fs/iomap/direct-io.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2010 Red Hat, Inc.
   4  * Copyright (c) 2016-2018 Christoph Hellwig.
   5  */
   6 #include <linux/module.h>
   7 #include <linux/compiler.h>
   8 #include <linux/fs.h>
   9 #include <linux/iomap.h>
  10 #include <linux/backing-dev.h>
  11 #include <linux/uio.h>
  12 #include <linux/task_io_accounting_ops.h>
  13
  14 #include "../internal.h"
  15
  16 /*
  17  * Private flags for iomap_dio, must not overlap with the public ones in
  18  * iomap.h:
  19  */
  20 #define IOMAP_DIO_WRITE_FUA     (1 << 28)
  21 #define IOMAP_DIO_NEED_SYNC     (1 << 29)
  22 #define IOMAP_DIO_WRITE         (1 << 30)
  23 #define IOMAP_DIO_DIRTY         (1 << 31)
  24
  25 struct iomap_dio {
  26         struct kiocb            *iocb;
  27         const struct iomap_dio_ops *dops;
  28         loff_t                  i_size;
  29         loff_t                  size;
  30         atomic_t                ref;
  31         unsigned                flags;
  32         int                     error;
  33         bool                    wait_for_completion;
  34
  35         union {
  36                 /* used during submission and for synchronous completion: */
  37                 struct {
  38                         struct iov_iter         *iter;
  39                         struct task_struct      *waiter;
  40                         struct request_queue    *last_queue;
  41                         blk_qc_t                cookie;
  42                 } submit;
  43
  44                 /* used for aio completion: */
  45                 struct {
  46                         struct work_struct      work;
  47                 } aio;
  48         };
  49 };
  50
  51 int iomap_dio_iopoll(struct kiocb *kiocb, bool spin)
  52 {
  53         struct request_queue *q = READ_ONCE(kiocb->private);
  54
  55         if (!q)
  56                 return 0;
  57         return blk_poll(q, READ_ONCE(kiocb->ki_cookie), spin);
  58 }
  59 EXPORT_SYMBOL_GPL(iomap_dio_iopoll);
  60
  61 static void iomap_dio_submit_bio(struct iomap_dio *dio, struct iomap *iomap,
  62                 struct bio *bio)
  63 {
  64         atomic_inc(&dio->ref);
  65
  66         if (dio->iocb->ki_flags & IOCB_HIPRI)
  67                 bio_set_polled(bio, dio->iocb);
  68
  69         dio->submit.last_queue = bdev_get_queue(iomap->bdev);
  70         dio->submit.cookie = submit_bio(bio);
  71 }
  72
  73 static ssize_t iomap_dio_complete(struct iomap_dio *dio)
  74 {
  75         const struct iomap_dio_ops *dops = dio->dops;
  76         struct kiocb *iocb = dio->iocb;
  77         struct inode *inode = file_inode(iocb->ki_filp);
  78         loff_t offset = iocb->ki_pos;
  79         ssize_t ret = dio->error;
  80
  81         if (dops && dops->end_io)
  82                 ret = dops->end_io(iocb, dio->size, ret, dio->flags);
  83
  84         if (likely(!ret)) {
  85                 ret = dio->size;
  86                 /* check for short read */
  87                 if (offset + ret > dio->i_size &&
  88                     !(dio->flags & IOMAP_DIO_WRITE))
  89                         ret = dio->i_size - offset;
  90                 iocb->ki_pos += ret;
  91         }
  92
  93         /*
  94          * Try again to invalidate clean pages which might have been cached by
  95          * non-direct readahead, or faulted in by get_user_pages() if the source
  96          * of the write was an mmap'ed region of the file we're writing.  Either
  97          * one is a pretty crazy thing to do, so we don't support it 100%.  If
  98          * this invalidation fails, tough, the write still worked...
  99          *
 100          * And this page cache invalidation has to be after ->end_io(), as some
 101          * filesystems convert unwritten extents to real allocations in
 102          * ->end_io() when necessary, otherwise a racing buffer read would cache
 103          * zeros from unwritten extents.
 104          */
 105         if (!dio->error &&
 106             (dio->flags & IOMAP_DIO_WRITE) && inode->i_mapping->nrpages) {
 107                 int err;
 108                 err = invalidate_inode_pages2_range(inode->i_mapping,
 109                                 offset >> PAGE_SHIFT,
 110                                 (offset + dio->size - 1) >> PAGE_SHIFT);
 111                 if (err)
 112                         dio_warn_stale_pagecache(iocb->ki_filp);
 113         }
 114
 115         /*
 116          * If this is a DSYNC write, make sure we push it to stable storage now
 117          * that we've written data.
