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 34 #include <asm/kmap_types.h> 35 #include <asm/uaccess.h> 36 #include <asm/mmu_context.h> 37 38 #if DEBUG > 1 39 #define dprintk printk 40 #else 41 #define dprintk(x...) do { ; } while (0) 42 #endif 43 44 /*------ sysctl variables----*/ 45 static DEFINE_SPINLOCK(aio_nr_lock); 46 unsigned long aio_nr; /* current system wide number of aio requests */ 47 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */ 48 /*----end sysctl variables---*/ 49 50 static struct kmem_cache *kiocb_cachep; 51 static struct kmem_cache *kioctx_cachep; 52 53 static struct workqueue_struct *aio_wq; 54 55 /* Used for rare fput completion. */ 56 static void aio_fput_routine(struct work_struct *); 57 static DECLARE_WORK(fput_work, aio_fput_routine); 58 59 static DEFINE_SPINLOCK(fput_lock); 60 static LIST_HEAD(fput_head); 61 62 static void aio_kick_handler(struct work_struct *); 63 static void aio_queue_work(struct kioctx *); 64 65 /* aio_setup 66 * Creates the slab caches used by the aio routines, panic on 67 * failure as this is done early during the boot sequence. 68 */ 69 static int __init aio_setup(void) 70 { 71 kiocb_cachep = kmem_cache_create("kiocb", sizeof(struct kiocb), 72 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL); 73 kioctx_cachep = kmem_cache_create("kioctx", sizeof(struct kioctx), 74 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL); 75 76 aio_wq = create_workqueue("aio"); 77 78 pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page)); 79 80 return 0; 81 } 82 83 static void aio_free_ring(struct kioctx *ctx) 84 { 85 struct aio_ring_info *info = &ctx->ring_info; 86 long i; 87 88 for (i=0; i<info->nr_pages; i++) 89 put_page(info->ring_pages[i]); 90 91 if (info->mmap_size) { 92 down_write(&ctx->mm->mmap_sem); 93 do_munmap(ctx->mm, info->mmap_base, info->mmap_size); 94 up_write(&ctx->mm->mmap_sem); 95 } 96 97 if (info->ring_pages && info->ring_pages != info->internal_pages) 98 kfree(info->ring_pages); 99 info->ring_pages = NULL; 100 info->nr = 0; 101 } 102 103 static int aio_setup_ring(struct kioctx *ctx) 104 { 105 struct aio_ring *ring; 106 struct aio_ring_info *info = &ctx->ring_info; 107 unsigned nr_events = ctx->max_reqs; 108 unsigned long size; 109 int nr_pages; 110 111 /* Compensate for the ring buffer's head/tail overlap entry */ 112 nr_events += 2; /* 1 is required, 2 for good luck */ 113 114 size = sizeof(struct aio_ring); 115 size += sizeof(struct io_event) * nr_events; 116 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT; 117 118 if (nr_pages < 0) 119 return -EINVAL; 120 121 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event); 122 123 info->nr = 0; 124 info->ring_pages = info->internal_pages; 125 if (nr_pages > AIO_RING_PAGES) { 126 info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL); 127 if (!info->ring_pages) 128 return -ENOMEM; 129 } 130 131 info->mmap_size = nr_pages * PAGE_SIZE; 132 dprintk("attempting mmap of %lu bytes\n", info->mmap_size); 133 down_write(&ctx->mm->mmap_sem); 134 info->mmap_base = do_mmap(NULL, 0, info->mmap_size, 135 PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE, 136 0); 137 if (IS_ERR((void *)info->mmap_base)) { 138 up_write(&ctx->mm->mmap_sem); 139 printk("mmap err: %ld\n", -info->mmap_base); 140 info->mmap_size = 0; 141 aio_free_ring(ctx); 142 return -EAGAIN; 143 } 144 145 dprintk("mmap address: 0x%08lx\n", info->mmap_base); 146 info->nr_pages = get_user_pages(current, ctx->mm, 147 info->mmap_base, nr_pages, 148 1, 0, info->ring_pages, NULL); 149 up_write(&ctx->mm->mmap_sem); 150 151 if (unlikely(info->nr_pages != nr_pages)) { 152 aio_free_ring(ctx); 153 return -EAGAIN; 154 } 155 156 ctx->user_id = info->mmap_base; 157 158 info->nr = nr_events; /* trusted copy */ 159 160 ring = kmap_atomic(info->ring_pages[0], KM_USER0); 161 ring->nr = nr_events; /* user copy */ 162 ring->id = ctx->user_id; 163 ring->head = ring->tail = 0; 164 ring->magic = AIO_RING_MAGIC; 165 ring->compat_features = AIO_RING_COMPAT_FEATURES; 166 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES; 167 ring->header_length = sizeof(struct aio_ring); 168 kunmap_atomic(ring, KM_USER0); 169 170 return 0; 171 } 172 173 174 /* aio_ring_event: returns a pointer to the event at the given index from 175 * kmap_atomic(, km). Release the pointer with put_aio_ring_event(); 176 */ 177 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event)) 178 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event)) 179 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE) 