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 in the slab cache 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 /* dio_state only used in the submission path */ 59 60 struct dio_submit { 61 struct bio *bio; /* bio under assembly */ 62 unsigned blkbits; /* doesn't change */ 63 unsigned blkfactor; /* When we're using an alignment which 64 is finer than the filesystem's soft 65 blocksize, this specifies how much 66 finer. blkfactor=2 means 1/4-block 67 alignment. Does not change */ 68 unsigned start_zero_done; /* flag: sub-blocksize zeroing has 69 been performed at the start of a 70 write */ 71 int pages_in_io; /* approximate total IO pages */ 72 size_t size; /* total request size (doesn't change)*/ 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 unsigned first_block_in_page; /* doesn't change, Used only once */ 79 int boundary; /* prev block is at a boundary */ 80 get_block_t *get_block; /* block mapping function */ 81 dio_submit_t *submit_io; /* IO submition function */ 82 83 loff_t logical_offset_in_bio; /* current first logical block in bio */ 84 sector_t final_block_in_bio; /* current final block in bio + 1 */ 85 sector_t next_block_for_io; /* next block to be put under IO, 86 in dio_blocks units */ 87 88 /* 89 * Deferred addition of a page to the dio. These variables are 90 * private to dio_send_cur_page(), submit_page_section() and 91 * dio_bio_add_page(). 92 */ 93 struct page *cur_page; /* The page */ 94 unsigned cur_page_offset; /* Offset into it, in bytes */ 95 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */ 96 sector_t cur_page_block; /* Where it starts */ 97 loff_t cur_page_fs_offset; /* Offset in file */ 98 99 /* 100 * Page fetching state. These variables belong to dio_refill_pages(). 101 */ 102 int curr_page; /* changes */ 103 int total_pages; /* doesn't change */ 104 unsigned long curr_user_address;/* changes */ 105 106 /* 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 }; 113 114 /* dio_state communicated between submission path and end_io */ 115 struct dio { 116 int flags; /* doesn't change */ 117 int rw; 118 struct inode *inode; 119 loff_t i_size; /* i_size when submitted */ 120 dio_iodone_t *end_io; /* IO completion function */ 121 122 void *private; /* copy from map_bh.b_private */ 123 124 /* BIO completion state */ 125 spinlock_t bio_lock; /* protects BIO fields below */ 126 int page_errors; /* errno from get_user_pages() */ 127 int is_async; /* is IO async ? */ 128 int io_error; /* IO error in completion path */ 129 unsigned long refcount; /* direct_io_worker() and bios */ 130 struct bio *bio_list; /* singly linked via bi_private */ 131 struct task_struct *waiter; /* waiting task (NULL if none) */ 132 133 /* AIO related stuff */ 134 struct kiocb *iocb; /* kiocb */ 135 ssize_t result; /* IO result */ 136 137 /* 138 * pages[] (and any fields placed after it) are not zeroed out at 139 * allocation time. Don't add new fields after pages[] unless you 140 * wish that they not be zeroed. 141 */ 142 struct page *pages[DIO_PAGES]; /* page buffer */ 143 } ____cacheline_aligned_in_smp; 144 145 static struct kmem_cache *dio_cache __read_mostly; 146 147 static void __inode_dio_wait(struct inode *inode) 148 { 149 wait_queue_head_t *wq = bit_waitqueue(&inode->i_state, __I_DIO_WAKEUP); 150 DEFINE_WAIT_BIT(q, &inode->i_state, __I_DIO_WAKEUP); 151 152 do { 153 prepare_to_wait(wq, &q.wait, TASK_UNINTERRUPTIBLE); 154 if (atomic_read(&inode->i_dio_count)) 155 schedule(); 156 } while (atomic_read(&inode->i_dio_count)); 157 finish_wait(wq, &q.wait); 158 } 159 160 /** 161 * inode_dio_wait - wait for outstanding DIO requests to finish 162 * @inode: inode to wait for 163 * 164 * Waits for all pending direct I/O requests to finish so that we can 165 * proceed with a truncate or equivalent operation. 166 * 167 * Must be called under a lock that serializes taking new references 168 * to i_dio_count, usually by inode->i_mutex. 169 */ 170 void inode_dio_wait(struct inode *inode) 171 { 172 if (atomic_read(&inode->i_dio_count)) 173 __inode_dio_wait(inode); 174 } 175 EXPORT_SYMBOL_GPL(inode_dio_wait); 176 177 /* 178 * inode_dio_done - signal finish of a direct I/O requests 179 * @inode: inode the direct I/O happens on 180 * 181 * This is called once we've finished processing a direct I/O request, 182 * and is used to wake up callers waiting for direct I/O to be quiesced. 183 */ 184 void inode_dio_done(struct inode *inode) 185 { 186 if (atomic_dec_and_test(&inode->i_dio_count)) 187 wake_up_bit(&inode->i_state, __I_DIO_WAKEUP); 188 } 189 EXPORT_SYMBOL_GPL(inode_dio_done); 190 191 /* 192 * How many pages are in the queue? 