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/vmalloc.h>
30
31 #include <drm/drm_pci.h>
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 */
radeon_gart_table_ram_alloc(struct radeon_device * rdev)71 int radeon_gart_table_ram_alloc(struct radeon_device *rdev)
72 {
73 void *ptr;
74
75 ptr = pci_alloc_consistent(rdev->pdev, rdev->gart.table_size,
76 &rdev->gart.table_addr);
77 if (ptr == NULL) {
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((unsigned long)ptr,
84 rdev->gart.table_size >> PAGE_SHIFT);
85 }
86 #endif
87 rdev->gart.ptr = ptr;
88 memset((void *)rdev->gart.ptr, 0, rdev->gart.table_size);
89 return 0;
90 }
91
92 /**
93 * radeon_gart_table_ram_free - free system ram for gart page table
94 *
95 * @rdev: radeon_device pointer
96 *
97 * Free system memory for GART page table
98 * (r1xx-r3xx, non-pcie r4xx, rs400). These asics require the
99 * gart table to be in system memory.
100 */
radeon_gart_table_ram_free(struct radeon_device * rdev)101 void radeon_gart_table_ram_free(struct radeon_device *rdev)
102 {
103 if (rdev->gart.ptr == NULL) {
104 return;
105 }
106 #ifdef CONFIG_X86
107 if (rdev->family == CHIP_RS400 || rdev->family == CHIP_RS480 ||
108 rdev->family == CHIP_RS690 || rdev->family == CHIP_RS740) {
109 set_memory_wb((unsigned long)rdev->gart.ptr,
110 rdev->gart.table_size >> PAGE_SHIFT);
111 }
112 #endif
113 pci_free_consistent(rdev->pdev, rdev->gart.table_size,
114 (void *)rdev->gart.ptr,
115 rdev->gart.table_addr);
116 rdev->gart.ptr = NULL;
117 rdev->gart.table_addr = 0;
118 }
119
120 /**
121 * radeon_gart_table_vram_alloc - allocate vram for gart page table
122 *
123 * @rdev: radeon_device pointer
124 *
125 * Allocate video memory for GART page table
126 * (pcie r4xx, r5xx+). These asics require the
127 * gart table to be in video memory.
128 * Returns 0 for success, error for failure.
129 */
radeon_gart_table_vram_alloc(struct radeon_device * rdev)130 int radeon_gart_table_vram_alloc(struct radeon_device *rdev)
131 {
132 int r;
133
134 if (rdev->gart.robj == NULL) {
135 r = radeon_bo_create(rdev, rdev->gart.table_size,
136 PAGE_SIZE, true, RADEON_GEM_DOMAIN_VRAM,
137 0, NULL, NULL, &rdev->gart.robj);
138 if (r) {
139 return r;
140 }
141 }
142 return 0;
143 }
144
145 /**
146 * radeon_gart_table_vram_pin - pin gart page table in vram
147 *
148 * @rdev: radeon_device pointer
149 *
150 * Pin the GART page table in vram so it will not be moved
151 * by the memory manager (pcie r4xx, r5xx+). These asics require the
152 * gart table to be in video memory.
153 * Returns 0 for success, error for failure.
154 */
radeon_gart_table_vram_pin(struct radeon_device * rdev)155 int radeon_gart_table_vram_pin(struct radeon_device *rdev)
156 {
157 uint64_t gpu_addr;
158 int r;
159
160 r = radeon_bo_reserve(rdev->gart.robj, false);
161 if (unlikely(r != 0))
162 return r;
163 r = radeon_bo_pin(rdev->gart.robj,
164 RADEON_GEM_DOMAIN_VRAM, &gpu_addr);
165 if (r) {
166 radeon_bo_unreserve(rdev->gart.robj);
167 return r;
168 }
169 r = radeon_bo_kmap(rdev->gart.robj, &rdev->gart.ptr);
170 if (r)
171 radeon_bo_unpin(rdev->gart.robj);
172 radeon_bo_unreserve(rdev->gart.robj);
173 rdev->gart.table_addr = gpu_addr;
174
175 if (!r) {
176 int i;
177
178 /* We might have dropped some GART table updates while it wasn't
179 * mapped, restore all entries
180 */
181 for (i = 0; i < rdev->gart.num_gpu_pages; i++)
182 radeon_gart_set_page(rdev, i, rdev->gart.pages_entry[i]);
183 mb();
184 radeon_gart_tlb_flush(rdev);
185 }
186
187 return r;
188 }
189
190 /**
191 * radeon_gart_table_vram_unpin - unpin gart page table in vram
192 *
193 * @rdev: radeon_device pointer
194 *
195 * Unpin the GART page table in vram (pcie r4xx, r5xx+).
