Loading...
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 */
71int radeon_gart_table_ram_alloc(struct radeon_device *rdev)
72{
73 void *ptr;
74
75 ptr = dma_alloc_coherent(&rdev->pdev->dev, rdev->gart.table_size,
76 &rdev->gart.table_addr, GFP_KERNEL);
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 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 */
100void radeon_gart_table_ram_free(struct radeon_device *rdev)
101{
102 if (rdev->gart.ptr == NULL) {
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((unsigned long)rdev->gart.ptr,
109 rdev->gart.table_size >> PAGE_SHIFT);
110 }
111#endif
112 dma_free_coherent(&rdev->pdev->dev, rdev->gart.table_size,
113 (void *)rdev->gart.ptr, 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 */
128int 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, rdev->gart.table_size,
134 PAGE_SIZE, true, RADEON_GEM_DOMAIN_VRAM,
135 0, NULL, NULL, &rdev->gart.robj);
136 if (r) {
137 return r;
138 }
139 }
140 return 0;
141}
142
143/**
144 * radeon_gart_table_vram_pin - pin gart page table in vram
145 *
146 * @rdev: radeon_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 */
153int radeon_gart_table_vram_pin(struct radeon_device *rdev)
154{
155 uint64_t gpu_addr;
156 int r;
157
158 r = radeon_bo_reserve(rdev->gart.robj, false);
159 if (unlikely(r != 0))
160 return r;
161 r = radeon_bo_pin(rdev->gart.robj,
162 RADEON_GEM_DOMAIN_VRAM, &gpu_addr);
163 if (r) {
164 radeon_bo_unreserve(rdev->gart.robj);
165 return r;
166 }
167 r = radeon_bo_kmap(rdev->gart.robj, &rdev->gart.ptr);
168 if (r)
169 radeon_bo_unpin(rdev->gart.robj);
170 radeon_bo_unreserve(rdev->gart.robj);
171 rdev->gart.table_addr = gpu_addr;
172
173 if (!r) {
174 int i;
175
176 /* We might have dropped some GART table updates while it wasn't
177 * mapped, restore all entries
178 */
179 for (i = 0; i < rdev->gart.num_gpu_pages; i++)
180 radeon_gart_set_page(rdev, i, rdev->gart.pages_entry[i]);
181 mb();
182 radeon_gart_tlb_flush(rdev);
183 }
184
185 return r;
186}
187
188/**
189 * radeon_gart_table_vram_unpin - unpin gart page table in vram
190 *
191 * @rdev: radeon_device pointer
192 *
193 * Unpin the GART page table in vram (pcie r4xx, r5xx+).
194 * These asics require the gart table to be in video memory.
195 */
196void radeon_gart_table_vram_unpin(struct radeon_device *rdev)
197{
198 int r;
199
200 if (rdev->gart.robj == NULL) {
201 return;
202 }
203 r = radeon_bo_reserve(rdev->gart.robj, false);
204 if (likely(r == 0)) {
205 radeon_bo_kunmap(rdev->gart.robj);
206 radeon_bo_unpin(rdev->gart.robj);
207 radeon_bo_unreserve(rdev->gart.robj);
208 rdev->gart.ptr = NULL;
209 }
210}
211
212/**
213 * radeon_gart_table_vram_free - free gart page table vram
214 *
215 * @rdev: radeon_device pointer
216 *
217 * Free the video memory used for the GART page table
218 * (pcie r4xx, r5xx+). These asics require the gart table to
219 * be in video memory.
220 */
221void radeon_gart_table_vram_free(struct radeon_device *rdev)
222{
223 if (rdev->gart.robj == NULL) {
224 return;
225 }
226 radeon_bo_unref(&rdev->gart.robj);
227}
228
229/*
230 * Common gart functions.
231 */
232/**
233 * radeon_gart_unbind - unbind pages from the gart page table
234 *
235 * @rdev: radeon_device pointer
236 * @offset: offset into the GPU's gart aperture
237 * @pages: number of pages to unbind
238 *
239 * Unbinds the requested pages from the gart page table and
240 * replaces them with the dummy page (all asics).
