Merge tag 'drm-intel-gt-next-2022-09-09' of git://anongit.freedesktop.org/drm/drm...
[sfrench/cifs-2.6.git] / drivers / gpu / drm / i915 / gem / i915_gem_ttm.c
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2021 Intel Corporation
4  */
5
6 #include <linux/shmem_fs.h>
7
8 #include <drm/ttm/ttm_bo_driver.h>
9 #include <drm/ttm/ttm_placement.h>
10 #include <drm/drm_buddy.h>
11
12 #include "i915_drv.h"
13 #include "i915_ttm_buddy_manager.h"
14 #include "intel_memory_region.h"
15 #include "intel_region_ttm.h"
16
17 #include "gem/i915_gem_mman.h"
18 #include "gem/i915_gem_object.h"
19 #include "gem/i915_gem_region.h"
20 #include "gem/i915_gem_ttm.h"
21 #include "gem/i915_gem_ttm_move.h"
22 #include "gem/i915_gem_ttm_pm.h"
23 #include "gt/intel_gpu_commands.h"
24
25 #define I915_TTM_PRIO_PURGE     0
26 #define I915_TTM_PRIO_NO_PAGES  1
27 #define I915_TTM_PRIO_HAS_PAGES 2
28 #define I915_TTM_PRIO_NEEDS_CPU_ACCESS 3
29
30 /*
31  * Size of struct ttm_place vector in on-stack struct ttm_placement allocs
32  */
33 #define I915_TTM_MAX_PLACEMENTS INTEL_REGION_UNKNOWN
34
35 /**
36  * struct i915_ttm_tt - TTM page vector with additional private information
37  * @ttm: The base TTM page vector.
38  * @dev: The struct device used for dma mapping and unmapping.
39  * @cached_rsgt: The cached scatter-gather table.
40  * @is_shmem: Set if using shmem.
41  * @filp: The shmem file, if using shmem backend.
42  *
43  * Note that DMA may be going on right up to the point where the page-
44  * vector is unpopulated in delayed destroy. Hence keep the
45  * scatter-gather table mapped and cached up to that point. This is
46  * different from the cached gem object io scatter-gather table which
47  * doesn't have an associated dma mapping.
48  */
49 struct i915_ttm_tt {
50         struct ttm_tt ttm;
51         struct device *dev;
52         struct i915_refct_sgt cached_rsgt;
53
54         bool is_shmem;
55         struct file *filp;
56 };
57
58 static const struct ttm_place sys_placement_flags = {
59         .fpfn = 0,
60         .lpfn = 0,
61         .mem_type = I915_PL_SYSTEM,
62         .flags = 0,
63 };
64
65 static struct ttm_placement i915_sys_placement = {
66         .num_placement = 1,
67         .placement = &sys_placement_flags,
68         .num_busy_placement = 1,
69         .busy_placement = &sys_placement_flags,
70 };
71
72 /**
73  * i915_ttm_sys_placement - Return the struct ttm_placement to be
74  * used for an object in system memory.
75  *
76  * Rather than making the struct extern, use this
77  * function.
78  *
79  * Return: A pointer to a static variable for sys placement.
80  */
81 struct ttm_placement *i915_ttm_sys_placement(void)
82 {
83         return &i915_sys_placement;
84 }
85
86 static int i915_ttm_err_to_gem(int err)
87 {
88         /* Fastpath */
89         if (likely(!err))
90                 return 0;
91
92         switch (err) {
93         case -EBUSY:
94                 /*
95                  * TTM likes to convert -EDEADLK to -EBUSY, and wants us to
96                  * restart the operation, since we don't record the contending
97                  * lock. We use -EAGAIN to restart.
98                  */
99                 return -EAGAIN;
100         case -ENOSPC:
101                 /*
102                  * Memory type / region is full, and we can't evict.
103                  * Except possibly system, that returns -ENOMEM;
104                  */
105                 return -ENXIO;
106         default:
107                 break;
108         }
109
110         return err;
111 }
112
113 static enum ttm_caching
114 i915_ttm_select_tt_caching(const struct drm_i915_gem_object *obj)
115 {
116         /*
117          * Objects only allowed in system get cached cpu-mappings, or when
118          * evicting lmem-only buffers to system for swapping. Other objects get
119          * WC mapping for now. Even if in system.