 118          */
 119         if (ret > 0 && (dio->flags & IOMAP_DIO_NEED_SYNC))
 120                 ret = generic_write_sync(iocb, ret);
 121
 122         inode_dio_end(file_inode(iocb->ki_filp));
 123         kfree(dio);
 124
 125         return ret;
 126 }
 127
 128 static void iomap_dio_complete_work(struct work_struct *work)
 129 {
 130         struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work);
 131         struct kiocb *iocb = dio->iocb;
 132
 133         iocb->ki_complete(iocb, iomap_dio_complete(dio), 0);
 134 }
 135
 136 /*
 137  * Set an error in the dio if none is set yet.  We have to use cmpxchg
 138  * as the submission context and the completion context(s) can race to
 139  * update the error.
 140  */
 141 static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret)
 142 {
 143         cmpxchg(&dio->error, 0, ret);
 144 }
 145
 146 static void iomap_dio_bio_end_io(struct bio *bio)
 147 {
 148         struct iomap_dio *dio = bio->bi_private;
 149         bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
 150
 151         if (bio->bi_status)
 152                 iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
 153
 154         if (atomic_dec_and_test(&dio->ref)) {
 155                 if (dio->wait_for_completion) {
 156                         struct task_struct *waiter = dio->submit.waiter;
 157                         WRITE_ONCE(dio->submit.waiter, NULL);
 158                         blk_wake_io_task(waiter);
 159                 } else if (dio->flags & IOMAP_DIO_WRITE) {
 160                         struct inode *inode = file_inode(dio->iocb->ki_filp);
 161
 162                         INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
 163                         queue_work(inode->i_sb->s_dio_done_wq, &dio->aio.work);
 164                 } else {
 165                         iomap_dio_complete_work(&dio->aio.work);
 166                 }
 167         }
 168
 169         if (should_dirty) {
 170                 bio_check_pages_dirty(bio);
 171         } else {
 172                 bio_release_pages(bio, false);
 173                 bio_put(bio);
 174         }
 175 }
 176
 177 static void
 178 iomap_dio_zero(struct iomap_dio *dio, struct iomap *iomap, loff_t pos,
 179                 unsigned len)
 180 {
 181         struct page *page = ZERO_PAGE(0);
 182         int flags = REQ_SYNC | REQ_IDLE;
 183         struct bio *bio;
 184
 185         bio = bio_alloc(GFP_KERNEL, 1);
 186         bio_set_dev(bio, iomap->bdev);
 187         bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
 188         bio->bi_private = dio;
 189         bio->bi_end_io = iomap_dio_bio_end_io;
 190
 191         get_page(page);
 192         __bio_add_page(bio, page, len, 0);
 193         bio_set_op_attrs(bio, REQ_OP_WRITE, flags);
 194         iomap_dio_submit_bio(dio, iomap, bio);
 195 }
 196
 197 static loff_t
 198 iomap_dio_bio_actor(struct inode *inode, loff_t pos, loff_t length,
 199                 struct iomap_dio *dio, struct iomap *iomap)
 200 {
 201         unsigned int blkbits = blksize_bits(bdev_logical_block_size(iomap->bdev));
 202         unsigned int fs_block_size = i_blocksize(inode), pad;
 203         unsigned int align = iov_iter_alignment(dio->submit.iter);
 204         struct bio *bio;
 205         bool need_zeroout = false;
 206         bool use_fua = false;
 207         int nr_pages, ret = 0;
 208         size_t copied = 0;
 209         size_t orig_count;
 210
 211         if ((pos | length | align) & ((1 << blkbits) - 1))
 212                 return -EINVAL;
 213
 214         if (iomap->type == IOMAP_UNWRITTEN) {
 215                 dio->flags |= IOMAP_DIO_UNWRITTEN;
 216                 need_zeroout = true;
 217         }
 218
 219         if (iomap->flags & IOMAP_F_SHARED)
 220                 dio->flags |= IOMAP_DIO_COW;
 221
 222         if (iomap->flags & IOMAP_F_NEW) {
 223                 need_zeroout = true;
 224         } else if (iomap->type == IOMAP_MAPPED) {
 225                 /*
 226                  * Use a FUA write if we need datasync semantics, this is a pure
 227                  * data IO that doesn't require any metadata updates (including
 228                  * after IO completion such as unwritten extent conversion) and
 229                  * the underlying device supports FUA. This allows us to avoid
 230                  * cache flushes on IO completion.
