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 blk_qc_t bio_cookie; 113 struct block_device *bio_bdev; 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 bool should_dirty; /* if pages should be dirtied */ 126 int io_error; /* IO error in completion path */ 127 unsigned long refcount; /* direct_io_worker() and bios */ 128 struct bio *bio_list; /* singly linked via bi_private */ 129 struct task_struct *waiter; /* waiting task (NULL if none) */ 130 131 /* AIO related stuff */ 132 struct kiocb *iocb; /* kiocb */ 133 ssize_t result; /* IO result */ 134 135 /* 136 * pages[] (and any fields placed after it) are not zeroed out at 137 * allocation time. Don't add new fields after pages[] unless you 138 * wish that they not be zeroed. 139 */ 140 union { 141 struct page *pages[DIO_PAGES]; /* page buffer */ 142 struct work_struct complete_work;/* deferred AIO completion */ 143 }; 144 } ____cacheline_aligned_in_smp; 145 146 static struct kmem_cache *dio_cache __read_mostly; 147 148 /* 149 * How many pages are in the queue? 150 */ 151 static inline unsigned dio_pages_present(struct dio_submit *sdio) 152 { 153 return sdio->tail - sdio->head; 154 } 155 156 /* 157 * Go grab and pin some userspace pages. Typically we'll get 64 at a time. 158 */ 159 static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio) 160 { 161 ssize_t ret; 162 163 ret = iov_iter_get_pages(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES, 164 &sdio->from); 165 166 if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) { 167 struct page *page = ZERO_PAGE(0); 168 /* 169 * A memory fault, but the filesystem has some outstanding 170 * mapped blocks. We need to use those blocks up to avoid 171 * leaking stale data in the file. 172 */ 173 if (dio->page_errors == 0) 174 dio->page_errors = ret; 175 page_cache_get(page); 176 dio->pages[0] = page; 177 sdio->head = 0; 178 sdio->tail = 1; 179 sdio->from = 0; 180 sdio->to = PAGE_SIZE; 181 return 0; 182 } 183 184 if (ret >= 0) { 185 iov_iter_advance(sdio->iter, ret); 186 ret += sdio->from; 187 sdio->head = 0; 188 sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE; 189 sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1; 190 return 0; 191 } 192 return ret; 193 } 194 195 /* 196 * Get another userspace page. Returns an ERR_PTR on error. Pages are 197 * buffered inside the dio so that we can call get_user_pages() against a 198 * decent number of pages, less frequently. To provide nicer use of the 199 * L1 cache. 200 */ 201 static inline struct page *dio_get_page(struct dio *dio, 202 struct dio_submit *sdio) 203 { 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 return dio->pages[sdio->head]; 213 } 214 215 /** 216 * dio_complete() - called when all DIO BIO I/O has been completed 217 * @offset: the byte offset in the file of the completed operation 218 * 219 * This drops i_dio_count, lets interested parties know that a DIO operation 220 * has completed, and calculates the resulting return code for the operation. 221 * 222 * It lets the filesystem know if it registered an interest earlier via 223 * get_block. Pass the private field of the map buffer_head so that 224 * filesystems can use it to hold additional state between get_block calls and 225 * dio_complete. 226 */ 227 static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret, 228 bool is_async) 229 { 230 ssize_t transferred = 0; 231 232 /* 233 * AIO submission can race with bio completion to get here while 234 * expecting to have the last io completed by bio completion. 235 * In that case -EIOCBQUEUED is in fact not an error we want 236 * to preserve through this call. 237 */ 238 if (ret == -EIOCBQUEUED) 239 ret = 0; 240 241 if (dio->result) { 242 transferred = dio->result; 243 244 /* Check for short read case */ 245 if ((dio->rw == READ) && ((offset + transferred) > dio->i_size)) 246 transferred = dio->i_size - offset; 247 } 248 249 if (ret == 0) 250 ret = dio->page_errors; 251 if (ret == 0) 252 ret = dio->io_error; 253 if (ret == 0) 254 ret = transferred; 255 256 if (dio->end_io) { 257 int err; 258 259 err = dio->end_io(dio->iocb, offset, ret, dio->private); 260 if (err) 261 ret = err; 262 } 263 264 if (!(dio->flags & DIO_SKIP_DIO_COUNT)) 265 inode_dio_end(dio->inode); 266 267 if (is_async) { 268 if (dio->rw & WRITE) { 269 int err; 270 271 err = generic_write_sync(dio->iocb->ki_filp, offset, 272 transferred); 273 if (err < 0 && ret > 0) 274 ret = err; 275 } 276 277 dio->iocb->ki_complete(dio->iocb, ret, 0); 278 } 279 280 kmem_cache_free(dio_cache, dio); 281 return ret; 282 } 283 284 static void dio_aio_complete_work(struct work_struct *work) 285 { 286 struct dio *dio = container_of(work, struct dio, complete_work); 287 288 dio_complete(dio, dio->iocb->ki_pos, 0, true); 289 } 290 291 static int dio_bio_complete(struct dio *dio, struct bio *bio); 292 293 /* 294 * Asynchronous IO callback. 