193 */ 194 static inline unsigned dio_pages_present(struct dio_submit *sdio) 195 { 196 return sdio->tail - sdio->head; 197 } 198 199 /* 200 * Go grab and pin some userspace pages. Typically we'll get 64 at a time. 201 */ 202 static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio) 203 { 204 int ret; 205 int nr_pages; 206 207 nr_pages = min(sdio->total_pages - sdio->curr_page, DIO_PAGES); 208 ret = get_user_pages_fast( 209 sdio->curr_user_address, /* Where from? */ 210 nr_pages, /* How many pages? */ 211 dio->rw == READ, /* Write to memory? */ 212 &dio->pages[0]); /* Put results here */ 213 214 if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) { 215 struct page *page = ZERO_PAGE(0); 216 /* 217 * A memory fault, but the filesystem has some outstanding 218 * mapped blocks. We need to use those blocks up to avoid 219 * leaking stale data in the file. 220 */ 221 if (dio->page_errors == 0) 222 dio->page_errors = ret; 223 page_cache_get(page); 224 dio->pages[0] = page; 225 sdio->head = 0; 226 sdio->tail = 1; 227 ret = 0; 228 goto out; 229 } 230 231 if (ret >= 0) { 232 sdio->curr_user_address += ret * PAGE_SIZE; 233 sdio->curr_page += ret; 234 sdio->head = 0; 235 sdio->tail = ret; 236 ret = 0; 237 } 238 out: 239 return ret; 240 } 241 242 /* 243 * Get another userspace page. Returns an ERR_PTR on error. Pages are 244 * buffered inside the dio so that we can call get_user_pages() against a 245 * decent number of pages, less frequently. To provide nicer use of the 246 * L1 cache. 247 */ 248 static inline struct page *dio_get_page(struct dio *dio, 249 struct dio_submit *sdio) 250 { 251 if (dio_pages_present(sdio) == 0) { 252 int ret; 253 254 ret = dio_refill_pages(dio, sdio); 255 if (ret) 256 return ERR_PTR(ret); 257 BUG_ON(dio_pages_present(sdio) == 0); 258 } 259 return dio->pages[sdio->head++]; 260 } 261 262 /** 263 * dio_complete() - called when all DIO BIO I/O has been completed 264 * @offset: the byte offset in the file of the completed operation 265 * 266 * This releases locks as dictated by the locking type, lets interested parties 267 * know that a DIO operation has completed, and calculates the resulting return 268 * code for the operation. 269 * 270 * It lets the filesystem know if it registered an interest earlier via 271 * get_block. Pass the private field of the map buffer_head so that 272 * filesystems can use it to hold additional state between get_block calls and 273 * dio_complete. 274 */ 275 static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret, bool is_async) 276 { 277 ssize_t transferred = 0; 278 279 /* 280 * AIO submission can race with bio completion to get here while 281 * expecting to have the last io completed by bio completion. 282 * In that case -EIOCBQUEUED is in fact not an error we want 283 * to preserve through this call. 284 */ 285 if (ret == -EIOCBQUEUED) 286 ret = 0; 287 288 if (dio->result) { 289 transferred = dio->result; 290 291 /* Check for short read case */ 292 if ((dio->rw == READ) && ((offset + transferred) > dio->i_size)) 293 transferred = dio->i_size - offset; 294 } 295 296 if (ret == 0) 297 ret = dio->page_errors; 298 if (ret == 0) 299 ret = dio->io_error; 300 if (ret == 0) 301 ret = transferred; 302 303 if (dio->end_io && dio->result) { 304 dio->end_io(dio->iocb, offset, transferred, 305 dio->private, ret, is_async); 306 } else { 307 if (is_async) 308 aio_complete(dio->iocb, ret, 0); 309 inode_dio_done(dio->inode); 310 } 311 312 return ret; 313 } 314 315 static int dio_bio_complete(struct dio *dio, struct bio *bio); 316 /* 317 * Asynchronous IO callback. 318 */ 319 static void dio_bio_end_aio(struct bio *bio, int error) 320 { 321 struct dio *dio = bio->bi_private; 322 unsigned long remaining; 323 unsigned long flags; 324 325 /* cleanup the bio */ 326 dio_bio_complete(dio, bio); 327 328 spin_lock_irqsave(&dio->bio_lock, flags); 329 remaining = --dio->refcount; 330 if (remaining == 1 && dio->waiter) 331 wake_up_process(dio->waiter); 332 spin_unlock_irqrestore(&dio->bio_lock, flags); 333 334 if (remaining == 0) { 335 dio_complete(dio, dio->iocb->ki_pos, 0, true); 336 kmem_cache_free(dio_cache, dio); 337 } 338 } 339 340 /* 341 * The BIO completion handler simply queues the BIO up for the process-context 342 * handler. 343 * 344 * During I/O bi_private points at the dio. After I/O, bi_private is used to 345 * implement a singly-linked list of completed BIOs, at dio->bio_list. 346 */ 347 static void dio_bio_end_io(struct bio *bio, int error) 348 { 349 struct dio *dio = bio->bi_private; 350 unsigned long flags; 351 352 spin_lock_irqsave(&dio->bio_lock, flags); 353 bio->bi_private = dio->bio_list; 354 dio->bio_list = bio; 355 if (--dio->refcount == 1 && dio->waiter) 356 wake_up_process(dio->waiter); 357 spin_unlock_irqrestore(&dio->bio_lock, flags); 358 } 359 360 /** 361 * dio_end_io - handle the end io action for the given bio 362 * @bio: The direct io bio thats being completed 363 * @error: Error if there was one 364 * 365 * This is meant to be called by any filesystem that uses their own dio_submit_t 366 * so that the DIO specific endio actions are dealt with after the filesystem 367 * has done it's completion work. 368 */ 369 void dio_end_io(struct bio *bio, int error) 370 { 371 struct dio *dio = bio->bi_private; 372 373 if (dio->is_async) 374 dio_bio_end_aio(bio, error); 375 else 376 dio_bio_end_io(bio, error); 377 } 378 EXPORT_SYMBOL_GPL(dio_end_io); 379 380 static inline void 381 dio_bio_alloc(struct dio *dio, struct dio_submit *sdio, 382 struct block_device *bdev, 383 sector_t first_sector, int nr_vecs) 384 { 385 struct bio *bio; 386 387 /* 388 * bio_alloc() is guaranteed to return a bio when called with 389 * __GFP_WAIT and we request a valid number of vectors. 