241 */
242void radeon_gart_unbind(struct radeon_device *rdev, unsigned offset,
243 int pages)
244{
245 unsigned t;
246 unsigned p;
247 int i, j;
248
249 if (!rdev->gart.ready) {
250 WARN(1, "trying to unbind memory from uninitialized GART !\n");
251 return;
252 }
253 t = offset / RADEON_GPU_PAGE_SIZE;
254 p = t / (PAGE_SIZE / RADEON_GPU_PAGE_SIZE);
255 for (i = 0; i < pages; i++, p++) {
256 if (rdev->gart.pages[p]) {
257 rdev->gart.pages[p] = NULL;
258 for (j = 0; j < (PAGE_SIZE / RADEON_GPU_PAGE_SIZE); j++, t++) {
259 rdev->gart.pages_entry[t] = rdev->dummy_page.entry;
260 if (rdev->gart.ptr) {
261 radeon_gart_set_page(rdev, t,
262 rdev->dummy_page.entry);
263 }
264 }
265 }
266 }
267 if (rdev->gart.ptr) {
268 mb();
269 radeon_gart_tlb_flush(rdev);
270 }
271}
272
273/**
274 * radeon_gart_bind - bind pages into the gart page table
275 *
276 * @rdev: radeon_device pointer
277 * @offset: offset into the GPU's gart aperture
278 * @pages: number of pages to bind
279 * @pagelist: pages to bind
280 * @dma_addr: DMA addresses of pages
281 * @flags: RADEON_GART_PAGE_* flags
282 *
283 * Binds the requested pages to the gart page table
284 * (all asics).
285 * Returns 0 for success, -EINVAL for failure.
286 */
287int radeon_gart_bind(struct radeon_device *rdev, unsigned offset,
288 int pages, struct page **pagelist, dma_addr_t *dma_addr,
289 uint32_t flags)
290{
291 unsigned t;
292 unsigned p;
293 uint64_t page_base, page_entry;
294 int i, j;
295
296 if (!rdev->gart.ready) {
297 WARN(1, "trying to bind memory to uninitialized GART !\n");
298 return -EINVAL;
299 }
300 t = offset / RADEON_GPU_PAGE_SIZE;
301 p = t / (PAGE_SIZE / RADEON_GPU_PAGE_SIZE);
302
303 for (i = 0; i < pages; i++, p++) {
304 rdev->gart.pages[p] = pagelist ? pagelist[i] :
305 rdev->dummy_page.page;
306 page_base = dma_addr[i];
307 for (j = 0; j < (PAGE_SIZE / RADEON_GPU_PAGE_SIZE); j++, t++) {
308 page_entry = radeon_gart_get_page_entry(page_base, flags);
309 rdev->gart.pages_entry[t] = page_entry;
310 if (rdev->gart.ptr) {
311 radeon_gart_set_page(rdev, t, page_entry);
312 }
313 page_base += RADEON_GPU_PAGE_SIZE;
314 }
315 }
316 if (rdev->gart.ptr) {
317 mb();
318 radeon_gart_tlb_flush(rdev);
319 }
320 return 0;
321}
322
323/**
324 * radeon_gart_init - init the driver info for managing the gart
325 *
326 * @rdev: radeon_device pointer
327 *
328 * Allocate the dummy page and init the gart driver info (all asics).
329 * Returns 0 for success, error for failure.
330 */
331int radeon_gart_init(struct radeon_device *rdev)
332{
333 int r, i;
334
335 if (rdev->gart.pages) {
336 return 0;
337 }
338 /* We need PAGE_SIZE >= RADEON_GPU_PAGE_SIZE */
339 if (PAGE_SIZE < RADEON_GPU_PAGE_SIZE) {
340 DRM_ERROR("Page size is smaller than GPU page size!\n");
341 return -EINVAL;
342 }
343 r = radeon_dummy_page_init(rdev);
344 if (r)
345 return r;
346 /* Compute table size */
347 rdev->gart.num_cpu_pages = rdev->mc.gtt_size / PAGE_SIZE;
348 rdev->gart.num_gpu_pages = rdev->mc.gtt_size / RADEON_GPU_PAGE_SIZE;
349 DRM_INFO("GART: num cpu pages %u, num gpu pages %u\n",
350 rdev->gart.num_cpu_pages, rdev->gart.num_gpu_pages);
351 /* Allocate pages table */
352 rdev->gart.pages = vzalloc(array_size(sizeof(void *),
353 rdev->gart.num_cpu_pages));
354 if (rdev->gart.pages == NULL) {
355 radeon_gart_fini(rdev);
356 return -ENOMEM;
357 }
358 rdev->gart.pages_entry = vmalloc(array_size(sizeof(uint64_t),
359 rdev->gart.num_gpu_pages));
360 if (rdev->gart.pages_entry == NULL) {
361 radeon_gart_fini(rdev);
362 return -ENOMEM;
363 }
364 /* set GART entry to point to the dummy page by default */
365 for (i = 0; i < rdev->gart.num_gpu_pages; i++)
366 rdev->gart.pages_entry[i] = rdev->dummy_page.entry;
367 return 0;
368}
369
370/**
371 * radeon_gart_fini - tear down the driver info for managing the gart
372 *
373 * @rdev: radeon_device pointer
374 *
375 * Tear down the gart driver info and free the dummy page (all asics).
