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/uio.h> 19 20 #define DEBUG 0 21 22 #include <linux/sched.h> 23 #include <linux/fs.h> 24 #include <linux/file.h> 25 #include <linux/mm.h> 26 #include <linux/mman.h> 27 #include <linux/mmu_context.h> 28 #include <linux/slab.h> 29 #include <linux/timer.h> 30 #include <linux/aio.h> 31 #include <linux/highmem.h> 32 #include <linux/workqueue.h> 33 #include <linux/security.h> 34 #include <linux/eventfd.h> 35 36 #include <asm/kmap_types.h> 37 #include <asm/uaccess.h> 38 39 #if DEBUG > 1 40 #define dprintk printk 41 #else 42 #define dprintk(x...) do { ; } while (0) 43 #endif 44 45 /*------ sysctl variables----*/ 46 static DEFINE_SPINLOCK(aio_nr_lock); 47 unsigned long aio_nr; /* current system wide number of aio requests */ 48 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */ 49 /*----end sysctl variables---*/ 50 51 static struct kmem_cache *kiocb_cachep; 52 static struct kmem_cache *kioctx_cachep; 53 54 static struct workqueue_struct *aio_wq; 55 56 /* Used for rare fput completion. */ 57 static void aio_fput_routine(struct work_struct *); 58 static DECLARE_WORK(fput_work, aio_fput_routine); 59 60 static DEFINE_SPINLOCK(fput_lock); 61 static LIST_HEAD(fput_head); 62 63 static void aio_kick_handler(struct work_struct *); 64 static void aio_queue_work(struct kioctx *); 65 66 /* aio_setup 67 * Creates the slab caches used by the aio routines, panic on 68 * failure as this is done early during the boot sequence. 69 */ 70 static int __init aio_setup(void) 71 { 72 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC); 73 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC); 74 75 aio_wq = create_workqueue("aio"); 76 77 pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page)); 78 79 return 0; 80 } 81 __initcall(aio_setup); 82 83 static void aio_free_ring(struct kioctx *ctx) 84 { 85 struct aio_ring_info *info = &ctx->ring_info; 86 long i; 87 88 for (i=0; i<info->nr_pages; i++) 89 put_page(info->ring_pages[i]); 90 91 if (info->mmap_size) { 92 down_write(&ctx->mm->mmap_sem); 93 do_munmap(ctx->mm, info->mmap_base, info->mmap_size); 94 up_write(&ctx->mm->mmap_sem); 95 } 96 97 if (info->ring_pages && info->ring_pages != info->internal_pages) 98 kfree(info->ring_pages); 99 info->ring_pages = NULL; 100 info->nr = 0; 101 } 102 103 static int aio_setup_ring(struct kioctx *ctx) 104 { 105 struct aio_ring *ring; 106 struct aio_ring_info *info = &ctx->ring_info; 107 unsigned nr_events = ctx->max_reqs; 108 unsigned long size; 109 int nr_pages; 110 111 /* Compensate for the ring buffer's head/tail overlap entry */ 112 nr_events += 2; /* 1 is required, 2 for good luck */ 113 114 size = sizeof(struct aio_ring); 115 size += sizeof(struct io_event) * nr_events; 116 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT; 117 118 if (nr_pages < 0) 119 return -EINVAL; 120 121 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event); 122 123 info->nr = 0; 124 info->ring_pages = info->internal_pages; 125 if (nr_pages > AIO_RING_PAGES) { 126 info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL); 127 if (!info->ring_pages) 128 return -ENOMEM; 129 } 130 131 info->mmap_size = nr_pages * PAGE_SIZE; 132 dprintk("attempting mmap of %lu bytes\n", info->mmap_size); 133 down_write(&ctx->mm->mmap_sem); 134 info->mmap_base = do_mmap(NULL, 0, info->mmap_size, 135 PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE, 136 0); 137 if (IS_ERR((void *)info->mmap_base)) { 138 up_write(&ctx->mm->mmap_sem); 139 info->mmap_size = 0; 140 aio_free_ring(ctx); 141 return -EAGAIN; 142 } 143 144 dprintk("mmap address: 0x%08lx\n", info->mmap_base); 145 info->nr_pages = get_user_pages(current, ctx->mm, 146 info->mmap_base, nr_pages, 147 1, 0, info->ring_pages, NULL); 148 up_write(&ctx->mm->mmap_sem); 149 150 if (unlikely(info->nr_pages != nr_pages)) { 151 aio_free_ring(ctx); 152 return -EAGAIN; 153 } 154 155 ctx->user_id = info->mmap_base; 156 157 info->nr = nr_events; /* trusted copy */ 158 159 ring = kmap_atomic(info->ring_pages[0], KM_USER0); 160 ring->nr = nr_events; /* user copy */ 161 ring->id = ctx->user_id; 162 ring->head = ring->tail = 0; 163 ring->magic = AIO_RING_MAGIC; 164 ring->compat_features = AIO_RING_COMPAT_FEATURES; 165 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES; 166 ring->header_length = sizeof(struct aio_ring); 167 kunmap_atomic(ring, KM_USER0); 168 169 return 0; 170 } 171 172 173 /* aio_ring_event: returns a pointer to the event at the given index from 174 * kmap_atomic(, km). Release the pointer with put_aio_ring_event(); 175 */ 176 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event)) 177 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event)) 178 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE) 179 180 #define aio_ring_event(info, nr, km) ({ \ 181 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \ 182 struct io_event *__event; \ 183 __event = kmap_atomic( \ 184 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \ 185 __event += pos % AIO_EVENTS_PER_PAGE; \ 186 __event; \ 187 }) 188 189 #define put_aio_ring_event(event, km) do { \ 190 struct io_event *__event = (event); \ 191 (void)__event; \ 192 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \ 193 } while(0) 194 195 static void ctx_rcu_free(struct rcu_head *head) 196 { 197 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head); 198 unsigned nr_events = ctx->max_reqs; 199 200 kmem_cache_free(kioctx_cachep, ctx); 201 202 if (nr_events) { 203 spin_lock(&aio_nr_lock); 204 BUG_ON(aio_nr - nr_events > aio_nr); 205 aio_nr -= nr_events; 206 spin_unlock(&aio_nr_lock); 207 } 208 } 209 210 /* __put_ioctx 211 * Called when the last user of an aio context has gone away, 212 * and the struct needs to be freed. 