1/* 2 * Copyright 2011 (c) Oracle Corp. 3 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sub license, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice (including the 12 * next paragraph) shall be included in all copies or substantial portions 13 * 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 NON-INFRINGEMENT. IN NO EVENT SHALL 18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING 20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER 21 * DEALINGS IN THE SOFTWARE. 22 * 23 * Author: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> 24 */ 25 26/* 27 * A simple DMA pool losely based on dmapool.c. It has certain advantages 28 * over the DMA pools: 29 * - Pool collects resently freed pages for reuse (and hooks up to 30 * the shrinker). 31 * - Tracks currently in use pages 32 * - Tracks whether the page is UC, WB or cached (and reverts to WB 33 * when freed). 34 */ 35 36#if defined(CONFIG_SWIOTLB) || defined(CONFIG_INTEL_IOMMU) 37#define pr_fmt(fmt) "[TTM] " fmt 38 39#include <linux/dma-mapping.h> 40#include <linux/list.h> 41#include <linux/seq_file.h> /* for seq_printf */ 42#include <linux/slab.h> 43#include <linux/spinlock.h> 44#include <linux/highmem.h> 45#include <linux/mm_types.h> 46#include <linux/module.h> 47#include <linux/mm.h> 48#include <linux/atomic.h> 49#include <linux/device.h> 50#include <linux/kthread.h> 51#include <drm/ttm/ttm_bo_driver.h> 52#include <drm/ttm/ttm_page_alloc.h> 53#ifdef TTM_HAS_AGP 54#include <asm/agp.h> 55#endif 56 57#define NUM_PAGES_TO_ALLOC (PAGE_SIZE/sizeof(struct page *)) 58#define SMALL_ALLOCATION 4 59#define FREE_ALL_PAGES (~0U) 60/* times are in msecs */ 61#define IS_UNDEFINED (0) 62#define IS_WC (1<<1) 63#define IS_UC (1<<2) 64#define IS_CACHED (1<<3) 65#define IS_DMA32 (1<<4) 66 67enum pool_type { 68 POOL_IS_UNDEFINED, 69 POOL_IS_WC = IS_WC, 70 POOL_IS_UC = IS_UC, 71 POOL_IS_CACHED = IS_CACHED, 72 POOL_IS_WC_DMA32 = IS_WC | IS_DMA32, 73 POOL_IS_UC_DMA32 = IS_UC | IS_DMA32, 74 POOL_IS_CACHED_DMA32 = IS_CACHED | IS_DMA32, 75}; 76/* 77 * The pool structure. There are usually six pools: 78 * - generic (not restricted to DMA32): 79 * - write combined, uncached, cached. 80 * - dma32 (up to 2^32 - so up 4GB): 81 * - write combined, uncached, cached. 82 * for each 'struct device'. The 'cached' is for pages that are actively used. 83 * The other ones can be shrunk by the shrinker API if neccessary. 84 * @pools: The 'struct device->dma_pools' link. 85 * @type: Type of the pool 86 * @lock: Protects the inuse_list and free_list from concurrnet access. Must be 87 * used with irqsave/irqrestore variants because pool allocator maybe called 88 * from delayed work. 89 * @inuse_list: Pool of pages that are in use. The order is very important and 90 * it is in the order that the TTM pages that are put back are in. 91 * @free_list: Pool of pages that are free to be used. No order requirements. 92 * @dev: The device that is associated with these pools. 93 * @size: Size used during DMA allocation. 94 * @npages_free: Count of available pages for re-use. 95 * @npages_in_use: Count of pages that are in use. 96 * @nfrees: Stats when pool is shrinking. 97 * @nrefills: Stats when the pool is grown. 98 * @gfp_flags: Flags to pass for alloc_page. 99 * @name: Name of the pool. 100 * @dev_name: Name derieved from dev - similar to how dev_info works. 101 * Used during shutdown as the dev_info during release is unavailable. 102 */ 103struct dma_pool { 104 struct list_head pools; /* The 'struct device->dma_pools link */ 105 enum pool_type type; 106 spinlock_t lock; 107 struct list_head inuse_list; 108 struct list_head free_list; 109 struct device *dev; 110 unsigned size; 111 unsigned npages_free; 112 unsigned npages_in_use; 113 unsigned long nfrees; /* Stats when shrunk. */ 114 unsigned long nrefills; /* Stats when grown. */ 115 gfp_t gfp_flags; 116 char name[13]; /* "cached dma32" */ 117 char dev_name[64]; /* Constructed from dev */ 118}; 119 120/* 121 * The accounting page keeping track of the allocated page along with 122 * the DMA address. 123 * @page_list: The link to the 'page_list' in 'struct dma_pool'. 