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.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 <asm/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 kioctx { 81 struct percpu_ref users; 82 atomic_t dead; 83 84 struct percpu_ref reqs; 85 86 unsigned long user_id; 87 88 struct __percpu kioctx_cpu *cpu; 89 90 /* 91 * For percpu reqs_available, number of slots we move to/from global 92 * counter at a time: 93 */ 94 unsigned req_batch; 95 /* 96 * This is what userspace passed to io_setup(), it's not used for 97 * anything but counting against the global max_reqs quota. 98 * 99 * The real limit is nr_events - 1, which will be larger (see 100 * aio_setup_ring()) 101 */ 102 unsigned max_reqs; 103 104 /* Size of ringbuffer, in units of struct io_event */ 105 unsigned nr_events; 106 107 unsigned long mmap_base; 108 unsigned long mmap_size; 109 110 struct page **ring_pages; 111 long nr_pages; 112 113 struct work_struct free_work; 114 115 /* 116 * signals when all in-flight requests are done 117 */ 118 struct completion *requests_done; 119 120 struct { 121 /* 122 * This counts the number of available slots in the ringbuffer, 123 * so we avoid overflowing it: it's decremented (if positive) 124 * when allocating a kiocb and incremented when the resulting 125 * io_event is pulled off the ringbuffer. 126 * 127 * We batch accesses to it with a percpu version. 128 */ 129 atomic_t reqs_available; 130 } ____cacheline_aligned_in_smp; 131 132 struct { 133 spinlock_t ctx_lock; 134 struct list_head active_reqs; /* used for cancellation */ 135 } ____cacheline_aligned_in_smp; 136 137 struct { 138 struct mutex ring_lock; 139 wait_queue_head_t wait; 140 } ____cacheline_aligned_in_smp; 141 142 struct { 143 unsigned tail; 144 unsigned completed_events; 145 spinlock_t completion_lock; 146 } ____cacheline_aligned_in_smp; 147 148 struct page *internal_pages[AIO_RING_PAGES]; 149 struct file *aio_ring_file; 150 151 unsigned id; 152 }; 153 154 /*------ sysctl variables----*/ 155 static DEFINE_SPINLOCK(aio_nr_lock); 156 unsigned long aio_nr; /* current system wide number of aio requests */ 157 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */ 158 /*----end sysctl variables---*/ 159 160 static struct kmem_cache *kiocb_cachep; 161 static struct kmem_cache *kioctx_cachep; 162 163 static struct vfsmount *aio_mnt; 164 165 static const struct file_operations aio_ring_fops; 166 static const struct address_space_operations aio_ctx_aops; 167 168 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages) 169 { 170 struct qstr this = QSTR_INIT("[aio]", 5); 171 struct file *file; 172 struct path path; 173 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb); 174 if (IS_ERR(inode)) 175 return ERR_CAST(inode); 176 177 inode->i_mapping->a_ops = &aio_ctx_aops; 178 inode->i_mapping->private_data = ctx; 179 inode->i_size = PAGE_SIZE * nr_pages; 180 181 path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this); 182 if (!path.dentry) { 183 iput(inode); 184 return ERR_PTR(-ENOMEM); 185 } 186 path.mnt = mntget(aio_mnt); 187 188 d_instantiate(path.dentry, inode); 189 file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops); 190 if (IS_ERR(file)) { 191 path_put(&path); 192 return file; 193 } 194 195 file->f_flags = O_RDWR; 196 return file; 197 } 198 199 static struct dentry *aio_mount(struct file_system_type *fs_type, 200 int flags, const char *dev_name, void *data) 201 { 202 static const struct dentry_operations ops = { 203 .d_dname = simple_dname, 204 }; 205 return mount_pseudo(fs_type, "aio:", NULL, &ops, AIO_RING_MAGIC); 206 } 207 208 /* aio_setup 209 * Creates the slab caches used by the aio routines, panic on 210 * failure as this is done early during the boot sequence. 211 */ 212 static int __init aio_setup(void) 213 { 214 static struct file_system_type aio_fs = { 215 .name = "aio", 216 .mount = aio_mount, 217 .kill_sb = kill_anon_super, 218 }; 219 aio_mnt = kern_mount(&aio_fs); 220 if (IS_ERR(aio_mnt)) 221 panic("Failed to create aio fs mount."); 222 223 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC); 224 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC); 225 226 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page)); 227 228 return 0; 229 } 230 __initcall(aio_setup); 231 232 static void put_aio_ring_file(struct kioctx *ctx) 233 { 234 struct file *aio_ring_file = ctx->aio_ring_file; 235 if (aio_ring_file) { 236 truncate_setsize(aio_ring_file->f_inode, 0); 237 238 /* Prevent further access to the kioctx from migratepages */ 239 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock); 240 aio_ring_file->f_inode->i_mapping->private_data = NULL; 241 ctx->aio_ring_file = NULL; 242 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock); 243 244 fput(aio_ring_file); 245 } 246 } 247 248 static void aio_free_ring(struct kioctx *ctx) 249 { 250 int i; 251 252 /* Disconnect the kiotx from the ring file. This prevents future 253 * accesses to the kioctx from page migration. 