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_placement.h>
9 #include <drm/ttm/ttm_tt.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 WARN_ON(overflows_type(offset >> PAGE_SHIFT, place->fpfn));
144 place->fpfn = offset >> PAGE_SHIFT;
145 WARN_ON(overflows_type(place->fpfn + (size >> PAGE_SHIFT), place->lpfn));
146 place->lpfn = place->fpfn + (size >> PAGE_SHIFT);
147 } else if (mr->io_size && mr->io_size < mr->total) {
148 if (flags & I915_BO_ALLOC_GPU_ONLY) {
149 place->flags |= TTM_PL_FLAG_TOPDOWN;
150 } else {
151 place->fpfn = 0;
152 WARN_ON(overflows_type(mr->io_size >> PAGE_SHIFT, place->lpfn));
153 place->lpfn = mr->io_size >> PAGE_SHIFT;
154 }
155 }
156 }
157
158 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)159 i915_ttm_placement_from_obj(const struct drm_i915_gem_object *obj,
160 struct ttm_place *requested,
161 struct ttm_place *busy,
162 struct ttm_placement *placement)
163 {
164 unsigned int num_allowed = obj->mm.n_placements;
165 unsigned int flags = obj->flags;
166 unsigned int i;
167
168 placement->num_placement = 1;
169 i915_ttm_place_from_region(num_allowed ? obj->mm.placements[0] :
170 obj->mm.region, requested, obj->bo_offset,
171 obj->base.size, flags);
172
173 /* Cache this on object? */
174 placement->num_busy_placement = num_allowed;
175 for (i = 0; i < placement->num_busy_placement; ++i)
176 i915_ttm_place_from_region(obj->mm.placements[i], busy + i,
177 obj->bo_offset, obj->base.size, flags);
178
179 if (num_allowed == 0) {
180 *busy = *requested;
181 placement->num_busy_placement = 1;
182 }
183
184 placement->placement = requested;
185 placement->busy_placement = busy;
186 }
187
i915_ttm_tt_shmem_populate(struct ttm_device * bdev,struct ttm_tt * ttm,struct ttm_operation_ctx * ctx)188 static int i915_ttm_tt_shmem_populate(struct ttm_device *bdev,
189 struct ttm_tt *ttm,
190 struct ttm_operation_ctx *ctx)
191 {
192 struct drm_i915_private *i915 = container_of(bdev, typeof(*i915), bdev);
193 struct intel_memory_region *mr = i915->mm.regions[INTEL_MEMORY_SYSTEM];
194 struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm);
195 const unsigned int max_segment = i915_sg_segment_size(i915->drm.dev);
196 const size_t size = (size_t)ttm->num_pages << PAGE_SHIFT;
197 struct file *filp = i915_tt->filp;
198 struct sgt_iter sgt_iter;
199 struct sg_table *st;
200 struct page *page;
201 unsigned long i;
202 int err;
203
204 if (!filp) {
205 struct address_space *mapping;
206 gfp_t mask;
207
208 filp = shmem_file_setup("i915-shmem-tt", size, VM_NORESERVE);
209 if (IS_ERR(filp))
210 return PTR_ERR(filp);
211
212 mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
213
214 mapping = filp->f_mapping;
215 mapping_set_gfp_mask(mapping, mask);
216 GEM_BUG_ON(!(mapping_gfp_mask(mapping) & __GFP_RECLAIM));
217
218 i915_tt->filp = filp;
219 }
220
221 st = &i915_tt->cached_rsgt.table;
222 err = shmem_sg_alloc_table(i915, st, size, mr, filp->f_mapping,
223 max_segment);
224 if (err)
225 return err;
226
227 err = dma_map_sgtable(i915_tt->dev, st, DMA_BIDIRECTIONAL,
228 DMA_ATTR_SKIP_CPU_SYNC);
229 if (err)
230 goto err_free_st;
231
232 i = 0;
233 for_each_sgt_page(page, sgt_iter, st)
234 ttm->pages[i++] = page;
235
236 if (ttm->page_flags & TTM_TT_FLAG_SWAPPED)
237 ttm->page_flags &= ~TTM_TT_FLAG_SWAPPED;
238
239 return 0;
240
241 err_free_st:
242 shmem_sg_free_table(st, filp->f_mapping, false, false);
243
244 return err;
245 }
246
i915_ttm_tt_shmem_unpopulate(struct ttm_tt * ttm)247 static void i915_ttm_tt_shmem_unpopulate(struct ttm_tt *ttm)
248 {
249 struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm);
250 bool backup = ttm->page_flags & TTM_TT_FLAG_SWAPPED;
251 struct sg_table *st = &i915_tt->cached_rsgt.table;
252
253 shmem_sg_free_table(st, file_inode(i915_tt->filp)->i_mapping,
254 backup, backup);
255 }
256
i915_ttm_tt_release(struct kref * ref)257 static void i915_ttm_tt_release(struct kref *ref)
258 {
259 struct i915_ttm_tt *i915_tt =
260 container_of(ref, typeof(*i915_tt), cached_rsgt.kref);
261 struct sg_table *st = &i915_tt->cached_rsgt.table;
262
263 GEM_WARN_ON(st->sgl);
264
265 kfree(i915_tt);
266 }
267
268 static const struct i915_refct_sgt_ops tt_rsgt_ops = {
269 .release = i915_ttm_tt_release
270 };
271
i915_ttm_tt_create(struct ttm_buffer_object * bo,uint32_t page_flags)272 static struct ttm_tt *i915_ttm_tt_create(struct ttm_buffer_object *bo,
273 uint32_t page_flags)
274 {
275 struct drm_i915_private *i915 = container_of(bo->bdev, typeof(*i915),
276 bdev);
277 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
278 unsigned long ccs_pages = 0;
279 enum ttm_caching caching;
280 struct i915_ttm_tt *i915_tt;
281 int ret;
282
283 if (i915_ttm_is_ghost_object(bo))
284 return NULL;
285
286 i915_tt = kzalloc(sizeof(*i915_tt), GFP_KERNEL);
287 if (!i915_tt)
288 return NULL;
289
290 if (obj->flags & I915_BO_ALLOC_CPU_CLEAR && (!bo->resource ||
291 ttm_manager_type(bo->bdev, bo->resource->mem_type)->use_tt))
292 page_flags |= TTM_TT_FLAG_ZERO_ALLOC;
293
