1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * fs/direct-io.c 4 * 5 * Copyright (C) 2002, Linus Torvalds. 6 * 7 * O_DIRECT 8 * 9 * 04Jul2002 Andrew Morton 10 * Initial version 11 * 11Sep2002 janetinc@us.ibm.com 12 * added readv/writev support. 13 * 29Oct2002 Andrew Morton 14 * rewrote bio_add_page() support. 15 * 30Oct2002 pbadari@us.ibm.com 16 * added support for non-aligned IO. 17 * 06Nov2002 pbadari@us.ibm.com 18 * added asynchronous IO support. 19 * 21Jul2003 nathans@sgi.com 20 * added IO completion notifier. 21 */ 22 23 #include <linux/kernel.h> 24 #include <linux/module.h> 25 #include <linux/types.h> 26 #include <linux/fs.h> 27 #include <linux/mm.h> 28 #include <linux/slab.h> 29 #include <linux/highmem.h> 30 #include <linux/pagemap.h> 31 #include <linux/task_io_accounting_ops.h> 32 #include <linux/bio.h> 33 #include <linux/wait.h> 34 #include <linux/err.h> 35 #include <linux/blkdev.h> 36 #include <linux/buffer_head.h> 37 #include <linux/rwsem.h> 38 #include <linux/uio.h> 39 #include <linux/atomic.h> 40 41 #include "internal.h" 42 43 /* 44 * How many user pages to map in one call to iov_iter_extract_pages(). This 45 * determines the size of a structure in the slab cache 46 */ 47 #define DIO_PAGES 64 48 49 /* 50 * Flags for dio_complete() 51 */ 52 #define DIO_COMPLETE_ASYNC 0x01 /* This is async IO */ 53 #define DIO_COMPLETE_INVALIDATE 0x02 /* Can invalidate pages */ 54 55 /* 56 * This code generally works in units of "dio_blocks". A dio_block is 57 * somewhere between the hard sector size and the filesystem block size. it 58 * is determined on a per-invocation basis. When talking to the filesystem 59 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity 60 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted 61 * to bio_block quantities by shifting left by blkfactor. 62 * 63 * If blkfactor is zero then the user's request was aligned to the filesystem's 64 * blocksize. 65 */ 66 67 /* dio_state only used in the submission path */ 68 69 struct dio_submit { 70 struct bio *bio; /* bio under assembly */ 71 unsigned blkbits; /* doesn't change */ 72 unsigned blkfactor; /* When we're using an alignment which 73 is finer than the filesystem's soft 74 blocksize, this specifies how much 75 finer. blkfactor=2 means 1/4-block 76 alignment. Does not change */ 77 unsigned start_zero_done; /* flag: sub-blocksize zeroing has 78 been performed at the start of a 79 write */ 80 int pages_in_io; /* approximate total IO pages */ 81 sector_t block_in_file; /* Current offset into the underlying 82 file in dio_block units. */ 83 unsigned blocks_available; /* At block_in_file. changes */ 84 int reap_counter; /* rate limit reaping */ 85 sector_t final_block_in_request;/* doesn't change */ 86 int boundary; /* prev block is at a boundary */ 87 get_block_t *get_block; /* block mapping function */ 88 89 loff_t logical_offset_in_bio; /* current first logical block in bio */ 90 sector_t final_block_in_bio; /* current final block in bio + 1 */ 91 sector_t next_block_for_io; /* next block to be put under IO, 92 in dio_blocks units */ 93 94 /* 95 * Deferred addition of a page to the dio. These variables are 96 * private to dio_send_cur_page(), submit_page_section() and 97 * dio_bio_add_page(). 98 */ 99 struct page *cur_page; /* The page */ 100 unsigned cur_page_offset; /* Offset into it, in bytes */ 101 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */ 102 sector_t cur_page_block; /* Where it starts */ 103 loff_t cur_page_fs_offset; /* Offset in file */ 104 105 struct iov_iter *iter; 106 /* 107 * Page queue. These variables belong to dio_refill_pages() and 108 * dio_get_page(). 109 */ 110 unsigned head; /* next page to process */ 111 unsigned tail; /* last valid page + 1 */ 112 size_t from, to; 113 }; 114 115 /* dio_state communicated between submission path and end_io */ 116 struct dio { 117 int flags; /* doesn't change */ 118 blk_opf_t opf; /* request operation type and flags */ 119 struct gendisk *bio_disk; 120 struct inode *inode; 121 loff_t i_size; /* i_size when submitted */ 122 dio_iodone_t *end_io; /* IO completion function */ 123 bool is_pinned; /* T if we have pins on the pages */ 124 125 void *private; /* copy from map_bh.b_private */ 126 127 /* BIO completion state */ 128 spinlock_t bio_lock; /* protects BIO fields below */ 129 int page_errors; /* err from iov_iter_extract_pages() */ 130 int is_async; /* is IO async ? */ 131 bool defer_completion; /* defer AIO completion to workqueue? */ 132 bool should_dirty; /* if pages should be dirtied */ 133 int io_error; /* IO error in completion path */ 134 unsigned long refcount; /* direct_io_worker() and bios */ 135 struct bio *bio_list; /* singly linked via bi_private */ 136 struct task_struct *waiter; /* waiting task (NULL if none) */ 137 138 /* AIO related stuff */ 139 struct kiocb *iocb; /* kiocb */ 140 ssize_t result; /* IO result */ 141 142 /* 143 * pages[] (and any fields placed after it) are not zeroed out at 144 * allocation time. Don't add new fields after pages[] unless you 145 * wish that they not be zeroed. 