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/amdgpu_drm.h>
33 #ifdef CONFIG_X86
34 #include <asm/set_memory.h>
35 #endif
36 #include "amdgpu.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 /**
63  * amdgpu_gart_dummy_page_init - init dummy page used by the driver
64  *
65  * @adev: amdgpu_device pointer
66  *
67  * Allocate the dummy page used by the driver (all asics).
68  * This dummy page is used by the driver as a filler for gart entries
69  * when pages are taken out of the GART
70  * Returns 0 on sucess, -ENOMEM on failure.
71  */
72 static int amdgpu_gart_dummy_page_init(struct amdgpu_device *adev)
73 {
74 	struct page *dummy_page = ttm_glob.dummy_read_page;
75 
76 	if (adev->dummy_page_addr)
77 		return 0;
78 	adev->dummy_page_addr = dma_map_page(&adev->pdev->dev, dummy_page, 0,
79 					     PAGE_SIZE, DMA_BIDIRECTIONAL);
80 	if (dma_mapping_error(&adev->pdev->dev, adev->dummy_page_addr)) {
81 		dev_err(&adev->pdev->dev, "Failed to DMA MAP the dummy page\n");
82 		adev->dummy_page_addr = 0;
83 		return -ENOMEM;
84 	}
85 	return 0;
86 }
87 
88 /**
89  * amdgpu_gart_dummy_page_fini - free dummy page used by the driver
90  *
91  * @adev: amdgpu_device pointer
92  *
93  * Frees the dummy page used by the driver (all asics).
94  */
95 void amdgpu_gart_dummy_page_fini(struct amdgpu_device *adev)
96 {
97 	if (!adev->dummy_page_addr)
98 		return;
99 	dma_unmap_page(&adev->pdev->dev, adev->dummy_page_addr, PAGE_SIZE,
100 		       DMA_BIDIRECTIONAL);
101 	adev->dummy_page_addr = 0;
102 }
103 
104 /**
105  * amdgpu_gart_table_vram_alloc - allocate vram for gart page table
106  *
107  * @adev: amdgpu_device pointer
108  *
109  * Allocate video memory for GART page table
110  * (pcie r4xx, r5xx+).  These asics require the
111  * gart table to be in video memory.
112  * Returns 0 for success, error for failure.
113  */
114 int amdgpu_gart_table_vram_alloc(struct amdgpu_device *adev)
115 {
116 	int r;
117 
118 	if (adev->gart.bo == NULL) {
119 		struct amdgpu_bo_param bp;
120 
121 		memset(&bp, 0, sizeof(bp));
122 		bp.size = adev->gart.table_size;
123 		bp.byte_align = PAGE_SIZE;
124 		bp.domain = AMDGPU_GEM_DOMAIN_VRAM;
125 		bp.flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
126 			AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
127 		bp.type = ttm_bo_type_kernel;
128 		bp.resv = NULL;
129 		bp.bo_ptr_size = sizeof(struct amdgpu_bo);
130 
131 		r = amdgpu_bo_create(adev, &bp, &adev->gart.bo);
132 		if (r) {
133 			return r;
134 		}
135 	}
136 	return 0;
137 }
138 
139 /**
140  * amdgpu_gart_table_vram_pin - pin gart page table in vram
141  *
142  * @adev: amdgpu_device pointer
143  *
144  * Pin the GART page table in vram so it will not be moved
145  * by the memory manager (pcie r4xx, r5xx+).  These asics require the
146  * gart table to be in video memory.
147  * Returns 0 for success, error for failure.
148  */
149 int amdgpu_gart_table_vram_pin(struct amdgpu_device *adev)
150 {
151 	int r;
152 
153 	r = amdgpu_bo_reserve(adev->gart.bo, false);
154 	if (unlikely(r != 0))
155 		return r;
156 	r = amdgpu_bo_pin(adev->gart.bo, AMDGPU_GEM_DOMAIN_VRAM);
157 	if (r) {
158 		amdgpu_bo_unreserve(adev->gart.bo);
159 		return r;
160 	}
161 	r = amdgpu_bo_kmap(adev->gart.bo, &adev->gart.ptr);
162 	if (r)
163 		amdgpu_bo_unpin(adev->gart.bo);
164 	amdgpu_bo_unreserve(adev->gart.bo);
165 	return r;
166 }
167 
168 /**
169  * amdgpu_gart_table_vram_unpin - unpin gart page table in vram
170  *
171  * @adev: amdgpu_device pointer
172  *
173  * Unpin the GART page table in vram (pcie r4xx, r5xx+).
174  * These asics require the gart table to be in video memory.
175  */
176 void amdgpu_gart_table_vram_unpin(struct amdgpu_device *adev)
177 {
178 	int r;
179 
180 	if (adev->gart.bo == NULL) {
181 		return;
182 	}
183 	r = amdgpu_bo_reserve(adev->gart.bo, true);
184 	if (likely(r == 0)) {
185 		amdgpu_bo_kunmap(adev->gart.bo);
186 		amdgpu_bo_unpin(adev->gart.bo);
187 		amdgpu_bo_unreserve(adev->gart.bo);
188 		adev->gart.ptr = NULL;
189 	}
190 }
191 
192 /**
193  * amdgpu_gart_table_vram_free - free gart page table vram
194  *
195  * @adev: amdgpu_device pointer
196  *
197  * Free the video memory used for the GART page table
198  * (pcie r4xx, r5xx+).  These asics require the gart table to
199  * be in video memory.
200  */
201 void amdgpu_gart_table_vram_free(struct amdgpu_device *adev)
202 {
203 	if (adev->gart.bo == NULL) {
204 		return;
205 	}
206 	amdgpu_bo_unref(&adev->gart.bo);
207 	adev->gart.ptr = NULL;
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  * Returns 0 for success, -EINVAL for failure.
