xref: /openbmc/linux/drivers/gpu/drm/msm/msm_gem.c (revision 884caada)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2013 Red Hat
4  * Author: Rob Clark <robdclark@gmail.com>
5  */
6 
7 #include <linux/spinlock.h>
8 #include <linux/shmem_fs.h>
9 #include <linux/dma-buf.h>
10 #include <linux/pfn_t.h>
11 
12 #include <drm/drm_prime.h>
13 
14 #include "msm_drv.h"
15 #include "msm_fence.h"
16 #include "msm_gem.h"
17 #include "msm_gpu.h"
18 #include "msm_mmu.h"
19 
20 static void msm_gem_vunmap_locked(struct drm_gem_object *obj);
21 
22 
23 static dma_addr_t physaddr(struct drm_gem_object *obj)
24 {
25 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
26 	struct msm_drm_private *priv = obj->dev->dev_private;
27 	return (((dma_addr_t)msm_obj->vram_node->start) << PAGE_SHIFT) +
28 			priv->vram.paddr;
29 }
30 
31 static bool use_pages(struct drm_gem_object *obj)
32 {
33 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
34 	return !msm_obj->vram_node;
35 }
36 
37 /*
38  * Cache sync.. this is a bit over-complicated, to fit dma-mapping
39  * API.  Really GPU cache is out of scope here (handled on cmdstream)
40  * and all we need to do is invalidate newly allocated pages before
41  * mapping to CPU as uncached/writecombine.
42  *
43  * On top of this, we have the added headache, that depending on
44  * display generation, the display's iommu may be wired up to either
45  * the toplevel drm device (mdss), or to the mdp sub-node, meaning
46  * that here we either have dma-direct or iommu ops.
47  *
48  * Let this be a cautionary tail of abstraction gone wrong.
49  */
50 
51 static void sync_for_device(struct msm_gem_object *msm_obj)
52 {
53 	struct device *dev = msm_obj->base.dev->dev;
54 
55 	if (get_dma_ops(dev) && IS_ENABLED(CONFIG_ARM64)) {
56 		dma_sync_sg_for_device(dev, msm_obj->sgt->sgl,
57 			msm_obj->sgt->nents, DMA_BIDIRECTIONAL);
58 	} else {
59 		dma_map_sg(dev, msm_obj->sgt->sgl,
60 			msm_obj->sgt->nents, DMA_BIDIRECTIONAL);
61 	}
62 }
63 
64 static void sync_for_cpu(struct msm_gem_object *msm_obj)
65 {
66 	struct device *dev = msm_obj->base.dev->dev;
67 
68 	if (get_dma_ops(dev) && IS_ENABLED(CONFIG_ARM64)) {
69 		dma_sync_sg_for_cpu(dev, msm_obj->sgt->sgl,
70 			msm_obj->sgt->nents, DMA_BIDIRECTIONAL);
71 	} else {
72 		dma_unmap_sg(dev, msm_obj->sgt->sgl,
73 			msm_obj->sgt->nents, DMA_BIDIRECTIONAL);
74 	}
75 }
76 
77 /* allocate pages from VRAM carveout, used when no IOMMU: */
78 static struct page **get_pages_vram(struct drm_gem_object *obj, int npages)
79 {
80 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
81 	struct msm_drm_private *priv = obj->dev->dev_private;
82 	dma_addr_t paddr;
83 	struct page **p;
84 	int ret, i;
85 
86 	p = kvmalloc_array(npages, sizeof(struct page *), GFP_KERNEL);
87 	if (!p)
88 		return ERR_PTR(-ENOMEM);
89 
90 	spin_lock(&priv->vram.lock);
91 	ret = drm_mm_insert_node(&priv->vram.mm, msm_obj->vram_node, npages);
92 	spin_unlock(&priv->vram.lock);
93 	if (ret) {
94 		kvfree(p);
95 		return ERR_PTR(ret);
96 	}
97 
98 	paddr = physaddr(obj);
99 	for (i = 0; i < npages; i++) {
100 		p[i] = phys_to_page(paddr);
101 		paddr += PAGE_SIZE;
102 	}
103 
104 	return p;
105 }
106 
107 static struct page **get_pages(struct drm_gem_object *obj)
108 {
109 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
110 
111 	if (!msm_obj->pages) {
112 		struct drm_device *dev = obj->dev;
113 		struct page **p;
114 		int npages = obj->size >> PAGE_SHIFT;
115 
116 		if (use_pages(obj))
117 			p = drm_gem_get_pages(obj);
118 		else
119 			p = get_pages_vram(obj, npages);
120 
121 		if (IS_ERR(p)) {
122 			DRM_DEV_ERROR(dev->dev, "could not get pages: %ld\n",
123 					PTR_ERR(p));
124 			return p;
125 		}
126 
127 		msm_obj->pages = p;
128 
129 		msm_obj->sgt = drm_prime_pages_to_sg(p, npages);
130 		if (IS_ERR(msm_obj->sgt)) {
131 			void *ptr = ERR_CAST(msm_obj->sgt);
132 
133 			DRM_DEV_ERROR(dev->dev, "failed to allocate sgt\n");
134 			msm_obj->sgt = NULL;
135 			return ptr;
136 		}
137 
138 		/* For non-cached buffers, ensure the new pages are clean
139 		 * because display controller, GPU, etc. are not coherent:
140 		 */
141 		if (msm_obj->flags & (MSM_BO_WC|MSM_BO_UNCACHED))
142 			sync_for_device(msm_obj);
143 	}
144 
145 	return msm_obj->pages;
146 }
147 
