1 /* 2 * fs/direct-io.c 3 * 4 * Copyright (C) 2002, Linus Torvalds. 5 * 6 * O_DIRECT 7 * 8 * 04Jul2002 Andrew Morton 9 * Initial version 10 * 11Sep2002 janetinc@us.ibm.com 11 * added readv/writev support. 12 * 29Oct2002 Andrew Morton 13 * rewrote bio_add_page() support. 14 * 30Oct2002 pbadari@us.ibm.com 15 * added support for non-aligned IO. 16 * 06Nov2002 pbadari@us.ibm.com 17 * added asynchronous IO support. 18 * 21Jul2003 nathans@sgi.com 19 * added IO completion notifier. 20 */ 21 22 #include <linux/kernel.h> 23 #include <linux/module.h> 24 #include <linux/types.h> 25 #include <linux/fs.h> 26 #include <linux/mm.h> 27 #include <linux/slab.h> 28 #include <linux/highmem.h> 29 #include <linux/pagemap.h> 30 #include <linux/task_io_accounting_ops.h> 31 #include <linux/bio.h> 32 #include <linux/wait.h> 33 #include <linux/err.h> 34 #include <linux/blkdev.h> 35 #include <linux/buffer_head.h> 36 #include <linux/rwsem.h> 37 #include <linux/uio.h> 38 #include <linux/atomic.h> 39 #include <linux/prefetch.h> 40 41 /* 42 * How many user pages to map in one call to get_user_pages(). This determines 43 * the size of a structure in the slab cache 44 */ 45 #define DIO_PAGES 64 46 47 /* 48 * This code generally works in units of "dio_blocks". A dio_block is 49 * somewhere between the hard sector size and the filesystem block size. it 50 * is determined on a per-invocation basis. When talking to the filesystem 51 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity 52 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted 53 * to bio_block quantities by shifting left by blkfactor. 54 * 55 * If blkfactor is zero then the user's request was aligned to the filesystem's 56 * blocksize. 57 */ 58 59 /* dio_state only used in the submission path */ 60 61 struct dio_submit { 62 struct bio *bio; /* bio under assembly */ 63 unsigned blkbits; /* doesn't change */ 64 unsigned blkfactor; /* When we're using an alignment which 65 is finer than the filesystem's soft 66 blocksize, this specifies how much 67 finer. blkfactor=2 means 1/4-block 68 alignment. Does not change */ 69 unsigned start_zero_done; /* flag: sub-blocksize zeroing has 70 been performed at the start of a 71 write */ 72 int pages_in_io; /* approximate total IO pages */ 73 sector_t block_in_file; /* Current offset into the underlying 74 file in dio_block units. */ 75 unsigned blocks_available; /* At block_in_file. changes */ 76 int reap_counter; /* rate limit reaping */ 77 sector_t final_block_in_request;/* doesn't change */ 78 int boundary; /* prev block is at a boundary */ 79 get_block_t *get_block; /* block mapping function */ 80 dio_submit_t *submit_io; /* IO submition function */ 81 82 loff_t logical_offset_in_bio; /* current first logical block in bio */ 83 sector_t final_block_in_bio; /* current final block in bio + 1 */ 84 sector_t next_block_for_io; /* next block to be put under IO, 85 in dio_blocks units */ 86 87 /* 88 * Deferred addition of a page to the dio. These variables are 89 * private to dio_send_cur_page(), submit_page_section() and 90 * dio_bio_add_page(). 91 */ 92 struct page *cur_page; /* The page */ 93 unsigned cur_page_offset; /* Offset into it, in bytes */ 94 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */ 95 sector_t cur_page_block; /* Where it starts */ 96 loff_t cur_page_fs_offset; /* Offset in file */ 97 98 struct iov_iter *iter; 99 /* 100 * Page queue. These variables belong to dio_refill_pages() and 101 * dio_get_page(). 102 */ 103 unsigned head; /* next page to process */ 104 unsigned tail; /* last valid page + 1 */ 105 size_t from, to; 106 }; 107 108 /* dio_state communicated between submission path and end_io */ 109 struct dio { 110 int flags; /* doesn't change */ 111 int rw; 112 struct inode *inode; 113 loff_t i_size; /* i_size when submitted */ 114 dio_iodone_t *end_io; /* IO completion function */ 115 116 void *private; /* copy from map_bh.b_private */ 117 118 /* BIO completion state */ 119 spinlock_t bio_lock; /* protects BIO fields below */ 120 int page_errors; /* errno from get_user_pages() */ 121 int is_async; /* is IO async ? */ 122 bool defer_completion; /* defer AIO completion to workqueue? */ 123 bool should_dirty; /* if pages should be dirtied */ 124 int io_error; /* IO error in completion path */ 125 unsigned long refcount; /* direct_io_worker() and bios */ 126 struct bio *bio_list; /* singly linked via bi_private */ 127 struct task_struct *waiter; /* waiting task (NULL if none) */ 128 129 /* AIO related stuff */ 130 struct kiocb *iocb; /* kiocb */ 131 ssize_t result; /* IO result */ 132 133 /* 134 * pages[] (and any fields placed after it) are not zeroed out at 135 * allocation time. Don't add new fields after pages[] unless you 136 * wish that they not be zeroed. 137 */ 138 union { 139 struct page *pages[DIO_PAGES]; /* page buffer */ 140 struct work_struct complete_work;/* deferred AIO completion */ 141 }; 142 } ____cacheline_aligned_in_smp; 143 144 static struct kmem_cache *dio_cache __read_mostly; 145 146 /* 147 * How many pages are in the queue? 