294 caching = i915_ttm_select_tt_caching(obj);
295 if (i915_gem_object_is_shrinkable(obj) && caching == ttm_cached) {
296 page_flags |= TTM_TT_FLAG_EXTERNAL |
297 TTM_TT_FLAG_EXTERNAL_MAPPABLE;
298 i915_tt->is_shmem = true;
299 }
300
301 if (i915_gem_object_needs_ccs_pages(obj))
302 ccs_pages = DIV_ROUND_UP(DIV_ROUND_UP(bo->base.size,
303 NUM_BYTES_PER_CCS_BYTE),
304 PAGE_SIZE);
305
306 ret = ttm_tt_init(&i915_tt->ttm, bo, page_flags, caching, ccs_pages);
307 if (ret)
308 goto err_free;
309
310 __i915_refct_sgt_init(&i915_tt->cached_rsgt, bo->base.size,
311 &tt_rsgt_ops);
312
313 i915_tt->dev = obj->base.dev->dev;
314
315 return &i915_tt->ttm;
316
317 err_free:
318 kfree(i915_tt);
319 return NULL;
320 }
321
i915_ttm_tt_populate(struct ttm_device * bdev,struct ttm_tt * ttm,struct ttm_operation_ctx * ctx)322 static int i915_ttm_tt_populate(struct ttm_device *bdev,
323 struct ttm_tt *ttm,
324 struct ttm_operation_ctx *ctx)
325 {
326 struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm);
327
328 if (i915_tt->is_shmem)
329 return i915_ttm_tt_shmem_populate(bdev, ttm, ctx);
330
331 return ttm_pool_alloc(&bdev->pool, ttm, ctx);
332 }
333
i915_ttm_tt_unpopulate(struct ttm_device * bdev,struct ttm_tt * ttm)334 static void i915_ttm_tt_unpopulate(struct ttm_device *bdev, struct ttm_tt *ttm)
335 {
336 struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm);
337 struct sg_table *st = &i915_tt->cached_rsgt.table;
338
339 if (st->sgl)
340 dma_unmap_sgtable(i915_tt->dev, st, DMA_BIDIRECTIONAL, 0);
341
342 if (i915_tt->is_shmem) {
343 i915_ttm_tt_shmem_unpopulate(ttm);
344 } else {
345 sg_free_table(st);
346 ttm_pool_free(&bdev->pool, ttm);
347 }
348 }
349
i915_ttm_tt_destroy(struct ttm_device * bdev,struct ttm_tt * ttm)350 static void i915_ttm_tt_destroy(struct ttm_device *bdev, struct ttm_tt *ttm)
351 {
352 struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm);
353
354 if (i915_tt->filp)
355 fput(i915_tt->filp);
356
357 ttm_tt_fini(ttm);
358 i915_refct_sgt_put(&i915_tt->cached_rsgt);
359 }
360
i915_ttm_eviction_valuable(struct ttm_buffer_object * bo,const struct ttm_place * place)361 static bool i915_ttm_eviction_valuable(struct ttm_buffer_object *bo,
362 const struct ttm_place *place)
363 {
364 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
365
366 if (i915_ttm_is_ghost_object(bo))
367 return false;
368
369 /*
370 * EXTERNAL objects should never be swapped out by TTM, instead we need
371 * to handle that ourselves. TTM will already skip such objects for us,
372 * but we would like to avoid grabbing locks for no good reason.
373 */
374 if (bo->ttm && bo->ttm->page_flags & TTM_TT_FLAG_EXTERNAL)
375 return false;
376
377 /* Will do for now. Our pinned objects are still on TTM's LRU lists */
378 if (!i915_gem_object_evictable(obj))
379 return false;
380
381 return ttm_bo_eviction_valuable(bo, place);
382 }
383
i915_ttm_evict_flags(struct ttm_buffer_object * bo,struct ttm_placement * placement)384 static void i915_ttm_evict_flags(struct ttm_buffer_object *bo,
385 struct ttm_placement *placement)
386 {
387 *placement = i915_sys_placement;
388 }
389
390 /**
391 * i915_ttm_free_cached_io_rsgt - Free object cached LMEM information
392 * @obj: The GEM object
393 * This function frees any LMEM-related information that is cached on
394 * the object. For example the radix tree for fast page lookup and the
395 * cached refcounted sg-table
396 */
i915_ttm_free_cached_io_rsgt(struct drm_i915_gem_object * obj)397 void i915_ttm_free_cached_io_rsgt(struct drm_i915_gem_object *obj)
398 {
399 struct radix_tree_iter iter;
400 void __rcu **slot;
401
402 if (!obj->ttm.cached_io_rsgt)
403 return;
404
405 rcu_read_lock();
406 radix_tree_for_each_slot(slot, &obj->ttm.get_io_page.radix, &iter, 0)
407 radix_tree_delete(&obj->ttm.get_io_page.radix, iter.index);
408 rcu_read_unlock();
409
410 i915_refct_sgt_put(obj->ttm.cached_io_rsgt);
411 obj->ttm.cached_io_rsgt = NULL;
412 }
413
414 /**
415 * i915_ttm_purge - Clear an object of its memory
416 * @obj: The object
417 *
418 * This function is called to clear an object of it's memory when it is
419 * marked as not needed anymore.
420 *
421 * Return: 0 on success, negative error code on failure.
422 */
i915_ttm_purge(struct drm_i915_gem_object * obj)423 int i915_ttm_purge(struct drm_i915_gem_object *obj)
424 {
425 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
426 struct i915_ttm_tt *i915_tt =
427 container_of(bo->ttm, typeof(*i915_tt), ttm);
428 struct ttm_operation_ctx ctx = {
429 .interruptible = true,
430 .no_wait_gpu = false,
431 };
432 struct ttm_placement place = {};
433 int ret;
434
435 if (obj->mm.madv == __I915_MADV_PURGED)
436 return 0;
437
438 ret = ttm_bo_validate(bo, &place, &ctx);
439 if (ret)
440 return ret;
441
442 if (bo->ttm && i915_tt->filp) {
443 /*
444 * The below fput(which eventually calls shmem_truncate) might
445 * be delayed by worker, so when directly called to purge the
446 * pages(like by the shrinker) we should try to be more
447 * aggressive and release the pages immediately.