148 static void put_pages_vram(struct drm_gem_object *obj)
149 {
150 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
151 	struct msm_drm_private *priv = obj->dev->dev_private;
152 
153 	spin_lock(&priv->vram.lock);
154 	drm_mm_remove_node(msm_obj->vram_node);
155 	spin_unlock(&priv->vram.lock);
156 
157 	kvfree(msm_obj->pages);
158 }
159 
160 static void put_pages(struct drm_gem_object *obj)
161 {
162 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
163 
164 	if (msm_obj->pages) {
165 		if (msm_obj->sgt) {
166 			/* For non-cached buffers, ensure the new
167 			 * pages are clean because display controller,
168 			 * GPU, etc. are not coherent:
169 			 */
170 			if (msm_obj->flags & (MSM_BO_WC|MSM_BO_UNCACHED))
171 				sync_for_cpu(msm_obj);
172 
173 			sg_free_table(msm_obj->sgt);
174 			kfree(msm_obj->sgt);
175 		}
176 
177 		if (use_pages(obj))
178 			drm_gem_put_pages(obj, msm_obj->pages, true, false);
179 		else
180 			put_pages_vram(obj);
181 
182 		msm_obj->pages = NULL;
183 	}
184 }
185 
186 struct page **msm_gem_get_pages(struct drm_gem_object *obj)
187 {
188 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
189 	struct page **p;
190 
191 	mutex_lock(&msm_obj->lock);
192 
193 	if (WARN_ON(msm_obj->madv != MSM_MADV_WILLNEED)) {
194 		mutex_unlock(&msm_obj->lock);
195 		return ERR_PTR(-EBUSY);
196 	}
197 
198 	p = get_pages(obj);
199 	mutex_unlock(&msm_obj->lock);
200 	return p;
201 }
202 
203 void msm_gem_put_pages(struct drm_gem_object *obj)
204 {
205 	/* when we start tracking the pin count, then do something here */
206 }
207 
208 int msm_gem_mmap_obj(struct drm_gem_object *obj,
209 		struct vm_area_struct *vma)
210 {
211 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
212 
213 	vma->vm_flags &= ~VM_PFNMAP;
214 	vma->vm_flags |= VM_MIXEDMAP;
215 
216 	if (msm_obj->flags & MSM_BO_WC) {
217 		vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
218 	} else if (msm_obj->flags & MSM_BO_UNCACHED) {
219 		vma->vm_page_prot = pgprot_noncached(vm_get_page_prot(vma->vm_flags));
220 	} else {
221 		/*
222 		 * Shunt off cached objs to shmem file so they have their own
223 		 * address_space (so unmap_mapping_range does what we want,
224 		 * in particular in the case of mmap'd dmabufs)
225 		 */
226 		fput(vma->vm_file);
227 		get_file(obj->filp);
228 		vma->vm_pgoff = 0;
229 		vma->vm_file  = obj->filp;
230 
231 		vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
232 	}
233 
234 	return 0;
235 }
236 
237 int msm_gem_mmap(struct file *filp, struct vm_area_struct *vma)
238 {
239 	int ret;
240 
241 	ret = drm_gem_mmap(filp, vma);
242 	if (ret) {
243 		DBG("mmap failed: %d", ret);
244 		return ret;
245 	}
246 
247 	return msm_gem_mmap_obj(vma->vm_private_data, vma);
248 }
249 
250 vm_fault_t msm_gem_fault(struct vm_fault *vmf)
251 {
252 	struct vm_area_struct *vma = vmf->vma;
253 	struct drm_gem_object *obj = vma->vm_private_data;
254 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
255 	struct page **pages;
256 	unsigned long pfn;
257 	pgoff_t pgoff;
258 	int err;
259 	vm_fault_t ret;
260 
261 	/*
262 	 * vm_ops.open/drm_gem_mmap_obj and close get and put
263 	 * a reference on obj. So, we dont need to hold one here.
264 	 */
265 	err = mutex_lock_interruptible(&msm_obj->lock);
266 	if (err) {
267 		ret = VM_FAULT_NOPAGE;
268 		goto out;
269 	}
270 
271 	if (WARN_ON(msm_obj->madv != MSM_MADV_WILLNEED)) {
272 		mutex_unlock(&msm_obj->lock);
273 		return VM_FAULT_SIGBUS;
274 	}
275 
276 	/* make sure we have pages attached now */
277 	pages = get_pages(obj);
278 	if (IS_ERR(pages)) {
279 		ret = vmf_error(PTR_ERR(pages));
280 		goto out_unlock;
281 	}
282 
283 	/* We don't use vmf->pgoff since that has the fake offset: */
284 	pgoff = (vmf->address - vma->vm_start) >> PAGE_SHIFT;
285 
286 	pfn = page_to_pfn(pages[pgoff]);
287 
288 	VERB("Inserting %p pfn %lx, pa %lx", (void *)vmf->address,
289 			pfn, pfn << PAGE_SHIFT);
290 
291 	ret = vmf_insert_mixed(vma, vmf->address, __pfn_to_pfn_t(pfn, PFN_DEV));
292 out_unlock:
293 	mutex_unlock(&msm_obj->lock);
294 out:
295 	return ret;
296 }
297 
298 /** get mmap offset */
299 static uint64_t mmap_offset(struct drm_gem_object *obj)
300 {
301 	struct drm_device *dev = obj->dev;
302 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
303 	int ret;
304 
