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 info->mmap_size = 0; 140 aio_free_ring(ctx); 141 return -EAGAIN; 142 } 143 144 dprintk("mmap address: 0x%08lx\n", info->mmap_base); 145 info->nr_pages = get_user_pages(current, ctx->mm, 146 info->mmap_base, nr_pages, 147 1, 0, info->ring_pages, NULL); 148 up_write(&ctx->mm->mmap_sem); 149 150 if (unlikely(info->nr_pages != nr_pages)) { 151 aio_free_ring(ctx); 152 return -EAGAIN; 153 } 154 155 ctx->user_id = info->mmap_base; 156 157 info->nr = nr_events; /* trusted copy */ 158 159 ring = kmap_atomic(info->ring_pages[0], KM_USER0); 160 ring->nr = nr_events; /* user copy */ 161 ring->id = ctx->user_id; 162 ring->head = ring->tail = 0; 163 ring->magic = AIO_RING_MAGIC; 164 ring->compat_features = AIO_RING_COMPAT_FEATURES; 165 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES; 166 ring->header_length = sizeof(struct aio_ring); 167 kunmap_atomic(ring, KM_USER0); 168 169 return 0; 170 } 171 172 173 /* aio_ring_event: returns a pointer to the event at the given index from 174 * kmap_atomic(, km). Release the pointer with put_aio_ring_event(); 175 */ 176 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event)) 177 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event)) 178 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE) 179 180 #define aio_ring_event(info, nr, km) ({ \ 181 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \ 182 struct io_event *__event; \ 183 __event = kmap_atomic( \ 184 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \ 185 __event += pos % AIO_EVENTS_PER_PAGE; \ 186 __event; \ 187 }) 188 189 #define put_aio_ring_event(event, km) do { \ 190 struct io_event *__event = (event); \ 191 (void)__event; \ 192 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \ 193 } while(0) 194 195 /* ioctx_alloc 196 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed. 197 */ 198 static struct kioctx *ioctx_alloc(unsigned nr_events) 199 { 200 struct mm_struct *mm; 201 struct kioctx *ctx; 202 203 /* Prevent overflows */ 204 if ((nr_events > (0x10000000U / sizeof(struct io_event))) || 205 (nr_events > (0x10000000U / sizeof(struct kiocb)))) { 206 pr_debug("ENOMEM: nr_events too high\n"); 207 return ERR_PTR(-EINVAL); 208 } 209 210 if ((unsigned long)nr_events > aio_max_nr) 211 return ERR_PTR(-EAGAIN); 212 213 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL); 214 if (!ctx) 215 return ERR_PTR(-ENOMEM); 216 217 ctx->max_reqs = nr_events; 218 mm = ctx->mm = current->mm; 219 atomic_inc(&mm->mm_count); 220 221 atomic_set(&ctx->users, 1); 222 spin_lock_init(&ctx->ctx_lock); 223 spin_lock_init(&ctx->ring_info.ring_lock); 224 init_waitqueue_head(&ctx->wait); 225 226 INIT_LIST_HEAD(&ctx->active_reqs); 227 INIT_LIST_HEAD(&ctx->run_list); 228 INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler); 229 230 if (aio_setup_ring(ctx) < 0) 231 goto out_freectx; 232 233 /* limit the number of system wide aios */ 234 spin_lock(&aio_nr_lock); 235 if (aio_nr + ctx->max_reqs > aio_max_nr || 236 aio_nr + ctx->max_reqs < aio_nr) 237 ctx->max_reqs = 0; 238 else 239 aio_nr += ctx->max_reqs; 240 spin_unlock(&aio_nr_lock); 241 if (ctx->max_reqs == 0) 242 goto out_cleanup; 243 244 /* now link into global list. kludge. FIXME */ 245 write_lock(&mm->ioctx_list_lock); 246 ctx->next = mm->ioctx_list; 247 mm->ioctx_list = ctx; 248 write_unlock(&mm->ioctx_list_lock); 249 250 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n", 251 ctx, ctx->user_id, current->mm, ctx->ring_info.nr); 252 return ctx; 253 254 out_cleanup: 255 __put_ioctx(ctx); 256 return ERR_PTR(-EAGAIN); 257 258 out_freectx: 259 mmdrop(mm); 260 kmem_cache_free(kioctx_cachep, ctx); 261 ctx = ERR_PTR(-ENOMEM); 262 263 dprintk("aio: error allocating ioctx %p\n", ctx); 264 return ctx; 265 } 266 267 /* aio_cancel_all 268 * Cancels all outstanding aio requests on an aio context. Used 269 * when the processes owning a context have all exited to encourage 270 * the rapid destruction of the kioctx. 271 */ 272 static void aio_cancel_all(struct kioctx *ctx) 273 { 274 int (*cancel)(struct kiocb *, struct io_event *); 275 struct io_event res; 276 spin_lock_irq(&ctx->ctx_lock); 277 ctx->dead = 1; 278 while (!list_empty(&ctx->active_reqs)) { 279 struct list_head *pos = ctx->active_reqs.next; 280 struct kiocb *iocb = list_kiocb(pos); 281 list_del_init(&iocb->ki_list); 282 cancel = iocb->ki_cancel; 283 kiocbSetCancelled(iocb); 284 if (cancel) { 285 iocb->ki_users++; 286 spin_unlock_irq(&ctx->ctx_lock); 287 cancel(iocb, &res); 288 spin_lock_irq(&ctx->ctx_lock); 289 } 290 } 291 spin_unlock_irq(&ctx->ctx_lock); 292 } 293 294 static void wait_for_all_aios(struct kioctx *ctx) 295 { 296 struct task_struct *tsk = current; 297 DECLARE_WAITQUEUE(wait, tsk); 298 299 spin_lock_irq(&ctx->ctx_lock); 300 if (!ctx->reqs_active) 301 goto out; 302 303 add_wait_queue(&ctx->wait, &wait); 304 set_task_state(tsk, TASK_UNINTERRUPTIBLE); 305 while (ctx->reqs_active) { 306 spin_unlock_irq(&ctx->ctx_lock); 307 schedule(); 308 set_task_state(tsk, TASK_UNINTERRUPTIBLE); 309 spin_lock_irq(&ctx->ctx_lock); 310 } 311 __set_task_state(tsk, TASK_RUNNING); 312 remove_wait_queue(&ctx->wait, &wait); 313 314 out: 315 spin_unlock_irq(&ctx->ctx_lock); 316 } 317 318 /* wait_on_sync_kiocb: 319 * Waits on the given sync kiocb to complete. 320 */ 321 ssize_t fastcall wait_on_sync_kiocb(struct kiocb *iocb) 322 { 323 while (iocb->ki_users) { 324 set_current_state(TASK_UNINTERRUPTIBLE); 325 if (!iocb->ki_users) 326 break; 327 schedule(); 328 } 329 __set_current_state(TASK_RUNNING); 330 return iocb->ki_user_data; 331 } 332 333 /* exit_aio: called when the last user of mm goes away. At this point, 334 * there is no way for any new requests to be submited or any of the 335 * io_* syscalls to be called on the context. However, there may be 336 * outstanding requests which hold references to the context; as they 337 * go away, they will call put_ioctx and release any pinned memory 338 * associated with the request (held via struct page * references). 