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 #include <linux/kernel.h> 12 #include <linux/init.h> 13 #include <linux/errno.h> 14 #include <linux/time.h> 15 #include <linux/aio_abi.h> 16 #include <linux/module.h> 17 #include <linux/syscalls.h> 18 #include <linux/backing-dev.h> 19 #include <linux/uio.h> 20 21 #define DEBUG 0 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/slab.h> 30 #include <linux/timer.h> 31 #include <linux/aio.h> 32 #include <linux/highmem.h> 33 #include <linux/workqueue.h> 34 #include <linux/security.h> 35 #include <linux/eventfd.h> 36 #include <linux/blkdev.h> 37 #include <linux/mempool.h> 38 #include <linux/hash.h> 39 40 #include <asm/kmap_types.h> 41 #include <asm/uaccess.h> 42 43 #if DEBUG > 1 44 #define dprintk printk 45 #else 46 #define dprintk(x...) do { ; } while (0) 47 #endif 48 49 /*------ sysctl variables----*/ 50 static DEFINE_SPINLOCK(aio_nr_lock); 51 unsigned long aio_nr; /* current system wide number of aio requests */ 52 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */ 53 /*----end sysctl variables---*/ 54 55 static struct kmem_cache *kiocb_cachep; 56 static struct kmem_cache *kioctx_cachep; 57 58 static struct workqueue_struct *aio_wq; 59 60 /* Used for rare fput completion. */ 61 static void aio_fput_routine(struct work_struct *); 62 static DECLARE_WORK(fput_work, aio_fput_routine); 63 64 static DEFINE_SPINLOCK(fput_lock); 65 static LIST_HEAD(fput_head); 66 67 #define AIO_BATCH_HASH_BITS 3 /* allocated on-stack, so don't go crazy */ 68 #define AIO_BATCH_HASH_SIZE (1 << AIO_BATCH_HASH_BITS) 69 struct aio_batch_entry { 70 struct hlist_node list; 71 struct address_space *mapping; 72 }; 73 mempool_t *abe_pool; 74 75 static void aio_kick_handler(struct work_struct *); 76 static void aio_queue_work(struct kioctx *); 77 78 /* aio_setup 79 * Creates the slab caches used by the aio routines, panic on 80 * failure as this is done early during the boot sequence. 81 */ 82 static int __init aio_setup(void) 83 { 84 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC); 85 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC); 86 87 aio_wq = create_workqueue("aio"); 88 abe_pool = mempool_create_kmalloc_pool(1, sizeof(struct aio_batch_entry)); 89 BUG_ON(!abe_pool); 90 91 pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page)); 92 93 return 0; 94 } 95 __initcall(aio_setup); 96 97 static void aio_free_ring(struct kioctx *ctx) 98 { 99 struct aio_ring_info *info = &ctx->ring_info; 100 long i; 101 102 for (i=0; i<info->nr_pages; i++) 103 put_page(info->ring_pages[i]); 104 105 if (info->mmap_size) { 106 down_write(&ctx->mm->mmap_sem); 107 do_munmap(ctx->mm, info->mmap_base, info->mmap_size); 108 up_write(&ctx->mm->mmap_sem); 109 } 110 111 if (info->ring_pages && info->ring_pages != info->internal_pages) 112 kfree(info->ring_pages); 113 info->ring_pages = NULL; 114 info->nr = 0; 115 } 116 117 static int aio_setup_ring(struct kioctx *ctx) 118 { 119 struct aio_ring *ring; 120 struct aio_ring_info *info = &ctx->ring_info; 121 unsigned nr_events = ctx->max_reqs; 122 unsigned long size; 123 int nr_pages; 124 125 /* Compensate for the ring buffer's head/tail overlap entry */ 126 nr_events += 2; /* 1 is required, 2 for good luck */ 127 128 size = sizeof(struct aio_ring); 129 size += sizeof(struct io_event) * nr_events; 130 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT; 131 132 if (nr_pages < 0) 133 return -EINVAL; 134 135 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event); 136 137 info->nr = 0; 138 info->ring_pages = info->internal_pages; 139 if (nr_pages > AIO_RING_PAGES) { 140 info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL); 141 if (!info->ring_pages) 142 return -ENOMEM; 143 } 144 145 info->mmap_size = nr_pages * PAGE_SIZE; 146 dprintk("attempting mmap of %lu bytes\n", info->mmap_size); 147 down_write(&ctx->mm->mmap_sem); 148 info->mmap_base = do_mmap(NULL, 0, info->mmap_size, 149 PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE, 150 0); 151 if (IS_ERR((void *)info->mmap_base)) { 152 up_write(&ctx->mm->mmap_sem); 153 info->mmap_size = 0; 154 aio_free_ring(ctx); 155 return -EAGAIN; 156 } 157 158 dprintk("mmap address: 0x%08lx\n", info->mmap_base); 159 info->nr_pages = get_user_pages(current, ctx->mm, 160 info->mmap_base, nr_pages, 161 1, 0, info->ring_pages, NULL); 162 up_write(&ctx->mm->mmap_sem); 163 164 if (unlikely(info->nr_pages != nr_pages)) { 165 aio_free_ring(ctx); 166 return -EAGAIN; 167 } 168 169 ctx->user_id = info->mmap_base; 170 171 info->nr = nr_events; /* trusted copy */ 172 173 ring = kmap_atomic(info->ring_pages[0], KM_USER0); 174 ring->nr = nr_events; /* user copy */ 175 ring->id = ctx->user_id; 176 ring->head = ring->tail = 0; 177 ring->magic = AIO_RING_MAGIC; 178 ring->compat_features = AIO_RING_COMPAT_FEATURES; 179 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES; 180 ring->header_length = sizeof(struct aio_ring); 181 kunmap_atomic(ring, KM_USER0); 182 183 return 0; 184 } 185 186 187 /* aio_ring_event: returns a pointer to the event at the given index from 188 * kmap_atomic(, km). Release the pointer with put_aio_ring_event(); 189 */ 190 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event)) 191 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event)) 192 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE) 193 194 #define aio_ring_event(info, nr, km) ({ \ 195 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \ 196 struct io_event *__event; \ 197 __event = kmap_atomic( \ 198 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \ 199 __event += pos % AIO_EVENTS_PER_PAGE; \ 200 __event; \ 201 }) 202 203 #define put_aio_ring_event(event, km) do { \ 204 struct io_event *__event = (event); \ 205 (void)__event; \ 206 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \ 207 } while(0) 208 209 static void ctx_rcu_free(struct rcu_head *head) 210 { 211 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head); 212 unsigned nr_events = ctx->max_reqs; 213 214 kmem_cache_free(kioctx_cachep, ctx); 215 216 if (nr_events) { 217 spin_lock(&aio_nr_lock); 218 BUG_ON(aio_nr - nr_events > aio_nr); 219 aio_nr -= nr_events; 220 spin_unlock(&aio_nr_lock); 221 } 222 } 223 224 /* __put_ioctx 225 * Called when the last user of an aio context has gone away, 226 * and the struct needs to be freed. 