xref: /netbsd-src/sys/external/bsd/drm2/dist/drm/amd/amdgpu/amdgpu_gart.c (revision ec19d4d25b9b9f6a32643d67b87659e8001a02e2)
1 /*	$NetBSD: amdgpu_gart.c,v 1.12 2024/07/01 13:27:55 riastradh Exp $	*/
2 
3 /*
4  * Copyright 2008 Advanced Micro Devices, Inc.
5  * Copyright 2008 Red Hat Inc.
6  * Copyright 2009 Jerome Glisse.
7  *
8  * Permission is hereby granted, free of charge, to any person obtaining a
9  * copy of this software and associated documentation files (the "Software"),
10  * to deal in the Software without restriction, including without limitation
11  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12  * and/or sell copies of the Software, and to permit persons to whom the
13  * Software is furnished to do so, subject to the following conditions:
14  *
15  * The above copyright notice and this permission notice shall be included in
16  * all copies or substantial portions of the Software.
17  *
18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
21  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
22  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
23  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
24  * OTHER DEALINGS IN THE SOFTWARE.
25  *
26  * Authors: Dave Airlie
27  *          Alex Deucher
28  *          Jerome Glisse
29  */
30 
31 #include <sys/cdefs.h>
32 __KERNEL_RCSID(0, "$NetBSD: amdgpu_gart.c,v 1.12 2024/07/01 13:27:55 riastradh Exp $");
33 
34 #include <linux/pci.h>
35 #include <linux/vmalloc.h>
36 
37 #include <drm/amdgpu_drm.h>
38 #ifdef CONFIG_X86
39 #include <asm/set_memory.h>
40 #endif
41 #include "amdgpu.h"
42 
43 /*
44  * GART
45  * The GART (Graphics Aperture Remapping Table) is an aperture
46  * in the GPU's address space.  System pages can be mapped into
47  * the aperture and look like contiguous pages from the GPU's
48  * perspective.  A page table maps the pages in the aperture
49  * to the actual backing pages in system memory.
50  *
51  * Radeon GPUs support both an internal GART, as described above,
52  * and AGP.  AGP works similarly, but the GART table is configured
53  * and maintained by the northbridge rather than the driver.
54  * Radeon hw has a separate AGP aperture that is programmed to
55  * point to the AGP aperture provided by the northbridge and the
56  * requests are passed through to the northbridge aperture.
57  * Both AGP and internal GART can be used at the same time, however
58  * that is not currently supported by the driver.
59  *
60  * This file handles the common internal GART management.
61  */
62 
63 /*
64  * Common GART table functions.
65  */
66 
67 /**
68  * amdgpu_dummy_page_init - init dummy page used by the driver
69  *
70  * @adev: amdgpu_device pointer
71  *
72  * Allocate the dummy page used by the driver (all asics).
73  * This dummy page is used by the driver as a filler for gart entries
74  * when pages are taken out of the GART
75  * Returns 0 on sucess, -ENOMEM on failure.
