xref: /linux/drivers/gpu/drm/etnaviv/etnaviv_mmu.c (revision 4b660dbd9ee2059850fd30e0df420ca7a38a1856)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2015-2018 Etnaviv Project
4  */
5 
6 #include <linux/dma-mapping.h>
7 #include <linux/scatterlist.h>
8 
9 #include "common.xml.h"
10 #include "etnaviv_cmdbuf.h"
11 #include "etnaviv_drv.h"
12 #include "etnaviv_gem.h"
13 #include "etnaviv_gpu.h"
14 #include "etnaviv_mmu.h"
15 
16 static void etnaviv_context_unmap(struct etnaviv_iommu_context *context,
17 				 unsigned long iova, size_t size)
18 {
19 	size_t unmapped_page, unmapped = 0;
20 	size_t pgsize = SZ_4K;
21 
22 	if (!IS_ALIGNED(iova | size, pgsize)) {
23 		pr_err("unaligned: iova 0x%lx size 0x%zx min_pagesz 0x%zx\n",
24 		       iova, size, pgsize);
25 		return;
26 	}
27 
28 	while (unmapped < size) {
29 		unmapped_page = context->global->ops->unmap(context, iova,
30 							    pgsize);
31 		if (!unmapped_page)
32 			break;
33 
34 		iova += unmapped_page;
35 		unmapped += unmapped_page;
36 	}
37 }
38 
39 static int etnaviv_context_map(struct etnaviv_iommu_context *context,
40 			      unsigned long iova, phys_addr_t paddr,
41 			      size_t size, int prot)
42 {
43 	unsigned long orig_iova = iova;
44 	size_t pgsize = SZ_4K;
45 	size_t orig_size = size;
46 	int ret = 0;
47 
48 	if (!IS_ALIGNED(iova | paddr | size, pgsize)) {
49 		pr_err("unaligned: iova 0x%lx pa %pa size 0x%zx min_pagesz 0x%zx\n",
50 		       iova, &paddr, size, pgsize);
51 		return -EINVAL;
52 	}
53 
54 	while (size) {
55 		ret = context->global->ops->map(context, iova, paddr, pgsize,
56 						prot);
57 		if (ret)
58 			break;
59 
60 		iova += pgsize;
61 		paddr += pgsize;
62 		size -= pgsize;
63 	}
64 
65 	/* unroll mapping in case something went wrong */
66 	if (ret)
67 		etnaviv_context_unmap(context, orig_iova, orig_size - size);
68 
69 	return ret;
70 }
71 
72 static int etnaviv_iommu_map(struct etnaviv_iommu_context *context, u32 iova,
73 			     struct sg_table *sgt, int prot)
74 {	struct scatterlist *sg;
75 	unsigned int da = iova;
76 	unsigned int i;
77 	int ret;
78 
79 	if (!context || !sgt)
80 		return -EINVAL;
81 
82 	for_each_sgtable_dma_sg(sgt, sg, i) {
83 		phys_addr_t pa = sg_dma_address(sg) - sg->offset;
84 		size_t bytes = sg_dma_len(sg) + sg->offset;
85 
86 		VERB("map[%d]: %08x %pap(%zx)", i, iova, &pa, bytes);
87 
88 		ret = etnaviv_context_map(context, da, pa, bytes, prot);
89 		if (ret)
90 			goto fail;
91 
92 		da += bytes;
93 	}
94 
95 	context->flush_seq++;
96 
97 	return 0;
98 
99 fail:
100 	etnaviv_context_unmap(context, iova, da - iova);
101 	return ret;
102 }
103 
104 static void etnaviv_iommu_unmap(struct etnaviv_iommu_context *context, u32 iova,
105 				struct sg_table *sgt, unsigned len)
106 {
107 	struct scatterlist *sg;
108 	unsigned int da = iova;
109 	int i;
110 
111 	for_each_sgtable_dma_sg(sgt, sg, i) {
112 		size_t bytes = sg_dma_len(sg) + sg->offset;
113 
114 		etnaviv_context_unmap(context, da, bytes);
115 
116 		VERB("unmap[%d]: %08x(%zx)", i, iova, bytes);
117 
118 		BUG_ON(!PAGE_ALIGNED(bytes));
119 
120 		da += bytes;
121 	}
122 
123 	context->flush_seq++;
124 }
125 