305 	WARN_ON(!mutex_is_locked(&msm_obj->lock));
306 
307 	/* Make it mmapable */
308 	ret = drm_gem_create_mmap_offset(obj);
309 
310 	if (ret) {
311 		DRM_DEV_ERROR(dev->dev, "could not allocate mmap offset\n");
312 		return 0;
313 	}
314 
315 	return drm_vma_node_offset_addr(&obj->vma_node);
316 }
317 
318 uint64_t msm_gem_mmap_offset(struct drm_gem_object *obj)
319 {
320 	uint64_t offset;
321 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
322 
323 	mutex_lock(&msm_obj->lock);
324 	offset = mmap_offset(obj);
325 	mutex_unlock(&msm_obj->lock);
326 	return offset;
327 }
328 
329 static struct msm_gem_vma *add_vma(struct drm_gem_object *obj,
330 		struct msm_gem_address_space *aspace)
331 {
332 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
333 	struct msm_gem_vma *vma;
334 
335 	WARN_ON(!mutex_is_locked(&msm_obj->lock));
336 
337 	vma = kzalloc(sizeof(*vma), GFP_KERNEL);
338 	if (!vma)
339 		return ERR_PTR(-ENOMEM);
340 
341 	vma->aspace = aspace;
342 
343 	list_add_tail(&vma->list, &msm_obj->vmas);
344 
345 	return vma;
346 }
347 
348 static struct msm_gem_vma *lookup_vma(struct drm_gem_object *obj,
349 		struct msm_gem_address_space *aspace)
350 {
351 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
352 	struct msm_gem_vma *vma;
353 
354 	WARN_ON(!mutex_is_locked(&msm_obj->lock));
355 
356 	list_for_each_entry(vma, &msm_obj->vmas, list) {
357 		if (vma->aspace == aspace)
358 			return vma;
359 	}
360 
361 	return NULL;
362 }
363 
364 static void del_vma(struct msm_gem_vma *vma)
365 {
366 	if (!vma)
367 		return;
368 
369 	list_del(&vma->list);
370 	kfree(vma);
371 }
372 
373 /* Called with msm_obj->lock locked */
374 static void
375 put_iova(struct drm_gem_object *obj)
376 {
377 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
378 	struct msm_gem_vma *vma, *tmp;
379 
380 	WARN_ON(!mutex_is_locked(&msm_obj->lock));
381 
382 	list_for_each_entry_safe(vma, tmp, &msm_obj->vmas, list) {
383 		if (vma->aspace) {
384 			msm_gem_purge_vma(vma->aspace, vma);
385 			msm_gem_close_vma(vma->aspace, vma);
386 		}
387 		del_vma(vma);
388 	}
389 }
390 
391 static int msm_gem_get_iova_locked(struct drm_gem_object *obj,
392 		struct msm_gem_address_space *aspace, uint64_t *iova)
393 {
394 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
395 	struct msm_gem_vma *vma;
396 	int ret = 0;
397 
398 	WARN_ON(!mutex_is_locked(&msm_obj->lock));
399 
400 	vma = lookup_vma(obj, aspace);
401 
402 	if (!vma) {
403 		vma = add_vma(obj, aspace);
404 		if (IS_ERR(vma))
405 			return PTR_ERR(vma);
406 
407 		ret = msm_gem_init_vma(aspace, vma, obj->size >> PAGE_SHIFT);
408 		if (ret) {
409 			del_vma(vma);
410 			return ret;
411 		}
412 	}
413 
414 	*iova = vma->iova;
415 	return 0;
416 }
417 
418 static int msm_gem_pin_iova(struct drm_gem_object *obj,
419 		struct msm_gem_address_space *aspace)
420 {
421 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
422 	struct msm_gem_vma *vma;
423 	struct page **pages;
424 	int prot = IOMMU_READ;
425 
426 	if (!(msm_obj->flags & MSM_BO_GPU_READONLY))
427 		prot |= IOMMU_WRITE;
428 
429 	WARN_ON(!mutex_is_locked(&msm_obj->lock));
430 
431 	if (WARN_ON(msm_obj->madv != MSM_MADV_WILLNEED))
432 		return -EBUSY;
433 
434 	vma = lookup_vma(obj, aspace);
435 	if (WARN_ON(!vma))
436 		return -EINVAL;
437 
438 	pages = get_pages(obj);
439 	if (IS_ERR(pages))
440 		return PTR_ERR(pages);
441 
442 	return msm_gem_map_vma(aspace, vma, prot,
443 			msm_obj->sgt, obj->size >> PAGE_SHIFT);
444 }
445 
446 /* get iova and pin it. Should have a matching put */
447 int msm_gem_get_and_pin_iova(struct drm_gem_object *obj,
448 		struct msm_gem_address_space *aspace, uint64_t *iova)
449 {
450 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
451 	u64 local;
452 	int ret;
453 
454 	mutex_lock(&msm_obj->lock);
455 
456 	ret = msm_gem_get_iova_locked(obj, aspace, &local);
457 
458 	if (!ret)
459 		ret = msm_gem_pin_iova(obj, aspace);
460 
461 	if (!ret)
462 		*iova = local;
463 
464 	mutex_unlock(&msm_obj->lock);
465 	return ret;
466 }
467 
468 /*
469  * Get an iova but don't pin it. Doesn't need a put because iovas are currently
470  * valid for the life of the object
471  */
472 int msm_gem_get_iova(struct drm_gem_object *obj,
473 		struct msm_gem_address_space *aspace, uint64_t *iova)
474 {
