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 #include "amdgpu.h"
31 
32 /*
33  * GART
34  * The GART (Graphics Aperture Remapping Table) is an aperture
35  * in the GPU's address space.  System pages can be mapped into
36  * the aperture and look like contiguous pages from the GPU's
37  * perspective.  A page table maps the pages in the aperture
38  * to the actual backing pages in system memory.
39  *
40  * Radeon GPUs support both an internal GART, as described above,
41  * and AGP.  AGP works similarly, but the GART table is configured
42  * and maintained by the northbridge rather than the driver.
43  * Radeon hw has a separate AGP aperture that is programmed to
44  * point to the AGP aperture provided by the northbridge and the
45  * requests are passed through to the northbridge aperture.
46  * Both AGP and internal GART can be used at the same time, however
47  * that is not currently supported by the driver.
48  *
49  * This file handles the common internal GART management.
50  */
51 
52 /*
53  * Common GART table functions.
54  */
55 /**
56  * amdgpu_gart_table_ram_alloc - allocate system ram for gart page table
57  *
58  * @adev: amdgpu_device pointer
59  *
60  * Allocate system memory for GART page table
61  * (r1xx-r3xx, non-pcie r4xx, rs400).  These asics require the
62  * gart table to be in system memory.
63  * Returns 0 for success, -ENOMEM for failure.
64  */
65 int amdgpu_gart_table_ram_alloc(struct amdgpu_device *adev)
66 {
67 	void *ptr;
68 
69 	ptr = pci_alloc_consistent(adev->pdev, adev->gart.table_size,
70 				   &adev->gart.table_addr);
71 	if (ptr == NULL) {
72 		return -ENOMEM;
73 	}
74 #ifdef CONFIG_X86
75 	if (0) {
76 		set_memory_uc((unsigned long)ptr,
77 			      adev->gart.table_size >> PAGE_SHIFT);
78 	}
79 #endif
80 	adev->gart.ptr = ptr;
81 	memset((void *)adev->gart.ptr, 0, adev->gart.table_size);
82 	return 0;
83 }
84 
85 /**
86  * amdgpu_gart_table_ram_free - free system ram for gart page table
87  *
88  * @adev: amdgpu_device pointer
89  *
90  * Free system memory for GART page table
91  * (r1xx-r3xx, non-pcie r4xx, rs400).  These asics require the
92  * gart table to be in system memory.
93  */
94 void amdgpu_gart_table_ram_free(struct amdgpu_device *adev)
95 {
96 	if (adev->gart.ptr == NULL) {
97 		return;
98 	}
99 #ifdef CONFIG_X86
100 	if (0) {
101 		set_memory_wb((unsigned long)adev->gart.ptr,
102 			      adev->gart.table_size >> PAGE_SHIFT);
103 	}
104 #endif
105 	pci_free_consistent(adev->pdev, adev->gart.table_size,
106 			    (void *)adev->gart.ptr,
107 			    adev->gart.table_addr);
108 	adev->gart.ptr = NULL;
109 	adev->gart.table_addr = 0;
110 }
111 
112 /**
113  * amdgpu_gart_table_vram_alloc - allocate vram for gart page table
114  *
115  * @adev: amdgpu_device pointer
116  *
117  * Allocate video memory for GART page table
118  * (pcie r4xx, r5xx+).  These asics require the
119  * gart table to be in video memory.
120  * Returns 0 for success, error for failure.
121  */
122 int amdgpu_gart_table_vram_alloc(struct amdgpu_device *adev)
123 {
124 	int r;
125 
126 	if (adev->gart.robj == NULL) {
127 		r = amdgpu_bo_create(adev, adev->gart.table_size,
128 				     PAGE_SIZE, true, AMDGPU_GEM_DOMAIN_VRAM, 0,
129 				     NULL, &adev->gart.robj);
130 		if (r) {
131 			return r;
132 		}
133 	}
134 	return 0;
135 }
136 
137 /**
138  * amdgpu_gart_table_vram_pin - pin gart page table in vram
139  *
140  * @adev: amdgpu_device pointer
141  *
142  * Pin the GART page table in vram so it will not be moved
143  * by the memory manager (pcie r4xx, r5xx+).  These asics require the
144  * gart table to be in video memory.
145  * Returns 0 for success, error for failure.