254 */ 255 put_aio_ring_file(ctx); 256 257 for (i = 0; i < ctx->nr_pages; i++) { 258 struct page *page; 259 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i, 260 page_count(ctx->ring_pages[i])); 261 page = ctx->ring_pages[i]; 262 if (!page) 263 continue; 264 ctx->ring_pages[i] = NULL; 265 put_page(page); 266 } 267 268 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) { 269 kfree(ctx->ring_pages); 270 ctx->ring_pages = NULL; 271 } 272 } 273 274 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma) 275 { 276 vma->vm_flags |= VM_DONTEXPAND; 277 vma->vm_ops = &generic_file_vm_ops; 278 return 0; 279 } 280 281 static void aio_ring_remap(struct file *file, struct vm_area_struct *vma) 282 { 283 struct mm_struct *mm = vma->vm_mm; 284 struct kioctx_table *table; 285 int i; 286 287 spin_lock(&mm->ioctx_lock); 288 rcu_read_lock(); 289 table = rcu_dereference(mm->ioctx_table); 290 for (i = 0; i < table->nr; i++) { 291 struct kioctx *ctx; 292 293 ctx = table->table[i]; 294 if (ctx && ctx->aio_ring_file == file) { 295 ctx->user_id = ctx->mmap_base = vma->vm_start; 296 break; 297 } 298 } 299 300 rcu_read_unlock(); 301 spin_unlock(&mm->ioctx_lock); 302 } 303 304 static const struct file_operations aio_ring_fops = { 305 .mmap = aio_ring_mmap, 306 .mremap = aio_ring_remap, 307 }; 308 309 #if IS_ENABLED(CONFIG_MIGRATION) 310 static int aio_migratepage(struct address_space *mapping, struct page *new, 311 struct page *old, enum migrate_mode mode) 312 { 313 struct kioctx *ctx; 314 unsigned long flags; 315 pgoff_t idx; 316 int rc; 317 318 rc = 0; 319 320 /* mapping->private_lock here protects against the kioctx teardown. */ 321 spin_lock(&mapping->private_lock); 322 ctx = mapping->private_data; 323 if (!ctx) { 324 rc = -EINVAL; 325 goto out; 326 } 327 328 /* The ring_lock mutex. The prevents aio_read_events() from writing 329 * to the ring's head, and prevents page migration from mucking in 330 * a partially initialized kiotx. 331 */ 332 if (!mutex_trylock(&ctx->ring_lock)) { 333 rc = -EAGAIN; 334 goto out; 335 } 336 337 idx = old->index; 338 if (idx < (pgoff_t)ctx->nr_pages) { 339 /* Make sure the old page hasn't already been changed */ 340 if (ctx->ring_pages[idx] != old) 341 rc = -EAGAIN; 342 } else 343 rc = -EINVAL; 344 345 if (rc != 0) 346 goto out_unlock; 347 348 /* Writeback must be complete */ 349 BUG_ON(PageWriteback(old)); 350 get_page(new); 351 352 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1); 353 if (rc != MIGRATEPAGE_SUCCESS) { 354 put_page(new); 355 goto out_unlock; 356 } 357 358 /* Take completion_lock to prevent other writes to the ring buffer 359 * while the old page is copied to the new. This prevents new 360 * events from being lost. 361 */ 362 spin_lock_irqsave(&ctx->completion_lock, flags); 363 migrate_page_copy(new, old); 364 BUG_ON(ctx->ring_pages[idx] != old); 365 ctx->ring_pages[idx] = new; 366 spin_unlock_irqrestore(&ctx->completion_lock, flags); 367 368 /* The old page is no longer accessible. */ 369 put_page(old); 370 371 out_unlock: 372 mutex_unlock(&ctx->ring_lock); 373 out: 374 spin_unlock(&mapping->private_lock); 375 return rc; 376 } 377 #endif 378 379 static const struct address_space_operations aio_ctx_aops = { 380 .set_page_dirty = __set_page_dirty_no_writeback, 381 #if IS_ENABLED(CONFIG_MIGRATION) 382 .migratepage = aio_migratepage, 383 #endif 384 }; 385 386 static int aio_setup_ring(struct kioctx *ctx) 387 { 388 struct aio_ring *ring; 389 unsigned nr_events = ctx->max_reqs; 390 struct mm_struct *mm = current->mm; 391 unsigned long size, unused; 392 int nr_pages; 393 int i; 394 struct file *file; 395 396 /* Compensate for the ring buffer's head/tail overlap entry */ 397 nr_events += 2; /* 1 is required, 2 for good luck */ 398 399 size = sizeof(struct aio_ring); 400 size += sizeof(struct io_event) * nr_events; 401 402 nr_pages = PFN_UP(size); 403 if (nr_pages < 0) 404 return -EINVAL; 405 406 file = aio_private_file(ctx, nr_pages); 407 if (IS_ERR(file)) { 408 ctx->aio_ring_file = NULL; 409 return -ENOMEM; 410 } 411 412 ctx->aio_ring_file = file; 413 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) 414 / sizeof(struct io_event); 415 416 ctx->ring_pages = ctx->internal_pages; 417 if (nr_pages > AIO_RING_PAGES) { 418 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *), 419 GFP_KERNEL); 420 if (!ctx->ring_pages) { 421 put_aio_ring_file(ctx); 422 return -ENOMEM; 423 } 424 } 425 426 for (i = 0; i < nr_pages; i++) { 427 struct page *page; 428 page = find_or_create_page(file->f_inode->i_mapping, 429 i, GFP_HIGHUSER | __GFP_ZERO); 430 if (!page) 431 break; 432 pr_debug("pid(%d) page[%d]->count=%d\n", 433 current->pid, i, page_count(page)); 434 SetPageUptodate(page); 435 unlock_page(page); 436 437 ctx->ring_pages[i] = page; 438 } 439 ctx->nr_pages = i; 440 441 if (unlikely(i != nr_pages)) { 442 aio_free_ring(ctx); 443 return -ENOMEM; 444 } 445 446 ctx->mmap_size = nr_pages * PAGE_SIZE; 447 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size); 448 449 down_write(&mm->mmap_sem); 450 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size, 451 PROT_READ | PROT_WRITE, 452 MAP_SHARED, 0, &unused); 453 up_write(&mm->mmap_sem); 454 if (IS_ERR((void *)ctx->mmap_base)) { 455 ctx->mmap_size = 0; 456 aio_free_ring(ctx); 457 return -ENOMEM; 458 } 459 460 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base); 461 462 ctx->user_id = ctx->mmap_base; 463 ctx->nr_events = nr_events; /* trusted copy */ 464 465 ring = kmap_atomic(ctx->ring_pages[0]); 466 ring->nr = nr_events; /* user copy */ 467 ring->id = ~0U; 468 ring->head = ring->tail = 0; 469 ring->magic = AIO_RING_MAGIC; 470 ring->compat_features = AIO_RING_COMPAT_FEATURES; 471 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES; 472 ring->header_length = sizeof(struct aio_ring); 473 kunmap_atomic(ring); 474 flush_dcache_page(ctx->ring_pages[0]); 475 476 return 0; 477 } 478 479 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event)) 480 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event)) 481 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE) 482 483 void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel) 484 { 485 struct kioctx *ctx = req->ki_ctx; 486 unsigned long flags; 487 488 spin_lock_irqsave(&ctx->ctx_lock, flags); 489 490 if (!req->ki_list.next) 491 list_add(&req->ki_list, &ctx->active_reqs); 492 493 req->ki_cancel = cancel; 494 495 spin_unlock_irqrestore(&ctx->ctx_lock, flags); 496 } 497 EXPORT_SYMBOL(kiocb_set_cancel_fn); 498 499 static int kiocb_cancel(struct kiocb *kiocb) 500 { 501 kiocb_cancel_fn *old, *cancel; 502 503 /* 504 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it 505 * actually has a cancel