1 /* 2 * An async IO implementation for Linux 3 * Written by Benjamin LaHaise <bcrl@kvack.org> 4 * 5 * Implements an efficient asynchronous io interface. 6 * 7 * Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved. 8 * 9 * See ../COPYING for licensing terms. 10 */ 11 #define pr_fmt(fmt) "%s: " fmt, __func__ 12 13 #include <linux/kernel.h> 14 #include <linux/init.h> 15 #include <linux/errno.h> 16 #include <linux/time.h> 17 #include <linux/aio_abi.h> 18 #include <linux/export.h> 19 #include <linux/syscalls.h> 20 #include <linux/backing-dev.h> 21 #include <linux/uio.h> 22 23 #include <linux/sched/signal.h> 24 #include <linux/fs.h> 25 #include <linux/file.h> 26 #include <linux/mm.h> 27 #include <linux/mman.h> 28 #include <linux/mmu_context.h> 29 #include <linux/percpu.h> 30 #include <linux/slab.h> 31 #include <linux/timer.h> 32 #include <linux/aio.h> 33 #include <linux/highmem.h> 34 #include <linux/workqueue.h> 35 #include <linux/security.h> 36 #include <linux/eventfd.h> 37 #include <linux/blkdev.h> 38 #include <linux/compat.h> 39 #include <linux/migrate.h> 40 #include <linux/ramfs.h> 41 #include <linux/percpu-refcount.h> 42 #include <linux/mount.h> 43 44 #include <asm/kmap_types.h> 45 #include <linux/uaccess.h> 46 47 #include "internal.h" 48 49 #define AIO_RING_MAGIC 0xa10a10a1 50 #define AIO_RING_COMPAT_FEATURES 1 51 #define AIO_RING_INCOMPAT_FEATURES 0 52 struct aio_ring { 53 unsigned id; /* kernel internal index number */ 54 unsigned nr; /* number of io_events */ 55 unsigned head; /* Written to by userland or under ring_lock 56 * mutex by aio_read_events_ring(). */ 57 unsigned tail; 58 59 unsigned magic; 60 unsigned compat_features; 61 unsigned incompat_features; 62 unsigned header_length; /* size of aio_ring */ 63 64 65 struct io_event io_events[0]; 66 }; /* 128 bytes + ring size */ 67 68 #define AIO_RING_PAGES 8 69 70 struct kioctx_table { 71 struct rcu_head rcu; 72 unsigned nr; 73 struct kioctx *table[]; 74 }; 75 76 struct kioctx_cpu { 77 unsigned reqs_available; 78 }; 79 80 struct ctx_rq_wait { 81 struct completion comp; 82 atomic_t count; 83 }; 84 85 struct kioctx { 86 struct percpu_ref users; 87 atomic_t dead; 88 89 struct percpu_ref reqs; 90 91 unsigned long user_id; 92 93 struct __percpu kioctx_cpu *cpu; 94 95 /* 96 * For percpu reqs_available, number of slots we move to/from global 97 * counter at a time: 98 */ 99 unsigned req_batch; 100 /* 101 * This is what userspace passed to io_setup(), it's not used for 102 * anything but counting against the global max_reqs quota. 103 * 104 * The real limit is nr_events - 1, which will be larger (see 105 * aio_setup_ring()) 106 */ 107 unsigned max_reqs; 108 109 /* Size of ringbuffer, in units of struct io_event */ 110 unsigned nr_events; 111 112 unsigned long mmap_base; 113 unsigned long mmap_size; 114 115 struct page **ring_pages; 116 long nr_pages; 117 118 struct work_struct free_work; 119 120 /* 121 * signals when all in-flight requests are done 122 */ 123 struct ctx_rq_wait *rq_wait; 124 125 struct { 126 /* 127 * This counts the number of available slots in the ringbuffer, 128 * so we avoid overflowing it: it's decremented (if positive) 129 * when allocating a kiocb and incremented when the resulting 130 * io_event is pulled off the ringbuffer. 131 * 132 * We batch accesses to it with a percpu version. 133 */ 134 atomic_t reqs_available; 135 } ____cacheline_aligned_in_smp; 136 137 struct { 138 spinlock_t ctx_lock; 139 struct list_head active_reqs; /* used for cancellation */ 140 } ____cacheline_aligned_in_smp; 141 142 struct { 143 struct mutex ring_lock; 144 wait_queue_head_t wait; 145 } ____cacheline_aligned_in_smp; 146 147 struct { 148 unsigned tail; 149 unsigned completed_events; 150 spinlock_t completion_lock; 151 } ____cacheline_aligned_in_smp; 152 153 struct page *internal_pages[AIO_RING_PAGES]; 154 struct file *aio_ring_file; 155 156 unsigned id; 157 }; 158 159 /* 160 * We use ki_cancel == KIOCB_CANCELLED to indicate that a kiocb has been either 161 * cancelled or completed (this makes a certain amount of sense because 162 * successful cancellation - io_cancel() - does deliver the completion to 163 * userspace). 164 * 165 * And since most things don't implement kiocb cancellation and we'd really like 166 * kiocb completion to be lockless when possible, we use ki_cancel to 167 * synchronize cancellation and completion - we only set it to KIOCB_CANCELLED 168 * with xchg() or cmpxchg(), see batch_complete_aio() and kiocb_cancel(). 169 */ 170 #define KIOCB_CANCELLED ((void *) (~0ULL)) 171 172 struct aio_kiocb { 173 struct kiocb common; 174 175 struct kioctx *ki_ctx; 176 kiocb_cancel_fn *ki_cancel; 177 178 struct iocb __user *ki_user_iocb; /* user's aiocb */ 179 __u64 ki_user_data; /* user's data for completion */ 180 181 struct list_head ki_list; /* the aio core uses this 182 * for cancellation */ 183 184 /* 185 * If the aio_resfd field of the userspace iocb is not zero, 186 * this is the underlying eventfd context to deliver events to. 187 */ 188 struct eventfd_ctx *ki_eventfd; 189 }; 190 191 /*------ sysctl variables----*/ 192 static DEFINE_SPINLOCK(aio_nr_lock); 193 unsigned long aio_nr; /* current system wide number of aio requests */ 194 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */ 195 /*----end sysctl variables---*/ 196 197 static struct kmem_cache *kiocb_cachep; 198 static struct kmem_cache *kioctx_cachep; 199 200 static struct vfsmount *aio_mnt; 201 202 static const struct file_operations aio_ring_fops; 203 static const struct address_space_operations aio_ctx_aops; 204 205 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages) 206 { 207 struct qstr this = QSTR_INIT("[aio]", 5); 208 struct file *file; 209 struct path path; 210 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb); 211 if (IS_ERR(inode)) 212 return ERR_CAST(inode); 213 214 inode->i_mapping->a_ops = &aio_ctx_aops; 215 inode->i_mapping->private_data = ctx; 216 inode->i_size = PAGE_SIZE * nr_pages; 217 218 path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this); 219 if (!path.dentry) { 220 iput(inode); 221 return ERR_PTR(-ENOMEM); 222 } 223 path.mnt = mntget(aio_mnt); 224 225 d_instantiate(path.dentry, inode); 226 file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops); 227 if (IS_ERR(file)) { 228 path_put(&path); 229 return file; 230 } 231 232 file->f_flags = O_RDWR; 233 return file; 234 } 235 236 static struct dentry *aio_mount(struct file_system_type *fs_type, 237 int flags, const char *dev_name, void *data) 238 { 239 static const struct dentry_operations ops = { 240 .d_dname = simple_dname, 241 }; 242 struct dentry *root = mount_pseudo(fs_type, "aio:", NULL, &ops, 243 AIO_RING_MAGIC); 244 245 if (!IS_ERR(root)) 246 root->d_sb->s_iflags |= SB_I_NOEXEC; 247 return root; 248 } 249 250 /* aio_setup 251 * Creates the slab caches used by the aio routines, panic on 252 * failure as this is done early during the boot sequence. 