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/export.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/compat.h> 38 39 #include <asm/kmap_types.h> 40 #include <asm/uaccess.h> 41 42 #if DEBUG > 1 43 #define dprintk printk 44 #else 45 #define dprintk(x...) do { ; } while (0) 46 #endif 47 48 /*------ sysctl variables----*/ 49 static DEFINE_SPINLOCK(aio_nr_lock); 50 unsigned long aio_nr; /* current system wide number of aio requests */ 51 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */ 52 /*----end sysctl variables---*/ 53 54 static struct kmem_cache *kiocb_cachep; 55 static struct kmem_cache *kioctx_cachep; 56 57 static struct workqueue_struct *aio_wq; 58 59 /* Used for rare fput completion. */ 60 static void aio_fput_routine(struct work_struct *); 61 static DECLARE_WORK(fput_work, aio_fput_routine); 62 63 static DEFINE_SPINLOCK(fput_lock); 64 static LIST_HEAD(fput_head); 65 66 static void aio_kick_handler(struct work_struct *); 67 static void aio_queue_work(struct kioctx *); 68 69 /* aio_setup 70 * Creates the slab caches used by the aio routines, panic on 71 * failure as this is done early during the boot sequence. 72 */ 73 static int __init aio_setup(void) 74 { 75 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC); 76 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC); 77 78 aio_wq = alloc_workqueue("aio", 0, 1); /* used to limit concurrency */ 79 BUG_ON(!aio_wq); 80 81 pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page)); 82 83 return 0; 84 } 85 __initcall(aio_setup); 86 87 static void aio_free_ring(struct kioctx *ctx) 88 { 89 struct aio_ring_info *info = &ctx->ring_info; 90 long i; 91 92 for (i=0; i<info->nr_pages; i++) 93 put_page(info->ring_pages[i]); 94 95 if (info->mmap_size) { 96 down_write(&ctx->mm->mmap_sem); 97 do_munmap(ctx->mm, info->mmap_base, info->mmap_size); 98 up_write(&ctx->mm->mmap_sem); 99 } 100 101 if (info->ring_pages && info->ring_pages != info->internal_pages) 102 kfree(info->ring_pages); 103 info->ring_pages = NULL; 104 info->nr = 0; 105 } 106 107 static int aio_setup_ring(struct kioctx *ctx) 108 { 109 struct aio_ring *ring; 110 struct aio_ring_info *info = &ctx->ring_info; 111 unsigned nr_events = ctx->max_reqs; 112 unsigned long size; 113 int nr_pages; 114 115 /* Compensate for the ring buffer's head/tail overlap entry */ 116 nr_events += 2; /* 1 is required, 2 for good luck */ 117 118 size = sizeof(struct aio_ring); 119 size += sizeof(struct io_event) * nr_events; 120 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT; 121 122 if (nr_pages < 0) 123 return -EINVAL; 124 125 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event); 126 127 info->nr = 0; 128 info->ring_pages = info->internal_pages; 129 if (nr_pages > AIO_RING_PAGES) { 130 info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL); 131 if (!info->ring_pages) 132 return -ENOMEM; 133 } 134 135 info->mmap_size = nr_pages * PAGE_SIZE; 136 dprintk("attempting mmap of %lu bytes\n", info->mmap_size); 137 down_write(&ctx->mm->mmap_sem); 138 info->mmap_base = do_mmap(NULL, 0, info->mmap_size, 139 PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE, 140 0); 141 if (IS_ERR((void *)info->mmap_base)) { 142 up_write(&ctx->mm->mmap_sem); 143 info->mmap_size = 0; 144 aio_free_ring(ctx); 145 return -EAGAIN; 146 } 147 148 dprintk("mmap address: 0x%08lx\n", info->mmap_base); 149 info->nr_pages = get_user_pages(current, ctx->mm, 150 info->mmap_base, nr_pages, 151 1, 0, info->ring_pages, NULL); 152 up_write(&ctx->mm->mmap_sem); 153 154 if (unlikely(info->nr_pages != nr_pages)) { 155 aio_free_ring(ctx); 156 return -EAGAIN; 157 } 158 159 ctx->user_id = info->mmap_base; 160 161 info->nr = nr_events; /* trusted copy */ 162 163 ring = kmap_atomic(info->ring_pages[0]); 164 ring->nr = nr_events; /* user copy */ 165 ring->id = ctx->user_id; 166 ring->head = ring->tail = 0; 167 ring->magic = AIO_RING_MAGIC; 168 ring->compat_features = AIO_RING_COMPAT_FEATURES; 169 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES; 170 ring->header_length = sizeof(struct aio_ring); 171 kunmap_atomic(ring); 172 173 return 0; 174 } 175 176 177 /* aio_ring_event: returns a pointer to the event at the given index from 178 * kmap_atomic(). Release the pointer with put_aio_ring_event(); 179 */ 180 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event)) 181 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event)) 182 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE) 183 184 #define aio_ring_event(info, nr) ({ \ 185 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \ 186 struct io_event *__event; \ 187 __event = kmap_atomic( \ 188 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE]); \ 189 __event += pos % AIO_EVENTS_PER_PAGE; \ 190 __event; \ 191 }) 192 193 #define put_aio_ring_event(event) do { \ 194 struct io_event *__event = (event); \ 195 (void)__event; \ 196 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK)); \ 197 } while(0) 198 199 static void ctx_rcu_free(struct rcu_head *head) 200 { 201 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head); 202 kmem_cache_free(kioctx_cachep, ctx); 203 } 204 205 /* __put_ioctx 206 * Called when the last user of an aio context has gone away, 207 * and the struct needs to be freed. 208 */ 209 static void __put_ioctx(struct kioctx *ctx) 210 { 211 unsigned nr_events = ctx->max_reqs; 212 BUG_ON(ctx->reqs_active); 213 214 cancel_delayed_work_sync(&ctx->wq); 215 aio_free_ring(ctx); 216 mmdrop(ctx->mm); 217 ctx->mm = NULL; 218 if (nr_events) { 219 spin_lock(&aio_nr_lock); 220 BUG_ON(aio_nr - nr_events > aio_nr); 221 aio_nr -= nr_events; 222 spin_unlock(&aio_nr_lock); 223 } 224 pr_debug("__put_ioctx: freeing %p\n", ctx); 225 call_rcu(&ctx->rcu_head, ctx_rcu_free); 226 } 227 228 static inline int try_get_ioctx(struct kioctx *kioctx) 229 { 230 return atomic_inc_not_zero(&kioctx->users); 231 } 232 233 static inline void put_ioctx(struct kioctx *kioctx) 234 { 235 BUG_ON(atomic_read(&kioctx->users) <= 0); 236 if (unlikely(atomic_dec_and_test(&kioctx->users))) 237 __put_ioctx(kioctx); 238 } 239 240 /* ioctx_alloc 241 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed. 242 */ 243 static struct kioctx *ioctx_alloc(unsigned nr_events) 244 { 245 struct mm_struct *mm; 246 struct kioctx *ctx; 247 int err = -ENOMEM; 248 249 /* Prevent overflows */ 250 if ((nr_events > (0x10000000U / sizeof(struct io_event))) || 251 (nr_events > (0x10000000U / sizeof(struct kiocb)))) { 252 pr_debug("ENOMEM: nr_events too high\n"); 253 return ERR_PTR(-EINVAL); 254 } 255 256 if (!nr_events || (unsigned long)nr_events > aio_max_nr) 257 return ERR_PTR(-EAGAIN); 