475 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
476 	int ret;
477 
478 	mutex_lock(&msm_obj->lock);
479 	ret = msm_gem_get_iova_locked(obj, aspace, iova);
480 	mutex_unlock(&msm_obj->lock);
481 
482 	return ret;
483 }
484 
485 /* get iova without taking a reference, used in places where you have
486  * already done a 'msm_gem_get_and_pin_iova' or 'msm_gem_get_iova'
487  */
488 uint64_t msm_gem_iova(struct drm_gem_object *obj,
489 		struct msm_gem_address_space *aspace)
490 {
491 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
492 	struct msm_gem_vma *vma;
493 
494 	mutex_lock(&msm_obj->lock);
495 	vma = lookup_vma(obj, aspace);
496 	mutex_unlock(&msm_obj->lock);
497 	WARN_ON(!vma);
498 
499 	return vma ? vma->iova : 0;
500 }
501 
502 /*
503  * Unpin a iova by updating the reference counts. The memory isn't actually
504  * purged until something else (shrinker, mm_notifier, destroy, etc) decides
505  * to get rid of it
506  */
507 void msm_gem_unpin_iova(struct drm_gem_object *obj,
508 		struct msm_gem_address_space *aspace)
509 {
510 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
511 	struct msm_gem_vma *vma;
512 
513 	mutex_lock(&msm_obj->lock);
514 	vma = lookup_vma(obj, aspace);
515 
516 	if (!WARN_ON(!vma))
517 		msm_gem_unmap_vma(aspace, vma);
518 
519 	mutex_unlock(&msm_obj->lock);
520 }
521 
522 int msm_gem_dumb_create(struct drm_file *file, struct drm_device *dev,
523 		struct drm_mode_create_dumb *args)
524 {
525 	args->pitch = align_pitch(args->width, args->bpp);
526 	args->size  = PAGE_ALIGN(args->pitch * args->height);
527 	return msm_gem_new_handle(dev, file, args->size,
528 			MSM_BO_SCANOUT | MSM_BO_WC, &args->handle, "dumb");
529 }
530 
531 int msm_gem_dumb_map_offset(struct drm_file *file, struct drm_device *dev,
532 		uint32_t handle, uint64_t *offset)
533 {
534 	struct drm_gem_object *obj;
535 	int ret = 0;
536 
537 	/* GEM does all our handle to object mapping */
538 	obj = drm_gem_object_lookup(file, handle);
539 	if (obj == NULL) {
540 		ret = -ENOENT;
541 		goto fail;
542 	}
543 
544 	*offset = msm_gem_mmap_offset(obj);
545 
546 	drm_gem_object_put_unlocked(obj);
547 
548 fail:
549 	return ret;
550 }
551 
552 static void *get_vaddr(struct drm_gem_object *obj, unsigned madv)
553 {
554 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
555 	int ret = 0;
556 
557 	mutex_lock(&msm_obj->lock);
558 
559 	if (WARN_ON(msm_obj->madv > madv)) {
560 		DRM_DEV_ERROR(obj->dev->dev, "Invalid madv state: %u vs %u\n",
561 			msm_obj->madv, madv);
562 		mutex_unlock(&msm_obj->lock);
563 		return ERR_PTR(-EBUSY);
564 	}
565 
566 	/* increment vmap_count *before* vmap() call, so shrinker can
567 	 * check vmap_count (is_vunmapable()) outside of msm_obj->lock.
568 	 * This guarantees that we won't try to msm_gem_vunmap() this
569 	 * same object from within the vmap() call (while we already
570 	 * hold msm_obj->lock)
571 	 */
572 	msm_obj->vmap_count++;
573 
574 	if (!msm_obj->vaddr) {
575 		struct page **pages = get_pages(obj);
576 		if (IS_ERR(pages)) {
577 			ret = PTR_ERR(pages);
578 			goto fail;
579 		}
580 		msm_obj->vaddr = vmap(pages, obj->size >> PAGE_SHIFT,
581 				VM_MAP, pgprot_writecombine(PAGE_KERNEL));
582 		if (msm_obj->vaddr == NULL) {
583 			ret = -ENOMEM;
584 			goto fail;
585 		}
586 	}
587 
588 	mutex_unlock(&msm_obj->lock);
589 	return msm_obj->vaddr;
590 
591 fail:
592 	msm_obj->vmap_count--;
593 	mutex_unlock(&msm_obj->lock);
594 	return ERR_PTR(ret);
595 }
596 
597 void *msm_gem_get_vaddr(struct drm_gem_object *obj)
598 {
599 	return get_vaddr(obj, MSM_MADV_WILLNEED);
600 }
601 
602 /*
603  * Don't use this!  It is for the very special case of dumping
604  * submits from GPU hangs or faults, were the bo may already
605  * be MSM_MADV_DONTNEED, but we know the buffer is still on the
606  * active list.
607  */
608 void *msm_gem_get_vaddr_active(struct drm_gem_object *obj)
609 {
610 	return get_vaddr(obj, __MSM_MADV_PURGED);
611 }
612 
613 void msm_gem_put_vaddr(struct drm_gem_object *obj)
614 {
615 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
616 
617 	mutex_lock(&msm_obj->lock);
618 	WARN_ON(msm_obj->vmap_count < 1);
619 	msm_obj->vmap_count--;
620 	mutex_unlock(&msm_obj->lock);
621 }
622 
623 /* Update madvise status, returns true if not purged, else
624  * false or -errno.
