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