448 */
449 shmem_truncate_range(file_inode(i915_tt->filp),
450 0, (loff_t)-1);
451 fput(fetch_and_zero(&i915_tt->filp));
452 }
453
454 obj->write_domain = 0;
455 obj->read_domains = 0;
456 i915_ttm_adjust_gem_after_move(obj);
457 i915_ttm_free_cached_io_rsgt(obj);
458 obj->mm.madv = __I915_MADV_PURGED;
459
460 return 0;
461 }
462
i915_ttm_shrink(struct drm_i915_gem_object * obj,unsigned int flags)463 static int i915_ttm_shrink(struct drm_i915_gem_object *obj, unsigned int flags)
464 {
465 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
466 struct i915_ttm_tt *i915_tt =
467 container_of(bo->ttm, typeof(*i915_tt), ttm);
468 struct ttm_operation_ctx ctx = {
469 .interruptible = true,
470 .no_wait_gpu = flags & I915_GEM_OBJECT_SHRINK_NO_GPU_WAIT,
471 };
472 struct ttm_placement place = {};
473 int ret;
474
475 if (!bo->ttm || bo->resource->mem_type != TTM_PL_SYSTEM)
476 return 0;
477
478 GEM_BUG_ON(!i915_tt->is_shmem);
479
480 if (!i915_tt->filp)
481 return 0;
482
483 ret = ttm_bo_wait_ctx(bo, &ctx);
484 if (ret)
485 return ret;
486
487 switch (obj->mm.madv) {
488 case I915_MADV_DONTNEED:
489 return i915_ttm_purge(obj);
490 case __I915_MADV_PURGED:
491 return 0;
492 }
493
494 if (bo->ttm->page_flags & TTM_TT_FLAG_SWAPPED)
495 return 0;
496
497 bo->ttm->page_flags |= TTM_TT_FLAG_SWAPPED;
498 ret = ttm_bo_validate(bo, &place, &ctx);
499 if (ret) {
500 bo->ttm->page_flags &= ~TTM_TT_FLAG_SWAPPED;
501 return ret;
502 }
503
504 if (flags & I915_GEM_OBJECT_SHRINK_WRITEBACK)
505 __shmem_writeback(obj->base.size, i915_tt->filp->f_mapping);
506
507 return 0;
508 }
509
i915_ttm_delete_mem_notify(struct ttm_buffer_object * bo)510 static void i915_ttm_delete_mem_notify(struct ttm_buffer_object *bo)
511 {
512 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
513
514 if (bo->resource && !i915_ttm_is_ghost_object(bo)) {
515 __i915_gem_object_pages_fini(obj);
516 i915_ttm_free_cached_io_rsgt(obj);
517 }
518 }
519
i915_ttm_tt_get_st(struct ttm_tt * ttm)520 static struct i915_refct_sgt *i915_ttm_tt_get_st(struct ttm_tt *ttm)
521 {
522 struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm);
523 struct sg_table *st;
524 int ret;
525
526 if (i915_tt->cached_rsgt.table.sgl)
527 return i915_refct_sgt_get(&i915_tt->cached_rsgt);
528
529 st = &i915_tt->cached_rsgt.table;
530 ret = sg_alloc_table_from_pages_segment(st,
531 ttm->pages, ttm->num_pages,
532 0, (unsigned long)ttm->num_pages << PAGE_SHIFT,
533 i915_sg_segment_size(i915_tt->dev), GFP_KERNEL);
534 if (ret) {
535 st->sgl = NULL;
536 return ERR_PTR(ret);
537 }
538
539 ret = dma_map_sgtable(i915_tt->dev, st, DMA_BIDIRECTIONAL, 0);
540 if (ret) {
541 sg_free_table(st);
542 return ERR_PTR(ret);
543 }
544
545 return i915_refct_sgt_get(&i915_tt->cached_rsgt);
546 }
547
548 /**
549 * i915_ttm_resource_get_st - Get a refcounted sg-table pointing to the
550 * resource memory
551 * @obj: The GEM object used for sg-table caching
552 * @res: The struct ttm_resource for which an sg-table is requested.
553 *
554 * This function returns a refcounted sg-table representing the memory
555 * pointed to by @res. If @res is the object's current resource it may also
556 * cache the sg_table on the object or attempt to access an already cached
557 * sg-table. The refcounted sg-table needs to be put when no-longer in use.
558 *
559 * Return: A valid pointer to a struct i915_refct_sgt or error pointer on
560 * failure.
561 */
562 struct i915_refct_sgt *
i915_ttm_resource_get_st(struct drm_i915_gem_object * obj,struct ttm_resource * res)563 i915_ttm_resource_get_st(struct drm_i915_gem_object *obj,
564 struct ttm_resource *res)
565 {
566 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
567 u32 page_alignment;
568
569 if (!i915_ttm_gtt_binds_lmem(res))
570 return i915_ttm_tt_get_st(bo->ttm);
571
572 page_alignment = bo->page_alignment << PAGE_SHIFT;
573 if (!page_alignment)
574 page_alignment = obj->mm.region->min_page_size;
575
576 /*
577 * If CPU mapping differs, we need to add the ttm_tt pages to
578 * the resulting st. Might make sense for GGTT.
579 */
580 GEM_WARN_ON(!i915_ttm_cpu_maps_iomem(res));
581 if (bo->resource == res) {
582 if (!obj->ttm.cached_io_rsgt) {
583 struct i915_refct_sgt *rsgt;
584
585 rsgt = intel_region_ttm_resource_to_rsgt(obj->mm.region,
586 res,
587 page_alignment);
588 if (IS_ERR(rsgt))
589 return rsgt;
590
591 obj->ttm.cached_io_rsgt = rsgt;
592 }
593 return i915_refct_sgt_get(obj->ttm.cached_io_rsgt);
594 }
595
596 return intel_region_ttm_resource_to_rsgt(obj->mm.region, res,
597 page_alignment);
598 }
599
i915_ttm_truncate(struct drm_i915_gem_object * obj)600 static int i915_ttm_truncate(struct drm_i915_gem_object *obj)
601 {
602 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
603 long err;
604
605 WARN_ON_ONCE(obj->mm.madv == I915_MADV_WILLNEED);
606
607 err = dma_resv_wait_timeout(bo->base.resv, DMA_RESV_USAGE_BOOKKEEP,
608 true, 15 * HZ);
609 if (err < 0)
610 return err;
611 if (err == 0)
612 return -EBUSY;
613
614 err = i915_ttm_move_notify(bo);
615 if (err)
616 return err;
617
618 return i915_ttm_purge(obj);
619 }
620
i915_ttm_swap_notify(struct ttm_buffer_object * bo)621 static void i915_ttm_swap_notify(struct ttm_buffer_object *bo)
622 {
623 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
624 int ret;
625
626 if (i915_ttm_is_ghost_object(bo))
627 return;
628
629 ret = i915_ttm_move_notify(bo);
630 GEM_WARN_ON(ret);
631 GEM_WARN_ON(obj->ttm.cached_io_rsgt);
632 if (!ret && obj->mm.madv != I915_MADV_WILLNEED)
633 i915_ttm_purge(obj);
634 }
635
636 /**
637 * i915_ttm_resource_mappable - Return true if the ttm resource is CPU
638 * accessible.
