1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (C) 2010 Red Hat, Inc. 4 * Copyright (c) 2016-2021 Christoph Hellwig. 5 */ 6 #include <linux/module.h> 7 #include <linux/compiler.h> 8 #include <linux/fs.h> 9 #include <linux/fscrypt.h> 10 #include <linux/pagemap.h> 11 #include <linux/iomap.h> 12 #include <linux/backing-dev.h> 13 #include <linux/uio.h> 14 #include <linux/task_io_accounting_ops.h> 15 #include "trace.h" 16 17 #include "../internal.h" 18 19 /* 20 * Private flags for iomap_dio, must not overlap with the public ones in 21 * iomap.h: 22 */ 23 #define IOMAP_DIO_WRITE_FUA (1 << 28) 24 #define IOMAP_DIO_NEED_SYNC (1 << 29) 25 #define IOMAP_DIO_WRITE (1 << 30) 26 #define IOMAP_DIO_DIRTY (1 << 31) 27 28 struct iomap_dio { 29 struct kiocb *iocb; 30 const struct iomap_dio_ops *dops; 31 loff_t i_size; 32 loff_t size; 33 atomic_t ref; 34 unsigned flags; 35 int error; 36 size_t done_before; 37 bool wait_for_completion; 38 39 union { 40 /* used during submission and for synchronous completion: */ 41 struct { 42 struct iov_iter *iter; 43 struct task_struct *waiter; 44 struct bio *poll_bio; 45 } submit; 46 47 /* used for aio completion: */ 48 struct { 49 struct work_struct work; 50 } aio; 51 }; 52 }; 53 54 static void iomap_dio_submit_bio(const struct iomap_iter *iter, 55 struct iomap_dio *dio, struct bio *bio, loff_t pos) 56 { 57 atomic_inc(&dio->ref); 58 59 if (dio->iocb->ki_flags & IOCB_HIPRI) { 60 bio_set_polled(bio, dio->iocb); 61 dio->submit.poll_bio = bio; 62 } 63 64 if (dio->dops && dio->dops->submit_io) 65 dio->dops->submit_io(iter, bio, pos); 66 else 67 submit_bio(bio); 68 } 69 70 ssize_t iomap_dio_complete(struct iomap_dio *dio) 71 { 72 const struct iomap_dio_ops *dops = dio->dops; 73 struct kiocb *iocb = dio->iocb; 74 struct inode *inode = file_inode(iocb->ki_filp); 75 loff_t offset = iocb->ki_pos; 76 ssize_t ret = dio->error; 77 78 if (dops && dops->end_io) 79 ret = dops->end_io(iocb, dio->size, ret, dio->flags); 80 81 if (likely(!ret)) { 82 ret = dio->size; 83 /* check for short read */ 84 if (offset + ret > dio->i_size && 85 !(dio->flags & IOMAP_DIO_WRITE)) 86 ret = dio->i_size - offset; 87 iocb->ki_pos += ret; 88 } 89 90 /* 91 * Try again to invalidate clean pages which might have been cached by 92 * non-direct readahead, or faulted in by get_user_pages() if the source 93 * of the write was an mmap'ed region of the file we're writing. Either 94 * one is a pretty crazy thing to do, so we don't support it 100%. If 95 * this invalidation fails, tough, the write still worked... 96 * 97 * And this page cache invalidation has to be after ->end_io(), as some 98 * filesystems convert unwritten extents to real allocations in 99 * ->end_io() when necessary, otherwise a racing buffer read would cache 100 * zeros from unwritten extents. 101 */ 102 if (!dio->error && dio->size && 103 (dio->flags & IOMAP_DIO_WRITE) && inode->i_mapping->nrpages) { 104 int err; 105 err = invalidate_inode_pages2_range(inode->i_mapping, 106 offset >> PAGE_SHIFT, 107 (offset + dio->size - 1) >> PAGE_SHIFT); 108 if (err) 109 dio_warn_stale_pagecache(iocb->ki_filp); 110 } 111 112 inode_dio_end(file_inode(iocb->ki_filp)); 113 /* 114 * If this is a DSYNC write, make sure we push it to stable storage now 115 * that we've written data. 116 */ 117 if (ret > 0 && (dio->flags & IOMAP_DIO_NEED_SYNC)) 118 ret = generic_write_sync(iocb, ret); 119 120 if (ret > 0) 121 ret += dio->done_before; 122 123 kfree(dio); 124 125 return ret; 126 } 127 EXPORT_SYMBOL_GPL(iomap_dio_complete); 128 129 static void iomap_dio_complete_work(struct work_struct *work) 130 { 131 struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work); 132 struct kiocb *iocb = dio->iocb; 133 134 iocb->ki_complete(iocb, iomap_dio_complete(dio)); 135 } 136 137 /* 138 * Set an error in the dio if none is set yet. We have to use cmpxchg 139 * as the submission context and the completion context(s) can race to 140 * update the error. 