| 1 | // SPDX-License-Identifier: GPL-2.0 OR MIT |
| 2 | /* |
| 3 | * Copyright 2020 Advanced Micro Devices, Inc. |
| 4 | * |
| 5 | * Permission is hereby granted, free of charge, to any person obtaining a |
| 6 | * copy of this software and associated documentation files (the "Software"), |
| 7 | * to deal in the Software without restriction, including without limitation |
| 8 | * the rights to use, copy, modify, merge, publish, distribute, sublicense, |
| 9 | * and/or sell copies of the Software, and to permit persons to whom the |
| 10 | * Software is furnished to do so, subject to the following conditions: |
| 11 | * |
| 12 | * The above copyright notice and this permission notice shall be included in |
| 13 | * all copies or substantial portions of the Software. |
| 14 | * |
| 15 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 16 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 17 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL |
| 18 | * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR |
| 19 | * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, |
| 20 | * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR |
| 21 | * OTHER DEALINGS IN THE SOFTWARE. |
| 22 | * |
| 23 | * Authors: Christian König |
| 24 | */ |
| 25 | |
| 26 | /* Pooling of allocated pages is necessary because changing the caching |
| 27 | * attributes on x86 of the linear mapping requires a costly cross CPU TLB |
| 28 | * invalidate for those addresses. |
| 29 | * |
| 30 | * Additional to that allocations from the DMA coherent API are pooled as well |
| 31 | * cause they are rather slow compared to alloc_pages+map. |
| 32 | */ |
| 33 | |
| 34 | #include <linux/export.h> |
| 35 | #include <linux/module.h> |
| 36 | #include <linux/dma-mapping.h> |
| 37 | #include <linux/debugfs.h> |
| 38 | #include <linux/highmem.h> |
| 39 | #include <linux/sched/mm.h> |
| 40 | |
| 41 | #ifdef CONFIG_X86 |
| 42 | #include <asm/set_memory.h> |
| 43 | #endif |
| 44 | |
| 45 | #include <drm/ttm/ttm_backup.h> |
| 46 | #include <drm/ttm/ttm_pool.h> |
| 47 | #include <drm/ttm/ttm_tt.h> |
| 48 | #include <drm/ttm/ttm_bo.h> |
| 49 | |
| 50 | #include "ttm_module.h" |
| 51 | #include "ttm_pool_internal.h" |
| 52 | |
| 53 | #ifdef CONFIG_FAULT_INJECTION |
| 54 | #include <linux/fault-inject.h> |
| 55 | static DECLARE_FAULT_ATTR(backup_fault_inject); |
| 56 | #else |
| 57 | #define should_fail(...) false |
| 58 | #endif |
| 59 | |
| 60 | /** |
| 61 | * struct ttm_pool_dma - Helper object for coherent DMA mappings |
| 62 | * |
| 63 | * @addr: original DMA address returned for the mapping |
| 64 | * @vaddr: original vaddr return for the mapping and order in the lower bits |
| 65 | */ |
| 66 | struct ttm_pool_dma { |
| 67 | dma_addr_t addr; |
| 68 | unsigned long vaddr; |
| 69 | }; |
| 70 | |
| 71 | /** |
| 72 | * struct ttm_pool_alloc_state - Current state of the tt page allocation process |
| 73 | * @pages: Pointer to the next tt page pointer to populate. |
| 74 | * @caching_divide: Pointer to the first page pointer whose page has a staged but |
| 75 | * not committed caching transition from write-back to @tt_caching. |
| 76 | * @dma_addr: Pointer to the next tt dma_address entry to populate if any. |
| 77 | * @remaining_pages: Remaining pages to populate. |
| 78 | * @tt_caching: The requested cpu-caching for the pages allocated. |
| 79 | */ |
| 80 | struct ttm_pool_alloc_state { |
| 81 | struct page **pages; |
| 82 | struct page **caching_divide; |
| 83 | dma_addr_t *dma_addr; |
| 84 | pgoff_t remaining_pages; |
| 85 | enum ttm_caching tt_caching; |
| 86 | }; |
| 87 | |
| 88 | /** |
| 89 | * struct ttm_pool_tt_restore - State representing restore from backup |
| 90 | * @pool: The pool used for page allocation while restoring. |
| 91 | * @snapshot_alloc: A snapshot of the most recent struct ttm_pool_alloc_state. |
| 92 | * @alloced_page: Pointer to the page most recently allocated from a pool or system. |
| 93 | * @first_dma: The dma address corresponding to @alloced_page if dma_mapping |
| 94 | * is requested. |
| 95 | * @alloced_pages: The number of allocated pages present in the struct ttm_tt |
| 96 | * page vector from this restore session. |
| 97 | * @restored_pages: The number of 4K pages restored for @alloced_page (which |
| 98 | * is typically a multi-order page). |
| 99 | * @page_caching: The struct ttm_tt requested caching |
| 100 | * @order: The order of @alloced_page. |
| 101 | * |
| 102 | * Recovery from backup might fail when we've recovered less than the |
| 103 | * full ttm_tt. In order not to loose any data (yet), keep information |
| 104 | * around that allows us to restart a failed ttm backup recovery. |
| 105 | */ |
| 106 | struct ttm_pool_tt_restore { |
| 107 | struct ttm_pool *pool; |
| 108 | struct ttm_pool_alloc_state snapshot_alloc; |
| 109 | struct page *alloced_page; |
| 110 | dma_addr_t first_dma; |
| 111 | pgoff_t alloced_pages; |
| 112 | pgoff_t restored_pages; |
| 113 | enum ttm_caching page_caching; |
| 114 | unsigned int order; |
| 115 | }; |
| 116 | |
| 117 | static unsigned long page_pool_size; |
| 118 | |
| 119 | MODULE_PARM_DESC(page_pool_size, "Number of pages in the WC/UC/DMA pool" ); |
| 120 | module_param(page_pool_size, ulong, 0644); |
| 121 | |
| 122 | static atomic_long_t allocated_pages; |
| 123 | |
| 124 | static struct ttm_pool_type global_write_combined[NR_PAGE_ORDERS]; |
| 125 | static struct ttm_pool_type global_uncached[NR_PAGE_ORDERS]; |
| 126 | |
| 127 | static struct ttm_pool_type global_dma32_write_combined[NR_PAGE_ORDERS]; |
| 128 | static struct ttm_pool_type global_dma32_uncached[NR_PAGE_ORDERS]; |
| 129 | |
| 130 | static spinlock_t shrinker_lock; |
| 131 | static struct list_head shrinker_list; |
| 132 | static struct shrinker *mm_shrinker; |
| 133 | static DECLARE_RWSEM(pool_shrink_rwsem); |
| 134 | |
| 135 | /* Allocate pages of size 1 << order with the given gfp_flags */ |
| 136 | static struct page *ttm_pool_alloc_page(struct ttm_pool *pool, gfp_t gfp_flags, |
| 137 | unsigned int order) |
| 138 | { |
| 139 | const unsigned int beneficial_order = ttm_pool_beneficial_order(pool); |
| 140 | unsigned long attr = DMA_ATTR_FORCE_CONTIGUOUS; |
| 141 | struct ttm_pool_dma *dma; |
| 142 | struct page *p; |
| 143 | void *vaddr; |
| 144 | |
| 145 | /* Don't set the __GFP_COMP flag for higher order allocations. |
