1/*
2 * Copyright (C) 2013 Red Hat
3 * Author: Rob Clark <robdclark@gmail.com>
4 *
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 as published by
7 * the Free Software Foundation.
8 *
9 * This program is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
12 * more details.
13 *
14 * You should have received a copy of the GNU General Public License along with
15 * this program.  If not, see <http://www.gnu.org/licenses/>.
16 */
17
18#include <linux/spinlock.h>
19#include <linux/shmem_fs.h>
20#include <linux/dma-buf.h>
21
22#include "msm_drv.h"
23#include "msm_gem.h"
24#include "msm_gpu.h"
25#include "msm_mmu.h"
26
27static dma_addr_t physaddr(struct drm_gem_object *obj)
28{
29	struct msm_gem_object *msm_obj = to_msm_bo(obj);
30	struct msm_drm_private *priv = obj->dev->dev_private;
31	return (((dma_addr_t)msm_obj->vram_node->start) << PAGE_SHIFT) +
32			priv->vram.paddr;
33}
34
35static bool use_pages(struct drm_gem_object *obj)
36{
37	struct msm_gem_object *msm_obj = to_msm_bo(obj);
38	return !msm_obj->vram_node;
39}
40
41/* allocate pages from VRAM carveout, used when no IOMMU: */
42static struct page **get_pages_vram(struct drm_gem_object *obj,
43		int npages)
44{
45	struct msm_gem_object *msm_obj = to_msm_bo(obj);
46	struct msm_drm_private *priv = obj->dev->dev_private;
47	dma_addr_t paddr;
48	struct page **p;
49	int ret, i;
50
51	p = drm_malloc_ab(npages, sizeof(struct page *));
52	if (!p)
53		return ERR_PTR(-ENOMEM);
54
55	ret = drm_mm_insert_node(&priv->vram.mm, msm_obj->vram_node,
56			npages, 0, DRM_MM_SEARCH_DEFAULT);
57	if (ret) {
58		drm_free_large(p);
59		return ERR_PTR(ret);
60	}
61
62	paddr = physaddr(obj);
63	for (i = 0; i < npages; i++) {
64		p[i] = phys_to_page(paddr);
65		paddr += PAGE_SIZE;
66	}
67
68	return p;
69}
70
71/* called with dev->struct_mutex held */
72static struct page **get_pages(struct drm_gem_object *obj)
73{
74	struct msm_gem_object *msm_obj = to_msm_bo(obj);
75
76	if (!msm_obj->pages) {
77		struct drm_device *dev = obj->dev;
78		struct page **p;
79		int npages = obj->size >> PAGE_SHIFT;
80
81		if (use_pages(obj))
82			p = drm_gem_get_pages(obj);
83		else
84			p = get_pages_vram(obj, npages);
85
86		if (IS_ERR(p)) {
87			dev_err(dev->dev, "could not get pages: %ld\n",
88					PTR_ERR(p));
89			return p;
90		}
91
92		msm_obj->sgt = drm_prime_pages_to_sg(p, npages);
93		if (IS_ERR(msm_obj->sgt)) {
94			dev_err(dev->dev, "failed to allocate sgt\n");
95			return ERR_CAST(msm_obj->sgt);
96		}
97
98		msm_obj->pages = p;
99
100		/* For non-cached buffers, ensure the new pages are clean
101		 * because display controller, GPU, etc. are not coherent:
102		 */
103		if (msm_obj->flags & (MSM_BO_WC|MSM_BO_UNCACHED))
104			dma_map_sg(dev->dev, msm_obj->sgt->sgl,
105					msm_obj->sgt->nents, DMA_BIDIRECTIONAL);
106	}
107
108	return msm_obj->pages;
109}
110
111static void put_pages(struct drm_gem_object *obj)
112{
113	struct msm_gem_object *msm_obj = to_msm_bo(obj);
114
115	if (msm_obj->pages) {
116		/* For non-cached buffers, ensure the new pages are clean
117		 * because display controller, GPU, etc. are not coherent:
118		 */
119		if (msm_obj->flags & (MSM_BO_WC|MSM_BO_UNCACHED))
