1 /* 2 * An async IO implementation for Linux 3 * Written by Benjamin LaHaise <bcrl@kvack.org> 4 * 5 * Implements an efficient asynchronous io interface. 6 * 7 * Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved. 8 * 9 * See ../COPYING for licensing terms. 10 */ 11 #include <linux/kernel.h> 12 #include <linux/init.h> 13 #include <linux/errno.h> 14 #include <linux/time.h> 15 #include <linux/aio_abi.h> 16 #include <linux/module.h> 17 #include <linux/syscalls.h> 18 #include <linux/uio.h> 19 20 #define DEBUG 0 21 22 #include <linux/sched.h> 23 #include <linux/fs.h> 24 #include <linux/file.h> 25 #include <linux/mm.h> 26 #include <linux/mman.h> 27 #include <linux/slab.h> 28 #include <linux/timer.h> 29 #include <linux/aio.h> 30 #include <linux/highmem.h> 31 #include <linux/workqueue.h> 32 #include <linux/security.h> 33 #include <linux/eventfd.h> 34 35 #include <asm/kmap_types.h> 36 #include <asm/uaccess.h> 37 #include <asm/mmu_context.h> 38 39 #if DEBUG > 1 40 #define dprintk printk 41 #else 42 #define dprintk(x...) do { ; } while (0) 43 #endif 44 45 /*------ sysctl variables----*/ 46 static DEFINE_SPINLOCK(aio_nr_lock); 47 unsigned long aio_nr; /* current system wide number of aio requests */ 48 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */ 49 /*----end sysctl variables---*/ 50 51 static struct kmem_cache *kiocb_cachep; 52 static struct kmem_cache *kioctx_cachep; 53 54 static struct workqueue_struct *aio_wq; 55 56 /* Used for rare fput completion. */ 57 static void aio_fput_routine(struct work_struct *); 58 static DECLARE_WORK(fput_work, aio_fput_routine); 59 60 static DEFINE_SPINLOCK(fput_lock); 61 static LIST_HEAD(fput_head); 62 63 static void aio_kick_handler(struct work_struct *); 64 static void aio_queue_work(struct kioctx *); 65 66 /* aio_setup 67 * Creates the slab caches used by the aio routines, panic on 68 * failure as this is done early during the boot sequence. 69 */ 70 static int __init aio_setup(void) 71 { 72 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC); 73 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC); 74 75 aio_wq = create_workqueue("aio"); 76 77 pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page)); 78 79 return 0; 80 } 81 82 static void aio_free_ring(struct kioctx *ctx) 83 { 84 struct aio_ring_info *info = &ctx->ring_info; 85 long i; 86 87 for (i=0; i<info->nr_pages; i++) 88 put_page(info->ring_pages[i]); 89 90 if (info->mmap_size) { 91 down_write(&ctx->mm->mmap_sem); 92 do_munmap(ctx->mm, info->mmap_base, info->mmap_size); 93 up_write(&ctx->mm->mmap_sem); 94 } 95 96 if (info->ring_pages && info->ring_pages != info->internal_pages) 97 kfree(info->ring_pages); 98 info->ring_pages = NULL; 99 info->nr = 0; 100 } 101 102 static int aio_setup_ring(struct kioctx *ctx) 103 { 104 struct aio_ring *ring; 105 struct aio_ring_info *info = &ctx->ring_info; 106 unsigned nr_events = ctx->max_reqs; 107 unsigned long size; 108 int nr_pages; 109 110 /* Compensate for the ring buffer's head/tail overlap entry */ 111 nr_events += 2; /* 1 is required, 2 for good luck */ 112 113 size = sizeof(struct aio_ring); 114 size += sizeof(struct io_event) * nr_events; 115 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT; 116 117 if (nr_pages < 0) 118 return -EINVAL; 119 120 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event); 121 122 info->nr = 0; 123 info->ring_pages = info->internal_pages; 124 if (nr_pages > AIO_RING_PAGES) { 125 info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL); 126 if (!info->ring_pages) 127 return -ENOMEM; 128 } 129 130 info->mmap_size = nr_pages * PAGE_SIZE; 131 dprintk("attempting mmap of %lu bytes\n", info->mmap_size); 132 down_write(&ctx->mm->mmap_sem); 133 info->mmap_base = do_mmap(NULL, 0, info->mmap_size, 134 PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE, 135 0); 136 if (IS_ERR((void *)info->mmap_base)) { 137 up_write(&ctx->mm->mmap_sem); 138 info->mmap_size = 0; 139 aio_free_ring(ctx); 140 return -EAGAIN; 141 } 142 143 dprintk("mmap address: 0x%08lx\n", info->mmap_base); 144 info->nr_pages = get_user_pages(current, ctx->mm, 145 info->mmap_base, nr_pages, 146 1, 0, info->ring_pages, NULL); 147 up_write(&ctx->mm->mmap_sem); 148 149 if (unlikely(info->nr_pages != nr_pages)) { 150 aio_free_ring(ctx); 151 return -EAGAIN; 152 } 153 154 ctx->user_id = info->mmap_base; 155 156 info->nr = nr_events; /* trusted copy */ 157 158 ring = kmap_atomic(info->ring_pages[0], KM_USER0); 159 ring->nr = nr_events; /* user copy */ 160 ring->id = ctx->user_id; 161 ring->head = ring->tail = 0; 162 ring->magic = AIO_RING_MAGIC; 163 ring->compat_features = AIO_RING_COMPAT_FEATURES; 164 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES; 165 ring->header_length = sizeof(struct aio_ring); 166 kunmap_atomic(ring, KM_USER0); 167 168 return 0; 169 } 170 171 172 /* aio_ring_event: returns a pointer to the event at the given index from 173 * kmap_atomic(, km). Release the pointer with put_aio_ring_event(); 174 */ 175 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event)) 176 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event)) 177 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE) 178 179 #define aio_ring_event(info, nr, km) ({ \ 180 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \ 181 struct io_event *__event; \ 182 __event = kmap_atomic( \ 183 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \ 184 __event += pos % AIO_EVENTS_PER_PAGE; \ 185 __event; \ 186 }) 187 188 #define put_aio_ring_event(event, km) do { \ 189 struct io_event *__event = (event); \ 190 (void)__event; \ 191 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \ 192 } while(0) 193 194 /* ioctx_alloc 195 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed. 196 */ 197 static struct kioctx *ioctx_alloc(unsigned nr_events) 198 { 199 struct mm_struct *mm; 200 struct kioctx *ctx; 201 202 /* Prevent overflows */ 203 if ((nr_events > (0x10000000U / sizeof(struct io_event))) || 204 (nr_events > (0x10000000U / sizeof(struct kiocb)))) { 205 pr_debug("ENOMEM: nr_events too high\n"); 206 return ERR_PTR(-EINVAL); 207 } 208 209 if ((unsigned long)nr_events > aio_max_nr) 210 return ERR_PTR(-EAGAIN); 211 212 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL); 213 if (!ctx) 214 return ERR_PTR(-ENOMEM); 215 216 ctx->max_reqs = nr_events; 217 mm = ctx->mm = current->mm; 218 atomic_inc(&mm->mm_count); 219 220 atomic_set(&ctx->users, 1); 221 spin_lock_init(&ctx->ctx_lock); 222 