639 * @res: The TTM resource to check.
640 *
641 * This is interesting on small-BAR systems where we may encounter lmem objects
642 * that can't be accessed via the CPU.
643 */
i915_ttm_resource_mappable(struct ttm_resource * res)644 bool i915_ttm_resource_mappable(struct ttm_resource *res)
645 {
646 struct i915_ttm_buddy_resource *bman_res = to_ttm_buddy_resource(res);
647
648 if (!i915_ttm_cpu_maps_iomem(res))
649 return true;
650
651 return bman_res->used_visible_size == PFN_UP(bman_res->base.size);
652 }
653
i915_ttm_io_mem_reserve(struct ttm_device * bdev,struct ttm_resource * mem)654 static int i915_ttm_io_mem_reserve(struct ttm_device *bdev, struct ttm_resource *mem)
655 {
656 struct drm_i915_gem_object *obj = i915_ttm_to_gem(mem->bo);
657 bool unknown_state;
658
659 if (i915_ttm_is_ghost_object(mem->bo))
660 return -EINVAL;
661
662 if (!kref_get_unless_zero(&obj->base.refcount))
663 return -EINVAL;
664
665 assert_object_held(obj);
666
667 unknown_state = i915_gem_object_has_unknown_state(obj);
668 i915_gem_object_put(obj);
669 if (unknown_state)
670 return -EINVAL;
671
672 if (!i915_ttm_cpu_maps_iomem(mem))
673 return 0;
674
675 if (!i915_ttm_resource_mappable(mem))
676 return -EINVAL;
677
678 mem->bus.caching = ttm_write_combined;
679 mem->bus.is_iomem = true;
680
681 return 0;
682 }
683
i915_ttm_io_mem_pfn(struct ttm_buffer_object * bo,unsigned long page_offset)684 static unsigned long i915_ttm_io_mem_pfn(struct ttm_buffer_object *bo,
685 unsigned long page_offset)
686 {
687 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
688 struct scatterlist *sg;
689 unsigned long base;
690 unsigned int ofs;
691
692 GEM_BUG_ON(i915_ttm_is_ghost_object(bo));
693 GEM_WARN_ON(bo->ttm);
694
695 base = obj->mm.region->iomap.base - obj->mm.region->region.start;
696 sg = i915_gem_object_page_iter_get_sg(obj, &obj->ttm.get_io_page, page_offset, &ofs);
697
698 return ((base + sg_dma_address(sg)) >> PAGE_SHIFT) + ofs;
699 }
700
i915_ttm_access_memory(struct ttm_buffer_object * bo,unsigned long offset,void * buf,int len,int write)701 static int i915_ttm_access_memory(struct ttm_buffer_object *bo,
702 unsigned long offset, void *buf,
703 int len, int write)
704 {
705 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
706 resource_size_t iomap = obj->mm.region->iomap.base -
707 obj->mm.region->region.start;
708 unsigned long page = offset >> PAGE_SHIFT;
709 unsigned long bytes_left = len;
710
711 /*
712 * TODO: For now just let it fail if the resource is non-mappable,
713 * otherwise we need to perform the memcpy from the gpu here, without
714 * interfering with the object (like moving the entire thing).
715 */
716 if (!i915_ttm_resource_mappable(bo->resource))
717 return -EIO;
718
719 offset -= page << PAGE_SHIFT;
720 do {
721 unsigned long bytes = min(bytes_left, PAGE_SIZE - offset);
722 void __iomem *ptr;
723 dma_addr_t daddr;
724
725 daddr = i915_gem_object_get_dma_address(obj, page);
726 ptr = ioremap_wc(iomap + daddr + offset, bytes);
727 if (!ptr)
728 return -EIO;
729
730 if (write)
731 memcpy_toio(ptr, buf, bytes);
732 else
733 memcpy_fromio(buf, ptr, bytes);
734 iounmap(ptr);
735
736 page++;
737 buf += bytes;
738 bytes_left -= bytes;
739 offset = 0;
740 } while (bytes_left);
741
742 return len;
743 }
744
745 /*
746 * All callbacks need to take care not to downcast a struct ttm_buffer_object
747 * without checking its subclass, since it might be a TTM ghost object.
748 */
749 static struct ttm_device_funcs i915_ttm_bo_driver = {
750 .ttm_tt_create = i915_ttm_tt_create,
751 .ttm_tt_populate = i915_ttm_tt_populate,
752 .ttm_tt_unpopulate = i915_ttm_tt_unpopulate,
753 .ttm_tt_destroy = i915_ttm_tt_destroy,
754 .eviction_valuable = i915_ttm_eviction_valuable,
755 .evict_flags = i915_ttm_evict_flags,
756 .move = i915_ttm_move,
757 .swap_notify = i915_ttm_swap_notify,
758 .delete_mem_notify = i915_ttm_delete_mem_notify,
759 .io_mem_reserve = i915_ttm_io_mem_reserve,
760 .io_mem_pfn = i915_ttm_io_mem_pfn,
761 .access_memory = i915_ttm_access_memory,
762 };
763
764 /**
765 * i915_ttm_driver - Return a pointer to the TTM device funcs
766 *
767 * Return: Pointer to statically allocated TTM device funcs.
768 */
i915_ttm_driver(void)769 struct ttm_device_funcs *i915_ttm_driver(void)
770 {
771 return &i915_ttm_bo_driver;
772 }
773
__i915_ttm_get_pages(struct drm_i915_gem_object * obj,struct ttm_placement * placement)774 static int __i915_ttm_get_pages(struct drm_i915_gem_object *obj,
775 struct ttm_placement *placement)
776 {
777 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
778 struct ttm_operation_ctx ctx = {
779 .interruptible = true,
780 .no_wait_gpu = false,
781 };
782 int real_num_busy;
783 int ret;
784
785 /* First try only the requested placement. No eviction. */
786 real_num_busy = fetch_and_zero(&placement->num_busy_placement);
787 ret = ttm_bo_validate(bo, placement, &ctx);
788 if (ret) {
789 ret = i915_ttm_err_to_gem(ret);
790 /*
791 * Anything that wants to restart the operation gets to
792 * do that.