126 static void etnaviv_iommu_remove_mapping(struct etnaviv_iommu_context *context,
127 	struct etnaviv_vram_mapping *mapping)
128 {
129 	struct etnaviv_gem_object *etnaviv_obj = mapping->object;
130 
131 	lockdep_assert_held(&context->lock);
132 
133 	etnaviv_iommu_unmap(context, mapping->vram_node.start,
134 			    etnaviv_obj->sgt, etnaviv_obj->base.size);
135 	drm_mm_remove_node(&mapping->vram_node);
136 }
137 
138 void etnaviv_iommu_reap_mapping(struct etnaviv_vram_mapping *mapping)
139 {
140 	struct etnaviv_iommu_context *context = mapping->context;
141 
142 	lockdep_assert_held(&context->lock);
143 	WARN_ON(mapping->use);
144 
145 	etnaviv_iommu_remove_mapping(context, mapping);
146 	etnaviv_iommu_context_put(mapping->context);
147 	mapping->context = NULL;
148 	list_del_init(&mapping->mmu_node);
149 }
150 
151 static int etnaviv_iommu_find_iova(struct etnaviv_iommu_context *context,
152 				   struct drm_mm_node *node, size_t size)
153 {
154 	struct etnaviv_vram_mapping *free = NULL;
155 	enum drm_mm_insert_mode mode = DRM_MM_INSERT_LOW;
156 	int ret;
157 
158 	lockdep_assert_held(&context->lock);
159 
160 	while (1) {
161 		struct etnaviv_vram_mapping *m, *n;
162 		struct drm_mm_scan scan;
163 		struct list_head list;
164 		bool found;
165 
166 		ret = drm_mm_insert_node_in_range(&context->mm, node,
167 						  size, 0, 0, 0, U64_MAX, mode);
168 		if (ret != -ENOSPC)
169 			break;
170 
171 		/* Try to retire some entries */
172 		drm_mm_scan_init(&scan, &context->mm, size, 0, 0, mode);
173 
174 		found = 0;
175 		INIT_LIST_HEAD(&list);
176 		list_for_each_entry(free, &context->mappings, mmu_node) {
177 			/* If this vram node has not been used, skip this. */
178 			if (!free->vram_node.mm)
179 				continue;
180 
181 			/*
182 			 * If the iova is pinned, then it's in-use,
183 			 * so we must keep its mapping.
184 			 */
185 			if (free->use)
186 				continue;
187 
188 			list_add(&free->scan_node, &list);
189 			if (drm_mm_scan_add_block(&scan, &free->vram_node)) {
190 				found = true;
191 				break;
192 			}
193 		}
194 
195 		if (!found) {
196 			/* Nothing found, clean up and fail */
197 			list_for_each_entry_safe(m, n, &list, scan_node)
198 				BUG_ON(drm_mm_scan_remove_block(&scan, &m->vram_node));
199 			break;
200 		}
201 
202 		/*
203 		 * drm_mm does not allow any other operations while
204 		 * scanning, so we have to remove all blocks first.
205 		 * If drm_mm_scan_remove_block() returns false, we
206 		 * can leave the block pinned.
207 		 */
208 		list_for_each_entry_safe(m, n, &list, scan_node)
209 			if (!drm_mm_scan_remove_block(&scan, &m->vram_node))
210 				list_del_init(&m->scan_node);
211 
212 		/*
213 		 * Unmap the blocks which need to be reaped from the MMU.
214 		 * Clear the mmu pointer to prevent the mapping_get finding
215 		 * this mapping.
216 		 */
217 		list_for_each_entry_safe(m, n, &list, scan_node) {
218 			etnaviv_iommu_reap_mapping(m);
219 			list_del_init(&m->scan_node);
220 		}
221 
222 		mode = DRM_MM_INSERT_EVICT;
223 
224 		/*
225 		 * We removed enough mappings so that the new allocation will
226 		 * succeed, retry the allocation one more time.