339 */ 340 void fastcall exit_aio(struct mm_struct *mm) 341 { 342 struct kioctx *ctx = mm->ioctx_list; 343 mm->ioctx_list = NULL; 344 while (ctx) { 345 struct kioctx *next = ctx->next; 346 ctx->next = NULL; 347 aio_cancel_all(ctx); 348 349 wait_for_all_aios(ctx); 350 /* 351 * this is an overkill, but ensures we don't leave 352 * the ctx on the aio_wq 353 */ 354 flush_workqueue(aio_wq); 355 356 if (1 != atomic_read(&ctx->users)) 357 printk(KERN_DEBUG 358 "exit_aio:ioctx still alive: %d %d %d\n", 359 atomic_read(&ctx->users), ctx->dead, 360 ctx->reqs_active); 361 put_ioctx(ctx); 362 ctx = next; 363 } 364 } 365 366 /* __put_ioctx 367 * Called when the last user of an aio context has gone away, 368 * and the struct needs to be freed. 369 */ 370 void fastcall __put_ioctx(struct kioctx *ctx) 371 { 372 unsigned nr_events = ctx->max_reqs; 373 374 BUG_ON(ctx->reqs_active); 375 376 cancel_delayed_work(&ctx->wq); 377 flush_workqueue(aio_wq); 378 aio_free_ring(ctx); 379 mmdrop(ctx->mm); 380 ctx->mm = NULL; 381 pr_debug("__put_ioctx: freeing %p\n", ctx); 382 kmem_cache_free(kioctx_cachep, ctx); 383 384 if (nr_events) { 385 spin_lock(&aio_nr_lock); 386 BUG_ON(aio_nr - nr_events > aio_nr); 387 aio_nr -= nr_events; 388 spin_unlock(&aio_nr_lock); 389 } 390 } 391 392 /* aio_get_req 393 * Allocate a slot for an aio request. Increments the users count 394 * of the kioctx so that the kioctx stays around until all requests are 395 * complete. Returns NULL if no requests are free. 396 * 397 * Returns with kiocb->users set to 2. The io submit code path holds 398 * an extra reference while submitting the i/o. 399 * This prevents races between the aio code path referencing the 400 * req (after submitting it) and aio_complete() freeing the req. 401 */ 402 static struct kiocb *FASTCALL(__aio_get_req(struct kioctx *ctx)); 403 static struct kiocb fastcall *__aio_get_req(struct kioctx *ctx) 404 { 405 struct kiocb *req = NULL; 406 struct aio_ring *ring; 407 int okay = 0; 408 409 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL); 410 if (unlikely(!req)) 411 return NULL; 412 413 req->ki_flags = 0; 414 req->ki_users = 2; 415 req->ki_key = 0; 416 req->ki_ctx = ctx; 417 req->ki_cancel = NULL; 418 req->ki_retry = NULL; 419 req->ki_dtor = NULL; 420 req->private = NULL; 421 req->ki_iovec = NULL; 422 INIT_LIST_HEAD(&req->ki_run_list); 423 424 /* Check if the completion queue has enough free space to 425 * accept an event from this io. 426 */ 427 spin_lock_irq(&ctx->ctx_lock); 428 ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0); 429 if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) { 430 list_add(&req->ki_list, &ctx->active_reqs); 431 ctx->reqs_active++; 432 okay = 1; 433 } 434 kunmap_atomic(ring, KM_USER0); 435 spin_unlock_irq(&ctx->ctx_lock); 436 437 if (!okay) { 438 kmem_cache_free(kiocb_cachep, req); 439 req = NULL; 440 } 441 442 return req; 443 } 444 445 static inline struct kiocb *aio_get_req(struct kioctx *ctx) 446 { 447 struct kiocb *req; 448 /* Handle a potential starvation case -- should be exceedingly rare as 449 * requests will be stuck on fput_head only if the aio_fput_routine is 450 * delayed and the requests were the last user of the struct file. 451 */ 452 req = __aio_get_req(ctx); 453 if (unlikely(NULL == req)) { 454 aio_fput_routine(NULL); 455 req = __aio_get_req(ctx); 456 } 457 return req; 458 } 459 460 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req) 461 { 462 assert_spin_locked(&ctx->ctx_lock); 463 464 if (req->ki_dtor) 465 req->ki_dtor(req); 466 if (req->ki_iovec != &req->ki_inline_vec) 467 kfree(req->ki_iovec); 468 kmem_cache_free(kiocb_cachep, req); 469 ctx->reqs_active--; 470 471 if (unlikely(!ctx->reqs_active && ctx->dead)) 472 wake_up(&ctx->wait); 473 } 474 475 static void aio_fput_routine(struct work_struct *data) 476 { 477 spin_lock_irq(&fput_lock); 478 while (likely(!list_empty(&fput_head))) { 479 struct kiocb *req = list_kiocb(fput_head.next); 480 struct kioctx *ctx = req->ki_ctx; 481 482 list_del(&req->ki_list); 483 spin_unlock_irq(&fput_lock); 484 485 /* Complete the fput */ 486 __fput(req->ki_filp); 487 488 /* Link the iocb into the context's free list */ 489 spin_lock_irq(&ctx->ctx_lock); 490 really_put_req(ctx, req); 491 spin_unlock_irq(&ctx->ctx_lock); 492 493 put_ioctx(ctx); 494 spin_lock_irq(&fput_lock); 495 } 496 spin_unlock_irq(&fput_lock); 497 } 498 499 /* __aio_put_req 500 * Returns true if this put was the last user of the request. 501 */ 502 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req) 503 { 504 dprintk(KERN_DEBUG "aio_put(%p): f_count=%d\n", 505 req, atomic_read(&req->ki_filp->f_count)); 506 507 assert_spin_locked(&ctx->ctx_lock); 508 509 req->ki_users --; 510 BUG_ON(req->ki_users < 0); 511 if (likely(req->ki_users)) 512 return 0; 513 list_del(&req->ki_list); /* remove from active_reqs */ 514 req->ki_cancel = NULL; 515 req->ki_retry = NULL; 516 517 /* Must be done under the lock to serialise against cancellation. 518 * Call this aio_fput as it duplicates fput via the fput_work. 519 */ 520 if (unlikely(atomic_dec_and_test(&req->ki_filp->f_count))) { 521 get_ioctx(ctx); 522 spin_lock(&fput_lock); 523 list_add(&req->ki_list, &fput_head); 524 spin_unlock(&fput_lock); 525 queue_work(aio_wq, &fput_work); 526 } else 527 really_put_req(ctx, req); 528 return 1; 529 } 530 531 /* aio_put_req 532 * Returns true if this put was the last user of the kiocb, 533 * false if the request is still in use. 534 */ 535 int fastcall aio_put_req(struct kiocb *req) 536 { 537 struct kioctx *ctx = req->ki_ctx; 538 int ret; 539 spin_lock_irq(&ctx->ctx_lock); 540 ret = __aio_put_req(ctx, req); 541 spin_unlock_irq(&ctx->ctx_lock); 542 return ret; 543 } 544 545 /* Lookup an ioctx id. ioctx_list is lockless for reads. 546 * FIXME: this is O(n) and is only suitable for development. 