fs/iomap/direct-io.c

   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_CALLER_COMP   (1U << 26)
  24 #define IOMAP_DIO_INLINE_COMP   (1U << 27)
  25 #define IOMAP_DIO_WRITE_THROUGH (1U << 28)
  26 #define IOMAP_DIO_NEED_SYNC     (1U << 29)
  27 #define IOMAP_DIO_WRITE         (1U << 30)
  28 #define IOMAP_DIO_DIRTY         (1U << 31)
  29
  30 /*
  31  * Used for sub block zeroing in iomap_dio_zero()
  32  */
  33 #define IOMAP_ZERO_PAGE_SIZE (SZ_64K)
  34 #define IOMAP_ZERO_PAGE_ORDER (get_order(IOMAP_ZERO_PAGE_SIZE))
  35 static struct page *zero_page;
  36
  37 struct iomap_dio {
  38         struct kiocb            *iocb;
  39         const struct iomap_dio_ops *dops;
  40         loff_t                  i_size;
  41         loff_t                  size;
  42         atomic_t                ref;
  43         unsigned                flags;
  44         int                     error;
  45         size_t                  done_before;
  46         bool                    wait_for_completion;
  47
  48         union {
  49                 /* used during submission and for synchronous completion: */
  50                 struct {
  51                         struct iov_iter         *iter;
  52                         struct task_struct      *waiter;
  53                 } submit;
  54
  55                 /* used for aio completion: */
  56                 struct {
  57                         struct work_struct      work;
  58                 } aio;
  59         };
  60 };
  61
  62 static struct bio *iomap_dio_alloc_bio(const struct iomap_iter *iter,
  63                 struct iomap_dio *dio, unsigned short nr_vecs, blk_opf_t opf)
  64 {
  65         if (dio->dops && dio->dops->bio_set)
  66                 return bio_alloc_bioset(iter->iomap.bdev, nr_vecs, opf,
  67                                         GFP_KERNEL, dio->dops->bio_set);
  68         return bio_alloc(iter->iomap.bdev, nr_vecs, opf, GFP_KERNEL);
  69 }
  70
  71 static void iomap_dio_submit_bio(const struct iomap_iter *iter,
  72                 struct iomap_dio *dio, struct bio *bio, loff_t pos)
  73 {
  74         struct kiocb *iocb = dio->iocb;
  75
  76         atomic_inc(&dio->ref);
  77
  78         /* Sync dio can't be polled reliably */
  79         if ((iocb->ki_flags & IOCB_HIPRI) && !is_sync_kiocb(iocb)) {
  80                 bio_set_polled(bio, iocb);
  81                 WRITE_ONCE(iocb->private, bio);
  82         }
  83
  84         if (dio->dops && dio->dops->submit_io)
  85                 dio->dops->submit_io(iter, bio, pos);
  86         else
  87                 submit_bio(bio);
  88 }
  89
  90 ssize_t iomap_dio_complete(struct iomap_dio *dio)
  91 {
  92         const struct iomap_dio_ops *dops = dio->dops;
  93         struct kiocb *iocb = dio->iocb;
  94         loff_t offset = iocb->ki_pos;
  95         ssize_t ret = dio->error;
  96
  97         if (dops && dops->end_io)
  98                 ret = dops->end_io(iocb, dio->size, ret, dio->flags);
  99
 100         if (likely(!ret)) {
 101                 ret = dio->size;
 102                 /* check for short read */
 103                 if (offset + ret > dio->i_size &&
 104                     !(dio->flags & IOMAP_DIO_WRITE))
 105                         ret = dio->i_size - offset;
 106         }
 107
 108         /*
 109          * Try again to invalidate clean pages which might have been cached by
 110          * non-direct readahead, or faulted in by get_user_pages() if the source
 111          * of the write was an mmap'ed region of the file we're writing.  Either
 112          * one is a pretty crazy thing to do, so we don't support it 100%.  If
 113          * this invalidation fails, tough, the write still worked...