120          */
121         if (obj->mm.n_placements <= 1)
122                 return ttm_cached;
123
124         return ttm_write_combined;
125 }
126
127 static void
128 i915_ttm_place_from_region(const struct intel_memory_region *mr,
129                            struct ttm_place *place,
130                            resource_size_t offset,
131                            resource_size_t size,
132                            unsigned int flags)
133 {
134         memset(place, 0, sizeof(*place));
135         place->mem_type = intel_region_to_ttm_type(mr);
136
137         if (mr->type == INTEL_MEMORY_SYSTEM)
138                 return;
139
140         if (flags & I915_BO_ALLOC_CONTIGUOUS)
141                 place->flags |= TTM_PL_FLAG_CONTIGUOUS;
142         if (offset != I915_BO_INVALID_OFFSET) {
143                 place->fpfn = offset >> PAGE_SHIFT;
144                 place->lpfn = place->fpfn + (size >> PAGE_SHIFT);
145         } else if (mr->io_size && mr->io_size < mr->total) {
146                 if (flags & I915_BO_ALLOC_GPU_ONLY) {
147                         place->flags |= TTM_PL_FLAG_TOPDOWN;
148                 } else {
149                         place->fpfn = 0;
150                         place->lpfn = mr->io_size >> PAGE_SHIFT;
151                 }
152         }
153 }
154
155 static void
156 i915_ttm_placement_from_obj(const struct drm_i915_gem_object *obj,
157                             struct ttm_place *requested,
158                             struct ttm_place *busy,
159                             struct ttm_placement *placement)
160 {
161         unsigned int num_allowed = obj->mm.n_placements;
162         unsigned int flags = obj->flags;
163         unsigned int i;
164
165         placement->num_placement = 1;
166         i915_ttm_place_from_region(num_allowed ? obj->mm.placements[0] :
167                                    obj->mm.region, requested, obj->bo_offset,
168                                    obj->base.size, flags);
169
170         /* Cache this on object? */
171         placement->num_busy_placement = num_allowed;
172         for (i = 0; i < placement->num_busy_placement; ++i)
173                 i915_ttm_place_from_region(obj->mm.placements[i], busy + i,
174                                            obj->bo_offset, obj->base.size, flags);
175
176         if (num_allowed == 0) {
177                 *busy = *requested;
178                 placement->num_busy_placement = 1;
179         }
180
181         placement->placement = requested;
182         placement->busy_placement = busy;
183 }
184
185 static int i915_ttm_tt_shmem_populate(struct ttm_device *bdev,
186                                       struct ttm_tt *ttm,
187                                       struct ttm_operation_ctx *ctx)
188 {
189         struct drm_i915_private *i915 = container_of(bdev, typeof(*i915), bdev);
190         struct intel_memory_region *mr = i915->mm.regions[INTEL_MEMORY_SYSTEM];
191         struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm);
192         const unsigned int max_segment = i915_sg_segment_size();
193         const size_t size = (size_t)ttm->num_pages << PAGE_SHIFT;
194         struct file *filp = i915_tt->filp;
195         struct sgt_iter sgt_iter;
196         struct sg_table *st;
197         struct page *page;
198         unsigned long i;
199         int err;
200
201         if (!filp) {
202                 struct address_space *mapping;
203                 gfp_t mask;
204
205                 filp = shmem_file_setup("i915-shmem-tt", size, VM_NORESERVE);
206                 if (IS_ERR(filp))
207                         return PTR_ERR(filp);
208
209                 mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
210
211                 mapping = filp->f_mapping;
212                 mapping_set_gfp_mask(mapping, mask);
213                 GEM_BUG_ON(!(mapping_gfp_mask(mapping) & __GFP_RECLAIM));
214
215                 i915_tt->filp = filp;
216         }
217
218         st = &i915_tt->cached_rsgt.table;
219         err = shmem_sg_alloc_table(i915, st, size, mr, filp->f_mapping,
220                                    max_segment);
221         if (err)
222                 return err;
223
224         err = dma_map_sgtable(i915_tt->dev, st, DMA_BIDIRECTIONAL,
225                               DMA_ATTR_SKIP_CPU_SYNC);
226         if (err)
227                 goto err_free_st;
228
229         i = 0;
230         for_each_sgt_page(page, sgt_iter, st)
231                 ttm->pages[i++] = page;
232
233         if (ttm->page_flags & TTM_TT_FLAG_SWAPPED)
234                 ttm->page_flags &= ~TTM_TT_FLAG_SWAPPED;
235
236         return 0;
237
238 err_free_st:
239         shmem_sg_free_table(st, filp->f_mapping, false, false);
240
241         return err;
242 }
243
244 static void i915_ttm_tt_shmem_unpopulate(struct ttm_tt *ttm)
245 {
246         struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm);
247         bool backup = ttm->page_flags & TTM_TT_FLAG_SWAPPED;
248         struct sg_table *st = &i915_tt->cached_rsgt.table;
249
250         shmem_sg_free_table(st, file_inode(i915_tt->filp)->i_mapping,
251                             backup, backup);
252 }
253
254 static void i915_ttm_tt_release(struct kref *ref)
255 {
256         struct i915_ttm_tt *i915_tt =
257                 container_of(ref, typeof(*i915_tt), cached_rsgt.kref);
258         struct sg_table *st = &i915_tt->cached_rsgt.table;
259
260         GEM_WARN_ON(st->sgl);
261
262         kfree(i915_tt);
263 }
264
265 static const struct i915_refct_sgt_ops tt_rsgt_ops = {
266         .release = i915_ttm_tt_release
267 };
268
269 static struct ttm_tt *i915_ttm_tt_create(struct ttm_buffer_object *bo,
270                                          uint32_t page_flags)
271 {
272         struct drm_i915_private *i915 = container_of(bo->bdev, typeof(*i915),
273                                                      bdev);
274         struct ttm_resource_manager *man =
275                 ttm_manager_type(bo->bdev, bo->resource->mem_type);
276         struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
277         unsigned long ccs_pages = 0;
278         enum ttm_caching caching;
279         struct i915_ttm_tt *i915_tt;
280         int ret;
281
282         if (!obj)
283                 return NULL;
284
285         i915_tt = kzalloc(sizeof(*i915_tt), GFP_KERNEL);
286         if (!i915_tt)
287                 return NULL;
288
289         if (obj->flags & I915_BO_ALLOC_CPU_CLEAR &&
290             man->use_tt)
291                 page_flags |= TTM_TT_FLAG_ZERO_ALLOC;
292
293         caching = i915_ttm_select_tt_caching(obj);
294         if (i915_gem_object_is_shrinkable(obj) && caching == ttm_cached) {
295                 page_flags |= TTM_TT_FLAG_EXTERNAL |
296                               TTM_TT_FLAG_EXTERNAL_MAPPABLE;
297                 i915_tt->is_shmem = true;
298         }
299
300         if (i915_gem_object_needs_ccs_pages(obj))
301                 ccs_pages = DIV_ROUND_UP(DIV_ROUND_UP(bo->base.size,
302                                                       NUM_BYTES_PER_CCS_BYTE),
303                                          PAGE_SIZE);
304
305         ret = ttm_tt_init(&i915_tt->ttm, bo, page_flags, caching, ccs_pages);
306         if (ret)
307                 goto err_free;
308
309         __i915_refct_sgt_init(&i915_tt->cached_rsgt, bo->base.size,
310                               &tt_rsgt_ops);
311
312         i915_tt->dev = obj->base.dev->dev;
313
314         return &i915_tt->ttm;
315
316 err_free:
317         kfree(i915_tt);
318         return NULL;
319 }
320
321 static int i915_ttm_tt_populate(struct ttm_device *bdev,
322                                 struct ttm_tt *ttm,
323                                 struct ttm_operation_ctx *ctx)
324 {
325         struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm);
326
327         if (i915_tt->is_shmem)
328                 return i915_ttm_tt_shmem_populate(bdev, ttm, ctx);
329
330         return ttm_pool_alloc(&bdev->pool, ttm, ctx);
331 }
332
333 static void i915_ttm_tt_unpopulate(struct ttm_device *bdev, struct ttm_tt *ttm)
334 {
335         struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm);
336         struct sg_table *st = &i915_tt->cached_rsgt.table;
337
338         if (st->sgl)
339                 dma_unmap_sgtable(i915_tt->dev, st, DMA_BIDIRECTIONAL, 0);
340
341         if (i915_tt->is_shmem) {
342                 i915_ttm_tt_shmem_unpopulate(ttm);
343         } else {
344                 sg_free_table(st);
345                 ttm_pool_free(&bdev->pool, ttm);
346         }
347 }
348
349 static void i915_ttm_tt_destroy(struct ttm_device *bdev, struct ttm_tt *ttm)
350 {
351         struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm);
352
353         if (i915_tt->filp)
354                 fput(i915_tt->filp);
355
356         ttm_tt_fini(ttm);
357         i915_refct_sgt_put(&i915_tt->cached_rsgt);
358 }
359
360 static bool i915_ttm_eviction_valuable(struct ttm_buffer_object *bo,
361                                        const struct ttm_place *place)
362 {
363         struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
364
365         if (!obj)
366                 return false;
367
368         /*
369          * EXTERNAL objects should never be swapped out by TTM, instead we need
370          * to handle that ourselves. TTM will already skip such objects for us,
371          * but we would like to avoid grabbing locks for no good reason.
