xref: /openbsd-src/sys/dev/pci/drm/amd/amdgpu/amdgpu_gart.c (revision f556b3f0f57c35df9d181b99eddecf955f164ab6)
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 "amdgpu_reset.h"
38 #include <drm/drm_drv.h>
39 #include <drm/ttm/ttm_tt.h>
40 
41 /*
42  * GART
43  * The GART (Graphics Aperture Remapping Table) is an aperture
44  * in the GPU's address space.  System pages can be mapped into
45  * the aperture and look like contiguous pages from the GPU's
46  * perspective.  A page table maps the pages in the aperture
47  * to the actual backing pages in system memory.
48  *
49  * Radeon GPUs support both an internal GART, as described above,
50  * and AGP.  AGP works similarly, but the GART table is configured
51  * and maintained by the northbridge rather than the driver.
52  * Radeon hw has a separate AGP aperture that is programmed to
53  * point to the AGP aperture provided by the northbridge and the
54  * requests are passed through to the northbridge aperture.
55  * Both AGP and internal GART can be used at the same time, however
56  * that is not currently supported by the driver.
57  *
58  * This file handles the common internal GART management.
59  */
60 
61 /*
62  * Common GART table functions.
63  */
64 
65 /**
66  * amdgpu_gart_dummy_page_init - init dummy page used by the driver
67  *
68  * @adev: amdgpu_device pointer
69  *
70  * Allocate the dummy page used by the driver (all asics).
71  * This dummy page is used by the driver as a filler for gart entries
72  * when pages are taken out of the GART
73  * Returns 0 on sucess, -ENOMEM on failure.
74  */
75 static int amdgpu_gart_dummy_page_init(struct amdgpu_device *adev)
76 {
77 	struct vm_page *dummy_page = ttm_glob.dummy_read_page;
78 
79 	if (adev->dummy_page_addr)
80 		return 0;
81 	adev->dummy_page_addr = dma_map_page(&adev->pdev->dev, dummy_page, 0,
82 					     PAGE_SIZE, DMA_BIDIRECTIONAL);
83 	if (dma_mapping_error(&adev->pdev->dev, adev->dummy_page_addr)) {
84 		dev_err(&adev->pdev->dev, "Failed to DMA MAP the dummy page\n");
85 		adev->dummy_page_addr = 0;
86 		return -ENOMEM;
87 	}
88 	return 0;
89 }
90 
91 /**
92  * amdgpu_gart_dummy_page_fini - free dummy page used by the driver
93  *
94  * @adev: amdgpu_device pointer
95  *
96  * Frees the dummy page used by the driver (all asics).
97  */
98 void amdgpu_gart_dummy_page_fini(struct amdgpu_device *adev)
99 {
100 	if (!adev->dummy_page_addr)
101 		return;
102 	dma_unmap_page(&adev->pdev->dev, adev->dummy_page_addr, PAGE_SIZE,
103 		       DMA_BIDIRECTIONAL);
104 	adev->dummy_page_addr = 0;
105 }
106 
107 /**
108  * amdgpu_gart_table_ram_alloc - allocate system ram for gart page table
109  *
110  * @adev: amdgpu_device pointer
111  *
112  * Allocate system memory for GART page table for ASICs that don't have
113  * dedicated VRAM.
114  * Returns 0 for success, error for failure.
115  */
116 int amdgpu_gart_table_ram_alloc(struct amdgpu_device *adev)
117 {
118 	STUB();
119 	return -ENOSYS;
120 #ifdef notyet
121 	unsigned int order = get_order(adev->gart.table_size);
122 	gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
123 	struct amdgpu_bo *bo = NULL;
124 	struct sg_table *sg = NULL;
125 	struct amdgpu_bo_param bp;
126 	dma_addr_t dma_addr;
127 	struct vm_page *p;
128 	int ret;
129 
130 	if (adev->gart.bo != NULL)
131 		return 0;
132 
133 	p = alloc_pages(gfp_flags, order);
134 	if (!p)
135 		return -ENOMEM;
136 
137 	/* If the hardware does not support UTCL2 snooping of the CPU caches
138 	 * then set_memory_wc() could be used as a workaround to mark the pages
139 	 * as write combine memory.
