ASoC: Merge up v6.6-rc7
[linux-block.git] / drivers / gpu / drm / i915 / gem / i915_gem_pages.c
1 /*
2  * SPDX-License-Identifier: MIT
3  *
4  * Copyright © 2014-2016 Intel Corporation
5  */
6
7 #include <drm/drm_cache.h>
8
9 #include "gt/intel_gt.h"
10 #include "gt/intel_tlb.h"
11
12 #include "i915_drv.h"
13 #include "i915_gem_object.h"
14 #include "i915_scatterlist.h"
15 #include "i915_gem_lmem.h"
16 #include "i915_gem_mman.h"
17
18 void __i915_gem_object_set_pages(struct drm_i915_gem_object *obj,
19                                  struct sg_table *pages)
20 {
21         struct drm_i915_private *i915 = to_i915(obj->base.dev);
22         unsigned long supported = RUNTIME_INFO(i915)->page_sizes;
23         bool shrinkable;
24         int i;
25
26         assert_object_held_shared(obj);
27
28         if (i915_gem_object_is_volatile(obj))
29                 obj->mm.madv = I915_MADV_DONTNEED;
30
31         /* Make the pages coherent with the GPU (flushing any swapin). */
32         if (obj->cache_dirty) {
33                 WARN_ON_ONCE(IS_DGFX(i915));
34                 obj->write_domain = 0;
35                 if (i915_gem_object_has_struct_page(obj))
36                         drm_clflush_sg(pages);
37                 obj->cache_dirty = false;
38         }
39
40         obj->mm.get_page.sg_pos = pages->sgl;
41         obj->mm.get_page.sg_idx = 0;
42         obj->mm.get_dma_page.sg_pos = pages->sgl;
43         obj->mm.get_dma_page.sg_idx = 0;
44
45         obj->mm.pages = pages;
46
47         obj->mm.page_sizes.phys = i915_sg_dma_sizes(pages->sgl);
48         GEM_BUG_ON(!obj->mm.page_sizes.phys);
49
50         /*
51          * Calculate the supported page-sizes which fit into the given
52          * sg_page_sizes. This will give us the page-sizes which we may be able
53          * to use opportunistically when later inserting into the GTT. For
54          * example if phys=2G, then in theory we should be able to use 1G, 2M,
55          * 64K or 4K pages, although in practice this will depend on a number of
56          * other factors.
57          */
58         obj->mm.page_sizes.sg = 0;
59         for_each_set_bit(i, &supported, ilog2(I915_GTT_MAX_PAGE_SIZE) + 1) {
60                 if (obj->mm.page_sizes.phys & ~0u << i)
61                         obj->mm.page_sizes.sg |= BIT(i);
62         }
63         GEM_BUG_ON(!HAS_PAGE_SIZES(i915, obj->mm.page_sizes.sg));
64
65         shrinkable = i915_gem_object_is_shrinkable(obj);
66
67         if (i915_gem_object_is_tiled(obj) &&
68             i915->gem_quirks & GEM_QUIRK_PIN_SWIZZLED_PAGES) {
69                 GEM_BUG_ON(i915_gem_object_has_tiling_quirk(obj));
70                 i915_gem_object_set_tiling_quirk(obj);
71                 GEM_BUG_ON(!list_empty(&obj->mm.link));
72                 atomic_inc(&obj->mm.shrink_pin);
73                 shrinkable = false;
74         }
75
76         if (shrinkable && !i915_gem_object_has_self_managed_shrink_list(obj)) {
77                 struct list_head *list;
78                 unsigned long flags;
79
80                 assert_object_held(obj);
81                 spin_lock_irqsave(&i915->mm.obj_lock, flags);
82
83                 i915->mm.shrink_count++;
84                 i915->mm.shrink_memory += obj->base.size;
85
86                 if (obj->mm.madv != I915_MADV_WILLNEED)
87                         list = &i915->mm.purge_list;
88                 else
89                         list = &i915->mm.shrink_list;
90                 list_add_tail(&obj->mm.link, list);
91
92                 atomic_set(&obj->mm.shrink_pin, 0);
93                 spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
94         }
95 }
96
97 int ____i915_gem_object_get_pages(struct drm_i915_gem_object *obj)
98 {
99         struct drm_i915_private *i915 = to_i915(obj->base.dev);
100         int err;
101
102         assert_object_held_shared(obj);
103
104         if (unlikely(obj->mm.madv != I915_MADV_WILLNEED)) {
105                 drm_dbg(&i915->drm,
106                         "Attempting to obtain a purgeable object\n");
107                 return -EFAULT;
108         }
109
110         err = obj->ops->get_pages(obj);
111         GEM_BUG_ON(!err && !i915_gem_object_has_pages(obj));
112
113         return err;
114 }
115
116 /* Ensure that the associated pages are gathered from the backing storage
117  * and pinned into our object. i915_gem_object_pin_pages() may be called
118  * multiple times before they are released by a single call to
119  * i915_gem_object_unpin_pages() - once the pages are no longer referenced
120  * either as a result of memory pressure (reaping pages under the shrinker)
121  * or as the object is itself released.