spin_lock_init(&ctx->ring_info.ring_lock); 223 init_waitqueue_head(&ctx->wait); 224 225 INIT_LIST_HEAD(&ctx->active_reqs); 226 INIT_LIST_HEAD(&ctx->run_list); 227 INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler); 228 229 if (aio_setup_ring(ctx) < 0) 230 goto out_freectx; 231 232 /* limit the number of system wide aios */ 233 spin_lock(&aio_nr_lock); 234 if (aio_nr + ctx->max_reqs > aio_max_nr || 235 aio_nr + ctx->max_reqs < aio_nr) 236 ctx->max_reqs = 0; 237 else 238 aio_nr += ctx->max_reqs; 239 spin_unlock(&aio_nr_lock); 240 if (ctx->max_reqs == 0) 241 goto out_cleanup; 242 243 /* now link into global list. kludge. FIXME */ 244 write_lock(&mm->ioctx_list_lock); 245 ctx->next = mm->ioctx_list; 246 mm->ioctx_list = ctx; 247 write_unlock(&mm->ioctx_list_lock); 248 249 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n", 250 ctx, ctx->user_id, current->mm, ctx->ring_info.nr); 251 return ctx; 252 253 out_cleanup: 254 __put_ioctx(ctx); 255 return ERR_PTR(-EAGAIN); 256 257 out_freectx: 258 mmdrop(mm); 259 kmem_cache_free(kioctx_cachep, ctx); 260 ctx = ERR_PTR(-ENOMEM); 261 262 dprintk("aio: error allocating ioctx %p\n", ctx); 263 return ctx; 264 } 265 266 /* aio_cancel_all 267 * Cancels all outstanding aio requests on an aio context. Used 268 * when the processes owning a context have all exited to encourage 269 * the rapid destruction of the kioctx. 270 */ 271 static void aio_cancel_all(struct kioctx *ctx) 272 { 273 int (*cancel)(struct kiocb *, struct io_event *); 274 struct io_event res; 275 spin_lock_irq(&ctx->ctx_lock); 276 ctx->dead = 1; 277 while (!list_empty(&ctx->active_reqs)) { 278 struct list_head *pos = ctx->active_reqs.next; 279 struct kiocb *iocb = list_kiocb(pos); 280 list_del_init(&iocb->ki_list); 281 cancel = iocb->ki_cancel; 282 kiocbSetCancelled(iocb); 283 if (cancel) { 284 iocb->ki_users++; 285 spin_unlock_irq(&ctx->ctx_lock); 286 cancel(iocb, &res); 287 spin_lock_irq(&ctx->ctx_lock); 288 } 289 } 290 spin_unlock_irq(&ctx->ctx_lock); 291 } 292 293 static void wait_for_all_aios(struct kioctx *ctx) 294 { 295 struct task_struct *tsk = current; 296 DECLARE_WAITQUEUE(wait, tsk); 297 298 spin_lock_irq(&ctx->ctx_lock); 299 if (!ctx->reqs_active) 300 goto out; 301 302 add_wait_queue(&ctx->wait, &wait); 303 set_task_state(tsk, TASK_UNINTERRUPTIBLE); 304 while (ctx->reqs_active) { 305 spin_unlock_irq(&ctx->ctx_lock); 306 io_schedule(); 307 set_task_state(tsk, TASK_UNINTERRUPTIBLE); 308 spin_lock_irq(&ctx->ctx_lock); 309 } 310 __set_task_state(tsk, TASK_RUNNING); 311 remove_wait_queue(&ctx->wait, &wait); 312 313 out: 314 spin_unlock_irq(&ctx->ctx_lock); 315 } 316 317 /* wait_on_sync_kiocb: 318 * Waits on the given sync kiocb to complete. 319 */ 320 ssize_t wait_on_sync_kiocb(struct kiocb *iocb) 321 { 322 while (iocb->ki_users) { 323 set_current_state(TASK_UNINTERRUPTIBLE); 324 if (!iocb->ki_users) 325 break; 326 io_schedule(); 327 } 328 __set_current_state(TASK_RUNNING); 329 return iocb->ki_user_data; 330 } 331 332 /* exit_aio: called when the last user of mm goes away. At this point, 333 * there is no way for any new requests to be submited or any of the 334 * io_* syscalls to be called on the context. However, there may be 335 * outstanding requests which hold references to the context; as they 336 * go away, they will call put_ioctx and release any pinned memory 337 * associated with the request (held via struct page * references). 338 */ 339 void exit_aio(struct mm_struct *mm) 340 { 341 struct kioctx *ctx = mm->ioctx_list; 342 mm->ioctx_list = NULL; 343 while (ctx) { 344 struct kioctx *next = ctx->next; 345 ctx->next = NULL; 346 aio_cancel_all(ctx); 347 348 wait_for_all_aios(ctx); 349 /* 350 * Ensure we don't leave the ctx on the aio_wq 351 */ 352 cancel_work_sync(&ctx->wq.work); 353 354 if (1 != atomic_read(&ctx->users)) 355 printk(KERN_DEBUG 356 "exit_aio:ioctx still alive: %d %d %d\n", 357 atomic_read(&ctx->users), ctx->dead, 358 ctx->reqs_active); 359 put_ioctx(ctx); 360 ctx = next; 361 } 362 } 363 364 /* __put_ioctx 365 * Called when the last user of an aio context has gone away, 366 * and the struct needs to be freed. 367 */ 368 void __put_ioctx(struct kioctx *ctx) 369 { 370 unsigned nr_events = ctx->max_reqs; 371 372 BUG_ON(ctx->reqs_active); 373 374 cancel_delayed_work(&ctx->wq); 375 cancel_work_sync(&ctx->wq.work); 376 aio_free_ring(ctx); 377 mmdrop(ctx->mm); 378 ctx->mm = NULL; 379 pr_debug("__put_ioctx: freeing %p\n", ctx); 380 kmem_cache_free(kioctx_cachep, ctx); 381 382 if (nr_events) { 383 spin_lock(&aio_nr_lock); 384 BUG_ON(aio_nr - nr_events > aio_nr); 385 aio_nr -= nr_events; 386 spin_unlock(&aio_nr_lock); 387 } 388 } 389 390 /* aio_get_req 391 * Allocate a slot for an aio request. Increments the users count 392 * of the kioctx so that the kioctx stays around until all requests are 393 * complete. Returns NULL if no requests are free. 394 * 395 * Returns with kiocb->users set to 2. The io submit code path holds 396 * an extra reference while submitting the i/o. 397 * This prevents races between the aio code path referencing the 398 * req (after submitting it) and aio_complete() freeing the req. 399 */ 400 static struct kiocb *__aio_get_req(struct kioctx *ctx) 401 { 402 struct kiocb *req = NULL; 403 struct aio_ring *ring; 404 int okay = 0; 405 406 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL); 407 if (unlikely(!req)) 408 return NULL; 409 410 req->ki_flags = 0; 411 req->ki_users = 2; 412 req->ki_key = 0; 413 req->ki_ctx = ctx; 414 req->ki_cancel = NULL; 415 req->ki_retry = NULL; 416 req->ki_dtor = NULL; 417 req->private = NULL; 418 req->ki_iovec = NULL; 419 INIT_LIST_HEAD(&req->ki_run_list); 420 req->ki_eventfd = ERR_PTR(-EINVAL); 421 422 /* Check if the completion queue has enough free space to 423 * accept an event from this io. 424 */ 425 spin_lock_irq(&ctx->ctx_lock); 426 ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0); 427 if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) { 428 list_add(&req->ki_list, &ctx->active_reqs); 429 ctx->reqs_active++; 430 okay = 1; 431 } 432 kunmap_atomic(ring, KM_USER0); 433 spin_unlock_irq(&ctx->ctx_lock); 434 435 if (!okay) { 436 kmem_cache_free(kiocb_cachep, req); 437 req = NULL; 438 } 439 440 return req; 441 } 442 443 static inline struct kiocb *aio_get_req(struct kioctx *ctx) 444 { 445 struct kiocb *req; 446 /* Handle a potential starvation case -- should be exceedingly rare as 447 * requests will be stuck on fput_head only if the