1 /* 2 * An async IO implementation for Linux 3 * Written by Benjamin LaHaise <bcrl@kvack.org> 4 * 5 * Implements an efficient asynchronous io interface. 6 * 7 * Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved. 8 * 9 * See ../COPYING for licensing terms. 10 */ 11 #include <linux/kernel.h> 12 #include <linux/init.h> 13 #include <linux/errno.h> 14 #include <linux/time.h> 15 #include <linux/aio_abi.h> 16 #include <linux/module.h> 17 #include <linux/syscalls.h> 18 #include <linux/backing-dev.h> 19 #include <linux/uio.h> 20 21 #define DEBUG 0 22 23 #include <linux/sched.h> 24 #include <linux/fs.h> 25 #include <linux/file.h> 26 #include <linux/mm.h> 27 #include <linux/mman.h> 28 #include <linux/mmu_context.h> 29 #include <linux/slab.h> 30 #include <linux/timer.h> 31 #include <linux/aio.h> 32 #include <linux/highmem.h> 33 #include <linux/workqueue.h> 34 #include <linux/security.h> 35 #include <linux/eventfd.h> 36 #include <linux/blkdev.h> 37 #include <linux/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 /* aio_cancel_all 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 aio_cancel_all(struct kioctx *ctx) 314 { 315 int (*cancel)(struct kiocb *, struct io_event *); 316 struct io_event res; 317 spin_lock_irq(&ctx->ctx_lock); 318 ctx->dead = 1; 319 while (!list_empty(&ctx->active_reqs)) { 320 struct list_head *pos = ctx->active_reqs.next; 321 struct kiocb *iocb = list_kiocb(pos); 322 list_del_init(&iocb->ki_list); 323 cancel = iocb->ki_cancel; 324 kiocbSetCancelled(iocb); 325 if (cancel) { 326 iocb->ki_users++; 327 spin_unlock_irq(&ctx->ctx_lock); 328 cancel(iocb, &res); 329 spin_lock_irq(&ctx->ctx_lock); 330 } 331 } 332 spin_unlock_irq(&ctx->ctx_lock); 333 } 334 335 static void wait_for_all_aios(struct kioctx *ctx) 336 { 337 struct task_struct *tsk = current; 338 DECLARE_WAITQUEUE(wait, tsk); 339 340 spin_lock_irq(&ctx->ctx_lock); 341 if (!ctx->reqs_active) 342 goto out; 343 344 add_wait_queue(&ctx->wait, &wait); 345 set_task_state(tsk, TASK_UNINTERRUPTIBLE); 346 while (ctx->reqs_active) { 347 spin_unlock_irq(&ctx->ctx_lock); 348 io_schedule(); 349 set_task_state(tsk, TASK_UNINTERRUPTIBLE); 350 spin_lock_irq(&ctx->ctx_lock); 351 } 352 __set_task_state(tsk, TASK_RUNNING); 353 remove_wait_queue(&ctx->wait, &wait); 354 355 out: 356 spin_unlock_irq(&ctx->ctx_lock); 357 } 358 359 /* wait_on_sync_kiocb: 360 * Waits on the given sync kiocb to complete. 361 */ 362 ssize_t wait_on_sync_kiocb(struct kiocb *iocb) 363 { 364 while (iocb->ki_users) { 365 set_current_state(TASK_UNINTERRUPTIBLE); 366 if (!iocb->ki_users) 367 break; 368 io_schedule(); 369 } 370 __set_current_state(TASK_RUNNING); 371 return iocb->ki_user_data; 372 } 373 EXPORT_SYMBOL(wait_on_sync_kiocb); 374 375 /* exit_aio: called when the last user of mm goes away. At this point, 376 * there is no way for any new requests to be submited or any of the 377 * io_* syscalls to be called on the context. However, there may be 378 * outstanding requests which hold references to the context; as they 379 * go away, they will call put_ioctx and release any pinned memory 380 * associated with the request (held via struct page * references). 381 */ 382 void exit_aio(struct mm_struct *mm) 383 { 384 struct kioctx *ctx; 385 386 while (!hlist_empty(&mm->ioctx_list)) { 387 ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list); 388 hlist_del_rcu(&ctx->list); 389 390 aio_cancel_all(ctx); 391 392 wait_for_all_aios(ctx); 393 394 if (1 != atomic_read(&ctx->users)) 395 printk(KERN_DEBUG 396 "exit_aio:ioctx still alive: %d %d %d\n", 397 atomic_read(&ctx->users), ctx->dead, 398 ctx->reqs_active); 399 put_ioctx(ctx); 400 } 401 } 402 403 /* aio_get_req 404 * Allocate a slot for an aio request. Increments the users count 405 * of the kioctx so that the kioctx stays around until all requests are 406 * complete. Returns NULL if no requests are free. 407 * 408 * Returns with kiocb->users set to 2. The io submit code path holds 409 * an extra reference while submitting the i/o. 410 * This prevents races between the aio code path referencing the 411 * req (after submitting it) and aio_complete() freeing the req. 412 */ 413 static struct kiocb *__aio_get_req(struct kioctx *ctx) 414 { 415 struct kiocb *req = NULL; 416 417 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL); 418 if (unlikely(!req)) 419 return NULL; 420 421 req->ki_flags = 0; 422 req->ki_users = 2; 423 req->ki_key = 0; 424 req->ki_ctx = ctx; 425 req->ki_cancel = NULL; 426 req->ki_retry = NULL; 427 req->ki_dtor = NULL; 428 req->private = NULL; 429 req->ki_iovec = NULL; 430 INIT_LIST_HEAD(&req->ki_run_list); 431 req->ki_eventfd = NULL; 432 433 return req; 434 } 435 436 /* 437 * struct kiocb's are allocated in batches to reduce the number of 438 * times the ctx lock is acquired and released. 439 */ 440 #define KIOCB_BATCH_SIZE 32L 441 struct kiocb_batch { 442 struct list_head head; 443 long count; /* number of requests left to allocate */ 444 }; 445 446 static void kiocb_batch_init(struct kiocb_batch *batch, long total) 447 { 448 INIT_LIST_HEAD(&batch->head); 449 batch->count = total; 450 } 451 452 static void kiocb_batch_free(struct kioctx *ctx, struct kiocb_batch *batch) 453 { 454 struct kiocb *req, *n; 455 456 if (list_empty(&batch->head)) 457 return; 458 459 spin_lock_irq(&ctx->ctx_lock); 460 list_for_each_entry_safe(req, n, &batch->head, ki_batch) { 461 list_del(&req->ki_batch); 462 list_del(&req->ki_list); 463 kmem_cache_free(kiocb_cachep, req); 464 ctx->reqs_active--; 465 } 466 if (unlikely(!ctx->reqs_active && ctx->dead)) 467 wake_up_all(&ctx->wait); 468 spin_unlock_irq(&ctx->ctx_lock); 469 } 470 471 /* 472 * Allocate a batch of kiocbs. This avoids taking and dropping the 473 * context lock a lot during setup. 