625  */
626 int msm_gem_madvise(struct drm_gem_object *obj, unsigned madv)
627 {
628 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
629 
630 	mutex_lock(&msm_obj->lock);
631 
632 	WARN_ON(!mutex_is_locked(&obj->dev->struct_mutex));
633 
634 	if (msm_obj->madv != __MSM_MADV_PURGED)
635 		msm_obj->madv = madv;
636 
637 	madv = msm_obj->madv;
638 
639 	mutex_unlock(&msm_obj->lock);
640 
641 	return (madv != __MSM_MADV_PURGED);
642 }
643 
644 void msm_gem_purge(struct drm_gem_object *obj, enum msm_gem_lock subclass)
645 {
646 	struct drm_device *dev = obj->dev;
647 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
648 
649 	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
650 	WARN_ON(!is_purgeable(msm_obj));
651 	WARN_ON(obj->import_attach);
652 
653 	mutex_lock_nested(&msm_obj->lock, subclass);
654 
655 	put_iova(obj);
656 
657 	msm_gem_vunmap_locked(obj);
658 
659 	put_pages(obj);
660 
661 	msm_obj->madv = __MSM_MADV_PURGED;
662 
663 	drm_vma_node_unmap(&obj->vma_node, dev->anon_inode->i_mapping);
664 	drm_gem_free_mmap_offset(obj);
665 
666 	/* Our goal here is to return as much of the memory as
667 	 * is possible back to the system as we are called from OOM.
668 	 * To do this we must instruct the shmfs to drop all of its
669 	 * backing pages, *now*.
670 	 */
671 	shmem_truncate_range(file_inode(obj->filp), 0, (loff_t)-1);
672 
673 	invalidate_mapping_pages(file_inode(obj->filp)->i_mapping,
674 			0, (loff_t)-1);
675 
676 	mutex_unlock(&msm_obj->lock);
677 }
678 
679 static void msm_gem_vunmap_locked(struct drm_gem_object *obj)
680 {
681 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
682 
683 	WARN_ON(!mutex_is_locked(&msm_obj->lock));
684 
685 	if (!msm_obj->vaddr || WARN_ON(!is_vunmapable(msm_obj)))
686 		return;
687 
688 	vunmap(msm_obj->vaddr);
689 	msm_obj->vaddr = NULL;
690 }
691 
692 void msm_gem_vunmap(struct drm_gem_object *obj, enum msm_gem_lock subclass)
693 {
694 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
695 
696 	mutex_lock_nested(&msm_obj->lock, subclass);
697 	msm_gem_vunmap_locked(obj);
698 	mutex_unlock(&msm_obj->lock);
699 }
700 
701 /* must be called before _move_to_active().. */
702 int msm_gem_sync_object(struct drm_gem_object *obj,
703 		struct msm_fence_context *fctx, bool exclusive)
704 {
705 	struct dma_resv_list *fobj;
706 	struct dma_fence *fence;
707 	int i, ret;
708 
709 	fobj = dma_resv_get_list(obj->resv);
710 	if (!fobj || (fobj->shared_count == 0)) {
711 		fence = dma_resv_get_excl(obj->resv);
712 		/* don't need to wait on our own fences, since ring is fifo */
713 		if (fence && (fence->context != fctx->context)) {
714 			ret = dma_fence_wait(fence, true);
715 			if (ret)
716 				return ret;
717 		}
718 	}
719 
720 	if (!exclusive || !fobj)
721 		return 0;
722 
723 	for (i = 0; i < fobj->shared_count; i++) {
724 		fence = rcu_dereference_protected(fobj->shared[i],
725 						dma_resv_held(obj->resv));
726 		if (fence->context != fctx->context) {
727 			ret = dma_fence_wait(fence, true);
728 			if (ret)
729 				return ret;
730 		}
731 	}
732 
733 	return 0;
734 }
735 
736 void msm_gem_move_to_active(struct drm_gem_object *obj,
737 		struct msm_gpu *gpu, bool exclusive, struct dma_fence *fence)
738 {
739 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
740 	WARN_ON(msm_obj->madv != MSM_MADV_WILLNEED);
741 	msm_obj->gpu = gpu;
742 	if (exclusive)
743 		dma_resv_add_excl_fence(obj->resv, fence);
744 	else
745 		dma_resv_add_shared_fence(obj->resv, fence);
746 	list_del_init(&msm_obj->mm_list);
747 	list_add_tail(&msm_obj->mm_list, &gpu->active_list);
748 }
749 
750 void msm_gem_move_to_inactive(struct drm_gem_object *obj)
751 {
752 	struct drm_device *dev = obj->dev;
753 	struct msm_drm_private *priv = dev->dev_private;
754 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
755 
756 	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
757 
758 	msm_obj->gpu = NULL;
759 	list_del_init(&msm_obj->mm_list);
760 	list_add_tail(&msm_obj->mm_list, &priv->inactive_list);
761 }
762 
763 int msm_gem_cpu_prep(struct drm_gem_object *obj, uint32_t op, ktime_t *timeout)
764 {
765 	bool write = !!(op & MSM_PREP_WRITE);
766 	unsigned long remain =
767 		op & MSM_PREP_NOSYNC ? 0 : timeout_to_jiffies(timeout);
768 	long ret;
769 
770 	ret = dma_resv_wait_timeout_rcu(obj->resv, write,