122  */
123 int __i915_gem_object_get_pages(struct drm_i915_gem_object *obj)
124 {
125         int err;
126
127         assert_object_held(obj);
128
129         assert_object_held_shared(obj);
130
131         if (unlikely(!i915_gem_object_has_pages(obj))) {
132                 GEM_BUG_ON(i915_gem_object_has_pinned_pages(obj));
133
134                 err = ____i915_gem_object_get_pages(obj);
135                 if (err)
136                         return err;
137
138                 smp_mb__before_atomic();
139         }
140         atomic_inc(&obj->mm.pages_pin_count);
141
142         return 0;
143 }
144
145 int i915_gem_object_pin_pages_unlocked(struct drm_i915_gem_object *obj)
146 {
147         struct i915_gem_ww_ctx ww;
148         int err;
149
150         i915_gem_ww_ctx_init(&ww, true);
151 retry:
152         err = i915_gem_object_lock(obj, &ww);
153         if (!err)
154                 err = i915_gem_object_pin_pages(obj);
155
156         if (err == -EDEADLK) {
157                 err = i915_gem_ww_ctx_backoff(&ww);
158                 if (!err)
159                         goto retry;
160         }
161         i915_gem_ww_ctx_fini(&ww);
162         return err;
163 }
164
165 /* Immediately discard the backing storage */
166 int i915_gem_object_truncate(struct drm_i915_gem_object *obj)
167 {
168         if (obj->ops->truncate)
169                 return obj->ops->truncate(obj);
170
171         return 0;
172 }
173
174 static void __i915_gem_object_reset_page_iter(struct drm_i915_gem_object *obj)
175 {
176         struct radix_tree_iter iter;
177         void __rcu **slot;
178
179         rcu_read_lock();
180         radix_tree_for_each_slot(slot, &obj->mm.get_page.radix, &iter, 0)
181                 radix_tree_delete(&obj->mm.get_page.radix, iter.index);
182         radix_tree_for_each_slot(slot, &obj->mm.get_dma_page.radix, &iter, 0)
183                 radix_tree_delete(&obj->mm.get_dma_page.radix, iter.index);
184         rcu_read_unlock();
185 }
186
187 static void unmap_object(struct drm_i915_gem_object *obj, void *ptr)
188 {
189         if (is_vmalloc_addr(ptr))
190                 vunmap(ptr);
191 }
192
193 static void flush_tlb_invalidate(struct drm_i915_gem_object *obj)
194 {
195         struct drm_i915_private *i915 = to_i915(obj->base.dev);
196         struct intel_gt *gt;
197         int id;
198
199         for_each_gt(gt, i915, id) {
200                 if (!obj->mm.tlb[id])
201                         continue;
202
203                 intel_gt_invalidate_tlb_full(gt, obj->mm.tlb[id]);
204                 obj->mm.tlb[id] = 0;
205         }
206 }
207
208 struct sg_table *
209 __i915_gem_object_unset_pages(struct drm_i915_gem_object *obj)
210 {
211         struct sg_table *pages;
212
213         assert_object_held_shared(obj);
214
215         pages = fetch_and_zero(&obj->mm.pages);
216         if (IS_ERR_OR_NULL(pages))
217                 return pages;
218
219         if (i915_gem_object_is_volatile(obj))
220                 obj->mm.madv = I915_MADV_WILLNEED;
221
222         if (!i915_gem_object_has_self_managed_shrink_list(obj))
223                 i915_gem_object_make_unshrinkable(obj);