474 */ 475 static int kiocb_batch_refill(struct kioctx *ctx, struct kiocb_batch *batch) 476 { 477 unsigned short allocated, to_alloc; 478 long avail; 479 bool called_fput = false; 480 struct kiocb *req, *n; 481 struct aio_ring *ring; 482 483 to_alloc = min(batch->count, KIOCB_BATCH_SIZE); 484 for (allocated = 0; allocated < to_alloc; allocated++) { 485 req = __aio_get_req(ctx); 486 if (!req) 487 /* allocation failed, go with what we've got */ 488 break; 489 list_add(&req->ki_batch, &batch->head); 490 } 491 492 if (allocated == 0) 493 goto out; 494 495 retry: 496 spin_lock_irq(&ctx->ctx_lock); 497 ring = kmap_atomic(ctx->ring_info.ring_pages[0]); 498 499 avail = aio_ring_avail(&ctx->ring_info, ring) - ctx->reqs_active; 500 BUG_ON(avail < 0); 501 if (avail == 0 && !called_fput) { 502 /* 503 * Handle a potential starvation case. It is possible that 504 * we hold the last reference on a struct file, causing us 505 * to delay the final fput to non-irq context. In this case, 506 * ctx->reqs_active is artificially high. Calling the fput 507 * routine here may free up a slot in the event completion 508 * ring, allowing this allocation to succeed. 509 */ 510 kunmap_atomic(ring); 511 spin_unlock_irq(&ctx->ctx_lock); 512 aio_fput_routine(NULL); 513 called_fput = true; 514 goto retry; 515 } 516 517 if (avail < allocated) { 518 /* Trim back the number of requests. */ 519 list_for_each_entry_safe(req, n, &batch->head, ki_batch) { 520 list_del(&req->ki_batch); 521 kmem_cache_free(kiocb_cachep, req); 522 if (--allocated <= avail) 523 break; 524 } 525 } 526 527 batch->count -= allocated; 528 list_for_each_entry(req, &batch->head, ki_batch) { 529 list_add(&req->ki_list, &ctx->active_reqs); 530 ctx->reqs_active++; 531 } 532 533 kunmap_atomic(ring); 534 spin_unlock_irq(&ctx->ctx_lock); 535 536 out: 537 return allocated; 538 } 539 540 static inline struct kiocb *aio_get_req(struct kioctx *ctx, 541 struct kiocb_batch *batch) 542 { 543 struct kiocb *req; 544 545 if (list_empty(&batch->head)) 546 if (kiocb_batch_refill(ctx, batch) == 0) 547 return NULL; 548 req = list_first_entry(&batch->head, struct kiocb, ki_batch); 549 list_del(&req->ki_batch); 550 return req; 551 } 552 553 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req) 554 { 555 assert_spin_locked(&ctx->ctx_lock); 556 557 if (req->ki_eventfd != NULL) 558 eventfd_ctx_put(req->ki_eventfd); 559 if (req->ki_dtor) 560 req->ki_dtor(req); 561 if (req->ki_iovec != &req->ki_inline_vec) 562 kfree(req->ki_iovec); 563 kmem_cache_free(kiocb_cachep, req); 564 ctx->reqs_active--; 565 566 if (unlikely(!ctx->reqs_active && ctx->dead)) 567 wake_up_all(&ctx->wait); 568 } 569 570 static void aio_fput_routine(struct work_struct *data) 571 { 572 spin_lock_irq(&fput_lock); 573 while (likely(!list_empty(&fput_head))) { 574 struct kiocb *req = list_kiocb(fput_head.next); 575 struct kioctx *ctx = req->ki_ctx; 576 577 list_del(&req->ki_list); 578 spin_unlock_irq(&fput_lock); 579 580 /* Complete the fput(s) */ 581 if (req->ki_filp != NULL) 582 fput(req->ki_filp); 583 584 /* Link the iocb into the context's free list */ 585 rcu_read_lock(); 586 spin_lock_irq(&ctx->ctx_lock); 587 really_put_req(ctx, req); 588 /* 589 * at that point ctx might've been killed, but actual 590 * freeing is RCU'd 591 */ 592 spin_unlock_irq(&ctx->ctx_lock); 593 rcu_read_unlock(); 594 595 spin_lock_irq(&fput_lock); 596 } 597 spin_unlock_irq(&fput_lock); 598 } 599 600 /* __aio_put_req 601 * Returns true if this put was the last user of the request. 602 */ 603 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req) 604 { 605 dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n", 606 req, atomic_long_read(&req->ki_filp->f_count)); 607 608 assert_spin_locked(&ctx->ctx_lock); 609 610 req->ki_users--; 611 BUG_ON(req->ki_users < 0); 612 if (likely(req->ki_users)) 613 return 0; 614 list_del(&req->ki_list); /* remove from active_reqs */ 615 req->ki_cancel = NULL; 616 req->ki_retry = NULL; 617 618 /* 619 * Try to optimize the aio and eventfd file* puts, by avoiding to 620 * schedule work in case it is not final fput() time. In normal cases, 621 * we would not be holding the last reference to the file*, so 622 * this function will be executed w/out any aio kthread wakeup. 623 */ 624 if (unlikely(!fput_atomic(req->ki_filp))) { 625 spin_lock(&fput_lock); 626 list_add(&req->ki_list, &fput_head); 627 spin_unlock(&fput_lock); 628 schedule_work(&fput_work); 629 } else { 630 req->ki_filp = NULL; 631 really_put_req(ctx, req); 632 } 633 return 1; 634 } 635 636 /* aio_put_req 637 * Returns true if this put was the last user of the kiocb, 638 * false if the request is still in use. 639 */ 640 int aio_put_req(struct kiocb *req) 641 { 642 struct kioctx *ctx = req->ki_ctx; 643 int ret; 644 spin_lock_irq(&ctx->ctx_lock); 645 ret = __aio_put_req(ctx, req); 646 spin_unlock_irq(&ctx->ctx_lock); 647 return ret; 648 } 649 EXPORT_SYMBOL(aio_put_req); 650 651 static struct kioctx *lookup_ioctx(unsigned long ctx_id) 652 { 653 struct mm_struct *mm = current->mm; 654 struct kioctx *ctx, *ret = NULL; 655 struct hlist_node *n; 656 657 rcu_read_lock(); 658 659 hlist_for_each_entry_rcu(ctx, n, &mm->ioctx_list, list) { 660 /* 661 * RCU protects us against accessing freed memory but 662 * we have to be careful not to get a reference when the 663 * reference count already dropped to 0 (ctx->dead test 664 * is unreliable because of races). 665 */ 666 if (ctx->user_id == ctx_id && !ctx->dead && try_get_ioctx(ctx)){ 667 ret = ctx; 668 break; 669 } 670 } 671 672 rcu_read_unlock(); 673 return ret; 674 } 675 676 /* 677 * Queue up a kiocb to be retried. Assumes that the kiocb 678 * has already been marked as kicked, and places it on 679 * the retry run list for the corresponding ioctx, if it 680 * isn't already queued. Returns 1 if it actually queued 681 * the kiocb (to tell the caller to activate the work 682 * queue to process it), or 0, if it found that it was 683 * already queued. 