771 						  true,  remain);
772 	if (ret == 0)
773 		return remain == 0 ? -EBUSY : -ETIMEDOUT;
774 	else if (ret < 0)
775 		return ret;
776 
777 	/* TODO cache maintenance */
778 
779 	return 0;
780 }
781 
782 int msm_gem_cpu_fini(struct drm_gem_object *obj)
783 {
784 	/* TODO cache maintenance */
785 	return 0;
786 }
787 
788 #ifdef CONFIG_DEBUG_FS
789 static void describe_fence(struct dma_fence *fence, const char *type,
790 		struct seq_file *m)
791 {
792 	if (!dma_fence_is_signaled(fence))
793 		seq_printf(m, "\t%9s: %s %s seq %llu\n", type,
794 				fence->ops->get_driver_name(fence),
795 				fence->ops->get_timeline_name(fence),
796 				fence->seqno);
797 }
798 
799 void msm_gem_describe(struct drm_gem_object *obj, struct seq_file *m)
800 {
801 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
802 	struct dma_resv *robj = obj->resv;
803 	struct dma_resv_list *fobj;
804 	struct dma_fence *fence;
805 	struct msm_gem_vma *vma;
806 	uint64_t off = drm_vma_node_start(&obj->vma_node);
807 	const char *madv;
808 
809 	mutex_lock(&msm_obj->lock);
810 
811 	switch (msm_obj->madv) {
812 	case __MSM_MADV_PURGED:
813 		madv = " purged";
814 		break;
815 	case MSM_MADV_DONTNEED:
816 		madv = " purgeable";
817 		break;
818 	case MSM_MADV_WILLNEED:
819 	default:
820 		madv = "";
821 		break;
822 	}
823 
824 	seq_printf(m, "%08x: %c %2d (%2d) %08llx %p",
825 			msm_obj->flags, is_active(msm_obj) ? 'A' : 'I',
826 			obj->name, kref_read(&obj->refcount),
827 			off, msm_obj->vaddr);
828 
829 	seq_printf(m, " %08zu %9s %-32s\n", obj->size, madv, msm_obj->name);
830 
831 	if (!list_empty(&msm_obj->vmas)) {
832 
833 		seq_puts(m, "      vmas:");
834 
835 		list_for_each_entry(vma, &msm_obj->vmas, list)
836 			seq_printf(m, " [%s: %08llx,%s,inuse=%d]",
837 				vma->aspace != NULL ? vma->aspace->name : NULL,
838 				vma->iova, vma->mapped ? "mapped" : "unmapped",
839 				vma->inuse);
840 
841 		seq_puts(m, "\n");
842 	}
843 
844 	rcu_read_lock();
845 	fobj = rcu_dereference(robj->fence);
846 	if (fobj) {
847 		unsigned int i, shared_count = fobj->shared_count;
848 
849 		for (i = 0; i < shared_count; i++) {
850 			fence = rcu_dereference(fobj->shared[i]);
851 			describe_fence(fence, "Shared", m);
852 		}
853 	}
854 
855 	fence = rcu_dereference(robj->fence_excl);
856 	if (fence)
857 		describe_fence(fence, "Exclusive", m);
858 	rcu_read_unlock();
859 
860 	mutex_unlock(&msm_obj->lock);
861 }
862 
863 void msm_gem_describe_objects(struct list_head *list, struct seq_file *m)
864 {
865 	struct msm_gem_object *msm_obj;
866 	int count = 0;
867 	size_t size = 0;
868 
869 	seq_puts(m, "   flags       id ref  offset   kaddr            size     madv      name\n");
870 	list_for_each_entry(msm_obj, list, mm_list) {
871 		struct drm_gem_object *obj = &msm_obj->base;
872 		seq_puts(m, "   ");
873 		msm_gem_describe(obj, m);
874 		count++;
875 		size += obj->size;
876 	}
877 
878 	seq_printf(m, "Total %d objects, %zu bytes\n", count, size);
879 }
880 #endif
881 
882 /* don't call directly!  Use drm_gem_object_put() and friends */
883 void msm_gem_free_object(struct drm_gem_object *obj)
884 {
885 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
886 	struct drm_device *dev = obj->dev;
887 	struct msm_drm_private *priv = dev->dev_private;
888 
889 	if (llist_add(&msm_obj->freed, &priv->free_list))
890 		queue_work(priv->wq, &priv->free_work);
891 }
892 
893 static void free_object(struct msm_gem_object *msm_obj)
894 {
895 	struct drm_gem_object *obj = &msm_obj->base;
896 	struct drm_device *dev = obj->dev;
897 
898 	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
899 
900 	/* object should not be on active list: */
901 	WARN_ON(is_active(msm_obj));
902 
903 	list_del(&msm_obj->mm_list);
904 
905 	mutex_lock(&msm_obj->lock);
906 
907 	put_iova(obj);
908 
909 	if (obj->import_attach) {
910 		if (msm_obj->vaddr)
911 			dma_buf_vunmap(obj->import_attach->dmabuf, msm_obj->vaddr);
912 
913 		/* Don't drop the pages for imported dmabuf, as they are not
914 		 * ours, just free the array we allocated:
915 		 */
916 		if (msm_obj->pages)
917 			kvfree(msm_obj->pages);
918 
919 		drm_prime_gem_destroy(obj, msm_obj->sgt);
920 	} else {
921 		msm_gem_vunmap_locked(obj);
922 		put_pages(obj);
923 	}
924 
925 	drm_gem_object_release(obj);
926 
927 	mutex_unlock(&msm_obj->lock);