180 181 #define aio_ring_event(info, nr, km) ({ \ 182 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \ 183 struct io_event *__event; \ 184 __event = kmap_atomic( \ 185 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \ 186 __event += pos % AIO_EVENTS_PER_PAGE; \ 187 __event; \ 188 }) 189 190 #define put_aio_ring_event(event, km) do { \ 191 struct io_event *__event = (event); \ 192 (void)__event; \ 193 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \ 194 } while(0) 195 196 /* ioctx_alloc 197 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed. 198 */ 199 static struct kioctx *ioctx_alloc(unsigned nr_events) 200 { 201 struct mm_struct *mm; 202 struct kioctx *ctx; 203 204 /* Prevent overflows */ 205 if ((nr_events > (0x10000000U / sizeof(struct io_event))) || 206 (nr_events > (0x10000000U / sizeof(struct kiocb)))) { 207 pr_debug("ENOMEM: nr_events too high\n"); 208 return ERR_PTR(-EINVAL); 209 } 210 211 if ((unsigned long)nr_events > aio_max_nr) 212 return ERR_PTR(-EAGAIN); 213 214 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL); 215 if (!ctx) 216 return ERR_PTR(-ENOMEM); 217 218 ctx->max_reqs = nr_events; 219 mm = ctx->mm = current->mm; 220 atomic_inc(&mm->mm_count); 221 222 atomic_set(&ctx->users, 1); 223 spin_lock_init(&ctx->ctx_lock); 224 spin_lock_init(&ctx->ring_info.ring_lock); 225 init_waitqueue_head(&ctx->wait); 226 227 INIT_LIST_HEAD(&ctx->active_reqs); 228 INIT_LIST_HEAD(&ctx->run_list); 229 INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler); 230 231 if (aio_setup_ring(ctx) < 0) 232 goto out_freectx; 233 234 /* limit the number of system wide aios */ 235 spin_lock(&aio_nr_lock); 236 if (aio_nr + ctx->max_reqs > aio_max_nr || 237 aio_nr + ctx->max_reqs < aio_nr) 238 ctx->max_reqs = 0; 239 else 240 aio_nr += ctx->max_reqs; 241 spin_unlock(&aio_nr_lock); 242 if (ctx->max_reqs == 0) 243 goto out_cleanup; 244 245 /* now link into global list. kludge. FIXME */ 246 write_lock(&mm->ioctx_list_lock); 247 ctx->next = mm->ioctx_list; 248 mm->ioctx_list = ctx; 249 write_unlock(&mm->ioctx_list_lock); 250 251 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n", 252 ctx, ctx->user_id, current->mm, ctx->ring_info.nr); 253 return ctx; 254 255 out_cleanup: 256 __put_ioctx(ctx); 257 return ERR_PTR(-EAGAIN); 258 259 out_freectx: 260 mmdrop(mm); 261 kmem_cache_free(kioctx_cachep, ctx); 262 ctx = ERR_PTR(-ENOMEM); 263 264 dprintk("aio: error allocating ioctx %p\n", ctx); 265 return ctx; 266 } 267 268 /* aio_cancel_all 269 * Cancels all outstanding aio requests on an aio context. Used 270 * when the processes owning a context have all exited to encourage 271 * the rapid destruction of the kioctx. 272 */ 273 static void aio_cancel_all(struct kioctx *ctx) 274 { 275 int (*cancel)(struct kiocb *, struct io_event *); 276 struct io_event res; 277 spin_lock_irq(&ctx->ctx_lock); 278 ctx->dead = 1; 279 while (!list_empty(&ctx->active_reqs)) { 280 struct list_head *pos = ctx->active_reqs.next; 281 struct kiocb *iocb = list_kiocb(pos); 282 list_del_init(&iocb->ki_list); 283 cancel = iocb->ki_cancel; 284 kiocbSetCancelled(iocb); 285 if (cancel) { 286 iocb->ki_users++; 287 spin_unlock_irq(&ctx->ctx_lock); 288 cancel(iocb, &res); 289 spin_lock_irq(&ctx->ctx_lock); 290 } 291 } 292 spin_unlock_irq(&ctx->ctx_lock); 293 } 294 295 static void wait_for_all_aios(struct kioctx *ctx) 296 { 297 struct task_struct *tsk = current; 298 DECLARE_WAITQUEUE(wait, tsk); 299 300 spin_lock_irq(&ctx->ctx_lock); 301 if (!ctx->reqs_active) 302 goto out; 303 304 add_wait_queue(&ctx->wait, &wait); 305 set_task_state(tsk, TASK_UNINTERRUPTIBLE); 306 while (ctx->reqs_active) { 307 spin_unlock_irq(&ctx->ctx_lock); 308 schedule(); 309 set_task_state(tsk, TASK_UNINTERRUPTIBLE); 310 spin_lock_irq(&ctx->ctx_lock); 311 } 312 __set_task_state(tsk, TASK_RUNNING); 313 remove_wait_queue(&ctx->wait, &wait); 314 315 out: 316 spin_unlock_irq(&ctx->ctx_lock); 317 } 318 319 /* wait_on_sync_kiocb: 320 * Waits on the given sync kiocb to complete. 321 */ 322 ssize_t fastcall wait_on_sync_kiocb(struct kiocb *iocb) 323 { 324 while (iocb->ki_users) { 325 set_current_state(TASK_UNINTERRUPTIBLE); 326 if (!iocb->ki_users) 327 break; 328 schedule(); 329 } 330 __set_current_state(TASK_RUNNING); 331 return iocb->ki_user_data; 332 } 333 334 /* exit_aio: called when the last user of mm goes away. At this point, 335 * there is no way for any new requests to be submited or any of the 336 * io_* syscalls to be called on the context. However, there may be 337 * outstanding requests which hold references to the context; as they 338 * go away, they will call put_ioctx and release any pinned memory 339 * associated with the request (held via struct page * references). 