684 */ 685 static inline int __queue_kicked_iocb(struct kiocb *iocb) 686 { 687 struct kioctx *ctx = iocb->ki_ctx; 688 689 assert_spin_locked(&ctx->ctx_lock); 690 691 if (list_empty(&iocb->ki_run_list)) { 692 list_add_tail(&iocb->ki_run_list, 693 &ctx->run_list); 694 return 1; 695 } 696 return 0; 697 } 698 699 /* aio_run_iocb 700 * This is the core aio execution routine. It is 701 * invoked both for initial i/o submission and 702 * subsequent retries via the aio_kick_handler. 703 * Expects to be invoked with iocb->ki_ctx->lock 704 * already held. The lock is released and reacquired 705 * as needed during processing. 706 * 707 * Calls the iocb retry method (already setup for the 708 * iocb on initial submission) for operation specific 709 * handling, but takes care of most of common retry 710 * execution details for a given iocb. The retry method 711 * needs to be non-blocking as far as possible, to avoid 712 * holding up other iocbs waiting to be serviced by the 713 * retry kernel thread. 714 * 715 * The trickier parts in this code have to do with 716 * ensuring that only one retry instance is in progress 717 * for a given iocb at any time. Providing that guarantee 718 * simplifies the coding of individual aio operations as 719 * it avoids various potential races. 720 */ 721 static ssize_t aio_run_iocb(struct kiocb *iocb) 722 { 723 struct kioctx *ctx = iocb->ki_ctx; 724 ssize_t (*retry)(struct kiocb *); 725 ssize_t ret; 726 727 if (!(retry = iocb->ki_retry)) { 728 printk("aio_run_iocb: iocb->ki_retry = NULL\n"); 729 return 0; 730 } 731 732 /* 733 * We don't want the next retry iteration for this 734 * operation to start until this one has returned and 735 * updated the iocb state. However, wait_queue functions 736 * can trigger a kick_iocb from interrupt context in the 737 * meantime, indicating that data is available for the next 738 * iteration. We want to remember that and enable the 739 * next retry iteration _after_ we are through with 740 * this one. 741 * 742 * So, in order to be able to register a "kick", but 743 * prevent it from being queued now, we clear the kick 744 * flag, but make the kick code *think* that the iocb is 745 * still on the run list until we are actually done. 746 * When we are done with this iteration, we check if 747 * the iocb was kicked in the meantime and if so, queue 748 * it up afresh. 749 */ 750 751 kiocbClearKicked(iocb); 752 753 /* 754 * This is so that aio_complete knows it doesn't need to 755 * pull the iocb off the run list (We can't just call 756 * INIT_LIST_HEAD because we don't want a kick_iocb to 757 * queue this on the run list yet) 758 */ 759 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL; 760 spin_unlock_irq(&ctx->ctx_lock); 761 762 /* Quit retrying if the i/o has been cancelled */ 763 if (kiocbIsCancelled(iocb)) { 764 ret = -EINTR; 765 aio_complete(iocb, ret, 0); 766 /* must not access the iocb after this */ 767 goto out; 768 } 769 770 /* 771 * Now we are all set to call the retry method in async 772 * context. 773 */ 774 ret = retry(iocb); 775 776 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) { 777 /* 778 * There's no easy way to restart the syscall since other AIO's 779 * may be already running. Just fail this IO with EINTR. 780 */ 781 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR || 782 ret == -ERESTARTNOHAND || ret == -ERESTART_RESTARTBLOCK)) 783 ret = -EINTR; 784 aio_complete(iocb, ret, 0); 785 } 786 out: 787 spin_lock_irq(&ctx->ctx_lock); 788 789 if (-EIOCBRETRY == ret) { 790 /* 791 * OK, now that we are done with this iteration 792 * and know that there is more left to go, 793 * this is where we let go so that a subsequent 794 * "kick" can start the next iteration 795 */ 796 797 /* will make __queue_kicked_iocb succeed from here on */ 798 INIT_LIST_HEAD(&iocb->ki_run_list); 799 /* we must queue the next iteration ourselves, if it 800 * has already been kicked */ 801 if (kiocbIsKicked(iocb)) { 802 __queue_kicked_iocb(iocb); 803 804 /* 805 * __queue_kicked_iocb will always return 1 here, because 806 * iocb->ki_run_list is empty at this point so it should 807 * be safe to unconditionally queue the context into the 808 * work queue. 809 */ 810 aio_queue_work(ctx); 811 } 812 } 813 return ret; 814 } 815 816 /* 817 * __aio_run_iocbs: 818 * Process all pending retries queued on the ioctx 819 * run list. 820 * Assumes it is operating within the aio issuer's mm 821 * context. 822 */ 823 static int __aio_run_iocbs(struct kioctx *ctx) 824 { 825 struct kiocb *iocb; 826 struct list_head run_list; 827 828 assert_spin_locked(&ctx->ctx_lock); 829 830 list_replace_init(&ctx->run_list, &run_list); 831 while (!list_empty(&run_list)) { 832 iocb = list_entry(run_list.next, struct kiocb, 833 ki_run_list); 834 list_del(&iocb->ki_run_list); 835 /* 836 * Hold an extra reference while retrying i/o. 837 */ 838 iocb->ki_users++; /* grab extra reference */ 839 aio_run_iocb(iocb); 840 __aio_put_req(ctx, iocb); 841 } 842 if (!list_empty(&ctx->run_list)) 843 return 1; 844 return 0; 845 } 846 847 static void aio_queue_work(struct kioctx * ctx) 848 { 849 unsigned long timeout; 850 /* 851 * if someone is waiting, get the work started right 852 * away, otherwise, use a longer delay 853 */ 854 smp_mb(); 855 if (waitqueue_active(&ctx->wait)) 856 timeout = 1; 857 else 858 timeout = HZ/10; 859 queue_delayed_work(aio_wq, &ctx->wq, timeout); 860 } 861 862 /* 863 * aio_run_all_iocbs: 864 * Process all pending retries queued on the ioctx 865 * run list, and keep running them until the list 866 * stays empty. 