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 #include <drm/drm_drv.h>
38 
39 /*
40  * GART
41  * The GART (Graphics Aperture Remapping Table) is an aperture
42  * in the GPU's address space.  System pages can be mapped into
43  * the aperture and look like contiguous pages from the GPU's
44  * perspective.  A page table maps the pages in the aperture
45  * to the actual backing pages in system memory.
46  *
47  * Radeon GPUs support both an internal GART, as described above,
48  * and AGP.  AGP works similarly, but the GART table is configured
49  * and maintained by the northbridge rather than the driver.
50  * Radeon hw has a separate AGP aperture that is programmed to
51  * point to the AGP aperture provided by the northbridge and the
52  * requests are passed through to the northbridge aperture.
53  * Both AGP and internal GART can be used at the same time, however
54  * that is not currently supported by the driver.
55  *
56  * This file handles the common internal GART management.
57  */
58 
59 /*
60  * Common GART table functions.
61  */
62 
63 /**
64  * amdgpu_gart_dummy_page_init - init dummy page used by the driver
65  *
66  * @adev: amdgpu_device pointer
67  *
68  * Allocate the dummy page used by the driver (all asics).
69  * This dummy page is used by the driver as a filler for gart entries
70  * when pages are taken out of the GART
71  * Returns 0 on sucess, -ENOMEM on failure.
72  */
73 static int amdgpu_gart_dummy_page_init(struct amdgpu_device *adev)
74 {
75 	struct page *dummy_page = ttm_glob.dummy_read_page;
76 
77 	if (adev->dummy_page_addr)
78 		return 0;
79 	adev->dummy_page_addr = dma_map_page(&adev->pdev->dev, dummy_page, 0,
80 					     PAGE_SIZE, DMA_BIDIRECTIONAL);
81 	if (dma_mapping_error(&adev->pdev->dev, adev->dummy_page_addr)) {
82 		dev_err(&adev->pdev->dev, "Failed to DMA MAP the dummy page\n");
83 		adev->dummy_page_addr = 0;
84 		return -ENOMEM;
85 	}
86 	return 0;
87 }
88 
89 /**
90  * amdgpu_gart_dummy_page_fini - free dummy page used by the driver
91  *
92  * @adev: amdgpu_device pointer
93  *
94  * Frees the dummy page used by the driver (all asics).
95  */
96 void amdgpu_gart_dummy_page_fini(struct amdgpu_device *adev)
97 {
98 	if (!adev->dummy_page_addr)
99 		return;
100 	dma_unmap_page(&adev->pdev->dev, adev->dummy_page_addr, PAGE_SIZE,
101 		       DMA_BIDIRECTIONAL);
102 	adev->dummy_page_addr = 0;
103 }
104 
105 /**
106  * amdgpu_gart_table_vram_alloc - allocate vram for gart page table
107  *
108  * @adev: amdgpu_device pointer
109  *
110  * Allocate video memory for GART page table
111  * (pcie r4xx, r5xx+).  These asics require the
112  * gart table to be in video memory.
113  * Returns 0 for success, error for failure.
114  */
115 int amdgpu_gart_table_vram_alloc(struct amdgpu_device *adev)
116 {
117 	int r;
118 
119 	if (adev->gart.bo == NULL) {
120 		struct amdgpu_bo_param bp;
121 
122 		memset(&bp, 0, sizeof(bp));
123 		bp.size = adev->gart.table_size;
124 		bp.byte_align = PAGE_SIZE;
125 		bp.domain = AMDGPU_GEM_DOMAIN_VRAM;
126 		bp.flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
127 			AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
128 		bp.type = ttm_bo_type_kernel;
129 		bp.resv = NULL;
130 		bp.bo_ptr_size = sizeof(struct amdgpu_bo);
131 
132 		r = amdgpu_bo_create(adev, &bp, &adev->gart.bo);
133 		if (r) {
134 			return r;
135 		}
136 	}
137 	return 0;
138 }
139 
140 /**
141  * amdgpu_gart_table_vram_pin - pin gart page table in vram
142  *
143  * @adev: amdgpu_device pointer
144  *
145  * Pin the GART page table in vram so it will not be moved
146  * by the memory manager (pcie r4xx, r5xx+).  These asics require the
147  * gart table to be in video memory.