124 * @vaddr: The virtual address of the page 125 * @dma: The bus address of the page. If the page is not allocated 126 * via the DMA API, it will be -1. 127 */ 128struct dma_page { 129 struct list_head page_list; 130 void *vaddr; 131 struct page *p; 132 dma_addr_t dma; 133}; 134 135/* 136 * Limits for the pool. They are handled without locks because only place where 137 * they may change is in sysfs store. They won't have immediate effect anyway 138 * so forcing serialization to access them is pointless. 139 */ 140 141struct ttm_pool_opts { 142 unsigned alloc_size; 143 unsigned max_size; 144 unsigned small; 145}; 146 147/* 148 * Contains the list of all of the 'struct device' and their corresponding 149 * DMA pools. Guarded by _mutex->lock. 150 * @pools: The link to 'struct ttm_pool_manager->pools' 151 * @dev: The 'struct device' associated with the 'pool' 152 * @pool: The 'struct dma_pool' associated with the 'dev' 153 */ 154struct device_pools { 155 struct list_head pools; 156 struct device *dev; 157 struct dma_pool *pool; 158}; 159 160/* 161 * struct ttm_pool_manager - Holds memory pools for fast allocation 162 * 163 * @lock: Lock used when adding/removing from pools 164 * @pools: List of 'struct device' and 'struct dma_pool' tuples. 165 * @options: Limits for the pool. 166 * @npools: Total amount of pools in existence. 167 * @shrinker: The structure used by [un|]register_shrinker 168 */ 169struct ttm_pool_manager { 170 struct mutex lock; 171 struct list_head pools; 172 struct ttm_pool_opts options; 173 unsigned npools; 174 struct shrinker mm_shrink; 175 struct kobject kobj; 176}; 177 178static struct ttm_pool_manager *_manager; 179 180static struct attribute ttm_page_pool_max = { 181 .name = "pool_max_size", 182 .mode = S_IRUGO | S_IWUSR 183}; 184static struct attribute ttm_page_pool_small = { 185 .name = "pool_small_allocation", 186 .mode = S_IRUGO | S_IWUSR 187}; 188static struct attribute ttm_page_pool_alloc_size = { 189 .name = "pool_allocation_size", 190 .mode = S_IRUGO | S_IWUSR 191}; 192 193static struct attribute *ttm_pool_attrs[] = { 194 &ttm_page_pool_max, 195 &ttm_page_pool_small, 196 &ttm_page_pool_alloc_size, 197 NULL 198}; 199 200static void ttm_pool_kobj_release(struct kobject *kobj) 201{ 202 struct ttm_pool_manager *m = 203 container_of(kobj, struct ttm_pool_manager, kobj); 204 kfree(m); 205} 206 207static ssize_t ttm_pool_store(struct kobject *kobj, struct attribute *attr, 208 const char *buffer, size_t size) 209{ 210 struct ttm_pool_manager *m = 211 container_of(kobj, struct ttm_pool_manager, kobj); 212 int chars; 213 unsigned val; 214 chars = sscanf(buffer, "%u", &val); 215 if (chars == 0) 216 return size; 217 218 /* Convert kb to number of pages */ 219 val = val / (PAGE_SIZE >> 10); 220 221 if (attr == &ttm_page_pool_max) 222 m->options.max_size = val; 223 else if (attr == &ttm_page_pool_small) 224 m->options.small = val; 225 else if (attr == &ttm_page_pool_alloc_size) { 226 if (val > NUM_PAGES_TO_ALLOC*8) { 227 pr_err("Setting allocation size to %lu is not allowed. Recommended size is %lu\n", 228 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 7), 229 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10)); 230 return size; 231 } else if (val > NUM_PAGES_TO_ALLOC) { 232 pr_warn("Setting allocation size to larger than %lu is not recommended\n", 233 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10)); 234 } 235 m->options.alloc_size = val; 236 } 237 238 return size; 239} 240 241static ssize_t ttm_pool_show(struct kobject *kobj, struct attribute *attr, 242 char *buffer) 243{ 244 struct ttm_pool_manager *m = 245 container_of(kobj, struct ttm_pool_manager, kobj); 246 unsigned val = 0; 247 248 if (attr == &ttm_page_pool_max) 249 val = m->options.max_size; 250 else if (attr == &ttm_page_pool_small) 251 val = m->options.small; 252 else if (attr == &ttm_page_pool_alloc_size) 253 val = m->options.alloc_size; 254 255 val = val * (PAGE_SIZE >> 10); 256 257 return snprintf(buffer, PAGE_SIZE, "%u\n", val); 258} 259 260static const struct sysfs_ops ttm_pool_sysfs_ops = { 261 .show = &ttm_pool_show, 262 .store = &ttm_pool_store, 263}; 264 