| 146 | * Mapping pages directly into an userspace process and calling |
| 147 | * put_page() on a TTM allocated page is illegal. |
| 148 | */ |
| 149 | if (order) |
| 150 | gfp_flags |= __GFP_NOMEMALLOC | __GFP_NORETRY | __GFP_NOWARN | |
| 151 | __GFP_THISNODE; |
| 152 | |
| 153 | /* |
| 154 | * Do not add latency to the allocation path for allocations orders |
| 155 | * device tolds us do not bring them additional performance gains. |
| 156 | */ |
| 157 | if (beneficial_order && order > beneficial_order) |
| 158 | gfp_flags &= ~__GFP_DIRECT_RECLAIM; |
| 159 | |
| 160 | if (!ttm_pool_uses_dma_alloc(pool)) { |
| 161 | p = alloc_pages_node(pool->nid, gfp_flags, order); |
| 162 | if (p) |
| 163 | p->private = order; |
| 164 | return p; |
| 165 | } |
| 166 | |
| 167 | dma = kmalloc(sizeof(*dma), GFP_KERNEL); |
| 168 | if (!dma) |
| 169 | return NULL; |
| 170 | |
| 171 | if (order) |
| 172 | attr |= DMA_ATTR_NO_WARN; |
| 173 | |
| 174 | vaddr = dma_alloc_attrs(dev: pool->dev, size: (1ULL << order) * PAGE_SIZE, |
| 175 | dma_handle: &dma->addr, flag: gfp_flags, attrs: attr); |
| 176 | if (!vaddr) |
| 177 | goto error_free; |
| 178 | |
| 179 | /* TODO: This is an illegal abuse of the DMA API, but we need to rework |
| 180 | * TTM page fault handling and extend the DMA API to clean this up. |
| 181 | */ |
| 182 | if (is_vmalloc_addr(x: vaddr)) |
| 183 | p = vmalloc_to_page(addr: vaddr); |
| 184 | else |
| 185 | p = virt_to_page(vaddr); |
| 186 | |
| 187 | dma->vaddr = (unsigned long)vaddr | order; |
| 188 | p->private = (unsigned long)dma; |
| 189 | return p; |
| 190 | |
| 191 | error_free: |
| 192 | kfree(objp: dma); |
| 193 | return NULL; |
| 194 | } |
| 195 | |
| 196 | /* Reset the caching and pages of size 1 << order */ |
| 197 | static void ttm_pool_free_page(struct ttm_pool *pool, enum ttm_caching caching, |
| 198 | unsigned int order, struct page *p) |
| 199 | { |
| 200 | unsigned long attr = DMA_ATTR_FORCE_CONTIGUOUS; |
| 201 | struct ttm_pool_dma *dma; |
| 202 | void *vaddr; |
| 203 | |
| 204 | #ifdef CONFIG_X86 |
| 205 | /* We don't care that set_pages_wb is inefficient here. This is only |
| 206 | * used when we have to shrink and CPU overhead is irrelevant then. |
| 207 | */ |
| 208 | if (caching != ttm_cached && !PageHighMem(page: p)) |
| 209 | set_pages_wb(page: p, numpages: 1 << order); |
| 210 | #endif |
| 211 | |
| 212 | if (!pool || !ttm_pool_uses_dma_alloc(pool)) { |
| 213 | __free_pages(page: p, order); |
| 214 | return; |
| 215 | } |
| 216 | |
| 217 | if (order) |
| 218 | attr |= DMA_ATTR_NO_WARN; |
| 219 | |
| 220 | dma = (void *)p->private; |
| 221 | vaddr = (void *)(dma->vaddr & PAGE_MASK); |
| 222 | dma_free_attrs(dev: pool->dev, size: (1UL << order) * PAGE_SIZE, cpu_addr: vaddr, dma_handle: dma->addr, |
| 223 | attrs: attr); |
| 224 | kfree(objp: dma); |
| 225 | } |
| 226 | |
| 227 | /* Apply any cpu-caching deferred during page allocation */ |
| 228 | static int ttm_pool_apply_caching(struct ttm_pool_alloc_state *alloc) |
| 229 | { |
| 230 | #ifdef CONFIG_X86 |
| 231 | unsigned int num_pages = alloc->pages - alloc->caching_divide; |
| 232 | |
| 233 | if (!num_pages) |
| 234 | return 0; |
| 235 | |
| 236 | switch (alloc->tt_caching) { |
| 237 | case ttm_cached: |
| 238 | break; |
| 239 | case ttm_write_combined: |
| 240 | return set_pages_array_wc(pages: alloc->caching_divide, addrinarray: num_pages); |
| 241 | case ttm_uncached: |
| 242 | return set_pages_array_uc(pages: alloc->caching_divide, addrinarray: num_pages); |
| 243 | } |
| 244 | #endif |
| 245 | alloc->caching_divide = alloc->pages; |
| 246 | return 0; |
| 247 | } |
| 248 | |
| 249 | /* DMA Map pages of 1 << order size and return the resulting dma_address. */ |
| 250 | static int ttm_pool_map(struct ttm_pool *pool, unsigned int order, |
| 251 | struct page *p, dma_addr_t *dma_addr) |
| 252 | { |
| 253 | dma_addr_t addr; |
| 254 | |
| 255 | if (ttm_pool_uses_dma_alloc(pool)) { |
| 256 | struct ttm_pool_dma *dma = (void *)p->private; |
| 257 | |
| 258 | addr = dma->addr; |
| 259 | } else { |
| 260 | size_t size = (1ULL << order) * PAGE_SIZE; |
| 261 | |
| 262 | addr = dma_map_page(pool->dev, p, 0, size, DMA_BIDIRECTIONAL); |
| 263 | if (dma_mapping_error(dev: pool->dev, dma_addr: addr)) |
| 264 | return -EFAULT; |
| 265 | } |
| 266 | |
| 267 | *dma_addr = addr; |
| 268 | |
| 269 | return 0; |
| 270 | } |
| 271 | |
| 272 | /* Unmap pages of 1 << order size */ |
| 273 | static void ttm_pool_unmap(struct ttm_pool *pool, dma_addr_t dma_addr, |
| 274 | unsigned int num_pages) |
| 275 | { |
| 276 | /* Unmapped while freeing the page */ |
| 277 | if (ttm_pool_uses_dma_alloc(pool)) |
| 278 | return; |
| 279 | |
| 280 | dma_unmap_page(pool->dev, dma_addr, (long)num_pages << PAGE_SHIFT, |
| 281 | DMA_BIDIRECTIONAL); |
| 282 | } |
| 283 | |
| 284 | /* Give pages into a specific pool_type */ |
| 285 | static void ttm_pool_type_give(struct ttm_pool_type *pt, struct page *p) |
| 286 | { |
| 287 | unsigned int i, num_pages = 1 << pt->order; |
| 288 | |
| 289 | for (i = 0; i < num_pages; ++i) { |
| 290 | if (PageHighMem(page: p)) |
| 291 | clear_highpage(page: p + i); |
| 292 | else |
| 293 | clear_page(page_address(p + i)); |
| 294 | } |
| 295 | |
| 296 | spin_lock(lock: &pt->lock); |
| 297 | list_add(new: &p->lru, head: &pt->pages); |
| 298 | spin_unlock(lock: &pt->lock); |
| 299 | atomic_long_add(i: 1 << pt->order, v: &allocated_pages); |
| 300 | } |
| 301 | |
| 302 | /* Take pages from a specific pool_type, return NULL when nothing available */ |
| 303 | static struct page *ttm_pool_type_take(struct ttm_pool_type *pt) |
| 304 | { |
| 305 | struct page *p; |
| 306 | |
| 307 | spin_lock(lock: &pt->lock); |
| 308 | p = list_first_entry_or_null(&pt->pages, typeof(*p), lru); |
| 309 | if (p) { |
| 310 | atomic_long_sub(i: 1 << pt->order, v: &allocated_pages); |
| 311 | list_del(entry: &p->lru); |
| 312 | } |
| 313 | spin_unlock(lock: &pt->lock); |
| 314 | |
| 315 | return p; |
| 316 | } |
| 317 | |
| 318 | /* Initialize and add a pool type to the global shrinker list */ |
| 319 | static void ttm_pool_type_init(struct ttm_pool_type *pt, struct ttm_pool *pool, |
| 320 | enum ttm_caching caching, unsigned int order) |
| 321 | { |
| 322 | pt->pool = pool; |
| 323 | pt->caching = caching; |
| 324 | pt->order = order; |
| 325 | spin_lock_init(&pt->lock); |
