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 11 void __i915_gem_object_set_pages(struct drm_i915_gem_object *obj, 12 struct sg_table *pages, 13 unsigned int sg_page_sizes) 14 { 15 struct drm_i915_private *i915 = to_i915(obj->base.dev); 16 unsigned long supported = INTEL_INFO(i915)->page_sizes; 17 int i; 18 19 lockdep_assert_held(&obj->mm.lock); 20 21 /* Make the pages coherent with the GPU (flushing any swapin). */ 22 if (obj->cache_dirty) { 23 obj->write_domain = 0; 24 if (i915_gem_object_has_struct_page(obj)) 25 drm_clflush_sg(pages); 26 obj->cache_dirty = false; 27 } 28 29 obj->mm.get_page.sg_pos = pages->sgl; 30 obj->mm.get_page.sg_idx = 0; 31 32 obj->mm.pages = pages; 33 34 if (i915_gem_object_is_tiled(obj) && 35 i915->quirks & QUIRK_PIN_SWIZZLED_PAGES) { 36 GEM_BUG_ON(obj->mm.quirked); 37 __i915_gem_object_pin_pages(obj); 38 obj->mm.quirked = true; 39 } 40 41 GEM_BUG_ON(!sg_page_sizes); 42 obj->mm.page_sizes.phys = sg_page_sizes; 43 44 /* 45 * Calculate the supported page-sizes which fit into the given 46 * sg_page_sizes. This will give us the page-sizes which we may be able 47 * to use opportunistically when later inserting into the GTT. For 48 * example if phys=2G, then in theory we should be able to use 1G, 2M, 49 * 64K or 4K pages, although in practice this will depend on a number of 50 * other factors. 51 */ 52 obj->mm.page_sizes.sg = 0; 53 for_each_set_bit(i, &supported, ilog2(I915_GTT_MAX_PAGE_SIZE) + 1) { 54 if (obj->mm.page_sizes.phys & ~0u << i) 55 obj->mm.page_sizes.sg |= BIT(i); 56 } 57 GEM_BUG_ON(!HAS_PAGE_SIZES(i915, obj->mm.page_sizes.sg)); 58 59 if (i915_gem_object_is_shrinkable(obj)) { 60 struct list_head *list; 61 unsigned long flags; 62 63 spin_lock_irqsave(&i915->mm.obj_lock, flags); 64 65 i915->mm.shrink_count++; 66 i915->mm.shrink_memory += obj->base.size; 67 68 if (obj->mm.madv != I915_MADV_WILLNEED) 69 list = &i915->mm.purge_list; 70 else 71 list = &i915->mm.shrink_list; 72 list_add_tail(&obj->mm.link, list); 73 74 spin_unlock_irqrestore(&i915->mm.obj_lock, flags); 75 } 76 } 77 78 int ____i915_gem_object_get_pages(struct drm_i915_gem_object *obj) 79 { 80 int err; 81 82 if (unlikely(obj->mm.madv != I915_MADV_WILLNEED)) { 83 DRM_DEBUG("Attempting to obtain a purgeable object\n"); 84 return -EFAULT; 85 } 86 87 err = obj->ops->get_pages(obj); 88 GEM_BUG_ON(!err && !i915_gem_object_has_pages(obj)); 89 90 return err; 91 } 92 93 /* Ensure that the associated pages are gathered from the backing storage 94 * and pinned into our object. i915_gem_object_pin_pages() may be called 95 * multiple times before they are released by a single call to 96 * i915_gem_object_unpin_pages() - once the pages are no longer referenced 97 * either as a result of memory pressure (reaping pages under the shrinker) 98 * or as the object is itself released. 99 */ 100 int __i915_gem_object_get_pages(struct drm_i915_gem_object *obj) 101 { 102 int err; 103 104 err = mutex_lock_interruptible(&obj->mm.lock); 105 if (err) 106 return err; 107 108 if (unlikely(!i915_gem_object_has_pages(obj))) { 109 GEM_BUG_ON(i915_gem_object_has_pinned_pages(obj)); 110 111 err = ____i915_gem_object_get_pages(obj); 112 if (err) 113 goto unlock; 114 115 smp_mb__before_atomic(); 116 } 117 atomic_inc(&obj->mm.pages_pin_count); 118 119 unlock: 120 mutex_unlock(&obj->mm.lock); 121 return err; 122 } 123 124 /* Immediately discard the backing storage */ 125 void i915_gem_object_truncate(struct drm_i915_gem_object *obj) 126 { 127 drm_gem_free_mmap_offset(&obj->base); 128 if (obj->ops->truncate) 129 