376 */
377void radeon_gart_fini(struct radeon_device *rdev)
378{
379 if (rdev->gart.ready) {
380 /* unbind pages */
381 radeon_gart_unbind(rdev, 0, rdev->gart.num_cpu_pages);
382 }
383 rdev->gart.ready = false;
384 vfree(rdev->gart.pages);
385 vfree(rdev->gart.pages_entry);
386 rdev->gart.pages = NULL;
387 rdev->gart.pages_entry = NULL;
388
389 radeon_dummy_page_fini(rdev);
390}
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 "drmP.h"
29#include "radeon_drm.h"
30#include "radeon.h"
31#include "radeon_reg.h"
32
33/*
34 * Common GART table functions.
35 */
36int radeon_gart_table_ram_alloc(struct radeon_device *rdev)
37{
38 void *ptr;
39
40 ptr = pci_alloc_consistent(rdev->pdev, rdev->gart.table_size,
41 &rdev->gart.table_addr);
42 if (ptr == NULL) {
43 return -ENOMEM;
44 }
45#ifdef CONFIG_X86
46 if (rdev->family == CHIP_RS400 || rdev->family == CHIP_RS480 ||
47 rdev->family == CHIP_RS690 || rdev->family == CHIP_RS740) {
48 set_memory_uc((unsigned long)ptr,
49 rdev->gart.table_size >> PAGE_SHIFT);
50 }
51#endif
52 rdev->gart.ptr = ptr;
53 memset((void *)rdev->gart.ptr, 0, rdev->gart.table_size);
54 return 0;
55}
56
57void radeon_gart_table_ram_free(struct radeon_device *rdev)
58{
59 if (rdev->gart.ptr == NULL) {
60 return;
61 }
62#ifdef CONFIG_X86
63 if (rdev->family == CHIP_RS400 || rdev->family == CHIP_RS480 ||
64 rdev->family == CHIP_RS690 || rdev->family == CHIP_RS740) {
65 set_memory_wb((unsigned long)rdev->gart.ptr,
66 rdev->gart.table_size >> PAGE_SHIFT);
67 }
68#endif
69 pci_free_consistent(rdev->pdev, rdev->gart.table_size,
70 (void *)rdev->gart.ptr,
71 rdev->gart.table_addr);
72 rdev->gart.ptr = NULL;
73 rdev->gart.table_addr = 0;
74}
75
76int radeon_gart_table_vram_alloc(struct radeon_device *rdev)
77{
78 int r;
79
80 if (rdev->gart.robj == NULL) {
81 r = radeon_bo_create(rdev, rdev->gart.table_size,
82 PAGE_SIZE, true, RADEON_GEM_DOMAIN_VRAM,
83 NULL, &rdev->gart.robj);
84 if (r) {
85 return r;
86 }
87 }
88 return 0;
89}
90
91int radeon_gart_table_vram_pin(struct radeon_device *rdev)
92{
93 uint64_t gpu_addr;
94 int r;
95
96 r = radeon_bo_reserve(rdev->gart.robj, false);
97 if (unlikely(r != 0))
98 return r;
99 r = radeon_bo_pin(rdev->gart.robj,
100 RADEON_GEM_DOMAIN_VRAM, &gpu_addr);
101 if (r) {
102 radeon_bo_unreserve(rdev->gart.robj);
103 return r;
104 }
105 r = radeon_bo_kmap(rdev->gart.robj, &rdev->gart.ptr);
106 if (r)
107 radeon_bo_unpin(rdev->gart.robj);
108 radeon_bo_unreserve(rdev->gart.robj);
109 rdev->gart.table_addr = gpu_addr;
110 return r;
111}
112
113void radeon_gart_table_vram_unpin(struct radeon_device *rdev)
114{
115 int r;
116
117 if (rdev->gart.robj == NULL) {
118 return;
119 }
120 r = radeon_bo_reserve(rdev->gart.robj, false);
121 if (likely(r == 0)) {
122 radeon_bo_kunmap(rdev->gart.robj);
123 radeon_bo_unpin(rdev->gart.robj);
124 radeon_bo_unreserve(rdev->gart.robj);
125 rdev->gart.ptr = NULL;
126 }
127}
128
129void radeon_gart_table_vram_free(struct radeon_device *rdev)
130{
131 if (rdev->gart.robj == NULL) {
132 return;
133 }
134 radeon_gart_table_vram_unpin(rdev);
135 radeon_bo_unref(&rdev->gart.robj);
136}
137
138
139
140
141/*
142 * Common gart functions.