76  */
amdgpu_gart_dummy_page_init(struct amdgpu_device * adev)77 static int amdgpu_gart_dummy_page_init(struct amdgpu_device *adev)
78 {
79 #ifdef __NetBSD__
80 	int rsegs;
81 	void *p;
82 	int error;
83 
84 	/* XXX Can this be called more than once??  */
85 	if (adev->dummy_page_map != NULL)
86 		return 0;
87 
88 	error = bus_dmamem_alloc(adev->ddev->dmat, PAGE_SIZE, PAGE_SIZE, 0,
89 	    &adev->dummy_page_seg, 1, &rsegs, BUS_DMA_WAITOK);
90 	if (error)
91 		goto fail0;
92 	KASSERT(rsegs == 1);
93 	error = bus_dmamem_map(adev->ddev->dmat, &adev->dummy_page_seg, 1,
94 	    PAGE_SIZE, &p, BUS_DMA_WAITOK);
95 	if (error)
96 		goto fail1;
97 	memset(p, 0, PAGE_SIZE);
98 	bus_dmamem_unmap(adev->ddev->dmat, p, PAGE_SIZE);
99 	error = bus_dmamap_create(adev->ddev->dmat, PAGE_SIZE, 1, PAGE_SIZE, 0,
100 	    BUS_DMA_WAITOK, &adev->dummy_page_map);
101 	if (error)
102 		goto fail1;
103 	error = bus_dmamap_load_raw(adev->ddev->dmat, adev->dummy_page_map,
104 	    &adev->dummy_page_seg, 1, PAGE_SIZE, BUS_DMA_WAITOK);
105 	if (error)
106 		goto fail2;
107 
108 	bus_dmamap_sync(adev->ddev->dmat, adev->dummy_page_map, 0, PAGE_SIZE,
109 	    BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
110 
111 	/* Success!  */
112 	adev->dummy_page_addr = adev->dummy_page_map->dm_segs[0].ds_addr;
113 	return 0;
114 
115 fail3: __unused
116 	bus_dmamap_unload(adev->ddev->dmat, adev->dummy_page_map);
117 fail2:	bus_dmamap_destroy(adev->ddev->dmat, adev->dummy_page_map);
118 fail1:	bus_dmamem_free(adev->ddev->dmat, &adev->dummy_page_seg, 1);
119 fail0:	KASSERT(error);
120 	adev->dummy_page_map = NULL;
121 	adev->dummy_page_addr = 0; /* paranoia */
122 	/* XXX errno NetBSD->Linux */
123 	return -error;
124 #else  /* __NetBSD__ */
125 	struct page *dummy_page = ttm_bo_glob.dummy_read_page;
126 
127 	if (adev->dummy_page_addr)
128 		return 0;
129 	adev->dummy_page_addr = pci_map_page(adev->pdev, dummy_page, 0,
130 					     PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
131 	if (pci_dma_mapping_error(adev->pdev, adev->dummy_page_addr)) {
132 		dev_err(&adev->pdev->dev, "Failed to DMA MAP the dummy page\n");
133 		adev->dummy_page_addr = 0;
134 		return -ENOMEM;
135 	}
136 	return 0;
137 #endif	/* __NetBSD__ */
138 }
139 
140 /**
141  * amdgpu_dummy_page_fini - free dummy page used by the driver
142  *
143  * @adev: amdgpu_device pointer
144  *
145  * Frees the dummy page used by the driver (all asics).
146  */
amdgpu_gart_dummy_page_fini(struct amdgpu_device * adev)147 static void amdgpu_gart_dummy_page_fini(struct amdgpu_device *adev)
148 {
149 	if (!adev->dummy_page_addr)
150 		return;
151 #ifdef __NetBSD__
152 	bus_dmamap_unload(adev->ddev->dmat, adev->dummy_page_map);
153 	bus_dmamap_destroy(adev->ddev->dmat, adev->dummy_page_map);
154 	bus_dmamem_free(adev->ddev->dmat, &adev->dummy_page_seg, 1);
155 	adev->dummy_page_map = NULL;
156 #else
157 	pci_unmap_page(adev->pdev, adev->dummy_page_addr,
158 		       PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
159 #endif
160 	adev->dummy_page_addr = 0;
161 }
162 
163 /**
164  * amdgpu_gart_table_vram_alloc - allocate vram for gart page table
165  *
166  * @adev: amdgpu_device pointer
167  *
168  * Allocate video memory for GART page table
169  * (pcie r4xx, r5xx+).  These asics require the
170  * gart table to be in video memory.
171  * Returns 0 for success, error for failure.