148 */ 149 static inline unsigned dio_pages_present(struct dio_submit *sdio) 150 { 151 return sdio->tail - sdio->head; 152 } 153 154 /* 155 * Go grab and pin some userspace pages. Typically we'll get 64 at a time. 156 */ 157 static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio) 158 { 159 ssize_t ret; 160 161 ret = iov_iter_get_pages(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES, 162 &sdio->from); 163 164 if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) { 165 struct page *page = ZERO_PAGE(0); 166 /* 167 * A memory fault, but the filesystem has some outstanding 168 * mapped blocks. We need to use those blocks up to avoid 169 * leaking stale data in the file. 170 */ 171 if (dio->page_errors == 0) 172 dio->page_errors = ret; 173 page_cache_get(page); 174 dio->pages[0] = page; 175 sdio->head = 0; 176 sdio->tail = 1; 177 sdio->from = 0; 178 sdio->to = PAGE_SIZE; 179 return 0; 180 } 181 182 if (ret >= 0) { 183 iov_iter_advance(sdio->iter, ret); 184 ret += sdio->from; 185 sdio->head = 0; 186 sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE; 187 sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1; 188 return 0; 189 } 190 return ret; 191 } 192 193 /* 194 * Get another userspace page. Returns an ERR_PTR on error. Pages are 195 * buffered inside the dio so that we can call get_user_pages() against a 196 * decent number of pages, less frequently. To provide nicer use of the 197 * L1 cache. 198 */ 199 static inline struct page *dio_get_page(struct dio *dio, 200 struct dio_submit *sdio) 201 { 202 if (dio_pages_present(sdio) == 0) { 203 int ret; 204 205 ret = dio_refill_pages(dio, sdio); 206 if (ret) 207 return ERR_PTR(ret); 208 BUG_ON(dio_pages_present(sdio) == 0); 209 } 210 return dio->pages[sdio->head]; 211 } 212 213 /** 214 * dio_complete() - called when all DIO BIO I/O has been completed 215 * @offset: the byte offset in the file of the completed operation 216 * 217 * This drops i_dio_count, lets interested parties know that a DIO operation 218 * has completed, and calculates the resulting return code for the operation. 219 * 220 * It lets the filesystem know if it registered an interest earlier via 221 * get_block. Pass the private field of the map buffer_head so that 222 * filesystems can use it to hold additional state between get_block calls and 223 * dio_complete. 224 */ 225 static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret, 226 bool is_async) 227 { 228 ssize_t transferred = 0; 229 230 /* 231 * AIO submission can race with bio completion to get here while 232 * expecting to have the last io completed by bio completion. 233 * In that case -EIOCBQUEUED is in fact not an error we want 234 * to preserve through this call. 235 */ 236 if (ret == -EIOCBQUEUED) 237 ret = 0; 238 239 if (dio->result) { 240 transferred = dio->result; 241 242 /* Check for short read case */ 243 if ((dio->rw == READ) && ((offset + transferred) > dio->i_size)) 244 transferred = dio->i_size - offset; 245 } 246 247 if (ret == 0) 248 ret = dio->page_errors; 249 if (ret == 0) 250 ret = dio->io_error; 251 if (ret == 0) 252 ret = transferred; 253 254 if (dio->end_io && dio->result) 255 dio->end_io(dio->iocb, offset, transferred, dio->private); 256 257 if (!(dio->flags & DIO_SKIP_DIO_COUNT)) 258 inode_dio_end(dio->inode); 259 260 if (is_async) { 261 if (dio->rw & WRITE) { 262 int err; 263 264 err = generic_write_sync(dio->iocb->ki_filp, offset, 265 transferred); 266 if (err < 0 && ret > 0) 267 ret = err; 268 } 269 270 dio->iocb->ki_complete(dio->iocb, ret, 0); 271 } 272 273 kmem_cache_free(dio_cache, dio); 274 return ret; 275 } 276 277 static void dio_aio_complete_work(struct work_struct *work) 278 { 279 struct dio *dio = container_of(work, struct dio, complete_work); 280 281 dio_complete(dio, dio->iocb->ki_pos, 0, true); 282 } 283 284 static int dio_bio_complete(struct dio *dio, struct bio *bio); 285 286 /* 287 * Asynchronous IO callback. 288 */ 289 static void dio_bio_end_aio(struct bio *bio) 290 { 291 struct dio *dio = bio->bi_private; 292 unsigned long remaining; 293 unsigned long flags; 294 295 /* cleanup the bio */ 296 dio_bio_complete(dio, bio); 297 298 spin_lock_irqsave(&dio->bio_lock, flags); 299 remaining = --dio->refcount; 300 if (remaining == 1 && dio->waiter) 301 wake_up_process(dio->waiter); 302 spin_unlock_irqrestore(&dio->bio_lock, flags); 303 304 if (remaining == 0) { 305 if (dio->result && dio->defer_completion) { 306 INIT_WORK(&dio->complete_work, dio_aio_complete_work); 307 queue_work(dio->inode->i_sb->s_dio_done_wq, 308 &dio->complete_work); 309 } else { 310 dio_complete(dio, dio->iocb->ki_pos, 0, true); 311 } 312 } 313 } 314 315 /* 316 * The BIO completion handler simply queues the BIO up for the process-context 317 * handler. 