120			dma_unmap_sg(obj->dev->dev, msm_obj->sgt->sgl,
121					msm_obj->sgt->nents, DMA_BIDIRECTIONAL);
122		sg_free_table(msm_obj->sgt);
123		kfree(msm_obj->sgt);
124
125		if (use_pages(obj))
126			drm_gem_put_pages(obj, msm_obj->pages, true, false);
127		else {
128			drm_mm_remove_node(msm_obj->vram_node);
129			drm_free_large(msm_obj->pages);
130		}
131
132		msm_obj->pages = NULL;
133	}
134}
135
136struct page **msm_gem_get_pages(struct drm_gem_object *obj)
137{
138	struct drm_device *dev = obj->dev;
139	struct page **p;
140	mutex_lock(&dev->struct_mutex);
141	p = get_pages(obj);
142	mutex_unlock(&dev->struct_mutex);
143	return p;
144}
145
146void msm_gem_put_pages(struct drm_gem_object *obj)
147{
148	/* when we start tracking the pin count, then do something here */
149}
150
151int msm_gem_mmap_obj(struct drm_gem_object *obj,
152		struct vm_area_struct *vma)
153{
154	struct msm_gem_object *msm_obj = to_msm_bo(obj);
155
156	vma->vm_flags &= ~VM_PFNMAP;
157	vma->vm_flags |= VM_MIXEDMAP;
158
159	if (msm_obj->flags & MSM_BO_WC) {
160		vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
161	} else if (msm_obj->flags & MSM_BO_UNCACHED) {
162		vma->vm_page_prot = pgprot_noncached(vm_get_page_prot(vma->vm_flags));
163	} else {
164		/*
165		 * Shunt off cached objs to shmem file so they have their own
166		 * address_space (so unmap_mapping_range does what we want,
167		 * in particular in the case of mmap'd dmabufs)
168		 */
169		fput(vma->vm_file);
170		get_file(obj->filp);
171		vma->vm_pgoff = 0;
172		vma->vm_file  = obj->filp;
173
174		vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
175	}
176
177	return 0;
178}
179
180int msm_gem_mmap(struct file *filp, struct vm_area_struct *vma)
181{
182	int ret;
183
184	ret = drm_gem_mmap(filp, vma);
185	if (ret) {
186		DBG("mmap failed: %d", ret);
187		return ret;
188	}
189
190	return msm_gem_mmap_obj(vma->vm_private_data, vma);
191}
192
193int msm_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
194{
195	struct drm_gem_object *obj = vma->vm_private_data;
196	struct drm_device *dev = obj->dev;
197	struct page **pages;
198	unsigned long pfn;
199	pgoff_t pgoff;
200	int ret;
201
202	/* Make sure we don't parallel update on a fault, nor move or remove
203	 * something from beneath our feet
204	 */
205	ret = mutex_lock_interruptible(&dev->struct_mutex);
206	if (ret)
207		goto out;
208
209	/* make sure we have pages attached now */
210	pages = get_pages(obj);
211	if (IS_ERR(pages)) {
212		ret = PTR_ERR(pages);
213		goto out_unlock;
214	}
215
216	/* We don't use vmf->pgoff since that has the fake offset: */
217	pgoff = ((unsigned long)vmf->virtual_address -
218			vma->vm_start) >> PAGE_SHIFT;
219
220	pfn = page_to_pfn(pages[pgoff]);
221
222	VERB("Inserting %p pfn %lx, pa %lx", vmf->virtual_address,
223			pfn, pfn << PAGE_SHIFT);
224
225	ret = vm_insert_mixed(vma, (unsigned long)vmf->virtual_address, pfn);
226
227out_unlock:
228	mutex_unlock(&dev->struct_mutex);
229out:
230	switch (ret) {
231	case -EAGAIN:
232	case 0:
233	case -ERESTARTSYS:
234	case -EINTR:
235	case -EBUSY:
236		/*
237		 * EBUSY is ok: this just means that another thread
238		 * already did the job.
