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 */
i915_ttm_sys_placement(void)81 struct ttm_placement *i915_ttm_sys_placement(void)
82 {
83 return &i915_sys_placement;
84 }
85
i915_ttm_err_to_gem(int err)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
i915_ttm_select_tt_caching(const struct drm_i915_gem_object * obj)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
i915_ttm_place_from_region(const struct intel_memory_region * mr,struct ttm_place * place,resource_size_t offset,resource_size_t size,unsigned int flags)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
i915_ttm_placement_from_obj(const struct drm_i915_gem_object * obj,struct ttm_place * requested,struct ttm_place * busy,struct ttm_placement * placement)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
i915_ttm_tt_shmem_populate(struct ttm_device * bdev,struct ttm_tt * ttm,struct ttm_operation_ctx * ctx)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(i915->drm.dev);
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
i915_ttm_tt_shmem_unpopulate(struct ttm_tt * ttm)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
i915_ttm_tt_release(struct kref * ref)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
i915_ttm_tt_create(struct ttm_buffer_object * bo,uint32_t page_flags)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 (i915_ttm_is_ghost_object(bo))
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
i915_ttm_tt_populate(struct ttm_device * bdev,struct ttm_tt * ttm,struct ttm_operation_ctx * ctx)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
i915_ttm_tt_unpopulate(struct ttm_device * bdev,struct ttm_tt * ttm)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
i915_ttm_tt_destroy(struct ttm_device * bdev,struct ttm_tt * ttm)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
i915_ttm_eviction_valuable(struct ttm_buffer_object * bo,const struct ttm_place * place)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 (i915_ttm_is_ghost_object(bo))
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
i915_ttm_evict_flags(struct ttm_buffer_object * bo,struct ttm_placement * placement)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 */
i915_ttm_free_cached_io_rsgt(struct drm_i915_gem_object * obj)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 */
i915_ttm_purge(struct drm_i915_gem_object * obj)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
i915_ttm_shrink(struct drm_i915_gem_object * obj,unsigned int flags)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
i915_ttm_delete_mem_notify(struct ttm_buffer_object * bo)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 (bo->resource && !i915_ttm_is_ghost_object(bo)) {
514 __i915_gem_object_pages_fini(obj);
515 i915_ttm_free_cached_io_rsgt(obj);
516 }
517 }
518
i915_ttm_tt_get_st(struct ttm_tt * ttm)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(i915_tt->dev), 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 *
i915_ttm_resource_get_st(struct drm_i915_gem_object * obj,struct ttm_resource * res)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
i915_ttm_truncate(struct drm_i915_gem_object * obj)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 = ttm_bo_wait(bo, true, false);
607 if (err)
608 return err;
609
610 err = i915_ttm_move_notify(bo);
611 if (err)
612 return err;
613
614 return i915_ttm_purge(obj);
615 }
616
i915_ttm_swap_notify(struct ttm_buffer_object * bo)617 static void i915_ttm_swap_notify(struct ttm_buffer_object *bo)
618 {
619 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
620 int ret;
621
622 if (i915_ttm_is_ghost_object(bo))
623 return;
624
625 ret = i915_ttm_move_notify(bo);
626 GEM_WARN_ON(ret);
627 GEM_WARN_ON(obj->ttm.cached_io_rsgt);
628 if (!ret && obj->mm.madv != I915_MADV_WILLNEED)
629 i915_ttm_purge(obj);
630 }
631
632 /**
633 * i915_ttm_resource_mappable - Return true if the ttm resource is CPU
634 * accessible.
635 * @res: The TTM resource to check.
636 *
637 * This is interesting on small-BAR systems where we may encounter lmem objects
638 * that can't be accessed via the CPU.
