Merge drm/drm-next into drm-intel-gt-next
[sfrench/cifs-2.6.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_gt_pm.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                                  unsigned int sg_page_sizes)
21 {
22         struct drm_i915_private *i915 = to_i915(obj->base.dev);
23         unsigned long supported = RUNTIME_INFO(i915)->page_sizes;
24         bool shrinkable;
25         int i;
26
27         assert_object_held_shared(obj);
28
29         if (i915_gem_object_is_volatile(obj))
30                 obj->mm.madv = I915_MADV_DONTNEED;
31
32         /* Make the pages coherent with the GPU (flushing any swapin). */
33         if (obj->cache_dirty) {
34                 WARN_ON_ONCE(IS_DGFX(i915));
35                 obj->write_domain = 0;
36                 if (i915_gem_object_has_struct_page(obj))
37                         drm_clflush_sg(pages);
38                 obj->cache_dirty = false;
39         }
40
41         obj->mm.get_page.sg_pos = pages->sgl;
42         obj->mm.get_page.sg_idx = 0;
43         obj->mm.get_dma_page.sg_pos = pages->sgl;
44         obj->mm.get_dma_page.sg_idx = 0;
45
46         obj->mm.pages = pages;
47
48         GEM_BUG_ON(!sg_page_sizes);
49         obj->mm.page_sizes.phys = sg_page_sizes;
50
51         /*
52          * Calculate the supported page-sizes which fit into the given
53          * sg_page_sizes. This will give us the page-sizes which we may be able
54          * to use opportunistically when later inserting into the GTT. For
55          * example if phys=2G, then in theory we should be able to use 1G, 2M,
56          * 64K or 4K pages, although in practice this will depend on a number of
57          * other factors.
58          */
59         obj->mm.page_sizes.sg = 0;
60         for_each_set_bit(i, &supported, ilog2(I915_GTT_MAX_PAGE_SIZE) + 1) {
61                 if (obj->mm.page_sizes.phys & ~0u << i)
62                         obj->mm.page_sizes.sg |= BIT(i);
63         }
64         GEM_BUG_ON(!HAS_PAGE_SIZES(i915, obj->mm.page_sizes.sg));
65
66         shrinkable = i915_gem_object_is_shrinkable(obj);
67
68         if (i915_gem_object_is_tiled(obj) &&
69             i915->gem_quirks & GEM_QUIRK_PIN_SWIZZLED_PAGES) {
70                 GEM_BUG_ON(i915_gem_object_has_tiling_quirk(obj));
71                 i915_gem_object_set_tiling_quirk(obj);
72                 GEM_BUG_ON(!list_empty(&obj->mm.link));
73                 atomic_inc(&obj->mm.shrink_pin);
74                 shrinkable = false;
75         }
76
77         if (shrinkable && !i915_gem_object_has_self_managed_shrink_list(obj)) {
78                 struct list_head *list;
79                 unsigned long flags;
80
81                 assert_object_held(obj);
82                 spin_lock_irqsave(&i915->mm.obj_lock, flags);
83
84                 i915->mm.shrink_count++;
85                 i915->mm.shrink_memory += obj->base.size;
86
87                 if (obj->mm.madv != I915_MADV_WILLNEED)
88                         list = &i915->mm.purge_list;
89                 else
90                         list = &i915->mm.shrink_list;
91                 list_add_tail(&obj->mm.link, list);
92
93                 atomic_set(&obj->mm.shrink_pin, 0);
94                 spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
95         }
96 }
97
98 int ____i915_gem_object_get_pages(struct drm_i915_gem_object *obj)
99 {
100         struct drm_i915_private *i915 = to_i915(obj->base.dev);
101         int err;
102
103         assert_object_held_shared(obj);
104
105         if (unlikely(obj->mm.madv != I915_MADV_WILLNEED)) {
106                 drm_dbg(&i915->drm,
107                         "Attempting to obtain a purgeable object\n");
108                 return -EFAULT;
109         }
110
111         err = obj->ops->get_pages(obj);
112         GEM_BUG_ON(!err && !i915_gem_object_has_pages(obj));
113
114         return err;
115 }
116
117 /* Ensure that the associated pages are gathered from the backing storage
118  * and pinned into our object. i915_gem_object_pin_pages() may be called
119  * multiple times before they are released by a single call to
120  * i915_gem_object_unpin_pages() - once the pages are no longer referenced
121  * either as a result of memory pressure (reaping pages under the shrinker)
122  * or as the object is itself released.