295 */ 296 static void dio_bio_end_aio(struct bio *bio) 297 { 298 struct dio *dio = bio->bi_private; 299 unsigned long remaining; 300 unsigned long flags; 301 302 /* cleanup the bio */ 303 dio_bio_complete(dio, bio); 304 305 spin_lock_irqsave(&dio->bio_lock, flags); 306 remaining = --dio->refcount; 307 if (remaining == 1 && dio->waiter) 308 wake_up_process(dio->waiter); 309 spin_unlock_irqrestore(&dio->bio_lock, flags); 310 311 if (remaining == 0) { 312 if (dio->result && dio->defer_completion) { 313 INIT_WORK(&dio->complete_work, dio_aio_complete_work); 314 queue_work(dio->inode->i_sb->s_dio_done_wq, 315 &dio->complete_work); 316 } else { 317 dio_complete(dio, dio->iocb->ki_pos, 0, true); 318 } 319 } 320 } 321 322 /* 323 * The BIO completion handler simply queues the BIO up for the process-context 324 * handler. 325 * 326 * During I/O bi_private points at the dio. After I/O, bi_private is used to 327 * implement a singly-linked list of completed BIOs, at dio->bio_list. 328 */ 329 static void dio_bio_end_io(struct bio *bio) 330 { 331 struct dio *dio = bio->bi_private; 332 unsigned long flags; 333 334 spin_lock_irqsave(&dio->bio_lock, flags); 335 bio->bi_private = dio->bio_list; 336 dio->bio_list = bio; 337 if (--dio->refcount == 1 && dio->waiter) 338 wake_up_process(dio->waiter); 339 spin_unlock_irqrestore(&dio->bio_lock, flags); 340 } 341 342 /** 343 * dio_end_io - handle the end io action for the given bio 344 * @bio: The direct io bio thats being completed 345 * @error: Error if there was one 346 * 347 * This is meant to be called by any filesystem that uses their own dio_submit_t 348 * so that the DIO specific endio actions are dealt with after the filesystem 349 * has done it's completion work. 350 */ 351 void dio_end_io(struct bio *bio, int error) 352 { 353 struct dio *dio = bio->bi_private; 354 355 if (dio->is_async) 356 dio_bio_end_aio(bio); 357 else 358 dio_bio_end_io(bio); 359 } 360 EXPORT_SYMBOL_GPL(dio_end_io); 361 362 static inline void 363 dio_bio_alloc(struct dio *dio, struct dio_submit *sdio, 364 struct block_device *bdev, 365 sector_t first_sector, int nr_vecs) 366 { 367 struct bio *bio; 368 369 /* 370 * bio_alloc() is guaranteed to return a bio when called with 371 * __GFP_RECLAIM and we request a valid number of vectors. 372 */ 373 bio = bio_alloc(GFP_KERNEL, nr_vecs); 374 375 bio->bi_bdev = bdev; 376 bio->bi_iter.bi_sector = first_sector; 377 if (dio->is_async) 378 bio->bi_end_io = dio_bio_end_aio; 379 else 380 bio->bi_end_io = dio_bio_end_io; 381 382 sdio->bio = bio; 383 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset; 384 } 385 386 /* 387 * In the AIO read case we speculatively dirty the pages before starting IO. 388 * During IO completion, any of these pages which happen to have been written 389 * back will be redirtied by bio_check_pages_dirty(). 390 * 391 * bios hold a dio reference between submit_bio and ->end_io. 392 */ 393 static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio) 394 { 395 struct bio *bio = sdio->bio; 396 unsigned long flags; 397 398 bio->bi_private = dio; 399 400 spin_lock_irqsave(&dio->bio_lock, flags); 401 dio->refcount++; 402 spin_unlock_irqrestore(&dio->bio_lock, flags); 403 404 if (dio->is_async && dio->rw == READ && dio->should_dirty) 405 bio_set_pages_dirty(bio); 406 407 dio->bio_bdev = bio->bi_bdev; 408 409 if (sdio->submit_io) { 410 sdio->submit_io(dio->rw, bio, dio->inode, 411 sdio->logical_offset_in_bio); 412 dio->bio_cookie = BLK_QC_T_NONE; 413 } else 414 dio->bio_cookie = submit_bio(dio->rw, bio); 415 416 sdio->bio = NULL; 417 sdio->boundary = 0; 418 sdio->logical_offset_in_bio = 0; 419 } 420 421 /* 422 * Release any resources in case of a failure 423 */ 424 static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio) 425 { 426 while (sdio->head < sdio->tail) 427 page_cache_release(dio->pages[sdio->head++]); 428 } 429 430 /* 431 * Wait for the next BIO to complete. Remove it and return it. NULL is 432 * returned once all BIOs have been completed. This must only be called once 433 * all bios have been issued so that dio->refcount can only decrease. This 434 * requires that that the caller hold a reference on the dio. 435 */ 436 static struct bio *dio_await_one(struct dio *dio) 437 { 438 unsigned long flags; 439 struct bio *bio = NULL; 440 441 spin_lock_irqsave(&dio->bio_lock, flags); 442 443 /* 444 * Wait as long as the list is empty and there are bios in flight. bio 445 * completion drops the count, maybe adds to the list, and wakes while 446 * holding the bio_lock so we don't need set_current_state()'s barrier 447 * and can call it after testing our condition. 