 114          *
 115          * And this page cache invalidation has to be after ->end_io(), as some
 116          * filesystems convert unwritten extents to real allocations in
 117          * ->end_io() when necessary, otherwise a racing buffer read would cache
 118          * zeros from unwritten extents.
 119          */
 120         if (!dio->error && dio->size && (dio->flags & IOMAP_DIO_WRITE))
 121                 kiocb_invalidate_post_direct_write(iocb, dio->size);
 122
 123         inode_dio_end(file_inode(iocb->ki_filp));
 124
 125         if (ret > 0) {
 126                 iocb->ki_pos += ret;
 127
 128                 /*
 129                  * If this is a DSYNC write, make sure we push it to stable
 130                  * storage now that we've written data.
 131                  */
 132                 if (dio->flags & IOMAP_DIO_NEED_SYNC)
 133                         ret = generic_write_sync(iocb, ret);
 134                 if (ret > 0)
 135                         ret += dio->done_before;
 136         }
 137         trace_iomap_dio_complete(iocb, dio->error, ret);
 138         kfree(dio);
 139         return ret;
 140 }
 141 EXPORT_SYMBOL_GPL(iomap_dio_complete);
 142
 143 static ssize_t iomap_dio_deferred_complete(void *data)
 144 {
 145         return iomap_dio_complete(data);
 146 }
 147
 148 static void iomap_dio_complete_work(struct work_struct *work)
 149 {
 150         struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work);
 151         struct kiocb *iocb = dio->iocb;
 152
 153         iocb->ki_complete(iocb, iomap_dio_complete(dio));
 154 }
 155
 156 /*
 157  * Set an error in the dio if none is set yet.  We have to use cmpxchg
 158  * as the submission context and the completion context(s) can race to
 159  * update the error.
 160  */
 161 static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret)
 162 {
 163         cmpxchg(&dio->error, 0, ret);
 164 }
 165
 166 void iomap_dio_bio_end_io(struct bio *bio)
 167 {
 168         struct iomap_dio *dio = bio->bi_private;
 169         bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
 170         struct kiocb *iocb = dio->iocb;
 171
 172         if (bio->bi_status)
 173                 iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
 174         if (!atomic_dec_and_test(&dio->ref))
 175                 goto release_bio;
 176
 177         /*
 178          * Synchronous dio, task itself will handle any completion work
 179          * that needs after IO. All we need to do is wake the task.
 180          */
 181         if (dio->wait_for_completion) {
 182                 struct task_struct *waiter = dio->submit.waiter;
 183
 184                 WRITE_ONCE(dio->submit.waiter, NULL);
 185                 blk_wake_io_task(waiter);
 186                 goto release_bio;
 187         }
 188
 189         /*
 190          * Flagged with IOMAP_DIO_INLINE_COMP, we can complete it inline
 191          */
 192         if (dio->flags & IOMAP_DIO_INLINE_COMP) {
 193                 WRITE_ONCE(iocb->private, NULL);
 194                 iomap_dio_complete_work(&dio->aio.work);
 195                 goto release_bio;
 196         }
 197
 198         /*
 199          * If this dio is flagged with IOMAP_DIO_CALLER_COMP, then schedule
 200          * our completion that way to avoid an async punt to a workqueue.
 201          */
 202         if (dio->flags & IOMAP_DIO_CALLER_COMP) {
 203                 /* only polled IO cares about private cleared */
 204                 iocb->private = dio;
 205                 iocb->dio_complete = iomap_dio_deferred_complete;
 206
 207                 /*
 208                  * Invoke ->ki_complete() directly. We've assigned our
 209                  * dio_complete callback handler, and since the issuer set
 210                  * IOCB_DIO_CALLER_COMP, we know their ki_complete handler will
 211                  * notice ->dio_complete being set and will defer calling that
 212                  * handler until it can be done from a safe task context.