390 */ 391 bio = bio_alloc(GFP_KERNEL, nr_vecs); 392 393 bio->bi_bdev = bdev; 394 bio->bi_sector = first_sector; 395 if (dio->is_async) 396 bio->bi_end_io = dio_bio_end_aio; 397 else 398 bio->bi_end_io = dio_bio_end_io; 399 400 sdio->bio = bio; 401 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset; 402 } 403 404 /* 405 * In the AIO read case we speculatively dirty the pages before starting IO. 406 * During IO completion, any of these pages which happen to have been written 407 * back will be redirtied by bio_check_pages_dirty(). 408 * 409 * bios hold a dio reference between submit_bio and ->end_io. 410 */ 411 static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio) 412 { 413 struct bio *bio = sdio->bio; 414 unsigned long flags; 415 416 bio->bi_private = dio; 417 418 spin_lock_irqsave(&dio->bio_lock, flags); 419 dio->refcount++; 420 spin_unlock_irqrestore(&dio->bio_lock, flags); 421 422 if (dio->is_async && dio->rw == READ) 423 bio_set_pages_dirty(bio); 424 425 if (sdio->submit_io) 426 sdio->submit_io(dio->rw, bio, dio->inode, 427 sdio->logical_offset_in_bio); 428 else 429 submit_bio(dio->rw, bio); 430 431 sdio->bio = NULL; 432 sdio->boundary = 0; 433 sdio->logical_offset_in_bio = 0; 434 } 435 436 /* 437 * Release any resources in case of a failure 438 */ 439 static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio) 440 { 441 while (dio_pages_present(sdio)) 442 page_cache_release(dio_get_page(dio, sdio)); 443 } 444 445 /* 446 * Wait for the next BIO to complete. Remove it and return it. NULL is 447 * returned once all BIOs have been completed. This must only be called once 448 * all bios have been issued so that dio->refcount can only decrease. This 449 * requires that that the caller hold a reference on the dio. 450 */ 451 static struct bio *dio_await_one(struct dio *dio) 452 { 453 unsigned long flags; 454 struct bio *bio = NULL; 455 456 spin_lock_irqsave(&dio->bio_lock, flags); 457 458 /* 459 * Wait as long as the list is empty and there are bios in flight. bio 460 * completion drops the count, maybe adds to the list, and wakes while 461 * holding the bio_lock so we don't need set_current_state()'s barrier 462 * and can call it after testing our condition. 463 */ 464 while (dio->refcount > 1 && dio->bio_list == NULL) { 465 __set_current_state(TASK_UNINTERRUPTIBLE); 466 dio->waiter = current; 467 spin_unlock_irqrestore(&dio->bio_lock, flags); 468 io_schedule(); 469 /* wake up sets us TASK_RUNNING */ 470 spin_lock_irqsave(&dio->bio_lock, flags); 471 dio->waiter = NULL; 472 } 473 if (dio->bio_list) { 474 bio = dio->bio_list; 475 dio->bio_list = bio->bi_private; 476 } 477 spin_unlock_irqrestore(&dio->bio_lock, flags); 478 return bio; 479 } 480 481 /* 482 * Process one completed BIO. No locks are held. 483 */ 484 static int dio_bio_complete(struct dio *dio, struct bio *bio) 485 { 486 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); 487 struct bio_vec *bvec = bio->bi_io_vec; 488 int page_no; 489 490 if (!uptodate) 491 dio->io_error = -EIO; 492 493 if (dio->is_async && dio->rw == READ) { 494 bio_check_pages_dirty(bio); /* transfers ownership */ 495 } else { 496 for (page_no = 0; page_no < bio->bi_vcnt; page_no++) { 497 struct page *page = bvec[page_no].bv_page; 498 499 if (dio->rw == READ && !PageCompound(page)) 500 set_page_dirty_lock(page); 501 page_cache_release(page); 502 } 503 bio_put(bio); 504 } 505 return uptodate ? 0 : -EIO; 506 } 507 508 /* 509 * Wait on and process all in-flight BIOs. This must only be called once 510 * all bios have been issued so that the refcount can only decrease. 511 * This just waits for all bios to make it through dio_bio_complete. IO 512 * errors are propagated through dio->io_error and should be propagated via 513 * dio_complete(). 514 */ 515 static void dio_await_completion(struct dio *dio) 516 { 517 struct bio *bio; 518 do { 519 bio = dio_await_one(dio); 520 if (bio) 521 dio_bio_complete(dio, bio); 522 } while (bio); 523 } 524 525 /* 526 * A really large O_DIRECT read or write can generate a lot of BIOs. So 527 * to keep the memory consumption sane we periodically reap any completed BIOs 528 * during the BIO generation phase. 529 * 530 * This also helps to limit the peak amount of pinned userspace memory. 