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 struct bio *iomap_dio_alloc_bio(const struct iomap_iter *iter, 55 struct iomap_dio *dio, unsigned short nr_vecs, blk_opf_t opf) 56 { 57 if (dio->dops && dio->dops->bio_set) 58 return bio_alloc_bioset(iter->iomap.bdev, nr_vecs, opf, 59 GFP_KERNEL, dio->dops->bio_set); 60 return bio_alloc(iter->iomap.bdev, nr_vecs, opf, GFP_KERNEL); 61 } 62 63 static void iomap_dio_submit_bio(const struct iomap_iter *iter, 64 struct iomap_dio *dio, struct bio *bio, loff_t pos) 65 { 66 atomic_inc(&dio->ref); 67 68 /* Sync dio can't be polled reliably */ 69 if ((dio->iocb->ki_flags & IOCB_HIPRI) && !is_sync_kiocb(dio->iocb)) { 70 bio_set_polled(bio, dio->iocb); 71 dio->submit.poll_bio = bio; 72 } 73 74 if (dio->dops && dio->dops->submit_io) 75 dio->dops->submit_io(iter, bio, pos); 76 else 77 submit_bio(bio); 78 } 79 80 ssize_t iomap_dio_complete(struct iomap_dio *dio) 81 { 82 const struct iomap_dio_ops *dops = dio->dops; 83 struct kiocb *iocb = dio->iocb; 84 loff_t offset = iocb->ki_pos; 85 ssize_t ret = dio->error; 86 87 if (dops && dops->end_io) 88 ret = dops->end_io(iocb, dio->size, ret, dio->flags); 89 90 if (likely(!ret)) { 91 ret = dio->size; 92 /* check for short read */ 93 if (offset + ret > dio->i_size && 94 !(dio->flags & IOMAP_DIO_WRITE)) 95 ret = dio->i_size - offset; 96 } 97 98 /* 99 * Try again to invalidate clean pages which might have been cached by 100 * non-direct readahead, or faulted in by get_user_pages() if the source 101 * of the write was an mmap'ed region of the file we're writing. Either 102 * one is a pretty crazy thing to do, so we don't support it 100%. If 103 * this invalidation fails, tough, the write still worked... 104 * 105 * And this page cache invalidation has to be after ->end_io(), as some 106 * filesystems convert unwritten extents to real allocations in 107 * ->end_io() when necessary, otherwise a racing buffer read would cache 108 * zeros from unwritten extents. 109 */ 110 if (!dio->error && dio->size && (dio->flags & IOMAP_DIO_WRITE)) 111 kiocb_invalidate_post_direct_write(iocb, dio->size); 112 113 inode_dio_end(file_inode(iocb->ki_filp)); 114 115 if (ret > 0) { 116 iocb->ki_pos += ret; 117 118 /* 119 * If this is a DSYNC write, make sure we push it to stable 120 * storage now that we've written data. 121 */ 122 if (dio->flags & IOMAP_DIO_NEED_SYNC) 123 ret = generic_write_sync(iocb, ret); 124 if (ret > 0) 125 ret += dio->done_before; 126 } 127 trace_iomap_dio_complete(iocb, dio->error, ret); 128 kfree(dio); 129 return ret; 130 } 131 EXPORT_SYMBOL_GPL(iomap_dio_complete); 132 133 static void iomap_dio_complete_work(struct work_struct *work) 134 { 135 struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work); 136 struct kiocb *iocb = dio->iocb; 137 138 iocb->ki_complete(iocb, iomap_dio_complete(dio)); 139 } 140 141 /* 142 * Set an error in the dio if none is set yet. We have to use cmpxchg 143 * as the submission context and the completion context(s) can race to 144 * update the error. 