148  * Returns 0 for success, error for failure.
149  */
150 int amdgpu_gart_table_vram_pin(struct amdgpu_device *adev)
151 {
152 	int r;
153 
154 	r = amdgpu_bo_reserve(adev->gart.bo, false);
155 	if (unlikely(r != 0))
156 		return r;
157 	r = amdgpu_bo_pin(adev->gart.bo, AMDGPU_GEM_DOMAIN_VRAM);
158 	if (r) {
159 		amdgpu_bo_unreserve(adev->gart.bo);
160 		return r;
161 	}
162 	r = amdgpu_bo_kmap(adev->gart.bo, &adev->gart.ptr);
163 	if (r)
164 		amdgpu_bo_unpin(adev->gart.bo);
165 	amdgpu_bo_unreserve(adev->gart.bo);
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.bo == NULL) {
182 		return;
183 	}
184 	r = amdgpu_bo_reserve(adev->gart.bo, true);
185 	if (likely(r == 0)) {
186 		amdgpu_bo_kunmap(adev->gart.bo);
187 		amdgpu_bo_unpin(adev->gart.bo);
188 		amdgpu_bo_unreserve(adev->gart.bo);
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.bo == NULL) {
205 		return;
206 	}
207 	amdgpu_bo_unref(&adev->gart.bo);
208 	adev->gart.ptr = NULL;
209 }
210 
211 /*
212  * Common gart functions.
213  */
214 /**
215  * amdgpu_gart_unbind - unbind pages from the gart page table
216  *
217  * @adev: amdgpu_device pointer
218  * @offset: offset into the GPU's gart aperture
219  * @pages: number of pages to unbind
220  *
221  * Unbinds the requested pages from the gart page table and
222  * replaces them with the dummy page (all asics).
223  * Returns 0 for success, -EINVAL for failure.
224  */
225 int amdgpu_gart_unbind(struct amdgpu_device *adev, uint64_t offset,
226 			int pages)
227 {
228 	unsigned t;
229 	unsigned p;
230 	int i, j;
231 	u64 page_base;
232 	/* Starting from VEGA10, system bit must be 0 to mean invalid. */
233 	uint64_t flags = 0;
234 	int idx;
235 
236 	if (!adev->gart.ready) {
237 		WARN(1, "trying to unbind memory from uninitialized GART !\n");
238 		return -EINVAL;
239 	}
240 
241 	if (!drm_dev_enter(adev_to_drm(adev), &idx))
242 		return 0;
243 
244 	t = offset / AMDGPU_GPU_PAGE_SIZE;
245 	p = t / AMDGPU_GPU_PAGES_IN_CPU_PAGE;
246 	for (i = 0; i < pages; i++, p++) {
247 		page_base = adev->dummy_page_addr;
248 		if (!adev->gart.ptr)
249 			continue;
250 
251 		for (j = 0; j < AMDGPU_GPU_PAGES_IN_CPU_PAGE; j++, t++) {
252 			amdgpu_gmc_set_pte_pde(adev, adev->gart.ptr,
253 					       t, page_base, flags);
254 			page_base += AMDGPU_GPU_PAGE_SIZE;
255 		}
256 	}
257 	mb();
258 	amdgpu_device_flush_hdp(adev, NULL);
259 	for (i = 0; i < adev->num_vmhubs; i++)
260 		amdgpu_gmc_flush_gpu_tlb(adev, 0, i, 0);
261 
262 	drm_dev_exit(idx);
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  * @flags: page table entry flags
274  * @dst: CPU address of the gart table
275  *
276  * Map the dma_addresses into GART entries (all asics).
277  * Returns 0 for success, -EINVAL for failure.
