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 19 #define DEBUG 0 20 21 #include <linux/sched.h> 22 #include <linux/fs.h> 23 #include <linux/file.h> 24 #include <linux/mm.h> 25 #include <linux/mman.h> 26 #include <linux/slab.h> 27 #include <linux/timer.h> 28 #include <linux/aio.h> 29 #include <linux/highmem.h> 30 #include <linux/workqueue.h> 31 #include <linux/security.h> 32 #include <linux/rcuref.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 atomic_t aio_nr = ATOMIC_INIT(0); /* current system wide number of aio requests */ 46 unsigned aio_max_nr = 0x10000; /* system wide maximum number of aio requests */ 47 /*----end sysctl variables---*/ 48 49 static kmem_cache_t *kiocb_cachep; 50 static kmem_cache_t *kioctx_cachep; 51 52 static struct workqueue_struct *aio_wq; 53 54 /* Used for rare fput completion. */ 55 static void aio_fput_routine(void *); 56 static DECLARE_WORK(fput_work, aio_fput_routine, NULL); 57 58 static DEFINE_SPINLOCK(fput_lock); 59 static LIST_HEAD(fput_head); 60 61 static void aio_kick_handler(void *); 62 static void aio_queue_work(struct kioctx *); 63 64 /* aio_setup 65 * Creates the slab caches used by the aio routines, panic on 66 * failure as this is done early during the boot sequence. 67 */ 68 static int __init aio_setup(void) 69 { 70 kiocb_cachep = kmem_cache_create("kiocb", sizeof(struct kiocb), 71 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL); 72 kioctx_cachep = kmem_cache_create("kioctx", sizeof(struct kioctx), 73 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL); 74 75 aio_wq = create_workqueue("aio"); 76 77 pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page)); 78 79 return 0; 80 } 81 82 static void aio_free_ring(struct kioctx *ctx) 83 { 84 struct aio_ring_info *info = &ctx->ring_info; 85 long i; 86 87 for (i=0; i<info->nr_pages; i++) 88 put_page(info->ring_pages[i]); 89 90 if (info->mmap_size) { 91 down_write(&ctx->mm->mmap_sem); 92 do_munmap(ctx->mm, info->mmap_base, info->mmap_size); 93 up_write(&ctx->mm->mmap_sem); 94 } 95 96 if (info->ring_pages && info->ring_pages != info->internal_pages) 97 kfree(info->ring_pages); 98 info->ring_pages = NULL; 99 info->nr = 0; 100 } 101 102 static int aio_setup_ring(struct kioctx *ctx) 103 { 104 struct aio_ring *ring; 105 struct aio_ring_info *info = &ctx->ring_info; 106 unsigned nr_events = ctx->max_reqs; 107 unsigned long size; 108 int nr_pages; 109 110 /* Compensate for the ring buffer's head/tail overlap entry */ 111 nr_events += 2; /* 1 is required, 2 for good luck */ 112 113 size = sizeof(struct aio_ring); 114 size += sizeof(struct io_event) * nr_events; 115 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT; 116 117 if (nr_pages < 0) 118 return -EINVAL; 119 120 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event); 121 122 info->nr = 0; 123 info->ring_pages = info->internal_pages; 124 if (nr_pages > AIO_RING_PAGES) { 125 info->ring_pages = kmalloc(sizeof(struct page *) * nr_pages, GFP_KERNEL); 126 if (!info->ring_pages) 127 return -ENOMEM; 128 memset(info->ring_pages, 0, sizeof(struct page *) * nr_pages); 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_ANON|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 (nr_events > aio_max_nr) 212 return ERR_PTR(-EAGAIN); 213 214 ctx = kmem_cache_alloc(kioctx_cachep, GFP_KERNEL); 215 if (!ctx) 216 return ERR_PTR(-ENOMEM); 217 218 memset(ctx, 0, sizeof(*ctx)); 219 ctx->max_reqs = nr_events; 220 mm = ctx->mm = current->mm; 221 atomic_inc(&mm->mm_count); 222 223 atomic_set(&ctx->users, 1); 224 spin_lock_init(&ctx->ctx_lock); 225 spin_lock_init(&ctx->ring_info.ring_lock); 226 init_waitqueue_head(&ctx->wait); 227 228 INIT_LIST_HEAD(&ctx->active_reqs); 229 INIT_LIST_HEAD(&ctx->run_list); 230 INIT_WORK(&ctx->wq, aio_kick_handler, ctx); 231 232 if (aio_setup_ring(ctx) < 0) 233 goto out_freectx; 234 235 /* limit the number of system wide aios */ 236 atomic_add(ctx->max_reqs, &aio_nr); /* undone by __put_ioctx */ 237 if (unlikely(atomic_read(&aio_nr) > aio_max_nr)) 238 goto out_cleanup; 239 240 /* now link into global list. kludge. FIXME */ 241 write_lock(&mm->ioctx_list_lock); 242 ctx->next = mm->ioctx_list; 243 mm->ioctx_list = ctx; 244 write_unlock(&mm->ioctx_list_lock); 245 246 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n", 247 ctx, ctx->user_id, current->mm, ctx->ring_info.nr); 248 return ctx; 249 250 out_cleanup: 251 atomic_sub(ctx->max_reqs, &aio_nr); 252 ctx->max_reqs = 0; /* prevent __put_ioctx from sub'ing aio_nr */ 253 __put_ioctx(ctx); 254 return ERR_PTR(-EAGAIN); 255 256 out_freectx: 257 mmdrop(mm); 258 kmem_cache_free(kioctx_cachep, ctx); 259 ctx = ERR_PTR(-ENOMEM); 260 261 dprintk("aio: error allocating ioctx %p\n", ctx); 262 return ctx; 263 } 264 265 /* aio_cancel_all 266 * Cancels all outstanding aio requests on an aio context. Used 267 * when the processes owning a context have all exited to encourage 268 * the rapid destruction of the kioctx. 