265static struct kobj_type ttm_pool_kobj_type = { 266 .release = &ttm_pool_kobj_release, 267 .sysfs_ops = &ttm_pool_sysfs_ops, 268 .default_attrs = ttm_pool_attrs, 269}; 270 271#ifndef CONFIG_X86 272static int set_pages_array_wb(struct page **pages, int addrinarray) 273{ 274#ifdef TTM_HAS_AGP 275 int i; 276 277 for (i = 0; i < addrinarray; i++) 278 unmap_page_from_agp(pages[i]); 279#endif 280 return 0; 281} 282 283static int set_pages_array_wc(struct page **pages, int addrinarray) 284{ 285#ifdef TTM_HAS_AGP 286 int i; 287 288 for (i = 0; i < addrinarray; i++) 289 map_page_into_agp(pages[i]); 290#endif 291 return 0; 292} 293 294static int set_pages_array_uc(struct page **pages, int addrinarray) 295{ 296#ifdef TTM_HAS_AGP 297 int i; 298 299 for (i = 0; i < addrinarray; i++) 300 map_page_into_agp(pages[i]); 301#endif 302 return 0; 303} 304#endif /* for !CONFIG_X86 */ 305 306static int ttm_set_pages_caching(struct dma_pool *pool, 307 struct page **pages, unsigned cpages) 308{ 309 int r = 0; 310 /* Set page caching */ 311 if (pool->type & IS_UC) { 312 r = set_pages_array_uc(pages, cpages); 313 if (r) 314 pr_err("%s: Failed to set %d pages to uc!\n", 315 pool->dev_name, cpages); 316 } 317 if (pool->type & IS_WC) { 318 r = set_pages_array_wc(pages, cpages); 319 if (r) 320 pr_err("%s: Failed to set %d pages to wc!\n", 321 pool->dev_name, cpages); 322 } 323 return r; 324} 325 326static void __ttm_dma_free_page(struct dma_pool *pool, struct dma_page *d_page) 327{ 328 dma_addr_t dma = d_page->dma; 329 dma_free_coherent(pool->dev, pool->size, d_page->vaddr, dma); 330 331 kfree(d_page); 332 d_page = NULL; 333} 334static struct dma_page *__ttm_dma_alloc_page(struct dma_pool *pool) 335{ 336 struct dma_page *d_page; 337 338 d_page = kmalloc(sizeof(struct dma_page), GFP_KERNEL); 339 if (!d_page) 340 return NULL; 341 342 d_page->vaddr = dma_alloc_coherent(pool->dev, pool->size, 343 &d_page->dma, 344 pool->gfp_flags); 345 if (d_page->vaddr) 346 d_page->p = virt_to_page(d_page->vaddr); 347 else { 348 kfree(d_page); 349 d_page = NULL; 350 } 351 return d_page; 352} 353static enum pool_type ttm_to_type(int flags, enum ttm_caching_state cstate) 354{ 355 enum pool_type type = IS_UNDEFINED; 356 357 if (flags & TTM_PAGE_FLAG_DMA32) 358 type |= IS_DMA32; 359 if (cstate == tt_cached) 360 type |= IS_CACHED; 361 else if (cstate == tt_uncached) 362 type |= IS_UC; 363 else 364 type |= IS_WC; 365 366 return type; 367} 368 369static void ttm_pool_update_free_locked(struct dma_pool *pool, 370 unsigned freed_pages) 371{ 372 pool->npages_free -= freed_pages; 373 pool->nfrees += freed_pages; 374 375} 376 377/* set memory back to wb and free the pages. */ 378static void ttm_dma_pages_put(struct dma_pool *pool, struct list_head *d_pages, 379 struct page *pages[], unsigned npages) 380{ 381 struct dma_page *d_page, *tmp; 382 383 /* Don't set WB on WB page pool. */ 384 if (npages && !(pool->type & IS_CACHED) && 385 set_pages_array_wb(pages, npages)) 386 pr_err("%s: Failed to set %d pages to wb!\n", 387 pool->dev_name, npages); 388 389 list_for_each_entry_safe(d_page, tmp, d_pages, page_list) { 390 list_del(&d_page->page_list); 391 __ttm_dma_free_page(pool, d_page); 392 } 393} 394 395static void ttm_dma_page_put(struct dma_pool *pool, struct dma_page *d_page) 396{ 397 /* Don't set WB on WB page pool. */ 398 if (!(pool->type & IS_CACHED) && set_pages_array_wb(&d_page->p, 1)) 399 pr_err("%s: Failed to set %d pages to wb!\n", 400 pool->dev_name, 1); 401 402 list_del(&d_page->page_list); 403 __ttm_dma_free_page(pool, d_page); 404} 405 406/* 407 * Free pages from pool. 408 * 409 * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC 410 * number of pages in one go. 411 * 412 * @pool: to free the pages from 413 * @nr_free: If set to true will free all pages in pool 414 * @use_static: Safe to use static buffer 415 **/ 416static unsigned ttm_dma_page_pool_free(struct dma_pool *pool, unsigned nr_free, 417 bool use_static) 418{ 419 static struct page *static_buf[NUM_PAGES_TO_ALLOC]; 420 unsigned long irq_flags; 421 struct dma_page *dma_p, *tmp; 422 struct page **pages_to_free; 423 struct list_head d_pages; 