xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_gart.c (revision e77a8005748547fb1f10645097f13ccdd804d7e5)
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 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_attrs(&adev->pdev->dev, dummy_page, 0,
82 							PAGE_SIZE, DMA_BIDIRECTIONAL,
83 							DMA_ATTR_SKIP_CPU_SYNC);
84 	if (dma_mapping_error(&adev->pdev->dev, adev->dummy_page_addr)) {
85 		dev_err(&adev->pdev->dev, "Failed to DMA MAP the dummy page\n");
86 		adev->dummy_page_addr = 0;
87 		return -ENOMEM;
88 	}
89 	return 0;
90 }
91 
92 /**
93  * amdgpu_gart_dummy_page_fini - free dummy page used by the driver
94  *
95  * @adev: amdgpu_device pointer
96  *
97  * Frees the dummy page used by the driver (all asics).
98  */
99 void amdgpu_gart_dummy_page_fini(struct amdgpu_device *adev)
100 {
101 	if (!adev->dummy_page_addr)
102 		return;
103 	dma_unmap_page_attrs(&adev->pdev->dev, adev->dummy_page_addr, PAGE_SIZE,
104 				DMA_BIDIRECTIONAL,
105 				DMA_ATTR_SKIP_CPU_SYNC);
106 	adev->dummy_page_addr = 0;
107 }
108 
109 /**
110  * amdgpu_gart_table_ram_alloc - allocate system ram for gart page table
111  *
112  * @adev: amdgpu_device pointer
113  *
114  * Allocate system memory for GART page table for ASICs that don't have
115  * dedicated VRAM.
116  * Returns 0 for success, error for failure.
117  */
118 int amdgpu_gart_table_ram_alloc(struct amdgpu_device *adev)
119 {
120 	unsigned int order = get_order(adev->gart.table_size);
121 	gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
122 	struct amdgpu_bo *bo = NULL;
123 	struct sg_table *sg = NULL;
124 	struct amdgpu_bo_param bp;
125 	dma_addr_t dma_addr;
126 	struct page *p;
127 	unsigned long x;
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 	/* assign pages to this device */
138 	for (x = 0; x < (1UL << order); x++)
139 		p[x].mapping = adev->mman.bdev.dev_mapping;
140 
141 	/* If the hardware does not support UTCL2 snooping of the CPU caches
142 	 * then set_memory_wc() could be used as a workaround to mark the pages
143 	 * as write combine memory.
144 	 */
145 	dma_addr = dma_map_page(&adev->pdev->dev, p, 0, adev->gart.table_size,
146 				DMA_BIDIRECTIONAL);
147 	if (dma_mapping_error(&adev->pdev->dev, dma_addr)) {
148 		dev_err(&adev->pdev->dev, "Failed to DMA MAP the GART BO page\n");
149 		__free_pages(p, order);
150 		p = NULL;
151 		return -EFAULT;
152 	}
153 
154 	dev_info(adev->dev, "%s dma_addr:%pad\n", __func__, &dma_addr);
155 	/* Create SG table */
156 	sg = kmalloc(sizeof(*sg), GFP_KERNEL);
157 	if (!sg) {
158 		ret = -ENOMEM;
159 		goto error;
160 	}
161 	ret = sg_alloc_table(sg, 1, GFP_KERNEL);
162 	if (ret)
163 		goto error;
164 
165 	sg_dma_address(sg->sgl) = dma_addr;
166 	sg->sgl->length = adev->gart.table_size;
167 #ifdef CONFIG_NEED_SG_DMA_LENGTH
168 	sg->sgl->dma_length = adev->gart.table_size;
169 #endif
170 	/* Create SG BO */
171 	memset(&bp, 0, sizeof(bp));
172 	bp.size = adev->gart.table_size;
173 	bp.byte_align = PAGE_SIZE;
174 	bp.domain = AMDGPU_GEM_DOMAIN_CPU;
175 	bp.type = ttm_bo_type_sg;
176 	bp.resv = NULL;
177 	bp.bo_ptr_size = sizeof(struct amdgpu_bo);
178 	bp.flags = 0;
179 	ret = amdgpu_bo_create(adev, &bp, &bo);
180 	if (ret)
181 		goto error;
182 
183 	bo->tbo.sg = sg;
184 	bo->tbo.ttm->sg = sg;
185 	bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
186 	bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
187 
188 	ret = amdgpu_bo_reserve(bo, true);
189 	if (ret) {
190 		dev_err(adev->dev, "(%d) failed to reserve bo for GART system bo\n", ret);
191 		goto error;
192 	}
193 
194 	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
195 	WARN(ret, "Pinning the GART table failed");
196 	if (ret)
197 		goto error_resv;
198 
199 	adev->gart.bo = bo;
200 	adev->gart.ptr = page_to_virt(p);
201 	/* Make GART table accessible in VMID0 */
202 	ret = amdgpu_ttm_alloc_gart(&adev->gart.bo->tbo);
203 	if (ret)
204 		amdgpu_gart_table_ram_free(adev);
205 	amdgpu_bo_unreserve(bo);
206 
207 	return 0;
208 
209 error_resv:
210 	amdgpu_bo_unreserve(bo);
211 error:
212 	amdgpu_bo_unref(&bo);
213 	if (sg) {
214 		sg_free_table(sg);
215 		kfree(sg);
216 	}
217 	__free_pages(p, order);
218 	return ret;
219 }
220 
221 /**
222  * amdgpu_gart_table_ram_free - free gart page table system ram
223  *
224  * @adev: amdgpu_device pointer
225  *
226  * Free the system memory used for the GART page tableon ASICs that don't
227  * have dedicated VRAM.