143 */
144void radeon_gart_unbind(struct radeon_device *rdev, unsigned offset,
145 int pages)
146{
147 unsigned t;
148 unsigned p;
149 int i, j;
150 u64 page_base;
151
152 if (!rdev->gart.ready) {
153 WARN(1, "trying to unbind memory from uninitialized GART !\n");
154 return;
155 }
156 t = offset / RADEON_GPU_PAGE_SIZE;
157 p = t / (PAGE_SIZE / RADEON_GPU_PAGE_SIZE);
158 for (i = 0; i < pages; i++, p++) {
159 if (rdev->gart.pages[p]) {
160 rdev->gart.pages[p] = NULL;
161 rdev->gart.pages_addr[p] = rdev->dummy_page.addr;
162 page_base = rdev->gart.pages_addr[p];
163 for (j = 0; j < (PAGE_SIZE / RADEON_GPU_PAGE_SIZE); j++, t++) {
164 if (rdev->gart.ptr) {
165 radeon_gart_set_page(rdev, t, page_base);
166 }
167 page_base += RADEON_GPU_PAGE_SIZE;
168 }
169 }
170 }
171 mb();
172 radeon_gart_tlb_flush(rdev);
173}
174
175int radeon_gart_bind(struct radeon_device *rdev, unsigned offset,
176 int pages, struct page **pagelist, dma_addr_t *dma_addr)
177{
178 unsigned t;
179 unsigned p;
180 uint64_t page_base;
181 int i, j;
182
183 if (!rdev->gart.ready) {
184 WARN(1, "trying to bind memory to uninitialized GART !\n");
185 return -EINVAL;
186 }
187 t = offset / RADEON_GPU_PAGE_SIZE;
188 p = t / (PAGE_SIZE / RADEON_GPU_PAGE_SIZE);
189
190 for (i = 0; i < pages; i++, p++) {
191 rdev->gart.pages_addr[p] = dma_addr[i];
192 rdev->gart.pages[p] = pagelist[i];
193 if (rdev->gart.ptr) {
194 page_base = rdev->gart.pages_addr[p];
195 for (j = 0; j < (PAGE_SIZE / RADEON_GPU_PAGE_SIZE); j++, t++) {
196 radeon_gart_set_page(rdev, t, page_base);
197 page_base += RADEON_GPU_PAGE_SIZE;
198 }
199 }
200 }
201 mb();
202 radeon_gart_tlb_flush(rdev);
203 return 0;
204}
205
206void radeon_gart_restore(struct radeon_device *rdev)
207{
208 int i, j, t;
209 u64 page_base;
210
211 if (!rdev->gart.ptr) {
212 return;
213 }
214 for (i = 0, t = 0; i < rdev->gart.num_cpu_pages; i++) {
215 page_base = rdev->gart.pages_addr[i];
216 for (j = 0; j < (PAGE_SIZE / RADEON_GPU_PAGE_SIZE); j++, t++) {
217 radeon_gart_set_page(rdev, t, page_base);
218 page_base += RADEON_GPU_PAGE_SIZE;
219 }
220 }
221 mb();
222 radeon_gart_tlb_flush(rdev);
223}
224
225int radeon_gart_init(struct radeon_device *rdev)
226{
227 int r, i;
228
229 if (rdev->gart.pages) {
230 return 0;
231 }
232 /* We need PAGE_SIZE >= RADEON_GPU_PAGE_SIZE */
233 if (PAGE_SIZE < RADEON_GPU_PAGE_SIZE) {
234 DRM_ERROR("Page size is smaller than GPU page size!\n");
235 return -EINVAL;
236 }
237 r = radeon_dummy_page_init(rdev);
238 if (r)
239 return r;
240 /* Compute table size */
241 rdev->gart.num_cpu_pages = rdev->mc.gtt_size / PAGE_SIZE;
242 rdev->gart.num_gpu_pages = rdev->mc.gtt_size / RADEON_GPU_PAGE_SIZE;
243 DRM_INFO("GART: num cpu pages %u, num gpu pages %u\n",
244 rdev->gart.num_cpu_pages, rdev->gart.num_gpu_pages);
245 /* Allocate pages table */
246 rdev->gart.pages = kzalloc(sizeof(void *) * rdev->gart.num_cpu_pages,
247 GFP_KERNEL);
248 if (rdev->gart.pages == NULL) {
249 radeon_gart_fini(rdev);
250 return -ENOMEM;
251 }
252 rdev->gart.pages_addr = kzalloc(sizeof(dma_addr_t) *
253 rdev->gart.num_cpu_pages, GFP_KERNEL);
254 if (rdev->gart.pages_addr == NULL) {