372          */
373         if (bo->ttm && bo->ttm->page_flags & TTM_TT_FLAG_EXTERNAL)
374                 return false;
375
376         /* Will do for now. Our pinned objects are still on TTM's LRU lists */
377         if (!i915_gem_object_evictable(obj))
378                 return false;
379
380         return ttm_bo_eviction_valuable(bo, place);
381 }
382
383 static void i915_ttm_evict_flags(struct ttm_buffer_object *bo,
384                                  struct ttm_placement *placement)
385 {
386         *placement = i915_sys_placement;
387 }
388
389 /**
390  * i915_ttm_free_cached_io_rsgt - Free object cached LMEM information
391  * @obj: The GEM object
392  * This function frees any LMEM-related information that is cached on
393  * the object. For example the radix tree for fast page lookup and the
394  * cached refcounted sg-table
395  */
396 void i915_ttm_free_cached_io_rsgt(struct drm_i915_gem_object *obj)
397 {
398         struct radix_tree_iter iter;
399         void __rcu **slot;
400
401         if (!obj->ttm.cached_io_rsgt)
402                 return;
403
404         rcu_read_lock();
405         radix_tree_for_each_slot(slot, &obj->ttm.get_io_page.radix, &iter, 0)
406                 radix_tree_delete(&obj->ttm.get_io_page.radix, iter.index);
407         rcu_read_unlock();
408
409         i915_refct_sgt_put(obj->ttm.cached_io_rsgt);
410         obj->ttm.cached_io_rsgt = NULL;
411 }
412
413 /**
414  * i915_ttm_purge - Clear an object of its memory
415  * @obj: The object
416  *
417  * This function is called to clear an object of it's memory when it is
418  * marked as not needed anymore.
419  *
420  * Return: 0 on success, negative error code on failure.
421  */
422 int i915_ttm_purge(struct drm_i915_gem_object *obj)
423 {
424         struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
425         struct i915_ttm_tt *i915_tt =
426                 container_of(bo->ttm, typeof(*i915_tt), ttm);
427         struct ttm_operation_ctx ctx = {
428                 .interruptible = true,
429                 .no_wait_gpu = false,
430         };
431         struct ttm_placement place = {};
432         int ret;
433
434         if (obj->mm.madv == __I915_MADV_PURGED)
435                 return 0;
436
437         ret = ttm_bo_validate(bo, &place, &ctx);
438         if (ret)
439                 return ret;
440
441         if (bo->ttm && i915_tt->filp) {
442                 /*
443                  * The below fput(which eventually calls shmem_truncate) might
444                  * be delayed by worker, so when directly called to purge the
445                  * pages(like by the shrinker) we should try to be more
446                  * aggressive and release the pages immediately.