269 */ 270 static void aio_cancel_all(struct kioctx *ctx) 271 { 272 int (*cancel)(struct kiocb *, struct io_event *); 273 struct io_event res; 274 spin_lock_irq(&ctx->ctx_lock); 275 ctx->dead = 1; 276 while (!list_empty(&ctx->active_reqs)) { 277 struct list_head *pos = ctx->active_reqs.next; 278 struct kiocb *iocb = list_kiocb(pos); 279 list_del_init(&iocb->ki_list); 280 cancel = iocb->ki_cancel; 281 kiocbSetCancelled(iocb); 282 if (cancel) { 283 iocb->ki_users++; 284 spin_unlock_irq(&ctx->ctx_lock); 285 cancel(iocb, &res); 286 spin_lock_irq(&ctx->ctx_lock); 287 } 288 } 289 spin_unlock_irq(&ctx->ctx_lock); 290 } 291 292 static void wait_for_all_aios(struct kioctx *ctx) 293 { 294 struct task_struct *tsk = current; 295 DECLARE_WAITQUEUE(wait, tsk); 296 297 if (!ctx->reqs_active) 298 return; 299 300 add_wait_queue(&ctx->wait, &wait); 301 set_task_state(tsk, TASK_UNINTERRUPTIBLE); 302 while (ctx->reqs_active) { 303 schedule(); 304 set_task_state(tsk, TASK_UNINTERRUPTIBLE); 305 } 306 __set_task_state(tsk, TASK_RUNNING); 307 remove_wait_queue(&ctx->wait, &wait); 308 } 309 310 /* wait_on_sync_kiocb: 311 * Waits on the given sync kiocb to complete. 312 */ 313 ssize_t fastcall wait_on_sync_kiocb(struct kiocb *iocb) 314 { 315 while (iocb->ki_users) { 316 set_current_state(TASK_UNINTERRUPTIBLE); 317 if (!iocb->ki_users) 318 break; 319 schedule(); 320 } 321 __set_current_state(TASK_RUNNING); 322 return iocb->ki_user_data; 323 } 324 325 /* exit_aio: called when the last user of mm goes away. At this point, 326 * there is no way for any new requests to be submited or any of the 327 * io_* syscalls to be called on the context. However, there may be 328 * outstanding requests which hold references to the context; as they 329 * go away, they will call put_ioctx and release any pinned memory 330 * associated with the request (held via struct page * references). 331 */ 332 void fastcall exit_aio(struct mm_struct *mm) 333 { 334 struct kioctx *ctx = mm->ioctx_list; 335 mm->ioctx_list = NULL; 336 while (ctx) { 337 struct kioctx *next = ctx->next; 338 ctx->next = NULL; 339 aio_cancel_all(ctx); 340 341 wait_for_all_aios(ctx); 342 /* 343 * this is an overkill, but ensures we don't leave 344 * the ctx on the aio_wq 345 */ 346 flush_workqueue(aio_wq); 347 348 if (1 != atomic_read(&ctx->users)) 349 printk(KERN_DEBUG 350 "exit_aio:ioctx still alive: %d %d %d\n", 351 atomic_read(&ctx->users), ctx->dead, 352 ctx->reqs_active); 353 put_ioctx(ctx); 354 ctx = next; 355 } 356 } 357 358 /* __put_ioctx 359 * Called when the last user of an aio context has gone away, 360 * and the struct needs to be freed. 361 */ 362 void fastcall __put_ioctx(struct kioctx *ctx) 363 { 364 unsigned nr_events = ctx->max_reqs; 365 366 if (unlikely(ctx->reqs_active)) 367 BUG(); 368 369 cancel_delayed_work(&ctx->wq); 370 flush_workqueue(aio_wq); 371 aio_free_ring(ctx); 372 mmdrop(ctx->mm); 373 ctx->mm = NULL; 374 pr_debug("__put_ioctx: freeing %p\n", ctx); 375 kmem_cache_free(kioctx_cachep, ctx); 376 377 atomic_sub(nr_events, &aio_nr); 378 } 379 380 /* aio_get_req 381 * Allocate a slot for an aio request. Increments the users count 382 * of the kioctx so that the kioctx stays around until all requests are 383 * complete. Returns NULL if no requests are free. 384 * 385 * Returns with kiocb->users set to 2. The io submit code path holds 386 * an extra reference while submitting the i/o. 387 * This prevents races between the aio code path referencing the 388 * req (after submitting it) and aio_complete() freeing the req. 389 */ 390 static struct kiocb *FASTCALL(__aio_get_req(struct kioctx *ctx)); 391 static struct kiocb fastcall *__aio_get_req(struct kioctx *ctx) 392 { 393 struct kiocb *req = NULL; 394 struct aio_ring *ring; 395 int okay = 0; 396 397 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL); 398 if (unlikely(!req)) 399 return NULL; 400 401 req->ki_flags = 0; 402 req->ki_users = 2; 403 req->ki_key = 0; 404 req->ki_ctx = ctx; 405 req->ki_cancel = NULL; 406 req->ki_retry = NULL; 407 req->ki_dtor = NULL; 408 req->private = NULL; 409 INIT_LIST_HEAD(&req->ki_run_list); 410 411 /* Check if the completion queue has enough free space to 412 * accept an event from this io. 413 */ 414 spin_lock_irq(&ctx->ctx_lock); 415 ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0); 416 if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) { 417 list_add(&req->ki_list, &ctx->active_reqs); 418 get_ioctx(ctx); 419 ctx->reqs_active++; 420 okay = 1; 421 } 422 kunmap_atomic(ring, KM_USER0); 423 spin_unlock_irq(&ctx->ctx_lock); 424 425 if (!okay) { 426 kmem_cache_free(kiocb_cachep, req); 427 req = NULL; 428 } 429 430 return req; 431 } 432 433 static inline struct kiocb *aio_get_req(struct kioctx *ctx) 434 { 435 struct kiocb *req; 436 /* Handle a potential starvation case -- should be exceedingly rare as 437 * requests will be stuck on fput_head only if the aio_fput_routine is 438 * delayed and the requests were the last user of the struct file. 439 */ 440 req = __aio_get_req(ctx); 441 if (unlikely(NULL == req)) { 442 aio_fput_routine(NULL); 443 req = __aio_get_req(ctx); 444 } 445 return req; 446 } 447 448 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req) 449 { 450 if (req->ki_dtor) 451 req->ki_dtor(req); 452 kmem_cache_free(kiocb_cachep, req); 453 ctx->reqs_active--; 454 455 if (unlikely(!ctx->reqs_active && ctx->dead)) 456 wake_up(&ctx->wait); 457 } 458 459 static void aio_fput_routine(void *data) 460 { 461 spin_lock_irq(&fput_lock); 462 while (likely(!list_empty(&fput_head))) { 463 struct kiocb *req = list_kiocb(fput_head.next); 464 struct kioctx *ctx = req->ki_ctx; 465 466 list_del(&req->ki_list); 467 spin_unlock_irq(&fput_lock); 468 469 /* Complete the fput */ 470 __fput(req->ki_filp); 471 472 /* Link the iocb into the context's free list */ 473 spin_lock_irq(&ctx->ctx_lock); 474 really_put_req(ctx, req); 475 spin_unlock_irq(&ctx->ctx_lock); 476 477 put_ioctx(ctx); 478 spin_lock_irq(&fput_lock); 479 } 480 spin_unlock_irq(&fput_lock); 481 } 482 483 /* __aio_put_req 484 * Returns true if this put was the last user of the request. 