123  */
124 int __i915_gem_object_get_pages(struct drm_i915_gem_object *obj)
125 {
126         int err;
127
128         assert_object_held(obj);
129
130         assert_object_held_shared(obj);
131
132         if (unlikely(!i915_gem_object_has_pages(obj))) {
133                 GEM_BUG_ON(i915_gem_object_has_pinned_pages(obj));
134
135                 err = ____i915_gem_object_get_pages(obj);
136                 if (err)
137                         return err;
138
139                 smp_mb__before_atomic();
140         }
141         atomic_inc(&obj->mm.pages_pin_count);
142
143         return 0;
144 }
145
146 int i915_gem_object_pin_pages_unlocked(struct drm_i915_gem_object *obj)
147 {
148         struct i915_gem_ww_ctx ww;
149         int err;
150
151         i915_gem_ww_ctx_init(&ww, true);
152 retry:
153         err = i915_gem_object_lock(obj, &ww);
154         if (!err)
155                 err = i915_gem_object_pin_pages(obj);
156
157         if (err == -EDEADLK) {
158                 err = i915_gem_ww_ctx_backoff(&ww);
159                 if (!err)
160                         goto retry;
161         }
162         i915_gem_ww_ctx_fini(&ww);
163         return err;
164 }
165
166 /* Immediately discard the backing storage */
167 int i915_gem_object_truncate(struct drm_i915_gem_object *obj)
168 {
169         if (obj->ops->truncate)
170                 return obj->ops->truncate(obj);
171
172         return 0;
173 }
174
175 static void __i915_gem_object_reset_page_iter(struct drm_i915_gem_object *obj)
176 {
177         struct radix_tree_iter iter;
178         void __rcu **slot;
179
180         rcu_read_lock();
181         radix_tree_for_each_slot(slot, &obj->mm.get_page.radix, &iter, 0)
182                 radix_tree_delete(&obj->mm.get_page.radix, iter.index);
183         radix_tree_for_each_slot(slot, &obj->mm.get_dma_page.radix, &iter, 0)
184                 radix_tree_delete(&obj->mm.get_dma_page.radix, iter.index);
185         rcu_read_unlock();
186 }
187
188 static void unmap_object(struct drm_i915_gem_object *obj, void *ptr)
189 {
190         if (is_vmalloc_addr(ptr))
191                 vunmap(ptr);
192 }
193
194 static void flush_tlb_invalidate(struct drm_i915_gem_object *obj)
195 {
196         struct drm_i915_private *i915 = to_i915(obj->base.dev);
197         struct intel_gt *gt = to_gt(i915);
198
199         if (!obj->mm.tlb)
200                 return;
201
202         intel_gt_invalidate_tlb(gt, obj->mm.tlb);
203         obj->mm.tlb = 0;
204 }
205
206 struct sg_table *
207 __i915_gem_object_unset_pages(struct drm_i915_gem_object *obj)
208 {
209         struct sg_table *pages;
210
211         assert_object_held_shared(obj);
212
213         pages = fetch_and_zero(&obj->mm.pages);
214         if (IS_ERR_OR_NULL(pages))
215                 return pages;
216
217         if (i915_gem_object_is_volatile(obj))
218                 obj->mm.madv = I915_MADV_WILLNEED;
219
220         if (!i915_gem_object_has_self_managed_shrink_list(obj))
221                 i915_gem_object_make_unshrinkable(obj);
222
223         if (obj->mm.mapping) {
224                 unmap_object(obj, page_mask_bits(obj->mm.mapping));
225                 obj->mm.mapping = NULL;
226         }
227
228         __i915_gem_object_reset_page_iter(obj);
229         obj->mm.page_sizes.phys = obj->mm.page_sizes.sg = 0;
230
231         flush_tlb_invalidate(obj);
232
233         return pages;
234 }
235
236 int __i915_gem_object_put_pages(struct drm_i915_gem_object *obj)
237 {
238         struct sg_table *pages;
239
240         if (i915_gem_object_has_pinned_pages(obj))
241                 return -EBUSY;
242
243         /* May be called by shrinker from within get_pages() (on another bo) */
244         assert_object_held_shared(obj);
245
246         i915_gem_object_release_mmap_offset(obj);
247
248         /*
249          * ->put_pages might need to allocate memory for the bit17 swizzle
250          * array, hence protect them from being reaped by removing them from gtt
251          * lists early.