531 */ 532 static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio) 533 { 534 int ret = 0; 535 536 if (sdio->reap_counter++ >= 64) { 537 while (dio->bio_list) { 538 unsigned long flags; 539 struct bio *bio; 540 int ret2; 541 542 spin_lock_irqsave(&dio->bio_lock, flags); 543 bio = dio->bio_list; 544 dio->bio_list = bio->bi_private; 545 spin_unlock_irqrestore(&dio->bio_lock, flags); 546 ret2 = dio_bio_complete(dio, bio); 547 if (ret == 0) 548 ret = ret2; 549 } 550 sdio->reap_counter = 0; 551 } 552 return ret; 553 } 554 555 /* 556 * Call into the fs to map some more disk blocks. We record the current number 557 * of available blocks at sdio->blocks_available. These are in units of the 558 * fs blocksize, (1 << inode->i_blkbits). 559 * 560 * The fs is allowed to map lots of blocks at once. If it wants to do that, 561 * it uses the passed inode-relative block number as the file offset, as usual. 562 * 563 * get_block() is passed the number of i_blkbits-sized blocks which direct_io 564 * has remaining to do. The fs should not map more than this number of blocks. 565 * 566 * If the fs has mapped a lot of blocks, it should populate bh->b_size to 567 * indicate how much contiguous disk space has been made available at 568 * bh->b_blocknr. 569 * 570 * If *any* of the mapped blocks are new, then the fs must set buffer_new(). 571 * This isn't very efficient... 572 * 573 * In the case of filesystem holes: the fs may return an arbitrarily-large 574 * hole by returning an appropriate value in b_size and by clearing 575 * buffer_mapped(). However the direct-io code will only process holes one 576 * block at a time - it will repeatedly call get_block() as it walks the hole. 577 */ 578 static int get_more_blocks(struct dio *dio, struct dio_submit *sdio, 579 struct buffer_head *map_bh) 580 { 581 int ret; 582 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */ 583 unsigned long fs_count; /* Number of filesystem-sized blocks */ 584 unsigned long dio_count;/* Number of dio_block-sized blocks */ 585 unsigned long blkmask; 586 int create; 587 588 /* 589 * If there was a memory error and we've overwritten all the 590 * mapped blocks then we can now return that memory error 591 */ 592 ret = dio->page_errors; 593 if (ret == 0) { 594 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request); 595 fs_startblk = sdio->block_in_file >> sdio->blkfactor; 596 dio_count = sdio->final_block_in_request - sdio->block_in_file; 597 fs_count = dio_count >> sdio->blkfactor; 598 blkmask = (1 << sdio->blkfactor) - 1; 599 if (dio_count & blkmask) 600 fs_count++; 601 602 map_bh->b_state = 0; 603 map_bh->b_size = fs_count << dio->inode->i_blkbits; 604 605 /* 606 * For writes inside i_size on a DIO_SKIP_HOLES filesystem we 607 * forbid block creations: only overwrites are permitted. 608 * We will return early to the caller once we see an 609 * unmapped buffer head returned, and the caller will fall 610 * back to buffered I/O. 611 * 612 * Otherwise the decision is left to the get_blocks method, 613 * which may decide to handle it or also return an unmapped 614 * buffer head. 615 */ 616 create = dio->rw & WRITE; 617 if (dio->flags & DIO_SKIP_HOLES) { 618 if (sdio->block_in_file < (i_size_read(dio->inode) >> 619 sdio->blkbits)) 620 create = 0; 621 } 622 623 ret = (*sdio->get_block)(dio->inode, fs_startblk, 624 map_bh, create); 625 626 /* Store for completion */ 627 dio->private = map_bh->b_private; 628 } 629 return ret; 630 } 631 632 /* 633 * There is no bio. Make one now. 634 */ 635 static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio, 636 sector_t start_sector, struct buffer_head *map_bh) 637 { 638 sector_t sector; 639 int ret, nr_pages; 640 641 ret = dio_bio_reap(dio, sdio); 642 if (ret) 643 goto out; 644 sector = start_sector << (sdio->blkbits - 9); 645 nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(map_bh->b_bdev)); 646 nr_pages = min(nr_pages, BIO_MAX_PAGES); 647 BUG_ON(nr_pages <= 0); 648 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages); 649 sdio->boundary = 0; 650 out: 651 return ret; 652 } 653 654 /* 655 * Attempt to put the current chunk of 'cur_page' into the current BIO. If 656 * that was successful then update final_block_in_bio and take a ref against 657 * the just-added page. 658 * 659 * Return zero on success. Non-zero means the caller needs to start a new BIO. 660 */ 661 static inline int dio_bio_add_page(struct dio_submit *sdio) 662 { 663 int ret; 664 665 ret = bio_add_page(sdio->bio, sdio->cur_page, 666 sdio->cur_page_len, sdio->cur_page_offset); 667 if (ret == sdio->cur_page_len) { 668 /* 669 * Decrement count only, if we are done with this page 670 */ 671 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE) 672 sdio->pages_in_io--; 673 page_cache_get(sdio->cur_page); 674 sdio->final_block_in_bio = sdio->cur_page_block + 675 (sdio->cur_page_len >> sdio->blkbits); 676 ret = 0; 677 } else { 678 ret = 1; 679 } 680 return ret; 681 } 682 683 /* 684 * Put cur_page under IO. The section of cur_page which is described by 685 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page 686 * starts on-disk at cur_page_block. 687 * 688 * We take a ref against the page here (on behalf of its presence in the bio). 