obj->ops->truncate(obj); 130 } 131 132 /* Try to discard unwanted pages */ 133 void i915_gem_object_writeback(struct drm_i915_gem_object *obj) 134 { 135 lockdep_assert_held(&obj->mm.lock); 136 GEM_BUG_ON(i915_gem_object_has_pages(obj)); 137 138 if (obj->ops->writeback) 139 obj->ops->writeback(obj); 140 } 141 142 static void __i915_gem_object_reset_page_iter(struct drm_i915_gem_object *obj) 143 { 144 struct radix_tree_iter iter; 145 void __rcu **slot; 146 147 rcu_read_lock(); 148 radix_tree_for_each_slot(slot, &obj->mm.get_page.radix, &iter, 0) 149 radix_tree_delete(&obj->mm.get_page.radix, iter.index); 150 rcu_read_unlock(); 151 } 152 153 struct sg_table * 154 __i915_gem_object_unset_pages(struct drm_i915_gem_object *obj) 155 { 156 struct sg_table *pages; 157 158 pages = fetch_and_zero(&obj->mm.pages); 159 if (IS_ERR_OR_NULL(pages)) 160 return pages; 161 162 i915_gem_object_make_unshrinkable(obj); 163 164 if (obj->mm.mapping) { 165 void *ptr; 166 167 ptr = page_mask_bits(obj->mm.mapping); 168 if (is_vmalloc_addr(ptr)) 169 vunmap(ptr); 170 else 171 kunmap(kmap_to_page(ptr)); 172 173 obj->mm.mapping = NULL; 174 } 175 176 __i915_gem_object_reset_page_iter(obj); 177 obj->mm.page_sizes.phys = obj->mm.page_sizes.sg = 0; 178 179 return pages; 180 } 181 182 int __i915_gem_object_put_pages(struct drm_i915_gem_object *obj, 183 enum i915_mm_subclass subclass) 184 { 185 struct sg_table *pages; 186 int err; 187 188 if (i915_gem_object_has_pinned_pages(obj)) 189 return -EBUSY; 190 191 GEM_BUG_ON(atomic_read(&obj->bind_count)); 192 193 /* May be called by shrinker from within get_pages() (on another bo) */ 194 mutex_lock_nested(&obj->mm.lock, subclass); 195 if (unlikely(atomic_read(&obj->mm.pages_pin_count))) { 196 err = -EBUSY; 197 goto unlock; 198 } 199 200 /* 201 * ->put_pages might need to allocate memory for the bit17 swizzle 202 * array, hence protect them from being reaped by removing them from gtt 203 * lists early. 204 */ 205 pages = __i915_gem_object_unset_pages(obj); 206 207 /* 208 * XXX Temporary hijinx to avoid updating all backends to handle 209 * NULL pages. In the future, when we have more asynchronous 210 * get_pages backends we should be better able to handle the 211 * cancellation of the async task in a more uniform manner. 212 */ 213 if (!pages && !i915_gem_object_needs_async_cancel(obj)) 214 pages = ERR_PTR(-EINVAL); 215 216 if (!IS_ERR(pages)) 217 obj->ops->put_pages(obj, pages); 218 219 err = 0; 220 unlock: 221 mutex_unlock(&obj->mm.lock); 222 223 return err; 224 } 225 226 /* The 'mapping' part of i915_gem_object_pin_map() below */ 227 static void *i915_gem_object_map(const struct drm_i915_gem_object *obj, 228 enum i915_map_type type) 229 { 230 unsigned long n_pages = obj->base.size >> PAGE_SHIFT; 231 struct sg_table *sgt = obj->mm.pages; 232 struct sgt_iter sgt_iter; 233 struct page *page; 234 struct page *stack_pages[32]; 235 struct page **pages = stack_pages; 236 unsigned long i = 0; 237 pgprot_t pgprot; 238 void *addr; 239 240 /* A single page can always be kmapped */ 241 if (n_pages == 1 && type == I915_MAP_WB) 242 return kmap(sg_page(sgt->sgl)); 243 244 if (n_pages > ARRAY_SIZE(stack_pages)) { 245 /* Too big for stack -- allocate temporary array instead */ 246 pages = kvmalloc_array(n_pages, sizeof(*pages), GFP_KERNEL); 247 if (!pages) 248 return NULL; 249 } 250 251 for_each_sgt_page(page, sgt_iter, sgt) 252 pages[i++] = page; 253 254 /* Check that we have the expected number of pages */ 255 GEM_BUG_ON(i != n_pages); 256 257 switch (type) { 258 default: 259 