485 */ 486 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req) 487 { 488 dprintk(KERN_DEBUG "aio_put(%p): f_count=%d\n", 489 req, atomic_read(&req->ki_filp->f_count)); 490 491 req->ki_users --; 492 if (unlikely(req->ki_users < 0)) 493 BUG(); 494 if (likely(req->ki_users)) 495 return 0; 496 list_del(&req->ki_list); /* remove from active_reqs */ 497 req->ki_cancel = NULL; 498 req->ki_retry = NULL; 499 500 /* Must be done under the lock to serialise against cancellation. 501 * Call this aio_fput as it duplicates fput via the fput_work. 502 */ 503 if (unlikely(rcuref_dec_and_test(&req->ki_filp->f_count))) { 504 get_ioctx(ctx); 505 spin_lock(&fput_lock); 506 list_add(&req->ki_list, &fput_head); 507 spin_unlock(&fput_lock); 508 queue_work(aio_wq, &fput_work); 509 } else 510 really_put_req(ctx, req); 511 return 1; 512 } 513 514 /* aio_put_req 515 * Returns true if this put was the last user of the kiocb, 516 * false if the request is still in use. 517 */ 518 int fastcall aio_put_req(struct kiocb *req) 519 { 520 struct kioctx *ctx = req->ki_ctx; 521 int ret; 522 spin_lock_irq(&ctx->ctx_lock); 523 ret = __aio_put_req(ctx, req); 524 spin_unlock_irq(&ctx->ctx_lock); 525 if (ret) 526 put_ioctx(ctx); 527 return ret; 528 } 529 530 /* Lookup an ioctx id. ioctx_list is lockless for reads. 531 * FIXME: this is O(n) and is only suitable for development. 532 */ 533 struct kioctx *lookup_ioctx(unsigned long ctx_id) 534 { 535 struct kioctx *ioctx; 536 struct mm_struct *mm; 537 538 mm = current->mm; 539 read_lock(&mm->ioctx_list_lock); 540 for (ioctx = mm->ioctx_list; ioctx; ioctx = ioctx->next) 541 if (likely(ioctx->user_id == ctx_id && !ioctx->dead)) { 542 get_ioctx(ioctx); 543 break; 544 } 545 read_unlock(&mm->ioctx_list_lock); 546 547 return ioctx; 548 } 549 550 /* 551 * use_mm 552 * Makes the calling kernel thread take on the specified 553 * mm context. 554 * Called by the retry thread execute retries within the 555 * iocb issuer's mm context, so that copy_from/to_user 556 * operations work seamlessly for aio. 557 * (Note: this routine is intended to be called only 558 * from a kernel thread context) 559 */ 560 static void use_mm(struct mm_struct *mm) 561 { 562 struct mm_struct *active_mm; 563 struct task_struct *tsk = current; 564 565 task_lock(tsk); 566 tsk->flags |= PF_BORROWED_MM; 567 active_mm = tsk->active_mm; 568 atomic_inc(&mm->mm_count); 569 tsk->mm = mm; 570 tsk->active_mm = mm; 571 /* 572 * Note that on UML this *requires* PF_BORROWED_MM to be set, otherwise 573 * it won't work. Update it accordingly if you change it here 574 */ 575 activate_mm(active_mm, mm); 576 task_unlock(tsk); 577 578 mmdrop(active_mm); 579 } 580 581 /* 582 * unuse_mm 583 * Reverses the effect of use_mm, i.e. releases the 584 * specified mm context which was earlier taken on 585 * by the calling kernel thread 586 * (Note: this routine is intended to be called only 587 * from a kernel thread context) 588 * 589 * Comments: Called with ctx->ctx_lock held. This nests 590 * task_lock instead ctx_lock. 591 */ 592 static void unuse_mm(struct mm_struct *mm) 593 { 594 struct task_struct *tsk = current; 595 596 task_lock(tsk); 597 tsk->flags &= ~PF_BORROWED_MM; 598 tsk->mm = NULL; 599 /* active_mm is still 'mm' */ 600 enter_lazy_tlb(mm, tsk); 601 task_unlock(tsk); 602 } 603 604 /* 605 * Queue up a kiocb to be retried. Assumes that the kiocb 606 * has already been marked as kicked, and places it on 607 * the retry run list for the corresponding ioctx, if it 608 * isn't already queued. Returns 1 if it actually queued 609 * the kiocb (to tell the caller to activate the work 610 * queue to process it), or 0, if it found that it was 611 * already queued. 612 * 613 * Should be called with the spin lock iocb->ki_ctx->ctx_lock 614 * held 615 */ 616 static inline int __queue_kicked_iocb(struct kiocb *iocb) 617 { 618 struct kioctx *ctx = iocb->ki_ctx; 619 620 if (list_empty(&iocb->ki_run_list)) { 621 list_add_tail(&iocb->ki_run_list, 622 &ctx->run_list); 623 return 1; 624 } 625 return 0; 626 } 627 628 /* aio_run_iocb 629 * This is the core aio execution routine. It is 630 * invoked both for initial i/o submission and 631 * subsequent retries via the aio_kick_handler. 632 * Expects to be invoked with iocb->ki_ctx->lock 633 * already held. The lock is released and reaquired 634 * as needed during processing. 635 * 636 * Calls the iocb retry method (already setup for the 637 * iocb on initial submission) for operation specific 638 * handling, but takes care of most of common retry 639 * execution details for a given iocb. The retry method 640 * needs to be non-blocking as far as possible, to avoid 641 * holding up other iocbs waiting to be serviced by the 642 * retry kernel thread. 643 * 644 * The trickier parts in this code have to do with 645 * ensuring that only one retry instance is in progress 646 * for a given iocb at any time. Providing that guarantee 647 * simplifies the coding of individual aio operations as 648 * it avoids various potential races. 649 */ 650 static ssize_t aio_run_iocb(struct kiocb *iocb) 651 { 652 struct kioctx *ctx = iocb->ki_ctx; 653 ssize_t (*retry)(struct kiocb *); 654 ssize_t ret; 655 656 if (iocb->ki_retried++ > 1024*1024) { 657 printk("Maximal retry count. Bytes done %Zd\n", 658 iocb->ki_nbytes - iocb->ki_left); 659 return -EAGAIN; 660 } 661 662 if (!