Loading...
Note: File does not exist in v3.1.
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/dma-map-ops.h>
8#include <linux/vmalloc.h>
9#include <linux/spinlock.h>
10#include <linux/shmem_fs.h>
11#include <linux/dma-buf.h>
12#include <linux/pfn_t.h>
13
14#include <drm/drm_prime.h>
15#include <drm/drm_file.h>
16
17#include <trace/events/gpu_mem.h>
18
19#include "msm_drv.h"
20#include "msm_fence.h"
21#include "msm_gem.h"
22#include "msm_gpu.h"
23#include "msm_mmu.h"
24
25static dma_addr_t physaddr(struct drm_gem_object *obj)
26{
27 struct msm_gem_object *msm_obj = to_msm_bo(obj);
28 struct msm_drm_private *priv = obj->dev->dev_private;
29 return (((dma_addr_t)msm_obj->vram_node->start) << PAGE_SHIFT) +
30 priv->vram.paddr;
31}
32
33static bool use_pages(struct drm_gem_object *obj)
34{
35 struct msm_gem_object *msm_obj = to_msm_bo(obj);
36 return !msm_obj->vram_node;
37}
38
39static void update_device_mem(struct msm_drm_private *priv, ssize_t size)
40{
41 uint64_t total_mem = atomic64_add_return(size, &priv->total_mem);
42 trace_gpu_mem_total(0, 0, total_mem);
43}
44
45static void update_ctx_mem(struct drm_file *file, ssize_t size)
46{
47 struct msm_file_private *ctx = file->driver_priv;
48 uint64_t ctx_mem = atomic64_add_return(size, &ctx->ctx_mem);
49
50 rcu_read_lock(); /* Locks file->pid! */
51 trace_gpu_mem_total(0, pid_nr(rcu_dereference(file->pid)), ctx_mem);
52 rcu_read_unlock();
53
54}
55
56static int msm_gem_open(struct drm_gem_object *obj, struct drm_file *file)
57{
58 update_ctx_mem(file, obj->size);
59 return 0;
60}
61
62static void msm_gem_close(struct drm_gem_object *obj, struct drm_file *file)
63{
64 update_ctx_mem(file, -obj->size);
65}
66
67/*
68 * Cache sync.. this is a bit over-complicated, to fit dma-mapping
69 * API. Really GPU cache is out of scope here (handled on cmdstream)
70 * and all we need to do is invalidate newly allocated pages before
71 * mapping to CPU as uncached/writecombine.
72 *
73 * On top of this, we have the added headache, that depending on
74 * display generation, the display's iommu may be wired up to either
75 * the toplevel drm device (mdss), or to the mdp sub-node, meaning
76 * that here we either have dma-direct or iommu ops.
77 *
78 * Let this be a cautionary tail of abstraction gone wrong.
79 */
80
81static void sync_for_device(struct msm_gem_object *msm_obj)
82{
83 struct device *dev = msm_obj->base.dev->dev;
84
85 dma_map_sgtable(dev, msm_obj->sgt, DMA_BIDIRECTIONAL, 0);
86}
87
88static void sync_for_cpu(struct msm_gem_object *msm_obj)
89{
90 struct device *dev = msm_obj->base.dev->dev;
91
92 dma_unmap_sgtable(dev, msm_obj->sgt, DMA_BIDIRECTIONAL, 0);
93}
94
95static void update_lru_active(struct drm_gem_object *obj)
96{
97 struct msm_drm_private *priv = obj->dev->dev_private;
98 struct msm_gem_object *msm_obj = to_msm_bo(obj);
99
100 GEM_WARN_ON(!msm_obj->pages);
101
102 if (msm_obj->pin_count) {
103 drm_gem_lru_move_tail_locked(&priv->lru.pinned, obj);
104 } else if (msm_obj->madv == MSM_MADV_WILLNEED) {
105 drm_gem_lru_move_tail_locked(&priv->lru.willneed, obj);
106 } else {
107 GEM_WARN_ON(msm_obj->madv != MSM_MADV_DONTNEED);
108
109 drm_gem_lru_move_tail_locked(&priv->lru.dontneed, obj);
110 }
111}
112
113static void update_lru_locked(struct drm_gem_object *obj)
114{
115 struct msm_drm_private *priv = obj->dev->dev_private;
116 struct msm_gem_object *msm_obj = to_msm_bo(obj);
117
118 msm_gem_assert_locked(&msm_obj->base);
119
120 if (!msm_obj->pages) {
121 GEM_WARN_ON(msm_obj->pin_count);
122
123 drm_gem_lru_move_tail_locked(&priv->lru.unbacked, obj);
124 } else {
125 update_lru_active(obj);
126 }
127}
128
129static void update_lru(struct drm_gem_object *obj)
130{
131 struct msm_drm_private *priv = obj->dev->dev_private;
132
133 mutex_lock(&priv->lru.lock);
134 update_lru_locked(obj);
135 mutex_unlock(&priv->lru.lock);
136}
137
138/* allocate pages from VRAM carveout, used when no IOMMU: */
139static struct page **get_pages_vram(struct drm_gem_object *obj, int npages)
140{
141 struct msm_gem_object *msm_obj = to_msm_bo(obj);
142 struct msm_drm_private *priv = obj->dev->dev_private;
143 dma_addr_t paddr;
144 struct page **p;
145 int ret, i;
146
147 p = kvmalloc_array(npages, sizeof(struct page *), GFP_KERNEL);
148 if (!p)
149 return ERR_PTR(-ENOMEM);
150
151 spin_lock(&priv->vram.lock);
152 ret = drm_mm_insert_node(&priv->vram.mm, msm_obj->vram_node, npages);