 213                  *
 214                  * Note that the 'res' being passed in here is not important
 215                  * for this case. The actual completion value of the request
 216                  * will be gotten from dio_complete when that is run by the
 217                  * issuer.
 218                  */
 219                 iocb->ki_complete(iocb, 0);
 220                 goto release_bio;
 221         }
 222
 223         /*
 224          * Async DIO completion that requires filesystem level completion work
 225          * gets punted to a work queue to complete as the operation may require
 226          * more IO to be issued to finalise filesystem metadata changes or
 227          * guarantee data integrity.
 228          */
 229         INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
 230         queue_work(file_inode(iocb->ki_filp)->i_sb->s_dio_done_wq,
 231                         &dio->aio.work);
 232 release_bio:
 233         if (should_dirty) {
 234                 bio_check_pages_dirty(bio);
 235         } else {
 236                 bio_release_pages(bio, false);
 237                 bio_put(bio);
 238         }
 239 }
 240 EXPORT_SYMBOL_GPL(iomap_dio_bio_end_io);
 241
 242 static int iomap_dio_zero(const struct iomap_iter *iter, struct iomap_dio *dio,
 243                 loff_t pos, unsigned len)
 244 {
 245         struct inode *inode = file_inode(dio->iocb->ki_filp);
 246         struct bio *bio;
 247
 248         if (!len)
 249                 return 0;
 250         /*
 251          * Max block size supported is 64k
 252          */
 253         if (WARN_ON_ONCE(len > IOMAP_ZERO_PAGE_SIZE))
 254                 return -EINVAL;
 255
 256         bio = iomap_dio_alloc_bio(iter, dio, 1, REQ_OP_WRITE | REQ_SYNC | REQ_IDLE);
 257         fscrypt_set_bio_crypt_ctx(bio, inode, pos >> inode->i_blkbits,
 258                                   GFP_KERNEL);
 259         bio->bi_iter.bi_sector = iomap_sector(&iter->iomap, pos);
 260         bio->bi_private = dio;
 261         bio->bi_end_io = iomap_dio_bio_end_io;
 262
 263         __bio_add_page(bio, zero_page, len, 0);
 264         iomap_dio_submit_bio(iter, dio, bio, pos);
 265         return 0;
 266 }
 267
 268 /*
 269  * Figure out the bio's operation flags from the dio request, the
 270  * mapping, and whether or not we want FUA.  Note that we can end up
 271  * clearing the WRITE_THROUGH flag in the dio request.
 272  */
 273 static inline blk_opf_t iomap_dio_bio_opflags(struct iomap_dio *dio,
 274                 const struct iomap *iomap, bool use_fua, bool atomic)
 275 {
 276         blk_opf_t opflags = REQ_SYNC | REQ_IDLE;
 277
 278         if (!(dio->flags & IOMAP_DIO_WRITE))
 279                 return REQ_OP_READ;
 280
 281         opflags |= REQ_OP_WRITE;
 282         if (use_fua)
 283                 opflags |= REQ_FUA;
 284         else
 285                 dio->flags &= ~IOMAP_DIO_WRITE_THROUGH;
 286         if (atomic)
 287                 opflags |= REQ_ATOMIC;
 288
 289         return opflags;
 290 }
 291
 292 static loff_t iomap_dio_bio_iter(const struct iomap_iter *iter,
 293                 struct iomap_dio *dio)
 294 {
 295         const struct iomap *iomap = &iter->iomap;
 296         struct inode *inode = iter->inode;
 297         unsigned int fs_block_size = i_blocksize(inode), pad;
 298         const loff_t length = iomap_length(iter);
 299         bool atomic = iter->flags & IOMAP_ATOMIC;