141 */ 142 static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret) 143 { 144 cmpxchg(&dio->error, 0, ret); 145 } 146 147 static void iomap_dio_bio_end_io(struct bio *bio) 148 { 149 struct iomap_dio *dio = bio->bi_private; 150 bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY); 151 152 if (bio->bi_status) 153 iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status)); 154 155 if (atomic_dec_and_test(&dio->ref)) { 156 if (dio->wait_for_completion) { 157 struct task_struct *waiter = dio->submit.waiter; 158 WRITE_ONCE(dio->submit.waiter, NULL); 159 blk_wake_io_task(waiter); 160 } else if (dio->flags & IOMAP_DIO_WRITE) { 161 struct inode *inode = file_inode(dio->iocb->ki_filp); 162 163 WRITE_ONCE(dio->iocb->private, NULL); 164 INIT_WORK(&dio->aio.work, iomap_dio_complete_work); 165 queue_work(inode->i_sb->s_dio_done_wq, &dio->aio.work); 166 } else { 167 WRITE_ONCE(dio->iocb->private, NULL); 168 iomap_dio_complete_work(&dio->aio.work); 169 } 170 } 171 172 if (should_dirty) { 173 bio_check_pages_dirty(bio); 174 } else { 175 bio_release_pages(bio, false); 176 bio_put(bio); 177 } 178 } 179 180 static void iomap_dio_zero(const struct iomap_iter *iter, struct iomap_dio *dio, 181 loff_t pos, unsigned len) 182 { 183 struct inode *inode = file_inode(dio->iocb->ki_filp); 184 struct page *page = ZERO_PAGE(0); 185 int flags = REQ_SYNC | REQ_IDLE; 186 struct bio *bio; 187 188 bio = bio_alloc(iter->iomap.bdev, 1, REQ_OP_WRITE | flags, GFP_KERNEL); 189 fscrypt_set_bio_crypt_ctx(bio, inode, pos >> inode->i_blkbits, 190 GFP_KERNEL); 191 bio->bi_iter.bi_sector = iomap_sector(&iter->iomap, pos); 192 bio->bi_private = dio; 193 bio->bi_end_io = iomap_dio_bio_end_io; 194 195 get_page(page); 196 __bio_add_page(bio, page, len, 0); 197 iomap_dio_submit_bio(iter, dio, bio, pos); 198 } 199 200 /* 201 * Figure out the bio's operation flags from the dio request, the 202 * mapping, and whether or not we want FUA. Note that we can end up 203 * clearing the WRITE_FUA flag in the dio request. 204 */ 205 static inline unsigned int iomap_dio_bio_opflags(struct iomap_dio *dio, 206 const struct iomap *iomap, bool use_fua) 207 { 208 unsigned int opflags = REQ_SYNC | REQ_IDLE; 209 210 if (!(dio->flags & IOMAP_DIO_WRITE)) { 211 WARN_ON_ONCE(iomap->flags & IOMAP_F_ZONE_APPEND); 212 return REQ_OP_READ; 213 } 214 215 if (iomap->flags & IOMAP_F_ZONE_APPEND) 216 opflags |= REQ_OP_ZONE_APPEND; 217 else 218 opflags |= REQ_OP_WRITE; 219 220 if (use_fua) 221 opflags |= REQ_FUA; 222 else 223 dio->flags &= ~IOMAP_DIO_WRITE_FUA; 224 225 return opflags; 226 } 227 228 static loff_t iomap_dio_bio_iter(const struct iomap_iter *iter, 229 struct iomap_dio *dio) 230 { 231 const struct iomap *iomap = &iter->iomap; 232 struct inode *inode = iter->inode; 233 unsigned int blkbits = blksize_bits(bdev_logical_block_size(iomap->bdev)); 234 unsigned int fs_block_size = i_blocksize(inode), pad; 235 unsigned int align = iov_iter_alignment(dio->submit.iter); 236 loff_t length = iomap_length(iter); 237 loff_t pos = iter->pos; 238 unsigned int bio_opf; 239 struct bio *bio; 240 bool need_zeroout = false; 241 bool use_fua = false; 242 int nr_pages, ret = 0; 243 size_t copied = 0; 244 size_t orig_count; 245 246 if ((pos | length | align) & ((1 << blkbits) - 1)) 247 return -EINVAL; 248 249 if (iomap->type == IOMAP_UNWRITTEN) { 250 dio->flags |= IOMAP_DIO_UNWRITTEN; 251 need_zeroout = true; 252 } 253 254 if (iomap->flags & IOMAP_F_SHARED) 255 dio->flags |= IOMAP_DIO_COW; 256 257 if (iomap->flags & IOMAP_F_NEW) { 258 need_zeroout = true; 259 } else if (iomap->type == IOMAP_MAPPED) { 260 /* 261 * Use a FUA write if we need datasync semantics, this is a pure 262 * data IO that doesn't require any metadata updates (including 263 * after IO completion such as unwritten extent conversion) and 264 * the underlying device supports FUA. This allows us to avoid 265 * cache flushes on IO completion. 266 */ 267 if (!(iomap->flags & (IOMAP_F_SHARED|IOMAP_F_DIRTY)) && 268 (dio->flags & IOMAP_DIO_WRITE_FUA) && 269 blk_queue_fua(bdev_get_queue(iomap->bdev))) 270 use_fua = true; 271 } 272 273 /* 274 * Save the original count and trim the iter to just the extent we 275 * are operating on right now. The iter will be re-expanded once 276 * we are done. 277 */ 278 orig_count = iov_iter_count(dio->submit.iter); 279 iov_iter_truncate(dio->submit.iter, length); 280 281 if (!iov_iter_count(dio->submit.iter)) 282 goto out; 283 284 /* 285 * We can only poll for single bio I/Os. 286 */ 287 if (need_zeroout || 288 ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode))) 289 dio->iocb->ki_flags &= ~IOCB_HIPRI; 290 291 if (need_zeroout) { 292 /* zero out from the start of the block to the write offset */ 293 pad = pos & (fs_block_size - 1); 294 if (pad) 295 iomap_dio_zero(iter, dio, pos - pad, pad); 296 } 297 298 /* 299 * Set the operation flags early so that bio_iov_iter_get_pages 300 * can set up the page vector appropriately for a ZONE_APPEND 301 * operation. 302 */ 303 bio_opf = iomap_dio_bio_opflags(dio, iomap, use_fua); 304 305 nr_pages = bio_iov_vecs_to_alloc(dio->submit.iter, BIO_MAX_VECS); 306 do { 307 size_t n; 308 if (dio->error) { 309 iov_iter_revert(dio->submit.iter, copied); 310 copied = ret = 0; 311 goto out; 312 } 313 314 bio = bio_alloc(iomap->bdev, nr_pages, bio_opf, GFP_KERNEL); 315 fscrypt_set_bio_crypt_ctx(bio, inode, pos >> inode->i_blkbits, 316 GFP_KERNEL); 317 bio->bi_iter.bi_sector = iomap_sector(iomap, pos); 318 bio->bi_ioprio = dio->iocb->ki_ioprio; 319 bio->bi_private = dio; 320 bio->bi_end_io = iomap_dio_bio_end_io; 321 322 ret = bio_iov_iter_get_pages(bio, dio->submit.iter); 323 if (unlikely(ret)) { 324 /* 325 * We have to stop part way through an IO. We must fall 326 * through to the sub-block tail zeroing here, otherwise 327 * this short IO may expose stale data in the tail of 328 * the block we haven't written data to. 329 */ 330 bio_put(bio); 331 goto zero_tail; 332 } 333 334 n = bio->bi_iter.bi_size; 335 if (dio->flags & IOMAP_DIO_WRITE) { 336 task_io_account_write(n); 337 } else { 338 if (dio->flags & IOMAP_DIO_DIRTY) 339 bio_set_pages_dirty(bio); 340 } 341 342 dio->size += n; 343 copied += n; 344 345 nr_pages = bio_iov_vecs_to_alloc(dio->submit.iter, 346 BIO_MAX_VECS); 347 /* 348 * We can only poll for single bio I/Os. 349 */ 350 if (nr_pages) 351 dio->iocb->ki_flags &= ~IOCB_HIPRI; 352 iomap_dio_submit_bio(iter, dio, bio, pos); 353 pos += n; 354 } while (nr_pages); 355 356 /* 357 * We need to zeroout the tail of a sub-block write if the extent type 358 * requires zeroing or the write extends beyond EOF. If we don't zero 359 * the block tail in the latter case, we can expose stale data via mmap 360 * reads of the EOF block. 