255 radeon_gart_fini(rdev);
256 return -ENOMEM;
257 }
258 /* set GART entry to point to the dummy page by default */
259 for (i = 0; i < rdev->gart.num_cpu_pages; i++) {
260 rdev->gart.pages_addr[i] = rdev->dummy_page.addr;
261 }
262 return 0;
263}
264
265void radeon_gart_fini(struct radeon_device *rdev)
266{
267 if (rdev->gart.pages && rdev->gart.pages_addr && rdev->gart.ready) {
268 /* unbind pages */
269 radeon_gart_unbind(rdev, 0, rdev->gart.num_cpu_pages);
270 }
271 rdev->gart.ready = false;
272 kfree(rdev->gart.pages);
273 kfree(rdev->gart.pages_addr);
274 rdev->gart.pages = NULL;
275 rdev->gart.pages_addr = NULL;
276
277 radeon_dummy_page_fini(rdev);
278}
279
280/*
281 * vm helpers
282 *
283 * TODO bind a default page at vm initialization for default address
284 */
285int radeon_vm_manager_init(struct radeon_device *rdev)
286{
287 int r;
288
289 rdev->vm_manager.enabled = false;
290
291 /* mark first vm as always in use, it's the system one */
292 /* allocate enough for 2 full VM pts */
293 r = radeon_sa_bo_manager_init(rdev, &rdev->vm_manager.sa_manager,
294 rdev->vm_manager.max_pfn * 8 * 2,
295 RADEON_GEM_DOMAIN_VRAM);
296 if (r) {
297 dev_err(rdev->dev, "failed to allocate vm bo (%dKB)\n",
298 (rdev->vm_manager.max_pfn * 8) >> 10);
299 return r;
300 }
301
302 r = rdev->vm_manager.funcs->init(rdev);
303 if (r == 0)
304 rdev->vm_manager.enabled = true;
305
306 return r;
307}
308
309/* cs mutex must be lock */
310static void radeon_vm_unbind_locked(struct radeon_device *rdev,
311 struct radeon_vm *vm)
312{
313 struct radeon_bo_va *bo_va;
314
315 if (vm->id == -1) {
316 return;
317 }
318
319 /* wait for vm use to end */
320 if (vm->fence) {
321 radeon_fence_wait(vm->fence, false);
322 radeon_fence_unref(&vm->fence);
323 }
324
325 /* hw unbind */
326 rdev->vm_manager.funcs->unbind(rdev, vm);
327 rdev->vm_manager.use_bitmap &= ~(1 << vm->id);
328 list_del_init(&vm->list);
329 vm->id = -1;
330 radeon_sa_bo_free(rdev, &vm->sa_bo, NULL);
331 vm->pt = NULL;
332
333 list_for_each_entry(bo_va, &vm->va, vm_list) {
334 bo_va->valid = false;
335 }
336}
337
338void radeon_vm_manager_fini(struct radeon_device *rdev)
339{
340 if (rdev->vm_manager.sa_manager.bo == NULL)
341 return;
342 radeon_vm_manager_suspend(rdev);
343 rdev->vm_manager.funcs->fini(rdev);
344 radeon_sa_bo_manager_fini(rdev, &rdev->vm_manager.sa_manager);
345 rdev->vm_manager.enabled = false;
346}
347
348int radeon_vm_manager_start(struct radeon_device *rdev)
349{
350 if (rdev->vm_manager.sa_manager.bo == NULL) {
351 return -EINVAL;
352 }
353 return radeon_sa_bo_manager_start(rdev, &rdev->vm_manager.sa_manager);
354}
355
356int radeon_vm_manager_suspend(struct radeon_device *rdev)
357{
358 struct radeon_vm *vm, *tmp;
359
360 radeon_mutex_lock(&rdev->cs_mutex);
361 /* unbind all active vm */
362 list_for_each_entry_safe(vm, tmp, &rdev->vm_manager.lru_vm, list) {
363 radeon_vm_unbind_locked(rdev, vm);
364 }
365 rdev->vm_manager.funcs->fini(rdev);
366 radeon_mutex_unlock(&rdev->cs_mutex);
367 return radeon_sa_bo_manager_suspend(rdev, &rdev->vm_manager.sa_manager);
368}
369
370/* cs mutex must be lock */
