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 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 fastcall 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 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 fastcall 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 fastcall __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 *FASTCALL(__aio_get_req(struct kioctx *ctx)); 401 static struct kiocb fastcall *__aio_get_req(struct kioctx *ctx) 402 { 403 struct kiocb *req = NULL; 404 struct aio_ring *ring; 405 int okay = 0; 406 407 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL); 408 if (unlikely(!req)) 409 return NULL; 410 411 req->ki_flags = 0; 412 req->ki_users = 2; 413 req->ki_key = 0; 414 req->ki_ctx = ctx; 415 req->ki_cancel = NULL; 416 req->ki_retry = NULL; 417 req->ki_dtor = NULL; 418 req->private = NULL; 419 req->ki_iovec = NULL; 420 INIT_LIST_HEAD(&req->ki_run_list); 421 req->ki_eventfd = ERR_PTR(-EINVAL); 422 423 /* Check if the completion queue has enough free space to 424 * accept an event from this io. 425 */ 426 spin_lock_irq(&ctx->ctx_lock); 427 ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0); 428 if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) { 429 list_add(&req->ki_list, &ctx->active_reqs); 430 ctx->reqs_active++; 431 okay = 1; 432 } 433 kunmap_atomic(ring, KM_USER0); 434 spin_unlock_irq(&ctx->ctx_lock); 435 436 if (!okay) { 437 kmem_cache_free(kiocb_cachep, req); 438 req = NULL; 439 } 440 441 return req; 442 } 443 444 static inline struct kiocb *aio_get_req(struct kioctx *ctx) 445 { 446 struct kiocb *req; 447 /* Handle a potential starvation case -- should be exceedingly rare as 448 * requests will be stuck on fput_head only if the aio_fput_routine is 449 * delayed and the requests were the last user of the struct file. 450 */ 451 req = __aio_get_req(ctx); 452 if (unlikely(NULL == req)) { 453 aio_fput_routine(NULL); 454 req = __aio_get_req(ctx); 455 } 456 return req; 457 } 458 459 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req) 460 { 461 assert_spin_locked(&ctx->ctx_lock); 462 463 if (!IS_ERR(req->ki_eventfd)) 464 fput(req->ki_eventfd); 465 if (req->ki_dtor) 466 req->ki_dtor(req); 467 if (req->ki_iovec != &req->ki_inline_vec) 468 kfree(req->ki_iovec); 469 kmem_cache_free(kiocb_cachep, req); 470 ctx->reqs_active--; 471 472 if (unlikely(!ctx->reqs_active && ctx->dead)) 473 wake_up(&ctx->wait); 474 } 475 476 static void aio_fput_routine(struct work_struct *data) 477 { 478 spin_lock_irq(&fput_lock); 479 while (likely(!list_empty(&fput_head))) { 480 struct kiocb *req = list_kiocb(fput_head.next); 481 struct kioctx *ctx = req->ki_ctx; 482 483 list_del(&req->ki_list); 484 spin_unlock_irq(&fput_lock); 485 486 /* Complete the fput */ 487 __fput(req->ki_filp); 488 489 /* Link the iocb into the context's free list */ 490 spin_lock_irq(&ctx->ctx_lock); 491 really_put_req(ctx, req); 492 spin_unlock_irq(&ctx->ctx_lock); 493 494 put_ioctx(ctx); 495 spin_lock_irq(&fput_lock); 496 } 497 spin_unlock_irq(&fput_lock); 498 } 499 500 /* __aio_put_req 501 * Returns true if this put was the last user of the request. 502 */ 503 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req) 504 { 505 dprintk(KERN_DEBUG "aio_put(%p): f_count=%d\n", 506 req, atomic_read(&req->ki_filp->f_count)); 507 508 assert_spin_locked(&ctx->ctx_lock); 509 510 req->ki_users --; 511 BUG_ON(req->ki_users < 0); 512 if (likely(req->ki_users)) 513 return 0; 514 list_del(&req->ki_list); /* remove from active_reqs */ 515 req->ki_cancel = NULL; 516 req->ki_retry = NULL; 517 518 /* Must be done under the lock to serialise against cancellation. 519 * Call this aio_fput as it duplicates fput via the fput_work. 520 */ 521 if (unlikely(atomic_dec_and_test(&req->ki_filp->f_count))) { 522 get_ioctx(ctx); 523 spin_lock(&fput_lock); 524 list_add(&req->ki_list, &fput_head); 525 spin_unlock(&fput_lock); 526 queue_work(aio_wq, &fput_work); 527 } else 528 really_put_req(ctx, req); 529 return 1; 530 } 531 532 /* aio_put_req 533 * Returns true if this put was the last user of the kiocb, 534 * false if the request is still in use. 535 */ 536 int fastcall aio_put_req(struct kiocb *req) 537 { 538 struct kioctx *ctx = req->ki_ctx; 539 int ret; 540 spin_lock_irq(&ctx->ctx_lock); 541 ret = __aio_put_req(ctx, req); 542 spin_unlock_irq(&ctx->ctx_lock); 543 return ret; 544 } 545 546 /* Lookup an ioctx id. ioctx_list is lockless for reads. 547 * FIXME: this is O(n) and is only suitable for development. 548 */ 549 struct kioctx *lookup_ioctx(unsigned long ctx_id) 550 { 551 struct kioctx *ioctx; 552 struct mm_struct *mm; 553 554 mm = current->mm; 555 read_lock(&mm->ioctx_list_lock); 556 for (ioctx = mm->ioctx_list; ioctx; ioctx = ioctx->next) 557 if (likely(ioctx->user_id == ctx_id && !ioctx->dead)) { 558 get_ioctx(ioctx); 559 break; 560 } 561 read_unlock(&mm->ioctx_list_lock); 562 563 return ioctx; 564 } 565 566 /* 567 * use_mm 568 * Makes the calling kernel thread take on the specified 569 * mm context. 