639 */
i915_ttm_resource_mappable(struct ttm_resource * res)640 bool i915_ttm_resource_mappable(struct ttm_resource *res)
641 {
642 struct i915_ttm_buddy_resource *bman_res = to_ttm_buddy_resource(res);
643
644 if (!i915_ttm_cpu_maps_iomem(res))
645 return true;
646
647 return bman_res->used_visible_size == bman_res->base.num_pages;
648 }
649
i915_ttm_io_mem_reserve(struct ttm_device * bdev,struct ttm_resource * mem)650 static int i915_ttm_io_mem_reserve(struct ttm_device *bdev, struct ttm_resource *mem)
651 {
652 struct drm_i915_gem_object *obj = i915_ttm_to_gem(mem->bo);
653 bool unknown_state;
654
655 if (i915_ttm_is_ghost_object(mem->bo))
656 return -EINVAL;
657
658 if (!kref_get_unless_zero(&obj->base.refcount))
659 return -EINVAL;
660
661 assert_object_held(obj);
662
663 unknown_state = i915_gem_object_has_unknown_state(obj);
664 i915_gem_object_put(obj);
665 if (unknown_state)
666 return -EINVAL;
667
668 if (!i915_ttm_cpu_maps_iomem(mem))
669 return 0;
670
671 if (!i915_ttm_resource_mappable(mem))
672 return -EINVAL;
673
674 mem->bus.caching = ttm_write_combined;
675 mem->bus.is_iomem = true;
676
677 return 0;
678 }
679
i915_ttm_io_mem_pfn(struct ttm_buffer_object * bo,unsigned long page_offset)680 static unsigned long i915_ttm_io_mem_pfn(struct ttm_buffer_object *bo,
681 unsigned long page_offset)
682 {
683 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
684 struct scatterlist *sg;
685 unsigned long base;
686 unsigned int ofs;
687
688 GEM_BUG_ON(i915_ttm_is_ghost_object(bo));
689 GEM_WARN_ON(bo->ttm);
690
691 base = obj->mm.region->iomap.base - obj->mm.region->region.start;
692 sg = __i915_gem_object_get_sg(obj, &obj->ttm.get_io_page, page_offset, &ofs, true);
693
694 return ((base + sg_dma_address(sg)) >> PAGE_SHIFT) + ofs;
695 }
696
697 /*
698 * All callbacks need to take care not to downcast a struct ttm_buffer_object
699 * without checking its subclass, since it might be a TTM ghost object.
700 */
701 static struct ttm_device_funcs i915_ttm_bo_driver = {
702 .ttm_tt_create = i915_ttm_tt_create,
703 .ttm_tt_populate = i915_ttm_tt_populate,
704 .ttm_tt_unpopulate = i915_ttm_tt_unpopulate,
705 .ttm_tt_destroy = i915_ttm_tt_destroy,
706 .eviction_valuable = i915_ttm_eviction_valuable,
707 .evict_flags = i915_ttm_evict_flags,
708 .move = i915_ttm_move,
709 .swap_notify = i915_ttm_swap_notify,
710 .delete_mem_notify = i915_ttm_delete_mem_notify,
711 .io_mem_reserve = i915_ttm_io_mem_reserve,
712 .io_mem_pfn = i915_ttm_io_mem_pfn,
713 };
714
715 /**
716 * i915_ttm_driver - Return a pointer to the TTM device funcs
717 *
718 * Return: Pointer to statically allocated TTM device funcs.
719 */
i915_ttm_driver(void)720 struct ttm_device_funcs *i915_ttm_driver(void)
721 {
722 return &i915_ttm_bo_driver;
723 }
724
__i915_ttm_get_pages(struct drm_i915_gem_object * obj,struct ttm_placement * placement)725 static int __i915_ttm_get_pages(struct drm_i915_gem_object *obj,
726 struct ttm_placement *placement)
727 {
728 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
729 struct ttm_operation_ctx ctx = {
730 .interruptible = true,
731 .no_wait_gpu = false,
732 };
733 int real_num_busy;
734 int ret;
735
736 /* First try only the requested placement. No eviction. */
737 real_num_busy = fetch_and_zero(&placement->num_busy_placement);
738 ret = ttm_bo_validate(bo, placement, &ctx);
739 if (ret) {
740 ret = i915_ttm_err_to_gem(ret);
741 /*
742 * Anything that wants to restart the operation gets to
743 * do that.