424 unsigned freed_pages = 0, 425 npages_to_free = nr_free; 426 427 if (NUM_PAGES_TO_ALLOC < nr_free) 428 npages_to_free = NUM_PAGES_TO_ALLOC; 429#if 0 430 if (nr_free > 1) { 431 pr_debug("%s: (%s:%d) Attempting to free %d (%d) pages\n", 432 pool->dev_name, pool->name, current->pid, 433 npages_to_free, nr_free); 434 } 435#endif 436 if (use_static) 437 pages_to_free = static_buf; 438 else 439 pages_to_free = kmalloc(npages_to_free * sizeof(struct page *), 440 GFP_KERNEL); 441 442 if (!pages_to_free) { 443 pr_err("%s: Failed to allocate memory for pool free operation\n", 444 pool->dev_name); 445 return 0; 446 } 447 INIT_LIST_HEAD(&d_pages); 448restart: 449 spin_lock_irqsave(&pool->lock, irq_flags); 450 451 /* We picking the oldest ones off the list */ 452 list_for_each_entry_safe_reverse(dma_p, tmp, &pool->free_list, 453 page_list) { 454 if (freed_pages >= npages_to_free) 455 break; 456 457 /* Move the dma_page from one list to another. */ 458 list_move(&dma_p->page_list, &d_pages); 459 460 pages_to_free[freed_pages++] = dma_p->p; 461 /* We can only remove NUM_PAGES_TO_ALLOC at a time. */ 462 if (freed_pages >= NUM_PAGES_TO_ALLOC) { 463 464 ttm_pool_update_free_locked(pool, freed_pages); 465 /** 466 * Because changing page caching is costly 467 * we unlock the pool to prevent stalling. 468 */ 469 spin_unlock_irqrestore(&pool->lock, irq_flags); 470 471 ttm_dma_pages_put(pool, &d_pages, pages_to_free, 472 freed_pages); 473 474 INIT_LIST_HEAD(&d_pages); 475 476 if (likely(nr_free != FREE_ALL_PAGES)) 477 nr_free -= freed_pages; 478 479 if (NUM_PAGES_TO_ALLOC >= nr_free) 480 npages_to_free = nr_free; 481 else 482 npages_to_free = NUM_PAGES_TO_ALLOC; 483 484 freed_pages = 0; 485 486 /* free all so restart the processing */ 487 if (nr_free) 488 goto restart; 489 490 /* Not allowed to fall through or break because 491 * following context is inside spinlock while we are 492 * outside here. 493 */ 494 goto out; 495 496 } 497 } 498 499 /* remove range of pages from the pool */ 500 if (freed_pages) { 501 ttm_pool_update_free_locked(pool, freed_pages); 502 nr_free -= freed_pages; 503 } 504 505 spin_unlock_irqrestore(&pool->lock, irq_flags); 506 507 if (freed_pages) 508 ttm_dma_pages_put(pool, &d_pages, pages_to_free, freed_pages); 509out: 510 if (pages_to_free != static_buf) 511 kfree(pages_to_free); 512 return nr_free; 513} 514 515static void ttm_dma_free_pool(struct device *dev, enum pool_type type) 516{ 517 struct device_pools *p; 518 struct dma_pool *pool; 519 520 if (!dev) 521 return; 522 523 mutex_lock(&_manager->lock); 524 list_for_each_entry_reverse(p, &_manager->pools, pools) { 525 if (p->dev != dev) 526 continue; 527 pool = p->pool; 528 if (pool->type != type) 529 continue; 530 531 list_del(&p->pools); 532 kfree(p); 533 _manager->npools--; 534 break; 535 } 536 list_for_each_entry_reverse(pool, &dev->dma_pools, pools) { 537 if (pool->type != type) 538 continue; 539 /* Takes a spinlock.. */ 540 /* OK to use static buffer since global mutex is held. */ 541 ttm_dma_page_pool_free(pool, FREE_ALL_PAGES, true); 542 WARN_ON(((pool->npages_in_use + pool->npages_free) != 0)); 543 /* This code path is called after _all_ references to the 544 * struct device has been dropped - so nobody should be 545 * touching it. In case somebody is trying to _add_ we are 546 * guarded by the mutex. */ 547 list_del(&pool->pools); 548 kfree(pool); 549 break; 550 } 551 mutex_unlock(&_manager->lock); 552} 553 554/* 555 * On free-ing of the 'struct device' this deconstructor is run. 556 * Albeit the pool might have already been freed earlier. 