146  */
147 int amdgpu_gart_table_vram_pin(struct amdgpu_device *adev)
148 {
149 	uint64_t gpu_addr;
150 	int r;
151 
152 	r = amdgpu_bo_reserve(adev->gart.robj, false);
153 	if (unlikely(r != 0))
154 		return r;
155 	r = amdgpu_bo_pin(adev->gart.robj,
156 				AMDGPU_GEM_DOMAIN_VRAM, &gpu_addr);
157 	if (r) {
158 		amdgpu_bo_unreserve(adev->gart.robj);
159 		return r;
160 	}
161 	r = amdgpu_bo_kmap(adev->gart.robj, &adev->gart.ptr);
162 	if (r)
163 		amdgpu_bo_unpin(adev->gart.robj);
164 	amdgpu_bo_unreserve(adev->gart.robj);
165 	adev->gart.table_addr = gpu_addr;
166 	return r;
167 }
168 
169 /**
170  * amdgpu_gart_table_vram_unpin - unpin gart page table in vram
171  *
172  * @adev: amdgpu_device pointer
173  *
174  * Unpin the GART page table in vram (pcie r4xx, r5xx+).
175  * These asics require the gart table to be in video memory.
176  */
177 void amdgpu_gart_table_vram_unpin(struct amdgpu_device *adev)
178 {
179 	int r;
180 
181 	if (adev->gart.robj == NULL) {
182 		return;
183 	}
184 	r = amdgpu_bo_reserve(adev->gart.robj, false);
185 	if (likely(r == 0)) {
186 		amdgpu_bo_kunmap(adev->gart.robj);
187 		amdgpu_bo_unpin(adev->gart.robj);
188 		amdgpu_bo_unreserve(adev->gart.robj);
189 		adev->gart.ptr = NULL;
190 	}
191 }
192 
193 /**
194  * amdgpu_gart_table_vram_free - free gart page table vram
195  *
196  * @adev: amdgpu_device pointer
197  *
198  * Free the video memory used for the GART page table
199  * (pcie r4xx, r5xx+).  These asics require the gart table to
200  * be in video memory.
201  */
202 void amdgpu_gart_table_vram_free(struct amdgpu_device *adev)
203 {
204 	if (adev->gart.robj == NULL) {
205 		return;
206 	}
207 	amdgpu_bo_unref(&adev->gart.robj);
208 }
209 
210 /*
211  * Common gart functions.
212  */
213 /**
214  * amdgpu_gart_unbind - unbind pages from the gart page table
215  *
216  * @adev: amdgpu_device pointer
217  * @offset: offset into the GPU's gart aperture
218  * @pages: number of pages to unbind
219  *
220  * Unbinds the requested pages from the gart page table and
221  * replaces them with the dummy page (all asics).
222  */
223 void amdgpu_gart_unbind(struct amdgpu_device *adev, unsigned offset,
224 			int pages)
225 {
226 	unsigned t;
227 	unsigned p;
228 	int i, j;
229 	u64 page_base;
230 	uint32_t flags = AMDGPU_PTE_SYSTEM;
231 
232 	if (!adev->gart.ready) {
233 		WARN(1, "trying to unbind memory from uninitialized GART !\n");
234 		return;
235 	}
236 
237 	t = offset / AMDGPU_GPU_PAGE_SIZE;
238 	p = t / (PAGE_SIZE / AMDGPU_GPU_PAGE_SIZE);
239 	for (i = 0; i < pages; i++, p++) {
240 		if (adev->gart.pages[p]) {
241 			adev->gart.pages[p] = NULL;
242 			adev->gart.pages_addr[p] = adev->dummy_page.addr;
243 			page_base = adev->gart.pages_addr[p];
244 			if (!adev->gart.ptr)
245 				continue;
246 
247 			for (j = 0; j < (PAGE_SIZE / AMDGPU_GPU_PAGE_SIZE); j++, t++) {
248 				amdgpu_gart_set_pte_pde(adev, adev->gart.ptr,
249 							t, page_base, flags);
250 				page_base += AMDGPU_GPU_PAGE_SIZE;
251 			}
252 		}
253 	}
254 	mb();
255 	amdgpu_gart_flush_gpu_tlb(adev, 0);
256 }
257 
258 /**
259  * amdgpu_gart_bind - bind pages into the gart page table
260  *
261  * @adev: amdgpu_device pointer
262  * @offset: offset into the GPU's gart aperture
263  * @pages: number of pages to bind
264  * @pagelist: pages to bind
265  * @dma_addr: DMA addresses of pages
266  *
267  * Binds the requested pages to the gart page table
268  * (all asics).