 300         loff_t pos = iter->pos;
 301         blk_opf_t bio_opf;
 302         struct bio *bio;
 303         bool need_zeroout = false;
 304         bool use_fua = false;
 305         int nr_pages, ret = 0;
 306         size_t copied = 0;
 307         size_t orig_count;
 308
 309         if (atomic && length != fs_block_size)
 310                 return -EINVAL;
 311
 312         if ((pos | length) & (bdev_logical_block_size(iomap->bdev) - 1) ||
 313             !bdev_iter_is_aligned(iomap->bdev, dio->submit.iter))
 314                 return -EINVAL;
 315
 316         if (iomap->type == IOMAP_UNWRITTEN) {
 317                 dio->flags |= IOMAP_DIO_UNWRITTEN;
 318                 need_zeroout = true;
 319         }
 320
 321         if (iomap->flags & IOMAP_F_SHARED)
 322                 dio->flags |= IOMAP_DIO_COW;
 323
 324         if (iomap->flags & IOMAP_F_NEW) {
 325                 need_zeroout = true;
 326         } else if (iomap->type == IOMAP_MAPPED) {
 327                 /*
 328                  * Use a FUA write if we need datasync semantics, this is a pure
 329                  * data IO that doesn't require any metadata updates (including
 330                  * after IO completion such as unwritten extent conversion) and
 331                  * the underlying device either supports FUA or doesn't have
 332                  * a volatile write cache. This allows us to avoid cache flushes
 333                  * on IO completion. If we can't use writethrough and need to
 334                  * sync, disable in-task completions as dio completion will
 335                  * need to call generic_write_sync() which will do a blocking
 336                  * fsync / cache flush call.
 337                  */
 338                 if (!(iomap->flags & (IOMAP_F_SHARED|IOMAP_F_DIRTY)) &&
 339                     (dio->flags & IOMAP_DIO_WRITE_THROUGH) &&
 340                     (bdev_fua(iomap->bdev) || !bdev_write_cache(iomap->bdev)))
 341                         use_fua = true;
 342                 else if (dio->flags & IOMAP_DIO_NEED_SYNC)
 343                         dio->flags &= ~IOMAP_DIO_CALLER_COMP;
 344         }
 345
 346         /*
 347          * Save the original count and trim the iter to just the extent we
 348          * are operating on right now.  The iter will be re-expanded once
 349          * we are done.
 350          */
 351         orig_count = iov_iter_count(dio->submit.iter);
 352         iov_iter_truncate(dio->submit.iter, length);
 353
 354         if (!iov_iter_count(dio->submit.iter))
 355                 goto out;
 356
 357         /*
 358          * We can only do deferred completion for pure overwrites that
 359          * don't require additional IO at completion. This rules out
 360          * writes that need zeroing or extent conversion, extend
 361          * the file size, or issue journal IO or cache flushes
 362          * during completion processing.
 363          */
 364         if (need_zeroout ||
 365             ((dio->flags & IOMAP_DIO_NEED_SYNC) && !use_fua) ||
 366             ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode)))
 367                 dio->flags &= ~IOMAP_DIO_CALLER_COMP;
 368
 369         /*
 370          * The rules for polled IO completions follow the guidelines as the
 371          * ones we set for inline and deferred completions. If none of those
 372          * are available for this IO, clear the polled flag.