867 * Assumes it is operating within the aio issuer's mm context. 868 */ 869 static inline void aio_run_all_iocbs(struct kioctx *ctx) 870 { 871 spin_lock_irq(&ctx->ctx_lock); 872 while (__aio_run_iocbs(ctx)) 873 ; 874 spin_unlock_irq(&ctx->ctx_lock); 875 } 876 877 /* 878 * aio_kick_handler: 879 * Work queue handler triggered to process pending 880 * retries on an ioctx. Takes on the aio issuer's 881 * mm context before running the iocbs, so that 882 * copy_xxx_user operates on the issuer's address 883 * space. 884 * Run on aiod's context. 885 */ 886 static void aio_kick_handler(struct work_struct *work) 887 { 888 struct kioctx *ctx = container_of(work, struct kioctx, wq.work); 889 mm_segment_t oldfs = get_fs(); 890 struct mm_struct *mm; 891 int requeue; 892 893 set_fs(USER_DS); 894 use_mm(ctx->mm); 895 spin_lock_irq(&ctx->ctx_lock); 896 requeue =__aio_run_iocbs(ctx); 897 mm = ctx->mm; 898 spin_unlock_irq(&ctx->ctx_lock); 899 unuse_mm(mm); 900 set_fs(oldfs); 901 /* 902 * we're in a worker thread already; no point using non-zero delay 903 */ 904 if (requeue) 905 queue_delayed_work(aio_wq, &ctx->wq, 0); 906 } 907 908 909 /* 910 * Called by kick_iocb to queue the kiocb for retry 911 * and if required activate the aio work queue to process 912 * it 913 */ 914 static void try_queue_kicked_iocb(struct kiocb *iocb) 915 { 916 struct kioctx *ctx = iocb->ki_ctx; 917 unsigned long flags; 918 int run = 0; 919 920 spin_lock_irqsave(&ctx->ctx_lock, flags); 921 /* set this inside the lock so that we can't race with aio_run_iocb() 922 * testing it and putting the iocb on the run list under the lock */ 923 if (!kiocbTryKick(iocb)) 924 run = __queue_kicked_iocb(iocb); 925 spin_unlock_irqrestore(&ctx->ctx_lock, flags); 926 if (run) 927 aio_queue_work(ctx); 928 } 929 930 /* 931 * kick_iocb: 932 * Called typically from a wait queue callback context 933 * to trigger a retry of the iocb. 934 * The retry is usually executed by aio workqueue 935 * threads (See aio_kick_handler). 936 */ 937 void kick_iocb(struct kiocb *iocb) 938 { 939 /* sync iocbs are easy: they can only ever be executing from a 940 * single context. */ 941 if (is_sync_kiocb(iocb)) { 942 kiocbSetKicked(iocb); 943 wake_up_process(iocb->ki_obj.tsk); 944 return; 945 } 946 947 try_queue_kicked_iocb(iocb); 948 } 949 EXPORT_SYMBOL(kick_iocb); 950 951 /* aio_complete 952 * Called when the io request on the given iocb is complete. 953 * Returns true if this is the last user of the request. The 954 * only other user of the request can be the cancellation code. 955 */ 956 int aio_complete(struct kiocb *iocb, long res, long res2) 957 { 958 struct kioctx *ctx = iocb->ki_ctx; 959 struct aio_ring_info *info; 960 struct aio_ring *ring; 961 struct io_event *event; 962 unsigned long flags; 963 unsigned long tail; 964 int ret; 965 966 /* 967 * Special case handling for sync iocbs: 968 * - events go directly into the iocb for fast handling 969 * - the sync task with the iocb in its stack holds the single iocb 970 * ref, no other paths have a way to get another ref 971 * - the sync task helpfully left a reference to itself in the iocb 972 */ 973 if (is_sync_kiocb(iocb)) { 974 BUG_ON(iocb->ki_users != 1); 975 iocb->ki_user_data = res; 976 iocb->ki_users = 0; 977 wake_up_process(iocb->ki_obj.tsk); 978 return 1; 979 } 980 981 info = &ctx->ring_info; 982 983 /* add a completion event to the ring buffer. 984 * must be done holding ctx->ctx_lock to prevent 985 * other code from messing with the tail 986 * pointer since we might be called from irq 987 * context. 988 */ 989 spin_lock_irqsave(&ctx->ctx_lock, flags); 990 991 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list)) 992 list_del_init(&iocb->ki_run_list); 993 994 /* 995 * cancelled requests don't get events, userland was given one 996 * when the event got cancelled. 997 */ 998 if (kiocbIsCancelled(iocb)) 999 goto put_rq; 1000 1001 ring = kmap_atomic(info->ring_pages[0]); 1002 1003 tail = info->tail; 1004 event = aio_ring_event(info, tail); 1005 if (++tail >= info->nr) 1006 tail = 0; 1007 1008 event->obj = (u64)(unsigned long)iocb->ki_obj.user; 1009 event->data = iocb->ki_user_data; 1010 event->res = res; 1011 event->res2 = res2; 1012 1013 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n", 1014 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data, 1015 res, res2); 1016 1017 /* after flagging the request as done, we 1018 * must never even look at it again 1019 */ 1020 smp_wmb(); /* make event visible before updating tail */ 1021 1022 info->tail = tail; 1023 ring->tail = tail; 1024 1025 put_aio_ring_event(event); 1026 kunmap_atomic(ring); 1027 1028 pr_debug("added to ring %p at [%lu]\n", iocb, tail); 1029 1030 /* 1031 * Check if the user asked us to deliver the result through an 1032 * eventfd. The eventfd_signal() function is safe to be called 1033 * from IRQ context. 1034 */ 1035 if (iocb->ki_eventfd != NULL) 1036 eventfd_signal(iocb->ki_eventfd, 1); 1037 1038 put_rq: 1039 /* everything turned out well, dispose of the aiocb. */ 1040 ret = __aio_put_req(ctx, iocb); 1041 1042 /* 1043 * We have to order our ring_info tail store above and test 1044 * of the wait list below outside the wait lock. This is 1045 * like in wake_up_bit() where clearing a bit has to be 1046 * ordered with the unlocked test. 1047 */ 1048 smp_mb(); 1049 1050 if (waitqueue_active(&ctx->wait)) 1051 wake_up(&ctx->wait); 1052 1053 spin_unlock_irqrestore(&ctx->ctx_lock, flags); 1054 return ret; 1055 } 1056 EXPORT_SYMBOL(aio_complete); 1057 1058 /* aio_read_evt 1059 * Pull an event off of the ioctx's event ring. Returns the number of 1060 * events fetched (0 or 1 ;-) 1061 * FIXME: make this use cmpxchg. 1062 * TODO: make the ringbuffer user mmap()able (requires FIXME). 