258 259 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL); 260 if (!ctx) 261 return ERR_PTR(-ENOMEM); 262 263 ctx->max_reqs = nr_events; 264 mm = ctx->mm = current->mm; 265 atomic_inc(&mm->mm_count); 266 267 atomic_set(&ctx->users, 2); 268 spin_lock_init(&ctx->ctx_lock); 269 spin_lock_init(&ctx->ring_info.ring_lock); 270 init_waitqueue_head(&ctx->wait); 271 272 INIT_LIST_HEAD(&ctx->active_reqs); 273 INIT_LIST_HEAD(&ctx->run_list); 274 INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler); 275 276 if (aio_setup_ring(ctx) < 0) 277 goto out_freectx; 278 279 /* limit the number of system wide aios */ 280 spin_lock(&aio_nr_lock); 281 if (aio_nr + nr_events > aio_max_nr || 282 aio_nr + nr_events < aio_nr) { 283 spin_unlock(&aio_nr_lock); 284 goto out_cleanup; 285 } 286 aio_nr += ctx->max_reqs; 287 spin_unlock(&aio_nr_lock); 288 289 /* now link into global list. */ 290 spin_lock(&mm->ioctx_lock); 291 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list); 292 spin_unlock(&mm->ioctx_lock); 293 294 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n", 295 ctx, ctx->user_id, current->mm, ctx->ring_info.nr); 296 return ctx; 297 298 out_cleanup: 299 err = -EAGAIN; 300 aio_free_ring(ctx); 301 out_freectx: 302 mmdrop(mm); 303 kmem_cache_free(kioctx_cachep, ctx); 304 dprintk("aio: error allocating ioctx %d\n", err); 305 return ERR_PTR(err); 306 } 307 308 /* kill_ctx 309 * Cancels all outstanding aio requests on an aio context. Used 310 * when the processes owning a context have all exited to encourage 311 * the rapid destruction of the kioctx. 312 */ 313 static void kill_ctx(struct kioctx *ctx) 314 { 315 int (*cancel)(struct kiocb *, struct io_event *); 316 struct task_struct *tsk = current; 317 DECLARE_WAITQUEUE(wait, tsk); 318 struct io_event res; 319 320 spin_lock_irq(&ctx->ctx_lock); 321 ctx->dead = 1; 322 while (!list_empty(&ctx->active_reqs)) { 323 struct list_head *pos = ctx->active_reqs.next; 324 struct kiocb *iocb = list_kiocb(pos); 325 list_del_init(&iocb->ki_list); 326 cancel = iocb->ki_cancel; 327 kiocbSetCancelled(iocb); 328 if (cancel) { 329 iocb->ki_users++; 330 spin_unlock_irq(&ctx->ctx_lock); 331 cancel(iocb, &res); 332 spin_lock_irq(&ctx->ctx_lock); 333 } 334 } 335 336 if (!ctx->reqs_active) 337 goto out; 338 339 add_wait_queue(&ctx->wait, &wait); 340 set_task_state(tsk, TASK_UNINTERRUPTIBLE); 341 while (ctx->reqs_active) { 342 spin_unlock_irq(&ctx->ctx_lock); 343 io_schedule(); 344 set_task_state(tsk, TASK_UNINTERRUPTIBLE); 345 spin_lock_irq(&ctx->ctx_lock); 346 } 347 __set_task_state(tsk, TASK_RUNNING); 348 remove_wait_queue(&ctx->wait, &wait); 349 350 out: 351 spin_unlock_irq(&ctx->ctx_lock); 352 } 353 354 /* wait_on_sync_kiocb: 355 * Waits on the given sync kiocb to complete. 356 */ 357 ssize_t wait_on_sync_kiocb(struct kiocb *iocb) 358 { 359 while (iocb->ki_users) { 360 set_current_state(TASK_UNINTERRUPTIBLE); 361 if (!iocb->ki_users) 362 break; 363 io_schedule(); 364 } 365 __set_current_state(TASK_RUNNING); 366 return iocb->ki_user_data; 367 } 368 EXPORT_SYMBOL(wait_on_sync_kiocb); 369 370 /* exit_aio: called when the last user of mm goes away. At this point, 371 * there is no way for any new requests to be submited or any of the 372 * io_* syscalls to be called on the context. However, there may be 373 * outstanding requests which hold references to the context; as they 374 * go away, they will call put_ioctx and release any pinned memory 375 * associated with the request (held via struct page * references). 376 */ 377 void exit_aio(struct mm_struct *mm) 378 { 379 struct kioctx *ctx; 380 381 while (!hlist_empty(&mm->ioctx_list)) { 382 ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list); 383 hlist_del_rcu(&ctx->list); 384 385 kill_ctx(ctx); 386 387 if (1 != atomic_read(&ctx->users)) 388 printk(KERN_DEBUG 389 "exit_aio:ioctx still alive: %d %d %d\n", 390 atomic_read(&ctx->users), ctx->dead, 391 ctx->reqs_active); 392 put_ioctx(ctx); 393 } 394 } 395 396 /* aio_get_req 397 * Allocate a slot for an aio request. Increments the users count 398 * of the kioctx so that the kioctx stays around until all requests are 399 * complete. Returns NULL if no requests are free. 400 * 401 * Returns with kiocb->users set to 2. The io submit code path holds 402 * an extra reference while submitting the i/o. 403 * This prevents races between the aio code path referencing the 404 * req (after submitting it) and aio_complete() freeing the req. 405 */ 406 static struct kiocb *__aio_get_req(struct kioctx *ctx) 407 { 408 struct kiocb *req = NULL; 409 410 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL); 411 if (unlikely(!req)) 412 return NULL; 413 414 req->ki_flags = 0; 415 req->ki_users = 2; 416 req->ki_key = 0; 417 req->ki_ctx = ctx; 418 req->ki_cancel = NULL; 419 req->ki_retry = NULL; 420 req->ki_dtor = NULL; 421 req->private = NULL; 422 req->ki_iovec = NULL; 423 INIT_LIST_HEAD(&req->ki_run_list); 424 req->ki_eventfd = NULL; 425 426 return req; 427 } 428 429 /* 430 * struct kiocb's are allocated in batches to reduce the number of 431 * times the ctx lock is acquired and released. 432 */ 433 #define KIOCB_BATCH_SIZE 32L 434 struct kiocb_batch { 435 struct list_head head; 436 long count; /* number of requests left to allocate */ 437 }; 438 439 static void kiocb_batch_init(struct kiocb_batch *batch, long total) 440 { 441 INIT_LIST_HEAD(&batch->head); 442 batch->count = total; 443 } 444 445 static void kiocb_batch_free(struct kioctx *ctx, struct kiocb_batch *batch) 446 { 447 struct kiocb *req, *n; 448 449 if (list_empty(&batch->head)) 450 return; 451 452 spin_lock_irq(&ctx->ctx_lock); 453 list_for_each_entry_safe(req, n, &batch->head, ki_batch) { 454 list_del(&req->ki_batch); 455 list_del(&req->ki_list); 456 kmem_cache_free(kiocb_cachep, req); 457 ctx->reqs_active--; 458 } 459 if (unlikely(!ctx->reqs_active && ctx->dead)) 460 wake_up_all(&ctx->wait); 461 spin_unlock_irq(&ctx->ctx_lock); 462 } 463 464 /* 465 * Allocate a batch of kiocbs. This avoids taking and dropping the 466 * context lock a lot during setup. 