MISSING_CASE(type); 260 /* fallthrough - to use PAGE_KERNEL anyway */ 261 case I915_MAP_WB: 262 pgprot = PAGE_KERNEL; 263 break; 264 case I915_MAP_WC: 265 pgprot = pgprot_writecombine(PAGE_KERNEL_IO); 266 break; 267 } 268 addr = vmap(pages, n_pages, 0, pgprot); 269 270 if (pages != stack_pages) 271 kvfree(pages); 272 273 return addr; 274 } 275 276 /* get, pin, and map the pages of the object into kernel space */ 277 void *i915_gem_object_pin_map(struct drm_i915_gem_object *obj, 278 enum i915_map_type type) 279 { 280 enum i915_map_type has_type; 281 bool pinned; 282 void *ptr; 283 int err; 284 285 if (unlikely(!i915_gem_object_has_struct_page(obj))) 286 return ERR_PTR(-ENXIO); 287 288 err = mutex_lock_interruptible(&obj->mm.lock); 289 if (err) 290 return ERR_PTR(err); 291 292 pinned = !(type & I915_MAP_OVERRIDE); 293 type &= ~I915_MAP_OVERRIDE; 294 295 if (!atomic_inc_not_zero(&obj->mm.pages_pin_count)) { 296 if (unlikely(!i915_gem_object_has_pages(obj))) { 297 GEM_BUG_ON(i915_gem_object_has_pinned_pages(obj)); 298 299 err = ____i915_gem_object_get_pages(obj); 300 if (err) 301 goto err_unlock; 302 303 smp_mb__before_atomic(); 304 } 305 atomic_inc(&obj->mm.pages_pin_count); 306 pinned = false; 307 } 308 GEM_BUG_ON(!i915_gem_object_has_pages(obj)); 309 310 ptr = page_unpack_bits(obj->mm.mapping, &has_type); 311 if (ptr && has_type != type) { 312 if (pinned) { 313 err = -EBUSY; 314 goto err_unpin; 315 } 316 317 if (is_vmalloc_addr(ptr)) 318 vunmap(ptr); 319 else 320 kunmap(kmap_to_page(ptr)); 321 322 ptr = obj->mm.mapping = NULL; 323 } 324 325 if (!ptr) { 326 ptr = i915_gem_object_map(obj, type); 327 if (!ptr) { 328 err = -ENOMEM; 329 goto err_unpin; 330 } 331 332 obj->mm.mapping = page_pack_bits(ptr, type); 333 } 334 335 out_unlock: 336 mutex_unlock(&obj->mm.lock); 337 return ptr; 338 339 err_unpin: 340 atomic_dec(&obj->mm.pages_pin_count); 341 err_unlock: 342 ptr = ERR_PTR(err); 343 goto out_unlock; 344 } 345 346 void __i915_gem_object_flush_map(struct drm_i915_gem_object *obj, 347 unsigned long offset, 348 unsigned long size) 349 { 350 enum i915_map_type has_type; 351 void *ptr; 352 353 GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj)); 354 GEM_BUG_ON(range_overflows_t(typeof(obj->base.size), 355 offset, size, obj->base.size)); 356 357 obj->mm.dirty = true; 358 359 if (obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_WRITE) 360 return; 361 362 ptr = page_unpack_bits(obj->mm.mapping, &has_type); 363 if (has_type == I915_MAP_WC) 364 return; 365 366 drm_clflush_virt_range(ptr + offset, size); 367 if (size == obj->base.size) { 368 obj->write_domain &= ~I915_GEM_DOMAIN_CPU; 369 obj->cache_dirty = false; 370 } 371 } 372 373 struct scatterlist * 374 i915_gem_object_get_sg(struct drm_i915_gem_object *obj, 375 unsigned int n, 376 unsigned int *offset) 377 { 378 struct i915_gem_object_page_iter *iter = &obj->mm.get_page; 379 struct scatterlist *sg; 380 unsigned int idx, count; 381 382 might_sleep(); 383 GEM_BUG_ON(n >= obj->base.size >> PAGE_SHIFT); 384 GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj)); 385 386 /* As we iterate forward through the sg, we record each entry in a 387 * radixtree for quick repeated (backwards) lookups. If we have seen 388 * this index previously, we will have an entry for it. 389 * 390 * Initial lookup is O(N), but this is amortized to O(1) for 391 * sequential page access (where each new request is consecutive 392 * to the previous one). Repeated lookups are O(lg(obj->base.size)), 393 * i.e. O(1) with a large constant! 