| 326 | INIT_LIST_HEAD(list: &pt->pages); |
| 327 | |
| 328 | spin_lock(lock: &shrinker_lock); |
| 329 | list_add_tail(new: &pt->shrinker_list, head: &shrinker_list); |
| 330 | spin_unlock(lock: &shrinker_lock); |
| 331 | } |
| 332 | |
| 333 | /* Remove a pool_type from the global shrinker list and free all pages */ |
| 334 | static void ttm_pool_type_fini(struct ttm_pool_type *pt) |
| 335 | { |
| 336 | struct page *p; |
| 337 | |
| 338 | spin_lock(lock: &shrinker_lock); |
| 339 | list_del(entry: &pt->shrinker_list); |
| 340 | spin_unlock(lock: &shrinker_lock); |
| 341 | |
| 342 | while ((p = ttm_pool_type_take(pt))) |
| 343 | ttm_pool_free_page(pool: pt->pool, caching: pt->caching, order: pt->order, p); |
| 344 | } |
| 345 | |
| 346 | /* Return the pool_type to use for the given caching and order */ |
| 347 | static struct ttm_pool_type *ttm_pool_select_type(struct ttm_pool *pool, |
| 348 | enum ttm_caching caching, |
| 349 | unsigned int order) |
| 350 | { |
| 351 | if (ttm_pool_uses_dma_alloc(pool)) |
| 352 | return &pool->caching[caching].orders[order]; |
| 353 | |
| 354 | #ifdef CONFIG_X86 |
| 355 | switch (caching) { |
| 356 | case ttm_write_combined: |
| 357 | if (pool->nid != NUMA_NO_NODE) |
| 358 | return &pool->caching[caching].orders[order]; |
| 359 | |
| 360 | if (ttm_pool_uses_dma32(pool)) |
| 361 | return &global_dma32_write_combined[order]; |
| 362 | |
| 363 | return &global_write_combined[order]; |
| 364 | case ttm_uncached: |
| 365 | if (pool->nid != NUMA_NO_NODE) |
| 366 | return &pool->caching[caching].orders[order]; |
| 367 | |
| 368 | if (ttm_pool_uses_dma32(pool)) |
| 369 | return &global_dma32_uncached[order]; |
| 370 | |
| 371 | return &global_uncached[order]; |
| 372 | default: |
| 373 | break; |
| 374 | } |
| 375 | #endif |
| 376 | |
| 377 | return NULL; |
| 378 | } |
| 379 | |
| 380 | /* Free pages using the global shrinker list */ |
| 381 | static unsigned int ttm_pool_shrink(void) |
| 382 | { |
| 383 | struct ttm_pool_type *pt; |
| 384 | unsigned int num_pages; |
| 385 | struct page *p; |
| 386 | |
| 387 | down_read(sem: &pool_shrink_rwsem); |
| 388 | spin_lock(lock: &shrinker_lock); |
| 389 | pt = list_first_entry(&shrinker_list, typeof(*pt), shrinker_list); |
| 390 | list_move_tail(list: &pt->shrinker_list, head: &shrinker_list); |
| 391 | spin_unlock(lock: &shrinker_lock); |
| 392 | |
| 393 | p = ttm_pool_type_take(pt); |
| 394 | if (p) { |
| 395 | ttm_pool_free_page(pool: pt->pool, caching: pt->caching, order: pt->order, p); |
| 396 | num_pages = 1 << pt->order; |
| 397 | } else { |
| 398 | num_pages = 0; |
| 399 | } |
| 400 | up_read(sem: &pool_shrink_rwsem); |
| 401 | |
| 402 | return num_pages; |
| 403 | } |
| 404 | |
| 405 | /* Return the allocation order based for a page */ |
| 406 | static unsigned int ttm_pool_page_order(struct ttm_pool *pool, struct page *p) |
| 407 | { |
| 408 | if (ttm_pool_uses_dma_alloc(pool)) { |
| 409 | struct ttm_pool_dma *dma = (void *)p->private; |
| 410 | |
| 411 | return dma->vaddr & ~PAGE_MASK; |
| 412 | } |
| 413 | |
| 414 | return p->private; |
| 415 | } |
| 416 | |
| 417 | /* |
| 418 | * Split larger pages so that we can free each PAGE_SIZE page as soon |
| 419 | * as it has been backed up, in order to avoid memory pressure during |
| 420 | * reclaim. |
| 421 | */ |
| 422 | static void ttm_pool_split_for_swap(struct ttm_pool *pool, struct page *p) |
| 423 | { |
| 424 | unsigned int order = ttm_pool_page_order(pool, p); |
| 425 | pgoff_t nr; |
| 426 | |
| 427 | if (!order) |
| 428 | return; |
| 429 | |
| 430 | split_page(page: p, order); |
| 431 | nr = 1UL << order; |
| 432 | while (nr--) |
| 433 | (p++)->private = 0; |
| 434 | } |
| 435 | |
| 436 | /** |
| 437 | * DOC: Partial backup and restoration of a struct ttm_tt. |
| 438 | * |
| 439 | * Swapout using ttm_backup_backup_page() and swapin using |
| 440 | * ttm_backup_copy_page() may fail. |
| 441 | * The former most likely due to lack of swap-space or memory, the latter due |
| 442 | * to lack of memory or because of signal interruption during waits. |
| 443 | * |
| 444 | * Backup failure is easily handled by using a ttm_tt pages vector that holds |
| 445 | * both backup handles and page pointers. This has to be taken into account when |
| 446 | * restoring such a ttm_tt from backup, and when freeing it while backed up. |
| 447 | * When restoring, for simplicity, new pages are actually allocated from the |
| 448 | * pool and the contents of any old pages are copied in and then the old pages |
| 449 | * are released. |
| 450 | * |
| 451 | * For restoration failures, the struct ttm_pool_tt_restore holds sufficient state |
| 452 | * to be able to resume an interrupted restore, and that structure is freed once |
| 453 | * the restoration is complete. If the struct ttm_tt is destroyed while there |
| 454 | * is a valid struct ttm_pool_tt_restore attached, that is also properly taken |
| 455 | * care of. |
| 456 | */ |
| 457 | |
| 458 | /* Is restore ongoing for the currently allocated page? */ |
| 459 | static bool ttm_pool_restore_valid(const struct ttm_pool_tt_restore *restore) |
| 460 | { |
| 461 | return restore && restore->restored_pages < (1 << restore->order); |
| 462 | } |
| 463 | |
| 464 | /* DMA unmap and free a multi-order page, either to the relevant pool or to system. */ |
| 465 | static pgoff_t ttm_pool_unmap_and_free(struct ttm_pool *pool, struct page *page, |
| 466 | const dma_addr_t *dma_addr, enum ttm_caching caching) |
| 467 | { |
| 468 | struct ttm_pool_type *pt = NULL; |
| 469 | unsigned int order; |
| 470 | pgoff_t nr; |
| 471 | |
| 472 | if (pool) { |
| 473 | order = ttm_pool_page_order(pool, p: page); |
| 474 | nr = (1UL << order); |
| 475 | if (dma_addr) |
| 476 | ttm_pool_unmap(pool, dma_addr: *dma_addr, num_pages: nr); |
| 477 | |
| 478 | pt = ttm_pool_select_type(pool, caching, order); |
| 479 | } else { |
| 480 | order = page->private; |
| 481 | nr = (1UL << order); |
| 482 | } |
| 483 | |
| 484 | if (pt) |
| 485 | ttm_pool_type_give(pt, p: page); |
| 486 | else |
| 487 | ttm_pool_free_page(pool, caching, order, p: page); |
| 488 | |
| 489 | return nr; |
| 490 | } |
| 491 | |
| 492 | /* Populate the page-array using the most recent allocated multi-order page. */ |
| 493 | static void ttm_pool_allocated_page_commit(struct page *allocated, |
| 494 | dma_addr_t first_dma, |
