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 "i915_drv.h"
   8 #include "i915_gem_object.h"
   9 #include "i915_scatterlist.h"
  10 #include "i915_gem_lmem.h"
  11 #include "i915_gem_mman.h"
  12
  13 void __i915_gem_object_set_pages(struct drm_i915_gem_object *obj,
  14                                  struct sg_table *pages,
  15                                  unsigned int sg_page_sizes)
  16 {
  17         struct drm_i915_private *i915 = to_i915(obj->base.dev);
  18         unsigned long supported = INTEL_INFO(i915)->page_sizes;
  19         int i;
  20
  21         lockdep_assert_held(&obj->mm.lock);
  22
  23         if (i915_gem_object_is_volatile(obj))
  24                 obj->mm.madv = I915_MADV_DONTNEED;
  25
  26         /* Make the pages coherent with the GPU (flushing any swapin). */
  27         if (obj->cache_dirty) {
  28                 obj->write_domain = 0;
  29                 if (i915_gem_object_has_struct_page(obj))
  30                         drm_clflush_sg(pages);
  31                 obj->cache_dirty = false;
  32         }
  33
  34         obj->mm.get_page.sg_pos = pages->sgl;
  35         obj->mm.get_page.sg_idx = 0;
  36
  37         obj->mm.pages = pages;
  38
  39         if (i915_gem_object_is_tiled(obj) &&
  40             i915->quirks & QUIRK_PIN_SWIZZLED_PAGES) {
  41                 GEM_BUG_ON(obj->mm.quirked);
  42                 __i915_gem_object_pin_pages(obj);
  43                 obj->mm.quirked = true;
  44         }
  45
  46         GEM_BUG_ON(!sg_page_sizes);
  47         obj->mm.page_sizes.phys = sg_page_sizes;
  48
  49         /*
  50          * Calculate the supported page-sizes which fit into the given
  51          * sg_page_sizes. This will give us the page-sizes which we may be able
  52          * to use opportunistically when later inserting into the GTT. For
  53          * example if phys=2G, then in theory we should be able to use 1G, 2M,
  54          * 64K or 4K pages, although in practice this will depend on a number of
  55          * other factors.
  56          */
  57         obj->mm.page_sizes.sg = 0;
  58         for_each_set_bit(i, &supported, ilog2(I915_GTT_MAX_PAGE_SIZE) + 1) {
  59                 if (obj->mm.page_sizes.phys & ~0u << i)
  60                         obj->mm.page_sizes.sg |= BIT(i);
  61         }
  62         GEM_BUG_ON(!HAS_PAGE_SIZES(i915, obj->mm.page_sizes.sg));
  63
  64         if (i915_gem_object_is_shrinkable(obj)) {
  65                 struct list_head *list;
  66                 unsigned long flags;
  67
  68                 spin_lock_irqsave(&i915->mm.obj_lock, flags);
  69
  70                 i915->mm.shrink_count++;
  71                 i915->mm.shrink_memory += obj->base.size;
  72
  73                 if (obj->mm.madv != I915_MADV_WILLNEED)
  74                         list = &i915->mm.purge_list;
  75                 else
  76                         list = &i915->mm.shrink_list;
  77                 list_add_tail(&obj->mm.link, list);
  78
  79                 atomic_set(&obj->mm.shrink_pin, 0);
  80                 spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
  81         }
  82 }
  83
  84 int ____i915_gem_object_get_pages(struct drm_i915_gem_object *obj)
  85 {
  86         int err;
  87
  88         if (unlikely(obj->mm.madv != I915_MADV_WILLNEED)) {
  89                 DRM_DEBUG("Attempting to obtain a purgeable object\n");
  90                 return -EFAULT;
  91         }
  92
  93         err = obj->ops->get_pages(obj);
  94         GEM_BUG_ON(!err && !i915_gem_object_has_pages(obj));
  95
  96         return err;
  97 }
  98
  99 /* Ensure that the associated pages are gathered from the backing storage
 100  * and pinned into our object. i915_gem_object_pin_pages() may be called
 101  * multiple times before they are released by a single call to
 102  * i915_gem_object_unpin_pages() - once the pages are no longer referenced
 103  * either as a result of memory pressure (reaping pages under the shrinker)
 104  * or as the object is itself released.