213 */ 214 static void __put_ioctx(struct kioctx *ctx) 215 { 216 BUG_ON(ctx->reqs_active); 217 218 cancel_delayed_work(&ctx->wq); 219 cancel_work_sync(&ctx->wq.work); 220 aio_free_ring(ctx); 221 mmdrop(ctx->mm); 222 ctx->mm = NULL; 223 pr_debug("__put_ioctx: freeing %p\n", ctx); 224 call_rcu(&ctx->rcu_head, ctx_rcu_free); 225 } 226 227 #define get_ioctx(kioctx) do { \ 228 BUG_ON(atomic_read(&(kioctx)->users) <= 0); \ 229 atomic_inc(&(kioctx)->users); \ 230 } while (0) 231 #define put_ioctx(kioctx) do { \ 232 BUG_ON(atomic_read(&(kioctx)->users) <= 0); \ 233 if (unlikely(atomic_dec_and_test(&(kioctx)->users))) \ 234 __put_ioctx(kioctx); \ 235 } while (0) 236 237 /* ioctx_alloc 238 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed. 239 */ 240 static struct kioctx *ioctx_alloc(unsigned nr_events) 241 { 242 struct mm_struct *mm; 243 struct kioctx *ctx; 244 int did_sync = 0; 245 246 /* Prevent overflows */ 247 if ((nr_events > (0x10000000U / sizeof(struct io_event))) || 248 (nr_events > (0x10000000U / sizeof(struct kiocb)))) { 249 pr_debug("ENOMEM: nr_events too high\n"); 250 return ERR_PTR(-EINVAL); 251 } 252 253 if ((unsigned long)nr_events > aio_max_nr) 254 return ERR_PTR(-EAGAIN); 255 256 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL); 257 if (!ctx) 258 return ERR_PTR(-ENOMEM); 259 260 ctx->max_reqs = nr_events; 261 mm = ctx->mm = current->mm; 262 atomic_inc(&mm->mm_count); 263 264 atomic_set(&ctx->users, 1); 265 spin_lock_init(&ctx->ctx_lock); 266 spin_lock_init(&ctx->ring_info.ring_lock); 267 init_waitqueue_head(&ctx->wait); 268 269 INIT_LIST_HEAD(&ctx->active_reqs); 270 INIT_LIST_HEAD(&ctx->run_list); 271 INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler); 272 273 if (aio_setup_ring(ctx) < 0) 274 goto out_freectx; 275 276 /* limit the number of system wide aios */ 277 do { 278 spin_lock_bh(&aio_nr_lock); 279 if (aio_nr + nr_events > aio_max_nr || 280 aio_nr + nr_events < aio_nr) 281 ctx->max_reqs = 0; 282 else 283 aio_nr += ctx->max_reqs; 284 spin_unlock_bh(&aio_nr_lock); 285 if (ctx->max_reqs || did_sync) 286 break; 287 288 /* wait for rcu callbacks to have completed before giving up */ 289 synchronize_rcu(); 290 did_sync = 1; 291 ctx->max_reqs = nr_events; 292 } while (1); 293 294 if (ctx->max_reqs == 0) 295 goto out_cleanup; 296 297 /* now link into global list. */ 298 spin_lock(&mm->ioctx_lock); 299 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list); 300 spin_unlock(&mm->ioctx_lock); 301 302 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n", 303 ctx, ctx->user_id, current->mm, ctx->ring_info.nr); 304 return ctx; 305 306 out_cleanup: 307 __put_ioctx(ctx); 308 return ERR_PTR(-EAGAIN); 309 310 out_freectx: 311 mmdrop(mm); 312 kmem_cache_free(kioctx_cachep, ctx); 313 ctx = ERR_PTR(-ENOMEM); 314 315 dprintk("aio: error allocating ioctx %p\n", ctx); 316 return ctx; 317 } 318 319 /* aio_cancel_all 320 * Cancels all outstanding aio requests on an aio context. Used 321 * when the processes owning a context have all exited to encourage 322 * the rapid destruction of the kioctx. 323 */ 324 static void aio_cancel_all(struct kioctx *ctx) 325 { 326 int (*cancel)(struct kiocb *, struct io_event *); 327 struct io_event res; 328 spin_lock_irq(&ctx->ctx_lock); 329 ctx->dead = 1; 330 while (!list_empty(&ctx->active_reqs)) { 331 struct list_head *pos = ctx->active_reqs.next; 332 struct kiocb *iocb = list_kiocb(pos); 333 list_del_init(&iocb->ki_list); 334 cancel = iocb->ki_cancel; 335 kiocbSetCancelled(iocb); 336 if (cancel) { 337 iocb->ki_users++; 338 spin_unlock_irq(&ctx->ctx_lock); 339 cancel(iocb, &res); 340 spin_lock_irq(&ctx->ctx_lock); 341 } 342 } 343 spin_unlock_irq(&ctx->ctx_lock); 344 } 345 346 static void wait_for_all_aios(struct kioctx *ctx) 347 { 348 struct task_struct *tsk = current; 349 DECLARE_WAITQUEUE(wait, tsk); 350 351 spin_lock_irq(&ctx->ctx_lock); 352 if (!ctx->reqs_active) 353 goto out; 354 355 add_wait_queue(&ctx->wait, &wait); 356 set_task_state(tsk, TASK_UNINTERRUPTIBLE); 357 while (ctx->reqs_active) { 358 spin_unlock_irq(&ctx->ctx_lock); 359 io_schedule(); 360 set_task_state(tsk, TASK_UNINTERRUPTIBLE); 361 spin_lock_irq(&ctx->ctx_lock); 362 } 363 __set_task_state(tsk, TASK_RUNNING); 364 remove_wait_queue(&ctx->wait, &wait); 365 366 out: 367 spin_unlock_irq(&ctx->ctx_lock); 368 } 369 370 /* wait_on_sync_kiocb: 371 * Waits on the given sync kiocb to complete. 372 */ 373 ssize_t wait_on_sync_kiocb(struct kiocb *iocb) 374 { 375 while (iocb->ki_users) { 376 set_current_state(TASK_UNINTERRUPTIBLE); 377 if (!iocb->ki_users) 378 break; 379 io_schedule(); 380 } 381 __set_current_state(TASK_RUNNING); 382 return iocb->ki_user_data; 383 } 384 EXPORT_SYMBOL(wait_on_sync_kiocb); 385 386 /* exit_aio: called when the last user of mm goes away. At this point, 387 * there is no way for any new requests to be submited or any of the 388 * io_* syscalls to be called on the context. However, there may be 389 * outstanding requests which hold references to the context; as they 390 * go away, they will call put_ioctx and release any pinned memory 391 * associated with the request (held via struct page * references). 392 */ 393 void exit_aio(struct mm_struct *mm) 394 { 395 struct kioctx *ctx; 396 397 while (!hlist_empty(&mm->ioctx_list)) { 398 ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list); 399 hlist_del_rcu(&ctx->list); 400 401 aio_cancel_all(ctx); 402 403 wait_for_all_aios(ctx); 404 /* 405 * Ensure we don't leave the ctx on the aio_wq 406 */ 407 cancel_work_sync(&ctx->wq.work); 408 409 if (1 != atomic_read(&ctx->users)) 410 printk(KERN_DEBUG 411 "exit_aio:ioctx still alive: %d %d %d\n", 412 atomic_read(&ctx->users), ctx->dead, 413 ctx->reqs_active); 414 put_ioctx(ctx); 415 } 416 } 417 418 /* aio_get_req 419 * Allocate a slot for an aio request. Increments the users count 420 * of the kioctx so that the kioctx stays around until all requests are 421 * complete. Returns NULL if no requests are free. 422 * 423 * Returns with kiocb->users set to 2. The io submit code path holds 424 * an extra reference while submitting the i/o. 425 * This prevents races between the aio code path referencing the 426 * req (after submitting it) and aio_complete() freeing the req. 427 */ 428 static struct kiocb *__aio_get_req(struct kioctx *ctx) 429 { 430 struct kiocb *req = NULL; 431 struct aio_ring *ring; 432 int okay = 0; 433 434 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL); 435 if (unlikely(!req)) 436 return NULL; 437 438 req->ki_flags = 0; 439 req->ki_users = 2; 440 req->ki_key = 0; 441 req->ki_ctx = ctx; 442 req->ki_cancel = NULL; 443 req->ki_retry = NULL; 444 req->ki_dtor = NULL; 445 req->private = NULL; 446 req->ki_iovec = NULL; 447 INIT_LIST_HEAD(&req->ki_run_list); 448 req->ki_eventfd = NULL; 449 450 /* Check if the completion queue has enough free space to 451 * accept an event from this io. 452 */ 453 spin_lock_irq(&ctx->ctx_lock); 454 ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0); 455 if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) { 456 