744 */
745 if (ret == -EDEADLK || ret == -EINTR || ret == -ERESTARTSYS ||
746 ret == -EAGAIN)
747 return ret;
748
749 /*
750 * If the initial attempt fails, allow all accepted placements,
751 * evicting if necessary.
752 */
753 placement->num_busy_placement = real_num_busy;
754 ret = ttm_bo_validate(bo, placement, &ctx);
755 if (ret)
756 return i915_ttm_err_to_gem(ret);
757 }
758
759 if (bo->ttm && !ttm_tt_is_populated(bo->ttm)) {
760 ret = ttm_tt_populate(bo->bdev, bo->ttm, &ctx);
761 if (ret)
762 return ret;
763
764 i915_ttm_adjust_domains_after_move(obj);
765 i915_ttm_adjust_gem_after_move(obj);
766 }
767
768 if (!i915_gem_object_has_pages(obj)) {
769 struct i915_refct_sgt *rsgt =
770 i915_ttm_resource_get_st(obj, bo->resource);
771
772 if (IS_ERR(rsgt))
773 return PTR_ERR(rsgt);
774
775 GEM_BUG_ON(obj->mm.rsgt);
776 obj->mm.rsgt = rsgt;
777 __i915_gem_object_set_pages(obj, &rsgt->table,
778 i915_sg_dma_sizes(rsgt->table.sgl));
779 }
780
781 GEM_BUG_ON(bo->ttm && ((obj->base.size >> PAGE_SHIFT) < bo->ttm->num_pages));
782 i915_ttm_adjust_lru(obj);
783 return ret;
784 }
785
i915_ttm_get_pages(struct drm_i915_gem_object * obj)786 static int i915_ttm_get_pages(struct drm_i915_gem_object *obj)
787 {
788 struct ttm_place requested, busy[I915_TTM_MAX_PLACEMENTS];
789 struct ttm_placement placement;
790
791 GEM_BUG_ON(obj->mm.n_placements > I915_TTM_MAX_PLACEMENTS);
792
793 /* Move to the requested placement. */
794 i915_ttm_placement_from_obj(obj, &requested, busy, &placement);
795
796 return __i915_ttm_get_pages(obj, &placement);
797 }
798
799 /**
800 * DOC: Migration vs eviction
801 *
802 * GEM migration may not be the same as TTM migration / eviction. If
803 * the TTM core decides to evict an object it may be evicted to a
804 * TTM memory type that is not in the object's allowable GEM regions, or
805 * in fact theoretically to a TTM memory type that doesn't correspond to
806 * a GEM memory region. In that case the object's GEM region is not
807 * updated, and the data is migrated back to the GEM region at
808 * get_pages time. TTM may however set up CPU ptes to the object even
809 * when it is evicted.
810 * Gem forced migration using the i915_ttm_migrate() op, is allowed even
811 * to regions that are not in the object's list of allowable placements.
812 */
__i915_ttm_migrate(struct drm_i915_gem_object * obj,struct intel_memory_region * mr,unsigned int flags)813 static int __i915_ttm_migrate(struct drm_i915_gem_object *obj,
814 struct intel_memory_region *mr,
815 unsigned int flags)
816 {
817 struct ttm_place requested;
818 struct ttm_placement placement;
819 int ret;
820
821 i915_ttm_place_from_region(mr, &requested, obj->bo_offset,
822 obj->base.size, flags);
823 placement.num_placement = 1;
824 placement.num_busy_placement = 1;
825 placement.placement = &requested;
826 placement.busy_placement = &requested;
827
828 ret = __i915_ttm_get_pages(obj, &placement);
829 if (ret)
830 return ret;
831
832 /*
833 * Reinitialize the region bindings. This is primarily
834 * required for objects where the new region is not in
835 * its allowable placements.
