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, loff_t pos)
  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         if (dio->dops && dio->dops->submit_io)
  71                 dio->submit.cookie = dio->dops->submit_io(
  72                                 file_inode(dio->iocb->ki_filp),
  73                                 iomap, bio, pos);
  74         else
  75                 dio->submit.cookie = submit_bio(bio);
  76 }
  77
  78 static ssize_t iomap_dio_complete(struct iomap_dio *dio)
  79 {
  80         const struct iomap_dio_ops *dops = dio->dops;
  81         struct kiocb *iocb = dio->iocb;
  82         struct inode *inode = file_inode(iocb->ki_filp);
  83         loff_t offset = iocb->ki_pos;
  84         ssize_t ret = dio->error;
  85
  86         if (dops && dops->end_io)
  87                 ret = dops->end_io(iocb, dio->size, ret, dio->flags);
  88
  89         if (likely(!ret)) {
  90                 ret = dio->size;
  91                 /* check for short read */
  92                 if (offset + ret > dio->i_size &&
  93                     !(dio->flags & IOMAP_DIO_WRITE))
  94                         ret = dio->i_size - offset;
  95                 iocb->ki_pos += ret;
  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 &&
 111             (dio->flags & IOMAP_DIO_WRITE) && inode->i_mapping->nrpages) {
 112                 int err;
 113                 err = invalidate_inode_pages2_range(inode->i_mapping,
 114                                 offset >> PAGE_SHIFT,
 115                                 (offset + dio->size - 1) >> PAGE_SHIFT);
 116                 if (err)
 117                         dio_warn_stale_pagecache(iocb->ki_filp);
 118         }
 119
 120         /*
 121          * If this is a DSYNC write, make sure we push it to stable storage now
 122          * that we've written data.
 123          */
 124         if (ret > 0 && (dio->flags & IOMAP_DIO_NEED_SYNC))
 125                 ret = generic_write_sync(iocb, ret);
 126
 127         inode_dio_end(file_inode(iocb->ki_filp));
 128         kfree(dio);
 129
 130         return ret;
 131 }
 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), 0);
 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 static 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
 156         if (bio->bi_status)
 157                 iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
 158
 159         if (atomic_dec_and_test(&dio->ref)) {
 160                 if (dio->wait_for_completion) {
 161                         struct task_struct *waiter = dio->submit.waiter;
 162                         WRITE_ONCE(dio->submit.waiter, NULL);
 163                         blk_wake_io_task(waiter);
 164                 } else if (dio->flags & IOMAP_DIO_WRITE) {
 165                         struct inode *inode = file_inode(dio->iocb->ki_filp);
 166
 167                         INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
 168                         queue_work(inode->i_sb->s_dio_done_wq, &dio->aio.work);
 169                 } else {
 170                         iomap_dio_complete_work(&dio->aio.work);
 171                 }
 172         }
 173
 174         if (should_dirty) {
 175                 bio_check_pages_dirty(bio);
 176         } else {
 177                 bio_release_pages(bio, false);
 178                 bio_put(bio);
 179         }
 180 }
 181
 182 static void
 183 iomap_dio_zero(struct iomap_dio *dio, struct iomap *iomap, loff_t pos,
 184                 unsigned len)
 185 {
 186         struct page *page = ZERO_PAGE(0);
 187         int flags = REQ_SYNC | REQ_IDLE;
 188         struct bio *bio;
 189
 190         bio = bio_alloc(GFP_KERNEL, 1);
 191         bio_set_dev(bio, iomap->bdev);
 192         bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
 193         bio->bi_private = dio;
 194         bio->bi_end_io = iomap_dio_bio_end_io;
 195
 196         get_page(page);
 197         __bio_add_page(bio, page, len, 0);
 198         bio_set_op_attrs(bio, REQ_OP_WRITE, flags);
 199         iomap_dio_submit_bio(dio, iomap, bio, pos);
 200 }
 201
 202 static loff_t
 203 iomap_dio_bio_actor(struct inode *inode, loff_t pos, loff_t length,
 204                 struct iomap_dio *dio, struct iomap *iomap)
 205 {
 206         unsigned int blkbits = blksize_bits(bdev_logical_block_size(iomap->bdev));
 207         unsigned int fs_block_size = i_blocksize(inode), pad;
 208         unsigned int align = iov_iter_alignment(dio->submit.iter);
 209         struct bio *bio;
 210         bool need_zeroout = false;
 211         bool use_fua = false;
 212         int nr_pages, ret = 0;
 213         size_t copied = 0;
 214         size_t orig_count;
 215
 216         if ((pos | length | align) & ((1 << blkbits) - 1))
 217                 return -EINVAL;
 218
 219         if (iomap->type == IOMAP_UNWRITTEN) {
 220                 dio->flags |= IOMAP_DIO_UNWRITTEN;
 221                 need_zeroout = true;
 222         }
 223
 224         if (iomap->flags & IOMAP_F_SHARED)
 225                 dio->flags |= IOMAP_DIO_COW;
 226
 227         if (iomap->flags & IOMAP_F_NEW) {
 228                 need_zeroout = true;
 229         } else if (iomap->type == IOMAP_MAPPED) {
 230                 /*
 231                  * Use a FUA write if we need datasync semantics, this is a pure
 232                  * data IO that doesn't require any metadata updates (including
 233                  * after IO completion such as unwritten extent conversion) and
 234                  * the underlying device supports FUA. This allows us to avoid
 235                  * cache flushes on IO completion.