 231                  */
 232                 if (!(iomap->flags & (IOMAP_F_SHARED|IOMAP_F_DIRTY)) &&
 233                     (dio->flags & IOMAP_DIO_WRITE_FUA) &&
 234                     blk_queue_fua(bdev_get_queue(iomap->bdev)))
 235                         use_fua = true;
 236         }
 237
 238         /*
 239          * Save the original count and trim the iter to just the extent we
 240          * are operating on right now.  The iter will be re-expanded once
 241          * we are done.
 242          */
 243         orig_count = iov_iter_count(dio->submit.iter);
 244         iov_iter_truncate(dio->submit.iter, length);
 245
 246         nr_pages = iov_iter_npages(dio->submit.iter, BIO_MAX_PAGES);
 247         if (nr_pages <= 0) {
 248                 ret = nr_pages;
 249                 goto out;
 250         }
 251
 252         if (need_zeroout) {
 253                 /* zero out from the start of the block to the write offset */
 254                 pad = pos & (fs_block_size - 1);
 255                 if (pad)
 256                         iomap_dio_zero(dio, iomap, pos - pad, pad);
 257         }
 258
 259         do {
 260                 size_t n;
 261                 if (dio->error) {
 262                         iov_iter_revert(dio->submit.iter, copied);
 263                         copied = ret = 0;
 264                         goto out;
 265                 }
 266
 267                 bio = bio_alloc(GFP_KERNEL, nr_pages);
 268                 bio_set_dev(bio, iomap->bdev);
 269                 bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
 270                 bio->bi_write_hint = dio->iocb->ki_hint;
 271                 bio->bi_ioprio = dio->iocb->ki_ioprio;
 272                 bio->bi_private = dio;
 273                 bio->bi_end_io = iomap_dio_bio_end_io;
 274
 275                 ret = bio_iov_iter_get_pages(bio, dio->submit.iter);
 276                 if (unlikely(ret)) {
 277                         /*
 278                          * We have to stop part way through an IO. We must fall
 279                          * through to the sub-block tail zeroing here, otherwise
 280                          * this short IO may expose stale data in the tail of
 281                          * the block we haven't written data to.
 282                          */
 283                         bio_put(bio);
 284                         goto zero_tail;
 285                 }
 286
 287                 n = bio->bi_iter.bi_size;
 288                 if (dio->flags & IOMAP_DIO_WRITE) {
 289                         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
 290                         if (use_fua)
 291                                 bio->bi_opf |= REQ_FUA;
 292                         else
 293                                 dio->flags &= ~IOMAP_DIO_WRITE_FUA;
 294                         task_io_account_write(n);
 295                 } else {
 296                         bio->bi_opf = REQ_OP_READ;
 297                         if (dio->flags & IOMAP_DIO_DIRTY)
 298                                 bio_set_pages_dirty(bio);
 299                 }
 300
 301                 dio->size += n;
 302                 pos += n;
 303                 copied += n;
 304
 305                 nr_pages = iov_iter_npages(dio->submit.iter, BIO_MAX_PAGES);
 306                 iomap_dio_submit_bio(dio, iomap, bio);
 307         } while (nr_pages);
 308
 309         /*
 310          * We need to zeroout the tail of a sub-block write if the extent type
 311          * requires zeroing or the write extends beyond EOF. If we don't zero
 312          * the block tail in the latter case, we can expose stale data via mmap
 313          * reads of the EOF block.
