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, unsigned len, 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 		u32 pa = sg_dma_address(sg) - sg->offset;
84 		size_t bytes = sg_dma_len(sg) + sg->offset;
85 
86 		VERB("map[%d]: %08x %08x(%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 	return 0;
96 
97 fail:
98 	etnaviv_context_unmap(context, iova, da - iova);
99 	return ret;
100 }
101 
102 static void etnaviv_iommu_unmap(struct etnaviv_iommu_context *context, u32 iova,
103 				struct sg_table *sgt, unsigned len)
104 {
105 	struct scatterlist *sg;
106 	unsigned int da = iova;
107 	int i;
108 
109 	for_each_sgtable_dma_sg(sgt, sg, i) {
110 		size_t bytes = sg_dma_len(sg) + sg->offset;
111 
112 		etnaviv_context_unmap(context, da, bytes);
113 
114 		VERB("unmap[%d]: %08x(%zx)", i, iova, bytes);
115 
116 		BUG_ON(!PAGE_ALIGNED(bytes));
117 
118 		da += bytes;
119 	}
120 }
121 
122 static void etnaviv_iommu_remove_mapping(struct etnaviv_iommu_context *context,
123 	struct etnaviv_vram_mapping *mapping)
124 {
125 	struct etnaviv_gem_object *etnaviv_obj = mapping->object;
126 
127 	lockdep_assert_held(&context->lock);
128 
129 	etnaviv_iommu_unmap(context, mapping->vram_node.start,
130 			    etnaviv_obj->sgt, etnaviv_obj->base.size);
131 	drm_mm_remove_node(&mapping->vram_node);
132 }
133 
134 static int etnaviv_iommu_find_iova(struct etnaviv_iommu_context *context,
135 				   struct drm_mm_node *node, size_t size)
136 {
137 	struct etnaviv_vram_mapping *free = NULL;
138 	enum drm_mm_insert_mode mode = DRM_MM_INSERT_LOW;
139 	int ret;
140 
141 	lockdep_assert_held(&context->lock);
142 
143 	while (1) {
144 		struct etnaviv_vram_mapping *m, *n;
145 		struct drm_mm_scan scan;
146 		struct list_head list;
147 		bool found;
148 
149 		ret = drm_mm_insert_node_in_range(&context->mm, node,
150 						  size, 0, 0, 0, U64_MAX, mode);
151 		if (ret != -ENOSPC)
152 			break;
153 
154 		/* Try to retire some entries */
155 		drm_mm_scan_init(&scan, &context->mm, size, 0, 0, mode);
156 
157 		found = 0;
158 		INIT_LIST_HEAD(&list);
159 		list_for_each_entry(free, &context->mappings, mmu_node) {
160 			/* If this vram node has not been used, skip this. */
161 			if (!free->vram_node.mm)
162 				continue;
163 
164 			/*
165 			 * If the iova is pinned, then it's in-use,
166 			 * so we must keep its mapping.
167 			 */
168 			if (free->use)
169 				continue;
170 
171 			list_add(&free->scan_node, &list);
172 			if (drm_mm_scan_add_block(&scan, &free->vram_node)) {
173 				found = true;
174 				break;
175 			}
176 		}
177 
178 		if (!found) {
179 			/* Nothing found, clean up and fail */
180 			list_for_each_entry_safe(m, n, &list, scan_node)
181 				BUG_ON(drm_mm_scan_remove_block(&scan, &m->vram_node));
182 			break;
183 		}
184 
185 		/*
186 		 * drm_mm does not allow any other operations while
187 		 * scanning, so we have to remove all blocks first.
188 		 * If drm_mm_scan_remove_block() returns false, we
189 		 * can leave the block pinned.
190 		 */
191 		list_for_each_entry_safe(m, n, &list, scan_node)
192 			if (!drm_mm_scan_remove_block(&scan, &m->vram_node))
193 				list_del_init(&m->scan_node);
194 
195 		/*
196 		 * Unmap the blocks which need to be reaped from the MMU.