447                  */
448                 shmem_truncate_range(file_inode(i915_tt->filp),
449                                      0, (loff_t)-1);
450                 fput(fetch_and_zero(&i915_tt->filp));
451         }
452
453         obj->write_domain = 0;
454         obj->read_domains = 0;
455         i915_ttm_adjust_gem_after_move(obj);
456         i915_ttm_free_cached_io_rsgt(obj);
457         obj->mm.madv = __I915_MADV_PURGED;
458
459         return 0;
460 }
461
462 static int i915_ttm_shrink(struct drm_i915_gem_object *obj, unsigned int flags)
463 {
464         struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
465         struct i915_ttm_tt *i915_tt =
466                 container_of(bo->ttm, typeof(*i915_tt), ttm);
467         struct ttm_operation_ctx ctx = {
468                 .interruptible = true,
469                 .no_wait_gpu = flags & I915_GEM_OBJECT_SHRINK_NO_GPU_WAIT,
470         };
471         struct ttm_placement place = {};
472         int ret;
473
474         if (!bo->ttm || bo->resource->mem_type != TTM_PL_SYSTEM)
475                 return 0;
476
477         GEM_BUG_ON(!i915_tt->is_shmem);
478
479         if (!i915_tt->filp)
480                 return 0;
481
482         ret = ttm_bo_wait_ctx(bo, &ctx);
483         if (ret)
484                 return ret;
485
486         switch (obj->mm.madv) {
487         case I915_MADV_DONTNEED:
488                 return i915_ttm_purge(obj);
489         case __I915_MADV_PURGED:
490                 return 0;
491         }
492
493         if (bo->ttm->page_flags & TTM_TT_FLAG_SWAPPED)
494                 return 0;
495
496         bo->ttm->page_flags |= TTM_TT_FLAG_SWAPPED;
497         ret = ttm_bo_validate(bo, &place, &ctx);
498         if (ret) {
499                 bo->ttm->page_flags &= ~TTM_TT_FLAG_SWAPPED;
500                 return ret;
501         }
502
503         if (flags & I915_GEM_OBJECT_SHRINK_WRITEBACK)
504                 __shmem_writeback(obj->base.size, i915_tt->filp->f_mapping);
505
506         return 0;
507 }
508
509 static void i915_ttm_delete_mem_notify(struct ttm_buffer_object *bo)
510 {
511         struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
512
513         if (likely(obj)) {
514                 __i915_gem_object_pages_fini(obj);
515                 i915_ttm_free_cached_io_rsgt(obj);
516         }
517 }
518
519 static struct i915_refct_sgt *i915_ttm_tt_get_st(struct ttm_tt *ttm)
520 {
521         struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm);
522         struct sg_table *st;
523         int ret;
524
525         if (i915_tt->cached_rsgt.table.sgl)
526                 return i915_refct_sgt_get(&i915_tt->cached_rsgt);
527
528         st = &i915_tt->cached_rsgt.table;
529         ret = sg_alloc_table_from_pages_segment(st,
530                         ttm->pages, ttm->num_pages,
531                         0, (unsigned long)ttm->num_pages << PAGE_SHIFT,
532                         i915_sg_segment_size(), GFP_KERNEL);
533         if (ret) {
534                 st->sgl = NULL;
535                 return ERR_PTR(ret);
536         }
537
538         ret = dma_map_sgtable(i915_tt->dev, st, DMA_BIDIRECTIONAL, 0);
539         if (ret) {
540                 sg_free_table(st);
541                 return ERR_PTR(ret);
542         }
543
544         return i915_refct_sgt_get(&i915_tt->cached_rsgt);
545 }
546
547 /**
548  * i915_ttm_resource_get_st - Get a refcounted sg-table pointing to the
549  * resource memory
550  * @obj: The GEM object used for sg-table caching
551  * @res: The struct ttm_resource for which an sg-table is requested.
552  *
553  * This function returns a refcounted sg-table representing the memory
554  * pointed to by @res. If @res is the object's current resource it may also
555  * cache the sg_table on the object or attempt to access an already cached
556  * sg-table. The refcounted sg-table needs to be put when no-longer in use.
557  *
558  * Return: A valid pointer to a struct i915_refct_sgt or error pointer on
559  * failure.
560  */
561 struct i915_refct_sgt *
562 i915_ttm_resource_get_st(struct drm_i915_gem_object *obj,
563                          struct ttm_resource *res)
564 {
565         struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
566         u32 page_alignment;
567
568         if (!i915_ttm_gtt_binds_lmem(res))
569                 return i915_ttm_tt_get_st(bo->ttm);
570
571         page_alignment = bo->page_alignment << PAGE_SHIFT;
572         if (!page_alignment)
573                 page_alignment = obj->mm.region->min_page_size;
574
575         /*
576          * If CPU mapping differs, we need to add the ttm_tt pages to
577          * the resulting st. Might make sense for GGTT.
578          */
579         GEM_WARN_ON(!i915_ttm_cpu_maps_iomem(res));
580         if (bo->resource == res) {
581                 if (!obj->ttm.cached_io_rsgt) {
582                         struct i915_refct_sgt *rsgt;
583
584                         rsgt = intel_region_ttm_resource_to_rsgt(obj->mm.region,
585                                                                  res,
586                                                                  page_alignment);
587                         if (IS_ERR(rsgt))
588                                 return rsgt;
589
590                         obj->ttm.cached_io_rsgt = rsgt;
591                 }
592                 return i915_refct_sgt_get(obj->ttm.cached_io_rsgt);
593         }
594
595         return intel_region_ttm_resource_to_rsgt(obj->mm.region, res,
596                                                  page_alignment);
597 }
598
599 static int i915_ttm_truncate(struct drm_i915_gem_object *obj)
600 {
601         struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
602         int err;
603
604         WARN_ON_ONCE(obj->mm.madv == I915_MADV_WILLNEED);
605
606         err = i915_ttm_move_notify(bo);
607         if (err)
608                 return err;
609
610         return i915_ttm_purge(obj);
611 }
612
613 static void i915_ttm_swap_notify(struct ttm_buffer_object *bo)
614 {
615         struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
616         int ret;
617
618         if (!obj)
619                 return;
620
621         ret = i915_ttm_move_notify(bo);
622         GEM_WARN_ON(ret);
623         GEM_WARN_ON(obj->ttm.cached_io_rsgt);
624         if (!ret && obj->mm.madv != I915_MADV_WILLNEED)
625                 i915_ttm_purge(obj);
626 }
627
628 /**
629  * i915_ttm_resource_mappable - Return true if the ttm resource is CPU
630  * accessible.
631  * @res: The TTM resource to check.
632  *
633  * This is interesting on small-BAR systems where we may encounter lmem objects
634  * that can't be accessed via the CPU.