239		 */
240		return VM_FAULT_NOPAGE;
241	case -ENOMEM:
242		return VM_FAULT_OOM;
243	default:
244		return VM_FAULT_SIGBUS;
245	}
246}
247
248/** get mmap offset */
249static uint64_t mmap_offset(struct drm_gem_object *obj)
250{
251	struct drm_device *dev = obj->dev;
252	int ret;
253
254	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
255
256	/* Make it mmapable */
257	ret = drm_gem_create_mmap_offset(obj);
258
259	if (ret) {
260		dev_err(dev->dev, "could not allocate mmap offset\n");
261		return 0;
262	}
263
264	return drm_vma_node_offset_addr(&obj->vma_node);
265}
266
267uint64_t msm_gem_mmap_offset(struct drm_gem_object *obj)
268{
269	uint64_t offset;
270	mutex_lock(&obj->dev->struct_mutex);
271	offset = mmap_offset(obj);
272	mutex_unlock(&obj->dev->struct_mutex);
273	return offset;
274}
275
276/* should be called under struct_mutex.. although it can be called
277 * from atomic context without struct_mutex to acquire an extra
278 * iova ref if you know one is already held.
279 *
280 * That means when I do eventually need to add support for unpinning
281 * the refcnt counter needs to be atomic_t.
282 */
283int msm_gem_get_iova_locked(struct drm_gem_object *obj, int id,
284		uint32_t *iova)
285{
286	struct msm_gem_object *msm_obj = to_msm_bo(obj);
287	int ret = 0;
288
289	if (!msm_obj->domain[id].iova) {
290		struct msm_drm_private *priv = obj->dev->dev_private;
291		struct page **pages = get_pages(obj);
292
293		if (IS_ERR(pages))
294			return PTR_ERR(pages);
295
296		if (iommu_present(&platform_bus_type)) {
297			struct msm_mmu *mmu = priv->mmus[id];
298			uint32_t offset;
299
300			if (WARN_ON(!mmu))
301				return -EINVAL;
302
303			offset = (uint32_t)mmap_offset(obj);
304			ret = mmu->funcs->map(mmu, offset, msm_obj->sgt,
305					obj->size, IOMMU_READ | IOMMU_WRITE);
306			msm_obj->domain[id].iova = offset;
307		} else {
308			msm_obj->domain[id].iova = physaddr(obj);
309		}
310	}
311
312	if (!ret)
313		*iova = msm_obj->domain[id].iova;
314
315	return ret;
316}
317
318/* get iova, taking a reference.  Should have a matching put */
319int msm_gem_get_iova(struct drm_gem_object *obj, int id, uint32_t *iova)
320{
321	struct msm_gem_object *msm_obj = to_msm_bo(obj);
322	int ret;
323
324	/* this is safe right now because we don't unmap until the
325	 * bo is deleted:
326	 */
327	if (msm_obj->domain[id].iova) {
328		*iova = msm_obj->domain[id].iova;
329		return 0;
330	}
331
332	mutex_lock(&obj->dev->struct_mutex);
333	ret = msm_gem_get_iova_locked(obj, id, iova);
334	mutex_unlock(&obj->dev->struct_mutex);
335	return ret;
336}
337
338/* get iova without taking a reference, used in places where you have
339 * already done a 'msm_gem_get_iova()'.
340 */
341uint32_t msm_gem_iova(struct drm_gem_object *obj, int id)
342{
343	struct msm_gem_object *msm_obj = to_msm_bo(obj);
344	WARN_ON(!msm_obj->domain[id].iova);
345	return msm_obj->domain[id].iova;
346}
347
348void msm_gem_put_iova(struct drm_gem_object *obj, int id)
349{
350	// XXX TODO ..
351	// NOTE: probably don't need a _locked() version.. we wouldn't
352	// normally unmap here, but instead just mark that it could be
353	// unmapped (if the iova refcnt drops to zero), but then later
354	// if another _get_iova_locked() fails we can start unmapping
355	// things that are no longer needed..