172  */
amdgpu_gart_table_vram_alloc(struct amdgpu_device * adev)173 int amdgpu_gart_table_vram_alloc(struct amdgpu_device *adev)
174 {
175 	int r;
176 
177 	if (adev->gart.bo == NULL) {
178 		struct amdgpu_bo_param bp;
179 
180 		memset(&bp, 0, sizeof(bp));
181 		bp.size = adev->gart.table_size;
182 		bp.byte_align = PAGE_SIZE;
183 		bp.domain = AMDGPU_GEM_DOMAIN_VRAM;
184 		bp.flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
185 			AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
186 		bp.type = ttm_bo_type_kernel;
187 		bp.resv = NULL;
188 		r = amdgpu_bo_create(adev, &bp, &adev->gart.bo);
189 		if (r) {
190 			return r;
191 		}
192 	}
193 	return 0;
194 }
195 
196 /**
197  * amdgpu_gart_table_vram_pin - pin gart page table in vram
198  *
199  * @adev: amdgpu_device pointer
200  *
201  * Pin the GART page table in vram so it will not be moved
202  * by the memory manager (pcie r4xx, r5xx+).  These asics require the
203  * gart table to be in video memory.
204  * Returns 0 for success, error for failure.
205  */
amdgpu_gart_table_vram_pin(struct amdgpu_device * adev)206 int amdgpu_gart_table_vram_pin(struct amdgpu_device *adev)
207 {
208 	int r;
209 
210 	r = amdgpu_bo_reserve(adev->gart.bo, false);
211 	if (unlikely(r != 0))
212 		return r;
213 	r = amdgpu_bo_pin(adev->gart.bo, AMDGPU_GEM_DOMAIN_VRAM);
214 	if (r) {
215 		amdgpu_bo_unreserve(adev->gart.bo);
216 		return r;
217 	}
218 	r = amdgpu_bo_kmap(adev->gart.bo, &adev->gart.ptr);
219 	if (r)
220 		amdgpu_bo_unpin(adev->gart.bo);
221 	amdgpu_bo_unreserve(adev->gart.bo);
222 	return r;
223 }
224 
225 /**
226  * amdgpu_gart_table_vram_unpin - unpin gart page table in vram
227  *
228  * @adev: amdgpu_device pointer
229  *
230  * Unpin the GART page table in vram (pcie r4xx, r5xx+).
231  * These asics require the gart table to be in video memory.
232  */
amdgpu_gart_table_vram_unpin(struct amdgpu_device * adev)233 void amdgpu_gart_table_vram_unpin(struct amdgpu_device *adev)
234 {
235 	int r;
236 
237 	if (adev->gart.bo == NULL) {
238 		return;
239 	}
240 	r = amdgpu_bo_reserve(adev->gart.bo, true);
241 	if (likely(r == 0)) {
242 		amdgpu_bo_kunmap(adev->gart.bo);
243 		amdgpu_bo_unpin(adev->gart.bo);
244 		amdgpu_bo_unreserve(adev->gart.bo);
245 		adev->gart.ptr = NULL;
246 	}
247 }
248 
249 /**
250  * amdgpu_gart_table_vram_free - free gart page table vram
251  *
252  * @adev: amdgpu_device pointer
253  *
254  * Free the video memory used for the GART page table
255  * (pcie r4xx, r5xx+).  These asics require the gart table to
256  * be in video memory.
257  */
amdgpu_gart_table_vram_free(struct amdgpu_device * adev)258 void amdgpu_gart_table_vram_free(struct amdgpu_device *adev)
259 {
260 	if (adev->gart.bo == NULL) {
261 		return;
262 	}
263 	amdgpu_bo_unref(&adev->gart.bo);
264 }
265 
266 #ifdef __NetBSD__
267 static void
amdgpu_gart_pre_update(struct amdgpu_device * adev,unsigned gpu_pgstart,unsigned gpu_npages)268 amdgpu_gart_pre_update(struct amdgpu_device *adev, unsigned gpu_pgstart,
269     unsigned gpu_npages)
270 {
271 
272 	if (adev->gart.ag_table_map != NULL) {
273 		const unsigned entsize =
274 		    adev->gart.table_size / adev->gart.num_gpu_pages;
275 
276 		bus_dmamap_sync(adev->ddev->dmat, adev->gart.ag_table_map,
277 		    gpu_pgstart*entsize, gpu_npages*entsize,
278 		    BUS_DMASYNC_POSTWRITE);
279 	}
280 }
281 
282 static void
amdgpu_gart_post_update(struct amdgpu_device * adev,unsigned gpu_pgstart,unsigned gpu_npages)283 amdgpu_gart_post_update(struct amdgpu_device *adev, unsigned gpu_pgstart,
284     unsigned gpu_npages)
285 {
286 	unsigned i;
287 
288 	if (adev->gart.ag_table_map != NULL) {
289 		const unsigned entsize =
290 		    adev->gart.table_size / adev->gart.num_gpu_pages;
291 
292 		bus_dmamap_sync(adev->ddev->dmat, adev->gart.ag_table_map,
293 		    gpu_pgstart*entsize, gpu_npages*entsize,
294 		    BUS_DMASYNC_PREWRITE);
295 	}
296 	mb();			/* XXX why is bus_dmamap_sync not enough? */
297 	amdgpu_asic_flush_hdp(adev, NULL);
298 	for (i = 0; i < adev->num_vmhubs; i++)
299 		amdgpu_gmc_flush_gpu_tlb(adev, 0, i, 0);
300 }
301 #endif
302 
303 /*
304  * Common gart functions.