1063 */ 1064 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent) 1065 { 1066 struct aio_ring_info *info = &ioctx->ring_info; 1067 struct aio_ring *ring; 1068 unsigned long head; 1069 int ret = 0; 1070 1071 ring = kmap_atomic(info->ring_pages[0]); 1072 dprintk("in aio_read_evt h%lu t%lu m%lu\n", 1073 (unsigned long)ring->head, (unsigned long)ring->tail, 1074 (unsigned long)ring->nr); 1075 1076 if (ring->head == ring->tail) 1077 goto out; 1078 1079 spin_lock(&info->ring_lock); 1080 1081 head = ring->head % info->nr; 1082 if (head != ring->tail) { 1083 struct io_event *evp = aio_ring_event(info, head); 1084 *ent = *evp; 1085 head = (head + 1) % info->nr; 1086 smp_mb(); /* finish reading the event before updatng the head */ 1087 ring->head = head; 1088 ret = 1; 1089 put_aio_ring_event(evp); 1090 } 1091 spin_unlock(&info->ring_lock); 1092 1093 out: 1094 kunmap_atomic(ring); 1095 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret, 1096 (unsigned long)ring->head, (unsigned long)ring->tail); 1097 return ret; 1098 } 1099 1100 struct aio_timeout { 1101 struct timer_list timer; 1102 int timed_out; 1103 struct task_struct *p; 1104 }; 1105 1106 static void timeout_func(unsigned long data) 1107 { 1108 struct aio_timeout *to = (struct aio_timeout *)data; 1109 1110 to->timed_out = 1; 1111 wake_up_process(to->p); 1112 } 1113 1114 static inline void init_timeout(struct aio_timeout *to) 1115 { 1116 setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to); 1117 to->timed_out = 0; 1118 to->p = current; 1119 } 1120 1121 static inline void set_timeout(long start_jiffies, struct aio_timeout *to, 1122 const struct timespec *ts) 1123 { 1124 to->timer.expires = start_jiffies + timespec_to_jiffies(ts); 1125 if (time_after(to->timer.expires, jiffies)) 1126 add_timer(&to->timer); 1127 else 1128 to->timed_out = 1; 1129 } 1130 1131 static inline void clear_timeout(struct aio_timeout *to) 1132 { 1133 del_singleshot_timer_sync(&to->timer); 1134 } 1135 1136 static int read_events(struct kioctx *ctx, 1137 long min_nr, long nr, 1138 struct io_event __user *event, 1139 struct timespec __user *timeout) 1140 { 1141 long start_jiffies = jiffies; 1142 struct task_struct *tsk = current; 1143 DECLARE_WAITQUEUE(wait, tsk); 1144 int ret; 1145 int i = 0; 1146 struct io_event ent; 1147 struct aio_timeout to; 1148 int retry = 0; 1149 1150 /* needed to zero any padding within an entry (there shouldn't be 1151 * any, but C is fun! 1152 */ 1153 memset(&ent, 0, sizeof(ent)); 1154 retry: 1155 ret = 0; 1156 while (likely(i < nr)) { 1157 ret = aio_read_evt(ctx, &ent); 1158 if (unlikely(ret <= 0)) 1159 break; 1160 1161 dprintk("read event: %Lx %Lx %Lx %Lx\n", 1162 ent.data, ent.obj, ent.res, ent.res2); 1163 1164 /* Could we split the check in two? */ 1165 ret = -EFAULT; 1166 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) { 1167 dprintk("aio: lost an event due to EFAULT.\n"); 1168 break; 1169 } 1170 ret = 0; 1171 1172 /* Good, event copied to userland, update counts. */ 1173 event ++; 1174 i ++; 1175 } 1176 1177 if (min_nr <= i) 1178 return i; 1179 if (ret) 1180 return ret; 1181 1182 /* End fast path */ 1183 1184 /* racey check, but it gets redone */ 1185 if (!retry && unlikely(!list_empty(&ctx->run_list))) { 1186 retry = 1; 1187 aio_run_all_iocbs(ctx); 1188 goto retry; 1189 } 1190 1191 init_timeout(&to); 1192 if (timeout) { 1193 struct timespec ts; 1194 ret = -EFAULT; 1195 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts)))) 1196 goto out; 1197 1198 set_timeout(start_jiffies, &to, &ts); 1199 } 1200 1201 while (likely(i < nr)) { 1202 add_wait_queue_exclusive(&ctx->wait, &wait); 1203 do { 1204 set_task_state(tsk, TASK_INTERRUPTIBLE); 1205 ret = aio_read_evt(ctx, &ent); 1206 if (ret) 1207 break; 1208 if (min_nr <= i) 1209 break; 1210 if (unlikely(ctx->dead)) { 1211 ret = -EINVAL; 1212 break; 1213 } 1214 if (to.timed_out) /* Only check after read evt */ 1215 break; 1216 /* Try to only show up in io wait if there are ops 1217 * in flight */ 1218 if (ctx->reqs_active) 1219 io_schedule(); 1220 else 1221 schedule(); 1222 if (signal_pending(tsk)) { 1223 ret = -EINTR; 1224 break; 1225 } 1226 /*ret = aio_read_evt(ctx, &ent);*/ 1227 } while (1) ; 1228 1229 set_task_state(tsk, TASK_RUNNING); 1230 remove_wait_queue(&ctx->wait, &wait); 1231 1232 if (unlikely(ret <= 0)) 1233 break; 1234 1235 ret = -EFAULT; 1236 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) { 1237 dprintk("aio: lost an event due to EFAULT.\n"); 1238 break; 1239 } 1240 1241 /* Good, event copied to userland, update counts. */ 1242 event ++; 1243 i ++; 1244 } 1245 1246 if (timeout) 1247 clear_timeout(&to); 1248 out: 1249 destroy_timer_on_stack(&to.timer); 1250 return i ? i : ret; 1251 } 1252 1253 /* Take an ioctx and remove it from the list of ioctx's. Protects 1254 * against races with itself via ->dead. 1255 */ 1256 static void io_destroy(struct kioctx *ioctx) 1257 { 1258 struct mm_struct *mm = current->mm; 1259 int was_dead; 1260 1261 /* delete the entry from the list is someone else hasn't already */ 1262 spin_lock(&mm->ioctx_lock); 1263 was_dead = ioctx->dead; 1264 ioctx->dead = 1; 1265 hlist_del_rcu(&ioctx->list); 1266 spin_unlock(&mm->ioctx_lock); 1267 1268 dprintk("aio_release(%p)\n", ioctx); 1269 if (likely(!was_dead)) 1270 put_ioctx(ioctx); /* twice for the list */ 1271 1272 aio_cancel_all(ioctx); 1273 wait_for_all_aios(ioctx); 1274 1275 /* 1276 * Wake up any waiters. The setting of ctx->dead must be seen 1277 * by other CPUs at this point. Right now, we rely on the 1278 * locking done by the above calls to ensure this consistency. 1279 */ 1280 wake_up_all(&ioctx->wait); 1281 put_ioctx(ioctx); /* once for the lookup */ 1282 } 1283 1284 /* sys_io_setup: 1285 * Create an aio_context capable of receiving at least nr_events. 