aio_fput_routine is 448 * delayed and the requests were the last user of the struct file. 449 */ 450 req = __aio_get_req(ctx); 451 if (unlikely(NULL == req)) { 452 aio_fput_routine(NULL); 453 req = __aio_get_req(ctx); 454 } 455 return req; 456 } 457 458 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req) 459 { 460 assert_spin_locked(&ctx->ctx_lock); 461 462 if (!IS_ERR(req->ki_eventfd)) 463 fput(req->ki_eventfd); 464 if (req->ki_dtor) 465 req->ki_dtor(req); 466 if (req->ki_iovec != &req->ki_inline_vec) 467 kfree(req->ki_iovec); 468 kmem_cache_free(kiocb_cachep, req); 469 ctx->reqs_active--; 470 471 if (unlikely(!ctx->reqs_active && ctx->dead)) 472 wake_up(&ctx->wait); 473 } 474 475 static void aio_fput_routine(struct work_struct *data) 476 { 477 spin_lock_irq(&fput_lock); 478 while (likely(!list_empty(&fput_head))) { 479 struct kiocb *req = list_kiocb(fput_head.next); 480 struct kioctx *ctx = req->ki_ctx; 481 482 list_del(&req->ki_list); 483 spin_unlock_irq(&fput_lock); 484 485 /* Complete the fput */ 486 __fput(req->ki_filp); 487 488 /* Link the iocb into the context's free list */ 489 spin_lock_irq(&ctx->ctx_lock); 490 really_put_req(ctx, req); 491 spin_unlock_irq(&ctx->ctx_lock); 492 493 put_ioctx(ctx); 494 spin_lock_irq(&fput_lock); 495 } 496 spin_unlock_irq(&fput_lock); 497 } 498 499 /* __aio_put_req 500 * Returns true if this put was the last user of the request. 501 */ 502 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req) 503 { 504 dprintk(KERN_DEBUG "aio_put(%p): f_count=%d\n", 505 req, atomic_read(&req->ki_filp->f_count)); 506 507 assert_spin_locked(&ctx->ctx_lock); 508 509 req->ki_users --; 510 BUG_ON(req->ki_users < 0); 511 if (likely(req->ki_users)) 512 return 0; 513 list_del(&req->ki_list); /* remove from active_reqs */ 514 req->ki_cancel = NULL; 515 req->ki_retry = NULL; 516 517 /* Must be done under the lock to serialise against cancellation. 518 * Call this aio_fput as it duplicates fput via the fput_work. 519 */ 520 if (unlikely(atomic_dec_and_test(&req->ki_filp->f_count))) { 521 get_ioctx(ctx); 522 spin_lock(&fput_lock); 523 list_add(&req->ki_list, &fput_head); 524 spin_unlock(&fput_lock); 525 queue_work(aio_wq, &fput_work); 526 } else 527 really_put_req(ctx, req); 528 return 1; 529 } 530 531 /* aio_put_req 532 * Returns true if this put was the last user of the kiocb, 533 * false if the request is still in use. 534 */ 535 int aio_put_req(struct kiocb *req) 536 { 537 struct kioctx *ctx = req->ki_ctx; 538 int ret; 539 spin_lock_irq(&ctx->ctx_lock); 540 ret = __aio_put_req(ctx, req); 541 spin_unlock_irq(&ctx->ctx_lock); 542 return ret; 543 } 544 545 /* Lookup an ioctx id. ioctx_list is lockless for reads. 546 * FIXME: this is O(n) and is only suitable for development. 547 */ 548 struct kioctx *lookup_ioctx(unsigned long ctx_id) 549 { 550 struct kioctx *ioctx; 551 struct mm_struct *mm; 552 553 mm = current->mm; 554 read_lock(&mm->ioctx_list_lock); 555 for (ioctx = mm->ioctx_list; ioctx; ioctx = ioctx->next) 556 if (likely(ioctx->user_id == ctx_id && !ioctx->dead)) { 557 get_ioctx(ioctx); 558 break; 559 } 560 read_unlock(&mm->ioctx_list_lock); 561 562 return ioctx; 563 } 564 565 /* 566 * use_mm 567 * Makes the calling kernel thread take on the specified 568 * mm context. 569 * Called by the retry thread execute retries within the 570 * iocb issuer's mm context, so that copy_from/to_user 571 * operations work seamlessly for aio. 572 * (Note: this routine is intended to be called only 573 * from a kernel thread context) 574 */ 575 static void use_mm(struct mm_struct *mm) 576 { 577 struct mm_struct *active_mm; 578 struct task_struct *tsk = current; 579 580 task_lock(tsk); 581 tsk->flags |= PF_BORROWED_MM; 582 active_mm = tsk->active_mm; 583 atomic_inc(&mm->mm_count); 584 tsk->mm = mm; 585 tsk->active_mm = mm; 586 /* 587 * Note that on UML this *requires* PF_BORROWED_MM to be set, otherwise 588 * it won't work. Update it accordingly if you change it here 589 */ 590 switch_mm(active_mm, mm, tsk); 591 task_unlock(tsk); 592 593 mmdrop(active_mm); 594 } 595 596 /* 597 * unuse_mm 598 * Reverses the effect of use_mm, i.e. releases the 599 * specified mm context which was earlier taken on 600 * by the calling kernel thread 601 * (Note: this routine is intended to be called only 602 * from a kernel thread context) 603 */ 604 static void unuse_mm(struct mm_struct *mm) 605 { 606 struct task_struct *tsk = current; 607 608 task_lock(tsk); 609 tsk->flags &= ~PF_BORROWED_MM; 610 tsk->mm = NULL; 611 /* active_mm is still 'mm' */ 612 enter_lazy_tlb(mm, tsk); 613 task_unlock(tsk); 614 } 615 616 /* 617 * Queue up a kiocb to be retried. Assumes that the kiocb 618 * has already been marked as kicked, and places it on 619 * the retry run list for the corresponding ioctx, if it 620 * isn't already queued. Returns 1 if it actually queued 621 * the kiocb (to tell the caller to activate the work 622 * queue to process it), or 0, if it found that it was 623 * already queued. 624 */ 625 static inline int __queue_kicked_iocb(struct kiocb *iocb) 626 { 627 struct kioctx *ctx = iocb->ki_ctx; 628 629 assert_spin_locked(&ctx->ctx_lock); 630 631 if (list_empty(&iocb->ki_run_list)) { 632 list_add_tail(&iocb->ki_run_list, 633 &ctx->run_list); 634 return 1; 635 } 636 return 0; 637 } 638 639 /* aio_run_iocb 640 * This is the core aio execution routine. It is 641 * invoked both for initial i/o submission and 642 * subsequent retries via the aio_kick_handler. 643 * Expects to be invoked with iocb->ki_ctx->lock 644 * already held. The lock is released and reacquired 645 * as needed during processing. 646 * 647 * Calls the iocb retry method (already setup for the 648 * iocb on initial submission) for operation specific 649 * handling, but takes care of most of common retry 650 * execution details for a given iocb. The retry method 651 * needs to be non-blocking as far as possible, to avoid 652 * holding up other iocbs waiting to be serviced by the 653 * retry kernel thread. 654 * 655 * The trickier parts in this code have to do with 656 * ensuring that only one retry instance is in progress 657 * for a given iocb at any time. Providing that guarantee 658 * simplifies the coding of individual aio operations as 659 * it avoids various potential races. 660 */ 661 static ssize_t aio_run_iocb(struct kiocb *iocb) 662 { 663 struct kioctx *ctx = iocb->ki_ctx; 664 ssize_t (*retry)(struct kiocb *); 665 ssize_t ret; 666 667 if (!