196 * These asics require the gart table to be in video memory.
197 */
radeon_gart_table_vram_unpin(struct radeon_device * rdev)198 void radeon_gart_table_vram_unpin(struct radeon_device *rdev)
199 {
200 int r;
201
202 if (rdev->gart.robj == NULL) {
203 return;
204 }
205 r = radeon_bo_reserve(rdev->gart.robj, false);
206 if (likely(r == 0)) {
207 radeon_bo_kunmap(rdev->gart.robj);
208 radeon_bo_unpin(rdev->gart.robj);
209 radeon_bo_unreserve(rdev->gart.robj);
210 rdev->gart.ptr = NULL;
211 }
212 }
213
214 /**
215 * radeon_gart_table_vram_free - free gart page table vram
216 *
217 * @rdev: radeon_device pointer
218 *
219 * Free the video memory used for the GART page table
220 * (pcie r4xx, r5xx+). These asics require the gart table to
221 * be in video memory.
222 */
radeon_gart_table_vram_free(struct radeon_device * rdev)223 void radeon_gart_table_vram_free(struct radeon_device *rdev)
224 {
225 if (rdev->gart.robj == NULL) {
226 return;
227 }
228 radeon_bo_unref(&rdev->gart.robj);
229 }
230
231 /*
232 * Common gart functions.
233 */
234 /**
235 * radeon_gart_unbind - unbind pages from the gart page table
236 *
237 * @rdev: radeon_device pointer
238 * @offset: offset into the GPU's gart aperture
239 * @pages: number of pages to unbind
240 *
241 * Unbinds the requested pages from the gart page table and
242 * replaces them with the dummy page (all asics).
243 */
radeon_gart_unbind(struct radeon_device * rdev,unsigned offset,int pages)244 void radeon_gart_unbind(struct radeon_device *rdev, unsigned offset,
245 int pages)
246 {
247 unsigned t;
248 unsigned p;
249 int i, j;
250
251 if (!rdev->gart.ready) {
252 WARN(1, "trying to unbind memory from uninitialized GART !\n");
253 return;
254 }
255 t = offset / RADEON_GPU_PAGE_SIZE;
256 p = t / (PAGE_SIZE / RADEON_GPU_PAGE_SIZE);
257 for (i = 0; i < pages; i++, p++) {
258 if (rdev->gart.pages[p]) {
259 rdev->gart.pages[p] = NULL;
260 for (j = 0; j < (PAGE_SIZE / RADEON_GPU_PAGE_SIZE); j++, t++) {
261 rdev->gart.pages_entry[t] = rdev->dummy_page.entry;
262 if (rdev->gart.ptr) {
263 radeon_gart_set_page(rdev, t,
264 rdev->dummy_page.entry);
265 }
266 }
267 }
268 }
269 if (rdev->gart.ptr) {
270 mb();
271 radeon_gart_tlb_flush(rdev);
272 }
273 }
274
275 /**
276 * radeon_gart_bind - bind pages into the gart page table
277 *
278 * @rdev: radeon_device pointer
279 * @offset: offset into the GPU's gart aperture
280 * @pages: number of pages to bind
281 * @pagelist: pages to bind
282 * @dma_addr: DMA addresses of pages
283 * @flags: RADEON_GART_PAGE_* flags
284 *
285 * Binds the requested pages to the gart page table
286 * (all asics).
287 * Returns 0 for success, -EINVAL for failure.