140 	 */
141 	dma_addr = dma_map_page(&adev->pdev->dev, p, 0, adev->gart.table_size,
142 				DMA_BIDIRECTIONAL);
143 	if (dma_mapping_error(&adev->pdev->dev, dma_addr)) {
144 		dev_err(&adev->pdev->dev, "Failed to DMA MAP the GART BO page\n");
145 		__free_pages(p, order);
146 		p = NULL;
147 		return -EFAULT;
148 	}
149 
150 	dev_info(adev->dev, "%s dma_addr:%pad\n", __func__, &dma_addr);
151 	/* Create SG table */
152 	sg = kmalloc(sizeof(*sg), GFP_KERNEL);
153 	if (!sg) {
154 		ret = -ENOMEM;
155 		goto error;
156 	}
157 	ret = sg_alloc_table(sg, 1, GFP_KERNEL);
158 	if (ret)
159 		goto error;
160 
161 	sg_dma_address(sg->sgl) = dma_addr;
162 	sg->sgl->length = adev->gart.table_size;
163 #ifdef CONFIG_NEED_SG_DMA_LENGTH
164 	sg->sgl->dma_length = adev->gart.table_size;
165 #endif
166 	/* Create SG BO */
167 	memset(&bp, 0, sizeof(bp));
168 	bp.size = adev->gart.table_size;
169 	bp.byte_align = PAGE_SIZE;
170 	bp.domain = AMDGPU_GEM_DOMAIN_CPU;
171 	bp.type = ttm_bo_type_sg;
172 	bp.resv = NULL;
173 	bp.bo_ptr_size = sizeof(struct amdgpu_bo);
174 	bp.flags = 0;
175 	ret = amdgpu_bo_create(adev, &bp, &bo);
176 	if (ret)
177 		goto error;
178 
179 	bo->tbo.sg = sg;
180 	bo->tbo.ttm->sg = sg;
181 	bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
182 	bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
183 
184 	ret = amdgpu_bo_reserve(bo, true);
185 	if (ret) {
186 		dev_err(adev->dev, "(%d) failed to reserve bo for GART system bo\n", ret);
187 		goto error;
188 	}
189 
190 	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
191 	WARN(ret, "Pinning the GART table failed");
192 	if (ret)
193 		goto error_resv;
194 
195 	adev->gart.bo = bo;
196 	adev->gart.ptr = page_to_virt(p);
197 	/* Make GART table accessible in VMID0 */
198 	ret = amdgpu_ttm_alloc_gart(&adev->gart.bo->tbo);
199 	if (ret)
200 		amdgpu_gart_table_ram_free(adev);
201 	amdgpu_bo_unreserve(bo);
202 
203 	return 0;
204 
205 error_resv:
206 	amdgpu_bo_unreserve(bo);
207 error:
208 	amdgpu_bo_unref(&bo);
209 	if (sg) {
210 		sg_free_table(sg);
211 		kfree(sg);
212 	}
213 	__free_pages(p, order);
214 	return ret;
215 #endif
216 }
217 
218 /**
219  * amdgpu_gart_table_ram_free - free gart page table system ram
220  *
221  * @adev: amdgpu_device pointer
222  *
223  * Free the system memory used for the GART page tableon ASICs that don't
224  * have dedicated VRAM.
225  */
226 void amdgpu_gart_table_ram_free(struct amdgpu_device *adev)
227 {
228 	unsigned int order = get_order(adev->gart.table_size);
229 	struct sg_table *sg = adev->gart.bo->tbo.sg;
230 	struct vm_page *p;
231 	int ret;
232 
233 	ret = amdgpu_bo_reserve(adev->gart.bo, false);
234 	if (!ret) {
235 		amdgpu_bo_unpin(adev->gart.bo);
236 		amdgpu_bo_unreserve(adev->gart.bo);
237 	}
238 	amdgpu_bo_unref(&adev->gart.bo);
239 	sg_free_table(sg);
240 	kfree(sg);
241 	p = virt_to_page(adev->gart.ptr);
242 	__free_pages(p, order);
243 
244 	adev->gart.ptr = NULL;
245 }
246 
247 /**
248  * amdgpu_gart_table_vram_alloc - allocate vram for gart page table
249  *
250  * @adev: amdgpu_device pointer
251  *
252  * Allocate video memory for GART page table
253  * (pcie r4xx, r5xx+).  These asics require the
254  * gart table to be in video memory.