635  */
636 bool i915_ttm_resource_mappable(struct ttm_resource *res)
637 {
638         struct i915_ttm_buddy_resource *bman_res = to_ttm_buddy_resource(res);
639
640         if (!i915_ttm_cpu_maps_iomem(res))
641                 return true;
642
643         return bman_res->used_visible_size == bman_res->base.num_pages;
644 }
645
646 static int i915_ttm_io_mem_reserve(struct ttm_device *bdev, struct ttm_resource *mem)
647 {
648         struct drm_i915_gem_object *obj = i915_ttm_to_gem(mem->bo);
649         bool unknown_state;
650
651         if (!obj)
652                 return -EINVAL;
653
654         if (!kref_get_unless_zero(&obj->base.refcount))
655                 return -EINVAL;
656
657         assert_object_held(obj);
658
659         unknown_state = i915_gem_object_has_unknown_state(obj);
660         i915_gem_object_put(obj);
661         if (unknown_state)
662                 return -EINVAL;
663
664         if (!i915_ttm_cpu_maps_iomem(mem))
665                 return 0;
666
667         if (!i915_ttm_resource_mappable(mem))
668                 return -EINVAL;
669
670         mem->bus.caching = ttm_write_combined;
671         mem->bus.is_iomem = true;
672
673         return 0;
674 }
675
676 static unsigned long i915_ttm_io_mem_pfn(struct ttm_buffer_object *bo,
677                                          unsigned long page_offset)
678 {
679         struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
680         struct scatterlist *sg;
681         unsigned long base;
682         unsigned int ofs;
683
684         GEM_BUG_ON(!obj);
685         GEM_WARN_ON(bo->ttm);
686
687         base = obj->mm.region->iomap.base - obj->mm.region->region.start;
688         sg = __i915_gem_object_get_sg(obj, &obj->ttm.get_io_page, page_offset, &ofs, true);
689
690         return ((base + sg_dma_address(sg)) >> PAGE_SHIFT) + ofs;
691 }
692
693 /*
694  * All callbacks need to take care not to downcast a struct ttm_buffer_object
695  * without checking its subclass, since it might be a TTM ghost object.
696  */
697 static struct ttm_device_funcs i915_ttm_bo_driver = {
698         .ttm_tt_create = i915_ttm_tt_create,
699         .ttm_tt_populate = i915_ttm_tt_populate,
700         .ttm_tt_unpopulate = i915_ttm_tt_unpopulate,
701         .ttm_tt_destroy = i915_ttm_tt_destroy,
702         .eviction_valuable = i915_ttm_eviction_valuable,
703         .evict_flags = i915_ttm_evict_flags,
704         .move = i915_ttm_move,
705         .swap_notify = i915_ttm_swap_notify,
706         .delete_mem_notify = i915_ttm_delete_mem_notify,
707         .io_mem_reserve = i915_ttm_io_mem_reserve,
708         .io_mem_pfn = i915_ttm_io_mem_pfn,
709 };
710
711 /**
712  * i915_ttm_driver - Return a pointer to the TTM device funcs
713  *
714  * Return: Pointer to statically allocated TTM device funcs.
715  */
716 struct ttm_device_funcs *i915_ttm_driver(void)
717 {
718         return &i915_ttm_bo_driver;
719 }
720
721 static int __i915_ttm_get_pages(struct drm_i915_gem_object *obj,
722                                 struct ttm_placement *placement)
723 {
724         struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
725         struct ttm_operation_ctx ctx = {
726                 .interruptible = true,
727                 .no_wait_gpu = false,
728         };
729         int real_num_busy;
730         int ret;
731
732         /* First try only the requested placement. No eviction. */
733         real_num_busy = fetch_and_zero(&placement->num_busy_placement);
734         ret = ttm_bo_validate(bo, placement, &ctx);
735         if (ret) {
736                 ret = i915_ttm_err_to_gem(ret);
737                 /*
738                  * Anything that wants to restart the operation gets to
739                  * do that.
740                  */
741                 if (ret == -EDEADLK || ret == -EINTR || ret == -ERESTARTSYS ||
742                     ret == -EAGAIN)
743                         return ret;
744
745                 /*
746                  * If the initial attempt fails, allow all accepted placements,
747                  * evicting if necessary.
748                  */
749                 placement->num_busy_placement = real_num_busy;
750                 ret = ttm_bo_validate(bo, placement, &ctx);
751                 if (ret)
752                         return i915_ttm_err_to_gem(ret);
753         }
754
755         if (bo->ttm && !ttm_tt_is_populated(bo->ttm)) {
756                 ret = ttm_tt_populate(bo->bdev, bo->ttm, &ctx);
757                 if (ret)
758                         return ret;
759
760                 i915_ttm_adjust_domains_after_move(obj);
761                 i915_ttm_adjust_gem_after_move(obj);
762         }
763
764         if (!i915_gem_object_has_pages(obj)) {
765                 struct i915_refct_sgt *rsgt =
766                         i915_ttm_resource_get_st(obj, bo->resource);
767
768                 if (IS_ERR(rsgt))
769                         return PTR_ERR(rsgt);
770
771                 GEM_BUG_ON(obj->mm.rsgt);
772                 obj->mm.rsgt = rsgt;
773                 __i915_gem_object_set_pages(obj, &rsgt->table,
774                                             i915_sg_dma_sizes(rsgt->table.sgl));
775         }
776
777         GEM_BUG_ON(bo->ttm && ((obj->base.size >> PAGE_SHIFT) < bo->ttm->num_pages));
778         i915_ttm_adjust_lru(obj);
779         return ret;
780 }
781
782 static int i915_ttm_get_pages(struct drm_i915_gem_object *obj)
783 {
784         struct ttm_place requested, busy[I915_TTM_MAX_PLACEMENTS];
785         struct ttm_placement placement;
786
787         GEM_BUG_ON(obj->mm.n_placements > I915_TTM_MAX_PLACEMENTS);
788
789         /* Move to the requested placement. */
790         i915_ttm_placement_from_obj(obj, &requested, busy, &placement);
791
792         return __i915_ttm_get_pages(obj, &placement);
793 }
794
795 /**
796  * DOC: Migration vs eviction
797  *
798  * GEM migration may not be the same as TTM migration / eviction. If
799  * the TTM core decides to evict an object it may be evicted to a
800  * TTM memory type that is not in the object's allowable GEM regions, or
801  * in fact theoretically to a TTM memory type that doesn't correspond to
802  * a GEM memory region. In that case the object's GEM region is not
803  * updated, and the data is migrated back to the GEM region at
804  * get_pages time. TTM may however set up CPU ptes to the object even
805  * when it is evicted.