467 */ 468 static int kiocb_batch_refill(struct kioctx *ctx, struct kiocb_batch *batch) 469 { 470 unsigned short allocated, to_alloc; 471 long avail; 472 bool called_fput = false; 473 struct kiocb *req, *n; 474 struct aio_ring *ring; 475 476 to_alloc = min(batch->count, KIOCB_BATCH_SIZE); 477 for (allocated = 0; allocated < to_alloc; allocated++) { 478 req = __aio_get_req(ctx); 479 if (!req) 480 /* allocation failed, go with what we've got */ 481 break; 482 list_add(&req->ki_batch, &batch->head); 483 } 484 485 if (allocated == 0) 486 goto out; 487 488 retry: 489 spin_lock_irq(&ctx->ctx_lock); 490 ring = kmap_atomic(ctx->ring_info.ring_pages[0]); 491 492 avail = aio_ring_avail(&ctx->ring_info, ring) - ctx->reqs_active; 493 BUG_ON(avail < 0); 494 if (avail == 0 && !called_fput) { 495 /* 496 * Handle a potential starvation case. It is possible that 497 * we hold the last reference on a struct file, causing us 498 * to delay the final fput to non-irq context. In this case, 499 * ctx->reqs_active is artificially high. Calling the fput 500 * routine here may free up a slot in the event completion 501 * ring, allowing this allocation to succeed. 502 */ 503 kunmap_atomic(ring); 504 spin_unlock_irq(&ctx->ctx_lock); 505 aio_fput_routine(NULL); 506 called_fput = true; 507 goto retry; 508 } 509 510 if (avail < allocated) { 511 /* Trim back the number of requests. */ 512 list_for_each_entry_safe(req, n, &batch->head, ki_batch) { 513 list_del(&req->ki_batch); 514 kmem_cache_free(kiocb_cachep, req); 515 if (--allocated <= avail) 516 break; 517 } 518 } 519 520 batch->count -= allocated; 521 list_for_each_entry(req, &batch->head, ki_batch) { 522 list_add(&req->ki_list, &ctx->active_reqs); 523 ctx->reqs_active++; 524 } 525 526 kunmap_atomic(ring); 527 spin_unlock_irq(&ctx->ctx_lock); 528 529 out: 530 return allocated; 531 } 532 533 static inline struct kiocb *aio_get_req(struct kioctx *ctx, 534 struct kiocb_batch *batch) 535 { 536 struct kiocb *req; 537 538 if (list_empty(&batch->head)) 539 if (kiocb_batch_refill(ctx, batch) == 0) 540 return NULL; 541 req = list_first_entry(&batch->head, struct kiocb, ki_batch); 542 list_del(&req->ki_batch); 543 return req; 544 } 545 546 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req) 547 { 548 assert_spin_locked(&ctx->ctx_lock); 549 550 if (req->ki_eventfd != NULL) 551 eventfd_ctx_put(req->ki_eventfd); 552 if (req->ki_dtor) 553 req->ki_dtor(req); 554 if (req->ki_iovec != &req->ki_inline_vec) 555 kfree(req->ki_iovec); 556 kmem_cache_free(kiocb_cachep, req); 557 ctx->reqs_active--; 558 559 if (unlikely(!ctx->reqs_active && ctx->dead)) 560 wake_up_all(&ctx->wait); 561 } 562 563 static void aio_fput_routine(struct work_struct *data) 564 { 565 spin_lock_irq(&fput_lock); 566 while (likely(!list_empty(&fput_head))) { 567 struct kiocb *req = list_kiocb(fput_head.next); 568 struct kioctx *ctx = req->ki_ctx; 569 570 list_del(&req->ki_list); 571 spin_unlock_irq(&fput_lock); 572 573 /* Complete the fput(s) */ 574 if (req->ki_filp != NULL) 575 fput(req->ki_filp); 576 577 /* Link the iocb into the context's free list */ 578 rcu_read_lock(); 579 spin_lock_irq(&ctx->ctx_lock); 580 really_put_req(ctx, req); 581 /* 582 * at that point ctx might've been killed, but actual 583 * freeing is RCU'd 584 */ 585 spin_unlock_irq(&ctx->ctx_lock); 586 rcu_read_unlock(); 587 588 spin_lock_irq(&fput_lock); 589 } 590 spin_unlock_irq(&fput_lock); 591 } 592 593 /* __aio_put_req 594 * Returns true if this put was the last user of the request. 595 */ 596 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req) 597 { 598 dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n", 599 req, atomic_long_read(&req->ki_filp->f_count)); 600 601 assert_spin_locked(&ctx->ctx_lock); 602 603 req->ki_users--; 604 BUG_ON(req->ki_users < 0); 605 if (likely(req->ki_users)) 606 return 0; 607 list_del(&req->ki_list); /* remove from active_reqs */ 608 req->ki_cancel = NULL; 609 req->ki_retry = NULL; 610 611 /* 612 * Try to optimize the aio and eventfd file* puts, by avoiding to 613 * schedule work in case it is not final fput() time. In normal cases, 614 * we would not be holding the last reference to the file*, so 615 * this function will be executed w/out any aio kthread wakeup. 616 */ 617 if (unlikely(!fput_atomic(req->ki_filp))) { 618 spin_lock(&fput_lock); 619 list_add(&req->ki_list, &fput_head); 620 spin_unlock(&fput_lock); 621 schedule_work(&fput_work); 622 } else { 623 req->ki_filp = NULL; 624 really_put_req(ctx, req); 625 } 626 return 1; 627 } 628 629 /* aio_put_req 630 * Returns true if this put was the last user of the kiocb, 631 * false if the request is still in use. 632 */ 633 int aio_put_req(struct kiocb *req) 634 { 635 struct kioctx *ctx = req->ki_ctx; 636 int ret; 637 spin_lock_irq(&ctx->ctx_lock); 638 ret = __aio_put_req(ctx, req); 639 spin_unlock_irq(&ctx->ctx_lock); 640 return ret; 641 } 642 EXPORT_SYMBOL(aio_put_req); 643 644 static struct kioctx *lookup_ioctx(unsigned long ctx_id) 645 { 646 struct mm_struct *mm = current->mm; 647 struct kioctx *ctx, *ret = NULL; 648 struct hlist_node *n; 649 650 rcu_read_lock(); 651 652 hlist_for_each_entry_rcu(ctx, n, &mm->ioctx_list, list) { 653 /* 654 * RCU protects us against accessing freed memory but 655 * we have to be careful not to get a reference when the 656 * reference count already dropped to 0 (ctx->dead test 657 * is unreliable because of races). 658 */ 659 if (ctx->user_id == ctx_id && !ctx->dead && try_get_ioctx(ctx)){ 660 ret = ctx; 661 break; 662 } 663 } 664 665 rcu_read_unlock(); 666 return ret; 667 } 668 669 /* 670 * Queue up a kiocb to be retried. Assumes that the kiocb 671 * has already been marked as kicked, and places it on 672 * the retry run list for the corresponding ioctx, if it 673 * isn't already queued. Returns 1 if it actually queued 674 * the kiocb (to tell the caller to activate the work 675 * queue to process it), or 0, if it found that it was 676 * already queued. 677 */ 678 static inline int __queue_kicked_iocb(struct kiocb *iocb) 679 { 680 struct kioctx *ctx = iocb->ki_ctx; 681 682 assert_spin_locked(&ctx->ctx_lock); 683 684 if (list_empty(&iocb->ki_run_list)) { 685 list_add_tail(&iocb->ki_run_list, 686 &ctx->run_list); 687 return 1; 688 } 689 return 0; 690 } 691 692 /* aio_run_iocb 693 * This is the core aio execution routine. It is 694 * invoked both for initial i/o submission and 695 * subsequent retries via the aio_kick_handler. 696 * Expects to be invoked with iocb->ki_ctx->lock 697 * already held. The lock is released and reacquired 698 * as needed during processing. 699 * 700 * Calls the iocb retry method (already setup for the 701 * iocb on initial submission) for operation specific 702 * handling, but takes care of most of common retry 703 * execution details for a given iocb. The retry method 704 * needs to be non-blocking as far as possible, to avoid 705 * holding up other iocbs waiting to be serviced by the 706 * retry kernel thread. 707 * 708 * The trickier parts in this code have to do with 709 * ensuring that only one retry instance is in progress 710 * for a given iocb at any time. Providing that guarantee 711 * simplifies the coding of individual aio operations as 712 * it avoids various potential races. 