(iocb->ki_retried & 0xff)) { 663 pr_debug("%ld retry: %d of %d\n", iocb->ki_retried, 664 iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes); 665 } 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. Expects to be called with ctx->ctx_lock held 765 */ 766 static int __aio_run_iocbs(struct kioctx *ctx) 767 { 768 struct kiocb *iocb; 769 LIST_HEAD(run_list); 770 771 list_splice_init(&ctx->run_list, &run_list); 772 while (!list_empty(&run_list)) { 773 iocb = list_entry(run_list.next, struct kiocb, 774 ki_run_list); 775 list_del(&iocb->ki_run_list); 776 /* 777 * Hold an extra reference while retrying i/o. 778 */ 779 iocb->ki_users++; /* grab extra reference */ 780 aio_run_iocb(iocb); 781 if (__aio_put_req(ctx, iocb)) /* drop extra ref */ 782 put_ioctx(ctx); 783 } 784 if (!list_empty(&ctx->run_list)) 785 return 1; 786 return 0; 787 } 788 789 static void aio_queue_work(struct kioctx * ctx) 790 { 791 unsigned long timeout; 792 /* 793 * if someone is waiting, get the work started right 794 * away, otherwise, use a longer delay 795 */ 796 smp_mb(); 797 if (waitqueue_active(&ctx->wait)) 798 timeout = 1; 799 else 800 timeout = HZ/10; 801 queue_delayed_work(aio_wq, &ctx->wq, timeout); 802 } 803 804 805 /* 806 * aio_run_iocbs: 807 * Process all pending retries queued on the ioctx 808 * run list. 809 * Assumes it is operating within the aio issuer's mm 810 * context. 811 */ 812 static inline void aio_run_iocbs(struct kioctx *ctx) 813 { 814 int requeue; 815 816 spin_lock_irq(&ctx->ctx_lock); 817 818 requeue = __aio_run_iocbs(ctx); 819 spin_unlock_irq(&ctx->ctx_lock); 820 if (requeue) 821 aio_queue_work(ctx); 822 } 823 824 /* 825 * just like aio_run_iocbs, but keeps running them until 826 * the list stays empty 827 */ 828 static inline void aio_run_all_iocbs(struct kioctx *ctx) 829 { 830 spin_lock_irq(&ctx->ctx_lock); 831 while (__aio_run_iocbs(ctx)) 832 ; 833 spin_unlock_irq(&ctx->ctx_lock); 834 } 835 836 /* 837 * aio_kick_handler: 838 * Work queue handler triggered to process pending 839 * retries on an ioctx. Takes on the aio issuer's 840 * mm context before running the iocbs, so that 841 * copy_xxx_user operates on the issuer's address 842 * space. 843 * Run on aiod's context. 844 */ 845 static void aio_kick_handler(void *data) 846 { 847 struct kioctx *ctx = data; 848 mm_segment_t oldfs = get_fs(); 849 int requeue; 850 851 set_fs(USER_DS); 852 use_mm(ctx->mm); 853 spin_lock_irq(&ctx->ctx_lock); 854 requeue =__aio_run_iocbs(ctx); 855 unuse_mm(ctx->mm); 856 spin_unlock_irq(&ctx->ctx_lock); 857 set_fs(oldfs); 858 /* 859 * we're in a worker thread already, don't use queue_delayed_work, 860 */ 861 if (requeue) 862 queue_work(aio_wq, &ctx->wq); 863 } 864 865 866 /* 867 * Called by kick_iocb to queue the kiocb for retry 868 * and if required activate the aio work queue to process 869 * it 870 */ 871 static void try_queue_kicked_iocb(struct kiocb *iocb) 872 { 873 struct kioctx *ctx = iocb->ki_ctx; 874 unsigned long flags; 875 int run = 0; 876 877 /* We're supposed to be the only path putting the iocb back on the run 878 * list. If we find that the iocb is *back* on a wait queue already 879 * than retry has happened before we could queue the iocb. This also 880 * means that the retry could have completed and freed our iocb, no 881 * good. */ 882 BUG_ON((!list_empty(&iocb->ki_wait.task_list))); 883 884 spin_lock_irqsave(&ctx->ctx_lock, flags); 885 /* set this inside the lock so that we can't race with aio_run_iocb() 886 * testing it and putting the iocb on the run list under the lock */ 887 if (!kiocbTryKick(iocb)) 888 run = __queue_kicked_iocb(iocb); 889 spin_unlock_irqrestore(&ctx->ctx_lock, flags); 890 if (run) 891 aio_queue_work(ctx); 892 } 893 894 /* 895 * kick_iocb: 896 * Called typically from a wait queue callback context 897 * (aio_wake_function) to trigger a retry of the iocb. 898 * The retry is usually executed by aio workqueue 899 * threads (See aio_kick_handler). 900 */ 901 void fastcall kick_iocb(struct kiocb *iocb) 902 { 903 /* sync iocbs are easy: they can only ever be executing from a 904 * single context. */ 905 if (is_sync_kiocb(iocb)) { 906 kiocbSetKicked(iocb); 907 wake_up_process(iocb->ki_obj.tsk); 908 return; 909 } 910 911 try_queue_kicked_iocb(iocb); 912 } 913 EXPORT_SYMBOL(kick_iocb); 914 915 /* aio_complete 916 * Called when the io request on the given iocb is complete. 917 * Returns true if this is the last user of the request. The 918 * only other user of the request can be the cancellation code. 919 */ 920 int fastcall aio_complete(struct kiocb *iocb, long res, long res2) 921 { 922 struct kioctx *ctx = iocb->ki_ctx; 923 struct aio_ring_info *info; 924 struct aio_ring *ring; 925 struct io_event *event; 926 unsigned long flags; 927 unsigned long tail; 928 int ret; 929 930 /* Special case handling for sync iocbs: events go directly 931 * into the iocb for fast handling. Note that this will not 932 * work if we allow sync kiocbs to be cancelled. in which 933 * case the usage count checks will have to move under ctx_lock 934 * for all cases. 935 */ 936 if (is_sync_kiocb(iocb)) { 937 int ret; 938 939 iocb->ki_user_data = res; 940 if (iocb->ki_users == 1) { 941 iocb->ki_users = 0; 942 ret = 1; 943 } else { 944 spin_lock_irq(&ctx->ctx_lock); 945 iocb->ki_users--; 946 ret = (0 == iocb->ki_users); 947 spin_unlock_irq(&ctx->ctx_lock); 948 } 949 /* sync iocbs put the task here for us */ 950 wake_up_process(iocb->ki_obj.tsk); 951 return ret; 952 } 953 954 info = &ctx->ring_info; 955 956 /* add a completion event to the ring buffer. 