371void radeon_vm_unbind(struct radeon_device *rdev, struct radeon_vm *vm)
372{
373 mutex_lock(&vm->mutex);
374 radeon_vm_unbind_locked(rdev, vm);
375 mutex_unlock(&vm->mutex);
376}
377
378/* cs mutex must be lock & vm mutex must be lock */
379int radeon_vm_bind(struct radeon_device *rdev, struct radeon_vm *vm)
380{
381 struct radeon_vm *vm_evict;
382 unsigned i;
383 int id = -1, r;
384
385 if (vm == NULL) {
386 return -EINVAL;
387 }
388
389 if (vm->id != -1) {
390 /* update lru */
391 list_del_init(&vm->list);
392 list_add_tail(&vm->list, &rdev->vm_manager.lru_vm);
393 return 0;
394 }
395
396retry:
397 r = radeon_sa_bo_new(rdev, &rdev->vm_manager.sa_manager, &vm->sa_bo,
398 RADEON_GPU_PAGE_ALIGN(vm->last_pfn * 8),
399 RADEON_GPU_PAGE_SIZE, false);
400 if (r) {
401 if (list_empty(&rdev->vm_manager.lru_vm)) {
402 return r;
403 }
404 vm_evict = list_first_entry(&rdev->vm_manager.lru_vm, struct radeon_vm, list);
405 radeon_vm_unbind(rdev, vm_evict);
406 goto retry;
407 }
408 vm->pt = radeon_sa_bo_cpu_addr(vm->sa_bo);
409 vm->pt_gpu_addr = radeon_sa_bo_gpu_addr(vm->sa_bo);
410 memset(vm->pt, 0, RADEON_GPU_PAGE_ALIGN(vm->last_pfn * 8));
411
412retry_id:
413 /* search for free vm */
414 for (i = 0; i < rdev->vm_manager.nvm; i++) {
415 if (!(rdev->vm_manager.use_bitmap & (1 << i))) {
416 id = i;
417 break;
418 }
419 }
420 /* evict vm if necessary */
421 if (id == -1) {
422 vm_evict = list_first_entry(&rdev->vm_manager.lru_vm, struct radeon_vm, list);
423 radeon_vm_unbind(rdev, vm_evict);
424 goto retry_id;
425 }
426
427 /* do hw bind */
428 r = rdev->vm_manager.funcs->bind(rdev, vm, id);
429 if (r) {
430 radeon_sa_bo_free(rdev, &vm->sa_bo, NULL);
431 return r;
432 }
433 rdev->vm_manager.use_bitmap |= 1 << id;
434 vm->id = id;
435 list_add_tail(&vm->list, &rdev->vm_manager.lru_vm);
436 return radeon_vm_bo_update_pte(rdev, vm, rdev->ring_tmp_bo.bo,
437 &rdev->ring_tmp_bo.bo->tbo.mem);
438}
439
440/* object have to be reserved */
441int radeon_vm_bo_add(struct radeon_device *rdev,
442 struct radeon_vm *vm,
443 struct radeon_bo *bo,
444 uint64_t offset,
445 uint32_t flags)
446{
447 struct radeon_bo_va *bo_va, *tmp;
448 struct list_head *head;
449 uint64_t size = radeon_bo_size(bo), last_offset = 0;
450 unsigned last_pfn;
451
452 bo_va = kzalloc(sizeof(struct radeon_bo_va), GFP_KERNEL);
453 if (bo_va == NULL) {
454 return -ENOMEM;
455 }
456 bo_va->vm = vm;
457 bo_va->bo = bo;
458 bo_va->soffset = offset;
459 bo_va->eoffset = offset + size;
460 bo_va->flags = flags;
461 bo_va->valid = false;
462 INIT_LIST_HEAD(&bo_va->bo_list);
463 INIT_LIST_HEAD(&bo_va->vm_list);
464 /* make sure object fit at this offset */
465 if (bo_va->soffset >= bo_va->eoffset) {
466 kfree(bo_va);
467 return -EINVAL;
468 }
469
470 last_pfn = bo_va->eoffset / RADEON_GPU_PAGE_SIZE;
471 if (last_pfn > rdev->vm_manager.max_pfn) {
472 kfree(bo_va);
473 dev_err(rdev->dev, "va above limit (0x%08X > 0x%08X)\n",
474 last_pfn, rdev->vm_manager.max_pfn);
475 return -EINVAL;
476 }
477
478 mutex_lock(&vm->mutex);
479 if (last_pfn > vm->last_pfn) {
480 /* release mutex and lock in right order */
481 mutex_unlock(&vm->mutex);
482 radeon_mutex_lock(&rdev->cs_mutex);