356}
357
358int msm_gem_dumb_create(struct drm_file *file, struct drm_device *dev,
359		struct drm_mode_create_dumb *args)
360{
361	args->pitch = align_pitch(args->width, args->bpp);
362	args->size  = PAGE_ALIGN(args->pitch * args->height);
363	return msm_gem_new_handle(dev, file, args->size,
364			MSM_BO_SCANOUT | MSM_BO_WC, &args->handle);
365}
366
367int msm_gem_dumb_map_offset(struct drm_file *file, struct drm_device *dev,
368		uint32_t handle, uint64_t *offset)
369{
370	struct drm_gem_object *obj;
371	int ret = 0;
372
373	/* GEM does all our handle to object mapping */
374	obj = drm_gem_object_lookup(dev, file, handle);
375	if (obj == NULL) {
376		ret = -ENOENT;
377		goto fail;
378	}
379
380	*offset = msm_gem_mmap_offset(obj);
381
382	drm_gem_object_unreference_unlocked(obj);
383
384fail:
385	return ret;
386}
387
388void *msm_gem_vaddr_locked(struct drm_gem_object *obj)
389{
390	struct msm_gem_object *msm_obj = to_msm_bo(obj);
391	WARN_ON(!mutex_is_locked(&obj->dev->struct_mutex));
392	if (!msm_obj->vaddr) {
393		struct page **pages = get_pages(obj);
394		if (IS_ERR(pages))
395			return ERR_CAST(pages);
396		msm_obj->vaddr = vmap(pages, obj->size >> PAGE_SHIFT,
397				VM_MAP, pgprot_writecombine(PAGE_KERNEL));
398	}
399	return msm_obj->vaddr;
400}
401
402void *msm_gem_vaddr(struct drm_gem_object *obj)
403{
404	void *ret;
405	mutex_lock(&obj->dev->struct_mutex);
406	ret = msm_gem_vaddr_locked(obj);
407	mutex_unlock(&obj->dev->struct_mutex);
408	return ret;
409}
410
411/* setup callback for when bo is no longer busy..
412 * TODO probably want to differentiate read vs write..
413 */
414int msm_gem_queue_inactive_cb(struct drm_gem_object *obj,
415		struct msm_fence_cb *cb)
416{
417	struct msm_gem_object *msm_obj = to_msm_bo(obj);
418	uint32_t fence = msm_gem_fence(msm_obj,
419			MSM_PREP_READ | MSM_PREP_WRITE);
420	return msm_queue_fence_cb(obj->dev, cb, fence);
421}
422
423void msm_gem_move_to_active(struct drm_gem_object *obj,
424		struct msm_gpu *gpu, bool write, uint32_t fence)
425{
426	struct msm_gem_object *msm_obj = to_msm_bo(obj);
427	msm_obj->gpu = gpu;
428	if (write)
429		msm_obj->write_fence = fence;
430	else
431		msm_obj->read_fence = fence;
432	list_del_init(&msm_obj->mm_list);
433	list_add_tail(&msm_obj->mm_list, &gpu->active_list);
434}
435
436void msm_gem_move_to_inactive(struct drm_gem_object *obj)
437{
438	struct drm_device *dev = obj->dev;
439	struct msm_drm_private *priv = dev->dev_private;
440	struct msm_gem_object *msm_obj = to_msm_bo(obj);
441
442	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
443
444	msm_obj->gpu = NULL;
445	msm_obj->read_fence = 0;
446	msm_obj->write_fence = 0;
447	list_del_init(&msm_obj->mm_list);
448	list_add_tail(&msm_obj->mm_list, &priv->inactive_list);
449}
450
451int msm_gem_cpu_prep(struct drm_gem_object *obj, uint32_t op, ktime_t *timeout)
452{
453	struct drm_device *dev = obj->dev;
454	struct msm_gem_object *msm_obj = to_msm_bo(obj);
455	int ret = 0;
456
457	if (is_active(msm_obj)) {
458		uint32_t fence = msm_gem_fence(msm_obj, op);
459
460		if (op & MSM_PREP_NOSYNC)
461			timeout = NULL;