| 495 | struct ttm_pool_alloc_state *alloc, |
| 496 | pgoff_t nr) |
| 497 | { |
| 498 | pgoff_t i; |
| 499 | |
| 500 | for (i = 0; i < nr; ++i) |
| 501 | *alloc->pages++ = allocated++; |
| 502 | |
| 503 | alloc->remaining_pages -= nr; |
| 504 | |
| 505 | if (!alloc->dma_addr) |
| 506 | return; |
| 507 | |
| 508 | for (i = 0; i < nr; ++i) { |
| 509 | *alloc->dma_addr++ = first_dma; |
| 510 | first_dma += PAGE_SIZE; |
| 511 | } |
| 512 | } |
| 513 | |
| 514 | /* |
| 515 | * When restoring, restore backed-up content to the newly allocated page and |
| 516 | * if successful, populate the page-table and dma-address arrays. |
| 517 | */ |
| 518 | static int ttm_pool_restore_commit(struct ttm_pool_tt_restore *restore, |
| 519 | struct file *backup, |
| 520 | const struct ttm_operation_ctx *ctx, |
| 521 | struct ttm_pool_alloc_state *alloc) |
| 522 | |
| 523 | { |
| 524 | pgoff_t i, nr = 1UL << restore->order; |
| 525 | struct page **first_page = alloc->pages; |
| 526 | struct page *p; |
| 527 | int ret = 0; |
| 528 | |
| 529 | for (i = restore->restored_pages; i < nr; ++i) { |
| 530 | p = first_page[i]; |
| 531 | if (ttm_backup_page_ptr_is_handle(page: p)) { |
| 532 | unsigned long handle = ttm_backup_page_ptr_to_handle(page: p); |
| 533 | |
| 534 | if (IS_ENABLED(CONFIG_FAULT_INJECTION) && ctx->interruptible && |
| 535 | should_fail(attr: &backup_fault_inject, size: 1)) { |
| 536 | ret = -EINTR; |
| 537 | break; |
| 538 | } |
| 539 | |
| 540 | if (handle == 0) { |
| 541 | restore->restored_pages++; |
| 542 | continue; |
| 543 | } |
| 544 | |
| 545 | ret = ttm_backup_copy_page(backup, dst: restore->alloced_page + i, |
| 546 | handle, intr: ctx->interruptible); |
| 547 | if (ret) |
| 548 | break; |
| 549 | |
| 550 | ttm_backup_drop(backup, handle); |
| 551 | } else if (p) { |
| 552 | /* |
| 553 | * We could probably avoid splitting the old page |
| 554 | * using clever logic, but ATM we don't care, as |
| 555 | * we prioritize releasing memory ASAP. Note that |
| 556 | * here, the old retained page is always write-back |
| 557 | * cached. |
| 558 | */ |
| 559 | ttm_pool_split_for_swap(pool: restore->pool, p); |
| 560 | copy_highpage(to: restore->alloced_page + i, from: p); |
| 561 | __free_pages(page: p, order: 0); |
| 562 | } |
| 563 | |
| 564 | restore->restored_pages++; |
| 565 | first_page[i] = ttm_backup_handle_to_page_ptr(handle: 0); |
| 566 | } |
| 567 | |
| 568 | if (ret) { |
| 569 | if (!restore->restored_pages) { |
| 570 | dma_addr_t *dma_addr = alloc->dma_addr ? &restore->first_dma : NULL; |
| 571 | |
| 572 | ttm_pool_unmap_and_free(pool: restore->pool, page: restore->alloced_page, |
| 573 | dma_addr, caching: restore->page_caching); |
| 574 | restore->restored_pages = nr; |
| 575 | } |
| 576 | return ret; |
| 577 | } |
| 578 | |
| 579 | ttm_pool_allocated_page_commit(allocated: restore->alloced_page, first_dma: restore->first_dma, |
| 580 | alloc, nr); |
| 581 | if (restore->page_caching == alloc->tt_caching || PageHighMem(page: restore->alloced_page)) |
| 582 | alloc->caching_divide = alloc->pages; |
| 583 | restore->snapshot_alloc = *alloc; |
| 584 | restore->alloced_pages += nr; |
| 585 | |
| 586 | return 0; |
| 587 | } |
| 588 | |
| 589 | /* If restoring, save information needed for ttm_pool_restore_commit(). */ |
| 590 | static void |
| 591 | ttm_pool_page_allocated_restore(struct ttm_pool *pool, unsigned int order, |
| 592 | struct page *p, |
| 593 | enum ttm_caching page_caching, |
| 594 | dma_addr_t first_dma, |
| 595 | struct ttm_pool_tt_restore *restore, |
| 596 | const struct ttm_pool_alloc_state *alloc) |
| 597 | { |
| 598 | restore->pool = pool; |
| 599 | restore->order = order; |
| 600 | restore->restored_pages = 0; |
| 601 | restore->page_caching = page_caching; |
| 602 | restore->first_dma = first_dma; |
| 603 | restore->alloced_page = p; |
| 604 | restore->snapshot_alloc = *alloc; |
| 605 | } |
| 606 | |
| 607 | /* |
| 608 | * Called when we got a page, either from a pool or newly allocated. |
| 609 | * if needed, dma map the page and populate the dma address array. |
| 610 | * Populate the page address array. |
| 611 | * If the caching is consistent, update any deferred caching. Otherwise |
| 612 | * stage this page for an upcoming deferred caching update. |
| 613 | */ |
| 614 | static int ttm_pool_page_allocated(struct ttm_pool *pool, unsigned int order, |
| 615 | struct page *p, enum ttm_caching page_caching, |
| 616 | struct ttm_pool_alloc_state *alloc, |
| 617 | struct ttm_pool_tt_restore *restore) |
| 618 | { |
| 619 | bool caching_consistent; |
| 620 | dma_addr_t first_dma; |
| 621 | int r = 0; |
| 622 | |
| 623 | caching_consistent = (page_caching == alloc->tt_caching) || PageHighMem(page: p); |
| 624 | |
| 625 | if (caching_consistent) { |
| 626 | r = ttm_pool_apply_caching(alloc); |
| 627 | if (r) |
| 628 | return r; |
| 629 | } |
| 630 | |
| 631 | if (alloc->dma_addr) { |
| 632 | r = ttm_pool_map(pool, order, p, dma_addr: &first_dma); |
| 633 | if (r) |
| 634 | return r; |
| 635 | } |
| 636 | |
| 637 | if (restore) { |
| 638 | ttm_pool_page_allocated_restore(pool, order, p, page_caching, |
| 639 | first_dma, restore, alloc); |
| 640 | } else { |
| 641 | ttm_pool_allocated_page_commit(allocated: p, first_dma, alloc, nr: 1UL << order); |
| 642 | |
| 643 | if (caching_consistent) |
| 644 | alloc->caching_divide = alloc->pages; |
| 645 | } |
| 646 | |
| 647 | return 0; |
| 648 | } |
| 649 | |
| 650 | /** |
| 651 | * ttm_pool_free_range() - Free a range of TTM pages |
| 652 | * @pool: The pool used for allocating. |
| 653 | * @tt: The struct ttm_tt holding the page pointers. |
| 654 | * @caching: The page caching mode used by the range. |
| 655 | * @start_page: index for first page to free. |
| 656 | * @end_page: index for last page to free + 1. |
| 657 | * |
| 658 | * During allocation the ttm_tt page-vector may be populated with ranges of |
| 659 | * pages with different attributes if allocation hit an error without being |
| 660 | * able to completely fulfill the allocation. This function can be used |
| 661 | * to free these individual ranges. |
| 662 | */ |
| 663 | static void ttm_pool_free_range(struct ttm_pool *pool, struct ttm_tt *tt, |
| 664 | enum ttm_caching caching, |
| 665 | pgoff_t start_page, pgoff_t end_page) |
| 666 | { |
| 667 | struct page **pages = &tt->pages[start_page]; |