list_add(&req->ki_list, &ctx->active_reqs); 457 ctx->reqs_active++; 458 okay = 1; 459 } 460 kunmap_atomic(ring, KM_USER0); 461 spin_unlock_irq(&ctx->ctx_lock); 462 463 if (!okay) { 464 kmem_cache_free(kiocb_cachep, req); 465 req = NULL; 466 } 467 468 return req; 469 } 470 471 static inline struct kiocb *aio_get_req(struct kioctx *ctx) 472 { 473 struct kiocb *req; 474 /* Handle a potential starvation case -- should be exceedingly rare as 475 * requests will be stuck on fput_head only if the aio_fput_routine is 476 * delayed and the requests were the last user of the struct file. 477 */ 478 req = __aio_get_req(ctx); 479 if (unlikely(NULL == req)) { 480 aio_fput_routine(NULL); 481 req = __aio_get_req(ctx); 482 } 483 return req; 484 } 485 486 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req) 487 { 488 assert_spin_locked(&ctx->ctx_lock); 489 490 if (req->ki_eventfd != NULL) 491 eventfd_ctx_put(req->ki_eventfd); 492 if (req->ki_dtor) 493 req->ki_dtor(req); 494 if (req->ki_iovec != &req->ki_inline_vec) 495 kfree(req->ki_iovec); 496 kmem_cache_free(kiocb_cachep, req); 497 ctx->reqs_active--; 498 499 if (unlikely(!ctx->reqs_active && ctx->dead)) 500 wake_up(&ctx->wait); 501 } 502 503 static void aio_fput_routine(struct work_struct *data) 504 { 505 spin_lock_irq(&fput_lock); 506 while (likely(!list_empty(&fput_head))) { 507 struct kiocb *req = list_kiocb(fput_head.next); 508 struct kioctx *ctx = req->ki_ctx; 509 510 list_del(&req->ki_list); 511 spin_unlock_irq(&fput_lock); 512 513 /* Complete the fput(s) */ 514 if (req->ki_filp != NULL) 515 __fput(req->ki_filp); 516 517 /* Link the iocb into the context's free list */ 518 spin_lock_irq(&ctx->ctx_lock); 519 really_put_req(ctx, req); 520 spin_unlock_irq(&ctx->ctx_lock); 521 522 put_ioctx(ctx); 523 spin_lock_irq(&fput_lock); 524 } 525 spin_unlock_irq(&fput_lock); 526 } 527 528 /* __aio_put_req 529 * Returns true if this put was the last user of the request. 530 */ 531 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req) 532 { 533 dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n", 534 req, atomic_long_read(&req->ki_filp->f_count)); 535 536 assert_spin_locked(&ctx->ctx_lock); 537 538 req->ki_users--; 539 BUG_ON(req->ki_users < 0); 540 if (likely(req->ki_users)) 541 return 0; 542 list_del(&req->ki_list); /* remove from active_reqs */ 543 req->ki_cancel = NULL; 544 req->ki_retry = NULL; 545 546 /* 547 * Try to optimize the aio and eventfd file* puts, by avoiding to 548 * schedule work in case it is not __fput() time. In normal cases, 549 * we would not be holding the last reference to the file*, so 550 * this function will be executed w/out any aio kthread wakeup. 551 */ 552 if (unlikely(atomic_long_dec_and_test(&req->ki_filp->f_count))) { 553 get_ioctx(ctx); 554 spin_lock(&fput_lock); 555 list_add(&req->ki_list, &fput_head); 556 spin_unlock(&fput_lock); 557 queue_work(aio_wq, &fput_work); 558 } else { 559 req->ki_filp = NULL; 560 really_put_req(ctx, req); 561 } 562 return 1; 563 } 564 565 /* aio_put_req 566 * Returns true if this put was the last user of the kiocb, 567 * false if the request is still in use. 568 */ 569 int aio_put_req(struct kiocb *req) 570 { 571 struct kioctx *ctx = req->ki_ctx; 572 int ret; 573 spin_lock_irq(&ctx->ctx_lock); 574 ret = __aio_put_req(ctx, req); 575 spin_unlock_irq(&ctx->ctx_lock); 576 return ret; 577 } 578 EXPORT_SYMBOL(aio_put_req); 579 580 static struct kioctx *lookup_ioctx(unsigned long ctx_id) 581 { 582 struct mm_struct *mm = current->mm; 583 struct kioctx *ctx, *ret = NULL; 584 struct hlist_node *n; 585 586 rcu_read_lock(); 587 588 hlist_for_each_entry_rcu(ctx, n, &mm->ioctx_list, list) { 589 if (ctx->user_id == ctx_id && !ctx->dead) { 590 get_ioctx(ctx); 591 ret = ctx; 592 break; 593 } 594 } 595 596 rcu_read_unlock(); 597 return ret; 598 } 599 600 /* 601 * Queue up a kiocb to be retried. Assumes that the kiocb 602 * has already been marked as kicked, and places it on 603 * the retry run list for the corresponding ioctx, if it 604 * isn't already queued. Returns 1 if it actually queued 605 * the kiocb (to tell the caller to activate the work 606 * queue to process it), or 0, if it found that it was 607 * already queued. 608 */ 609 static inline int __queue_kicked_iocb(struct kiocb *iocb) 610 { 611 struct kioctx *ctx = iocb->ki_ctx; 612 613 assert_spin_locked(&ctx->ctx_lock); 614 615 if (list_empty(&iocb->ki_run_list)) { 616 list_add_tail(&iocb->ki_run_list, 617 &ctx->run_list); 618 return 1; 619 } 620 return 0; 621 } 622 623 /* aio_run_iocb 624 * This is the core aio execution routine. It is 625 * invoked both for initial i/o submission and 626 * subsequent retries via the aio_kick_handler. 627 * Expects to be invoked with iocb->ki_ctx->lock 628 * already held. The lock is released and reacquired 629 * as needed during processing. 630 * 631 * Calls the iocb retry method (already setup for the 632 * iocb on initial submission) for operation specific 633 * handling, but takes care of most of common retry 634 * execution details for a given iocb. The retry method 635 * needs to be non-blocking as far as possible, to avoid 636 * holding up other iocbs waiting to be serviced by the 637 * retry kernel thread. 638 * 639 * The trickier parts in this code have to do with 640 * ensuring that only one retry instance is in progress 641 * for a given iocb at any time. Providing that guarantee 642 * simplifies the coding of individual aio operations as 643 * it avoids various potential races. 644 */ 645 static ssize_t aio_run_iocb(struct kiocb *iocb) 646 { 647 struct kioctx *ctx = iocb->ki_ctx; 648 ssize_t (*retry)(struct kiocb *); 649 ssize_t ret; 650 651 if (!(retry = iocb->ki_retry)) { 652 printk("aio_run_iocb: iocb->ki_retry = NULL\n"); 653 return 0; 654 } 655 656 /* 657 * We don't want the next retry iteration for this 658 * operation to start until this one has returned and 659 * updated the iocb state. However, wait_queue functions 660 * can trigger a kick_iocb from interrupt context in the 661 * meantime, indicating that data is available for the next 662 * iteration. We want to remember that and enable the 663 * next retry iteration _after_ we are through with 664 * this one. 665 * 666 * So, in order to be able to register a "kick", but 667 * prevent it from being queued now, we clear the kick 668 * flag, but make the kick code *think* that the iocb is 669 * still on the run list until we are actually done. 670 * When we are done with this iteration, we check if 671 * the iocb was kicked in the meantime and if so, queue 672 * it up afresh. 