570 * Called by the retry thread execute retries within the 571 * iocb issuer's mm context, so that copy_from/to_user 572 * operations work seamlessly for aio. 573 * (Note: this routine is intended to be called only 574 * from a kernel thread context) 575 */ 576 static void use_mm(struct mm_struct *mm) 577 { 578 struct mm_struct *active_mm; 579 struct task_struct *tsk = current; 580 581 task_lock(tsk); 582 tsk->flags |= PF_BORROWED_MM; 583 active_mm = tsk->active_mm; 584 atomic_inc(&mm->mm_count); 585 tsk->mm = mm; 586 tsk->active_mm = mm; 587 /* 588 * Note that on UML this *requires* PF_BORROWED_MM to be set, otherwise 589 * it won't work. Update it accordingly if you change it here 590 */ 591 switch_mm(active_mm, mm, tsk); 592 task_unlock(tsk); 593 594 mmdrop(active_mm); 595 } 596 597 /* 598 * unuse_mm 599 * Reverses the effect of use_mm, i.e. releases the 600 * specified mm context which was earlier taken on 601 * by the calling kernel thread 602 * (Note: this routine is intended to be called only 603 * from a kernel thread context) 604 */ 605 static void unuse_mm(struct mm_struct *mm) 606 { 607 struct task_struct *tsk = current; 608 609 task_lock(tsk); 610 tsk->flags &= ~PF_BORROWED_MM; 611 tsk->mm = NULL; 612 /* active_mm is still 'mm' */ 613 enter_lazy_tlb(mm, tsk); 614 task_unlock(tsk); 615 } 616 617 /* 618 * Queue up a kiocb to be retried. Assumes that the kiocb 619 * has already been marked as kicked, and places it on 620 * the retry run list for the corresponding ioctx, if it 621 * isn't already queued. Returns 1 if it actually queued 622 * the kiocb (to tell the caller to activate the work 623 * queue to process it), or 0, if it found that it was 624 * already queued. 625 */ 626 static inline int __queue_kicked_iocb(struct kiocb *iocb) 627 { 628 struct kioctx *ctx = iocb->ki_ctx; 629 630 assert_spin_locked(&ctx->ctx_lock); 631 632 if (list_empty(&iocb->ki_run_list)) { 633 list_add_tail(&iocb->ki_run_list, 634 &ctx->run_list); 635 return 1; 636 } 637 return 0; 638 } 639 640 /* aio_run_iocb 641 * This is the core aio execution routine. It is 642 * invoked both for initial i/o submission and 643 * subsequent retries via the aio_kick_handler. 644 * Expects to be invoked with iocb->ki_ctx->lock 645 * already held. The lock is released and reacquired 646 * as needed during processing. 647 * 648 * Calls the iocb retry method (already setup for the 649 * iocb on initial submission) for operation specific 650 * handling, but takes care of most of common retry 651 * execution details for a given iocb. The retry method 652 * needs to be non-blocking as far as possible, to avoid 653 * holding up other iocbs waiting to be serviced by the 654 * retry kernel thread. 655 * 656 * The trickier parts in this code have to do with 657 * ensuring that only one retry instance is in progress 658 * for a given iocb at any time. Providing that guarantee 659 * simplifies the coding of individual aio operations as 660 * it avoids various potential races. 661 */ 662 static ssize_t aio_run_iocb(struct kiocb *iocb) 663 { 664 struct kioctx *ctx = iocb->ki_ctx; 665 ssize_t (*retry)(struct kiocb *); 666 ssize_t ret; 667 668 if (!(retry = iocb->ki_retry)) { 669 printk("aio_run_iocb: iocb->ki_retry = NULL\n"); 670 return 0; 671 } 672 673 /* 674 * We don't want the next retry iteration for this 675 * operation to start until this one has returned and 676 * updated the iocb state. However, wait_queue functions 677 * can trigger a kick_iocb from interrupt context in the 678 * meantime, indicating that data is available for the next 679 * iteration. We want to remember that and enable the 680 * next retry iteration _after_ we are through with 681 * this one. 682 * 683 * So, in order to be able to register a "kick", but 684 * prevent it from being queued now, we clear the kick 685 * flag, but make the kick code *think* that the iocb is 686 * still on the run list until we are actually done. 687 * When we are done with this iteration, we check if 688 * the iocb was kicked in the meantime and if so, queue 689 * it up afresh. 690 */ 691 692 kiocbClearKicked(iocb); 693 694 /* 695 * This is so that aio_complete knows it doesn't need to 696 * pull the iocb off the run list (We can't just call 697 * INIT_LIST_HEAD because we don't want a kick_iocb to 698 * queue this on the run list yet) 699 */ 700 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL; 701 spin_unlock_irq(&ctx->ctx_lock); 702 703 /* Quit retrying if the i/o has been cancelled */ 704 if (kiocbIsCancelled(iocb)) { 705 ret = -EINTR; 706 aio_complete(iocb, ret, 0); 707 /* must not access the iocb after this */ 708 goto out; 709 } 710 711 /* 712 * Now we are all set to call the retry method in async 713 * context. By setting this thread's io_wait context 714 * to point to the wait queue entry inside the currently 715 * running iocb for the duration of the retry, we ensure 716 * that async notification wakeups are queued by the 717 * operation instead of blocking waits, and when notified, 718 * cause the iocb to be kicked for continuation (through 719 * the aio_wake_function callback). 