function, hence the cmpxchg() 506 */ 507 508 cancel = ACCESS_ONCE(kiocb->ki_cancel); 509 do { 510 if (!cancel || cancel == KIOCB_CANCELLED) 511 return -EINVAL; 512 513 old = cancel; 514 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED); 515 } while (cancel != old); 516 517 return cancel(kiocb); 518 } 519 520 static void free_ioctx(struct work_struct *work) 521 { 522 struct kioctx *ctx = container_of(work, struct kioctx, free_work); 523 524 pr_debug("freeing %p\n", ctx); 525 526 aio_free_ring(ctx); 527 free_percpu(ctx->cpu); 528 percpu_ref_exit(&ctx->reqs); 529 percpu_ref_exit(&ctx->users); 530 kmem_cache_free(kioctx_cachep, ctx); 531 } 532 533 static void free_ioctx_reqs(struct percpu_ref *ref) 534 { 535 struct kioctx *ctx = container_of(ref, struct kioctx, reqs); 536 537 /* At this point we know that there are no any in-flight requests */ 538 if (ctx->requests_done) 539 complete(ctx->requests_done); 540 541 INIT_WORK(&ctx->free_work, free_ioctx); 542 schedule_work(&ctx->free_work); 543 } 544 545 /* 546 * When this function runs, the kioctx has been removed from the "hash table" 547 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted - 548 * now it's safe to cancel any that need to be. 549 */ 550 static void free_ioctx_users(struct percpu_ref *ref) 551 { 552 struct kioctx *ctx = container_of(ref, struct kioctx, users); 553 struct kiocb *req; 554 555 spin_lock_irq(&ctx->ctx_lock); 556 557 while (!list_empty(&ctx->active_reqs)) { 558 req = list_first_entry(&ctx->active_reqs, 559 struct kiocb, ki_list); 560 561 list_del_init(&req->ki_list); 562 kiocb_cancel(req); 563 } 564 565 spin_unlock_irq(&ctx->ctx_lock); 566 567 percpu_ref_kill(&ctx->reqs); 568 percpu_ref_put(&ctx->reqs); 569 } 570 571 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm) 572 { 573 unsigned i, new_nr; 574 struct kioctx_table *table, *old; 575 struct aio_ring *ring; 576 577 spin_lock(&mm->ioctx_lock); 578 table = rcu_dereference_raw(mm->ioctx_table); 579 580 while (1) { 581 if (table) 582 for (i = 0; i < table->nr; i++) 583 if (!table->table[i]) { 584 ctx->id = i; 585 table->table[i] = ctx; 586 spin_unlock(&mm->ioctx_lock); 587 588 /* While kioctx setup is in progress, 589 * we are protected from page migration 590 * changes ring_pages by ->ring_lock. 591 */ 592 ring = kmap_atomic(ctx->ring_pages[0]); 593 ring->id = ctx->id; 594 kunmap_atomic(ring); 595 return 0; 596 } 597 598 new_nr = (table ? table->nr : 1) * 4; 599 spin_unlock(&mm->ioctx_lock); 600 601 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) * 602 new_nr, GFP_KERNEL); 603 if (!table) 604 return -ENOMEM; 605 606 table->nr = new_nr; 607 608 spin_lock(&mm->ioctx_lock); 609 old = rcu_dereference_raw(mm->ioctx_table); 610 611 if (!old) { 612 rcu_assign_pointer(mm->ioctx_table, table); 613 } else if (table->nr > old->nr) { 614 memcpy(table->table, old->table, 615 old->nr * sizeof(struct kioctx *)); 616 617 rcu_assign_pointer(mm->ioctx_table, table); 618 kfree_rcu(old, rcu); 619 } else { 620 kfree(table); 621 table = old; 622 } 623 } 624 } 625 626 static void aio_nr_sub(unsigned nr) 627 { 628 spin_lock(&aio_nr_lock); 629 if (WARN_ON(aio_nr - nr > aio_nr)) 630 aio_nr = 0; 631 else 632 aio_nr -= nr; 633 spin_unlock(&aio_nr_lock); 634 } 635 636 /* ioctx_alloc 637 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed. 638 */ 639 static struct kioctx *ioctx_alloc(unsigned nr_events) 640 { 641 struct mm_struct *mm = current->mm; 642 struct kioctx *ctx; 643 int err = -ENOMEM; 644 645 /* 646 * We keep track of the number of available ringbuffer slots, to prevent 647 * overflow (reqs_available), and we also use percpu counters for this. 648 * 649 * So since up to half the slots might be on other cpu's percpu counters 650 * and unavailable, double nr_events so userspace sees what they 651 * expected: additionally, we move req_batch slots to/from percpu 652 * counters at a time, so make sure that isn't 0: 653 */ 654 nr_events = max(nr_events, num_possible_cpus() * 4); 655 nr_events *= 2; 656 657 /* Prevent overflows */ 658 if ((nr_events > (0x10000000U / sizeof(struct io_event))) || 659 (nr_events > (0x10000000U / sizeof(struct kiocb)))) { 660 pr_debug("ENOMEM: nr_events too high\n"); 661 return ERR_PTR(-EINVAL); 662 } 663 664 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL)) 665 return ERR_PTR(-EAGAIN); 666 667 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL); 668 if (!ctx) 669 return ERR_PTR(-ENOMEM); 670 671 ctx->max_reqs = nr_events; 672 673 spin_lock_init(&ctx->ctx_lock); 674 spin_lock_init(&ctx->completion_lock); 675 mutex_init(&ctx->ring_lock); 676 /* Protect against page migration throughout kiotx setup by keeping 677 * the ring_lock mutex held until setup is complete. */ 678 mutex_lock(&ctx->ring_lock); 679 init_waitqueue_head(&ctx->wait); 680 681 INIT_LIST_HEAD(&ctx->active_reqs); 682 683 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL)) 684 goto err; 685 686 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL)) 687 goto err; 688 689 ctx->cpu = alloc_percpu(struct kioctx_cpu); 690 if (!ctx->cpu) 691 goto err; 692 693 err = aio_setup_ring(ctx); 694 if (err < 0) 695 goto err; 696 697 atomic_set(&ctx->reqs_available, ctx->nr_events - 1); 698 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4); 699 if (ctx->req_batch < 1) 700 ctx->req_batch = 1; 701 702 /* limit the number of system wide aios */ 703 spin_lock(&aio_nr_lock); 704 if (aio_nr + nr_events > (aio_max_nr * 2UL) || 705 aio_nr + nr_events < aio_nr) { 706 spin_unlock(&aio_nr_lock); 707 err = -EAGAIN; 708 goto err_ctx; 709 } 710 aio_nr += ctx->max_reqs; 711 spin_unlock(&aio_nr_lock); 712 713 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */ 714 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */ 715 716 err = ioctx_add_table(ctx, mm); 717 if (err) 718 goto err_cleanup; 719 720 /* Release the ring_lock mutex now that all setup is complete. */ 721 mutex_unlock(&ctx->ring_lock); 722 723 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n", 724 ctx, ctx->user_id, mm, ctx->nr_events); 725 return ctx; 726 727 err_cleanup: 728 aio_nr_sub(ctx->max_reqs); 729 err_ctx: 730 aio_free_ring(ctx); 731 err: 732 mutex_unlock(&ctx->ring_lock); 733 free_percpu(ctx->cpu); 734 percpu_ref_exit(&ctx->reqs); 735 percpu_ref_exit(&ctx->users); 736 kmem_cache_free(kioctx_cachep, ctx); 737 pr_debug("error allocating ioctx %d\n", err); 738 return ERR_PTR(err); 739 } 740 741 /* kill_ioctx 742 * Cancels all outstanding aio requests on an aio context. Used 743 * when the processes owning a context have all exited to encourage 744 * the rapid destruction of the kioctx. 