197 		 * Clear the mmu pointer to prevent the mapping_get finding
198 		 * this mapping.
199 		 */
200 		list_for_each_entry_safe(m, n, &list, scan_node) {
201 			etnaviv_iommu_remove_mapping(context, m);
202 			etnaviv_iommu_context_put(m->context);
203 			m->context = NULL;
204 			list_del_init(&m->mmu_node);
205 			list_del_init(&m->scan_node);
206 		}
207 
208 		mode = DRM_MM_INSERT_EVICT;
209 
210 		/*
211 		 * We removed enough mappings so that the new allocation will
212 		 * succeed, retry the allocation one more time.
213 		 */
214 	}
215 
216 	return ret;
217 }
218 
219 static int etnaviv_iommu_insert_exact(struct etnaviv_iommu_context *context,
220 		   struct drm_mm_node *node, size_t size, u64 va)
221 {
222 	lockdep_assert_held(&context->lock);
223 
224 	return drm_mm_insert_node_in_range(&context->mm, node, size, 0, 0, va,
225 					   va + size, DRM_MM_INSERT_LOWEST);
226 }
227 
228 int etnaviv_iommu_map_gem(struct etnaviv_iommu_context *context,
229 	struct etnaviv_gem_object *etnaviv_obj, u32 memory_base,
230 	struct etnaviv_vram_mapping *mapping, u64 va)
231 {
232 	struct sg_table *sgt = etnaviv_obj->sgt;
233 	struct drm_mm_node *node;
234 	int ret;
235 
236 	lockdep_assert_held(&etnaviv_obj->lock);
237 
238 	mutex_lock(&context->lock);
239 
240 	/* v1 MMU can optimize single entry (contiguous) scatterlists */
241 	if (context->global->version == ETNAVIV_IOMMU_V1 &&
242 	    sgt->nents == 1 && !(etnaviv_obj->flags & ETNA_BO_FORCE_MMU)) {
243 		u32 iova;
244 
245 		iova = sg_dma_address(sgt->sgl) - memory_base;
246 		if (iova < 0x80000000 - sg_dma_len(sgt->sgl)) {
247 			mapping->iova = iova;
248 			list_add_tail(&mapping->mmu_node, &context->mappings);
249 			ret = 0;
250 			goto unlock;
251 		}
252 	}
253 
254 	node = &mapping->vram_node;
255 
256 	if (va)
257 		ret = etnaviv_iommu_insert_exact(context, node,
258 						 etnaviv_obj->base.size, va);
259 	else
260 		ret = etnaviv_iommu_find_iova(context, node,
261 					      etnaviv_obj->base.size);
262 	if (ret < 0)
263 		goto unlock;
264 
265 	mapping->iova = node->start;
266 	ret = etnaviv_iommu_map(context, node->start, sgt, etnaviv_obj->base.size,
267 				ETNAVIV_PROT_READ | ETNAVIV_PROT_WRITE);
268 
269 	if (ret < 0) {
270 		drm_mm_remove_node(node);
271 		goto unlock;
272 	}
273 
274 	list_add_tail(&mapping->mmu_node, &context->mappings);
275 	context->flush_seq++;
276 unlock:
277 	mutex_unlock(&context->lock);
278 
279 	return ret;
280 }
281 
282 void etnaviv_iommu_unmap_gem(struct etnaviv_iommu_context *context,
283 	struct etnaviv_vram_mapping *mapping)
284 {
285 	WARN_ON(mapping->use);
286 
287 	mutex_lock(&context->lock);
288 
289 	/* If the vram node is on the mm, unmap and remove the node */
290 	if (mapping->vram_node.mm == &context->mm)
291 		etnaviv_iommu_remove_mapping(context, mapping);
292 
293 	list_del(&mapping->mmu_node);
294 	context->flush_seq++;
295 	mutex_unlock(&context->lock);
296 }
297 
298 static void etnaviv_iommu_context_free(struct kref *kref)
299 {