448 */ 449 while (dio->refcount > 1 && dio->bio_list == NULL) { 450 __set_current_state(TASK_UNINTERRUPTIBLE); 451 dio->waiter = current; 452 spin_unlock_irqrestore(&dio->bio_lock, flags); 453 if (!(dio->iocb->ki_flags & IOCB_HIPRI) || 454 !blk_poll(bdev_get_queue(dio->bio_bdev), dio->bio_cookie)) 455 io_schedule(); 456 /* wake up sets us TASK_RUNNING */ 457 spin_lock_irqsave(&dio->bio_lock, flags); 458 dio->waiter = NULL; 459 } 460 if (dio->bio_list) { 461 bio = dio->bio_list; 462 dio->bio_list = bio->bi_private; 463 } 464 spin_unlock_irqrestore(&dio->bio_lock, flags); 465 return bio; 466 } 467 468 /* 469 * Process one completed BIO. No locks are held. 470 */ 471 static int dio_bio_complete(struct dio *dio, struct bio *bio) 472 { 473 struct bio_vec *bvec; 474 unsigned i; 475 int err; 476 477 if (bio->bi_error) 478 dio->io_error = -EIO; 479 480 if (dio->is_async && dio->rw == READ && dio->should_dirty) { 481 err = bio->bi_error; 482 bio_check_pages_dirty(bio); /* transfers ownership */ 483 } else { 484 bio_for_each_segment_all(bvec, bio, i) { 485 struct page *page = bvec->bv_page; 486 487 if (dio->rw == READ && !PageCompound(page) && 488 dio->should_dirty) 489 set_page_dirty_lock(page); 490 page_cache_release(page); 491 } 492 err = bio->bi_error; 493 bio_put(bio); 494 } 495 return err; 496 } 497 498 /* 499 * Wait on and process all in-flight BIOs. This must only be called once 500 * all bios have been issued so that the refcount can only decrease. 501 * This just waits for all bios to make it through dio_bio_complete. IO 502 * errors are propagated through dio->io_error and should be propagated via 503 * dio_complete(). 504 */ 505 static void dio_await_completion(struct dio *dio) 506 { 507 struct bio *bio; 508 do { 509 bio = dio_await_one(dio); 510 if (bio) 511 dio_bio_complete(dio, bio); 512 } while (bio); 513 } 514 515 /* 516 * A really large O_DIRECT read or write can generate a lot of BIOs. So 517 * to keep the memory consumption sane we periodically reap any completed BIOs 518 * during the BIO generation phase. 519 * 520 * This also helps to limit the peak amount of pinned userspace memory. 521 */ 522 static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio) 523 { 524 int ret = 0; 525 526 if (sdio->reap_counter++ >= 64) { 527 while (dio->bio_list) { 528 unsigned long flags; 529 struct bio *bio; 530 int ret2; 531 532 spin_lock_irqsave(&dio->bio_lock, flags); 533 bio = dio->bio_list; 534 dio->bio_list = bio->bi_private; 535 spin_unlock_irqrestore(&dio->bio_lock, flags); 536 ret2 = dio_bio_complete(dio, bio); 537 if (ret == 0) 538 ret = ret2; 539 } 540 sdio->reap_counter = 0; 541 } 542 return ret; 543 } 544 545 /* 546 * Create workqueue for deferred direct IO completions. We allocate the 547 * workqueue when it's first needed. This avoids creating workqueue for 548 * filesystems that don't need it and also allows us to create the workqueue 549 * late enough so the we can include s_id in the name of the workqueue. 550 */ 551 static int sb_init_dio_done_wq(struct super_block *sb) 552 { 553 struct workqueue_struct *old; 554 struct workqueue_struct *wq = alloc_workqueue("dio/%s", 555 WQ_MEM_RECLAIM, 0, 556 sb->s_id); 557 if (!wq) 558 return -ENOMEM; 559 /* 560 * This has to be atomic as more DIOs can race to create the workqueue 561 */ 562 old = cmpxchg(&sb->s_dio_done_wq, NULL, wq); 563 /* Someone created workqueue before us? Free ours... */ 564 if (old) 565 destroy_workqueue(wq); 566 return 0; 567 } 568 569 static int dio_set_defer_completion(struct dio *dio) 570 { 571 struct super_block *sb = dio->inode->i_sb; 572 573 if (dio->defer_completion) 574 return 0; 575 dio->defer_completion = true; 576 if (!sb->s_dio_done_wq) 577 return sb_init_dio_done_wq(sb); 578 return 0; 579 } 580 581 /* 582 * Call into the fs to map some more disk blocks. We record the current number 583 * of available blocks at sdio->blocks_available. These are in units of the 584 * fs blocksize, (1 << inode->i_blkbits). 585 * 586 * The fs is allowed to map lots of blocks at once. If it wants to do that, 587 * it uses the passed inode-relative block number as the file offset, as usual. 588 * 589 * get_block() is passed the number of i_blkbits-sized blocks which direct_io 590 * has remaining to do. The fs should not map more than this number of blocks. 591 * 592 * If the fs has mapped a lot of blocks, it should populate bh->b_size to 593 * indicate how much contiguous disk space has been made available at 594 * bh->b_blocknr. 595 * 596 * If *any* of the mapped blocks are new, then the fs must set buffer_new(). 597 * This isn't very efficient... 