252          */
253         pages = __i915_gem_object_unset_pages(obj);
254
255         /*
256          * XXX Temporary hijinx to avoid updating all backends to handle
257          * NULL pages. In the future, when we have more asynchronous
258          * get_pages backends we should be better able to handle the
259          * cancellation of the async task in a more uniform manner.
260          */
261         if (!IS_ERR_OR_NULL(pages))
262                 obj->ops->put_pages(obj, pages);
263
264         return 0;
265 }
266
267 /* The 'mapping' part of i915_gem_object_pin_map() below */
268 static void *i915_gem_object_map_page(struct drm_i915_gem_object *obj,
269                                       enum i915_map_type type)
270 {
271         unsigned long n_pages = obj->base.size >> PAGE_SHIFT, i;
272         struct page *stack[32], **pages = stack, *page;
273         struct sgt_iter iter;
274         pgprot_t pgprot;
275         void *vaddr;
276
277         switch (type) {
278         default:
279                 MISSING_CASE(type);
280                 fallthrough;    /* to use PAGE_KERNEL anyway */
281         case I915_MAP_WB:
282                 /*
283                  * On 32b, highmem using a finite set of indirect PTE (i.e.
284                  * vmap) to provide virtual mappings of the high pages.
285                  * As these are finite, map_new_virtual() must wait for some
286                  * other kmap() to finish when it runs out. If we map a large
287                  * number of objects, there is no method for it to tell us
288                  * to release the mappings, and we deadlock.
289                  *
290                  * However, if we make an explicit vmap of the page, that
291                  * uses a larger vmalloc arena, and also has the ability
292                  * to tell us to release unwanted mappings. Most importantly,
293                  * it will fail and propagate an error instead of waiting
294                  * forever.
295                  *
296                  * So if the page is beyond the 32b boundary, make an explicit
297                  * vmap.
298                  */
299                 if (n_pages == 1 && !PageHighMem(sg_page(obj->mm.pages->sgl)))
300                         return page_address(sg_page(obj->mm.pages->sgl));
301                 pgprot = PAGE_KERNEL;
302                 break;
303         case I915_MAP_WC:
304                 pgprot = pgprot_writecombine(PAGE_KERNEL_IO);
305                 break;
306         }
307
308         if (n_pages > ARRAY_SIZE(stack)) {
309                 /* Too big for stack -- allocate temporary array instead */
310                 pages = kvmalloc_array(n_pages, sizeof(*pages), GFP_KERNEL);
311                 if (!pages)
312                         return ERR_PTR(-ENOMEM);
313         }
314
315         i = 0;
316         for_each_sgt_page(page, iter, obj->mm.pages)
317                 pages[i++] = page;
318         vaddr = vmap(pages, n_pages, 0, pgprot);
319         if (pages != stack)
320                 kvfree(pages);
321
322         return vaddr ?: ERR_PTR(-ENOMEM);
323 }
324
325 static void *i915_gem_object_map_pfn(struct drm_i915_gem_object *obj,
326                                      enum i915_map_type type)
327 {
328         resource_size_t iomap = obj->mm.region->iomap.base -
329                 obj->mm.region->region.start;
330         unsigned long n_pfn = obj->base.size >> PAGE_SHIFT;
331         unsigned long stack[32], *pfns = stack, i;
332         struct sgt_iter iter;
333         dma_addr_t addr;
334         void *vaddr;
335
336         GEM_BUG_ON(type != I915_MAP_WC);
337
338         if (n_pfn > ARRAY_SIZE(stack)) {
339                 /* Too big for stack -- allocate temporary array instead */
340                 pfns = kvmalloc_array(n_pfn, sizeof(*pfns), GFP_KERNEL);
341                 if (!pfns)
342                         return ERR_PTR(-ENOMEM);
343         }
344
345         i = 0;
346         for_each_sgt_daddr(addr, iter, obj->mm.pages)