557 */ 558static void ttm_dma_pool_release(struct device *dev, void *res) 559{ 560 struct dma_pool *pool = *(struct dma_pool **)res; 561 562 if (pool) 563 ttm_dma_free_pool(dev, pool->type); 564} 565 566static int ttm_dma_pool_match(struct device *dev, void *res, void *match_data) 567{ 568 return *(struct dma_pool **)res == match_data; 569} 570 571static struct dma_pool *ttm_dma_pool_init(struct device *dev, gfp_t flags, 572 enum pool_type type) 573{ 574 char *n[] = {"wc", "uc", "cached", " dma32", "unknown",}; 575 enum pool_type t[] = {IS_WC, IS_UC, IS_CACHED, IS_DMA32, IS_UNDEFINED}; 576 struct device_pools *sec_pool = NULL; 577 struct dma_pool *pool = NULL, **ptr; 578 unsigned i; 579 int ret = -ENODEV; 580 char *p; 581 582 if (!dev) 583 return NULL; 584 585 ptr = devres_alloc(ttm_dma_pool_release, sizeof(*ptr), GFP_KERNEL); 586 if (!ptr) 587 return NULL; 588 589 ret = -ENOMEM; 590 591 pool = kmalloc_node(sizeof(struct dma_pool), GFP_KERNEL, 592 dev_to_node(dev)); 593 if (!pool) 594 goto err_mem; 595 596 sec_pool = kmalloc_node(sizeof(struct device_pools), GFP_KERNEL, 597 dev_to_node(dev)); 598 if (!sec_pool) 599 goto err_mem; 600 601 INIT_LIST_HEAD(&sec_pool->pools); 602 sec_pool->dev = dev; 603 sec_pool->pool = pool; 604 605 INIT_LIST_HEAD(&pool->free_list); 606 INIT_LIST_HEAD(&pool->inuse_list); 607 INIT_LIST_HEAD(&pool->pools); 608 spin_lock_init(&pool->lock); 609 pool->dev = dev; 610 pool->npages_free = pool->npages_in_use = 0; 611 pool->nfrees = 0; 612 pool->gfp_flags = flags; 613 pool->size = PAGE_SIZE; 614 pool->type = type; 615 pool->nrefills = 0; 616 p = pool->name; 617 for (i = 0; i < 5; i++) { 618 if (type & t[i]) { 619 p += snprintf(p, sizeof(pool->name) - (p - pool->name), 620 "%s", n[i]); 621 } 622 } 623 *p = 0; 624 /* We copy the name for pr_ calls b/c when dma_pool_destroy is called 625 * - the kobj->name has already been deallocated.*/ 626 snprintf(pool->dev_name, sizeof(pool->dev_name), "%s %s", 627 dev_driver_string(dev), dev_name(dev)); 628 mutex_lock(&_manager->lock); 629 /* You can get the dma_pool from either the global: */ 630 list_add(&sec_pool->pools, &_manager->pools); 631 _manager->npools++; 632 /* or from 'struct device': */ 633 list_add(&pool->pools, &dev->dma_pools); 634 mutex_unlock(&_manager->lock); 635 636 *ptr = pool; 637 devres_add(dev, ptr); 638 639 return pool; 640err_mem: 641 devres_free(ptr); 642 kfree(sec_pool); 643 kfree(pool); 644 return ERR_PTR(ret); 645} 646 647static struct dma_pool *ttm_dma_find_pool(struct device *dev, 648 enum pool_type type) 649{ 650 struct dma_pool *pool, *tmp, *found = NULL; 651 652 if (type == IS_UNDEFINED) 653 return found; 654 655 /* NB: We iterate on the 'struct dev' which has no spinlock, but 656 * it does have a kref which we have taken. The kref is taken during 657 * graphic driver loading - in the drm_pci_init it calls either 658 * pci_dev_get or pci_register_driver which both end up taking a kref 659 * on 'struct device'. 660 * 661 * On teardown, the graphic drivers end up quiescing the TTM (put_pages) 662 * and calls the dev_res deconstructors: ttm_dma_pool_release. The nice 663 * thing is at that point of time there are no pages associated with the 664 * driver so this function will not be called. 665 */ 666 list_for_each_entry_safe(pool, tmp, &dev->dma_pools, pools) { 667 if (pool->type != type) 668 continue; 669 found = pool; 670 break; 671 } 672 return found; 673} 674 675/* 676 * Free pages the pages that failed to change the caching state. If there 677 * are pages that have changed their caching state already put them to the 678 * pool. 679 */ 680static void ttm_dma_handle_caching_state_failure(struct dma_pool *pool, 681 struct list_head *d_pages, 682 struct page **failed_pages, 683 unsigned cpages) 684{ 685 struct dma_page *d_page, *tmp; 686 struct page *p; 687 unsigned i = 0; 688 689 p = failed_pages[0]; 690 if (!p) 691 return; 692 /* Find the failed page. */ 693 list_for_each_entry_safe(d_page, tmp, d_pages, page_list) { 694 if (d_page->p != p) 695 continue; 696 /* .. and then progress over the full list. */ 697 list_del(&d_page->page_list); 698 __ttm_dma_free_page(pool, d_page); 699 if (++i < cpages) 700 p = failed_pages[i]; 701 else 702 break; 703 } 704 705} 706 707/* 708 * Allocate 'count' pages, and put 'need' number of them on the 709 * 'pages' and as well on the 'dma_address' starting at 'dma_offset' offset. 710 * The full list of pages should also be on 'd_pages'. 711 * We return zero for success, and negative numbers as errors. 