720 */ 721 BUG_ON(current->io_wait != NULL); 722 current->io_wait = &iocb->ki_wait; 723 ret = retry(iocb); 724 current->io_wait = NULL; 725 726 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) { 727 BUG_ON(!list_empty(&iocb->ki_wait.task_list)); 728 aio_complete(iocb, ret, 0); 729 } 730 out: 731 spin_lock_irq(&ctx->ctx_lock); 732 733 if (-EIOCBRETRY == ret) { 734 /* 735 * OK, now that we are done with this iteration 736 * and know that there is more left to go, 737 * this is where we let go so that a subsequent 738 * "kick" can start the next iteration 739 */ 740 741 /* will make __queue_kicked_iocb succeed from here on */ 742 INIT_LIST_HEAD(&iocb->ki_run_list); 743 /* we must queue the next iteration ourselves, if it 744 * has already been kicked */ 745 if (kiocbIsKicked(iocb)) { 746 __queue_kicked_iocb(iocb); 747 748 /* 749 * __queue_kicked_iocb will always return 1 here, because 750 * iocb->ki_run_list is empty at this point so it should 751 * be safe to unconditionally queue the context into the 752 * work queue. 753 */ 754 aio_queue_work(ctx); 755 } 756 } 757 return ret; 758 } 759 760 /* 761 * __aio_run_iocbs: 762 * Process all pending retries queued on the ioctx 763 * run list. 764 * Assumes it is operating within the aio issuer's mm 765 * context. 766 */ 767 static int __aio_run_iocbs(struct kioctx *ctx) 768 { 769 struct kiocb *iocb; 770 struct list_head run_list; 771 772 assert_spin_locked(&ctx->ctx_lock); 773 774 list_replace_init(&ctx->run_list, &run_list); 775 while (!list_empty(&run_list)) { 776 iocb = list_entry(run_list.next, struct kiocb, 777 ki_run_list); 778 list_del(&iocb->ki_run_list); 779 /* 780 * Hold an extra reference while retrying i/o. 781 */ 782 iocb->ki_users++; /* grab extra reference */ 783 aio_run_iocb(iocb); 784 __aio_put_req(ctx, iocb); 785 } 786 if (!list_empty(&ctx->run_list)) 787 return 1; 788 return 0; 789 } 790 791 static void aio_queue_work(struct kioctx * ctx) 792 { 793 unsigned long timeout; 794 /* 795 * if someone is waiting, get the work started right 796 * away, otherwise, use a longer delay 797 */ 798 smp_mb(); 799 if (waitqueue_active(&ctx->wait)) 800 timeout = 1; 801 else 802 timeout = HZ/10; 803 queue_delayed_work(aio_wq, &ctx->wq, timeout); 804 } 805 806 807 /* 808 * aio_run_iocbs: 809 * Process all pending retries queued on the ioctx 810 * run list. 811 * Assumes it is operating within the aio issuer's mm 812 * context. 813 */ 814 static inline void aio_run_iocbs(struct kioctx *ctx) 815 { 816 int requeue; 817 818 spin_lock_irq(&ctx->ctx_lock); 819 820 requeue = __aio_run_iocbs(ctx); 821 spin_unlock_irq(&ctx->ctx_lock); 822 if (requeue) 823 aio_queue_work(ctx); 824 } 825 826 /* 827 * just like aio_run_iocbs, but keeps running them until 828 * the list stays empty 829 */ 830 static inline void aio_run_all_iocbs(struct kioctx *ctx) 831 { 832 spin_lock_irq(&ctx->ctx_lock); 833 while (__aio_run_iocbs(ctx)) 834 ; 835 spin_unlock_irq(&ctx->ctx_lock); 836 } 837 838 /* 839 * aio_kick_handler: 840 * Work queue handler triggered to process pending 841 * retries on an ioctx. Takes on the aio issuer's 842 * mm context before running the iocbs, so that 843 * copy_xxx_user operates on the issuer's address 844 * space. 845 * Run on aiod's context. 846 */ 847 static void aio_kick_handler(struct work_struct *work) 848 { 849 struct kioctx *ctx = container_of(work, struct kioctx, wq.work); 850 mm_segment_t oldfs = get_fs(); 851 struct mm_struct *mm; 852 int requeue; 853 854 set_fs(USER_DS); 855 use_mm(ctx->mm); 856 spin_lock_irq(&ctx->ctx_lock); 857 requeue =__aio_run_iocbs(ctx); 858 mm = ctx->mm; 859 spin_unlock_irq(&ctx->ctx_lock); 860 unuse_mm(mm); 861 set_fs(oldfs); 862 /* 863 * we're in a worker thread already, don't use queue_delayed_work, 864 */ 865 if (requeue) 866 queue_delayed_work(aio_wq, &ctx->wq, 0); 867 } 868 869 870 /* 871 * Called by kick_iocb to queue the kiocb for retry 872 * and if required activate the aio work queue to process 873 * it 874 */ 875 static void try_queue_kicked_iocb(struct kiocb *iocb) 876 { 877 struct kioctx *ctx = iocb->ki_ctx; 878 unsigned long flags; 879 int run = 0; 880 881 /* We're supposed to be the only path putting the iocb back on the run 882 * list. If we find that the iocb is *back* on a wait queue already 883 * than retry has happened before we could queue the iocb. This also 884 * means that the retry could have completed and freed our iocb, no 885 * good. */ 886 BUG_ON((!list_empty(&iocb->ki_wait.task_list))); 887 888 spin_lock_irqsave(&ctx->ctx_lock, flags); 889 /* set this inside the lock so that we can't race with aio_run_iocb() 890 * testing it and putting the iocb on the run list under the lock */ 891 if (!kiocbTryKick(iocb)) 892 run = __queue_kicked_iocb(iocb); 893 spin_unlock_irqrestore(&ctx->ctx_lock, flags); 894 if (run) 895 aio_queue_work(ctx); 896 } 897 898 /* 899 * kick_iocb: 900 * Called typically from a wait queue callback context 901 * (aio_wake_function) to trigger a retry of the iocb. 902 * The retry is usually executed by aio workqueue 903 * threads (See aio_kick_handler). 