 105  */
 106 int __i915_gem_object_get_pages(struct drm_i915_gem_object *obj)
 107 {
 108         int err;
 109
 110         err = mutex_lock_interruptible_nested(&obj->mm.lock, I915_MM_GET_PAGES);
 111         if (err)
 112                 return err;
 113
 114         if (unlikely(!i915_gem_object_has_pages(obj))) {
 115                 GEM_BUG_ON(i915_gem_object_has_pinned_pages(obj));
 116
 117                 err = ____i915_gem_object_get_pages(obj);
 118                 if (err)
 119                         goto unlock;
 120
 121                 smp_mb__before_atomic();
 122         }
 123         atomic_inc(&obj->mm.pages_pin_count);
 124
 125 unlock:
 126         mutex_unlock(&obj->mm.lock);
 127         return err;
 128 }
 129
 130 /* Immediately discard the backing storage */
 131 void i915_gem_object_truncate(struct drm_i915_gem_object *obj)
 132 {
 133         drm_gem_free_mmap_offset(&obj->base);
 134         if (obj->ops->truncate)
 135                 obj->ops->truncate(obj);
 136 }
 137
 138 /* Try to discard unwanted pages */
 139 void i915_gem_object_writeback(struct drm_i915_gem_object *obj)
 140 {
 141         lockdep_assert_held(&obj->mm.lock);
 142         GEM_BUG_ON(i915_gem_object_has_pages(obj));
 143
 144         if (obj->ops->writeback)
 145                 obj->ops->writeback(obj);
 146 }
 147
 148 static void __i915_gem_object_reset_page_iter(struct drm_i915_gem_object *obj)
 149 {
 150         struct radix_tree_iter iter;
 151         void __rcu **slot;
 152
 153         rcu_read_lock();
 154         radix_tree_for_each_slot(slot, &obj->mm.get_page.radix, &iter, 0)
 155                 radix_tree_delete(&obj->mm.get_page.radix, iter.index);
 156         rcu_read_unlock();
 157 }
 158
 159 static void unmap_object(struct drm_i915_gem_object *obj, void *ptr)
 160 {
 161         if (is_vmalloc_addr(ptr))
 162                 vunmap(ptr);
 163         else
 164                 kunmap(kmap_to_page(ptr));
 165 }
 166
 167 struct sg_table *
 168 __i915_gem_object_unset_pages(struct drm_i915_gem_object *obj)
 169 {
 170         struct sg_table *pages;
 171
 172         pages = fetch_and_zero(&obj->mm.pages);
 173         if (IS_ERR_OR_NULL(pages))
 174                 return pages;
 175
 176         if (i915_gem_object_is_volatile(obj))
 177                 obj->mm.madv = I915_MADV_WILLNEED;
 178
 179         i915_gem_object_make_unshrinkable(obj);
 180
 181         if (obj->mm.mapping) {
 182                 unmap_object(obj, page_mask_bits(obj->mm.mapping));
 183                 obj->mm.mapping = NULL;
 184         }
 185
 186         __i915_gem_object_reset_page_iter(obj);
 187         obj->mm.page_sizes.phys = obj->mm.page_sizes.sg = 0;
 188
 189         return pages;
 190 }
 191
 192 int __i915_gem_object_put_pages(struct drm_i915_gem_object *obj)
 193 {
 194         struct sg_table *pages;
 195         int err;
 196
 197         if (i915_gem_object_has_pinned_pages(obj))
 198                 return -EBUSY;
 199
 200         GEM_BUG_ON(atomic_read(&obj->bind_count));
 201
 202         /* May be called by shrinker from within get_pages() (on another bo) */
 203         mutex_lock(&obj->mm.lock);
 204         if (unlikely(atomic_read(&obj->mm.pages_pin_count))) {
 205                 err = -EBUSY;
 206                 goto unlock;
 207         }
 208
 209         i915_gem_object_release_mmap_offset(obj);
 210
 211         /*
 212          * ->put_pages might need to allocate memory for the bit17 swizzle
 213          * array, hence protect them from being reaped by removing them from gtt
 214          * lists early.