318 * 319 * During I/O bi_private points at the dio. After I/O, bi_private is used to 320 * implement a singly-linked list of completed BIOs, at dio->bio_list. 321 */ 322 static void dio_bio_end_io(struct bio *bio) 323 { 324 struct dio *dio = bio->bi_private; 325 unsigned long flags; 326 327 spin_lock_irqsave(&dio->bio_lock, flags); 328 bio->bi_private = dio->bio_list; 329 dio->bio_list = bio; 330 if (--dio->refcount == 1 && dio->waiter) 331 wake_up_process(dio->waiter); 332 spin_unlock_irqrestore(&dio->bio_lock, flags); 333 } 334 335 /** 336 * dio_end_io - handle the end io action for the given bio 337 * @bio: The direct io bio thats being completed 338 * @error: Error if there was one 339 * 340 * This is meant to be called by any filesystem that uses their own dio_submit_t 341 * so that the DIO specific endio actions are dealt with after the filesystem 342 * has done it's completion work. 343 */ 344 void dio_end_io(struct bio *bio, int error) 345 { 346 struct dio *dio = bio->bi_private; 347 348 if (dio->is_async) 349 dio_bio_end_aio(bio); 350 else 351 dio_bio_end_io(bio); 352 } 353 EXPORT_SYMBOL_GPL(dio_end_io); 354 355 static inline void 356 dio_bio_alloc(struct dio *dio, struct dio_submit *sdio, 357 struct block_device *bdev, 358 sector_t first_sector, int nr_vecs) 359 { 360 struct bio *bio; 361 362 /* 363 * bio_alloc() is guaranteed to return a bio when called with 364 * __GFP_RECLAIM and we request a valid number of vectors. 365 */ 366 bio = bio_alloc(GFP_KERNEL, nr_vecs); 367 368 bio->bi_bdev = bdev; 369 bio->bi_iter.bi_sector = first_sector; 370 if (dio->is_async) 371 bio->bi_end_io = dio_bio_end_aio; 372 else 373 bio->bi_end_io = dio_bio_end_io; 374 375 sdio->bio = bio; 376 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset; 377 } 378 379 /* 380 * In the AIO read case we speculatively dirty the pages before starting IO. 381 * During IO completion, any of these pages which happen to have been written 382 * back will be redirtied by bio_check_pages_dirty(). 383 * 384 * bios hold a dio reference between submit_bio and ->end_io. 385 */ 386 static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio) 387 { 388 struct bio *bio = sdio->bio; 389 unsigned long flags; 390 391 bio->bi_private = dio; 392 393 spin_lock_irqsave(&dio->bio_lock, flags); 394 dio->refcount++; 395 spin_unlock_irqrestore(&dio->bio_lock, flags); 396 397 if (dio->is_async && dio->rw == READ && dio->should_dirty) 398 bio_set_pages_dirty(bio); 399 400 if (sdio->submit_io) 401 sdio->submit_io(dio->rw, bio, dio->inode, 402 sdio->logical_offset_in_bio); 403 else 404 submit_bio(dio->rw, bio); 405 406 sdio->bio = NULL; 407 sdio->boundary = 0; 408 sdio->logical_offset_in_bio = 0; 409 } 410 411 /* 412 * Release any resources in case of a failure 413 */ 414 static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio) 415 { 416 while (sdio->head < sdio->tail) 417 page_cache_release(dio->pages[sdio->head++]); 418 } 419 420 /* 421 * Wait for the next BIO to complete. Remove it and return it. NULL is 422 * returned once all BIOs have been completed. This must only be called once 423 * all bios have been issued so that dio->refcount can only decrease. This 424 * requires that that the caller hold a reference on the dio. 425 */ 426 static struct bio *dio_await_one(struct dio *dio) 427 { 428 unsigned long flags; 429 struct bio *bio = NULL; 430 431 spin_lock_irqsave(&dio->bio_lock, flags); 432 433 /* 434 * Wait as long as the list is empty and there are bios in flight. bio 435 * completion drops the count, maybe adds to the list, and wakes while 436 * holding the bio_lock so we don't need set_current_state()'s barrier 437 * and can call it after testing our condition. 438 */ 439 while (dio->refcount > 1 && dio->bio_list == NULL) { 440 __set_current_state(TASK_UNINTERRUPTIBLE); 441 dio->waiter = current; 442 spin_unlock_irqrestore(&dio->bio_lock, flags); 443 io_schedule(); 444 /* wake up sets us TASK_RUNNING */ 445 spin_lock_irqsave(&dio->bio_lock, flags); 446 dio->waiter = NULL; 447 } 448 if (dio->bio_list) { 449 bio = dio->bio_list; 450 dio->bio_list = bio->bi_private; 451 } 452 spin_unlock_irqrestore(&dio->bio_lock, flags); 453 return bio; 454 } 455 456 /* 457 * Process one completed BIO. No locks are held. 458 */ 459 static int dio_bio_complete(struct dio *dio, struct bio *bio) 460 { 461 struct bio_vec *bvec; 462 unsigned i; 463 int err; 464 465 if (bio->bi_error) 466 dio->io_error = -EIO; 467 468 if (dio->is_async && dio->rw == READ && dio->should_dirty) { 469 bio_check_pages_dirty(bio); /* transfers ownership */ 470 err = bio->bi_error; 471 } else { 472 bio_for_each_segment_all(bvec, bio, i) { 473 struct page *page = bvec->bv_page; 474 475 if (dio->rw == READ && !PageCompound(page) && 476 dio->should_dirty) 477 set_page_dirty_lock(page); 478 page_cache_release(page); 479 } 480 err = bio->bi_error; 481 bio_put(bio); 482 } 483 return err; 484 } 485 486 /* 487 * Wait on and process all in-flight BIOs. This must only be called once 488 * all bios have been issued so that the refcount can only decrease. 489 * This just waits for all bios to make it through dio_bio_complete. IO 490 * errors are propagated through dio->io_error and should be propagated via 491 * dio_complete(). 492 */ 493 static void dio_await_completion(struct dio *dio) 494 { 495 struct bio *bio; 496 do { 497 bio = dio_await_one(dio); 498 if (bio) 499 dio_bio_complete(dio, bio); 500 } while (bio); 501 } 502 503 /* 504 * A really large O_DIRECT read or write can generate a lot of BIOs. So 505 * to keep the memory consumption sane we periodically reap any completed BIOs 506 * during the BIO generation phase. 507 * 508 * This also helps to limit the peak amount of pinned userspace memory. 