146 */ 147 union { 148 struct page *pages[DIO_PAGES]; /* page buffer */ 149 struct work_struct complete_work;/* deferred AIO completion */ 150 }; 151 } ____cacheline_aligned_in_smp; 152 153 static struct kmem_cache *dio_cache __ro_after_init; 154 155 /* 156 * How many pages are in the queue? 157 */ 158 static inline unsigned dio_pages_present(struct dio_submit *sdio) 159 { 160 return sdio->tail - sdio->head; 161 } 162 163 /* 164 * Go grab and pin some userspace pages. Typically we'll get 64 at a time. 165 */ 166 static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio) 167 { 168 struct page **pages = dio->pages; 169 const enum req_op dio_op = dio->opf & REQ_OP_MASK; 170 ssize_t ret; 171 172 ret = iov_iter_extract_pages(sdio->iter, &pages, LONG_MAX, 173 DIO_PAGES, 0, &sdio->from); 174 175 if (ret < 0 && sdio->blocks_available && dio_op == REQ_OP_WRITE) { 176 /* 177 * A memory fault, but the filesystem has some outstanding 178 * mapped blocks. We need to use those blocks up to avoid 179 * leaking stale data in the file. 180 */ 181 if (dio->page_errors == 0) 182 dio->page_errors = ret; 183 dio->pages[0] = ZERO_PAGE(0); 184 sdio->head = 0; 185 sdio->tail = 1; 186 sdio->from = 0; 187 sdio->to = PAGE_SIZE; 188 return 0; 189 } 190 191 if (ret >= 0) { 192 ret += sdio->from; 193 sdio->head = 0; 194 sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE; 195 sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1; 196 return 0; 197 } 198 return ret; 199 } 200 201 /* 202 * Get another userspace page. Returns an ERR_PTR on error. Pages are 203 * buffered inside the dio so that we can call iov_iter_extract_pages() 204 * against a decent number of pages, less frequently. To provide nicer use of 205 * the L1 cache. 206 */ 207 static inline struct page *dio_get_page(struct dio *dio, 208 struct dio_submit *sdio) 209 { 210 if (dio_pages_present(sdio) == 0) { 211 int ret; 212 213 ret = dio_refill_pages(dio, sdio); 214 if (ret) 215 return ERR_PTR(ret); 216 BUG_ON(dio_pages_present(sdio) == 0); 217 } 218 return dio->pages[sdio->head]; 219 } 220 221 static void dio_pin_page(struct dio *dio, struct page *page) 222 { 223 if (dio->is_pinned) 224 folio_add_pin(page_folio(page)); 225 } 226 227 static void dio_unpin_page(struct dio *dio, struct page *page) 228 { 229 if (dio->is_pinned) 230 unpin_user_page(page); 231 } 232 233 /* 234 * dio_complete() - called when all DIO BIO I/O has been completed 235 * 236 * This drops i_dio_count, lets interested parties know that a DIO operation 237 * has completed, and calculates the resulting return code for the operation. 238 * 239 * It lets the filesystem know if it registered an interest earlier via 240 * get_block. Pass the private field of the map buffer_head so that 241 * filesystems can use it to hold additional state between get_block calls and 242 * dio_complete. 243 */ 244 static ssize_t dio_complete(struct dio *dio, ssize_t ret, unsigned int flags) 245 { 246 const enum req_op dio_op = dio->opf & REQ_OP_MASK; 247 loff_t offset = dio->iocb->ki_pos; 248 ssize_t transferred = 0; 249 int err; 250 251 /* 252 * AIO submission can race with bio completion to get here while 253 * expecting to have the last io completed by bio completion. 254 * In that case -EIOCBQUEUED is in fact not an error we want 255 * to preserve through this call. 256 */ 257 if (ret == -EIOCBQUEUED) 258 ret = 0; 259 260 if (dio->result) { 261 transferred = dio->result; 262 263 /* Check for short read case */ 264 if (dio_op == REQ_OP_READ && 265 ((offset + transferred) > dio->i_size)) 266 transferred = dio->i_size - offset; 267 /* ignore EFAULT if some IO has been done */ 268 if (unlikely(ret == -EFAULT) && transferred) 269 ret = 0; 270 } 271 272 if (ret == 0) 273 ret = dio->page_errors; 274 if (ret == 0) 275 ret = dio->io_error; 276 if (ret == 0) 277 ret = transferred; 278 279 if (dio->end_io) { 280 // XXX: ki_pos?? 281 err = dio->end_io(dio->iocb, offset, ret, dio->private); 282 if (err) 283 ret = err; 284 } 285 286 /* 287 * Try again to invalidate clean pages which might have been cached by 288 * non-direct readahead, or faulted in by get_user_pages() if the source 289 * of the write was an mmap'ed region of the file we're writing. Either 290 * one is a pretty crazy thing to do, so we don't support it 100%. If 291 * this invalidation fails, tough, the write still worked... 