253 */ 254 static int __init aio_setup(void) 255 { 256 static struct file_system_type aio_fs = { 257 .name = "aio", 258 .mount = aio_mount, 259 .kill_sb = kill_anon_super, 260 }; 261 aio_mnt = kern_mount(&aio_fs); 262 if (IS_ERR(aio_mnt)) 263 panic("Failed to create aio fs mount."); 264 265 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC); 266 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC); 267 268 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page)); 269 270 return 0; 271 } 272 __initcall(aio_setup); 273 274 static void put_aio_ring_file(struct kioctx *ctx) 275 { 276 struct file *aio_ring_file = ctx->aio_ring_file; 277 struct address_space *i_mapping; 278 279 if (aio_ring_file) { 280 truncate_setsize(file_inode(aio_ring_file), 0); 281 282 /* Prevent further access to the kioctx from migratepages */ 283 i_mapping = aio_ring_file->f_mapping; 284 spin_lock(&i_mapping->private_lock); 285 i_mapping->private_data = NULL; 286 ctx->aio_ring_file = NULL; 287 spin_unlock(&i_mapping->private_lock); 288 289 fput(aio_ring_file); 290 } 291 } 292 293 static void aio_free_ring(struct kioctx *ctx) 294 { 295 int i; 296 297 /* Disconnect the kiotx from the ring file. This prevents future 298 * accesses to the kioctx from page migration. 299 */ 300 put_aio_ring_file(ctx); 301 302 for (i = 0; i < ctx->nr_pages; i++) { 303 struct page *page; 304 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i, 305 page_count(ctx->ring_pages[i])); 306 page = ctx->ring_pages[i]; 307 if (!page) 308 continue; 309 ctx->ring_pages[i] = NULL; 310 put_page(page); 311 } 312 313 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) { 314 kfree(ctx->ring_pages); 315 ctx->ring_pages = NULL; 316 } 317 } 318 319 static int aio_ring_mremap(struct vm_area_struct *vma) 320 { 321 struct file *file = vma->vm_file; 322 struct mm_struct *mm = vma->vm_mm; 323 struct kioctx_table *table; 324 int i, res = -EINVAL; 325 326 spin_lock(&mm->ioctx_lock); 327 rcu_read_lock(); 328 table = rcu_dereference(mm->ioctx_table); 329 for (i = 0; i < table->nr; i++) { 330 struct kioctx *ctx; 331 332 ctx = table->table[i]; 333 if (ctx && ctx->aio_ring_file == file) { 334 if (!atomic_read(&ctx->dead)) { 335 ctx->user_id = ctx->mmap_base = vma->vm_start; 336 res = 0; 337 } 338 break; 339 } 340 } 341 342 rcu_read_unlock(); 343 spin_unlock(&mm->ioctx_lock); 344 return res; 345 } 346 347 static const struct vm_operations_struct aio_ring_vm_ops = { 348 .mremap = aio_ring_mremap, 349 #if IS_ENABLED(CONFIG_MMU) 350 .fault = filemap_fault, 351 .map_pages = filemap_map_pages, 352 .page_mkwrite = filemap_page_mkwrite, 353 #endif 354 }; 355 356 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma) 357 { 358 vma->vm_flags |= VM_DONTEXPAND; 359 vma->vm_ops = &aio_ring_vm_ops; 360 return 0; 361 } 362 363 static const struct file_operations aio_ring_fops = { 364 .mmap = aio_ring_mmap, 365 }; 366 367 #if IS_ENABLED(CONFIG_MIGRATION) 368 static int aio_migratepage(struct address_space *mapping, struct page *new, 369 struct page *old, enum migrate_mode mode) 370 { 371 struct kioctx *ctx; 372 unsigned long flags; 373 pgoff_t idx; 374 int rc; 375 376 /* 377 * We cannot support the _NO_COPY case here, because copy needs to 378 * happen under the ctx->completion_lock. That does not work with the 379 * migration workflow of MIGRATE_SYNC_NO_COPY. 380 */ 381 if (mode == MIGRATE_SYNC_NO_COPY) 382 return -EINVAL; 383 384 rc = 0; 385 386 /* mapping->private_lock here protects against the kioctx teardown. */ 387 spin_lock(&mapping->private_lock); 388 ctx = mapping->private_data; 389 if (!ctx) { 390 rc = -EINVAL; 391 goto out; 392 } 393 394 /* The ring_lock mutex. The prevents aio_read_events() from writing 395 * to the ring's head, and prevents page migration from mucking in 396 * a partially initialized kiotx. 397 */ 398 if (!mutex_trylock(&ctx->ring_lock)) { 399 rc = -EAGAIN; 400 goto out; 401 } 402 403 idx = old->index; 404 if (idx < (pgoff_t)ctx->nr_pages) { 405 /* Make sure the old page hasn't already been changed */ 406 if (ctx->ring_pages[idx] != old) 407 rc = -EAGAIN; 408 } else 409 rc = -EINVAL; 410 411 if (rc != 0) 412 goto out_unlock; 413 414 /* Writeback must be complete */ 415 BUG_ON(PageWriteback(old)); 416 get_page(new); 417 418 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1); 419 if (rc != MIGRATEPAGE_SUCCESS) { 420 put_page(new); 421 goto out_unlock; 422 } 423 424 /* Take completion_lock to prevent other writes to the ring buffer 425 * while the old page is copied to the new. This prevents new 426 * events from being lost. 427 */ 428 spin_lock_irqsave(&ctx->completion_lock, flags); 429 migrate_page_copy(new, old); 430 BUG_ON(ctx->ring_pages[idx] != old); 431 ctx->ring_pages[idx] = new; 432 spin_unlock_irqrestore(&ctx->completion_lock, flags); 433 434 /* The old page is no longer accessible. */ 435 put_page(old); 436 437 out_unlock: 438 mutex_unlock(&ctx->ring_lock); 439 out: 440 spin_unlock(&mapping->private_lock); 441 return rc; 442 } 443 #endif 444 445 static const struct address_space_operations aio_ctx_aops = { 446 .set_page_dirty = __set_page_dirty_no_writeback, 447 #if IS_ENABLED(CONFIG_MIGRATION) 448 .migratepage = aio_migratepage, 449 #endif 450 }; 451 452 static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events) 453 { 454 struct aio_ring *ring; 455 struct mm_struct *mm = current->mm; 456 unsigned long size, unused; 457 int nr_pages; 458 int i; 459 struct file *file; 460 461 /* Compensate for the ring buffer's head/tail overlap entry */ 462 nr_events += 2; /* 1 is required, 2 for good luck */ 463 464 size = sizeof(struct aio_ring); 465 size += sizeof(struct io_event) * nr_events; 466 467 nr_pages = PFN_UP(size); 468 if (nr_pages < 0) 469 return -EINVAL; 470 471 file = aio_private_file(ctx, nr_pages); 472 if (IS_ERR(file)) { 473 ctx->aio_ring_file = NULL; 474 return -ENOMEM; 475 } 476 477 ctx->aio_ring_file = file; 478 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) 479 / sizeof(struct io_event); 480 481 ctx->ring_pages = ctx->internal_pages; 482 if (nr_pages > AIO_RING_PAGES) { 483 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *), 484 GFP_KERNEL); 485 if (!ctx->ring_pages) { 486 put_aio_ring_file(ctx); 487 return -ENOMEM; 488 } 489 } 490 491 for (i = 0; i < nr_pages; i++) { 492 struct page *page; 493 page = find_or_create_page(file->f_mapping, 494 i, GFP_HIGHUSER | __GFP_ZERO); 495 if (!page) 496 break; 497 pr_debug("pid(%d) page[%d]->count=%d\n", 498 current->pid, i, page_count(page)); 499 SetPageUptodate(page); 500 unlock_page(page); 501 502 ctx->ring_pages[i] = page; 503 } 504 ctx->nr_pages = i; 505 506 if (unlikely(i != nr_pages)) { 507 aio_free_ring(ctx); 508 return -ENOMEM; 509 } 510 511 ctx->mmap_size = nr_pages * PAGE_SIZE; 512 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size); 513 514 if (down_write_killable(&mm->mmap_sem)) { 515 ctx->mmap_size = 0; 516 aio_free_ring(ctx); 517 return -EINTR; 518 } 519 520 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size, 521 PROT_READ | PROT_WRITE, 522 MAP_SHARED, 0, &unused, NULL); 523 up_write(&mm->mmap_sem); 524 if (IS_ERR((void *)ctx->mmap_base)) { 525 ctx->mmap_size = 0; 526 aio_free_ring(ctx); 527 return -ENOMEM; 528 } 529 530 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base); 531 532 ctx->user_id = ctx->mmap_base; 533 ctx->nr_events = nr_events; /* trusted copy */ 534 535 ring = kmap_atomic(ctx->ring_pages[0]); 536 ring->nr = nr_events; /* user copy */ 537 ring->id = ~0U; 538 ring->head = ring->tail = 0; 539 ring->magic = AIO_RING_MAGIC; 540 ring->compat_features = AIO_RING_COMPAT_FEATURES; 541 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES; 542 ring->header_length = sizeof(struct aio_ring); 543 kunmap_atomic(ring); 544 flush_dcache_page(ctx->ring_pages[0]); 545 546 return 0; 547 } 548 549 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event)) 550 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event)) 551 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE) 552 553 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel) 554 { 555 struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, common); 556 struct kioctx *ctx = req->ki_ctx; 557 unsigned long flags; 558 559 spin_lock_irqsave(&ctx->ctx_lock, flags); 560 561 if (!req->ki_list.next) 562 list_add(&req->ki_list, &ctx->active_reqs); 563 564 req->ki_cancel = cancel; 565 566 spin_unlock_irqrestore(&ctx->ctx_lock, flags); 567 } 568 EXPORT_SYMBOL(kiocb_set_cancel_fn); 569 570 static int kiocb_cancel(struct aio_kiocb *kiocb) 571 { 572 kiocb_cancel_fn *old, *cancel; 573 574 /* 575 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it 576 * actually has a cancel function, hence the cmpxchg() 577 */ 578 579 cancel = ACCESS_ONCE(kiocb->ki_cancel); 580 do { 581 if (!cancel || cancel == KIOCB_CANCELLED) 582 return -EINVAL; 583 584 old = cancel; 585 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED); 586 } while (cancel != old); 587 588 return cancel(&kiocb->common); 589 } 590 591 static void free_ioctx(struct work_struct *work) 592 { 593 struct kioctx *ctx = container_of(work, struct kioctx, free_work); 594 595 pr_debug("freeing %p\n", ctx); 596 597 aio_free_ring(ctx); 598 free_percpu(ctx->cpu); 599 percpu_ref_exit(&ctx->reqs); 600 percpu_ref_exit(&ctx->users); 601 kmem_cache_free(kioctx_cachep, ctx); 602 } 603 604 static void free_ioctx_reqs(struct percpu_ref *ref) 605 { 606 struct kioctx *ctx = container_of(ref, struct kioctx, reqs); 607 608 /* At this point we know that there are no any in-flight requests */ 609 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count)) 610 complete(&ctx->rq_wait->comp); 611 612 INIT_WORK(&ctx->free_work, free_ioctx); 613 schedule_work(&ctx->free_work); 614 } 615 616 /* 617 * When this function runs, the kioctx has been removed from the "hash table" 618 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted - 619 * now it's safe to cancel any that need to be. 620 */ 621 static void free_ioctx_users(struct percpu_ref *ref) 622 { 623 struct kioctx *ctx = container_of(ref, struct kioctx, users); 624 struct aio_kiocb *req; 625 626 spin_lock_irq(&ctx->ctx_lock); 627 628 while (!list_empty(&ctx->active_reqs)) { 629 req = list_first_entry(&ctx->active_reqs, 630 struct aio_kiocb, ki_list); 631 632 list_del_init(&req->ki_list); 633 kiocb_cancel(req); 634 } 635 636 spin_unlock_irq(&ctx->ctx_lock); 637 638 percpu_ref_kill(&ctx->reqs); 639 percpu_ref_put(&ctx->reqs); 640 } 641 642 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm) 643 { 644 unsigned i, new_nr; 645 struct kioctx_table *table, *old; 646 struct aio_ring *ring; 647 648 spin_lock(&mm->ioctx_lock); 649 table = rcu_dereference_raw(mm->ioctx_table); 650 651 while (1) { 652 if (table) 653 for (i = 0; i < table->nr; i++) 654 if (!table->table[i]) { 655 ctx->id = i; 656 table->table[i] = ctx; 657 spin_unlock(&mm->ioctx_lock); 658 659 /* While kioctx setup is in progress, 660 * we are protected from page migration 661 * changes ring_pages by ->ring_lock. 662 */ 663 ring = kmap_atomic(ctx->ring_pages[0]); 664 ring->id = ctx->id; 665 kunmap_atomic(ring); 666 return 0; 667 } 668 669 new_nr = (table ? table->nr : 1) * 4; 670 spin_unlock(&mm->ioctx_lock); 671 672 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) * 673 new_nr, GFP_KERNEL); 674 if (!table) 675 return -ENOMEM; 676 677 table->nr = new_nr; 678 679 spin_lock(&mm->ioctx_lock); 680 old = rcu_dereference_raw(mm->ioctx_table); 681 682 if (!old) { 683 rcu_assign_pointer(mm->ioctx_table, table); 684 } else if (table->nr > old->nr) { 685 memcpy(table->table, old->table, 686 old->nr * sizeof(struct kioctx *)); 687 688 rcu_assign_pointer(mm->ioctx_table, table); 689 kfree_rcu(old, rcu); 690 } else { 691 kfree(table); 692 table = old; 693 } 694 } 695 } 696 697 static void aio_nr_sub(unsigned nr) 698 { 699 spin_lock(&aio_nr_lock); 700 if (WARN_ON(aio_nr - nr > aio_nr)) 701 aio_nr = 0; 702 else 703 aio_nr -= nr; 704 spin_unlock(&aio_nr_lock); 705 } 706 707 /* ioctx_alloc 708 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed. 709 */ 710 static struct kioctx *ioctx_alloc(unsigned nr_events) 711 { 712 struct mm_struct *mm = current->mm; 713 struct kioctx *ctx; 714 int err = -ENOMEM; 715 716 /* 717 * Store the original nr_events -- what userspace passed to io_setup(), 718 * for counting against the global limit -- before it changes. 