 215          */
 216         pages = __i915_gem_object_unset_pages(obj);
 217
 218         /*
 219          * XXX Temporary hijinx to avoid updating all backends to handle
 220          * NULL pages. In the future, when we have more asynchronous
 221          * get_pages backends we should be better able to handle the
 222          * cancellation of the async task in a more uniform manner.
 223          */
 224         if (!pages && !i915_gem_object_needs_async_cancel(obj))
 225                 pages = ERR_PTR(-EINVAL);
 226
 227         if (!IS_ERR(pages))
 228                 obj->ops->put_pages(obj, pages);
 229
 230         err = 0;
 231 unlock:
 232         mutex_unlock(&obj->mm.lock);
 233
 234         return err;
 235 }
 236
 237 static inline pte_t iomap_pte(resource_size_t base,
 238                               dma_addr_t offset,
 239                               pgprot_t prot)
 240 {
 241         return pte_mkspecial(pfn_pte((base + offset) >> PAGE_SHIFT, prot));
 242 }
 243
 244 /* The 'mapping' part of i915_gem_object_pin_map() below */
 245 static void *i915_gem_object_map(struct drm_i915_gem_object *obj,
 246                                  enum i915_map_type type)
 247 {
 248         unsigned long n_pte = obj->base.size >> PAGE_SHIFT;
 249         struct sg_table *sgt = obj->mm.pages;
 250         pte_t *stack[32], **mem;
 251         struct vm_struct *area;
 252         pgprot_t pgprot;
 253
 254         if (!i915_gem_object_has_struct_page(obj) && type != I915_MAP_WC)
 255                 return NULL;
 256
 257         /* A single page can always be kmapped */
 258         if (n_pte == 1 && type == I915_MAP_WB)
 259                 return kmap(sg_page(sgt->sgl));
 260
 261         mem = stack;
 262         if (n_pte > ARRAY_SIZE(stack)) {
 263                 /* Too big for stack -- allocate temporary array instead */
 264                 mem = kvmalloc_array(n_pte, sizeof(*mem), GFP_KERNEL);
 265                 if (!mem)
 266                         return NULL;
 267         }
 268
 269         area = alloc_vm_area(obj->base.size, mem);
 270         if (!area) {
 271                 if (mem != stack)
 272                         kvfree(mem);
 273                 return NULL;
 274         }
 275
 276         switch (type) {
 277         default:
 278                 MISSING_CASE(type);
 279                 /* fallthrough - to use PAGE_KERNEL anyway */
 280         case I915_MAP_WB:
 281                 pgprot = PAGE_KERNEL;
 282                 break;
 283         case I915_MAP_WC:
 284                 pgprot = pgprot_writecombine(PAGE_KERNEL_IO);
 285                 break;
 286         }
 287
 288         if (i915_gem_object_has_struct_page(obj)) {
 289                 struct sgt_iter iter;
 290                 struct page *page;
 291                 pte_t **ptes = mem;
 292
 293                 for_each_sgt_page(page, iter, sgt)
 294                         **ptes++ = mk_pte(page, pgprot);
 295         } else {
 296                 resource_size_t iomap;
 297                 struct sgt_iter iter;
 298                 pte_t **ptes = mem;
 299                 dma_addr_t addr;
 300
 301                 iomap = obj->mm.region->iomap.base;
 302                 iomap -= obj->mm.region->region.start;
 303
 304                 for_each_sgt_daddr(addr, iter, sgt)
 305                         **ptes++ = iomap_pte(iomap, addr, pgprot);
 306         }
 307
 308         if (mem != stack)
 309                 kvfree(mem);
 310
 311         return area->addr;
 312 }
 313
 314 /* get, pin, and map the pages of the object into kernel space */
 315 void *i915_gem_object_pin_map(struct drm_i915_gem_object *obj,
 316                               enum i915_map_type type)
 317 {
 318         enum i915_map_type has_type;
 319         unsigned int flags;
 320         bool pinned;
 321         void *ptr;
 322         int err;
 323
 324         flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE | I915_GEM_OBJECT_HAS_IOMEM;
 325         if (!i915_gem_object_type_has(obj, flags))
 326                 return ERR_PTR(-ENXIO);
 327
 328         err = mutex_lock_interruptible_nested(&obj->mm.lock, I915_MM_GET_PAGES);