(retry = iocb->ki_retry)) { 668 printk("aio_run_iocb: iocb->ki_retry = NULL\n"); 669 return 0; 670 } 671 672 /* 673 * We don't want the next retry iteration for this 674 * operation to start until this one has returned and 675 * updated the iocb state. However, wait_queue functions 676 * can trigger a kick_iocb from interrupt context in the 677 * meantime, indicating that data is available for the next 678 * iteration. We want to remember that and enable the 679 * next retry iteration _after_ we are through with 680 * this one. 681 * 682 * So, in order to be able to register a "kick", but 683 * prevent it from being queued now, we clear the kick 684 * flag, but make the kick code *think* that the iocb is 685 * still on the run list until we are actually done. 686 * When we are done with this iteration, we check if 687 * the iocb was kicked in the meantime and if so, queue 688 * it up afresh. 689 */ 690 691 kiocbClearKicked(iocb); 692 693 /* 694 * This is so that aio_complete knows it doesn't need to 695 * pull the iocb off the run list (We can't just call 696 * INIT_LIST_HEAD because we don't want a kick_iocb to 697 * queue this on the run list yet) 698 */ 699 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL; 700 spin_unlock_irq(&ctx->ctx_lock); 701 702 /* Quit retrying if the i/o has been cancelled */ 703 if (kiocbIsCancelled(iocb)) { 704 ret = -EINTR; 705 aio_complete(iocb, ret, 0); 706 /* must not access the iocb after this */ 707 goto out; 708 } 709 710 /* 711 * Now we are all set to call the retry method in async 712 * context. 713 */ 714 ret = retry(iocb); 715 716 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) { 717 BUG_ON(!list_empty(&iocb->ki_wait.task_list)); 718 aio_complete(iocb, ret, 0); 719 } 720 out: 721 spin_lock_irq(&ctx->ctx_lock); 722 723 if (-EIOCBRETRY == ret) { 724 /* 725 * OK, now that we are done with this iteration 726 * and know that there is more left to go, 727 * this is where we let go so that a subsequent 728 * "kick" can start the next iteration 729 */ 730 731 /* will make __queue_kicked_iocb succeed from here on */ 732 INIT_LIST_HEAD(&iocb->ki_run_list); 733 /* we must queue the next iteration ourselves, if it 734 * has already been kicked */ 735 if (kiocbIsKicked(iocb)) { 736 __queue_kicked_iocb(iocb); 737 738 /* 739 * __queue_kicked_iocb will always return 1 here, because 740 * iocb->ki_run_list is empty at this point so it should 741 * be safe to unconditionally queue the context into the 742 * work queue. 743 */ 744 aio_queue_work(ctx); 745 } 746 } 747 return ret; 748 } 749 750 /* 751 * __aio_run_iocbs: 752 * Process all pending retries queued on the ioctx 753 * run list. 754 * Assumes it is operating within the aio issuer's mm 755 * context. 756 */ 757 static int __aio_run_iocbs(struct kioctx *ctx) 758 { 759 struct kiocb *iocb; 760 struct list_head run_list; 761 762 assert_spin_locked(&ctx->ctx_lock); 763 764 list_replace_init(&ctx->run_list, &run_list); 765 while (!list_empty(&run_list)) { 766 iocb = list_entry(run_list.next, struct kiocb, 767 ki_run_list); 768 list_del(&iocb->ki_run_list); 769 /* 770 * Hold an extra reference while retrying i/o. 771 */ 772 iocb->ki_users++; /* grab extra reference */ 773 aio_run_iocb(iocb); 774 __aio_put_req(ctx, iocb); 775 } 776 if (!list_empty(&ctx->run_list)) 777 return 1; 778 return 0; 779 } 780 781 static void aio_queue_work(struct kioctx * ctx) 782 { 783 unsigned long timeout; 784 /* 785 * if someone is waiting, get the work started right 786 * away, otherwise, use a longer delay 787 */ 788 smp_mb(); 789 if (waitqueue_active(&ctx->wait)) 790 timeout = 1; 791 else 792 timeout = HZ/10; 793 queue_delayed_work(aio_wq, &ctx->wq, timeout); 794 } 795 796 797 /* 798 * aio_run_iocbs: 799 * Process all pending retries queued on the ioctx 800 * run list. 801 * Assumes it is operating within the aio issuer's mm 802 * context. 803 */ 804 static inline void aio_run_iocbs(struct kioctx *ctx) 805 { 806 int requeue; 807 808 spin_lock_irq(&ctx->ctx_lock); 809 810 requeue = __aio_run_iocbs(ctx); 811 spin_unlock_irq(&ctx->ctx_lock); 812 if (requeue) 813 aio_queue_work(ctx); 814 } 815 816 /* 817 * just like aio_run_iocbs, but keeps running them until 818 * the list stays empty 819 */ 820 static inline void aio_run_all_iocbs(struct kioctx *ctx) 821 { 822 spin_lock_irq(&ctx->ctx_lock); 823 while (__aio_run_iocbs(ctx)) 824 ; 825 spin_unlock_irq(&ctx->ctx_lock); 826 } 827 828 /* 829 * aio_kick_handler: 830 * Work queue handler triggered to process pending 831 * retries on an ioctx. Takes on the aio issuer's 832 * mm context before running the iocbs, so that 833 * copy_xxx_user operates on the issuer's address 834 * space. 835 * Run on aiod's context. 836 */ 837 static void aio_kick_handler(struct work_struct *work) 838 { 839 struct kioctx *ctx = container_of(work, struct kioctx, wq.work); 840 mm_segment_t oldfs = get_fs(); 841 struct mm_struct *mm; 842 int requeue; 843 844 set_fs(USER_DS); 845 use_mm(ctx->mm); 846 spin_lock_irq(&ctx->ctx_lock); 847 requeue =__aio_run_iocbs(ctx); 848 mm = ctx->mm; 849 spin_unlock_irq(&ctx->ctx_lock); 850 unuse_mm(mm); 851 set_fs(oldfs); 852 /* 853 * we're in a worker thread already, don't use queue_delayed_work, 854 */ 855 if (requeue) 856 queue_delayed_work(aio_wq, &ctx->wq, 0); 857 } 858 859 860 /* 861 * Called by kick_iocb to queue the kiocb for retry 862 * and if required activate the aio work queue to process 863 * it 864 */ 865 static void try_queue_kicked_iocb(struct kiocb *iocb) 866 { 867 struct kioctx *ctx = iocb->ki_ctx; 868 unsigned long flags; 869 int run = 0; 870 871 /* We're supposed to be the only path putting the iocb back on the run 872 * list. If we find that the iocb is *back* on a wait queue already 873 * than retry has happened before we could queue the iocb. This also 874 * means that the retry could have completed and freed our iocb, no 875 * good. */ 876 BUG_ON((!list_empty(&iocb->ki_wait.task_list))); 877 878 spin_lock_irqsave(&ctx->ctx_lock, flags); 879 /* set this inside the lock so that we can't race with aio_run_iocb() 880 * testing it and putting the iocb on the run list under the lock */ 881 if (!kiocbTryKick(iocb)) 882 run = __queue_kicked_iocb(iocb); 883 spin_unlock_irqrestore(&ctx->ctx_lock, flags); 884 if (run) 885 aio_queue_work(ctx); 886 } 887 888 /* 889 * kick_iocb: 890 * Called typically from a wait queue callback context 891 * (aio_wake_function) to trigger a retry of the iocb. 892 * The retry is usually executed by aio workqueue 893 * threads (See aio_kick_handler). 