745 */ 746 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx, 747 struct completion *requests_done) 748 { 749 struct kioctx_table *table; 750 751 if (atomic_xchg(&ctx->dead, 1)) 752 return -EINVAL; 753 754 755 spin_lock(&mm->ioctx_lock); 756 table = rcu_dereference_raw(mm->ioctx_table); 757 WARN_ON(ctx != table->table[ctx->id]); 758 table->table[ctx->id] = NULL; 759 spin_unlock(&mm->ioctx_lock); 760 761 /* percpu_ref_kill() will do the necessary call_rcu() */ 762 wake_up_all(&ctx->wait); 763 764 /* 765 * It'd be more correct to do this in free_ioctx(), after all 766 * the outstanding kiocbs have finished - but by then io_destroy 767 * has already returned, so io_setup() could potentially return 768 * -EAGAIN with no ioctxs actually in use (as far as userspace 769 * could tell). 770 */ 771 aio_nr_sub(ctx->max_reqs); 772 773 if (ctx->mmap_size) 774 vm_munmap(ctx->mmap_base, ctx->mmap_size); 775 776 ctx->requests_done = requests_done; 777 percpu_ref_kill(&ctx->users); 778 return 0; 779 } 780 781 /* wait_on_sync_kiocb: 782 * Waits on the given sync kiocb to complete. 783 */ 784 ssize_t wait_on_sync_kiocb(struct kiocb *req) 785 { 786 while (!req->ki_ctx) { 787 set_current_state(TASK_UNINTERRUPTIBLE); 788 if (req->ki_ctx) 789 break; 790 io_schedule(); 791 } 792 __set_current_state(TASK_RUNNING); 793 return req->ki_user_data; 794 } 795 EXPORT_SYMBOL(wait_on_sync_kiocb); 796 797 /* 798 * exit_aio: called when the last user of mm goes away. At this point, there is 799 * no way for any new requests to be submited or any of the io_* syscalls to be 800 * called on the context. 801 * 802 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on 803 * them. 804 */ 805 void exit_aio(struct mm_struct *mm) 806 { 807 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table); 808 int i; 809 810 if (!table) 811 return; 812 813 for (i = 0; i < table->nr; ++i) { 814 struct kioctx *ctx = table->table[i]; 815 struct completion requests_done = 816 COMPLETION_INITIALIZER_ONSTACK(requests_done); 817 818 if (!ctx) 819 continue; 820 /* 821 * We don't need to bother with munmap() here - exit_mmap(mm) 822 * is coming and it'll unmap everything. And we simply can't, 823 * this is not necessarily our ->mm. 824 * Since kill_ioctx() uses non-zero ->mmap_size as indicator 825 * that it needs to unmap the area, just set it to 0. 826 */ 827 ctx->mmap_size = 0; 828 kill_ioctx(mm, ctx, &requests_done); 829 830 /* Wait until all IO for the context are done. */ 831 wait_for_completion(&requests_done); 832 } 833 834 RCU_INIT_POINTER(mm->ioctx_table, NULL); 835 kfree(table); 836 } 837 838 static void put_reqs_available(struct kioctx *ctx, unsigned nr) 839 { 840 struct kioctx_cpu *kcpu; 841 unsigned long flags; 842 843 local_irq_save(flags); 844 kcpu = this_cpu_ptr(ctx->cpu); 845 kcpu->reqs_available += nr; 846 847 while (kcpu->reqs_available >= ctx->req_batch * 2) { 848 kcpu->reqs_available -= ctx->req_batch; 849 atomic_add(ctx->req_batch, &ctx->reqs_available); 850 } 851 852 local_irq_restore(flags); 853 } 854 855 static bool get_reqs_available(struct kioctx *ctx) 856 { 857 struct kioctx_cpu *kcpu; 858 bool ret = false; 859 unsigned long flags; 860 861 local_irq_save(flags); 862 kcpu = this_cpu_ptr(ctx->cpu); 863 if (!kcpu->reqs_available) { 864 int old, avail = atomic_read(&ctx->reqs_available); 865 866 do { 867 if (avail < ctx->req_batch) 868 goto out; 869 870 old = avail; 871 avail = atomic_cmpxchg(&ctx->reqs_available, 872 avail, avail - ctx->req_batch); 873 } while (avail != old); 874 875 kcpu->reqs_available += ctx->req_batch; 876 } 877 878 ret = true; 879 kcpu->reqs_available--; 880 out: 881 local_irq_restore(flags); 882 return ret; 883 } 884 885 /* refill_reqs_available 886 * Updates the reqs_available reference counts used for tracking the 887 * number of free slots in the completion ring. This can be called 888 * from aio_complete() (to optimistically update reqs_available) or 889 * from aio_get_req() (the we're out of events case). It must be 890 * called holding ctx->completion_lock. 891 */ 892 static void refill_reqs_available(struct kioctx *ctx, unsigned head, 893 unsigned tail) 894 { 895 unsigned events_in_ring, completed; 896 897 /* Clamp head since userland can write to it. */ 898 head %= ctx->nr_events; 899 if (head <= tail) 900 events_in_ring = tail - head; 901 else 902 events_in_ring = ctx->nr_events - (head - tail); 903 904 completed = ctx->completed_events; 905 if (events_in_ring < completed) 906 completed -= events_in_ring; 907 else 908 completed = 0; 909 910 if (!completed) 911 return; 912 913 ctx->completed_events -= completed; 914 put_reqs_available(ctx, completed); 915 } 916 917 /* user_refill_reqs_available 918 * Called to refill reqs_available when aio_get_req() encounters an 919 * out of space in the completion ring. 920 */ 921 static void user_refill_reqs_available(struct kioctx *ctx) 922 { 923 spin_lock_irq(&ctx->completion_lock); 924 if (ctx->completed_events) { 925 struct aio_ring *ring; 926 unsigned head; 927 928 /* Access of ring->head may race with aio_read_events_ring() 929 * here, but that's okay since whether we read the old version 930 * or the new version, and either will be valid. The important 931 * part is that head cannot pass tail since we prevent 932 * aio_complete() from updating tail by holding 933 * ctx->completion_lock. Even if head is invalid, the check 934 * against ctx->completed_events below will make sure we do the 935 * safe/right thing. 