227 		 */
228 	}
229 
230 	return ret;
231 }
232 
233 static int etnaviv_iommu_insert_exact(struct etnaviv_iommu_context *context,
234 		   struct drm_mm_node *node, size_t size, u64 va)
235 {
236 	struct etnaviv_vram_mapping *m, *n;
237 	struct drm_mm_node *scan_node;
238 	LIST_HEAD(scan_list);
239 	int ret;
240 
241 	lockdep_assert_held(&context->lock);
242 
243 	ret = drm_mm_insert_node_in_range(&context->mm, node, size, 0, 0, va,
244 					  va + size, DRM_MM_INSERT_LOWEST);
245 	if (ret != -ENOSPC)
246 		return ret;
247 
248 	/*
249 	 * When we can't insert the node, due to a existing mapping blocking
250 	 * the address space, there are two possible reasons:
251 	 * 1. Userspace genuinely messed up and tried to reuse address space
252 	 * before the last job using this VMA has finished executing.
253 	 * 2. The existing buffer mappings are idle, but the buffers are not
254 	 * destroyed yet (likely due to being referenced by another context) in
255 	 * which case the mappings will not be cleaned up and we must reap them
256 	 * here to make space for the new mapping.
257 	 */
258 
259 	drm_mm_for_each_node_in_range(scan_node, &context->mm, va, va + size) {
260 		m = container_of(scan_node, struct etnaviv_vram_mapping,
261 				 vram_node);
262 
263 		if (m->use)
264 			return -ENOSPC;
265 
266 		list_add(&m->scan_node, &scan_list);
267 	}
268 
269 	list_for_each_entry_safe(m, n, &scan_list, scan_node) {
270 		etnaviv_iommu_reap_mapping(m);
271 		list_del_init(&m->scan_node);
272 	}
273 
274 	return drm_mm_insert_node_in_range(&context->mm, node, size, 0, 0, va,
275 					   va + size, DRM_MM_INSERT_LOWEST);
276 }
277 
278 int etnaviv_iommu_map_gem(struct etnaviv_iommu_context *context,
279 	struct etnaviv_gem_object *etnaviv_obj, u32 memory_base,
280 	struct etnaviv_vram_mapping *mapping, u64 va)
281 {
282 	struct sg_table *sgt = etnaviv_obj->sgt;
283 	struct drm_mm_node *node;
284 	int ret;
285 
286 	lockdep_assert_held(&etnaviv_obj->lock);
287 
288 	mutex_lock(&context->lock);
289 
290 	/* v1 MMU can optimize single entry (contiguous) scatterlists */
291 	if (context->global->version == ETNAVIV_IOMMU_V1 &&
292 	    sgt->nents == 1 && !(etnaviv_obj->flags & ETNA_BO_FORCE_MMU)) {
293 		u32 iova;
294 
295 		iova = sg_dma_address(sgt->sgl) - memory_base;
296 		if (iova < 0x80000000 - sg_dma_len(sgt->sgl)) {
297 			mapping->iova = iova;
298 			mapping->context = etnaviv_iommu_context_get(context);
299 			list_add_tail(&mapping->mmu_node, &context->mappings);
300 			ret = 0;
301 			goto unlock;
302 		}
303 	}
304 
305 	node = &mapping->vram_node;
306 
307 	if (va)
308 		ret = etnaviv_iommu_insert_exact(context, node,
309 						 etnaviv_obj->base.size, va);
310 	else
311 		ret = etnaviv_iommu_find_iova(context, node,
312 					      etnaviv_obj->base.size);
313 	if (ret < 0)
314 		goto unlock;
315 
316 	mapping->iova = node->start;
317 	ret = etnaviv_iommu_map(context, node->start, sgt,
318 				ETNAVIV_PROT_READ | ETNAVIV_PROT_WRITE);
319 
320 	if (ret < 0) {
321 		drm_mm_remove_node(node);
322 		goto unlock;