292 * 293 * And this page cache invalidation has to be after dio->end_io(), as 294 * some filesystems convert unwritten extents to real allocations in 295 * end_io() when necessary, otherwise a racing buffer read would cache 296 * zeros from unwritten extents. 297 */ 298 if (flags & DIO_COMPLETE_INVALIDATE && 299 ret > 0 && dio_op == REQ_OP_WRITE) 300 kiocb_invalidate_post_direct_write(dio->iocb, ret); 301 302 inode_dio_end(dio->inode); 303 304 if (flags & DIO_COMPLETE_ASYNC) { 305 /* 306 * generic_write_sync expects ki_pos to have been updated 307 * already, but the submission path only does this for 308 * synchronous I/O. 309 */ 310 dio->iocb->ki_pos += transferred; 311 312 if (ret > 0 && dio_op == REQ_OP_WRITE) 313 ret = generic_write_sync(dio->iocb, ret); 314 dio->iocb->ki_complete(dio->iocb, ret); 315 } 316 317 kmem_cache_free(dio_cache, dio); 318 return ret; 319 } 320 321 static void dio_aio_complete_work(struct work_struct *work) 322 { 323 struct dio *dio = container_of(work, struct dio, complete_work); 324 325 dio_complete(dio, 0, DIO_COMPLETE_ASYNC | DIO_COMPLETE_INVALIDATE); 326 } 327 328 static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio); 329 330 /* 331 * Asynchronous IO callback. 332 */ 333 static void dio_bio_end_aio(struct bio *bio) 334 { 335 struct dio *dio = bio->bi_private; 336 const enum req_op dio_op = dio->opf & REQ_OP_MASK; 337 unsigned long remaining; 338 unsigned long flags; 339 bool defer_completion = false; 340 341 /* cleanup the bio */ 342 dio_bio_complete(dio, bio); 343 344 spin_lock_irqsave(&dio->bio_lock, flags); 345 remaining = --dio->refcount; 346 if (remaining == 1 && dio->waiter) 347 wake_up_process(dio->waiter); 348 spin_unlock_irqrestore(&dio->bio_lock, flags); 349 350 if (remaining == 0) { 351 /* 352 * Defer completion when defer_completion is set or 353 * when the inode has pages mapped and this is AIO write. 354 * We need to invalidate those pages because there is a 355 * chance they contain stale data in the case buffered IO 356 * went in between AIO submission and completion into the 357 * same region. 358 */ 359 if (dio->result) 360 defer_completion = dio->defer_completion || 361 (dio_op == REQ_OP_WRITE && 362 dio->inode->i_mapping->nrpages); 363 if (defer_completion) { 364 INIT_WORK(&dio->complete_work, dio_aio_complete_work); 365 queue_work(dio->inode->i_sb->s_dio_done_wq, 366 &dio->complete_work); 367 } else { 368 dio_complete(dio, 0, DIO_COMPLETE_ASYNC); 369 } 370 } 371 } 372 373 /* 374 * The BIO completion handler simply queues the BIO up for the process-context 375 * handler. 376 * 377 * During I/O bi_private points at the dio. After I/O, bi_private is used to 378 * implement a singly-linked list of completed BIOs, at dio->bio_list. 379 */ 380 static void dio_bio_end_io(struct bio *bio) 381 { 382 struct dio *dio = bio->bi_private; 383 unsigned long flags; 384 385 spin_lock_irqsave(&dio->bio_lock, flags); 386 bio->bi_private = dio->bio_list; 387 dio->bio_list = bio; 388 if (--dio->refcount == 1 && dio->waiter) 389 wake_up_process(dio->waiter); 390 spin_unlock_irqrestore(&dio->bio_lock, flags); 391 } 392 393 static inline void 394 dio_bio_alloc(struct dio *dio, struct dio_submit *sdio, 395 struct block_device *bdev, 396 sector_t first_sector, int nr_vecs) 397 { 398 struct bio *bio; 399 400 /* 401 * bio_alloc() is guaranteed to return a bio when allowed to sleep and 402 * we request a valid number of vectors. 403 */ 404 bio = bio_alloc(bdev, nr_vecs, dio->opf, GFP_KERNEL); 405 bio->bi_iter.bi_sector = first_sector; 406 if (dio->is_async) 407 bio->bi_end_io = dio_bio_end_aio; 408 else 409 bio->bi_end_io = dio_bio_end_io; 410 if (dio->is_pinned) 411 bio_set_flag(bio, BIO_PAGE_PINNED); 412 bio->bi_write_hint = file_inode(dio->iocb->ki_filp)->i_write_hint; 413 414 sdio->bio = bio; 415 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset; 416 } 417 418 /* 419 * In the AIO read case we speculatively dirty the pages before starting IO. 420 * During IO completion, any of these pages which happen to have been written 421 * back will be redirtied by bio_check_pages_dirty(). 