936 */ 937 ring = kmap_atomic(ctx->ring_pages[0]); 938 head = ring->head; 939 kunmap_atomic(ring); 940 941 refill_reqs_available(ctx, head, ctx->tail); 942 } 943 944 spin_unlock_irq(&ctx->completion_lock); 945 } 946 947 /* aio_get_req 948 * Allocate a slot for an aio request. 949 * Returns NULL if no requests are free. 950 */ 951 static inline struct kiocb *aio_get_req(struct kioctx *ctx) 952 { 953 struct kiocb *req; 954 955 if (!get_reqs_available(ctx)) { 956 user_refill_reqs_available(ctx); 957 if (!get_reqs_available(ctx)) 958 return NULL; 959 } 960 961 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO); 962 if (unlikely(!req)) 963 goto out_put; 964 965 percpu_ref_get(&ctx->reqs); 966 967 req->ki_ctx = ctx; 968 return req; 969 out_put: 970 put_reqs_available(ctx, 1); 971 return NULL; 972 } 973 974 static void kiocb_free(struct kiocb *req) 975 { 976 if (req->ki_filp) 977 fput(req->ki_filp); 978 if (req->ki_eventfd != NULL) 979 eventfd_ctx_put(req->ki_eventfd); 980 kmem_cache_free(kiocb_cachep, req); 981 } 982 983 static struct kioctx *lookup_ioctx(unsigned long ctx_id) 984 { 985 struct aio_ring __user *ring = (void __user *)ctx_id; 986 struct mm_struct *mm = current->mm; 987 struct kioctx *ctx, *ret = NULL; 988 struct kioctx_table *table; 989 unsigned id; 990 991 if (get_user(id, &ring->id)) 992 return NULL; 993 994 rcu_read_lock(); 995 table = rcu_dereference(mm->ioctx_table); 996 997 if (!table || id >= table->nr) 998 goto out; 999 1000 ctx = table->table[id]; 1001 if (ctx && ctx->user_id == ctx_id) { 1002 percpu_ref_get(&ctx->users); 1003 ret = ctx; 1004 } 1005 out: 1006 rcu_read_unlock(); 1007 return ret; 1008 } 1009 1010 /* aio_complete 1011 * Called when the io request on the given iocb is complete. 1012 */ 1013 void aio_complete(struct kiocb *iocb, long res, long res2) 1014 { 1015 struct kioctx *ctx = iocb->ki_ctx; 1016 struct aio_ring *ring; 1017 struct io_event *ev_page, *event; 1018 unsigned tail, pos, head; 1019 unsigned long flags; 1020 1021 /* 1022 * Special case handling for sync iocbs: 1023 * - events go directly into the iocb for fast handling 1024 * - the sync task with the iocb in its stack holds the single iocb 1025 * ref, no other paths have a way to get another ref 1026 * - the sync task helpfully left a reference to itself in the iocb 1027 */ 1028 if (is_sync_kiocb(iocb)) { 1029 iocb->ki_user_data = res; 1030 smp_wmb(); 1031 iocb->ki_ctx = ERR_PTR(-EXDEV); 1032 wake_up_process(iocb->ki_obj.tsk); 1033 return; 1034 } 1035 1036 if (iocb->ki_list.next) { 1037 unsigned long flags; 1038 1039 spin_lock_irqsave(&ctx->ctx_lock, flags); 1040 list_del(&iocb->ki_list); 1041 spin_unlock_irqrestore(&ctx->ctx_lock, flags); 1042 } 1043 1044 /* 1045 * Add a completion event to the ring buffer. Must be done holding 1046 * ctx->completion_lock to prevent other code from messing with the tail 1047 * pointer since we might be called from irq context. 1048 */ 1049 spin_lock_irqsave(&ctx->completion_lock, flags); 1050 1051 tail = ctx->tail; 1052 pos = tail + AIO_EVENTS_OFFSET; 1053 1054 if (++tail >= ctx->nr_events) 1055 tail = 0; 1056 1057 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]); 1058 event = ev_page + pos % AIO_EVENTS_PER_PAGE; 1059 1060 event->obj = (u64)(unsigned long)iocb->ki_obj.user; 1061 event->data = iocb->ki_user_data; 1062 event->res = res; 1063 event->res2 = res2; 1064 1065 kunmap_atomic(ev_page); 1066 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]); 1067 1068 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n", 1069 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data, 1070 res, res2); 1071 1072 /* after flagging the request as done, we 1073 * must never even look at it again 1074 */ 1075 smp_wmb(); /* make event visible before updating tail */ 1076 1077 ctx->tail = tail; 1078 1079 ring = kmap_atomic(ctx->ring_pages[0]); 1080 head = ring->head; 1081 ring->tail = tail; 1082 kunmap_atomic(ring); 1083 flush_dcache_page(ctx->ring_pages[0]); 1084 1085 ctx->completed_events++; 1086 if (ctx->completed_events > 1) 1087 refill_reqs_available(ctx, head, tail); 1088 spin_unlock_irqrestore(&ctx->completion_lock, flags); 1089 1090 pr_debug("added to ring %p at [%u]\n", iocb, tail); 1091 1092 /* 1093 * Check if the user asked us to deliver the result through an 1094 * eventfd. The eventfd_signal() function is safe to be called 1095 * from IRQ context. 1096 */ 1097 if (iocb->ki_eventfd != NULL) 1098 eventfd_signal(iocb->ki_eventfd, 1); 1099 1100 /* everything turned out well, dispose of the aiocb. */ 1101 kiocb_free(iocb); 1102 1103 /* 1104 * We have to order our ring_info tail store above and test 1105 * of the wait list below outside the wait lock. This is 1106 * like in wake_up_bit() where clearing a bit has to be 1107 * ordered with the unlocked test. 1108 */ 1109 smp_mb(); 1110 1111 if (waitqueue_active(&ctx->wait)) 1112 wake_up(&ctx->wait); 1113 1114 percpu_ref_put(&ctx->reqs); 1115 } 1116 EXPORT_SYMBOL(aio_complete); 1117 1118 /* aio_read_events_ring 1119 * Pull an event off of the ioctx's event ring. Returns the number of 1120 * events fetched 1121 */ 1122 static long aio_read_events_ring(struct kioctx *ctx, 1123 struct io_event __user *event, long nr) 1124 { 1125 struct aio_ring *ring; 1126 unsigned head, tail, pos; 1127 long ret = 0; 1128 int copy_ret; 1129 1130 /* 1131 * The mutex can block and wake us up and that will cause 1132 * wait_event_interruptible_hrtimeout() to schedule without sleeping 1133 * and repeat. This should be rare enough that it doesn't cause 1134 * peformance issues. See the comment in read_events() for more detail. 