904 */ 905 void fastcall kick_iocb(struct kiocb *iocb) 906 { 907 /* sync iocbs are easy: they can only ever be executing from a 908 * single context. */ 909 if (is_sync_kiocb(iocb)) { 910 kiocbSetKicked(iocb); 911 wake_up_process(iocb->ki_obj.tsk); 912 return; 913 } 914 915 try_queue_kicked_iocb(iocb); 916 } 917 EXPORT_SYMBOL(kick_iocb); 918 919 /* aio_complete 920 * Called when the io request on the given iocb is complete. 921 * Returns true if this is the last user of the request. The 922 * only other user of the request can be the cancellation code. 923 */ 924 int fastcall aio_complete(struct kiocb *iocb, long res, long res2) 925 { 926 struct kioctx *ctx = iocb->ki_ctx; 927 struct aio_ring_info *info; 928 struct aio_ring *ring; 929 struct io_event *event; 930 unsigned long flags; 931 unsigned long tail; 932 int ret; 933 934 /* 935 * Special case handling for sync iocbs: 936 * - events go directly into the iocb for fast handling 937 * - the sync task with the iocb in its stack holds the single iocb 938 * ref, no other paths have a way to get another ref 939 * - the sync task helpfully left a reference to itself in the iocb 940 */ 941 if (is_sync_kiocb(iocb)) { 942 BUG_ON(iocb->ki_users != 1); 943 iocb->ki_user_data = res; 944 iocb->ki_users = 0; 945 wake_up_process(iocb->ki_obj.tsk); 946 return 1; 947 } 948 949 /* 950 * Check if the user asked us to deliver the result through an 951 * eventfd. The eventfd_signal() function is safe to be called 952 * from IRQ context. 953 */ 954 if (!IS_ERR(iocb->ki_eventfd)) 955 eventfd_signal(iocb->ki_eventfd, 1); 956 957 info = &ctx->ring_info; 958 959 /* add a completion event to the ring buffer. 960 * must be done holding ctx->ctx_lock to prevent 961 * other code from messing with the tail 962 * pointer since we might be called from irq 963 * context. 964 */ 965 spin_lock_irqsave(&ctx->ctx_lock, flags); 966 967 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list)) 968 list_del_init(&iocb->ki_run_list); 969 970 /* 971 * cancelled requests don't get events, userland was given one 972 * when the event got cancelled. 973 */ 974 if (kiocbIsCancelled(iocb)) 975 goto put_rq; 976 977 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1); 978 979 tail = info->tail; 980 event = aio_ring_event(info, tail, KM_IRQ0); 981 if (++tail >= info->nr) 982 tail = 0; 983 984 event->obj = (u64)(unsigned long)iocb->ki_obj.user; 985 event->data = iocb->ki_user_data; 986 event->res = res; 987 event->res2 = res2; 988 989 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n", 990 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data, 991 res, res2); 992 993 /* after flagging the request as done, we 994 * must never even look at it again 995 */ 996 smp_wmb(); /* make event visible before updating tail */ 997 998 info->tail = tail; 999 ring->tail = tail; 1000 1001 put_aio_ring_event(event, KM_IRQ0); 1002 kunmap_atomic(ring, KM_IRQ1); 1003 1004 pr_debug("added to ring %p at [%lu]\n", iocb, tail); 1005 put_rq: 1006 /* everything turned out well, dispose of the aiocb. */ 1007 ret = __aio_put_req(ctx, iocb); 1008 1009 if (waitqueue_active(&ctx->wait)) 1010 wake_up(&ctx->wait); 1011 1012 spin_unlock_irqrestore(&ctx->ctx_lock, flags); 1013 return ret; 1014 } 1015 1016 /* aio_read_evt 1017 * Pull an event off of the ioctx's event ring. Returns the number of 1018 * events fetched (0 or 1 ;-) 1019 * FIXME: make this use cmpxchg. 1020 * TODO: make the ringbuffer user mmap()able (requires FIXME). 1021 */ 1022 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent) 1023 { 1024 struct aio_ring_info *info = &ioctx->ring_info; 1025 struct aio_ring *ring; 1026 unsigned long head; 1027 int ret = 0; 1028 1029 ring = kmap_atomic(info->ring_pages[0], KM_USER0); 1030 dprintk("in aio_read_evt h%lu t%lu m%lu\n", 1031 (unsigned long)ring->head, (unsigned long)ring->tail, 1032 (unsigned long)ring->nr); 1033 1034 if (ring->head == ring->tail) 1035 goto out; 1036 1037 spin_lock(&info->ring_lock); 1038 1039 head = ring->head % info->nr; 1040 if (head != ring->tail) { 1041 struct io_event *evp = aio_ring_event(info, head, KM_USER1); 1042 *ent = *evp; 1043 head = (head + 1) % info->nr; 1044 smp_mb(); /* finish reading the event before updatng the head */ 1045 ring->head = head; 1046 ret = 1; 1047 put_aio_ring_event(evp, KM_USER1); 1048 } 1049 spin_unlock(&info->ring_lock); 1050 1051 out: 1052 kunmap_atomic(ring, KM_USER0); 1053 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret, 1054 (unsigned long)ring->head, (unsigned long)ring->tail); 1055 return ret; 1056 } 1057 1058 struct aio_timeout { 1059 struct timer_list timer; 1060 int timed_out; 1061 struct task_struct *p; 1062 }; 1063 1064 static void timeout_func(unsigned long data) 1065 { 1066 struct aio_timeout *to = (struct aio_timeout *)data; 1067 1068 to->timed_out = 1; 1069 wake_up_process(to->p); 1070 } 1071 1072 static inline void init_timeout(struct aio_timeout *to) 1073 { 1074 init_timer(&to->timer); 1075 to->timer.data = (unsigned long)to; 1076 to->timer.function = timeout_func; 1077 to->timed_out = 0; 1078 to->p = current; 1079 } 1080 1081 static inline void set_timeout(long start_jiffies, struct