 314          */
 315 zero_tail:
 316         if (need_zeroout ||
 317             ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode))) {
 318                 /* zero out from the end of the write to the end of the block */
 319                 pad = pos & (fs_block_size - 1);
 320                 if (pad)
 321                         iomap_dio_zero(dio, iomap, pos, fs_block_size - pad);
 322         }
 323 out:
 324         /* Undo iter limitation to current extent */
 325         iov_iter_reexpand(dio->submit.iter, orig_count - copied);
 326         if (copied)
 327                 return copied;
 328         return ret;
 329 }
 330
 331 static loff_t
 332 iomap_dio_hole_actor(loff_t length, struct iomap_dio *dio)
 333 {
 334         length = iov_iter_zero(length, dio->submit.iter);
 335         dio->size += length;
 336         return length;
 337 }
 338
 339 static loff_t
 340 iomap_dio_inline_actor(struct inode *inode, loff_t pos, loff_t length,
 341                 struct iomap_dio *dio, struct iomap *iomap)
 342 {
 343         struct iov_iter *iter = dio->submit.iter;
 344         size_t copied;
 345
 346         BUG_ON(pos + length > PAGE_SIZE - offset_in_page(iomap->inline_data));
 347
 348         if (dio->flags & IOMAP_DIO_WRITE) {
 349                 loff_t size = inode->i_size;
 350
 351                 if (pos > size)
 352                         memset(iomap->inline_data + size, 0, pos - size);
 353                 copied = copy_from_iter(iomap->inline_data + pos, length, iter);
 354                 if (copied) {
 355                         if (pos + copied > size)
 356                                 i_size_write(inode, pos + copied);
 357                         mark_inode_dirty(inode);
 358                 }
 359         } else {
 360                 copied = copy_to_iter(iomap->inline_data + pos, length, iter);
 361         }
 362         dio->size += copied;
 363         return copied;
 364 }
 365
 366 static loff_t
 367 iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
 368                 void *data, struct iomap *iomap, struct iomap *srcmap)
 369 {
 370         struct iomap_dio *dio = data;
 371
 372         switch (iomap->type) {
 373         case IOMAP_HOLE:
 374                 if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
 375                         return -EIO;
 376                 return iomap_dio_hole_actor(length, dio);
 377         case IOMAP_UNWRITTEN:
 378                 if (!(dio->flags & IOMAP_DIO_WRITE))
 379                         return iomap_dio_hole_actor(length, dio);
 380                 return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
 381         case IOMAP_MAPPED:
 382                 return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
 383         case IOMAP_INLINE:
 384                 return iomap_dio_inline_actor(inode, pos, length, dio, iomap);
 385         default:
 386                 WARN_ON_ONCE(1);
 387                 return -EIO;
 388         }
 389 }
 390
 391 /*
 392  * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
 393  * is being issued as AIO or not.  This allows us to optimise pure data writes
 394  * to use REQ_FUA rather than requiring generic_write_sync() to issue a
 395  * REQ_FLUSH post write. This is slightly tricky because a single request here
 396  * can be mapped into multiple disjoint IOs and only a subset of the IOs issued
 397  * may be pure data writes. In that case, we still need to do a full data sync
 398  * completion.
 399  */
 400 ssize_t
 401 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 402                 const struct iomap_ops *ops, const struct iomap_dio_ops *dops,
 403                 bool wait_for_completion)
 404 {
 405         struct address_space *mapping = iocb->ki_filp->f_mapping;
 406         struct inode *inode = file_inode(iocb->ki_filp);
 407         size_t count = iov_iter_count(iter);
 408         loff_t pos = iocb->ki_pos;
 409         loff_t end = iocb->ki_pos + count - 1, ret = 0;
 410         unsigned int flags = IOMAP_DIRECT;
 411         struct blk_plug plug;
 412         struct iomap_dio *dio;
 413
 414         lockdep_assert_held(&inode->i_rwsem);
 415
 416         if (!count)
 417                 return 0;
 418
 419         if (WARN_ON(is_sync_kiocb(iocb) && !wait_for_completion))
 420                 return -EIO;
 421
 422         dio = kmalloc(sizeof(*dio), GFP_KERNEL);
 423         if (!dio)
 424                 return -ENOMEM;
 425
 426         dio->iocb = iocb;
 427         atomic_set(&dio->ref, 1);
 428         dio->size = 0;
 429         dio->i_size = i_size_read(inode);
 430         dio->dops = dops;
 431         dio->error = 0;
 432         dio->flags = 0;
 433
 434         dio->submit.iter = iter;
 435         dio->submit.waiter = current;
 436         dio->submit.cookie = BLK_QC_T_NONE;
 437         dio->submit.last_queue = NULL;
 438
 439         if (iov_iter_rw(iter) == READ) {
 440                 if (pos >= dio->i_size)
 441                         goto out_free_dio;
 442
 443                 if (iter_is_iovec(iter))
 444                         dio->flags |= IOMAP_DIO_DIRTY;
 445         } else {
 446                 flags |= IOMAP_WRITE;
 447                 dio->flags |= IOMAP_DIO_WRITE;
 448
 449                 /* for data sync or sync, we need sync completion processing */
 450                 if (iocb->ki_flags & IOCB_DSYNC)
 451                         dio->flags |= IOMAP_DIO_NEED_SYNC;
 452
 453                 /*
 454                  * For datasync only writes, we optimistically try using FUA for
 455                  * this IO.  Any non-FUA write that occurs will clear this flag,
 456                  * hence we know before completion whether a cache flush is
 457                  * necessary.
