1 /* SPDX-License-Identifier: MIT */
2 /*
3 * Copyright © 2016 Intel Corporation
4 */
5
6 #ifndef __I915_GEM_OBJECT_TYPES_H__
7 #define __I915_GEM_OBJECT_TYPES_H__
8
9 #include <linux/mmu_notifier.h>
10
11 #include <drm/drm_gem.h>
12 #include <drm/ttm/ttm_bo.h>
13 #include <uapi/drm/i915_drm.h>
14
15 #include "i915_active.h"
16 #include "i915_selftest.h"
17 #include "i915_vma_resource.h"
18
19 #include "gt/intel_gt_defines.h"
20
21 struct drm_i915_gem_object;
22 struct intel_fronbuffer;
23 struct intel_memory_region;
24
25 /*
26 * struct i915_lut_handle tracks the fast lookups from handle to vma used
27 * for execbuf. Although we use a radixtree for that mapping, in order to
28 * remove them as the object or context is closed, we need a secondary list
29 * and a translation entry (i915_lut_handle).
30 */
31 struct i915_lut_handle {
32 struct list_head obj_link;
33 struct i915_gem_context *ctx;
34 u32 handle;
35 };
36
37 struct drm_i915_gem_object_ops {
38 unsigned int flags;
39 #define I915_GEM_OBJECT_IS_SHRINKABLE BIT(1)
40 /* Skip the shrinker management in set_pages/unset_pages */
41 #define I915_GEM_OBJECT_SELF_MANAGED_SHRINK_LIST BIT(2)
42 #define I915_GEM_OBJECT_IS_PROXY BIT(3)
43 #define I915_GEM_OBJECT_NO_MMAP BIT(4)
44
45 /* Interface between the GEM object and its backing storage.
46 * get_pages() is called once prior to the use of the associated set
47 * of pages before to binding them into the GTT, and put_pages() is
48 * called after we no longer need them. As we expect there to be
49 * associated cost with migrating pages between the backing storage
50 * and making them available for the GPU (e.g. clflush), we may hold
51 * onto the pages after they are no longer referenced by the GPU
52 * in case they may be used again shortly (for example migrating the
53 * pages to a different memory domain within the GTT). put_pages()
54 * will therefore most likely be called when the object itself is
55 * being released or under memory pressure (where we attempt to
56 * reap pages for the shrinker).
57 */
58 int (*get_pages)(struct drm_i915_gem_object *obj);
59 void (*put_pages)(struct drm_i915_gem_object *obj,
60 struct sg_table *pages);
61 int (*truncate)(struct drm_i915_gem_object *obj);
62 /**
63 * shrink - Perform further backend specific actions to facilate
64 * shrinking.
65 * @obj: The gem object
66 * @flags: Extra flags to control shrinking behaviour in the backend
67 *
68 * Possible values for @flags:
69 *
70 * I915_GEM_OBJECT_SHRINK_WRITEBACK - Try to perform writeback of the
71 * backing pages, if supported.
72 *
73 * I915_GEM_OBJECT_SHRINK_NO_GPU_WAIT - Don't wait for the object to
74 * idle. Active objects can be considered later. The TTM backend for
75 * example might have aync migrations going on, which don't use any
76 * i915_vma to track the active GTT binding, and hence having an unbound
77 * object might not be enough.
78 */
79 #define I915_GEM_OBJECT_SHRINK_WRITEBACK BIT(0)
80 #define I915_GEM_OBJECT_SHRINK_NO_GPU_WAIT BIT(1)
81 int (*shrink)(struct drm_i915_gem_object *obj, unsigned int flags);
82
83 int (*pread)(struct drm_i915_gem_object *obj,
84 const struct drm_i915_gem_pread *arg);
85 int (*pwrite)(struct drm_i915_gem_object *obj,
86 const struct drm_i915_gem_pwrite *arg);
87 u64 (*mmap_offset)(struct drm_i915_gem_object *obj);
88 void (*unmap_virtual)(struct drm_i915_gem_object *obj);
89
90 int (*dmabuf_export)(struct drm_i915_gem_object *obj);
91
92 /**
93 * adjust_lru - notify that the madvise value was updated
94 * @obj: The gem object
95 *
96 * The madvise value may have been updated, or object was recently
97 * referenced so act accordingly (Perhaps changing an LRU list etc).
