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 #include <linux/aio.h> 41 42 /* 43 * How many user pages to map in one call to get_user_pages(). This determines 44 * the size of a structure in the slab cache 45 */ 46 #define DIO_PAGES 64 47 48 /* 49 * This code generally works in units of "dio_blocks". A dio_block is 50 * somewhere between the hard sector size and the filesystem block size. it 51 * is determined on a per-invocation basis. When talking to the filesystem 52 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity 53 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted 54 * to bio_block quantities by shifting left by blkfactor. 55 * 56 * If blkfactor is zero then the user's request was aligned to the filesystem's 57 * blocksize. 58 */ 59 60 /* dio_state only used in the submission path */ 61 62 struct dio_submit { 63 struct bio *bio; /* bio under assembly */ 64 unsigned blkbits; /* doesn't change */ 65 unsigned blkfactor; /* When we're using an alignment which 66 is finer than the filesystem's soft 67 blocksize, this specifies how much 68 finer. blkfactor=2 means 1/4-block 69 alignment. Does not change */ 70 unsigned start_zero_done; /* flag: sub-blocksize zeroing has 71 been performed at the start of a 72 write */ 73 int pages_in_io; /* approximate total IO pages */ 74 size_t size; /* total request size (doesn't change)*/ 75 sector_t block_in_file; /* Current offset into the underlying 76 file in dio_block units. */ 77 unsigned blocks_available; /* At block_in_file. changes */ 78 int reap_counter; /* rate limit reaping */ 79 sector_t final_block_in_request;/* doesn't change */ 80 int boundary; /* prev block is at a boundary */ 81 get_block_t *get_block; /* block mapping function */ 82 dio_submit_t *submit_io; /* IO submition function */ 83 84 loff_t logical_offset_in_bio; /* current first logical block in bio */ 85 sector_t final_block_in_bio; /* current final block in bio + 1 */ 86 sector_t next_block_for_io; /* next block to be put under IO, 87 in dio_blocks units */ 88 89 /* 90 * Deferred addition of a page to the dio. These variables are 91 * private to dio_send_cur_page(), submit_page_section() and 92 * dio_bio_add_page(). 93 */ 94 struct page *cur_page; /* The page */ 95 unsigned cur_page_offset; /* Offset into it, in bytes */ 96 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */ 97 sector_t cur_page_block; /* Where it starts */ 98 loff_t cur_page_fs_offset; /* Offset in file */ 99 100 struct iov_iter *iter; 101 /* 102 * Page queue. These variables belong to dio_refill_pages() and 103 * dio_get_page(). 104 */ 105 unsigned head; /* next page to process */ 106 unsigned tail; /* last valid page + 1 */ 107 size_t from, to; 108 }; 109 110 /* dio_state communicated between submission path and end_io */ 111 struct dio { 112 int flags; /* doesn't change */ 113 int rw; 114 struct inode *inode; 115 loff_t i_size; /* i_size when submitted */ 116 dio_iodone_t *end_io; /* IO completion function */ 117 118 void *private; /* copy from map_bh.b_private */ 119 120 /* BIO completion state */ 121 spinlock_t bio_lock; /* protects BIO fields below */ 122 int page_errors; /* errno from get_user_pages() */ 123 int is_async; /* is IO async ? */ 124 bool defer_completion; /* defer AIO completion to workqueue? */ 125 int io_error; /* IO error in completion path */ 126 unsigned long refcount; /* direct_io_worker() and bios */ 127 struct bio *bio_list; /* singly linked via bi_private */ 128 struct task_struct *waiter; /* waiting task (NULL if none) */ 129 130 /* AIO related stuff */ 131 struct kiocb *iocb; /* kiocb */ 132 ssize_t result; /* IO result */ 133 134 /* 135 * pages[] (and any fields placed after it) are not zeroed out at 136 * allocation time. Don't add new fields after pages[] unless you 137 * wish that they not be zeroed. 