928 	kfree(msm_obj);
929 }
930 
931 void msm_gem_free_work(struct work_struct *work)
932 {
933 	struct msm_drm_private *priv =
934 		container_of(work, struct msm_drm_private, free_work);
935 	struct drm_device *dev = priv->dev;
936 	struct llist_node *freed;
937 	struct msm_gem_object *msm_obj, *next;
938 
939 	while ((freed = llist_del_all(&priv->free_list))) {
940 
941 		mutex_lock(&dev->struct_mutex);
942 
943 		llist_for_each_entry_safe(msm_obj, next,
944 					  freed, freed)
945 			free_object(msm_obj);
946 
947 		mutex_unlock(&dev->struct_mutex);
948 
949 		if (need_resched())
950 			break;
951 	}
952 }
953 
954 /* convenience method to construct a GEM buffer object, and userspace handle */
955 int msm_gem_new_handle(struct drm_device *dev, struct drm_file *file,
956 		uint32_t size, uint32_t flags, uint32_t *handle,
957 		char *name)
958 {
959 	struct drm_gem_object *obj;
960 	int ret;
961 
962 	obj = msm_gem_new(dev, size, flags);
963 
964 	if (IS_ERR(obj))
965 		return PTR_ERR(obj);
966 
967 	if (name)
968 		msm_gem_object_set_name(obj, "%s", name);
969 
970 	ret = drm_gem_handle_create(file, obj, handle);
971 
972 	/* drop reference from allocate - handle holds it now */
973 	drm_gem_object_put_unlocked(obj);
974 
975 	return ret;
976 }
977 
978 static int msm_gem_new_impl(struct drm_device *dev,
979 		uint32_t size, uint32_t flags,
980 		struct drm_gem_object **obj,
981 		bool struct_mutex_locked)
982 {
983 	struct msm_drm_private *priv = dev->dev_private;
984 	struct msm_gem_object *msm_obj;
985 
986 	switch (flags & MSM_BO_CACHE_MASK) {
987 	case MSM_BO_UNCACHED:
988 	case MSM_BO_CACHED:
989 	case MSM_BO_WC:
990 		break;
991 	default:
992 		DRM_DEV_ERROR(dev->dev, "invalid cache flag: %x\n",
993 				(flags & MSM_BO_CACHE_MASK));
994 		return -EINVAL;
995 	}
996 
997 	msm_obj = kzalloc(sizeof(*msm_obj), GFP_KERNEL);
998 	if (!msm_obj)
999 		return -ENOMEM;
1000 
1001 	mutex_init(&msm_obj->lock);
1002 
1003 	msm_obj->flags = flags;
1004 	msm_obj->madv = MSM_MADV_WILLNEED;
1005 
1006 	INIT_LIST_HEAD(&msm_obj->submit_entry);
1007 	INIT_LIST_HEAD(&msm_obj->vmas);
1008 
1009 	if (struct_mutex_locked) {
1010 		WARN_ON(!mutex_is_locked(&dev->struct_mutex));
1011 		list_add_tail(&msm_obj->mm_list, &priv->inactive_list);
1012 	} else {
1013 		mutex_lock(&dev->struct_mutex);
1014 		list_add_tail(&msm_obj->mm_list, &priv->inactive_list);
1015 		mutex_unlock(&dev->struct_mutex);
1016 	}
1017 
1018 	*obj = &msm_obj->base;
1019 
1020 	return 0;
1021 }
1022 
1023 static struct drm_gem_object *_msm_gem_new(struct drm_device *dev,
1024 		uint32_t size, uint32_t flags, bool struct_mutex_locked)
1025 {
1026 	struct msm_drm_private *priv = dev->dev_private;
1027 	struct drm_gem_object *obj = NULL;
1028 	bool use_vram = false;
1029 	int ret;
1030 
1031 	size = PAGE_ALIGN(size);
1032 
1033 	if (!msm_use_mmu(dev))
1034 		use_vram = true;
1035 	else if ((flags & (MSM_BO_STOLEN | MSM_BO_SCANOUT)) && priv->vram.size)
1036 		use_vram = true;
1037 
1038 	if (WARN_ON(use_vram && !priv->vram.size))
1039 		return ERR_PTR(-EINVAL);
1040 
1041 	/* Disallow zero sized objects as they make the underlying
1042 	 * infrastructure grumpy
1043 	 */
1044 	if (size == 0)
1045 		return ERR_PTR(-EINVAL);
1046 
1047 	ret = msm_gem_new_impl(dev, size, flags, &obj, struct_mutex_locked);
1048 	if (ret)
1049 		goto fail;
1050 
1051 	if (use_vram) {
1052 		struct msm_gem_vma *vma;
1053 		struct page **pages;
1054 		struct msm_gem_object *msm_obj = to_msm_bo(obj);
1055 
1056 		mutex_lock(&msm_obj->lock);
1057 
1058 		vma = add_vma(obj, NULL);
1059 		mutex_unlock(&msm_obj->lock);
1060 		if (IS_ERR(vma)) {
1061 			ret = PTR_ERR(vma);
1062 			goto fail;
1063 		}
1064 
1065 		to_msm_bo(obj)->vram_node = &vma->node;
1066 
1067 		drm_gem_private_object_init(dev, obj, size);
1068 
1069 		pages = get_pages(obj);
1070 		if (IS_ERR(pages)) {
1071 			ret = PTR_ERR(pages);
1072 			goto fail;
1073 		}
1074 
1075 		vma->iova = physaddr(obj);
1076 	} else {
1077 		ret = drm_gem_object_init(dev, obj, size);
1078 		if (ret)
1079 			goto fail;
1080 		/*
1081 		 * Our buffers are kept pinned, so allocating them from the
1082 		 * MOVABLE zone is a really bad idea, and conflicts with CMA.
1083 		 * See comments above new_inode() why this is required _and_
1084 		 * expected if you're going to pin these pages.