153 spin_unlock(&priv->vram.lock);
154 if (ret) {
155 kvfree(p);
156 return ERR_PTR(ret);
157 }
158
159 paddr = physaddr(obj);
160 for (i = 0; i < npages; i++) {
161 p[i] = pfn_to_page(__phys_to_pfn(paddr));
162 paddr += PAGE_SIZE;
163 }
164
165 return p;
166}
167
168static struct page **get_pages(struct drm_gem_object *obj)
169{
170 struct msm_gem_object *msm_obj = to_msm_bo(obj);
171
172 msm_gem_assert_locked(obj);
173
174 if (!msm_obj->pages) {
175 struct drm_device *dev = obj->dev;
176 struct page **p;
177 int npages = obj->size >> PAGE_SHIFT;
178
179 if (use_pages(obj))
180 p = drm_gem_get_pages(obj);
181 else
182 p = get_pages_vram(obj, npages);
183
184 if (IS_ERR(p)) {
185 DRM_DEV_ERROR(dev->dev, "could not get pages: %ld\n",
186 PTR_ERR(p));
187 return p;
188 }
189
190 update_device_mem(dev->dev_private, obj->size);
191
192 msm_obj->pages = p;
193
194 msm_obj->sgt = drm_prime_pages_to_sg(obj->dev, p, npages);
195 if (IS_ERR(msm_obj->sgt)) {
196 void *ptr = ERR_CAST(msm_obj->sgt);
197
198 DRM_DEV_ERROR(dev->dev, "failed to allocate sgt\n");
199 msm_obj->sgt = NULL;
200 return ptr;
201 }
202
203 /* For non-cached buffers, ensure the new pages are clean
204 * because display controller, GPU, etc. are not coherent:
205 */
206 if (msm_obj->flags & MSM_BO_WC)
207 sync_for_device(msm_obj);
208
209 update_lru(obj);
210 }
211
212 return msm_obj->pages;
213}
214
215static void put_pages_vram(struct drm_gem_object *obj)
216{
217 struct msm_gem_object *msm_obj = to_msm_bo(obj);
218 struct msm_drm_private *priv = obj->dev->dev_private;
219
220 spin_lock(&priv->vram.lock);
221 drm_mm_remove_node(msm_obj->vram_node);
222 spin_unlock(&priv->vram.lock);
223
224 kvfree(msm_obj->pages);
225}
226
227static void put_pages(struct drm_gem_object *obj)
228{
229 struct msm_gem_object *msm_obj = to_msm_bo(obj);
230
231 if (msm_obj->pages) {
232 if (msm_obj->sgt) {
233 /* For non-cached buffers, ensure the new
234 * pages are clean because display controller,
235 * GPU, etc. are not coherent:
236 */
237 if (msm_obj->flags & MSM_BO_WC)
238 sync_for_cpu(msm_obj);
239
240 sg_free_table(msm_obj->sgt);
241 kfree(msm_obj->sgt);
242 msm_obj->sgt = NULL;
243 }
244
245 update_device_mem(obj->dev->dev_private, -obj->size);
246
247 if (use_pages(obj))
248 drm_gem_put_pages(obj, msm_obj->pages, true, false);
249 else
250 put_pages_vram(obj);
251
252 msm_obj->pages = NULL;
253 update_lru(obj);
254 }
255}
256
257static struct page **msm_gem_get_pages_locked(struct drm_gem_object *obj,
258 unsigned madv)
259{
260 struct msm_gem_object *msm_obj = to_msm_bo(obj);
261
262 msm_gem_assert_locked(obj);
263
264 if (msm_obj->madv > madv) {
265 DRM_DEV_DEBUG_DRIVER(obj->dev->dev, "Invalid madv state: %u vs %u\n",
266 msm_obj->madv, madv);
267 return ERR_PTR(-EBUSY);
268 }
269
270 return get_pages(obj);
271}
272
273/*
274 * Update the pin count of the object, call under lru.lock
275 */
276void msm_gem_pin_obj_locked(struct drm_gem_object *obj)
277{
278 struct msm_drm_private *priv = obj->dev->dev_private;
279
280 msm_gem_assert_locked(obj);
281
282 to_msm_bo(obj)->pin_count++;
283 drm_gem_lru_move_tail_locked(&priv->lru.pinned, obj);
284}
285
286static void pin_obj_locked(struct drm_gem_object *obj)
287{
288 struct msm_drm_private *priv = obj->dev->dev_private;
289
290 mutex_lock(&priv->lru.lock);
291 msm_gem_pin_obj_locked(obj);
292 mutex_unlock(&priv->lru.lock);
293}
294
295struct page **msm_gem_pin_pages_locked(struct drm_gem_object *obj)
296{
297 struct page **p;
298
299 msm_gem_assert_locked(obj);
300
301 p = msm_gem_get_pages_locked(obj, MSM_MADV_WILLNEED);
302 if (!IS_ERR(p))
303 pin_obj_locked(obj);
304
305 return p;
306}
307
308void msm_gem_unpin_pages_locked(struct drm_gem_object *obj)
309{
310 msm_gem_assert_locked(obj);
311
312 msm_gem_unpin_locked(obj);
313}
314
315static pgprot_t msm_gem_pgprot(struct msm_gem_object *msm_obj, pgprot_t prot)
316{
317 if (msm_obj->flags & MSM_BO_WC)
318 return pgprot_writecombine(prot);
319 return prot;
320}
321
322static vm_fault_t msm_gem_fault(struct vm_fault *vmf)
323{
324 struct vm_area_struct *vma = vmf->vma;
325 struct drm_gem_object *obj = vma->vm_private_data;
326 struct msm_gem_object *msm_obj = to_msm_bo(obj);
327 struct page **pages;
328 unsigned long pfn;
329 pgoff_t pgoff;
330 int err;
331 vm_fault_t ret;
332
333 /*
334 * vm_ops.open/drm_gem_mmap_obj and close get and put
335 * a reference on obj. So, we dont need to hold one here.