145 */ 146 static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret) 147 { 148 cmpxchg(&dio->error, 0, ret); 149 } 150 151 void iomap_dio_bio_end_io(struct bio *bio) 152 { 153 struct iomap_dio *dio = bio->bi_private; 154 bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY); 155 struct kiocb *iocb = dio->iocb; 156 157 if (bio->bi_status) 158 iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status)); 159 if (!atomic_dec_and_test(&dio->ref)) 160 goto release_bio; 161 162 /* 163 * Synchronous dio, task itself will handle any completion work 164 * that needs after IO. All we need to do is wake the task. 165 */ 166 if (dio->wait_for_completion) { 167 struct task_struct *waiter = dio->submit.waiter; 168 169 WRITE_ONCE(dio->submit.waiter, NULL); 170 blk_wake_io_task(waiter); 171 goto release_bio; 172 } 173 174 /* Read completion can always complete inline. */ 175 if (!(dio->flags & IOMAP_DIO_WRITE)) { 176 WRITE_ONCE(iocb->private, NULL); 177 iomap_dio_complete_work(&dio->aio.work); 178 goto release_bio; 179 } 180 181 /* 182 * Async DIO completion that requires filesystem level completion work 183 * gets punted to a work queue to complete as the operation may require 184 * more IO to be issued to finalise filesystem metadata changes or 185 * guarantee data integrity. 186 */ 187 WRITE_ONCE(iocb->private, NULL); 188 INIT_WORK(&dio->aio.work, iomap_dio_complete_work); 189 queue_work(file_inode(iocb->ki_filp)->i_sb->s_dio_done_wq, 190 &dio->aio.work); 191 release_bio: 192 if (should_dirty) { 193 bio_check_pages_dirty(bio); 194 } else { 195 bio_release_pages(bio, false); 196 bio_put(bio); 197 } 198 } 199 EXPORT_SYMBOL_GPL(iomap_dio_bio_end_io); 200 201 static void iomap_dio_zero(const struct iomap_iter *iter, struct iomap_dio *dio, 202 loff_t pos, unsigned len) 203 { 204 struct inode *inode = file_inode(dio->iocb->ki_filp); 205 struct page *page = ZERO_PAGE(0); 206 struct bio *bio; 207 208 bio = iomap_dio_alloc_bio(iter, dio, 1, REQ_OP_WRITE | REQ_SYNC | REQ_IDLE); 209 fscrypt_set_bio_crypt_ctx(bio, inode, pos >> inode->i_blkbits, 210 GFP_KERNEL); 211 bio->bi_iter.bi_sector = iomap_sector(&iter->iomap, pos); 212 bio->bi_private = dio; 213 bio->bi_end_io = iomap_dio_bio_end_io; 214 215 __bio_add_page(bio, page, len, 0); 216 iomap_dio_submit_bio(iter, dio, bio, pos); 217 } 218 219 /* 220 * Figure out the bio's operation flags from the dio request, the 221 * mapping, and whether or not we want FUA. Note that we can end up 222 * clearing the WRITE_FUA flag in the dio request. 223 */ 224 static inline blk_opf_t iomap_dio_bio_opflags(struct iomap_dio *dio, 225 const struct iomap *iomap, bool use_fua) 226 { 227 blk_opf_t opflags = REQ_SYNC | REQ_IDLE; 228 229 if (!(dio->flags & IOMAP_DIO_WRITE)) 230 return REQ_OP_READ; 231 232 opflags |= REQ_OP_WRITE; 233 if (use_fua) 234 opflags |= REQ_FUA; 235 else 236 dio->flags &= ~IOMAP_DIO_WRITE_FUA; 237 238 return opflags; 239 } 240 241 static loff_t iomap_dio_bio_iter(const struct iomap_iter *iter, 242 struct iomap_dio *dio) 243 { 244 const struct iomap *iomap = &iter->iomap; 245 struct inode *inode = iter->inode; 246 unsigned int fs_block_size = i_blocksize(inode), pad; 247 loff_t length = iomap_length(iter); 248 loff_t pos = iter->pos; 249 blk_opf_t bio_opf; 250 struct bio *bio; 251 bool need_zeroout = false; 252 bool use_fua = false; 253 int nr_pages, ret = 0; 254 size_t copied = 0; 255 size_t orig_count; 256 257 if ((pos | length) & (bdev_logical_block_size(iomap->bdev) - 1) || 258 !bdev_iter_is_aligned(iomap->bdev, dio->submit.iter)) 259 return -EINVAL; 260 261 if (iomap->type == IOMAP_UNWRITTEN) { 262 dio->flags |= IOMAP_DIO_UNWRITTEN; 263 need_zeroout = true; 264 } 265 266 if (iomap->flags & IOMAP_F_SHARED) 267 dio->flags |= IOMAP_DIO_COW; 268 269 if (iomap->flags & IOMAP_F_NEW) { 270 need_zeroout = true; 271 } else if (iomap->type == IOMAP_MAPPED) { 272 /* 273 * Use a FUA write if we need datasync semantics, this is a pure 274 * data IO that doesn't require any metadata updates (including 275 * after IO completion such as unwritten extent conversion) and 276 * the underlying device supports FUA. This allows us to avoid 277 * cache flushes on IO completion. 