98 */
99 void (*adjust_lru)(struct drm_i915_gem_object *obj);
100
101 /**
102 * delayed_free - Override the default delayed free implementation
103 */
104 void (*delayed_free)(struct drm_i915_gem_object *obj);
105
106 /**
107 * migrate - Migrate object to a different region either for
108 * pinning or for as long as the object lock is held.
109 */
110 int (*migrate)(struct drm_i915_gem_object *obj,
111 struct intel_memory_region *mr,
112 unsigned int flags);
113
114 void (*release)(struct drm_i915_gem_object *obj);
115
116 const struct vm_operations_struct *mmap_ops;
117 const char *name; /* friendly name for debug, e.g. lockdep classes */
118 };
119
120 /**
121 * enum i915_cache_level - The supported GTT caching values for system memory
122 * pages.
123 *
124 * These translate to some special GTT PTE bits when binding pages into some
125 * address space. It also determines whether an object, or rather its pages are
126 * coherent with the GPU, when also reading or writing through the CPU cache
127 * with those pages.
128 *
129 * Userspace can also control this through struct drm_i915_gem_caching.
130 */
131 enum i915_cache_level {
132 /**
133 * @I915_CACHE_NONE:
134 *
135 * GPU access is not coherent with the CPU cache. If the cache is dirty
136 * and we need the underlying pages to be coherent with some later GPU
137 * access then we need to manually flush the pages.
138 *
139 * On shared LLC platforms reads and writes through the CPU cache are
140 * still coherent even with this setting. See also
141 * &drm_i915_gem_object.cache_coherent for more details. Due to this we
142 * should only ever use uncached for scanout surfaces, otherwise we end
143 * up over-flushing in some places.
144 *
145 * This is the default on non-LLC platforms.
146 */
147 I915_CACHE_NONE = 0,
148 /**
149 * @I915_CACHE_LLC:
150 *
151 * GPU access is coherent with the CPU cache. If the cache is dirty,
152 * then the GPU will ensure that access remains coherent, when both
153 * reading and writing through the CPU cache. GPU writes can dirty the
154 * CPU cache.
155 *
156 * Not used for scanout surfaces.
157 *
158 * Applies to both platforms with shared LLC(HAS_LLC), and snooping
159 * based platforms(HAS_SNOOP).
160 *
161 * This is the default on shared LLC platforms. The only exception is
162 * scanout objects, where the display engine is not coherent with the
163 * CPU cache. For such objects I915_CACHE_NONE or I915_CACHE_WT is
164 * automatically applied by the kernel in pin_for_display, if userspace
165 * has not done so already.
166 */
167 I915_CACHE_LLC,
168 /**
169 * @I915_CACHE_L3_LLC:
170 *
171 * Explicitly enable the Gfx L3 cache, with coherent LLC.
172 *
173 * The Gfx L3 sits between the domain specific caches, e.g
174 * sampler/render caches, and the larger LLC. LLC is coherent with the
175 * GPU, but L3 is only visible to the GPU, so likely needs to be flushed
176 * when the workload completes.
177 *
178 * Not used for scanout surfaces.
179 *
180 * Only exposed on some gen7 + GGTT. More recent hardware has dropped
181 * this explicit setting, where it should now be enabled by default.
182 */
183 I915_CACHE_L3_LLC,
184 /**
185 * @I915_CACHE_WT:
186 *
187 * Write-through. Used for scanout surfaces.
188 *
189 * The GPU can utilise the caches, while still having the display engine
190 * be coherent with GPU writes, as a result we don't need to flush the
191 * CPU caches when moving out of the render domain. This is the default
192 * setting chosen by the kernel, if supported by the HW, otherwise we
193 * fallback to I915_CACHE_NONE. On the CPU side writes through the CPU
194 * cache still need to be flushed, to remain coherent with the display
195 * engine.
196 */
197 I915_CACHE_WT,
198 /**
199 * @I915_MAX_CACHE_LEVEL:
200 *
201 * Mark the last entry in the enum. Used for defining cachelevel_to_pat
202 * array for cache_level to pat translation table.