336 */
337 err = msm_gem_lock_interruptible(obj);
338 if (err) {
339 ret = VM_FAULT_NOPAGE;
340 goto out;
341 }
342
343 if (GEM_WARN_ON(msm_obj->madv != MSM_MADV_WILLNEED)) {
344 msm_gem_unlock(obj);
345 return VM_FAULT_SIGBUS;
346 }
347
348 /* make sure we have pages attached now */
349 pages = get_pages(obj);
350 if (IS_ERR(pages)) {
351 ret = vmf_error(PTR_ERR(pages));
352 goto out_unlock;
353 }
354
355 /* We don't use vmf->pgoff since that has the fake offset: */
356 pgoff = (vmf->address - vma->vm_start) >> PAGE_SHIFT;
357
358 pfn = page_to_pfn(pages[pgoff]);
359
360 VERB("Inserting %p pfn %lx, pa %lx", (void *)vmf->address,
361 pfn, pfn << PAGE_SHIFT);
362
363 ret = vmf_insert_pfn(vma, vmf->address, pfn);
364
365out_unlock:
366 msm_gem_unlock(obj);
367out:
368 return ret;
369}
370
371/** get mmap offset */
372static uint64_t mmap_offset(struct drm_gem_object *obj)
373{
374 struct drm_device *dev = obj->dev;
375 int ret;
376
377 msm_gem_assert_locked(obj);
378
379 /* Make it mmapable */
380 ret = drm_gem_create_mmap_offset(obj);
381
382 if (ret) {
383 DRM_DEV_ERROR(dev->dev, "could not allocate mmap offset\n");
384 return 0;
385 }
386
387 return drm_vma_node_offset_addr(&obj->vma_node);
388}
389
390uint64_t msm_gem_mmap_offset(struct drm_gem_object *obj)
391{
392 uint64_t offset;
393
394 msm_gem_lock(obj);
395 offset = mmap_offset(obj);
396 msm_gem_unlock(obj);
397 return offset;
398}
399
400static struct msm_gem_vma *add_vma(struct drm_gem_object *obj,
401 struct msm_gem_address_space *aspace)
402{
403 struct msm_gem_object *msm_obj = to_msm_bo(obj);
404 struct msm_gem_vma *vma;
405
406 msm_gem_assert_locked(obj);
407
408 vma = msm_gem_vma_new(aspace);
409 if (!vma)
410 return ERR_PTR(-ENOMEM);
411
412 list_add_tail(&vma->list, &msm_obj->vmas);
413
414 return vma;
415}
416
417static struct msm_gem_vma *lookup_vma(struct drm_gem_object *obj,
418 struct msm_gem_address_space *aspace)
419{
420 struct msm_gem_object *msm_obj = to_msm_bo(obj);
421 struct msm_gem_vma *vma;
422
423 msm_gem_assert_locked(obj);
424
425 list_for_each_entry(vma, &msm_obj->vmas, list) {
426 if (vma->aspace == aspace)
427 return vma;
428 }
429
430 return NULL;
431}
432
433static void del_vma(struct msm_gem_vma *vma)
434{
435 if (!vma)
436 return;
437
438 list_del(&vma->list);
439 kfree(vma);
440}
441
442/*
443 * If close is true, this also closes the VMA (releasing the allocated
444 * iova range) in addition to removing the iommu mapping. In the eviction
445 * case (!close), we keep the iova allocated, but only remove the iommu
446 * mapping.
447 */
448static void
449put_iova_spaces(struct drm_gem_object *obj, bool close)
450{
451 struct msm_gem_object *msm_obj = to_msm_bo(obj);
452 struct msm_gem_vma *vma;
453
454 msm_gem_assert_locked(obj);
455
456 list_for_each_entry(vma, &msm_obj->vmas, list) {
457 if (vma->aspace) {
458 msm_gem_vma_purge(vma);
459 if (close)
460 msm_gem_vma_close(vma);
461 }
462 }
463}
464
465/* Called with msm_obj locked */
466static void
467put_iova_vmas(struct drm_gem_object *obj)
468{
469 struct msm_gem_object *msm_obj = to_msm_bo(obj);
470 struct msm_gem_vma *vma, *tmp;
471
472 msm_gem_assert_locked(obj);
473
474 list_for_each_entry_safe(vma, tmp, &msm_obj->vmas, list) {
475 del_vma(vma);
476 }
477}
478
479static struct msm_gem_vma *get_vma_locked(struct drm_gem_object *obj,
480 struct msm_gem_address_space *aspace,
481 u64 range_start, u64 range_end)
482{
483 struct msm_gem_vma *vma;
484
485 msm_gem_assert_locked(obj);
486
487 vma = lookup_vma(obj, aspace);
488
489 if (!vma) {
490 int ret;
491
492 vma = add_vma(obj, aspace);
493 if (IS_ERR(vma))
494 return vma;
495
496 ret = msm_gem_vma_init(vma, obj->size,
497 range_start, range_end);
498 if (ret) {
499 del_vma(vma);
500 return ERR_PTR(ret);
501 }
502 } else {
503 GEM_WARN_ON(vma->iova < range_start);
504 GEM_WARN_ON((vma->iova + obj->size) > range_end);
505 }
506
507 return vma;
508}
509
510int msm_gem_pin_vma_locked(struct drm_gem_object *obj, struct msm_gem_vma *vma)
511{
512 struct msm_gem_object *msm_obj = to_msm_bo(obj);
513 struct page **pages;
514 int prot = IOMMU_READ;
515
516 if (!(msm_obj->flags & MSM_BO_GPU_READONLY))
517 prot |= IOMMU_WRITE;
518
519 if (msm_obj->flags & MSM_BO_MAP_PRIV)
520 prot |= IOMMU_PRIV;
521
522 if (msm_obj->flags & MSM_BO_CACHED_COHERENT)
523 prot |= IOMMU_CACHE;
524
525 msm_gem_assert_locked(obj);
526
527 pages = msm_gem_get_pages_locked(obj, MSM_MADV_WILLNEED);
528 if (IS_ERR(pages))
529 return PTR_ERR(pages);
530
531 return msm_gem_vma_map(vma, prot, msm_obj->sgt, obj->size);
532}
533
534void msm_gem_unpin_locked(struct drm_gem_object *obj)
535{
536 struct msm_drm_private *priv = obj->dev->dev_private;
537 struct msm_gem_object *msm_obj = to_msm_bo(obj);
538
539 msm_gem_assert_locked(obj);
540
541 mutex_lock(&priv->lru.lock);
542 msm_obj->pin_count--;
543 GEM_WARN_ON(msm_obj->pin_count < 0);
544 update_lru_locked(obj);
545 mutex_unlock(&priv->lru.lock);
546}
547
548/* Special unpin path for use in fence-signaling path, avoiding the need
549 * to hold the obj lock by only depending on things that a protected by
550 * the LRU lock. In particular we know that that we already have backing
551 * and and that the object's dma_resv has the fence for the current
552 * submit/job which will prevent us racing against page eviction.