719 */ 720 unsigned int max_reqs = nr_events; 721 722 /* 723 * We keep track of the number of available ringbuffer slots, to prevent 724 * overflow (reqs_available), and we also use percpu counters for this. 725 * 726 * So since up to half the slots might be on other cpu's percpu counters 727 * and unavailable, double nr_events so userspace sees what they 728 * expected: additionally, we move req_batch slots to/from percpu 729 * counters at a time, so make sure that isn't 0: 730 */ 731 nr_events = max(nr_events, num_possible_cpus() * 4); 732 nr_events *= 2; 733 734 /* Prevent overflows */ 735 if (nr_events > (0x10000000U / sizeof(struct io_event))) { 736 pr_debug("ENOMEM: nr_events too high\n"); 737 return ERR_PTR(-EINVAL); 738 } 739 740 if (!nr_events || (unsigned long)max_reqs > aio_max_nr) 741 return ERR_PTR(-EAGAIN); 742 743 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL); 744 if (!ctx) 745 return ERR_PTR(-ENOMEM); 746 747 ctx->max_reqs = max_reqs; 748 749 spin_lock_init(&ctx->ctx_lock); 750 spin_lock_init(&ctx->completion_lock); 751 mutex_init(&ctx->ring_lock); 752 /* Protect against page migration throughout kiotx setup by keeping 753 * the ring_lock mutex held until setup is complete. */ 754 mutex_lock(&ctx->ring_lock); 755 init_waitqueue_head(&ctx->wait); 756 757 INIT_LIST_HEAD(&ctx->active_reqs); 758 759 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL)) 760 goto err; 761 762 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL)) 763 goto err; 764 765 ctx->cpu = alloc_percpu(struct kioctx_cpu); 766 if (!ctx->cpu) 767 goto err; 768 769 err = aio_setup_ring(ctx, nr_events); 770 if (err < 0) 771 goto err; 772 773 atomic_set(&ctx->reqs_available, ctx->nr_events - 1); 774 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4); 775 if (ctx->req_batch < 1) 776 ctx->req_batch = 1; 777 778 /* limit the number of system wide aios */ 779 spin_lock(&aio_nr_lock); 780 if (aio_nr + ctx->max_reqs > aio_max_nr || 781 aio_nr + ctx->max_reqs < aio_nr) { 782 spin_unlock(&aio_nr_lock); 783 err = -EAGAIN; 784 goto err_ctx; 785 } 786 aio_nr += ctx->max_reqs; 787 spin_unlock(&aio_nr_lock); 788 789 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */ 790 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */ 791 792 err = ioctx_add_table(ctx, mm); 793 if (err) 794 goto err_cleanup; 795 796 /* Release the ring_lock mutex now that all setup is complete. */ 797 mutex_unlock(&ctx->ring_lock); 798 799 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n", 800 ctx, ctx->user_id, mm, ctx->nr_events); 801 return ctx; 802 803 err_cleanup: 804 aio_nr_sub(ctx->max_reqs); 805 err_ctx: 806 atomic_set(&ctx->dead, 1); 807 if (ctx->mmap_size) 808 vm_munmap(ctx->mmap_base, ctx->mmap_size); 809 aio_free_ring(ctx); 810 err: 811 mutex_unlock(&ctx->ring_lock); 812 free_percpu(ctx->cpu); 813 percpu_ref_exit(&ctx->reqs); 814 percpu_ref_exit(&ctx->users); 815 kmem_cache_free(kioctx_cachep, ctx); 816 pr_debug("error allocating ioctx %d\n", err); 817 return ERR_PTR(err); 818 } 819 820 /* kill_ioctx 821 * Cancels all outstanding aio requests on an aio context. Used 822 * when the processes owning a context have all exited to encourage 823 * the rapid destruction of the kioctx. 824 */ 825 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx, 826 struct ctx_rq_wait *wait) 827 { 828 struct kioctx_table *table; 829 830 spin_lock(&mm->ioctx_lock); 831 if (atomic_xchg(&ctx->dead, 1)) { 832 spin_unlock(&mm->ioctx_lock); 833 return -EINVAL; 834 } 835 836 table = rcu_dereference_raw(mm->ioctx_table); 837 WARN_ON(ctx != table->table[ctx->id]); 838 table->table[ctx->id] = NULL; 839 spin_unlock(&mm->ioctx_lock); 840 841 /* percpu_ref_kill() will do the necessary call_rcu() */ 842 wake_up_all(&ctx->wait); 843 844 /* 845 * It'd be more correct to do this in free_ioctx(), after all 846 * the outstanding kiocbs have finished - but by then io_destroy 847 * has already returned, so io_setup() could potentially return 848 * -EAGAIN with no ioctxs actually in use (as far as userspace 849 * could tell). 850 */ 851 aio_nr_sub(ctx->max_reqs); 852 853 if (ctx->mmap_size) 854 vm_munmap(ctx->mmap_base, ctx->mmap_size); 855 856 ctx->rq_wait = wait; 857 percpu_ref_kill(&ctx->users); 858 return 0; 859 } 860 861 /* 862 * exit_aio: called when the last user of mm goes away. At this point, there is 863 * no way for any new requests to be submited or any of the io_* syscalls to be 864 * called on the context. 865 * 866 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on 867 * them. 868 */ 869 void exit_aio(struct mm_struct *mm) 870 { 871 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table); 872 struct ctx_rq_wait wait; 873 int i, skipped; 874 875 if (!table) 876 return; 877 878 atomic_set(&wait.count, table->nr); 879 init_completion(&wait.comp); 880 881 skipped = 0; 882 for (i = 0; i < table->nr; ++i) { 883 struct kioctx *ctx = table->table[i]; 884 885 if (!ctx) { 886 skipped++; 887 continue; 888 } 889 890 /* 891 * We don't need to bother with munmap() here - exit_mmap(mm) 892 * is coming and it'll unmap everything. And we simply can't, 893 * this is not necessarily our ->mm. 894 * Since kill_ioctx() uses non-zero ->mmap_size as indicator 895 * that it needs to unmap the area, just set it to 0. 896 */ 897 ctx->mmap_size = 0; 898 kill_ioctx(mm, ctx, &wait); 899 } 900 901 if (!atomic_sub_and_test(skipped, &wait.count)) { 902 /* Wait until all IO for the context are done. */ 903 wait_for_completion(&wait.comp); 904 } 905 906 RCU_INIT_POINTER(mm->ioctx_table, NULL); 907 kfree(table); 908 } 909 910 static void put_reqs_available(struct kioctx *ctx, unsigned nr) 911 { 912 struct kioctx_cpu *kcpu; 913 unsigned long flags; 914 915 local_irq_save(flags); 916 kcpu = this_cpu_ptr(ctx->cpu); 917 kcpu->reqs_available += nr; 918 919 while (kcpu->reqs_available >= ctx->req_batch * 2) { 920 kcpu->reqs_available -= ctx->req_batch; 921 atomic_add(ctx->req_batch, &ctx->reqs_available); 922 } 923 924 local_irq_restore(flags); 925 } 926 927 static bool get_reqs_available(struct kioctx *ctx) 928 { 929 struct kioctx_cpu *kcpu; 930 bool ret = false; 931 unsigned long flags; 932 933 local_irq_save(flags); 934 kcpu = this_cpu_ptr(ctx->cpu); 935 if (!kcpu->reqs_available) { 936 int old, avail = atomic_read(&ctx->reqs_available); 937 938 do { 939 if (avail < ctx->req_batch) 940 goto out; 941 942 old = avail; 943 avail = atomic_cmpxchg(&ctx->reqs_available, 944 avail, avail - ctx->req_batch); 945 } while (avail != old); 946 947 kcpu->reqs_available += ctx->req_batch; 948 } 949 950 ret = true; 951 kcpu->reqs_available--; 952 out: 953 local_irq_restore(flags); 954 return ret; 955 } 956 957 /* refill_reqs_available 958 * Updates the reqs_available reference counts used for tracking the 959 * number of free slots in the completion ring. This can be called 960 * from aio_complete() (to optimistically update reqs_available) or 961 * from aio_get_req() (the we're out of events case). It must be 962 * called holding ctx->completion_lock. 963 */ 964 static void refill_reqs_available(struct kioctx *ctx, unsigned head, 965 unsigned tail) 966 { 967 unsigned events_in_ring, completed; 968 969 /* Clamp head since userland can write to it. */ 970 head %= ctx->nr_events; 971 if (head <= tail) 972 events_in_ring = tail - head; 973 else 974 events_in_ring = ctx->nr_events - (head - tail); 975 976 completed = ctx->completed_events; 977 if (events_in_ring < completed) 978 completed -= events_in_ring; 979 else 980 completed = 0; 981 982 if (!completed) 983 return; 984 985 ctx->completed_events -= completed; 986 put_reqs_available(ctx, completed); 987 } 988 989 /* user_refill_reqs_available 990 * Called to refill reqs_available when aio_get_req() encounters an 991 * out of space in the completion ring. 992 */ 993 static void user_refill_reqs_available(struct kioctx *ctx) 994 { 995 spin_lock_irq(&ctx->completion_lock); 996 if (ctx->completed_events) { 997 struct aio_ring *ring; 998 unsigned head; 999 1000 /* Access of ring->head may race with aio_read_events_ring() 1001 * here, but that's okay since whether we read the old version 1002 * or the new version, and either will be valid. The important 1003 * part is that head cannot pass tail since we prevent 1004 * aio_complete() from updating tail by holding 1005 * ctx->completion_lock. Even if head is invalid, the check 1006 * against ctx->completed_events below will make sure we do the 1007 * safe/right thing. 1008 */ 1009 ring = kmap_atomic(ctx->ring_pages[0]); 1010 head = ring->head; 1011 kunmap_atomic(ring); 1012 1013 refill_reqs_available(ctx, head, ctx->tail); 1014 } 1015 1016 spin_unlock_irq(&ctx->completion_lock); 1017 } 1018 1019 /* aio_get_req 1020 * Allocate a slot for an aio request. 1021 * Returns NULL if no requests are free. 1022 */ 1023 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx) 1024 { 1025 struct aio_kiocb *req; 1026 1027 if (!get_reqs_available(ctx)) { 1028 user_refill_reqs_available(ctx); 1029 if (!get_reqs_available(ctx)) 1030 return NULL; 1031 } 1032 1033 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO); 1034 if (unlikely(!req)) 1035 goto out_put; 1036 1037 percpu_ref_get(&ctx->reqs); 1038 1039 req->ki_ctx = ctx; 1040 return req; 1041 out_put: 1042 put_reqs_available(ctx, 1); 1043 return NULL; 1044 } 1045 1046 static void kiocb_free(struct aio_kiocb *req) 1047 { 1048 if (req->common.ki_filp) 1049 fput(req->common.ki_filp); 1050 if (req->ki_eventfd != NULL) 1051 eventfd_ctx_put(req->ki_eventfd); 1052 kmem_cache_free(kiocb_cachep, req); 1053 } 1054 1055 static struct kioctx *lookup_ioctx(unsigned long ctx_id) 1056 { 1057 struct aio_ring __user *ring = (void __user *)ctx_id; 1058 struct mm_struct *mm = current->mm; 1059 struct kioctx *ctx, *ret = NULL; 1060 struct kioctx_table *table; 1061 unsigned id; 1062 1063 if (get_user(id, &ring->id)) 1064 return NULL; 1065 1066 rcu_read_lock(); 1067 table = rcu_dereference(mm->ioctx_table); 1068 1069 if (!table || id >= table->nr) 1070 goto out; 1071 1072 ctx = table->table[id]; 1073 if (ctx && ctx->user_id == ctx_id) { 1074 percpu_ref_get(&ctx->users); 1075 ret = ctx; 1076 } 1077 out: 1078 rcu_read_unlock(); 1079 return ret; 1080 } 1081 1082 /* aio_complete 1083 * Called when the io request on the given iocb is complete. 1084 */ 1085 static void aio_complete(struct kiocb *kiocb, long res, long res2) 1086 { 1087 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, common); 1088 struct kioctx *ctx = iocb->ki_ctx; 1089 struct aio_ring *ring; 1090 struct io_event *ev_page, *event; 1091 unsigned tail, pos, head; 1092 unsigned long flags; 1093 1094 if (kiocb->ki_flags & IOCB_WRITE) { 1095 struct file *file = kiocb->ki_filp; 1096 1097 /* 1098 * Tell lockdep we inherited freeze protection from submission 1099 * thread. 1100 */ 1101 if (S_ISREG(file_inode(file)->i_mode)) 1102 __sb_writers_acquired(file_inode(file)->i_sb, SB_FREEZE_WRITE); 1103 file_end_write(file); 1104 } 1105 1106 /* 1107 * Special case handling for sync iocbs: 1108 * - events go directly into the iocb for fast handling 1109 * - the sync task with the iocb in its stack holds the single iocb 1110 * ref, no other paths have a way to get another ref 1111 * - the sync task helpfully left a reference to itself in the iocb 1112 */ 1113 BUG_ON(is_sync_kiocb(kiocb)); 1114 1115 if (iocb->ki_list.next) { 1116 unsigned long flags; 1117 1118 spin_lock_irqsave(&ctx->ctx_lock, flags); 1119 list_del(&iocb->ki_list); 1120 spin_unlock_irqrestore(&ctx->ctx_lock, flags); 1121 } 1122 1123 /* 1124 * Add a completion event to the ring buffer. Must be done holding 1125 * ctx->completion_lock to prevent other code from messing with the tail 1126 * pointer since we might be called from irq context. 1127 */ 1128 spin_lock_irqsave(&ctx->completion_lock, flags); 1129 1130 tail = ctx->tail; 1131 pos = tail + AIO_EVENTS_OFFSET; 1132 1133 if (++tail >= ctx->nr_events) 1134 tail = 0; 1135 1136 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]); 1137 event = ev_page + pos % AIO_EVENTS_PER_PAGE; 1138 1139 event->obj = (u64)(unsigned long)iocb->ki_user_iocb; 1140 event->data = iocb->ki_user_data; 1141 event->res = res; 1142 event->res2 = res2; 1143 1144 kunmap_atomic(ev_page); 1145 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]); 1146 1147 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n", 1148 ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data, 1149 res, res2); 1150 1151 /* after flagging the request as done, we 1152 * must never even look at it again 1153 */ 1154 smp_wmb(); /* make event visible before updating tail */ 1155 1156 ctx->tail = tail; 1157 1158 ring = kmap_atomic(ctx->ring_pages[0]); 1159 head = ring->head; 1160 ring->tail = tail; 1161 kunmap_atomic(ring); 1162 flush_dcache_page(ctx->ring_pages[0]); 1163 1164 ctx->completed_events++; 1165 if (ctx->completed_events > 1) 1166 refill_reqs_available(ctx, head, tail); 1167 spin_unlock_irqrestore(&ctx->completion_lock, flags); 1168 1169 pr_debug("added to ring %p at [%u]\n", iocb, tail); 1170 1171 /* 1172 * Check if the user asked us to deliver the result through an 1173 * eventfd. The eventfd_signal() function is safe to be called 1174 * from IRQ context. 1175 */ 1176 if (iocb->ki_eventfd != NULL) 1177 eventfd_signal(iocb->ki_eventfd, 1); 1178 1179 /* everything turned out well, dispose of the aiocb. */ 1180 kiocb_free(iocb); 1181 1182 /* 1183 * We have to order our ring_info tail store above and test 1184 * of the wait list below outside the wait lock. This is 1185 * like in wake_up_bit() where clearing a bit has to be 1186 * ordered with the unlocked test. 1187 */ 1188 smp_mb(); 1189 1190 if (waitqueue_active(&ctx->wait)) 1191 wake_up(&ctx->wait); 1192 1193 percpu_ref_put(&ctx->reqs); 1194 } 1195 1196 /* aio_read_events_ring 1197 * Pull an event off of the ioctx's event ring. Returns the number of 1198 * events fetched 1199 */ 1200 static long aio_read_events_ring(struct kioctx *ctx, 1201 struct io_event __user *event, long nr) 1202 { 1203 struct aio_ring *ring; 1204 unsigned head, tail, pos; 1205 long ret = 0; 1206 int copy_ret; 1207 1208 /* 1209 * The mutex can block and wake us up and that will cause 1210 * wait_event_interruptible_hrtimeout() to schedule without sleeping 1211 * and repeat. This should be rare enough that it doesn't cause 1212 * peformance issues. See the comment in read_events() for more detail. 