361 */ 362 zero_tail: 363 if (need_zeroout || 364 ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode))) { 365 /* zero out from the end of the write to the end of the block */ 366 pad = pos & (fs_block_size - 1); 367 if (pad) 368 iomap_dio_zero(iter, dio, pos, fs_block_size - pad); 369 } 370 out: 371 /* Undo iter limitation to current extent */ 372 iov_iter_reexpand(dio->submit.iter, orig_count - copied); 373 if (copied) 374 return copied; 375 return ret; 376 } 377 378 static loff_t iomap_dio_hole_iter(const struct iomap_iter *iter, 379 struct iomap_dio *dio) 380 { 381 loff_t length = iov_iter_zero(iomap_length(iter), dio->submit.iter); 382 383 dio->size += length; 384 if (!length) 385 return -EFAULT; 386 return length; 387 } 388 389 static loff_t iomap_dio_inline_iter(const struct iomap_iter *iomi, 390 struct iomap_dio *dio) 391 { 392 const struct iomap *iomap = &iomi->iomap; 393 struct iov_iter *iter = dio->submit.iter; 394 void *inline_data = iomap_inline_data(iomap, iomi->pos); 395 loff_t length = iomap_length(iomi); 396 loff_t pos = iomi->pos; 397 size_t copied; 398 399 if (WARN_ON_ONCE(!iomap_inline_data_valid(iomap))) 400 return -EIO; 401 402 if (dio->flags & IOMAP_DIO_WRITE) { 403 loff_t size = iomi->inode->i_size; 404 405 if (pos > size) 406 memset(iomap_inline_data(iomap, size), 0, pos - size); 407 copied = copy_from_iter(inline_data, length, iter); 408 if (copied) { 409 if (pos + copied > size) 410 i_size_write(iomi->inode, pos + copied); 411 mark_inode_dirty(iomi->inode); 412 } 413 } else { 414 copied = copy_to_iter(inline_data, length, iter); 415 } 416 dio->size += copied; 417 if (!copied) 418 return -EFAULT; 419 return copied; 420 } 421 422 static loff_t iomap_dio_iter(const struct iomap_iter *iter, 423 struct iomap_dio *dio) 424 { 425 switch (iter->iomap.type) { 426 case IOMAP_HOLE: 427 if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE)) 428 return -EIO; 429 return iomap_dio_hole_iter(iter, dio); 430 case IOMAP_UNWRITTEN: 431 if (!(dio->flags & IOMAP_DIO_WRITE)) 432 return iomap_dio_hole_iter(iter, dio); 433 return iomap_dio_bio_iter(iter, dio); 434 case IOMAP_MAPPED: 435 return iomap_dio_bio_iter(iter, dio); 436 case IOMAP_INLINE: 437 return iomap_dio_inline_iter(iter, dio); 438 case IOMAP_DELALLOC: 439 /* 440 * DIO is not serialised against mmap() access at all, and so 441 * if the page_mkwrite occurs between the writeback and the 442 * iomap_iter() call in the DIO path, then it will see the 443 * DELALLOC block that the page-mkwrite allocated. 444 */ 445 pr_warn_ratelimited("Direct I/O collision with buffered writes! File: %pD4 Comm: %.20s\n", 446 dio->iocb->ki_filp, current->comm); 447 return -EIO; 448 default: 449 WARN_ON_ONCE(1); 450 return -EIO; 451 } 452 } 453 454 /* 455 * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO 456 * is being issued as AIO or not. This allows us to optimise pure data writes 457 * to use REQ_FUA rather than requiring generic_write_sync() to issue a 458 * REQ_FLUSH post write. This is slightly tricky because a single request here 459 * can be mapped into multiple disjoint IOs and only a subset of the IOs issued 460 * may be pure data writes. In that case, we still need to do a full data sync 461 * completion. 