| 668 | struct file *backup = tt->backup; |
| 669 | pgoff_t i, nr; |
| 670 | |
| 671 | for (i = start_page; i < end_page; i += nr, pages += nr) { |
| 672 | struct page *p = *pages; |
| 673 | |
| 674 | nr = 1; |
| 675 | if (ttm_backup_page_ptr_is_handle(page: p)) { |
| 676 | unsigned long handle = ttm_backup_page_ptr_to_handle(page: p); |
| 677 | |
| 678 | if (handle != 0) |
| 679 | ttm_backup_drop(backup, handle); |
| 680 | } else if (p) { |
| 681 | dma_addr_t *dma_addr = tt->dma_address ? |
| 682 | tt->dma_address + i : NULL; |
| 683 | |
| 684 | nr = ttm_pool_unmap_and_free(pool, page: p, dma_addr, caching); |
| 685 | } |
| 686 | } |
| 687 | } |
| 688 | |
| 689 | static void ttm_pool_alloc_state_init(const struct ttm_tt *tt, |
| 690 | struct ttm_pool_alloc_state *alloc) |
| 691 | { |
| 692 | alloc->pages = tt->pages; |
| 693 | alloc->caching_divide = tt->pages; |
| 694 | alloc->dma_addr = tt->dma_address; |
| 695 | alloc->remaining_pages = tt->num_pages; |
| 696 | alloc->tt_caching = tt->caching; |
| 697 | } |
| 698 | |
| 699 | /* |
| 700 | * Find a suitable allocation order based on highest desired order |
| 701 | * and number of remaining pages |
| 702 | */ |
| 703 | static unsigned int ttm_pool_alloc_find_order(unsigned int highest, |
| 704 | const struct ttm_pool_alloc_state *alloc) |
| 705 | { |
| 706 | return min_t(unsigned int, highest, __fls(alloc->remaining_pages)); |
| 707 | } |
| 708 | |
| 709 | static int __ttm_pool_alloc(struct ttm_pool *pool, struct ttm_tt *tt, |
| 710 | const struct ttm_operation_ctx *ctx, |
| 711 | struct ttm_pool_alloc_state *alloc, |
| 712 | struct ttm_pool_tt_restore *restore) |
| 713 | { |
| 714 | enum ttm_caching page_caching; |
| 715 | gfp_t gfp_flags = GFP_USER; |
| 716 | pgoff_t caching_divide; |
| 717 | unsigned int order; |
| 718 | bool allow_pools; |
| 719 | struct page *p; |
| 720 | int r; |
| 721 | |
| 722 | WARN_ON(!alloc->remaining_pages || ttm_tt_is_populated(tt)); |
| 723 | WARN_ON(alloc->dma_addr && !pool->dev); |
| 724 | |
| 725 | if (tt->page_flags & TTM_TT_FLAG_ZERO_ALLOC) |
| 726 | gfp_flags |= __GFP_ZERO; |
| 727 | |
| 728 | if (ctx->gfp_retry_mayfail) |
| 729 | gfp_flags |= __GFP_RETRY_MAYFAIL; |
| 730 | |
| 731 | if (ttm_pool_uses_dma32(pool)) |
| 732 | gfp_flags |= GFP_DMA32; |
| 733 | else |
| 734 | gfp_flags |= GFP_HIGHUSER; |
| 735 | |
| 736 | page_caching = tt->caching; |
| 737 | allow_pools = true; |
| 738 | for (order = ttm_pool_alloc_find_order(MAX_PAGE_ORDER, alloc); |
| 739 | alloc->remaining_pages; |
| 740 | order = ttm_pool_alloc_find_order(highest: order, alloc)) { |
| 741 | struct ttm_pool_type *pt; |
| 742 | |
| 743 | /* First, try to allocate a page from a pool if one exists. */ |
| 744 | p = NULL; |
| 745 | pt = ttm_pool_select_type(pool, caching: page_caching, order); |
| 746 | if (pt && allow_pools) |
| 747 | p = ttm_pool_type_take(pt); |
| 748 | /* |
| 749 | * If that fails or previously failed, allocate from system. |
| 750 | * Note that this also disallows additional pool allocations using |
| 751 | * write-back cached pools of the same order. Consider removing |
| 752 | * that behaviour. |
| 753 | */ |
| 754 | if (!p) { |
| 755 | page_caching = ttm_cached; |
| 756 | allow_pools = false; |
| 757 | p = ttm_pool_alloc_page(pool, gfp_flags, order); |
| 758 | } |
| 759 | /* If that fails, lower the order if possible and retry. */ |
| 760 | if (!p) { |
| 761 | if (order) { |
| 762 | --order; |
| 763 | page_caching = tt->caching; |
| 764 | allow_pools = true; |
| 765 | continue; |
| 766 | } |
| 767 | r = -ENOMEM; |
| 768 | goto error_free_all; |
| 769 | } |
| 770 | r = ttm_pool_page_allocated(pool, order, p, page_caching, alloc, |
| 771 | restore); |
| 772 | if (r) |
| 773 | goto error_free_page; |
| 774 | |
| 775 | if (ttm_pool_restore_valid(restore)) { |
| 776 | r = ttm_pool_restore_commit(restore, backup: tt->backup, ctx, alloc); |
| 777 | if (r) |
| 778 | goto error_free_all; |
| 779 | } |
| 780 | } |
| 781 | |
| 782 | r = ttm_pool_apply_caching(alloc); |
| 783 | if (r) |
| 784 | goto error_free_all; |
| 785 | |
| 786 | kfree(objp: tt->restore); |
| 787 | tt->restore = NULL; |
| 788 | |
| 789 | return 0; |
| 790 | |
| 791 | error_free_page: |
| 792 | ttm_pool_free_page(pool, caching: page_caching, order, p); |
| 793 | |
| 794 | error_free_all: |
| 795 | if (tt->restore) |
| 796 | return r; |
| 797 | |
| 798 | caching_divide = alloc->caching_divide - tt->pages; |
| 799 | ttm_pool_free_range(pool, tt, caching: tt->caching, start_page: 0, end_page: caching_divide); |
| 800 | ttm_pool_free_range(pool, tt, caching: ttm_cached, start_page: caching_divide, |
| 801 | end_page: tt->num_pages - alloc->remaining_pages); |
| 802 | |
| 803 | return r; |
| 804 | } |
| 805 | |
| 806 | /** |
| 807 | * ttm_pool_alloc - Fill a ttm_tt object |
| 808 | * |
| 809 | * @pool: ttm_pool to use |
| 810 | * @tt: ttm_tt object to fill |
| 811 | * @ctx: operation context |
| 812 | * |
| 813 | * Fill the ttm_tt object with pages and also make sure to DMA map them when |
| 814 | * necessary. |
| 815 | * |
| 816 | * Returns: 0 on successe, negative error code otherwise. |
| 817 | */ |
| 818 | int ttm_pool_alloc(struct ttm_pool *pool, struct ttm_tt *tt, |
| 819 | struct ttm_operation_ctx *ctx) |
| 820 | { |
| 821 | struct ttm_pool_alloc_state alloc; |
| 822 | |
| 823 | if (WARN_ON(ttm_tt_is_backed_up(tt))) |
| 824 | return -EINVAL; |
| 825 | |
| 826 | ttm_pool_alloc_state_init(tt, alloc: &alloc); |
| 827 | |
| 828 | return __ttm_pool_alloc(pool, tt, ctx, alloc: &alloc, NULL); |
| 829 | } |
| 830 | EXPORT_SYMBOL(ttm_pool_alloc); |
| 831 | |
| 832 | /** |
| 833 | * ttm_pool_restore_and_alloc - Fill a ttm_tt, restoring previously backed-up |
| 834 | * content. |
| 835 | * |
| 836 | * @pool: ttm_pool to use |
| 837 | * @tt: ttm_tt object to fill |
| 838 | * @ctx: operation context |
| 839 | * |
| 840 | * Fill the ttm_tt object with pages and also make sure to DMA map them when |
| 841 | * necessary. Read in backed-up content. |
| 842 | * |
| 843 | * Returns: 0 on successe, negative error code otherwise. |
| 844 | */ |
| 845 | int ttm_pool_restore_and_alloc(struct ttm_pool *pool, struct ttm_tt *tt, |
| 846 | const struct ttm_operation_ctx *ctx) |
| 847 | { |
| 848 | struct ttm_pool_alloc_state alloc; |
| 849 | |
| 850 | if (WARN_ON(!ttm_tt_is_backed_up(tt))) |