394 */ 395 if (n < READ_ONCE(iter->sg_idx)) 396 goto lookup; 397 398 mutex_lock(&iter->lock); 399 400 /* We prefer to reuse the last sg so that repeated lookup of this 401 * (or the subsequent) sg are fast - comparing against the last 402 * sg is faster than going through the radixtree. 403 */ 404 405 sg = iter->sg_pos; 406 idx = iter->sg_idx; 407 count = __sg_page_count(sg); 408 409 while (idx + count <= n) { 410 void *entry; 411 unsigned long i; 412 int ret; 413 414 /* If we cannot allocate and insert this entry, or the 415 * individual pages from this range, cancel updating the 416 * sg_idx so that on this lookup we are forced to linearly 417 * scan onwards, but on future lookups we will try the 418 * insertion again (in which case we need to be careful of 419 * the error return reporting that we have already inserted 420 * this index). 421 */ 422 ret = radix_tree_insert(&iter->radix, idx, sg); 423 if (ret && ret != -EEXIST) 424 goto scan; 425 426 entry = xa_mk_value(idx); 427 for (i = 1; i < count; i++) { 428 ret = radix_tree_insert(&iter->radix, idx + i, entry); 429 if (ret && ret != -EEXIST) 430 goto scan; 431 } 432 433 idx += count; 434 sg = ____sg_next(sg); 435 count = __sg_page_count(sg); 436 } 437 438 scan: 439 iter->sg_pos = sg; 440 iter->sg_idx = idx; 441 442 mutex_unlock(&iter->lock); 443 444 if (unlikely(n < idx)) /* insertion completed by another thread */ 445 goto lookup; 446 447 /* In case we failed to insert the entry into the radixtree, we need 448 * to look beyond the current sg. 449 */ 450 while (idx + count <= n) { 451 idx += count; 452 sg = ____sg_next(sg); 453 count = __sg_page_count(sg); 454 } 455 456 *offset = n - idx; 457 return sg; 458 459 lookup: 460 rcu_read_lock(); 461 462 sg = radix_tree_lookup(&iter->radix, n); 463 GEM_BUG_ON(!sg); 464 465 /* If this index is in the middle of multi-page sg entry, 466 * the radix tree will contain a value entry that points 467 * to the start of that range. We will return the pointer to 468 * the base page and the offset of this page within the 469 * sg entry's range. 470 */ 471 *offset = 0; 472 if (unlikely(xa_is_value(sg))) { 473 unsigned long base = xa_to_value(sg); 474 475 sg = radix_tree_lookup(&iter->radix, base); 476 GEM_BUG_ON(!sg); 477 478 *offset = n - base; 479 } 480 481 rcu_read_unlock(); 482 483 return sg; 484 } 485 486 struct page * 487 i915_gem_object_get_page(struct drm_i915_gem_object *obj, unsigned int n) 488 { 489 struct scatterlist *sg; 490 unsigned int offset; 491 492 GEM_BUG_ON(!i915_gem_object_has_struct_page(obj)); 493 494 sg = i915_gem_object_get_sg(obj, n, &offset); 495 return nth_page(sg_page(sg), offset); 496 } 497 498 /* Like i915_gem_object_get_page(), but mark the returned page dirty */ 499 struct page * 500 i915_gem_object_get_dirty_page(struct drm_i915_gem_object *obj, 501 unsigned int n) 502 { 503 struct page *page; 504 505 page = i915_gem_object_get_page(obj, n); 506 if (!obj->mm.dirty) 507 set_page_dirty(page); 508 509 return page; 510 } 511 512 dma_addr_t 513 i915_gem_object_get_dma_address_len(struct drm_i915_gem_object *obj, 514 unsigned long n, 515 unsigned int *len) 516 { 517 struct scatterlist *sg; 518 unsigned int offset; 519 520 sg = i915_gem_object_get_sg(obj, n, &offset); 521 522 if (len) 523 *len = sg_dma_len(sg) - (offset << PAGE_SHIFT); 524 525 return sg_dma_address(sg) + (offset << PAGE_SHIFT); 526 } 527 528 dma_addr_t 529 i915_gem_object_get_dma_address(struct drm_i915_gem_object *obj, 530 unsigned long n) 531 { 532 return i915_gem_object_get_dma_address_len(obj, n, NULL); 533 } 534