138 */ 139 union { 140 struct page *pages[DIO_PAGES]; /* page buffer */ 141 struct work_struct complete_work;/* deferred AIO completion */ 142 }; 143 } ____cacheline_aligned_in_smp; 144 145 static struct kmem_cache *dio_cache __read_mostly; 146 147 /* 148 * How many pages are in the queue? 149 */ 150 static inline unsigned dio_pages_present(struct dio_submit *sdio) 151 { 152 return sdio->tail - sdio->head; 153 } 154 155 /* 156 * Go grab and pin some userspace pages. Typically we'll get 64 at a time. 157 */ 158 static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio) 159 { 160 ssize_t ret; 161 162 ret = iov_iter_get_pages(sdio->iter, dio->pages, DIO_PAGES * PAGE_SIZE, 163 &sdio->from); 164 165 if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) { 166 struct page *page = ZERO_PAGE(0); 167 /* 168 * A memory fault, but the filesystem has some outstanding 169 * mapped blocks. We need to use those blocks up to avoid 170 * leaking stale data in the file. 171 */ 172 if (dio->page_errors == 0) 173 dio->page_errors = ret; 174 page_cache_get(page); 175 dio->pages[0] = page; 176 sdio->head = 0; 177 sdio->tail = 1; 178 sdio->from = 0; 179 sdio->to = PAGE_SIZE; 180 return 0; 181 } 182 183 if (ret >= 0) { 184 iov_iter_advance(sdio->iter, ret); 185 ret += sdio->from; 186 sdio->head = 0; 187 sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE; 188 sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1; 189 return 0; 190 } 191 return ret; 192 } 193 194 /* 195 * Get another userspace page. Returns an ERR_PTR on error. Pages are 196 * buffered inside the dio so that we can call get_user_pages() against a 197 * decent number of pages, less frequently. To provide nicer use of the 198 * L1 cache. 199 */ 200 static inline struct page *dio_get_page(struct dio *dio, 201 struct dio_submit *sdio, size_t *from, size_t *to) 202 { 203 int n; 204 if (dio_pages_present(sdio) == 0) { 205 int ret; 206 207 ret = dio_refill_pages(dio, sdio); 208 if (ret) 209 return ERR_PTR(ret); 210 BUG_ON(dio_pages_present(sdio) == 0); 211 } 212 n = sdio->head++; 213 *from = n ? 0 : sdio->from; 214 *to = (n == sdio->tail - 1) ? sdio->to : PAGE_SIZE; 215 return dio->pages[n]; 216 } 217 218 /** 219 * dio_complete() - called when all DIO BIO I/O has been completed 220 * @offset: the byte offset in the file of the completed operation 221 * 222 * This drops i_dio_count, lets interested parties know that a DIO operation 223 * has completed, and calculates the resulting return code for the operation. 224 * 225 * It lets the filesystem know if it registered an interest earlier via 226 * get_block. Pass the private field of the map buffer_head so that 227 * filesystems can use it to hold additional state between get_block calls and 228 * dio_complete. 229 */ 230 static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret, 231 bool is_async) 232 { 233 ssize_t transferred = 0; 234 235 /* 236 * AIO submission can race with bio completion to get here while 237 * expecting to have the last io completed by bio completion. 238 * In that case -EIOCBQUEUED is in fact not an error we want 239 * to preserve through this call. 240 */ 241 if (ret == -EIOCBQUEUED) 242 ret = 0; 243 244 if (dio->result) { 245 transferred = dio->result; 246 247 /* Check for short read case */ 248 if ((dio->rw == READ) && ((offset + transferred) > dio->i_size)) 249 transferred = dio->i_size - offset; 250 } 251 252 if (ret == 0) 253 ret = dio->page_errors; 254 if (ret == 0) 255 ret = dio->io_error; 256 if (ret == 0) 257 ret = transferred; 258 259 if (dio->end_io && dio->result) 260 dio->end_io(dio->iocb, offset, transferred, dio->private); 261 262 inode_dio_done(dio->inode); 263 if (is_async) { 264 if (dio->rw & WRITE) { 265 int err; 266 267 err = generic_write_sync(dio->iocb->ki_filp, offset, 268 transferred); 269 if (err < 0 && ret > 0) 270 ret = err; 271 } 272 273 aio_complete(dio->iocb, ret, 0); 274 } 275 276 kmem_cache_free(dio_cache, dio); 277 return ret; 278 } 279 280 static void dio_aio_complete_work(struct work_struct *work) 281 { 282 struct dio *dio = container_of(work, struct dio, complete_work); 283 284 dio_complete(dio, dio->iocb->ki_pos, 0, true); 285 } 286 287 static int dio_bio_complete(struct dio *dio, struct bio *bio); 288 289 /* 290 * Asynchronous IO callback. 