| 851 | return -EINVAL; |
| 852 | |
| 853 | if (!tt->restore) { |
| 854 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
| 855 | |
| 856 | ttm_pool_alloc_state_init(tt, alloc: &alloc); |
| 857 | if (ctx->gfp_retry_mayfail) |
| 858 | gfp |= __GFP_RETRY_MAYFAIL; |
| 859 | |
| 860 | tt->restore = kzalloc(sizeof(*tt->restore), gfp); |
| 861 | if (!tt->restore) |
| 862 | return -ENOMEM; |
| 863 | |
| 864 | tt->restore->snapshot_alloc = alloc; |
| 865 | tt->restore->pool = pool; |
| 866 | tt->restore->restored_pages = 1; |
| 867 | } else { |
| 868 | struct ttm_pool_tt_restore *restore = tt->restore; |
| 869 | int ret; |
| 870 | |
| 871 | alloc = restore->snapshot_alloc; |
| 872 | if (ttm_pool_restore_valid(restore: tt->restore)) { |
| 873 | ret = ttm_pool_restore_commit(restore, backup: tt->backup, ctx, alloc: &alloc); |
| 874 | if (ret) |
| 875 | return ret; |
| 876 | } |
| 877 | if (!alloc.remaining_pages) |
| 878 | return 0; |
| 879 | } |
| 880 | |
| 881 | return __ttm_pool_alloc(pool, tt, ctx, alloc: &alloc, restore: tt->restore); |
| 882 | } |
| 883 | |
| 884 | /** |
| 885 | * ttm_pool_free - Free the backing pages from a ttm_tt object |
| 886 | * |
| 887 | * @pool: Pool to give pages back to. |
| 888 | * @tt: ttm_tt object to unpopulate |
| 889 | * |
| 890 | * Give the packing pages back to a pool or free them |
| 891 | */ |
| 892 | void ttm_pool_free(struct ttm_pool *pool, struct ttm_tt *tt) |
| 893 | { |
| 894 | ttm_pool_free_range(pool, tt, caching: tt->caching, start_page: 0, end_page: tt->num_pages); |
| 895 | |
| 896 | while (atomic_long_read(v: &allocated_pages) > page_pool_size) |
| 897 | ttm_pool_shrink(); |
| 898 | } |
| 899 | EXPORT_SYMBOL(ttm_pool_free); |
| 900 | |
| 901 | /** |
| 902 | * ttm_pool_drop_backed_up() - Release content of a swapped-out struct ttm_tt |
| 903 | * @tt: The struct ttm_tt. |
| 904 | * |
| 905 | * Release handles with associated content or any remaining pages of |
| 906 | * a backed-up struct ttm_tt. |
| 907 | */ |
| 908 | void ttm_pool_drop_backed_up(struct ttm_tt *tt) |
| 909 | { |
| 910 | struct ttm_pool_tt_restore *restore; |
| 911 | pgoff_t start_page = 0; |
| 912 | |
| 913 | WARN_ON(!ttm_tt_is_backed_up(tt)); |
| 914 | |
| 915 | restore = tt->restore; |
| 916 | |
| 917 | /* |
| 918 | * Unmap and free any uncommitted restore page. |
| 919 | * any tt page-array backup entries already read back has |
| 920 | * been cleared already |
| 921 | */ |
| 922 | if (ttm_pool_restore_valid(restore)) { |
| 923 | dma_addr_t *dma_addr = tt->dma_address ? &restore->first_dma : NULL; |
| 924 | |
| 925 | ttm_pool_unmap_and_free(pool: restore->pool, page: restore->alloced_page, |
| 926 | dma_addr, caching: restore->page_caching); |
| 927 | restore->restored_pages = 1UL << restore->order; |
| 928 | } |
| 929 | |
| 930 | /* |
| 931 | * If a restore is ongoing, part of the tt pages may have a |
| 932 | * caching different than writeback. |
| 933 | */ |
| 934 | if (restore) { |
| 935 | pgoff_t mid = restore->snapshot_alloc.caching_divide - tt->pages; |
| 936 | |
| 937 | start_page = restore->alloced_pages; |
| 938 | WARN_ON(mid > start_page); |
| 939 | /* Pages that might be dma-mapped and non-cached */ |
| 940 | ttm_pool_free_range(pool: restore->pool, tt, caching: tt->caching, |
| 941 | start_page: 0, end_page: mid); |
| 942 | /* Pages that might be dma-mapped but cached */ |
| 943 | ttm_pool_free_range(pool: restore->pool, tt, caching: ttm_cached, |
| 944 | start_page: mid, end_page: restore->alloced_pages); |
| 945 | kfree(objp: restore); |
| 946 | tt->restore = NULL; |
| 947 | } |
| 948 | |
| 949 | ttm_pool_free_range(NULL, tt, caching: ttm_cached, start_page, end_page: tt->num_pages); |
| 950 | } |
| 951 | |
| 952 | /** |
| 953 | * ttm_pool_backup() - Back up or purge a struct ttm_tt |
| 954 | * @pool: The pool used when allocating the struct ttm_tt. |
| 955 | * @tt: The struct ttm_tt. |
| 956 | * @flags: Flags to govern the backup behaviour. |
| 957 | * |
| 958 | * Back up or purge a struct ttm_tt. If @purge is true, then |
| 959 | * all pages will be freed directly to the system rather than to the pool |
| 960 | * they were allocated from, making the function behave similarly to |
| 961 | * ttm_pool_free(). If @purge is false the pages will be backed up instead, |
| 962 | * exchanged for handles. |
| 963 | * A subsequent call to ttm_pool_restore_and_alloc() will then read back the content and |
| 964 | * a subsequent call to ttm_pool_drop_backed_up() will drop it. |
| 965 | * If backup of a page fails for whatever reason, @ttm will still be |
| 966 | * partially backed up, retaining those pages for which backup fails. |
| 967 | * In that case, this function can be retried, possibly after freeing up |
| 968 | * memory resources. |
| 969 | * |
| 970 | * Return: Number of pages actually backed up or freed, or negative |
| 971 | * error code on error. |
| 972 | */ |
| 973 | long ttm_pool_backup(struct ttm_pool *pool, struct ttm_tt *tt, |
| 974 | const struct ttm_backup_flags *flags) |
| 975 | { |
| 976 | struct file *backup = tt->backup; |
| 977 | struct page *page; |
| 978 | unsigned long handle; |
| 979 | gfp_t alloc_gfp; |
| 980 | gfp_t gfp; |
| 981 | int ret = 0; |
| 982 | pgoff_t shrunken = 0; |
| 983 | pgoff_t i, num_pages; |
| 984 | |
| 985 | if (WARN_ON(ttm_tt_is_backed_up(tt))) |
| 986 | return -EINVAL; |
| 987 | |
| 988 | if ((!ttm_backup_bytes_avail() && !flags->purge) || |
| 989 | ttm_pool_uses_dma_alloc(pool) || ttm_tt_is_backed_up(tt)) |
| 990 | return -EBUSY; |
| 991 | |
| 992 | #ifdef CONFIG_X86 |
| 993 | /* Anything returned to the system needs to be cached. */ |
| 994 | if (tt->caching != ttm_cached) |
| 995 | set_pages_array_wb(pages: tt->pages, addrinarray: tt->num_pages); |
| 996 | #endif |
| 997 | |
| 998 | if (tt->dma_address || flags->purge) { |
| 999 | for (i = 0; i < tt->num_pages; i += num_pages) { |
| 1000 | unsigned int order; |
| 1001 | |
| 1002 | page = tt->pages[i]; |
| 1003 | if (unlikely(!page)) { |
| 1004 | num_pages = 1; |
| 1005 | continue; |
| 1006 | } |
| 1007 | |
| 1008 | order = ttm_pool_page_order(pool, p: page); |
| 1009 | num_pages = 1UL << order; |
| 1010 | if (tt->dma_address) |
| 1011 | ttm_pool_unmap(pool, dma_addr: tt->dma_address[i], |
| 1012 | num_pages); |
| 1013 | if (flags->purge) { |
| 1014 | shrunken += num_pages; |
| 1015 | page->private = 0; |