553 */
554void msm_gem_unpin_active(struct drm_gem_object *obj)
555{
556 struct msm_gem_object *msm_obj = to_msm_bo(obj);
557
558 msm_obj->pin_count--;
559 GEM_WARN_ON(msm_obj->pin_count < 0);
560 update_lru_active(obj);
561}
562
563struct msm_gem_vma *msm_gem_get_vma_locked(struct drm_gem_object *obj,
564 struct msm_gem_address_space *aspace)
565{
566 return get_vma_locked(obj, aspace, 0, U64_MAX);
567}
568
569static int get_and_pin_iova_range_locked(struct drm_gem_object *obj,
570 struct msm_gem_address_space *aspace, uint64_t *iova,
571 u64 range_start, u64 range_end)
572{
573 struct msm_gem_vma *vma;
574 int ret;
575
576 msm_gem_assert_locked(obj);
577
578 vma = get_vma_locked(obj, aspace, range_start, range_end);
579 if (IS_ERR(vma))
580 return PTR_ERR(vma);
581
582 ret = msm_gem_pin_vma_locked(obj, vma);
583 if (!ret) {
584 *iova = vma->iova;
585 pin_obj_locked(obj);
586 }
587
588 return ret;
589}
590
591/*
592 * get iova and pin it. Should have a matching put
593 * limits iova to specified range (in pages)
594 */
595int msm_gem_get_and_pin_iova_range(struct drm_gem_object *obj,
596 struct msm_gem_address_space *aspace, uint64_t *iova,
597 u64 range_start, u64 range_end)
598{
599 int ret;
600
601 msm_gem_lock(obj);
602 ret = get_and_pin_iova_range_locked(obj, aspace, iova, range_start, range_end);
603 msm_gem_unlock(obj);
604
605 return ret;
606}
607
608/* get iova and pin it. Should have a matching put */
609int msm_gem_get_and_pin_iova(struct drm_gem_object *obj,
610 struct msm_gem_address_space *aspace, uint64_t *iova)
611{
612 return msm_gem_get_and_pin_iova_range(obj, aspace, iova, 0, U64_MAX);
613}
614
615/*
616 * Get an iova but don't pin it. Doesn't need a put because iovas are currently
617 * valid for the life of the object
618 */
619int msm_gem_get_iova(struct drm_gem_object *obj,
620 struct msm_gem_address_space *aspace, uint64_t *iova)
621{
622 struct msm_gem_vma *vma;
623 int ret = 0;
624
625 msm_gem_lock(obj);
626 vma = get_vma_locked(obj, aspace, 0, U64_MAX);
627 if (IS_ERR(vma)) {
628 ret = PTR_ERR(vma);
629 } else {
630 *iova = vma->iova;
631 }
632 msm_gem_unlock(obj);
633
634 return ret;
635}
636
637static int clear_iova(struct drm_gem_object *obj,
638 struct msm_gem_address_space *aspace)
639{
640 struct msm_gem_vma *vma = lookup_vma(obj, aspace);
641
642 if (!vma)
643 return 0;
644
645 msm_gem_vma_purge(vma);
646 msm_gem_vma_close(vma);
647 del_vma(vma);
648
649 return 0;
650}
651
652/*
653 * Get the requested iova but don't pin it. Fails if the requested iova is
654 * not available. Doesn't need a put because iovas are currently valid for
655 * the life of the object.
656 *
657 * Setting an iova of zero will clear the vma.
658 */
659int msm_gem_set_iova(struct drm_gem_object *obj,
660 struct msm_gem_address_space *aspace, uint64_t iova)
661{
662 int ret = 0;
663
664 msm_gem_lock(obj);
665 if (!iova) {
666 ret = clear_iova(obj, aspace);
667 } else {
668 struct msm_gem_vma *vma;
669 vma = get_vma_locked(obj, aspace, iova, iova + obj->size);
670 if (IS_ERR(vma)) {
671 ret = PTR_ERR(vma);
672 } else if (GEM_WARN_ON(vma->iova != iova)) {
673 clear_iova(obj, aspace);
674 ret = -EBUSY;
675 }
676 }
677 msm_gem_unlock(obj);
678
679 return ret;
680}
681
682/*
683 * Unpin a iova by updating the reference counts. The memory isn't actually
684 * purged until something else (shrinker, mm_notifier, destroy, etc) decides
685 * to get rid of it
686 */
687void msm_gem_unpin_iova(struct drm_gem_object *obj,
688 struct msm_gem_address_space *aspace)
689{
690 struct msm_gem_vma *vma;
691
692 msm_gem_lock(obj);
693 vma = lookup_vma(obj, aspace);
694 if (!GEM_WARN_ON(!vma)) {
695 msm_gem_unpin_locked(obj);
696 }
697 msm_gem_unlock(obj);
698}
699
700int msm_gem_dumb_create(struct drm_file *file, struct drm_device *dev,
701 struct drm_mode_create_dumb *args)
702{
703 args->pitch = align_pitch(args->width, args->bpp);
704 args->size = PAGE_ALIGN(args->pitch * args->height);
705 return msm_gem_new_handle(dev, file, args->size,
706 MSM_BO_SCANOUT | MSM_BO_WC, &args->handle, "dumb");
707}
708
709int msm_gem_dumb_map_offset(struct drm_file *file, struct drm_device *dev,
710 uint32_t handle, uint64_t *offset)
711{
712 struct drm_gem_object *obj;
713 int ret = 0;
714
715 /* GEM does all our handle to object mapping */
716 obj = drm_gem_object_lookup(file, handle);
717 if (obj == NULL) {
718 ret = -ENOENT;
719 goto fail;
720 }
721
722 *offset = msm_gem_mmap_offset(obj);
723
724 drm_gem_object_put(obj);
725
726fail:
727 return ret;
728}
729
730static void *get_vaddr(struct drm_gem_object *obj, unsigned madv)
731{
732 struct msm_gem_object *msm_obj = to_msm_bo(obj);
733 struct page **pages;
734 int ret = 0;
735
736 msm_gem_assert_locked(obj);
737
738 if (obj->import_attach)
739 return ERR_PTR(-ENODEV);
740
741 pages = msm_gem_get_pages_locked(obj, madv);
742 if (IS_ERR(pages))
743 return ERR_CAST(pages);
744
745 pin_obj_locked(obj);
746
747 /* increment vmap_count *before* vmap() call, so shrinker can
748 * check vmap_count (is_vunmapable()) outside of msm_obj lock.