1135 */ 1136 sched_annotate_sleep(); 1137 mutex_lock(&ctx->ring_lock); 1138 1139 /* Access to ->ring_pages here is protected by ctx->ring_lock. */ 1140 ring = kmap_atomic(ctx->ring_pages[0]); 1141 head = ring->head; 1142 tail = ring->tail; 1143 kunmap_atomic(ring); 1144 1145 /* 1146 * Ensure that once we've read the current tail pointer, that 1147 * we also see the events that were stored up to the tail. 1148 */ 1149 smp_rmb(); 1150 1151 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events); 1152 1153 if (head == tail) 1154 goto out; 1155 1156 head %= ctx->nr_events; 1157 tail %= ctx->nr_events; 1158 1159 while (ret < nr) { 1160 long avail; 1161 struct io_event *ev; 1162 struct page *page; 1163 1164 avail = (head <= tail ? tail : ctx->nr_events) - head; 1165 if (head == tail) 1166 break; 1167 1168 avail = min(avail, nr - ret); 1169 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - 1170 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE)); 1171 1172 pos = head + AIO_EVENTS_OFFSET; 1173 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]; 1174 pos %= AIO_EVENTS_PER_PAGE; 1175 1176 ev = kmap(page); 1177 copy_ret = copy_to_user(event + ret, ev + pos, 1178 sizeof(*ev) * avail); 1179 kunmap(page); 1180 1181 if (unlikely(copy_ret)) { 1182 ret = -EFAULT; 1183 goto out; 1184 } 1185 1186 ret += avail; 1187 head += avail; 1188 head %= ctx->nr_events; 1189 } 1190 1191 ring = kmap_atomic(ctx->ring_pages[0]); 1192 ring->head = head; 1193 kunmap_atomic(ring); 1194 flush_dcache_page(ctx->ring_pages[0]); 1195 1196 pr_debug("%li h%u t%u\n", ret, head, tail); 1197 out: 1198 mutex_unlock(&ctx->ring_lock); 1199 1200 return ret; 1201 } 1202 1203 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr, 1204 struct io_event __user *event, long *i) 1205 { 1206 long ret = aio_read_events_ring(ctx, event + *i, nr - *i); 1207 1208 if (ret > 0) 1209 *i += ret; 1210 1211 if (unlikely(atomic_read(&ctx->dead))) 1212 ret = -EINVAL; 1213 1214 if (!*i) 1215 *i = ret; 1216 1217 return ret < 0 || *i >= min_nr; 1218 } 1219 1220 static long read_events(struct kioctx *ctx, long min_nr, long nr, 1221 struct io_event __user *event, 1222 struct timespec __user *timeout) 1223 { 1224 ktime_t until = { .tv64 = KTIME_MAX }; 1225 long ret = 0; 1226 1227 if (timeout) { 1228 struct timespec ts; 1229 1230 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts)))) 1231 return -EFAULT; 1232 1233 until = timespec_to_ktime(ts); 1234 } 1235 1236 /* 1237 * Note that aio_read_events() is being called as the conditional - i.e. 1238 * we're calling it after prepare_to_wait() has set task state to 1239 * TASK_INTERRUPTIBLE. 1240 * 1241 * But aio_read_events() can block, and if it blocks it's going to flip 1242 * the task state back to TASK_RUNNING. 1243 * 1244 * This should be ok, provided it doesn't flip the state back to 1245 * TASK_RUNNING and return 0 too much - that causes us to spin. That 1246 * will only happen if the mutex_lock() call blocks, and we then find 1247 * the ringbuffer empty. So in practice we should be ok, but it's 1248 * something to be aware of when touching this code. 1249 */ 1250 if (until.tv64 == 0) 1251 aio_read_events(ctx, min_nr, nr, event, &ret); 1252 else 1253 wait_event_interruptible_hrtimeout(ctx->wait, 1254 aio_read_events(ctx, min_nr, nr, event, &ret), 1255 until); 1256 1257 if (!ret && signal_pending(current)) 1258 ret = -EINTR; 1259 1260 return ret; 1261 } 1262 1263 /* sys_io_setup: 1264 * Create an aio_context capable of receiving at least nr_events. 1265 * ctxp must not point to an aio_context that already exists, and 1266 * must be initialized to 0 prior to the call. On successful 1267 * creation of the aio_context, *ctxp is filled in with the resulting 1268 * handle. May fail with -EINVAL if *ctxp is not initialized, 1269 * if the specified nr_events exceeds internal limits. May fail 1270 * with -EAGAIN if the specified nr_events exceeds the user's limit 1271 * of available events. May fail with -ENOMEM if insufficient kernel 1272 * resources are available. May fail with -EFAULT if an invalid 1273 * pointer is passed for ctxp. Will fail with -ENOSYS if not 1274 * implemented. 1275 */ 1276 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp) 1277 { 1278 struct kioctx *ioctx = NULL; 1279 unsigned long ctx; 1280 long ret; 1281 1282 ret = get_user(ctx, ctxp); 1283 if (unlikely(ret)) 1284 goto out; 1285 1286 ret = -EINVAL; 1287 if (unlikely(ctx || nr_events == 0)) { 1288 pr_debug("EINVAL: ctx %lu nr_events %u\n", 1289 ctx, nr_events); 1290 goto out; 1291 } 1292 1293 ioctx = ioctx_alloc(nr_events); 1294 ret = PTR_ERR(ioctx); 1295 if (!IS_ERR(ioctx)) { 1296 ret = put_user(ioctx->user_id, ctxp); 1297 if (ret) 1298 kill_ioctx(current->mm, ioctx, NULL); 1299 percpu_ref_put(&ioctx->users); 1300 } 1301 1302 out: 1303 return ret; 1304 } 1305 1306 /* sys_io_destroy: 1307 * Destroy the aio_context specified. May cancel any outstanding 1308 * AIOs and block on completion. Will fail with -ENOSYS if not 1309 * implemented. May fail with -EINVAL if the context pointed to 1310 * is invalid. 1311 */ 1312 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx) 1313 { 1314 struct kioctx *ioctx = lookup_ioctx(ctx); 1315 if (likely(NULL != ioctx)) { 1316 struct completion requests_done = 1317 COMPLETION_INITIALIZER_ONSTACK(requests_done); 1318 int ret; 1319 1320 /* Pass requests_done to kill_ioctx() where it can be set 1321 * in a thread-safe way. If we try to set it here then we have 1322 * a race condition if two io_destroy() called simultaneously. 