957 * must be done holding ctx->ctx_lock to prevent 958 * other code from messing with the tail 959 * pointer since we might be called from irq 960 * context. 961 */ 962 spin_lock_irqsave(&ctx->ctx_lock, flags); 963 964 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list)) 965 list_del_init(&iocb->ki_run_list); 966 967 /* 968 * cancelled requests don't get events, userland was given one 969 * when the event got cancelled. 970 */ 971 if (kiocbIsCancelled(iocb)) 972 goto put_rq; 973 974 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1); 975 976 tail = info->tail; 977 event = aio_ring_event(info, tail, KM_IRQ0); 978 if (++tail >= info->nr) 979 tail = 0; 980 981 event->obj = (u64)(unsigned long)iocb->ki_obj.user; 982 event->data = iocb->ki_user_data; 983 event->res = res; 984 event->res2 = res2; 985 986 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n", 987 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data, 988 res, res2); 989 990 /* after flagging the request as done, we 991 * must never even look at it again 992 */ 993 smp_wmb(); /* make event visible before updating tail */ 994 995 info->tail = tail; 996 ring->tail = tail; 997 998 put_aio_ring_event(event, KM_IRQ0); 999 kunmap_atomic(ring, KM_IRQ1); 1000 1001 pr_debug("added to ring %p at [%lu]\n", iocb, tail); 1002 1003 pr_debug("%ld retries: %d of %d\n", iocb->ki_retried, 1004 iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes); 1005 put_rq: 1006 /* everything turned out well, dispose of the aiocb. */ 1007 ret = __aio_put_req(ctx, iocb); 1008 1009 spin_unlock_irqrestore(&ctx->ctx_lock, flags); 1010 1011 if (waitqueue_active(&ctx->wait)) 1012 wake_up(&ctx->wait); 1013 1014 if (ret) 1015 put_ioctx(ctx); 1016 1017 return ret; 1018 } 1019 1020 /* aio_read_evt 1021 * Pull an event off of the ioctx's event ring. Returns the number of 1022 * events fetched (0 or 1 ;-) 1023 * FIXME: make this use cmpxchg. 1024 * TODO: make the ringbuffer user mmap()able (requires FIXME). 1025 */ 1026 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent) 1027 { 1028 struct aio_ring_info *info = &ioctx->ring_info; 1029 struct aio_ring *ring; 1030 unsigned long head; 1031 int ret = 0; 1032 1033 ring = kmap_atomic(info->ring_pages[0], KM_USER0); 1034 dprintk("in aio_read_evt h%lu t%lu m%lu\n", 1035 (unsigned long)ring->head, (unsigned long)ring->tail, 1036 (unsigned long)ring->nr); 1037 1038 if (ring->head == ring->tail) 1039 goto out; 1040 1041 spin_lock(&info->ring_lock); 1042 1043 head = ring->head % info->nr; 1044 if (head != ring->tail) { 1045 struct io_event *evp = aio_ring_event(info, head, KM_USER1); 1046 *ent = *evp; 1047 head = (head + 1) % info->nr; 1048 smp_mb(); /* finish reading the event before updatng the head */ 1049 ring->head = head; 1050 ret = 1; 1051 put_aio_ring_event(evp, KM_USER1); 1052 } 1053 spin_unlock(&info->ring_lock); 1054 1055 out: 1056 kunmap_atomic(ring, KM_USER0); 1057 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret, 1058 (unsigned long)ring->head, (unsigned long)ring->tail); 1059 return ret; 1060 } 1061 1062 struct aio_timeout { 1063 struct timer_list timer; 1064 int timed_out; 1065 struct task_struct *p; 1066 }; 1067 1068 static void timeout_func(unsigned long data) 1069 { 1070 struct aio_timeout *to = (struct aio_timeout *)data; 1071 1072 to->timed_out = 1; 1073 wake_up_process(to->p); 1074 } 1075 1076 static inline void init_timeout(struct aio_timeout *to) 1077 { 1078 init_timer(&to->timer); 1079 to->timer.data = (unsigned long)to; 1080 to->timer.function = timeout_func; 1081 to->timed_out = 0; 1082 to->p = current; 1083 } 1084 1085 static inline void set_timeout(long start_jiffies, struct aio_timeout *to, 1086 const struct timespec *ts) 1087 { 1088 to->timer.expires = start_jiffies + timespec_to_jiffies(ts); 1089 if (time_after(to->timer.expires, jiffies)) 1090 add_timer(&to->timer); 1091 else 1092 to->timed_out = 1; 1093 } 1094 1095 static inline void clear_timeout(struct aio_timeout *to) 1096 { 1097 del_singleshot_timer_sync(&to->timer); 1098 } 1099 1100 static int read_events(struct kioctx *ctx, 1101 long min_nr, long nr, 1102 struct io_event __user *event, 1103 struct timespec __user *timeout) 1104 { 1105 long start_jiffies = jiffies; 1106 struct task_struct *tsk = current; 1107 DECLARE_WAITQUEUE(wait, tsk); 1108 int ret; 1109 int i = 0; 1110 struct io_event ent; 1111 struct aio_timeout to; 1112 int retry = 0; 1113 1114 /* needed to zero any padding within an entry (there shouldn't be 1115 * any, but C is fun! 1116 */ 1117 memset(&ent, 0, sizeof(ent)); 1118 retry: 1119 ret = 0; 1120 while (likely(i < nr)) { 1121 ret = aio_read_evt(ctx, &ent); 1122 if (unlikely(ret <= 0)) 1123 break; 1124 1125 dprintk("read event: %Lx %Lx %Lx %Lx\n", 1126 ent.data, ent.obj, ent.res, ent.res2); 1127 1128 /* Could we split the check in two? */ 1129 ret = -EFAULT; 1130 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) { 1131 dprintk("aio: lost an event due to EFAULT.