509 */ 510 static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio) 511 { 512 int ret = 0; 513 514 if (sdio->reap_counter++ >= 64) { 515 while (dio->bio_list) { 516 unsigned long flags; 517 struct bio *bio; 518 int ret2; 519 520 spin_lock_irqsave(&dio->bio_lock, flags); 521 bio = dio->bio_list; 522 dio->bio_list = bio->bi_private; 523 spin_unlock_irqrestore(&dio->bio_lock, flags); 524 ret2 = dio_bio_complete(dio, bio); 525 if (ret == 0) 526 ret = ret2; 527 } 528 sdio->reap_counter = 0; 529 } 530 return ret; 531 } 532 533 /* 534 * Create workqueue for deferred direct IO completions. We allocate the 535 * workqueue when it's first needed. This avoids creating workqueue for 536 * filesystems that don't need it and also allows us to create the workqueue 537 * late enough so the we can include s_id in the name of the workqueue. 538 */ 539 static int sb_init_dio_done_wq(struct super_block *sb) 540 { 541 struct workqueue_struct *old; 542 struct workqueue_struct *wq = alloc_workqueue("dio/%s", 543 WQ_MEM_RECLAIM, 0, 544 sb->s_id); 545 if (!wq) 546 return -ENOMEM; 547 /* 548 * This has to be atomic as more DIOs can race to create the workqueue 549 */ 550 old = cmpxchg(&sb->s_dio_done_wq, NULL, wq); 551 /* Someone created workqueue before us? Free ours... */ 552 if (old) 553 destroy_workqueue(wq); 554 return 0; 555 } 556 557 static int dio_set_defer_completion(struct dio *dio) 558 { 559 struct super_block *sb = dio->inode->i_sb; 560 561 if (dio->defer_completion) 562 return 0; 563 dio->defer_completion = true; 564 if (!sb->s_dio_done_wq) 565 return sb_init_dio_done_wq(sb); 566 return 0; 567 } 568 569 /* 570 * Call into the fs to map some more disk blocks. We record the current number 571 * of available blocks at sdio->blocks_available. These are in units of the 572 * fs blocksize, (1 << inode->i_blkbits). 573 * 574 * The fs is allowed to map lots of blocks at once. If it wants to do that, 575 * it uses the passed inode-relative block number as the file offset, as usual. 576 * 577 * get_block() is passed the number of i_blkbits-sized blocks which direct_io 578 * has remaining to do. The fs should not map more than this number of blocks. 579 * 580 * If the fs has mapped a lot of blocks, it should populate bh->b_size to 581 * indicate how much contiguous disk space has been made available at 582 * bh->b_blocknr. 583 * 584 * If *any* of the mapped blocks are new, then the fs must set buffer_new(). 585 * This isn't very efficient... 586 * 587 * In the case of filesystem holes: the fs may return an arbitrarily-large 588 * hole by returning an appropriate value in b_size and by clearing 589 * buffer_mapped(). However the direct-io code will only process holes one 590 * block at a time - it will repeatedly call get_block() as it walks the hole. 591 */ 592 static int get_more_blocks(struct dio *dio, struct dio_submit *sdio, 593 struct buffer_head *map_bh) 594 { 595 int ret; 596 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */ 597 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */ 598 unsigned long fs_count; /* Number of filesystem-sized blocks */ 599 int create; 600 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor; 601 602 /* 603 * If there was a memory error and we've overwritten all the 604 * mapped blocks then we can now return that memory error 605 */ 606 ret = dio->page_errors; 607 if (ret == 0) { 608 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request); 609 fs_startblk = sdio->block_in_file >> sdio->blkfactor; 610 fs_endblk = (sdio->final_block_in_request - 1) >> 611 sdio->blkfactor; 612 fs_count = fs_endblk - fs_startblk + 1; 613 614 map_bh->b_state = 0; 615 map_bh->b_size = fs_count << i_blkbits; 616 617 /* 618 * For writes inside i_size on a DIO_SKIP_HOLES filesystem we 619 * forbid block creations: only overwrites are permitted. 620 * We will return early to the caller once we see an 621 * unmapped buffer head returned, and the caller will fall 622 * back to buffered I/O. 623 * 624 * Otherwise the decision is left to the get_blocks method, 625 * which may decide to handle it or also return an unmapped 626 * buffer head. 627 */ 628 create = dio->rw & WRITE; 629 if (dio->flags & DIO_SKIP_HOLES) { 630 if (sdio->block_in_file < (i_size_read(dio->inode) >> 631 sdio->blkbits)) 632 create = 0; 633 } 634 635 ret = (*sdio->get_block)(dio->inode, fs_startblk, 636 map_bh, create); 637 638 /* Store for completion */ 639 dio->private = map_bh->b_private; 640 641 if (ret == 0 && buffer_defer_completion(map_bh)) 642 ret = dio_set_defer_completion(dio); 643 } 644 return ret; 645 } 646 647 /* 648 * There is no bio. Make one now. 649 */ 650 static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio, 651 sector_t start_sector, struct buffer_head *map_bh) 652 { 653 sector_t sector; 654 int ret, nr_pages; 655 656 ret = dio_bio_reap(dio, sdio); 657 if (ret) 658 goto out; 659 sector = start_sector << (sdio->blkbits - 9); 660 nr_pages = min(sdio->pages_in_io, BIO_MAX_PAGES); 661 BUG_ON(nr_pages <= 0); 662 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages); 663 sdio->boundary = 0; 664 out: 665 return ret; 666 } 667 668 /* 669 * Attempt to put the current chunk of 'cur_page' into the current BIO. If 670 * that was successful then update final_block_in_bio and take a ref against 671 * the just-added page. 672 * 673 * Return zero on success. Non-zero means