894 */ 895 void kick_iocb(struct kiocb *iocb) 896 { 897 /* sync iocbs are easy: they can only ever be executing from a 898 * single context. */ 899 if (is_sync_kiocb(iocb)) { 900 kiocbSetKicked(iocb); 901 wake_up_process(iocb->ki_obj.tsk); 902 return; 903 } 904 905 try_queue_kicked_iocb(iocb); 906 } 907 EXPORT_SYMBOL(kick_iocb); 908 909 /* aio_complete 910 * Called when the io request on the given iocb is complete. 911 * Returns true if this is the last user of the request. The 912 * only other user of the request can be the cancellation code. 913 */ 914 int aio_complete(struct kiocb *iocb, long res, long res2) 915 { 916 struct kioctx *ctx = iocb->ki_ctx; 917 struct aio_ring_info *info; 918 struct aio_ring *ring; 919 struct io_event *event; 920 unsigned long flags; 921 unsigned long tail; 922 int ret; 923 924 /* 925 * Special case handling for sync iocbs: 926 * - events go directly into the iocb for fast handling 927 * - the sync task with the iocb in its stack holds the single iocb 928 * ref, no other paths have a way to get another ref 929 * - the sync task helpfully left a reference to itself in the iocb 930 */ 931 if (is_sync_kiocb(iocb)) { 932 BUG_ON(iocb->ki_users != 1); 933 iocb->ki_user_data = res; 934 iocb->ki_users = 0; 935 wake_up_process(iocb->ki_obj.tsk); 936 return 1; 937 } 938 939 /* 940 * Check if the user asked us to deliver the result through an 941 * eventfd. The eventfd_signal() function is safe to be called 942 * from IRQ context. 943 */ 944 if (!IS_ERR(iocb->ki_eventfd)) 945 eventfd_signal(iocb->ki_eventfd, 1); 946 947 info = &ctx->ring_info; 948 949 /* add a completion event to the ring buffer. 950 * must be done holding ctx->ctx_lock to prevent 951 * other code from messing with the tail 952 * pointer since we might be called from irq 953 * context. 954 */ 955 spin_lock_irqsave(&ctx->ctx_lock, flags); 956 957 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list)) 958 list_del_init(&iocb->ki_run_list); 959 960 /* 961 * cancelled requests don't get events, userland was given one 962 * when the event got cancelled. 963 */ 964 if (kiocbIsCancelled(iocb)) 965 goto put_rq; 966 967 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1); 968 969 tail = info->tail; 970 event = aio_ring_event(info, tail, KM_IRQ0); 971 if (++tail >= info->nr) 972 tail = 0; 973 974 event->obj = (u64)(unsigned long)iocb->ki_obj.user; 975 event->data = iocb->ki_user_data; 976 event->res = res; 977 event->res2 = res2; 978 979 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n", 980 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data, 981 res, res2); 982 983 /* after flagging the request as done, we 984 * must never even look at it again 985 */ 986 smp_wmb(); /* make event visible before updating tail */ 987 988 info->tail = tail; 989 ring->tail = tail; 990 991 put_aio_ring_event(event, KM_IRQ0); 992 kunmap_atomic(ring, KM_IRQ1); 993 994 pr_debug("added to ring %p at [%lu]\n", iocb, tail); 995 put_rq: 996 /* everything turned out well, dispose of the aiocb. */ 997 ret = __aio_put_req(ctx, iocb); 998 999 if (waitqueue_active(&ctx->wait)) 1000 wake_up(&ctx->wait); 1001 1002 spin_unlock_irqrestore(&ctx->ctx_lock, flags); 1003 return ret; 1004 } 1005 1006 /* aio_read_evt 1007 * Pull an event off of the ioctx's event ring. Returns the number of 1008 * events fetched (0 or 1 ;-) 1009 * FIXME: make this use cmpxchg. 1010 * TODO: make the ringbuffer user mmap()able (requires FIXME). 1011 */ 1012 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent) 1013 { 1014 struct aio_ring_info *info = &ioctx->ring_info; 1015 struct aio_ring *ring; 1016 unsigned long head; 1017 int ret = 0; 1018 1019 ring = kmap_atomic(info->ring_pages[0], KM_USER0); 1020 dprintk("in aio_read_evt h%lu t%lu m%lu\n", 1021 (unsigned long)ring->head, (unsigned long)ring->tail, 1022 (unsigned long)ring->nr); 1023 1024 if (ring->head == ring->tail) 1025 goto out; 1026 1027 spin_lock(&info->ring_lock); 1028 1029 head = ring->head % info->nr; 1030 if (head != ring->tail) { 1031 struct io_event *evp = aio_ring_event(info, head, KM_USER1); 1032 *ent = *evp; 1033 head = (head + 1) % info->nr; 1034 smp_mb(); /* finish reading the event before updatng the head */ 1035 ring->head = head; 1036 ret = 1; 1037 put_aio_ring_event(evp, KM_USER1); 1038 } 1039 spin_unlock(&info->ring_lock); 1040 1041 out: 1042 kunmap_atomic(ring, KM_USER0); 1043 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret, 1044 (unsigned long)ring->head, (unsigned long)ring->tail); 1045 return ret; 1046 } 1047 1048 struct aio_timeout { 1049 struct timer_list timer; 1050 int timed_out; 1051 struct task_struct *p; 1052 }; 1053 1054 static void timeout_func(unsigned long data) 1055 { 1056 struct aio_timeout *to = (struct aio_timeout *)data; 1057 1058 to->timed_out = 1; 1059 wake_up_process(to->p); 1060 } 1061 1062 static inline void init_timeout(struct aio_timeout *to) 1063 { 1064 init_timer(&to->timer); 1065 to->timer.data = (unsigned long)to; 1066 to->timer.function = timeout_func; 1067 to->timed_out = 0; 1068 to->p = current; 1069 } 1070 1071 static inline void set_timeout(long start_jiffies, struct aio_timeout *to, 1072 const struct timespec *ts) 1073 { 1074 to->timer.expires = start_jiffies + timespec_to_jiffies(ts); 1075 if (time_after(to->timer.expires, jiffies)) 1076 add_timer(&to->timer); 1077 else 1078 to->timed_out = 1; 1079 } 1080 1081 static inline void clear_timeout(struct aio_timeout *to) 1082 { 1083 del_singleshot_timer_sync(&to->timer); 1084 } 1085 1086 static int read_events(struct kioctx *ctx, 1087 long min_nr, long nr, 1088 struct io_event __user *event, 1089 struct timespec __user *timeout) 1090 { 1091 long start_jiffies = jiffies; 1092 struct task_struct *tsk = current; 1093 DECLARE_WAITQUEUE(wait, tsk); 1094 int ret; 1095 int i = 0; 1096 struct io_event ent; 1097 struct aio_timeout to; 1098 int retry = 0; 1099 1100 /* needed to zero any padding within an entry (there shouldn't be 1101 * any, but C is fun! 1102 */ 1103 memset(&ent, 0, sizeof(ent)); 1104 retry: 1105 ret = 0; 1106 while (likely(i < nr)) { 1107 ret = aio_read_evt(ctx, &ent); 1108 if (unlikely(ret <= 0)) 1109 break; 1110 1111 dprintk("read event: %Lx %Lx %Lx %Lx\n", 1112 ent.data, ent.obj, ent.res, ent.res2); 1113 1114 /* Could we split the check in two? */ 1115 ret = -EFAULT; 1116 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) { 1117 dprintk("aio: lost an event due to EFAULT.