305  */
306 #ifdef __NetBSD__
307 int
amdgpu_gart_unbind(struct amdgpu_device * adev,uint64_t gpu_start,unsigned npages)308 amdgpu_gart_unbind(struct amdgpu_device *adev, uint64_t gpu_start,
309     unsigned npages)
310 {
311 	const unsigned gpu_per_cpu = AMDGPU_GPU_PAGES_IN_CPU_PAGE;
312 	const unsigned gpu_npages = (npages * gpu_per_cpu);
313 	const uint64_t gpu_pgstart = (gpu_start / AMDGPU_GPU_PAGE_SIZE);
314 	const uint64_t pgstart __diagused = (gpu_pgstart / gpu_per_cpu);
315 	uint64_t pgno, gpu_pgno;
316 	uint32_t flags = AMDGPU_PTE_SYSTEM;
317 
318 	KASSERT(pgstart == (gpu_start / PAGE_SIZE));
319 	KASSERT(npages <= adev->gart.num_cpu_pages);
320 	KASSERT(gpu_npages <= adev->gart.num_cpu_pages);
321 
322 	if (!adev->gart.ready) {
323 		WARN(1, "trying to bind memory to uninitialized GART !\n");
324 		return -EINVAL;
325 	}
326 
327 	amdgpu_gart_pre_update(adev, gpu_pgstart, gpu_npages);
328 	for (pgno = 0; pgno < npages; pgno++) {
329 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
330 		adev->gart.pages[pgstart + pgno] = NULL;
331 #endif
332 
333 		if (adev->gart.ptr == NULL)
334 			continue;
335 		for (gpu_pgno = 0; gpu_pgno < gpu_per_cpu; gpu_pgno++) {
336 			amdgpu_gmc_set_pte_pde(adev, adev->gart.ptr,
337 			    gpu_pgstart + gpu_per_cpu*pgno + gpu_pgno,
338 			    adev->dummy_page_addr, flags);
339 		}
340 	}
341 	amdgpu_gart_post_update(adev, gpu_pgstart, gpu_npages);
342 
343 	return 0;
344 }
345 #else  /* __NetBSD__ */
346 /**
347  * amdgpu_gart_unbind - unbind pages from the gart page table
348  *
349  * @adev: amdgpu_device pointer
350  * @offset: offset into the GPU's gart aperture
351  * @pages: number of pages to unbind
352  *
353  * Unbinds the requested pages from the gart page table and
354  * replaces them with the dummy page (all asics).
355  * Returns 0 for success, -EINVAL for failure.