340 */ 341 void fastcall exit_aio(struct mm_struct *mm) 342 { 343 struct kioctx *ctx = mm->ioctx_list; 344 mm->ioctx_list = NULL; 345 while (ctx) { 346 struct kioctx *next = ctx->next; 347 ctx->next = NULL; 348 aio_cancel_all(ctx); 349 350 wait_for_all_aios(ctx); 351 /* 352 * this is an overkill, but ensures we don't leave 353 * the ctx on the aio_wq 354 */ 355 flush_workqueue(aio_wq); 356 357 if (1 != atomic_read(&ctx->users)) 358 printk(KERN_DEBUG 359 "exit_aio:ioctx still alive: %d %d %d\n", 360 atomic_read(&ctx->users), ctx->dead, 361 ctx->reqs_active); 362 put_ioctx(ctx); 363 ctx = next; 364 } 365 } 366 367 /* __put_ioctx 368 * Called when the last user of an aio context has gone away, 369 * and the struct needs to be freed. 370 */ 371 void fastcall __put_ioctx(struct kioctx *ctx) 372 { 373 unsigned nr_events = ctx->max_reqs; 374 375 BUG_ON(ctx->reqs_active); 376 377 cancel_delayed_work(&ctx->wq); 378 flush_workqueue(aio_wq); 379 aio_free_ring(ctx); 380 mmdrop(ctx->mm); 381 ctx->mm = NULL; 382 pr_debug("__put_ioctx: freeing %p\n", ctx); 383 kmem_cache_free(kioctx_cachep, ctx); 384 385 if (nr_events) { 386 spin_lock(&aio_nr_lock); 387 BUG_ON(aio_nr - nr_events > aio_nr); 388 aio_nr -= nr_events; 389 spin_unlock(&aio_nr_lock); 390 } 391 } 392 393 /* aio_get_req 394 * Allocate a slot for an aio request. Increments the users count 395 * of the kioctx so that the kioctx stays around until all requests are 396 * complete. Returns NULL if no requests are free. 397 * 398 * Returns with kiocb->users set to 2. The io submit code path holds 399 * an extra reference while submitting the i/o. 400 * This prevents races between the aio code path referencing the 401 * req (after submitting it) and aio_complete() freeing the req. 402 */ 403 static struct kiocb *FASTCALL(__aio_get_req(struct kioctx *ctx)); 404 static struct kiocb fastcall *__aio_get_req(struct kioctx *ctx) 405 { 406 struct kiocb *req = NULL; 407 struct aio_ring *ring; 408 int okay = 0; 409 410 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL); 411 if (unlikely(!req)) 412 return NULL; 413 414 req->ki_flags = 0; 415 req->ki_users = 2; 416 req->ki_key = 0; 417 req->ki_ctx = ctx; 418 req->ki_cancel = NULL; 419 req->ki_retry = NULL; 420 req->ki_dtor = NULL; 421 req->private = NULL; 422 req->ki_iovec = NULL; 423 INIT_LIST_HEAD(&req->ki_run_list); 424 425 /* Check if the completion queue has enough free space to 426 * accept an event from this io. 427 */ 428 spin_lock_irq(&ctx->ctx_lock); 429 ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0); 430 if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) { 431 list_add(&req->ki_list, &ctx->active_reqs); 432 ctx->reqs_active++; 433 okay = 1; 434 } 435 kunmap_atomic(ring, KM_USER0); 436 spin_unlock_irq(&ctx->ctx_lock); 437 438 if (!okay) { 439 kmem_cache_free(kiocb_cachep, req); 440 req = NULL; 441 } 442 443 return req; 444 } 445 446 static inline struct kiocb *aio_get_req(struct kioctx *ctx) 447 { 448 struct kiocb *req; 449 /* Handle a potential starvation case -- should be exceedingly rare as 450 * requests will be stuck on fput_head only if the aio_fput_routine is 451 * delayed and the requests were the last user of the struct file. 452 */ 453 req = __aio_get_req(ctx); 454 if (unlikely(NULL == req)) { 455 aio_fput_routine(NULL); 456 req = __aio_get_req(ctx); 457 } 458 return req; 459 } 460 461 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req) 462 { 463 assert_spin_locked(&ctx->ctx_lock); 464 465 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 info = &ctx->ring_info; 950 951 /* add a completion event to the ring buffer. 952 * must be done holding ctx->ctx_lock to prevent 953 * other code from messing with the tail 954 * pointer since we might be called from irq 955 * context. 956 */ 957 spin_lock_irqsave(&ctx->ctx_lock, flags); 958 959 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list)) 960 list_del_init(&iocb->ki_run_list); 961 962 /* 963 * cancelled requests don't get events, userland was given one 964 * when the event got cancelled. 