1213 */ 1214 sched_annotate_sleep(); 1215 mutex_lock(&ctx->ring_lock); 1216 1217 /* Access to ->ring_pages here is protected by ctx->ring_lock. */ 1218 ring = kmap_atomic(ctx->ring_pages[0]); 1219 head = ring->head; 1220 tail = ring->tail; 1221 kunmap_atomic(ring); 1222 1223 /* 1224 * Ensure that once we've read the current tail pointer, that 1225 * we also see the events that were stored up to the tail. 1226 */ 1227 smp_rmb(); 1228 1229 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events); 1230 1231 if (head == tail) 1232 goto out; 1233 1234 head %= ctx->nr_events; 1235 tail %= ctx->nr_events; 1236 1237 while (ret < nr) { 1238 long avail; 1239 struct io_event *ev; 1240 struct page *page; 1241 1242 avail = (head <= tail ? tail : ctx->nr_events) - head; 1243 if (head == tail) 1244 break; 1245 1246 avail = min(avail, nr - ret); 1247 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - 1248 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE)); 1249 1250 pos = head + AIO_EVENTS_OFFSET; 1251 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]; 1252 pos %= AIO_EVENTS_PER_PAGE; 1253 1254 ev = kmap(page); 1255 copy_ret = copy_to_user(event + ret, ev + pos, 1256 sizeof(*ev) * avail); 1257 kunmap(page); 1258 1259 if (unlikely(copy_ret)) { 1260 ret = -EFAULT; 1261 goto out; 1262 } 1263 1264 ret += avail; 1265 head += avail; 1266 head %= ctx->nr_events; 1267 } 1268 1269 ring = kmap_atomic(ctx->ring_pages[0]); 1270 ring->head = head; 1271 kunmap_atomic(ring); 1272 flush_dcache_page(ctx->ring_pages[0]); 1273 1274 pr_debug("%li h%u t%u\n", ret, head, tail); 1275 out: 1276 mutex_unlock(&ctx->ring_lock); 1277 1278 return ret; 1279 } 1280 1281 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr, 1282 struct io_event __user *event, long *i) 1283 { 1284 long ret = aio_read_events_ring(ctx, event + *i, nr - *i); 1285 1286 if (ret > 0) 1287 *i += ret; 1288 1289 if (unlikely(atomic_read(&ctx->dead))) 1290 ret = -EINVAL; 1291 1292 if (!*i) 1293 *i = ret; 1294 1295 return ret < 0 || *i >= min_nr; 1296 } 1297 1298 static long read_events(struct kioctx *ctx, long min_nr, long nr, 1299 struct io_event __user *event, 1300 struct timespec __user *timeout) 1301 { 1302 ktime_t until = KTIME_MAX; 1303 long ret = 0; 1304 1305 if (timeout) { 1306 struct timespec ts; 1307 1308 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts)))) 1309 return -EFAULT; 1310 1311 until = timespec_to_ktime(ts); 1312 } 1313 1314 /* 1315 * Note that aio_read_events() is being called as the conditional - i.e. 1316 * we're calling it after prepare_to_wait() has set task state to 1317 * TASK_INTERRUPTIBLE. 1318 * 1319 * But aio_read_events() can block, and if it blocks it's going to flip 1320 * the task state back to TASK_RUNNING. 1321 * 1322 * This should be ok, provided it doesn't flip the state back to 1323 * TASK_RUNNING and return 0 too much - that causes us to spin. That 1324 * will only happen if the mutex_lock() call blocks, and we then find 1325 * the ringbuffer empty. So in practice we should be ok, but it's 1326 * something to be aware of when touching this code. 1327 */ 1328 if (until == 0) 1329 aio_read_events(ctx, min_nr, nr, event, &ret); 1330 else 1331 wait_event_interruptible_hrtimeout(ctx->wait, 1332 aio_read_events(ctx, min_nr, nr, event, &ret), 1333 until); 1334 1335 if (!ret && signal_pending(current)) 1336 ret = -EINTR; 1337 1338 return ret; 1339 } 1340 1341 /* sys_io_setup: 1342 * Create an aio_context capable of receiving at least nr_events. 1343 * ctxp must not point to an aio_context that already exists, and 1344 * must be initialized to 0 prior to the call. On successful 1345 * creation of the aio_context, *ctxp is filled in with the resulting 1346 * handle. May fail with -EINVAL if *ctxp is not initialized, 1347 * if the specified nr_events exceeds internal limits. May fail 1348 * with -EAGAIN if the specified nr_events exceeds the user's limit 1349 * of available events. May fail with -ENOMEM if insufficient kernel 1350 * resources are available. May fail with -EFAULT if an invalid 1351 * pointer is passed for ctxp. Will fail with -ENOSYS if not 1352 * implemented. 1353 */ 1354 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp) 1355 { 1356 struct kioctx *ioctx = NULL; 1357 unsigned long ctx; 1358 long ret; 1359 1360 ret = get_user(ctx, ctxp); 1361 if (unlikely(ret)) 1362 goto out; 1363 1364 ret = -EINVAL; 1365 if (unlikely(ctx || nr_events == 0)) { 1366 pr_debug("EINVAL: ctx %lu nr_events %u\n", 1367 ctx, nr_events); 1368 goto out; 1369 } 1370 1371 ioctx = ioctx_alloc(nr_events); 1372 ret = PTR_ERR(ioctx); 1373 if (!IS_ERR(ioctx)) { 1374 ret = put_user(ioctx->user_id, ctxp); 1375 if (ret) 1376 kill_ioctx(current->mm, ioctx, NULL); 1377 percpu_ref_put(&ioctx->users); 1378 } 1379 1380 out: 1381 return ret; 1382 } 1383 1384 #ifdef CONFIG_COMPAT 1385 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p) 1386 { 1387 struct kioctx *ioctx = NULL; 1388 unsigned long ctx; 1389 long ret; 1390 1391 ret = get_user(ctx, ctx32p); 1392 if (unlikely(ret)) 1393 goto out; 1394 1395 ret = -EINVAL; 1396 if (unlikely(ctx || nr_events == 0)) { 1397 pr_debug("EINVAL: ctx %lu nr_events %u\n", 1398 ctx, nr_events); 1399 goto out; 1400 } 1401 1402 ioctx = ioctx_alloc(nr_events); 1403 ret = PTR_ERR(ioctx); 1404 if (!IS_ERR(ioctx)) { 1405 /* truncating is ok because it's a user address */ 1406 ret = put_user((u32)ioctx->user_id, ctx32p); 1407 if (ret) 1408 kill_ioctx(current->mm, ioctx, NULL); 1409 percpu_ref_put(&ioctx->users); 1410 } 1411 1412 out: 1413 return ret; 1414 } 1415 #endif 1416 1417 /* sys_io_destroy: 1418 * Destroy the aio_context specified. May cancel any outstanding 1419 * AIOs and block on completion. Will fail with -ENOSYS if not 1420 * implemented. May fail with -EINVAL if the context pointed to 1421 * is invalid. 