228  */
229 void amdgpu_gart_table_ram_free(struct amdgpu_device *adev)
230 {
231 	unsigned int order = get_order(adev->gart.table_size);
232 	struct sg_table *sg = adev->gart.bo->tbo.sg;
233 	struct page *p;
234 	unsigned long x;
235 	int ret;
236 
237 	ret = amdgpu_bo_reserve(adev->gart.bo, false);
238 	if (!ret) {
239 		amdgpu_bo_unpin(adev->gart.bo);
240 		amdgpu_bo_unreserve(adev->gart.bo);
241 	}
242 	amdgpu_bo_unref(&adev->gart.bo);
243 	sg_free_table(sg);
244 	kfree(sg);
245 	p = virt_to_page(adev->gart.ptr);
246 	for (x = 0; x < (1UL << order); x++)
247 		p[x].mapping = NULL;
248 	__free_pages(p, order);
249 
250 	adev->gart.ptr = NULL;
251 }
252 
253 /**
254  * amdgpu_gart_table_vram_alloc - allocate vram for gart page table
255  *
256  * @adev: amdgpu_device pointer
257  *
258  * Allocate video memory for GART page table
259  * (pcie r4xx, r5xx+).  These asics require the
260  * gart table to be in video memory.
261  * Returns 0 for success, error for failure.
262  */
263 int amdgpu_gart_table_vram_alloc(struct amdgpu_device *adev)
264 {
265 	if (adev->gart.bo != NULL)
266 		return 0;
267 
268 	return amdgpu_bo_create_kernel(adev,  adev->gart.table_size, PAGE_SIZE,
269 				       AMDGPU_GEM_DOMAIN_VRAM, &adev->gart.bo,
270 				       NULL, (void *)&adev->gart.ptr);
271 }
272 
273 /**
274  * amdgpu_gart_table_vram_free - free gart page table vram
275  *
276  * @adev: amdgpu_device pointer
277  *
278  * Free the video memory used for the GART page table
279  * (pcie r4xx, r5xx+).  These asics require the gart table to
280  * be in video memory.
281  */
282 void amdgpu_gart_table_vram_free(struct amdgpu_device *adev)
283 {
284 	amdgpu_bo_free_kernel(&adev->gart.bo, NULL, (void *)&adev->gart.ptr);
285 }
286 
287 /*
288  * Common gart functions.
289  */
290 /**
291  * amdgpu_gart_unbind - unbind pages from the gart page table
292  *
293  * @adev: amdgpu_device pointer
294  * @offset: offset into the GPU's gart aperture
295  * @pages: number of pages to unbind
296  *
297  * Unbinds the requested pages from the gart page table and
298  * replaces them with the dummy page (all asics).
299  * Returns 0 for success, -EINVAL for failure.
300  */
301 void amdgpu_gart_unbind(struct amdgpu_device *adev, uint64_t offset,
302 			int pages)
303 {
304 	unsigned t;
305 	unsigned p;
306 	int i, j;
307 	u64 page_base;
308 	/* Starting from VEGA10, system bit must be 0 to mean invalid. */
309 	uint64_t flags = 0;
310 	int idx;
311 
312 	if (!adev->gart.ptr)
313 		return;
314 
315 	if (!drm_dev_enter(adev_to_drm(adev), &idx))
316 		return;
317 
318 	t = offset / AMDGPU_GPU_PAGE_SIZE;
319 	p = t / AMDGPU_GPU_PAGES_IN_CPU_PAGE;
320 	for (i = 0; i < pages; i++, p++) {
321 		page_base = adev->dummy_page_addr;
322 		if (!adev->gart.ptr)
323 			continue;
324 
325 		for (j = 0; j < AMDGPU_GPU_PAGES_IN_CPU_PAGE; j++, t++) {
326 			amdgpu_gmc_set_pte_pde(adev, adev->gart.ptr,
327 					       t, page_base, flags);
328 			page_base += AMDGPU_GPU_PAGE_SIZE;
329 		}
330 	}
331 	amdgpu_gart_invalidate_tlb(adev);
332 
333 	drm_dev_exit(idx);
334 }
335 
336 /**
337  * amdgpu_gart_map - map dma_addresses into GART entries
338  *
339  * @adev: amdgpu_device pointer
340  * @offset: offset into the GPU's gart aperture
341  * @pages: number of pages to bind
342  * @dma_addr: DMA addresses of pages
343  * @flags: page table entry flags
344  * @dst: CPU address of the gart table
345  *
346  * Map the dma_addresses into GART entries (all asics).