227 */ 228 static void __put_ioctx(struct kioctx *ctx) 229 { 230 BUG_ON(ctx->reqs_active); 231 232 cancel_delayed_work(&ctx->wq); 233 cancel_work_sync(&ctx->wq.work); 234 aio_free_ring(ctx); 235 mmdrop(ctx->mm); 236 ctx->mm = NULL; 237 pr_debug("__put_ioctx: freeing %p\n", ctx); 238 call_rcu(&ctx->rcu_head, ctx_rcu_free); 239 } 240 241 #define get_ioctx(kioctx) do { \ 242 BUG_ON(atomic_read(&(kioctx)->users) <= 0); \ 243 atomic_inc(&(kioctx)->users); \ 244 } while (0) 245 #define put_ioctx(kioctx) do { \ 246 BUG_ON(atomic_read(&(kioctx)->users) <= 0); \ 247 if (unlikely(atomic_dec_and_test(&(kioctx)->users))) \ 248 __put_ioctx(kioctx); \ 249 } while (0) 250 251 /* ioctx_alloc 252 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed. 253 */ 254 static struct kioctx *ioctx_alloc(unsigned nr_events) 255 { 256 struct mm_struct *mm; 257 struct kioctx *ctx; 258 int did_sync = 0; 259 260 /* Prevent overflows */ 261 if ((nr_events > (0x10000000U / sizeof(struct io_event))) || 262 (nr_events > (0x10000000U / sizeof(struct kiocb)))) { 263 pr_debug("ENOMEM: nr_events too high\n"); 264 return ERR_PTR(-EINVAL); 265 } 266 267 if ((unsigned long)nr_events > aio_max_nr) 268 return ERR_PTR(-EAGAIN); 269 270 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL); 271 if (!ctx) 272 return ERR_PTR(-ENOMEM); 273 274 ctx->max_reqs = nr_events; 275 mm = ctx->mm = current->mm; 276 atomic_inc(&mm->mm_count); 277 278 atomic_set(&ctx->users, 1); 279 spin_lock_init(&ctx->ctx_lock); 280 spin_lock_init(&ctx->ring_info.ring_lock); 281 init_waitqueue_head(&ctx->wait); 282 283 INIT_LIST_HEAD(&ctx->active_reqs); 284 INIT_LIST_HEAD(&ctx->run_list); 285 INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler); 286 287 if (aio_setup_ring(ctx) < 0) 288 goto out_freectx; 289 290 /* limit the number of system wide aios */ 291 do { 292 spin_lock_bh(&aio_nr_lock); 293 if (aio_nr + nr_events > aio_max_nr || 294 aio_nr + nr_events < aio_nr) 295 ctx->max_reqs = 0; 296 else 297 aio_nr += ctx->max_reqs; 298 spin_unlock_bh(&aio_nr_lock); 299 if (ctx->max_reqs || did_sync) 300 break; 301 302 /* wait for rcu callbacks to have completed before giving up */ 303 synchronize_rcu(); 304 did_sync = 1; 305 ctx->max_reqs = nr_events; 306 } while (1); 307 308 if (ctx->max_reqs == 0) 309 goto out_cleanup; 310 311 /* now link into global list. */ 312 spin_lock(&mm->ioctx_lock); 313 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list); 314 spin_unlock(&mm->ioctx_lock); 315 316 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n", 317 ctx, ctx->user_id, current->mm, ctx->ring_info.nr); 318 return ctx; 319 320 out_cleanup: 321 __put_ioctx(ctx); 322 return ERR_PTR(-EAGAIN); 323 324 out_freectx: 325 mmdrop(mm); 326 kmem_cache_free(kioctx_cachep, ctx); 327 ctx = ERR_PTR(-ENOMEM); 328 329 dprintk("aio: error allocating ioctx %p\n", ctx); 330 return ctx; 331 } 332 333 /* aio_cancel_all 334 * Cancels all outstanding aio requests on an aio context. Used 335 * when the processes owning a context have all exited to encourage 336 * the rapid destruction of the kioctx. 337 */ 338 static void aio_cancel_all(struct kioctx *ctx) 339 { 340 int (*cancel)(struct kiocb *, struct io_event *); 341 struct io_event res; 342 spin_lock_irq(&ctx->ctx_lock); 343 ctx->dead = 1; 344 while (!list_empty(&ctx->active_reqs)) { 345 struct list_head *pos = ctx->active_reqs.next; 346 struct kiocb *iocb = list_kiocb(pos); 347 list_del_init(&iocb->ki_list); 348 cancel = iocb->ki_cancel; 349 kiocbSetCancelled(iocb); 350 if (cancel) { 351 iocb->ki_users++; 352 spin_unlock_irq(&ctx->ctx_lock); 353 cancel(iocb, &res); 354 spin_lock_irq(&ctx->ctx_lock); 355 } 356 } 357 spin_unlock_irq(&ctx->ctx_lock); 358 } 359 360 static void wait_for_all_aios(struct kioctx *ctx) 361 { 362 struct task_struct *tsk = current; 363 DECLARE_WAITQUEUE(wait, tsk); 364 365 spin_lock_irq(&ctx->ctx_lock); 366 if (!ctx->reqs_active) 367 goto out; 368 369 add_wait_queue(&ctx->wait, &wait); 370 set_task_state(tsk, TASK_UNINTERRUPTIBLE); 371 while (ctx->reqs_active) { 372 spin_unlock_irq(&ctx->ctx_lock); 373 io_schedule(); 374 set_task_state(tsk, TASK_UNINTERRUPTIBLE); 375 spin_lock_irq(&ctx->ctx_lock); 376 } 377 __set_task_state(tsk, TASK_RUNNING); 378 remove_wait_queue(&ctx->wait, &wait); 379 380 out: 381 spin_unlock_irq(&ctx->ctx_lock); 382 } 383 384 /* wait_on_sync_kiocb: 385 * Waits on the given sync kiocb to complete. 386 */ 387 ssize_t wait_on_sync_kiocb(struct kiocb *iocb) 388 { 389 while (iocb->ki_users) { 390 set_current_state(TASK_UNINTERRUPTIBLE); 391 if (!iocb->ki_users) 392 break; 393 io_schedule(); 394 } 395 __set_current_state(TASK_RUNNING); 396 return iocb->ki_user_data; 397 } 398 EXPORT_SYMBOL(wait_on_sync_kiocb); 399 400 /* exit_aio: called when the last user of mm goes away. At this point, 401 * there is no way for any new requests to be submited or any of the 402 * io_* syscalls to be called on the context. However, there may be 403 * outstanding requests which hold references to the context; as they 404 * go away, they will call put_ioctx and release any pinned memory 405 * associated with the request (held via struct page * references). 406 */ 407 void exit_aio(struct mm_struct *mm) 408 { 409 struct kioctx *ctx; 410 411 while (!hlist_empty(&mm->ioctx_list)) { 412 ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list); 413 hlist_del_rcu(&ctx->list); 414 415 aio_cancel_all(ctx); 416 417 wait_for_all_aios(ctx); 418 /* 419 * Ensure we don't leave the ctx on the aio_wq 420 */ 421 cancel_work_sync(&ctx->wq.work); 422 423 if (1 != atomic_read(&ctx->users)) 424 printk(KERN_DEBUG 425 "exit_aio:ioctx still alive: %d %d %d\n", 426 atomic_read(&ctx->users), ctx->dead, 427 ctx->reqs_active); 428 put_ioctx(ctx); 429 } 430 } 431 432 /* aio_get_req 433 * Allocate a slot for an aio request. Increments the users count 434 * of the kioctx so that the kioctx stays around until all requests are 435 * complete. Returns NULL if no requests are free. 436 * 437 * Returns with kiocb->users set to 2. The io submit code path holds 438 * an extra reference while submitting the i/o. 439 * This prevents races between the aio code path referencing the 440 * req (after submitting it) and aio_complete() freeing the req. 441 */ 442 static struct kiocb *__aio_get_req(struct kioctx *ctx) 443 { 444 struct kiocb *req = NULL; 445 struct aio_ring *ring; 446 int okay = 0; 447 448 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL); 449 if (unlikely(!req)) 450 return NULL; 451 452 req->ki_flags = 0; 453 req->ki_users = 2; 454 req->ki_key = 0; 455 req->ki_ctx = ctx; 456 req->ki_cancel = NULL; 457 req->ki_retry = NULL; 458 req->ki_dtor = NULL; 459 req->private = NULL; 460 req->ki_iovec = NULL; 461 INIT_LIST_HEAD(&req->ki_run_list); 462 req->ki_eventfd = NULL; 463 464 /* Check if the completion queue has enough free space to 465 * accept an event from this io. 466 */ 467 spin_lock_irq(&ctx->ctx_lock); 468 ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0); 