1323 */ 1324 ret = kill_ioctx(current->mm, ioctx, &requests_done); 1325 percpu_ref_put(&ioctx->users); 1326 1327 /* Wait until all IO for the context are done. Otherwise kernel 1328 * keep using user-space buffers even if user thinks the context 1329 * is destroyed. 1330 */ 1331 if (!ret) 1332 wait_for_completion(&requests_done); 1333 1334 return ret; 1335 } 1336 pr_debug("EINVAL: invalid context id\n"); 1337 return -EINVAL; 1338 } 1339 1340 typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *, 1341 unsigned long, loff_t); 1342 typedef ssize_t (rw_iter_op)(struct kiocb *, struct iov_iter *); 1343 1344 static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb, 1345 int rw, char __user *buf, 1346 unsigned long *nr_segs, 1347 struct iovec **iovec, 1348 bool compat) 1349 { 1350 ssize_t ret; 1351 1352 *nr_segs = kiocb->ki_nbytes; 1353 1354 #ifdef CONFIG_COMPAT 1355 if (compat) 1356 ret = compat_rw_copy_check_uvector(rw, 1357 (struct compat_iovec __user *)buf, 1358 *nr_segs, UIO_FASTIOV, *iovec, iovec); 1359 else 1360 #endif 1361 ret = rw_copy_check_uvector(rw, 1362 (struct iovec __user *)buf, 1363 *nr_segs, UIO_FASTIOV, *iovec, iovec); 1364 if (ret < 0) 1365 return ret; 1366 1367 /* ki_nbytes now reflect bytes instead of segs */ 1368 kiocb->ki_nbytes = ret; 1369 return 0; 1370 } 1371 1372 static ssize_t aio_setup_single_vector(struct kiocb *kiocb, 1373 int rw, char __user *buf, 1374 unsigned long *nr_segs, 1375 struct iovec *iovec) 1376 { 1377 if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes))) 1378 return -EFAULT; 1379 1380 iovec->iov_base = buf; 1381 iovec->iov_len = kiocb->ki_nbytes; 1382 *nr_segs = 1; 1383 return 0; 1384 } 1385 1386 /* 1387 * aio_run_iocb: 1388 * Performs the initial checks and io submission. 1389 */ 1390 static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode, 1391 char __user *buf, bool compat) 1392 { 1393 struct file *file = req->ki_filp; 1394 ssize_t ret; 1395 unsigned long nr_segs; 1396 int rw; 1397 fmode_t mode; 1398 aio_rw_op *rw_op; 1399 rw_iter_op *iter_op; 1400 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; 1401 struct iov_iter iter; 1402 1403 switch (opcode) { 1404 case IOCB_CMD_PREAD: 1405 case IOCB_CMD_PREADV: 1406 mode = FMODE_READ; 1407 rw = READ; 1408 rw_op = file->f_op->aio_read; 1409 iter_op = file->f_op->read_iter; 1410 goto rw_common; 1411 1412 case IOCB_CMD_PWRITE: 1413 case IOCB_CMD_PWRITEV: 1414 mode = FMODE_WRITE; 1415 rw = WRITE; 1416 rw_op = file->f_op->aio_write; 1417 iter_op = file->f_op->write_iter; 1418 goto rw_common; 1419 rw_common: 1420 if (unlikely(!(file->f_mode & mode))) 1421 return -EBADF; 1422 1423 if (!rw_op && !iter_op) 1424 return -EINVAL; 1425 1426 ret = (opcode == IOCB_CMD_PREADV || 1427 opcode == IOCB_CMD_PWRITEV) 1428 ? aio_setup_vectored_rw(req, rw, buf, &nr_segs, 1429 &iovec, compat) 1430 : aio_setup_single_vector(req, rw, buf, &nr_segs, 1431 iovec); 1432 if (!ret) 1433 ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes); 1434 if (ret < 0) { 1435 if (iovec != inline_vecs) 1436 kfree(iovec); 1437 return ret; 1438 } 1439 1440 req->ki_nbytes = ret; 1441 1442 /* XXX: move/kill - rw_verify_area()? */ 1443 /* This matches the pread()/pwrite() logic */ 1444 if (req->ki_pos < 0) { 1445 ret = -EINVAL; 1446 break; 1447 } 1448 1449 if (rw == WRITE) 1450 file_start_write(file); 1451 1452 if (iter_op) { 1453 iov_iter_init(&iter, rw, iovec, nr_segs, req->ki_nbytes); 1454 ret = iter_op(req, &iter); 1455 } else { 1456 ret = rw_op(req, iovec, nr_segs, req->ki_pos); 1457 } 1458 1459 if (rw == WRITE) 1460 file_end_write(file); 1461 break; 1462 1463 case IOCB_CMD_FDSYNC: 1464 if (!file->f_op->aio_fsync) 1465 return -EINVAL; 1466 1467 ret = file->f_op->aio_fsync(req, 1); 1468 break; 1469 1470 case IOCB_CMD_FSYNC: 1471 if (!file->f_op->aio_fsync) 1472 return -EINVAL; 1473 1474 ret = file->f_op->aio_fsync(req, 0); 1475 break; 1476 1477 default: 1478 pr_debug("EINVAL: no operation provided\n"); 1479 return -EINVAL; 1480 } 1481 1482 if (iovec != inline_vecs) 1483 kfree(iovec); 1484 1485 if (ret != -EIOCBQUEUED) { 1486 /* 1487 * There's no easy way to restart the syscall since other AIO's 1488 * may be already running. Just fail this IO with EINTR. 1489 */ 1490 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR || 1491 ret == -ERESTARTNOHAND || 1492 ret == -ERESTART_RESTARTBLOCK)) 1493 ret = -EINTR; 1494 aio_complete(req, ret, 0); 1495 } 1496 1497 return 0; 1498 } 1499 1500 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb, 1501 struct iocb *iocb, bool compat) 1502 { 1503 struct kiocb *req; 1504 ssize_t ret; 1505 1506 /* enforce forwards compatibility on users */ 1507 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) { 1508 pr_debug("EINVAL: reserve field set\n"); 1509 return -EINVAL; 1510 } 1511 1512 /* prevent overflows */ 1513 if (unlikely( 1514 (iocb->aio_buf != (unsigned long)iocb->aio_buf) || 1515 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) || 1516 ((ssize_t)iocb->aio_nbytes < 0) 1517 )) { 1518 pr_debug("EINVAL: overflow check\n"); 1519 return -EINVAL; 1520 } 1521 1522 req = aio_get_req(ctx); 1523 if (unlikely(!req)) 1524 return -EAGAIN; 1525 1526 req->ki_filp = fget(iocb->aio_fildes); 1527 if (unlikely(!req->ki_filp)) { 1528 ret = -EBADF; 1529 goto out_put_req; 1530 } 1531 1532 if (iocb->aio_flags & IOCB_FLAG_RESFD) { 1533 /* 1534 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an 1535 * instance of the file* now. The file descriptor must be 1536 * an eventfd() fd, and will be signaled for each completed 1537 * event using the eventfd_signal() function. 