300 	struct etnaviv_iommu_context *context =
301 		container_of(kref, struct etnaviv_iommu_context, refcount);
302 
303 	etnaviv_cmdbuf_suballoc_unmap(context, &context->cmdbuf_mapping);
304 
305 	context->global->ops->free(context);
306 }
307 void etnaviv_iommu_context_put(struct etnaviv_iommu_context *context)
308 {
309 	kref_put(&context->refcount, etnaviv_iommu_context_free);
310 }
311 
312 struct etnaviv_iommu_context *
313 etnaviv_iommu_context_init(struct etnaviv_iommu_global *global,
314 			   struct etnaviv_cmdbuf_suballoc *suballoc)
315 {
316 	struct etnaviv_iommu_context *ctx;
317 	int ret;
318 
319 	if (global->version == ETNAVIV_IOMMU_V1)
320 		ctx = etnaviv_iommuv1_context_alloc(global);
321 	else
322 		ctx = etnaviv_iommuv2_context_alloc(global);
323 
324 	if (!ctx)
325 		return NULL;
326 
327 	ret = etnaviv_cmdbuf_suballoc_map(suballoc, ctx, &ctx->cmdbuf_mapping,
328 					  global->memory_base);
329 	if (ret)
330 		goto out_free;
331 
332 	if (global->version == ETNAVIV_IOMMU_V1 &&
333 	    ctx->cmdbuf_mapping.iova > 0x80000000) {
334 		dev_err(global->dev,
335 		        "command buffer outside valid memory window\n");
336 		goto out_unmap;
337 	}
338 
339 	return ctx;
340 
341 out_unmap:
342 	etnaviv_cmdbuf_suballoc_unmap(ctx, &ctx->cmdbuf_mapping);
343 out_free:
344 	global->ops->free(ctx);
345 	return NULL;
346 }
347 
348 void etnaviv_iommu_restore(struct etnaviv_gpu *gpu,
349 			   struct etnaviv_iommu_context *context)
350 {
351 	context->global->ops->restore(gpu, context);
352 }
353 
354 int etnaviv_iommu_get_suballoc_va(struct etnaviv_iommu_context *context,
355 				  struct etnaviv_vram_mapping *mapping,
356 				  u32 memory_base, dma_addr_t paddr,
357 				  size_t size)
358 {
359 	mutex_lock(&context->lock);
360 
361 	if (mapping->use > 0) {
362 		mapping->use++;
363 		mutex_unlock(&context->lock);
364 		return 0;
365 	}
366 
367 	/*
368 	 * For MMUv1 we don't add the suballoc region to the pagetables, as
369 	 * those GPUs can only work with cmdbufs accessed through the linear
370 	 * window. Instead we manufacture a mapping to make it look uniform
371 	 * to the upper layers.
372 	 */
373 	if (context->global->version == ETNAVIV_IOMMU_V1) {
374 		mapping->iova = paddr - memory_base;
375 	} else {
376 		struct drm_mm_node *node = &mapping->vram_node;
377 		int ret;
378 
379 		ret = etnaviv_iommu_find_iova(context, node, size);
380 		if (ret < 0) {
381 			mutex_unlock(&context->lock);
382 			return ret;
383 		}
384 
385 		mapping->iova = node->start;
386 		ret = etnaviv_context_map(context, node->start, paddr, size,
387 					  ETNAVIV_PROT_READ);
388 		if (ret < 0) {
389 			drm_mm_remove_node(node);
390 			mutex_unlock(&context->lock);
391 			return ret;
392 		}
393 
394 		context->flush_seq++;
395 	}
396 
397 	list_add_tail(&mapping->mmu_node, &context->mappings);
398 	mapping->use = 1;
399 
400 	mutex_unlock(&context->lock);
401 
402 	return 0;
403 }
404 
405 void etnaviv_iommu_put_suballoc_va(struct etnaviv_iommu_context *context,