965 */ 966 if (kiocbIsCancelled(iocb)) 967 goto put_rq; 968 969 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1); 970 971 tail = info->tail; 972 event = aio_ring_event(info, tail, KM_IRQ0); 973 if (++tail >= info->nr) 974 tail = 0; 975 976 event->obj = (u64)(unsigned long)iocb->ki_obj.user; 977 event->data = iocb->ki_user_data; 978 event->res = res; 979 event->res2 = res2; 980 981 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n", 982 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data, 983 res, res2); 984 985 /* after flagging the request as done, we 986 * must never even look at it again 987 */ 988 smp_wmb(); /* make event visible before updating tail */ 989 990 info->tail = tail; 991 ring->tail = tail; 992 993 put_aio_ring_event(event, KM_IRQ0); 994 kunmap_atomic(ring, KM_IRQ1); 995 996 pr_debug("added to ring %p at [%lu]\n", iocb, tail); 997 put_rq: 998 /* everything turned out well, dispose of the aiocb. */ 999 ret = __aio_put_req(ctx, iocb); 1000 1001 if (waitqueue_active(&ctx->wait)) 1002 wake_up(&ctx->wait); 1003 1004 spin_unlock_irqrestore(&ctx->ctx_lock, flags); 1005 return ret; 1006 } 1007 1008 /* aio_read_evt 1009 * Pull an event off of the ioctx's event ring. Returns the number of 1010 * events fetched (0 or 1 ;-) 1011 * FIXME: make this use cmpxchg. 1012 * TODO: make the ringbuffer user mmap()able (requires FIXME). 1013 */ 1014 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent) 1015 { 1016 struct aio_ring_info *info = &ioctx->ring_info; 1017 struct aio_ring *ring; 1018 unsigned long head; 1019 int ret = 0; 1020 1021 ring = kmap_atomic(info->ring_pages[0], KM_USER0); 1022 dprintk("in aio_read_evt h%lu t%lu m%lu\n", 1023 (unsigned long)ring->head, (unsigned long)ring->tail, 1024 (unsigned long)ring->nr); 1025 1026 if (ring->head == ring->tail) 1027 goto out; 1028 1029 spin_lock(&info->ring_lock); 1030 1031 head = ring->head % info->nr; 1032 if (head != ring->tail) { 1033 struct io_event *evp = aio_ring_event(info, head, KM_USER1); 1034 *ent = *evp; 1035 head = (head + 1) % info->nr; 1036 smp_mb(); /* finish reading the event before updatng the head */ 1037 ring->head = head; 1038 ret = 1; 1039 put_aio_ring_event(evp, KM_USER1); 1040 } 1041 spin_unlock(&info->ring_lock); 1042 1043 out: 1044 kunmap_atomic(ring, KM_USER0); 1045 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret, 1046 (unsigned long)ring->head, (unsigned long)ring->tail); 1047 return ret; 1048 } 1049 1050 struct aio_timeout { 1051 struct timer_list timer; 1052 int timed_out; 1053 struct task_struct *p; 1054 }; 1055 1056 static void timeout_func(unsigned long data) 1057 { 1058 struct aio_timeout *to = (struct aio_timeout *)data; 1059 1060 to->timed_out = 1; 1061 wake_up_process(to->p); 1062 } 1063 1064 static inline void init_timeout(struct aio_timeout *to) 1065 { 1066 init_timer(&to->timer); 1067 to->timer.data = (unsigned long)to; 1068 to->timer.function = timeout_func; 1069 to->timed_out = 0; 1070 to->p = current; 1071 } 1072 1073 static inline void set_timeout(long start_jiffies, struct aio_timeout *to, 1074 const struct timespec *ts) 1075 { 1076 to->timer.expires = start_jiffies + timespec_to_jiffies(ts); 1077 if (time_after(to->timer.expires, jiffies)) 1078 add_timer(&to->timer); 1079 else 1080 to->timed_out = 1; 1081 } 1082 1083 static inline void clear_timeout(struct aio_timeout *to) 1084 { 1085 del_singleshot_timer_sync(&to->timer); 1086 } 1087 1088 static int read_events(struct kioctx *ctx, 1089 long min_nr, long nr, 1090 struct io_event __user *event, 1091 struct timespec __user *timeout) 1092 { 1093 long start_jiffies = jiffies; 1094 struct task_struct *tsk = current; 1095 DECLARE_WAITQUEUE(wait, tsk); 1096 int ret; 1097 int i = 0; 1098 struct io_event ent; 1099 struct aio_timeout to; 1100 int retry = 0; 1101 1102 /* needed to zero any padding within an entry (there shouldn't be 1103 * any, but C is fun! 1104 */ 1105 memset(&ent, 0, sizeof(ent)); 1106 retry: 1107 ret = 0; 1108 while (likely(i < nr)) { 1109 ret = aio_read_evt(ctx, &ent); 1110 if (unlikely(ret <= 0)) 1111 break; 1112 1113 dprintk("read event: %Lx %Lx %Lx %Lx\n", 1114 ent.data, ent.obj, ent.res, ent.res2); 1115 1116 /* Could we split the check in two? */ 1117 ret = -EFAULT; 1118 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) { 1119 dprintk("aio: lost an event due to EFAULT.