422 * 423 * bios hold a dio reference between submit_bio and ->end_io. 424 */ 425 static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio) 426 { 427 const enum req_op dio_op = dio->opf & REQ_OP_MASK; 428 struct bio *bio = sdio->bio; 429 unsigned long flags; 430 431 bio->bi_private = dio; 432 433 spin_lock_irqsave(&dio->bio_lock, flags); 434 dio->refcount++; 435 spin_unlock_irqrestore(&dio->bio_lock, flags); 436 437 if (dio->is_async && dio_op == REQ_OP_READ && dio->should_dirty) 438 bio_set_pages_dirty(bio); 439 440 dio->bio_disk = bio->bi_bdev->bd_disk; 441 442 submit_bio(bio); 443 444 sdio->bio = NULL; 445 sdio->boundary = 0; 446 sdio->logical_offset_in_bio = 0; 447 } 448 449 /* 450 * Release any resources in case of a failure 451 */ 452 static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio) 453 { 454 if (dio->is_pinned) 455 unpin_user_pages(dio->pages + sdio->head, 456 sdio->tail - sdio->head); 457 sdio->head = sdio->tail; 458 } 459 460 /* 461 * Wait for the next BIO to complete. Remove it and return it. NULL is 462 * returned once all BIOs have been completed. This must only be called once 463 * all bios have been issued so that dio->refcount can only decrease. This 464 * requires that the caller hold a reference on the dio. 465 */ 466 static struct bio *dio_await_one(struct dio *dio) 467 { 468 unsigned long flags; 469 struct bio *bio = NULL; 470 471 spin_lock_irqsave(&dio->bio_lock, flags); 472 473 /* 474 * Wait as long as the list is empty and there are bios in flight. bio 475 * completion drops the count, maybe adds to the list, and wakes while 476 * holding the bio_lock so we don't need set_current_state()'s barrier 477 * and can call it after testing our condition. 478 */ 479 while (dio->refcount > 1 && dio->bio_list == NULL) { 480 __set_current_state(TASK_UNINTERRUPTIBLE); 481 dio->waiter = current; 482 spin_unlock_irqrestore(&dio->bio_lock, flags); 483 blk_io_schedule(); 484 /* wake up sets us TASK_RUNNING */ 485 spin_lock_irqsave(&dio->bio_lock, flags); 486 dio->waiter = NULL; 487 } 488 if (dio->bio_list) { 489 bio = dio->bio_list; 490 dio->bio_list = bio->bi_private; 491 } 492 spin_unlock_irqrestore(&dio->bio_lock, flags); 493 return bio; 494 } 495 496 /* 497 * Process one completed BIO. No locks are held. 498 */ 499 static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio) 500 { 501 blk_status_t err = bio->bi_status; 502 const enum req_op dio_op = dio->opf & REQ_OP_MASK; 503 bool should_dirty = dio_op == REQ_OP_READ && dio->should_dirty; 504 505 if (err) 506 dio->io_error = -EIO; 507 508 if (dio->is_async && should_dirty) { 509 bio_check_pages_dirty(bio); /* transfers ownership */ 510 } else { 511 bio_release_pages(bio, should_dirty); 512 bio_put(bio); 513 } 514 return err; 515 } 516 517 /* 518 * Wait on and process all in-flight BIOs. This must only be called once 519 * all bios have been issued so that the refcount can only decrease. 520 * This just waits for all bios to make it through dio_bio_complete. IO 521 * errors are propagated through dio->io_error and should be propagated via 522 * dio_complete(). 523 */ 524 static void dio_await_completion(struct dio *dio) 525 { 526 struct bio *bio; 527 do { 528 bio = dio_await_one(dio); 529 if (bio) 530 dio_bio_complete(dio, bio); 531 } while (bio); 532 } 533 534 /* 535 * A really large O_DIRECT read or write can generate a lot of BIOs. So 536 * to keep the memory consumption sane we periodically reap any completed BIOs 537 * during the BIO generation phase. 538 * 539 * This also helps to limit the peak amount of pinned userspace memory. 540 */ 541 static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio) 542 { 543 int ret = 0; 544 545 if (sdio->reap_counter++ >= 64) { 546 while (dio->bio_list) { 547 unsigned long flags; 548 struct bio *bio; 549 int ret2; 550 551 spin_lock_irqsave(&dio->bio_lock, flags); 552 bio = dio->bio_list; 553 dio->bio_list = bio->bi_private; 554 spin_unlock_irqrestore(&dio->bio_lock, flags); 555 ret2 = blk_status_to_errno(dio_bio_complete(dio, bio)); 556 if (ret == 0) 557 ret = ret2; 558 } 559 sdio->reap_counter = 0; 560 } 561 return ret; 562 } 563 564 static int dio_set_defer_completion(struct dio *dio) 565 { 566 struct super_block *sb = dio->inode->i_sb; 567 568 if (dio->defer_completion) 569 return 0; 570 dio->defer_completion = true; 571 if (!sb->s_dio_done_wq) 572 return sb_init_dio_done_wq(sb); 573 return 0; 574 } 575 576 /* 577 * Call into the fs to map some more disk blocks. We record the current number 578 * of available blocks at sdio->blocks_available. These are in units of the 579 * fs blocksize, i_blocksize(inode). 