469 if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) { 470 list_add(&req->ki_list, &ctx->active_reqs); 471 ctx->reqs_active++; 472 okay = 1; 473 } 474 kunmap_atomic(ring, KM_USER0); 475 spin_unlock_irq(&ctx->ctx_lock); 476 477 if (!okay) { 478 kmem_cache_free(kiocb_cachep, req); 479 req = NULL; 480 } 481 482 return req; 483 } 484 485 static inline struct kiocb *aio_get_req(struct kioctx *ctx) 486 { 487 struct kiocb *req; 488 /* Handle a potential starvation case -- should be exceedingly rare as 489 * requests will be stuck on fput_head only if the aio_fput_routine is 490 * delayed and the requests were the last user of the struct file. 491 */ 492 req = __aio_get_req(ctx); 493 if (unlikely(NULL == req)) { 494 aio_fput_routine(NULL); 495 req = __aio_get_req(ctx); 496 } 497 return req; 498 } 499 500 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req) 501 { 502 assert_spin_locked(&ctx->ctx_lock); 503 504 if (req->ki_eventfd != NULL) 505 eventfd_ctx_put(req->ki_eventfd); 506 if (req->ki_dtor) 507 req->ki_dtor(req); 508 if (req->ki_iovec != &req->ki_inline_vec) 509 kfree(req->ki_iovec); 510 kmem_cache_free(kiocb_cachep, req); 511 ctx->reqs_active--; 512 513 if (unlikely(!ctx->reqs_active && ctx->dead)) 514 wake_up(&ctx->wait); 515 } 516 517 static void aio_fput_routine(struct work_struct *data) 518 { 519 spin_lock_irq(&fput_lock); 520 while (likely(!list_empty(&fput_head))) { 521 struct kiocb *req = list_kiocb(fput_head.next); 522 struct kioctx *ctx = req->ki_ctx; 523 524 list_del(&req->ki_list); 525 spin_unlock_irq(&fput_lock); 526 527 /* Complete the fput(s) */ 528 if (req->ki_filp != NULL) 529 __fput(req->ki_filp); 530 531 /* Link the iocb into the context's free list */ 532 spin_lock_irq(&ctx->ctx_lock); 533 really_put_req(ctx, req); 534 spin_unlock_irq(&ctx->ctx_lock); 535 536 put_ioctx(ctx); 537 spin_lock_irq(&fput_lock); 538 } 539 spin_unlock_irq(&fput_lock); 540 } 541 542 /* __aio_put_req 543 * Returns true if this put was the last user of the request. 544 */ 545 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req) 546 { 547 dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n", 548 req, atomic_long_read(&req->ki_filp->f_count)); 549 550 assert_spin_locked(&ctx->ctx_lock); 551 552 req->ki_users--; 553 BUG_ON(req->ki_users < 0); 554 if (likely(req->ki_users)) 555 return 0; 556 list_del(&req->ki_list); /* remove from active_reqs */ 557 req->ki_cancel = NULL; 558 req->ki_retry = NULL; 559 560 /* 561 * Try to optimize the aio and eventfd file* puts, by avoiding to 562 * schedule work in case it is not __fput() time. In normal cases, 563 * we would not be holding the last reference to the file*, so 564 * this function will be executed w/out any aio kthread wakeup. 565 */ 566 if (unlikely(atomic_long_dec_and_test(&req->ki_filp->f_count))) { 567 get_ioctx(ctx); 568 spin_lock(&fput_lock); 569 list_add(&req->ki_list, &fput_head); 570 spin_unlock(&fput_lock); 571 queue_work(aio_wq, &fput_work); 572 } else { 573 req->ki_filp = NULL; 574 really_put_req(ctx, req); 575 } 576 return 1; 577 } 578 579 /* aio_put_req 580 * Returns true if this put was the last user of the kiocb, 581 * false if the request is still in use. 582 */ 583 int aio_put_req(struct kiocb *req) 584 { 585 struct kioctx *ctx = req->ki_ctx; 586 int ret; 587 spin_lock_irq(&ctx->ctx_lock); 588 ret = __aio_put_req(ctx, req); 589 spin_unlock_irq(&ctx->ctx_lock); 590 return ret; 591 } 592 EXPORT_SYMBOL(aio_put_req); 593 594 static struct kioctx *lookup_ioctx(unsigned long ctx_id) 595 { 596 struct mm_struct *mm = current->mm; 597 struct kioctx *ctx, *ret = NULL; 598 struct hlist_node *n; 599 600 rcu_read_lock(); 601 602 hlist_for_each_entry_rcu(ctx, n, &mm->ioctx_list, list) { 603 if (ctx->user_id == ctx_id && !ctx->dead) { 604 get_ioctx(ctx); 605 ret = ctx; 606 break; 607 } 608 } 609 610 rcu_read_unlock(); 611 return ret; 612 } 613 614 /* 615 * Queue up a kiocb to be retried. Assumes that the kiocb 616 * has already been marked as kicked, and places it on 617 * the retry run list for the corresponding ioctx, if it 618 * isn't already queued. Returns 1 if it actually queued 619 * the kiocb (to tell the caller to activate the work 620 * queue to process it), or 0, if it found that it was 621 * already queued. 622 */ 623 static inline int __queue_kicked_iocb(struct kiocb *iocb) 624 { 625 struct kioctx *ctx = iocb->ki_ctx; 626 627 assert_spin_locked(&ctx->ctx_lock); 628 629 if (list_empty(&iocb->ki_run_list)) { 630 list_add_tail(&iocb->ki_run_list, 631 &ctx->run_list); 632 return 1; 633 } 634 return 0; 635 } 636 637 /* aio_run_iocb 638 * This is the core aio execution routine. It is 639 * invoked both for initial i/o submission and 640 * subsequent retries via the aio_kick_handler. 641 * Expects to be invoked with iocb->ki_ctx->lock 642 * already held. The lock is released and reacquired 643 * as needed during processing. 644 * 645 * Calls the iocb retry method (already setup for the 646 * iocb on initial submission) for operation specific 647 * handling, but takes care of most of common retry 648 * execution details for a given iocb. The retry method 649 * needs to be non-blocking as far as possible, to avoid 650 * holding up other iocbs waiting to be serviced by the 651 * retry kernel thread. 652 * 653 * The trickier parts in this code have to do with 654 * ensuring that only one retry instance is in progress 655 * for a given iocb at any time. Providing that guarantee 656 * simplifies the coding of individual aio operations as 657 * it avoids various potential races. 658 */ 659 static ssize_t aio_run_iocb(struct kiocb *iocb) 660 { 661 struct kioctx *ctx = iocb->ki_ctx; 662 ssize_t (*retry)(struct kiocb *); 663 ssize_t ret; 664 665 if (!(retry = iocb->ki_retry)) { 666 printk("aio_run_iocb: iocb->ki_retry = NULL\n"); 667 return 0; 668 } 669 670 /* 671 * We don't want the next retry iteration for this 672 * operation to start until this one has returned and 673 * updated the iocb state. However, wait_queue functions 674 * can trigger a kick_iocb from interrupt context in the 675 * meantime, indicating that data is available for the next 676 * iteration. We want to remember that and enable the 677 * next retry iteration _after_ we are through with 678 * this one. 679 * 680 * So, in order to be able to register a "kick", but 681 * prevent it from being queued now, we clear the kick 682 * flag, but make the kick code *think* that the iocb is 683 * still on the run list until we are actually done. 684 * When we are done with this iteration, we check if 685 * the iocb was kicked in the meantime and if so, queue 686 * it up afresh. 