224
225         if (obj->mm.mapping) {
226                 unmap_object(obj, page_mask_bits(obj->mm.mapping));
227                 obj->mm.mapping = NULL;
228         }
229
230         __i915_gem_object_reset_page_iter(obj);
231         obj->mm.page_sizes.phys = obj->mm.page_sizes.sg = 0;
232
233         flush_tlb_invalidate(obj);
234
235         return pages;
236 }
237
238 int __i915_gem_object_put_pages(struct drm_i915_gem_object *obj)
239 {
240         struct sg_table *pages;
241
242         if (i915_gem_object_has_pinned_pages(obj))
243                 return -EBUSY;
244
245         /* May be called by shrinker from within get_pages() (on another bo) */
246         assert_object_held_shared(obj);
247
248         i915_gem_object_release_mmap_offset(obj);
249
250         /*
251          * ->put_pages might need to allocate memory for the bit17 swizzle
252          * array, hence protect them from being reaped by removing them from gtt
253          * lists early.
254          */
255         pages = __i915_gem_object_unset_pages(obj);
256
257         /*
258          * XXX Temporary hijinx to avoid updating all backends to handle
259          * NULL pages. In the future, when we have more asynchronous
260          * get_pages backends we should be better able to handle the
261          * cancellation of the async task in a more uniform manner.
262          */
263         if (!IS_ERR_OR_NULL(pages))
264                 obj->ops->put_pages(obj, pages);
265
266         return 0;
267 }
268
269 /* The 'mapping' part of i915_gem_object_pin_map() below */
270 static void *i915_gem_object_map_page(struct drm_i915_gem_object *obj,
271                                       enum i915_map_type type)
272 {
273         unsigned long n_pages = obj->base.size >> PAGE_SHIFT, i;
274         struct page *stack[32], **pages = stack, *page;
275         struct sgt_iter iter;
276         pgprot_t pgprot;
277         void *vaddr;
278
279         switch (type) {
280         default:
281                 MISSING_CASE(type);
282                 fallthrough;    /* to use PAGE_KERNEL anyway */
283         case I915_MAP_WB:
284                 /*
285                  * On 32b, highmem using a finite set of indirect PTE (i.e.
286                  * vmap) to provide virtual mappings of the high pages.
287                  * As these are finite, map_new_virtual() must wait for some
288                  * other kmap() to finish when it runs out. If we map a large
289                  * number of objects, there is no method for it to tell us
290                  * to release the mappings, and we deadlock.
291                  *
292                  * However, if we make an explicit vmap of the page, that
293                  * uses a larger vmalloc arena, and also has the ability
294                  * to tell us to release unwanted mappings. Most importantly,
295                  * it will fail and propagate an error instead of waiting
296                  * forever.
297                  *
298                  * So if the page is beyond the 32b boundary, make an explicit
299                  * vmap.