\n"); 1120 break; 1121 } 1122 ret = 0; 1123 1124 /* Good, event copied to userland, update counts. */ 1125 event ++; 1126 i ++; 1127 } 1128 1129 if (min_nr <= i) 1130 return i; 1131 if (ret) 1132 return ret; 1133 1134 /* End fast path */ 1135 1136 /* racey check, but it gets redone */ 1137 if (!retry && unlikely(!list_empty(&ctx->run_list))) { 1138 retry = 1; 1139 aio_run_all_iocbs(ctx); 1140 goto retry; 1141 } 1142 1143 init_timeout(&to); 1144 if (timeout) { 1145 struct timespec ts; 1146 ret = -EFAULT; 1147 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts)))) 1148 goto out; 1149 1150 set_timeout(start_jiffies, &to, &ts); 1151 } 1152 1153 while (likely(i < nr)) { 1154 add_wait_queue_exclusive(&ctx->wait, &wait); 1155 do { 1156 set_task_state(tsk, TASK_INTERRUPTIBLE); 1157 ret = aio_read_evt(ctx, &ent); 1158 if (ret) 1159 break; 1160 if (min_nr <= i) 1161 break; 1162 ret = 0; 1163 if (to.timed_out) /* Only check after read evt */ 1164 break; 1165 schedule(); 1166 if (signal_pending(tsk)) { 1167 ret = -EINTR; 1168 break; 1169 } 1170 /*ret = aio_read_evt(ctx, &ent);*/ 1171 } while (1) ; 1172 1173 set_task_state(tsk, TASK_RUNNING); 1174 remove_wait_queue(&ctx->wait, &wait); 1175 1176 if (unlikely(ret <= 0)) 1177 break; 1178 1179 ret = -EFAULT; 1180 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) { 1181 dprintk("aio: lost an event due to EFAULT.\n"); 1182 break; 1183 } 1184 1185 /* Good, event copied to userland, update counts. */ 1186 event ++; 1187 i ++; 1188 } 1189 1190 if (timeout) 1191 clear_timeout(&to); 1192 out: 1193 return i ? i : ret; 1194 } 1195 1196 /* Take an ioctx and remove it from the list of ioctx's. Protects 1197 * against races with itself via ->dead. 1198 */ 1199 static void io_destroy(struct kioctx *ioctx) 1200 { 1201 struct mm_struct *mm = current->mm; 1202 struct kioctx **tmp; 1203 int was_dead; 1204 1205 /* delete the entry from the list is someone else hasn't already */ 1206 write_lock(&mm->ioctx_list_lock); 1207 was_dead = ioctx->dead; 1208 ioctx->dead = 1; 1209 for (tmp = &mm->ioctx_list; *tmp && *tmp != ioctx; 1210 tmp = &(*tmp)->next) 1211 ; 1212 if (*tmp) 1213 *tmp = ioctx->next; 1214 write_unlock(&mm->ioctx_list_lock); 1215 1216 dprintk("aio_release(%p)\n", ioctx); 1217 if (likely(!was_dead)) 1218 put_ioctx(ioctx); /* twice for the list */ 1219 1220 aio_cancel_all(ioctx); 1221 wait_for_all_aios(ioctx); 1222 put_ioctx(ioctx); /* once for the lookup */ 1223 } 1224 1225 /* sys_io_setup: 1226 * Create an aio_context capable of receiving at least nr_events. 1227 * ctxp must not point to an aio_context that already exists, and 1228 * must be initialized to 0 prior to the call. On successful 1229 * creation of the aio_context, *ctxp is filled in with the resulting 1230 * handle. May fail with -EINVAL if *ctxp is not initialized, 1231 * if the specified nr_events exceeds internal limits. May fail 1232 * with -EAGAIN if the specified nr_events exceeds the user's limit 1233 * of available events. May fail with -ENOMEM if insufficient kernel 1234 * resources are available. May fail with -EFAULT if an invalid 1235 * pointer is passed for ctxp. Will fail with -ENOSYS if not 1236 * implemented. 1237 */ 1238 asmlinkage long sys_io_setup(unsigned nr_events, aio_context_t __user *ctxp) 1239 { 1240 struct kioctx *ioctx = NULL; 1241 unsigned long ctx; 1242 long ret; 1243 1244 ret = get_user(ctx, ctxp); 1245 if (unlikely(ret)) 1246 goto out; 1247 1248 ret = -EINVAL; 1249 if (unlikely(ctx || nr_events == 0)) { 1250 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n", 1251 ctx, nr_events); 1252 goto out; 1253 } 1254 1255 ioctx = ioctx_alloc(nr_events); 1256 ret = PTR_ERR(ioctx); 1257 if (!IS_ERR(ioctx)) { 1258 ret = put_user(ioctx->user_id, ctxp); 1259 if (!ret) 1260 return 0; 1261 1262 get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */ 1263 io_destroy(ioctx); 1264 } 1265 1266 out: 1267 return ret; 1268 } 1269 1270 /* sys_io_destroy: 1271 * Destroy the aio_context specified. May cancel any outstanding 1272 * AIOs and block on completion. Will fail with -ENOSYS if not 1273 * implemented. May fail with -EFAULT if the context pointed to 1274 * is invalid. 1275 */ 1276 asmlinkage long sys_io_destroy(aio_context_t ctx) 1277 { 1278 struct kioctx *ioctx = lookup_ioctx(ctx); 1279 if (likely(NULL != ioctx)) { 1280 io_destroy(ioctx); 1281 return 0; 1282 } 1283 pr_debug("EINVAL: io_destroy: invalid context id\n"); 1284 return -EINVAL; 1285 } 1286 1287 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret) 1288 { 1289 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg]; 1290 1291 BUG_ON(ret <= 0); 1292 1293 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) { 1294 ssize_t this = min((ssize_t)iov->iov_len, ret); 1295 iov->iov_base += this; 1296 iov->iov_len -= this; 1297 iocb->ki_left -= this; 1298 ret -= this; 1299 if (iov->iov_len == 0) { 1300 iocb->ki_cur_seg++; 1301 iov++; 1302 } 1303 } 1304 1305 /* the caller should not have done more io than what fit in 1306 * the remaining iovecs */ 1307 BUG_ON(ret > 0 && iocb->ki_left == 0); 1308 } 1309 1310 static ssize_t aio_rw_vect_retry(struct kiocb *iocb) 1311 { 1312 struct file *file = iocb->ki_filp; 1313 struct address_space *mapping = file->f_mapping; 1314 struct inode *inode = mapping->host; 1315 ssize_t (*rw_op)(struct kiocb *, const struct iovec *, 1316 unsigned long, loff_t); 1317 ssize_t ret = 0; 1318 unsigned short opcode; 1319 1320 if ((iocb->ki_opcode == IOCB_CMD_PREADV) || 1321 (iocb->ki_opcode == IOCB_CMD_PREAD)) { 1322 rw_op = file->f_op->aio_read; 1323 opcode = IOCB_CMD_PREADV; 1324 } else { 1325 rw_op = file->f_op->aio_write; 1326 opcode = IOCB_CMD_PWRITEV; 1327 } 1328 1329 do { 1330 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg], 1331 iocb->ki_nr_segs - iocb->ki_cur_seg, 1332 iocb->ki_pos); 1333 if (ret > 0) 1334 aio_advance_iovec(iocb, ret); 1335 1336 /* retry all partial writes. retry partial reads as long as its a 1337 * regular file. */ 1338 } while (ret > 0 && iocb->ki_left > 0 && 1339 (opcode == IOCB_CMD_PWRITEV || 1340 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode)))); 1341 1342 /* This means we must have transferred all that we could */ 1343 /* No need to retry anymore */ 1344 if ((ret == 0) || (iocb->ki_left == 0)) 1345 ret = iocb->ki_nbytes - iocb->ki_left; 1346 1347 return ret; 1348 } 1349 1350 static ssize_t aio_fdsync(struct kiocb *iocb) 1351 { 1352 struct file *file = iocb->ki_filp; 1353 ssize_t ret = -EINVAL; 1354 1355 if (file->f_op->aio_fsync) 1356 ret = file->f_op->aio_fsync(iocb, 1); 1357 return ret; 1358 } 1359 1360 static ssize_t aio_fsync(struct kiocb *iocb) 1361 { 1362 struct file *file = iocb->ki_filp; 1363 ssize_t ret = -EINVAL; 1364 1365 if (file->f_op->aio_fsync) 1366 ret = file->f_op->aio_fsync(iocb, 0); 1367 return ret; 1368 } 1369 1370 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb) 1371 { 1372 ssize_t ret; 1373 1374 ret = rw_copy_check_uvector(type, (struct iovec __user *)kiocb->ki_buf, 1375 kiocb->ki_nbytes, 1, 1376 &kiocb->ki_inline_vec, &kiocb->ki_iovec); 1377 if (ret < 0) 1378 goto out; 1379 1380 kiocb->ki_nr_segs = kiocb->ki_nbytes; 1381 kiocb->ki_cur_seg = 0; 1382 /* ki_nbytes/left now reflect bytes instead of segs */ 1383 kiocb->ki_nbytes = ret; 1384 kiocb->ki_left = ret; 1385 1386 ret = 0; 1387 out: 1388 return ret; 1389 } 1390 1391 static ssize_t aio_setup_single_vector(struct kiocb *kiocb) 1392 { 1393 kiocb->ki_iovec = &kiocb->ki_inline_vec; 1394 kiocb->ki_iovec->iov_base = kiocb->ki_buf; 1395 kiocb->ki_iovec->iov_len = kiocb->ki_left; 1396 kiocb->ki_nr_segs = 1; 1397 kiocb->ki_cur_seg = 0; 1398 return 0; 1399 } 1400 1401 /* 1402 * aio_setup_iocb: 1403 * Performs the initial checks and aio retry method 1404 * setup for the kiocb at the time of io submission. 1405 */ 1406 static ssize_t aio_setup_iocb(struct kiocb *kiocb) 1407 { 1408 struct file *file = kiocb->ki_filp; 1409 ssize_t ret = 0; 1410 1411 switch (kiocb->ki_opcode) { 1412 case IOCB_CMD_PREAD: 1413 ret = -EBADF; 1414 if (unlikely(!(file->f_mode & FMODE_READ))) 1415 break; 1416 ret = -EFAULT; 1417 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf, 1418 kiocb->ki_left))) 1419 break; 1420 ret = security_file_permission(file, MAY_READ); 1421 if (unlikely(ret)) 1422 break; 1423 ret = aio_setup_single_vector(kiocb); 1424 if (ret) 1425 break; 1426 ret = -EINVAL; 1427 if (file->f_op->aio_read) 1428 kiocb->ki_retry = aio_rw_vect_retry; 1429 break; 1430 case IOCB_CMD_PWRITE: 1431 ret = -EBADF; 1432 if (unlikely(!(file->f_mode & FMODE_WRITE))) 1433 break; 1434 ret = -EFAULT; 1435 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf, 1436 kiocb->ki_left))) 1437 break; 1438 ret = security_file_permission(file, MAY_WRITE); 1439 if (unlikely(ret)) 1440 break; 1441 ret = aio_setup_single_vector(kiocb); 1442 if (ret) 1443 break; 1444 ret = -EINVAL; 1445 if (file->f_op->aio_write) 1446 kiocb->ki_retry = aio_rw_vect_retry; 1447 break; 1448 case IOCB_CMD_PREADV: 1449 ret = -EBADF; 1450 if (unlikely(!(file->f_mode & FMODE_READ))) 1451 break; 1452 ret = security_file_permission(file, MAY_READ); 1453 if (unlikely(ret)) 1454 break; 1455 ret = aio_setup_vectored_rw(READ, kiocb); 1456 if (ret) 1457 break; 1458 ret = -EINVAL; 1459 if (file->f_op->aio_read) 1460 kiocb->ki_retry = aio_rw_vect_retry; 1461 break; 1462 case IOCB_CMD_PWRITEV: 1463 ret = -EBADF; 1464 if (unlikely(!