712 */ 713static int ttm_dma_pool_alloc_new_pages(struct dma_pool *pool, 714 struct list_head *d_pages, 715 unsigned count) 716{ 717 struct page **caching_array; 718 struct dma_page *dma_p; 719 struct page *p; 720 int r = 0; 721 unsigned i, cpages; 722 unsigned max_cpages = min(count, 723 (unsigned)(PAGE_SIZE/sizeof(struct page *))); 724 725 /* allocate array for page caching change */ 726 caching_array = kmalloc(max_cpages*sizeof(struct page *), GFP_KERNEL); 727 728 if (!caching_array) { 729 pr_err("%s: Unable to allocate table for new pages\n", 730 pool->dev_name); 731 return -ENOMEM; 732 } 733 734 if (count > 1) { 735 pr_debug("%s: (%s:%d) Getting %d pages\n", 736 pool->dev_name, pool->name, current->pid, count); 737 } 738 739 for (i = 0, cpages = 0; i < count; ++i) { 740 dma_p = __ttm_dma_alloc_page(pool); 741 if (!dma_p) { 742 pr_err("%s: Unable to get page %u\n", 743 pool->dev_name, i); 744 745 /* store already allocated pages in the pool after 746 * setting the caching state */ 747 if (cpages) { 748 r = ttm_set_pages_caching(pool, caching_array, 749 cpages); 750 if (r) 751 ttm_dma_handle_caching_state_failure( 752 pool, d_pages, caching_array, 753 cpages); 754 } 755 r = -ENOMEM; 756 goto out; 757 } 758 p = dma_p->p; 759#ifdef CONFIG_HIGHMEM 760 /* gfp flags of highmem page should never be dma32 so we 761 * we should be fine in such case 762 */ 763 if (!PageHighMem(p)) 764#endif 765 { 766 caching_array[cpages++] = p; 767 if (cpages == max_cpages) { 768 /* Note: Cannot hold the spinlock */ 769 r = ttm_set_pages_caching(pool, caching_array, 770 cpages); 771 if (r) { 772 ttm_dma_handle_caching_state_failure( 773 pool, d_pages, caching_array, 774 cpages); 775 goto out; 776 } 777 cpages = 0; 778 } 779 } 780 list_add(&dma_p->page_list, d_pages); 781 } 782 783 if (cpages) { 784 r = ttm_set_pages_caching(pool, caching_array, cpages); 785 if (r) 786 ttm_dma_handle_caching_state_failure(pool, d_pages, 787 caching_array, cpages); 788 } 789out: 790 kfree(caching_array); 791 return r; 792} 793 794/* 795 * @return count of pages still required to fulfill the request. 796 */ 797static int ttm_dma_page_pool_fill_locked(struct dma_pool *pool, 798 unsigned long *irq_flags) 799{ 800 unsigned count = _manager->options.small; 801 int r = pool->npages_free; 802 803 if (count > pool->npages_free) { 804 struct list_head d_pages; 805 806 INIT_LIST_HEAD(&d_pages); 807 808 spin_unlock_irqrestore(&pool->lock, *irq_flags); 809 810 /* Returns how many more are neccessary to fulfill the 811 * request. */ 812 r = ttm_dma_pool_alloc_new_pages(pool, &d_pages, count); 813 814 spin_lock_irqsave(&pool->lock, *irq_flags); 815 if (!r) { 816 /* Add the fresh to the end.. */ 817 list_splice(&d_pages, &pool->free_list); 818 ++pool->nrefills; 819 pool->npages_free += count; 820 r = count; 821 } else { 822 struct dma_page *d_page; 823 unsigned cpages = 0; 824 825 pr_err("%s: Failed to fill %s pool (r:%d)!\n", 826 pool->dev_name, pool->name, r); 827 828 list_for_each_entry(d_page, &d_pages, page_list) { 829 cpages++; 830 } 831 list_splice_tail(&d_pages, &pool->free_list); 832 pool->npages_free += cpages; 833 r = cpages; 834 } 835 } 836 return r; 837} 838 839/* 840 * @return count of pages still required to fulfill the request. 841 * The populate list is actually a stack (not that is matters as TTM 842 * allocates one page at a time. 843 */ 844static int ttm_dma_pool_get_pages(struct dma_pool *pool, 845 struct ttm_dma_tt *ttm_dma, 846 unsigned index) 847{ 848 struct dma_page *d_page; 849 struct ttm_tt *ttm = &ttm_dma->ttm; 850 unsigned long irq_flags; 851 int count, r = -ENOMEM; 852 853 spin_lock_irqsave(&pool->lock, irq_flags); 854 count = ttm_dma_page_pool_fill_locked(pool, &irq_flags); 855 if (count) { 856 d_page = list_first_entry(&pool->free_list, struct dma_page, page_list); 857 ttm->pages[index] = d_page->p; 858 ttm_dma->cpu_address[index] = d_page->vaddr; 859 ttm_dma->dma_address[index] = d_page->dma; 860 list_move_tail(&d_page->page_list, &ttm_dma->pages_list); 861 r = 0; 862 pool->npages_in_use += 1; 863 pool->npages_free -= 1; 864 } 865 spin_unlock_irqrestore(&pool->lock, irq_flags); 866 return r; 867} 868 869/* 870 * On success pages list will hold count number of correctly 871 * cached pages. On failure will hold the negative return value (-ENOMEM, etc). 