the caller needs to start a new BIO. 674 */ 675 static inline int dio_bio_add_page(struct dio_submit *sdio) 676 { 677 int ret; 678 679 ret = bio_add_page(sdio->bio, sdio->cur_page, 680 sdio->cur_page_len, sdio->cur_page_offset); 681 if (ret == sdio->cur_page_len) { 682 /* 683 * Decrement count only, if we are done with this page 684 */ 685 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE) 686 sdio->pages_in_io--; 687 page_cache_get(sdio->cur_page); 688 sdio->final_block_in_bio = sdio->cur_page_block + 689 (sdio->cur_page_len >> sdio->blkbits); 690 ret = 0; 691 } else { 692 ret = 1; 693 } 694 return ret; 695 } 696 697 /* 698 * Put cur_page under IO. The section of cur_page which is described by 699 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page 700 * starts on-disk at cur_page_block. 701 * 702 * We take a ref against the page here (on behalf of its presence in the bio). 703 * 704 * The caller of this function is responsible for removing cur_page from the 705 * dio, and for dropping the refcount which came from that presence. 706 */ 707 static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio, 708 struct buffer_head *map_bh) 709 { 710 int ret = 0; 711 712 if (sdio->bio) { 713 loff_t cur_offset = sdio->cur_page_fs_offset; 714 loff_t bio_next_offset = sdio->logical_offset_in_bio + 715 sdio->bio->bi_iter.bi_size; 716 717 /* 718 * See whether this new request is contiguous with the old. 719 * 720 * Btrfs cannot handle having logically non-contiguous requests 721 * submitted. For example if you have 722 * 723 * Logical: [0-4095][HOLE][8192-12287] 724 * Physical: [0-4095] [4096-8191] 725 * 726 * We cannot submit those pages together as one BIO. So if our 727 * current logical offset in the file does not equal what would 728 * be the next logical offset in the bio, submit the bio we 729 * have. 730 */ 731 if (sdio->final_block_in_bio != sdio->cur_page_block || 732 cur_offset != bio_next_offset) 733 dio_bio_submit(dio, sdio); 734 } 735 736 if (sdio->bio == NULL) { 737 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh); 738 if (ret) 739 goto out; 740 } 741 742 if (dio_bio_add_page(sdio) != 0) { 743 dio_bio_submit(dio, sdio); 744 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh); 745 if (ret == 0) { 746 ret = dio_bio_add_page(sdio); 747 BUG_ON(ret != 0); 748 } 749 } 750 out: 751 return ret; 752 } 753 754 /* 755 * An autonomous function to put a chunk of a page under deferred IO. 756 * 757 * The caller doesn't actually know (or care) whether this piece of page is in 758 * a BIO, or is under IO or whatever. We just take care of all possible 759 * situations here. The separation between the logic of do_direct_IO() and 760 * that of submit_page_section() is important for clarity. Please don't break. 761 * 762 * The chunk of page starts on-disk at blocknr. 763 * 764 * We perform deferred IO, by recording the last-submitted page inside our 765 * private part of the dio structure. If possible, we just expand the IO 766 * across that page here. 767 * 768 * If that doesn't work out then we put the old page into the bio and add this 769 * page to the dio instead. 770 */ 771 static inline int 772 submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page, 773 unsigned offset, unsigned len, sector_t blocknr, 774 struct buffer_head *map_bh) 775 { 776 int ret = 0; 777 778 if (dio->rw & WRITE) { 779 /* 780 * Read accounting is performed in submit_bio() 781 */ 782 task_io_account_write(len); 783 } 784 785 /* 786 * Can we just grow the current page's presence in the dio? 787 */ 788 if (sdio->cur_page == page && 789 sdio->cur_page_offset + sdio->cur_page_len == offset && 790 sdio->cur_page_block + 791 (sdio->cur_page_len >> sdio->blkbits) == blocknr) { 792 sdio->cur_page_len += len; 793 goto out; 794 } 795 796 /* 797 * If there's a deferred page already there then send it. 798 */ 799 if (sdio->cur_page) { 800 ret = dio_send_cur_page(dio, sdio, map_bh); 801 page_cache_release(sdio->cur_page); 802 sdio->cur_page = NULL; 803 if (ret) 804 return ret; 805 } 806 807 page_cache_get(page); /* It is in dio */ 808 sdio->cur_page = page; 809 sdio->cur_page_offset = offset; 810 sdio->cur_page_len = len; 811 sdio->cur_page_block = blocknr; 812 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits; 813 out: 814 /* 815 * If sdio->boundary then we want to schedule the IO now to 816 * avoid metadata seeks. 817 */ 818 if (sdio->boundary) { 819 ret = dio_send_cur_page(dio, sdio, map_bh); 820 dio_bio_submit(dio, sdio); 821 page_cache_release(sdio->cur_page); 822 sdio->cur_page = NULL; 823 } 824 return ret; 825 } 826 827 /* 828 * Clean any dirty buffers in the blockdev mapping which alias newly-created 829 * file blocks. Only called for S_ISREG files - blockdevs do not set 830 * buffer_new 831 */ 832 static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh) 833 { 834 unsigned i; 835 unsigned nblocks; 836 837 nblocks = map_bh->b_size >> dio->inode->i_blkbits; 838 839 for (i = 0; i < nblocks; i++) { 840 unmap_underlying_metadata(map_bh->b_bdev, 841 map_bh->b_blocknr + i); 842 } 843 } 844 845 /* 846 * If we are not writing the entire block and get_block() allocated 847 * the block for us, we need to fill-in the unused portion of the 848 * block with zeros. This happens only if user-buffer, fileoffset or 849 * io length is not filesystem block-size multiple. 