713 */ 714 static ssize_t aio_run_iocb(struct kiocb *iocb) 715 { 716 struct kioctx *ctx = iocb->ki_ctx; 717 ssize_t (*retry)(struct kiocb *); 718 ssize_t ret; 719 720 if (!(retry = iocb->ki_retry)) { 721 printk("aio_run_iocb: iocb->ki_retry = NULL\n"); 722 return 0; 723 } 724 725 /* 726 * We don't want the next retry iteration for this 727 * operation to start until this one has returned and 728 * updated the iocb state. However, wait_queue functions 729 * can trigger a kick_iocb from interrupt context in the 730 * meantime, indicating that data is available for the next 731 * iteration. We want to remember that and enable the 732 * next retry iteration _after_ we are through with 733 * this one. 734 * 735 * So, in order to be able to register a "kick", but 736 * prevent it from being queued now, we clear the kick 737 * flag, but make the kick code *think* that the iocb is 738 * still on the run list until we are actually done. 739 * When we are done with this iteration, we check if 740 * the iocb was kicked in the meantime and if so, queue 741 * it up afresh. 742 */ 743 744 kiocbClearKicked(iocb); 745 746 /* 747 * This is so that aio_complete knows it doesn't need to 748 * pull the iocb off the run list (We can't just call 749 * INIT_LIST_HEAD because we don't want a kick_iocb to 750 * queue this on the run list yet) 751 */ 752 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL; 753 spin_unlock_irq(&ctx->ctx_lock); 754 755 /* Quit retrying if the i/o has been cancelled */ 756 if (kiocbIsCancelled(iocb)) { 757 ret = -EINTR; 758 aio_complete(iocb, ret, 0); 759 /* must not access the iocb after this */ 760 goto out; 761 } 762 763 /* 764 * Now we are all set to call the retry method in async 765 * context. 766 */ 767 ret = retry(iocb); 768 769 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) { 770 /* 771 * There's no easy way to restart the syscall since other AIO's 772 * may be already running. Just fail this IO with EINTR. 773 */ 774 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR || 775 ret == -ERESTARTNOHAND || ret == -ERESTART_RESTARTBLOCK)) 776 ret = -EINTR; 777 aio_complete(iocb, ret, 0); 778 } 779 out: 780 spin_lock_irq(&ctx->ctx_lock); 781 782 if (-EIOCBRETRY == ret) { 783 /* 784 * OK, now that we are done with this iteration 785 * and know that there is more left to go, 786 * this is where we let go so that a subsequent 787 * "kick" can start the next iteration 788 */ 789 790 /* will make __queue_kicked_iocb succeed from here on */ 791 INIT_LIST_HEAD(&iocb->ki_run_list); 792 /* we must queue the next iteration ourselves, if it 793 * has already been kicked */ 794 if (kiocbIsKicked(iocb)) { 795 __queue_kicked_iocb(iocb); 796 797 /* 798 * __queue_kicked_iocb will always return 1 here, because 799 * iocb->ki_run_list is empty at this point so it should 800 * be safe to unconditionally queue the context into the 801 * work queue. 802 */ 803 aio_queue_work(ctx); 804 } 805 } 806 return ret; 807 } 808 809 /* 810 * __aio_run_iocbs: 811 * Process all pending retries queued on the ioctx 812 * run list. 813 * Assumes it is operating within the aio issuer's mm 814 * context. 815 */ 816 static int __aio_run_iocbs(struct kioctx *ctx) 817 { 818 struct kiocb *iocb; 819 struct list_head run_list; 820 821 assert_spin_locked(&ctx->ctx_lock); 822 823 list_replace_init(&ctx->run_list, &run_list); 824 while (!list_empty(&run_list)) { 825 iocb = list_entry(run_list.next, struct kiocb, 826 ki_run_list); 827 list_del(&iocb->ki_run_list); 828 /* 829 * Hold an extra reference while retrying i/o. 830 */ 831 iocb->ki_users++; /* grab extra reference */ 832 aio_run_iocb(iocb); 833 __aio_put_req(ctx, iocb); 834 } 835 if (!list_empty(&ctx->run_list)) 836 return 1; 837 return 0; 838 } 839 840 static void aio_queue_work(struct kioctx * ctx) 841 { 842 unsigned long timeout; 843 /* 844 * if someone is waiting, get the work started right 845 * away, otherwise, use a longer delay 846 */ 847 smp_mb(); 848 if (waitqueue_active(&ctx->wait)) 849 timeout = 1; 850 else 851 timeout = HZ/10; 852 queue_delayed_work(aio_wq, &ctx->wq, timeout); 853 } 854 855 /* 856 * aio_run_all_iocbs: 857 * Process all pending retries queued on the ioctx 858 * run list, and keep running them until the list 859 * stays empty. 860 * Assumes it is operating within the aio issuer's mm context. 861 */ 862 static inline void aio_run_all_iocbs(struct kioctx *ctx) 863 { 864 spin_lock_irq(&ctx->ctx_lock); 865 while (__aio_run_iocbs(ctx)) 866 ; 867 spin_unlock_irq(&ctx->ctx_lock); 868 } 869 870 /* 871 * aio_kick_handler: 872 * Work queue handler triggered to process pending 873 * retries on an ioctx. Takes on the aio issuer's 874 * mm context before running the iocbs, so that 875 * copy_xxx_user operates on the issuer's address 876 * space. 877 * Run on aiod's context. 878 */ 879 static void aio_kick_handler(struct work_struct *work) 880 { 881 struct kioctx *ctx = container_of(work, struct kioctx, wq.work); 882 mm_segment_t oldfs = get_fs(); 883 struct mm_struct *mm; 884 int requeue; 885 886 set_fs(USER_DS); 887 use_mm(ctx->mm); 888 spin_lock_irq(&ctx->ctx_lock); 889 requeue =__aio_run_iocbs(ctx); 890 mm = ctx->mm; 891 spin_unlock_irq(&ctx->ctx_lock); 892 unuse_mm(mm); 893 set_fs(oldfs); 894 /* 895 * we're in a worker thread already; no point using non-zero delay 896 */ 897 if (requeue) 898 queue_delayed_work(aio_wq, &ctx->wq, 0); 899 } 900 901 902 /* 903 * Called by kick_iocb to queue the kiocb for retry 904 * and if required activate the aio work queue to process 905 * it 906 */ 907 static void try_queue_kicked_iocb(struct kiocb *iocb) 908 { 909 struct kioctx *ctx = iocb->ki_ctx; 910 unsigned long flags; 911 int run = 0; 912 913 spin_lock_irqsave(&ctx->ctx_lock, flags); 914 /* set this inside the lock so that we can't race with aio_run_iocb() 915 * testing it and putting the iocb on the run list under the lock */ 916 if (!kiocbTryKick(iocb)) 917 run = __queue_kicked_iocb(iocb); 918 spin_unlock_irqrestore(&ctx->ctx_lock, flags); 919 if (run) 920 aio_queue_work(ctx); 921 } 922 923 /* 924 * kick_iocb: 925 * Called typically from a wait queue callback context 926 * to trigger a retry of the iocb. 927 * The retry is usually executed by aio workqueue 928 * threads (See aio_kick_handler). 929 */ 930 void kick_iocb(struct kiocb *iocb) 931 { 932 /* sync iocbs are easy: they can only ever be executing from a 933 * single context. */ 934 if (is_sync_kiocb(iocb)) { 935 kiocbSetKicked(iocb); 936 wake_up_process(iocb->ki_obj.tsk); 937 return; 938 } 939 940 try_queue_kicked_iocb(iocb); 941 } 942 EXPORT_SYMBOL(kick_iocb); 943 944 /* aio_complete 945 * Called when the io request on the given iocb is complete. 