836 */
837 if (obj->mm.region != mr) {
838 i915_gem_object_release_memory_region(obj);
839 i915_gem_object_init_memory_region(obj, mr);
840 }
841
842 return 0;
843 }
844
i915_ttm_migrate(struct drm_i915_gem_object * obj,struct intel_memory_region * mr,unsigned int flags)845 static int i915_ttm_migrate(struct drm_i915_gem_object *obj,
846 struct intel_memory_region *mr,
847 unsigned int flags)
848 {
849 return __i915_ttm_migrate(obj, mr, flags);
850 }
851
i915_ttm_put_pages(struct drm_i915_gem_object * obj,struct sg_table * st)852 static void i915_ttm_put_pages(struct drm_i915_gem_object *obj,
853 struct sg_table *st)
854 {
855 /*
856 * We're currently not called from a shrinker, so put_pages()
857 * typically means the object is about to destroyed, or called
858 * from move_notify(). So just avoid doing much for now.
859 * If the object is not destroyed next, The TTM eviction logic
860 * and shrinkers will move it out if needed.
861 */
862
863 if (obj->mm.rsgt)
864 i915_refct_sgt_put(fetch_and_zero(&obj->mm.rsgt));
865 }
866
867 /**
868 * i915_ttm_adjust_lru - Adjust an object's position on relevant LRU lists.
869 * @obj: The object
870 */
i915_ttm_adjust_lru(struct drm_i915_gem_object * obj)871 void i915_ttm_adjust_lru(struct drm_i915_gem_object *obj)
872 {
873 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
874 struct i915_ttm_tt *i915_tt =
875 container_of(bo->ttm, typeof(*i915_tt), ttm);
876 bool shrinkable =
877 bo->ttm && i915_tt->filp && ttm_tt_is_populated(bo->ttm);
878
879 /*
880 * Don't manipulate the TTM LRUs while in TTM bo destruction.
881 * We're called through i915_ttm_delete_mem_notify().
882 */
883 if (!kref_read(&bo->kref))
884 return;
885
886 /*
887 * We skip managing the shrinker LRU in set_pages() and just manage
888 * everything here. This does at least solve the issue with having
889 * temporary shmem mappings(like with evicted lmem) not being visible to
890 * the shrinker. Only our shmem objects are shrinkable, everything else
891 * we keep as unshrinkable.
892 *
893 * To make sure everything plays nice we keep an extra shrink pin in TTM
894 * if the underlying pages are not currently shrinkable. Once we release
895 * our pin, like when the pages are moved to shmem, the pages will then
896 * be added to the shrinker LRU, assuming the caller isn't also holding
897 * a pin.
898 *
899 * TODO: consider maybe also bumping the shrinker list here when we have
900 * already unpinned it, which should give us something more like an LRU.
901 *
902 * TODO: There is a small window of opportunity for this function to
903 * get called from eviction after we've dropped the last GEM refcount,
904 * but before the TTM deleted flag is set on the object. Avoid
905 * adjusting the shrinker list in such cases, since the object is
906 * not available to the shrinker anyway due to its zero refcount.
907 * To fix this properly we should move to a TTM shrinker LRU list for
908 * these objects.
909 */
910 if (kref_get_unless_zero(&obj->base.refcount)) {
911 if (shrinkable != obj->mm.ttm_shrinkable) {
912 if (shrinkable) {
913 if (obj->mm.madv == I915_MADV_WILLNEED)
914 __i915_gem_object_make_shrinkable(obj);
915 else
916 __i915_gem_object_make_purgeable(obj);
917 } else {
918 i915_gem_object_make_unshrinkable(obj);
919 }
920
921 obj->mm.ttm_shrinkable = shrinkable;
922 }
923 i915_gem_object_put(obj);
924 }
925
926 /*
927 * Put on the correct LRU list depending on the MADV status
928 */
929 spin_lock(&bo->bdev->lru_lock);
930 if (shrinkable) {
931 /* Try to keep shmem_tt from being considered for shrinking. */
932 bo->priority = TTM_MAX_BO_PRIORITY - 1;
933 } else if (obj->mm.madv != I915_MADV_WILLNEED) {
934 bo->priority = I915_TTM_PRIO_PURGE;
935 } else if (!i915_gem_object_has_pages(obj)) {
936 bo->priority = I915_TTM_PRIO_NO_PAGES;
937 } else {
938 struct ttm_resource_manager *man =
939 ttm_manager_type(bo->bdev, bo->resource->mem_type);
940
941 /*
942 * If we need to place an LMEM resource which doesn't need CPU
943 * access then we should try not to victimize mappable objects
944 * first, since we likely end up stealing more of the mappable
945 * portion. And likewise when we try to find space for a mappble
946 * object, we know not to ever victimize objects that don't
947 * occupy any mappable pages.