547 */ 548 struct kioctx *lookup_ioctx(unsigned long ctx_id) 549 { 550 struct kioctx *ioctx; 551 struct mm_struct *mm; 552 553 mm = current->mm; 554 read_lock(&mm->ioctx_list_lock); 555 for (ioctx = mm->ioctx_list; ioctx; ioctx = ioctx->next) 556 if (likely(ioctx->user_id == ctx_id && !ioctx->dead)) { 557 get_ioctx(ioctx); 558 break; 559 } 560 read_unlock(&mm->ioctx_list_lock); 561 562 return ioctx; 563 } 564 565 /* 566 * use_mm 567 * Makes the calling kernel thread take on the specified 568 * mm context. 569 * Called by the retry thread execute retries within the 570 * iocb issuer's mm context, so that copy_from/to_user 571 * operations work seamlessly for aio. 572 * (Note: this routine is intended to be called only 573 * from a kernel thread context) 574 */ 575 static void use_mm(struct mm_struct *mm) 576 { 577 struct mm_struct *active_mm; 578 struct task_struct *tsk = current; 579 580 task_lock(tsk); 581 tsk->flags |= PF_BORROWED_MM; 582 active_mm = tsk->active_mm; 583 atomic_inc(&mm->mm_count); 584 tsk->mm = mm; 585 tsk->active_mm = mm; 586 /* 587 * Note that on UML this *requires* PF_BORROWED_MM to be set, otherwise 588 * it won't work. Update it accordingly if you change it here 589 */ 590 switch_mm(active_mm, mm, tsk); 591 task_unlock(tsk); 592 593 mmdrop(active_mm); 594 } 595 596 /* 597 * unuse_mm 598 * Reverses the effect of use_mm, i.e. releases the 599 * specified mm context which was earlier taken on 600 * by the calling kernel thread 601 * (Note: this routine is intended to be called only 602 * from a kernel thread context) 603 */ 604 static void unuse_mm(struct mm_struct *mm) 605 { 606 struct task_struct *tsk = current; 607 608 task_lock(tsk); 609 tsk->flags &= ~PF_BORROWED_MM; 610 tsk->mm = NULL; 611 /* active_mm is still 'mm' */ 612 enter_lazy_tlb(mm, tsk); 613 task_unlock(tsk); 614 } 615 616 /* 617 * Queue up a kiocb to be retried. Assumes that the kiocb 618 * has already been marked as kicked, and places it on 619 * the retry run list for the corresponding ioctx, if it 620 * isn't already queued. Returns 1 if it actually queued 621 * the kiocb (to tell the caller to activate the work 622 * queue to process it), or 0, if it found that it was 623 * already queued. 624 */ 625 static inline int __queue_kicked_iocb(struct kiocb *iocb) 626 { 627 struct kioctx *ctx = iocb->ki_ctx; 628 629 assert_spin_locked(&ctx->ctx_lock); 630 631 if (list_empty(&iocb->ki_run_list)) { 632 list_add_tail(&iocb->ki_run_list, 633 &ctx->run_list); 634 return 1; 635 } 636 return 0; 637 } 638 639 /* aio_run_iocb 640 * This is the core aio execution routine. It is 641 * invoked both for initial i/o submission and 642 * subsequent retries via the aio_kick_handler. 643 * Expects to be invoked with iocb->ki_ctx->lock 644 * already held. The lock is released and reacquired 645 * as needed during processing. 646 * 647 * Calls the iocb retry method (already setup for the 648 * iocb on initial submission) for operation specific 649 * handling, but takes care of most of common retry 650 * execution details for a given iocb. The retry method 651 * needs to be non-blocking as far as possible, to avoid 652 * holding up other iocbs waiting to be serviced by the 653 * retry kernel thread. 654 * 655 * The trickier parts in this code have to do with 656 * ensuring that only one retry instance is in progress 657 * for a given iocb at any time. Providing that guarantee 658 * simplifies the coding of individual aio operations as 659 * it avoids various potential races. 660 */ 661 static ssize_t aio_run_iocb(struct kiocb *iocb) 662 { 663 struct kioctx *ctx = iocb->ki_ctx; 664 ssize_t (*retry)(struct kiocb *); 665 ssize_t ret; 666 667 if (!(retry = iocb->ki_retry)) { 668 printk("aio_run_iocb: iocb->ki_retry = NULL\n"); 669 return 0; 670 } 671 672 /* 673 * We don't want the next retry iteration for this 674 * operation to start until this one has returned and 675 * updated the iocb state. However, wait_queue functions 676 * can trigger a kick_iocb from interrupt context in the 677 * meantime, indicating that data is available for the next 678 * iteration. We want to remember that and enable the 679 * next retry iteration _after_ we are through with 680 * this one. 681 * 682 * So, in order to be able to register a "kick", but 683 * prevent it from being queued now, we clear the kick 684 * flag, but make the kick code *think* that the iocb is 685 * still on the run list until we are actually done. 686 * When we are done with this iteration, we check if 687 * the iocb was kicked in the meantime and if so, queue 688 * it up afresh. 689 */ 690 691 kiocbClearKicked(iocb); 692 693 /* 694 * This is so that aio_complete knows it doesn't need to 695 * pull the iocb off the run list (We can't just call 696 * INIT_LIST_HEAD because we don't want a kick_iocb to 697 * queue this on the run list yet) 698 */ 699 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL; 700 spin_unlock_irq(&ctx->ctx_lock); 701 702 /* Quit retrying if the i/o has been cancelled */ 703 if (kiocbIsCancelled(iocb)) { 704 ret = -EINTR; 705 aio_complete(iocb, ret, 0); 706 /* must not access the iocb after this */ 707 goto out; 708 } 709 710 /* 711 * Now we are all set to call the retry method in async 712 * context. By setting this thread's io_wait context 713 * to point to the wait queue entry inside the currently 714 * running iocb for the duration of the retry, we ensure 715 * that async notification wakeups are queued by the 716 * operation instead of blocking waits, and when notified, 717 * cause the iocb to be kicked for continuation (through 718 * the aio_wake_function callback). 