946 * Returns true if this is the last user of the request. The 947 * only other user of the request can be the cancellation code. 948 */ 949 int aio_complete(struct kiocb *iocb, long res, long res2) 950 { 951 struct kioctx *ctx = iocb->ki_ctx; 952 struct aio_ring_info *info; 953 struct aio_ring *ring; 954 struct io_event *event; 955 unsigned long flags; 956 unsigned long tail; 957 int ret; 958 959 /* 960 * Special case handling for sync iocbs: 961 * - events go directly into the iocb for fast handling 962 * - the sync task with the iocb in its stack holds the single iocb 963 * ref, no other paths have a way to get another ref 964 * - the sync task helpfully left a reference to itself in the iocb 965 */ 966 if (is_sync_kiocb(iocb)) { 967 BUG_ON(iocb->ki_users != 1); 968 iocb->ki_user_data = res; 969 iocb->ki_users = 0; 970 wake_up_process(iocb->ki_obj.tsk); 971 return 1; 972 } 973 974 info = &ctx->ring_info; 975 976 /* add a completion event to the ring buffer. 977 * must be done holding ctx->ctx_lock to prevent 978 * other code from messing with the tail 979 * pointer since we might be called from irq 980 * context. 981 */ 982 spin_lock_irqsave(&ctx->ctx_lock, flags); 983 984 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list)) 985 list_del_init(&iocb->ki_run_list); 986 987 /* 988 * cancelled requests don't get events, userland was given one 989 * when the event got cancelled. 990 */ 991 if (kiocbIsCancelled(iocb)) 992 goto put_rq; 993 994 ring = kmap_atomic(info->ring_pages[0]); 995 996 tail = info->tail; 997 event = aio_ring_event(info, tail); 998 if (++tail >= info->nr) 999 tail = 0; 1000 1001 event->obj = (u64)(unsigned long)iocb->ki_obj.user; 1002 event->data = iocb->ki_user_data; 1003 event->res = res; 1004 event->res2 = res2; 1005 1006 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n", 1007 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data, 1008 res, res2); 1009 1010 /* after flagging the request as done, we 1011 * must never even look at it again 1012 */ 1013 smp_wmb(); /* make event visible before updating tail */ 1014 1015 info->tail = tail; 1016 ring->tail = tail; 1017 1018 put_aio_ring_event(event); 1019 kunmap_atomic(ring); 1020 1021 pr_debug("added to ring %p at [%lu]\n", iocb, tail); 1022 1023 /* 1024 * Check if the user asked us to deliver the result through an 1025 * eventfd. The eventfd_signal() function is safe to be called 1026 * from IRQ context. 1027 */ 1028 if (iocb->ki_eventfd != NULL) 1029 eventfd_signal(iocb->ki_eventfd, 1); 1030 1031 put_rq: 1032 /* everything turned out well, dispose of the aiocb. */ 1033 ret = __aio_put_req(ctx, iocb); 1034 1035 /* 1036 * We have to order our ring_info tail store above and test 1037 * of the wait list below outside the wait lock. This is 1038 * like in wake_up_bit() where clearing a bit has to be 1039 * ordered with the unlocked test. 1040 */ 1041 smp_mb(); 1042 1043 if (waitqueue_active(&ctx->wait)) 1044 wake_up(&ctx->wait); 1045 1046 spin_unlock_irqrestore(&ctx->ctx_lock, flags); 1047 return ret; 1048 } 1049 EXPORT_SYMBOL(aio_complete); 1050 1051 /* aio_read_evt 1052 * Pull an event off of the ioctx's event ring. Returns the number of 1053 * events fetched (0 or 1 ;-) 1054 * FIXME: make this use cmpxchg. 1055 * TODO: make the ringbuffer user mmap()able (requires FIXME). 1056 */ 1057 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent) 1058 { 1059 struct aio_ring_info *info = &ioctx->ring_info; 1060 struct aio_ring *ring; 1061 unsigned long head; 1062 int ret = 0; 1063 1064 ring = kmap_atomic(info->ring_pages[0]); 1065 dprintk("in aio_read_evt h%lu t%lu m%lu\n", 1066 (unsigned long)ring->head, (unsigned long)ring->tail, 1067 (unsigned long)ring->nr); 1068 1069 if (ring->head == ring->tail) 1070 goto out; 1071 1072 spin_lock(&info->ring_lock); 1073 1074 head = ring->head % info->nr; 1075 if (head != ring->tail) { 1076 struct io_event *evp = aio_ring_event(info, head); 1077 *ent = *evp; 1078 head = (head + 1) % info->nr; 1079 smp_mb(); /* finish reading the event before updatng the head */ 1080 ring->head = head; 1081 ret = 1; 1082 put_aio_ring_event(evp); 1083 } 1084 spin_unlock(&info->ring_lock); 1085 1086 out: 1087 kunmap_atomic(ring); 1088 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret, 1089 (unsigned long)ring->head, (unsigned long)ring->tail); 1090 return ret; 1091 } 1092 1093 struct aio_timeout { 1094 struct timer_list timer; 1095 int timed_out; 1096 struct task_struct *p; 1097 }; 1098 1099 static void timeout_func(unsigned long data) 1100 { 1101 struct aio_timeout *to = (struct aio_timeout *)data; 1102 1103 to->timed_out = 1; 1104 wake_up_process(to->p); 1105 } 1106 1107 static inline void init_timeout(struct aio_timeout *to) 1108 { 1109 setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to); 1110 to->timed_out = 0; 1111 to->p = current; 1112 } 1113 1114 static inline void set_timeout(long start_jiffies, struct aio_timeout *to, 1115 const struct timespec *ts) 1116 { 1117 to->timer.expires = start_jiffies + timespec_to_jiffies(ts); 1118 if (time_after(to->timer.expires, jiffies)) 1119 add_timer(&to->timer); 1120 else 1121 to->timed_out = 1; 1122 } 1123 1124 static inline void clear_timeout(struct aio_timeout *to) 1125 { 1126 del_singleshot_timer_sync(&to->timer); 1127 } 1128 1129 static int read_events(struct kioctx *ctx, 1130 long min_nr, long nr, 1131 struct io_event __user *event, 1132 struct timespec __user *timeout) 1133 { 1134 long start_jiffies = jiffies; 1135 struct task_struct *tsk = current; 1136 DECLARE_WAITQUEUE(wait, tsk); 1137 int ret; 1138 int i = 0; 1139 struct io_event ent; 1140 struct aio_timeout to; 1141 int retry = 0; 1142 1143 /* needed to zero any padding within an entry (there shouldn't be 1144 * any, but C is fun! 1145 */ 1146 memset(&ent, 0, sizeof(ent)); 1147 retry: 1148 ret = 0; 1149 while (likely(i < nr)) { 1150 ret = aio_read_evt(ctx, &ent); 1151 if (unlikely(ret <= 0)) 1152 break; 1153 1154 dprintk("read event: %Lx %Lx %Lx %Lx\n", 1155 ent.data, ent.obj, ent.res, ent.res2); 1156 1157 /* Could we split the check in two? */ 1158 ret = -EFAULT; 1159 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) { 1160 dprintk("aio: lost an event due to EFAULT.