689 * 690 * The caller of this function is responsible for removing cur_page from the 691 * dio, and for dropping the refcount which came from that presence. 692 */ 693 static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio, 694 struct buffer_head *map_bh) 695 { 696 int ret = 0; 697 698 if (sdio->bio) { 699 loff_t cur_offset = sdio->cur_page_fs_offset; 700 loff_t bio_next_offset = sdio->logical_offset_in_bio + 701 sdio->bio->bi_size; 702 703 /* 704 * See whether this new request is contiguous with the old. 705 * 706 * Btrfs cannot handle having logically non-contiguous requests 707 * submitted. For example if you have 708 * 709 * Logical: [0-4095][HOLE][8192-12287] 710 * Physical: [0-4095] [4096-8191] 711 * 712 * We cannot submit those pages together as one BIO. So if our 713 * current logical offset in the file does not equal what would 714 * be the next logical offset in the bio, submit the bio we 715 * have. 716 */ 717 if (sdio->final_block_in_bio != sdio->cur_page_block || 718 cur_offset != bio_next_offset) 719 dio_bio_submit(dio, sdio); 720 /* 721 * Submit now if the underlying fs is about to perform a 722 * metadata read 723 */ 724 else if (sdio->boundary) 725 dio_bio_submit(dio, sdio); 726 } 727 728 if (sdio->bio == NULL) { 729 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh); 730 if (ret) 731 goto out; 732 } 733 734 if (dio_bio_add_page(sdio) != 0) { 735 dio_bio_submit(dio, sdio); 736 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh); 737 if (ret == 0) { 738 ret = dio_bio_add_page(sdio); 739 BUG_ON(ret != 0); 740 } 741 } 742 out: 743 return ret; 744 } 745 746 /* 747 * An autonomous function to put a chunk of a page under deferred IO. 748 * 749 * The caller doesn't actually know (or care) whether this piece of page is in 750 * a BIO, or is under IO or whatever. We just take care of all possible 751 * situations here. The separation between the logic of do_direct_IO() and 752 * that of submit_page_section() is important for clarity. Please don't break. 753 * 754 * The chunk of page starts on-disk at blocknr. 755 * 756 * We perform deferred IO, by recording the last-submitted page inside our 757 * private part of the dio structure. If possible, we just expand the IO 758 * across that page here. 759 * 760 * If that doesn't work out then we put the old page into the bio and add this 761 * page to the dio instead. 762 */ 763 static inline int 764 submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page, 765 unsigned offset, unsigned len, sector_t blocknr, 766 struct buffer_head *map_bh) 767 { 768 int ret = 0; 769 770 if (dio->rw & WRITE) { 771 /* 772 * Read accounting is performed in submit_bio() 773 */ 774 task_io_account_write(len); 775 } 776 777 /* 778 * Can we just grow the current page's presence in the dio? 779 */ 780 if (sdio->cur_page == page && 781 sdio->cur_page_offset + sdio->cur_page_len == offset && 782 sdio->cur_page_block + 783 (sdio->cur_page_len >> sdio->blkbits) == blocknr) { 784 sdio->cur_page_len += len; 785 786 /* 787 * If sdio->boundary then we want to schedule the IO now to 788 * avoid metadata seeks. 789 */ 790 if (sdio->boundary) { 791 ret = dio_send_cur_page(dio, sdio, map_bh); 792 page_cache_release(sdio->cur_page); 793 sdio->cur_page = NULL; 794 } 795 goto out; 796 } 797 798 /* 799 * If there's a deferred page already there then send it. 800 */ 801 if (sdio->cur_page) { 802 ret = dio_send_cur_page(dio, sdio, map_bh); 803 page_cache_release(sdio->cur_page); 804 sdio->cur_page = NULL; 805 if (ret) 806 goto out; 807 } 808 809 page_cache_get(page); /* It is in dio */ 810 sdio->cur_page = page; 811 sdio->cur_page_offset = offset; 812 sdio->cur_page_len = len; 813 sdio->cur_page_block = blocknr; 814 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits; 815 out: 816 return ret; 817 } 818 819 /* 820 * Clean any dirty buffers in the blockdev mapping which alias newly-created 821 * file blocks. Only called for S_ISREG files - blockdevs do not set 822 * buffer_new 823 */ 824 static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh) 825 { 826 unsigned i; 827 unsigned nblocks; 828 829 nblocks = map_bh->b_size >> dio->inode->i_blkbits; 830 831 for (i = 0; i < nblocks; i++) { 832 unmap_underlying_metadata(map_bh->b_bdev, 833 map_bh->b_blocknr + i); 834 } 835 } 836 837 /* 838 * If we are not writing the entire block and get_block() allocated 839 * the block for us, we need to fill-in the unused portion of the 840 * block with zeros. This happens only if user-buffer, fileoffset or 841 * io length is not filesystem block-size multiple. 842 * 843 * `end' is zero if we're doing the start of the IO, 1 at the end of the 844 * IO. 845 */ 846 static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio, 847 int end, struct buffer_head *map_bh) 848 { 849 unsigned dio_blocks_per_fs_block; 850 unsigned this_chunk_blocks; /* In dio_blocks */ 851 unsigned this_chunk_bytes; 852 struct page *page; 853 854 sdio->start_zero_done = 1; 855 if (!sdio->blkfactor || !buffer_new(map_bh)) 856 return; 857 858 dio_blocks_per_fs_block = 1 << sdio->blkfactor; 859 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1); 860 861 if (!this_chunk_blocks) 862 return; 863 864 /* 865 * We need to zero out part of an fs block. It is either at the 866 * beginning or the end of the fs block. 