598 * 599 * In the case of filesystem holes: the fs may return an arbitrarily-large 600 * hole by returning an appropriate value in b_size and by clearing 601 * buffer_mapped(). However the direct-io code will only process holes one 602 * block at a time - it will repeatedly call get_block() as it walks the hole. 603 */ 604 static int get_more_blocks(struct dio *dio, struct dio_submit *sdio, 605 struct buffer_head *map_bh) 606 { 607 int ret; 608 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */ 609 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */ 610 unsigned long fs_count; /* Number of filesystem-sized blocks */ 611 int create; 612 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor; 613 614 /* 615 * If there was a memory error and we've overwritten all the 616 * mapped blocks then we can now return that memory error 617 */ 618 ret = dio->page_errors; 619 if (ret == 0) { 620 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request); 621 fs_startblk = sdio->block_in_file >> sdio->blkfactor; 622 fs_endblk = (sdio->final_block_in_request - 1) >> 623 sdio->blkfactor; 624 fs_count = fs_endblk - fs_startblk + 1; 625 626 map_bh->b_state = 0; 627 map_bh->b_size = fs_count << i_blkbits; 628 629 /* 630 * For writes inside i_size on a DIO_SKIP_HOLES filesystem we 631 * forbid block creations: only overwrites are permitted. 632 * We will return early to the caller once we see an 633 * unmapped buffer head returned, and the caller will fall 634 * back to buffered I/O. 635 * 636 * Otherwise the decision is left to the get_blocks method, 637 * which may decide to handle it or also return an unmapped 638 * buffer head. 639 */ 640 create = dio->rw & WRITE; 641 if (dio->flags & DIO_SKIP_HOLES) { 642 if (sdio->block_in_file < (i_size_read(dio->inode) >> 643 sdio->blkbits)) 644 create = 0; 645 } 646 647 ret = (*sdio->get_block)(dio->inode, fs_startblk, 648 map_bh, create); 649 650 /* Store for completion */ 651 dio->private = map_bh->b_private; 652 653 if (ret == 0 && buffer_defer_completion(map_bh)) 654 ret = dio_set_defer_completion(dio); 655 } 656 return ret; 657 } 658 659 /* 660 * There is no bio. Make one now. 661 */ 662 static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio, 663 sector_t start_sector, struct buffer_head *map_bh) 664 { 665 sector_t sector; 666 int ret, nr_pages; 667 668 ret = dio_bio_reap(dio, sdio); 669 if (ret) 670 goto out; 671 sector = start_sector << (sdio->blkbits - 9); 672 nr_pages = min(sdio->pages_in_io, BIO_MAX_PAGES); 673 BUG_ON(nr_pages <= 0); 674 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages); 675 sdio->boundary = 0; 676 out: 677 return ret; 678 } 679 680 /* 681 * Attempt to put the current chunk of 'cur_page' into the current BIO. If 682 * that was successful then update final_block_in_bio and take a ref against 683 * the just-added page. 684 * 685 * Return zero on success. Non-zero means the caller needs to start a new BIO. 686 */ 687 static inline int dio_bio_add_page(struct dio_submit *sdio) 688 { 689 int ret; 690 691 ret = bio_add_page(sdio->bio, sdio->cur_page, 692 sdio->cur_page_len, sdio->cur_page_offset); 693 if (ret == sdio->cur_page_len) { 694 /* 695 * Decrement count only, if we are done with this page 696 */ 697 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE) 698 sdio->pages_in_io--; 699 page_cache_get(sdio->cur_page); 700 sdio->final_block_in_bio = sdio->cur_page_block + 701 (sdio->cur_page_len >> sdio->blkbits); 702 ret = 0; 703 } else { 704 ret = 1; 705 } 706 return ret; 707 } 708 709 /* 710 * Put cur_page under IO. The section of cur_page which is described by 711 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page 712 * starts on-disk at cur_page_block. 713 * 714 * We take a ref against the page here (on behalf of its presence in the bio). 715 * 716 * The caller of this function is responsible for removing cur_page from the 717 * dio, and for dropping the refcount which came from that presence. 718 */ 719 static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio, 720 struct buffer_head *map_bh) 721 { 722 int ret = 0; 723 724 if (sdio->bio) { 725 loff_t cur_offset = sdio->cur_page_fs_offset; 726 loff_t bio_next_offset = sdio->logical_offset_in_bio + 727 sdio->bio->bi_iter.bi_size; 728 729 /* 730 * See whether this new request is contiguous with the old. 731 * 732 * Btrfs cannot handle having logically non-contiguous requests 733 * submitted. For example if you have 734 * 735 * Logical: [0-4095][HOLE][8192-12287] 736 * Physical: [0-4095] [4096-8191] 737 * 738 * We cannot submit those pages together as one BIO. So if our 739 * current logical offset in the file does not equal what would 740 * be the next logical offset in the bio, submit the bio we 741 * have. 742 */ 743 if (sdio->final_block_in_bio != sdio->cur_page_block || 744 cur_offset != bio_next_offset) 745 dio_bio_submit(dio, sdio); 746 } 747 748 if (sdio->bio == NULL) { 749 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh); 750 if (ret) 751 goto out; 752 } 753 754 if (dio_bio_add_page(sdio) != 0) { 755 dio_bio_submit(dio, sdio); 756 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh); 757 if (ret == 0) { 758 ret = dio_bio_add_page(sdio); 759 BUG_ON(ret != 0); 760 } 761 } 762 out: 763 return ret; 764 } 765 766 /* 767 * An autonomous function to put a chunk of a page under deferred IO. 768 * 769 * The caller doesn't actually know (or care) whether this piece of page is in 770 * a BIO, or is under IO or whatever. We just take care of all possible 771 * situations here. The separation between the logic of do_direct_IO() and 772 * that of submit_page_section() is important for clarity. Please don't break. 