203 */
204 I915_MAX_CACHE_LEVEL,
205 };
206
207 enum i915_map_type {
208 I915_MAP_WB = 0,
209 I915_MAP_WC,
210 #define I915_MAP_OVERRIDE BIT(31)
211 I915_MAP_FORCE_WB = I915_MAP_WB | I915_MAP_OVERRIDE,
212 I915_MAP_FORCE_WC = I915_MAP_WC | I915_MAP_OVERRIDE,
213 };
214
215 enum i915_mmap_type {
216 I915_MMAP_TYPE_GTT = 0,
217 I915_MMAP_TYPE_WC,
218 I915_MMAP_TYPE_WB,
219 I915_MMAP_TYPE_UC,
220 I915_MMAP_TYPE_FIXED,
221 };
222
223 struct i915_mmap_offset {
224 struct drm_vma_offset_node vma_node;
225 struct drm_i915_gem_object *obj;
226 enum i915_mmap_type mmap_type;
227
228 struct rb_node offset;
229 };
230
231 struct i915_gem_object_page_iter {
232 struct scatterlist *sg_pos;
233 unsigned int sg_idx; /* in pages, but 32bit eek! */
234
235 struct radix_tree_root radix;
236 struct mutex lock; /* protects this cache */
237 };
238
239 struct drm_i915_gem_object {
240 /*
241 * We might have reason to revisit the below since it wastes
242 * a lot of space for non-ttm gem objects.
243 * In any case, always use the accessors for the ttm_buffer_object
244 * when accessing it.
245 */
246 union {
247 struct drm_gem_object base;
248 struct ttm_buffer_object __do_not_access;
249 };
250
251 const struct drm_i915_gem_object_ops *ops;
252
253 struct {
254 /**
255 * @vma.lock: protect the list/tree of vmas
256 */
257 spinlock_t lock;
258
259 /**
260 * @vma.list: List of VMAs backed by this object
261 *
262 * The VMA on this list are ordered by type, all GGTT vma are
263 * placed at the head and all ppGTT vma are placed at the tail.
264 * The different types of GGTT vma are unordered between
265 * themselves, use the @vma.tree (which has a defined order
266 * between all VMA) to quickly find an exact match.
267 */
268 struct list_head list;
269
270 /**
271 * @vma.tree: Ordered tree of VMAs backed by this object
272 *
273 * All VMA created for this object are placed in the @vma.tree
274 * for fast retrieval via a binary search in
275 * i915_vma_instance(). They are also added to @vma.list for
276 * easy iteration.
277 */
278 struct rb_root tree;
279 } vma;
280
281 /**
282 * @lut_list: List of vma lookup entries in use for this object.
283 *
284 * If this object is closed, we need to remove all of its VMA from
285 * the fast lookup index in associated contexts; @lut_list provides
286 * this translation from object to context->handles_vma.
287 */
288 struct list_head lut_list;
289 spinlock_t lut_lock; /* guards lut_list */
290
291 /**
292 * @obj_link: Link into @i915_gem_ww_ctx.obj_list
293 *
294 * When we lock this object through i915_gem_object_lock() with a
295 * context, we add it to the list to ensure we can unlock everything
296 * when i915_gem_ww_ctx_backoff() or i915_gem_ww_ctx_fini() are called.
297 */
298 struct list_head obj_link;
299 /**
300 * @shared_resv_from: The object shares the resv from this vm.
301 */
302 struct i915_address_space *shares_resv_from;
303
304 #ifdef CONFIG_PROC_FS
305 /**
306 * @client: @i915_drm_client which created the object
307 */
308 struct i915_drm_client *client;
309
310 /**
311 * @client_link: Link into @i915_drm_client.objects_list
312 */
313 struct list_head client_link;
314 #endif
315
316 union {
317 struct rcu_head rcu;
318 struct llist_node freed;
319 };
320
321 /**
322 * Whether the object is currently in the GGTT or any other supported
323 * fake offset mmap backed by lmem.
324 */
325 unsigned int userfault_count;
326 struct list_head userfault_link;
327
328 struct {
329 spinlock_t lock; /* Protects access to mmo offsets */
330 struct rb_root offsets;
331 } mmo;
332
333 I915_SELFTEST_DECLARE(struct list_head st_link);
334
335 unsigned long flags;
336 #define I915_BO_ALLOC_CONTIGUOUS BIT(0)
337 #define I915_BO_ALLOC_VOLATILE BIT(1)
338 #define I915_BO_ALLOC_CPU_CLEAR BIT(2)
339 #define I915_BO_ALLOC_USER BIT(3)
340 /* Object is allowed to lose its contents on suspend / resume, even if pinned */
341 #define I915_BO_ALLOC_PM_VOLATILE BIT(4)
342 /* Object needs to be restored early using memcpy during resume */
343 #define I915_BO_ALLOC_PM_EARLY BIT(5)
344 /*
345 * Object is likely never accessed by the CPU. This will prioritise the BO to be
346 * allocated in the non-mappable portion of lmem. This is merely a hint, and if
347 * dealing with userspace objects the CPU fault handler is free to ignore this.