749 * This guarantees that we won't try to msm_gem_vunmap() this
750 * same object from within the vmap() call (while we already
751 * hold msm_obj lock)
752 */
753 msm_obj->vmap_count++;
754
755 if (!msm_obj->vaddr) {
756 msm_obj->vaddr = vmap(pages, obj->size >> PAGE_SHIFT,
757 VM_MAP, msm_gem_pgprot(msm_obj, PAGE_KERNEL));
758 if (msm_obj->vaddr == NULL) {
759 ret = -ENOMEM;
760 goto fail;
761 }
762 }
763
764 return msm_obj->vaddr;
765
766fail:
767 msm_obj->vmap_count--;
768 msm_gem_unpin_locked(obj);
769 return ERR_PTR(ret);
770}
771
772void *msm_gem_get_vaddr_locked(struct drm_gem_object *obj)
773{
774 return get_vaddr(obj, MSM_MADV_WILLNEED);
775}
776
777void *msm_gem_get_vaddr(struct drm_gem_object *obj)
778{
779 void *ret;
780
781 msm_gem_lock(obj);
782 ret = msm_gem_get_vaddr_locked(obj);
783 msm_gem_unlock(obj);
784
785 return ret;
786}
787
788/*
789 * Don't use this! It is for the very special case of dumping
790 * submits from GPU hangs or faults, were the bo may already
791 * be MSM_MADV_DONTNEED, but we know the buffer is still on the
792 * active list.
793 */
794void *msm_gem_get_vaddr_active(struct drm_gem_object *obj)
795{
796 return get_vaddr(obj, __MSM_MADV_PURGED);
797}
798
799void msm_gem_put_vaddr_locked(struct drm_gem_object *obj)
800{
801 struct msm_gem_object *msm_obj = to_msm_bo(obj);
802
803 msm_gem_assert_locked(obj);
804 GEM_WARN_ON(msm_obj->vmap_count < 1);
805
806 msm_obj->vmap_count--;
807 msm_gem_unpin_locked(obj);
808}
809
810void msm_gem_put_vaddr(struct drm_gem_object *obj)
811{
812 msm_gem_lock(obj);
813 msm_gem_put_vaddr_locked(obj);
814 msm_gem_unlock(obj);
815}
816
817/* Update madvise status, returns true if not purged, else
818 * false or -errno.
819 */
820int msm_gem_madvise(struct drm_gem_object *obj, unsigned madv)
821{
822 struct msm_drm_private *priv = obj->dev->dev_private;
823 struct msm_gem_object *msm_obj = to_msm_bo(obj);
824
825 msm_gem_lock(obj);
826
827 mutex_lock(&priv->lru.lock);
828
829 if (msm_obj->madv != __MSM_MADV_PURGED)
830 msm_obj->madv = madv;
831
832 madv = msm_obj->madv;
833
834 /* If the obj is inactive, we might need to move it
835 * between inactive lists
836 */
837 update_lru_locked(obj);
838
839 mutex_unlock(&priv->lru.lock);
840
841 msm_gem_unlock(obj);
842
843 return (madv != __MSM_MADV_PURGED);
844}
845
846void msm_gem_purge(struct drm_gem_object *obj)
847{
848 struct drm_device *dev = obj->dev;
849 struct msm_drm_private *priv = obj->dev->dev_private;
850 struct msm_gem_object *msm_obj = to_msm_bo(obj);
851
852 msm_gem_assert_locked(obj);
853 GEM_WARN_ON(!is_purgeable(msm_obj));
854
855 /* Get rid of any iommu mapping(s): */
856 put_iova_spaces(obj, true);
857
858 msm_gem_vunmap(obj);
859
860 drm_vma_node_unmap(&obj->vma_node, dev->anon_inode->i_mapping);
861
862 put_pages(obj);
863
864 put_iova_vmas(obj);
865
866 mutex_lock(&priv->lru.lock);
867 /* A one-way transition: */
868 msm_obj->madv = __MSM_MADV_PURGED;
869 mutex_unlock(&priv->lru.lock);
870
871 drm_gem_free_mmap_offset(obj);
872
873 /* Our goal here is to return as much of the memory as
874 * is possible back to the system as we are called from OOM.
875 * To do this we must instruct the shmfs to drop all of its
876 * backing pages, *now*.
877 */
878 shmem_truncate_range(file_inode(obj->filp), 0, (loff_t)-1);
879
880 invalidate_mapping_pages(file_inode(obj->filp)->i_mapping,
881 0, (loff_t)-1);
882}
883
884/*
885 * Unpin the backing pages and make them available to be swapped out.