278 */ 279 if (!(iomap->flags & (IOMAP_F_SHARED|IOMAP_F_DIRTY)) && 280 (dio->flags & IOMAP_DIO_WRITE_FUA) && bdev_fua(iomap->bdev)) 281 use_fua = true; 282 } 283 284 /* 285 * Save the original count and trim the iter to just the extent we 286 * are operating on right now. The iter will be re-expanded once 287 * we are done. 288 */ 289 orig_count = iov_iter_count(dio->submit.iter); 290 iov_iter_truncate(dio->submit.iter, length); 291 292 if (!iov_iter_count(dio->submit.iter)) 293 goto out; 294 295 /* 296 * We can only poll for single bio I/Os. 297 */ 298 if (need_zeroout || 299 ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode))) 300 dio->iocb->ki_flags &= ~IOCB_HIPRI; 301 302 if (need_zeroout) { 303 /* zero out from the start of the block to the write offset */ 304 pad = pos & (fs_block_size - 1); 305 if (pad) 306 iomap_dio_zero(iter, dio, pos - pad, pad); 307 } 308 309 /* 310 * Set the operation flags early so that bio_iov_iter_get_pages 311 * can set up the page vector appropriately for a ZONE_APPEND 312 * operation. 313 */ 314 bio_opf = iomap_dio_bio_opflags(dio, iomap, use_fua); 315 316 nr_pages = bio_iov_vecs_to_alloc(dio->submit.iter, BIO_MAX_VECS); 317 do { 318 size_t n; 319 if (dio->error) { 320 iov_iter_revert(dio->submit.iter, copied); 321 copied = ret = 0; 322 goto out; 323 } 324 325 bio = iomap_dio_alloc_bio(iter, dio, nr_pages, bio_opf); 326 fscrypt_set_bio_crypt_ctx(bio, inode, pos >> inode->i_blkbits, 327 GFP_KERNEL); 328 bio->bi_iter.bi_sector = iomap_sector(iomap, pos); 329 bio->bi_ioprio = dio->iocb->ki_ioprio; 330 bio->bi_private = dio; 331 bio->bi_end_io = iomap_dio_bio_end_io; 332 333 ret = bio_iov_iter_get_pages(bio, dio->submit.iter); 334 if (unlikely(ret)) { 335 /* 336 * We have to stop part way through an IO. We must fall 337 * through to the sub-block tail zeroing here, otherwise 338 * this short IO may expose stale data in the tail of 339 * the block we haven't written data to. 340 */ 341 bio_put(bio); 342 goto zero_tail; 343 } 344 345 n = bio->bi_iter.bi_size; 346 if (dio->flags & IOMAP_DIO_WRITE) { 347 task_io_account_write(n); 348 } else { 349 if (dio->flags & IOMAP_DIO_DIRTY) 350 bio_set_pages_dirty(bio); 351 } 352 353 dio->size += n; 354 copied += n; 355 356 nr_pages = bio_iov_vecs_to_alloc(dio->submit.iter, 357 BIO_MAX_VECS); 358 /* 359 * We can only poll for single bio I/Os. 360 */ 361 if (nr_pages) 362 dio->iocb->ki_flags &= ~IOCB_HIPRI; 363 iomap_dio_submit_bio(iter, dio, bio, pos); 364 pos += n; 365 } while (nr_pages); 366 367 /* 368 * We need to zeroout the tail of a sub-block write if the extent type 369 * requires zeroing or the write extends beyond EOF. If we don't zero 370 * the block tail in the latter case, we can expose stale data via mmap 371 * reads of the EOF block. 