\n"); 1161 break; 1162 } 1163 ret = 0; 1164 1165 /* Good, event copied to userland, update counts. */ 1166 event ++; 1167 i ++; 1168 } 1169 1170 if (min_nr <= i) 1171 return i; 1172 if (ret) 1173 return ret; 1174 1175 /* End fast path */ 1176 1177 /* racey check, but it gets redone */ 1178 if (!retry && unlikely(!list_empty(&ctx->run_list))) { 1179 retry = 1; 1180 aio_run_all_iocbs(ctx); 1181 goto retry; 1182 } 1183 1184 init_timeout(&to); 1185 if (timeout) { 1186 struct timespec ts; 1187 ret = -EFAULT; 1188 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts)))) 1189 goto out; 1190 1191 set_timeout(start_jiffies, &to, &ts); 1192 } 1193 1194 while (likely(i < nr)) { 1195 add_wait_queue_exclusive(&ctx->wait, &wait); 1196 do { 1197 set_task_state(tsk, TASK_INTERRUPTIBLE); 1198 ret = aio_read_evt(ctx, &ent); 1199 if (ret) 1200 break; 1201 if (min_nr <= i) 1202 break; 1203 if (unlikely(ctx->dead)) { 1204 ret = -EINVAL; 1205 break; 1206 } 1207 if (to.timed_out) /* Only check after read evt */ 1208 break; 1209 /* Try to only show up in io wait if there are ops 1210 * in flight */ 1211 if (ctx->reqs_active) 1212 io_schedule(); 1213 else 1214 schedule(); 1215 if (signal_pending(tsk)) { 1216 ret = -EINTR; 1217 break; 1218 } 1219 /*ret = aio_read_evt(ctx, &ent);*/ 1220 } while (1) ; 1221 1222 set_task_state(tsk, TASK_RUNNING); 1223 remove_wait_queue(&ctx->wait, &wait); 1224 1225 if (unlikely(ret <= 0)) 1226 break; 1227 1228 ret = -EFAULT; 1229 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) { 1230 dprintk("aio: lost an event due to EFAULT.\n"); 1231 break; 1232 } 1233 1234 /* Good, event copied to userland, update counts. */ 1235 event ++; 1236 i ++; 1237 } 1238 1239 if (timeout) 1240 clear_timeout(&to); 1241 out: 1242 destroy_timer_on_stack(&to.timer); 1243 return i ? i : ret; 1244 } 1245 1246 /* Take an ioctx and remove it from the list of ioctx's. Protects 1247 * against races with itself via ->dead. 1248 */ 1249 static void io_destroy(struct kioctx *ioctx) 1250 { 1251 struct mm_struct *mm = current->mm; 1252 int was_dead; 1253 1254 /* delete the entry from the list is someone else hasn't already */ 1255 spin_lock(&mm->ioctx_lock); 1256 was_dead = ioctx->dead; 1257 ioctx->dead = 1; 1258 hlist_del_rcu(&ioctx->list); 1259 spin_unlock(&mm->ioctx_lock); 1260 1261 dprintk("aio_release(%p)\n", ioctx); 1262 if (likely(!was_dead)) 1263 put_ioctx(ioctx); /* twice for the list */ 1264 1265 kill_ctx(ioctx); 1266 1267 /* 1268 * Wake up any waiters. The setting of ctx->dead must be seen 1269 * by other CPUs at this point. Right now, we rely on the 1270 * locking done by the above calls to ensure this consistency. 1271 */ 1272 wake_up_all(&ioctx->wait); 1273 } 1274 1275 /* sys_io_setup: 1276 * Create an aio_context capable of receiving at least nr_events. 1277 * ctxp must not point to an aio_context that already exists, and 1278 * must be initialized to 0 prior to the call. On successful 1279 * creation of the aio_context, *ctxp is filled in with the resulting 1280 * handle. May fail with -EINVAL if *ctxp is not initialized, 1281 * if the specified nr_events exceeds internal limits. May fail 1282 * with -EAGAIN if the specified nr_events exceeds the user's limit 1283 * of available events. May fail with -ENOMEM if insufficient kernel 1284 * resources are available. May fail with -EFAULT if an invalid 1285 * pointer is passed for ctxp. Will fail with -ENOSYS if not 1286 * implemented. 1287 */ 1288 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp) 1289 { 1290 struct kioctx *ioctx = NULL; 1291 unsigned long ctx; 1292 long ret; 1293 1294 ret = get_user(ctx, ctxp); 1295 if (unlikely(ret)) 1296 goto out; 1297 1298 ret = -EINVAL; 1299 if (unlikely(ctx || nr_events == 0)) { 1300 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n", 1301 ctx, nr_events); 1302 goto out; 1303 } 1304 1305 ioctx = ioctx_alloc(nr_events); 1306 ret = PTR_ERR(ioctx); 1307 if (!IS_ERR(ioctx)) { 1308 ret = put_user(ioctx->user_id, ctxp); 1309 if (ret) 1310 io_destroy(ioctx); 1311 put_ioctx(ioctx); 1312 } 1313 1314 out: 1315 return ret; 1316 } 1317 1318 /* sys_io_destroy: 1319 * Destroy the aio_context specified. May cancel any outstanding 1320 * AIOs and block on completion. Will fail with -ENOSYS if not 1321 * implemented. May fail with -EINVAL if the context pointed to 1322 * is invalid. 1323 */ 1324 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx) 1325 { 1326 struct kioctx *ioctx = lookup_ioctx(ctx); 1327 if (likely(NULL != ioctx)) { 1328 io_destroy(ioctx); 1329 put_ioctx(ioctx); 1330 return 0; 1331 } 1332 pr_debug("EINVAL: io_destroy: invalid context id\n"); 1333 return -EINVAL; 1334 } 1335 1336 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret) 1337 { 1338 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg]; 1339 1340 BUG_ON(ret <= 0); 1341 1342 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) { 1343 ssize_t this = min((ssize_t)iov->iov_len, ret); 1344 iov->iov_base += this; 1345 iov->iov_len -= this; 1346 iocb->ki_left -= this; 1347 ret -= this; 1348 if (iov->iov_len == 0) { 1349 iocb->ki_cur_seg++; 1350 iov++; 1351 } 1352 } 1353 1354 /* the caller should not have done more io than what fit in 1355 * the remaining iovecs */ 1356 BUG_ON(ret > 0 && iocb->ki_left == 0); 1357 } 1358 1359 static ssize_t aio_rw_vect_retry(struct kiocb *iocb) 1360 { 1361 struct file *file = iocb->ki_filp; 1362 struct address_space *mapping = file->f_mapping; 1363 struct inode *inode = mapping->host; 1364 ssize_t (*rw_op)(struct kiocb *, const struct iovec *, 1365 unsigned long, loff_t); 1366 ssize_t ret = 0; 1367 unsigned short opcode; 1368 1369 if ((iocb->ki_opcode == IOCB_CMD_PREADV) || 1370 (iocb->ki_opcode == IOCB_CMD_PREAD)) { 1371 rw_op = file->f_op->aio_read; 1372 opcode = IOCB_CMD_PREADV; 1373 } else { 1374 rw_op = file->f_op->aio_write; 1375 opcode = IOCB_CMD_PWRITEV; 1376 } 1377 1378 /* This matches the pread()/pwrite() logic */ 1379 if (iocb->ki_pos < 0) 1380 return -EINVAL; 1381 1382 do { 1383 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg], 1384 iocb->ki_nr_segs - iocb->ki_cur_seg, 1385 iocb->ki_pos); 1386 if (ret > 0) 1387 aio_advance_iovec(iocb, ret); 1388 1389 /* retry all partial writes. retry partial reads as long as its a 1390 * regular file. */ 1391 } while (ret > 0 && iocb->ki_left > 0 && 1392 (opcode == IOCB_CMD_PWRITEV || 1393 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode)))); 1394 1395 /* This means we must have transferred all that we could */ 1396 /* No need to retry anymore */ 1397 if ((ret == 0) || (iocb->ki_left == 0)) 1398 ret = iocb->ki_nbytes - iocb->ki_left; 1399 1400 /* If we managed to write some out we return that, rather than 1401 * the eventual error. */ 1402 if (opcode == IOCB_CMD_PWRITEV 1403 && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY 1404 && iocb->ki_nbytes - iocb->ki_left) 1405 ret = iocb->ki_nbytes - iocb->ki_left; 1406 