773 * 774 * The chunk of page starts on-disk at blocknr. 775 * 776 * We perform deferred IO, by recording the last-submitted page inside our 777 * private part of the dio structure. If possible, we just expand the IO 778 * across that page here. 779 * 780 * If that doesn't work out then we put the old page into the bio and add this 781 * page to the dio instead. 782 */ 783 static inline int 784 submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page, 785 unsigned offset, unsigned len, sector_t blocknr, 786 struct buffer_head *map_bh) 787 { 788 int ret = 0; 789 790 if (dio->rw & WRITE) { 791 /* 792 * Read accounting is performed in submit_bio() 793 */ 794 task_io_account_write(len); 795 } 796 797 /* 798 * Can we just grow the current page's presence in the dio? 799 */ 800 if (sdio->cur_page == page && 801 sdio->cur_page_offset + sdio->cur_page_len == offset && 802 sdio->cur_page_block + 803 (sdio->cur_page_len >> sdio->blkbits) == blocknr) { 804 sdio->cur_page_len += len; 805 goto out; 806 } 807 808 /* 809 * If there's a deferred page already there then send it. 810 */ 811 if (sdio->cur_page) { 812 ret = dio_send_cur_page(dio, sdio, map_bh); 813 page_cache_release(sdio->cur_page); 814 sdio->cur_page = NULL; 815 if (ret) 816 return ret; 817 } 818 819 page_cache_get(page); /* It is in dio */ 820 sdio->cur_page = page; 821 sdio->cur_page_offset = offset; 822 sdio->cur_page_len = len; 823 sdio->cur_page_block = blocknr; 824 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits; 825 out: 826 /* 827 * If sdio->boundary then we want to schedule the IO now to 828 * avoid metadata seeks. 829 */ 830 if (sdio->boundary) { 831 ret = dio_send_cur_page(dio, sdio, map_bh); 832 dio_bio_submit(dio, sdio); 833 page_cache_release(sdio->cur_page); 834 sdio->cur_page = NULL; 835 } 836 return ret; 837 } 838 839 /* 840 * Clean any dirty buffers in the blockdev mapping which alias newly-created 841 * file blocks. Only called for S_ISREG files - blockdevs do not set 842 * buffer_new 843 */ 844 static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh) 845 { 846 unsigned i; 847 unsigned nblocks; 848 849 nblocks = map_bh->b_size >> dio->inode->i_blkbits; 850 851 for (i = 0; i < nblocks; i++) { 852 unmap_underlying_metadata(map_bh->b_bdev, 853 map_bh->b_blocknr + i); 854 } 855 } 856 857 /* 858 * If we are not writing the entire block and get_block() allocated 859 * the block for us, we need to fill-in the unused portion of the 860 * block with zeros. This happens only if user-buffer, fileoffset or 861 * io length is not filesystem block-size multiple. 862 * 863 * `end' is zero if we're doing the start of the IO, 1 at the end of the 864 * IO. 865 */ 866 static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio, 867 int end, struct buffer_head *map_bh) 868 { 869 unsigned dio_blocks_per_fs_block; 870 unsigned this_chunk_blocks; /* In dio_blocks */ 871 unsigned this_chunk_bytes; 872 struct page *page; 873 874 sdio->start_zero_done = 1; 875 if (!sdio->blkfactor || !buffer_new(map_bh)) 876 return; 877 878 dio_blocks_per_fs_block = 1 << sdio->blkfactor; 879 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1); 880 881 if (!this_chunk_blocks) 882 return; 883 884 /* 885 * We need to zero out part of an fs block. It is either at the 886 * beginning or the end of the fs block. 887 */ 888 if (end) 889 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks; 890 891 this_chunk_bytes = this_chunk_blocks << sdio->blkbits; 892 893 page = ZERO_PAGE(0); 894 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes, 895 sdio->next_block_for_io, map_bh)) 896 return; 897 898 sdio->next_block_for_io += this_chunk_blocks; 899 } 900 901 /* 902 * Walk the user pages, and the file, mapping blocks to disk and generating 903 * a sequence of (page,offset,len,block) mappings. These mappings are injected 904 * into submit_page_section(), which takes care of the next stage of submission 905 * 906 * Direct IO against a blockdev is different from a file. Because we can 907 * happily perform page-sized but 512-byte aligned IOs. It is important that 908 * blockdev IO be able to have fine alignment and large sizes. 