 458                  */
 459                 if ((iocb->ki_flags & (IOCB_DSYNC | IOCB_SYNC)) == IOCB_DSYNC)
 460                         dio->flags |= IOMAP_DIO_WRITE_FUA;
 461         }
 462
 463         if (iocb->ki_flags & IOCB_NOWAIT) {
 464                 if (filemap_range_has_page(mapping, pos, end)) {
 465                         ret = -EAGAIN;
 466                         goto out_free_dio;
 467                 }
 468                 flags |= IOMAP_NOWAIT;
 469         }
 470
 471         ret = filemap_write_and_wait_range(mapping, pos, end);
 472         if (ret)
 473                 goto out_free_dio;
 474
 475         /*
 476          * Try to invalidate cache pages for the range we're direct
 477          * writing.  If this invalidation fails, tough, the write will
 478          * still work, but racing two incompatible write paths is a
 479          * pretty crazy thing to do, so we don't support it 100%.
 480          */
 481         ret = invalidate_inode_pages2_range(mapping,
 482                         pos >> PAGE_SHIFT, end >> PAGE_SHIFT);
 483         if (ret)
 484                 dio_warn_stale_pagecache(iocb->ki_filp);
 485         ret = 0;
 486
 487         if (iov_iter_rw(iter) == WRITE && !wait_for_completion &&
 488             !inode->i_sb->s_dio_done_wq) {
 489                 ret = sb_init_dio_done_wq(inode->i_sb);
 490                 if (ret < 0)
 491                         goto out_free_dio;
 492         }
 493
 494         inode_dio_begin(inode);
 495
 496         blk_start_plug(&plug);
 497         do {
 498                 ret = iomap_apply(inode, pos, count, flags, ops, dio,
 499                                 iomap_dio_actor);
 500                 if (ret <= 0) {
 501                         /* magic error code to fall back to buffered I/O */
 502                         if (ret == -ENOTBLK) {
 503                                 wait_for_completion = true;
 504                                 ret = 0;
 505                         }
 506                         break;
 507                 }
 508                 pos += ret;
 509
 510                 if (iov_iter_rw(iter) == READ && pos >= dio->i_size) {
 511                         /*
 512                          * We only report that we've read data up to i_size.
 513                          * Revert iter to a state corresponding to that as
 514                          * some callers (such as splice code) rely on it.
 515                          */
 516                         iov_iter_revert(iter, pos - dio->i_size);
 517                         break;
 518                 }
 519         } while ((count = iov_iter_count(iter)) > 0);
 520         blk_finish_plug(&plug);
 521
 522         if (ret < 0)
 523                 iomap_dio_set_error(dio, ret);
 524
 525         /*
 526          * If all the writes we issued were FUA, we don't need to flush the
 527          * cache on IO completion. Clear the sync flag for this case.
 528          */
 529         if (dio->flags & IOMAP_DIO_WRITE_FUA)
 530                 dio->flags &= ~IOMAP_DIO_NEED_SYNC;
 531
 532         WRITE_ONCE(iocb->ki_cookie, dio->submit.cookie);
 533         WRITE_ONCE(iocb->private, dio->submit.last_queue);
 534
 535         /*
 536          * We are about to drop our additional submission reference, which
 537          * might be the last reference to the dio.  There are three different
 538          * ways we can progress here:
 539          *
 540          *  (a) If this is the last reference we will always complete and free
 541          *      the dio ourselves.
 542          *  (b) If this is not the last reference, and we serve an asynchronous
 543          *      iocb, we must never touch the dio after the decrement, the
 544          *      I/O completion handler will complete and free it.
 545          *  (c) If this is not the last reference, but we serve a synchronous
 546          *      iocb, the I/O completion handler will wake us up on the drop
 547          *      of the final reference, and we will complete and free it here
 548          *      after we got woken by the I/O completion handler.
 549          */
 550         dio->wait_for_completion = wait_for_completion;
 551         if (!atomic_dec_and_test(&dio->ref)) {
 552                 if (!wait_for_completion)
 553                         return -EIOCBQUEUED;
 554
 555                 for (;;) {
 556                         set_current_state(TASK_UNINTERRUPTIBLE);
 557                         if (!READ_ONCE(dio->submit.waiter))
 558                                 break;
 559
 560                         if (!(iocb->ki_flags & IOCB_HIPRI) ||
 561                             !dio->submit.last_queue ||
 562                             !blk_poll(dio->submit.last_queue,
 563                                          dio->submit.cookie, true))
 564                                 io_schedule();
 565                 }
 566                 __set_current_state(TASK_RUNNING);
 567         }
 568
 569         return iomap_dio_complete(dio);
 570
 571 out_free_dio:
 572         kfree(dio);
 573         return ret;
 574 }
 575 EXPORT_SYMBOL_GPL(iomap_dio_rw);