948 */
949 if (i915_ttm_cpu_maps_iomem(bo->resource) &&
950 i915_ttm_buddy_man_visible_size(man) < man->size &&
951 !(obj->flags & I915_BO_ALLOC_GPU_ONLY))
952 bo->priority = I915_TTM_PRIO_NEEDS_CPU_ACCESS;
953 else
954 bo->priority = I915_TTM_PRIO_HAS_PAGES;
955 }
956
957 ttm_bo_move_to_lru_tail(bo);
958 spin_unlock(&bo->bdev->lru_lock);
959 }
960
961 /*
962 * TTM-backed gem object destruction requires some clarification.
963 * Basically we have two possibilities here. We can either rely on the
964 * i915 delayed destruction and put the TTM object when the object
965 * is idle. This would be detected by TTM which would bypass the
966 * TTM delayed destroy handling. The other approach is to put the TTM
967 * object early and rely on the TTM destroyed handling, and then free
968 * the leftover parts of the GEM object once TTM's destroyed list handling is
969 * complete. For now, we rely on the latter for two reasons:
970 * a) TTM can evict an object even when it's on the delayed destroy list,
971 * which in theory allows for complete eviction.
972 * b) There is work going on in TTM to allow freeing an object even when
973 * it's not idle, and using the TTM destroyed list handling could help us
974 * benefit from that.
975 */
i915_ttm_delayed_free(struct drm_i915_gem_object * obj)976 static void i915_ttm_delayed_free(struct drm_i915_gem_object *obj)
977 {
978 GEM_BUG_ON(!obj->ttm.created);
979
980 ttm_bo_put(i915_gem_to_ttm(obj));
981 }
982
vm_fault_ttm(struct vm_fault * vmf)983 static vm_fault_t vm_fault_ttm(struct vm_fault *vmf)
984 {
985 struct vm_area_struct *area = vmf->vma;
986 struct ttm_buffer_object *bo = area->vm_private_data;
987 struct drm_device *dev = bo->base.dev;
988 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
989 intel_wakeref_t wakeref = 0;
990 vm_fault_t ret;
991 int idx;
992
993 if (i915_ttm_is_ghost_object(bo))
994 return VM_FAULT_SIGBUS;
995
996 /* Sanity check that we allow writing into this object */
997 if (unlikely(i915_gem_object_is_readonly(obj) &&
998 area->vm_flags & VM_WRITE))
999 return VM_FAULT_SIGBUS;
1000
1001 ret = ttm_bo_vm_reserve(bo, vmf);
1002 if (ret)
1003 return ret;
1004
1005 if (obj->mm.madv != I915_MADV_WILLNEED) {
1006 dma_resv_unlock(bo->base.resv);
1007 return VM_FAULT_SIGBUS;
1008 }
1009
1010 if (!i915_ttm_resource_mappable(bo->resource)) {
1011 int err = -ENODEV;
1012 int i;
1013
1014 for (i = 0; i < obj->mm.n_placements; i++) {
1015 struct intel_memory_region *mr = obj->mm.placements[i];
1016 unsigned int flags;
1017
1018 if (!mr->io_size && mr->type != INTEL_MEMORY_SYSTEM)
1019 continue;
1020
1021 flags = obj->flags;
1022 flags &= ~I915_BO_ALLOC_GPU_ONLY;
1023 err = __i915_ttm_migrate(obj, mr, flags);
1024 if (!err)
1025 break;
1026 }
1027
1028 if (err) {
1029 drm_dbg(dev, "Unable to make resource CPU accessible\n");
1030 dma_resv_unlock(bo->base.resv);
1031 ret = VM_FAULT_SIGBUS;
1032 goto out_rpm;
1033 }
1034 }
1035
1036 if (i915_ttm_cpu_maps_iomem(bo->resource))
1037 wakeref = intel_runtime_pm_get(&to_i915(obj->base.dev)->runtime_pm);
1038
1039 if (drm_dev_enter(dev, &idx)) {
1040 ret = ttm_bo_vm_fault_reserved(vmf, vmf->vma->vm_page_prot,
1041 TTM_BO_VM_NUM_PREFAULT);
1042 drm_dev_exit(idx);
1043 } else {
1044 ret = ttm_bo_vm_dummy_page(vmf, vmf->vma->vm_page_prot);
1045 }
1046
1047 if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
1048 goto out_rpm;
1049
1050 /*
1051 * ttm_bo_vm_reserve() already has dma_resv_lock.