aio_timeout *to, 1082 const struct timespec *ts) 1083 { 1084 to->timer.expires = start_jiffies + timespec_to_jiffies(ts); 1085 if (time_after(to->timer.expires, jiffies)) 1086 add_timer(&to->timer); 1087 else 1088 to->timed_out = 1; 1089 } 1090 1091 static inline void clear_timeout(struct aio_timeout *to) 1092 { 1093 del_singleshot_timer_sync(&to->timer); 1094 } 1095 1096 static int read_events(struct kioctx *ctx, 1097 long min_nr, long nr, 1098 struct io_event __user *event, 1099 struct timespec __user *timeout) 1100 { 1101 long start_jiffies = jiffies; 1102 struct task_struct *tsk = current; 1103 DECLARE_WAITQUEUE(wait, tsk); 1104 int ret; 1105 int i = 0; 1106 struct io_event ent; 1107 struct aio_timeout to; 1108 int retry = 0; 1109 1110 /* needed to zero any padding within an entry (there shouldn't be 1111 * any, but C is fun! 1112 */ 1113 memset(&ent, 0, sizeof(ent)); 1114 retry: 1115 ret = 0; 1116 while (likely(i < nr)) { 1117 ret = aio_read_evt(ctx, &ent); 1118 if (unlikely(ret <= 0)) 1119 break; 1120 1121 dprintk("read event: %Lx %Lx %Lx %Lx\n", 1122 ent.data, ent.obj, ent.res, ent.res2); 1123 1124 /* Could we split the check in two? */ 1125 ret = -EFAULT; 1126 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) { 1127 dprintk("aio: lost an event due to EFAULT.\n"); 1128 break; 1129 } 1130 ret = 0; 1131 1132 /* Good, event copied to userland, update counts. */ 1133 event ++; 1134 i ++; 1135 } 1136 1137 if (min_nr <= i) 1138 return i; 1139 if (ret) 1140 return ret; 1141 1142 /* End fast path */ 1143 1144 /* racey check, but it gets redone */ 1145 if (!retry && unlikely(!list_empty(&ctx->run_list))) { 1146 retry = 1; 1147 aio_run_all_iocbs(ctx); 1148 goto retry; 1149 } 1150 1151 init_timeout(&to); 1152 if (timeout) { 1153 struct timespec ts; 1154 ret = -EFAULT; 1155 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts)))) 1156 goto out; 1157 1158 set_timeout(start_jiffies, &to, &ts); 1159 } 1160 1161 while (likely(i < nr)) { 1162 add_wait_queue_exclusive(&ctx->wait, &wait); 1163 do { 1164 set_task_state(tsk, TASK_INTERRUPTIBLE); 1165 ret = aio_read_evt(ctx, &ent); 1166 if (ret) 1167 break; 1168 if (min_nr <= i) 1169 break; 1170 ret = 0; 1171 if (to.timed_out) /* Only check after read evt */ 1172 break; 1173 schedule(); 1174 if (signal_pending(tsk)) { 1175 ret = -EINTR; 1176 break; 1177 } 1178 /*ret = aio_read_evt(ctx, &ent);*/ 1179 } while (1) ; 1180 1181 set_task_state(tsk, TASK_RUNNING); 1182 remove_wait_queue(&ctx->wait, &wait); 1183 1184 if (unlikely(ret <= 0)) 1185 break; 1186 1187 ret = -EFAULT; 1188 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) { 1189 dprintk("aio: lost an event due to EFAULT.\n"); 1190 break; 1191 } 1192 1193 /* Good, event copied to userland, update counts. */ 1194 event ++; 1195 i ++; 1196 } 1197 1198 if (timeout) 1199 clear_timeout(&to); 1200 out: 1201 return i ? i : ret; 1202 } 1203 1204 /* Take an ioctx and remove it from the list of ioctx's. Protects 1205 * against races with itself via ->dead. 1206 */ 1207 static void io_destroy(struct kioctx *ioctx) 1208 { 1209 struct mm_struct *mm = current->mm; 1210 struct kioctx **tmp; 1211 int was_dead; 1212 1213 /* delete the entry from the list is someone else hasn't already */ 1214 write_lock(&mm->ioctx_list_lock); 1215 was_dead = ioctx->dead; 1216 ioctx->dead = 1; 1217 for (tmp = &mm->ioctx_list; *tmp && *tmp != ioctx; 1218 tmp = &(*tmp)->next) 1219 ; 1220 if (*tmp) 1221 *tmp = ioctx->next; 1222 write_unlock(&mm->ioctx_list_lock); 1223 1224 dprintk("aio_release(%p)\n", ioctx); 1225 if (likely(!was_dead)) 1226 put_ioctx(ioctx); /* twice for the list */ 1227 1228 aio_cancel_all(ioctx); 1229 wait_for_all_aios(ioctx); 1230 put_ioctx(ioctx); /* once for the lookup */ 1231 } 1232 1233 /* sys_io_setup: 1234 * Create an aio_context capable of receiving at least nr_events. 1235 * ctxp must not point to an aio_context that already exists, and 1236 * must be initialized to 0 prior to the call. On successful 1237 * creation of the aio_context, *ctxp is filled in with the resulting 1238 * handle. May fail with -EINVAL if *ctxp is not initialized, 1239 * if the specified nr_events exceeds internal limits. May fail 1240 * with -EAGAIN if the specified nr_events exceeds the user's limit 1241 * of available events. May fail with -ENOMEM if insufficient kernel 1242 * resources are available. May fail with -EFAULT if an invalid 1243 * pointer is passed for ctxp. Will fail with -ENOSYS if not 1244 * implemented. 1245 */ 1246 asmlinkage long sys_io_setup(unsigned nr_events, aio_context_t __user *ctxp) 1247 { 1248 struct kioctx *ioctx = NULL; 1249 unsigned long ctx; 1250 long ret; 1251 1252 ret = get_user(ctx, ctxp); 1253 if (unlikely(ret)) 1254 goto out; 1255 1256 ret = -EINVAL; 1257 if (unlikely(ctx || nr_events == 0)) { 1258 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n", 1259 ctx, nr_events); 1260 goto out; 1261 } 1262 1263 ioctx = ioctx_alloc(nr_events); 1264 ret = PTR_ERR(ioctx); 1265 if (!IS_ERR(ioctx)) { 1266 ret = put_user(ioctx->user_id, ctxp); 1267 if (!ret) 1268 return 0; 1269 1270 get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */ 1271 io_destroy(ioctx); 1272 } 1273 1274 out: 1275 return ret; 1276 } 1277 1278 /* sys_io_destroy: 1279 * Destroy the aio_context specified. May cancel any outstanding 1280 * AIOs and block on completion. Will fail with -ENOSYS if not 1281 * implemented. May fail with -EFAULT if the context pointed to 1282 * is invalid. 