462 * 463 * When page faults are disabled and @dio_flags includes IOMAP_DIO_PARTIAL, 464 * __iomap_dio_rw can return a partial result if it encounters a non-resident 465 * page in @iter after preparing a transfer. In that case, the non-resident 466 * pages can be faulted in and the request resumed with @done_before set to the 467 * number of bytes previously transferred. The request will then complete with 468 * the correct total number of bytes transferred; this is essential for 469 * completing partial requests asynchronously. 470 * 471 * Returns -ENOTBLK In case of a page invalidation invalidation failure for 472 * writes. The callers needs to fall back to buffered I/O in this case. 473 */ 474 struct iomap_dio * 475 __iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter, 476 const struct iomap_ops *ops, const struct iomap_dio_ops *dops, 477 unsigned int dio_flags, size_t done_before) 478 { 479 struct address_space *mapping = iocb->ki_filp->f_mapping; 480 struct inode *inode = file_inode(iocb->ki_filp); 481 struct iomap_iter iomi = { 482 .inode = inode, 483 .pos = iocb->ki_pos, 484 .len = iov_iter_count(iter), 485 .flags = IOMAP_DIRECT, 486 }; 487 loff_t end = iomi.pos + iomi.len - 1, ret = 0; 488 bool wait_for_completion = 489 is_sync_kiocb(iocb) || (dio_flags & IOMAP_DIO_FORCE_WAIT); 490 struct blk_plug plug; 491 struct iomap_dio *dio; 492 493 if (!iomi.len) 494 return NULL; 495 496 dio = kmalloc(sizeof(*dio), GFP_KERNEL); 497 if (!dio) 498 return ERR_PTR(-ENOMEM); 499 500 dio->iocb = iocb; 501 atomic_set(&dio->ref, 1); 502 dio->size = 0; 503 dio->i_size = i_size_read(inode); 504 dio->dops = dops; 505 dio->error = 0; 506 dio->flags = 0; 507 dio->done_before = done_before; 508 509 dio->submit.iter = iter; 510 dio->submit.waiter = current; 511 dio->submit.poll_bio = NULL; 512 513 if (iov_iter_rw(iter) == READ) { 514 if (iomi.pos >= dio->i_size) 515 goto out_free_dio; 516 517 if (iocb->ki_flags & IOCB_NOWAIT) { 518 if (filemap_range_needs_writeback(mapping, iomi.pos, 519 end)) { 520 ret = -EAGAIN; 521 goto out_free_dio; 522 } 523 iomi.flags |= IOMAP_NOWAIT; 524 } 525 526 if (iter_is_iovec(iter)) 527 dio->flags |= IOMAP_DIO_DIRTY; 528 } else { 529 iomi.flags |= IOMAP_WRITE; 530 dio->flags |= IOMAP_DIO_WRITE; 531 532 if (iocb->ki_flags & IOCB_NOWAIT) { 533 if (filemap_range_has_page(mapping, iomi.pos, end)) { 534 ret = -EAGAIN; 535 goto out_free_dio; 536 } 537 iomi.flags |= IOMAP_NOWAIT; 538 } 539 540 /* for data sync or sync, we need sync completion processing */ 541 if (iocb->ki_flags & IOCB_DSYNC) 542 dio->flags |= IOMAP_DIO_NEED_SYNC; 543 544 /* 545 * For datasync only writes, we optimistically try using FUA for 546 * this IO. Any non-FUA write that occurs will clear this flag, 547 * hence we know before completion whether a cache flush is 548 * necessary. 549 */ 550 if ((iocb->ki_flags & (IOCB_DSYNC | IOCB_SYNC)) == IOCB_DSYNC) 551 dio->flags |= IOMAP_DIO_WRITE_FUA; 552 } 553 554 if (dio_flags & IOMAP_DIO_OVERWRITE_ONLY) { 555 ret = -EAGAIN; 556 if (iomi.pos >= dio->i_size || 557 iomi.pos + iomi.len > dio->i_size) 558 goto out_free_dio; 559 iomi.flags |= IOMAP_OVERWRITE_ONLY; 560 } 561 562 ret = filemap_write_and_wait_range(mapping, iomi.pos, end); 563 if (ret) 564 goto out_free_dio; 565 566 if (iov_iter_rw(iter) == WRITE) { 567 /* 568 * Try to invalidate cache pages for the range we are writing. 