406 		  struct etnaviv_vram_mapping *mapping)
407 {
408 	struct drm_mm_node *node = &mapping->vram_node;
409 
410 	mutex_lock(&context->lock);
411 	mapping->use--;
412 
413 	if (mapping->use > 0 || context->global->version == ETNAVIV_IOMMU_V1) {
414 		mutex_unlock(&context->lock);
415 		return;
416 	}
417 
418 	etnaviv_context_unmap(context, node->start, node->size);
419 	drm_mm_remove_node(node);
420 	mutex_unlock(&context->lock);
421 }
422 
423 size_t etnaviv_iommu_dump_size(struct etnaviv_iommu_context *context)
424 {
425 	return context->global->ops->dump_size(context);
426 }
427 
428 void etnaviv_iommu_dump(struct etnaviv_iommu_context *context, void *buf)
429 {
430 	context->global->ops->dump(context, buf);
431 }
432 
433 int etnaviv_iommu_global_init(struct etnaviv_gpu *gpu)
434 {
435 	enum etnaviv_iommu_version version = ETNAVIV_IOMMU_V1;
436 	struct etnaviv_drm_private *priv = gpu->drm->dev_private;
437 	struct etnaviv_iommu_global *global;
438 	struct device *dev = gpu->drm->dev;
439 
440 	if (gpu->identity.minor_features1 & chipMinorFeatures1_MMU_VERSION)
441 		version = ETNAVIV_IOMMU_V2;
442 
443 	if (priv->mmu_global) {
444 		if (priv->mmu_global->version != version) {
445 			dev_err(gpu->dev,
446 				"MMU version doesn't match global version\n");
447 			return -ENXIO;
448 		}
449 
450 		priv->mmu_global->use++;
451 		return 0;
452 	}
453 
454 	global = kzalloc(sizeof(*global), GFP_KERNEL);
455 	if (!global)
456 		return -ENOMEM;
457 
458 	global->bad_page_cpu = dma_alloc_wc(dev, SZ_4K, &global->bad_page_dma,
459 					    GFP_KERNEL);
460 	if (!global->bad_page_cpu)
461 		goto free_global;
462 
463 	memset32(global->bad_page_cpu, 0xdead55aa, SZ_4K / sizeof(u32));
464 
465 	if (version == ETNAVIV_IOMMU_V2) {
466 		global->v2.pta_cpu = dma_alloc_wc(dev, ETNAVIV_PTA_SIZE,
467 					       &global->v2.pta_dma, GFP_KERNEL);
468 		if (!global->v2.pta_cpu)
469 			goto free_bad_page;
470 	}
471 
472 	global->dev = dev;
473 	global->version = version;
474 	global->use = 1;
475 	mutex_init(&global->lock);
476 
477 	if (version == ETNAVIV_IOMMU_V1)
478 		global->ops = &etnaviv_iommuv1_ops;
479 	else
480 		global->ops = &etnaviv_iommuv2_ops;
481 
482 	priv->mmu_global = global;
483 
484 	return 0;
485 
486 free_bad_page:
487 	dma_free_wc(dev, SZ_4K, global->bad_page_cpu, global->bad_page_dma);
488 free_global:
489 	kfree(global);
490 
491 	return -ENOMEM;
492 }
493 
494 void etnaviv_iommu_global_fini(struct etnaviv_gpu *gpu)
495 {
496 	struct etnaviv_drm_private *priv = gpu->drm->dev_private;
497 	struct etnaviv_iommu_global *global = priv->mmu_global;
498 
499 	if (--global->use > 0)
500 		return;
501 
502 	if (global->v2.pta_cpu)
503 		dma_free_wc(global->dev, ETNAVIV_PTA_SIZE,
504 			    global->v2.pta_cpu, global->v2.pta_dma);
505 
506 	if (global->bad_page_cpu)
507 		dma_free_wc(global->dev, SZ_4K,
508 			    global->bad_page_cpu, global->bad_page_dma);
509 
510 	mutex_destroy(&global->lock);
511 	kfree(global);
512 
513 	priv->mmu_global = NULL;
514 }
515