580 * 581 * The fs is allowed to map lots of blocks at once. If it wants to do that, 582 * it uses the passed inode-relative block number as the file offset, as usual. 583 * 584 * get_block() is passed the number of i_blkbits-sized blocks which direct_io 585 * has remaining to do. The fs should not map more than this number of blocks. 586 * 587 * If the fs has mapped a lot of blocks, it should populate bh->b_size to 588 * indicate how much contiguous disk space has been made available at 589 * bh->b_blocknr. 590 * 591 * If *any* of the mapped blocks are new, then the fs must set buffer_new(). 592 * This isn't very efficient... 593 * 594 * In the case of filesystem holes: the fs may return an arbitrarily-large 595 * hole by returning an appropriate value in b_size and by clearing 596 * buffer_mapped(). However the direct-io code will only process holes one 597 * block at a time - it will repeatedly call get_block() as it walks the hole. 598 */ 599 static int get_more_blocks(struct dio *dio, struct dio_submit *sdio, 600 struct buffer_head *map_bh) 601 { 602 const enum req_op dio_op = dio->opf & REQ_OP_MASK; 603 int ret; 604 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */ 605 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */ 606 unsigned long fs_count; /* Number of filesystem-sized blocks */ 607 int create; 608 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor; 609 loff_t i_size; 610 611 /* 612 * If there was a memory error and we've overwritten all the 613 * mapped blocks then we can now return that memory error 614 */ 615 ret = dio->page_errors; 616 if (ret == 0) { 617 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request); 618 fs_startblk = sdio->block_in_file >> sdio->blkfactor; 619 fs_endblk = (sdio->final_block_in_request - 1) >> 620 sdio->blkfactor; 621 fs_count = fs_endblk - fs_startblk + 1; 622 623 map_bh->b_state = 0; 624 map_bh->b_size = fs_count << i_blkbits; 625 626 /* 627 * For writes that could fill holes inside i_size on a 628 * DIO_SKIP_HOLES filesystem we forbid block creations: only 629 * overwrites are permitted. We will return early to the caller 630 * once we see an unmapped buffer head returned, and the caller 631 * will fall back to buffered I/O. 632 * 633 * Otherwise the decision is left to the get_blocks method, 634 * which may decide to handle it or also return an unmapped 635 * buffer head. 636 */ 637 create = dio_op == REQ_OP_WRITE; 638 if (dio->flags & DIO_SKIP_HOLES) { 639 i_size = i_size_read(dio->inode); 640 if (i_size && fs_startblk <= (i_size - 1) >> i_blkbits) 641 create = 0; 642 } 643 644 ret = (*sdio->get_block)(dio->inode, fs_startblk, 645 map_bh, create); 646 647 /* Store for completion */ 648 dio->private = map_bh->b_private; 649 650 if (ret == 0 && buffer_defer_completion(map_bh)) 651 ret = dio_set_defer_completion(dio); 652 } 653 return ret; 654 } 655 656 /* 657 * There is no bio. Make one now. 658 */ 659 static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio, 660 sector_t start_sector, struct buffer_head *map_bh) 661 { 662 sector_t sector; 663 int ret, nr_pages; 664 665 ret = dio_bio_reap(dio, sdio); 666 if (ret) 667 goto out; 668 sector = start_sector << (sdio->blkbits - 9); 669 nr_pages = bio_max_segs(sdio->pages_in_io); 670 BUG_ON(nr_pages <= 0); 671 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages); 672 sdio->boundary = 0; 673 out: 674 return ret; 675 } 676 677 /* 678 * Attempt to put the current chunk of 'cur_page' into the current BIO. If 679 * that was successful then update final_block_in_bio and take a ref against 680 * the just-added page. 681 * 682 * Return zero on success. Non-zero means the caller needs to start a new BIO. 683 */ 684 static inline int dio_bio_add_page(struct dio *dio, struct dio_submit *sdio) 685 { 686 int ret; 687 688 ret = bio_add_page(sdio->bio, sdio->cur_page, 689 sdio->cur_page_len, sdio->cur_page_offset); 690 if (ret == sdio->cur_page_len) { 691 /* 692 * Decrement count only, if we are done with this page 693 */ 694 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE) 695 sdio->pages_in_io--; 696 dio_pin_page(dio, sdio->cur_page); 697 sdio->final_block_in_bio = sdio->cur_page_block + 698 (sdio->cur_page_len >> sdio->blkbits); 699 ret = 0; 700 } else { 701 ret = 1; 702 } 703 return ret; 704 } 705 706 /* 707 * Put cur_page under IO. The section of cur_page which is described by 708 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page 709 * starts on-disk at cur_page_block. 