356  */
amdgpu_gart_unbind(struct amdgpu_device * adev,uint64_t offset,int pages)357 int amdgpu_gart_unbind(struct amdgpu_device *adev, uint64_t offset,
358 			int pages)
359 {
360 	unsigned t;
361 	unsigned p;
362 	int i, j;
363 	u64 page_base;
364 	/* Starting from VEGA10, system bit must be 0 to mean invalid. */
365 	uint64_t flags = 0;
366 
367 	if (!adev->gart.ready) {
368 		WARN(1, "trying to unbind memory from uninitialized GART !\n");
369 		return -EINVAL;
370 	}
371 
372 	t = offset / AMDGPU_GPU_PAGE_SIZE;
373 	p = t / AMDGPU_GPU_PAGES_IN_CPU_PAGE;
374 	for (i = 0; i < pages; i++, p++) {
375 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
376 		adev->gart.pages[p] = NULL;
377 #endif
378 		page_base = adev->dummy_page_addr;
379 		if (!adev->gart.ptr)
380 			continue;
381 
382 		for (j = 0; j < AMDGPU_GPU_PAGES_IN_CPU_PAGE; j++, t++) {
383 			amdgpu_gmc_set_pte_pde(adev, adev->gart.ptr,
384 					       t, page_base, flags);
385 			page_base += AMDGPU_GPU_PAGE_SIZE;
386 		}
387 	}
388 	mb();
389 	amdgpu_asic_flush_hdp(adev, NULL);
390 	for (i = 0; i < adev->num_vmhubs; i++)
391 		amdgpu_gmc_flush_gpu_tlb(adev, 0, i, 0);
392 
393 	return 0;
394 }
395 #endif	/* __NetBSD__ */
396 
397 /**
398  * amdgpu_gart_map - map dma_addresses into GART entries
399  *
400  * @adev: amdgpu_device pointer
401  * @offset: offset into the GPU's gart aperture
402  * @pages: number of pages to bind
403  * @dma_addr: DMA addresses of pages
404  * @flags: page table entry flags
405  * @dst: CPU address of the gart table
406  *
407  * Map the dma_addresses into GART entries (all asics).
408  * Returns 0 for success, -EINVAL for failure.
409  */
410 #ifdef __NetBSD__
amdgpu_gart_map(struct amdgpu_device * adev,uint64_t gpu_start,unsigned npages,bus_size_t map_start,bus_dmamap_t dmamap,uint32_t flags,void * dst)411 int amdgpu_gart_map(struct amdgpu_device *adev, uint64_t gpu_start,
412     unsigned npages, bus_size_t map_start, bus_dmamap_t dmamap, uint32_t flags,
413     void *dst)
414 {
415 	bus_size_t seg_off = 0;
416 	unsigned i, j, t;
417 
418 	CTASSERT(AMDGPU_GPU_PAGE_SIZE <= PAGE_SIZE);
419 	CTASSERT((PAGE_SIZE % AMDGPU_GPU_PAGE_SIZE) == 0);
420 
421 	KASSERT((gpu_start & (PAGE_SIZE - 1)) == 0);
422 
423 	if (!adev->gart.ready) {
424 		WARN(1, "trying to bind memory to uninitialized GART !\n");
425 		return -EINVAL;
426 	}
427 
428 	for (i = 0; i < dmamap->dm_nsegs; i++) {
429 		KASSERT((dmamap->dm_segs[i].ds_len & (PAGE_SIZE - 1)) == 0);
430 		if (map_start == 0)
431 			break;
432 		if (map_start < dmamap->dm_segs[i].ds_len) {
433 			seg_off = map_start;
434 			break;
435 		}
436 		map_start -= dmamap->dm_segs[i].ds_len;
437 	}
438 	KASSERT(i < dmamap->dm_nsegs);
439 
440 	t = gpu_start / AMDGPU_GPU_PAGE_SIZE;
441 
442 	for (i = 0; npages --> 0;) {
443 		for (j = 0; j < AMDGPU_GPU_PAGES_IN_CPU_PAGE; j++, t++) {
444 			KASSERT(i < dmamap->dm_nsegs);
445 			KASSERT(seg_off < dmamap->dm_segs[i].ds_len);
446 			amdgpu_gmc_set_pte_pde(adev, dst, t,
447 			    dmamap->dm_segs[i].ds_addr + seg_off, flags);
448 			seg_off += AMDGPU_GPU_PAGE_SIZE;