719 */ 720 BUG_ON(current->io_wait != NULL); 721 current->io_wait = &iocb->ki_wait; 722 ret = retry(iocb); 723 current->io_wait = NULL; 724 725 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) { 726 BUG_ON(!list_empty(&iocb->ki_wait.task_list)); 727 aio_complete(iocb, ret, 0); 728 } 729 out: 730 spin_lock_irq(&ctx->ctx_lock); 731 732 if (-EIOCBRETRY == ret) { 733 /* 734 * OK, now that we are done with this iteration 735 * and know that there is more left to go, 736 * this is where we let go so that a subsequent 737 * "kick" can start the next iteration 738 */ 739 740 /* will make __queue_kicked_iocb succeed from here on */ 741 INIT_LIST_HEAD(&iocb->ki_run_list); 742 /* we must queue the next iteration ourselves, if it 743 * has already been kicked */ 744 if (kiocbIsKicked(iocb)) { 745 __queue_kicked_iocb(iocb); 746 747 /* 748 * __queue_kicked_iocb will always return 1 here, because 749 * iocb->ki_run_list is empty at this point so it should 750 * be safe to unconditionally queue the context into the 751 * work queue. 752 */ 753 aio_queue_work(ctx); 754 } 755 } 756 return ret; 757 } 758 759 /* 760 * __aio_run_iocbs: 761 * Process all pending retries queued on the ioctx 762 * run list. 763 * Assumes it is operating within the aio issuer's mm 764 * context. 765 */ 766 static int __aio_run_iocbs(struct kioctx *ctx) 767 { 768 struct kiocb *iocb; 769 struct list_head run_list; 770 771 assert_spin_locked(&ctx->ctx_lock); 772 773 list_replace_init(&ctx->run_list, &run_list); 774 while (!list_empty(&run_list)) { 775 iocb = list_entry(run_list.next, struct kiocb, 776 ki_run_list); 777 list_del(&iocb->ki_run_list); 778 /* 779 * Hold an extra reference while retrying i/o. 780 */ 781 iocb->ki_users++; /* grab extra reference */ 782 aio_run_iocb(iocb); 783 __aio_put_req(ctx, iocb); 784 } 785 if (!list_empty(&ctx->run_list)) 786 return 1; 787 return 0; 788 } 789 790 static void aio_queue_work(struct kioctx * ctx) 791 { 792 unsigned long timeout; 793 /* 794 * if someone is waiting, get the work started right 795 * away, otherwise, use a longer delay 796 */ 797 smp_mb(); 798 if (waitqueue_active(&ctx->wait)) 799 timeout = 1; 800 else 801 timeout = HZ/10; 802 queue_delayed_work(aio_wq, &ctx->wq, timeout); 803 } 804 805 806 /* 807 * aio_run_iocbs: 808 * Process all pending retries queued on the ioctx 809 * run list. 810 * Assumes it is operating within the aio issuer's mm 811 * context. 812 */ 813 static inline void aio_run_iocbs(struct kioctx *ctx) 814 { 815 int requeue; 816 817 spin_lock_irq(&ctx->ctx_lock); 818 819 requeue = __aio_run_iocbs(ctx); 820 spin_unlock_irq(&ctx->ctx_lock); 821 if (requeue) 822 aio_queue_work(ctx); 823 } 824 825 /* 826 * just like aio_run_iocbs, but keeps running them until 827 * the list stays empty 828 */ 829 static inline void aio_run_all_iocbs(struct kioctx *ctx) 830 { 831 spin_lock_irq(&ctx->ctx_lock); 832 while (__aio_run_iocbs(ctx)) 833 ; 834 spin_unlock_irq(&ctx->ctx_lock); 835 } 836 837 /* 838 * aio_kick_handler: 839 * Work queue handler triggered to process pending 840 * retries on an ioctx. Takes on the aio issuer's 841 * mm context before running the iocbs, so that 842 * copy_xxx_user operates on the issuer's address 843 * space. 844 * Run on aiod's context. 845 */ 846 static void aio_kick_handler(struct work_struct *work) 847 { 848 struct kioctx *ctx = container_of(work, struct kioctx, wq.work); 849 mm_segment_t oldfs = get_fs(); 850 struct mm_struct *mm; 851 int requeue; 852 853 set_fs(USER_DS); 854 use_mm(ctx->mm); 855 spin_lock_irq(&ctx->ctx_lock); 856 requeue =__aio_run_iocbs(ctx); 857 mm = ctx->mm; 858 spin_unlock_irq(&ctx->ctx_lock); 859 unuse_mm(mm); 860 set_fs(oldfs); 861 /* 862 * we're in a worker thread already, don't use queue_delayed_work, 863 */ 864 if (requeue) 865 queue_delayed_work(aio_wq, &ctx->wq, 0); 866 } 867 868 869 /* 870 * Called by kick_iocb to queue the kiocb for retry 871 * and if required activate the aio work queue to process 872 * it 873 */ 874 static void try_queue_kicked_iocb(struct kiocb *iocb) 875 { 876 struct kioctx *ctx = iocb->ki_ctx; 877 unsigned long flags; 878 int run = 0; 879 880 /* We're supposed to be the only path putting the iocb back on the run 881 * list. If we find that the iocb is *back* on a wait queue already 882 * than retry has happened before we could queue the iocb. This also 883 * means that the retry could have completed and freed our iocb, no 884 * good. */ 885 BUG_ON((!list_empty(&iocb->ki_wait.task_list))); 886 887 spin_lock_irqsave(&ctx->ctx_lock, flags); 888 /* set this inside the lock so that we can't race with aio_run_iocb() 889 * testing it and putting the iocb on the run list under the lock */ 890 if (!kiocbTryKick(iocb)) 891 run = __queue_kicked_iocb(iocb); 892 spin_unlock_irqrestore(&ctx->ctx_lock, flags); 893 if (run) 894 aio_queue_work(ctx); 895 } 896 897 /* 898 * kick_iocb: 899 * Called typically from a wait queue callback context 900 * (aio_wake_function) to trigger a retry of the iocb. 901 * The retry is usually executed by aio workqueue 902 * threads (See aio_kick_handler). 903 */ 904 void fastcall kick_iocb(struct kiocb *iocb) 905 { 906 /* sync iocbs are easy: they can only ever be executing from a 907 * single context. */ 908 if (is_sync_kiocb(iocb)) { 909 kiocbSetKicked(iocb); 910 wake_up_process(iocb->ki_obj.tsk); 911 return; 912 } 913 914 try_queue_kicked_iocb(iocb); 915 } 916 EXPORT_SYMBOL(kick_iocb); 917 918 /* aio_complete 919 * Called when the io request on the given iocb is complete. 920 * Returns true if this is the last user of the request. The 921 * only other user of the request can be the cancellation code. 922 */ 923 int fastcall aio_complete(struct kiocb *iocb, long res, long res2) 924 { 925 struct kioctx *ctx = iocb->ki_ctx; 926 struct aio_ring_info *info; 927 struct aio_ring *ring; 928 struct io_event *event; 929 unsigned long flags; 930 unsigned long tail; 931 int ret; 932 933 /* 934 * Special case handling for sync iocbs: 935 * - events go directly into the iocb for fast handling 936 * - the sync task with the iocb in its stack holds the single iocb 937 * ref, no other paths have a way to get another ref 938 * - the sync task helpfully left a reference to itself in the iocb 939 */ 940 if (is_sync_kiocb(iocb)) { 941 BUG_ON(iocb->ki_users != 1); 942 iocb->ki_user_data = res; 943 iocb->ki_users = 0; 944 wake_up_process(iocb->ki_obj.tsk); 945 return 1; 946 } 947 948 info = &ctx->ring_info; 949 950 /* add a completion event to the ring buffer. 