\n"); 1132 break; 1133 } 1134 ret = 0; 1135 1136 /* Good, event copied to userland, update counts. */ 1137 event ++; 1138 i ++; 1139 } 1140 1141 if (min_nr <= i) 1142 return i; 1143 if (ret) 1144 return ret; 1145 1146 /* End fast path */ 1147 1148 /* racey check, but it gets redone */ 1149 if (!retry && unlikely(!list_empty(&ctx->run_list))) { 1150 retry = 1; 1151 aio_run_all_iocbs(ctx); 1152 goto retry; 1153 } 1154 1155 init_timeout(&to); 1156 if (timeout) { 1157 struct timespec ts; 1158 ret = -EFAULT; 1159 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts)))) 1160 goto out; 1161 1162 set_timeout(start_jiffies, &to, &ts); 1163 } 1164 1165 while (likely(i < nr)) { 1166 add_wait_queue_exclusive(&ctx->wait, &wait); 1167 do { 1168 set_task_state(tsk, TASK_INTERRUPTIBLE); 1169 ret = aio_read_evt(ctx, &ent); 1170 if (ret) 1171 break; 1172 if (min_nr <= i) 1173 break; 1174 ret = 0; 1175 if (to.timed_out) /* Only check after read evt */ 1176 break; 1177 schedule(); 1178 if (signal_pending(tsk)) { 1179 ret = -EINTR; 1180 break; 1181 } 1182 /*ret = aio_read_evt(ctx, &ent);*/ 1183 } while (1) ; 1184 1185 set_task_state(tsk, TASK_RUNNING); 1186 remove_wait_queue(&ctx->wait, &wait); 1187 1188 if (unlikely(ret <= 0)) 1189 break; 1190 1191 ret = -EFAULT; 1192 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) { 1193 dprintk("aio: lost an event due to EFAULT.\n"); 1194 break; 1195 } 1196 1197 /* Good, event copied to userland, update counts. */ 1198 event ++; 1199 i ++; 1200 } 1201 1202 if (timeout) 1203 clear_timeout(&to); 1204 out: 1205 return i ? i : ret; 1206 } 1207 1208 /* Take an ioctx and remove it from the list of ioctx's. Protects 1209 * against races with itself via ->dead. 1210 */ 1211 static void io_destroy(struct kioctx *ioctx) 1212 { 1213 struct mm_struct *mm = current->mm; 1214 struct kioctx **tmp; 1215 int was_dead; 1216 1217 /* delete the entry from the list is someone else hasn't already */ 1218 write_lock(&mm->ioctx_list_lock); 1219 was_dead = ioctx->dead; 1220 ioctx->dead = 1; 1221 for (tmp = &mm->ioctx_list; *tmp && *tmp != ioctx; 1222 tmp = &(*tmp)->next) 1223 ; 1224 if (*tmp) 1225 *tmp = ioctx->next; 1226 write_unlock(&mm->ioctx_list_lock); 1227 1228 dprintk("aio_release(%p)\n", ioctx); 1229 if (likely(!was_dead)) 1230 put_ioctx(ioctx); /* twice for the list */ 1231 1232 aio_cancel_all(ioctx); 1233 wait_for_all_aios(ioctx); 1234 put_ioctx(ioctx); /* once for the lookup */ 1235 } 1236 1237 /* sys_io_setup: 1238 * Create an aio_context capable of receiving at least nr_events. 1239 * ctxp must not point to an aio_context that already exists, and 1240 * must be initialized to 0 prior to the call. On successful 1241 * creation of the aio_context, *ctxp is filled in with the resulting 1242 * handle. May fail with -EINVAL if *ctxp is not initialized, 1243 * if the specified nr_events exceeds internal limits. May fail 1244 * with -EAGAIN if the specified nr_events exceeds the user's limit 1245 * of available events. May fail with -ENOMEM if insufficient kernel 1246 * resources are available. May fail with -EFAULT if an invalid 1247 * pointer is passed for ctxp. Will fail with -ENOSYS if not 1248 * implemented. 1249 */ 1250 asmlinkage long sys_io_setup(unsigned nr_events, aio_context_t __user *ctxp) 1251 { 1252 struct kioctx *ioctx = NULL; 1253 unsigned long ctx; 1254 long ret; 1255 1256 ret = get_user(ctx, ctxp); 1257 if (unlikely(ret)) 1258 goto out; 1259 1260 ret = -EINVAL; 1261 if (unlikely(ctx || (int)nr_events <= 0)) { 1262 pr_debug("EINVAL: io_setup: ctx or nr_events > max\n"); 1263 goto out; 1264 } 1265 1266 ioctx = ioctx_alloc(nr_events); 1267 ret = PTR_ERR(ioctx); 1268 if (!IS_ERR(ioctx)) { 1269 ret = put_user(ioctx->user_id, ctxp); 1270 if (!ret) 1271 return 0; 1272 1273 get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */ 1274 io_destroy(ioctx); 1275 } 1276 1277 out: 1278 return ret; 1279 } 1280 1281 /* sys_io_destroy: 1282 * Destroy the aio_context specified. May cancel any outstanding 1283 * AIOs and block on completion. Will fail with -ENOSYS if not 1284 * implemented. May fail with -EFAULT if the context pointed to 1285 * is invalid. 1286 */ 1287 asmlinkage long sys_io_destroy(aio_context_t ctx) 1288 { 1289 struct kioctx *ioctx = lookup_ioctx(ctx); 1290 if (likely(NULL != ioctx)) { 1291 io_destroy(ioctx); 1292 return 0; 1293 } 1294 pr_debug("EINVAL: io_destroy: invalid context id\n"); 1295 return -EINVAL; 1296 } 1297 1298 /* 1299 * aio_p{read,write} are the default ki_retry methods for 1300 * IO_CMD_P{READ,WRITE}. They maintains kiocb retry state around potentially 1301 * multiple calls to f_op->aio_read(). They loop around partial progress 1302 * instead of returning -EIOCBRETRY because they don't have the means to call 1303 * kick_iocb(). 1304 */ 1305 static ssize_t aio_pread(struct kiocb *iocb) 1306 { 1307 struct file *file = iocb->ki_filp; 1308 struct address_space *mapping = file->f_mapping; 1309 struct inode *inode = mapping->host; 1310 ssize_t ret = 0; 1311 1312 do { 1313 ret = file->f_op->aio_read(iocb, iocb->ki_buf, 1314 iocb->ki_left, iocb->ki_pos); 1315 /* 1316 * Can't just depend on iocb->ki_left to determine 1317 * whether we are done. This may have been a short read. 1318 */ 1319 if (ret > 0) { 1320 iocb->ki_buf += ret; 1321 iocb->ki_left -= ret; 1322 } 1323 1324 /* 1325 * For pipes and sockets we return once we have some data; for 1326 * regular files we retry till we complete the entire read or 1327 * find that we can't read any more data (e.g short reads). 