462
463		ret = msm_wait_fence(dev, fence, timeout, true);
464	}
465
466	/* TODO cache maintenance */
467
468	return ret;
469}
470
471int msm_gem_cpu_fini(struct drm_gem_object *obj)
472{
473	/* TODO cache maintenance */
474	return 0;
475}
476
477#ifdef CONFIG_DEBUG_FS
478void msm_gem_describe(struct drm_gem_object *obj, struct seq_file *m)
479{
480	struct drm_device *dev = obj->dev;
481	struct msm_gem_object *msm_obj = to_msm_bo(obj);
482	uint64_t off = drm_vma_node_start(&obj->vma_node);
483
484	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
485	seq_printf(m, "%08x: %c(r=%u,w=%u) %2d (%2d) %08llx %p %zu\n",
486			msm_obj->flags, is_active(msm_obj) ? 'A' : 'I',
487			msm_obj->read_fence, msm_obj->write_fence,
488			obj->name, obj->refcount.refcount.counter,
489			off, msm_obj->vaddr, obj->size);
490}
491
492void msm_gem_describe_objects(struct list_head *list, struct seq_file *m)
493{
494	struct msm_gem_object *msm_obj;
495	int count = 0;
496	size_t size = 0;
497
498	list_for_each_entry(msm_obj, list, mm_list) {
499		struct drm_gem_object *obj = &msm_obj->base;
500		seq_printf(m, "   ");
501		msm_gem_describe(obj, m);
502		count++;
503		size += obj->size;
504	}
505
506	seq_printf(m, "Total %d objects, %zu bytes\n", count, size);
507}
508#endif
509
510void msm_gem_free_object(struct drm_gem_object *obj)
511{
512	struct drm_device *dev = obj->dev;
513	struct msm_drm_private *priv = obj->dev->dev_private;
514	struct msm_gem_object *msm_obj = to_msm_bo(obj);
515	int id;
516
517	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
518
519	/* object should not be on active list: */
520	WARN_ON(is_active(msm_obj));
521
522	list_del(&msm_obj->mm_list);
523
524	for (id = 0; id < ARRAY_SIZE(msm_obj->domain); id++) {
525		struct msm_mmu *mmu = priv->mmus[id];
526		if (mmu && msm_obj->domain[id].iova) {
527			uint32_t offset = msm_obj->domain[id].iova;
528			mmu->funcs->unmap(mmu, offset, msm_obj->sgt, obj->size);
529		}
530	}
531
532	if (obj->import_attach) {
533		if (msm_obj->vaddr)
534			dma_buf_vunmap(obj->import_attach->dmabuf, msm_obj->vaddr);
535
536		/* Don't drop the pages for imported dmabuf, as they are not
537		 * ours, just free the array we allocated:
538		 */
539		if (msm_obj->pages)
540			drm_free_large(msm_obj->pages);
541
542		drm_prime_gem_destroy(obj, msm_obj->sgt);
543	} else {
544		vunmap(msm_obj->vaddr);
545		put_pages(obj);
546	}
547
548	if (msm_obj->resv == &msm_obj->_resv)
549		reservation_object_fini(msm_obj->resv);
550
551	drm_gem_object_release(obj);
552
553	kfree(msm_obj);
554}
555
556/* convenience method to construct a GEM buffer object, and userspace handle */
557int msm_gem_new_handle(struct drm_device *dev, struct drm_file *file,
558		uint32_t size, uint32_t flags, uint32_t *handle)
559{
560	struct drm_gem_object *obj;
561	int ret;
562
563	ret = mutex_lock_interruptible(&dev->struct_mutex);
564	if (ret)
565		return ret;
566
567	obj = msm_gem_new(dev, size, flags);
568
569	mutex_unlock(&dev->struct_mutex);
570
571	if (IS_ERR(obj))
572		return PTR_ERR(obj);
573
574	ret = drm_gem_handle_create(file, obj, handle);