1407 return ret; 1408 } 1409 1410 static ssize_t aio_fdsync(struct kiocb *iocb) 1411 { 1412 struct file *file = iocb->ki_filp; 1413 ssize_t ret = -EINVAL; 1414 1415 if (file->f_op->aio_fsync) 1416 ret = file->f_op->aio_fsync(iocb, 1); 1417 return ret; 1418 } 1419 1420 static ssize_t aio_fsync(struct kiocb *iocb) 1421 { 1422 struct file *file = iocb->ki_filp; 1423 ssize_t ret = -EINVAL; 1424 1425 if (file->f_op->aio_fsync) 1426 ret = file->f_op->aio_fsync(iocb, 0); 1427 return ret; 1428 } 1429 1430 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat) 1431 { 1432 ssize_t ret; 1433 1434 #ifdef CONFIG_COMPAT 1435 if (compat) 1436 ret = compat_rw_copy_check_uvector(type, 1437 (struct compat_iovec __user *)kiocb->ki_buf, 1438 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec, 1439 &kiocb->ki_iovec, 1); 1440 else 1441 #endif 1442 ret = rw_copy_check_uvector(type, 1443 (struct iovec __user *)kiocb->ki_buf, 1444 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec, 1445 &kiocb->ki_iovec, 1); 1446 if (ret < 0) 1447 goto out; 1448 1449 kiocb->ki_nr_segs = kiocb->ki_nbytes; 1450 kiocb->ki_cur_seg = 0; 1451 /* ki_nbytes/left now reflect bytes instead of segs */ 1452 kiocb->ki_nbytes = ret; 1453 kiocb->ki_left = ret; 1454 1455 ret = 0; 1456 out: 1457 return ret; 1458 } 1459 1460 static ssize_t aio_setup_single_vector(struct kiocb *kiocb) 1461 { 1462 kiocb->ki_iovec = &kiocb->ki_inline_vec; 1463 kiocb->ki_iovec->iov_base = kiocb->ki_buf; 1464 kiocb->ki_iovec->iov_len = kiocb->ki_left; 1465 kiocb->ki_nr_segs = 1; 1466 kiocb->ki_cur_seg = 0; 1467 return 0; 1468 } 1469 1470 /* 1471 * aio_setup_iocb: 1472 * Performs the initial checks and aio retry method 1473 * setup for the kiocb at the time of io submission. 1474 */ 1475 static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat) 1476 { 1477 struct file *file = kiocb->ki_filp; 1478 ssize_t ret = 0; 1479 1480 switch (kiocb->ki_opcode) { 1481 case IOCB_CMD_PREAD: 1482 ret = -EBADF; 1483 if (unlikely(!(file->f_mode & FMODE_READ))) 1484 break; 1485 ret = -EFAULT; 1486 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf, 1487 kiocb->ki_left))) 1488 break; 1489 ret = security_file_permission(file, MAY_READ); 1490 if (unlikely(ret)) 1491 break; 1492 ret = aio_setup_single_vector(kiocb); 1493 if (ret) 1494 break; 1495 ret = -EINVAL; 1496 if (file->f_op->aio_read) 1497 kiocb->ki_retry = aio_rw_vect_retry; 1498 break; 1499 case IOCB_CMD_PWRITE: 1500 ret = -EBADF; 1501 if (unlikely(!(file->f_mode & FMODE_WRITE))) 1502 break; 1503 ret = -EFAULT; 1504 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf, 1505 kiocb->ki_left))) 1506 break; 1507 ret = security_file_permission(file, MAY_WRITE); 1508 if (unlikely(ret)) 1509 break; 1510 ret = aio_setup_single_vector(kiocb); 1511 if (ret) 1512 break; 1513 ret = -EINVAL; 1514 if (file->f_op->aio_write) 1515 kiocb->ki_retry = aio_rw_vect_retry; 1516 break; 1517 case IOCB_CMD_PREADV: 1518 ret = -EBADF; 1519 if (unlikely(!(file->f_mode & FMODE_READ))) 1520 break; 1521 ret = security_file_permission(file, MAY_READ); 1522 if (unlikely(ret)) 1523 break; 1524 ret = aio_setup_vectored_rw(READ, kiocb, compat); 1525 if (ret) 1526 break; 1527 ret = -EINVAL; 1528 if (file->f_op->aio_read) 1529 kiocb->ki_retry = aio_rw_vect_retry; 1530 break; 1531 case IOCB_CMD_PWRITEV: 1532 ret = -EBADF; 1533 if (unlikely(!(file->f_mode & FMODE_WRITE))) 1534 break; 1535 ret = security_file_permission(file, MAY_WRITE); 1536 if (unlikely(ret)) 1537 break; 1538 ret = aio_setup_vectored_rw(WRITE, kiocb, compat); 1539 if (ret) 1540 break; 1541 ret = -EINVAL; 1542 if (file->f_op->aio_write) 1543 kiocb->ki_retry = aio_rw_vect_retry; 1544 break; 1545 case IOCB_CMD_FDSYNC: 1546 ret = -EINVAL; 1547 if (file->f_op->aio_fsync) 1548 kiocb->ki_retry = aio_fdsync; 1549 break; 1550 case IOCB_CMD_FSYNC: 1551 ret = -EINVAL; 1552 if (file->f_op->aio_fsync) 1553 kiocb->ki_retry = aio_fsync; 1554 break; 1555 default: 1556 dprintk("EINVAL: io_submit: no operation provided\n"); 1557 ret = -EINVAL; 1558 } 1559 1560 if (!kiocb->ki_retry) 1561 return ret; 1562 1563 return 0; 1564 } 1565 1566 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb, 1567 struct iocb *iocb, struct kiocb_batch *batch, 1568 bool compat) 1569 { 1570 struct kiocb *req; 1571 struct file *file; 1572 ssize_t ret; 1573 1574 /* enforce forwards compatibility on users */ 1575 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) { 1576 pr_debug("EINVAL: io_submit: reserve field set\n"); 1577 return -EINVAL; 1578 } 1579 1580 /* prevent overflows */ 1581 if (unlikely( 1582 (iocb->aio_buf != (unsigned long)iocb->aio_buf) || 1583 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) || 1584 ((ssize_t)iocb->aio_nbytes < 0) 1585 )) { 1586 pr_debug("EINVAL: io_submit: overflow check\n"); 1587 return -EINVAL; 1588 } 1589 1590 file = fget(iocb->aio_fildes); 1591 if (unlikely(!file)) 1592 return -EBADF; 1593 1594 req = aio_get_req(ctx, batch); /* returns with 2 references to req */ 1595 if (unlikely(!req)) { 1596 fput(file); 1597 return -EAGAIN; 1598 } 1599 req->ki_filp = file; 1600 if (iocb->aio_flags & IOCB_FLAG_RESFD) { 1601 /* 1602 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an 1603 * instance of the file* now. The file descriptor must be 1604 * an eventfd() fd, and will be signaled for each completed 1605 * event using the eventfd_signal() function. 1606 */ 1607 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd); 1608 if (IS_ERR(req->ki_eventfd)) { 1609 ret = PTR_ERR(req->ki_eventfd); 1610 req->ki_eventfd = NULL; 1611 goto out_put_req; 1612 } 1613 } 1614 1615 ret = put_user(req->ki_key, &user_iocb->aio_key); 1616 if (unlikely(ret)) { 1617 dprintk("EFAULT: aio_key\n"); 1618 goto out_put_req; 1619 } 1620 1621 req->ki_obj.user = user_iocb; 1622 req->ki_user_data = iocb->aio_data; 1623 req->ki_pos = iocb->aio_offset; 1624 1625 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf; 1626 req->ki_left = req->ki_nbytes = iocb->aio_nbytes; 1627 req->ki_opcode = iocb->aio_lio_opcode; 1628 1629 ret = aio_setup_iocb(req, compat); 1630 1631 if (ret) 1632 goto out_put_req; 1633 1634 spin_lock_irq(&ctx->ctx_lock); 1635 /* 1636 * We could have raced with io_destroy() and are currently holding a 1637 * reference to ctx which should be destroyed. We cannot submit IO 1638 * since ctx gets freed as soon as io_submit() puts its reference. The 1639 * check here is reliable: io_destroy() sets ctx->dead before waiting 1640 * for outstanding IO and the barrier between these two is realized by 1641 * unlock of mm->ioctx_lock and lock of ctx->ctx_lock. Analogously we 1642 * increment ctx->reqs_active before checking for ctx->dead and the 1643 * barrier is realized by unlock and lock of ctx->ctx_lock. Thus if we 1644 * don't see ctx->dead set here, io_destroy() waits for our IO to 1645 * finish. 