867 */ 868 if (end) 869 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks; 870 871 this_chunk_bytes = this_chunk_blocks << sdio->blkbits; 872 873 page = ZERO_PAGE(0); 874 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes, 875 sdio->next_block_for_io, map_bh)) 876 return; 877 878 sdio->next_block_for_io += this_chunk_blocks; 879 } 880 881 /* 882 * Walk the user pages, and the file, mapping blocks to disk and generating 883 * a sequence of (page,offset,len,block) mappings. These mappings are injected 884 * into submit_page_section(), which takes care of the next stage of submission 885 * 886 * Direct IO against a blockdev is different from a file. Because we can 887 * happily perform page-sized but 512-byte aligned IOs. It is important that 888 * blockdev IO be able to have fine alignment and large sizes. 889 * 890 * So what we do is to permit the ->get_block function to populate bh.b_size 891 * with the size of IO which is permitted at this offset and this i_blkbits. 892 * 893 * For best results, the blockdev should be set up with 512-byte i_blkbits and 894 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives 895 * fine alignment but still allows this function to work in PAGE_SIZE units. 896 */ 897 static int do_direct_IO(struct dio *dio, struct dio_submit *sdio, 898 struct buffer_head *map_bh) 899 { 900 const unsigned blkbits = sdio->blkbits; 901 const unsigned blocks_per_page = PAGE_SIZE >> blkbits; 902 struct page *page; 903 unsigned block_in_page; 904 int ret = 0; 905 906 /* The I/O can start at any block offset within the first page */ 907 block_in_page = sdio->first_block_in_page; 908 909 while (sdio->block_in_file < sdio->final_block_in_request) { 910 page = dio_get_page(dio, sdio); 911 if (IS_ERR(page)) { 912 ret = PTR_ERR(page); 913 goto out; 914 } 915 916 while (block_in_page < blocks_per_page) { 917 unsigned offset_in_page = block_in_page << blkbits; 918 unsigned this_chunk_bytes; /* # of bytes mapped */ 919 unsigned this_chunk_blocks; /* # of blocks */ 920 unsigned u; 921 922 if (sdio->blocks_available == 0) { 923 /* 924 * Need to go and map some more disk 925 */ 926 unsigned long blkmask; 927 unsigned long dio_remainder; 928 929 ret = get_more_blocks(dio, sdio, map_bh); 930 if (ret) { 931 page_cache_release(page); 932 goto out; 933 } 934 if (!buffer_mapped(map_bh)) 935 goto do_holes; 936 937 sdio->blocks_available = 938 map_bh->b_size >> sdio->blkbits; 939 sdio->next_block_for_io = 940 map_bh->b_blocknr << sdio->blkfactor; 941 if (buffer_new(map_bh)) 942 clean_blockdev_aliases(dio, map_bh); 943 944 if (!sdio->blkfactor) 945 goto do_holes; 946 947 blkmask = (1 << sdio->blkfactor) - 1; 948 dio_remainder = (sdio->block_in_file & blkmask); 949 950 /* 951 * If we are at the start of IO and that IO 952 * starts partway into a fs-block, 953 * dio_remainder will be non-zero. If the IO 954 * is a read then we can simply advance the IO 955 * cursor to the first block which is to be 956 * read. But if the IO is a write and the 957 * block was newly allocated we cannot do that; 958 * the start of the fs block must be zeroed out 959 * on-disk 960 */ 961 if (!buffer_new(map_bh)) 962 sdio->next_block_for_io += dio_remainder; 963 sdio->blocks_available -= dio_remainder; 964 } 965 do_holes: 966 /* Handle holes */ 967 if (!buffer_mapped(map_bh)) { 968 loff_t i_size_aligned; 969 970 /* AKPM: eargh, -ENOTBLK is a hack */ 971 if (dio->rw & WRITE) { 972 page_cache_release(page); 973 return -ENOTBLK; 974 } 975 976 /* 977 * Be sure to account for a partial block as the 978 * last block in the file 979 */ 980 i_size_aligned = ALIGN(i_size_read(dio->inode), 981 1 << blkbits); 982 if (sdio->block_in_file >= 983 i_size_aligned >> blkbits) { 984 /* We hit eof */ 985 page_cache_release(page); 986 goto out; 987 } 988 zero_user(page, block_in_page << blkbits, 989 1 << blkbits); 990 sdio->block_in_file++; 991 block_in_page++; 992 goto next_block; 993 } 994 995 /* 996 * If we're performing IO which has an alignment which 997 * is finer than the underlying fs, go check to see if 998 * we must zero out the start of this block. 