1538 */ 1539 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd); 1540 if (IS_ERR(req->ki_eventfd)) { 1541 ret = PTR_ERR(req->ki_eventfd); 1542 req->ki_eventfd = NULL; 1543 goto out_put_req; 1544 } 1545 } 1546 1547 ret = put_user(KIOCB_KEY, &user_iocb->aio_key); 1548 if (unlikely(ret)) { 1549 pr_debug("EFAULT: aio_key\n"); 1550 goto out_put_req; 1551 } 1552 1553 req->ki_obj.user = user_iocb; 1554 req->ki_user_data = iocb->aio_data; 1555 req->ki_pos = iocb->aio_offset; 1556 req->ki_nbytes = iocb->aio_nbytes; 1557 1558 ret = aio_run_iocb(req, iocb->aio_lio_opcode, 1559 (char __user *)(unsigned long)iocb->aio_buf, 1560 compat); 1561 if (ret) 1562 goto out_put_req; 1563 1564 return 0; 1565 out_put_req: 1566 put_reqs_available(ctx, 1); 1567 percpu_ref_put(&ctx->reqs); 1568 kiocb_free(req); 1569 return ret; 1570 } 1571 1572 long do_io_submit(aio_context_t ctx_id, long nr, 1573 struct iocb __user *__user *iocbpp, bool compat) 1574 { 1575 struct kioctx *ctx; 1576 long ret = 0; 1577 int i = 0; 1578 struct blk_plug plug; 1579 1580 if (unlikely(nr < 0)) 1581 return -EINVAL; 1582 1583 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp))) 1584 nr = LONG_MAX/sizeof(*iocbpp); 1585 1586 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp))))) 1587 return -EFAULT; 1588 1589 ctx = lookup_ioctx(ctx_id); 1590 if (unlikely(!ctx)) { 1591 pr_debug("EINVAL: invalid context id\n"); 1592 return -EINVAL; 1593 } 1594 1595 blk_start_plug(&plug); 1596 1597 /* 1598 * AKPM: should this return a partial result if some of the IOs were 1599 * successfully submitted? 1600 */ 1601 for (i=0; i<nr; i++) { 1602 struct iocb __user *user_iocb; 1603 struct iocb tmp; 1604 1605 if (unlikely(__get_user(user_iocb, iocbpp + i))) { 1606 ret = -EFAULT; 1607 break; 1608 } 1609 1610 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) { 1611 ret = -EFAULT; 1612 break; 1613 } 1614 1615 ret = io_submit_one(ctx, user_iocb, &tmp, compat); 1616 if (ret) 1617 break; 1618 } 1619 blk_finish_plug(&plug); 1620 1621 percpu_ref_put(&ctx->users); 1622 return i ? i : ret; 1623 } 1624 1625 /* sys_io_submit: 1626 * Queue the nr iocbs pointed to by iocbpp for processing. Returns 1627 * the number of iocbs queued. May return -EINVAL if the aio_context 1628 * specified by ctx_id is invalid, if nr is < 0, if the iocb at 1629 * *iocbpp[0] is not properly initialized, if the operation specified 1630 * is invalid for the file descriptor in the iocb. May fail with 1631 * -EFAULT if any of the data structures point to invalid data. May 1632 * fail with -EBADF if the file descriptor specified in the first 1633 * iocb is invalid. May fail with -EAGAIN if insufficient resources 1634 * are available to queue any iocbs. Will return 0 if nr is 0. Will 1635 * fail with -ENOSYS if not implemented. 1636 */ 1637 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr, 1638 struct iocb __user * __user *, iocbpp) 1639 { 1640 return do_io_submit(ctx_id, nr, iocbpp, 0); 1641 } 1642 1643 /* lookup_kiocb 1644 * Finds a given iocb for cancellation. 1645 */ 1646 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, 1647 u32 key) 1648 { 1649 struct list_head *pos; 1650 1651 assert_spin_locked(&ctx->ctx_lock); 1652 1653 if (key != KIOCB_KEY) 1654 return NULL; 1655 1656 /* TODO: use a hash or array, this sucks. */ 1657 list_for_each(pos, &ctx->active_reqs) { 1658 struct kiocb *kiocb = list_kiocb(pos); 1659 if (kiocb->ki_obj.user == iocb) 1660 return kiocb; 1661 } 1662 return NULL; 1663 } 1664 1665 /* sys_io_cancel: 1666 * Attempts to cancel an iocb previously passed to io_submit. If 1667 * the operation is successfully cancelled, the resulting event is 1668 * copied into the memory pointed to by result without being placed 1669 * into the completion queue and 0 is returned. May fail with 1670 * -EFAULT if any of the data structures pointed to are invalid. 1671 * May fail with -EINVAL if aio_context specified by ctx_id is 1672 * invalid. May fail with -EAGAIN if the iocb specified was not 1673 * cancelled. Will fail with -ENOSYS if not implemented. 1674 */ 1675 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb, 1676 struct io_event __user *, result) 1677 { 1678 struct kioctx *ctx; 1679 struct kiocb *kiocb; 1680 u32 key; 1681 int ret; 1682 1683 ret = get_user(key, &iocb->aio_key); 1684 if (unlikely(ret)) 1685 return -EFAULT; 1686 1687 ctx = lookup_ioctx(ctx_id); 1688 if (unlikely(!ctx)) 1689 return -EINVAL; 1690 1691 spin_lock_irq(&ctx->ctx_lock); 1692 1693 kiocb = lookup_kiocb(ctx, iocb, key); 1694 if (kiocb) 1695 ret = kiocb_cancel(kiocb); 1696 else 1697 ret = -EINVAL; 1698 1699 spin_unlock_irq(&ctx->ctx_lock); 1700 1701 if (!ret) { 1702 /* 1703 * The result argument is no longer used - the io_event is 1704 * always delivered via the ring buffer. -EINPROGRESS indicates 1705 * cancellation is progress: 1706 */ 1707 ret = -EINPROGRESS; 1708 } 1709 1710 percpu_ref_put(&ctx->users); 1711 1712 return ret; 1713 } 1714 1715 /* io_getevents: 1716 * Attempts to read at least min_nr events and up to nr events from 1717 * the completion queue for the aio_context specified by ctx_id. If 1718 * it succeeds, the number of read events is returned. May fail with 1719 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is 1720 * out of range, if timeout is out of range. May fail with -EFAULT 1721 * if any of the memory specified is invalid. May return 0 or 1722 * < min_nr if the timeout specified by timeout has elapsed 1723 * before sufficient events are available, where timeout == NULL 1724 * specifies an infinite timeout. Note that the timeout pointed to by 1725 * timeout is relative. Will fail with -ENOSYS if not implemented. 1726 */ 1727 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id, 1728 long, min_nr, 1729 long, nr, 1730 struct io_event __user *, events, 1731 struct timespec __user *, timeout) 1732 { 1733 struct kioctx *ioctx = lookup_ioctx(ctx_id); 1734 long ret = -EINVAL; 1735 1736 if (likely(ioctx)) { 1737 if (likely(min_nr <= nr && min_nr >= 0)) 1738 ret = read_events(ioctx, min_nr, nr, events, timeout); 1739 percpu_ref_put(&ioctx->users); 1740 } 1741 return ret; 1742 } 1743