\n"); 1118 break; 1119 } 1120 ret = 0; 1121 1122 /* Good, event copied to userland, update counts. */ 1123 event ++; 1124 i ++; 1125 } 1126 1127 if (min_nr <= i) 1128 return i; 1129 if (ret) 1130 return ret; 1131 1132 /* End fast path */ 1133 1134 /* racey check, but it gets redone */ 1135 if (!retry && unlikely(!list_empty(&ctx->run_list))) { 1136 retry = 1; 1137 aio_run_all_iocbs(ctx); 1138 goto retry; 1139 } 1140 1141 init_timeout(&to); 1142 if (timeout) { 1143 struct timespec ts; 1144 ret = -EFAULT; 1145 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts)))) 1146 goto out; 1147 1148 set_timeout(start_jiffies, &to, &ts); 1149 } 1150 1151 while (likely(i < nr)) { 1152 add_wait_queue_exclusive(&ctx->wait, &wait); 1153 do { 1154 set_task_state(tsk, TASK_INTERRUPTIBLE); 1155 ret = aio_read_evt(ctx, &ent); 1156 if (ret) 1157 break; 1158 if (min_nr <= i) 1159 break; 1160 ret = 0; 1161 if (to.timed_out) /* Only check after read evt */ 1162 break; 1163 /* Try to only show up in io wait if there are ops 1164 * in flight */ 1165 if (ctx->reqs_active) 1166 io_schedule(); 1167 else 1168 schedule(); 1169 if (signal_pending(tsk)) { 1170 ret = -EINTR; 1171 break; 1172 } 1173 /*ret = aio_read_evt(ctx, &ent);*/ 1174 } while (1) ; 1175 1176 set_task_state(tsk, TASK_RUNNING); 1177 remove_wait_queue(&ctx->wait, &wait); 1178 1179 if (unlikely(ret <= 0)) 1180 break; 1181 1182 ret = -EFAULT; 1183 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) { 1184 dprintk("aio: lost an event due to EFAULT.\n"); 1185 break; 1186 } 1187 1188 /* Good, event copied to userland, update counts. */ 1189 event ++; 1190 i ++; 1191 } 1192 1193 if (timeout) 1194 clear_timeout(&to); 1195 out: 1196 return i ? i : ret; 1197 } 1198 1199 /* Take an ioctx and remove it from the list of ioctx's. Protects 1200 * against races with itself via ->dead. 1201 */ 1202 static void io_destroy(struct kioctx *ioctx) 1203 { 1204 struct mm_struct *mm = current->mm; 1205 struct kioctx **tmp; 1206 int was_dead; 1207 1208 /* delete the entry from the list is someone else hasn't already */ 1209 write_lock(&mm->ioctx_list_lock); 1210 was_dead = ioctx->dead; 1211 ioctx->dead = 1; 1212 for (tmp = &mm->ioctx_list; *tmp && *tmp != ioctx; 1213 tmp = &(*tmp)->next) 1214 ; 1215 if (*tmp) 1216 *tmp = ioctx->next; 1217 write_unlock(&mm->ioctx_list_lock); 1218 1219 dprintk("aio_release(%p)\n", ioctx); 1220 if (likely(!was_dead)) 1221 put_ioctx(ioctx); /* twice for the list */ 1222 1223 aio_cancel_all(ioctx); 1224 wait_for_all_aios(ioctx); 1225 put_ioctx(ioctx); /* once for the lookup */ 1226 } 1227 1228 /* sys_io_setup: 1229 * Create an aio_context capable of receiving at least nr_events. 1230 * ctxp must not point to an aio_context that already exists, and 1231 * must be initialized to 0 prior to the call. On successful 1232 * creation of the aio_context, *ctxp is filled in with the resulting 1233 * handle. May fail with -EINVAL if *ctxp is not initialized, 1234 * if the specified nr_events exceeds internal limits. May fail 1235 * with -EAGAIN if the specified nr_events exceeds the user's limit 1236 * of available events. May fail with -ENOMEM if insufficient kernel 1237 * resources are available. May fail with -EFAULT if an invalid 1238 * pointer is passed for ctxp. Will fail with -ENOSYS if not 1239 * implemented. 1240 */ 1241 asmlinkage long sys_io_setup(unsigned nr_events, aio_context_t __user *ctxp) 1242 { 1243 struct kioctx *ioctx = NULL; 1244 unsigned long ctx; 1245 long ret; 1246 1247 ret = get_user(ctx, ctxp); 1248 if (unlikely(ret)) 1249 goto out; 1250 1251 ret = -EINVAL; 1252 if (unlikely(ctx || nr_events == 0)) { 1253 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n", 1254 ctx, nr_events); 1255 goto out; 1256 } 1257 1258 ioctx = ioctx_alloc(nr_events); 1259 ret = PTR_ERR(ioctx); 1260 if (!IS_ERR(ioctx)) { 1261 ret = put_user(ioctx->user_id, ctxp); 1262 if (!ret) 1263 return 0; 1264 1265 get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */ 1266 io_destroy(ioctx); 1267 } 1268 1269 out: 1270 return ret; 1271 } 1272 1273 /* sys_io_destroy: 1274 * Destroy the aio_context specified. May cancel any outstanding 1275 * AIOs and block on completion. Will fail with -ENOSYS if not 1276 * implemented. May fail with -EFAULT if the context pointed to 1277 * is invalid. 1278 */ 1279 asmlinkage long sys_io_destroy(aio_context_t ctx) 1280 { 1281 struct kioctx *ioctx = lookup_ioctx(ctx); 1282 if (likely(NULL != ioctx)) { 1283 io_destroy(ioctx); 1284 return 0; 1285 } 1286 pr_debug("EINVAL: io_destroy: invalid context id\n"); 1287 return -EINVAL; 1288 } 1289 1290 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret) 1291 { 1292 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg]; 1293 1294 BUG_ON(ret <= 0); 1295 1296 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) { 1297 ssize_t this = min((ssize_t)iov->iov_len, ret); 1298 iov->iov_base += this; 1299 iov->iov_len -= this; 1300 iocb->ki_left -= this; 1301 ret -= this; 1302 if (iov->iov_len == 0) { 1303 iocb->ki_cur_seg++; 1304 iov++; 1305 } 1306 } 1307 1308 /* the caller should not have done more io than what fit in 1309 * the remaining iovecs */ 1310 BUG_ON(ret > 0 && iocb->ki_left == 0); 1311 } 1312 1313 static ssize_t aio_rw_vect_retry(struct kiocb *iocb) 1314 { 1315 struct file *file = iocb->ki_filp; 1316 struct address_space *mapping = file->f_mapping; 1317 struct inode *inode = mapping->host; 1318 ssize_t (*rw_op)(struct kiocb *, const struct iovec *, 1319 unsigned long, loff_t); 1320 ssize_t ret = 0; 1321 unsigned short opcode; 1322 1323 if ((iocb->ki_opcode == IOCB_CMD_PREADV) || 1324 (iocb->ki_opcode == IOCB_CMD_PREAD)) { 1325 rw_op = file->f_op->aio_read; 1326 opcode = IOCB_CMD_PREADV; 1327 } else { 1328 rw_op = file->f_op->aio_write; 1329 opcode = IOCB_CMD_PWRITEV; 1330 } 1331 1332 /* This matches the pread()/pwrite() logic */ 1333 if (iocb->ki_pos < 0) 1334 return -EINVAL; 1335 1336 do { 1337 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg], 1338 iocb->ki_nr_segs - iocb->ki_cur_seg, 1339 iocb->ki_pos); 1340 if (ret > 0) 1341 aio_advance_iovec(iocb, ret); 1342 1343 /* retry all partial writes. retry partial reads as long as its a 1344 * regular file. */ 1345 } while (ret > 0 && iocb->ki_left > 0 && 1346 (opcode == IOCB_CMD_PWRITEV || 1347 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode)))); 1348 1349 /* This means we must have transferred all that we could */ 1350 /* No need to retry anymore */ 1351 if ((ret == 0) || (iocb->ki_left == 0)) 1352 ret = iocb->ki_nbytes - iocb->ki_left; 1353 1354 /* If we managed to write some out we return that, rather than 1355 * the eventual error. */ 1356 if (opcode == IOCB_CMD_PWRITEV 1357 && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY 1358 && iocb->ki_nbytes - iocb->ki_left) 1359 ret = iocb->ki_nbytes - iocb->ki_left; 1360 1361 return ret; 1362 } 1363 1364 static ssize_t aio_fdsync(struct kiocb *iocb) 1365 { 1366 struct file *file = iocb->ki_filp; 1367 ssize_t ret = -EINVAL; 1368 1369 if (file->f_op->aio_fsync) 1370 ret = file->f_op->aio_fsync(iocb, 1); 1371 return ret; 1372 } 1373 1374 static ssize_t aio_fsync(struct kiocb *iocb) 1375 { 1376 struct file *file = iocb->ki_filp; 1377 ssize_t ret = -EINVAL; 1378 1379 if (file->f_op->aio_fsync) 1380 ret = file->f_op->aio_fsync(iocb, 0); 1381 return ret; 1382 } 1383 1384 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb) 1385 { 1386 ssize_t ret; 1387 1388 ret = rw_copy_check_uvector(type, (struct iovec __user *)kiocb->ki_buf, 1389 kiocb->ki_nbytes, 1, 1390 &kiocb->ki_inline_vec, &kiocb->ki_iovec); 1391 if (ret < 0) 1392 goto out; 1393 1394 kiocb->ki_nr_segs = kiocb->ki_nbytes; 1395 kiocb->ki_cur_seg = 0; 1396 /* ki_nbytes/left now reflect bytes instead of segs */ 1397 kiocb->ki_nbytes = ret; 1398 kiocb->ki_left = ret; 1399 1400 ret = 0; 1401 out: 1402 return ret; 1403 } 1404 1405 static ssize_t aio_setup_single_vector(struct kiocb *kiocb) 1406 { 1407 kiocb->ki_iovec = &kiocb->ki_inline_vec; 1408 kiocb->ki_iovec->iov_base = kiocb->ki_buf; 1409 kiocb->ki_iovec->iov_len = kiocb->ki_left; 1410 kiocb->ki_nr_segs = 1; 1411 kiocb->ki_cur_seg = 0; 1412 return 0; 1413 } 1414 1415 /* 1416 * aio_setup_iocb: 1417 * Performs the initial checks and aio retry method 1418 * setup for the kiocb at the time of io submission. 1419 */ 1420 static ssize_t aio_setup_iocb(struct kiocb *kiocb) 1421 { 1422 struct file *file = kiocb->ki_filp; 1423 ssize_t ret = 0; 1424 1425 switch (kiocb->ki_opcode) { 1426 case IOCB_CMD_PREAD: 1427 ret = -EBADF; 1428 if (unlikely(!(file->f_mode & FMODE_READ))) 1429 break; 1430 ret = -EFAULT; 1431 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf, 1432 kiocb->ki_left))) 1433 break; 1434 ret = security_file_permission(file, MAY_READ); 1435 if (unlikely(ret)) 1436 break; 1437 ret = aio_setup_single_vector(kiocb); 1438 if (ret) 1439 break; 1440 ret = -EINVAL; 1441 if (file->f_op->aio_read) 1442 kiocb->ki_retry = aio_rw_vect_retry; 1443 break; 1444 case IOCB_CMD_PWRITE: 1445 ret = -EBADF; 1446 if (unlikely(!(file->f_mode & FMODE_WRITE))) 1447 break; 1448 ret = -EFAULT; 1449 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf, 1450 kiocb->ki_left))) 1451 break; 1452 ret = security_file_permission(file, MAY_WRITE); 1453 if (unlikely(ret)) 1454 break; 1455 ret = aio_setup_single_vector(kiocb); 1456 if (ret) 1457 break; 1458 ret = -EINVAL; 1459 if (file->f_op->aio_write) 1460 kiocb->ki_retry = aio_rw_vect_retry; 1461 break; 1462 case IOCB_CMD_PREADV: 1463 ret = -EBADF; 1464 if (unlikely(!(file->f_mode & FMODE_READ))) 1465 break; 1466 ret = security_file_permission(file, MAY_READ); 1467 if (unlikely(ret)) 1468 break; 1469 ret = aio_setup_vectored_rw(READ, kiocb); 1470 if (ret) 1471 break; 1472 ret = -EINVAL; 1473 if (file->f_op->aio_read) 1474 kiocb->ki_retry = aio_rw_vect_retry; 1475 break; 1476 case IOCB_CMD_PWRITEV: 1477 ret = -EBADF; 1478 if (unlikely(!(file->f_mode & FMODE_WRITE))) 1479 break; 1480 ret = security_file_permission(file, MAY_WRITE); 1481 if (unlikely(ret)) 1482 break; 1483 ret = aio_setup_vectored_rw(WRITE, kiocb); 1484 if (ret) 1485 break; 1486 ret = -EINVAL; 1487 if (file->f_op->aio_write) 1488 kiocb->ki_retry = aio_rw_vect_retry; 1489 break; 1490 case IOCB_CMD_FDSYNC: 1491 ret = -EINVAL; 1492 if (file->f_op->aio_fsync) 1493 kiocb->ki_retry = aio_fdsync; 1494 break; 1495 case IOCB_CMD_FSYNC: 1496 ret = -EINVAL; 1497 if (file->f_op->aio_fsync) 1498 kiocb->ki_retry = aio_fsync; 1499 break; 1500 default: 1501 dprintk("EINVAL: io_submit: no operation provided\n"); 1502 ret = -EINVAL; 1503 } 1504 1505 if (!kiocb->ki_retry) 1506 return ret; 1507 1508 return 0; 1509 } 1510 1511 /* 1512 * aio_wake_function: 1513 * wait queue callback function for aio notification, 1514 * Simply triggers a retry of the operation via kick_iocb. 1515 * 1516 * This callback is specified in the wait queue entry in 1517 * a kiocb. 1518 * 1519 * Note: 1520 * This routine is executed with the wait queue lock held. 1521 * Since kick_iocb acquires iocb->ctx->ctx_lock, it nests 1522 * the ioctx lock inside the wait queue lock. This is safe 1523 * because this callback isn't used for wait queues which 1524 * are nested inside ioctx lock (i.e. ctx->wait) 1525 */ 1526 static int aio_wake_function(wait_queue_t *wait, unsigned mode, 1527 int sync, void *key) 1528 { 1529 struct kiocb *iocb = container_of(wait, struct kiocb, ki_wait); 1530 1531 list_del_init(&wait->task_list); 1532 kick_iocb(iocb); 1533 return 1; 1534 } 1535 1536 int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb, 1537 struct iocb *iocb) 1538 { 1539 struct kiocb *req; 1540 struct file *file; 1541 ssize_t ret; 1542 1543 /* enforce forwards compatibility on users */ 1544 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) { 1545 pr_debug("EINVAL: io_submit: reserve field set\n"); 1546 return -EINVAL; 1547 } 1548 1549 /* prevent overflows */ 1550 if (unlikely( 1551 (iocb->aio_buf != (unsigned long)iocb->aio_buf) || 1552 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) || 1553 ((ssize_t)iocb->aio_nbytes < 0) 1554 )) { 1555 pr_debug("EINVAL: io_submit: overflow check\n"); 1556 return -EINVAL; 1557 } 1558 1559 file = fget(iocb->aio_fildes); 1560 if (unlikely(!file)) 1561 return -EBADF; 1562 1563 req = aio_get_req(ctx); /* returns with 2 references to req */ 1564 if (unlikely(!req)) { 1565 fput(file); 1566 return -EAGAIN; 1567 } 1568 req->ki_filp = file; 1569 if (iocb->aio_flags & IOCB_FLAG_RESFD) { 1570 /* 1571 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an 1572 * instance of the file* now. The file descriptor must be 1573 * an eventfd() fd, and will be signaled for each completed 1574 * event using the eventfd_signal() function. 