1328 */ 1329 } while (ret > 0 && iocb->ki_left > 0 && 1330 !S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode)); 1331 1332 /* This means we must have transferred all that we could */ 1333 /* No need to retry anymore */ 1334 if ((ret == 0) || (iocb->ki_left == 0)) 1335 ret = iocb->ki_nbytes - iocb->ki_left; 1336 1337 return ret; 1338 } 1339 1340 /* see aio_pread() */ 1341 static ssize_t aio_pwrite(struct kiocb *iocb) 1342 { 1343 struct file *file = iocb->ki_filp; 1344 ssize_t ret = 0; 1345 1346 do { 1347 ret = file->f_op->aio_write(iocb, iocb->ki_buf, 1348 iocb->ki_left, iocb->ki_pos); 1349 if (ret > 0) { 1350 iocb->ki_buf += ret; 1351 iocb->ki_left -= ret; 1352 } 1353 } while (ret > 0 && iocb->ki_left > 0); 1354 1355 if ((ret == 0) || (iocb->ki_left == 0)) 1356 ret = iocb->ki_nbytes - iocb->ki_left; 1357 1358 return ret; 1359 } 1360 1361 static ssize_t aio_fdsync(struct kiocb *iocb) 1362 { 1363 struct file *file = iocb->ki_filp; 1364 ssize_t ret = -EINVAL; 1365 1366 if (file->f_op->aio_fsync) 1367 ret = file->f_op->aio_fsync(iocb, 1); 1368 return ret; 1369 } 1370 1371 static ssize_t aio_fsync(struct kiocb *iocb) 1372 { 1373 struct file *file = iocb->ki_filp; 1374 ssize_t ret = -EINVAL; 1375 1376 if (file->f_op->aio_fsync) 1377 ret = file->f_op->aio_fsync(iocb, 0); 1378 return ret; 1379 } 1380 1381 /* 1382 * aio_setup_iocb: 1383 * Performs the initial checks and aio retry method 1384 * setup for the kiocb at the time of io submission. 1385 */ 1386 static ssize_t aio_setup_iocb(struct kiocb *kiocb) 1387 { 1388 struct file *file = kiocb->ki_filp; 1389 ssize_t ret = 0; 1390 1391 switch (kiocb->ki_opcode) { 1392 case IOCB_CMD_PREAD: 1393 ret = -EBADF; 1394 if (unlikely(!(file->f_mode & FMODE_READ))) 1395 break; 1396 ret = -EFAULT; 1397 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf, 1398 kiocb->ki_left))) 1399 break; 1400 ret = security_file_permission(file, MAY_READ); 1401 if (unlikely(ret)) 1402 break; 1403 ret = -EINVAL; 1404 if (file->f_op->aio_read) 1405 kiocb->ki_retry = aio_pread; 1406 break; 1407 case IOCB_CMD_PWRITE: 1408 ret = -EBADF; 1409 if (unlikely(!(file->f_mode & FMODE_WRITE))) 1410 break; 1411 ret = -EFAULT; 1412 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf, 1413 kiocb->ki_left))) 1414 break; 1415 ret = security_file_permission(file, MAY_WRITE); 1416 if (unlikely(ret)) 1417 break; 1418 ret = -EINVAL; 1419 if (file->f_op->aio_write) 1420 kiocb->ki_retry = aio_pwrite; 1421 break; 1422 case IOCB_CMD_FDSYNC: 1423 ret = -EINVAL; 1424 if (file->f_op->aio_fsync) 1425 kiocb->ki_retry = aio_fdsync; 1426 break; 1427 case IOCB_CMD_FSYNC: 1428 ret = -EINVAL; 1429 if (file->f_op->aio_fsync) 1430 kiocb->ki_retry = aio_fsync; 1431 break; 1432 default: 1433 dprintk("EINVAL: io_submit: no operation provided\n"); 1434 ret = -EINVAL; 1435 } 1436 1437 if (!kiocb->ki_retry) 1438 return ret; 1439 1440 return 0; 1441 } 1442 1443 /* 1444 * aio_wake_function: 1445 * wait queue callback function for aio notification, 1446 * Simply triggers a retry of the operation via kick_iocb. 1447 * 1448 * This callback is specified in the wait queue entry in 1449 * a kiocb (current->io_wait points to this wait queue 1450 * entry when an aio operation executes; it is used 1451 * instead of a synchronous wait when an i/o blocking 1452 * condition is encountered during aio). 1453 * 1454 * Note: 1455 * This routine is executed with the wait queue lock held. 1456 * Since kick_iocb acquires iocb->ctx->ctx_lock, it nests 1457 * the ioctx lock inside the wait queue lock. This is safe 1458 * because this callback isn't used for wait queues which 1459 * are nested inside ioctx lock (i.e. ctx->wait) 1460 */ 1461 static int aio_wake_function(wait_queue_t *wait, unsigned mode, 1462 int sync, void *key) 1463 { 1464 struct kiocb *iocb = container_of(wait, struct kiocb, ki_wait); 1465 1466 list_del_init(&wait->task_list); 1467 kick_iocb(iocb); 1468 return 1; 1469 } 1470 1471 int fastcall io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb, 1472 struct iocb *iocb) 1473 { 1474 struct kiocb *req; 1475 struct file *file; 1476 ssize_t ret; 1477 1478 /* enforce forwards compatibility on users */ 1479 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2 || 1480 iocb->aio_reserved3)) { 1481 pr_debug("EINVAL: io_submit: reserve field set\n"); 1482 return -EINVAL; 1483 } 1484 1485 /* prevent overflows */ 1486 if (unlikely( 1487 (iocb->aio_buf != (unsigned long)iocb->aio_buf) || 1488 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) || 1489 ((ssize_t)iocb->aio_nbytes < 0) 1490 )) { 1491 pr_debug("EINVAL: io_submit: overflow check\n"); 1492 return -EINVAL; 1493 } 1494 1495 file = fget(iocb->aio_fildes); 1496 if (unlikely(!file)) 1497 return -EBADF; 1498 1499 req = aio_get_req(ctx); /* returns with 2 references to req */ 1500 if (unlikely(!req)) { 1501 fput(file); 1502 return -EAGAIN; 1503 } 1504 1505 req->ki_filp = file; 1506 ret = put_user(req->ki_key, &user_iocb->aio_key); 1507 if (unlikely(ret)) { 1508 dprintk("EFAULT: aio_key\n"); 1509 goto out_put_req; 1510 } 1511 1512 req->ki_obj.user = user_iocb; 1513 req->ki_user_data = iocb->aio_data; 1514 req->ki_pos = iocb->aio_offset; 1515 1516 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf; 1517 req->ki_left = req->ki_nbytes = iocb->aio_nbytes; 1518 req->ki_opcode = iocb->aio_lio_opcode; 1519 init_waitqueue_func_entry(&req->ki_wait, aio_wake_function); 1520 INIT_LIST_HEAD(&req->ki_wait.task_list); 1521 req->ki_retried = 0; 1522 1523 ret = aio_setup_iocb(req); 1524 1525 if (ret) 1526 goto out_put_req; 1527 1528 spin_lock_irq(&ctx->ctx_lock); 1529 aio_run_iocb(req); 1530 if (!list_empty(&ctx->run_list)) { 1531 /* drain the run list */ 1532 while (__aio_run_iocbs(ctx)) 1533 ; 1534 } 1535 spin_unlock_irq(&ctx->ctx_lock); 1536 aio_put_req(req); /* drop extra ref to req */ 1537 return 0; 1538 1539 out_put_req: 1540 aio_put_req(req); /* drop extra ref to req */ 1541 aio_put_req(req); /* drop i/o ref to req */ 1542 return ret; 1543 } 1544 1545 /* sys_io_submit: 1546 * Queue the nr iocbs pointed to by iocbpp for processing. Returns 1547 * the number of iocbs queued. May return -EINVAL if the aio_context 1548 * specified by ctx_id is invalid, if nr is < 0, if the iocb at 1549 * *iocbpp[0] is not properly initialized, if the operation specified 1550 * is invalid for the file descriptor in the iocb. May fail with 1551 * -EFAULT if any of the data structures point to invalid data. May 1552 * fail with -EBADF if the file descriptor specified in the first 1553 * iocb is invalid. May fail with -EAGAIN if insufficient resources 1554 * are available to queue any iocbs. Will return 0 if nr is 0. Will 1555 * fail with -ENOSYS if not implemented. 