951 * must be done holding ctx->ctx_lock to prevent 952 * other code from messing with the tail 953 * pointer since we might be called from irq 954 * context. 955 */ 956 spin_lock_irqsave(&ctx->ctx_lock, flags); 957 958 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list)) 959 list_del_init(&iocb->ki_run_list); 960 961 /* 962 * cancelled requests don't get events, userland was given one 963 * when the event got cancelled. 964 */ 965 if (kiocbIsCancelled(iocb)) 966 goto put_rq; 967 968 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1); 969 970 tail = info->tail; 971 event = aio_ring_event(info, tail, KM_IRQ0); 972 if (++tail >= info->nr) 973 tail = 0; 974 975 event->obj = (u64)(unsigned long)iocb->ki_obj.user; 976 event->data = iocb->ki_user_data; 977 event->res = res; 978 event->res2 = res2; 979 980 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n", 981 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data, 982 res, res2); 983 984 /* after flagging the request as done, we 985 * must never even look at it again 986 */ 987 smp_wmb(); /* make event visible before updating tail */ 988 989 info->tail = tail; 990 ring->tail = tail; 991 992 put_aio_ring_event(event, KM_IRQ0); 993 kunmap_atomic(ring, KM_IRQ1); 994 995 pr_debug("added to ring %p at [%lu]\n", iocb, tail); 996 put_rq: 997 /* everything turned out well, dispose of the aiocb. */ 998 ret = __aio_put_req(ctx, iocb); 999 1000 if (waitqueue_active(&ctx->wait)) 1001 wake_up(&ctx->wait); 1002 1003 spin_unlock_irqrestore(&ctx->ctx_lock, flags); 1004 return ret; 1005 } 1006 1007 /* aio_read_evt 1008 * Pull an event off of the ioctx's event ring. Returns the number of 1009 * events fetched (0 or 1 ;-) 1010 * FIXME: make this use cmpxchg. 1011 * TODO: make the ringbuffer user mmap()able (requires FIXME). 1012 */ 1013 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent) 1014 { 1015 struct aio_ring_info *info = &ioctx->ring_info; 1016 struct aio_ring *ring; 1017 unsigned long head; 1018 int ret = 0; 1019 1020 ring = kmap_atomic(info->ring_pages[0], KM_USER0); 1021 dprintk("in aio_read_evt h%lu t%lu m%lu\n", 1022 (unsigned long)ring->head, (unsigned long)ring->tail, 1023 (unsigned long)ring->nr); 1024 1025 if (ring->head == ring->tail) 1026 goto out; 1027 1028 spin_lock(&info->ring_lock); 1029 1030 head = ring->head % info->nr; 1031 if (head != ring->tail) { 1032 struct io_event *evp = aio_ring_event(info, head, KM_USER1); 1033 *ent = *evp; 1034 head = (head + 1) % info->nr; 1035 smp_mb(); /* finish reading the event before updatng the head */ 1036 ring->head = head; 1037 ret = 1; 1038 put_aio_ring_event(evp, KM_USER1); 1039 } 1040 spin_unlock(&info->ring_lock); 1041 1042 out: 1043 kunmap_atomic(ring, KM_USER0); 1044 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret, 1045 (unsigned long)ring->head, (unsigned long)ring->tail); 1046 return ret; 1047 } 1048 1049 struct aio_timeout { 1050 struct timer_list timer; 1051 int timed_out; 1052 struct task_struct *p; 1053 }; 1054 1055 static void timeout_func(unsigned long data) 1056 { 1057 struct aio_timeout *to = (struct aio_timeout *)data; 1058 1059 to->timed_out = 1; 1060 wake_up_process(to->p); 1061 } 1062 1063 static inline void init_timeout(struct aio_timeout *to) 1064 { 1065 init_timer(&to->timer); 1066 to->timer.data = (unsigned long)to; 1067 to->timer.function = timeout_func; 1068 to->timed_out = 0; 1069 to->p = current; 1070 } 1071 1072 static inline void set_timeout(long start_jiffies, struct aio_timeout *to, 1073 const struct timespec *ts) 1074 { 1075 to->timer.expires = start_jiffies + timespec_to_jiffies(ts); 1076 if (time_after(to->timer.expires, jiffies)) 1077 add_timer(&to->timer); 1078 else 1079 to->timed_out = 1; 1080 } 1081 1082 static inline void clear_timeout(struct aio_timeout *to) 1083 { 1084 del_singleshot_timer_sync(&to->timer); 1085 } 1086 1087 static int read_events(struct kioctx *ctx, 1088 long min_nr, long nr, 1089 struct io_event __user *event, 1090 struct timespec __user *timeout) 1091 { 1092 long start_jiffies = jiffies; 1093 struct task_struct *tsk = current; 1094 DECLARE_WAITQUEUE(wait, tsk); 1095 int ret; 1096 int i = 0; 1097 struct io_event ent; 1098 struct aio_timeout to; 1099 int retry = 0; 1100 1101 /* needed to zero any padding within an entry (there shouldn't be 1102 * any, but C is fun! 1103 */ 1104 memset(&ent, 0, sizeof(ent)); 1105 retry: 1106 ret = 0; 1107 while (likely(i < nr)) { 1108 ret = aio_read_evt(ctx, &ent); 1109 if (unlikely(ret <= 0)) 1110 break; 1111 1112 dprintk("read event: %Lx %Lx %Lx %Lx\n", 1113 ent.data, ent.obj, ent.res, ent.res2); 1114 1115 /* Could we split the check in two? */ 1116 ret = -EFAULT; 1117 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) { 1118 dprintk("aio: lost an event due to EFAULT.\n"); 1119 break; 1120 } 1121 ret = 0; 1122 1123 /* Good, event copied to userland, update counts. */ 1124 event ++; 1125 i ++; 1126 } 1127 1128 if (min_nr <= i) 1129 return i; 1130 if (ret) 1131 return ret; 1132 1133 /* End fast path */ 1134 1135 /* racey check, but it gets redone */ 1136 if (!retry && unlikely(!list_empty(&ctx->run_list))) { 1137 retry = 1; 1138 aio_run_all_iocbs(ctx); 1139 goto retry; 1140 } 1141 1142 init_timeout(&to); 1143 if (timeout) { 1144 struct timespec ts; 1145 ret = -EFAULT; 1146 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts)))) 1147 goto out; 1148 1149 set_timeout(start_jiffies, &to, &ts); 1150 } 1151 1152 while (likely(i < nr)) { 1153 add_wait_queue_exclusive(&ctx->wait, &wait); 1154 do { 1155 set_task_state(tsk, TASK_INTERRUPTIBLE); 1156 ret = aio_read_evt(ctx, &ent); 1157 if (ret) 1158 break; 1159 if (min_nr <= i) 1160 break; 1161 ret = 0; 1162 if (to.timed_out) /* Only check after read evt */ 1163 break; 1164 schedule(); 1165 if (signal_pending(tsk)) { 1166 ret = -EINTR; 1167 break; 1168 } 1169 /*ret = aio_read_evt(ctx, &ent);*/ 1170 } while (1) ; 1171 1172 set_task_state(tsk, TASK_RUNNING); 1173 remove_wait_queue(&ctx->wait, &wait); 1174 1175 if (unlikely(ret <= 0)) 1176 break; 1177 1178 ret = -EFAULT; 1179 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) { 1180 dprintk("aio: lost an event due to EFAULT.