223  */
224 int amdgpu_gart_unbind(struct amdgpu_device *adev, uint64_t offset,
225 			int pages)
226 {
227 	unsigned t;
228 	unsigned p;
229 	int i, j;
230 	u64 page_base;
231 	/* Starting from VEGA10, system bit must be 0 to mean invalid. */
232 	uint64_t flags = 0;
233 
234 	if (!adev->gart.ready) {
235 		WARN(1, "trying to unbind memory from uninitialized GART !\n");
236 		return -EINVAL;
237 	}
238 
239 	t = offset / AMDGPU_GPU_PAGE_SIZE;
240 	p = t / AMDGPU_GPU_PAGES_IN_CPU_PAGE;
241 	for (i = 0; i < pages; i++, p++) {
242 		page_base = adev->dummy_page_addr;
243 		if (!adev->gart.ptr)
244 			continue;
245 
246 		for (j = 0; j < AMDGPU_GPU_PAGES_IN_CPU_PAGE; j++, t++) {
247 			amdgpu_gmc_set_pte_pde(adev, adev->gart.ptr,
248 					       t, page_base, flags);
249 			page_base += AMDGPU_GPU_PAGE_SIZE;
250 		}
251 	}
252 	mb();
253 	amdgpu_device_flush_hdp(adev, NULL);
254 	for (i = 0; i < adev->num_vmhubs; i++)
255 		amdgpu_gmc_flush_gpu_tlb(adev, 0, i, 0);
256 
257 	return 0;
258 }
259 
260 /**
261  * amdgpu_gart_map - map dma_addresses into GART entries
262  *
263  * @adev: amdgpu_device pointer
264  * @offset: offset into the GPU's gart aperture
265  * @pages: number of pages to bind
266  * @dma_addr: DMA addresses of pages
267  * @flags: page table entry flags
268  * @dst: CPU address of the gart table
269  *
270  * Map the dma_addresses into GART entries (all asics).
271  * Returns 0 for success, -EINVAL for failure.
272  */
273 int amdgpu_gart_map(struct amdgpu_device *adev, uint64_t offset,
274 		    int pages, dma_addr_t *dma_addr, uint64_t flags,
275 		    void *dst)
276 {
277 	uint64_t page_base;
278 	unsigned i, j, t;
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 
287 	for (i = 0; i < pages; i++) {
288 		page_base = dma_addr[i];
289 		for (j = 0; j < AMDGPU_GPU_PAGES_IN_CPU_PAGE; j++, t++) {
290 			amdgpu_gmc_set_pte_pde(adev, dst, t, page_base, flags);
291 			page_base += AMDGPU_GPU_PAGE_SIZE;
292 		}
293 	}
294 	return 0;
295 }
296 
297 /**
298  * amdgpu_gart_bind - bind pages into the gart page table
299  *
300  * @adev: amdgpu_device pointer
301  * @offset: offset into the GPU's gart aperture
302  * @pages: number of pages to bind
303  * @dma_addr: DMA addresses of pages
304  * @flags: page table entry flags
305  *
306  * Binds the requested pages to the gart page table
307  * (all asics).
308  * Returns 0 for success, -EINVAL for failure.
309  */
310 int amdgpu_gart_bind(struct amdgpu_device *adev, uint64_t offset,
311 		     int pages, dma_addr_t *dma_addr,
312 		     uint64_t flags)
313 {
314 	if (!adev->gart.ready) {
315 		WARN(1, "trying to bind memory to uninitialized GART !\n");
316 		return -EINVAL;
317 	}
318 
319 	if (!adev->gart.ptr)
320 		return 0;
321 
322 	return amdgpu_gart_map(adev, offset, pages, dma_addr, flags,
323 			       adev->gart.ptr);
324 }
325 
326 /**
327  * amdgpu_gart_invalidate_tlb - invalidate gart TLB
328  *
329  * @adev: amdgpu device driver pointer
330  *
331  * Invalidate gart TLB which can be use as a way to flush gart changes
332  *
333  */
334 void amdgpu_gart_invalidate_tlb(struct amdgpu_device *adev)
335 {
336 	int i;
337 
338 	mb();
339 	amdgpu_device_flush_hdp(adev, NULL);
340 	for (i = 0; i < adev->num_vmhubs; i++)
341 		amdgpu_gmc_flush_gpu_tlb(adev, 0, i, 0);
342 }
343 
344 /**
345  * amdgpu_gart_init - init the driver info for managing the gart
346  *
347  * @adev: amdgpu_device pointer
348  *
349  * Allocate the dummy page and init the gart driver info (all asics).
350  * Returns 0 for success, error for failure.
351  */
352 int amdgpu_gart_init(struct amdgpu_device *adev)
353 {
354 	int r;
355 
356 	if (adev->dummy_page_addr)
357 		return 0;
358 
359 	/* We need PAGE_SIZE >= AMDGPU_GPU_PAGE_SIZE */
360 	if (PAGE_SIZE < AMDGPU_GPU_PAGE_SIZE) {
361 		DRM_ERROR("Page size is smaller than GPU page size!\n");
362 		return -EINVAL;
363 	}
364 	r = amdgpu_gart_dummy_page_init(adev);
365 	if (r)
366 		return r;
367 	/* Compute table size */
368 	adev->gart.num_cpu_pages = adev->gmc.gart_size / PAGE_SIZE;
369 	adev->gart.num_gpu_pages = adev->gmc.gart_size / AMDGPU_GPU_PAGE_SIZE;
370 	DRM_INFO("GART: num cpu pages %u, num gpu pages %u\n",
371 		 adev->gart.num_cpu_pages, adev->gart.num_gpu_pages);
372 
373 	return 0;
374 }
375