347                 pfns[i++] = (iomap + addr) >> PAGE_SHIFT;
348         vaddr = vmap_pfn(pfns, n_pfn, pgprot_writecombine(PAGE_KERNEL_IO));
349         if (pfns != stack)
350                 kvfree(pfns);
351
352         return vaddr ?: ERR_PTR(-ENOMEM);
353 }
354
355 /* get, pin, and map the pages of the object into kernel space */
356 void *i915_gem_object_pin_map(struct drm_i915_gem_object *obj,
357                               enum i915_map_type type)
358 {
359         enum i915_map_type has_type;
360         bool pinned;
361         void *ptr;
362         int err;
363
364         if (!i915_gem_object_has_struct_page(obj) &&
365             !i915_gem_object_has_iomem(obj))
366                 return ERR_PTR(-ENXIO);
367
368         if (WARN_ON_ONCE(obj->flags & I915_BO_ALLOC_GPU_ONLY))
369                 return ERR_PTR(-EINVAL);
370
371         assert_object_held(obj);
372
373         pinned = !(type & I915_MAP_OVERRIDE);
374         type &= ~I915_MAP_OVERRIDE;
375
376         if (!atomic_inc_not_zero(&obj->mm.pages_pin_count)) {
377                 if (unlikely(!i915_gem_object_has_pages(obj))) {
378                         GEM_BUG_ON(i915_gem_object_has_pinned_pages(obj));
379
380                         err = ____i915_gem_object_get_pages(obj);
381                         if (err)
382                                 return ERR_PTR(err);
383
384                         smp_mb__before_atomic();
385                 }
386                 atomic_inc(&obj->mm.pages_pin_count);
387                 pinned = false;
388         }
389         GEM_BUG_ON(!i915_gem_object_has_pages(obj));
390
391         /*
392          * For discrete our CPU mappings needs to be consistent in order to
393          * function correctly on !x86. When mapping things through TTM, we use
394          * the same rules to determine the caching type.
395          *
396          * The caching rules, starting from DG1:
397          *
398          *      - If the object can be placed in device local-memory, then the
399          *        pages should be allocated and mapped as write-combined only.
400          *
401          *      - Everything else is always allocated and mapped as write-back,
402          *        with the guarantee that everything is also coherent with the
403          *        GPU.
404          *
405          * Internal users of lmem are already expected to get this right, so no
406          * fudging needed there.
407          */
408         if (i915_gem_object_placement_possible(obj, INTEL_MEMORY_LOCAL)) {
409                 if (type != I915_MAP_WC && !obj->mm.n_placements) {
410                         ptr = ERR_PTR(-ENODEV);
411                         goto err_unpin;
412                 }
413
414                 type = I915_MAP_WC;
415         } else if (IS_DGFX(to_i915(obj->base.dev))) {
416                 type = I915_MAP_WB;
417         }
418
419         ptr = page_unpack_bits(obj->mm.mapping, &has_type);
420         if (ptr && has_type != type) {
421                 if (pinned) {
422                         ptr = ERR_PTR(-EBUSY);
423                         goto err_unpin;
424                 }
425
426                 unmap_object(obj, ptr);
427
428                 ptr = obj->mm.mapping = NULL;
429         }
430
431         if (!ptr) {
432                 err = i915_gem_object_wait_moving_fence(obj, true);
433                 if (err) {
434                         ptr = ERR_PTR(err);
435                         goto err_unpin;
436                 }
437
438                 if (GEM_WARN_ON(type == I915_MAP_WC && !pat_enabled()))
439                         ptr = ERR_PTR(-ENODEV);
440                 else if (i915_gem_object_has_struct_page(obj))
441                         ptr = i915_gem_object_map_page(obj, type);
442                 else
443                         ptr = i915_gem_object_map_pfn(obj, type);
444                 if (IS_ERR(ptr))
445                         goto err_unpin;
446
447                 obj->mm.mapping = page_pack_bits(ptr, type);