806  * Gem forced migration using the i915_ttm_migrate() op, is allowed even
807  * to regions that are not in the object's list of allowable placements.
808  */
809 static int __i915_ttm_migrate(struct drm_i915_gem_object *obj,
810                               struct intel_memory_region *mr,
811                               unsigned int flags)
812 {
813         struct ttm_place requested;
814         struct ttm_placement placement;
815         int ret;
816
817         i915_ttm_place_from_region(mr, &requested, obj->bo_offset,
818                                    obj->base.size, flags);
819         placement.num_placement = 1;
820         placement.num_busy_placement = 1;
821         placement.placement = &requested;
822         placement.busy_placement = &requested;
823
824         ret = __i915_ttm_get_pages(obj, &placement);
825         if (ret)
826                 return ret;
827
828         /*
829          * Reinitialize the region bindings. This is primarily
830          * required for objects where the new region is not in
831          * its allowable placements.
832          */
833         if (obj->mm.region != mr) {
834                 i915_gem_object_release_memory_region(obj);
835                 i915_gem_object_init_memory_region(obj, mr);
836         }
837
838         return 0;
839 }
840
841 static int i915_ttm_migrate(struct drm_i915_gem_object *obj,
842                             struct intel_memory_region *mr)
843 {
844         return __i915_ttm_migrate(obj, mr, obj->flags);
845 }
846
847 static void i915_ttm_put_pages(struct drm_i915_gem_object *obj,
848                                struct sg_table *st)
849 {
850         /*
851          * We're currently not called from a shrinker, so put_pages()
852          * typically means the object is about to destroyed, or called
853          * from move_notify(). So just avoid doing much for now.
854          * If the object is not destroyed next, The TTM eviction logic
855          * and shrinkers will move it out if needed.
856          */
857
858         if (obj->mm.rsgt)
859                 i915_refct_sgt_put(fetch_and_zero(&obj->mm.rsgt));
860 }
861
862 /**
863  * i915_ttm_adjust_lru - Adjust an object's position on relevant LRU lists.
864  * @obj: The object
865  */
866 void i915_ttm_adjust_lru(struct drm_i915_gem_object *obj)
867 {
868         struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
869         struct i915_ttm_tt *i915_tt =
870                 container_of(bo->ttm, typeof(*i915_tt), ttm);
871         bool shrinkable =
872                 bo->ttm && i915_tt->filp && ttm_tt_is_populated(bo->ttm);
873
874         /*
875          * Don't manipulate the TTM LRUs while in TTM bo destruction.
876          * We're called through i915_ttm_delete_mem_notify().
877          */
878         if (!kref_read(&bo->kref))
879                 return;
880
881         /*
882          * We skip managing the shrinker LRU in set_pages() and just manage
883          * everything here. This does at least solve the issue with having
884          * temporary shmem mappings(like with evicted lmem) not being visible to
885          * the shrinker. Only our shmem objects are shrinkable, everything else
886          * we keep as unshrinkable.
887          *
888          * To make sure everything plays nice we keep an extra shrink pin in TTM
889          * if the underlying pages are not currently shrinkable. Once we release
890          * our pin, like when the pages are moved to shmem, the pages will then
891          * be added to the shrinker LRU, assuming the caller isn't also holding
892          * a pin.
893          *
894          * TODO: consider maybe also bumping the shrinker list here when we have
895          * already unpinned it, which should give us something more like an LRU.
896          *
897          * TODO: There is a small window of opportunity for this function to
898          * get called from eviction after we've dropped the last GEM refcount,
899          * but before the TTM deleted flag is set on the object. Avoid
900          * adjusting the shrinker list in such cases, since the object is
901          * not available to the shrinker anyway due to its zero refcount.
902          * To fix this properly we should move to a TTM shrinker LRU list for
903          * these objects.
904          */
905         if (kref_get_unless_zero(&obj->base.refcount)) {
906                 if (shrinkable != obj->mm.ttm_shrinkable) {
907                         if (shrinkable) {
908                                 if (obj->mm.madv == I915_MADV_WILLNEED)
909                                         __i915_gem_object_make_shrinkable(obj);
910                                 else
911                                         __i915_gem_object_make_purgeable(obj);
912                         } else {
913                                 i915_gem_object_make_unshrinkable(obj);
914                         }
915
916                         obj->mm.ttm_shrinkable = shrinkable;
917                 }
918                 i915_gem_object_put(obj);
919         }
920
921         /*
922          * Put on the correct LRU list depending on the MADV status
923          */
924         spin_lock(&bo->bdev->lru_lock);
925         if (shrinkable) {
926                 /* Try to keep shmem_tt from being considered for shrinking. */
927                 bo->priority = TTM_MAX_BO_PRIORITY - 1;
928         } else if (obj->mm.madv != I915_MADV_WILLNEED) {
929                 bo->priority = I915_TTM_PRIO_PURGE;
930         } else if (!i915_gem_object_has_pages(obj)) {
931                 bo->priority = I915_TTM_PRIO_NO_PAGES;
932         } else {
933                 struct ttm_resource_manager *man =
934                         ttm_manager_type(bo->bdev, bo->resource->mem_type);
935
936                 /*
937                  * If we need to place an LMEM resource which doesn't need CPU
938                  * access then we should try not to victimize mappable objects
939                  * first, since we likely end up stealing more of the mappable
940                  * portion. And likewise when we try to find space for a mappble
941                  * object, we know not to ever victimize objects that don't
942                  * occupy any mappable pages.
943                  */
944                 if (i915_ttm_cpu_maps_iomem(bo->resource) &&
945                     i915_ttm_buddy_man_visible_size(man) < man->size &&
946                     !(obj->flags & I915_BO_ALLOC_GPU_ONLY))
947                         bo->priority = I915_TTM_PRIO_NEEDS_CPU_ACCESS;
948                 else
949                         bo->priority = I915_TTM_PRIO_HAS_PAGES;
950         }
951
952         ttm_bo_move_to_lru_tail(bo);
953         spin_unlock(&bo->bdev->lru_lock);
954 }
955
956 /*
957  * TTM-backed gem object destruction requires some clarification.