1286 * ctxp must not point to an aio_context that already exists, and 1287 * must be initialized to 0 prior to the call. On successful 1288 * creation of the aio_context, *ctxp is filled in with the resulting 1289 * handle. May fail with -EINVAL if *ctxp is not initialized, 1290 * if the specified nr_events exceeds internal limits. May fail 1291 * with -EAGAIN if the specified nr_events exceeds the user's limit 1292 * of available events. May fail with -ENOMEM if insufficient kernel 1293 * resources are available. May fail with -EFAULT if an invalid 1294 * pointer is passed for ctxp. Will fail with -ENOSYS if not 1295 * implemented. 1296 */ 1297 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp) 1298 { 1299 struct kioctx *ioctx = NULL; 1300 unsigned long ctx; 1301 long ret; 1302 1303 ret = get_user(ctx, ctxp); 1304 if (unlikely(ret)) 1305 goto out; 1306 1307 ret = -EINVAL; 1308 if (unlikely(ctx || nr_events == 0)) { 1309 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n", 1310 ctx, nr_events); 1311 goto out; 1312 } 1313 1314 ioctx = ioctx_alloc(nr_events); 1315 ret = PTR_ERR(ioctx); 1316 if (!IS_ERR(ioctx)) { 1317 ret = put_user(ioctx->user_id, ctxp); 1318 if (!ret) { 1319 put_ioctx(ioctx); 1320 return 0; 1321 } 1322 io_destroy(ioctx); 1323 } 1324 1325 out: 1326 return ret; 1327 } 1328 1329 /* sys_io_destroy: 1330 * Destroy the aio_context specified. May cancel any outstanding 1331 * AIOs and block on completion. Will fail with -ENOSYS if not 1332 * implemented. May fail with -EINVAL if the context pointed to 1333 * is invalid. 1334 */ 1335 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx) 1336 { 1337 struct kioctx *ioctx = lookup_ioctx(ctx); 1338 if (likely(NULL != ioctx)) { 1339 io_destroy(ioctx); 1340 return 0; 1341 } 1342 pr_debug("EINVAL: io_destroy: invalid context id\n"); 1343 return -EINVAL; 1344 } 1345 1346 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret) 1347 { 1348 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg]; 1349 1350 BUG_ON(ret <= 0); 1351 1352 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) { 1353 ssize_t this = min((ssize_t)iov->iov_len, ret); 1354 iov->iov_base += this; 1355 iov->iov_len -= this; 1356 iocb->ki_left -= this; 1357 ret -= this; 1358 if (iov->iov_len == 0) { 1359 iocb->ki_cur_seg++; 1360 iov++; 1361 } 1362 } 1363 1364 /* the caller should not have done more io than what fit in 1365 * the remaining iovecs */ 1366 BUG_ON(ret > 0 && iocb->ki_left == 0); 1367 } 1368 1369 static ssize_t aio_rw_vect_retry(struct kiocb *iocb) 1370 { 1371 struct file *file = iocb->ki_filp; 1372 struct address_space *mapping = file->f_mapping; 1373 struct inode *inode = mapping->host; 1374 ssize_t (*rw_op)(struct kiocb *, const struct iovec *, 1375 unsigned long, loff_t); 1376 ssize_t ret = 0; 1377 unsigned short opcode; 1378 1379 if ((iocb->ki_opcode == IOCB_CMD_PREADV) || 1380 (iocb->ki_opcode == IOCB_CMD_PREAD)) { 1381 rw_op = file->f_op->aio_read; 1382 opcode = IOCB_CMD_PREADV; 1383 } else { 1384 rw_op = file->f_op->aio_write; 1385 opcode = IOCB_CMD_PWRITEV; 1386 } 1387 1388 /* This matches the pread()/pwrite() logic */ 1389 if (iocb->ki_pos < 0) 1390 return -EINVAL; 1391 1392 do { 1393 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg], 1394 iocb->ki_nr_segs - iocb->ki_cur_seg, 1395 iocb->ki_pos); 1396 if (ret > 0) 1397 aio_advance_iovec(iocb, ret); 1398 1399 /* retry all partial writes. retry partial reads as long as its a 1400 * regular file. */ 1401 } while (ret > 0 && iocb->ki_left > 0 && 1402 (opcode == IOCB_CMD_PWRITEV || 1403 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode)))); 1404 1405 /* This means we must have transferred all that we could */ 1406 /* No need to retry anymore */ 1407 if ((ret == 0) || (iocb->ki_left == 0)) 1408 ret = iocb->ki_nbytes - iocb->ki_left; 1409 1410 /* If we managed to write some out we return that, rather than 1411 * the eventual error. */ 1412 if (opcode == IOCB_CMD_PWRITEV 1413 && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY 1414 && iocb->ki_nbytes - iocb->ki_left) 1415 ret = iocb->ki_nbytes - iocb->ki_left; 1416 1417 return ret; 1418 } 1419 1420 static ssize_t aio_fdsync(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, 1); 1427 return ret; 1428 } 1429 1430 static ssize_t aio_fsync(struct kiocb *iocb) 1431 { 1432 struct file *file = iocb->ki_filp; 1433 ssize_t ret = -EINVAL; 1434 1435 if (file->f_op->aio_fsync) 1436 ret = file->f_op->aio_fsync(iocb, 0); 1437 return ret; 1438 } 1439 1440 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat) 1441 { 1442 ssize_t ret; 1443 1444 #ifdef CONFIG_COMPAT 1445 if (compat) 1446 ret = compat_rw_copy_check_uvector(type, 1447 (struct compat_iovec __user *)kiocb->ki_buf, 1448 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec, 1449 &kiocb->ki_iovec, 1); 1450 else 1451 #endif 1452 ret = rw_copy_check_uvector(type, 1453 (struct iovec __user *)kiocb->ki_buf, 1454 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec, 1455 &kiocb->ki_iovec, 1); 1456 if (ret < 0) 1457 goto out; 1458 1459 kiocb->ki_nr_segs = kiocb->ki_nbytes; 1460 kiocb->ki_cur_seg = 0; 1461 /* ki_nbytes/left now reflect bytes instead of segs */ 1462 kiocb->ki_nbytes = ret; 1463 kiocb->ki_left = ret; 1464 1465 ret = 0; 1466 out: 1467 return ret; 1468 } 1469 1470 static ssize_t aio_setup_single_vector(struct kiocb *kiocb) 1471 { 1472 kiocb->ki_iovec = &kiocb->ki_inline_vec; 1473 kiocb->ki_iovec->iov_base = kiocb->ki_buf; 1474 kiocb->ki_iovec->iov_len = kiocb->ki_left; 1475 kiocb->ki_nr_segs = 1; 1476 kiocb->ki_cur_seg = 0; 1477 return 0; 1478 } 1479 1480 /* 1481 * aio_setup_iocb: 1482 * Performs the initial checks and aio retry method 1483 * setup for the kiocb at the time of io submission. 