(file->f_mode & FMODE_WRITE))) 1465 break; 1466 ret = security_file_permission(file, MAY_WRITE); 1467 if (unlikely(ret)) 1468 break; 1469 ret = aio_setup_vectored_rw(WRITE, kiocb); 1470 if (ret) 1471 break; 1472 ret = -EINVAL; 1473 if (file->f_op->aio_write) 1474 kiocb->ki_retry = aio_rw_vect_retry; 1475 break; 1476 case IOCB_CMD_FDSYNC: 1477 ret = -EINVAL; 1478 if (file->f_op->aio_fsync) 1479 kiocb->ki_retry = aio_fdsync; 1480 break; 1481 case IOCB_CMD_FSYNC: 1482 ret = -EINVAL; 1483 if (file->f_op->aio_fsync) 1484 kiocb->ki_retry = aio_fsync; 1485 break; 1486 default: 1487 dprintk("EINVAL: io_submit: no operation provided\n"); 1488 ret = -EINVAL; 1489 } 1490 1491 if (!kiocb->ki_retry) 1492 return ret; 1493 1494 return 0; 1495 } 1496 1497 /* 1498 * aio_wake_function: 1499 * wait queue callback function for aio notification, 1500 * Simply triggers a retry of the operation via kick_iocb. 1501 * 1502 * This callback is specified in the wait queue entry in 1503 * a kiocb (current->io_wait points to this wait queue 1504 * entry when an aio operation executes; it is used 1505 * instead of a synchronous wait when an i/o blocking 1506 * condition is encountered during aio). 1507 * 1508 * Note: 1509 * This routine is executed with the wait queue lock held. 1510 * Since kick_iocb acquires iocb->ctx->ctx_lock, it nests 1511 * the ioctx lock inside the wait queue lock. This is safe 1512 * because this callback isn't used for wait queues which 1513 * are nested inside ioctx lock (i.e. ctx->wait) 1514 */ 1515 static int aio_wake_function(wait_queue_t *wait, unsigned mode, 1516 int sync, void *key) 1517 { 1518 struct kiocb *iocb = container_of(wait, struct kiocb, ki_wait); 1519 1520 list_del_init(&wait->task_list); 1521 kick_iocb(iocb); 1522 return 1; 1523 } 1524 1525 int fastcall io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb, 1526 struct iocb *iocb) 1527 { 1528 struct kiocb *req; 1529 struct file *file; 1530 ssize_t ret; 1531 1532 /* enforce forwards compatibility on users */ 1533 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2 || 1534 iocb->aio_reserved3)) { 1535 pr_debug("EINVAL: io_submit: reserve field set\n"); 1536 return -EINVAL; 1537 } 1538 1539 /* prevent overflows */ 1540 if (unlikely( 1541 (iocb->aio_buf != (unsigned long)iocb->aio_buf) || 1542 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) || 1543 ((ssize_t)iocb->aio_nbytes < 0) 1544 )) { 1545 pr_debug("EINVAL: io_submit: overflow check\n"); 1546 return -EINVAL; 1547 } 1548 1549 file = fget(iocb->aio_fildes); 1550 if (unlikely(!file)) 1551 return -EBADF; 1552 1553 req = aio_get_req(ctx); /* returns with 2 references to req */ 1554 if (unlikely(!req)) { 1555 fput(file); 1556 return -EAGAIN; 1557 } 1558 1559 req->ki_filp = file; 1560 ret = put_user(req->ki_key, &user_iocb->aio_key); 1561 if (unlikely(ret)) { 1562 dprintk("EFAULT: aio_key\n"); 1563 goto out_put_req; 1564 } 1565 1566 req->ki_obj.user = user_iocb; 1567 req->ki_user_data = iocb->aio_data; 1568 req->ki_pos = iocb->aio_offset; 1569 1570 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf; 1571 req->ki_left = req->ki_nbytes = iocb->aio_nbytes; 1572 req->ki_opcode = iocb->aio_lio_opcode; 1573 init_waitqueue_func_entry(&req->ki_wait, aio_wake_function); 1574 INIT_LIST_HEAD(&req->ki_wait.task_list); 1575 1576 ret = aio_setup_iocb(req); 1577 1578 if (ret) 1579 goto out_put_req; 1580 1581 spin_lock_irq(&ctx->ctx_lock); 1582 aio_run_iocb(req); 1583 if (!list_empty(&ctx->run_list)) { 1584 /* drain the run list */ 1585 while (__aio_run_iocbs(ctx)) 1586 ; 1587 } 1588 spin_unlock_irq(&ctx->ctx_lock); 1589 aio_put_req(req); /* drop extra ref to req */ 1590 return 0; 1591 1592 out_put_req: 1593 aio_put_req(req); /* drop extra ref to req */ 1594 aio_put_req(req); /* drop i/o ref to req */ 1595 return ret; 1596 } 1597 1598 /* sys_io_submit: 1599 * Queue the nr iocbs pointed to by iocbpp for processing. Returns 1600 * the number of iocbs queued. May return -EINVAL if the aio_context 1601 * specified by ctx_id is invalid, if nr is < 0, if the iocb at 1602 * *iocbpp[0] is not properly initialized, if the operation specified 1603 * is invalid for the file descriptor in the iocb. May fail with 1604 * -EFAULT if any of the data structures point to invalid data. May 1605 * fail with -EBADF if the file descriptor specified in the first 1606 * iocb is invalid. May fail with -EAGAIN if insufficient resources 1607 * are available to queue any iocbs. Will return 0 if nr is 0. Will 1608 * fail with -ENOSYS if not implemented. 