372 */ 373 zero_tail: 374 if (need_zeroout || 375 ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode))) { 376 /* zero out from the end of the write to the end of the block */ 377 pad = pos & (fs_block_size - 1); 378 if (pad) 379 iomap_dio_zero(iter, dio, pos, fs_block_size - pad); 380 } 381 out: 382 /* Undo iter limitation to current extent */ 383 iov_iter_reexpand(dio->submit.iter, orig_count - copied); 384 if (copied) 385 return copied; 386 return ret; 387 } 388 389 static loff_t iomap_dio_hole_iter(const struct iomap_iter *iter, 390 struct iomap_dio *dio) 391 { 392 loff_t length = iov_iter_zero(iomap_length(iter), dio->submit.iter); 393 394 dio->size += length; 395 if (!length) 396 return -EFAULT; 397 return length; 398 } 399 400 static loff_t iomap_dio_inline_iter(const struct iomap_iter *iomi, 401 struct iomap_dio *dio) 402 { 403 const struct iomap *iomap = &iomi->iomap; 404 struct iov_iter *iter = dio->submit.iter; 405 void *inline_data = iomap_inline_data(iomap, iomi->pos); 406 loff_t length = iomap_length(iomi); 407 loff_t pos = iomi->pos; 408 size_t copied; 409 410 if (WARN_ON_ONCE(!iomap_inline_data_valid(iomap))) 411 return -EIO; 412 413 if (dio->flags & IOMAP_DIO_WRITE) { 414 loff_t size = iomi->inode->i_size; 415 416 if (pos > size) 417 memset(iomap_inline_data(iomap, size), 0, pos - size); 418 copied = copy_from_iter(inline_data, length, iter); 419 if (copied) { 420 if (pos + copied > size) 421 i_size_write(iomi->inode, pos + copied); 422 mark_inode_dirty(iomi->inode); 423 } 424 } else { 425 copied = copy_to_iter(inline_data, length, iter); 426 } 427 dio->size += copied; 428 if (!copied) 429 return -EFAULT; 430 return copied; 431 } 432 433 static loff_t iomap_dio_iter(const struct iomap_iter *iter, 434 struct iomap_dio *dio) 435 { 436 switch (iter->iomap.type) { 437 case IOMAP_HOLE: 438 if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE)) 439 return -EIO; 440 return iomap_dio_hole_iter(iter, dio); 441 case IOMAP_UNWRITTEN: 442 if (!(dio->flags & IOMAP_DIO_WRITE)) 443 return iomap_dio_hole_iter(iter, dio); 444 return iomap_dio_bio_iter(iter, dio); 445 case IOMAP_MAPPED: 446 return iomap_dio_bio_iter(iter, dio); 447 case IOMAP_INLINE: 448 return iomap_dio_inline_iter(iter, dio); 449 case IOMAP_DELALLOC: 450 /* 451 * DIO is not serialised against mmap() access at all, and so 452 * if the page_mkwrite occurs between the writeback and the 453 * iomap_iter() call in the DIO path, then it will see the 454 * DELALLOC block that the page-mkwrite allocated. 455 */ 456 pr_warn_ratelimited("Direct I/O collision with buffered writes! File: %pD4 Comm: %.20s\n", 457 dio->iocb->ki_filp, current->comm); 458 return -EIO; 459 default: 460 WARN_ON_ONCE(1); 461 return -EIO; 462 } 463 } 464 465 /* 466 * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO 467 * is being issued as AIO or not. This allows us to optimise pure data writes 468 * to use REQ_FUA rather than requiring generic_write_sync() to issue a 469 * REQ_FLUSH post write. This is slightly tricky because a single request here 470 * can be mapped into multiple disjoint IOs and only a subset of the IOs issued 471 * may be pure data writes. In that case, we still need to do a full data sync 472 * completion. 473 * 474 * When page faults are disabled and @dio_flags includes IOMAP_DIO_PARTIAL, 475 * __iomap_dio_rw can return a partial result if it encounters a non-resident 476 * page in @iter after preparing a transfer. In that case, the non-resident 477 * pages can be faulted in and the request resumed with @done_before set to the 478 * number of bytes previously transferred. The request will then complete with 479 * the correct total number of bytes transferred; this is essential for 480 * completing partial requests asynchronously. 