909 * 910 * So what we do is to permit the ->get_block function to populate bh.b_size 911 * with the size of IO which is permitted at this offset and this i_blkbits. 912 * 913 * For best results, the blockdev should be set up with 512-byte i_blkbits and 914 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives 915 * fine alignment but still allows this function to work in PAGE_SIZE units. 916 */ 917 static int do_direct_IO(struct dio *dio, struct dio_submit *sdio, 918 struct buffer_head *map_bh) 919 { 920 const unsigned blkbits = sdio->blkbits; 921 int ret = 0; 922 923 while (sdio->block_in_file < sdio->final_block_in_request) { 924 struct page *page; 925 size_t from, to; 926 927 page = dio_get_page(dio, sdio); 928 if (IS_ERR(page)) { 929 ret = PTR_ERR(page); 930 goto out; 931 } 932 from = sdio->head ? 0 : sdio->from; 933 to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE; 934 sdio->head++; 935 936 while (from < to) { 937 unsigned this_chunk_bytes; /* # of bytes mapped */ 938 unsigned this_chunk_blocks; /* # of blocks */ 939 unsigned u; 940 941 if (sdio->blocks_available == 0) { 942 /* 943 * Need to go and map some more disk 944 */ 945 unsigned long blkmask; 946 unsigned long dio_remainder; 947 948 ret = get_more_blocks(dio, sdio, map_bh); 949 if (ret) { 950 page_cache_release(page); 951 goto out; 952 } 953 if (!buffer_mapped(map_bh)) 954 goto do_holes; 955 956 sdio->blocks_available = 957 map_bh->b_size >> sdio->blkbits; 958 sdio->next_block_for_io = 959 map_bh->b_blocknr << sdio->blkfactor; 960 if (buffer_new(map_bh)) 961 clean_blockdev_aliases(dio, map_bh); 962 963 if (!sdio->blkfactor) 964 goto do_holes; 965 966 blkmask = (1 << sdio->blkfactor) - 1; 967 dio_remainder = (sdio->block_in_file & blkmask); 968 969 /* 970 * If we are at the start of IO and that IO 971 * starts partway into a fs-block, 972 * dio_remainder will be non-zero. If the IO 973 * is a read then we can simply advance the IO 974 * cursor to the first block which is to be 975 * read. But if the IO is a write and the 976 * block was newly allocated we cannot do that; 977 * the start of the fs block must be zeroed out 978 * on-disk 979 */ 980 if (!buffer_new(map_bh)) 981 sdio->next_block_for_io += dio_remainder; 982 sdio->blocks_available -= dio_remainder; 983 } 984 do_holes: 985 /* Handle holes */ 986 if (!buffer_mapped(map_bh)) { 987 loff_t i_size_aligned; 988 989 /* AKPM: eargh, -ENOTBLK is a hack */ 990 if (dio->rw & WRITE) { 991 page_cache_release(page); 992 return -ENOTBLK; 993 } 994 995 /* 996 * Be sure to account for a partial block as the 997 * last block in the file 998 */ 999 i_size_aligned = ALIGN(i_size_read(dio->inode), 1000 1 << blkbits); 1001 if (sdio->block_in_file >= 1002 i_size_aligned >> blkbits) { 1003 /* We hit eof */ 1004 page_cache_release(page); 1005 goto out; 1006 } 1007 zero_user(page, from, 1 << blkbits); 1008 sdio->block_in_file++; 1009 from += 1 << blkbits; 1010 dio->result += 1 << blkbits; 1011 goto next_block; 1012 } 1013 1014 /* 1015 * If we're performing IO which has an alignment which 1016 * is finer than the underlying fs, go check to see if 1017 * we must zero out the start of this block. 1018 */ 1019 if (unlikely(sdio->blkfactor && !sdio->start_zero_done)) 1020 dio_zero_block(dio, sdio, 0, map_bh); 1021 1022 /* 1023 * Work out, in this_chunk_blocks, how much disk we 1024 * can add to this page 1025 */ 1026 this_chunk_blocks = sdio->blocks_available; 1027 u = (to - from) >> blkbits; 1028 if (this_chunk_blocks > u) 1029 this_chunk_blocks = u; 1030 u = sdio->final_block_in_request - sdio->block_in_file; 1031 if (this_chunk_blocks > u) 1032 this_chunk_blocks = u; 1033 this_chunk_bytes = this_chunk_blocks << blkbits; 1034 BUG_ON(this_chunk_bytes == 0); 1035 1036 if (this_chunk_blocks == sdio->blocks_available) 1037 sdio->boundary = buffer_boundary(map_bh); 1038 ret = submit_page_section(dio, sdio, page, 1039 from, 1040 this_chunk_bytes, 1041 sdio->next_block_for_io, 1042 map_bh); 1043 if (ret) { 1044 page_cache_release(page); 1045 goto out; 1046 } 1047 sdio->next_block_for_io += this_chunk_blocks; 1048 1049 sdio->block_in_file += this_chunk_blocks; 1050 from += this_chunk_bytes; 1051 dio->result += this_chunk_bytes; 1052 sdio->blocks_available -= this_chunk_blocks; 1053 next_block: 1054 BUG_ON(sdio->block_in_file > sdio->final_block_in_request); 1055 if (sdio->block_in_file == sdio->final_block_in_request) 1056 break; 1057 } 1058 1059 /* Drop the ref which was taken in get_user_pages() */ 1060 page_cache_release(page); 1061 } 1062 out: 1063 return ret; 1064 } 1065 1066 static inline int drop_refcount(struct dio *dio) 1067 { 1068 int ret2; 1069 unsigned long flags; 1070 1071 /* 1072 * Sync will always be dropping the final ref and completing the 1073 * operation. AIO can if it was a broken operation described above or 1074 * in fact if all the bios race to complete before we get here. In 1075 * that case dio_complete() translates the EIOCBQUEUED into the proper 1076 * return code that the caller will hand to ->complete(). 