448         }
449
450         return ptr;
451
452 err_unpin:
453         atomic_dec(&obj->mm.pages_pin_count);
454         return ptr;
455 }
456
457 void *i915_gem_object_pin_map_unlocked(struct drm_i915_gem_object *obj,
458                                        enum i915_map_type type)
459 {
460         void *ret;
461
462         i915_gem_object_lock(obj, NULL);
463         ret = i915_gem_object_pin_map(obj, type);
464         i915_gem_object_unlock(obj);
465
466         return ret;
467 }
468
469 enum i915_map_type i915_coherent_map_type(struct drm_i915_private *i915,
470                                           struct drm_i915_gem_object *obj,
471                                           bool always_coherent)
472 {
473         if (i915_gem_object_is_lmem(obj))
474                 return I915_MAP_WC;
475         if (HAS_LLC(i915) || always_coherent)
476                 return I915_MAP_WB;
477         else
478                 return I915_MAP_WC;
479 }
480
481 void __i915_gem_object_flush_map(struct drm_i915_gem_object *obj,
482                                  unsigned long offset,
483                                  unsigned long size)
484 {
485         enum i915_map_type has_type;
486         void *ptr;
487
488         GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
489         GEM_BUG_ON(range_overflows_t(typeof(obj->base.size),
490                                      offset, size, obj->base.size));
491
492         wmb(); /* let all previous writes be visible to coherent partners */
493         obj->mm.dirty = true;
494
495         if (obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_WRITE)
496                 return;
497
498         ptr = page_unpack_bits(obj->mm.mapping, &has_type);
499         if (has_type == I915_MAP_WC)
500                 return;
501
502         drm_clflush_virt_range(ptr + offset, size);
503         if (size == obj->base.size) {
504                 obj->write_domain &= ~I915_GEM_DOMAIN_CPU;
505                 obj->cache_dirty = false;
506         }
507 }
508
509 void __i915_gem_object_release_map(struct drm_i915_gem_object *obj)
510 {
511         GEM_BUG_ON(!obj->mm.mapping);
512
513         /*
514          * We allow removing the mapping from underneath pinned pages!
515          *
516          * Furthermore, since this is an unsafe operation reserved only
517          * for construction time manipulation, we ignore locking prudence.
518          */
519         unmap_object(obj, page_mask_bits(fetch_and_zero(&obj->mm.mapping)));
520
521         i915_gem_object_unpin_map(obj);
522 }
523
524 struct scatterlist *
525 __i915_gem_object_get_sg(struct drm_i915_gem_object *obj,
526                          struct i915_gem_object_page_iter *iter,
527                          unsigned int n,
528                          unsigned int *offset,
529                          bool dma)
530 {
531         struct scatterlist *sg;
532         unsigned int idx, count;
533
534         might_sleep();
535         GEM_BUG_ON(n >= obj->base.size >> PAGE_SHIFT);
536         if (!i915_gem_object_has_pinned_pages(obj))
537                 assert_object_held(obj);
538
539         /* As we iterate forward through the sg, we record each entry in a
540          * radixtree for quick repeated (backwards) lookups. If we have seen
541          * this index previously, we will have an entry for it.
542          *
543          * Initial lookup is O(N), but this is amortized to O(1) for
544          * sequential page access (where each new request is consecutive
545          * to the previous one). Repeated lookups are O(lg(obj->base.size)),
546          * i.e. O(1) with a large constant!
547          */
548         if (n < READ_ONCE(iter->sg_idx))
549                 goto lookup;
550
551         mutex_lock(&iter->lock);
552
553         /* We prefer to reuse the last sg so that repeated lookup of this
554          * (or the subsequent) sg are fast - comparing against the last
555          * sg is faster than going through the radixtree.