481 * 482 * Returns -ENOTBLK In case of a page invalidation invalidation failure for 483 * writes. The callers needs to fall back to buffered I/O in this case. 484 */ 485 struct iomap_dio * 486 __iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter, 487 const struct iomap_ops *ops, const struct iomap_dio_ops *dops, 488 unsigned int dio_flags, void *private, size_t done_before) 489 { 490 struct inode *inode = file_inode(iocb->ki_filp); 491 struct iomap_iter iomi = { 492 .inode = inode, 493 .pos = iocb->ki_pos, 494 .len = iov_iter_count(iter), 495 .flags = IOMAP_DIRECT, 496 .private = private, 497 }; 498 bool wait_for_completion = 499 is_sync_kiocb(iocb) || (dio_flags & IOMAP_DIO_FORCE_WAIT); 500 struct blk_plug plug; 501 struct iomap_dio *dio; 502 loff_t ret = 0; 503 504 trace_iomap_dio_rw_begin(iocb, iter, dio_flags, done_before); 505 506 if (!iomi.len) 507 return NULL; 508 509 dio = kmalloc(sizeof(*dio), GFP_KERNEL); 510 if (!dio) 511 return ERR_PTR(-ENOMEM); 512 513 dio->iocb = iocb; 514 atomic_set(&dio->ref, 1); 515 dio->size = 0; 516 dio->i_size = i_size_read(inode); 517 dio->dops = dops; 518 dio->error = 0; 519 dio->flags = 0; 520 dio->done_before = done_before; 521 522 dio->submit.iter = iter; 523 dio->submit.waiter = current; 524 dio->submit.poll_bio = NULL; 525 526 if (iocb->ki_flags & IOCB_NOWAIT) 527 iomi.flags |= IOMAP_NOWAIT; 528 529 if (iov_iter_rw(iter) == READ) { 530 if (iomi.pos >= dio->i_size) 531 goto out_free_dio; 532 533 if (user_backed_iter(iter)) 534 dio->flags |= IOMAP_DIO_DIRTY; 535 536 ret = kiocb_write_and_wait(iocb, iomi.len); 537 if (ret) 538 goto out_free_dio; 539 } else { 540 iomi.flags |= IOMAP_WRITE; 541 dio->flags |= IOMAP_DIO_WRITE; 542 543 if (dio_flags & IOMAP_DIO_OVERWRITE_ONLY) { 544 ret = -EAGAIN; 545 if (iomi.pos >= dio->i_size || 546 iomi.pos + iomi.len > dio->i_size) 547 goto out_free_dio; 548 iomi.flags |= IOMAP_OVERWRITE_ONLY; 549 } 550 551 /* for data sync or sync, we need sync completion processing */ 552 if (iocb_is_dsync(iocb)) { 553 dio->flags |= IOMAP_DIO_NEED_SYNC; 554 555 /* 556 * For datasync only writes, we optimistically try 557 * using FUA for this IO. Any non-FUA write that 558 * occurs will clear this flag, hence we know before 559 * completion whether a cache flush is necessary. 560 */ 561 if (!(iocb->ki_flags & IOCB_SYNC)) 562 dio->flags |= IOMAP_DIO_WRITE_FUA; 563 } 564 565 /* 566 * Try to invalidate cache pages for the range we are writing. 567 * If this invalidation fails, let the caller fall back to 568 * buffered I/O. 569 */ 570 ret = kiocb_invalidate_pages(iocb, iomi.len); 571 if (ret) { 572 if (ret != -EAGAIN) { 573 trace_iomap_dio_invalidate_fail(inode, iomi.pos, 574 iomi.len); 575 ret = -ENOTBLK; 576 } 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 trace_iomap_dio_rw_queued(inode, iomi.pos, iomi.len); 651 return ERR_PTR(-EIOCBQUEUED); 652 } 653 654 for (;;) { 655 set_current_state(TASK_UNINTERRUPTIBLE); 656 if (!READ_ONCE(dio->submit.waiter)) 657 break; 658 659 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, void *private, size_t done_before) 678 { 679 struct iomap_dio *dio; 680 681 dio = __iomap_dio_rw(iocb, iter, ops, dops, dio_flags, private, 682 done_before); 683 if (IS_ERR_OR_NULL(dio)) 684 return PTR_ERR_OR_ZERO(dio); 685 return iomap_dio_complete(dio); 686 } 687 EXPORT_SYMBOL_GPL(iomap_dio_rw); 688