1609 */ 1610 asmlinkage long sys_io_submit(aio_context_t ctx_id, long nr, 1611 struct iocb __user * __user *iocbpp) 1612 { 1613 struct kioctx *ctx; 1614 long ret = 0; 1615 int i; 1616 1617 if (unlikely(nr < 0)) 1618 return -EINVAL; 1619 1620 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp))))) 1621 return -EFAULT; 1622 1623 ctx = lookup_ioctx(ctx_id); 1624 if (unlikely(!ctx)) { 1625 pr_debug("EINVAL: io_submit: invalid context id\n"); 1626 return -EINVAL; 1627 } 1628 1629 /* 1630 * AKPM: should this return a partial result if some of the IOs were 1631 * successfully submitted? 1632 */ 1633 for (i=0; i<nr; i++) { 1634 struct iocb __user *user_iocb; 1635 struct iocb tmp; 1636 1637 if (unlikely(__get_user(user_iocb, iocbpp + i))) { 1638 ret = -EFAULT; 1639 break; 1640 } 1641 1642 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) { 1643 ret = -EFAULT; 1644 break; 1645 } 1646 1647 ret = io_submit_one(ctx, user_iocb, &tmp); 1648 if (ret) 1649 break; 1650 } 1651 1652 put_ioctx(ctx); 1653 return i ? i : ret; 1654 } 1655 1656 /* lookup_kiocb 1657 * Finds a given iocb for cancellation. 1658 */ 1659 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, 1660 u32 key) 1661 { 1662 struct list_head *pos; 1663 1664 assert_spin_locked(&ctx->ctx_lock); 1665 1666 /* TODO: use a hash or array, this sucks. */ 1667 list_for_each(pos, &ctx->active_reqs) { 1668 struct kiocb *kiocb = list_kiocb(pos); 1669 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key) 1670 return kiocb; 1671 } 1672 return NULL; 1673 } 1674 1675 /* sys_io_cancel: 1676 * Attempts to cancel an iocb previously passed to io_submit. If 1677 * the operation is successfully cancelled, the resulting event is 1678 * copied into the memory pointed to by result without being placed 1679 * into the completion queue and 0 is returned. May fail with 1680 * -EFAULT if any of the data structures pointed to are invalid. 1681 * May fail with -EINVAL if aio_context specified by ctx_id is 1682 * invalid. May fail with -EAGAIN if the iocb specified was not 1683 * cancelled. Will fail with -ENOSYS if not implemented. 1684 */ 1685 asmlinkage long sys_io_cancel(aio_context_t ctx_id, struct iocb __user *iocb, 1686 struct io_event __user *result) 1687 { 1688 int (*cancel)(struct kiocb *iocb, struct io_event *res); 1689 struct kioctx *ctx; 1690 struct kiocb *kiocb; 1691 u32 key; 1692 int ret; 1693 1694 ret = get_user(key, &iocb->aio_key); 1695 if (unlikely(ret)) 1696 return -EFAULT; 1697 1698 ctx = lookup_ioctx(ctx_id); 1699 if (unlikely(!ctx)) 1700 return -EINVAL; 1701 1702 spin_lock_irq(&ctx->ctx_lock); 1703 ret = -EAGAIN; 1704 kiocb = lookup_kiocb(ctx, iocb, key); 1705 if (kiocb && kiocb->ki_cancel) { 1706 cancel = kiocb->ki_cancel; 1707 kiocb->ki_users ++; 1708 kiocbSetCancelled(kiocb); 1709 } else 1710 cancel = NULL; 1711 spin_unlock_irq(&ctx->ctx_lock); 1712 1713 if (NULL != cancel) { 1714 struct io_event tmp; 1715 pr_debug("calling cancel\n"); 1716 memset(&tmp, 0, sizeof(tmp)); 1717 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user; 1718 tmp.data = kiocb->ki_user_data; 1719 ret = cancel(kiocb, &tmp); 1720 if (!ret) { 1721 /* Cancellation succeeded -- copy the result 1722 * into the user's buffer. 1723 */ 1724 if (copy_to_user(result, &tmp, sizeof(tmp))) 1725 ret = -EFAULT; 1726 } 1727 } else 1728 ret = -EINVAL; 1729 1730 put_ioctx(ctx); 1731 1732 return ret; 1733 } 1734 1735 /* io_getevents: 1736 * Attempts to read at least min_nr events and up to nr events from 1737 * the completion queue for the aio_context specified by ctx_id. May 1738 * fail with -EINVAL if ctx_id is invalid, if min_nr is out of range, 1739 * if nr is out of range, if when is out of range. May fail with 1740 * -EFAULT if any of the memory specified to is invalid. May return 1741 * 0 or < min_nr if no events are available and the timeout specified 1742 * by when has elapsed, where when == NULL specifies an infinite 1743 * timeout. Note that the timeout pointed to by when is relative and 1744 * will be updated if not NULL and the operation blocks. Will fail 1745 * with -ENOSYS if not implemented. 1746 */ 1747 asmlinkage long sys_io_getevents(aio_context_t ctx_id, 1748 long min_nr, 1749 long nr, 1750 struct io_event __user *events, 1751 struct timespec __user *timeout) 1752 { 1753 struct kioctx *ioctx = lookup_ioctx(ctx_id); 1754 long ret = -EINVAL; 1755 1756 if (likely(ioctx)) { 1757 if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0)) 1758 ret = read_events(ioctx, min_nr, nr, events, timeout); 1759 put_ioctx(ioctx); 1760 } 1761 1762 return ret; 1763 } 1764 1765 __initcall(aio_setup); 1766 1767 EXPORT_SYMBOL(aio_complete); 1768 EXPORT_SYMBOL(aio_put_req); 1769 EXPORT_SYMBOL(wait_on_sync_kiocb); 1770