999 */ 1000 if (unlikely(sdio->blkfactor && !sdio->start_zero_done)) 1001 dio_zero_block(dio, sdio, 0, map_bh); 1002 1003 /* 1004 * Work out, in this_chunk_blocks, how much disk we 1005 * can add to this page 1006 */ 1007 this_chunk_blocks = sdio->blocks_available; 1008 u = (PAGE_SIZE - offset_in_page) >> blkbits; 1009 if (this_chunk_blocks > u) 1010 this_chunk_blocks = u; 1011 u = sdio->final_block_in_request - sdio->block_in_file; 1012 if (this_chunk_blocks > u) 1013 this_chunk_blocks = u; 1014 this_chunk_bytes = this_chunk_blocks << blkbits; 1015 BUG_ON(this_chunk_bytes == 0); 1016 1017 sdio->boundary = buffer_boundary(map_bh); 1018 ret = submit_page_section(dio, sdio, page, 1019 offset_in_page, 1020 this_chunk_bytes, 1021 sdio->next_block_for_io, 1022 map_bh); 1023 if (ret) { 1024 page_cache_release(page); 1025 goto out; 1026 } 1027 sdio->next_block_for_io += this_chunk_blocks; 1028 1029 sdio->block_in_file += this_chunk_blocks; 1030 block_in_page += this_chunk_blocks; 1031 sdio->blocks_available -= this_chunk_blocks; 1032 next_block: 1033 BUG_ON(sdio->block_in_file > sdio->final_block_in_request); 1034 if (sdio->block_in_file == sdio->final_block_in_request) 1035 break; 1036 } 1037 1038 /* Drop the ref which was taken in get_user_pages() */ 1039 page_cache_release(page); 1040 block_in_page = 0; 1041 } 1042 out: 1043 return ret; 1044 } 1045 1046 static inline int drop_refcount(struct dio *dio) 1047 { 1048 int ret2; 1049 unsigned long flags; 1050 1051 /* 1052 * Sync will always be dropping the final ref and completing the 1053 * operation. AIO can if it was a broken operation described above or 1054 * in fact if all the bios race to complete before we get here. In 1055 * that case dio_complete() translates the EIOCBQUEUED into the proper 1056 * return code that the caller will hand to aio_complete(). 1057 * 1058 * This is managed by the bio_lock instead of being an atomic_t so that 1059 * completion paths can drop their ref and use the remaining count to 1060 * decide to wake the submission path atomically. 1061 */ 1062 spin_lock_irqsave(&dio->bio_lock, flags); 1063 ret2 = --dio->refcount; 1064 spin_unlock_irqrestore(&dio->bio_lock, flags); 1065 return ret2; 1066 } 1067 1068 /* 1069 * This is a library function for use by filesystem drivers. 1070 * 1071 * The locking rules are governed by the flags parameter: 1072 * - if the flags value contains DIO_LOCKING we use a fancy locking 1073 * scheme for dumb filesystems. 1074 * For writes this function is called under i_mutex and returns with 1075 * i_mutex held, for reads, i_mutex is not held on entry, but it is 1076 * taken and dropped again before returning. 1077 * - if the flags value does NOT contain DIO_LOCKING we don't use any 1078 * internal locking but rather rely on the filesystem to synchronize 1079 * direct I/O reads/writes versus each other and truncate. 1080 * 1081 * To help with locking against truncate we incremented the i_dio_count 1082 * counter before starting direct I/O, and decrement it once we are done. 1083 * Truncate can wait for it to reach zero to provide exclusion. It is 1084 * expected that filesystem provide exclusion between new direct I/O 1085 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex, 1086 * but other filesystems need to take care of this on their own. 1087 * 1088 * NOTE: if you pass "sdio" to anything by pointer make sure that function 1089 * is always inlined. Otherwise gcc is unable to split the structure into 1090 * individual fields and will generate much worse code. This is important 1091 * for the whole file. 1092 */ 1093 ssize_t 1094 __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode, 1095 struct block_device *bdev, const struct iovec *iov, loff_t offset, 1096 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io, 1097 dio_submit_t submit_io, int flags) 1098 { 1099 int seg; 1100 size_t size; 1101 unsigned long addr; 1102 unsigned blkbits = inode->i_blkbits; 1103 unsigned bdev_blkbits = 0; 1104 unsigned blocksize_mask = (1 << blkbits) - 1; 1105 ssize_t retval = -EINVAL; 1106 loff_t end = offset; 1107 struct dio *dio; 1108 struct dio_submit sdio = { 0, }; 1109 unsigned long user_addr; 1110 size_t bytes; 1111 struct buffer_head map_bh = { 0, }; 1112 1113 if (rw & WRITE) 1114 rw = WRITE_ODIRECT; 1115 1116 if (bdev) 1117 bdev_blkbits = blksize_bits(bdev_logical_block_size(bdev)); 1118 1119 if (offset & blocksize_mask) { 1120 if (bdev) 1121 blkbits = bdev_blkbits; 1122 blocksize_mask = (1 << blkbits) - 1; 1123 if (offset & blocksize_mask) 1124 goto out; 1125 } 1126 1127 /* Check the memory alignment. Blocks cannot straddle pages */ 1128 for (seg = 0; seg < nr_segs; seg++) { 1129 addr = (unsigned long)iov[seg].iov_base; 1130 size = iov[seg].iov_len; 1131 end += size; 1132 if ((addr & blocksize_mask) || (size & blocksize_mask)) { 1133 if (bdev) 1134 blkbits = bdev_blkbits; 1135 blocksize_mask = (1 << blkbits) - 1; 1136 if ((addr & blocksize_mask) || (size & blocksize_mask)) 1137 goto out; 1138 } 1139 } 1140 1141 /* watch out for a 0 len io from a tricksy fs */ 1142 if (rw == READ && end == offset) 1143 return 0; 1144 1145 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL); 1146 retval = -ENOMEM; 1147 if (!dio) 1148 goto out; 1149 /* 1150 * Believe it or not, zeroing out the page array caused a .5% 1151 * performance regression in a database benchmark. So, we take 1152 * care to only zero out what's needed. 