850 * 851 * `end' is zero if we're doing the start of the IO, 1 at the end of the 852 * IO. 853 */ 854 static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio, 855 int end, struct buffer_head *map_bh) 856 { 857 unsigned dio_blocks_per_fs_block; 858 unsigned this_chunk_blocks; /* In dio_blocks */ 859 unsigned this_chunk_bytes; 860 struct page *page; 861 862 sdio->start_zero_done = 1; 863 if (!sdio->blkfactor || !buffer_new(map_bh)) 864 return; 865 866 dio_blocks_per_fs_block = 1 << sdio->blkfactor; 867 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1); 868 869 if (!this_chunk_blocks) 870 return; 871 872 /* 873 * We need to zero out part of an fs block. It is either at the 874 * beginning or the end of the fs block. 875 */ 876 if (end) 877 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks; 878 879 this_chunk_bytes = this_chunk_blocks << sdio->blkbits; 880 881 page = ZERO_PAGE(0); 882 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes, 883 sdio->next_block_for_io, map_bh)) 884 return; 885 886 sdio->next_block_for_io += this_chunk_blocks; 887 } 888 889 /* 890 * Walk the user pages, and the file, mapping blocks to disk and generating 891 * a sequence of (page,offset,len,block) mappings. These mappings are injected 892 * into submit_page_section(), which takes care of the next stage of submission 893 * 894 * Direct IO against a blockdev is different from a file. Because we can 895 * happily perform page-sized but 512-byte aligned IOs. It is important that 896 * blockdev IO be able to have fine alignment and large sizes. 897 * 898 * So what we do is to permit the ->get_block function to populate bh.b_size 899 * with the size of IO which is permitted at this offset and this i_blkbits. 900 * 901 * For best results, the blockdev should be set up with 512-byte i_blkbits and 902 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives 903 * fine alignment but still allows this function to work in PAGE_SIZE units. 904 */ 905 static int do_direct_IO(struct dio *dio, struct dio_submit *sdio, 906 struct buffer_head *map_bh) 907 { 908 const unsigned blkbits = sdio->blkbits; 909 int ret = 0; 910 911 while (sdio->block_in_file < sdio->final_block_in_request) { 912 struct page *page; 913 size_t from, to; 914 915 page = dio_get_page(dio, sdio); 916 if (IS_ERR(page)) { 917 ret = PTR_ERR(page); 918 goto out; 919 } 920 from = sdio->head ? 0 : sdio->from; 921 to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE; 922 sdio->head++; 923 924 while (from < to) { 925 unsigned this_chunk_bytes; /* # of bytes mapped */ 926 unsigned this_chunk_blocks; /* # of blocks */ 927 unsigned u; 928 929 if (sdio->blocks_available == 0) { 930 /* 931 * Need to go and map some more disk 932 */ 933 unsigned long blkmask; 934 unsigned long dio_remainder; 935 936 ret = get_more_blocks(dio, sdio, map_bh); 937 if (ret) { 938 page_cache_release(page); 939 goto out; 940 } 941 if (!buffer_mapped(map_bh)) 942 goto do_holes; 943 944 sdio->blocks_available = 945 map_bh->b_size >> sdio->blkbits; 946 sdio->next_block_for_io = 947 map_bh->b_blocknr << sdio->blkfactor; 948 if (buffer_new(map_bh)) 949 clean_blockdev_aliases(dio, map_bh); 950 951 if (!sdio->blkfactor) 952 goto do_holes; 953 954 blkmask = (1 << sdio->blkfactor) - 1; 955 dio_remainder = (sdio->block_in_file & blkmask); 956 957 /* 958 * If we are at the start of IO and that IO 959 * starts partway into a fs-block, 960 * dio_remainder will be non-zero. If the IO 961 * is a read then we can simply advance the IO 962 * cursor to the first block which is to be 963 * read. But if the IO is a write and the 964 * block was newly allocated we cannot do that; 965 * the start of the fs block must be zeroed out 966 * on-disk 967 */ 968 if (!buffer_new(map_bh)) 969 sdio->next_block_for_io += dio_remainder; 970 sdio->blocks_available -= dio_remainder; 971 } 972 do_holes: 973 /* Handle holes */ 974 if (!buffer_mapped(map_bh)) { 975 loff_t i_size_aligned; 976 977 /* AKPM: eargh, -ENOTBLK is a hack */ 978 if (dio->rw & WRITE) { 979 page_cache_release(page); 980 return -ENOTBLK; 981 } 982 983 /* 984 * Be sure to account for a partial block as the 985 * last block in the file 986 */ 987 i_size_aligned = ALIGN(i_size_read(dio->inode), 988 1 << blkbits); 989 if (sdio->block_in_file >= 990 i_size_aligned >> blkbits) { 991 /* We hit eof */ 992 page_cache_release(page); 993 goto out; 994 } 995 zero_user(page, from, 1 << blkbits); 996 sdio->block_in_file++; 997 from += 1 << blkbits; 998 dio->result += 1 << blkbits; 999 goto next_block; 1000 } 1001 1002 /* 1003 * If we're performing IO which has an alignment which 1004 * is finer than the underlying fs, go check to see if 1005 * we must zero out the start of this block. 