348 */
349 #define I915_BO_ALLOC_GPU_ONLY BIT(6)
350 #define I915_BO_ALLOC_CCS_AUX BIT(7)
351 /*
352 * Object is allowed to retain its initial data and will not be cleared on first
353 * access if used along with I915_BO_ALLOC_USER. This is mainly to keep
354 * preallocated framebuffer data intact while transitioning it to i915drmfb.
355 */
356 #define I915_BO_PREALLOC BIT(8)
357 #define I915_BO_ALLOC_FLAGS (I915_BO_ALLOC_CONTIGUOUS | \
358 I915_BO_ALLOC_VOLATILE | \
359 I915_BO_ALLOC_CPU_CLEAR | \
360 I915_BO_ALLOC_USER | \
361 I915_BO_ALLOC_PM_VOLATILE | \
362 I915_BO_ALLOC_PM_EARLY | \
363 I915_BO_ALLOC_GPU_ONLY | \
364 I915_BO_ALLOC_CCS_AUX | \
365 I915_BO_PREALLOC)
366 #define I915_BO_READONLY BIT(9)
367 #define I915_TILING_QUIRK_BIT 10 /* unknown swizzling; do not release! */
368 #define I915_BO_PROTECTED BIT(11)
369 /**
370 * @mem_flags - Mutable placement-related flags
371 *
372 * These are flags that indicate specifics of the memory region
373 * the object is currently in. As such they are only stable
374 * either under the object lock or if the object is pinned.
375 */
376 unsigned int mem_flags;
377 #define I915_BO_FLAG_STRUCT_PAGE BIT(0) /* Object backed by struct pages */
378 #define I915_BO_FLAG_IOMEM BIT(1) /* Object backed by IO memory */
379 /**
380 * @pat_index: The desired PAT index.
381 *
382 * See hardware specification for valid PAT indices for each platform.
383 * This field replaces the @cache_level that contains a value of enum
384 * i915_cache_level since PAT indices are being used by both userspace
385 * and kernel mode driver for caching policy control after GEN12.
386 * In the meantime platform specific tables are created to translate
387 * i915_cache_level into pat index, for more details check the macros
388 * defined i915/i915_pci.c, e.g. TGL_CACHELEVEL.
389 * For backward compatibility, this field contains values exactly match
390 * the entries of enum i915_cache_level for pre-GEN12 platforms (See
391 * LEGACY_CACHELEVEL), so that the PTE encode functions for these
392 * legacy platforms can stay the same.
393 */
394 unsigned int pat_index:6;
395 /**
396 * @pat_set_by_user: Indicate whether pat_index is set by user space
397 *
398 * This field is set to false by default, only set to true if the
399 * pat_index is set by user space. By design, user space is capable of
400 * managing caching behavior by setting pat_index, in which case this
401 * kernel mode driver should never touch the pat_index.
402 */
403 unsigned int pat_set_by_user:1;
404 /**
405 * @cache_coherent:
406 *
407 * Note: with the change above which replaced @cache_level with pat_index,
408 * the use of @cache_coherent is limited to the objects created by kernel
409 * or by userspace without pat index specified.
410 * Check for @pat_set_by_user to find out if an object has pat index set
411 * by userspace. The ioctl's to change cache settings have also been
412 * disabled for the objects with pat index set by userspace. Please don't
413 * assume @cache_coherent having the flags set as describe here. A helper
414 * function i915_gem_object_has_cache_level() provides one way to bypass
415 * the use of this field.
416 *
417 * Track whether the pages are coherent with the GPU if reading or
418 * writing through the CPU caches. The largely depends on the
419 * @cache_level setting.