886 */
887void msm_gem_evict(struct drm_gem_object *obj)
888{
889 struct drm_device *dev = obj->dev;
890 struct msm_gem_object *msm_obj = to_msm_bo(obj);
891
892 msm_gem_assert_locked(obj);
893 GEM_WARN_ON(is_unevictable(msm_obj));
894
895 /* Get rid of any iommu mapping(s): */
896 put_iova_spaces(obj, false);
897
898 drm_vma_node_unmap(&obj->vma_node, dev->anon_inode->i_mapping);
899
900 put_pages(obj);
901}
902
903void msm_gem_vunmap(struct drm_gem_object *obj)
904{
905 struct msm_gem_object *msm_obj = to_msm_bo(obj);
906
907 msm_gem_assert_locked(obj);
908
909 if (!msm_obj->vaddr || GEM_WARN_ON(!is_vunmapable(msm_obj)))
910 return;
911
912 vunmap(msm_obj->vaddr);
913 msm_obj->vaddr = NULL;
914}
915
916bool msm_gem_active(struct drm_gem_object *obj)
917{
918 msm_gem_assert_locked(obj);
919
920 if (to_msm_bo(obj)->pin_count)
921 return true;
922
923 return !dma_resv_test_signaled(obj->resv, dma_resv_usage_rw(true));
924}
925
926int msm_gem_cpu_prep(struct drm_gem_object *obj, uint32_t op, ktime_t *timeout)
927{
928 bool write = !!(op & MSM_PREP_WRITE);
929 unsigned long remain =
930 op & MSM_PREP_NOSYNC ? 0 : timeout_to_jiffies(timeout);
931 long ret;
932
933 if (op & MSM_PREP_BOOST) {
934 dma_resv_set_deadline(obj->resv, dma_resv_usage_rw(write),
935 ktime_get());
936 }
937
938 ret = dma_resv_wait_timeout(obj->resv, dma_resv_usage_rw(write),
939 true, remain);
940 if (ret == 0)
941 return remain == 0 ? -EBUSY : -ETIMEDOUT;
942 else if (ret < 0)
943 return ret;
944
945 /* TODO cache maintenance */
946
947 return 0;
948}
949
950int msm_gem_cpu_fini(struct drm_gem_object *obj)
951{
952 /* TODO cache maintenance */
953 return 0;
954}
955
956#ifdef CONFIG_DEBUG_FS
957void msm_gem_describe(struct drm_gem_object *obj, struct seq_file *m,
958 struct msm_gem_stats *stats)
959{
960 struct msm_gem_object *msm_obj = to_msm_bo(obj);
961 struct dma_resv *robj = obj->resv;
962 struct msm_gem_vma *vma;
963 uint64_t off = drm_vma_node_start(&obj->vma_node);
964 const char *madv;
965
966 msm_gem_lock(obj);
967
968 stats->all.count++;
969 stats->all.size += obj->size;
970
971 if (msm_gem_active(obj)) {
972 stats->active.count++;
973 stats->active.size += obj->size;
974 }
975
976 if (msm_obj->pages) {
977 stats->resident.count++;
978 stats->resident.size += obj->size;
979 }
980
981 switch (msm_obj->madv) {
982 case __MSM_MADV_PURGED:
983 stats->purged.count++;
984 stats->purged.size += obj->size;
985 madv = " purged";
986 break;
987 case MSM_MADV_DONTNEED:
988 stats->purgeable.count++;
989 stats->purgeable.size += obj->size;
990 madv = " purgeable";
991 break;
992 case MSM_MADV_WILLNEED:
993 default:
994 madv = "";
995 break;
996 }
997
998 seq_printf(m, "%08x: %c %2d (%2d) %08llx %p",
999 msm_obj->flags, msm_gem_active(obj) ? 'A' : 'I',
1000 obj->name, kref_read(&obj->refcount),
1001 off, msm_obj->vaddr);
1002
1003 seq_printf(m, " %08zu %9s %-32s\n", obj->size, madv, msm_obj->name);
1004
1005 if (!list_empty(&msm_obj->vmas)) {
1006
1007 seq_puts(m, " vmas:");
1008
1009 list_for_each_entry(vma, &msm_obj->vmas, list) {
1010 const char *name, *comm;
1011 if (vma->aspace) {
1012 struct msm_gem_address_space *aspace = vma->aspace;
1013 struct task_struct *task =
1014 get_pid_task(aspace->pid, PIDTYPE_PID);
1015 if (task) {
1016 comm = kstrdup(task->comm, GFP_KERNEL);
1017 put_task_struct(task);
1018 } else {
1019 comm = NULL;
1020 }
1021 name = aspace->name;
1022 } else {
1023 name = comm = NULL;
1024 }
1025 seq_printf(m, " [%s%s%s: aspace=%p, %08llx,%s]",
1026 name, comm ? ":" : "", comm ? comm : "",
1027 vma->aspace, vma->iova,
1028 vma->mapped ? "mapped" : "unmapped");
1029 kfree(comm);
1030 }
1031
1032 seq_puts(m, "\n");
1033 }
1034
1035 dma_resv_describe(robj, m);
1036 msm_gem_unlock(obj);
1037}
1038
1039void msm_gem_describe_objects(struct list_head *list, struct seq_file *m)
1040{
1041 struct msm_gem_stats stats = {};
1042 struct msm_gem_object *msm_obj;
1043
1044 seq_puts(m, " flags id ref offset kaddr size madv name\n");
1045 list_for_each_entry(msm_obj, list, node) {
1046 struct drm_gem_object *obj = &msm_obj->base;
1047 seq_puts(m, " ");
1048 msm_gem_describe(obj, m, &stats);
1049 }
1050
1051 seq_printf(m, "Total: %4d objects, %9zu bytes\n",
1052 stats.all.count, stats.all.size);
1053 seq_printf(m, "Active: %4d objects, %9zu bytes\n",
1054 stats.active.count, stats.active.size);
1055 seq_printf(m, "Resident: %4d objects, %9zu bytes\n",
1056 stats.resident.count, stats.resident.size);
1057 seq_printf(m, "Purgeable: %4d objects, %9zu bytes\n",
1058 stats.purgeable.count, stats.purgeable.size);
1059 seq_printf(m, "Purged: %4d objects, %9zu bytes\n",
1060 stats.purged.count, stats.purged.size);
1061}
1062#endif
1063
1064/* don't call directly! Use drm_gem_object_put() */
1065static void msm_gem_free_object(struct drm_gem_object *obj)
1066{
1067 struct msm_gem_object *msm_obj = to_msm_bo(obj);
1068 struct drm_device *dev = obj->dev;
1069 struct msm_drm_private *priv = dev->dev_private;
1070
1071 mutex_lock(&priv->obj_lock);
1072 list_del(&msm_obj->node);
1073 mutex_unlock(&priv->obj_lock);