575
576	/* drop reference from allocate - handle holds it now */
577	drm_gem_object_unreference_unlocked(obj);
578
579	return ret;
580}
581
582static int msm_gem_new_impl(struct drm_device *dev,
583		uint32_t size, uint32_t flags,
584		struct drm_gem_object **obj)
585{
586	struct msm_drm_private *priv = dev->dev_private;
587	struct msm_gem_object *msm_obj;
588	unsigned sz;
589	bool use_vram = false;
590
591	switch (flags & MSM_BO_CACHE_MASK) {
592	case MSM_BO_UNCACHED:
593	case MSM_BO_CACHED:
594	case MSM_BO_WC:
595		break;
596	default:
597		dev_err(dev->dev, "invalid cache flag: %x\n",
598				(flags & MSM_BO_CACHE_MASK));
599		return -EINVAL;
600	}
601
602	if (!iommu_present(&platform_bus_type))
603		use_vram = true;
604	else if ((flags & MSM_BO_STOLEN) && priv->vram.size)
605		use_vram = true;
606
607	if (WARN_ON(use_vram && !priv->vram.size))
608		return -EINVAL;
609
610	sz = sizeof(*msm_obj);
611	if (use_vram)
612		sz += sizeof(struct drm_mm_node);
613
614	msm_obj = kzalloc(sz, GFP_KERNEL);
615	if (!msm_obj)
616		return -ENOMEM;
617
618	if (use_vram)
619		msm_obj->vram_node = (void *)&msm_obj[1];
620
621	msm_obj->flags = flags;
622
623	msm_obj->resv = &msm_obj->_resv;
624	reservation_object_init(msm_obj->resv);
625
626	INIT_LIST_HEAD(&msm_obj->submit_entry);
627	list_add_tail(&msm_obj->mm_list, &priv->inactive_list);
628
629	*obj = &msm_obj->base;
630
631	return 0;
632}
633
634struct drm_gem_object *msm_gem_new(struct drm_device *dev,
635		uint32_t size, uint32_t flags)
636{
637	struct drm_gem_object *obj = NULL;
638	int ret;
639
640	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
641
642	size = PAGE_ALIGN(size);
643
644	ret = msm_gem_new_impl(dev, size, flags, &obj);
645	if (ret)
646		goto fail;
647
648	if (use_pages(obj)) {
649		ret = drm_gem_object_init(dev, obj, size);
650		if (ret)
651			goto fail;
652	} else {
653		drm_gem_private_object_init(dev, obj, size);
654	}
655
656	return obj;
657
658fail:
659	if (obj)
660		drm_gem_object_unreference(obj);
661
662	return ERR_PTR(ret);
663}
664
665struct drm_gem_object *msm_gem_import(struct drm_device *dev,
666		uint32_t size, struct sg_table *sgt)
667{
668	struct msm_gem_object *msm_obj;
669	struct drm_gem_object *obj;
670	int ret, npages;
671
672	/* if we don't have IOMMU, don't bother pretending we can import: */
673	if (!iommu_present(&platform_bus_type)) {
674		dev_err(dev->dev, "cannot import without IOMMU\n");
675		return ERR_PTR(-EINVAL);
676	}
677
678	size = PAGE_ALIGN(size);
679
680	ret = msm_gem_new_impl(dev, size, MSM_BO_WC, &obj);
681	if (ret)
682		goto fail;
683
684	drm_gem_private_object_init(dev, obj, size);
685
686	npages = size / PAGE_SIZE;
687
688	msm_obj = to_msm_bo(obj);
689	msm_obj->sgt = sgt;
690	msm_obj->pages = drm_malloc_ab(npages, sizeof(struct page *));
691	if (!msm_obj->pages) {
692		ret = -ENOMEM;
693		goto fail;
694	}
695
696	ret = drm_prime_sg_to_page_addr_arrays(sgt, msm_obj->pages, NULL, npages);
697	if (ret)
698		goto fail;
699
700	return obj;
701
702fail:
703	if (obj)
704		drm_gem_object_unreference_unlocked(obj);
705
706	return ERR_PTR(ret);
707}
708