269  * Returns 0 for success, -EINVAL for failure.
270  */
271 int amdgpu_gart_bind(struct amdgpu_device *adev, unsigned offset,
272 		     int pages, struct page **pagelist, dma_addr_t *dma_addr,
273 		     uint32_t flags)
274 {
275 	unsigned t;
276 	unsigned p;
277 	uint64_t page_base;
278 	int i, j;
279 
280 	if (!adev->gart.ready) {
281 		WARN(1, "trying to bind memory to uninitialized GART !\n");
282 		return -EINVAL;
283 	}
284 
285 	t = offset / AMDGPU_GPU_PAGE_SIZE;
286 	p = t / (PAGE_SIZE / AMDGPU_GPU_PAGE_SIZE);
287 
288 	for (i = 0; i < pages; i++, p++) {
289 		adev->gart.pages_addr[p] = dma_addr[i];
290 		adev->gart.pages[p] = pagelist[i];
291 		if (adev->gart.ptr) {
292 			page_base = adev->gart.pages_addr[p];
293 			for (j = 0; j < (PAGE_SIZE / AMDGPU_GPU_PAGE_SIZE); j++, t++) {
294 				amdgpu_gart_set_pte_pde(adev, adev->gart.ptr, t, page_base, flags);
295 				page_base += AMDGPU_GPU_PAGE_SIZE;
296 			}
297 		}
298 	}
299 	mb();
300 	amdgpu_gart_flush_gpu_tlb(adev, 0);
301 	return 0;
302 }
303 
304 /**
305  * amdgpu_gart_init - init the driver info for managing the gart
306  *
307  * @adev: amdgpu_device pointer
308  *
309  * Allocate the dummy page and init the gart driver info (all asics).
310  * Returns 0 for success, error for failure.
311  */
312 int amdgpu_gart_init(struct amdgpu_device *adev)
313 {
314 	int r, i;
315 
316 	if (adev->gart.pages) {
317 		return 0;
318 	}
319 	/* We need PAGE_SIZE >= AMDGPU_GPU_PAGE_SIZE */
320 	if (PAGE_SIZE < AMDGPU_GPU_PAGE_SIZE) {
321 		DRM_ERROR("Page size is smaller than GPU page size!\n");
322 		return -EINVAL;
323 	}
324 	r = amdgpu_dummy_page_init(adev);
325 	if (r)
326 		return r;
327 	/* Compute table size */
328 	adev->gart.num_cpu_pages = adev->mc.gtt_size / PAGE_SIZE;
329 	adev->gart.num_gpu_pages = adev->mc.gtt_size / AMDGPU_GPU_PAGE_SIZE;
330 	DRM_INFO("GART: num cpu pages %u, num gpu pages %u\n",
331 		 adev->gart.num_cpu_pages, adev->gart.num_gpu_pages);
332 	/* Allocate pages table */
333 	adev->gart.pages = vzalloc(sizeof(void *) * adev->gart.num_cpu_pages);
334 	if (adev->gart.pages == NULL) {
335 		amdgpu_gart_fini(adev);
336 		return -ENOMEM;
337 	}
338 	adev->gart.pages_addr = vzalloc(sizeof(dma_addr_t) *
339 					adev->gart.num_cpu_pages);
340 	if (adev->gart.pages_addr == NULL) {
341 		amdgpu_gart_fini(adev);
342 		return -ENOMEM;
343 	}
344 	/* set GART entry to point to the dummy page by default */
345 	for (i = 0; i < adev->gart.num_cpu_pages; i++) {
346 		adev->gart.pages_addr[i] = adev->dummy_page.addr;
347 	}
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 {
360 	if (adev->gart.pages && adev->gart.pages_addr && adev->gart.ready) {
361 		/* unbind pages */
362 		amdgpu_gart_unbind(adev, 0, adev->gart.num_cpu_pages);
363 	}
364 	adev->gart.ready = false;
365 	vfree(adev->gart.pages);
366 	vfree(adev->gart.pages_addr);
367 	adev->gart.pages = NULL;
368 	adev->gart.pages_addr = NULL;
369 
370 	amdgpu_dummy_page_fini(adev);
371 }
372