673 */ 674 675 kiocbClearKicked(iocb); 676 677 /* 678 * This is so that aio_complete knows it doesn't need to 679 * pull the iocb off the run list (We can't just call 680 * INIT_LIST_HEAD because we don't want a kick_iocb to 681 * queue this on the run list yet) 682 */ 683 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL; 684 spin_unlock_irq(&ctx->ctx_lock); 685 686 /* Quit retrying if the i/o has been cancelled */ 687 if (kiocbIsCancelled(iocb)) { 688 ret = -EINTR; 689 aio_complete(iocb, ret, 0); 690 /* must not access the iocb after this */ 691 goto out; 692 } 693 694 /* 695 * Now we are all set to call the retry method in async 696 * context. 697 */ 698 ret = retry(iocb); 699 700 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) { 701 BUG_ON(!list_empty(&iocb->ki_wait.task_list)); 702 aio_complete(iocb, ret, 0); 703 } 704 out: 705 spin_lock_irq(&ctx->ctx_lock); 706 707 if (-EIOCBRETRY == ret) { 708 /* 709 * OK, now that we are done with this iteration 710 * and know that there is more left to go, 711 * this is where we let go so that a subsequent 712 * "kick" can start the next iteration 713 */ 714 715 /* will make __queue_kicked_iocb succeed from here on */ 716 INIT_LIST_HEAD(&iocb->ki_run_list); 717 /* we must queue the next iteration ourselves, if it 718 * has already been kicked */ 719 if (kiocbIsKicked(iocb)) { 720 __queue_kicked_iocb(iocb); 721 722 /* 723 * __queue_kicked_iocb will always return 1 here, because 724 * iocb->ki_run_list is empty at this point so it should 725 * be safe to unconditionally queue the context into the 726 * work queue. 727 */ 728 aio_queue_work(ctx); 729 } 730 } 731 return ret; 732 } 733 734 /* 735 * __aio_run_iocbs: 736 * Process all pending retries queued on the ioctx 737 * run list. 738 * Assumes it is operating within the aio issuer's mm 739 * context. 740 */ 741 static int __aio_run_iocbs(struct kioctx *ctx) 742 { 743 struct kiocb *iocb; 744 struct list_head run_list; 745 746 assert_spin_locked(&ctx->ctx_lock); 747 748 list_replace_init(&ctx->run_list, &run_list); 749 while (!list_empty(&run_list)) { 750 iocb = list_entry(run_list.next, struct kiocb, 751 ki_run_list); 752 list_del(&iocb->ki_run_list); 753 /* 754 * Hold an extra reference while retrying i/o. 755 */ 756 iocb->ki_users++; /* grab extra reference */ 757 aio_run_iocb(iocb); 758 __aio_put_req(ctx, iocb); 759 } 760 if (!list_empty(&ctx->run_list)) 761 return 1; 762 return 0; 763 } 764 765 static void aio_queue_work(struct kioctx * ctx) 766 { 767 unsigned long timeout; 768 /* 769 * if someone is waiting, get the work started right 770 * away, otherwise, use a longer delay 771 */ 772 smp_mb(); 773 if (waitqueue_active(&ctx->wait)) 774 timeout = 1; 775 else 776 timeout = HZ/10; 777 queue_delayed_work(aio_wq, &ctx->wq, timeout); 778 } 779 780 781 /* 782 * aio_run_iocbs: 783 * Process all pending retries queued on the ioctx 784 * run list. 785 * Assumes it is operating within the aio issuer's mm 786 * context. 787 */ 788 static inline void aio_run_iocbs(struct kioctx *ctx) 789 { 790 int requeue; 791 792 spin_lock_irq(&ctx->ctx_lock); 793 794 requeue = __aio_run_iocbs(ctx); 795 spin_unlock_irq(&ctx->ctx_lock); 796 if (requeue) 797 aio_queue_work(ctx); 798 } 799 800 /* 801 * just like aio_run_iocbs, but keeps running them until 802 * the list stays empty 803 */ 804 static inline void aio_run_all_iocbs(struct kioctx *ctx) 805 { 806 spin_lock_irq(&ctx->ctx_lock); 807 while (__aio_run_iocbs(ctx)) 808 ; 809 spin_unlock_irq(&ctx->ctx_lock); 810 } 811 812 /* 813 * aio_kick_handler: 814 * Work queue handler triggered to process pending 815 * retries on an ioctx. Takes on the aio issuer's 816 * mm context before running the iocbs, so that 817 * copy_xxx_user operates on the issuer's address 818 * space. 819 * Run on aiod's context. 820 */ 821 static void aio_kick_handler(struct work_struct *work) 822 { 823 struct kioctx *ctx = container_of(work, struct kioctx, wq.work); 824 mm_segment_t oldfs = get_fs(); 825 struct mm_struct *mm; 826 int requeue; 827 828 set_fs(USER_DS); 829 use_mm(ctx->mm); 830 spin_lock_irq(&ctx->ctx_lock); 831 requeue =__aio_run_iocbs(ctx); 832 mm = ctx->mm; 833 spin_unlock_irq(&ctx->ctx_lock); 834 unuse_mm(mm); 835 set_fs(oldfs); 836 /* 837 * we're in a worker thread already, don't use queue_delayed_work, 838 */ 839 if (requeue) 840 queue_delayed_work(aio_wq, &ctx->wq, 0); 841 } 842 843 844 /* 845 * Called by kick_iocb to queue the kiocb for retry 846 * and if required activate the aio work queue to process 847 * it 848 */ 849 static void try_queue_kicked_iocb(struct kiocb *iocb) 850 { 851 struct kioctx *ctx = iocb->ki_ctx; 852 unsigned long flags; 853 int run = 0; 854 855 /* We're supposed to be the only path putting the iocb back on the run 856 * list. If we find that the iocb is *back* on a wait queue already 857 * than retry has happened before we could queue the iocb. This also 858 * means that the retry could have completed and freed our iocb, no 859 * good. */ 860 BUG_ON((!list_empty(&iocb->ki_wait.task_list))); 861 862 spin_lock_irqsave(&ctx->ctx_lock, flags); 863 /* set this inside the lock so that we can't race with aio_run_iocb() 864 * testing it and putting the iocb on the run list under the lock */ 865 if (!kiocbTryKick(iocb)) 866 run = __queue_kicked_iocb(iocb); 867 spin_unlock_irqrestore(&ctx->ctx_lock, flags); 868 if (run) 869 aio_queue_work(ctx); 870 } 871 872 /* 873 * kick_iocb: 874 * Called typically from a wait queue callback context 875 * (aio_wake_function) to trigger a retry of the iocb. 876 * The retry is usually executed by aio workqueue 877 * threads (See aio_kick_handler). 878 */ 879 void kick_iocb(struct kiocb *iocb) 880 { 881 /* sync iocbs are easy: they can only ever be executing from a 882 * single context. */ 883 if (is_sync_kiocb(iocb)) { 884 kiocbSetKicked(iocb); 885 wake_up_process(iocb->ki_obj.tsk); 886 return; 887 } 888 889 try_queue_kicked_iocb(iocb); 890 } 891 EXPORT_SYMBOL(kick_iocb); 892 893 /* aio_complete 894 * Called when the io request on the given iocb is complete. 