i,t,p;
320 #endif
321 int r;
322
323 if (!adev->gart.ready) {
324 WARN(1, "trying to bind memory to uninitialized GART !\n");
325 return -EINVAL;
326 }
327
328 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
329 t = offset / AMDGPU_GPU_PAGE_SIZE;
330 p = t / (PAGE_SIZE / AMDGPU_GPU_PAGE_SIZE);
331 for (i = 0; i < pages; i++, p++)
332 adev->gart.pages[p] = pagelist[i];
333 #endif
334
335 if (adev->gart.ptr) {
336 r = amdgpu_gart_map(adev, offset, pages, dma_addr, flags,
337 adev->gart.ptr);
338 if (r)
339 return r;
340 }
341
342 mb();
343 amdgpu_gart_flush_gpu_tlb(adev, 0);
344 return 0;
345 }
346
347 /**
348 * amdgpu_gart_init - init the driver info for managing the gart
349 *
350 * @adev: amdgpu_device pointer
351 *
352 * Allocate the dummy page and init the gart driver info (all asics).
353 * Returns 0 for success, error for failure.
354 */
355 int amdgpu_gart_init(struct amdgpu_device *adev)
356 {
357 int r;
358
359 if (adev->dummy_page.page)
360 return 0;
361
362 /* We need PAGE_SIZE >= AMDGPU_GPU_PAGE_SIZE */
363 if (PAGE_SIZE < AMDGPU_GPU_PAGE_SIZE) {
364 DRM_ERROR("Page size is smaller than GPU page size!\n");
365 return -EINVAL;
366 }
367 r = amdgpu_dummy_page_init(adev);
368 if (r)
369 return r;
370 /* Compute table size */
371 adev->gart.num_cpu_pages = adev->mc.gart_size / PAGE_SIZE;
372 adev->gart.num_gpu_pages = adev->mc.gart_size / AMDGPU_GPU_PAGE_SIZE;
373 DRM_INFO("GART: num cpu pages %u, num gpu pages %u\n",
374 adev->gart.num_cpu_pages, adev->gart.num_gpu_pages);
375
376 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
377 /* Allocate pages table */
378 adev->gart.pages = vzalloc(sizeof(void *) * adev->gart.num_cpu_pages);
379 if (adev->gart.pages == NULL) {
380 amdgpu_gart_fini(adev);
381 return -ENOMEM;
382 }
383 #endif
384
385 return 0;
386 }
387
388 /**
389 * amdgpu_gart_fini - tear down the driver info for managing the gart
390 *
391 * @adev: amdgpu_device pointer
392 *
393 * Tear down the gart driver info and free the dummy page (all asics).
394 */
395 void amdgpu_gart_fini(struct amdgpu_device *adev)
396 {
397 if (adev->gart.ready) {
398 /* unbind pages */
399 amdgpu_gart_unbind(adev, 0, adev->gart.num_cpu_pages);
400 }
401 adev->gart.ready = false;
402 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
403 vfree(adev->gart.pages);
404 adev->gart.pages = NULL;
405 #endif
406 amdgpu_dummy_page_fini(adev);
407 }