| 1016 | __free_pages(page, order); |
| 1017 | memset(tt->pages + i, 0, |
| 1018 | num_pages * sizeof(*tt->pages)); |
| 1019 | } |
| 1020 | } |
| 1021 | } |
| 1022 | |
| 1023 | if (flags->purge) |
| 1024 | return shrunken; |
| 1025 | |
| 1026 | if (ttm_pool_uses_dma32(pool)) |
| 1027 | gfp = GFP_DMA32; |
| 1028 | else |
| 1029 | gfp = GFP_HIGHUSER; |
| 1030 | |
| 1031 | alloc_gfp = GFP_KERNEL | __GFP_HIGH | __GFP_NOWARN | __GFP_RETRY_MAYFAIL; |
| 1032 | |
| 1033 | num_pages = tt->num_pages; |
| 1034 | |
| 1035 | /* Pretend doing fault injection by shrinking only half of the pages. */ |
| 1036 | if (IS_ENABLED(CONFIG_FAULT_INJECTION) && should_fail(attr: &backup_fault_inject, size: 1)) |
| 1037 | num_pages = DIV_ROUND_UP(num_pages, 2); |
| 1038 | |
| 1039 | for (i = 0; i < num_pages; ++i) { |
| 1040 | s64 shandle; |
| 1041 | |
| 1042 | page = tt->pages[i]; |
| 1043 | if (unlikely(!page)) |
| 1044 | continue; |
| 1045 | |
| 1046 | ttm_pool_split_for_swap(pool, p: page); |
| 1047 | |
| 1048 | shandle = ttm_backup_backup_page(backup, page, writeback: flags->writeback, idx: i, |
| 1049 | page_gfp: gfp, alloc_gfp); |
| 1050 | if (shandle < 0) { |
| 1051 | /* We allow partially shrunken tts */ |
| 1052 | ret = shandle; |
| 1053 | break; |
| 1054 | } |
| 1055 | handle = shandle; |
| 1056 | tt->pages[i] = ttm_backup_handle_to_page_ptr(handle); |
| 1057 | put_page(page); |
| 1058 | shrunken++; |
| 1059 | } |
| 1060 | |
| 1061 | return shrunken ? shrunken : ret; |
| 1062 | } |
| 1063 | |
| 1064 | /** |
| 1065 | * ttm_pool_init - Initialize a pool |
| 1066 | * |
| 1067 | * @pool: the pool to initialize |
| 1068 | * @dev: device for DMA allocations and mappings |
| 1069 | * @nid: NUMA node to use for allocations |
| 1070 | * @alloc_flags: TTM_ALLOCATION_POOL_* flags |
| 1071 | * |
| 1072 | * Initialize the pool and its pool types. |
| 1073 | */ |
| 1074 | void ttm_pool_init(struct ttm_pool *pool, struct device *dev, |
| 1075 | int nid, unsigned int alloc_flags) |
| 1076 | { |
| 1077 | unsigned int i, j; |
| 1078 | |
| 1079 | WARN_ON(!dev && ttm_pool_uses_dma_alloc(pool)); |
| 1080 | |
| 1081 | pool->dev = dev; |
| 1082 | pool->nid = nid; |
| 1083 | pool->alloc_flags = alloc_flags; |
| 1084 | |
| 1085 | for (i = 0; i < TTM_NUM_CACHING_TYPES; ++i) { |
| 1086 | for (j = 0; j < NR_PAGE_ORDERS; ++j) { |
| 1087 | struct ttm_pool_type *pt; |
| 1088 | |
| 1089 | /* Initialize only pool types which are actually used */ |
| 1090 | pt = ttm_pool_select_type(pool, caching: i, order: j); |
| 1091 | if (pt != &pool->caching[i].orders[j]) |
| 1092 | continue; |
| 1093 | |
| 1094 | ttm_pool_type_init(pt, pool, caching: i, order: j); |
| 1095 | } |
| 1096 | } |
| 1097 | } |
| 1098 | EXPORT_SYMBOL(ttm_pool_init); |
| 1099 | |
| 1100 | /** |
| 1101 | * ttm_pool_synchronize_shrinkers - Wait for all running shrinkers to complete. |
| 1102 | * |
| 1103 | * This is useful to guarantee that all shrinker invocations have seen an |
| 1104 | * update, before freeing memory, similar to rcu. |
| 1105 | */ |
| 1106 | static void ttm_pool_synchronize_shrinkers(void) |
| 1107 | { |
| 1108 | down_write(sem: &pool_shrink_rwsem); |
| 1109 | up_write(sem: &pool_shrink_rwsem); |
| 1110 | } |
| 1111 | |
| 1112 | /** |
| 1113 | * ttm_pool_fini - Cleanup a pool |
| 1114 | * |
| 1115 | * @pool: the pool to clean up |
| 1116 | * |
| 1117 | * Free all pages in the pool and unregister the types from the global |
| 1118 | * shrinker. |
| 1119 | */ |
| 1120 | void ttm_pool_fini(struct ttm_pool *pool) |
| 1121 | { |
| 1122 | unsigned int i, j; |
| 1123 | |
| 1124 | for (i = 0; i < TTM_NUM_CACHING_TYPES; ++i) { |
| 1125 | for (j = 0; j < NR_PAGE_ORDERS; ++j) { |
| 1126 | struct ttm_pool_type *pt; |
| 1127 | |
| 1128 | pt = ttm_pool_select_type(pool, caching: i, order: j); |
| 1129 | if (pt != &pool->caching[i].orders[j]) |
| 1130 | continue; |
| 1131 | |
| 1132 | ttm_pool_type_fini(pt); |
| 1133 | } |
| 1134 | } |
| 1135 | |
| 1136 | /* We removed the pool types from the LRU, but we need to also make sure |
| 1137 | * that no shrinker is concurrently freeing pages from the pool. |
| 1138 | */ |
| 1139 | ttm_pool_synchronize_shrinkers(); |
| 1140 | } |
| 1141 | EXPORT_SYMBOL(ttm_pool_fini); |
| 1142 | |
| 1143 | /* Free average pool number of pages. */ |
| 1144 | #define TTM_SHRINKER_BATCH ((1 << (MAX_PAGE_ORDER / 2)) * NR_PAGE_ORDERS) |
| 1145 | |
| 1146 | static unsigned long ttm_pool_shrinker_scan(struct shrinker *shrink, |
| 1147 | struct shrink_control *sc) |
| 1148 | { |
| 1149 | unsigned long num_freed = 0; |
| 1150 | |
| 1151 | do |
| 1152 | num_freed += ttm_pool_shrink(); |
| 1153 | while (num_freed < sc->nr_to_scan && |
| 1154 | atomic_long_read(v: &allocated_pages)); |
| 1155 | |
| 1156 | sc->nr_scanned = num_freed; |
| 1157 | |
| 1158 | return num_freed ?: SHRINK_STOP; |
| 1159 | } |
| 1160 | |
| 1161 | /* Return the number of pages available or SHRINK_EMPTY if we have none */ |
| 1162 | static unsigned long ttm_pool_shrinker_count(struct shrinker *shrink, |
| 1163 | struct shrink_control *sc) |
| 1164 | { |
| 1165 | unsigned long num_pages = atomic_long_read(v: &allocated_pages); |
| 1166 | |
| 1167 | return num_pages ? num_pages : SHRINK_EMPTY; |
| 1168 | } |
| 1169 | |
| 1170 | #ifdef CONFIG_DEBUG_FS |
| 1171 | /* Count the number of pages available in a pool_type */ |
| 1172 | static unsigned int ttm_pool_type_count(struct ttm_pool_type *pt) |
| 1173 | { |
| 1174 | unsigned int count = 0; |
| 1175 | struct page *p; |
| 1176 | |
| 1177 | spin_lock(lock: &pt->lock); |
| 1178 | /* Only used for debugfs, the overhead doesn't matter */ |
| 1179 | list_for_each_entry(p, &pt->pages, lru) |
| 1180 | ++count; |
| 1181 | spin_unlock(lock: &pt->lock); |
| 1182 | |
| 1183 | return count; |
| 1184 | } |
| 1185 | |
| 1186 | /* Print a nice header for the order */ |
| 1187 | static void (struct seq_file *m) |
| 1188 | { |
| 1189 | unsigned int i; |
| 1190 | |
| 1191 | seq_puts(m, s: "\t " ); |
| 1192 | for (i = 0; i < NR_PAGE_ORDERS; ++i) |
| 1193 | seq_printf(m, fmt: " ---%2u---" , i); |
| 1194 | seq_puts(m, s: "\n" ); |
| 1195 | } |
| 1196 | |
| 1197 | /* Dump information about the different pool types */ |
| 1198 | static void ttm_pool_debugfs_orders(struct ttm_pool_type *pt, |
| 1199 | struct seq_file *m) |
| 1200 | { |
| 1201 | unsigned int i; |
| 1202 | |