300                  */
301                 if (n_pages == 1 && !PageHighMem(sg_page(obj->mm.pages->sgl)))
302                         return page_address(sg_page(obj->mm.pages->sgl));
303                 pgprot = PAGE_KERNEL;
304                 break;
305         case I915_MAP_WC:
306                 pgprot = pgprot_writecombine(PAGE_KERNEL_IO);
307                 break;
308         }
309
310         if (n_pages > ARRAY_SIZE(stack)) {
311                 /* Too big for stack -- allocate temporary array instead */
312                 pages = kvmalloc_array(n_pages, sizeof(*pages), GFP_KERNEL);
313                 if (!pages)
314                         return ERR_PTR(-ENOMEM);
315         }
316
317         i = 0;
318         for_each_sgt_page(page, iter, obj->mm.pages)
319                 pages[i++] = page;
320         vaddr = vmap(pages, n_pages, 0, pgprot);
321         if (pages != stack)
322                 kvfree(pages);
323
324         return vaddr ?: ERR_PTR(-ENOMEM);
325 }
326
327 static void *i915_gem_object_map_pfn(struct drm_i915_gem_object *obj,
328                                      enum i915_map_type type)
329 {
330         resource_size_t iomap = obj->mm.region->iomap.base -
331                 obj->mm.region->region.start;
332         unsigned long n_pfn = obj->base.size >> PAGE_SHIFT;
333         unsigned long stack[32], *pfns = stack, i;
334         struct sgt_iter iter;
335         dma_addr_t addr;
336         void *vaddr;
337
338         GEM_BUG_ON(type != I915_MAP_WC);
339
340         if (n_pfn > ARRAY_SIZE(stack)) {
341                 /* Too big for stack -- allocate temporary array instead */
342                 pfns = kvmalloc_array(n_pfn, sizeof(*pfns), GFP_KERNEL);
343                 if (!pfns)
344                         return ERR_PTR(-ENOMEM);
345         }
346
347         i = 0;
348         for_each_sgt_daddr(addr, iter, obj->mm.pages)
349                 pfns[i++] = (iomap + addr) >> PAGE_SHIFT;
350         vaddr = vmap_pfn(pfns, n_pfn, pgprot_writecombine(PAGE_KERNEL_IO));
351         if (pfns != stack)
352                 kvfree(pfns);
353
354         return vaddr ?: ERR_PTR(-ENOMEM);
355 }
356
357 /* get, pin, and map the pages of the object into kernel space */
358 void *i915_gem_object_pin_map(struct drm_i915_gem_object *obj,
359                               enum i915_map_type type)
360 {
361         enum i915_map_type has_type;
362         bool pinned;
363         void *ptr;
364         int err;
365
366         if (!i915_gem_object_has_struct_page(obj) &&
367             !i915_gem_object_has_iomem(obj))
368                 return ERR_PTR(-ENXIO);
369
370         if (WARN_ON_ONCE(obj->flags & I915_BO_ALLOC_GPU_ONLY))
371                 return ERR_PTR(-EINVAL);
372
373         assert_object_held(obj);
374
375         pinned = !(type & I915_MAP_OVERRIDE);
376         type &= ~I915_MAP_OVERRIDE;
377
378         if (!atomic_inc_not_zero(&obj->mm.pages_pin_count)) {
379                 if (unlikely(!i915_gem_object_has_pages(obj))) {
380                         GEM_BUG_ON(i915_gem_object_has_pinned_pages(obj));
381
382                         err = ____i915_gem_object_get_pages(obj);
383                         if (err)
384                                 return ERR_PTR(err);
385
386                         smp_mb__before_atomic();
387                 }
388                 atomic_inc(&obj->mm.pages_pin_count);
389                 pinned = false;
390         }
391         GEM_BUG_ON(!i915_gem_object_has_pages(obj));
392
393         /*
394          * For discrete our CPU mappings needs to be consistent in order to
395          * function correctly on !x86. When mapping things through TTM, we use
396          * the same rules to determine the caching type.
397          *
398          * The caching rules, starting from DG1:
399          *
400          *      - If the object can be placed in device local-memory, then the
401          *        pages should be allocated and mapped as write-combined only.
402          *
403          *      - Everything else is always allocated and mapped as write-back,
404          *        with the guarantee that everything is also coherent with the
405          *        GPU.
406          *
407          * Internal users of lmem are already expected to get this right, so no
408          * fudging needed there.