420 *
421 * On platforms which don't have the shared LLC(HAS_SNOOP), like on Atom
422 * platforms, coherency must be explicitly requested with some special
423 * GTT caching bits(see enum i915_cache_level). When enabling coherency
424 * it does come at a performance and power cost on such platforms. On
425 * the flip side the kernel does not need to manually flush any buffers
426 * which need to be coherent with the GPU, if the object is not coherent
427 * i.e @cache_coherent is zero.
428 *
429 * On platforms that share the LLC with the CPU(HAS_LLC), all GT memory
430 * access will automatically snoop the CPU caches(even with CACHE_NONE).
431 * The one exception is when dealing with the display engine, like with
432 * scanout surfaces. To handle this the kernel will always flush the
433 * surface out of the CPU caches when preparing it for scanout. Also
434 * note that since scanout surfaces are only ever read by the display
435 * engine we only need to care about flushing any writes through the CPU
436 * cache, reads on the other hand will always be coherent.
437 *
438 * Something strange here is why @cache_coherent is not a simple
439 * boolean, i.e coherent vs non-coherent. The reasoning for this is back
440 * to the display engine not being fully coherent. As a result scanout
441 * surfaces will either be marked as I915_CACHE_NONE or I915_CACHE_WT.
442 * In the case of seeing I915_CACHE_NONE the kernel makes the assumption
443 * that this is likely a scanout surface, and will set @cache_coherent
444 * as only I915_BO_CACHE_COHERENT_FOR_READ, on platforms with the shared
445 * LLC. The kernel uses this to always flush writes through the CPU
446 * cache as early as possible, where it can, in effect keeping
447 * @cache_dirty clean, so we can potentially avoid stalling when
448 * flushing the surface just before doing the scanout. This does mean
449 * we might unnecessarily flush non-scanout objects in some places, but
450 * the default assumption is that all normal objects should be using
451 * I915_CACHE_LLC, at least on platforms with the shared LLC.
452 *
453 * Supported values:
454 *
455 * I915_BO_CACHE_COHERENT_FOR_READ:
456 *
457 * On shared LLC platforms, we use this for special scanout surfaces,
458 * where the display engine is not coherent with the CPU cache. As such
459 * we need to ensure we flush any writes before doing the scanout. As an
460 * optimisation we try to flush any writes as early as possible to avoid
461 * stalling later.
462 *
463 * Thus for scanout surfaces using I915_CACHE_NONE, on shared LLC
464 * platforms, we use:
465 *
466 * cache_coherent = I915_BO_CACHE_COHERENT_FOR_READ
467 *
468 * While for normal objects that are fully coherent, including special
469 * scanout surfaces marked as I915_CACHE_WT, we use:
470 *
471 * cache_coherent = I915_BO_CACHE_COHERENT_FOR_READ |
472 * I915_BO_CACHE_COHERENT_FOR_WRITE
473 *
474 * And then for objects that are not coherent at all we use:
475 *
476 * cache_coherent = 0
477 *
478 * I915_BO_CACHE_COHERENT_FOR_WRITE:
479 *
480 * When writing through the CPU cache, the GPU is still coherent. Note
481 * that this also implies I915_BO_CACHE_COHERENT_FOR_READ.
482 */
483 #define I915_BO_CACHE_COHERENT_FOR_READ BIT(0)
484 #define I915_BO_CACHE_COHERENT_FOR_WRITE BIT(1)
485 unsigned int cache_coherent:2;
486
487 /**
488 * @cache_dirty:
489 *
490 * Note: with the change above which replaced cache_level with pat_index,
491 * the use of @cache_dirty is limited to the objects created by kernel
492 * or by userspace without pat index specified.
493 * Check for @pat_set_by_user to find out if an object has pat index set
494 * by userspace. The ioctl's to change cache settings have also been
495 * disabled for the objects with pat_index set by userspace. Please don't
496 * assume @cache_dirty is set as describe here. Also see helper function
497 * i915_gem_object_has_cache_level() for possible ways to bypass the use
498 * of this field.
499 *
500 * Track if we are we dirty with writes through the CPU cache for this
501 * object. As a result reading directly from main memory might yield
502 * stale data.
503 *
504 * This also ties into whether the kernel is tracking the object as
505 * coherent with the GPU, as per @cache_coherent, as it determines if
506 * flushing might be needed at various points.