291 */ 292 static void dio_bio_end_aio(struct bio *bio, int error) 293 { 294 struct dio *dio = bio->bi_private; 295 unsigned long remaining; 296 unsigned long flags; 297 298 /* cleanup the bio */ 299 dio_bio_complete(dio, bio); 300 301 spin_lock_irqsave(&dio->bio_lock, flags); 302 remaining = --dio->refcount; 303 if (remaining == 1 && dio->waiter) 304 wake_up_process(dio->waiter); 305 spin_unlock_irqrestore(&dio->bio_lock, flags); 306 307 if (remaining == 0) { 308 if (dio->result && dio->defer_completion) { 309 INIT_WORK(&dio->complete_work, dio_aio_complete_work); 310 queue_work(dio->inode->i_sb->s_dio_done_wq, 311 &dio->complete_work); 312 } else { 313 dio_complete(dio, dio->iocb->ki_pos, 0, true); 314 } 315 } 316 } 317 318 /* 319 * The BIO completion handler simply queues the BIO up for the process-context 320 * handler. 321 * 322 * During I/O bi_private points at the dio. After I/O, bi_private is used to 323 * implement a singly-linked list of completed BIOs, at dio->bio_list. 324 */ 325 static void dio_bio_end_io(struct bio *bio, int error) 326 { 327 struct dio *dio = bio->bi_private; 328 unsigned long flags; 329 330 spin_lock_irqsave(&dio->bio_lock, flags); 331 bio->bi_private = dio->bio_list; 332 dio->bio_list = bio; 333 if (--dio->refcount == 1 && dio->waiter) 334 wake_up_process(dio->waiter); 335 spin_unlock_irqrestore(&dio->bio_lock, flags); 336 } 337 338 /** 339 * dio_end_io - handle the end io action for the given bio 340 * @bio: The direct io bio thats being completed 341 * @error: Error if there was one 342 * 343 * This is meant to be called by any filesystem that uses their own dio_submit_t 344 * so that the DIO specific endio actions are dealt with after the filesystem 345 * has done it's completion work. 346 */ 347 void dio_end_io(struct bio *bio, int error) 348 { 349 struct dio *dio = bio->bi_private; 350 351 if (dio->is_async) 352 dio_bio_end_aio(bio, error); 353 else 354 dio_bio_end_io(bio, error); 355 } 356 EXPORT_SYMBOL_GPL(dio_end_io); 357 358 static inline void 359 dio_bio_alloc(struct dio *dio, struct dio_submit *sdio, 360 struct block_device *bdev, 361 sector_t first_sector, int nr_vecs) 362 { 363 struct bio *bio; 364 365 /* 366 * bio_alloc() is guaranteed to return a bio when called with 367 * __GFP_WAIT and we request a valid number of vectors. 368 */ 369 bio = bio_alloc(GFP_KERNEL, nr_vecs); 370 371 bio->bi_bdev = bdev; 372 bio->bi_iter.bi_sector = first_sector; 373 if (dio->is_async) 374 bio->bi_end_io = dio_bio_end_aio; 375 else 376 bio->bi_end_io = dio_bio_end_io; 377 378 sdio->bio = bio; 379 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset; 380 } 381 382 /* 383 * In the AIO read case we speculatively dirty the pages before starting IO. 384 * During IO completion, any of these pages which happen to have been written 385 * back will be redirtied by bio_check_pages_dirty(). 386 * 387 * bios hold a dio reference between submit_bio and ->end_io. 388 */ 389 static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio) 390 { 391 struct bio *bio = sdio->bio; 392 unsigned long flags; 393 394 bio->bi_private = dio; 395 396 spin_lock_irqsave(&dio->bio_lock, flags); 397 dio->refcount++; 398 spin_unlock_irqrestore(&dio->bio_lock, flags); 399 400 if (dio->is_async && dio->rw == READ) 401 bio_set_pages_dirty(bio); 402 403 if (sdio->submit_io) 404 sdio->submit_io(dio->rw, bio, dio->inode, 405 sdio->logical_offset_in_bio); 406 else 407 submit_bio(dio->rw, bio); 