\n"); 1181 break; 1182 } 1183 1184 /* Good, event copied to userland, update counts. */ 1185 event ++; 1186 i ++; 1187 } 1188 1189 if (timeout) 1190 clear_timeout(&to); 1191 out: 1192 return i ? i : ret; 1193 } 1194 1195 /* Take an ioctx and remove it from the list of ioctx's. Protects 1196 * against races with itself via ->dead. 1197 */ 1198 static void io_destroy(struct kioctx *ioctx) 1199 { 1200 struct mm_struct *mm = current->mm; 1201 struct kioctx **tmp; 1202 int was_dead; 1203 1204 /* delete the entry from the list is someone else hasn't already */ 1205 write_lock(&mm->ioctx_list_lock); 1206 was_dead = ioctx->dead; 1207 ioctx->dead = 1; 1208 for (tmp = &mm->ioctx_list; *tmp && *tmp != ioctx; 1209 tmp = &(*tmp)->next) 1210 ; 1211 if (*tmp) 1212 *tmp = ioctx->next; 1213 write_unlock(&mm->ioctx_list_lock); 1214 1215 dprintk("aio_release(%p)\n", ioctx); 1216 if (likely(!was_dead)) 1217 put_ioctx(ioctx); /* twice for the list */ 1218 1219 aio_cancel_all(ioctx); 1220 wait_for_all_aios(ioctx); 1221 put_ioctx(ioctx); /* once for the lookup */ 1222 } 1223 1224 /* sys_io_setup: 1225 * Create an aio_context capable of receiving at least nr_events. 1226 * ctxp must not point to an aio_context that already exists, and 1227 * must be initialized to 0 prior to the call. On successful 1228 * creation of the aio_context, *ctxp is filled in with the resulting 1229 * handle. May fail with -EINVAL if *ctxp is not initialized, 1230 * if the specified nr_events exceeds internal limits. May fail 1231 * with -EAGAIN if the specified nr_events exceeds the user's limit 1232 * of available events. May fail with -ENOMEM if insufficient kernel 1233 * resources are available. May fail with -EFAULT if an invalid 1234 * pointer is passed for ctxp. Will fail with -ENOSYS if not 1235 * implemented. 1236 */ 1237 asmlinkage long sys_io_setup(unsigned nr_events, aio_context_t __user *ctxp) 1238 { 1239 struct kioctx *ioctx = NULL; 1240 unsigned long ctx; 1241 long ret; 1242 1243 ret = get_user(ctx, ctxp); 1244 if (unlikely(ret)) 1245 goto out; 1246 1247 ret = -EINVAL; 1248 if (unlikely(ctx || nr_events == 0)) { 1249 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n", 1250 ctx, nr_events); 1251 goto out; 1252 } 1253 1254 ioctx = ioctx_alloc(nr_events); 1255 ret = PTR_ERR(ioctx); 1256 if (!IS_ERR(ioctx)) { 1257 ret = put_user(ioctx->user_id, ctxp); 1258 if (!ret) 1259 return 0; 1260 1261 get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */ 1262 io_destroy(ioctx); 1263 } 1264 1265 out: 1266 return ret; 1267 } 1268 1269 /* sys_io_destroy: 1270 * Destroy the aio_context specified. May cancel any outstanding 1271 * AIOs and block on completion. Will fail with -ENOSYS if not 1272 * implemented. May fail with -EFAULT if the context pointed to 1273 * is invalid. 1274 */ 1275 asmlinkage long sys_io_destroy(aio_context_t ctx) 1276 { 1277 struct kioctx *ioctx = lookup_ioctx(ctx); 1278 if (likely(NULL != ioctx)) { 1279 io_destroy(ioctx); 1280 return 0; 1281 } 1282 pr_debug("EINVAL: io_destroy: invalid context id\n"); 1283 return -EINVAL; 1284 } 1285 1286 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret) 1287 { 1288 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg]; 1289 1290 BUG_ON(ret <= 0); 1291 1292 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) { 1293 ssize_t this = min((ssize_t)iov->iov_len, ret); 1294 iov->iov_base += this; 1295 iov->iov_len -= this; 1296 iocb->ki_left -= this; 1297 ret -= this; 1298 if (iov->iov_len == 0) { 1299 iocb->ki_cur_seg++; 1300 iov++; 1301 } 1302 } 1303 1304 /* the caller should not have done more io than what fit in 1305 * the remaining iovecs */ 1306 BUG_ON(ret > 0 && iocb->ki_left == 0); 1307 } 1308 1309 static ssize_t aio_rw_vect_retry(struct kiocb *iocb) 1310 { 1311 struct file *file = iocb->ki_filp; 1312 struct address_space *mapping = file->f_mapping; 1313 struct inode *inode = mapping->host; 1314 ssize_t (*rw_op)(struct kiocb *, const struct iovec *, 1315 unsigned long, loff_t); 1316 ssize_t ret = 0; 1317 unsigned short opcode; 1318 1319 if ((iocb->ki_opcode == IOCB_CMD_PREADV) || 1320 (iocb->ki_opcode == IOCB_CMD_PREAD)) { 1321 rw_op = file->f_op->aio_read; 1322 opcode = IOCB_CMD_PREADV; 1323 } else { 1324 rw_op = file->f_op->aio_write; 1325 opcode = IOCB_CMD_PWRITEV; 1326 } 1327 1328 do { 1329 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg], 1330 iocb->ki_nr_segs - iocb->ki_cur_seg, 1331 iocb->ki_pos); 1332 if (ret > 0) 1333 aio_advance_iovec(iocb, ret); 1334 1335 /* retry all partial writes. retry partial reads as long as its a 1336 * regular file. */ 1337 } while (ret > 0 && iocb->ki_left > 0 && 1338 (opcode == IOCB_CMD_PWRITEV || 1339 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode)))); 1340 1341 /* This means we must have transferred all that we could */ 1342 /* No need to retry anymore */ 1343 if ((ret == 0) || (iocb->ki_left == 0)) 1344 ret = iocb->ki_nbytes - iocb->ki_left; 1345 1346 return ret; 1347 } 1348 1349 static ssize_t aio_fdsync(struct kiocb *iocb) 1350 { 1351 struct file *file = iocb->ki_filp; 1352 ssize_t ret = -EINVAL; 1353 1354 if (file->f_op->aio_fsync) 1355 ret = file->f_op->aio_fsync(iocb, 1); 1356 return ret; 1357 } 1358 1359 static ssize_t aio_fsync(struct kiocb *iocb) 1360 { 1361 struct file *file = iocb->ki_filp; 1362 ssize_t ret = -EINVAL; 1363 1364 if (file->f_op->aio_fsync) 1365 ret = file->f_op->aio_fsync(iocb, 