793 */
794 if (ret == -EDEADLK || ret == -EINTR || ret == -ERESTARTSYS ||
795 ret == -EAGAIN)
796 return ret;
797
798 /*
799 * If the initial attempt fails, allow all accepted placements,
800 * evicting if necessary.
801 */
802 placement->num_busy_placement = real_num_busy;
803 ret = ttm_bo_validate(bo, placement, &ctx);
804 if (ret)
805 return i915_ttm_err_to_gem(ret);
806 }
807
808 if (bo->ttm && !ttm_tt_is_populated(bo->ttm)) {
809 ret = ttm_tt_populate(bo->bdev, bo->ttm, &ctx);
810 if (ret)
811 return ret;
812
813 i915_ttm_adjust_domains_after_move(obj);
814 i915_ttm_adjust_gem_after_move(obj);
815 }
816
817 if (!i915_gem_object_has_pages(obj)) {
818 struct i915_refct_sgt *rsgt =
819 i915_ttm_resource_get_st(obj, bo->resource);
820
821 if (IS_ERR(rsgt))
822 return PTR_ERR(rsgt);
823
824 GEM_BUG_ON(obj->mm.rsgt);
825 obj->mm.rsgt = rsgt;
826 __i915_gem_object_set_pages(obj, &rsgt->table);
827 }
828
829 GEM_BUG_ON(bo->ttm && ((obj->base.size >> PAGE_SHIFT) < bo->ttm->num_pages));
830 i915_ttm_adjust_lru(obj);
831 return ret;
832 }
833
i915_ttm_get_pages(struct drm_i915_gem_object * obj)834 static int i915_ttm_get_pages(struct drm_i915_gem_object *obj)
835 {
836 struct ttm_place requested, busy[I915_TTM_MAX_PLACEMENTS];
837 struct ttm_placement placement;
838
839 /* restricted by sg_alloc_table */
840 if (overflows_type(obj->base.size >> PAGE_SHIFT, unsigned int))
841 return -E2BIG;
842
843 GEM_BUG_ON(obj->mm.n_placements > I915_TTM_MAX_PLACEMENTS);
844
845 /* Move to the requested placement. */
846 i915_ttm_placement_from_obj(obj, &requested, busy, &placement);
847
848 return __i915_ttm_get_pages(obj, &placement);
849 }
850
851 /**
852 * DOC: Migration vs eviction
853 *
854 * GEM migration may not be the same as TTM migration / eviction. If
855 * the TTM core decides to evict an object it may be evicted to a
856 * TTM memory type that is not in the object's allowable GEM regions, or
857 * in fact theoretically to a TTM memory type that doesn't correspond to
858 * a GEM memory region. In that case the object's GEM region is not
859 * updated, and the data is migrated back to the GEM region at
860 * get_pages time. TTM may however set up CPU ptes to the object even
861 * when it is evicted.
862 * Gem forced migration using the i915_ttm_migrate() op, is allowed even
863 * to regions that are not in the object's list of allowable placements.
864 */
__i915_ttm_migrate(struct drm_i915_gem_object * obj,struct intel_memory_region * mr,unsigned int flags)865 static int __i915_ttm_migrate(struct drm_i915_gem_object *obj,
866 struct intel_memory_region *mr,
867 unsigned int flags)
868 {
869 struct ttm_place requested;
870 struct ttm_placement placement;
871 int ret;
872
873 i915_ttm_place_from_region(mr, &requested, obj->bo_offset,
874 obj->base.size, flags);
875 placement.num_placement = 1;
876 placement.num_busy_placement = 1;
877 placement.placement = &requested;
878 placement.busy_placement = &requested;
879
880 ret = __i915_ttm_get_pages(obj, &placement);
881 if (ret)
882 return ret;
883
884 /*
885 * Reinitialize the region bindings. This is primarily
886 * required for objects where the new region is not in
887 * its allowable placements.
888 */
889 if (obj->mm.region != mr) {
890 i915_gem_object_release_memory_region(obj);
891 i915_gem_object_init_memory_region(obj, mr);
892 }
893
894 return 0;
895 }
896
i915_ttm_migrate(struct drm_i915_gem_object * obj,struct intel_memory_region * mr,unsigned int flags)897 static int i915_ttm_migrate(struct drm_i915_gem_object *obj,
898 struct intel_memory_region *mr,
899 unsigned int flags)
900 {
901 return __i915_ttm_migrate(obj, mr, flags);
902 }
903
i915_ttm_put_pages(struct drm_i915_gem_object * obj,struct sg_table * st)904 static void i915_ttm_put_pages(struct drm_i915_gem_object *obj,
905 struct sg_table *st)
906 {
907 /*
908 * We're currently not called from a shrinker, so put_pages()
909 * typically means the object is about to destroyed, or called
910 * from move_notify(). So just avoid doing much for now.
911 * If the object is not destroyed next, The TTM eviction logic
912 * and shrinkers will move it out if needed.
913 */
914
915 if (obj->mm.rsgt)
916 i915_refct_sgt_put(fetch_and_zero(&obj->mm.rsgt));
917 }
918
919 /**
920 * i915_ttm_adjust_lru - Adjust an object's position on relevant LRU lists.
921 * @obj: The object
922 */
i915_ttm_adjust_lru(struct drm_i915_gem_object * obj)923 void i915_ttm_adjust_lru(struct drm_i915_gem_object *obj)
924 {
925 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
926 struct i915_ttm_tt *i915_tt =
927 container_of(bo->ttm, typeof(*i915_tt), ttm);
928 bool shrinkable =
929 bo->ttm && i915_tt->filp && ttm_tt_is_populated(bo->ttm);
930
931 /*
932 * Don't manipulate the TTM LRUs while in TTM bo destruction.
933 * We're called through i915_ttm_delete_mem_notify().
934 */
935 if (!kref_read(&bo->kref))
936 return;
937
938 /*
939 * We skip managing the shrinker LRU in set_pages() and just manage
940 * everything here. This does at least solve the issue with having
941 * temporary shmem mappings(like with evicted lmem) not being visible to
942 * the shrinker. Only our shmem objects are shrinkable, everything else
943 * we keep as unshrinkable.