569 * If this invalidation fails, let the caller fall back to 570 * buffered I/O. 571 */ 572 if (invalidate_inode_pages2_range(mapping, 573 iomi.pos >> PAGE_SHIFT, end >> PAGE_SHIFT)) { 574 trace_iomap_dio_invalidate_fail(inode, iomi.pos, 575 iomi.len); 576 ret = -ENOTBLK; 577 goto out_free_dio; 578 } 579 580 if (!wait_for_completion && !inode->i_sb->s_dio_done_wq) { 581 ret = sb_init_dio_done_wq(inode->i_sb); 582 if (ret < 0) 583 goto out_free_dio; 584 } 585 } 586 587 inode_dio_begin(inode); 588 589 blk_start_plug(&plug); 590 while ((ret = iomap_iter(&iomi, ops)) > 0) { 591 iomi.processed = iomap_dio_iter(&iomi, dio); 592 593 /* 594 * We can only poll for single bio I/Os. 595 */ 596 iocb->ki_flags &= ~IOCB_HIPRI; 597 } 598 599 blk_finish_plug(&plug); 600 601 /* 602 * We only report that we've read data up to i_size. 603 * Revert iter to a state corresponding to that as some callers (such 604 * as the splice code) rely on it. 605 */ 606 if (iov_iter_rw(iter) == READ && iomi.pos >= dio->i_size) 607 iov_iter_revert(iter, iomi.pos - dio->i_size); 608 609 if (ret == -EFAULT && dio->size && (dio_flags & IOMAP_DIO_PARTIAL)) { 610 if (!(iocb->ki_flags & IOCB_NOWAIT)) 611 wait_for_completion = true; 612 ret = 0; 613 } 614 615 /* magic error code to fall back to buffered I/O */ 616 if (ret == -ENOTBLK) { 617 wait_for_completion = true; 618 ret = 0; 619 } 620 if (ret < 0) 621 iomap_dio_set_error(dio, ret); 622 623 /* 624 * If all the writes we issued were FUA, we don't need to flush the 625 * cache on IO completion. Clear the sync flag for this case. 626 */ 627 if (dio->flags & IOMAP_DIO_WRITE_FUA) 628 dio->flags &= ~IOMAP_DIO_NEED_SYNC; 629 630 WRITE_ONCE(iocb->private, dio->submit.poll_bio); 631 632 /* 633 * We are about to drop our additional submission reference, which 634 * might be the last reference to the dio. There are three different 635 * ways we can progress here: 636 * 637 * (a) If this is the last reference we will always complete and free 638 * the dio ourselves. 639 * (b) If this is not the last reference, and we serve an asynchronous 640 * iocb, we must never touch the dio after the decrement, the 641 * I/O completion handler will complete and free it. 642 * (c) If this is not the last reference, but we serve a synchronous 643 * iocb, the I/O completion handler will wake us up on the drop 644 * of the final reference, and we will complete and free it here 645 * after we got woken by the I/O completion handler. 646 */ 647 dio->wait_for_completion = wait_for_completion; 648 if (!atomic_dec_and_test(&dio->ref)) { 649 if (!wait_for_completion) 650 return ERR_PTR(-EIOCBQUEUED); 651 652 for (;;) { 653 set_current_state(TASK_UNINTERRUPTIBLE); 654 if (!READ_ONCE(dio->submit.waiter)) 655 break; 656 657 if (!dio->submit.poll_bio || 658 !bio_poll(dio->submit.poll_bio, NULL, 0)) 659 blk_io_schedule(); 660 } 661 __set_current_state(TASK_RUNNING); 662 } 663 664 return dio; 665 666 out_free_dio: 667 kfree(dio); 668 if (ret) 669 return ERR_PTR(ret); 670 return NULL; 671 } 672 EXPORT_SYMBOL_GPL(__iomap_dio_rw); 673 674 ssize_t 675 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter, 676 const struct iomap_ops *ops, const struct iomap_dio_ops *dops, 677 unsigned int dio_flags, size_t done_before) 678 { 679 struct iomap_dio *dio; 680 681 dio = __iomap_dio_rw(iocb, iter, ops, dops, dio_flags, done_before); 682 if (IS_ERR_OR_NULL(dio)) 683 return PTR_ERR_OR_ZERO(dio); 684 return iomap_dio_complete(dio); 685 } 686 EXPORT_SYMBOL_GPL(iomap_dio_rw); 687