323 	}
324 
325 	mapping->context = etnaviv_iommu_context_get(context);
326 	list_add_tail(&mapping->mmu_node, &context->mappings);
327 unlock:
328 	mutex_unlock(&context->lock);
329 
330 	return ret;
331 }
332 
333 void etnaviv_iommu_unmap_gem(struct etnaviv_iommu_context *context,
334 	struct etnaviv_vram_mapping *mapping)
335 {
336 	WARN_ON(mapping->use);
337 
338 	mutex_lock(&context->lock);
339 
340 	/* Bail if the mapping has been reaped by another thread */
341 	if (!mapping->context) {
342 		mutex_unlock(&context->lock);
343 		return;
344 	}
345 
346 	/* If the vram node is on the mm, unmap and remove the node */
347 	if (mapping->vram_node.mm == &context->mm)
348 		etnaviv_iommu_remove_mapping(context, mapping);
349 
350 	list_del(&mapping->mmu_node);
351 	mutex_unlock(&context->lock);
352 	etnaviv_iommu_context_put(context);
353 }
354 
355 static void etnaviv_iommu_context_free(struct kref *kref)
356 {
357 	struct etnaviv_iommu_context *context =
358 		container_of(kref, struct etnaviv_iommu_context, refcount);
359 
360 	etnaviv_cmdbuf_suballoc_unmap(context, &context->cmdbuf_mapping);
361 
362 	context->global->ops->free(context);
363 }
364 void etnaviv_iommu_context_put(struct etnaviv_iommu_context *context)
365 {
366 	kref_put(&context->refcount, etnaviv_iommu_context_free);
367 }
368 
369 struct etnaviv_iommu_context *
370 etnaviv_iommu_context_init(struct etnaviv_iommu_global *global,
371 			   struct etnaviv_cmdbuf_suballoc *suballoc)
372 {
373 	struct etnaviv_iommu_context *ctx;
374 	int ret;
375 
376 	if (global->version == ETNAVIV_IOMMU_V1)
377 		ctx = etnaviv_iommuv1_context_alloc(global);
378 	else
379 		ctx = etnaviv_iommuv2_context_alloc(global);
380 
381 	if (!ctx)
382 		return NULL;
383 
384 	ret = etnaviv_cmdbuf_suballoc_map(suballoc, ctx, &ctx->cmdbuf_mapping,
385 					  global->memory_base);
386 	if (ret)
387 		goto out_free;
388 
389 	if (global->version == ETNAVIV_IOMMU_V1 &&
390 	    ctx->cmdbuf_mapping.iova > 0x80000000) {
391 		dev_err(global->dev,
392 		        "command buffer outside valid memory window\n");
393 		goto out_unmap;
394 	}
395 
396 	return ctx;
397 
398 out_unmap:
399 	etnaviv_cmdbuf_suballoc_unmap(ctx, &ctx->cmdbuf_mapping);
400 out_free:
401 	global->ops->free(ctx);
402 	return NULL;
403 }
404 
405 void etnaviv_iommu_restore(struct etnaviv_gpu *gpu,
406 			   struct etnaviv_iommu_context *context)
407 {
408 	context->global->ops->restore(gpu, context);
409 }
410 
411 int etnaviv_iommu_get_suballoc_va(struct etnaviv_iommu_context *context,
412 				  struct etnaviv_vram_mapping *mapping,
413 				  u32 memory_base, dma_addr_t paddr,
414 				  size_t size)
415 {
416 	mutex_lock(&context->lock);
417 
418 	if (mapping->use > 0) {
419 		mapping->use++;
420 		mutex_unlock(&context->lock);
421 		return 0;
422 	}
423 
424 	/*
425 	 * For MMUv1 we don't add the suballoc region to the pagetables, as
426 	 * those GPUs can only work with cmdbufs accessed through the linear
427 	 * window. Instead we manufacture a mapping to make it look uniform
428 	 * to the upper layers.