409          */
410         if (i915_gem_object_placement_possible(obj, INTEL_MEMORY_LOCAL)) {
411                 if (type != I915_MAP_WC && !obj->mm.n_placements) {
412                         ptr = ERR_PTR(-ENODEV);
413                         goto err_unpin;
414                 }
415
416                 type = I915_MAP_WC;
417         } else if (IS_DGFX(to_i915(obj->base.dev))) {
418                 type = I915_MAP_WB;
419         }
420
421         ptr = page_unpack_bits(obj->mm.mapping, &has_type);
422         if (ptr && has_type != type) {
423                 if (pinned) {
424                         ptr = ERR_PTR(-EBUSY);
425                         goto err_unpin;
426                 }
427
428                 unmap_object(obj, ptr);
429
430                 ptr = obj->mm.mapping = NULL;
431         }
432
433         if (!ptr) {
434                 err = i915_gem_object_wait_moving_fence(obj, true);
435                 if (err) {
436                         ptr = ERR_PTR(err);
437                         goto err_unpin;
438                 }
439
440                 if (GEM_WARN_ON(type == I915_MAP_WC && !pat_enabled()))
441                         ptr = ERR_PTR(-ENODEV);
442                 else if (i915_gem_object_has_struct_page(obj))
443                         ptr = i915_gem_object_map_page(obj, type);
444                 else
445                         ptr = i915_gem_object_map_pfn(obj, type);
446                 if (IS_ERR(ptr))
447                         goto err_unpin;
448
449                 obj->mm.mapping = page_pack_bits(ptr, type);
450         }
451
452         return ptr;
453
454 err_unpin:
455         atomic_dec(&obj->mm.pages_pin_count);
456         return ptr;
457 }
458
459 void *i915_gem_object_pin_map_unlocked(struct drm_i915_gem_object *obj,
460                                        enum i915_map_type type)
461 {
462         void *ret;
463
464         i915_gem_object_lock(obj, NULL);
465         ret = i915_gem_object_pin_map(obj, type);
466         i915_gem_object_unlock(obj);
467
468         return ret;
469 }
470
471 void __i915_gem_object_flush_map(struct drm_i915_gem_object *obj,
472                                  unsigned long offset,
473                                  unsigned long size)
474 {
475         enum i915_map_type has_type;
476         void *ptr;
477
478         GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
479         GEM_BUG_ON(range_overflows_t(typeof(obj->base.size),
480                                      offset, size, obj->base.size));
481
482         wmb(); /* let all previous writes be visible to coherent partners */
483         obj->mm.dirty = true;
484
485         if (obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_WRITE)
486                 return;
487
488         ptr = page_unpack_bits(obj->mm.mapping, &has_type);
489         if (has_type == I915_MAP_WC)
490                 return;
491
492         drm_clflush_virt_range(ptr + offset, size);
493         if (size == obj->base.size) {
494                 obj->write_domain &= ~I915_GEM_DOMAIN_CPU;
495                 obj->cache_dirty = false;
496         }
497 }
498
499 void __i915_gem_object_release_map(struct drm_i915_gem_object *obj)
500 {
501         GEM_BUG_ON(!obj->mm.mapping);
502
503         /*
504          * We allow removing the mapping from underneath pinned pages!
505          *
506          * Furthermore, since this is an unsafe operation reserved only
507          * for construction time manipulation, we ignore locking prudence.
508          */
509         unmap_object(obj, page_mask_bits(fetch_and_zero(&obj->mm.mapping)));
510
511         i915_gem_object_unpin_map(obj);
512 }
513
514 struct scatterlist *
515 __i915_gem_object_page_iter_get_sg(struct drm_i915_gem_object *obj,
516                                    struct i915_gem_object_page_iter *iter,
517                                    pgoff_t n,
518                                    unsigned int *offset)
519
520 {
521         const bool dma = iter == &obj->mm.get_dma_page ||
522                          iter == &obj->ttm.get_io_page;
523         unsigned int idx, count;
524         struct scatterlist *sg;
525
526         might_sleep();
527         GEM_BUG_ON(n >= obj->base.size >> PAGE_SHIFT);
528         if (!i915_gem_object_has_pinned_pages(obj))
529                 assert_object_held(obj);
530
531         /* As we iterate forward through the sg, we record each entry in a
532          * radixtree for quick repeated (backwards) lookups. If we have seen
533          * this index previously, we will have an entry for it.
534          *
535          * Initial lookup is O(N), but this is amortized to O(1) for
536          * sequential page access (where each new request is consecutive
537          * to the previous one). Repeated lookups are O(lg(obj->base.size)),
538          * i.e. O(1) with a large constant!
539          */
540         if (n < READ_ONCE(iter->sg_idx))
541                 goto lookup;
542
543         mutex_lock(&iter->lock);
544
545         /* We prefer to reuse the last sg so that repeated lookup of this
546          * (or the subsequent) sg are fast - comparing against the last
547          * sg is faster than going through the radixtree.