 373          */
 374         if (!(dio->flags & (IOMAP_DIO_INLINE_COMP|IOMAP_DIO_CALLER_COMP)))
 375                 dio->iocb->ki_flags &= ~IOCB_HIPRI;
 376
 377         if (need_zeroout) {
 378                 /* zero out from the start of the block to the write offset */
 379                 pad = pos & (fs_block_size - 1);
 380
 381                 ret = iomap_dio_zero(iter, dio, pos - pad, pad);
 382                 if (ret)
 383                         goto out;
 384         }
 385
 386         bio_opf = iomap_dio_bio_opflags(dio, iomap, use_fua, atomic);
 387
 388         nr_pages = bio_iov_vecs_to_alloc(dio->submit.iter, BIO_MAX_VECS);
 389         do {
 390                 size_t n;
 391                 if (dio->error) {
 392                         iov_iter_revert(dio->submit.iter, copied);
 393                         copied = ret = 0;
 394                         goto out;
 395                 }
 396
 397                 bio = iomap_dio_alloc_bio(iter, dio, nr_pages, bio_opf);
 398                 fscrypt_set_bio_crypt_ctx(bio, inode, pos >> inode->i_blkbits,
 399                                           GFP_KERNEL);
 400                 bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
 401                 bio->bi_write_hint = inode->i_write_hint;
 402                 bio->bi_ioprio = dio->iocb->ki_ioprio;
 403                 bio->bi_private = dio;
 404                 bio->bi_end_io = iomap_dio_bio_end_io;
 405
 406                 ret = bio_iov_iter_get_pages(bio, dio->submit.iter);
 407                 if (unlikely(ret)) {
 408                         /*
 409                          * We have to stop part way through an IO. We must fall
 410                          * through to the sub-block tail zeroing here, otherwise
 411                          * this short IO may expose stale data in the tail of
 412                          * the block we haven't written data to.
 413                          */
 414                         bio_put(bio);
 415                         goto zero_tail;
 416                 }
 417
 418                 n = bio->bi_iter.bi_size;
 419                 if (WARN_ON_ONCE(atomic && n != length)) {
 420                         /*
 421                          * This bio should have covered the complete length,
 422                          * which it doesn't, so error. We may need to zero out
 423                          * the tail (complete FS block), similar to when
 424                          * bio_iov_iter_get_pages() returns an error, above.
 425                          */
 426                         ret = -EINVAL;
 427                         bio_put(bio);
 428                         goto zero_tail;
 429                 }
 430                 if (dio->flags & IOMAP_DIO_WRITE) {
 431                         task_io_account_write(n);
 432                 } else {
 433                         if (dio->flags & IOMAP_DIO_DIRTY)
 434                                 bio_set_pages_dirty(bio);
 435                 }
 436
 437                 dio->size += n;
 438                 copied += n;
 439
 440                 nr_pages = bio_iov_vecs_to_alloc(dio->submit.iter,
 441                                                  BIO_MAX_VECS);
 442                 /*
 443                  * We can only poll for single bio I/Os.
 444                  */
 445                 if (nr_pages)
 446                         dio->iocb->ki_flags &= ~IOCB_HIPRI;
 447                 iomap_dio_submit_bio(iter, dio, bio, pos);
 448                 pos += n;
 449         } while (nr_pages);
 450
 451         /*
 452          * We need to zeroout the tail of a sub-block write if the extent type
 453          * requires zeroing or the write extends beyond EOF. If we don't zero
 454          * the block tail in the latter case, we can expose stale data via mmap
 455          * reads of the EOF block.