1077 * 1078 * This is managed by the bio_lock instead of being an atomic_t so that 1079 * completion paths can drop their ref and use the remaining count to 1080 * decide to wake the submission path atomically. 1081 */ 1082 spin_lock_irqsave(&dio->bio_lock, flags); 1083 ret2 = --dio->refcount; 1084 spin_unlock_irqrestore(&dio->bio_lock, flags); 1085 return ret2; 1086 } 1087 1088 /* 1089 * This is a library function for use by filesystem drivers. 1090 * 1091 * The locking rules are governed by the flags parameter: 1092 * - if the flags value contains DIO_LOCKING we use a fancy locking 1093 * scheme for dumb filesystems. 1094 * For writes this function is called under i_mutex and returns with 1095 * i_mutex held, for reads, i_mutex is not held on entry, but it is 1096 * taken and dropped again before returning. 1097 * - if the flags value does NOT contain DIO_LOCKING we don't use any 1098 * internal locking but rather rely on the filesystem to synchronize 1099 * direct I/O reads/writes versus each other and truncate. 1100 * 1101 * To help with locking against truncate we incremented the i_dio_count 1102 * counter before starting direct I/O, and decrement it once we are done. 1103 * Truncate can wait for it to reach zero to provide exclusion. It is 1104 * expected that filesystem provide exclusion between new direct I/O 1105 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex, 1106 * but other filesystems need to take care of this on their own. 1107 * 1108 * NOTE: if you pass "sdio" to anything by pointer make sure that function 1109 * is always inlined. Otherwise gcc is unable to split the structure into 1110 * individual fields and will generate much worse code. This is important 1111 * for the whole file. 1112 */ 1113 static inline ssize_t 1114 do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode, 1115 struct block_device *bdev, struct iov_iter *iter, 1116 loff_t offset, get_block_t get_block, dio_iodone_t end_io, 1117 dio_submit_t submit_io, int flags) 1118 { 1119 unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits); 1120 unsigned blkbits = i_blkbits; 1121 unsigned blocksize_mask = (1 << blkbits) - 1; 1122 ssize_t retval = -EINVAL; 1123 size_t count = iov_iter_count(iter); 1124 loff_t end = offset + count; 1125 struct dio *dio; 1126 struct dio_submit sdio = { 0, }; 1127 struct buffer_head map_bh = { 0, }; 1128 struct blk_plug plug; 1129 unsigned long align = offset | iov_iter_alignment(iter); 1130 1131 /* 1132 * Avoid references to bdev if not absolutely needed to give 1133 * the early prefetch in the caller enough time. 1134 */ 1135 1136 if (align & blocksize_mask) { 1137 if (bdev) 1138 blkbits = blksize_bits(bdev_logical_block_size(bdev)); 1139 blocksize_mask = (1 << blkbits) - 1; 1140 if (align & blocksize_mask) 1141 goto out; 1142 } 1143 1144 /* watch out for a 0 len io from a tricksy fs */ 1145 if (iov_iter_rw(iter) == READ && !iov_iter_count(iter)) 1146 return 0; 1147 1148 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL); 1149 retval = -ENOMEM; 1150 if (!dio) 1151 goto out; 1152 /* 1153 * Believe it or not, zeroing out the page array caused a .5% 1154 * performance regression in a database benchmark. So, we take 1155 * care to only zero out what's needed. 1156 */ 1157 memset(dio, 0, offsetof(struct dio, pages)); 1158 1159 dio->flags = flags; 1160 if (dio->flags & DIO_LOCKING) { 1161 if (iov_iter_rw(iter) == READ) { 1162 struct address_space *mapping = 1163 iocb->ki_filp->f_mapping; 1164 1165 /* will be released by direct_io_worker */ 1166 inode_lock(inode); 1167 1168 retval = filemap_write_and_wait_range(mapping, offset, 1169 end - 1); 1170 if (retval) { 1171 inode_unlock(inode); 1172 kmem_cache_free(dio_cache, dio); 1173 goto out; 1174 } 1175 } 1176 } 1177 1178 /* Once we sampled i_size check for reads beyond EOF */ 1179 dio->i_size = i_size_read(inode); 1180 if (iov_iter_rw(iter) == READ && offset >= dio->i_size) { 1181 if (dio->flags & DIO_LOCKING) 1182 inode_unlock(inode); 1183 kmem_cache_free(dio_cache, dio); 1184 retval = 0; 1185 goto out; 1186 } 1187 1188 /* 1189 * For file extending writes updating i_size before data writeouts 1190 * complete can expose uninitialized blocks in dumb filesystems. 1191 * In that case we need to wait for I/O completion even if asked 1192 * for an asynchronous write. 1193 */ 1194 if (is_sync_kiocb(iocb)) 1195 dio->is_async = false; 1196 else if (!