 236                  */
 237                 if (!(iomap->flags & (IOMAP_F_SHARED|IOMAP_F_DIRTY)) &&
 238                     (dio->flags & IOMAP_DIO_WRITE_FUA) &&
 239                     blk_queue_fua(bdev_get_queue(iomap->bdev)))
 240                         use_fua = true;
 241         }
 242
 243         /*
 244          * Save the original count and trim the iter to just the extent we
 245          * are operating on right now.  The iter will be re-expanded once
 246          * we are done.
 247          */
 248         orig_count = iov_iter_count(dio->submit.iter);
 249         iov_iter_truncate(dio->submit.iter, length);
 250
 251         nr_pages = iov_iter_npages(dio->submit.iter, BIO_MAX_PAGES);
 252         if (nr_pages <= 0) {
 253                 ret = nr_pages;
 254                 goto out;
 255         }
 256
 257         if (need_zeroout) {
 258                 /* zero out from the start of the block to the write offset */
 259                 pad = pos & (fs_block_size - 1);
 260                 if (pad)
 261                         iomap_dio_zero(dio, iomap, pos - pad, pad);
 262         }
 263
 264         do {
 265                 size_t n;
 266                 if (dio->error) {
 267                         iov_iter_revert(dio->submit.iter, copied);
 268                         copied = ret = 0;
 269                         goto out;
 270                 }
 271
 272                 bio = bio_alloc(GFP_KERNEL, nr_pages);
 273                 bio_set_dev(bio, iomap->bdev);
 274                 bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
 275                 bio->bi_write_hint = dio->iocb->ki_hint;
 276                 bio->bi_ioprio = dio->iocb->ki_ioprio;
 277                 bio->bi_private = dio;
 278                 bio->bi_end_io = iomap_dio_bio_end_io;
 279
 280                 ret = bio_iov_iter_get_pages(bio, dio->submit.iter);
 281                 if (unlikely(ret)) {
 282                         /*
 283                          * We have to stop part way through an IO. We must fall
 284                          * through to the sub-block tail zeroing here, otherwise
 285                          * this short IO may expose stale data in the tail of
 286                          * the block we haven't written data to.
 287                          */
 288                         bio_put(bio);
 289                         goto zero_tail;
 290                 }
 291
 292                 n = bio->bi_iter.bi_size;
 293                 if (dio->flags & IOMAP_DIO_WRITE) {
 294                         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
 295                         if (use_fua)
 296                                 bio->bi_opf |= REQ_FUA;
 297                         else
 298                                 dio->flags &= ~IOMAP_DIO_WRITE_FUA;
 299                         task_io_account_write(n);
 300                 } else {
 301                         bio->bi_opf = REQ_OP_READ;
 302                         if (dio->flags & IOMAP_DIO_DIRTY)
 303                                 bio_set_pages_dirty(bio);
 304                 }
 305
 306                 dio->size += n;
 307                 copied += n;
 308
 309                 nr_pages = iov_iter_npages(dio->submit.iter, BIO_MAX_PAGES);
 310                 iomap_dio_submit_bio(dio, iomap, bio, pos);
 311                 pos += n;
 312         } while (nr_pages);
 313
 314         /*
 315          * We need to zeroout the tail of a sub-block write if the extent type
 316          * requires zeroing or the write extends beyond EOF. If we don't zero
 317          * the block tail in the latter case, we can expose stale data via mmap
 318          * reads of the EOF block.