944 *
945 * To make sure everything plays nice we keep an extra shrink pin in TTM
946 * if the underlying pages are not currently shrinkable. Once we release
947 * our pin, like when the pages are moved to shmem, the pages will then
948 * be added to the shrinker LRU, assuming the caller isn't also holding
949 * a pin.
950 *
951 * TODO: consider maybe also bumping the shrinker list here when we have
952 * already unpinned it, which should give us something more like an LRU.
953 *
954 * TODO: There is a small window of opportunity for this function to
955 * get called from eviction after we've dropped the last GEM refcount,
956 * but before the TTM deleted flag is set on the object. Avoid
957 * adjusting the shrinker list in such cases, since the object is
958 * not available to the shrinker anyway due to its zero refcount.
959 * To fix this properly we should move to a TTM shrinker LRU list for
960 * these objects.
961 */
962 if (kref_get_unless_zero(&obj->base.refcount)) {
963 if (shrinkable != obj->mm.ttm_shrinkable) {
964 if (shrinkable) {
965 if (obj->mm.madv == I915_MADV_WILLNEED)
966 __i915_gem_object_make_shrinkable(obj);
967 else
968 __i915_gem_object_make_purgeable(obj);
969 } else {
970 i915_gem_object_make_unshrinkable(obj);
971 }
972
973 obj->mm.ttm_shrinkable = shrinkable;
974 }
975 i915_gem_object_put(obj);
976 }
977
978 /*
979 * Put on the correct LRU list depending on the MADV status
980 */
981 spin_lock(&bo->bdev->lru_lock);
982 if (shrinkable) {
983 /* Try to keep shmem_tt from being considered for shrinking. */
984 bo->priority = TTM_MAX_BO_PRIORITY - 1;
985 } else if (obj->mm.madv != I915_MADV_WILLNEED) {
986 bo->priority = I915_TTM_PRIO_PURGE;
987 } else if (!i915_gem_object_has_pages(obj)) {
988 bo->priority = I915_TTM_PRIO_NO_PAGES;
989 } else {
990 struct ttm_resource_manager *man =
991 ttm_manager_type(bo->bdev, bo->resource->mem_type);
992
993 /*
994 * If we need to place an LMEM resource which doesn't need CPU
995 * access then we should try not to victimize mappable objects
996 * first, since we likely end up stealing more of the mappable
997 * portion. And likewise when we try to find space for a mappble
998 * object, we know not to ever victimize objects that don't
999 * occupy any mappable pages.
1000 */
1001 if (i915_ttm_cpu_maps_iomem(bo->resource) &&
1002 i915_ttm_buddy_man_visible_size(man) < man->size &&
1003 !(obj->flags & I915_BO_ALLOC_GPU_ONLY))
1004 bo->priority = I915_TTM_PRIO_NEEDS_CPU_ACCESS;
1005 else
1006 bo->priority = I915_TTM_PRIO_HAS_PAGES;
1007 }
1008
1009 ttm_bo_move_to_lru_tail(bo);
1010 spin_unlock(&bo->bdev->lru_lock);
1011 }
1012
1013 /*
1014 * TTM-backed gem object destruction requires some clarification.
1015 * Basically we have two possibilities here. We can either rely on the
1016 * i915 delayed destruction and put the TTM object when the object
1017 * is idle. This would be detected by TTM which would bypass the
1018 * TTM delayed destroy handling. The other approach is to put the TTM
1019 * object early and rely on the TTM destroyed handling, and then free
1020 * the leftover parts of the GEM object once TTM's destroyed list handling is
1021 * complete. For now, we rely on the latter for two reasons:
1022 * a) TTM can evict an object even when it's on the delayed destroy list,
1023 * which in theory allows for complete eviction.
1024 * b) There is work going on in TTM to allow freeing an object even when
1025 * it's not idle, and using the TTM destroyed list handling could help us
1026 * benefit from that.
1027 */
i915_ttm_delayed_free(struct drm_i915_gem_object * obj)1028 static void i915_ttm_delayed_free(struct drm_i915_gem_object *obj)
1029 {
1030 GEM_BUG_ON(!obj->ttm.created);
1031
1032 ttm_bo_put(i915_gem_to_ttm(obj));
1033 }
1034
vm_fault_ttm(struct vm_fault * vmf)1035 static vm_fault_t vm_fault_ttm(struct vm_fault *vmf)
1036 {
1037 struct vm_area_struct *area = vmf->vma;
1038 struct ttm_buffer_object *bo = area->vm_private_data;
1039 struct drm_device *dev = bo->base.dev;
1040 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
1041 intel_wakeref_t wakeref = 0;
1042 vm_fault_t ret;
1043 int idx;
1044
1045 /* Sanity check that we allow writing into this object */
1046 if (unlikely(i915_gem_object_is_readonly(obj) &&
1047 area->vm_flags & VM_WRITE))
1048 return VM_FAULT_SIGBUS;
1049
1050 ret = ttm_bo_vm_reserve(bo, vmf);
1051 if (ret)
1052 return ret;
1053
1054 if (obj->mm.madv != I915_MADV_WILLNEED) {
1055 dma_resv_unlock(bo->base.resv);
1056 return VM_FAULT_SIGBUS;
1057 }
1058
1059 /*
1060 * This must be swapped out with shmem ttm_tt (pipeline-gutting).
1061 * Calling ttm_bo_validate() here with TTM_PL_SYSTEM should only go as
1062 * far as far doing a ttm_bo_move_null(), which should skip all the
1063 * other junk.