1422 */ 1423 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx) 1424 { 1425 struct kioctx *ioctx = lookup_ioctx(ctx); 1426 if (likely(NULL != ioctx)) { 1427 struct ctx_rq_wait wait; 1428 int ret; 1429 1430 init_completion(&wait.comp); 1431 atomic_set(&wait.count, 1); 1432 1433 /* Pass requests_done to kill_ioctx() where it can be set 1434 * in a thread-safe way. If we try to set it here then we have 1435 * a race condition if two io_destroy() called simultaneously. 1436 */ 1437 ret = kill_ioctx(current->mm, ioctx, &wait); 1438 percpu_ref_put(&ioctx->users); 1439 1440 /* Wait until all IO for the context are done. Otherwise kernel 1441 * keep using user-space buffers even if user thinks the context 1442 * is destroyed. 1443 */ 1444 if (!ret) 1445 wait_for_completion(&wait.comp); 1446 1447 return ret; 1448 } 1449 pr_debug("EINVAL: invalid context id\n"); 1450 return -EINVAL; 1451 } 1452 1453 static int aio_setup_rw(int rw, struct iocb *iocb, struct iovec **iovec, 1454 bool vectored, bool compat, struct iov_iter *iter) 1455 { 1456 void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf; 1457 size_t len = iocb->aio_nbytes; 1458 1459 if (!vectored) { 1460 ssize_t ret = import_single_range(rw, buf, len, *iovec, iter); 1461 *iovec = NULL; 1462 return ret; 1463 } 1464 #ifdef CONFIG_COMPAT 1465 if (compat) 1466 return compat_import_iovec(rw, buf, len, UIO_FASTIOV, iovec, 1467 iter); 1468 #endif 1469 return import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter); 1470 } 1471 1472 static inline ssize_t aio_ret(struct kiocb *req, ssize_t ret) 1473 { 1474 switch (ret) { 1475 case -EIOCBQUEUED: 1476 return ret; 1477 case -ERESTARTSYS: 1478 case -ERESTARTNOINTR: 1479 case -ERESTARTNOHAND: 1480 case -ERESTART_RESTARTBLOCK: 1481 /* 1482 * There's no easy way to restart the syscall since other AIO's 1483 * may be already running. Just fail this IO with EINTR. 1484 */ 1485 ret = -EINTR; 1486 /*FALLTHRU*/ 1487 default: 1488 aio_complete(req, ret, 0); 1489 return 0; 1490 } 1491 } 1492 1493 static ssize_t aio_read(struct kiocb *req, struct iocb *iocb, bool vectored, 1494 bool compat) 1495 { 1496 struct file *file = req->ki_filp; 1497 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; 1498 struct iov_iter iter; 1499 ssize_t ret; 1500 1501 if (unlikely(!(file->f_mode & FMODE_READ))) 1502 return -EBADF; 1503 if (unlikely(!file->f_op->read_iter)) 1504 return -EINVAL; 1505 1506 ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter); 1507 if (ret) 1508 return ret; 1509 ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter)); 1510 if (!ret) 1511 ret = aio_ret(req, call_read_iter(file, req, &iter)); 1512 kfree(iovec); 1513 return ret; 1514 } 1515 1516 static ssize_t aio_write(struct kiocb *req, struct iocb *iocb, bool vectored, 1517 bool compat) 1518 { 1519 struct file *file = req->ki_filp; 1520 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; 1521 struct iov_iter iter; 1522 ssize_t ret; 1523 1524 if (unlikely(!(file->f_mode & FMODE_WRITE))) 1525 return -EBADF; 1526 if (unlikely(!file->f_op->write_iter)) 1527 return -EINVAL; 1528 1529 ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter); 1530 if (ret) 1531 return ret; 1532 ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter)); 1533 if (!ret) { 1534 req->ki_flags |= IOCB_WRITE; 1535 file_start_write(file); 1536 ret = aio_ret(req, call_write_iter(file, req, &iter)); 1537 /* 1538 * We release freeze protection in aio_complete(). Fool lockdep 1539 * by telling it the lock got released so that it doesn't 1540 * complain about held lock when we return to userspace. 1541 */ 1542 if (S_ISREG(file_inode(file)->i_mode)) 1543 __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE); 1544 } 1545 kfree(iovec); 1546 return ret; 1547 } 1548 1549 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb, 1550 struct iocb *iocb, bool compat) 1551 { 1552 struct aio_kiocb *req; 1553 struct file *file; 1554 ssize_t ret; 1555 1556 /* enforce forwards compatibility on users */ 1557 if (unlikely(iocb->aio_reserved2)) { 1558 pr_debug("EINVAL: reserve field set\n"); 1559 return -EINVAL; 1560 } 1561 1562 /* prevent overflows */ 1563 if (unlikely( 1564 (iocb->aio_buf != (unsigned long)iocb->aio_buf) || 1565 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) || 1566 ((ssize_t)iocb->aio_nbytes < 0) 1567 )) { 1568 pr_debug("EINVAL: overflow check\n"); 1569 return -EINVAL; 1570 } 1571 1572 req = aio_get_req(ctx); 1573 if (unlikely(!req)) 1574 return -EAGAIN; 1575 1576 req->common.ki_filp = file = fget(iocb->aio_fildes); 1577 if (unlikely(!req->common.ki_filp)) { 1578 ret = -EBADF; 1579 goto out_put_req; 1580 } 1581 req->common.ki_pos = iocb->aio_offset; 1582 req->common.ki_complete = aio_complete; 1583 req->common.ki_flags = iocb_flags(req->common.ki_filp); 1584 req->common.ki_hint = file_write_hint(file); 1585 1586 if (iocb->aio_flags & IOCB_FLAG_RESFD) { 1587 /* 1588 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an 1589 * instance of the file* now. The file descriptor must be 1590 * an eventfd() fd, and will be signaled for each completed 1591 * event using the eventfd_signal() function. 1592 */ 1593 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd); 1594 if (IS_ERR(req->ki_eventfd)) { 1595 ret = PTR_ERR(req->ki_eventfd); 1596 req->ki_eventfd = NULL; 1597 goto out_put_req; 1598 } 1599 1600 req->common.ki_flags |= IOCB_EVENTFD; 1601 } 1602 1603 ret = kiocb_set_rw_flags(&req->common, iocb->aio_rw_flags); 1604 if (unlikely(ret)) { 1605 pr_debug("EINVAL: aio_rw_flags\n"); 1606 goto out_put_req; 1607 } 1608 1609 ret = put_user(KIOCB_KEY, &user_iocb->aio_key); 1610 if (unlikely(ret)) { 1611 pr_debug("EFAULT: aio_key\n"); 1612 goto out_put_req; 1613 } 1614 1615 req->ki_user_iocb = user_iocb; 1616 req->ki_user_data = iocb->aio_data; 1617 1618 get_file(file); 1619 switch (iocb->aio_lio_opcode) { 1620 case IOCB_CMD_PREAD: 1621 ret = aio_read(&req->common, iocb, false, compat); 1622 break; 1623 case IOCB_CMD_PWRITE: 1624 ret = aio_write(&req->common, iocb, false, compat); 1625 break; 1626 case IOCB_CMD_PREADV: 1627 ret = aio_read(&req->common, iocb, true, compat); 1628 break; 1629 case IOCB_CMD_PWRITEV: 1630 ret = aio_write(&req->common, iocb, true, compat); 1631 break; 1632 default: 1633 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode); 1634 ret = -EINVAL; 1635 break; 1636 } 1637 fput(file); 1638 1639 if (ret && ret != -EIOCBQUEUED) 1640 goto out_put_req; 1641 return 0; 1642 out_put_req: 1643 put_reqs_available(ctx, 1); 1644 percpu_ref_put(&ctx->reqs); 1645 kiocb_free(req); 1646 return ret; 1647 } 1648 1649 static long do_io_submit(aio_context_t ctx_id, long nr, 1650 struct iocb __user *__user *iocbpp, bool compat) 1651 { 1652 struct kioctx *ctx; 1653 long ret = 0; 1654 int i = 0; 1655 struct blk_plug plug; 1656 1657 if (unlikely(nr < 0)) 1658 return -EINVAL; 1659 1660 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp))) 1661 nr = LONG_MAX/sizeof(*iocbpp); 1662 1663 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp))))) 1664 return -EFAULT; 1665 1666 ctx = lookup_ioctx(ctx_id); 1667 if (unlikely(!ctx)) { 1668 pr_debug("EINVAL: invalid context id\n"); 1669 return -EINVAL; 1670 } 1671 1672 blk_start_plug(&plug); 1673 1674 /* 1675 * AKPM: should this return a partial result if some of the IOs were 1676 * successfully submitted? 