710 * 711 * We take a ref against the page here (on behalf of its presence in the bio). 712 * 713 * The caller of this function is responsible for removing cur_page from the 714 * dio, and for dropping the refcount which came from that presence. 715 */ 716 static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio, 717 struct buffer_head *map_bh) 718 { 719 int ret = 0; 720 721 if (sdio->bio) { 722 loff_t cur_offset = sdio->cur_page_fs_offset; 723 loff_t bio_next_offset = sdio->logical_offset_in_bio + 724 sdio->bio->bi_iter.bi_size; 725 726 /* 727 * See whether this new request is contiguous with the old. 728 * 729 * Btrfs cannot handle having logically non-contiguous requests 730 * submitted. For example if you have 731 * 732 * Logical: [0-4095][HOLE][8192-12287] 733 * Physical: [0-4095] [4096-8191] 734 * 735 * We cannot submit those pages together as one BIO. So if our 736 * current logical offset in the file does not equal what would 737 * be the next logical offset in the bio, submit the bio we 738 * have. 739 */ 740 if (sdio->final_block_in_bio != sdio->cur_page_block || 741 cur_offset != bio_next_offset) 742 dio_bio_submit(dio, sdio); 743 } 744 745 if (sdio->bio == NULL) { 746 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh); 747 if (ret) 748 goto out; 749 } 750 751 if (dio_bio_add_page(dio, sdio) != 0) { 752 dio_bio_submit(dio, sdio); 753 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh); 754 if (ret == 0) { 755 ret = dio_bio_add_page(dio, sdio); 756 BUG_ON(ret != 0); 757 } 758 } 759 out: 760 return ret; 761 } 762 763 /* 764 * An autonomous function to put a chunk of a page under deferred IO. 765 * 766 * The caller doesn't actually know (or care) whether this piece of page is in 767 * a BIO, or is under IO or whatever. We just take care of all possible 768 * situations here. The separation between the logic of do_direct_IO() and 769 * that of submit_page_section() is important for clarity. Please don't break. 770 * 771 * The chunk of page starts on-disk at blocknr. 772 * 773 * We perform deferred IO, by recording the last-submitted page inside our 774 * private part of the dio structure. If possible, we just expand the IO 775 * across that page here. 776 * 777 * If that doesn't work out then we put the old page into the bio and add this 778 * page to the dio instead. 779 */ 780 static inline int 781 submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page, 782 unsigned offset, unsigned len, sector_t blocknr, 783 struct buffer_head *map_bh) 784 { 785 const enum req_op dio_op = dio->opf & REQ_OP_MASK; 786 int ret = 0; 787 int boundary = sdio->boundary; /* dio_send_cur_page may clear it */ 788 789 if (dio_op == REQ_OP_WRITE) { 790 /* 791 * Read accounting is performed in submit_bio() 792 */ 793 task_io_account_write(len); 794 } 795 796 /* 797 * Can we just grow the current page's presence in the dio? 798 */ 799 if (sdio->cur_page == page && 800 sdio->cur_page_offset + sdio->cur_page_len == offset && 801 sdio->cur_page_block + 802 (sdio->cur_page_len >> sdio->blkbits) == blocknr) { 803 sdio->cur_page_len += len; 804 goto out; 805 } 806 807 /* 808 * If there's a deferred page already there then send it. 809 */ 810 if (sdio->cur_page) { 811 ret = dio_send_cur_page(dio, sdio, map_bh); 812 dio_unpin_page(dio, sdio->cur_page); 813 sdio->cur_page = NULL; 814 if (ret) 815 return ret; 816 } 817 818 dio_pin_page(dio, page); /* It is in dio */ 819 sdio->cur_page = page; 820 sdio->cur_page_offset = offset; 821 sdio->cur_page_len = len; 822 sdio->cur_page_block = blocknr; 823 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits; 824 out: 825 /* 826 * If boundary then we want to schedule the IO now to 827 * avoid metadata seeks. 