483 mutex_lock(&vm->mutex);
484 /* and check again */
485 if (last_pfn > vm->last_pfn) {
486 /* grow va space 32M by 32M */
487 unsigned align = ((32 << 20) >> 12) - 1;
488 radeon_vm_unbind_locked(rdev, vm);
489 vm->last_pfn = (last_pfn + align) & ~align;
490 }
491 radeon_mutex_unlock(&rdev->cs_mutex);
492 }
493 head = &vm->va;
494 last_offset = 0;
495 list_for_each_entry(tmp, &vm->va, vm_list) {
496 if (bo_va->soffset >= last_offset && bo_va->eoffset < tmp->soffset) {
497 /* bo can be added before this one */
498 break;
499 }
500 if (bo_va->soffset >= tmp->soffset && bo_va->soffset < tmp->eoffset) {
501 /* bo and tmp overlap, invalid offset */
502 dev_err(rdev->dev, "bo %p va 0x%08X conflict with (bo %p 0x%08X 0x%08X)\n",
503 bo, (unsigned)bo_va->soffset, tmp->bo,
504 (unsigned)tmp->soffset, (unsigned)tmp->eoffset);
505 kfree(bo_va);
506 mutex_unlock(&vm->mutex);
507 return -EINVAL;
508 }
509 last_offset = tmp->eoffset;
510 head = &tmp->vm_list;
511 }
512 list_add(&bo_va->vm_list, head);
513 list_add_tail(&bo_va->bo_list, &bo->va);
514 mutex_unlock(&vm->mutex);
515 return 0;
516}
517
518static u64 radeon_vm_get_addr(struct radeon_device *rdev,
519 struct ttm_mem_reg *mem,
520 unsigned pfn)
521{
522 u64 addr = 0;
523
524 switch (mem->mem_type) {
525 case TTM_PL_VRAM:
526 addr = (mem->start << PAGE_SHIFT);
527 addr += pfn * RADEON_GPU_PAGE_SIZE;
528 addr += rdev->vm_manager.vram_base_offset;
529 break;
530 case TTM_PL_TT:
531 /* offset inside page table */
532 addr = mem->start << PAGE_SHIFT;
533 addr += pfn * RADEON_GPU_PAGE_SIZE;
534 addr = addr >> PAGE_SHIFT;
535 /* page table offset */
536 addr = rdev->gart.pages_addr[addr];
537 /* in case cpu page size != gpu page size*/
538 addr += (pfn * RADEON_GPU_PAGE_SIZE) & (~PAGE_MASK);
539 break;
540 default:
541 break;
542 }
543 return addr;
544}
545
546/* object have to be reserved & cs mutex took & vm mutex took */
547int radeon_vm_bo_update_pte(struct radeon_device *rdev,
548 struct radeon_vm *vm,
549 struct radeon_bo *bo,
550 struct ttm_mem_reg *mem)
551{
552 struct radeon_bo_va *bo_va;
553 unsigned ngpu_pages, i;
554 uint64_t addr = 0, pfn;
555 uint32_t flags;
556
557 /* nothing to do if vm isn't bound */
558 if (vm->id == -1)
559 return 0;
560
561 bo_va = radeon_bo_va(bo, vm);
562 if (bo_va == NULL) {
563 dev_err(rdev->dev, "bo %p not in vm %p\n", bo, vm);
564 return -EINVAL;
565 }
566
567 if (bo_va->valid && mem)
568 return 0;
569
570 ngpu_pages = radeon_bo_ngpu_pages(bo);
571 bo_va->flags &= ~RADEON_VM_PAGE_VALID;
572 bo_va->flags &= ~RADEON_VM_PAGE_SYSTEM;
573 if (mem) {
574 if (mem->mem_type != TTM_PL_SYSTEM) {
575 bo_va->flags |= RADEON_VM_PAGE_VALID;
576 bo_va->valid = true;
577 }
578 if (mem->mem_type == TTM_PL_TT) {
579 bo_va->flags |= RADEON_VM_PAGE_SYSTEM;
580 }
581 }
582 pfn = bo_va->soffset / RADEON_GPU_PAGE_SIZE;
583 flags = rdev->vm_manager.funcs->page_flags(rdev, bo_va->vm, bo_va->flags);
584 for (i = 0, addr = 0; i < ngpu_pages; i++) {
585 if (mem && bo_va->valid) {
586 addr = radeon_vm_get_addr(rdev, mem, i);
587 }
588 rdev->vm_manager.funcs->set_page(rdev, bo_va->vm, i + pfn, addr, flags);
589 }