 319          */
 320 zero_tail:
 321         if (need_zeroout ||
 322             ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode))) {
 323                 /* zero out from the end of the write to the end of the block */
 324                 pad = pos & (fs_block_size - 1);
 325                 if (pad)
 326                         iomap_dio_zero(dio, iomap, pos, fs_block_size - pad);
 327         }
 328 out:
 329         /* Undo iter limitation to current extent */
 330         iov_iter_reexpand(dio->submit.iter, orig_count - copied);
 331         if (copied)
 332                 return copied;
 333         return ret;
 334 }
 335
 336 static loff_t
 337 iomap_dio_hole_actor(loff_t length, struct iomap_dio *dio)
 338 {
 339         length = iov_iter_zero(length, dio->submit.iter);
 340         dio->size += length;
 341         return length;
 342 }
 343
 344 static loff_t
 345 iomap_dio_inline_actor(struct inode *inode, loff_t pos, loff_t length,
 346                 struct iomap_dio *dio, struct iomap *iomap)
 347 {
 348         struct iov_iter *iter = dio->submit.iter;
 349         size_t copied;
 350
 351         BUG_ON(pos + length > PAGE_SIZE - offset_in_page(iomap->inline_data));
 352
 353         if (dio->flags & IOMAP_DIO_WRITE) {
 354                 loff_t size = inode->i_size;
 355
 356                 if (pos > size)
 357                         memset(iomap->inline_data + size, 0, pos - size);
 358                 copied = copy_from_iter(iomap->inline_data + pos, length, iter);
 359                 if (copied) {
 360                         if (pos + copied > size)
 361                                 i_size_write(inode, pos + copied);
 362                         mark_inode_dirty(inode);
 363                 }
 364         } else {
 365                 copied = copy_to_iter(iomap->inline_data + pos, length, iter);
 366         }
 367         dio->size += copied;
 368         return copied;
 369 }
 370
 371 static loff_t
 372 iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
 373                 void *data, struct iomap *iomap, struct iomap *srcmap)
 374 {
 375         struct iomap_dio *dio = data;
 376
 377         switch (iomap->type) {
 378         case IOMAP_HOLE:
 379                 if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
 380                         return -EIO;
 381                 return iomap_dio_hole_actor(length, dio);
 382         case IOMAP_UNWRITTEN:
 383                 if (!(dio->flags & IOMAP_DIO_WRITE))
 384                         return iomap_dio_hole_actor(length, dio);
 385                 return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
 386         case IOMAP_MAPPED:
 387                 return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
 388         case IOMAP_INLINE:
 389                 return iomap_dio_inline_actor(inode, pos, length, dio, iomap);
 390         default:
 391                 WARN_ON_ONCE(1);
 392                 return -EIO;
 393         }
 394 }
 395
 396 /*
 397  * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
 398  * is being issued as AIO or not.  This allows us to optimise pure data writes
 399  * to use REQ_FUA rather than requiring generic_write_sync() to issue a
 400  * REQ_FLUSH post write. This is slightly tricky because a single request here
 401  * can be mapped into multiple disjoint IOs and only a subset of the IOs issued
 402  * may be pure data writes. In that case, we still need to do a full data sync
 403  * completion.
 404  */
 405 ssize_t
 406 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 407                 const struct iomap_ops *ops, const struct iomap_dio_ops *dops,
 408                 bool wait_for_completion)
 409 {
 410         struct address_space *mapping = iocb->ki_filp->f_mapping;
 411         struct inode *inode = file_inode(iocb->ki_filp);
 412         size_t count = iov_iter_count(iter);
 413         loff_t pos = iocb->ki_pos;
 414         loff_t end = iocb->ki_pos + count - 1, ret = 0;
 415         unsigned int flags = IOMAP_DIRECT;
 416         struct blk_plug plug;
 417         struct iomap_dio *dio;
 418
 419         if (!count)
 420                 return 0;
 421
 422         if (WARN_ON(is_sync_kiocb(iocb) && !wait_for_completion))
 423                 return -EIO;
 424
 425         dio = kmalloc(sizeof(*dio), GFP_KERNEL);
 426         if (!dio)
 427                 return -ENOMEM;
 428
 429         dio->iocb = iocb;
 430         atomic_set(&dio->ref, 1);
 431         dio->size = 0;
 432         dio->i_size = i_size_read(inode);
 433         dio->dops = dops;
 434         dio->error = 0;
 435         dio->flags = 0;
 436
 437         dio->submit.iter = iter;
 438         dio->submit.waiter = current;
 439         dio->submit.cookie = BLK_QC_T_NONE;
 440         dio->submit.last_queue = NULL;
 441
 442         if (iov_iter_rw(iter) == READ) {
 443                 if (pos >= dio->i_size)
 444                         goto out_free_dio;
 445
 446                 if (iter_is_iovec(iter))
 447                         dio->flags |= IOMAP_DIO_DIRTY;
 448         } else {
 449                 flags |= IOMAP_WRITE;
 450                 dio->flags |= IOMAP_DIO_WRITE;
 451
 452                 /* for data sync or sync, we need sync completion processing */
 453                 if (iocb->ki_flags & IOCB_DSYNC)
 454                         dio->flags |= IOMAP_DIO_NEED_SYNC;
 455
 456                 /*
 457                  * For datasync only writes, we optimistically try using FUA for
 458                  * this IO.  Any non-FUA write that occurs will clear this flag,
 459                  * hence we know before completion whether a cache flush is
 460                  * necessary.