| 1203 | for (i = 0; i < NR_PAGE_ORDERS; ++i) |
| 1204 | seq_printf(m, fmt: " %8u" , ttm_pool_type_count(pt: &pt[i])); |
| 1205 | seq_puts(m, s: "\n" ); |
| 1206 | } |
| 1207 | |
| 1208 | /* Dump the total amount of allocated pages */ |
| 1209 | static void (struct seq_file *m) |
| 1210 | { |
| 1211 | seq_printf(m, fmt: "\ntotal\t: %8lu of %8lu\n" , |
| 1212 | atomic_long_read(v: &allocated_pages), page_pool_size); |
| 1213 | } |
| 1214 | |
| 1215 | /* Dump the information for the global pools */ |
| 1216 | static int ttm_pool_debugfs_globals_show(struct seq_file *m, void *data) |
| 1217 | { |
| 1218 | ttm_pool_debugfs_header(m); |
| 1219 | |
| 1220 | spin_lock(lock: &shrinker_lock); |
| 1221 | seq_puts(m, s: "wc\t:" ); |
| 1222 | ttm_pool_debugfs_orders(pt: global_write_combined, m); |
| 1223 | seq_puts(m, s: "uc\t:" ); |
| 1224 | ttm_pool_debugfs_orders(pt: global_uncached, m); |
| 1225 | seq_puts(m, s: "wc 32\t:" ); |
| 1226 | ttm_pool_debugfs_orders(pt: global_dma32_write_combined, m); |
| 1227 | seq_puts(m, s: "uc 32\t:" ); |
| 1228 | ttm_pool_debugfs_orders(pt: global_dma32_uncached, m); |
| 1229 | spin_unlock(lock: &shrinker_lock); |
| 1230 | |
| 1231 | ttm_pool_debugfs_footer(m); |
| 1232 | |
| 1233 | return 0; |
| 1234 | } |
| 1235 | DEFINE_SHOW_ATTRIBUTE(ttm_pool_debugfs_globals); |
| 1236 | |
| 1237 | /** |
| 1238 | * ttm_pool_debugfs - Debugfs dump function for a pool |
| 1239 | * |
| 1240 | * @pool: the pool to dump the information for |
| 1241 | * @m: seq_file to dump to |
| 1242 | * |
| 1243 | * Make a debugfs dump with the per pool and global information. |
| 1244 | */ |
| 1245 | int ttm_pool_debugfs(struct ttm_pool *pool, struct seq_file *m) |
| 1246 | { |
| 1247 | unsigned int i; |
| 1248 | |
| 1249 | if (!ttm_pool_uses_dma_alloc(pool) && pool->nid == NUMA_NO_NODE) { |
| 1250 | seq_puts(m, s: "unused\n" ); |
| 1251 | return 0; |
| 1252 | } |
| 1253 | |
| 1254 | ttm_pool_debugfs_header(m); |
| 1255 | |
| 1256 | spin_lock(lock: &shrinker_lock); |
| 1257 | for (i = 0; i < TTM_NUM_CACHING_TYPES; ++i) { |
| 1258 | if (!ttm_pool_select_type(pool, caching: i, order: 0)) |
| 1259 | continue; |
| 1260 | if (ttm_pool_uses_dma_alloc(pool)) |
| 1261 | seq_puts(m, s: "DMA " ); |
| 1262 | else |
| 1263 | seq_printf(m, fmt: "N%d " , pool->nid); |
| 1264 | switch (i) { |
| 1265 | case ttm_cached: |
| 1266 | seq_puts(m, s: "\t:" ); |
| 1267 | break; |
| 1268 | case ttm_write_combined: |
| 1269 | seq_puts(m, s: "wc\t:" ); |
| 1270 | break; |
| 1271 | case ttm_uncached: |
| 1272 | seq_puts(m, s: "uc\t:" ); |
| 1273 | break; |
| 1274 | } |
| 1275 | ttm_pool_debugfs_orders(pt: pool->caching[i].orders, m); |
| 1276 | } |
| 1277 | spin_unlock(lock: &shrinker_lock); |
| 1278 | |
| 1279 | ttm_pool_debugfs_footer(m); |
| 1280 | return 0; |
| 1281 | } |
| 1282 | EXPORT_SYMBOL(ttm_pool_debugfs); |
| 1283 | |
| 1284 | /* Test the shrinker functions and dump the result */ |
| 1285 | static int ttm_pool_debugfs_shrink_show(struct seq_file *m, void *data) |
| 1286 | { |
| 1287 | struct shrink_control sc = { |
| 1288 | .gfp_mask = GFP_NOFS, |
| 1289 | .nr_to_scan = TTM_SHRINKER_BATCH, |
| 1290 | }; |
| 1291 | unsigned long count; |
| 1292 | |
| 1293 | fs_reclaim_acquire(GFP_KERNEL); |
| 1294 | count = ttm_pool_shrinker_count(shrink: mm_shrinker, sc: &sc); |
| 1295 | seq_printf(m, fmt: "%lu/%lu\n" , count, |
| 1296 | ttm_pool_shrinker_scan(shrink: mm_shrinker, sc: &sc)); |
| 1297 | fs_reclaim_release(GFP_KERNEL); |
| 1298 | |
| 1299 | return 0; |
| 1300 | } |
| 1301 | DEFINE_SHOW_ATTRIBUTE(ttm_pool_debugfs_shrink); |
| 1302 | |
| 1303 | #endif |
| 1304 | |
| 1305 | /** |
| 1306 | * ttm_pool_mgr_init - Initialize globals |
| 1307 | * |
| 1308 | * @num_pages: default number of pages |
| 1309 | * |
| 1310 | * Initialize the global locks and lists for the MM shrinker. |
| 1311 | */ |
| 1312 | int ttm_pool_mgr_init(unsigned long num_pages) |
| 1313 | { |
| 1314 | unsigned int i; |
| 1315 | |
| 1316 | if (!page_pool_size) |
| 1317 | page_pool_size = num_pages; |
| 1318 | |
| 1319 | spin_lock_init(&shrinker_lock); |
| 1320 | INIT_LIST_HEAD(list: &shrinker_list); |
| 1321 | |
| 1322 | for (i = 0; i < NR_PAGE_ORDERS; ++i) { |
| 1323 | ttm_pool_type_init(pt: &global_write_combined[i], NULL, |
| 1324 | caching: ttm_write_combined, order: i); |
| 1325 | ttm_pool_type_init(pt: &global_uncached[i], NULL, caching: ttm_uncached, order: i); |
| 1326 | |
| 1327 | ttm_pool_type_init(pt: &global_dma32_write_combined[i], NULL, |
| 1328 | caching: ttm_write_combined, order: i); |
| 1329 | ttm_pool_type_init(pt: &global_dma32_uncached[i], NULL, |
| 1330 | caching: ttm_uncached, order: i); |
| 1331 | } |
| 1332 | |
| 1333 | #ifdef CONFIG_DEBUG_FS |
| 1334 | debugfs_create_file("page_pool" , 0444, ttm_debugfs_root, NULL, |
| 1335 | &ttm_pool_debugfs_globals_fops); |
| 1336 | debugfs_create_file("page_pool_shrink" , 0400, ttm_debugfs_root, NULL, |
| 1337 | &ttm_pool_debugfs_shrink_fops); |
| 1338 | #ifdef CONFIG_FAULT_INJECTION |
| 1339 | fault_create_debugfs_attr(name: "backup_fault_inject" , parent: ttm_debugfs_root, |
| 1340 | attr: &backup_fault_inject); |
| 1341 | #endif |
| 1342 | #endif |
| 1343 | |
| 1344 | mm_shrinker = shrinker_alloc(flags: 0, fmt: "drm-ttm_pool" ); |
| 1345 | if (!mm_shrinker) |
| 1346 | return -ENOMEM; |
| 1347 | |
| 1348 | mm_shrinker->count_objects = ttm_pool_shrinker_count; |
| 1349 | mm_shrinker->scan_objects = ttm_pool_shrinker_scan; |
| 1350 | mm_shrinker->batch = TTM_SHRINKER_BATCH; |
| 1351 | mm_shrinker->seeks = 1; |
| 1352 | |
| 1353 | shrinker_register(shrinker: mm_shrinker); |
| 1354 | |
| 1355 | return 0; |
| 1356 | } |
| 1357 | |
| 1358 | /** |
| 1359 | * ttm_pool_mgr_fini - Finalize globals |
| 1360 | * |
| 1361 | * Cleanup the global pools and unregister the MM shrinker. |
| 1362 | */ |
| 1363 | void ttm_pool_mgr_fini(void) |
| 1364 | { |
| 1365 | unsigned int i; |
| 1366 | |
| 1367 | for (i = 0; i < NR_PAGE_ORDERS; ++i) { |
| 1368 | ttm_pool_type_fini(pt: &global_write_combined[i]); |
| 1369 | ttm_pool_type_fini(pt: &global_uncached[i]); |
| 1370 | |
| 1371 | ttm_pool_type_fini(pt: &global_dma32_write_combined[i]); |
| 1372 | ttm_pool_type_fini(pt: &global_dma32_uncached[i]); |
| 1373 | } |
| 1374 | |
| 1375 | shrinker_free(shrinker: mm_shrinker); |
| 1376 | WARN_ON(!list_empty(&shrinker_list)); |
| 1377 | } |
| 1378 | |