 329         if (err)
 330                 return ERR_PTR(err);
 331
 332         pinned = !(type & I915_MAP_OVERRIDE);
 333         type &= ~I915_MAP_OVERRIDE;
 334
 335         if (!atomic_inc_not_zero(&obj->mm.pages_pin_count)) {
 336                 if (unlikely(!i915_gem_object_has_pages(obj))) {
 337                         GEM_BUG_ON(i915_gem_object_has_pinned_pages(obj));
 338
 339                         err = ____i915_gem_object_get_pages(obj);
 340                         if (err)
 341                                 goto err_unlock;
 342
 343                         smp_mb__before_atomic();
 344                 }
 345                 atomic_inc(&obj->mm.pages_pin_count);
 346                 pinned = false;
 347         }
 348         GEM_BUG_ON(!i915_gem_object_has_pages(obj));
 349
 350         ptr = page_unpack_bits(obj->mm.mapping, &has_type);
 351         if (ptr && has_type != type) {
 352                 if (pinned) {
 353                         err = -EBUSY;
 354                         goto err_unpin;
 355                 }
 356
 357                 unmap_object(obj, ptr);
 358
 359                 ptr = obj->mm.mapping = NULL;
 360         }
 361
 362         if (!ptr) {
 363                 ptr = i915_gem_object_map(obj, type);
 364                 if (!ptr) {
 365                         err = -ENOMEM;
 366                         goto err_unpin;
 367                 }
 368
 369                 obj->mm.mapping = page_pack_bits(ptr, type);
 370         }
 371
 372 out_unlock:
 373         mutex_unlock(&obj->mm.lock);
 374         return ptr;
 375
 376 err_unpin:
 377         atomic_dec(&obj->mm.pages_pin_count);
 378 err_unlock:
 379         ptr = ERR_PTR(err);
 380         goto out_unlock;
 381 }
 382
 383 void __i915_gem_object_flush_map(struct drm_i915_gem_object *obj,
 384                                  unsigned long offset,
 385                                  unsigned long size)
 386 {
 387         enum i915_map_type has_type;
 388         void *ptr;
 389
 390         GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
 391         GEM_BUG_ON(range_overflows_t(typeof(obj->base.size),
 392                                      offset, size, obj->base.size));
 393
 394         obj->mm.dirty = true;
 395
 396         if (obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_WRITE)
 397                 return;
 398
 399         ptr = page_unpack_bits(obj->mm.mapping, &has_type);
 400         if (has_type == I915_MAP_WC)
 401                 return;
 402
 403         drm_clflush_virt_range(ptr + offset, size);
 404         if (size == obj->base.size) {
 405                 obj->write_domain &= ~I915_GEM_DOMAIN_CPU;
 406                 obj->cache_dirty = false;
 407         }
 408 }
 409
 410 struct scatterlist *
 411 i915_gem_object_get_sg(struct drm_i915_gem_object *obj,
 412                        unsigned int n,
 413                        unsigned int *offset)
 414 {
 415         struct i915_gem_object_page_iter *iter = &obj->mm.get_page;
 416         struct scatterlist *sg;
 417         unsigned int idx, count;
 418
 419         might_sleep();
 420         GEM_BUG_ON(n >= obj->base.size >> PAGE_SHIFT);
 421         GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
 422
 423         /* As we iterate forward through the sg, we record each entry in a
 424          * radixtree for quick repeated (backwards) lookups. If we have seen
 425          * this index previously, we will have an entry for it.
 426          *
 427          * Initial lookup is O(N), but this is amortized to O(1) for
 428          * sequential page access (where each new request is consecutive
 429          * to the previous one). Repeated lookups are O(lg(obj->base.size)),
 430          * i.e. O(1) with a large constant!
 431          */
 432         if (n < READ_ONCE(iter->sg_idx))
 433                 goto lookup;
 434
 435         mutex_lock(&iter->lock);
 436
 437         /* We prefer to reuse the last sg so that repeated lookup of this
 438          * (or the subsequent) sg are fast - comparing against the last
 439          * sg is faster than going through the radixtree.