687 */ 688 689 kiocbClearKicked(iocb); 690 691 /* 692 * This is so that aio_complete knows it doesn't need to 693 * pull the iocb off the run list (We can't just call 694 * INIT_LIST_HEAD because we don't want a kick_iocb to 695 * queue this on the run list yet) 696 */ 697 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL; 698 spin_unlock_irq(&ctx->ctx_lock); 699 700 /* Quit retrying if the i/o has been cancelled */ 701 if (kiocbIsCancelled(iocb)) { 702 ret = -EINTR; 703 aio_complete(iocb, ret, 0); 704 /* must not access the iocb after this */ 705 goto out; 706 } 707 708 /* 709 * Now we are all set to call the retry method in async 710 * context. 711 */ 712 ret = retry(iocb); 713 714 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) { 715 BUG_ON(!list_empty(&iocb->ki_wait.task_list)); 716 aio_complete(iocb, ret, 0); 717 } 718 out: 719 spin_lock_irq(&ctx->ctx_lock); 720 721 if (-EIOCBRETRY == ret) { 722 /* 723 * OK, now that we are done with this iteration 724 * and know that there is more left to go, 725 * this is where we let go so that a subsequent 726 * "kick" can start the next iteration 727 */ 728 729 /* will make __queue_kicked_iocb succeed from here on */ 730 INIT_LIST_HEAD(&iocb->ki_run_list); 731 /* we must queue the next iteration ourselves, if it 732 * has already been kicked */ 733 if (kiocbIsKicked(iocb)) { 734 __queue_kicked_iocb(iocb); 735 736 /* 737 * __queue_kicked_iocb will always return 1 here, because 738 * iocb->ki_run_list is empty at this point so it should 739 * be safe to unconditionally queue the context into the 740 * work queue. 741 */ 742 aio_queue_work(ctx); 743 } 744 } 745 return ret; 746 } 747 748 /* 749 * __aio_run_iocbs: 750 * Process all pending retries queued on the ioctx 751 * run list. 752 * Assumes it is operating within the aio issuer's mm 753 * context. 754 */ 755 static int __aio_run_iocbs(struct kioctx *ctx) 756 { 757 struct kiocb *iocb; 758 struct list_head run_list; 759 760 assert_spin_locked(&ctx->ctx_lock); 761 762 list_replace_init(&ctx->run_list, &run_list); 763 while (!list_empty(&run_list)) { 764 iocb = list_entry(run_list.next, struct kiocb, 765 ki_run_list); 766 list_del(&iocb->ki_run_list); 767 /* 768 * Hold an extra reference while retrying i/o. 769 */ 770 iocb->ki_users++; /* grab extra reference */ 771 aio_run_iocb(iocb); 772 __aio_put_req(ctx, iocb); 773 } 774 if (!list_empty(&ctx->run_list)) 775 return 1; 776 return 0; 777 } 778 779 static void aio_queue_work(struct kioctx * ctx) 780 { 781 unsigned long timeout; 782 /* 783 * if someone is waiting, get the work started right 784 * away, otherwise, use a longer delay 785 */ 786 smp_mb(); 787 if (waitqueue_active(&ctx->wait)) 788 timeout = 1; 789 else 790 timeout = HZ/10; 791 queue_delayed_work(aio_wq, &ctx->wq, timeout); 792 } 793 794 795 /* 796 * aio_run_iocbs: 797 * Process all pending retries queued on the ioctx 798 * run list. 799 * Assumes it is operating within the aio issuer's mm 800 * context. 801 */ 802 static inline void aio_run_iocbs(struct kioctx *ctx) 803 { 804 int requeue; 805 806 spin_lock_irq(&ctx->ctx_lock); 807 808 requeue = __aio_run_iocbs(ctx); 809 spin_unlock_irq(&ctx->ctx_lock); 810 if (requeue) 811 aio_queue_work(ctx); 812 } 813 814 /* 815 * just like aio_run_iocbs, but keeps running them until 816 * the list stays empty 817 */ 818 static inline void aio_run_all_iocbs(struct kioctx *ctx) 819 { 820 spin_lock_irq(&ctx->ctx_lock); 821 while (__aio_run_iocbs(ctx)) 822 ; 823 spin_unlock_irq(&ctx->ctx_lock); 824 } 825 826 /* 827 * aio_kick_handler: 828 * Work queue handler triggered to process pending 829 * retries on an ioctx. Takes on the aio issuer's 830 * mm context before running the iocbs, so that 831 * copy_xxx_user operates on the issuer's address 832 * space. 833 * Run on aiod's context. 834 */ 835 static void aio_kick_handler(struct work_struct *work) 836 { 837 struct kioctx *ctx = container_of(work, struct kioctx, wq.work); 838 mm_segment_t oldfs = get_fs(); 839 struct mm_struct *mm; 840 int requeue; 841 842 set_fs(USER_DS); 843 use_mm(ctx->mm); 844 spin_lock_irq(&ctx->ctx_lock); 845 requeue =__aio_run_iocbs(ctx); 846 mm = ctx->mm; 847 spin_unlock_irq(&ctx->ctx_lock); 848 unuse_mm(mm); 849 set_fs(oldfs); 850 /* 851 * we're in a worker thread already, don't use queue_delayed_work, 852 */ 853 if (requeue) 854 queue_delayed_work(aio_wq, &ctx->wq, 0); 855 } 856 857 858 /* 859 * Called by kick_iocb to queue the kiocb for retry 860 * and if required activate the aio work queue to process 861 * it 862 */ 863 static void try_queue_kicked_iocb(struct kiocb *iocb) 864 { 865 struct kioctx *ctx = iocb->ki_ctx; 866 unsigned long flags; 867 int run = 0; 868 869 /* We're supposed to be the only path putting the iocb back on the run 870 * list. If we find that the iocb is *back* on a wait queue already 871 * than retry has happened before we could queue the iocb. This also 872 * means that the retry could have completed and freed our iocb, no 873 * good. */ 874 BUG_ON((!list_empty(&iocb->ki_wait.task_list))); 875 876 spin_lock_irqsave(&ctx->ctx_lock, flags); 877 /* set this inside the lock so that we can't race with aio_run_iocb() 878 * testing it and putting the iocb on the run list under the lock */ 879 if (!kiocbTryKick(iocb)) 880 run = __queue_kicked_iocb(iocb); 881 spin_unlock_irqrestore(&ctx->ctx_lock, flags); 882 if (run) 883 aio_queue_work(ctx); 884 } 885 886 /* 887 * kick_iocb: 888 * Called typically from a wait queue callback context 889 * (aio_wake_function) to trigger a retry of the iocb. 890 * The retry is usually executed by aio workqueue 891 * threads (See aio_kick_handler). 892 */ 893 void kick_iocb(struct kiocb *iocb) 894 { 895 /* sync iocbs are easy: they can only ever be executing from a 896 * single context. */ 897 if (is_sync_kiocb(iocb)) { 898 kiocbSetKicked(iocb); 899 wake_up_process(iocb->ki_obj.tsk); 900 return; 901 } 902 903 try_queue_kicked_iocb(iocb); 904 } 905 EXPORT_SYMBOL(kick_iocb); 906 907 /* aio_complete 908 * Called when the io request on the given iocb is complete. 909 * Returns true if this is the last user of the request. The 910 * only other user of the request can be the cancellation code. 911 */ 912 int aio_complete(struct kiocb *iocb, long res, long res2) 913 { 914 struct kioctx *ctx = iocb->ki_ctx; 915 struct aio_ring_info *info; 916 struct aio_ring *ring; 917 struct io_event *event; 918 unsigned long flags; 919 unsigned long tail; 920 int ret; 921 922 /* 923 * Special case handling for sync iocbs: 924 * - events go directly into the iocb for fast handling 925 * - the sync task with the iocb in its stack holds the single iocb 926 * ref, no other paths have a way to get another ref 927 * - the sync task helpfully left a reference to itself in the iocb 928 */ 929 if (is_sync_kiocb(iocb)) { 930 BUG_ON(iocb->ki_users != 1); 931 iocb->ki_user_data = res; 932 iocb->ki_users = 0; 933 wake_up_process(iocb->ki_obj.tsk); 934 return 1; 935 } 936 937 info = &ctx->ring_info; 938 939 /* add a completion event to the ring buffer. 