1646 */ 1647 if (ctx->dead) { 1648 spin_unlock_irq(&ctx->ctx_lock); 1649 ret = -EINVAL; 1650 goto out_put_req; 1651 } 1652 aio_run_iocb(req); 1653 if (!list_empty(&ctx->run_list)) { 1654 /* drain the run list */ 1655 while (__aio_run_iocbs(ctx)) 1656 ; 1657 } 1658 spin_unlock_irq(&ctx->ctx_lock); 1659 1660 aio_put_req(req); /* drop extra ref to req */ 1661 return 0; 1662 1663 out_put_req: 1664 aio_put_req(req); /* drop extra ref to req */ 1665 aio_put_req(req); /* drop i/o ref to req */ 1666 return ret; 1667 } 1668 1669 long do_io_submit(aio_context_t ctx_id, long nr, 1670 struct iocb __user *__user *iocbpp, bool compat) 1671 { 1672 struct kioctx *ctx; 1673 long ret = 0; 1674 int i = 0; 1675 struct blk_plug plug; 1676 struct kiocb_batch batch; 1677 1678 if (unlikely(nr < 0)) 1679 return -EINVAL; 1680 1681 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp))) 1682 nr = LONG_MAX/sizeof(*iocbpp); 1683 1684 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp))))) 1685 return -EFAULT; 1686 1687 ctx = lookup_ioctx(ctx_id); 1688 if (unlikely(!ctx)) { 1689 pr_debug("EINVAL: io_submit: invalid context id\n"); 1690 return -EINVAL; 1691 } 1692 1693 kiocb_batch_init(&batch, nr); 1694 1695 blk_start_plug(&plug); 1696 1697 /* 1698 * AKPM: should this return a partial result if some of the IOs were 1699 * successfully submitted? 1700 */ 1701 for (i=0; i<nr; i++) { 1702 struct iocb __user *user_iocb; 1703 struct iocb tmp; 1704 1705 if (unlikely(__get_user(user_iocb, iocbpp + i))) { 1706 ret = -EFAULT; 1707 break; 1708 } 1709 1710 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) { 1711 ret = -EFAULT; 1712 break; 1713 } 1714 1715 ret = io_submit_one(ctx, user_iocb, &tmp, &batch, compat); 1716 if (ret) 1717 break; 1718 } 1719 blk_finish_plug(&plug); 1720 1721 kiocb_batch_free(ctx, &batch); 1722 put_ioctx(ctx); 1723 return i ? i : ret; 1724 } 1725 1726 /* sys_io_submit: 1727 * Queue the nr iocbs pointed to by iocbpp for processing. Returns 1728 * the number of iocbs queued. May return -EINVAL if the aio_context 1729 * specified by ctx_id is invalid, if nr is < 0, if the iocb at 1730 * *iocbpp[0] is not properly initialized, if the operation specified 1731 * is invalid for the file descriptor in the iocb. May fail with 1732 * -EFAULT if any of the data structures point to invalid data. May 1733 * fail with -EBADF if the file descriptor specified in the first 1734 * iocb is invalid. May fail with -EAGAIN if insufficient resources 1735 * are available to queue any iocbs. Will return 0 if nr is 0. Will 1736 * fail with -ENOSYS if not implemented. 1737 */ 1738 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr, 1739 struct iocb __user * __user *, iocbpp) 1740 { 1741 return do_io_submit(ctx_id, nr, iocbpp, 0); 1742 } 1743 1744 /* lookup_kiocb 1745 * Finds a given iocb for cancellation. 1746 */ 1747 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, 1748 u32 key) 1749 { 1750 struct list_head *pos; 1751 1752 assert_spin_locked(&ctx->ctx_lock); 1753 1754 /* TODO: use a hash or array, this sucks. */ 1755 list_for_each(pos, &ctx->active_reqs) { 1756 struct kiocb *kiocb = list_kiocb(pos); 1757 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key) 1758 return kiocb; 1759 } 1760 return NULL; 1761 } 1762 1763 /* sys_io_cancel: 1764 * Attempts to cancel an iocb previously passed to io_submit. If 1765 * the operation is successfully cancelled, the resulting event is 1766 * copied into the memory pointed to by result without being placed 1767 * into the completion queue and 0 is returned. May fail with 1768 * -EFAULT if any of the data structures pointed to are invalid. 1769 * May fail with -EINVAL if aio_context specified by ctx_id is 1770 * invalid. May fail with -EAGAIN if the iocb specified was not 1771 * cancelled. Will fail with -ENOSYS if not implemented. 1772 */ 1773 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb, 1774 struct io_event __user *, result) 1775 { 1776 int (*cancel)(struct kiocb *iocb, struct io_event *res); 1777 struct kioctx *ctx; 1778 struct kiocb *kiocb; 1779 u32 key; 1780 int ret; 1781 1782 ret = get_user(key, &iocb->aio_key); 1783 if (unlikely(ret)) 1784 return -EFAULT; 1785 1786 ctx = lookup_ioctx(ctx_id); 1787 if (unlikely(!ctx)) 1788 return -EINVAL; 1789 1790 spin_lock_irq(&ctx->ctx_lock); 1791 ret = -EAGAIN; 1792 kiocb = lookup_kiocb(ctx, iocb, key); 1793 if (kiocb && kiocb->ki_cancel) { 1794 cancel = kiocb->ki_cancel; 1795 kiocb->ki_users ++; 1796 kiocbSetCancelled(kiocb); 1797 } else 1798 cancel = NULL; 1799 spin_unlock_irq(&ctx->ctx_lock); 1800 1801 if (NULL != cancel) { 1802 struct io_event tmp; 1803 pr_debug("calling cancel\n"); 1804 memset(&tmp, 0, sizeof(tmp)); 1805 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user; 1806 tmp.data = kiocb->ki_user_data; 1807 ret = cancel(kiocb, &tmp); 1808 if (!ret) { 1809 /* Cancellation succeeded -- copy the result 1810 * into the user's buffer. 1811 */ 1812 if (copy_to_user(result, &tmp, sizeof(tmp))) 1813 ret = -EFAULT; 1814 } 1815 } else 1816 ret = -EINVAL; 1817 1818 put_ioctx(ctx); 1819 1820 return ret; 1821 } 1822 1823 /* io_getevents: 1824 * Attempts to read at least min_nr events and up to nr events from 1825 * the completion queue for the aio_context specified by ctx_id. If 1826 * it succeeds, the number of read events is returned. May fail with 1827 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is 1828 * out of range, if timeout is out of range. May fail with -EFAULT 1829 * if any of the memory specified is invalid. May return 0 or 1830 * < min_nr if the timeout specified by timeout has elapsed 1831 * before sufficient events are available, where timeout == NULL 1832 * specifies an infinite timeout. Note that the timeout pointed to by 1833 * timeout is relative and will be updated if not NULL and the 1834 * operation blocks. Will fail with -ENOSYS if not implemented. 1835 */ 1836 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id, 1837 long, min_nr, 1838 long, nr, 1839 struct io_event __user *, events, 1840 struct timespec __user *, timeout) 1841 { 1842 struct kioctx *ioctx = lookup_ioctx(ctx_id); 1843 long ret = -EINVAL; 1844 1845 if (likely(ioctx)) { 1846 if (likely(min_nr <= nr && min_nr >= 0)) 1847 ret = read_events(ioctx, min_nr, nr, events, timeout); 1848 put_ioctx(ioctx); 1849 } 1850 1851 asmlinkage_protect(5, ret, ctx_id, min_nr, nr, events, timeout); 1852 return ret; 1853 } 1854