 456          */
 457 zero_tail:
 458         if (need_zeroout ||
 459             ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode))) {
 460                 /* zero out from the end of the write to the end of the block */
 461                 pad = pos & (fs_block_size - 1);
 462                 if (pad)
 463                         ret = iomap_dio_zero(iter, dio, pos,
 464                                              fs_block_size - pad);
 465         }
 466 out:
 467         /* Undo iter limitation to current extent */
 468         iov_iter_reexpand(dio->submit.iter, orig_count - copied);
 469         if (copied)
 470                 return copied;
 471         return ret;
 472 }
 473
 474 static loff_t iomap_dio_hole_iter(const struct iomap_iter *iter,
 475                 struct iomap_dio *dio)
 476 {
 477         loff_t length = iov_iter_zero(iomap_length(iter), dio->submit.iter);
 478
 479         dio->size += length;
 480         if (!length)
 481                 return -EFAULT;
 482         return length;
 483 }
 484
 485 static loff_t iomap_dio_inline_iter(const struct iomap_iter *iomi,
 486                 struct iomap_dio *dio)
 487 {
 488         const struct iomap *iomap = &iomi->iomap;
 489         struct iov_iter *iter = dio->submit.iter;
 490         void *inline_data = iomap_inline_data(iomap, iomi->pos);
 491         loff_t length = iomap_length(iomi);
 492         loff_t pos = iomi->pos;
 493         size_t copied;
 494
 495         if (WARN_ON_ONCE(!iomap_inline_data_valid(iomap)))
 496                 return -EIO;
 497
 498         if (dio->flags & IOMAP_DIO_WRITE) {
 499                 loff_t size = iomi->inode->i_size;
 500
 501                 if (pos > size)
 502                         memset(iomap_inline_data(iomap, size), 0, pos - size);
 503                 copied = copy_from_iter(inline_data, length, iter);
 504                 if (copied) {
 505                         if (pos + copied > size)
 506                                 i_size_write(iomi->inode, pos + copied);
 507                         mark_inode_dirty(iomi->inode);
 508                 }
 509         } else {
 510                 copied = copy_to_iter(inline_data, length, iter);
 511         }
 512         dio->size += copied;
 513         if (!copied)
 514                 return -EFAULT;
 515         return copied;
 516 }
 517
 518 static loff_t iomap_dio_iter(const struct iomap_iter *iter,
 519                 struct iomap_dio *dio)
 520 {
 521         switch (iter->iomap.type) {
 522         case IOMAP_HOLE:
 523                 if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
 524                         return -EIO;
 525                 return iomap_dio_hole_iter(iter, dio);
 526         case IOMAP_UNWRITTEN:
 527                 if (!(dio->flags & IOMAP_DIO_WRITE))
 528                         return iomap_dio_hole_iter(iter, dio);
 529                 return iomap_dio_bio_iter(iter, dio);
 530         case IOMAP_MAPPED:
 531                 return iomap_dio_bio_iter(iter, dio);
 532         case IOMAP_INLINE:
 533                 return iomap_dio_inline_iter(iter, dio);
 534         case IOMAP_DELALLOC:
 535                 /*
 536                  * DIO is not serialised against mmap() access at all, and so
 537                  * if the page_mkwrite occurs between the writeback and the
 538                  * iomap_iter() call in the DIO path, then it will see the
 539                  * DELALLOC block that the page-mkwrite allocated.
 540                  */
 541                 pr_warn_ratelimited("Direct I/O collision with buffered writes! File: %pD4 Comm: %.20s\n",
 542                                     dio->iocb->ki_filp, current->comm);
 543                 return -EIO;
 544         default:
 545                 WARN_ON_ONCE(1);
 546                 return -EIO;
 547         }
 548 }
 549
 550 /*
 551  * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
 552  * is being issued as AIO or not.  This allows us to optimise pure data writes
 553  * to use REQ_FUA rather than requiring generic_write_sync() to issue a
 554  * REQ_FLUSH post write. This is slightly tricky because a single request here
 555  * can be mapped into multiple disjoint IOs and only a subset of the IOs issued
 556  * may be pure data writes. In that case, we still need to do a full data sync
 557  * completion.
 558  *
 559  * When page faults are disabled and @dio_flags includes IOMAP_DIO_PARTIAL,
 560  * __iomap_dio_rw can return a partial result if it encounters a non-resident
 561  * page in @iter after preparing a transfer.  In that case, the non-resident
 562  * pages can be faulted in and the request resumed with @done_before set to the
 563  * number of bytes previously transferred.  The request will then complete with
 564  * the correct total number of bytes transferred; this is essential for
 565  * completing partial requests asynchronously.
 566  *
 567  * Returns -ENOTBLK In case of a page invalidation invalidation failure for
 568  * writes.  The callers needs to fall back to buffered I/O in this case.