895 * Returns true if this is the last user of the request. The 896 * only other user of the request can be the cancellation code. 897 */ 898 int aio_complete(struct kiocb *iocb, long res, long res2) 899 { 900 struct kioctx *ctx = iocb->ki_ctx; 901 struct aio_ring_info *info; 902 struct aio_ring *ring; 903 struct io_event *event; 904 unsigned long flags; 905 unsigned long tail; 906 int ret; 907 908 /* 909 * Special case handling for sync iocbs: 910 * - events go directly into the iocb for fast handling 911 * - the sync task with the iocb in its stack holds the single iocb 912 * ref, no other paths have a way to get another ref 913 * - the sync task helpfully left a reference to itself in the iocb 914 */ 915 if (is_sync_kiocb(iocb)) { 916 BUG_ON(iocb->ki_users != 1); 917 iocb->ki_user_data = res; 918 iocb->ki_users = 0; 919 wake_up_process(iocb->ki_obj.tsk); 920 return 1; 921 } 922 923 info = &ctx->ring_info; 924 925 /* add a completion event to the ring buffer. 926 * must be done holding ctx->ctx_lock to prevent 927 * other code from messing with the tail 928 * pointer since we might be called from irq 929 * context. 930 */ 931 spin_lock_irqsave(&ctx->ctx_lock, flags); 932 933 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list)) 934 list_del_init(&iocb->ki_run_list); 935 936 /* 937 * cancelled requests don't get events, userland was given one 938 * when the event got cancelled. 939 */ 940 if (kiocbIsCancelled(iocb)) 941 goto put_rq; 942 943 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1); 944 945 tail = info->tail; 946 event = aio_ring_event(info, tail, KM_IRQ0); 947 if (++tail >= info->nr) 948 tail = 0; 949 950 event->obj = (u64)(unsigned long)iocb->ki_obj.user; 951 event->data = iocb->ki_user_data; 952 event->res = res; 953 event->res2 = res2; 954 955 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n", 956 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data, 957 res, res2); 958 959 /* after flagging the request as done, we 960 * must never even look at it again 961 */ 962 smp_wmb(); /* make event visible before updating tail */ 963 964 info->tail = tail; 965 ring->tail = tail; 966 967 put_aio_ring_event(event, KM_IRQ0); 968 kunmap_atomic(ring, KM_IRQ1); 969 970 pr_debug("added to ring %p at [%lu]\n", iocb, tail); 971 972 /* 973 * Check if the user asked us to deliver the result through an 974 * eventfd. The eventfd_signal() function is safe to be called 975 * from IRQ context. 976 */ 977 if (iocb->ki_eventfd != NULL) 978 eventfd_signal(iocb->ki_eventfd, 1); 979 980 put_rq: 981 /* everything turned out well, dispose of the aiocb. */ 982 ret = __aio_put_req(ctx, iocb); 983 984 /* 985 * We have to order our ring_info tail store above and test 986 * of the wait list below outside the wait lock. This is 987 * like in wake_up_bit() where clearing a bit has to be 988 * ordered with the unlocked test. 989 */ 990 smp_mb(); 991 992 if (waitqueue_active(&ctx->wait)) 993 wake_up(&ctx->wait); 994 995 spin_unlock_irqrestore(&ctx->ctx_lock, flags); 996 return ret; 997 } 998 EXPORT_SYMBOL(aio_complete); 999 1000 /* aio_read_evt 1001 * Pull an event off of the ioctx's event ring. Returns the number of 1002 * events fetched (0 or 1 ;-) 1003 * FIXME: make this use cmpxchg. 1004 * TODO: make the ringbuffer user mmap()able (requires FIXME). 1005 */ 1006 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent) 1007 { 1008 struct aio_ring_info *info = &ioctx->ring_info; 1009 struct aio_ring *ring; 1010 unsigned long head; 1011 int ret = 0; 1012 1013 ring = kmap_atomic(info->ring_pages[0], KM_USER0); 1014 dprintk("in aio_read_evt h%lu t%lu m%lu\n", 1015 (unsigned long)ring->head, (unsigned long)ring->tail, 1016 (unsigned long)ring->nr); 1017 1018 if (ring->head == ring->tail) 1019 goto out; 1020 1021 spin_lock(&info->ring_lock); 1022 1023 head = ring->head % info->nr; 1024 if (head != ring->tail) { 1025 struct io_event *evp = aio_ring_event(info, head, KM_USER1); 1026 *ent = *evp; 1027 head = (head + 1) % info->nr; 1028 smp_mb(); /* finish reading the event before updatng the head */ 1029 ring->head = head; 1030 ret = 1; 1031 put_aio_ring_event(evp, KM_USER1); 1032 } 1033 spin_unlock(&info->ring_lock); 1034 1035 out: 1036 kunmap_atomic(ring, KM_USER0); 1037 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret, 1038 (unsigned long)ring->head, (unsigned long)ring->tail); 1039 return ret; 1040 } 1041 1042 struct aio_timeout { 1043 struct timer_list timer; 1044 int timed_out; 1045 struct task_struct *p; 1046 }; 1047 1048 static void timeout_func(unsigned long data) 1049 { 1050 struct aio_timeout *to = (struct aio_timeout *)data; 1051 1052 to->timed_out = 1; 1053 wake_up_process(to->p); 1054 } 1055 1056 static inline void init_timeout(struct aio_timeout *to) 1057 { 1058 setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to); 1059 to->timed_out = 0; 1060 to->p = current; 1061 } 1062 1063 static inline void set_timeout(long start_jiffies, struct aio_timeout *to, 1064 const struct timespec *ts) 1065 { 1066 to->timer.expires = start_jiffies + timespec_to_jiffies(ts); 1067 if (time_after(to->timer.expires, jiffies)) 1068 add_timer(&to->timer); 1069 else 1070 to->timed_out = 1; 1071 } 1072 1073 static inline void clear_timeout(struct aio_timeout *to) 1074 { 1075 del_singleshot_timer_sync(&to->timer); 1076 } 1077 1078 static int read_events(struct kioctx *ctx, 1079 long min_nr, long nr, 1080 struct io_event __user *event, 1081 struct timespec __user *timeout) 1082 { 1083 long start_jiffies = jiffies; 1084 struct task_struct *tsk = current; 1085 DECLARE_WAITQUEUE(wait, tsk); 1086 int ret; 1087 int i = 0; 1088 struct io_event ent; 1089 struct aio_timeout to; 1090 int retry = 0; 1091 1092 /* needed to zero any padding within an entry (there shouldn't be 1093 * any, but C is fun! 1094 */ 1095 memset(&ent, 0, sizeof(ent)); 1096 retry: 1097 ret = 0; 1098 while (likely(i < nr)) { 1099 ret = aio_read_evt(ctx, &ent); 1100 if (unlikely(ret <= 0)) 1101 break; 1102 1103 dprintk("read event: %Lx %Lx %Lx %Lx\n", 1104 ent.data, ent.obj, ent.res, ent.res2); 1105 1106 /* Could we split the check in two? */ 1107 ret = -EFAULT; 1108 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) { 1109 dprintk("aio: lost an event due to EFAULT.