940 * must be done holding ctx->ctx_lock to prevent 941 * other code from messing with the tail 942 * pointer since we might be called from irq 943 * context. 944 */ 945 spin_lock_irqsave(&ctx->ctx_lock, flags); 946 947 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list)) 948 list_del_init(&iocb->ki_run_list); 949 950 /* 951 * cancelled requests don't get events, userland was given one 952 * when the event got cancelled. 953 */ 954 if (kiocbIsCancelled(iocb)) 955 goto put_rq; 956 957 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1); 958 959 tail = info->tail; 960 event = aio_ring_event(info, tail, KM_IRQ0); 961 if (++tail >= info->nr) 962 tail = 0; 963 964 event->obj = (u64)(unsigned long)iocb->ki_obj.user; 965 event->data = iocb->ki_user_data; 966 event->res = res; 967 event->res2 = res2; 968 969 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n", 970 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data, 971 res, res2); 972 973 /* after flagging the request as done, we 974 * must never even look at it again 975 */ 976 smp_wmb(); /* make event visible before updating tail */ 977 978 info->tail = tail; 979 ring->tail = tail; 980 981 put_aio_ring_event(event, KM_IRQ0); 982 kunmap_atomic(ring, KM_IRQ1); 983 984 pr_debug("added to ring %p at [%lu]\n", iocb, tail); 985 986 /* 987 * Check if the user asked us to deliver the result through an 988 * eventfd. The eventfd_signal() function is safe to be called 989 * from IRQ context. 990 */ 991 if (iocb->ki_eventfd != NULL) 992 eventfd_signal(iocb->ki_eventfd, 1); 993 994 put_rq: 995 /* everything turned out well, dispose of the aiocb. */ 996 ret = __aio_put_req(ctx, iocb); 997 998 /* 999 * We have to order our ring_info tail store above and test 1000 * of the wait list below outside the wait lock. This is 1001 * like in wake_up_bit() where clearing a bit has to be 1002 * ordered with the unlocked test. 1003 */ 1004 smp_mb(); 1005 1006 if (waitqueue_active(&ctx->wait)) 1007 wake_up(&ctx->wait); 1008 1009 spin_unlock_irqrestore(&ctx->ctx_lock, flags); 1010 return ret; 1011 } 1012 EXPORT_SYMBOL(aio_complete); 1013 1014 /* aio_read_evt 1015 * Pull an event off of the ioctx's event ring. Returns the number of 1016 * events fetched (0 or 1 ;-) 1017 * FIXME: make this use cmpxchg. 1018 * TODO: make the ringbuffer user mmap()able (requires FIXME). 1019 */ 1020 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent) 1021 { 1022 struct aio_ring_info *info = &ioctx->ring_info; 1023 struct aio_ring *ring; 1024 unsigned long head; 1025 int ret = 0; 1026 1027 ring = kmap_atomic(info->ring_pages[0], KM_USER0); 1028 dprintk("in aio_read_evt h%lu t%lu m%lu\n", 1029 (unsigned long)ring->head, (unsigned long)ring->tail, 1030 (unsigned long)ring->nr); 1031 1032 if (ring->head == ring->tail) 1033 goto out; 1034 1035 spin_lock(&info->ring_lock); 1036 1037 head = ring->head % info->nr; 1038 if (head != ring->tail) { 1039 struct io_event *evp = aio_ring_event(info, head, KM_USER1); 1040 *ent = *evp; 1041 head = (head + 1) % info->nr; 1042 smp_mb(); /* finish reading the event before updatng the head */ 1043 ring->head = head; 1044 ret = 1; 1045 put_aio_ring_event(evp, KM_USER1); 1046 } 1047 spin_unlock(&info->ring_lock); 1048 1049 out: 1050 kunmap_atomic(ring, KM_USER0); 1051 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret, 1052 (unsigned long)ring->head, (unsigned long)ring->tail); 1053 return ret; 1054 } 1055 1056 struct aio_timeout { 1057 struct timer_list timer; 1058 int timed_out; 1059 struct task_struct *p; 1060 }; 1061 1062 static void timeout_func(unsigned long data) 1063 { 1064 struct aio_timeout *to = (struct aio_timeout *)data; 1065 1066 to->timed_out = 1; 1067 wake_up_process(to->p); 1068 } 1069 1070 static inline void init_timeout(struct aio_timeout *to) 1071 { 1072 setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to); 1073 to->timed_out = 0; 1074 to->p = current; 1075 } 1076 1077 static inline void set_timeout(long start_jiffies, struct aio_timeout *to, 1078 const struct timespec *ts) 1079 { 1080 to->timer.expires = start_jiffies + timespec_to_jiffies(ts); 1081 if (time_after(to->timer.expires, jiffies)) 1082 add_timer(&to->timer); 1083 else 1084 to->timed_out = 1; 1085 } 1086 1087 static inline void clear_timeout(struct aio_timeout *to) 1088 { 1089 del_singleshot_timer_sync(&to->timer); 1090 } 1091 1092 static int read_events(struct kioctx *ctx, 1093 long min_nr, long nr, 1094 struct io_event __user *event, 1095 struct timespec __user *timeout) 1096 { 1097 long start_jiffies = jiffies; 1098 struct task_struct *tsk = current; 1099 DECLARE_WAITQUEUE(wait, tsk); 1100 int ret; 1101 int i = 0; 1102 struct io_event ent; 1103 struct aio_timeout to; 1104 int retry = 0; 1105 1106 /* needed to zero any padding within an entry (there shouldn't be 1107 * any, but C is fun! 1108 */ 1109 memset(&ent, 0, sizeof(ent)); 1110 retry: 1111 ret = 0; 1112 while (likely(i < nr)) { 1113 ret = aio_read_evt(ctx, &ent); 1114 if (unlikely(ret <= 0)) 1115 break; 1116 1117 dprintk("read event: %Lx %Lx %Lx %Lx\n", 1118 ent.data, ent.obj, ent.res, ent.res2); 1119 1120 /* Could we split the check in two? */ 1121 ret = -EFAULT; 1122 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) { 1123 dprintk("aio: lost an event due to EFAULT.\n"); 1124 break; 1125 } 1126 ret = 0; 1127 1128 /* Good, event copied to userland, update counts. */ 1129 event ++; 1130 i ++; 1131 } 1132 1133 if (min_nr <= i) 1134 return i; 1135 if (ret) 1136 return ret; 1137 1138 /* End fast path */ 1139 1140 /* racey check, but it gets redone */ 1141 if (!retry && unlikely(!list_empty(&ctx->run_list))) { 1142 retry = 1; 1143 aio_run_all_iocbs(ctx); 1144 goto retry; 1145 } 1146 1147 init_timeout(&to); 1148 if (timeout) { 1149 struct timespec ts; 1150 ret = -EFAULT; 1151 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts)))) 1152 goto out; 1153 1154 set_timeout(start_jiffies, &to, &ts); 1155 } 1156 1157 while (likely(i < nr)) { 1158 add_wait_queue_exclusive(&ctx->wait, &wait); 1159 do { 1160 set_task_state(tsk, TASK_INTERRUPTIBLE); 1161 ret = aio_read_evt(ctx, &ent); 1162 if (ret) 1163 break; 1164 if (min_nr <= i) 1165 break; 1166 if (unlikely(ctx->dead)) { 1167 ret = -EINVAL; 1168 break; 1169 } 1170 if (to.timed_out) /* Only check after read evt */ 1171 break; 1172 /* Try to only show up in io wait if there are ops 1173 * in flight */ 1174 if (ctx->reqs_active) 1175 io_schedule(); 1176 else 1177 schedule(); 1178 if (signal_pending(tsk)) { 1179 ret = -EINTR; 1180 break; 1181 } 1182 /*ret = aio_read_evt(ctx, &ent);*/ 1183 } while (1) ; 1184 1185 set_task_state(tsk, TASK_RUNNING); 1186 remove_wait_queue(&ctx->wait, &wait); 1187 1188 if (unlikely(ret <= 0)) 1189 break; 1190 1191 ret = -EFAULT; 1192 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) { 1193 dprintk("aio: lost an event due to EFAULT.