0); 1366 return ret; 1367 } 1368 1369 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb) 1370 { 1371 ssize_t ret; 1372 1373 ret = rw_copy_check_uvector(type, (struct iovec __user *)kiocb->ki_buf, 1374 kiocb->ki_nbytes, 1, 1375 &kiocb->ki_inline_vec, &kiocb->ki_iovec); 1376 if (ret < 0) 1377 goto out; 1378 1379 kiocb->ki_nr_segs = kiocb->ki_nbytes; 1380 kiocb->ki_cur_seg = 0; 1381 /* ki_nbytes/left now reflect bytes instead of segs */ 1382 kiocb->ki_nbytes = ret; 1383 kiocb->ki_left = ret; 1384 1385 ret = 0; 1386 out: 1387 return ret; 1388 } 1389 1390 static ssize_t aio_setup_single_vector(struct kiocb *kiocb) 1391 { 1392 kiocb->ki_iovec = &kiocb->ki_inline_vec; 1393 kiocb->ki_iovec->iov_base = kiocb->ki_buf; 1394 kiocb->ki_iovec->iov_len = kiocb->ki_left; 1395 kiocb->ki_nr_segs = 1; 1396 kiocb->ki_cur_seg = 0; 1397 return 0; 1398 } 1399 1400 /* 1401 * aio_setup_iocb: 1402 * Performs the initial checks and aio retry method 1403 * setup for the kiocb at the time of io submission. 1404 */ 1405 static ssize_t aio_setup_iocb(struct kiocb *kiocb) 1406 { 1407 struct file *file = kiocb->ki_filp; 1408 ssize_t ret = 0; 1409 1410 switch (kiocb->ki_opcode) { 1411 case IOCB_CMD_PREAD: 1412 ret = -EBADF; 1413 if (unlikely(!(file->f_mode & FMODE_READ))) 1414 break; 1415 ret = -EFAULT; 1416 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf, 1417 kiocb->ki_left))) 1418 break; 1419 ret = security_file_permission(file, MAY_READ); 1420 if (unlikely(ret)) 1421 break; 1422 ret = aio_setup_single_vector(kiocb); 1423 if (ret) 1424 break; 1425 ret = -EINVAL; 1426 if (file->f_op->aio_read) 1427 kiocb->ki_retry = aio_rw_vect_retry; 1428 break; 1429 case IOCB_CMD_PWRITE: 1430 ret = -EBADF; 1431 if (unlikely(!(file->f_mode & FMODE_WRITE))) 1432 break; 1433 ret = -EFAULT; 1434 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf, 1435 kiocb->ki_left))) 1436 break; 1437 ret = security_file_permission(file, MAY_WRITE); 1438 if (unlikely(ret)) 1439 break; 1440 ret = aio_setup_single_vector(kiocb); 1441 if (ret) 1442 break; 1443 ret = -EINVAL; 1444 if (file->f_op->aio_write) 1445 kiocb->ki_retry = aio_rw_vect_retry; 1446 break; 1447 case IOCB_CMD_PREADV: 1448 ret = -EBADF; 1449 if (unlikely(!(file->f_mode & FMODE_READ))) 1450 break; 1451 ret = security_file_permission(file, MAY_READ); 1452 if (unlikely(ret)) 1453 break; 1454 ret = aio_setup_vectored_rw(READ, kiocb); 1455 if (ret) 1456 break; 1457 ret = -EINVAL; 1458 if (file->f_op->aio_read) 1459 kiocb->ki_retry = aio_rw_vect_retry; 1460 break; 1461 case IOCB_CMD_PWRITEV: 1462 ret = -EBADF; 1463 if (unlikely(!(file->f_mode & FMODE_WRITE))) 1464 break; 1465 ret = security_file_permission(file, MAY_WRITE); 1466 if (unlikely(ret)) 1467 break; 1468 ret = aio_setup_vectored_rw(WRITE, kiocb); 1469 if (ret) 1470 break; 1471 ret = -EINVAL; 1472 if (file->f_op->aio_write) 1473 kiocb->ki_retry = aio_rw_vect_retry; 1474 break; 1475 case IOCB_CMD_FDSYNC: 1476 ret = -EINVAL; 1477 if (file->f_op->aio_fsync) 1478 kiocb->ki_retry = aio_fdsync; 1479 break; 1480 case IOCB_CMD_FSYNC: 1481 ret = -EINVAL; 1482 if (file->f_op->aio_fsync) 1483 kiocb->ki_retry = aio_fsync; 1484 break; 1485 default: 1486 dprintk("EINVAL: io_submit: no operation provided\n"); 1487 ret = -EINVAL; 1488 } 1489 1490 if (!kiocb->ki_retry) 1491 return ret; 1492 1493 return 0; 1494 } 1495 1496 /* 1497 * aio_wake_function: 1498 * wait queue callback function for aio notification, 1499 * Simply triggers a retry of the operation via kick_iocb. 1500 * 1501 * This callback is specified in the wait queue entry in 1502 * a kiocb (current->io_wait points to this wait queue 1503 * entry when an aio operation executes; it is used 1504 * instead of a synchronous wait when an i/o blocking 1505 * condition is encountered during aio). 1506 * 1507 * Note: 1508 * This routine is executed with the wait queue lock held. 1509 * Since kick_iocb acquires iocb->ctx->ctx_lock, it nests 1510 * the ioctx lock inside the wait queue lock. This is safe 1511 * because this callback isn't used for wait queues which 1512 * are nested inside ioctx lock (i.e. ctx->wait) 1513 */ 1514 static int aio_wake_function(wait_queue_t *wait, unsigned mode, 1515 int sync, void *key) 1516 { 1517 struct kiocb *iocb = container_of(wait, struct kiocb, ki_wait); 1518 1519 list_del_init(&wait->task_list); 1520 kick_iocb(iocb); 1521 return 1; 1522 } 1523 1524 int fastcall io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb, 1525 struct iocb *iocb) 1526 { 1527 struct kiocb *req; 1528 struct file *file; 1529 ssize_t ret; 1530 1531 /* enforce forwards compatibility on users */ 1532 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2 || 1533 iocb->aio_reserved3)) { 1534 pr_debug("EINVAL: io_submit: reserve field set\n"); 1535 return -EINVAL; 1536 } 1537 1538 /* prevent overflows */ 1539 if (unlikely( 1540 (iocb->aio_buf != (unsigned long)iocb->aio_buf) || 1541 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) || 1542 ((ssize_t)iocb->aio_nbytes < 0) 1543 )) { 1544 pr_debug("EINVAL: io_submit: overflow check\n"); 1545 return -EINVAL; 1546 } 1547 1548 file = fget(iocb->aio_fildes); 1549 if (unlikely(!file)) 1550 return -EBADF; 1551 1552 req = aio_get_req(ctx); /* returns with 2 references to req */ 1553 if (unlikely(!req)) { 1554 fput(file); 1555 return -EAGAIN; 1556 } 1557 1558 req->ki_filp = file; 1559 ret = put_user(req->ki_key, &user_iocb->aio_key); 1560 if (unlikely(ret)) { 1561 dprintk("EFAULT: aio_key\n"); 1562 goto out_put_req; 1563 } 1564 1565 req->ki_obj.user = user_iocb; 1566 req->ki_user_data = iocb->aio_data; 1567 req->ki_pos = iocb->aio_offset; 1568 1569 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf; 1570 req->ki_left = req->ki_nbytes = iocb->aio_nbytes; 1571 req->ki_opcode = iocb->aio_lio_opcode; 1572 init_waitqueue_func_entry(&req->ki_wait, aio_wake_function); 1573 INIT_LIST_HEAD(&req->ki_wait.task_list); 1574 1575 ret = aio_setup_iocb(req); 