278  */
279 int amdgpu_gart_map(struct amdgpu_device *adev, uint64_t offset,
280 		    int pages, dma_addr_t *dma_addr, uint64_t flags,
281 		    void *dst)
282 {
283 	uint64_t page_base;
284 	unsigned i, j, t;
285 	int idx;
286 
287 	if (!adev->gart.ready) {
288 		WARN(1, "trying to bind memory to uninitialized GART !\n");
289 		return -EINVAL;
290 	}
291 
292 	if (!drm_dev_enter(adev_to_drm(adev), &idx))
293 		return 0;
294 
295 	t = offset / AMDGPU_GPU_PAGE_SIZE;
296 
297 	for (i = 0; i < pages; i++) {
298 		page_base = dma_addr[i];
299 		for (j = 0; j < AMDGPU_GPU_PAGES_IN_CPU_PAGE; j++, t++) {
300 			amdgpu_gmc_set_pte_pde(adev, dst, t, page_base, flags);
301 			page_base += AMDGPU_GPU_PAGE_SIZE;
302 		}
303 	}
304 	drm_dev_exit(idx);
305 	return 0;
306 }
307 
308 /**
309  * amdgpu_gart_bind - bind pages into the gart page table
310  *
311  * @adev: amdgpu_device pointer
312  * @offset: offset into the GPU's gart aperture
313  * @pages: number of pages to bind
314  * @dma_addr: DMA addresses of pages
315  * @flags: page table entry flags
316  *
317  * Binds the requested pages to the gart page table
318  * (all asics).
319  * Returns 0 for success, -EINVAL for failure.
320  */
321 int amdgpu_gart_bind(struct amdgpu_device *adev, uint64_t offset,
322 		     int pages, dma_addr_t *dma_addr,
323 		     uint64_t flags)
324 {
325 	if (!adev->gart.ready) {
326 		WARN(1, "trying to bind memory to uninitialized GART !\n");
327 		return -EINVAL;
328 	}
329 
330 	if (!adev->gart.ptr)
331 		return 0;
332 
333 	return amdgpu_gart_map(adev, offset, pages, dma_addr, flags,
334 			       adev->gart.ptr);
335 }
336 
337 /**
338  * amdgpu_gart_invalidate_tlb - invalidate gart TLB
339  *
340  * @adev: amdgpu device driver pointer
341  *
342  * Invalidate gart TLB which can be use as a way to flush gart changes
343  *
344  */
345 void amdgpu_gart_invalidate_tlb(struct amdgpu_device *adev)
346 {
347 	int i;
348 
349 	mb();
350 	amdgpu_device_flush_hdp(adev, NULL);
351 	for (i = 0; i < adev->num_vmhubs; i++)
352 		amdgpu_gmc_flush_gpu_tlb(adev, 0, i, 0);
353 }
354 
355 /**
356  * amdgpu_gart_init - init the driver info for managing the gart
357  *
358  * @adev: amdgpu_device pointer
359  *
360  * Allocate the dummy page and init the gart driver info (all asics).
361  * Returns 0 for success, error for failure.
362  */
363 int amdgpu_gart_init(struct amdgpu_device *adev)
364 {
365 	int r;
366 
367 	if (adev->dummy_page_addr)
368 		return 0;
369 
370 	/* We need PAGE_SIZE >= AMDGPU_GPU_PAGE_SIZE */
371 	if (PAGE_SIZE < AMDGPU_GPU_PAGE_SIZE) {
372 		DRM_ERROR("Page size is smaller than GPU page size!\n");
373 		return -EINVAL;
374 	}
375 	r = amdgpu_gart_dummy_page_init(adev);
376 	if (r)
377 		return r;
378 	/* Compute table size */
379 	adev->gart.num_cpu_pages = adev->gmc.gart_size / PAGE_SIZE;
380 	adev->gart.num_gpu_pages = adev->gmc.gart_size / AMDGPU_GPU_PAGE_SIZE;
381 	DRM_INFO("GART: num cpu pages %u, num gpu pages %u\n",
382 		 adev->gart.num_cpu_pages, adev->gart.num_gpu_pages);
383 
384 	return 0;
385 }
386