1484 */ 1485 static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat) 1486 { 1487 struct file *file = kiocb->ki_filp; 1488 ssize_t ret = 0; 1489 1490 switch (kiocb->ki_opcode) { 1491 case IOCB_CMD_PREAD: 1492 ret = -EBADF; 1493 if (unlikely(!(file->f_mode & FMODE_READ))) 1494 break; 1495 ret = -EFAULT; 1496 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf, 1497 kiocb->ki_left))) 1498 break; 1499 ret = security_file_permission(file, MAY_READ); 1500 if (unlikely(ret)) 1501 break; 1502 ret = aio_setup_single_vector(kiocb); 1503 if (ret) 1504 break; 1505 ret = -EINVAL; 1506 if (file->f_op->aio_read) 1507 kiocb->ki_retry = aio_rw_vect_retry; 1508 break; 1509 case IOCB_CMD_PWRITE: 1510 ret = -EBADF; 1511 if (unlikely(!(file->f_mode & FMODE_WRITE))) 1512 break; 1513 ret = -EFAULT; 1514 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf, 1515 kiocb->ki_left))) 1516 break; 1517 ret = security_file_permission(file, MAY_WRITE); 1518 if (unlikely(ret)) 1519 break; 1520 ret = aio_setup_single_vector(kiocb); 1521 if (ret) 1522 break; 1523 ret = -EINVAL; 1524 if (file->f_op->aio_write) 1525 kiocb->ki_retry = aio_rw_vect_retry; 1526 break; 1527 case IOCB_CMD_PREADV: 1528 ret = -EBADF; 1529 if (unlikely(!(file->f_mode & FMODE_READ))) 1530 break; 1531 ret = security_file_permission(file, MAY_READ); 1532 if (unlikely(ret)) 1533 break; 1534 ret = aio_setup_vectored_rw(READ, kiocb, compat); 1535 if (ret) 1536 break; 1537 ret = -EINVAL; 1538 if (file->f_op->aio_read) 1539 kiocb->ki_retry = aio_rw_vect_retry; 1540 break; 1541 case IOCB_CMD_PWRITEV: 1542 ret = -EBADF; 1543 if (unlikely(!(file->f_mode & FMODE_WRITE))) 1544 break; 1545 ret = security_file_permission(file, MAY_WRITE); 1546 if (unlikely(ret)) 1547 break; 1548 ret = aio_setup_vectored_rw(WRITE, kiocb, compat); 1549 if (ret) 1550 break; 1551 ret = -EINVAL; 1552 if (file->f_op->aio_write) 1553 kiocb->ki_retry = aio_rw_vect_retry; 1554 break; 1555 case IOCB_CMD_FDSYNC: 1556 ret = -EINVAL; 1557 if (file->f_op->aio_fsync) 1558 kiocb->ki_retry = aio_fdsync; 1559 break; 1560 case IOCB_CMD_FSYNC: 1561 ret = -EINVAL; 1562 if (file->f_op->aio_fsync) 1563 kiocb->ki_retry = aio_fsync; 1564 break; 1565 default: 1566 dprintk("EINVAL: io_submit: no operation provided\n"); 1567 ret = -EINVAL; 1568 } 1569 1570 if (!kiocb->ki_retry) 1571 return ret; 1572 1573 return 0; 1574 } 1575 1576 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb, 1577 struct iocb *iocb, struct kiocb_batch *batch, 1578 bool compat) 1579 { 1580 struct kiocb *req; 1581 struct file *file; 1582 ssize_t ret; 1583 1584 /* enforce forwards compatibility on users */ 1585 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) { 1586 pr_debug("EINVAL: io_submit: reserve field set\n"); 1587 return -EINVAL; 1588 } 1589 1590 /* prevent overflows */ 1591 if (unlikely( 1592 (iocb->aio_buf != (unsigned long)iocb->aio_buf) || 1593 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) || 1594 ((ssize_t)iocb->aio_nbytes < 0) 1595 )) { 1596 pr_debug("EINVAL: io_submit: overflow check\n"); 1597 return -EINVAL; 1598 } 1599 1600 file = fget(iocb->aio_fildes); 1601 if (unlikely(!file)) 1602 return -EBADF; 1603 1604 req = aio_get_req(ctx, batch); /* returns with 2 references to req */ 1605 if (unlikely(!req)) { 1606 fput(file); 1607 return -EAGAIN; 1608 } 1609 req->ki_filp = file; 1610 if (iocb->aio_flags & IOCB_FLAG_RESFD) { 1611 /* 1612 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an 1613 * instance of the file* now. The file descriptor must be 1614 * an eventfd() fd, and will be signaled for each completed 1615 * event using the eventfd_signal() function. 1616 */ 1617 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd); 1618 if (IS_ERR(req->ki_eventfd)) { 1619 ret = PTR_ERR(req->ki_eventfd); 1620 req->ki_eventfd = NULL; 1621 goto out_put_req; 1622 } 1623 } 1624 1625 ret = put_user(req->ki_key, &user_iocb->aio_key); 1626 if (unlikely(ret)) { 1627 dprintk("EFAULT: aio_key\n"); 1628 goto out_put_req; 1629 } 1630 1631 req->ki_obj.user = user_iocb; 1632 req->ki_user_data = iocb->aio_data; 1633 req->ki_pos = iocb->aio_offset; 1634 1635 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf; 1636 req->ki_left = req->ki_nbytes = iocb->aio_nbytes; 1637 req->ki_opcode = iocb->aio_lio_opcode; 1638 1639 ret = aio_setup_iocb(req, compat); 1640 1641 if (ret) 1642 goto out_put_req; 1643 1644 spin_lock_irq(&ctx->ctx_lock); 1645 /* 1646 * We could have raced with io_destroy() and are currently holding a 1647 * reference to ctx which should be destroyed. We cannot submit IO 1648 * since ctx gets freed as soon as io_submit() puts its reference. The 1649 * check here is reliable: io_destroy() sets ctx->dead before waiting 1650 * for outstanding IO and the barrier between these two is realized by 1651 * unlock of mm->ioctx_lock and lock of ctx->ctx_lock. Analogously we 1652 * increment ctx->reqs_active before checking for ctx->dead and the 1653 * barrier is realized by unlock and lock of ctx->ctx_lock. Thus if we 1654 * don't see ctx->dead set here, io_destroy() waits for our IO to 1655 * finish. 