429 	 */
430 	if (context->global->version == ETNAVIV_IOMMU_V1) {
431 		mapping->iova = paddr - memory_base;
432 	} else {
433 		struct drm_mm_node *node = &mapping->vram_node;
434 		int ret;
435 
436 		ret = etnaviv_iommu_find_iova(context, node, size);
437 		if (ret < 0) {
438 			mutex_unlock(&context->lock);
439 			return ret;
440 		}
441 
442 		mapping->iova = node->start;
443 		ret = etnaviv_context_map(context, node->start, paddr, size,
444 					  ETNAVIV_PROT_READ);
445 		if (ret < 0) {
446 			drm_mm_remove_node(node);
447 			mutex_unlock(&context->lock);
448 			return ret;
449 		}
450 
451 		context->flush_seq++;
452 	}
453 
454 	list_add_tail(&mapping->mmu_node, &context->mappings);
455 	mapping->use = 1;
456 
457 	mutex_unlock(&context->lock);
458 
459 	return 0;
460 }
461 
462 void etnaviv_iommu_put_suballoc_va(struct etnaviv_iommu_context *context,
463 		  struct etnaviv_vram_mapping *mapping)
464 {
465 	struct drm_mm_node *node = &mapping->vram_node;
466 
467 	mutex_lock(&context->lock);
468 	mapping->use--;
469 
470 	if (mapping->use > 0 || context->global->version == ETNAVIV_IOMMU_V1) {
471 		mutex_unlock(&context->lock);
472 		return;
473 	}
474 
475 	etnaviv_context_unmap(context, node->start, node->size);
476 	drm_mm_remove_node(node);
477 	mutex_unlock(&context->lock);
478 }
479 
480 size_t etnaviv_iommu_dump_size(struct etnaviv_iommu_context *context)
481 {
482 	return context->global->ops->dump_size(context);
483 }
484 
485 void etnaviv_iommu_dump(struct etnaviv_iommu_context *context, void *buf)
486 {
487 	context->global->ops->dump(context, buf);
488 }
489 
490 int etnaviv_iommu_global_init(struct etnaviv_gpu *gpu)
491 {
492 	enum etnaviv_iommu_version version = ETNAVIV_IOMMU_V1;
493 	struct etnaviv_drm_private *priv = gpu->drm->dev_private;
494 	struct etnaviv_iommu_global *global;
495 	struct device *dev = gpu->drm->dev;
496 
497 	if (gpu->identity.minor_features1 & chipMinorFeatures1_MMU_VERSION)
498 		version = ETNAVIV_IOMMU_V2;
499 
500 	if (priv->mmu_global) {
501 		if (priv->mmu_global->version != version) {
502 			dev_err(gpu->dev,
503 				"MMU version doesn't match global version\n");
504 			return -ENXIO;
505 		}
506 
507 		priv->mmu_global->use++;
508 		return 0;
509 	}
510 
511 	global = kzalloc(sizeof(*global), GFP_KERNEL);
512 	if (!global)
513 		return -ENOMEM;
514 
515 	global->bad_page_cpu = dma_alloc_wc(dev, SZ_4K, &global->bad_page_dma,
516 					    GFP_KERNEL);
517 	if (!global->bad_page_cpu)
518 		goto free_global;
519 
520 	memset32(global->bad_page_cpu, 0xdead55aa, SZ_4K / sizeof(u32));
521 
522 	if (version == ETNAVIV_IOMMU_V2) {
523 		global->v2.pta_cpu = dma_alloc_wc(dev, ETNAVIV_PTA_SIZE,
524 					       &global->v2.pta_dma, GFP_KERNEL);
525 		if (!global->v2.pta_cpu)
526 			goto free_bad_page;
527 	}
528 
529 	global->dev = dev;
530 	global->version = version;
531 	global->use = 1;
532 	mutex_init(&global->lock);
533 
534 	if (version == ETNAVIV_IOMMU_V1)
535 		global->ops = &etnaviv_iommuv1_ops;
536 	else
537 		global->ops = &etnaviv_iommuv2_ops;
538 
539 	priv->mmu_global = global;
540 
541 	return 0;
542 
543 free_bad_page:
544 	dma_free_wc(dev, SZ_4K, global->bad_page_cpu, global->bad_page_dma);
545 free_global:
546 	kfree(global);
547 
548 	return -ENOMEM;
549 }
550 
551 void etnaviv_iommu_global_fini(struct etnaviv_gpu *gpu)
552 {
553 	struct etnaviv_drm_private *priv = gpu->drm->dev_private;
554 	struct etnaviv_iommu_global *global = priv->mmu_global;
555 
556 	if (!global)
557 		return;
558 
559 	if (--global->use > 0)
560 		return;
561 
562 	if (global->v2.pta_cpu)
563 		dma_free_wc(global->dev, ETNAVIV_PTA_SIZE,
564 			    global->v2.pta_cpu, global->v2.pta_dma);
565 
566 	if (global->bad_page_cpu)
567 		dma_free_wc(global->dev, SZ_4K,
568 			    global->bad_page_cpu, global->bad_page_dma);
569 
570 	mutex_destroy(&global->lock);
571 	kfree(global);
572 
573 	priv->mmu_global = NULL;
574 }
575