1064 */
1065 if (!bo->resource) {
1066 struct ttm_operation_ctx ctx = {
1067 .interruptible = true,
1068 .no_wait_gpu = true, /* should be idle already */
1069 };
1070
1071 GEM_BUG_ON(!bo->ttm || !(bo->ttm->page_flags & TTM_TT_FLAG_SWAPPED));
1072
1073 ret = ttm_bo_validate(bo, i915_ttm_sys_placement(), &ctx);
1074 if (ret) {
1075 dma_resv_unlock(bo->base.resv);
1076 return VM_FAULT_SIGBUS;
1077 }
1078 } else if (!i915_ttm_resource_mappable(bo->resource)) {
1079 int err = -ENODEV;
1080 int i;
1081
1082 for (i = 0; i < obj->mm.n_placements; i++) {
1083 struct intel_memory_region *mr = obj->mm.placements[i];
1084 unsigned int flags;
1085
1086 if (!mr->io_size && mr->type != INTEL_MEMORY_SYSTEM)
1087 continue;
1088
1089 flags = obj->flags;
1090 flags &= ~I915_BO_ALLOC_GPU_ONLY;
1091 err = __i915_ttm_migrate(obj, mr, flags);
1092 if (!err)
1093 break;
1094 }
1095
1096 if (err) {
1097 drm_dbg(dev, "Unable to make resource CPU accessible(err = %pe)\n",
1098 ERR_PTR(err));
1099 dma_resv_unlock(bo->base.resv);
1100 ret = VM_FAULT_SIGBUS;
1101 goto out_rpm;
1102 }
1103 }
1104
1105 if (i915_ttm_cpu_maps_iomem(bo->resource))
1106 wakeref = intel_runtime_pm_get(&to_i915(obj->base.dev)->runtime_pm);
1107
1108 if (drm_dev_enter(dev, &idx)) {
1109 ret = ttm_bo_vm_fault_reserved(vmf, vmf->vma->vm_page_prot,
1110 TTM_BO_VM_NUM_PREFAULT);
1111 drm_dev_exit(idx);
1112 } else {
1113 ret = ttm_bo_vm_dummy_page(vmf, vmf->vma->vm_page_prot);
1114 }
1115
1116 if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
1117 goto out_rpm;
1118
1119 /*
1120 * ttm_bo_vm_reserve() already has dma_resv_lock.
1121 * userfault_count is protected by dma_resv lock and rpm wakeref.
1122 */
1123 if (ret == VM_FAULT_NOPAGE && wakeref && !obj->userfault_count) {
1124 obj->userfault_count = 1;
1125 spin_lock(&to_i915(obj->base.dev)->runtime_pm.lmem_userfault_lock);
1126 list_add(&obj->userfault_link, &to_i915(obj->base.dev)->runtime_pm.lmem_userfault_list);
1127 spin_unlock(&to_i915(obj->base.dev)->runtime_pm.lmem_userfault_lock);
1128
1129 GEM_WARN_ON(!i915_ttm_cpu_maps_iomem(bo->resource));
1130 }
1131
1132 if (wakeref & CONFIG_DRM_I915_USERFAULT_AUTOSUSPEND)
1133 intel_wakeref_auto(&to_i915(obj->base.dev)->runtime_pm.userfault_wakeref,
1134 msecs_to_jiffies_timeout(CONFIG_DRM_I915_USERFAULT_AUTOSUSPEND));
1135
1136 i915_ttm_adjust_lru(obj);
1137
1138 dma_resv_unlock(bo->base.resv);
1139
1140 out_rpm:
1141 if (wakeref)
1142 intel_runtime_pm_put(&to_i915(obj->base.dev)->runtime_pm, wakeref);
1143
1144 return ret;
1145 }
1146
1147 static int
vm_access_ttm(struct vm_area_struct * area,unsigned long addr,void * buf,int len,int write)1148 vm_access_ttm(struct vm_area_struct *area, unsigned long addr,
1149 void *buf, int len, int write)
1150 {
1151 struct drm_i915_gem_object *obj =
1152 i915_ttm_to_gem(area->vm_private_data);
1153
1154 if (i915_gem_object_is_readonly(obj) && write)
1155 return -EACCES;
1156
1157 return ttm_bo_vm_access(area, addr, buf, len, write);
1158 }
1159
ttm_vm_open(struct vm_area_struct * vma)1160 static void ttm_vm_open(struct vm_area_struct *vma)
1161 {
1162 struct drm_i915_gem_object *obj =
1163 i915_ttm_to_gem(vma->vm_private_data);
1164
1165 GEM_BUG_ON(i915_ttm_is_ghost_object(vma->vm_private_data));
1166 i915_gem_object_get(obj);
1167 }
1168
ttm_vm_close(struct vm_area_struct * vma)1169 static void ttm_vm_close(struct vm_area_struct *vma)
1170 {
1171 struct drm_i915_gem_object *obj =
1172 i915_ttm_to_gem(vma->vm_private_data);
1173
1174 GEM_BUG_ON(i915_ttm_is_ghost_object(vma->vm_private_data));
1175 i915_gem_object_put(obj);
1176 }
1177
1178 static const struct vm_operations_struct vm_ops_ttm = {
1179 .fault = vm_fault_ttm,
1180 .access = vm_access_ttm,
1181 .open = ttm_vm_open,
1182 .close = ttm_vm_close,
1183 };
1184
i915_ttm_mmap_offset(struct drm_i915_gem_object * obj)1185 static u64 i915_ttm_mmap_offset(struct drm_i915_gem_object *obj)
1186 {
1187 /* The ttm_bo must be allocated with I915_BO_ALLOC_USER */
1188 GEM_BUG_ON(!drm_mm_node_allocated(&obj->base.vma_node.vm_node));
1189
1190 return drm_vma_node_offset_addr(&obj->base.vma_node);
1191 }
1192
i915_ttm_unmap_virtual(struct drm_i915_gem_object * obj)1193 static void i915_ttm_unmap_virtual(struct drm_i915_gem_object *obj)
1194 {
1195 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
1196 intel_wakeref_t wakeref = 0;
1197
1198 assert_object_held_shared(obj);
1199
1200 if (i915_ttm_cpu_maps_iomem(bo->resource)) {
1201 wakeref = intel_runtime_pm_get(&to_i915(obj->base.dev)->runtime_pm);
1202
1203 /* userfault_count is protected by obj lock and rpm wakeref. */
1204 if (obj->userfault_count) {
1205 spin_lock(&to_i915(obj->base.dev)->runtime_pm.lmem_userfault_lock);
1206 list_del(&obj->userfault_link);
1207 spin_unlock(&to_i915(obj->base.dev)->runtime_pm.lmem_userfault_lock);
1208 obj->userfault_count = 0;
1209 }
1210 }
1211
1212 GEM_WARN_ON(obj->userfault_count);
1213
1214 ttm_bo_unmap_virtual(i915_gem_to_ttm(obj));
1215
1216 if (wakeref)
1217 intel_runtime_pm_put(&to_i915(obj->base.dev)->runtime_pm, wakeref);
1218 }
1219
1220 static const struct drm_i915_gem_object_ops i915_gem_ttm_obj_ops = {
1221 .name = "i915_gem_object_ttm",
1222 .flags = I915_GEM_OBJECT_IS_SHRINKABLE |
1223 I915_GEM_OBJECT_SELF_MANAGED_SHRINK_LIST,
1224
1225 .get_pages = i915_ttm_get_pages,
1226 .put_pages = i915_ttm_put_pages,
1227 .truncate = i915_ttm_truncate,
1228 .shrink = i915_ttm_shrink,
1229
1230 .adjust_lru = i915_ttm_adjust_lru,
1231 .delayed_free = i915_ttm_delayed_free,
1232 .migrate = i915_ttm_migrate,
1233
1234 .mmap_offset = i915_ttm_mmap_offset,
1235 .unmap_virtual = i915_ttm_unmap_virtual,
1236 .mmap_ops = &vm_ops_ttm,
1237 };
1238
i915_ttm_bo_destroy(struct ttm_buffer_object * bo)1239 void i915_ttm_bo_destroy(struct ttm_buffer_object *bo)
1240 {
1241 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
1242
1243 i915_gem_object_release_memory_region(obj);
1244 mutex_destroy(&obj->ttm.get_io_page.lock);
1245
1246 if (obj->ttm.created) {
1247 /*
1248 * We freely manage the shrinker LRU outide of the mm.pages life
1249 * cycle. As a result when destroying the object we should be
1250 * extra paranoid and ensure we remove it from the LRU, before
1251 * we free the object.