1052 * userfault_count is protected by dma_resv lock and rpm wakeref.
1053 */
1054 if (ret == VM_FAULT_NOPAGE && wakeref && !obj->userfault_count) {
1055 obj->userfault_count = 1;
1056 spin_lock(&to_i915(obj->base.dev)->runtime_pm.lmem_userfault_lock);
1057 list_add(&obj->userfault_link, &to_i915(obj->base.dev)->runtime_pm.lmem_userfault_list);
1058 spin_unlock(&to_i915(obj->base.dev)->runtime_pm.lmem_userfault_lock);
1059 }
1060
1061 if (wakeref & CONFIG_DRM_I915_USERFAULT_AUTOSUSPEND)
1062 intel_wakeref_auto(&to_i915(obj->base.dev)->runtime_pm.userfault_wakeref,
1063 msecs_to_jiffies_timeout(CONFIG_DRM_I915_USERFAULT_AUTOSUSPEND));
1064
1065 i915_ttm_adjust_lru(obj);
1066
1067 dma_resv_unlock(bo->base.resv);
1068
1069 out_rpm:
1070 if (wakeref)
1071 intel_runtime_pm_put(&to_i915(obj->base.dev)->runtime_pm, wakeref);
1072
1073 return ret;
1074 }
1075
1076 static int
vm_access_ttm(struct vm_area_struct * area,unsigned long addr,void * buf,int len,int write)1077 vm_access_ttm(struct vm_area_struct *area, unsigned long addr,
1078 void *buf, int len, int write)
1079 {
1080 struct drm_i915_gem_object *obj =
1081 i915_ttm_to_gem(area->vm_private_data);
1082
1083 if (i915_gem_object_is_readonly(obj) && write)
1084 return -EACCES;
1085
1086 return ttm_bo_vm_access(area, addr, buf, len, write);
1087 }
1088
ttm_vm_open(struct vm_area_struct * vma)1089 static void ttm_vm_open(struct vm_area_struct *vma)
1090 {
1091 struct drm_i915_gem_object *obj =
1092 i915_ttm_to_gem(vma->vm_private_data);
1093
1094 GEM_BUG_ON(i915_ttm_is_ghost_object(vma->vm_private_data));
1095 i915_gem_object_get(obj);
1096 }
1097
ttm_vm_close(struct vm_area_struct * vma)1098 static void ttm_vm_close(struct vm_area_struct *vma)
1099 {
1100 struct drm_i915_gem_object *obj =
1101 i915_ttm_to_gem(vma->vm_private_data);
1102
1103 GEM_BUG_ON(i915_ttm_is_ghost_object(vma->vm_private_data));
1104 i915_gem_object_put(obj);
1105 }
1106
1107 static const struct vm_operations_struct vm_ops_ttm = {
1108 .fault = vm_fault_ttm,
1109 .access = vm_access_ttm,
1110 .open = ttm_vm_open,
1111 .close = ttm_vm_close,
1112 };
1113
i915_ttm_mmap_offset(struct drm_i915_gem_object * obj)1114 static u64 i915_ttm_mmap_offset(struct drm_i915_gem_object *obj)
1115 {
1116 /* The ttm_bo must be allocated with I915_BO_ALLOC_USER */
1117 GEM_BUG_ON(!drm_mm_node_allocated(&obj->base.vma_node.vm_node));
1118
1119 return drm_vma_node_offset_addr(&obj->base.vma_node);
1120 }
1121
i915_ttm_unmap_virtual(struct drm_i915_gem_object * obj)1122 static void i915_ttm_unmap_virtual(struct drm_i915_gem_object *obj)
1123 {
1124 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