449 			if (seg_off == dmamap->dm_segs[i].ds_len) {
450 				i++;
451 				seg_off = 0;
452 			}
453 		}
454 	}
455 
456 	return 0;
457 }
458 #else
amdgpu_gart_map(struct amdgpu_device * adev,uint64_t offset,int pages,dma_addr_t * dma_addr,uint64_t flags,void * dst)459 int amdgpu_gart_map(struct amdgpu_device *adev, uint64_t offset,
460 		    int pages, dma_addr_t *dma_addr, uint64_t flags,
461 		    void *dst)
462 {
463 	uint64_t page_base;
464 	unsigned i, j, t;
465 
466 	if (!adev->gart.ready) {
467 		WARN(1, "trying to bind memory to uninitialized GART !\n");
468 		return -EINVAL;
469 	}
470 
471 	t = offset / AMDGPU_GPU_PAGE_SIZE;
472 
473 	for (i = 0; i < pages; i++) {
474 		page_base = dma_addr[i];
475 		for (j = 0; j < AMDGPU_GPU_PAGES_IN_CPU_PAGE; j++, t++) {
476 			amdgpu_gmc_set_pte_pde(adev, dst, t, page_base, flags);
477 			page_base += AMDGPU_GPU_PAGE_SIZE;
478 		}
479 	}
480 	return 0;
481 }
482 #endif	/* __NetBSD__ */
483 
484 #ifdef __NetBSD__
485 int
amdgpu_gart_bind(struct amdgpu_device * adev,uint64_t gpu_start,unsigned npages,struct page ** pages,bus_dmamap_t dmamap,uint32_t flags)486 amdgpu_gart_bind(struct amdgpu_device *adev, uint64_t gpu_start,
487     unsigned npages, struct page **pages, bus_dmamap_t dmamap, uint32_t flags)
488 {
489 	const unsigned gpu_per_cpu = AMDGPU_GPU_PAGES_IN_CPU_PAGE;
490 	const unsigned gpu_npages = (npages * gpu_per_cpu);
491 	const uint64_t gpu_pgstart = (gpu_start / AMDGPU_GPU_PAGE_SIZE);
492 	const uint64_t pgstart __diagused = (gpu_pgstart / gpu_per_cpu);
493 	uint64_t pgno, gpu_pgno;
494 
495 	KASSERT(pgstart == (gpu_start / PAGE_SIZE));
496 	KASSERT(npages == dmamap->dm_nsegs);
497 	KASSERT(npages <= adev->gart.num_cpu_pages);
498 	KASSERT(gpu_npages <= adev->gart.num_cpu_pages);
499 
500 	if (!adev->gart.ready) {
501 		WARN(1, "trying to bind memory to uninitialized GART !\n");
502 		return -EINVAL;
503 	}
504 
505 	amdgpu_gart_pre_update(adev, gpu_pgstart, gpu_npages);
506 	for (pgno = 0; pgno < npages; pgno++) {
507 		const bus_addr_t addr = dmamap->dm_segs[pgno].ds_addr;
508 
509 		KASSERT(dmamap->dm_segs[pgno].ds_len == PAGE_SIZE);
510 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
511 		adev->gart.pages[pgstart + pgno] = NULL;
512 #endif
513 
514 		if (adev->gart.ptr == NULL)
515 			continue;
516 
517 		for (gpu_pgno = 0; gpu_pgno < gpu_per_cpu; gpu_pgno++) {
518 			amdgpu_gmc_set_pte_pde(adev, adev->gart.ptr,
519 			    gpu_pgstart + gpu_per_cpu*pgno + gpu_pgno,
520 			    addr + gpu_pgno*AMDGPU_GPU_PAGE_SIZE, flags);
521 		}
522 	}
523 	amdgpu_gart_post_update(adev, gpu_pgstart, gpu_npages);
524 
525 	return 0;
526 }
527 #else  /* __NetBSD__ */
528 /**
529  * amdgpu_gart_bind - bind pages into the gart page table
530  *
531  * @adev: amdgpu_device pointer
532  * @offset: offset into the GPU's gart aperture
533  * @pages: number of pages to bind
534  * @pagelist: pages to bind
535  * @dma_addr: DMA addresses of pages
536  * @flags: page table entry flags
537  *
538  * Binds the requested pages to the gart page table
539  * (all asics).