347  * Returns 0 for success, -EINVAL for failure.
348  */
349 void amdgpu_gart_map(struct amdgpu_device *adev, uint64_t offset,
350 		    int pages, dma_addr_t *dma_addr, uint64_t flags,
351 		    void *dst)
352 {
353 	uint64_t page_base;
354 	unsigned i, j, t;
355 	int idx;
356 
357 	if (!drm_dev_enter(adev_to_drm(adev), &idx))
358 		return;
359 
360 	t = offset / AMDGPU_GPU_PAGE_SIZE;
361 
362 	for (i = 0; i < pages; i++) {
363 		page_base = dma_addr[i];
364 		for (j = 0; j < AMDGPU_GPU_PAGES_IN_CPU_PAGE; j++, t++) {
365 			amdgpu_gmc_set_pte_pde(adev, dst, t, page_base, flags);
366 			page_base += AMDGPU_GPU_PAGE_SIZE;
367 		}
368 	}
369 	drm_dev_exit(idx);
370 }
371 
372 /**
373  * amdgpu_gart_bind - bind pages into the gart page table
374  *
375  * @adev: amdgpu_device pointer
376  * @offset: offset into the GPU's gart aperture
377  * @pages: number of pages to bind
378  * @dma_addr: DMA addresses of pages
379  * @flags: page table entry flags
380  *
381  * Binds the requested pages to the gart page table
382  * (all asics).
383  * Returns 0 for success, -EINVAL for failure.
384  */
385 void amdgpu_gart_bind(struct amdgpu_device *adev, uint64_t offset,
386 		     int pages, dma_addr_t *dma_addr,
387 		     uint64_t flags)
388 {
389 	if (!adev->gart.ptr)
390 		return;
391 
392 	amdgpu_gart_map(adev, offset, pages, dma_addr, flags, adev->gart.ptr);
393 }
394 
395 /**
396  * amdgpu_gart_invalidate_tlb - invalidate gart TLB
397  *
398  * @adev: amdgpu device driver pointer
399  *
400  * Invalidate gart TLB which can be use as a way to flush gart changes
401  *
402  */
403 void amdgpu_gart_invalidate_tlb(struct amdgpu_device *adev)
404 {
405 	int i;
406 
407 	if (!adev->gart.ptr)
408 		return;
409 
410 	mb();
411 	if (down_read_trylock(&adev->reset_domain->sem)) {
412 		amdgpu_device_flush_hdp(adev, NULL);
413 		up_read(&adev->reset_domain->sem);
414 	}
415 	for_each_set_bit(i, adev->vmhubs_mask, AMDGPU_MAX_VMHUBS)
416 		amdgpu_gmc_flush_gpu_tlb(adev, 0, i, 0);
417 }
418 
419 /**
420  * amdgpu_gart_init - init the driver info for managing the gart
421  *
422  * @adev: amdgpu_device pointer
423  *
424  * Allocate the dummy page and init the gart driver info (all asics).
425  * Returns 0 for success, error for failure.
426  */
427 int amdgpu_gart_init(struct amdgpu_device *adev)
428 {
429 	int r;
430 
431 	if (adev->dummy_page_addr)
432 		return 0;
433 
434 	/* We need PAGE_SIZE >= AMDGPU_GPU_PAGE_SIZE */
435 	if (PAGE_SIZE < AMDGPU_GPU_PAGE_SIZE) {
436 		DRM_ERROR("Page size is smaller than GPU page size!\n");
437 		return -EINVAL;
438 	}
439 	r = amdgpu_gart_dummy_page_init(adev);
440 	if (r)
441 		return r;
442 	/* Compute table size */
443 	adev->gart.num_cpu_pages = adev->gmc.gart_size / PAGE_SIZE;
444 	adev->gart.num_gpu_pages = adev->gmc.gart_size / AMDGPU_GPU_PAGE_SIZE;
445 	DRM_INFO("GART: num cpu pages %u, num gpu pages %u\n",
446 		 adev->gart.num_cpu_pages, adev->gart.num_gpu_pages);
447 
448 	return 0;
449 }
450