1283 */ 1284 asmlinkage long sys_io_destroy(aio_context_t ctx) 1285 { 1286 struct kioctx *ioctx = lookup_ioctx(ctx); 1287 if (likely(NULL != ioctx)) { 1288 io_destroy(ioctx); 1289 return 0; 1290 } 1291 pr_debug("EINVAL: io_destroy: invalid context id\n"); 1292 return -EINVAL; 1293 } 1294 1295 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret) 1296 { 1297 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg]; 1298 1299 BUG_ON(ret <= 0); 1300 1301 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) { 1302 ssize_t this = min((ssize_t)iov->iov_len, ret); 1303 iov->iov_base += this; 1304 iov->iov_len -= this; 1305 iocb->ki_left -= this; 1306 ret -= this; 1307 if (iov->iov_len == 0) { 1308 iocb->ki_cur_seg++; 1309 iov++; 1310 } 1311 } 1312 1313 /* the caller should not have done more io than what fit in 1314 * the remaining iovecs */ 1315 BUG_ON(ret > 0 && iocb->ki_left == 0); 1316 } 1317 1318 static ssize_t aio_rw_vect_retry(struct kiocb *iocb) 1319 { 1320 struct file *file = iocb->ki_filp; 1321 struct address_space *mapping = file->f_mapping; 1322 struct inode *inode = mapping->host; 1323 ssize_t (*rw_op)(struct kiocb *, const struct iovec *, 1324 unsigned long, loff_t); 1325 ssize_t ret = 0; 1326 unsigned short opcode; 1327 1328 if ((iocb->ki_opcode == IOCB_CMD_PREADV) || 1329 (iocb->ki_opcode == IOCB_CMD_PREAD)) { 1330 rw_op = file->f_op->aio_read; 1331 opcode = IOCB_CMD_PREADV; 1332 } else { 1333 rw_op = file->f_op->aio_write; 1334 opcode = IOCB_CMD_PWRITEV; 1335 } 1336 1337 do { 1338 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg], 1339 iocb->ki_nr_segs - iocb->ki_cur_seg, 1340 iocb->ki_pos); 1341 if (ret > 0) 1342 aio_advance_iovec(iocb, ret); 1343 1344 /* retry all partial writes. retry partial reads as long as its a 1345 * regular file. */ 1346 } while (ret > 0 && iocb->ki_left > 0 && 1347 (opcode == IOCB_CMD_PWRITEV || 1348 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode)))); 1349 1350 /* This means we must have transferred all that we could */ 1351 /* No need to retry anymore */ 1352 if ((ret == 0) || (iocb->ki_left == 0)) 1353 ret = iocb->ki_nbytes - iocb->ki_left; 1354 1355 return ret; 1356 } 1357 1358 static ssize_t aio_fdsync(struct kiocb *iocb) 1359 { 1360 struct file *file = iocb->ki_filp; 1361 ssize_t ret = -EINVAL; 1362 1363 if (file->f_op->aio_fsync) 1364 ret = file->f_op->aio_fsync(iocb, 1); 1365 return ret; 1366 } 1367 1368 static ssize_t aio_fsync(struct kiocb *iocb) 1369 { 1370 struct file *file = iocb->ki_filp; 1371 ssize_t ret = -EINVAL; 1372 1373 if (file->f_op->aio_fsync) 1374 ret = file->f_op->aio_fsync(iocb, 0); 1375 return ret; 1376 } 1377 1378 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb) 1379 { 1380 ssize_t ret; 1381 1382 ret = rw_copy_check_uvector(type, (struct iovec __user *)kiocb->ki_buf, 1383 kiocb->ki_nbytes, 1, 1384 &kiocb->ki_inline_vec, &kiocb->ki_iovec); 1385 if (ret < 0) 1386 goto out; 1387 1388 kiocb->ki_nr_segs = kiocb->ki_nbytes; 1389 kiocb->ki_cur_seg = 0; 1390 /* ki_nbytes/left now reflect bytes instead of segs */ 1391 kiocb->ki_nbytes = ret; 1392 kiocb->ki_left = ret; 1393 1394 ret = 0; 1395 out: 1396 return ret; 1397 } 1398 1399 static ssize_t aio_setup_single_vector(struct kiocb *kiocb) 1400 { 1401 kiocb->ki_iovec = &kiocb->ki_inline_vec; 1402 kiocb->ki_iovec->iov_base = kiocb->ki_buf; 1403 kiocb->ki_iovec->iov_len = kiocb->ki_left; 1404 kiocb->ki_nr_segs = 1; 1405 kiocb->ki_cur_seg = 0; 1406 return 0; 1407 } 1408 1409 /* 1410 * aio_setup_iocb: 1411 * Performs the initial checks and aio retry method 1412 * setup for the kiocb at the time of io submission. 1413 */ 1414 static ssize_t aio_setup_iocb(struct kiocb *kiocb) 1415 { 1416 struct file *file = kiocb->ki_filp; 1417 ssize_t ret = 0; 1418 1419 switch (kiocb->ki_opcode) { 1420 case IOCB_CMD_PREAD: 1421 ret = -EBADF; 1422 if (unlikely(!(file->f_mode & FMODE_READ))) 1423 break; 1424 ret = -EFAULT; 1425 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf, 1426 kiocb->ki_left))) 1427 break; 1428 ret = security_file_permission(file, MAY_READ); 1429 if (unlikely(ret)) 1430 break; 1431 ret = aio_setup_single_vector(kiocb); 1432 if (ret) 1433 break; 1434 ret = -EINVAL; 1435 if (file->f_op->aio_read) 1436 kiocb->ki_retry = aio_rw_vect_retry; 1437 break; 1438 case IOCB_CMD_PWRITE: 1439 ret = -EBADF; 1440 if (unlikely(!(file->f_mode & FMODE_WRITE))) 1441 break; 1442 ret = -EFAULT; 1443 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf, 1444 kiocb->ki_left))) 1445 break; 1446 ret = security_file_permission(file, MAY_WRITE); 1447 if (unlikely(ret)) 1448 break; 1449 ret = aio_setup_single_vector(kiocb); 1450 if (ret) 1451 break; 1452 ret = -EINVAL; 1453 if (file->f_op->aio_write) 1454 kiocb->ki_retry = aio_rw_vect_retry; 1455 break; 1456 case IOCB_CMD_PREADV: 1457 ret = -EBADF; 1458 if (unlikely(!