540  * Returns 0 for success, -EINVAL for failure.
541  */
amdgpu_gart_bind(struct amdgpu_device * adev,uint64_t offset,int pages,struct page ** pagelist,dma_addr_t * dma_addr,uint64_t flags)542 int amdgpu_gart_bind(struct amdgpu_device *adev, uint64_t offset,
543 		     int pages, struct page **pagelist, dma_addr_t *dma_addr,
544 		     uint64_t flags)
545 {
546 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
547 	unsigned t,p;
548 #endif
549 	int r, i;
550 
551 	if (!adev->gart.ready) {
552 		WARN(1, "trying to bind memory to uninitialized GART !\n");
553 		return -EINVAL;
554 	}
555 
556 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
557 	t = offset / AMDGPU_GPU_PAGE_SIZE;
558 	p = t / AMDGPU_GPU_PAGES_IN_CPU_PAGE;
559 	for (i = 0; i < pages; i++, p++)
560 		adev->gart.pages[p] = pagelist ? pagelist[i] : NULL;
561 #endif
562 
563 	if (!adev->gart.ptr)
564 		return 0;
565 
566 	r = amdgpu_gart_map(adev, offset, pages, dma_addr, flags,
567 		    adev->gart.ptr);
568 	if (r)
569 		return r;
570 
571 	mb();
572 	amdgpu_asic_flush_hdp(adev, NULL);
573 	for (i = 0; i < adev->num_vmhubs; i++)
574 		amdgpu_gmc_flush_gpu_tlb(adev, 0, i, 0);
575 	return 0;
576 }
577 #endif
578 
579 /**
580  * amdgpu_gart_init - init the driver info for managing the gart
581  *
582  * @adev: amdgpu_device pointer
583  *
584  * Allocate the dummy page and init the gart driver info (all asics).
585  * Returns 0 for success, error for failure.
586  */
amdgpu_gart_init(struct amdgpu_device * adev)587 int amdgpu_gart_init(struct amdgpu_device *adev)
588 {
589 	int r;
590 
591 	if (adev->dummy_page_addr)
592 		return 0;
593 
594 	/* We need PAGE_SIZE >= AMDGPU_GPU_PAGE_SIZE */
595 	if (PAGE_SIZE < AMDGPU_GPU_PAGE_SIZE) {
596 		DRM_ERROR("Page size is smaller than GPU page size!\n");
597 		return -EINVAL;
598 	}
599 	r = amdgpu_gart_dummy_page_init(adev);
600 	if (r)
601 		return r;
602 	/* Compute table size */
603 	adev->gart.num_cpu_pages = adev->gmc.gart_size / PAGE_SIZE;
604 	adev->gart.num_gpu_pages = adev->gmc.gart_size / AMDGPU_GPU_PAGE_SIZE;
605 	DRM_INFO("GART: num cpu pages %u, num gpu pages %u\n",
606 		 adev->gart.num_cpu_pages, adev->gart.num_gpu_pages);
607 
608 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
609 	/* Allocate pages table */
610 	adev->gart.pages = vzalloc(array_size(sizeof(void *),
611 					      adev->gart.num_cpu_pages));
612 	if (adev->gart.pages == NULL)
613 		return -ENOMEM;
614 #endif
615 
616 	return 0;
617 }
618 
619 /**
620  * amdgpu_gart_fini - tear down the driver info for managing the gart
621  *
622  * @adev: amdgpu_device pointer
623  *
624  * Tear down the gart driver info and free the dummy page (all asics).
625  */
amdgpu_gart_fini(struct amdgpu_device * adev)626 void amdgpu_gart_fini(struct amdgpu_device *adev)
627 {
628 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
629 	vfree(adev->gart.pages);
630 	adev->gart.pages = NULL;
631 #endif
632 	amdgpu_gart_dummy_page_fini(adev);
633 }
634