1006 */ 1007 if (unlikely(sdio->blkfactor && !sdio->start_zero_done)) 1008 dio_zero_block(dio, sdio, 0, map_bh); 1009 1010 /* 1011 * Work out, in this_chunk_blocks, how much disk we 1012 * can add to this page 1013 */ 1014 this_chunk_blocks = sdio->blocks_available; 1015 u = (to - from) >> blkbits; 1016 if (this_chunk_blocks > u) 1017 this_chunk_blocks = u; 1018 u = sdio->final_block_in_request - sdio->block_in_file; 1019 if (this_chunk_blocks > u) 1020 this_chunk_blocks = u; 1021 this_chunk_bytes = this_chunk_blocks << blkbits; 1022 BUG_ON(this_chunk_bytes == 0); 1023 1024 if (this_chunk_blocks == sdio->blocks_available) 1025 sdio->boundary = buffer_boundary(map_bh); 1026 ret = submit_page_section(dio, sdio, page, 1027 from, 1028 this_chunk_bytes, 1029 sdio->next_block_for_io, 1030 map_bh); 1031 if (ret) { 1032 page_cache_release(page); 1033 goto out; 1034 } 1035 sdio->next_block_for_io += this_chunk_blocks; 1036 1037 sdio->block_in_file += this_chunk_blocks; 1038 from += this_chunk_bytes; 1039 dio->result += this_chunk_bytes; 1040 sdio->blocks_available -= this_chunk_blocks; 1041 next_block: 1042 BUG_ON(sdio->block_in_file > sdio->final_block_in_request); 1043 if (sdio->block_in_file == sdio->final_block_in_request) 1044 break; 1045 } 1046 1047 /* Drop the ref which was taken in get_user_pages() */ 1048 page_cache_release(page); 1049 } 1050 out: 1051 return ret; 1052 } 1053 1054 static inline int drop_refcount(struct dio *dio) 1055 { 1056 int ret2; 1057 unsigned long flags; 1058 1059 /* 1060 * Sync will always be dropping the final ref and completing the 1061 * operation. AIO can if it was a broken operation described above or 1062 * in fact if all the bios race to complete before we get here. In 1063 * that case dio_complete() translates the EIOCBQUEUED into the proper 1064 * return code that the caller will hand to ->complete(). 1065 * 1066 * This is managed by the bio_lock instead of being an atomic_t so that 1067 * completion paths can drop their ref and use the remaining count to 1068 * decide to wake the submission path atomically. 1069 */ 1070 spin_lock_irqsave(&dio->bio_lock, flags); 1071 ret2 = --dio->refcount; 1072 spin_unlock_irqrestore(&dio->bio_lock, flags); 1073 return ret2; 1074 } 1075 1076 /* 1077 * This is a library function for use by filesystem drivers. 1078 * 1079 * The locking rules are governed by the flags parameter: 1080 * - if the flags value contains DIO_LOCKING we use a fancy locking 1081 * scheme for dumb filesystems. 1082 * For writes this function is called under i_mutex and returns with 1083 * i_mutex held, for reads, i_mutex is not held on entry, but it is 1084 * taken and dropped again before returning. 1085 * - if the flags value does NOT contain DIO_LOCKING we don't use any 1086 * internal locking but rather rely on the filesystem to synchronize 1087 * direct I/O reads/writes versus each other and truncate. 1088 * 1089 * To help with locking against truncate we incremented the i_dio_count 1090 * counter before starting direct I/O, and decrement it once we are done. 1091 * Truncate can wait for it to reach zero to provide exclusion. It is 1092 * expected that filesystem provide exclusion between new direct I/O 1093 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex, 1094 * but other filesystems need to take care of this on their own. 1095 * 1096 * NOTE: if you pass "sdio" to anything by pointer make sure that function 1097 * is always inlined. Otherwise gcc is unable to split the structure into 1098 * individual fields and will generate much worse code. This is important 1099 * for the whole file. 1100 */ 1101 static inline ssize_t 1102 do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode, 1103 struct block_device *bdev, struct iov_iter *iter, 1104 loff_t offset, get_block_t get_block, dio_iodone_t end_io, 1105 dio_submit_t submit_io, int flags) 1106 { 1107 unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits); 1108 unsigned blkbits = i_blkbits; 1109 unsigned blocksize_mask = (1 << blkbits) - 1; 1110 ssize_t retval = -EINVAL; 1111 size_t count = iov_iter_count(iter); 1112 loff_t end = offset + count; 1113 struct dio *dio; 1114 struct dio_submit sdio = { 0, }; 1115 struct buffer_head map_bh = { 0, }; 1116 struct blk_plug plug; 1117 unsigned long align = offset | iov_iter_alignment(iter); 1118 1119 /* 1120 * Avoid references to bdev if not absolutely needed to give 1121 * the early prefetch in the caller enough time. 1122 */ 1123 1124 if (align & blocksize_mask) { 1125 if (bdev) 1126 blkbits = blksize_bits(bdev_logical_block_size(bdev)); 1127 blocksize_mask = (1 << blkbits) - 1; 1128 if (align & blocksize_mask) 1129 goto out; 1130 } 1131 1132 /* watch out for a 0 len io from a tricksy fs */ 1133 if (iov_iter_rw(iter) == READ && !iov_iter_count(iter)) 1134 return 0; 1135 1136 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL); 1137 retval = -ENOMEM; 1138 if (!dio) 1139 goto out; 1140 /* 1141 * Believe it or not, zeroing out the page array caused a .5% 1142 * performance regression in a database benchmark. So, we take 1143 * care to only zero out what's needed. 1144 */ 1145 memset(dio, 0, offsetof(struct dio, pages)); 1146 1147 dio->flags = flags; 1148 if (dio->flags & DIO_LOCKING) { 1149 if (iov_iter_rw(iter) == READ) { 1150 struct address_space *mapping = 1151 iocb->ki_filp->f_mapping; 1152 1153 /* will be released by direct_io_worker */ 1154 mutex_lock(&inode->i_mutex); 1155 1156 retval = filemap_write_and_wait_range(mapping, offset, 1157 end - 1); 1158 if (retval) { 1159 mutex_unlock(&inode->i_mutex); 1160 kmem_cache_free(dio_cache, dio); 1161 goto out; 1162 } 1163 } 1164 } 1165 1166 /* 1167 * For file extending writes updating i_size before data writeouts 1168 * complete can expose uninitialized blocks in dumb filesystems. 