1252 *
1253 * Touching the ttm_shrinkable outside of the object lock here
1254 * should be safe now that the last GEM object ref was dropped.
1255 */
1256 if (obj->mm.ttm_shrinkable)
1257 i915_gem_object_make_unshrinkable(obj);
1258
1259 i915_ttm_backup_free(obj);
1260
1261 /* This releases all gem object bindings to the backend. */
1262 __i915_gem_free_object(obj);
1263
1264 call_rcu(&obj->rcu, __i915_gem_free_object_rcu);
1265 } else {
1266 __i915_gem_object_fini(obj);
1267 }
1268 }
1269
1270 /**
1271 * __i915_gem_ttm_object_init - Initialize a ttm-backed i915 gem object
1272 * @mem: The initial memory region for the object.
1273 * @obj: The gem object.
1274 * @size: Object size in bytes.
1275 * @flags: gem object flags.
1276 *
1277 * Return: 0 on success, negative error code on failure.
1278 */
__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)1279 int __i915_gem_ttm_object_init(struct intel_memory_region *mem,
1280 struct drm_i915_gem_object *obj,
1281 resource_size_t offset,
1282 resource_size_t size,
1283 resource_size_t page_size,
1284 unsigned int flags)
1285 {
1286 static struct lock_class_key lock_class;
1287 struct drm_i915_private *i915 = mem->i915;
1288 struct ttm_operation_ctx ctx = {
1289 .interruptible = true,
1290 .no_wait_gpu = false,
1291 };
1292 enum ttm_bo_type bo_type;
1293 int ret;
1294
1295 drm_gem_private_object_init(&i915->drm, &obj->base, size);
1296 i915_gem_object_init(obj, &i915_gem_ttm_obj_ops, &lock_class, flags);
1297
1298 obj->bo_offset = offset;
1299
1300 /* Don't put on a region list until we're either locked or fully initialized. */
1301 obj->mm.region = mem;
1302 INIT_LIST_HEAD(&obj->mm.region_link);
1303
1304 INIT_RADIX_TREE(&obj->ttm.get_io_page.radix, GFP_KERNEL | __GFP_NOWARN);
1305 mutex_init(&obj->ttm.get_io_page.lock);
1306 bo_type = (obj->flags & I915_BO_ALLOC_USER) ? ttm_bo_type_device :
1307 ttm_bo_type_kernel;
1308
1309 obj->base.vma_node.driver_private = i915_gem_to_ttm(obj);
1310
1311 /* Forcing the page size is kernel internal only */
1312 GEM_BUG_ON(page_size && obj->mm.n_placements);
1313
1314 /*
1315 * Keep an extra shrink pin to prevent the object from being made
1316 * shrinkable too early. If the ttm_tt is ever allocated in shmem, we
1317 * drop the pin. The TTM backend manages the shrinker LRU itself,
1318 * outside of the normal mm.pages life cycle.
1319 */
1320 i915_gem_object_make_unshrinkable(obj);
1321
1322 /*
1323 * If this function fails, it will call the destructor, but
1324 * our caller still owns the object. So no freeing in the
1325 * destructor until obj->ttm.created is true.
1326 * Similarly, in delayed_destroy, we can't call ttm_bo_put()
1327 * until successful initialization.
1328 */
1329 ret = ttm_bo_init_reserved(&i915->bdev, i915_gem_to_ttm(obj), bo_type,
1330 &i915_sys_placement, page_size >> PAGE_SHIFT,
1331 &ctx, NULL, NULL, i915_ttm_bo_destroy);
1332
1333 /*
1334 * XXX: The ttm_bo_init_reserved() functions returns -ENOSPC if the size
1335 * is too big to add vma. The direct function that returns -ENOSPC is
1336 * drm_mm_insert_node_in_range(). To handle the same error as other code
1337 * that returns -E2BIG when the size is too large, it converts -ENOSPC to
1338 * -E2BIG.
1339 */
1340 if (size >> PAGE_SHIFT > INT_MAX && ret == -ENOSPC)
1341 ret = -E2BIG;
1342
1343 if (ret)
1344 return i915_ttm_err_to_gem(ret);
1345
1346 obj->ttm.created = true;
1347 i915_gem_object_release_memory_region(obj);
1348 i915_gem_object_init_memory_region(obj, mem);
1349 i915_ttm_adjust_domains_after_move(obj);
1350 i915_ttm_adjust_gem_after_move(obj);
1351 i915_gem_object_unlock(obj);
1352
1353 return 0;
1354 }
1355
1356 static const struct intel_memory_region_ops ttm_system_region_ops = {
1357 .init_object = __i915_gem_ttm_object_init,
1358 .release = intel_region_ttm_fini,
1359 };
1360
1361 struct intel_memory_region *
i915_gem_ttm_system_setup(struct drm_i915_private * i915,u16 type,u16 instance)1362 i915_gem_ttm_system_setup(struct drm_i915_private *i915,
1363 u16 type, u16 instance)
1364 {
1365 struct intel_memory_region *mr;
1366
1367 mr = intel_memory_region_create(i915, 0,
1368 totalram_pages() << PAGE_SHIFT,
1369 PAGE_SIZE, 0, 0,
1370 type, instance,
1371 &ttm_system_region_ops);
1372 if (IS_ERR(mr))
1373 return mr;
1374
1375 intel_memory_region_set_name(mr, "system-ttm");
1376 return mr;
1377 }
1378