1153 */ 1154 memset(dio, 0, offsetof(struct dio, pages)); 1155 1156 dio->flags = flags; 1157 if (dio->flags & DIO_LOCKING) { 1158 if (rw == READ) { 1159 struct address_space *mapping = 1160 iocb->ki_filp->f_mapping; 1161 1162 /* will be released by direct_io_worker */ 1163 mutex_lock(&inode->i_mutex); 1164 1165 retval = filemap_write_and_wait_range(mapping, offset, 1166 end - 1); 1167 if (retval) { 1168 mutex_unlock(&inode->i_mutex); 1169 kmem_cache_free(dio_cache, dio); 1170 goto out; 1171 } 1172 } 1173 } 1174 1175 /* 1176 * Will be decremented at I/O completion time. 1177 */ 1178 atomic_inc(&inode->i_dio_count); 1179 1180 /* 1181 * For file extending writes updating i_size before data 1182 * writeouts complete can expose uninitialized blocks. So 1183 * even for AIO, we need to wait for i/o to complete before 1184 * returning in this case. 1185 */ 1186 dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) && 1187 (end > i_size_read(inode))); 1188 1189 retval = 0; 1190 1191 dio->inode = inode; 1192 dio->rw = rw; 1193 sdio.blkbits = blkbits; 1194 sdio.blkfactor = inode->i_blkbits - blkbits; 1195 sdio.block_in_file = offset >> blkbits; 1196 1197 sdio.get_block = get_block; 1198 dio->end_io = end_io; 1199 sdio.submit_io = submit_io; 1200 sdio.final_block_in_bio = -1; 1201 sdio.next_block_for_io = -1; 1202 1203 dio->iocb = iocb; 1204 dio->i_size = i_size_read(inode); 1205 1206 spin_lock_init(&dio->bio_lock); 1207 dio->refcount = 1; 1208 1209 /* 1210 * In case of non-aligned buffers, we may need 2 more 1211 * pages since we need to zero out first and last block. 1212 */ 1213 if (unlikely(sdio.blkfactor)) 1214 sdio.pages_in_io = 2; 1215 1216 for (seg = 0; seg < nr_segs; seg++) { 1217 user_addr = (unsigned long)iov[seg].iov_base; 1218 sdio.pages_in_io += 1219 ((user_addr + iov[seg].iov_len + PAGE_SIZE-1) / 1220 PAGE_SIZE - user_addr / PAGE_SIZE); 1221 } 1222 1223 for (seg = 0; seg < nr_segs; seg++) { 1224 user_addr = (unsigned long)iov[seg].iov_base; 1225 sdio.size += bytes = iov[seg].iov_len; 1226 1227 /* Index into the first page of the first block */ 1228 sdio.first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits; 1229 sdio.final_block_in_request = sdio.block_in_file + 1230 (bytes >> blkbits); 1231 /* Page fetching state */ 1232 sdio.head = 0; 1233 sdio.tail = 0; 1234 sdio.curr_page = 0; 1235 1236 sdio.total_pages = 0; 1237 if (user_addr & (PAGE_SIZE-1)) { 1238 sdio.total_pages++; 1239 bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1)); 1240 } 1241 sdio.total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE; 1242 sdio.curr_user_address = user_addr; 1243 1244 retval = do_direct_IO(dio, &sdio, &map_bh); 1245 1246 dio->result += iov[seg].iov_len - 1247 ((sdio.final_block_in_request - sdio.block_in_file) << 1248 blkbits); 1249 1250 if (retval) { 1251 dio_cleanup(dio, &sdio); 1252 break; 1253 } 1254 } /* end iovec loop */ 1255 1256 if (retval == -ENOTBLK) { 1257 /* 1258 * The remaining part of the request will be 1259 * be handled by buffered I/O when we return 1260 */ 1261 retval = 0; 1262 } 1263 /* 1264 * There may be some unwritten disk at the end of a part-written 1265 * fs-block-sized block. Go zero that now. 1266 */ 1267 dio_zero_block(dio, &sdio, 1, &map_bh); 1268 1269 if (sdio.cur_page) { 1270 ssize_t ret2; 1271 1272 ret2 = dio_send_cur_page(dio, &sdio, &map_bh); 1273 if (retval == 0) 1274 retval = ret2; 1275 page_cache_release(sdio.cur_page); 1276 sdio.cur_page = NULL; 1277 } 1278 if (sdio.bio) 1279 dio_bio_submit(dio, &sdio); 1280 1281 /* 1282 * It is possible that, we return short IO due to end of file. 1283 * In that case, we need to release all the pages we got hold on. 1284 */ 1285 dio_cleanup(dio, &sdio); 1286 1287 /* 1288 * All block lookups have been performed. For READ requests 1289 * we can let i_mutex go now that its achieved its purpose 1290 * of protecting us from looking up uninitialized blocks. 1291 */ 1292 if (rw == READ && (dio->flags & DIO_LOCKING)) 1293 mutex_unlock(&dio->inode->i_mutex); 1294 1295 /* 1296 * The only time we want to leave bios in flight is when a successful 1297 * partial aio read or full aio write have been setup. In that case 1298 * bio completion will call aio_complete. The only time it's safe to 1299 * call aio_complete is when we return -EIOCBQUEUED, so we key on that. 1300 * This had *better* be the only place that raises -EIOCBQUEUED. 1301 */ 1302 BUG_ON(retval == -EIOCBQUEUED); 1303 if (dio->is_async && retval == 0 && dio->result && 1304 ((rw & READ) || (dio->result == sdio.size))) 1305 retval = -EIOCBQUEUED; 1306 1307 if (retval != -EIOCBQUEUED) 1308 dio_await_completion(dio); 1309 1310 if (drop_refcount(dio) == 0) { 1311 retval = dio_complete(dio, offset, retval, false); 1312 kmem_cache_free(dio_cache, dio); 1313 } else 1314 BUG_ON(retval != -EIOCBQUEUED); 1315 1316 out: 1317 return retval; 1318 } 1319 EXPORT_SYMBOL(__blockdev_direct_IO); 1320 1321 static __init int dio_init(void) 1322 { 1323 dio_cache = KMEM_CACHE(dio, SLAB_PANIC); 1324 return 0; 1325 } 1326 module_init(dio_init) 1327