590 rdev->vm_manager.funcs->tlb_flush(rdev, bo_va->vm);
591 return 0;
592}
593
594/* object have to be reserved */
595int radeon_vm_bo_rmv(struct radeon_device *rdev,
596 struct radeon_vm *vm,
597 struct radeon_bo *bo)
598{
599 struct radeon_bo_va *bo_va;
600 int r;
601
602 bo_va = radeon_bo_va(bo, vm);
603 if (bo_va == NULL)
604 return 0;
605
606 /* wait for va use to end */
607 while (bo_va->fence) {
608 r = radeon_fence_wait(bo_va->fence, false);
609 if (r) {
610 DRM_ERROR("error while waiting for fence: %d\n", r);
611 }
612 if (r == -EDEADLK) {
613 r = radeon_gpu_reset(rdev);
614 if (!r)
615 continue;
616 }
617 break;
618 }
619 radeon_fence_unref(&bo_va->fence);
620
621 radeon_mutex_lock(&rdev->cs_mutex);
622 mutex_lock(&vm->mutex);
623 radeon_vm_bo_update_pte(rdev, vm, bo, NULL);
624 radeon_mutex_unlock(&rdev->cs_mutex);
625 list_del(&bo_va->vm_list);
626 mutex_unlock(&vm->mutex);
627 list_del(&bo_va->bo_list);
628
629 kfree(bo_va);
630 return 0;
631}
632
633void radeon_vm_bo_invalidate(struct radeon_device *rdev,
634 struct radeon_bo *bo)
635{
636 struct radeon_bo_va *bo_va;
637
638 BUG_ON(!atomic_read(&bo->tbo.reserved));
639 list_for_each_entry(bo_va, &bo->va, bo_list) {
640 bo_va->valid = false;
641 }
642}
643
644int radeon_vm_init(struct radeon_device *rdev, struct radeon_vm *vm)
645{
646 int r;
647
648 vm->id = -1;
649 vm->fence = NULL;
650 mutex_init(&vm->mutex);
651 INIT_LIST_HEAD(&vm->list);
652 INIT_LIST_HEAD(&vm->va);
653 /* SI requires equal sized PTs for all VMs, so always set
654 * last_pfn to max_pfn. cayman allows variable sized
655 * pts so we can grow then as needed. Once we switch
656 * to two level pts we can unify this again.
657 */
658 if (rdev->family >= CHIP_TAHITI)
659 vm->last_pfn = rdev->vm_manager.max_pfn;
660 else
661 vm->last_pfn = 0;
662 /* map the ib pool buffer at 0 in virtual address space, set
663 * read only
664 */
665 r = radeon_vm_bo_add(rdev, vm, rdev->ring_tmp_bo.bo, 0,
666 RADEON_VM_PAGE_READABLE | RADEON_VM_PAGE_SNOOPED);
667 return r;
668}
669
670void radeon_vm_fini(struct radeon_device *rdev, struct radeon_vm *vm)
671{
672 struct radeon_bo_va *bo_va, *tmp;
673 int r;
674
675 radeon_mutex_lock(&rdev->cs_mutex);
676 mutex_lock(&vm->mutex);
677 radeon_vm_unbind_locked(rdev, vm);
678 radeon_mutex_unlock(&rdev->cs_mutex);
679
680 /* remove all bo at this point non are busy any more because unbind
681 * waited for the last vm fence to signal
682 */
683 r = radeon_bo_reserve(rdev->ring_tmp_bo.bo, false);
684 if (!r) {
685 bo_va = radeon_bo_va(rdev->ring_tmp_bo.bo, vm);
686 list_del_init(&bo_va->bo_list);
687 list_del_init(&bo_va->vm_list);
688 radeon_fence_unref(&bo_va->fence);
689 radeon_bo_unreserve(rdev->ring_tmp_bo.bo);
690 kfree(bo_va);
691 }
692 if (!list_empty(&vm->va)) {
693 dev_err(rdev->dev, "still active bo inside vm\n");
694 }
695 list_for_each_entry_safe(bo_va, tmp, &vm->va, vm_list) {
696 list_del_init(&bo_va->vm_list);
697 r = radeon_bo_reserve(bo_va->bo, false);
698 if (!r) {
699 list_del_init(&bo_va->bo_list);
700 radeon_fence_unref(&bo_va->fence);
701 radeon_bo_unreserve(bo_va->bo);
702 kfree(bo_va);
703 }
704 }
705 mutex_unlock(&vm->mutex);
706}