1125 intel_wakeref_t wakeref = 0;
1126
1127 assert_object_held_shared(obj);
1128
1129 if (i915_ttm_cpu_maps_iomem(bo->resource)) {
1130 wakeref = intel_runtime_pm_get(&to_i915(obj->base.dev)->runtime_pm);
1131
1132 /* userfault_count is protected by obj lock and rpm wakeref. */
1133 if (obj->userfault_count) {
1134 spin_lock(&to_i915(obj->base.dev)->runtime_pm.lmem_userfault_lock);
1135 list_del(&obj->userfault_link);
1136 spin_unlock(&to_i915(obj->base.dev)->runtime_pm.lmem_userfault_lock);
1137 obj->userfault_count = 0;
1138 }
1139 }
1140
1141 ttm_bo_unmap_virtual(i915_gem_to_ttm(obj));
1142
1143 if (wakeref)
1144 intel_runtime_pm_put(&to_i915(obj->base.dev)->runtime_pm, wakeref);
1145 }
1146
1147 static const struct drm_i915_gem_object_ops i915_gem_ttm_obj_ops = {
1148 .name = "i915_gem_object_ttm",
1149 .flags = I915_GEM_OBJECT_IS_SHRINKABLE |
1150 I915_GEM_OBJECT_SELF_MANAGED_SHRINK_LIST,
1151
1152 .get_pages = i915_ttm_get_pages,
1153 .put_pages = i915_ttm_put_pages,
1154 .truncate = i915_ttm_truncate,
1155 .shrink = i915_ttm_shrink,
1156
1157 .adjust_lru = i915_ttm_adjust_lru,
1158 .delayed_free = i915_ttm_delayed_free,
1159 .migrate = i915_ttm_migrate,
1160
1161 .mmap_offset = i915_ttm_mmap_offset,
1162 .unmap_virtual = i915_ttm_unmap_virtual,
1163 .mmap_ops = &vm_ops_ttm,
1164 };
1165
i915_ttm_bo_destroy(struct ttm_buffer_object * bo)1166 void i915_ttm_bo_destroy(struct ttm_buffer_object *bo)
1167 {
1168 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
1169
1170 i915_gem_object_release_memory_region(obj);
1171 mutex_destroy(&obj->ttm.get_io_page.lock);
1172
1173 if (obj->ttm.created) {
1174 /*
1175 * We freely manage the shrinker LRU outide of the mm.pages life
1176 * cycle. As a result when destroying the object we should be
1177 * extra paranoid and ensure we remove it from the LRU, before
1178 * we free the object.
1179 *
1180 * Touching the ttm_shrinkable outside of the object lock here
1181 * should be safe now that the last GEM object ref was dropped.
1182 */
1183 if (obj->mm.ttm_shrinkable)
1184 i915_gem_object_make_unshrinkable(obj);
1185
1186 i915_ttm_backup_free(obj);
1187
1188 /* This releases all gem object bindings to the backend. */
1189 __i915_gem_free_object(obj);
1190
1191 call_rcu(&obj->rcu, __i915_gem_free_object_rcu);
1192 } else {
1193 __i915_gem_object_fini(obj);
1194 }
1195 }
1196
1197 /**
1198 * __i915_gem_ttm_object_init - Initialize a ttm-backed i915 gem object
1199 * @mem: The initial memory region for the object.
1200 * @obj: The gem object.
1201 * @size: Object size in bytes.
1202 * @flags: gem object flags.
1203 *
1204 * Return: 0 on success, negative error code on failure.