556          */
557
558         sg = iter->sg_pos;
559         idx = iter->sg_idx;
560         count = dma ? __sg_dma_page_count(sg) : __sg_page_count(sg);
561
562         while (idx + count <= n) {
563                 void *entry;
564                 unsigned long i;
565                 int ret;
566
567                 /* If we cannot allocate and insert this entry, or the
568                  * individual pages from this range, cancel updating the
569                  * sg_idx so that on this lookup we are forced to linearly
570                  * scan onwards, but on future lookups we will try the
571                  * insertion again (in which case we need to be careful of
572                  * the error return reporting that we have already inserted
573                  * this index).
574                  */
575                 ret = radix_tree_insert(&iter->radix, idx, sg);
576                 if (ret && ret != -EEXIST)
577                         goto scan;
578
579                 entry = xa_mk_value(idx);
580                 for (i = 1; i < count; i++) {
581                         ret = radix_tree_insert(&iter->radix, idx + i, entry);
582                         if (ret && ret != -EEXIST)
583                                 goto scan;
584                 }
585
586                 idx += count;
587                 sg = ____sg_next(sg);
588                 count = dma ? __sg_dma_page_count(sg) : __sg_page_count(sg);
589         }
590
591 scan:
592         iter->sg_pos = sg;
593         iter->sg_idx = idx;
594
595         mutex_unlock(&iter->lock);
596
597         if (unlikely(n < idx)) /* insertion completed by another thread */
598                 goto lookup;
599
600         /* In case we failed to insert the entry into the radixtree, we need
601          * to look beyond the current sg.
602          */
603         while (idx + count <= n) {
604                 idx += count;
605                 sg = ____sg_next(sg);
606                 count = dma ? __sg_dma_page_count(sg) : __sg_page_count(sg);
607         }
608
609         *offset = n - idx;
610         return sg;
611
612 lookup:
613         rcu_read_lock();
614
615         sg = radix_tree_lookup(&iter->radix, n);
616         GEM_BUG_ON(!sg);
617
618         /* If this index is in the middle of multi-page sg entry,
619          * the radix tree will contain a value entry that points
620          * to the start of that range. We will return the pointer to
621          * the base page and the offset of this page within the
622          * sg entry's range.
623          */
624         *offset = 0;
625         if (unlikely(xa_is_value(sg))) {
626                 unsigned long base = xa_to_value(sg);
627
628                 sg = radix_tree_lookup(&iter->radix, base);
629                 GEM_BUG_ON(!sg);
630
631                 *offset = n - base;
632         }
633
634         rcu_read_unlock();
635
636         return sg;
637 }
638
639 struct page *
640 i915_gem_object_get_page(struct drm_i915_gem_object *obj, unsigned int n)
641 {
642         struct scatterlist *sg;
643         unsigned int offset;
644
645         GEM_BUG_ON(!i915_gem_object_has_struct_page(obj));
646
647         sg = i915_gem_object_get_sg(obj, n, &offset);
648         return nth_page(sg_page(sg), offset);
649 }
650
651 /* Like i915_gem_object_get_page(), but mark the returned page dirty */
652 struct page *
653 i915_gem_object_get_dirty_page(struct drm_i915_gem_object *obj,
654                                unsigned int n)
655 {
656         struct page *page;
657
658         page = i915_gem_object_get_page(obj, n);
659         if (!obj->mm.dirty)
660                 set_page_dirty(page);
661
662         return page;
663 }
664
665 dma_addr_t
666 i915_gem_object_get_dma_address_len(struct drm_i915_gem_object *obj,
667                                     unsigned long n,
668                                     unsigned int *len)
669 {
670         struct scatterlist *sg;
671         unsigned int offset;
672
673         sg = i915_gem_object_get_sg_dma(obj, n, &offset);
674
675         if (len)
676                 *len = sg_dma_len(sg) - (offset << PAGE_SHIFT);
677
678         return sg_dma_address(sg) + (offset << PAGE_SHIFT);
679 }
680
681 dma_addr_t
682 i915_gem_object_get_dma_address(struct drm_i915_gem_object *obj,
683                                 unsigned long n)
684 {
685         return i915_gem_object_get_dma_address_len(obj, n, NULL);
686 }