 461                  */
 462                 if ((iocb->ki_flags & (IOCB_DSYNC | IOCB_SYNC)) == IOCB_DSYNC)
 463                         dio->flags |= IOMAP_DIO_WRITE_FUA;
 464         }
 465
 466         if (iocb->ki_flags & IOCB_NOWAIT) {
 467                 if (filemap_range_has_page(mapping, pos, end)) {
 468                         ret = -EAGAIN;
 469                         goto out_free_dio;
 470                 }
 471                 flags |= IOMAP_NOWAIT;
 472         }
 473
 474         ret = filemap_write_and_wait_range(mapping, pos, end);
 475         if (ret)
 476                 goto out_free_dio;
 477
 478         /*
 479          * Try to invalidate cache pages for the range we're direct
 480          * writing.  If this invalidation fails, tough, the write will
 481          * still work, but racing two incompatible write paths is a
 482          * pretty crazy thing to do, so we don't support it 100%.
 483          */
 484         ret = invalidate_inode_pages2_range(mapping,
 485                         pos >> PAGE_SHIFT, end >> PAGE_SHIFT);
 486         if (ret)
 487                 dio_warn_stale_pagecache(iocb->ki_filp);
 488         ret = 0;
 489
 490         if (iov_iter_rw(iter) == WRITE && !wait_for_completion &&
 491             !inode->i_sb->s_dio_done_wq) {
 492                 ret = sb_init_dio_done_wq(inode->i_sb);
 493                 if (ret < 0)
 494                         goto out_free_dio;
 495         }
 496
 497         inode_dio_begin(inode);
 498
 499         blk_start_plug(&plug);
 500         do {
 501                 ret = iomap_apply(inode, pos, count, flags, ops, dio,
 502                                 iomap_dio_actor);
 503                 if (ret <= 0) {
 504                         /* magic error code to fall back to buffered I/O */
 505                         if (ret == -ENOTBLK) {
 506                                 wait_for_completion = true;
 507                                 ret = 0;
 508                         }
 509                         break;
 510                 }
 511                 pos += ret;
 512
 513                 if (iov_iter_rw(iter) == READ && pos >= dio->i_size) {
 514                         /*
 515                          * We only report that we've read data up to i_size.
 516                          * Revert iter to a state corresponding to that as
 517                          * some callers (such as splice code) rely on it.
 518                          */
 519                         iov_iter_revert(iter, pos - dio->i_size);
 520                         break;
 521                 }
 522         } while ((count = iov_iter_count(iter)) > 0);
 523         blk_finish_plug(&plug);
 524
 525         if (ret < 0)
 526                 iomap_dio_set_error(dio, ret);
 527
 528         /*
 529          * If all the writes we issued were FUA, we don't need to flush the
 530          * cache on IO completion. Clear the sync flag for this case.
 531          */
 532         if (dio->flags & IOMAP_DIO_WRITE_FUA)
 533                 dio->flags &= ~IOMAP_DIO_NEED_SYNC;
 534
 535         WRITE_ONCE(iocb->ki_cookie, dio->submit.cookie);
 536         WRITE_ONCE(iocb->private, dio->submit.last_queue);
 537
 538         /*
 539          * We are about to drop our additional submission reference, which
 540          * might be the last reference to the dio.  There are three different
 541          * ways we can progress here:
 542          *
 543          *  (a) If this is the last reference we will always complete and free
 544          *      the dio ourselves.
 545          *  (b) If this is not the last reference, and we serve an asynchronous
 546          *      iocb, we must never touch the dio after the decrement, the
 547          *      I/O completion handler will complete and free it.
 548          *  (c) If this is not the last reference, but we serve a synchronous
 549          *      iocb, the I/O completion handler will wake us up on the drop
 550          *      of the final reference, and we will complete and free it here
 551          *      after we got woken by the I/O completion handler.
 552          */
 553         dio->wait_for_completion = wait_for_completion;
 554         if (!atomic_dec_and_test(&dio->ref)) {
 555                 if (!wait_for_completion)
 556                         return -EIOCBQUEUED;
 557
 558                 for (;;) {
 559                         set_current_state(TASK_UNINTERRUPTIBLE);
 560                         if (!READ_ONCE(dio->submit.waiter))
 561                                 break;
 562
 563                         if (!(iocb->ki_flags & IOCB_HIPRI) ||
 564                             !dio->submit.last_queue ||
 565                             !blk_poll(dio->submit.last_queue,
 566                                          dio->submit.cookie, true))
 567                                 blk_io_schedule();
 568                 }
 569                 __set_current_state(TASK_RUNNING);
 570         }
 571
 572         return iomap_dio_complete(dio);
 573
 574 out_free_dio:
 575         kfree(dio);
 576         return ret;
 577 }
 578 EXPORT_SYMBOL_GPL(iomap_dio_rw);