\n"); 1194 break; 1195 } 1196 1197 /* Good, event copied to userland, update counts. */ 1198 event ++; 1199 i ++; 1200 } 1201 1202 if (timeout) 1203 clear_timeout(&to); 1204 out: 1205 destroy_timer_on_stack(&to.timer); 1206 return i ? i : ret; 1207 } 1208 1209 /* Take an ioctx and remove it from the list of ioctx's. Protects 1210 * against races with itself via ->dead. 1211 */ 1212 static void io_destroy(struct kioctx *ioctx) 1213 { 1214 struct mm_struct *mm = current->mm; 1215 int was_dead; 1216 1217 /* delete the entry from the list is someone else hasn't already */ 1218 spin_lock(&mm->ioctx_lock); 1219 was_dead = ioctx->dead; 1220 ioctx->dead = 1; 1221 hlist_del_rcu(&ioctx->list); 1222 spin_unlock(&mm->ioctx_lock); 1223 1224 dprintk("aio_release(%p)\n", ioctx); 1225 if (likely(!was_dead)) 1226 put_ioctx(ioctx); /* twice for the list */ 1227 1228 aio_cancel_all(ioctx); 1229 wait_for_all_aios(ioctx); 1230 1231 /* 1232 * Wake up any waiters. The setting of ctx->dead must be seen 1233 * by other CPUs at this point. Right now, we rely on the 1234 * locking done by the above calls to ensure this consistency. 1235 */ 1236 wake_up(&ioctx->wait); 1237 put_ioctx(ioctx); /* once for the lookup */ 1238 } 1239 1240 /* sys_io_setup: 1241 * Create an aio_context capable of receiving at least nr_events. 1242 * ctxp must not point to an aio_context that already exists, and 1243 * must be initialized to 0 prior to the call. On successful 1244 * creation of the aio_context, *ctxp is filled in with the resulting 1245 * handle. May fail with -EINVAL if *ctxp is not initialized, 1246 * if the specified nr_events exceeds internal limits. May fail 1247 * with -EAGAIN if the specified nr_events exceeds the user's limit 1248 * of available events. May fail with -ENOMEM if insufficient kernel 1249 * resources are available. May fail with -EFAULT if an invalid 1250 * pointer is passed for ctxp. Will fail with -ENOSYS if not 1251 * implemented. 1252 */ 1253 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp) 1254 { 1255 struct kioctx *ioctx = NULL; 1256 unsigned long ctx; 1257 long ret; 1258 1259 ret = get_user(ctx, ctxp); 1260 if (unlikely(ret)) 1261 goto out; 1262 1263 ret = -EINVAL; 1264 if (unlikely(ctx || nr_events == 0)) { 1265 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n", 1266 ctx, nr_events); 1267 goto out; 1268 } 1269 1270 ioctx = ioctx_alloc(nr_events); 1271 ret = PTR_ERR(ioctx); 1272 if (!IS_ERR(ioctx)) { 1273 ret = put_user(ioctx->user_id, ctxp); 1274 if (!ret) 1275 return 0; 1276 1277 get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */ 1278 io_destroy(ioctx); 1279 } 1280 1281 out: 1282 return ret; 1283 } 1284 1285 /* sys_io_destroy: 1286 * Destroy the aio_context specified. May cancel any outstanding 1287 * AIOs and block on completion. Will fail with -ENOSYS if not 1288 * implemented. May fail with -EFAULT if the context pointed to 1289 * is invalid. 1290 */ 1291 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx) 1292 { 1293 struct kioctx *ioctx = lookup_ioctx(ctx); 1294 if (likely(NULL != ioctx)) { 1295 io_destroy(ioctx); 1296 return 0; 1297 } 1298 pr_debug("EINVAL: io_destroy: invalid context id\n"); 1299 return -EINVAL; 1300 } 1301 1302 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret) 1303 { 1304 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg]; 1305 1306 BUG_ON(ret <= 0); 1307 1308 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) { 1309 ssize_t this = min((ssize_t)iov->iov_len, ret); 1310 iov->iov_base += this; 1311 iov->iov_len -= this; 1312 iocb->ki_left -= this; 1313 ret -= this; 1314 if (iov->iov_len == 0) { 1315 iocb->ki_cur_seg++; 1316 iov++; 1317 } 1318 } 1319 1320 /* the caller should not have done more io than what fit in 1321 * the remaining iovecs */ 1322 BUG_ON(ret > 0 && iocb->ki_left == 0); 1323 } 1324 1325 static ssize_t aio_rw_vect_retry(struct kiocb *iocb) 1326 { 1327 struct file *file = iocb->ki_filp; 1328 struct address_space *mapping = file->f_mapping; 1329 struct inode *inode = mapping->host; 1330 ssize_t (*rw_op)(struct kiocb *, const struct iovec *, 1331 unsigned long, loff_t); 1332 ssize_t ret = 0; 1333 unsigned short opcode; 1334 1335 if ((iocb->ki_opcode == IOCB_CMD_PREADV) || 1336 (iocb->ki_opcode == IOCB_CMD_PREAD)) { 1337 rw_op = file->f_op->aio_read; 1338 opcode = IOCB_CMD_PREADV; 1339 } else { 1340 rw_op = file->f_op->aio_write; 1341 opcode = IOCB_CMD_PWRITEV; 1342 } 1343 1344 /* This matches the pread()/pwrite() logic */ 1345 if (iocb->ki_pos < 0) 1346 return -EINVAL; 1347 1348 do { 1349 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg], 1350 iocb->ki_nr_segs - iocb->ki_cur_seg, 1351 iocb->ki_pos); 1352 if (ret > 0) 1353 aio_advance_iovec(iocb, ret); 1354 1355 /* retry all partial writes. retry partial reads as long as its a 1356 * regular file. */ 1357 } while (ret > 0 && iocb->ki_left > 0 && 1358 (opcode == IOCB_CMD_PWRITEV || 1359 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode)))); 1360 1361 /* This means we must have transferred all that we could */ 1362 /* No need to retry anymore */ 1363 if ((ret == 0) || (iocb->ki_left == 0)) 1364 ret = iocb->ki_nbytes - iocb->ki_left; 1365 1366 /* If we managed to write some out we return that, rather than 1367 * the eventual error. */ 1368 if (opcode == IOCB_CMD_PWRITEV 1369 && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY 1370 && iocb->ki_nbytes - iocb->ki_left) 1371 ret = iocb->ki_nbytes - iocb->ki_left; 1372 1373 return ret; 1374 } 1375 1376 static ssize_t aio_fdsync(struct kiocb *iocb) 1377 { 1378 struct file *file = iocb->ki_filp; 1379 ssize_t ret = -EINVAL; 1380 1381 if (file->f_op->aio_fsync) 1382 ret = file->f_op->aio_fsync(iocb, 1); 1383 return ret; 1384 } 1385 1386 static ssize_t aio_fsync(struct kiocb *iocb) 1387 { 1388 struct file *file = iocb->ki_filp; 1389 ssize_t ret = -EINVAL; 1390 1391 if (file->f_op->aio_fsync) 1392 ret = file->f_op->aio_fsync(iocb, 0); 1393 return ret; 1394 } 1395 1396 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb) 1397 { 1398 ssize_t ret; 1399 1400 ret = rw_copy_check_uvector(type, (struct iovec __user *)kiocb->ki_buf, 1401 kiocb->ki_nbytes, 1, 1402 &kiocb->ki_inline_vec, &kiocb->ki_iovec); 1403 if (ret < 0) 1404 goto out; 1405 1406 kiocb->ki_nr_segs = kiocb->ki_nbytes; 1407 kiocb->ki_cur_seg = 0; 1408 /* ki_nbytes/left now reflect bytes instead of segs */ 1409 kiocb->ki_nbytes = ret; 1410 kiocb->ki_left = ret; 1411 1412 ret = 0; 1413 out: 1414 return ret; 1415 } 1416 1417 static ssize_t aio_setup_single_vector(struct kiocb *kiocb) 1418 { 1419 kiocb->ki_iovec = &kiocb->ki_inline_vec; 1420 kiocb->ki_iovec->iov_base = kiocb->ki_buf; 1421 kiocb->ki_iovec->iov_len = kiocb->ki_left; 1422 kiocb->ki_nr_segs = 1; 1423 kiocb->ki_cur_seg = 0; 1424 return 0; 1425 } 1426 1427 /* 1428 * aio_setup_iocb: 1429 * Performs the initial checks and aio retry method 1430 * setup for the kiocb at the time of io submission. 