1074
1075 put_iova_spaces(obj, true);
1076
1077 if (obj->import_attach) {
1078 GEM_WARN_ON(msm_obj->vaddr);
1079
1080 /* Don't drop the pages for imported dmabuf, as they are not
1081 * ours, just free the array we allocated:
1082 */
1083 kvfree(msm_obj->pages);
1084
1085 put_iova_vmas(obj);
1086
1087 drm_prime_gem_destroy(obj, msm_obj->sgt);
1088 } else {
1089 msm_gem_vunmap(obj);
1090 put_pages(obj);
1091 put_iova_vmas(obj);
1092 }
1093
1094 drm_gem_object_release(obj);
1095
1096 kfree(msm_obj->metadata);
1097 kfree(msm_obj);
1098}
1099
1100static int msm_gem_object_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma)
1101{
1102 struct msm_gem_object *msm_obj = to_msm_bo(obj);
1103
1104 vm_flags_set(vma, VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP);
1105 vma->vm_page_prot = msm_gem_pgprot(msm_obj, vm_get_page_prot(vma->vm_flags));
1106
1107 return 0;
1108}
1109
1110/* convenience method to construct a GEM buffer object, and userspace handle */
1111int msm_gem_new_handle(struct drm_device *dev, struct drm_file *file,
1112 uint32_t size, uint32_t flags, uint32_t *handle,
1113 char *name)
1114{
1115 struct drm_gem_object *obj;
1116 int ret;
1117
1118 obj = msm_gem_new(dev, size, flags);
1119
1120 if (IS_ERR(obj))
1121 return PTR_ERR(obj);
1122
1123 if (name)
1124 msm_gem_object_set_name(obj, "%s", name);
1125
1126 ret = drm_gem_handle_create(file, obj, handle);
1127
1128 /* drop reference from allocate - handle holds it now */
1129 drm_gem_object_put(obj);
1130
1131 return ret;
1132}
1133
1134static enum drm_gem_object_status msm_gem_status(struct drm_gem_object *obj)
1135{
1136 struct msm_gem_object *msm_obj = to_msm_bo(obj);
1137 enum drm_gem_object_status status = 0;
1138
1139 if (msm_obj->pages)
1140 status |= DRM_GEM_OBJECT_RESIDENT;
1141
1142 if (msm_obj->madv == MSM_MADV_DONTNEED)
1143 status |= DRM_GEM_OBJECT_PURGEABLE;
1144
1145 return status;
1146}
1147
1148static const struct vm_operations_struct vm_ops = {
1149 .fault = msm_gem_fault,
1150 .open = drm_gem_vm_open,
1151 .close = drm_gem_vm_close,
1152};
1153
1154static const struct drm_gem_object_funcs msm_gem_object_funcs = {
1155 .free = msm_gem_free_object,
1156 .open = msm_gem_open,
1157 .close = msm_gem_close,
1158 .pin = msm_gem_prime_pin,
1159 .unpin = msm_gem_prime_unpin,
1160 .get_sg_table = msm_gem_prime_get_sg_table,
1161 .vmap = msm_gem_prime_vmap,
1162 .vunmap = msm_gem_prime_vunmap,
1163 .mmap = msm_gem_object_mmap,
1164 .status = msm_gem_status,
1165 .vm_ops = &vm_ops,
1166};
1167
1168static int msm_gem_new_impl(struct drm_device *dev,
1169 uint32_t size, uint32_t flags,
1170 struct drm_gem_object **obj)
1171{
1172 struct msm_drm_private *priv = dev->dev_private;
1173 struct msm_gem_object *msm_obj;
1174
1175 switch (flags & MSM_BO_CACHE_MASK) {
1176 case MSM_BO_CACHED:
1177 case MSM_BO_WC:
1178 break;
1179 case MSM_BO_CACHED_COHERENT:
1180 if (priv->has_cached_coherent)
1181 break;
1182 fallthrough;
1183 default:
1184 DRM_DEV_DEBUG(dev->dev, "invalid cache flag: %x\n",
1185 (flags & MSM_BO_CACHE_MASK));
1186 return -EINVAL;
1187 }
1188
1189 msm_obj = kzalloc(sizeof(*msm_obj), GFP_KERNEL);
1190 if (!msm_obj)
1191 return -ENOMEM;
1192
1193 msm_obj->flags = flags;
1194 msm_obj->madv = MSM_MADV_WILLNEED;
1195
1196 INIT_LIST_HEAD(&msm_obj->node);
1197 INIT_LIST_HEAD(&msm_obj->vmas);
1198
1199 *obj = &msm_obj->base;
1200 (*obj)->funcs = &msm_gem_object_funcs;
1201
1202 return 0;
1203}
1204
1205struct drm_gem_object *msm_gem_new(struct drm_device *dev, uint32_t size, uint32_t flags)
1206{
1207 struct msm_drm_private *priv = dev->dev_private;
1208 struct msm_gem_object *msm_obj;
1209 struct drm_gem_object *obj = NULL;
1210 bool use_vram = false;
1211 int ret;
1212
1213 size = PAGE_ALIGN(size);
1214
1215 if (!msm_use_mmu(dev))
1216 use_vram = true;
1217 else if ((flags & (MSM_BO_STOLEN | MSM_BO_SCANOUT)) && priv->vram.size)
1218 use_vram = true;
1219
1220 if (GEM_WARN_ON(use_vram && !priv->vram.size))
1221 return ERR_PTR(-EINVAL);
1222
1223 /* Disallow zero sized objects as they make the underlying
1224 * infrastructure grumpy
1225 */
1226 if (size == 0)
1227 return ERR_PTR(-EINVAL);
1228
1229 ret = msm_gem_new_impl(dev, size, flags, &obj);
1230 if (ret)
1231 return ERR_PTR(ret);
1232
1233 msm_obj = to_msm_bo(obj);
1234
1235 if (use_vram) {
1236 struct msm_gem_vma *vma;
1237 struct page **pages;
1238
1239 drm_gem_private_object_init(dev, obj, size);
1240
1241 msm_gem_lock(obj);
1242
1243 vma = add_vma(obj, NULL);
1244 msm_gem_unlock(obj);
1245 if (IS_ERR(vma)) {
1246 ret = PTR_ERR(vma);
1247 goto fail;
1248 }
1249
1250 to_msm_bo(obj)->vram_node = &vma->node;
1251
1252 msm_gem_lock(obj);
1253 pages = get_pages(obj);
1254 msm_gem_unlock(obj);
1255 if (IS_ERR(pages)) {
1256 ret = PTR_ERR(pages);
1257 goto fail;
1258 }
1259
1260 vma->iova = physaddr(obj);
1261 } else {
1262 ret = drm_gem_object_init(dev, obj, size);
1263 if (ret)
1264 goto fail;
1265 /*
1266 * Our buffers are kept pinned, so allocating them from the
1267 * MOVABLE zone is a really bad idea, and conflicts with CMA.