1205 */
__i915_gem_ttm_object_init(struct intel_memory_region * mem,struct drm_i915_gem_object * obj,resource_size_t offset,resource_size_t size,resource_size_t page_size,unsigned int flags)1206 int __i915_gem_ttm_object_init(struct intel_memory_region *mem,
1207 struct drm_i915_gem_object *obj,
1208 resource_size_t offset,
1209 resource_size_t size,
1210 resource_size_t page_size,
1211 unsigned int flags)
1212 {
1213 static struct lock_class_key lock_class;
1214 struct drm_i915_private *i915 = mem->i915;
1215 struct ttm_operation_ctx ctx = {
1216 .interruptible = true,
1217 .no_wait_gpu = false,
1218 };
1219 enum ttm_bo_type bo_type;
1220 int ret;
1221
1222 drm_gem_private_object_init(&i915->drm, &obj->base, size);
1223 i915_gem_object_init(obj, &i915_gem_ttm_obj_ops, &lock_class, flags);
1224
1225 obj->bo_offset = offset;
1226
1227 /* Don't put on a region list until we're either locked or fully initialized. */
1228 obj->mm.region = mem;
1229 INIT_LIST_HEAD(&obj->mm.region_link);
1230
1231 INIT_RADIX_TREE(&obj->ttm.get_io_page.radix, GFP_KERNEL | __GFP_NOWARN);
1232 mutex_init(&obj->ttm.get_io_page.lock);
1233 bo_type = (obj->flags & I915_BO_ALLOC_USER) ? ttm_bo_type_device :
1234 ttm_bo_type_kernel;
1235
1236 obj->base.vma_node.driver_private = i915_gem_to_ttm(obj);
1237
1238 /* Forcing the page size is kernel internal only */
1239 GEM_BUG_ON(page_size && obj->mm.n_placements);
1240
1241 /*
1242 * Keep an extra shrink pin to prevent the object from being made
1243 * shrinkable too early. If the ttm_tt is ever allocated in shmem, we
1244 * drop the pin. The TTM backend manages the shrinker LRU itself,
1245 * outside of the normal mm.pages life cycle.
1246 */
1247 i915_gem_object_make_unshrinkable(obj);
1248
1249 /*
1250 * If this function fails, it will call the destructor, but
1251 * our caller still owns the object. So no freeing in the
1252 * destructor until obj->ttm.created is true.
1253 * Similarly, in delayed_destroy, we can't call ttm_bo_put()
1254 * until successful initialization.
1255 */
1256 ret = ttm_bo_init_reserved(&i915->bdev, i915_gem_to_ttm(obj), bo_type,
1257 &i915_sys_placement, page_size >> PAGE_SHIFT,
1258 &ctx, NULL, NULL, i915_ttm_bo_destroy);
1259 if (ret)
1260 return i915_ttm_err_to_gem(ret);
1261
1262 obj->ttm.created = true;
1263 i915_gem_object_release_memory_region(obj);
1264 i915_gem_object_init_memory_region(obj, mem);
1265 i915_ttm_adjust_domains_after_move(obj);
1266 i915_ttm_adjust_gem_after_move(obj);
1267 i915_gem_object_unlock(obj);
1268
1269 return 0;
1270 }
1271
1272 static const struct intel_memory_region_ops ttm_system_region_ops = {
1273 .init_object = __i915_gem_ttm_object_init,
1274 .release = intel_region_ttm_fini,
1275 };
1276
1277 struct intel_memory_region *
i915_gem_ttm_system_setup(struct drm_i915_private * i915,u16 type,u16 instance)1278 i915_gem_ttm_system_setup(struct drm_i915_private *i915,
1279 u16 type, u16 instance)
1280 {
1281 struct intel_memory_region *mr;
1282
1283 mr = intel_memory_region_create(i915, 0,
1284 totalram_pages() << PAGE_SHIFT,
1285 PAGE_SIZE, 0, 0,
1286 type, instance,
1287 &ttm_system_region_ops);
1288 if (IS_ERR(mr))
1289 return mr;
1290
1291 intel_memory_region_set_name(mr, "system-ttm");
1292 return mr;
1293 }
1294