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