1576 1577 if (ret) 1578 goto out_put_req; 1579 1580 spin_lock_irq(&ctx->ctx_lock); 1581 aio_run_iocb(req); 1582 if (!list_empty(&ctx->run_list)) { 1583 /* drain the run list */ 1584 while (__aio_run_iocbs(ctx)) 1585 ; 1586 } 1587 spin_unlock_irq(&ctx->ctx_lock); 1588 aio_put_req(req); /* drop extra ref to req */ 1589 return 0; 1590 1591 out_put_req: 1592 aio_put_req(req); /* drop extra ref to req */ 1593 aio_put_req(req); /* drop i/o ref to req */ 1594 return ret; 1595 } 1596 1597 /* sys_io_submit: 1598 * Queue the nr iocbs pointed to by iocbpp for processing. Returns 1599 * the number of iocbs queued. May return -EINVAL if the aio_context 1600 * specified by ctx_id is invalid, if nr is < 0, if the iocb at 1601 * *iocbpp[0] is not properly initialized, if the operation specified 1602 * is invalid for the file descriptor in the iocb. May fail with 1603 * -EFAULT if any of the data structures point to invalid data. May 1604 * fail with -EBADF if the file descriptor specified in the first 1605 * iocb is invalid. May fail with -EAGAIN if insufficient resources 1606 * are available to queue any iocbs. Will return 0 if nr is 0. Will 1607 * fail with -ENOSYS if not implemented. 1608 */ 1609 asmlinkage long sys_io_submit(aio_context_t ctx_id, long nr, 1610 struct iocb __user * __user *iocbpp) 1611 { 1612 struct kioctx *ctx; 1613 long ret = 0; 1614 int i; 1615 1616 if (unlikely(nr < 0)) 1617 return -EINVAL; 1618 1619 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp))))) 1620 return -EFAULT; 1621 1622 ctx = lookup_ioctx(ctx_id); 1623 if (unlikely(!ctx)) { 1624 pr_debug("EINVAL: io_submit: invalid context id\n"); 1625 return -EINVAL; 1626 } 1627 1628 /* 1629 * AKPM: should this return a partial result if some of the IOs were 1630 * successfully submitted? 1631 */ 1632 for (i=0; i<nr; i++) { 1633 struct iocb __user *user_iocb; 1634 struct iocb tmp; 1635 1636 if (unlikely(__get_user(user_iocb, iocbpp + i))) { 1637 ret = -EFAULT; 1638 break; 1639 } 1640 1641 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) { 1642 ret = -EFAULT; 1643 break; 1644 } 1645 1646 ret = io_submit_one(ctx, user_iocb, &tmp); 1647 if (ret) 1648 break; 1649 } 1650 1651 put_ioctx(ctx); 1652 return i ? i : ret; 1653 } 1654 1655 /* lookup_kiocb 1656 * Finds a given iocb for cancellation. 1657 */ 1658 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, 1659 u32 key) 1660 { 1661 struct list_head *pos; 1662 1663 assert_spin_locked(&ctx->ctx_lock); 1664 1665 /* TODO: use a hash or array, this sucks. */ 1666 list_for_each(pos, &ctx->active_reqs) { 1667 struct kiocb *kiocb = list_kiocb(pos); 1668 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key) 1669 return kiocb; 1670 } 1671 return NULL; 1672 } 1673 1674 /* sys_io_cancel: 1675 * Attempts to cancel an iocb previously passed to io_submit. If 1676 * the operation is successfully cancelled, the resulting event is 1677 * copied into the memory pointed to by result without being placed 1678 * into the completion queue and 0 is returned. May fail with 1679 * -EFAULT if any of the data structures pointed to are invalid. 1680 * May fail with -EINVAL if aio_context specified by ctx_id is 1681 * invalid. May fail with -EAGAIN if the iocb specified was not 1682 * cancelled. Will fail with -ENOSYS if not implemented. 1683 */ 1684 asmlinkage long sys_io_cancel(aio_context_t ctx_id, struct iocb __user *iocb, 1685 struct io_event __user *result) 1686 { 1687 int (*cancel)(struct kiocb *iocb, struct io_event *res); 1688 struct kioctx *ctx; 1689 struct kiocb *kiocb; 1690 u32 key; 1691 int ret; 1692 1693 ret = get_user(key, &iocb->aio_key); 1694 if (unlikely(ret)) 1695 return -EFAULT; 1696 1697 ctx = lookup_ioctx(ctx_id); 1698 if (unlikely(!ctx)) 1699 return -EINVAL; 1700 1701 spin_lock_irq(&ctx->ctx_lock); 1702 ret = -EAGAIN; 1703 kiocb = lookup_kiocb(ctx, iocb, key); 1704 if (kiocb && kiocb->ki_cancel) { 1705 cancel = kiocb->ki_cancel; 1706 kiocb->ki_users ++; 1707 kiocbSetCancelled(kiocb); 1708 } else 1709 cancel = NULL; 1710 spin_unlock_irq(&ctx->ctx_lock); 1711 1712 if (NULL != cancel) { 1713 struct io_event tmp; 1714 pr_debug("calling cancel\n"); 1715 memset(&tmp, 0, sizeof(tmp)); 1716 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user; 1717 tmp.data = kiocb->ki_user_data; 1718 ret = cancel(kiocb, &tmp); 1719 if (!ret) { 1720 /* Cancellation succeeded -- copy the result 1721 * into the user's buffer. 1722 */ 1723 if (copy_to_user(result, &tmp, sizeof(tmp))) 1724 ret = -EFAULT; 1725 } 1726 } else 1727 ret = -EINVAL; 1728 1729 put_ioctx(ctx); 1730 1731 return ret; 1732 } 1733 1734 /* io_getevents: 1735 * Attempts to read at least min_nr events and up to nr events from 1736 * the completion queue for the aio_context specified by ctx_id. May 1737 * fail with -EINVAL if ctx_id is invalid, if min_nr is out of range, 1738 * if nr is out of range, if when is out of range. May fail with 1739 * -EFAULT if any of the memory specified to is invalid. May return 1740 * 0 or < min_nr if no events are available and the timeout specified 1741 * by when has elapsed, where when == NULL specifies an infinite 1742 * timeout. Note that the timeout pointed to by when is relative and 1743 * will be updated if not NULL and the operation blocks. Will fail 1744 * with -ENOSYS if not implemented. 1745 */ 1746 asmlinkage long sys_io_getevents(aio_context_t ctx_id, 1747 long min_nr, 1748 long nr, 1749 struct io_event __user *events, 1750 struct timespec __user *timeout) 1751 { 1752 struct kioctx *ioctx = lookup_ioctx(ctx_id); 1753 long ret = -EINVAL; 1754 1755 if (likely(ioctx)) { 1756 if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0)) 1757 ret = read_events(ioctx, min_nr, nr, events, timeout); 1758 put_ioctx(ioctx); 1759 } 1760 1761 return ret; 1762 } 1763 1764 __initcall(aio_setup); 1765 1766 EXPORT_SYMBOL(aio_complete); 1767 EXPORT_SYMBOL(aio_put_req); 1768 EXPORT_SYMBOL(wait_on_sync_kiocb); 1769