(file->f_mode & FMODE_READ))) 1459 break; 1460 ret = security_file_permission(file, MAY_READ); 1461 if (unlikely(ret)) 1462 break; 1463 ret = aio_setup_vectored_rw(READ, kiocb); 1464 if (ret) 1465 break; 1466 ret = -EINVAL; 1467 if (file->f_op->aio_read) 1468 kiocb->ki_retry = aio_rw_vect_retry; 1469 break; 1470 case IOCB_CMD_PWRITEV: 1471 ret = -EBADF; 1472 if (unlikely(!(file->f_mode & FMODE_WRITE))) 1473 break; 1474 ret = security_file_permission(file, MAY_WRITE); 1475 if (unlikely(ret)) 1476 break; 1477 ret = aio_setup_vectored_rw(WRITE, kiocb); 1478 if (ret) 1479 break; 1480 ret = -EINVAL; 1481 if (file->f_op->aio_write) 1482 kiocb->ki_retry = aio_rw_vect_retry; 1483 break; 1484 case IOCB_CMD_FDSYNC: 1485 ret = -EINVAL; 1486 if (file->f_op->aio_fsync) 1487 kiocb->ki_retry = aio_fdsync; 1488 break; 1489 case IOCB_CMD_FSYNC: 1490 ret = -EINVAL; 1491 if (file->f_op->aio_fsync) 1492 kiocb->ki_retry = aio_fsync; 1493 break; 1494 default: 1495 dprintk("EINVAL: io_submit: no operation provided\n"); 1496 ret = -EINVAL; 1497 } 1498 1499 if (!kiocb->ki_retry) 1500 return ret; 1501 1502 return 0; 1503 } 1504 1505 /* 1506 * aio_wake_function: 1507 * wait queue callback function for aio notification, 1508 * Simply triggers a retry of the operation via kick_iocb. 1509 * 1510 * This callback is specified in the wait queue entry in 1511 * a kiocb (current->io_wait points to this wait queue 1512 * entry when an aio operation executes; it is used 1513 * instead of a synchronous wait when an i/o blocking 1514 * condition is encountered during aio). 1515 * 1516 * Note: 1517 * This routine is executed with the wait queue lock held. 1518 * Since kick_iocb acquires iocb->ctx->ctx_lock, it nests 1519 * the ioctx lock inside the wait queue lock. This is safe 1520 * because this callback isn't used for wait queues which 1521 * are nested inside ioctx lock (i.e. ctx->wait) 1522 */ 1523 static int aio_wake_function(wait_queue_t *wait, unsigned mode, 1524 int sync, void *key) 1525 { 1526 struct kiocb *iocb = container_of(wait, struct kiocb, ki_wait); 1527 1528 list_del_init(&wait->task_list); 1529 kick_iocb(iocb); 1530 return 1; 1531 } 1532 1533 int fastcall io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb, 1534 struct iocb *iocb) 1535 { 1536 struct kiocb *req; 1537 struct file *file; 1538 ssize_t ret; 1539 1540 /* enforce forwards compatibility on users */ 1541 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) { 1542 pr_debug("EINVAL: io_submit: reserve field set\n"); 1543 return -EINVAL; 1544 } 1545 1546 /* prevent overflows */ 1547 if (unlikely( 1548 (iocb->aio_buf != (unsigned long)iocb->aio_buf) || 1549 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) || 1550 ((ssize_t)iocb->aio_nbytes < 0) 1551 )) { 1552 pr_debug("EINVAL: io_submit: overflow check\n"); 1553 return -EINVAL; 1554 } 1555 1556 file = fget(iocb->aio_fildes); 1557 if (unlikely(!file)) 1558 return -EBADF; 1559 1560 req = aio_get_req(ctx); /* returns with 2 references to req */ 1561 if (unlikely(!req)) { 1562 fput(file); 1563 return -EAGAIN; 1564 } 1565 if (iocb->aio_flags & IOCB_FLAG_RESFD) { 1566 /* 1567 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an 1568 * instance of the file* now. The file descriptor must be 1569 * an eventfd() fd, and will be signaled for each completed 1570 * event using the eventfd_signal() function. 1571 */ 1572 req->ki_eventfd = eventfd_fget((int) iocb->aio_resfd); 1573 if (unlikely(IS_ERR(req->ki_eventfd))) { 1574 ret = PTR_ERR(req->ki_eventfd); 1575 goto out_put_req; 1576 } 1577 } 1578 1579 req->ki_filp = file; 1580 ret = put_user(req->ki_key, &user_iocb->aio_key); 1581 if (unlikely(ret)) { 1582 dprintk("EFAULT: aio_key\n"); 1583 goto out_put_req; 1584 } 1585 1586 req->ki_obj.user = user_iocb; 1587 req->ki_user_data = iocb->aio_data; 1588 req->ki_pos = iocb->aio_offset; 1589 1590 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf; 1591 req->ki_left = req->ki_nbytes = iocb->aio_nbytes; 1592 req->ki_opcode = iocb->aio_lio_opcode; 1593 init_waitqueue_func_entry(&req->ki_wait, aio_wake_function); 1594 INIT_LIST_HEAD(&req->ki_wait.task_list); 1595 1596 ret = aio_setup_iocb(req); 1597 1598 if (ret) 1599 goto out_put_req; 1600 1601 spin_lock_irq(&ctx->ctx_lock); 1602 aio_run_iocb(req); 1603 if (!list_empty(&ctx->run_list)) { 1604 /* drain the run list */ 1605 while (__aio_run_iocbs(ctx)) 1606 ; 1607 } 1608 spin_unlock_irq(&ctx->ctx_lock); 1609 aio_put_req(req); /* drop extra ref to req */ 1610 return 0; 1611 1612 out_put_req: 1613 aio_put_req(req); /* drop extra ref to req */ 1614 aio_put_req(req); /* drop i/o ref to req */ 1615 return ret; 1616 } 1617 1618 /* sys_io_submit: 1619 * Queue the nr iocbs pointed to by iocbpp for processing. Returns 1620 * the number of iocbs queued. May return -EINVAL if the aio_context 1621 * specified by ctx_id is invalid, if nr is < 0, if the iocb at 1622 * *iocbpp[0] is not properly initialized, if the operation specified 1623 * is invalid for the file descriptor in the iocb. May fail with 1624 * -EFAULT if any of the data structures point to invalid data. May 1625 * fail with -EBADF if the file descriptor specified in the first 1626 * iocb is invalid. May fail with -EAGAIN if insufficient resources 1627 * are available to queue any iocbs. Will return 0 if nr is 0. Will 1628 * fail with -ENOSYS if not implemented. 