1677 */ 1678 for (i=0; i<nr; i++) { 1679 struct iocb __user *user_iocb; 1680 struct iocb tmp; 1681 1682 if (unlikely(__get_user(user_iocb, iocbpp + i))) { 1683 ret = -EFAULT; 1684 break; 1685 } 1686 1687 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) { 1688 ret = -EFAULT; 1689 break; 1690 } 1691 1692 ret = io_submit_one(ctx, user_iocb, &tmp, compat); 1693 if (ret) 1694 break; 1695 } 1696 blk_finish_plug(&plug); 1697 1698 percpu_ref_put(&ctx->users); 1699 return i ? i : ret; 1700 } 1701 1702 /* sys_io_submit: 1703 * Queue the nr iocbs pointed to by iocbpp for processing. Returns 1704 * the number of iocbs queued. May return -EINVAL if the aio_context 1705 * specified by ctx_id is invalid, if nr is < 0, if the iocb at 1706 * *iocbpp[0] is not properly initialized, if the operation specified 1707 * is invalid for the file descriptor in the iocb. May fail with 1708 * -EFAULT if any of the data structures point to invalid data. May 1709 * fail with -EBADF if the file descriptor specified in the first 1710 * iocb is invalid. May fail with -EAGAIN if insufficient resources 1711 * are available to queue any iocbs. Will return 0 if nr is 0. Will 1712 * fail with -ENOSYS if not implemented. 1713 */ 1714 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr, 1715 struct iocb __user * __user *, iocbpp) 1716 { 1717 return do_io_submit(ctx_id, nr, iocbpp, 0); 1718 } 1719 1720 #ifdef CONFIG_COMPAT 1721 static inline long 1722 copy_iocb(long nr, u32 __user *ptr32, struct iocb __user * __user *ptr64) 1723 { 1724 compat_uptr_t uptr; 1725 int i; 1726 1727 for (i = 0; i < nr; ++i) { 1728 if (get_user(uptr, ptr32 + i)) 1729 return -EFAULT; 1730 if (put_user(compat_ptr(uptr), ptr64 + i)) 1731 return -EFAULT; 1732 } 1733 return 0; 1734 } 1735 1736 #define MAX_AIO_SUBMITS (PAGE_SIZE/sizeof(struct iocb *)) 1737 1738 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id, 1739 int, nr, u32 __user *, iocb) 1740 { 1741 struct iocb __user * __user *iocb64; 1742 long ret; 1743 1744 if (unlikely(nr < 0)) 1745 return -EINVAL; 1746 1747 if (nr > MAX_AIO_SUBMITS) 1748 nr = MAX_AIO_SUBMITS; 1749 1750 iocb64 = compat_alloc_user_space(nr * sizeof(*iocb64)); 1751 ret = copy_iocb(nr, iocb, iocb64); 1752 if (!ret) 1753 ret = do_io_submit(ctx_id, nr, iocb64, 1); 1754 return ret; 1755 } 1756 #endif 1757 1758 /* lookup_kiocb 1759 * Finds a given iocb for cancellation. 1760 */ 1761 static struct aio_kiocb * 1762 lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, u32 key) 1763 { 1764 struct aio_kiocb *kiocb; 1765 1766 assert_spin_locked(&ctx->ctx_lock); 1767 1768 if (key != KIOCB_KEY) 1769 return NULL; 1770 1771 /* TODO: use a hash or array, this sucks. */ 1772 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) { 1773 if (kiocb->ki_user_iocb == iocb) 1774 return kiocb; 1775 } 1776 return NULL; 1777 } 1778 1779 /* sys_io_cancel: 1780 * Attempts to cancel an iocb previously passed to io_submit. If 1781 * the operation is successfully cancelled, the resulting event is 1782 * copied into the memory pointed to by result without being placed 1783 * into the completion queue and 0 is returned. May fail with 1784 * -EFAULT if any of the data structures pointed to are invalid. 1785 * May fail with -EINVAL if aio_context specified by ctx_id is 1786 * invalid. May fail with -EAGAIN if the iocb specified was not 1787 * cancelled. Will fail with -ENOSYS if not implemented. 1788 */ 1789 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb, 1790 struct io_event __user *, result) 1791 { 1792 struct kioctx *ctx; 1793 struct aio_kiocb *kiocb; 1794 u32 key; 1795 int ret; 1796 1797 ret = get_user(key, &iocb->aio_key); 1798 if (unlikely(ret)) 1799 return -EFAULT; 1800 1801 ctx = lookup_ioctx(ctx_id); 1802 if (unlikely(!ctx)) 1803 return -EINVAL; 1804 1805 spin_lock_irq(&ctx->ctx_lock); 1806 1807 kiocb = lookup_kiocb(ctx, iocb, key); 1808 if (kiocb) 1809 ret = kiocb_cancel(kiocb); 1810 else 1811 ret = -EINVAL; 1812 1813 spin_unlock_irq(&ctx->ctx_lock); 1814 1815 if (!ret) { 1816 /* 1817 * The result argument is no longer used - the io_event is 1818 * always delivered via the ring buffer. -EINPROGRESS indicates 1819 * cancellation is progress: 1820 */ 1821 ret = -EINPROGRESS; 1822 } 1823 1824 percpu_ref_put(&ctx->users); 1825 1826 return ret; 1827 } 1828 1829 /* io_getevents: 1830 * Attempts to read at least min_nr events and up to nr events from 1831 * the completion queue for the aio_context specified by ctx_id. If 1832 * it succeeds, the number of read events is returned. May fail with 1833 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is 1834 * out of range, if timeout is out of range. May fail with -EFAULT 1835 * if any of the memory specified is invalid. May return 0 or 1836 * < min_nr if the timeout specified by timeout has elapsed 1837 * before sufficient events are available, where timeout == NULL 1838 * specifies an infinite timeout. Note that the timeout pointed to by 1839 * timeout is relative. Will fail with -ENOSYS if not implemented. 1840 */ 1841 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id, 1842 long, min_nr, 1843 long, nr, 1844 struct io_event __user *, events, 1845 struct timespec __user *, timeout) 1846 { 1847 struct kioctx *ioctx = lookup_ioctx(ctx_id); 1848 long ret = -EINVAL; 1849 1850 if (likely(ioctx)) { 1851 if (likely(min_nr <= nr && min_nr >= 0)) 1852 ret = read_events(ioctx, min_nr, nr, events, timeout); 1853 percpu_ref_put(&ioctx->users); 1854 } 1855 return ret; 1856 } 1857 1858 #ifdef CONFIG_COMPAT 1859 COMPAT_SYSCALL_DEFINE5(io_getevents, compat_aio_context_t, ctx_id, 1860 compat_long_t, min_nr, 1861 compat_long_t, nr, 1862 struct io_event __user *, events, 1863 struct compat_timespec __user *, timeout) 1864 { 1865 struct timespec t; 1866 struct timespec __user *ut = NULL; 1867 1868 if (timeout) { 1869 if (compat_get_timespec(&t, timeout)) 1870 return -EFAULT; 1871 1872 ut = compat_alloc_user_space(sizeof(*ut)); 1873 if (copy_to_user(ut, &t, sizeof(t))) 1874 return -EFAULT; 1875 } 1876 return sys_io_getevents(ctx_id, min_nr, nr, events, ut); 1877 } 1878 #endif 1879