958  * Basically we have two possibilities here. We can either rely on the
959  * i915 delayed destruction and put the TTM object when the object
960  * is idle. This would be detected by TTM which would bypass the
961  * TTM delayed destroy handling. The other approach is to put the TTM
962  * object early and rely on the TTM destroyed handling, and then free
963  * the leftover parts of the GEM object once TTM's destroyed list handling is
964  * complete. For now, we rely on the latter for two reasons:
965  * a) TTM can evict an object even when it's on the delayed destroy list,
966  * which in theory allows for complete eviction.
967  * b) There is work going on in TTM to allow freeing an object even when
968  * it's not idle, and using the TTM destroyed list handling could help us
969  * benefit from that.
970  */
971 static void i915_ttm_delayed_free(struct drm_i915_gem_object *obj)
972 {
973         GEM_BUG_ON(!obj->ttm.created);
974
975         ttm_bo_put(i915_gem_to_ttm(obj));
976 }
977
978 static vm_fault_t vm_fault_ttm(struct vm_fault *vmf)
979 {
980         struct vm_area_struct *area = vmf->vma;
981         struct ttm_buffer_object *bo = area->vm_private_data;
982         struct drm_device *dev = bo->base.dev;
983         struct drm_i915_gem_object *obj;
984         vm_fault_t ret;
985         int idx;
986
987         obj = i915_ttm_to_gem(bo);
988         if (!obj)
989                 return VM_FAULT_SIGBUS;
990
991         /* Sanity check that we allow writing into this object */
992         if (unlikely(i915_gem_object_is_readonly(obj) &&
993                      area->vm_flags & VM_WRITE))
994                 return VM_FAULT_SIGBUS;
995
996         ret = ttm_bo_vm_reserve(bo, vmf);
997         if (ret)
998                 return ret;
999
1000         if (obj->mm.madv != I915_MADV_WILLNEED) {
1001                 dma_resv_unlock(bo->base.resv);
1002                 return VM_FAULT_SIGBUS;
1003         }
1004
1005         if (!i915_ttm_resource_mappable(bo->resource)) {
1006                 int err = -ENODEV;
1007                 int i;
1008
1009                 for (i = 0; i < obj->mm.n_placements; i++) {
1010                         struct intel_memory_region *mr = obj->mm.placements[i];
1011                         unsigned int flags;
1012
1013                         if (!mr->io_size && mr->type != INTEL_MEMORY_SYSTEM)
1014                                 continue;
1015
1016                         flags = obj->flags;
1017                         flags &= ~I915_BO_ALLOC_GPU_ONLY;
1018                         err = __i915_ttm_migrate(obj, mr, flags);
1019                         if (!err)
1020                                 break;
1021                 }
1022
1023                 if (err) {
1024                         drm_dbg(dev, "Unable to make resource CPU accessible\n");
1025                         dma_resv_unlock(bo->base.resv);
1026                         return VM_FAULT_SIGBUS;
1027                 }
1028         }
1029
1030         if (drm_dev_enter(dev, &idx)) {
1031                 ret = ttm_bo_vm_fault_reserved(vmf, vmf->vma->vm_page_prot,
1032                                                TTM_BO_VM_NUM_PREFAULT);
1033                 drm_dev_exit(idx);
1034         } else {
1035                 ret = ttm_bo_vm_dummy_page(vmf, vmf->vma->vm_page_prot);
1036         }
1037         if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
1038                 return ret;
1039
1040         i915_ttm_adjust_lru(obj);
1041
1042         dma_resv_unlock(bo->base.resv);
1043         return ret;
1044 }
1045
1046 static int
1047 vm_access_ttm(struct vm_area_struct *area, unsigned long addr,
1048               void *buf, int len, int write)
1049 {
1050         struct drm_i915_gem_object *obj =
1051                 i915_ttm_to_gem(area->vm_private_data);
1052
1053         if (i915_gem_object_is_readonly(obj) && write)
1054                 return -EACCES;
1055
1056         return ttm_bo_vm_access(area, addr, buf, len, write);
1057 }
1058
1059 static void ttm_vm_open(struct vm_area_struct *vma)
1060 {
1061         struct drm_i915_gem_object *obj =
1062                 i915_ttm_to_gem(vma->vm_private_data);
1063
1064         GEM_BUG_ON(!obj);
1065         i915_gem_object_get(obj);
1066 }
1067
1068 static void ttm_vm_close(struct vm_area_struct *vma)
1069 {
1070         struct drm_i915_gem_object *obj =
1071                 i915_ttm_to_gem(vma->vm_private_data);
1072
1073         GEM_BUG_ON(!obj);
1074         i915_gem_object_put(obj);
1075 }
1076
1077 static const struct vm_operations_struct vm_ops_ttm = {
1078         .fault = vm_fault_ttm,
1079         .access = vm_access_ttm,
1080         .open = ttm_vm_open,
1081         .close = ttm_vm_close,
1082 };
1083
1084 static u64 i915_ttm_mmap_offset(struct drm_i915_gem_object *obj)
1085 {
1086         /* The ttm_bo must be allocated with I915_BO_ALLOC_USER */
1087         GEM_BUG_ON(!drm_mm_node_allocated(&obj->base.vma_node.vm_node));
1088
1089         return drm_vma_node_offset_addr(&obj->base.vma_node);
1090 }
1091
1092 static void i915_ttm_unmap_virtual(struct drm_i915_gem_object *obj)
1093 {
1094         ttm_bo_unmap_virtual(i915_gem_to_ttm(obj));
1095 }
1096
1097 static const struct drm_i915_gem_object_ops i915_gem_ttm_obj_ops = {
1098         .name = "i915_gem_object_ttm",
1099         .flags = I915_GEM_OBJECT_IS_SHRINKABLE |
1100                  I915_GEM_OBJECT_SELF_MANAGED_SHRINK_LIST,
1101
1102         .get_pages = i915_ttm_get_pages,
1103         .put_pages = i915_ttm_put_pages,
1104         .truncate = i915_ttm_truncate,
1105         .shrink = i915_ttm_shrink,
1106
1107         .adjust_lru = i915_ttm_adjust_lru,
1108         .delayed_free = i915_ttm_delayed_free,
1109         .migrate = i915_ttm_migrate,
1110
1111         .mmap_offset = i915_ttm_mmap_offset,
1112         .unmap_virtual = i915_ttm_unmap_virtual,
1113         .mmap_ops = &vm_ops_ttm,
1114 };
1115
1116 void i915_ttm_bo_destroy(struct ttm_buffer_object *bo)
1117 {
1118         struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
1119
1120         i915_gem_object_release_memory_region(obj);
1121         mutex_destroy(&obj->ttm.get_io_page.lock);
1122
1123         if (obj->ttm.created) {
1124                 /*
1125                  * We freely manage the shrinker LRU outide of the mm.pages life
1126                  * cycle. As a result when destroying the object we should be
1127                  * extra paranoid and ensure we remove it from the LRU, before
1128                  * we free the object.