288 */
radeon_gart_bind(struct radeon_device * rdev,unsigned offset,int pages,struct page ** pagelist,dma_addr_t * dma_addr,uint32_t flags)289 int radeon_gart_bind(struct radeon_device *rdev, unsigned offset,
290 int pages, struct page **pagelist, dma_addr_t *dma_addr,
291 uint32_t flags)
292 {
293 unsigned t;
294 unsigned p;
295 uint64_t page_base, page_entry;
296 int i, j;
297
298 if (!rdev->gart.ready) {
299 WARN(1, "trying to bind memory to uninitialized GART !\n");
300 return -EINVAL;
301 }
302 t = offset / RADEON_GPU_PAGE_SIZE;
303 p = t / (PAGE_SIZE / RADEON_GPU_PAGE_SIZE);
304
305 for (i = 0; i < pages; i++, p++) {
306 rdev->gart.pages[p] = pagelist[i];
307 page_base = dma_addr[i];
308 for (j = 0; j < (PAGE_SIZE / RADEON_GPU_PAGE_SIZE); j++, t++) {
309 page_entry = radeon_gart_get_page_entry(page_base, flags);
310 rdev->gart.pages_entry[t] = page_entry;
311 if (rdev->gart.ptr) {
312 radeon_gart_set_page(rdev, t, page_entry);
313 }
314 page_base += RADEON_GPU_PAGE_SIZE;
315 }
316 }
317 if (rdev->gart.ptr) {
318 mb();
319 radeon_gart_tlb_flush(rdev);
320 }
321 return 0;
322 }
323
324 /**
325 * radeon_gart_init - init the driver info for managing the gart
326 *
327 * @rdev: radeon_device pointer
328 *
329 * Allocate the dummy page and init the gart driver info (all asics).
330 * Returns 0 for success, error for failure.
331 */
radeon_gart_init(struct radeon_device * rdev)332 int radeon_gart_init(struct radeon_device *rdev)
333 {
334 int r, i;
335
336 if (rdev->gart.pages) {
337 return 0;
338 }
339 /* We need PAGE_SIZE >= RADEON_GPU_PAGE_SIZE */
340 if (PAGE_SIZE < RADEON_GPU_PAGE_SIZE) {
341 DRM_ERROR("Page size is smaller than GPU page size!\n");
342 return -EINVAL;
343 }
344 r = radeon_dummy_page_init(rdev);
345 if (r)
346 return r;
347 /* Compute table size */
348 rdev->gart.num_cpu_pages = rdev->mc.gtt_size / PAGE_SIZE;
349 rdev->gart.num_gpu_pages = rdev->mc.gtt_size / RADEON_GPU_PAGE_SIZE;
350 DRM_INFO("GART: num cpu pages %u, num gpu pages %u\n",
351 rdev->gart.num_cpu_pages, rdev->gart.num_gpu_pages);
352 /* Allocate pages table */
353 rdev->gart.pages = vzalloc(array_size(sizeof(void *),
354 rdev->gart.num_cpu_pages));
355 if (rdev->gart.pages == NULL) {
356 radeon_gart_fini(rdev);
357 return -ENOMEM;
358 }
359 rdev->gart.pages_entry = vmalloc(array_size(sizeof(uint64_t),
360 rdev->gart.num_gpu_pages));
361 if (rdev->gart.pages_entry == NULL) {
362 radeon_gart_fini(rdev);
363 return -ENOMEM;
364 }
365 /* set GART entry to point to the dummy page by default */
366 for (i = 0; i < rdev->gart.num_gpu_pages; i++)
367 rdev->gart.pages_entry[i] = rdev->dummy_page.entry;
368 return 0;
369 }
370
371 /**
372 * radeon_gart_fini - tear down the driver info for managing the gart
373 *
374 * @rdev: radeon_device pointer
375 *
376 * Tear down the gart driver info and free the dummy page (all asics).
377 */
radeon_gart_fini(struct radeon_device * rdev)378 void radeon_gart_fini(struct radeon_device *rdev)
379 {
380 if (rdev->gart.ready) {
381 /* unbind pages */
382 radeon_gart_unbind(rdev, 0, rdev->gart.num_cpu_pages);
383 }
384 rdev->gart.ready = false;
385 vfree(rdev->gart.pages);
386 vfree(rdev->gart.pages_entry);
387 rdev->gart.pages = NULL;
388 rdev->gart.pages_entry = NULL;
389
390 radeon_dummy_page_fini(rdev);
391 }
392