1085 		 */
1086 		mapping_set_gfp_mask(obj->filp->f_mapping, GFP_HIGHUSER);
1087 	}
1088 
1089 	return obj;
1090 
1091 fail:
1092 	drm_gem_object_put_unlocked(obj);
1093 	return ERR_PTR(ret);
1094 }
1095 
1096 struct drm_gem_object *msm_gem_new_locked(struct drm_device *dev,
1097 		uint32_t size, uint32_t flags)
1098 {
1099 	return _msm_gem_new(dev, size, flags, true);
1100 }
1101 
1102 struct drm_gem_object *msm_gem_new(struct drm_device *dev,
1103 		uint32_t size, uint32_t flags)
1104 {
1105 	return _msm_gem_new(dev, size, flags, false);
1106 }
1107 
1108 struct drm_gem_object *msm_gem_import(struct drm_device *dev,
1109 		struct dma_buf *dmabuf, struct sg_table *sgt)
1110 {
1111 	struct msm_gem_object *msm_obj;
1112 	struct drm_gem_object *obj;
1113 	uint32_t size;
1114 	int ret, npages;
1115 
1116 	/* if we don't have IOMMU, don't bother pretending we can import: */
1117 	if (!msm_use_mmu(dev)) {
1118 		DRM_DEV_ERROR(dev->dev, "cannot import without IOMMU\n");
1119 		return ERR_PTR(-EINVAL);
1120 	}
1121 
1122 	size = PAGE_ALIGN(dmabuf->size);
1123 
1124 	ret = msm_gem_new_impl(dev, size, MSM_BO_WC, &obj, false);
1125 	if (ret)
1126 		goto fail;
1127 
1128 	drm_gem_private_object_init(dev, obj, size);
1129 
1130 	npages = size / PAGE_SIZE;
1131 
1132 	msm_obj = to_msm_bo(obj);
1133 	mutex_lock(&msm_obj->lock);
1134 	msm_obj->sgt = sgt;
1135 	msm_obj->pages = kvmalloc_array(npages, sizeof(struct page *), GFP_KERNEL);
1136 	if (!msm_obj->pages) {
1137 		mutex_unlock(&msm_obj->lock);
1138 		ret = -ENOMEM;
1139 		goto fail;
1140 	}
1141 
1142 	ret = drm_prime_sg_to_page_addr_arrays(sgt, msm_obj->pages, NULL, npages);
1143 	if (ret) {
1144 		mutex_unlock(&msm_obj->lock);
1145 		goto fail;
1146 	}
1147 
1148 	mutex_unlock(&msm_obj->lock);
1149 	return obj;
1150 
1151 fail:
1152 	drm_gem_object_put_unlocked(obj);
1153 	return ERR_PTR(ret);
1154 }
1155 
1156 static void *_msm_gem_kernel_new(struct drm_device *dev, uint32_t size,
1157 		uint32_t flags, struct msm_gem_address_space *aspace,
1158 		struct drm_gem_object **bo, uint64_t *iova, bool locked)
1159 {
1160 	void *vaddr;
1161 	struct drm_gem_object *obj = _msm_gem_new(dev, size, flags, locked);
1162 	int ret;
1163 
1164 	if (IS_ERR(obj))
1165 		return ERR_CAST(obj);
1166 
1167 	if (iova) {
1168 		ret = msm_gem_get_and_pin_iova(obj, aspace, iova);
1169 		if (ret)
1170 			goto err;
1171 	}
1172 
1173 	vaddr = msm_gem_get_vaddr(obj);
1174 	if (IS_ERR(vaddr)) {
1175 		msm_gem_unpin_iova(obj, aspace);
1176 		ret = PTR_ERR(vaddr);
1177 		goto err;
1178 	}
1179 
1180 	if (bo)
1181 		*bo = obj;
1182 
1183 	return vaddr;
1184 err:
1185 	if (locked)
1186 		drm_gem_object_put(obj);
1187 	else
1188 		drm_gem_object_put_unlocked(obj);
1189 
1190 	return ERR_PTR(ret);
1191 
1192 }
1193 
1194 void *msm_gem_kernel_new(struct drm_device *dev, uint32_t size,
1195 		uint32_t flags, struct msm_gem_address_space *aspace,
1196 		struct drm_gem_object **bo, uint64_t *iova)
1197 {
1198 	return _msm_gem_kernel_new(dev, size, flags, aspace, bo, iova, false);
1199 }
1200 
1201 void *msm_gem_kernel_new_locked(struct drm_device *dev, uint32_t size,
1202 		uint32_t flags, struct msm_gem_address_space *aspace,
1203 		struct drm_gem_object **bo, uint64_t *iova)
1204 {
1205 	return _msm_gem_kernel_new(dev, size, flags, aspace, bo, iova, true);
1206 }
1207 
1208 void msm_gem_kernel_put(struct drm_gem_object *bo,
1209 		struct msm_gem_address_space *aspace, bool locked)
1210 {
1211 	if (IS_ERR_OR_NULL(bo))
1212 		return;
1213 
1214 	msm_gem_put_vaddr(bo);
1215 	msm_gem_unpin_iova(bo, aspace);
1216 
1217 	if (locked)
1218 		drm_gem_object_put(bo);
1219 	else
1220 		drm_gem_object_put_unlocked(bo);
1221 }
1222 
1223 void msm_gem_object_set_name(struct drm_gem_object *bo, const char *fmt, ...)
1224 {
1225 	struct msm_gem_object *msm_obj = to_msm_bo(bo);
1226 	va_list ap;
1227 
1228 	if (!fmt)
1229 		return;
1230 
1231 	va_start(ap, fmt);
1232 	vsnprintf(msm_obj->name, sizeof(msm_obj->name), fmt, ap);
1233 	va_end(ap);
1234 }
1235