1556 */ 1557 asmlinkage long sys_io_submit(aio_context_t ctx_id, long nr, 1558 struct iocb __user * __user *iocbpp) 1559 { 1560 struct kioctx *ctx; 1561 long ret = 0; 1562 int i; 1563 1564 if (unlikely(nr < 0)) 1565 return -EINVAL; 1566 1567 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp))))) 1568 return -EFAULT; 1569 1570 ctx = lookup_ioctx(ctx_id); 1571 if (unlikely(!ctx)) { 1572 pr_debug("EINVAL: io_submit: invalid context id\n"); 1573 return -EINVAL; 1574 } 1575 1576 /* 1577 * AKPM: should this return a partial result if some of the IOs were 1578 * successfully submitted? 1579 */ 1580 for (i=0; i<nr; i++) { 1581 struct iocb __user *user_iocb; 1582 struct iocb tmp; 1583 1584 if (unlikely(__get_user(user_iocb, iocbpp + i))) { 1585 ret = -EFAULT; 1586 break; 1587 } 1588 1589 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) { 1590 ret = -EFAULT; 1591 break; 1592 } 1593 1594 ret = io_submit_one(ctx, user_iocb, &tmp); 1595 if (ret) 1596 break; 1597 } 1598 1599 put_ioctx(ctx); 1600 return i ? i : ret; 1601 } 1602 1603 /* lookup_kiocb 1604 * Finds a given iocb for cancellation. 1605 * MUST be called with ctx->ctx_lock held. 1606 */ 1607 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, 1608 u32 key) 1609 { 1610 struct list_head *pos; 1611 /* TODO: use a hash or array, this sucks. */ 1612 list_for_each(pos, &ctx->active_reqs) { 1613 struct kiocb *kiocb = list_kiocb(pos); 1614 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key) 1615 return kiocb; 1616 } 1617 return NULL; 1618 } 1619 1620 /* sys_io_cancel: 1621 * Attempts to cancel an iocb previously passed to io_submit. If 1622 * the operation is successfully cancelled, the resulting event is 1623 * copied into the memory pointed to by result without being placed 1624 * into the completion queue and 0 is returned. May fail with 1625 * -EFAULT if any of the data structures pointed to are invalid. 1626 * May fail with -EINVAL if aio_context specified by ctx_id is 1627 * invalid. May fail with -EAGAIN if the iocb specified was not 1628 * cancelled. Will fail with -ENOSYS if not implemented. 1629 */ 1630 asmlinkage long sys_io_cancel(aio_context_t ctx_id, struct iocb __user *iocb, 1631 struct io_event __user *result) 1632 { 1633 int (*cancel)(struct kiocb *iocb, struct io_event *res); 1634 struct kioctx *ctx; 1635 struct kiocb *kiocb; 1636 u32 key; 1637 int ret; 1638 1639 ret = get_user(key, &iocb->aio_key); 1640 if (unlikely(ret)) 1641 return -EFAULT; 1642 1643 ctx = lookup_ioctx(ctx_id); 1644 if (unlikely(!ctx)) 1645 return -EINVAL; 1646 1647 spin_lock_irq(&ctx->ctx_lock); 1648 ret = -EAGAIN; 1649 kiocb = lookup_kiocb(ctx, iocb, key); 1650 if (kiocb && kiocb->ki_cancel) { 1651 cancel = kiocb->ki_cancel; 1652 kiocb->ki_users ++; 1653 kiocbSetCancelled(kiocb); 1654 } else 1655 cancel = NULL; 1656 spin_unlock_irq(&ctx->ctx_lock); 1657 1658 if (NULL != cancel) { 1659 struct io_event tmp; 1660 pr_debug("calling cancel\n"); 1661 memset(&tmp, 0, sizeof(tmp)); 1662 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user; 1663 tmp.data = kiocb->ki_user_data; 1664 ret = cancel(kiocb, &tmp); 1665 if (!ret) { 1666 /* Cancellation succeeded -- copy the result 1667 * into the user's buffer. 1668 */ 1669 if (copy_to_user(result, &tmp, sizeof(tmp))) 1670 ret = -EFAULT; 1671 } 1672 } else 1673 ret = -EINVAL; 1674 1675 put_ioctx(ctx); 1676 1677 return ret; 1678 } 1679 1680 /* io_getevents: 1681 * Attempts to read at least min_nr events and up to nr events from 1682 * the completion queue for the aio_context specified by ctx_id. May 1683 * fail with -EINVAL if ctx_id is invalid, if min_nr is out of range, 1684 * if nr is out of range, if when is out of range. May fail with 1685 * -EFAULT if any of the memory specified to is invalid. May return 1686 * 0 or < min_nr if no events are available and the timeout specified 1687 * by when has elapsed, where when == NULL specifies an infinite 1688 * timeout. Note that the timeout pointed to by when is relative and 1689 * will be updated if not NULL and the operation blocks. Will fail 1690 * with -ENOSYS if not implemented. 1691 */ 1692 asmlinkage long sys_io_getevents(aio_context_t ctx_id, 1693 long min_nr, 1694 long nr, 1695 struct io_event __user *events, 1696 struct timespec __user *timeout) 1697 { 1698 struct kioctx *ioctx = lookup_ioctx(ctx_id); 1699 long ret = -EINVAL; 1700 1701 if (likely(ioctx)) { 1702 if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0)) 1703 ret = read_events(ioctx, min_nr, nr, events, timeout); 1704 put_ioctx(ioctx); 1705 } 1706 1707 return ret; 1708 } 1709 1710 __initcall(aio_setup); 1711 1712 EXPORT_SYMBOL(aio_complete); 1713 EXPORT_SYMBOL(aio_put_req); 1714 EXPORT_SYMBOL(wait_on_sync_kiocb); 1715