1575 */ 1576 req->ki_eventfd = eventfd_fget((int) iocb->aio_resfd); 1577 if (unlikely(IS_ERR(req->ki_eventfd))) { 1578 ret = PTR_ERR(req->ki_eventfd); 1579 goto out_put_req; 1580 } 1581 } 1582 1583 ret = put_user(req->ki_key, &user_iocb->aio_key); 1584 if (unlikely(ret)) { 1585 dprintk("EFAULT: aio_key\n"); 1586 goto out_put_req; 1587 } 1588 1589 req->ki_obj.user = user_iocb; 1590 req->ki_user_data = iocb->aio_data; 1591 req->ki_pos = iocb->aio_offset; 1592 1593 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf; 1594 req->ki_left = req->ki_nbytes = iocb->aio_nbytes; 1595 req->ki_opcode = iocb->aio_lio_opcode; 1596 init_waitqueue_func_entry(&req->ki_wait, aio_wake_function); 1597 INIT_LIST_HEAD(&req->ki_wait.task_list); 1598 1599 ret = aio_setup_iocb(req); 1600 1601 if (ret) 1602 goto out_put_req; 1603 1604 spin_lock_irq(&ctx->ctx_lock); 1605 aio_run_iocb(req); 1606 if (!list_empty(&ctx->run_list)) { 1607 /* drain the run list */ 1608 while (__aio_run_iocbs(ctx)) 1609 ; 1610 } 1611 spin_unlock_irq(&ctx->ctx_lock); 1612 aio_put_req(req); /* drop extra ref to req */ 1613 return 0; 1614 1615 out_put_req: 1616 aio_put_req(req); /* drop extra ref to req */ 1617 aio_put_req(req); /* drop i/o ref to req */ 1618 return ret; 1619 } 1620 1621 /* sys_io_submit: 1622 * Queue the nr iocbs pointed to by iocbpp for processing. Returns 1623 * the number of iocbs queued. May return -EINVAL if the aio_context 1624 * specified by ctx_id is invalid, if nr is < 0, if the iocb at 1625 * *iocbpp[0] is not properly initialized, if the operation specified 1626 * is invalid for the file descriptor in the iocb. May fail with 1627 * -EFAULT if any of the data structures point to invalid data. May 1628 * fail with -EBADF if the file descriptor specified in the first 1629 * iocb is invalid. May fail with -EAGAIN if insufficient resources 1630 * are available to queue any iocbs. Will return 0 if nr is 0. Will 1631 * fail with -ENOSYS if not implemented. 1632 */ 1633 asmlinkage long sys_io_submit(aio_context_t ctx_id, long nr, 1634 struct iocb __user * __user *iocbpp) 1635 { 1636 struct kioctx *ctx; 1637 long ret = 0; 1638 int i; 1639 1640 if (unlikely(nr < 0)) 1641 return -EINVAL; 1642 1643 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp))))) 1644 return -EFAULT; 1645 1646 ctx = lookup_ioctx(ctx_id); 1647 if (unlikely(!ctx)) { 1648 pr_debug("EINVAL: io_submit: invalid context id\n"); 1649 return -EINVAL; 1650 } 1651 1652 /* 1653 * AKPM: should this return a partial result if some of the IOs were 1654 * successfully submitted? 1655 */ 1656 for (i=0; i<nr; i++) { 1657 struct iocb __user *user_iocb; 1658 struct iocb tmp; 1659 1660 if (unlikely(__get_user(user_iocb, iocbpp + i))) { 1661 ret = -EFAULT; 1662 break; 1663 } 1664 1665 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) { 1666 ret = -EFAULT; 1667 break; 1668 } 1669 1670 ret = io_submit_one(ctx, user_iocb, &tmp); 1671 if (ret) 1672 break; 1673 } 1674 1675 put_ioctx(ctx); 1676 return i ? i : ret; 1677 } 1678 1679 /* lookup_kiocb 1680 * Finds a given iocb for cancellation. 1681 */ 1682 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, 1683 u32 key) 1684 { 1685 struct list_head *pos; 1686 1687 assert_spin_locked(&ctx->ctx_lock); 1688 1689 /* TODO: use a hash or array, this sucks. */ 1690 list_for_each(pos, &ctx->active_reqs) { 1691 struct kiocb *kiocb = list_kiocb(pos); 1692 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key) 1693 return kiocb; 1694 } 1695 return NULL; 1696 } 1697 1698 /* sys_io_cancel: 1699 * Attempts to cancel an iocb previously passed to io_submit. If 1700 * the operation is successfully cancelled, the resulting event is 1701 * copied into the memory pointed to by result without being placed 1702 * into the completion queue and 0 is returned. May fail with 1703 * -EFAULT if any of the data structures pointed to are invalid. 1704 * May fail with -EINVAL if aio_context specified by ctx_id is 1705 * invalid. May fail with -EAGAIN if the iocb specified was not 1706 * cancelled. Will fail with -ENOSYS if not implemented. 1707 */ 1708 asmlinkage long sys_io_cancel(aio_context_t ctx_id, struct iocb __user *iocb, 1709 struct io_event __user *result) 1710 { 1711 int (*cancel)(struct kiocb *iocb, struct io_event *res); 1712 struct kioctx *ctx; 1713 struct kiocb *kiocb; 1714 u32 key; 1715 int ret; 1716 1717 ret = get_user(key, &iocb->aio_key); 1718 if (unlikely(ret)) 1719 return -EFAULT; 1720 1721 ctx = lookup_ioctx(ctx_id); 1722 if (unlikely(!ctx)) 1723 return -EINVAL; 1724 1725 spin_lock_irq(&ctx->ctx_lock); 1726 ret = -EAGAIN; 1727 kiocb = lookup_kiocb(ctx, iocb, key); 1728 if (kiocb && kiocb->ki_cancel) { 1729 cancel = kiocb->ki_cancel; 1730 kiocb->ki_users ++; 1731 kiocbSetCancelled(kiocb); 1732 } else 1733 cancel = NULL; 1734 spin_unlock_irq(&ctx->ctx_lock); 1735 1736 if (NULL != cancel) { 1737 struct io_event tmp; 1738 pr_debug("calling cancel\n"); 1739 memset(&tmp, 0, sizeof(tmp)); 1740 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user; 1741 tmp.data = kiocb->ki_user_data; 1742 ret = cancel(kiocb, &tmp); 1743 if (!ret) { 1744 /* Cancellation succeeded -- copy the result 1745 * into the user's buffer. 1746 */ 1747 if (copy_to_user(result, &tmp, sizeof(tmp))) 1748 ret = -EFAULT; 1749 } 1750 } else 1751 ret = -EINVAL; 1752 1753 put_ioctx(ctx); 1754 1755 return ret; 1756 } 1757 1758 /* io_getevents: 1759 * Attempts to read at least min_nr events and up to nr events from 1760 * the completion queue for the aio_context specified by ctx_id. May 1761 * fail with -EINVAL if ctx_id is invalid, if min_nr is out of range, 1762 * if nr is out of range, if when is out of range. May fail with 1763 * -EFAULT if any of the memory specified to is invalid. May return 1764 * 0 or < min_nr if no events are available and the timeout specified 1765 * by when has elapsed, where when == NULL specifies an infinite 1766 * timeout. Note that the timeout pointed to by when is relative and 1767 * will be updated if not NULL and the operation blocks. Will fail 1768 * with -ENOSYS if not implemented. 1769 */ 1770 asmlinkage long sys_io_getevents(aio_context_t ctx_id, 1771 long min_nr, 1772 long nr, 1773 struct io_event __user *events, 1774 struct timespec __user *timeout) 1775 { 1776 struct kioctx *ioctx = lookup_ioctx(ctx_id); 1777 long ret = -EINVAL; 1778 1779 if (likely(ioctx)) { 1780 if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0)) 1781 ret = read_events(ioctx, min_nr, nr, events, timeout); 1782 put_ioctx(ioctx); 1783 } 1784 1785 return ret; 1786 } 1787 1788 __initcall(aio_setup); 1789 1790 EXPORT_SYMBOL(aio_complete); 1791 EXPORT_SYMBOL(aio_put_req); 1792 EXPORT_SYMBOL(wait_on_sync_kiocb); 1793