872 */ 873int ttm_dma_populate(struct ttm_dma_tt *ttm_dma, struct device *dev) 874{ 875 struct ttm_tt *ttm = &ttm_dma->ttm; 876 struct ttm_mem_global *mem_glob = ttm->glob->mem_glob; 877 struct dma_pool *pool; 878 enum pool_type type; 879 unsigned i; 880 gfp_t gfp_flags; 881 int ret; 882 883 if (ttm->state != tt_unpopulated) 884 return 0; 885 886 type = ttm_to_type(ttm->page_flags, ttm->caching_state); 887 if (ttm->page_flags & TTM_PAGE_FLAG_DMA32) 888 gfp_flags = GFP_USER | GFP_DMA32; 889 else 890 gfp_flags = GFP_HIGHUSER; 891 if (ttm->page_flags & TTM_PAGE_FLAG_ZERO_ALLOC) 892 gfp_flags |= __GFP_ZERO; 893 894 pool = ttm_dma_find_pool(dev, type); 895 if (!pool) { 896 pool = ttm_dma_pool_init(dev, gfp_flags, type); 897 if (IS_ERR_OR_NULL(pool)) { 898 return -ENOMEM; 899 } 900 } 901 902 INIT_LIST_HEAD(&ttm_dma->pages_list); 903 for (i = 0; i < ttm->num_pages; ++i) { 904 ret = ttm_dma_pool_get_pages(pool, ttm_dma, i); 905 if (ret != 0) { 906 ttm_dma_unpopulate(ttm_dma, dev); 907 return -ENOMEM; 908 } 909 910 ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i], 911 false, false); 912 if (unlikely(ret != 0)) { 913 ttm_dma_unpopulate(ttm_dma, dev); 914 return -ENOMEM; 915 } 916 } 917 918 if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) { 919 ret = ttm_tt_swapin(ttm); 920 if (unlikely(ret != 0)) { 921 ttm_dma_unpopulate(ttm_dma, dev); 922 return ret; 923 } 924 } 925 926 ttm->state = tt_unbound; 927 return 0; 928} 929EXPORT_SYMBOL_GPL(ttm_dma_populate); 930 931/* Put all pages in pages list to correct pool to wait for reuse */ 932void ttm_dma_unpopulate(struct ttm_dma_tt *ttm_dma, struct device *dev) 933{ 934 struct ttm_tt *ttm = &ttm_dma->ttm; 935 struct dma_pool *pool; 936 struct dma_page *d_page, *next; 937 enum pool_type type; 938 bool is_cached = false; 939 unsigned count = 0, i, npages = 0; 940 unsigned long irq_flags; 941 942 type = ttm_to_type(ttm->page_flags, ttm->caching_state); 943 pool = ttm_dma_find_pool(dev, type); 944 if (!pool) 945 return; 946 947 is_cached = (ttm_dma_find_pool(pool->dev, 948 ttm_to_type(ttm->page_flags, tt_cached)) == pool); 949 950 /* make sure pages array match list and count number of pages */ 951 list_for_each_entry(d_page, &ttm_dma->pages_list, page_list) { 952 ttm->pages[count] = d_page->p; 953 count++; 954 } 955 956 spin_lock_irqsave(&pool->lock, irq_flags); 957 pool->npages_in_use -= count; 958 if (is_cached) { 959 pool->nfrees += count; 960 } else { 961 pool->npages_free += count; 962 list_splice(&ttm_dma->pages_list, &pool->free_list); 963 npages = count; 964 if (pool->npages_free > _manager->options.max_size) { 965 npages = pool->npages_free - _manager->options.max_size; 966 /* free at least NUM_PAGES_TO_ALLOC number of pages 967 * to reduce calls to set_memory_wb */ 968 if (npages < NUM_PAGES_TO_ALLOC) 969 npages = NUM_PAGES_TO_ALLOC; 970 } 971 } 972 spin_unlock_irqrestore(&pool->lock, irq_flags); 973 974 if (is_cached) { 975 list_for_each_entry_safe(d_page, next, &ttm_dma->pages_list, page_list) { 976 ttm_mem_global_free_page(ttm->glob->mem_glob, 977 d_page->p); 978 ttm_dma_page_put(pool, d_page); 979 } 980 } else { 981 for (i = 0; i < count; i++) { 982 ttm_mem_global_free_page(ttm->glob->mem_glob, 983 ttm->pages[i]); 984 } 985 } 986 987 INIT_LIST_HEAD(&ttm_dma->pages_list); 988 for (i = 0; i < ttm->num_pages; i++) { 989 ttm->pages[i] = NULL; 990 ttm_dma->cpu_address[i] = 0; 991 ttm_dma->dma_address[i] = 0; 992 } 993 994 /* shrink pool if necessary (only on !is_cached pools)*/ 995 if (npages) 996 ttm_dma_page_pool_free(pool, npages, false); 997 ttm->state = tt_unpopulated; 998} 999EXPORT_SYMBOL_GPL(ttm_dma_unpopulate); 1000 1001/** 1002 * Callback for mm to request pool to reduce number of page held. 1003 * 1004 * XXX: (dchinner) Deadlock warning! 1005 * 1006 * I'm getting sadder as I hear more pathetical whimpers about needing per-pool 1007 * shrinkers 1008 */ 1009static unsigned long 1010ttm_dma_pool_shrink_scan(struct shrinker *shrink, struct shrink_control *sc) 1011{ 1012 static unsigned start_pool; 1013 unsigned idx = 0; 1014 unsigned pool_offset; 1015 unsigned shrink_pages = sc->nr_to_scan; 1016 struct device_pools *p; 1017 unsigned long freed = 0; 1018 1019 if (list_empty(&_manager->pools)) 1020 return SHRINK_STOP; 1021 1022 if (!mutex_trylock(&_manager->lock)) 1023 return SHRINK_STOP; 1024 if (!