548          */
549
550         sg = iter->sg_pos;
551         idx = iter->sg_idx;
552         count = dma ? __sg_dma_page_count(sg) : __sg_page_count(sg);
553
554         while (idx + count <= n) {
555                 void *entry;
556                 unsigned long i;
557                 int ret;
558
559                 /* If we cannot allocate and insert this entry, or the
560                  * individual pages from this range, cancel updating the
561                  * sg_idx so that on this lookup we are forced to linearly
562                  * scan onwards, but on future lookups we will try the
563                  * insertion again (in which case we need to be careful of
564                  * the error return reporting that we have already inserted
565                  * this index).
566                  */
567                 ret = radix_tree_insert(&iter->radix, idx, sg);
568                 if (ret && ret != -EEXIST)
569                         goto scan;
570
571                 entry = xa_mk_value(idx);
572                 for (i = 1; i < count; i++) {
573                         ret = radix_tree_insert(&iter->radix, idx + i, entry);
574                         if (ret && ret != -EEXIST)
575                                 goto scan;
576                 }
577
578                 idx += count;
579                 sg = ____sg_next(sg);
580                 count = dma ? __sg_dma_page_count(sg) : __sg_page_count(sg);
581         }
582
583 scan:
584         iter->sg_pos = sg;
585         iter->sg_idx = idx;
586
587         mutex_unlock(&iter->lock);
588
589         if (unlikely(n < idx)) /* insertion completed by another thread */
590                 goto lookup;
591
592         /* In case we failed to insert the entry into the radixtree, we need
593          * to look beyond the current sg.
594          */
595         while (idx + count <= n) {
596                 idx += count;
597                 sg = ____sg_next(sg);
598                 count = dma ? __sg_dma_page_count(sg) : __sg_page_count(sg);
599         }
600
601         *offset = n - idx;
602         return sg;
603
604 lookup:
605         rcu_read_lock();
606
607         sg = radix_tree_lookup(&iter->radix, n);
608         GEM_BUG_ON(!sg);
609
610         /* If this index is in the middle of multi-page sg entry,
611          * the radix tree will contain a value entry that points
612          * to the start of that range. We will return the pointer to
613          * the base page and the offset of this page within the
614          * sg entry's range.
615          */
616         *offset = 0;
617         if (unlikely(xa_is_value(sg))) {
618                 unsigned long base = xa_to_value(sg);
619
620                 sg = radix_tree_lookup(&iter->radix, base);
621                 GEM_BUG_ON(!sg);
622
623                 *offset = n - base;
624         }
625
626         rcu_read_unlock();
627
628         return sg;
629 }
630
631 struct page *
632 __i915_gem_object_get_page(struct drm_i915_gem_object *obj, pgoff_t n)
633 {
634         struct scatterlist *sg;
635         unsigned int offset;
636
637         GEM_BUG_ON(!i915_gem_object_has_struct_page(obj));
638
639         sg = i915_gem_object_get_sg(obj, n, &offset);
640         return nth_page(sg_page(sg), offset);
641 }
642
643 /* Like i915_gem_object_get_page(), but mark the returned page dirty */
644 struct page *
645 __i915_gem_object_get_dirty_page(struct drm_i915_gem_object *obj, pgoff_t n)
646 {
647         struct page *page;
648
649         page = i915_gem_object_get_page(obj, n);
650         if (!obj->mm.dirty)
651                 set_page_dirty(page);
652
653         return page;
654 }
655
656 dma_addr_t
657 __i915_gem_object_get_dma_address_len(struct drm_i915_gem_object *obj,
658                                       pgoff_t n, unsigned int *len)
659 {
660         struct scatterlist *sg;
661         unsigned int offset;
662
663         sg = i915_gem_object_get_sg_dma(obj, n, &offset);
664
665         if (len)
666                 *len = sg_dma_len(sg) - (offset << PAGE_SHIFT);
667
668         return sg_dma_address(sg) + (offset << PAGE_SHIFT);
669 }
670
671 dma_addr_t
672 __i915_gem_object_get_dma_address(struct drm_i915_gem_object *obj, pgoff_t n)
673 {
674         return i915_gem_object_get_dma_address_len(obj, n, NULL);
675 }