1169 * In that case we need to wait for I/O completion even if asked 1170 * for an asynchronous write. 1171 */ 1172 if (is_sync_kiocb(iocb)) 1173 dio->is_async = false; 1174 else if (!(dio->flags & DIO_ASYNC_EXTEND) && 1175 iov_iter_rw(iter) == WRITE && end > i_size_read(inode)) 1176 dio->is_async = false; 1177 else 1178 dio->is_async = true; 1179 1180 dio->inode = inode; 1181 dio->rw = iov_iter_rw(iter) == WRITE ? WRITE_ODIRECT : READ; 1182 1183 /* 1184 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue 1185 * so that we can call ->fsync. 1186 */ 1187 if (dio->is_async && iov_iter_rw(iter) == WRITE && 1188 ((iocb->ki_filp->f_flags & O_DSYNC) || 1189 IS_SYNC(iocb->ki_filp->f_mapping->host))) { 1190 retval = dio_set_defer_completion(dio); 1191 if (retval) { 1192 /* 1193 * We grab i_mutex only for reads so we don't have 1194 * to release it here 1195 */ 1196 kmem_cache_free(dio_cache, dio); 1197 goto out; 1198 } 1199 } 1200 1201 /* 1202 * Will be decremented at I/O completion time. 1203 */ 1204 if (!(dio->flags & DIO_SKIP_DIO_COUNT)) 1205 inode_dio_begin(inode); 1206 1207 retval = 0; 1208 sdio.blkbits = blkbits; 1209 sdio.blkfactor = i_blkbits - blkbits; 1210 sdio.block_in_file = offset >> blkbits; 1211 1212 sdio.get_block = get_block; 1213 dio->end_io = end_io; 1214 sdio.submit_io = submit_io; 1215 sdio.final_block_in_bio = -1; 1216 sdio.next_block_for_io = -1; 1217 1218 dio->iocb = iocb; 1219 dio->i_size = i_size_read(inode); 1220 1221 spin_lock_init(&dio->bio_lock); 1222 dio->refcount = 1; 1223 1224 dio->should_dirty = (iter->type == ITER_IOVEC); 1225 sdio.iter = iter; 1226 sdio.final_block_in_request = 1227 (offset + iov_iter_count(iter)) >> blkbits; 1228 1229 /* 1230 * In case of non-aligned buffers, we may need 2 more 1231 * pages since we need to zero out first and last block. 1232 */ 1233 if (unlikely(sdio.blkfactor)) 1234 sdio.pages_in_io = 2; 1235 1236 sdio.pages_in_io += iov_iter_npages(iter, INT_MAX); 1237 1238 blk_start_plug(&plug); 1239 1240 retval = do_direct_IO(dio, &sdio, &map_bh); 1241 if (retval) 1242 dio_cleanup(dio, &sdio); 1243 1244 if (retval == -ENOTBLK) { 1245 /* 1246 * The remaining part of the request will be 1247 * be handled by buffered I/O when we return 1248 */ 1249 retval = 0; 1250 } 1251 /* 1252 * There may be some unwritten disk at the end of a part-written 1253 * fs-block-sized block. Go zero that now. 1254 */ 1255 dio_zero_block(dio, &sdio, 1, &map_bh); 1256 1257 if (sdio.cur_page) { 1258 ssize_t ret2; 1259 1260 ret2 = dio_send_cur_page(dio, &sdio, &map_bh); 1261 if (retval == 0) 1262 retval = ret2; 1263 page_cache_release(sdio.cur_page); 1264 sdio.cur_page = NULL; 1265 } 1266 if (sdio.bio) 1267 dio_bio_submit(dio, &sdio); 1268 1269 blk_finish_plug(&plug); 1270 1271 /* 1272 * It is possible that, we return short IO due to end of file. 1273 * In that case, we need to release all the pages we got hold on. 1274 */ 1275 dio_cleanup(dio, &sdio); 1276 1277 /* 1278 * All block lookups have been performed. For READ requests 1279 * we can let i_mutex go now that its achieved its purpose 1280 * of protecting us from looking up uninitialized blocks. 1281 */ 1282 if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING)) 1283 mutex_unlock(&dio->inode->i_mutex); 1284 1285 /* 1286 * The only time we want to leave bios in flight is when a successful 1287 * partial aio read or full aio write have been setup. In that case 1288 * bio completion will call aio_complete. The only time it's safe to 1289 * call aio_complete is when we return -EIOCBQUEUED, so we key on that. 1290 * This had *better* be the only place that raises -EIOCBQUEUED. 1291 */ 1292 BUG_ON(retval == -EIOCBQUEUED); 1293 if (dio->is_async && retval == 0 && dio->result && 1294 (iov_iter_rw(iter) == READ || dio->result == count)) 1295 retval = -EIOCBQUEUED; 1296 else 1297 dio_await_completion(dio); 1298 1299 if (drop_refcount(dio) == 0) { 1300 retval = dio_complete(dio, offset, retval, false); 1301 } else 1302 BUG_ON(retval != -EIOCBQUEUED); 1303 1304 out: 1305 return retval; 1306 } 1307 1308 ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode, 1309 struct block_device *bdev, struct iov_iter *iter, 1310 loff_t offset, get_block_t get_block, 1311 dio_iodone_t end_io, dio_submit_t submit_io, 1312 int flags) 1313 { 1314 /* 1315 * The block device state is needed in the end to finally 1316 * submit everything. Since it's likely to be cache cold 1317 * prefetch it here as first thing to hide some of the 1318 * latency. 1319 * 1320 * Attempt to prefetch the pieces we likely need later. 1321 */ 1322 prefetch(&bdev->bd_disk->part_tbl); 1323 prefetch(bdev->bd_queue); 1324 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES); 1325 1326 return do_blockdev_direct_IO(iocb, inode, bdev, iter, offset, get_block, 1327 end_io, submit_io, flags); 1328 } 1329 1330 EXPORT_SYMBOL(__blockdev_direct_IO); 1331 1332 static __init int dio_init(void) 1333 { 1334 dio_cache = KMEM_CACHE(dio, SLAB_PANIC); 1335 return 0; 1336 } 1337 module_init(dio_init) 1338