1629 */ 1630 asmlinkage long sys_io_submit(aio_context_t ctx_id, long nr, 1631 struct iocb __user * __user *iocbpp) 1632 { 1633 struct kioctx *ctx; 1634 long ret = 0; 1635 int i; 1636 1637 if (unlikely(nr < 0)) 1638 return -EINVAL; 1639 1640 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp))))) 1641 return -EFAULT; 1642 1643 ctx = lookup_ioctx(ctx_id); 1644 if (unlikely(!ctx)) { 1645 pr_debug("EINVAL: io_submit: invalid context id\n"); 1646 return -EINVAL; 1647 } 1648 1649 /* 1650 * AKPM: should this return a partial result if some of the IOs were 1651 * successfully submitted? 1652 */ 1653 for (i=0; i<nr; i++) { 1654 struct iocb __user *user_iocb; 1655 struct iocb tmp; 1656 1657 if (unlikely(__get_user(user_iocb, iocbpp + i))) { 1658 ret = -EFAULT; 1659 break; 1660 } 1661 1662 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) { 1663 ret = -EFAULT; 1664 break; 1665 } 1666 1667 ret = io_submit_one(ctx, user_iocb, &tmp); 1668 if (ret) 1669 break; 1670 } 1671 1672 put_ioctx(ctx); 1673 return i ? i : ret; 1674 } 1675 1676 /* lookup_kiocb 1677 * Finds a given iocb for cancellation. 1678 */ 1679 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, 1680 u32 key) 1681 { 1682 struct list_head *pos; 1683 1684 assert_spin_locked(&ctx->ctx_lock); 1685 1686 /* TODO: use a hash or array, this sucks. */ 1687 list_for_each(pos, &ctx->active_reqs) { 1688 struct kiocb *kiocb = list_kiocb(pos); 1689 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key) 1690 return kiocb; 1691 } 1692 return NULL; 1693 } 1694 1695 /* sys_io_cancel: 1696 * Attempts to cancel an iocb previously passed to io_submit. If 1697 * the operation is successfully cancelled, the resulting event is 1698 * copied into the memory pointed to by result without being placed 1699 * into the completion queue and 0 is returned. May fail with 1700 * -EFAULT if any of the data structures pointed to are invalid. 1701 * May fail with -EINVAL if aio_context specified by ctx_id is 1702 * invalid. May fail with -EAGAIN if the iocb specified was not 1703 * cancelled. Will fail with -ENOSYS if not implemented. 1704 */ 1705 asmlinkage long sys_io_cancel(aio_context_t ctx_id, struct iocb __user *iocb, 1706 struct io_event __user *result) 1707 { 1708 int (*cancel)(struct kiocb *iocb, struct io_event *res); 1709 struct kioctx *ctx; 1710 struct kiocb *kiocb; 1711 u32 key; 1712 int ret; 1713 1714 ret = get_user(key, &iocb->aio_key); 1715 if (unlikely(ret)) 1716 return -EFAULT; 1717 1718 ctx = lookup_ioctx(ctx_id); 1719 if (unlikely(!ctx)) 1720 return -EINVAL; 1721 1722 spin_lock_irq(&ctx->ctx_lock); 1723 ret = -EAGAIN; 1724 kiocb = lookup_kiocb(ctx, iocb, key); 1725 if (kiocb && kiocb->ki_cancel) { 1726 cancel = kiocb->ki_cancel; 1727 kiocb->ki_users ++; 1728 kiocbSetCancelled(kiocb); 1729 } else 1730 cancel = NULL; 1731 spin_unlock_irq(&ctx->ctx_lock); 1732 1733 if (NULL != cancel) { 1734 struct io_event tmp; 1735 pr_debug("calling cancel\n"); 1736 memset(&tmp, 0, sizeof(tmp)); 1737 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user; 1738 tmp.data = kiocb->ki_user_data; 1739 ret = cancel(kiocb, &tmp); 1740 if (!ret) { 1741 /* Cancellation succeeded -- copy the result 1742 * into the user's buffer. 1743 */ 1744 if (copy_to_user(result, &tmp, sizeof(tmp))) 1745 ret = -EFAULT; 1746 } 1747 } else 1748 ret = -EINVAL; 1749 1750 put_ioctx(ctx); 1751 1752 return ret; 1753 } 1754 1755 /* io_getevents: 1756 * Attempts to read at least min_nr events and up to nr events from 1757 * the completion queue for the aio_context specified by ctx_id. May 1758 * fail with -EINVAL if ctx_id is invalid, if min_nr is out of range, 1759 * if nr is out of range, if when is out of range. May fail with 1760 * -EFAULT if any of the memory specified to is invalid. May return 1761 * 0 or < min_nr if no events are available and the timeout specified 1762 * by when has elapsed, where when == NULL specifies an infinite 1763 * timeout. Note that the timeout pointed to by when is relative and 1764 * will be updated if not NULL and the operation blocks. Will fail 1765 * with -ENOSYS if not implemented. 1766 */ 1767 asmlinkage long sys_io_getevents(aio_context_t ctx_id, 1768 long min_nr, 1769 long nr, 1770 struct io_event __user *events, 1771 struct timespec __user *timeout) 1772 { 1773 struct kioctx *ioctx = lookup_ioctx(ctx_id); 1774 long ret = -EINVAL; 1775 1776 if (likely(ioctx)) { 1777 if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0)) 1778 ret = read_events(ioctx, min_nr, nr, events, timeout); 1779 put_ioctx(ioctx); 1780 } 1781 1782 return ret; 1783 } 1784 1785 __initcall(aio_setup); 1786 1787 EXPORT_SYMBOL(aio_complete); 1788 EXPORT_SYMBOL(aio_put_req); 1789 EXPORT_SYMBOL(wait_on_sync_kiocb); 1790