255  * Returns 0 for success, error for failure.
256  */
257 int amdgpu_gart_table_vram_alloc(struct amdgpu_device *adev)
258 {
259 	if (adev->gart.bo != NULL)
260 		return 0;
261 
262 	return amdgpu_bo_create_kernel(adev,  adev->gart.table_size, PAGE_SIZE,
263 				       AMDGPU_GEM_DOMAIN_VRAM, &adev->gart.bo,
264 				       NULL, (void *)&adev->gart.ptr);
265 }
266 
267 /**
268  * amdgpu_gart_table_vram_free - free gart page table vram
269  *
270  * @adev: amdgpu_device pointer
271  *
272  * Free the video memory used for the GART page table
273  * (pcie r4xx, r5xx+).  These asics require the gart table to
274  * be in video memory.
275  */
276 void amdgpu_gart_table_vram_free(struct amdgpu_device *adev)
277 {
278 	amdgpu_bo_free_kernel(&adev->gart.bo, NULL, (void *)&adev->gart.ptr);
279 }
280 
281 /*
282  * Common gart functions.
283  */
284 /**
285  * amdgpu_gart_unbind - unbind pages from the gart page table
286  *
287  * @adev: amdgpu_device pointer
288  * @offset: offset into the GPU's gart aperture
289  * @pages: number of pages to unbind
290  *
291  * Unbinds the requested pages from the gart page table and
292  * replaces them with the dummy page (all asics).
293  * Returns 0 for success, -EINVAL for failure.
294  */
295 void amdgpu_gart_unbind(struct amdgpu_device *adev, uint64_t offset,
296 			int pages)
297 {
298 	unsigned t;
299 	unsigned p;
300 	int i, j;
301 	u64 page_base;
302 	/* Starting from VEGA10, system bit must be 0 to mean invalid. */
303 	uint64_t flags = 0;
304 	int idx;
305 
306 	if (!adev->gart.ptr)
307 		return;
308 
309 	if (!drm_dev_enter(adev_to_drm(adev), &idx))
310 		return;
311 
312 	t = offset / AMDGPU_GPU_PAGE_SIZE;
313 	p = t / AMDGPU_GPU_PAGES_IN_CPU_PAGE;
314 	for (i = 0; i < pages; i++, p++) {
315 		page_base = adev->dummy_page_addr;
316 		if (!adev->gart.ptr)
317 			continue;
318 
319 		for (j = 0; j < AMDGPU_GPU_PAGES_IN_CPU_PAGE; j++, t++) {
320 			amdgpu_gmc_set_pte_pde(adev, adev->gart.ptr,
321 					       t, page_base, flags);
322 			page_base += AMDGPU_GPU_PAGE_SIZE;
323 		}
324 	}
325 	mb();
326 	amdgpu_device_flush_hdp(adev, NULL);
327 	for_each_set_bit(i, adev->vmhubs_mask, AMDGPU_MAX_VMHUBS)
328 		amdgpu_gmc_flush_gpu_tlb(adev, 0, i, 0);
329 
330 	drm_dev_exit(idx);
331 }
332 
333 /**
334  * amdgpu_gart_map - map dma_addresses into GART entries
335  *
336  * @adev: amdgpu_device pointer
337  * @offset: offset into the GPU's gart aperture
338  * @pages: number of pages to bind
339  * @dma_addr: DMA addresses of pages
340  * @flags: page table entry flags
341  * @dst: CPU address of the gart table
342  *
343  * Map the dma_addresses into GART entries (all asics).