408 409 sdio->bio = NULL; 410 sdio->boundary = 0; 411 sdio->logical_offset_in_bio = 0; 412 } 413 414 /* 415 * Release any resources in case of a failure 416 */ 417 static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio) 418 { 419 while (sdio->head < sdio->tail) 420 page_cache_release(dio->pages[sdio->head++]); 421 } 422 423 /* 424 * Wait for the next BIO to complete. Remove it and return it. NULL is 425 * returned once all BIOs have been completed. This must only be called once 426 * all bios have been issued so that dio->refcount can only decrease. This 427 * requires that that the caller hold a reference on the dio. 428 */ 429 static struct bio *dio_await_one(struct dio *dio) 430 { 431 unsigned long flags; 432 struct bio *bio = NULL; 433 434 spin_lock_irqsave(&dio->bio_lock, flags); 435 436 /* 437 * Wait as long as the list is empty and there are bios in flight. bio 438 * completion drops the count, maybe adds to the list, and wakes while 439 * holding the bio_lock so we don't need set_current_state()'s barrier 440 * and can call it after testing our condition. 441 */ 442 while (dio->refcount > 1 && dio->bio_list == NULL) { 443 __set_current_state(TASK_UNINTERRUPTIBLE); 444 dio->waiter = current; 445 spin_unlock_irqrestore(&dio->bio_lock, flags); 446 io_schedule(); 447 /* wake up sets us TASK_RUNNING */ 448 spin_lock_irqsave(&dio->bio_lock, flags); 449 dio->waiter = NULL; 450 } 451 if (dio->bio_list) { 452 bio = dio->bio_list; 453 dio->bio_list = bio->bi_private; 454 } 455 spin_unlock_irqrestore(&dio->bio_lock, flags); 456 return bio; 457 } 458 459 /* 460 * Process one completed BIO. No locks are held. 461 */ 462 static int dio_bio_complete(struct dio *dio, struct bio *bio) 463 { 464 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); 465 struct bio_vec *bvec; 466 unsigned i; 467 468 if (!uptodate) 469 dio->io_error = -EIO; 470 471 if (dio->is_async && dio->rw == READ) { 472 bio_check_pages_dirty(bio); /* transfers ownership */ 473 } else { 474 bio_for_each_segment_all(bvec, bio, i) { 475 struct page *page = bvec->bv_page; 476 477 if (dio->rw == READ && !PageCompound(page)) 478 set_page_dirty_lock(page); 479 page_cache_release(page); 480 } 481 bio_put(bio); 482 } 483 return uptodate ? 0 : -EIO; 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_get_nr_vecs(map_bh->b_bdev)); 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 page = dio_get_page(dio, sdio, &from, &to); 915 if (IS_ERR(page)) { 916 ret = PTR_ERR(page); 917 goto out; 918 } 919 920 while (from < to) { 921 unsigned this_chunk_bytes; /* # of bytes mapped */ 922 unsigned this_chunk_blocks; /* # of blocks */ 923 unsigned u; 924 925 if (sdio->blocks_available == 0) { 926 /* 927 * Need to go and map some more disk 928 */ 929 unsigned long blkmask; 930 unsigned long dio_remainder; 931 932 ret = get_more_blocks(dio, sdio, map_bh); 933 if (ret) { 934 page_cache_release(page); 935 goto out; 936 } 937 if (!buffer_mapped(map_bh)) 938 goto do_holes; 939 940 sdio->blocks_available = 941 map_bh->b_size >> sdio->blkbits; 942 sdio->next_block_for_io = 943 map_bh->b_blocknr << sdio->blkfactor; 944 if (buffer_new(map_bh)) 945 clean_blockdev_aliases(dio, map_bh); 946 947 if (!sdio->blkfactor) 948 goto do_holes; 949 950 blkmask = (1 << sdio->blkfactor) - 1; 951 dio_remainder = (sdio->block_in_file & blkmask); 952 953 /* 954 * If we are at the start of IO and that IO 955 * starts partway into a fs-block, 956 * dio_remainder will be non-zero. If the IO 957 * is a read then we can simply advance the IO 958 * cursor to the first block which is to be 959 * read. But if the IO is a write and the 960 * block was newly allocated we cannot do that; 961 * the start of the fs block must be zeroed out 962 * on-disk 963 */ 964 if (!buffer_new(map_bh)) 965 sdio->next_block_for_io += dio_remainder; 966 sdio->blocks_available -= dio_remainder; 967 } 968 do_holes: 969 /* Handle holes */ 970 if (!buffer_mapped(map_bh)) { 971 loff_t i_size_aligned; 972 973 /* AKPM: eargh, -ENOTBLK is a hack */ 974 if (dio->rw & WRITE) { 975 page_cache_release(page); 976 return -ENOTBLK; 977 } 978 979 /* 980 * Be sure to account for a partial block as the 981 * last block in the file 982 */ 983 i_size_aligned = ALIGN(i_size_read(dio->inode), 984 1 << blkbits); 985 if (sdio->block_in_file >= 986 i_size_aligned >> blkbits) { 987 /* We hit eof */ 988 page_cache_release(page); 989 goto out; 990 } 991 zero_user(page, from, 1 << blkbits); 992 sdio->block_in_file++; 993 from += 1 << blkbits; 994 dio->result += 1 << blkbits; 995 goto next_block; 996 } 997 998 /* 999 * If we're performing IO which has an alignment which 1000 * is finer than the underlying fs, go check to see if 1001 * we must zero out the start of this block. 1002 */ 1003 if (unlikely(sdio->blkfactor && !sdio->start_zero_done)) 1004 dio_zero_block(dio, sdio, 0, map_bh); 1005 1006 /* 1007 * Work out, in this_chunk_blocks, how much disk we 1008 * can add to this page 1009 */ 1010 this_chunk_blocks = sdio->blocks_available; 1011 u = (to - from) >> blkbits; 1012 if (this_chunk_blocks > u) 1013 this_chunk_blocks = u; 1014 u = sdio->final_block_in_request - sdio->block_in_file; 1015 if (this_chunk_blocks > u) 1016 this_chunk_blocks = u; 1017 this_chunk_bytes = this_chunk_blocks << blkbits; 1018 BUG_ON(this_chunk_bytes == 0); 1019 1020 if (this_chunk_blocks == sdio->blocks_available) 1021 sdio->boundary = buffer_boundary(map_bh); 1022 ret = submit_page_section(dio, sdio, page, 1023 from, 1024 this_chunk_bytes, 1025 sdio->next_block_for_io, 1026 map_bh); 1027 if (ret) { 1028 page_cache_release(page); 1029 goto out; 1030 } 1031 sdio->next_block_for_io += this_chunk_blocks; 1032 1033 sdio->block_in_file += this_chunk_blocks; 1034 from += this_chunk_bytes; 1035 dio->result += this_chunk_bytes; 1036 sdio->blocks_available -= this_chunk_blocks; 1037 next_block: 1038 BUG_ON(sdio->block_in_file > sdio->final_block_in_request); 1039 if (sdio->block_in_file == sdio->final_block_in_request) 1040 break; 1041 } 1042 1043 /* Drop the ref which was taken in get_user_pages() */ 1044 page_cache_release(page); 1045 } 1046 out: 1047 return ret; 1048 } 1049 1050 static inline int drop_refcount(struct dio *dio) 1051 { 1052 int ret2; 1053 unsigned long flags; 1054 1055 /* 1056 * Sync will always be dropping the final ref and completing the 1057 * operation. AIO can if it was a broken operation described above or 1058 * in fact if all the bios race to complete before we get here. In 1059 * that case dio_complete() translates the EIOCBQUEUED into the proper 1060 * return code that the caller will hand to aio_complete(). 1061 * 1062 * This is managed by the bio_lock instead of being an atomic_t so that 1063 * completion paths can drop their ref and use the remaining count to 1064 * decide to wake the submission path atomically. 1065 */ 1066 spin_lock_irqsave(&dio->bio_lock, flags); 1067 ret2 = --dio->refcount; 1068 spin_unlock_irqrestore(&dio->bio_lock, flags); 1069 return ret2; 1070 } 1071 1072 /* 1073 * This is a library function for use by filesystem drivers. 1074 * 1075 * The locking rules are governed by the flags parameter: 1076 * - if the flags value contains DIO_LOCKING we use a fancy locking 1077 * scheme for dumb filesystems. 1078 * For writes this function is called under i_mutex and returns with 1079 * i_mutex held, for reads, i_mutex is not held on entry, but it is 1080 * taken and dropped again before returning. 1081 * - if the flags value does NOT contain DIO_LOCKING we don't use any 1082 * internal locking but rather rely on the filesystem to synchronize 1083 * direct I/O reads/writes versus each other and truncate. 