\n"); 1110 break; 1111 } 1112 ret = 0; 1113 1114 /* Good, event copied to userland, update counts. */ 1115 event ++; 1116 i ++; 1117 } 1118 1119 if (min_nr <= i) 1120 return i; 1121 if (ret) 1122 return ret; 1123 1124 /* End fast path */ 1125 1126 /* racey check, but it gets redone */ 1127 if (!retry && unlikely(!list_empty(&ctx->run_list))) { 1128 retry = 1; 1129 aio_run_all_iocbs(ctx); 1130 goto retry; 1131 } 1132 1133 init_timeout(&to); 1134 if (timeout) { 1135 struct timespec ts; 1136 ret = -EFAULT; 1137 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts)))) 1138 goto out; 1139 1140 set_timeout(start_jiffies, &to, &ts); 1141 } 1142 1143 while (likely(i < nr)) { 1144 add_wait_queue_exclusive(&ctx->wait, &wait); 1145 do { 1146 set_task_state(tsk, TASK_INTERRUPTIBLE); 1147 ret = aio_read_evt(ctx, &ent); 1148 if (ret) 1149 break; 1150 if (min_nr <= i) 1151 break; 1152 if (unlikely(ctx->dead)) { 1153 ret = -EINVAL; 1154 break; 1155 } 1156 if (to.timed_out) /* Only check after read evt */ 1157 break; 1158 /* Try to only show up in io wait if there are ops 1159 * in flight */ 1160 if (ctx->reqs_active) 1161 io_schedule(); 1162 else 1163 schedule(); 1164 if (signal_pending(tsk)) { 1165 ret = -EINTR; 1166 break; 1167 } 1168 /*ret = aio_read_evt(ctx, &ent);*/ 1169 } while (1) ; 1170 1171 set_task_state(tsk, TASK_RUNNING); 1172 remove_wait_queue(&ctx->wait, &wait); 1173 1174 if (unlikely(ret <= 0)) 1175 break; 1176 1177 ret = -EFAULT; 1178 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) { 1179 dprintk("aio: lost an event due to EFAULT.\n"); 1180 break; 1181 } 1182 1183 /* Good, event copied to userland, update counts. */ 1184 event ++; 1185 i ++; 1186 } 1187 1188 if (timeout) 1189 clear_timeout(&to); 1190 out: 1191 destroy_timer_on_stack(&to.timer); 1192 return i ? i : ret; 1193 } 1194 1195 /* Take an ioctx and remove it from the list of ioctx's. Protects 1196 * against races with itself via ->dead. 1197 */ 1198 static void io_destroy(struct kioctx *ioctx) 1199 { 1200 struct mm_struct *mm = current->mm; 1201 int was_dead; 1202 1203 /* delete the entry from the list is someone else hasn't already */ 1204 spin_lock(&mm->ioctx_lock); 1205 was_dead = ioctx->dead; 1206 ioctx->dead = 1; 1207 hlist_del_rcu(&ioctx->list); 1208 spin_unlock(&mm->ioctx_lock); 1209 1210 dprintk("aio_release(%p)\n", ioctx); 1211 if (likely(!was_dead)) 1212 put_ioctx(ioctx); /* twice for the list */ 1213 1214 aio_cancel_all(ioctx); 1215 wait_for_all_aios(ioctx); 1216 1217 /* 1218 * Wake up any waiters. The setting of ctx->dead must be seen 1219 * by other CPUs at this point. Right now, we rely on the 1220 * locking done by the above calls to ensure this consistency. 1221 */ 1222 wake_up(&ioctx->wait); 1223 put_ioctx(ioctx); /* once for the lookup */ 1224 } 1225 1226 /* sys_io_setup: 1227 * Create an aio_context capable of receiving at least nr_events. 1228 * ctxp must not point to an aio_context that already exists, and 1229 * must be initialized to 0 prior to the call. On successful 1230 * creation of the aio_context, *ctxp is filled in with the resulting 1231 * handle. May fail with -EINVAL if *ctxp is not initialized, 1232 * if the specified nr_events exceeds internal limits. May fail 1233 * with -EAGAIN if the specified nr_events exceeds the user's limit 1234 * of available events. May fail with -ENOMEM if insufficient kernel 1235 * resources are available. May fail with -EFAULT if an invalid 1236 * pointer is passed for ctxp. Will fail with -ENOSYS if not 1237 * implemented. 1238 */ 1239 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp) 1240 { 1241 struct kioctx *ioctx = NULL; 1242 unsigned long ctx; 1243 long ret; 1244 1245 ret = get_user(ctx, ctxp); 1246 if (unlikely(ret)) 1247 goto out; 1248 1249 ret = -EINVAL; 1250 if (unlikely(ctx || nr_events == 0)) { 1251 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n", 1252 ctx, nr_events); 1253 goto out; 1254 } 1255 1256 ioctx = ioctx_alloc(nr_events); 1257 ret = PTR_ERR(ioctx); 1258 if (!IS_ERR(ioctx)) { 1259 ret = put_user(ioctx->user_id, ctxp); 1260 if (!ret) 1261 return 0; 1262 1263 get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */ 1264 io_destroy(ioctx); 1265 } 1266 1267 out: 1268 return ret; 1269 } 1270 1271 /* sys_io_destroy: 1272 * Destroy the aio_context specified. May cancel any outstanding 1273 * AIOs and block on completion. Will fail with -ENOSYS if not 1274 * implemented. May fail with -EFAULT if the context pointed to 1275 * is invalid. 1276 */ 1277 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx) 1278 { 1279 struct kioctx *ioctx = lookup_ioctx(ctx); 1280 if (likely(NULL != ioctx)) { 1281 io_destroy(ioctx); 1282 return 0; 1283 } 1284 pr_debug("EINVAL: io_destroy: invalid context id\n"); 1285 return -EINVAL; 1286 } 1287 1288 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret) 1289 { 1290 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg]; 1291 1292 BUG_ON(ret <= 0); 1293 1294 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) { 1295 ssize_t this = min((ssize_t)iov->iov_len, ret); 1296 iov->iov_base += this; 1297 iov->iov_len -= this; 1298 iocb->ki_left -= this; 1299 ret -= this; 1300 if (iov->iov_len == 0) { 1301 iocb->ki_cur_seg++; 1302 iov++; 1303 } 1304 } 1305 1306 /* the caller should not have done more io than what fit in 1307 * the remaining iovecs */ 1308 BUG_ON(ret > 0 && iocb->ki_left == 0); 1309 } 1310 1311 static ssize_t aio_rw_vect_retry(struct kiocb *iocb) 1312 { 1313 struct file *file = iocb->ki_filp; 1314 struct address_space *mapping = file->f_mapping; 1315 struct inode *inode = mapping->host; 1316 ssize_t (*rw_op)(struct kiocb *, const struct iovec *, 1317 unsigned long, loff_t); 1318 ssize_t ret = 0; 1319 unsigned short opcode; 1320 1321 if ((iocb->ki_opcode == IOCB_CMD_PREADV) || 1322 (iocb->ki_opcode == IOCB_CMD_PREAD)) { 1323 rw_op = file->f_op->aio_read; 1324 opcode = IOCB_CMD_PREADV; 1325 } else { 1326 rw_op = file->f_op->aio_write; 1327 opcode = IOCB_CMD_PWRITEV; 1328 } 1329 1330 /* This matches the pread()/pwrite() logic */ 1331 if (iocb->ki_pos < 0) 1332 return -EINVAL; 1333 1334 do { 1335 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg], 1336 iocb->ki_nr_segs - iocb->ki_cur_seg, 1337 iocb->ki_pos); 1338 if (ret > 0) 1339 aio_advance_iovec(iocb, ret); 1340 1341 /* retry all partial writes. retry partial reads as long as its a 1342 * regular file. */ 1343 } while (ret > 0 && iocb->ki_left > 0 && 1344 (opcode == IOCB_CMD_PWRITEV || 1345 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode)))); 1346 1347 /* This means we must have transferred all that we could */ 1348 /* No need to retry anymore */ 1349 if ((ret == 0) || (iocb->ki_left == 0)) 1350 ret = iocb->ki_nbytes - iocb->ki_left; 1351 1352 /* If we managed to write some out we return that, rather than 1353 * the eventual error. */ 1354 if (opcode == IOCB_CMD_PWRITEV 1355 && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY 1356 && iocb->ki_nbytes - iocb->ki_left) 1357 ret = iocb->ki_nbytes - iocb->ki_left; 1358 1359 return ret; 1360 } 1361 1362 static ssize_t aio_fdsync(struct kiocb *iocb) 1363 { 1364 struct file *file = iocb->ki_filp; 1365 ssize_t ret = -EINVAL; 1366 1367 if (file->f_op->aio_fsync) 1368 ret = file->f_op->aio_fsync(iocb, 1); 1369 return ret; 1370 } 1371 1372 static ssize_t aio_fsync(struct kiocb *iocb) 1373 { 1374 struct file *file = iocb->ki_filp; 1375 ssize_t ret = -EINVAL; 1376 1377 if (file->f_op->aio_fsync) 1378 ret = file->f_op->aio_fsync(iocb, 0); 1379 return ret; 1380 } 1381 1382 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb) 1383 { 1384 ssize_t ret; 1385 1386 ret = rw_copy_check_uvector(type, (struct iovec __user *)kiocb->ki_buf, 1387 kiocb->ki_nbytes, 1, 1388 &kiocb->ki_inline_vec, &kiocb->ki_iovec); 1389 if (ret < 0) 1390 goto out; 1391 1392 kiocb->ki_nr_segs = kiocb->ki_nbytes; 1393 kiocb->ki_cur_seg = 0; 1394 /* ki_nbytes/left now reflect bytes instead of segs */ 1395 kiocb->ki_nbytes = ret; 1396 kiocb->ki_left = ret; 1397 1398 ret = 0; 1399 out: 1400 return ret; 1401 } 1402 1403 static ssize_t aio_setup_single_vector(struct kiocb *kiocb) 1404 { 1405 kiocb->ki_iovec = &kiocb->ki_inline_vec; 1406 kiocb->ki_iovec->iov_base = kiocb->ki_buf; 1407 kiocb->ki_iovec->iov_len = kiocb->ki_left; 1408 kiocb->ki_nr_segs = 1; 1409 kiocb->ki_cur_seg = 0; 1410 return 0; 1411 } 1412 1413 /* 1414 * aio_setup_iocb: 1415 * Performs the initial checks and aio retry method 1416 * setup for the kiocb at the time of io submission. 1417 */ 1418 static ssize_t aio_setup_iocb(struct kiocb *kiocb) 1419 { 1420 struct file *file = kiocb->ki_filp; 1421 ssize_t ret = 0; 1422 1423 switch (kiocb->ki_opcode) { 1424 case IOCB_CMD_PREAD: 1425 ret = -EBADF; 1426 if (unlikely(!(file->f_mode & FMODE_READ))) 1427 break; 1428 ret = -EFAULT; 1429 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf, 1430 kiocb->ki_left))) 1431 break; 1432 ret = security_file_permission(file, MAY_READ); 1433 if (unlikely(ret)) 1434 break; 1435 ret = aio_setup_single_vector(kiocb); 1436 if (ret) 1437 break; 1438 ret = -EINVAL; 1439 if (file->f_op->aio_read) 1440 kiocb->ki_retry = aio_rw_vect_retry; 1441 break; 1442 case IOCB_CMD_PWRITE: 1443 ret = -EBADF; 1444 if (unlikely(!(file->f_mode & FMODE_WRITE))) 1445 break; 1446 ret = -EFAULT; 1447 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf, 1448 kiocb->ki_left))) 1449 break; 1450 ret = security_file_permission(file, MAY_WRITE); 1451 if (unlikely(ret)) 1452 break; 1453 ret = aio_setup_single_vector(kiocb); 1454 if (ret) 1455 break; 1456 ret = -EINVAL; 1457 if (file->f_op->aio_write) 1458 kiocb->ki_retry = aio_rw_vect_retry; 1459 break; 1460 case IOCB_CMD_PREADV: 1461 ret = -EBADF; 1462 if (unlikely(!(file->f_mode & FMODE_READ))) 1463 break; 1464 ret = security_file_permission(file, MAY_READ); 1465 if (unlikely(ret)) 1466 break; 1467 ret = aio_setup_vectored_rw(READ, kiocb); 1468 if (ret) 1469 break; 1470 ret = -EINVAL; 1471 if (file->f_op->aio_read) 1472 kiocb->ki_retry = aio_rw_vect_retry; 1473 break; 1474 case IOCB_CMD_PWRITEV: 1475 ret = -EBADF; 1476 if (unlikely(!(file->f_mode & FMODE_WRITE))) 1477 break; 1478 ret = security_file_permission(file, MAY_WRITE); 1479 if (unlikely(ret)) 1480 break; 1481 ret = aio_setup_vectored_rw(WRITE, kiocb); 1482 if (ret) 1483 break; 1484 ret = -EINVAL; 1485 if (file->f_op->aio_write) 1486 kiocb->ki_retry = aio_rw_vect_retry; 1487 break; 1488 case IOCB_CMD_FDSYNC: 1489 ret = -EINVAL; 1490 if (file->f_op->aio_fsync) 1491 kiocb->ki_retry = aio_fdsync; 1492 break; 1493 case IOCB_CMD_FSYNC: 1494 ret = -EINVAL; 1495 if (file->f_op->aio_fsync) 1496 kiocb->ki_retry = aio_fsync; 1497 break; 1498 default: 1499 dprintk("EINVAL: io_submit: no operation provided\n"); 1500 ret = -EINVAL; 1501 } 1502 1503 if (!kiocb->ki_retry) 1504 return ret; 1505 1506 return 0; 1507 } 1508 1509 /* 1510 * aio_wake_function: 1511 * wait queue callback function for aio notification, 1512 * Simply triggers a retry of the operation via kick_iocb. 1513 * 1514 * This callback is specified in the wait queue entry in 1515 * a kiocb. 1516 * 1517 * Note: 1518 * This routine is executed with the wait queue lock held. 1519 * Since kick_iocb acquires iocb->ctx->ctx_lock, it nests 1520 * the ioctx lock inside the wait queue lock. This is safe 1521 * because this callback isn't used for wait queues which 1522 * are nested inside ioctx lock (i.e. ctx->wait) 1523 */ 1524 static int aio_wake_function(wait_queue_t *wait, unsigned mode, 1525 int sync, void *key) 1526 { 1527 struct kiocb *iocb = container_of(wait, struct kiocb, ki_wait); 1528 1529 list_del_init(&wait->task_list); 1530 kick_iocb(iocb); 1531 return 1; 1532 } 1533 1534 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb, 1535 struct iocb *iocb) 1536 { 1537 struct kiocb *req; 1538 struct file *file; 1539 ssize_t ret; 1540 1541 /* enforce forwards compatibility on users */ 1542 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) { 1543 pr_debug("EINVAL: io_submit: reserve field set\n"); 1544 return -EINVAL; 1545 } 1546 1547 /* prevent overflows */ 1548 if (unlikely( 1549 (iocb->aio_buf != (unsigned long)iocb->aio_buf) || 1550 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) || 1551 ((ssize_t)iocb->aio_nbytes < 0) 1552 )) { 1553 pr_debug("EINVAL: io_submit: overflow check\n"); 1554 return -EINVAL; 1555 } 1556 1557 file = fget(iocb->aio_fildes); 1558 if (unlikely(!file)) 1559 return -EBADF; 1560 1561 req = aio_get_req(ctx); /* returns with 2 references to req */ 1562 if (unlikely(!req)) { 1563 fput(file); 1564 return -EAGAIN; 1565 } 1566 req->ki_filp = file; 1567 if (iocb->aio_flags & IOCB_FLAG_RESFD) { 1568 /* 1569 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an 1570 * instance of the file* now. The file descriptor must be 1571 * an eventfd() fd, and will be signaled for each completed 1572 * event using the eventfd_signal() function. 