344  * Returns 0 for success, -EINVAL for failure.
345  */
346 void amdgpu_gart_map(struct amdgpu_device *adev, uint64_t offset,
347 		    int pages, dma_addr_t *dma_addr, uint64_t flags,
348 		    void *dst)
349 {
350 	uint64_t page_base;
351 	unsigned i, j, t;
352 	int idx;
353 
354 	if (!drm_dev_enter(adev_to_drm(adev), &idx))
355 		return;
356 
357 	t = offset / AMDGPU_GPU_PAGE_SIZE;
358 
359 	for (i = 0; i < pages; i++) {
360 		page_base = dma_addr[i];
361 		for (j = 0; j < AMDGPU_GPU_PAGES_IN_CPU_PAGE; j++, t++) {
362 			amdgpu_gmc_set_pte_pde(adev, dst, t, page_base, flags);
363 			page_base += AMDGPU_GPU_PAGE_SIZE;
364 		}
365 	}
366 	drm_dev_exit(idx);
367 }
368 
369 /**
370  * amdgpu_gart_bind - bind pages into the gart page table
371  *
372  * @adev: amdgpu_device pointer
373  * @offset: offset into the GPU's gart aperture
374  * @pages: number of pages to bind
375  * @dma_addr: DMA addresses of pages
376  * @flags: page table entry flags
377  *
378  * Binds the requested pages to the gart page table
379  * (all asics).
380  * Returns 0 for success, -EINVAL for failure.
381  */
382 void amdgpu_gart_bind(struct amdgpu_device *adev, uint64_t offset,
383 		     int pages, dma_addr_t *dma_addr,
384 		     uint64_t flags)
385 {
386 	if (!adev->gart.ptr)
387 		return;
388 
389 	amdgpu_gart_map(adev, offset, pages, dma_addr, flags, adev->gart.ptr);
390 }
391 
392 /**
393  * amdgpu_gart_invalidate_tlb - invalidate gart TLB
394  *
395  * @adev: amdgpu device driver pointer
396  *
397  * Invalidate gart TLB which can be use as a way to flush gart changes
398  *
399  */
400 void amdgpu_gart_invalidate_tlb(struct amdgpu_device *adev)
401 {
402 	int i;
403 
404 	if (!adev->gart.ptr)
405 		return;
406 
407 	mb();
408 	if (down_read_trylock(&adev->reset_domain->sem)) {
409 		amdgpu_device_flush_hdp(adev, NULL);
410 		up_read(&adev->reset_domain->sem);
411 	}
412 	for_each_set_bit(i, adev->vmhubs_mask, AMDGPU_MAX_VMHUBS)
413 		amdgpu_gmc_flush_gpu_tlb(adev, 0, i, 0);
414 }
415 
416 /**
417  * amdgpu_gart_init - init the driver info for managing the gart
418  *
419  * @adev: amdgpu_device pointer
420  *
421  * Allocate the dummy page and init the gart driver info (all asics).
422  * Returns 0 for success, error for failure.
423  */
424 int amdgpu_gart_init(struct amdgpu_device *adev)
425 {
426 	int r;
427 
428 	if (adev->dummy_page_addr)
429 		return 0;
430 
431 	/* We need PAGE_SIZE >= AMDGPU_GPU_PAGE_SIZE */
432 	if (PAGE_SIZE < AMDGPU_GPU_PAGE_SIZE) {
433 		DRM_ERROR("Page size is smaller than GPU page size!\n");
434 		return -EINVAL;
435 	}
436 	r = amdgpu_gart_dummy_page_init(adev);
437 	if (r)
438 		return r;
439 	/* Compute table size */
440 	adev->gart.num_cpu_pages = adev->gmc.gart_size / PAGE_SIZE;
441 	adev->gart.num_gpu_pages = adev->gmc.gart_size / AMDGPU_GPU_PAGE_SIZE;
442 	DRM_INFO("GART: num cpu pages %u, num gpu pages %u\n",
443 		 adev->gart.num_cpu_pages, adev->gart.num_gpu_pages);
444 
445 	return 0;
446 }
447