1084 * 1085 * To help with locking against truncate we incremented the i_dio_count 1086 * counter before starting direct I/O, and decrement it once we are done. 1087 * Truncate can wait for it to reach zero to provide exclusion. It is 1088 * expected that filesystem provide exclusion between new direct I/O 1089 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex, 1090 * but other filesystems need to take care of this on their own. 1091 * 1092 * NOTE: if you pass "sdio" to anything by pointer make sure that function 1093 * is always inlined. Otherwise gcc is unable to split the structure into 1094 * individual fields and will generate much worse code. This is important 1095 * for the whole file. 1096 */ 1097 static inline ssize_t 1098 do_blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode, 1099 struct block_device *bdev, struct iov_iter *iter, loff_t offset, 1100 get_block_t get_block, dio_iodone_t end_io, 1101 dio_submit_t submit_io, int flags) 1102 { 1103 unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits); 1104 unsigned blkbits = i_blkbits; 1105 unsigned blocksize_mask = (1 << blkbits) - 1; 1106 ssize_t retval = -EINVAL; 1107 loff_t end = offset + iov_iter_count(iter); 1108 struct dio *dio; 1109 struct dio_submit sdio = { 0, }; 1110 struct buffer_head map_bh = { 0, }; 1111 struct blk_plug plug; 1112 unsigned long align = offset | iov_iter_alignment(iter); 1113 1114 if (rw & WRITE) 1115 rw = WRITE_ODIRECT; 1116 1117 /* 1118 * Avoid references to bdev if not absolutely needed to give 1119 * the early prefetch in the caller enough time. 1120 */ 1121 1122 if (align & blocksize_mask) { 1123 if (bdev) 1124 blkbits = blksize_bits(bdev_logical_block_size(bdev)); 1125 blocksize_mask = (1 << blkbits) - 1; 1126 if (align & blocksize_mask) 1127 goto out; 1128 } 1129 1130 /* watch out for a 0 len io from a tricksy fs */ 1131 if (rw == READ && !iov_iter_count(iter)) 1132 return 0; 1133 1134 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL); 1135 retval = -ENOMEM; 1136 if (!dio) 1137 goto out; 1138 /* 1139 * Believe it or not, zeroing out the page array caused a .5% 1140 * performance regression in a database benchmark. So, we take 1141 * care to only zero out what's needed. 1142 */ 1143 memset(dio, 0, offsetof(struct dio, pages)); 1144 1145 dio->flags = flags; 1146 if (dio->flags & DIO_LOCKING) { 1147 if (rw == READ) { 1148 struct address_space *mapping = 1149 iocb->ki_filp->f_mapping; 1150 1151 /* will be released by direct_io_worker */ 1152 mutex_lock(&inode->i_mutex); 1153 1154 retval = filemap_write_and_wait_range(mapping, offset, 1155 end - 1); 1156 if (retval) { 1157 mutex_unlock(&inode->i_mutex); 1158 kmem_cache_free(dio_cache, dio); 1159 goto out; 1160 } 1161 } 1162 } 1163 1164 /* 1165 * For file extending writes updating i_size before data writeouts 1166 * complete can expose uninitialized blocks in dumb filesystems. 1167 * In that case we need to wait for I/O completion even if asked 1168 * for an asynchronous write. 1169 */ 1170 if (is_sync_kiocb(iocb)) 1171 dio->is_async = false; 1172 else if (!(dio->flags & DIO_ASYNC_EXTEND) && 1173 (rw & WRITE) && end > i_size_read(inode)) 1174 dio->is_async = false; 1175 else 1176 dio->is_async = true; 1177 1178 dio->inode = inode; 1179 dio->rw = rw; 1180 1181 /* 1182 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue 1183 * so that we can call ->fsync. 