1129                  *
1130                  * Touching the ttm_shrinkable outside of the object lock here
1131                  * should be safe now that the last GEM object ref was dropped.
1132                  */
1133                 if (obj->mm.ttm_shrinkable)
1134                         i915_gem_object_make_unshrinkable(obj);
1135
1136                 i915_ttm_backup_free(obj);
1137
1138                 /* This releases all gem object bindings to the backend. */
1139                 __i915_gem_free_object(obj);
1140
1141                 call_rcu(&obj->rcu, __i915_gem_free_object_rcu);
1142         } else {
1143                 __i915_gem_object_fini(obj);
1144         }
1145 }
1146
1147 /**
1148  * __i915_gem_ttm_object_init - Initialize a ttm-backed i915 gem object
1149  * @mem: The initial memory region for the object.
1150  * @obj: The gem object.
1151  * @size: Object size in bytes.
1152  * @flags: gem object flags.
1153  *
1154  * Return: 0 on success, negative error code on failure.
1155  */
1156 int __i915_gem_ttm_object_init(struct intel_memory_region *mem,
1157                                struct drm_i915_gem_object *obj,
1158                                resource_size_t offset,
1159                                resource_size_t size,
1160                                resource_size_t page_size,
1161                                unsigned int flags)
1162 {
1163         static struct lock_class_key lock_class;
1164         struct drm_i915_private *i915 = mem->i915;
1165         struct ttm_operation_ctx ctx = {
1166                 .interruptible = true,
1167                 .no_wait_gpu = false,
1168         };
1169         enum ttm_bo_type bo_type;
1170         int ret;
1171
1172         drm_gem_private_object_init(&i915->drm, &obj->base, size);
1173         i915_gem_object_init(obj, &i915_gem_ttm_obj_ops, &lock_class, flags);
1174
1175         obj->bo_offset = offset;
1176
1177         /* Don't put on a region list until we're either locked or fully initialized. */
1178         obj->mm.region = mem;
1179         INIT_LIST_HEAD(&obj->mm.region_link);
1180
1181         INIT_RADIX_TREE(&obj->ttm.get_io_page.radix, GFP_KERNEL | __GFP_NOWARN);
1182         mutex_init(&obj->ttm.get_io_page.lock);
1183         bo_type = (obj->flags & I915_BO_ALLOC_USER) ? ttm_bo_type_device :
1184                 ttm_bo_type_kernel;
1185
1186         obj->base.vma_node.driver_private = i915_gem_to_ttm(obj);
1187
1188         /* Forcing the page size is kernel internal only */
1189         GEM_BUG_ON(page_size && obj->mm.n_placements);
1190
1191         /*
1192          * Keep an extra shrink pin to prevent the object from being made
1193          * shrinkable too early. If the ttm_tt is ever allocated in shmem, we
1194          * drop the pin. The TTM backend manages the shrinker LRU itself,
1195          * outside of the normal mm.pages life cycle.
1196          */
1197         i915_gem_object_make_unshrinkable(obj);
1198
1199         /*
1200          * If this function fails, it will call the destructor, but
1201          * our caller still owns the object. So no freeing in the
1202          * destructor until obj->ttm.created is true.
1203          * Similarly, in delayed_destroy, we can't call ttm_bo_put()
1204          * until successful initialization.
1205          */
1206         ret = ttm_bo_init_reserved(&i915->bdev, i915_gem_to_ttm(obj), bo_type,
1207                                    &i915_sys_placement, page_size >> PAGE_SHIFT,
1208                                    &ctx, NULL, NULL, i915_ttm_bo_destroy);
1209         if (ret)
1210                 return i915_ttm_err_to_gem(ret);
1211
1212         obj->ttm.created = true;
1213         i915_gem_object_release_memory_region(obj);
1214         i915_gem_object_init_memory_region(obj, mem);
1215         i915_ttm_adjust_domains_after_move(obj);
1216         i915_ttm_adjust_gem_after_move(obj);
1217         i915_gem_object_unlock(obj);
1218
1219         return 0;
1220 }
1221
1222 static const struct intel_memory_region_ops ttm_system_region_ops = {
1223         .init_object = __i915_gem_ttm_object_init,
1224         .release = intel_region_ttm_fini,
1225 };
1226
1227 struct intel_memory_region *
1228 i915_gem_ttm_system_setup(struct drm_i915_private *i915,
1229                           u16 type, u16 instance)
1230 {
1231         struct intel_memory_region *mr;
1232
1233         mr = intel_memory_region_create(i915, 0,
1234                                         totalram_pages() << PAGE_SHIFT,
1235                                         PAGE_SIZE, 0, 0,
1236                                         type, instance,
1237                                         &ttm_system_region_ops);
1238         if (IS_ERR(mr))
1239                 return mr;
1240
1241         intel_memory_region_set_name(mr, "system-ttm");
1242         return mr;
1243 }