1656 */ 1657 if (ctx->dead) { 1658 spin_unlock_irq(&ctx->ctx_lock); 1659 ret = -EINVAL; 1660 goto out_put_req; 1661 } 1662 aio_run_iocb(req); 1663 if (!list_empty(&ctx->run_list)) { 1664 /* drain the run list */ 1665 while (__aio_run_iocbs(ctx)) 1666 ; 1667 } 1668 spin_unlock_irq(&ctx->ctx_lock); 1669 1670 aio_put_req(req); /* drop extra ref to req */ 1671 return 0; 1672 1673 out_put_req: 1674 aio_put_req(req); /* drop extra ref to req */ 1675 aio_put_req(req); /* drop i/o ref to req */ 1676 return ret; 1677 } 1678 1679 long do_io_submit(aio_context_t ctx_id, long nr, 1680 struct iocb __user *__user *iocbpp, bool compat) 1681 { 1682 struct kioctx *ctx; 1683 long ret = 0; 1684 int i = 0; 1685 struct blk_plug plug; 1686 struct kiocb_batch batch; 1687 1688 if (unlikely(nr < 0)) 1689 return -EINVAL; 1690 1691 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp))) 1692 nr = LONG_MAX/sizeof(*iocbpp); 1693 1694 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp))))) 1695 return -EFAULT; 1696 1697 ctx = lookup_ioctx(ctx_id); 1698 if (unlikely(!ctx)) { 1699 pr_debug("EINVAL: io_submit: invalid context id\n"); 1700 return -EINVAL; 1701 } 1702 1703 kiocb_batch_init(&batch, nr); 1704 1705 blk_start_plug(&plug); 1706 1707 /* 1708 * AKPM: should this return a partial result if some of the IOs were 1709 * successfully submitted? 1710 */ 1711 for (i=0; i<nr; i++) { 1712 struct iocb __user *user_iocb; 1713 struct iocb tmp; 1714 1715 if (unlikely(__get_user(user_iocb, iocbpp + i))) { 1716 ret = -EFAULT; 1717 break; 1718 } 1719 1720 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) { 1721 ret = -EFAULT; 1722 break; 1723 } 1724 1725 ret = io_submit_one(ctx, user_iocb, &tmp, &batch, compat); 1726 if (ret) 1727 break; 1728 } 1729 blk_finish_plug(&plug); 1730 1731 kiocb_batch_free(ctx, &batch); 1732 put_ioctx(ctx); 1733 return i ? i : ret; 1734 } 1735 1736 /* sys_io_submit: 1737 * Queue the nr iocbs pointed to by iocbpp for processing. Returns 1738 * the number of iocbs queued. May return -EINVAL if the aio_context 1739 * specified by ctx_id is invalid, if nr is < 0, if the iocb at 1740 * *iocbpp[0] is not properly initialized, if the operation specified 1741 * is invalid for the file descriptor in the iocb. May fail with 1742 * -EFAULT if any of the data structures point to invalid data. May 1743 * fail with -EBADF if the file descriptor specified in the first 1744 * iocb is invalid. May fail with -EAGAIN if insufficient resources 1745 * are available to queue any iocbs. Will return 0 if nr is 0. Will 1746 * fail with -ENOSYS if not implemented. 1747 */ 1748 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr, 1749 struct iocb __user * __user *, iocbpp) 1750 { 1751 return do_io_submit(ctx_id, nr, iocbpp, 0); 1752 } 1753 1754 /* lookup_kiocb 1755 * Finds a given iocb for cancellation. 1756 */ 1757 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, 1758 u32 key) 1759 { 1760 struct list_head *pos; 1761 1762 assert_spin_locked(&ctx->ctx_lock); 1763 1764 /* TODO: use a hash or array, this sucks. */ 1765 list_for_each(pos, &ctx->active_reqs) { 1766 struct kiocb *kiocb = list_kiocb(pos); 1767 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key) 1768 return kiocb; 1769 } 1770 return NULL; 1771 } 1772 1773 /* sys_io_cancel: 1774 * Attempts to cancel an iocb previously passed to io_submit. If 1775 * the operation is successfully cancelled, the resulting event is 1776 * copied into the memory pointed to by result without being placed 1777 * into the completion queue and 0 is returned. May fail with 1778 * -EFAULT if any of the data structures pointed to are invalid. 1779 * May fail with -EINVAL if aio_context specified by ctx_id is 1780 * invalid. May fail with -EAGAIN if the iocb specified was not 1781 * cancelled. Will fail with -ENOSYS if not implemented. 1782 */ 1783 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb, 1784 struct io_event __user *, result) 1785 { 1786 int (*cancel)(struct kiocb *iocb, struct io_event *res); 1787 struct kioctx *ctx; 1788 struct kiocb *kiocb; 1789 u32 key; 1790 int ret; 1791 1792 ret = get_user(key, &iocb->aio_key); 1793 if (unlikely(ret)) 1794 return -EFAULT; 1795 1796 ctx = lookup_ioctx(ctx_id); 1797 if (unlikely(!ctx)) 1798 return -EINVAL; 1799 1800 spin_lock_irq(&ctx->ctx_lock); 1801 ret = -EAGAIN; 1802 kiocb = lookup_kiocb(ctx, iocb, key); 1803 if (kiocb && kiocb->ki_cancel) { 1804 cancel = kiocb->ki_cancel; 1805 kiocb->ki_users ++; 1806 kiocbSetCancelled(kiocb); 1807 } else 1808 cancel = NULL; 1809 spin_unlock_irq(&ctx->ctx_lock); 1810 1811 if (NULL != cancel) { 1812 struct io_event tmp; 1813 pr_debug("calling cancel\n"); 1814 memset(&tmp, 0, sizeof(tmp)); 1815 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user; 1816 tmp.data = kiocb->ki_user_data; 1817 ret = cancel(kiocb, &tmp); 1818 if (!ret) { 1819 /* Cancellation succeeded -- copy the result 1820 * into the user's buffer. 1821 */ 1822 if (copy_to_user(result, &tmp, sizeof(tmp))) 1823 ret = -EFAULT; 1824 } 1825 } else 1826 ret = -EINVAL; 1827 1828 put_ioctx(ctx); 1829 1830 return ret; 1831 } 1832 1833 /* io_getevents: 1834 * Attempts to read at least min_nr events and up to nr events from 1835 * the completion queue for the aio_context specified by ctx_id. If 1836 * it succeeds, the number of read events is returned. May fail with 1837 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is 1838 * out of range, if timeout is out of range. May fail with -EFAULT 1839 * if any of the memory specified is invalid. May return 0 or 1840 * < min_nr if the timeout specified by timeout has elapsed 1841 * before sufficient events are available, where timeout == NULL 1842 * specifies an infinite timeout. Note that the timeout pointed to by 1843 * timeout is relative and will be updated if not NULL and the 1844 * operation blocks. Will fail with -ENOSYS if not implemented. 1845 */ 1846 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id, 1847 long, min_nr, 1848 long, nr, 1849 struct io_event __user *, events, 1850 struct timespec __user *, timeout) 1851 { 1852 struct kioctx *ioctx = lookup_ioctx(ctx_id); 1853 long ret = -EINVAL; 1854 1855 if (likely(ioctx)) { 1856 if (likely(min_nr <= nr && min_nr >= 0)) 1857 ret = read_events(ioctx, min_nr, nr, events, timeout); 1858 put_ioctx(ioctx); 1859 } 1860 1861 asmlinkage_protect(5, ret, ctx_id, min_nr, nr, events, timeout); 1862 return ret; 1863 } 1864