 440          */
 441
 442         sg = iter->sg_pos;
 443         idx = iter->sg_idx;
 444         count = __sg_page_count(sg);
 445
 446         while (idx + count <= n) {
 447                 void *entry;
 448                 unsigned long i;
 449                 int ret;
 450
 451                 /* If we cannot allocate and insert this entry, or the
 452                  * individual pages from this range, cancel updating the
 453                  * sg_idx so that on this lookup we are forced to linearly
 454                  * scan onwards, but on future lookups we will try the
 455                  * insertion again (in which case we need to be careful of
 456                  * the error return reporting that we have already inserted
 457                  * this index).
 458                  */
 459                 ret = radix_tree_insert(&iter->radix, idx, sg);
 460                 if (ret && ret != -EEXIST)
 461                         goto scan;
 462
 463                 entry = xa_mk_value(idx);
 464                 for (i = 1; i < count; i++) {
 465                         ret = radix_tree_insert(&iter->radix, idx + i, entry);
 466                         if (ret && ret != -EEXIST)
 467                                 goto scan;
 468                 }
 469
 470                 idx += count;
 471                 sg = ____sg_next(sg);
 472                 count = __sg_page_count(sg);
 473         }
 474
 475 scan:
 476         iter->sg_pos = sg;
 477         iter->sg_idx = idx;
 478
 479         mutex_unlock(&iter->lock);
 480
 481         if (unlikely(n < idx)) /* insertion completed by another thread */
 482                 goto lookup;
 483
 484         /* In case we failed to insert the entry into the radixtree, we need
 485          * to look beyond the current sg.
 486          */
 487         while (idx + count <= n) {
 488                 idx += count;
 489                 sg = ____sg_next(sg);
 490                 count = __sg_page_count(sg);
 491         }
 492
 493         *offset = n - idx;
 494         return sg;
 495
 496 lookup:
 497         rcu_read_lock();
 498
 499         sg = radix_tree_lookup(&iter->radix, n);
 500         GEM_BUG_ON(!sg);
 501
 502         /* If this index is in the middle of multi-page sg entry,
 503          * the radix tree will contain a value entry that points
 504          * to the start of that range. We will return the pointer to
 505          * the base page and the offset of this page within the
 506          * sg entry's range.
 507          */
 508         *offset = 0;
 509         if (unlikely(xa_is_value(sg))) {
 510                 unsigned long base = xa_to_value(sg);
 511
 512                 sg = radix_tree_lookup(&iter->radix, base);
 513                 GEM_BUG_ON(!sg);
 514
 515                 *offset = n - base;
 516         }
 517
 518         rcu_read_unlock();
 519
 520         return sg;
 521 }
 522
 523 struct page *
 524 i915_gem_object_get_page(struct drm_i915_gem_object *obj, unsigned int n)
 525 {
 526         struct scatterlist *sg;
 527         unsigned int offset;
 528
 529         GEM_BUG_ON(!i915_gem_object_has_struct_page(obj));
 530
 531         sg = i915_gem_object_get_sg(obj, n, &offset);
 532         return nth_page(sg_page(sg), offset);
 533 }
 534
 535 /* Like i915_gem_object_get_page(), but mark the returned page dirty */
 536 struct page *
 537 i915_gem_object_get_dirty_page(struct drm_i915_gem_object *obj,
 538                                unsigned int n)
 539 {
 540         struct page *page;
 541
 542         page = i915_gem_object_get_page(obj, n);
 543         if (!obj->mm.dirty)
 544                 set_page_dirty(page);
 545
 546         return page;
 547 }
 548
 549 dma_addr_t
 550 i915_gem_object_get_dma_address_len(struct drm_i915_gem_object *obj,
 551                                     unsigned long n,
 552                                     unsigned int *len)
 553 {
 554         struct scatterlist *sg;
 555         unsigned int offset;
 556
 557         sg = i915_gem_object_get_sg(obj, n, &offset);
 558
 559         if (len)
 560                 *len = sg_dma_len(sg) - (offset << PAGE_SHIFT);
 561
 562         return sg_dma_address(sg) + (offset << PAGE_SHIFT);
 563 }
 564
 565 dma_addr_t
 566 i915_gem_object_get_dma_address(struct drm_i915_gem_object *obj,
 567                                 unsigned long n)
 568 {
 569         return i915_gem_object_get_dma_address_len(obj, n, NULL);
 570 }