507 *
508 * Another part of @cache_dirty is managing flushing when first
509 * acquiring the pages for system memory, at this point the pages are
510 * considered foreign, so the default assumption is that the cache is
511 * dirty, for example the page zeroing done by the kernel might leave
512 * writes though the CPU cache, or swapping-in, while the actual data in
513 * main memory is potentially stale. Note that this is a potential
514 * security issue when dealing with userspace objects and zeroing. Now,
515 * whether we actually need apply the big sledgehammer of flushing all
516 * the pages on acquire depends on if @cache_coherent is marked as
517 * I915_BO_CACHE_COHERENT_FOR_WRITE, i.e that the GPU will be coherent
518 * for both reads and writes though the CPU cache.
519 *
520 * Note that on shared LLC platforms we still apply the heavy flush for
521 * I915_CACHE_NONE objects, under the assumption that this is going to
522 * be used for scanout.
523 *
524 * Update: On some hardware there is now also the 'Bypass LLC' MOCS
525 * entry, which defeats our @cache_coherent tracking, since userspace
526 * can freely bypass the CPU cache when touching the pages with the GPU,
527 * where the kernel is completely unaware. On such platform we need
528 * apply the sledgehammer-on-acquire regardless of the @cache_coherent.
529 *
530 * Special care is taken on non-LLC platforms, to prevent potential
531 * information leak. The driver currently ensures:
532 *
533 * 1. All userspace objects, by default, have @cache_level set as
534 * I915_CACHE_NONE. The only exception is userptr objects, where we
535 * instead force I915_CACHE_LLC, but we also don't allow userspace to
536 * ever change the @cache_level for such objects. Another special case
537 * is dma-buf, which doesn't rely on @cache_dirty, but there we
538 * always do a forced flush when acquiring the pages, if there is a
539 * chance that the pages can be read directly from main memory with
540 * the GPU.
541 *
542 * 2. All I915_CACHE_NONE objects have @cache_dirty initially true.
543 *
544 * 3. All swapped-out objects(i.e shmem) have @cache_dirty set to
545 * true.
546 *
547 * 4. The @cache_dirty is never freely reset before the initial
548 * flush, even if userspace adjusts the @cache_level through the
549 * i915_gem_set_caching_ioctl.
550 *
551 * 5. All @cache_dirty objects(including swapped-in) are initially
552 * flushed with a synchronous call to drm_clflush_sg in
553 * __i915_gem_object_set_pages. The @cache_dirty can be freely reset
554 * at this point. All further asynchronous clfushes are never security
555 * critical, i.e userspace is free to race against itself.
556 */
557 unsigned int cache_dirty:1;
558
559 /* @is_dpt: Object houses a display page table (DPT) */
560 unsigned int is_dpt:1;
561
562 /**
563 * @read_domains: Read memory domains.
564 *
565 * These monitor which caches contain read/write data related to the
566 * object. When transitioning from one set of domains to another,
567 * the driver is called to ensure that caches are suitably flushed and
568 * invalidated.
569 */
570 u16 read_domains;
571
572 /**
573 * @write_domain: Corresponding unique write memory domain.
574 */
575 u16 write_domain;
576
577 struct intel_frontbuffer __rcu *frontbuffer;
578
579 /** Current tiling stride for the object, if it's tiled. */
580 unsigned int tiling_and_stride;
581 #define FENCE_MINIMUM_STRIDE 128 /* See i915_tiling_ok() */
582 #define TILING_MASK (FENCE_MINIMUM_STRIDE - 1)
583 #define STRIDE_MASK (~TILING_MASK)
584
585 struct {
586 /*
587 * Protects the pages and their use. Do not use directly, but
588 * instead go through the pin/unpin interfaces.
589 */
590 atomic_t pages_pin_count;
591
592 /**
593 * @shrink_pin: Prevents the pages from being made visible to
594 * the shrinker, while the shrink_pin is non-zero. Most users
595 * should pretty much never have to care about this, outside of
596 * some special use cases.
597 *
598 * By default most objects will start out as visible to the
599 * shrinker(if I915_GEM_OBJECT_IS_SHRINKABLE) as soon as the
600 * backing pages are attached to the object, like in
601 * __i915_gem_object_set_pages(). They will then be removed the
602 * shrinker list once the pages are released.
603 *
604 * The @shrink_pin is incremented by calling
605 * i915_gem_object_make_unshrinkable(), which will also remove
606 * the object from the shrinker list, if the pin count was zero.