1184 */ 1185 if (dio->is_async && (rw & WRITE) && 1186 ((iocb->ki_filp->f_flags & O_DSYNC) || 1187 IS_SYNC(iocb->ki_filp->f_mapping->host))) { 1188 retval = dio_set_defer_completion(dio); 1189 if (retval) { 1190 /* 1191 * We grab i_mutex only for reads so we don't have 1192 * to release it here 1193 */ 1194 kmem_cache_free(dio_cache, dio); 1195 goto out; 1196 } 1197 } 1198 1199 /* 1200 * Will be decremented at I/O completion time. 1201 */ 1202 atomic_inc(&inode->i_dio_count); 1203 1204 retval = 0; 1205 sdio.blkbits = blkbits; 1206 sdio.blkfactor = i_blkbits - blkbits; 1207 sdio.block_in_file = offset >> blkbits; 1208 1209 sdio.get_block = get_block; 1210 dio->end_io = end_io; 1211 sdio.submit_io = submit_io; 1212 sdio.final_block_in_bio = -1; 1213 sdio.next_block_for_io = -1; 1214 1215 dio->iocb = iocb; 1216 dio->i_size = i_size_read(inode); 1217 1218 spin_lock_init(&dio->bio_lock); 1219 dio->refcount = 1; 1220 1221 sdio.iter = iter; 1222 sdio.final_block_in_request = 1223 (offset + iov_iter_count(iter)) >> blkbits; 1224 1225 /* 1226 * In case of non-aligned buffers, we may need 2 more 1227 * pages since we need to zero out first and last block. 1228 */ 1229 if (unlikely(sdio.blkfactor)) 1230 sdio.pages_in_io = 2; 1231 1232 sdio.pages_in_io += iov_iter_npages(iter, INT_MAX); 1233 1234 blk_start_plug(&plug); 1235 1236 retval = do_direct_IO(dio, &sdio, &map_bh); 1237 if (retval) 1238 dio_cleanup(dio, &sdio); 1239 1240 if (retval == -ENOTBLK) { 1241 /* 1242 * The remaining part of the request will be 1243 * be handled by buffered I/O when we return 1244 */ 1245 retval = 0; 1246 } 1247 /* 1248 * There may be some unwritten disk at the end of a part-written 1249 * fs-block-sized block. Go zero that now. 1250 */ 1251 dio_zero_block(dio, &sdio, 1, &map_bh); 1252 1253 if (sdio.cur_page) { 1254 ssize_t ret2; 1255 1256 ret2 = dio_send_cur_page(dio, &sdio, &map_bh); 1257 if (retval == 0) 1258 retval = ret2; 1259 page_cache_release(sdio.cur_page); 1260 sdio.cur_page = NULL; 1261 } 1262 if (sdio.bio) 1263 dio_bio_submit(dio, &sdio); 1264 1265 blk_finish_plug(&plug); 1266 1267 /* 1268 * It is possible that, we return short IO due to end of file. 1269 * In that case, we need to release all the pages we got hold on. 1270 */ 1271 dio_cleanup(dio, &sdio); 1272 1273 /* 1274 * All block lookups have been performed. For READ requests 1275 * we can let i_mutex go now that its achieved its purpose 1276 * of protecting us from looking up uninitialized blocks. 1277 */ 1278 if (rw == READ && (dio->flags & DIO_LOCKING)) 1279 mutex_unlock(&dio->inode->i_mutex); 1280 1281 /* 1282 * The only time we want to leave bios in flight is when a successful 1283 * partial aio read or full aio write have been setup. In that case 1284 * bio completion will call aio_complete. The only time it's safe to 1285 * call aio_complete is when we return -EIOCBQUEUED, so we key on that. 1286 * This had *better* be the only place that raises -EIOCBQUEUED. 1287 */ 1288 BUG_ON(retval == -EIOCBQUEUED); 1289 if (dio->is_async && retval == 0 && dio->result && 1290 ((rw == READ) || (dio->result == sdio.size))) 1291 retval = -EIOCBQUEUED; 1292 1293 if (retval != -EIOCBQUEUED) 1294 dio_await_completion(dio); 1295 1296 if (drop_refcount(dio) == 0) { 1297 retval = dio_complete(dio, offset, retval, false); 1298 } else 1299 BUG_ON(retval != -EIOCBQUEUED); 1300 1301 out: 1302 return retval; 1303 } 1304 1305 ssize_t 1306 __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode, 1307 struct block_device *bdev, struct iov_iter *iter, loff_t offset, 1308 get_block_t get_block, dio_iodone_t end_io, 1309 dio_submit_t submit_io, int flags) 1310 { 1311 /* 1312 * The block device state is needed in the end to finally 1313 * submit everything. Since it's likely to be cache cold 1314 * prefetch it here as first thing to hide some of the 1315 * latency. 1316 * 1317 * Attempt to prefetch the pieces we likely need later. 1318 */ 1319 prefetch(&bdev->bd_disk->part_tbl); 1320 prefetch(bdev->bd_queue); 1321 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES); 1322 1323 return do_blockdev_direct_IO(rw, iocb, inode, bdev, iter, offset, 1324 get_block, end_io, submit_io, flags); 1325 } 1326 1327 EXPORT_SYMBOL(__blockdev_direct_IO); 1328 1329 static __init int dio_init(void) 1330 { 1331 dio_cache = KMEM_CACHE(dio, SLAB_PANIC); 1332 return 0; 1333 } 1334 module_init(dio_init) 1335