828 */ 829 if (boundary) { 830 ret = dio_send_cur_page(dio, sdio, map_bh); 831 if (sdio->bio) 832 dio_bio_submit(dio, sdio); 833 dio_unpin_page(dio, sdio->cur_page); 834 sdio->cur_page = NULL; 835 } 836 return ret; 837 } 838 839 /* 840 * If we are not writing the entire block and get_block() allocated 841 * the block for us, we need to fill-in the unused portion of the 842 * block with zeros. This happens only if user-buffer, fileoffset or 843 * io length is not filesystem block-size multiple. 844 * 845 * `end' is zero if we're doing the start of the IO, 1 at the end of the 846 * IO. 847 */ 848 static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio, 849 int end, struct buffer_head *map_bh) 850 { 851 unsigned dio_blocks_per_fs_block; 852 unsigned this_chunk_blocks; /* In dio_blocks */ 853 unsigned this_chunk_bytes; 854 struct page *page; 855 856 sdio->start_zero_done = 1; 857 if (!sdio->blkfactor || !buffer_new(map_bh)) 858 return; 859 860 dio_blocks_per_fs_block = 1 << sdio->blkfactor; 861 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1); 862 863 if (!this_chunk_blocks) 864 return; 865 866 /* 867 * We need to zero out part of an fs block. It is either at the 868 * beginning or the end of the fs block. 869 */ 870 if (end) 871 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks; 872 873 this_chunk_bytes = this_chunk_blocks << sdio->blkbits; 874 875 page = ZERO_PAGE(0); 876 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes, 877 sdio->next_block_for_io, map_bh)) 878 return; 879 880 sdio->next_block_for_io += this_chunk_blocks; 881 } 882 883 /* 884 * Walk the user pages, and the file, mapping blocks to disk and generating 885 * a sequence of (page,offset,len,block) mappings. These mappings are injected 886 * into submit_page_section(), which takes care of the next stage of submission 887 * 888 * Direct IO against a blockdev is different from a file. Because we can 889 * happily perform page-sized but 512-byte aligned IOs. It is important that 890 * blockdev IO be able to have fine alignment and large sizes. 891 * 892 * So what we do is to permit the ->get_block function to populate bh.b_size 893 * with the size of IO which is permitted at this offset and this i_blkbits. 894 * 895 * For best results, the blockdev should be set up with 512-byte i_blkbits and 896 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives 897 * fine alignment but still allows this function to work in PAGE_SIZE units. 898 */ 899 static int do_direct_IO(struct dio *dio, struct dio_submit *sdio, 900 struct buffer_head *map_bh) 901 { 902 const enum req_op dio_op = dio->opf & REQ_OP_MASK; 903 const unsigned blkbits = sdio->blkbits; 904 const unsigned i_blkbits = blkbits + sdio->blkfactor; 905 int ret = 0; 906 907 while (sdio->block_in_file < sdio->final_block_in_request) { 908 struct page *page; 909 size_t from, to; 910 911 page = dio_get_page(dio, sdio); 912 if (IS_ERR(page)) { 913 ret = PTR_ERR(page); 914 goto out; 915 } 916 from = sdio->head ? 0 : sdio->from; 917 to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE; 918 sdio->head++; 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 dio_unpin_page(dio, 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 >> blkbits; 942 sdio->next_block_for_io = 943 map_bh->b_blocknr << sdio->blkfactor; 944 if (buffer_new(map_bh)) { 945 clean_bdev_aliases( 946 map_bh->b_bdev, 947 map_bh->b_blocknr, 948 map_bh->b_size >> i_blkbits); 949 } 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_op == REQ_OP_WRITE) { 979 dio_unpin_page(dio, 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 dio_unpin_page(dio, page); 993 goto out; 994 } 995 memzero_page(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 dio_unpin_page(dio, 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 pin which was taken in get_user_pages() */ 1048 dio_unpin_page(dio, 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_rwsem and returns with 1083 * i_rwsem held, for reads, i_rwsem 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_rwsem, 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 ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode, 1102 struct block_device *bdev, struct iov_iter *iter, 1103 get_block_t get_block, dio_iodone_t end_io, 1104 int flags) 1105 { 1106 unsigned i_blkbits = READ_ONCE(inode->i_blkbits); 1107 unsigned blkbits = i_blkbits; 1108 unsigned blocksize_mask = (1 << blkbits) - 1; 1109 ssize_t retval = -EINVAL; 1110 const size_t count = iov_iter_count(iter); 1111 loff_t offset = iocb->ki_pos; 1112 const loff_t end = offset + count; 1113 struct dio *dio; 1114 struct dio_submit sdio = { NULL, }; 1115 struct buffer_head map_bh = { 0, }; 1116 struct blk_plug plug; 1117 unsigned long align = offset | iov_iter_alignment(iter); 1118 1119 /* watch out for a 0 len io from a tricksy fs */ 1120 if (iov_iter_rw(iter) == READ && !count) 1121 return 0; 1122 1123 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL); 1124 if (!dio) 1125 return -ENOMEM; 1126 /* 1127 * Believe it or not, zeroing out the page array caused a .5% 1128 * performance regression in a database benchmark. So, we take 1129 * care to only zero out what's needed. 