607 *
608 * Callers will then typically call
609 * i915_gem_object_make_shrinkable() or
610 * i915_gem_object_make_purgeable() to decrement the pin count,
611 * and make the pages visible again.
612 */
613 atomic_t shrink_pin;
614
615 /**
616 * @ttm_shrinkable: True when the object is using shmem pages
617 * underneath. Protected by the object lock.
618 */
619 bool ttm_shrinkable;
620
621 /**
622 * @unknown_state: Indicate that the object is effectively
623 * borked. This is write-once and set if we somehow encounter a
624 * fatal error when moving/clearing the pages, and we are not
625 * able to fallback to memcpy/memset, like on small-BAR systems.
626 * The GPU should also be wedged (or in the process) at this
627 * point.
628 *
629 * Only valid to read this after acquiring the dma-resv lock and
630 * waiting for all DMA_RESV_USAGE_KERNEL fences to be signalled,
631 * or if we otherwise know that the moving fence has signalled,
632 * and we are certain the pages underneath are valid for
633 * immediate access (under normal operation), like just prior to
634 * binding the object or when setting up the CPU fault handler.
635 * See i915_gem_object_has_unknown_state();
636 */
637 bool unknown_state;
638
639 /**
640 * Priority list of potential placements for this object.
641 */
642 struct intel_memory_region **placements;
643 int n_placements;
644
645 /**
646 * Memory region for this object.
647 */
648 struct intel_memory_region *region;
649
650 /**
651 * Memory manager resource allocated for this object. Only
652 * needed for the mock region.
653 */
654 struct ttm_resource *res;
655
656 /**
657 * Element within memory_region->objects or region->purgeable
658 * if the object is marked as DONTNEED. Access is protected by
659 * region->obj_lock.
660 */
661 struct list_head region_link;
662
663 struct i915_refct_sgt *rsgt;
664 struct sg_table *pages;
665 void *mapping;
666
667 struct i915_page_sizes page_sizes;
668
669 I915_SELFTEST_DECLARE(unsigned int page_mask);
670
671 struct i915_gem_object_page_iter get_page;
672 struct i915_gem_object_page_iter get_dma_page;
673
674 /**
675 * Element within i915->mm.shrink_list or i915->mm.purge_list,
676 * locked by i915->mm.obj_lock.
677 */
678 struct list_head link;
679
680 /**
681 * Advice: are the backing pages purgeable?
682 */
683 unsigned int madv:2;
684
685 /**
686 * This is set if the object has been written to since the
687 * pages were last acquired.
688 */
689 bool dirty:1;
690
691 u32 tlb[I915_MAX_GT];
692 } mm;
693
694 struct {
695 struct i915_refct_sgt *cached_io_rsgt;
696 struct i915_gem_object_page_iter get_io_page;
697 struct drm_i915_gem_object *backup;
698 bool created:1;
699 } ttm;
700
701 /*
702 * Record which PXP key instance this object was created against (if
703 * any), so we can use it to determine if the encryption is valid by
704 * comparing against the current key instance.
705 */
706 u32 pxp_key_instance;
707
708 /** Record of address bit 17 of each page at last unbind. */
709 unsigned long *bit_17;
710
711 union {
712 #ifdef CONFIG_MMU_NOTIFIER
713 struct i915_gem_userptr {
714 uintptr_t ptr;
715 unsigned long notifier_seq;
716
717 struct mmu_interval_notifier notifier;
718 struct page **pvec;
719 int page_ref;
720 } userptr;
721 #endif
722
723 struct drm_mm_node *stolen;
724
725 resource_size_t bo_offset;
726
727 unsigned long scratch;
728 u64 encode;
729
730 void *gvt_info;
731 };
732 };
733
734 #define intel_bo_to_drm_bo(bo) (&(bo)->base)
735 #define intel_bo_to_i915(bo) to_i915(intel_bo_to_drm_bo(bo)->dev)
736
737 static inline struct drm_i915_gem_object *
to_intel_bo(struct drm_gem_object * gem)738 to_intel_bo(struct drm_gem_object *gem)
739 {
740 /* Assert that to_intel_bo(NULL) == NULL */
741 BUILD_BUG_ON(offsetof(struct drm_i915_gem_object, base));
742
743 return container_of(gem, struct drm_i915_gem_object, base);
744 }
745
746 #endif
747