(dio->flags & DIO_ASYNC_EXTEND) && 1197 iov_iter_rw(iter) == WRITE && end > i_size_read(inode)) 1198 dio->is_async = false; 1199 else 1200 dio->is_async = true; 1201 1202 dio->inode = inode; 1203 dio->rw = iov_iter_rw(iter) == WRITE ? WRITE_ODIRECT : READ; 1204 1205 /* 1206 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue 1207 * so that we can call ->fsync. 1208 */ 1209 if (dio->is_async && iov_iter_rw(iter) == WRITE && 1210 ((iocb->ki_filp->f_flags & O_DSYNC) || 1211 IS_SYNC(iocb->ki_filp->f_mapping->host))) { 1212 retval = dio_set_defer_completion(dio); 1213 if (retval) { 1214 /* 1215 * We grab i_mutex only for reads so we don't have 1216 * to release it here 1217 */ 1218 kmem_cache_free(dio_cache, dio); 1219 goto out; 1220 } 1221 } 1222 1223 /* 1224 * Will be decremented at I/O completion time. 1225 */ 1226 if (!(dio->flags & DIO_SKIP_DIO_COUNT)) 1227 inode_dio_begin(inode); 1228 1229 retval = 0; 1230 sdio.blkbits = blkbits; 1231 sdio.blkfactor = i_blkbits - blkbits; 1232 sdio.block_in_file = offset >> blkbits; 1233 1234 sdio.get_block = get_block; 1235 dio->end_io = end_io; 1236 sdio.submit_io = submit_io; 1237 sdio.final_block_in_bio = -1; 1238 sdio.next_block_for_io = -1; 1239 1240 dio->iocb = iocb; 1241 1242 spin_lock_init(&dio->bio_lock); 1243 dio->refcount = 1; 1244 1245 dio->should_dirty = (iter->type == ITER_IOVEC); 1246 sdio.iter = iter; 1247 sdio.final_block_in_request = 1248 (offset + iov_iter_count(iter)) >> blkbits; 1249 1250 /* 1251 * In case of non-aligned buffers, we may need 2 more 1252 * pages since we need to zero out first and last block. 1253 */ 1254 if (unlikely(sdio.blkfactor)) 1255 sdio.pages_in_io = 2; 1256 1257 sdio.pages_in_io += iov_iter_npages(iter, INT_MAX); 1258 1259 blk_start_plug(&plug); 1260 1261 retval = do_direct_IO(dio, &sdio, &map_bh); 1262 if (retval) 1263 dio_cleanup(dio, &sdio); 1264 1265 if (retval == -ENOTBLK) { 1266 /* 1267 * The remaining part of the request will be 1268 * be handled by buffered I/O when we return 1269 */ 1270 retval = 0; 1271 } 1272 /* 1273 * There may be some unwritten disk at the end of a part-written 1274 * fs-block-sized block. Go zero that now. 1275 */ 1276 dio_zero_block(dio, &sdio, 1, &map_bh); 1277 1278 if (sdio.cur_page) { 1279 ssize_t ret2; 1280 1281 ret2 = dio_send_cur_page(dio, &sdio, &map_bh); 1282 if (retval == 0) 1283 retval = ret2; 1284 page_cache_release(sdio.cur_page); 1285 sdio.cur_page = NULL; 1286 } 1287 if (sdio.bio) 1288 dio_bio_submit(dio, &sdio); 1289 1290 blk_finish_plug(&plug); 1291 1292 /* 1293 * It is possible that, we return short IO due to end of file. 1294 * In that case, we need to release all the pages we got hold on. 1295 */ 1296 dio_cleanup(dio, &sdio); 1297 1298 /* 1299 * All block lookups have been performed. For READ requests 1300 * we can let i_mutex go now that its achieved its purpose 1301 * of protecting us from looking up uninitialized blocks. 1302 */ 1303 if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING)) 1304 inode_unlock(dio->inode); 1305 1306 /* 1307 * The only time we want to leave bios in flight is when a successful 1308 * partial aio read or full aio write have been setup. In that case 1309 * bio completion will call aio_complete. The only time it's safe to 1310 * call aio_complete is when we return -EIOCBQUEUED, so we key on that. 1311 * This had *better* be the only place that raises -EIOCBQUEUED. 1312 */ 1313 BUG_ON(retval == -EIOCBQUEUED); 1314 if (dio->is_async && retval == 0 && dio->result && 1315 (iov_iter_rw(iter) == READ || dio->result == count)) 1316 retval = -EIOCBQUEUED; 1317 else 1318 dio_await_completion(dio); 1319 1320 if (drop_refcount(dio) == 0) { 1321 retval = dio_complete(dio, offset, retval, false); 1322 } else 1323 BUG_ON(retval != -EIOCBQUEUED); 1324 1325 out: 1326 return retval; 1327 } 1328 1329 ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode, 1330 struct block_device *bdev, struct iov_iter *iter, 1331 loff_t offset, get_block_t get_block, 1332 dio_iodone_t end_io, dio_submit_t submit_io, 1333 int flags) 1334 { 1335 /* 1336 * The block device state is needed in the end to finally 1337 * submit everything. Since it's likely to be cache cold 1338 * prefetch it here as first thing to hide some of the 1339 * latency. 1340 * 1341 * Attempt to prefetch the pieces we likely need later. 1342 */ 1343 prefetch(&bdev->bd_disk->part_tbl); 1344 prefetch(bdev->bd_queue); 1345 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES); 1346 1347 return do_blockdev_direct_IO(iocb, inode, bdev, iter, offset, get_block, 1348 end_io, submit_io, flags); 1349 } 1350 1351 EXPORT_SYMBOL(__blockdev_direct_IO); 1352 1353 static __init int dio_init(void) 1354 { 1355 dio_cache = KMEM_CACHE(dio, SLAB_PANIC); 1356 return 0; 1357 } 1358 module_init(dio_init) 1359