1130 */ 1131 memset(dio, 0, offsetof(struct dio, pages)); 1132 1133 dio->flags = flags; 1134 if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ) { 1135 /* will be released by direct_io_worker */ 1136 inode_lock(inode); 1137 } 1138 dio->is_pinned = iov_iter_extract_will_pin(iter); 1139 1140 /* Once we sampled i_size check for reads beyond EOF */ 1141 dio->i_size = i_size_read(inode); 1142 if (iov_iter_rw(iter) == READ && offset >= dio->i_size) { 1143 retval = 0; 1144 goto fail_dio; 1145 } 1146 1147 if (align & blocksize_mask) { 1148 if (bdev) 1149 blkbits = blksize_bits(bdev_logical_block_size(bdev)); 1150 blocksize_mask = (1 << blkbits) - 1; 1151 if (align & blocksize_mask) 1152 goto fail_dio; 1153 } 1154 1155 if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ) { 1156 struct address_space *mapping = iocb->ki_filp->f_mapping; 1157 1158 retval = filemap_write_and_wait_range(mapping, offset, end - 1); 1159 if (retval) 1160 goto fail_dio; 1161 } 1162 1163 /* 1164 * For file extending writes updating i_size before data writeouts 1165 * complete can expose uninitialized blocks in dumb filesystems. 1166 * In that case we need to wait for I/O completion even if asked 1167 * for an asynchronous write. 1168 */ 1169 if (is_sync_kiocb(iocb)) 1170 dio->is_async = false; 1171 else if (iov_iter_rw(iter) == WRITE && end > i_size_read(inode)) 1172 dio->is_async = false; 1173 else 1174 dio->is_async = true; 1175 1176 dio->inode = inode; 1177 if (iov_iter_rw(iter) == WRITE) 1178 dio->opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE; 1179 else 1180 dio->opf = REQ_OP_READ; 1181 1182 /* 1183 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue 1184 * so that we can call ->fsync. 1185 */ 1186 if (dio->is_async && iov_iter_rw(iter) == WRITE) { 1187 retval = 0; 1188 if (iocb_is_dsync(iocb)) 1189 retval = dio_set_defer_completion(dio); 1190 else if (!dio->inode->i_sb->s_dio_done_wq) { 1191 /* 1192 * In case of AIO write racing with buffered read we 1193 * need to defer completion. We can't decide this now, 1194 * however the workqueue needs to be initialized here. 1195 */ 1196 retval = sb_init_dio_done_wq(dio->inode->i_sb); 1197 } 1198 if (retval) 1199 goto fail_dio; 1200 } 1201 1202 /* 1203 * Will be decremented at I/O completion time. 1204 */ 1205 inode_dio_begin(inode); 1206 1207 sdio.blkbits = blkbits; 1208 sdio.blkfactor = i_blkbits - blkbits; 1209 sdio.block_in_file = offset >> blkbits; 1210 1211 sdio.get_block = get_block; 1212 dio->end_io = end_io; 1213 sdio.final_block_in_bio = -1; 1214 sdio.next_block_for_io = -1; 1215 1216 dio->iocb = iocb; 1217 1218 spin_lock_init(&dio->bio_lock); 1219 dio->refcount = 1; 1220 1221 dio->should_dirty = user_backed_iter(iter) && iov_iter_rw(iter) == READ; 1222 sdio.iter = iter; 1223 sdio.final_block_in_request = end >> 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 * 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 dio_unpin_page(dio, 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_rwsem go now that its achieved its purpose 1276 * of protecting us from looking up uninitialized blocks. 1277 */ 1278 if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING)) 1279 inode_unlock(dio->inode); 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 (iov_iter_rw(iter) == READ || dio->result == count)) 1291 retval = -EIOCBQUEUED; 1292 else 1293 dio_await_completion(dio); 1294 1295 if (drop_refcount(dio) == 0) { 1296 retval = dio_complete(dio, retval, DIO_COMPLETE_INVALIDATE); 1297 } else 1298 BUG_ON(retval != -EIOCBQUEUED); 1299 1300 return retval; 1301 1302 fail_dio: 1303 if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ) 1304 inode_unlock(inode); 1305 1306 kmem_cache_free(dio_cache, dio); 1307 return retval; 1308 } 1309 EXPORT_SYMBOL(__blockdev_direct_IO); 1310 1311 static __init int dio_init(void) 1312 { 1313 dio_cache = KMEM_CACHE(dio, SLAB_PANIC); 1314 return 0; 1315 } 1316 module_init(dio_init) 1317