1431 */ 1432 static ssize_t aio_setup_iocb(struct kiocb *kiocb) 1433 { 1434 struct file *file = kiocb->ki_filp; 1435 ssize_t ret = 0; 1436 1437 switch (kiocb->ki_opcode) { 1438 case IOCB_CMD_PREAD: 1439 ret = -EBADF; 1440 if (unlikely(!(file->f_mode & FMODE_READ))) 1441 break; 1442 ret = -EFAULT; 1443 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf, 1444 kiocb->ki_left))) 1445 break; 1446 ret = security_file_permission(file, MAY_READ); 1447 if (unlikely(ret)) 1448 break; 1449 ret = aio_setup_single_vector(kiocb); 1450 if (ret) 1451 break; 1452 ret = -EINVAL; 1453 if (file->f_op->aio_read) 1454 kiocb->ki_retry = aio_rw_vect_retry; 1455 break; 1456 case IOCB_CMD_PWRITE: 1457 ret = -EBADF; 1458 if (unlikely(!(file->f_mode & FMODE_WRITE))) 1459 break; 1460 ret = -EFAULT; 1461 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf, 1462 kiocb->ki_left))) 1463 break; 1464 ret = security_file_permission(file, MAY_WRITE); 1465 if (unlikely(ret)) 1466 break; 1467 ret = aio_setup_single_vector(kiocb); 1468 if (ret) 1469 break; 1470 ret = -EINVAL; 1471 if (file->f_op->aio_write) 1472 kiocb->ki_retry = aio_rw_vect_retry; 1473 break; 1474 case IOCB_CMD_PREADV: 1475 ret = -EBADF; 1476 if (unlikely(!(file->f_mode & FMODE_READ))) 1477 break; 1478 ret = security_file_permission(file, MAY_READ); 1479 if (unlikely(ret)) 1480 break; 1481 ret = aio_setup_vectored_rw(READ, kiocb); 1482 if (ret) 1483 break; 1484 ret = -EINVAL; 1485 if (file->f_op->aio_read) 1486 kiocb->ki_retry = aio_rw_vect_retry; 1487 break; 1488 case IOCB_CMD_PWRITEV: 1489 ret = -EBADF; 1490 if (unlikely(!(file->f_mode & FMODE_WRITE))) 1491 break; 1492 ret = security_file_permission(file, MAY_WRITE); 1493 if (unlikely(ret)) 1494 break; 1495 ret = aio_setup_vectored_rw(WRITE, kiocb); 1496 if (ret) 1497 break; 1498 ret = -EINVAL; 1499 if (file->f_op->aio_write) 1500 kiocb->ki_retry = aio_rw_vect_retry; 1501 break; 1502 case IOCB_CMD_FDSYNC: 1503 ret = -EINVAL; 1504 if (file->f_op->aio_fsync) 1505 kiocb->ki_retry = aio_fdsync; 1506 break; 1507 case IOCB_CMD_FSYNC: 1508 ret = -EINVAL; 1509 if (file->f_op->aio_fsync) 1510 kiocb->ki_retry = aio_fsync; 1511 break; 1512 default: 1513 dprintk("EINVAL: io_submit: no operation provided\n"); 1514 ret = -EINVAL; 1515 } 1516 1517 if (!kiocb->ki_retry) 1518 return ret; 1519 1520 return 0; 1521 } 1522 1523 /* 1524 * aio_wake_function: 1525 * wait queue callback function for aio notification, 1526 * Simply triggers a retry of the operation via kick_iocb. 1527 * 1528 * This callback is specified in the wait queue entry in 1529 * a kiocb. 1530 * 1531 * Note: 1532 * This routine is executed with the wait queue lock held. 1533 * Since kick_iocb acquires iocb->ctx->ctx_lock, it nests 1534 * the ioctx lock inside the wait queue lock. This is safe 1535 * because this callback isn't used for wait queues which 1536 * are nested inside ioctx lock (i.e. ctx->wait) 1537 */ 1538 static int aio_wake_function(wait_queue_t *wait, unsigned mode, 1539 int sync, void *key) 1540 { 1541 struct kiocb *iocb = container_of(wait, struct kiocb, ki_wait); 1542 1543 list_del_init(&wait->task_list); 1544 kick_iocb(iocb); 1545 return 1; 1546 } 1547 1548 static void aio_batch_add(struct address_space *mapping, 1549 struct hlist_head *batch_hash) 1550 { 1551 struct aio_batch_entry *abe; 1552 struct hlist_node *pos; 1553 unsigned bucket; 1554 1555 bucket = hash_ptr(mapping, AIO_BATCH_HASH_BITS); 1556 hlist_for_each_entry(abe, pos, &batch_hash[bucket], list) { 1557 if (abe->mapping == mapping) 1558 return; 1559 } 1560 1561 abe = mempool_alloc(abe_pool, GFP_KERNEL); 1562 BUG_ON(!igrab(mapping->host)); 1563 abe->mapping = mapping; 1564 hlist_add_head(&abe->list, &batch_hash[bucket]); 1565 return; 1566 } 1567 1568 static void aio_batch_free(struct hlist_head *batch_hash) 1569 { 1570 struct aio_batch_entry *abe; 1571 struct hlist_node *pos, *n; 1572 int i; 1573 1574 for (i = 0; i < AIO_BATCH_HASH_SIZE; i++) { 1575 hlist_for_each_entry_safe(abe, pos, n, &batch_hash[i], list) { 1576 blk_run_address_space(abe->mapping); 1577 iput(abe->mapping->host); 1578 hlist_del(&abe->list); 1579 mempool_free(abe, abe_pool); 1580 } 1581 } 1582 } 1583 1584 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb, 1585 struct iocb *iocb, struct hlist_head *batch_hash) 1586 { 1587 struct kiocb *req; 1588 struct file *file; 1589 ssize_t ret; 1590 1591 /* enforce forwards compatibility on users */ 1592 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) { 1593 pr_debug("EINVAL: io_submit: reserve field set\n"); 1594 return -EINVAL; 1595 } 1596 1597 /* prevent overflows */ 1598 if (unlikely( 1599 (iocb->aio_buf != (unsigned long)iocb->aio_buf) || 1600 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) || 1601 ((ssize_t)iocb->aio_nbytes < 0) 1602 )) { 1603 pr_debug("EINVAL: io_submit: overflow check\n"); 1604 return -EINVAL; 1605 } 1606 1607 file = fget(iocb->aio_fildes); 1608 if (unlikely(!file)) 1609 return -EBADF; 1610 1611 req = aio_get_req(ctx); /* returns with 2 references to req */ 1612 if (unlikely(!req)) { 1613 fput(file); 1614 return -EAGAIN; 1615 } 1616 req->ki_filp = file; 1617 if (iocb->aio_flags & IOCB_FLAG_RESFD) { 1618 /* 1619 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an 1620 * instance of the file* now. The file descriptor must be 1621 * an eventfd() fd, and will be signaled for each completed 1622 * event using the eventfd_signal() function. 