_manager->npools) 1025 goto out; 1026 pool_offset = ++start_pool % _manager->npools; 1027 list_for_each_entry(p, &_manager->pools, pools) { 1028 unsigned nr_free; 1029 1030 if (!p->dev) 1031 continue; 1032 if (shrink_pages == 0) 1033 break; 1034 /* Do it in round-robin fashion. */ 1035 if (++idx < pool_offset) 1036 continue; 1037 nr_free = shrink_pages; 1038 /* OK to use static buffer since global mutex is held. */ 1039 shrink_pages = ttm_dma_page_pool_free(p->pool, nr_free, true); 1040 freed += nr_free - shrink_pages; 1041 1042 pr_debug("%s: (%s:%d) Asked to shrink %d, have %d more to go\n", 1043 p->pool->dev_name, p->pool->name, current->pid, 1044 nr_free, shrink_pages); 1045 } 1046out: 1047 mutex_unlock(&_manager->lock); 1048 return freed; 1049} 1050 1051static unsigned long 1052ttm_dma_pool_shrink_count(struct shrinker *shrink, struct shrink_control *sc) 1053{ 1054 struct device_pools *p; 1055 unsigned long count = 0; 1056 1057 if (!mutex_trylock(&_manager->lock)) 1058 return 0; 1059 list_for_each_entry(p, &_manager->pools, pools) 1060 count += p->pool->npages_free; 1061 mutex_unlock(&_manager->lock); 1062 return count; 1063} 1064 1065static void ttm_dma_pool_mm_shrink_init(struct ttm_pool_manager *manager) 1066{ 1067 manager->mm_shrink.count_objects = ttm_dma_pool_shrink_count; 1068 manager->mm_shrink.scan_objects = &ttm_dma_pool_shrink_scan; 1069 manager->mm_shrink.seeks = 1; 1070 register_shrinker(&manager->mm_shrink); 1071} 1072 1073static void ttm_dma_pool_mm_shrink_fini(struct ttm_pool_manager *manager) 1074{ 1075 unregister_shrinker(&manager->mm_shrink); 1076} 1077 1078int ttm_dma_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages) 1079{ 1080 int ret = -ENOMEM; 1081 1082 WARN_ON(_manager); 1083 1084 pr_info("Initializing DMA pool allocator\n"); 1085 1086 _manager = kzalloc(sizeof(*_manager), GFP_KERNEL); 1087 if (!_manager) 1088 goto err; 1089 1090 mutex_init(&_manager->lock); 1091 INIT_LIST_HEAD(&_manager->pools); 1092 1093 _manager->options.max_size = max_pages; 1094 _manager->options.small = SMALL_ALLOCATION; 1095 _manager->options.alloc_size = NUM_PAGES_TO_ALLOC; 1096 1097 /* This takes care of auto-freeing the _manager */ 1098 ret = kobject_init_and_add(&_manager->kobj, &ttm_pool_kobj_type, 1099 &glob->kobj, "dma_pool"); 1100 if (unlikely(ret != 0)) { 1101 kobject_put(&_manager->kobj); 1102 goto err; 1103 } 1104 ttm_dma_pool_mm_shrink_init(_manager); 1105 return 0; 1106err: 1107 return ret; 1108} 1109 1110void ttm_dma_page_alloc_fini(void) 1111{ 1112 struct device_pools *p, *t; 1113 1114 pr_info("Finalizing DMA pool allocator\n"); 1115 ttm_dma_pool_mm_shrink_fini(_manager); 1116 1117 list_for_each_entry_safe_reverse(p, t, &_manager->pools, pools) { 1118 dev_dbg(p->dev, "(%s:%d) Freeing.\n", p->pool->name, 1119 current->pid); 1120 WARN_ON(devres_destroy(p->dev, ttm_dma_pool_release, 1121 ttm_dma_pool_match, p->pool)); 1122 ttm_dma_free_pool(p->dev, p->pool->type); 1123 } 1124 kobject_put(&_manager->kobj); 1125 _manager = NULL; 1126} 1127 1128int ttm_dma_page_alloc_debugfs(struct seq_file *m, void *data) 1129{ 1130 struct device_pools *p; 1131 struct dma_pool *pool = NULL; 1132 char *h[] = {"pool", "refills", "pages freed", "inuse", "available", 1133 "name", "virt", "busaddr"}; 1134 1135 if (!_manager) { 1136 seq_printf(m, "No pool allocator running.\n"); 1137 return 0; 1138 } 1139 seq_printf(m, "%13s %12s %13s %8s %8s %8s\n", 1140 h[0], h[1], h[2], h[3], h[4], h[5]); 1141 mutex_lock(&_manager->lock); 1142 list_for_each_entry(p, &_manager->pools, pools) { 1143 struct device *dev = p->dev; 1144 if (!dev) 1145 continue; 1146 pool = p->pool; 1147 seq_printf(m, "%13s %12ld %13ld %8d %8d %8s\n", 1148 pool->name, pool->nrefills, 1149 pool->nfrees, pool->npages_in_use, 1150 pool->npages_free, 1151 pool->dev_name); 1152 } 1153 mutex_unlock(&_manager->lock); 1154 return 0; 1155} 1156EXPORT_SYMBOL_GPL(ttm_dma_page_alloc_debugfs); 1157 1158#endif 1159