1268 * See comments above new_inode() why this is required _and_
1269 * expected if you're going to pin these pages.
1270 */
1271 mapping_set_gfp_mask(obj->filp->f_mapping, GFP_HIGHUSER);
1272 }
1273
1274 drm_gem_lru_move_tail(&priv->lru.unbacked, obj);
1275
1276 mutex_lock(&priv->obj_lock);
1277 list_add_tail(&msm_obj->node, &priv->objects);
1278 mutex_unlock(&priv->obj_lock);
1279
1280 ret = drm_gem_create_mmap_offset(obj);
1281 if (ret)
1282 goto fail;
1283
1284 return obj;
1285
1286fail:
1287 drm_gem_object_put(obj);
1288 return ERR_PTR(ret);
1289}
1290
1291struct drm_gem_object *msm_gem_import(struct drm_device *dev,
1292 struct dma_buf *dmabuf, struct sg_table *sgt)
1293{
1294 struct msm_drm_private *priv = dev->dev_private;
1295 struct msm_gem_object *msm_obj;
1296 struct drm_gem_object *obj;
1297 uint32_t size;
1298 int ret, npages;
1299
1300 /* if we don't have IOMMU, don't bother pretending we can import: */
1301 if (!msm_use_mmu(dev)) {
1302 DRM_DEV_ERROR(dev->dev, "cannot import without IOMMU\n");
1303 return ERR_PTR(-EINVAL);
1304 }
1305
1306 size = PAGE_ALIGN(dmabuf->size);
1307
1308 ret = msm_gem_new_impl(dev, size, MSM_BO_WC, &obj);
1309 if (ret)
1310 return ERR_PTR(ret);
1311
1312 drm_gem_private_object_init(dev, obj, size);
1313
1314 npages = size / PAGE_SIZE;
1315
1316 msm_obj = to_msm_bo(obj);
1317 msm_gem_lock(obj);
1318 msm_obj->sgt = sgt;
1319 msm_obj->pages = kvmalloc_array(npages, sizeof(struct page *), GFP_KERNEL);
1320 if (!msm_obj->pages) {
1321 msm_gem_unlock(obj);
1322 ret = -ENOMEM;
1323 goto fail;
1324 }
1325
1326 ret = drm_prime_sg_to_page_array(sgt, msm_obj->pages, npages);
1327 if (ret) {
1328 msm_gem_unlock(obj);
1329 goto fail;
1330 }
1331
1332 msm_gem_unlock(obj);
1333
1334 drm_gem_lru_move_tail(&priv->lru.pinned, obj);
1335
1336 mutex_lock(&priv->obj_lock);
1337 list_add_tail(&msm_obj->node, &priv->objects);
1338 mutex_unlock(&priv->obj_lock);
1339
1340 ret = drm_gem_create_mmap_offset(obj);
1341 if (ret)
1342 goto fail;
1343
1344 return obj;
1345
1346fail:
1347 drm_gem_object_put(obj);
1348 return ERR_PTR(ret);
1349}
1350
1351void *msm_gem_kernel_new(struct drm_device *dev, uint32_t size,
1352 uint32_t flags, struct msm_gem_address_space *aspace,
1353 struct drm_gem_object **bo, uint64_t *iova)
1354{
1355 void *vaddr;
1356 struct drm_gem_object *obj = msm_gem_new(dev, size, flags);
1357 int ret;
1358
1359 if (IS_ERR(obj))
1360 return ERR_CAST(obj);
1361
1362 if (iova) {
1363 ret = msm_gem_get_and_pin_iova(obj, aspace, iova);
1364 if (ret)
1365 goto err;
1366 }
1367
1368 vaddr = msm_gem_get_vaddr(obj);
1369 if (IS_ERR(vaddr)) {
1370 msm_gem_unpin_iova(obj, aspace);
1371 ret = PTR_ERR(vaddr);
1372 goto err;
1373 }
1374
1375 if (bo)
1376 *bo = obj;
1377
1378 return vaddr;
1379err:
1380 drm_gem_object_put(obj);
1381
1382 return ERR_PTR(ret);
1383
1384}
1385
1386void msm_gem_kernel_put(struct drm_gem_object *bo,
1387 struct msm_gem_address_space *aspace)
1388{
1389 if (IS_ERR_OR_NULL(bo))
1390 return;
1391
1392 msm_gem_put_vaddr(bo);
1393 msm_gem_unpin_iova(bo, aspace);
1394 drm_gem_object_put(bo);
1395}
1396
1397void msm_gem_object_set_name(struct drm_gem_object *bo, const char *fmt, ...)
1398{
1399 struct msm_gem_object *msm_obj = to_msm_bo(bo);
1400 va_list ap;
1401
1402 if (!fmt)
1403 return;
1404
1405 va_start(ap, fmt);
1406 vsnprintf(msm_obj->name, sizeof(msm_obj->name), fmt, ap);
1407 va_end(ap);
1408}