1623 */ 1624 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd); 1625 if (IS_ERR(req->ki_eventfd)) { 1626 ret = PTR_ERR(req->ki_eventfd); 1627 req->ki_eventfd = NULL; 1628 goto out_put_req; 1629 } 1630 } 1631 1632 ret = put_user(req->ki_key, &user_iocb->aio_key); 1633 if (unlikely(ret)) { 1634 dprintk("EFAULT: aio_key\n"); 1635 goto out_put_req; 1636 } 1637 1638 req->ki_obj.user = user_iocb; 1639 req->ki_user_data = iocb->aio_data; 1640 req->ki_pos = iocb->aio_offset; 1641 1642 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf; 1643 req->ki_left = req->ki_nbytes = iocb->aio_nbytes; 1644 req->ki_opcode = iocb->aio_lio_opcode; 1645 init_waitqueue_func_entry(&req->ki_wait, aio_wake_function); 1646 INIT_LIST_HEAD(&req->ki_wait.task_list); 1647 1648 ret = aio_setup_iocb(req); 1649 1650 if (ret) 1651 goto out_put_req; 1652 1653 spin_lock_irq(&ctx->ctx_lock); 1654 aio_run_iocb(req); 1655 if (!list_empty(&ctx->run_list)) { 1656 /* drain the run list */ 1657 while (__aio_run_iocbs(ctx)) 1658 ; 1659 } 1660 spin_unlock_irq(&ctx->ctx_lock); 1661 if (req->ki_opcode == IOCB_CMD_PREAD || 1662 req->ki_opcode == IOCB_CMD_PREADV || 1663 req->ki_opcode == IOCB_CMD_PWRITE || 1664 req->ki_opcode == IOCB_CMD_PWRITEV) 1665 aio_batch_add(file->f_mapping, batch_hash); 1666 1667 aio_put_req(req); /* drop extra ref to req */ 1668 return 0; 1669 1670 out_put_req: 1671 aio_put_req(req); /* drop extra ref to req */ 1672 aio_put_req(req); /* drop i/o ref to req */ 1673 return ret; 1674 } 1675 1676 /* sys_io_submit: 1677 * Queue the nr iocbs pointed to by iocbpp for processing. Returns 1678 * the number of iocbs queued. May return -EINVAL if the aio_context 1679 * specified by ctx_id is invalid, if nr is < 0, if the iocb at 1680 * *iocbpp[0] is not properly initialized, if the operation specified 1681 * is invalid for the file descriptor in the iocb. May fail with 1682 * -EFAULT if any of the data structures point to invalid data. May 1683 * fail with -EBADF if the file descriptor specified in the first 1684 * iocb is invalid. May fail with -EAGAIN if insufficient resources 1685 * are available to queue any iocbs. Will return 0 if nr is 0. Will 1686 * fail with -ENOSYS if not implemented. 1687 */ 1688 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr, 1689 struct iocb __user * __user *, iocbpp) 1690 { 1691 struct kioctx *ctx; 1692 long ret = 0; 1693 int i; 1694 struct hlist_head batch_hash[AIO_BATCH_HASH_SIZE] = { { 0, }, }; 1695 1696 if (unlikely(nr < 0)) 1697 return -EINVAL; 1698 1699 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp))))) 1700 return -EFAULT; 1701 1702 ctx = lookup_ioctx(ctx_id); 1703 if (unlikely(!ctx)) { 1704 pr_debug("EINVAL: io_submit: invalid context id\n"); 1705 return -EINVAL; 1706 } 1707 1708 /* 1709 * AKPM: should this return a partial result if some of the IOs were 1710 * successfully submitted? 1711 */ 1712 for (i=0; i<nr; i++) { 1713 struct iocb __user *user_iocb; 1714 struct iocb tmp; 1715 1716 if (unlikely(__get_user(user_iocb, iocbpp + i))) { 1717 ret = -EFAULT; 1718 break; 1719 } 1720 1721 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) { 1722 ret = -EFAULT; 1723 break; 1724 } 1725 1726 ret = io_submit_one(ctx, user_iocb, &tmp, batch_hash); 1727 if (ret) 1728 break; 1729 } 1730 aio_batch_free(batch_hash); 1731 1732 put_ioctx(ctx); 1733 return i ? i : ret; 1734 } 1735 1736 /* lookup_kiocb 1737 * Finds a given iocb for cancellation. 1738 */ 1739 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, 1740 u32 key) 1741 { 1742 struct list_head *pos; 1743 1744 assert_spin_locked(&ctx->ctx_lock); 1745 1746 /* TODO: use a hash or array, this sucks. */ 1747 list_for_each(pos, &ctx->active_reqs) { 1748 struct kiocb *kiocb = list_kiocb(pos); 1749 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key) 1750 return kiocb; 1751 } 1752 return NULL; 1753 } 1754 1755 /* sys_io_cancel: 1756 * Attempts to cancel an iocb previously passed to io_submit. If 1757 * the operation is successfully cancelled, the resulting event is 1758 * copied into the memory pointed to by result without being placed 1759 * into the completion queue and 0 is returned. May fail with 1760 * -EFAULT if any of the data structures pointed to are invalid. 1761 * May fail with -EINVAL if aio_context specified by ctx_id is 1762 * invalid. May fail with -EAGAIN if the iocb specified was not 1763 * cancelled. Will fail with -ENOSYS if not implemented. 1764 */ 1765 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb, 1766 struct io_event __user *, result) 1767 { 1768 int (*cancel)(struct kiocb *iocb, struct io_event *res); 1769 struct kioctx *ctx; 1770 struct kiocb *kiocb; 1771 u32 key; 1772 int ret; 1773 1774 ret = get_user(key, &iocb->aio_key); 1775 if (unlikely(ret)) 1776 return -EFAULT; 1777 1778 ctx = lookup_ioctx(ctx_id); 1779 if (unlikely(!ctx)) 1780 return -EINVAL; 1781 1782 spin_lock_irq(&ctx->ctx_lock); 1783 ret = -EAGAIN; 1784 kiocb = lookup_kiocb(ctx, iocb, key); 1785 if (kiocb && kiocb->ki_cancel) { 1786 cancel = kiocb->ki_cancel; 1787 kiocb->ki_users ++; 1788 kiocbSetCancelled(kiocb); 1789 } else 1790 cancel = NULL; 1791 spin_unlock_irq(&ctx->ctx_lock); 1792 1793 if (NULL != cancel) { 1794 struct io_event tmp; 1795 pr_debug("calling cancel\n"); 1796 memset(&tmp, 0, sizeof(tmp)); 1797 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user; 1798 tmp.data = kiocb->ki_user_data; 1799 ret = cancel(kiocb, &tmp); 1800 if (!ret) { 1801 /* Cancellation succeeded -- copy the result 1802 * into the user's buffer. 1803 */ 1804 if (copy_to_user(result, &tmp, sizeof(tmp))) 1805 ret = -EFAULT; 1806 } 1807 } else 1808 ret = -EINVAL; 1809 1810 put_ioctx(ctx); 1811 1812 return ret; 1813 } 1814 1815 /* io_getevents: 1816 * Attempts to read at least min_nr events and up to nr events from 1817 * the completion queue for the aio_context specified by ctx_id. May 1818 * fail with -EINVAL if ctx_id is invalid, if min_nr is out of range, 1819 * if nr is out of range, if when is out of range. May fail with 1820 * -EFAULT if any of the memory specified to is invalid. May return 1821 * 0 or < min_nr if no events are available and the timeout specified 1822 * by when has elapsed, where when == NULL specifies an infinite 1823 * timeout. Note that the timeout pointed to by when is relative and 1824 * will be updated if not NULL and the operation blocks. Will fail 1825 * with -ENOSYS if not implemented. 1826 */ 1827 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id, 1828 long, min_nr, 1829 long, nr, 1830 struct io_event __user *, events, 1831 struct timespec __user *, timeout) 1832 { 1833 struct kioctx *ioctx = lookup_ioctx(ctx_id); 1834 long ret = -EINVAL; 1835 1836 if (likely(ioctx)) { 1837 if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0)) 1838 ret = read_events(ioctx, min_nr, nr, events, timeout); 1839 put_ioctx(ioctx); 1840 } 1841 1842 asmlinkage_protect(5, ret, ctx_id, min_nr, nr, events, timeout); 1843 return ret; 1844 } 1845