1573 */ 1574 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd); 1575 if (IS_ERR(req->ki_eventfd)) { 1576 ret = PTR_ERR(req->ki_eventfd); 1577 req->ki_eventfd = NULL; 1578 goto out_put_req; 1579 } 1580 } 1581 1582 ret = put_user(req->ki_key, &user_iocb->aio_key); 1583 if (unlikely(ret)) { 1584 dprintk("EFAULT: aio_key\n"); 1585 goto out_put_req; 1586 } 1587 1588 req->ki_obj.user = user_iocb; 1589 req->ki_user_data = iocb->aio_data; 1590 req->ki_pos = iocb->aio_offset; 1591 1592 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf; 1593 req->ki_left = req->ki_nbytes = iocb->aio_nbytes; 1594 req->ki_opcode = iocb->aio_lio_opcode; 1595 init_waitqueue_func_entry(&req->ki_wait, aio_wake_function); 1596 INIT_LIST_HEAD(&req->ki_wait.task_list); 1597 1598 ret = aio_setup_iocb(req); 1599 1600 if (ret) 1601 goto out_put_req; 1602 1603 spin_lock_irq(&ctx->ctx_lock); 1604 aio_run_iocb(req); 1605 if (!list_empty(&ctx->run_list)) { 1606 /* drain the run list */ 1607 while (__aio_run_iocbs(ctx)) 1608 ; 1609 } 1610 spin_unlock_irq(&ctx->ctx_lock); 1611 aio_put_req(req); /* drop extra ref to req */ 1612 return 0; 1613 1614 out_put_req: 1615 aio_put_req(req); /* drop extra ref to req */ 1616 aio_put_req(req); /* drop i/o ref to req */ 1617 return ret; 1618 } 1619 1620 /* sys_io_submit: 1621 * Queue the nr iocbs pointed to by iocbpp for processing. Returns 1622 * the number of iocbs queued. May return -EINVAL if the aio_context 1623 * specified by ctx_id is invalid, if nr is < 0, if the iocb at 1624 * *iocbpp[0] is not properly initialized, if the operation specified 1625 * is invalid for the file descriptor in the iocb. May fail with 1626 * -EFAULT if any of the data structures point to invalid data. May 1627 * fail with -EBADF if the file descriptor specified in the first 1628 * iocb is invalid. May fail with -EAGAIN if insufficient resources 1629 * are available to queue any iocbs. Will return 0 if nr is 0. Will 1630 * fail with -ENOSYS if not implemented. 1631 */ 1632 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr, 1633 struct iocb __user * __user *, iocbpp) 1634 { 1635 struct kioctx *ctx; 1636 long ret = 0; 1637 int i; 1638 1639 if (unlikely(nr < 0)) 1640 return -EINVAL; 1641 1642 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp))))) 1643 return -EFAULT; 1644 1645 ctx = lookup_ioctx(ctx_id); 1646 if (unlikely(!ctx)) { 1647 pr_debug("EINVAL: io_submit: invalid context id\n"); 1648 return -EINVAL; 1649 } 1650 1651 /* 1652 * AKPM: should this return a partial result if some of the IOs were 1653 * successfully submitted? 1654 */ 1655 for (i=0; i<nr; i++) { 1656 struct iocb __user *user_iocb; 1657 struct iocb tmp; 1658 1659 if (unlikely(__get_user(user_iocb, iocbpp + i))) { 1660 ret = -EFAULT; 1661 break; 1662 } 1663 1664 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) { 1665 ret = -EFAULT; 1666 break; 1667 } 1668 1669 ret = io_submit_one(ctx, user_iocb, &tmp); 1670 if (ret) 1671 break; 1672 } 1673 1674 put_ioctx(ctx); 1675 return i ? i : ret; 1676 } 1677 1678 /* lookup_kiocb 1679 * Finds a given iocb for cancellation. 1680 */ 1681 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, 1682 u32 key) 1683 { 1684 struct list_head *pos; 1685 1686 assert_spin_locked(&ctx->ctx_lock); 1687 1688 /* TODO: use a hash or array, this sucks. */ 1689 list_for_each(pos, &ctx->active_reqs) { 1690 struct kiocb *kiocb = list_kiocb(pos); 1691 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key) 1692 return kiocb; 1693 } 1694 return NULL; 1695 } 1696 1697 /* sys_io_cancel: 1698 * Attempts to cancel an iocb previously passed to io_submit. If 1699 * the operation is successfully cancelled, the resulting event is 1700 * copied into the memory pointed to by result without being placed 1701 * into the completion queue and 0 is returned. May fail with 1702 * -EFAULT if any of the data structures pointed to are invalid. 1703 * May fail with -EINVAL if aio_context specified by ctx_id is 1704 * invalid. May fail with -EAGAIN if the iocb specified was not 1705 * cancelled. Will fail with -ENOSYS if not implemented. 1706 */ 1707 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb, 1708 struct io_event __user *, result) 1709 { 1710 int (*cancel)(struct kiocb *iocb, struct io_event *res); 1711 struct kioctx *ctx; 1712 struct kiocb *kiocb; 1713 u32 key; 1714 int ret; 1715 1716 ret = get_user(key, &iocb->aio_key); 1717 if (unlikely(ret)) 1718 return -EFAULT; 1719 1720 ctx = lookup_ioctx(ctx_id); 1721 if (unlikely(!ctx)) 1722 return -EINVAL; 1723 1724 spin_lock_irq(&ctx->ctx_lock); 1725 ret = -EAGAIN; 1726 kiocb = lookup_kiocb(ctx, iocb, key); 1727 if (kiocb && kiocb->ki_cancel) { 1728 cancel = kiocb->ki_cancel; 1729 kiocb->ki_users ++; 1730 kiocbSetCancelled(kiocb); 1731 } else 1732 cancel = NULL; 1733 spin_unlock_irq(&ctx->ctx_lock); 1734 1735 if (NULL != cancel) { 1736 struct io_event tmp; 1737 pr_debug("calling cancel\n"); 1738 memset(&tmp, 0, sizeof(tmp)); 1739 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user; 1740 tmp.data = kiocb->ki_user_data; 1741 ret = cancel(kiocb, &tmp); 1742 if (!ret) { 1743 /* Cancellation succeeded -- copy the result 1744 * into the user's buffer. 1745 */ 1746 if (copy_to_user(result, &tmp, sizeof(tmp))) 1747 ret = -EFAULT; 1748 } 1749 } else 1750 ret = -EINVAL; 1751 1752 put_ioctx(ctx); 1753 1754 return ret; 1755 } 1756 1757 /* io_getevents: 1758 * Attempts to read at least min_nr events and up to nr events from 1759 * the completion queue for the aio_context specified by ctx_id. May 1760 * fail with -EINVAL if ctx_id is invalid, if min_nr is out of range, 1761 * if nr is out of range, if when is out of range. May fail with 1762 * -EFAULT if any of the memory specified to is invalid. May return 1763 * 0 or < min_nr if no events are available and the timeout specified 1764 * by when has elapsed, where when == NULL specifies an infinite 1765 * timeout. Note that the timeout pointed to by when is relative and 1766 * will be updated if not NULL and the operation blocks. Will fail 1767 * with -ENOSYS if not implemented. 1768 */ 1769 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id, 1770 long, min_nr, 1771 long, nr, 1772 struct io_event __user *, events, 1773 struct timespec __user *, timeout) 1774 { 1775 struct kioctx *ioctx = lookup_ioctx(ctx_id); 1776 long ret = -EINVAL; 1777 1778 if (likely(ioctx)) { 1779 if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0)) 1780 ret = read_events(ioctx, min_nr, nr, events, timeout); 1781 put_ioctx(ioctx); 1782 } 1783 1784 asmlinkage_protect(5, ret, ctx_id, min_nr, nr, events, timeout); 1785 return ret; 1786 } 1787