 569  */
 570 struct iomap_dio *
 571 __iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 572                 const struct iomap_ops *ops, const struct iomap_dio_ops *dops,
 573                 unsigned int dio_flags, void *private, size_t done_before)
 574 {
 575         struct inode *inode = file_inode(iocb->ki_filp);
 576         struct iomap_iter iomi = {
 577                 .inode          = inode,
 578                 .pos            = iocb->ki_pos,
 579                 .len            = iov_iter_count(iter),
 580                 .flags          = IOMAP_DIRECT,
 581                 .private        = private,
 582         };
 583         bool wait_for_completion =
 584                 is_sync_kiocb(iocb) || (dio_flags & IOMAP_DIO_FORCE_WAIT);
 585         struct blk_plug plug;
 586         struct iomap_dio *dio;
 587         loff_t ret = 0;
 588
 589         trace_iomap_dio_rw_begin(iocb, iter, dio_flags, done_before);
 590
 591         if (!iomi.len)
 592                 return NULL;
 593
 594         dio = kmalloc(sizeof(*dio), GFP_KERNEL);
 595         if (!dio)
 596                 return ERR_PTR(-ENOMEM);
 597
 598         dio->iocb = iocb;
 599         atomic_set(&dio->ref, 1);
 600         dio->size = 0;
 601         dio->i_size = i_size_read(inode);
 602         dio->dops = dops;
 603         dio->error = 0;
 604         dio->flags = 0;
 605         dio->done_before = done_before;
 606
 607         dio->submit.iter = iter;
 608         dio->submit.waiter = current;
 609
 610         if (iocb->ki_flags & IOCB_NOWAIT)
 611                 iomi.flags |= IOMAP_NOWAIT;
 612
 613         if (iocb->ki_flags & IOCB_ATOMIC)
 614                 iomi.flags |= IOMAP_ATOMIC;
 615
 616         if (iov_iter_rw(iter) == READ) {
 617                 /* reads can always complete inline */
 618                 dio->flags |= IOMAP_DIO_INLINE_COMP;
 619
 620                 if (iomi.pos >= dio->i_size)
 621                         goto out_free_dio;
 622
 623                 if (user_backed_iter(iter))
 624                         dio->flags |= IOMAP_DIO_DIRTY;
 625
 626                 ret = kiocb_write_and_wait(iocb, iomi.len);
 627                 if (ret)
 628                         goto out_free_dio;
 629         } else {
 630                 iomi.flags |= IOMAP_WRITE;
 631                 dio->flags |= IOMAP_DIO_WRITE;
 632
 633                 /*
 634                  * Flag as supporting deferred completions, if the issuer
 635                  * groks it. This can avoid a workqueue punt for writes.
 636                  * We may later clear this flag if we need to do other IO
 637                  * as part of this IO completion.
 638                  */
 639                 if (iocb->ki_flags & IOCB_DIO_CALLER_COMP)
 640                         dio->flags |= IOMAP_DIO_CALLER_COMP;
 641
 642                 if (dio_flags & IOMAP_DIO_OVERWRITE_ONLY) {
 643                         ret = -EAGAIN;
 644                         if (iomi.pos >= dio->i_size ||
 645                             iomi.pos + iomi.len > dio->i_size)
 646                                 goto out_free_dio;
 647                         iomi.flags |= IOMAP_OVERWRITE_ONLY;
 648                 }
 649
 650                 /* for data sync or sync, we need sync completion processing */
 651                 if (iocb_is_dsync(iocb)) {
 652                         dio->flags |= IOMAP_DIO_NEED_SYNC;
 653
 654                        /*
 655                         * For datasync only writes, we optimistically try using
 656                         * WRITE_THROUGH for this IO. This flag requires either
 657                         * FUA writes through the device's write cache, or a
 658                         * normal write to a device without a volatile write
 659                         * cache. For the former, Any non-FUA write that occurs
 660                         * will clear this flag, hence we know before completion
 661                         * whether a cache flush is necessary.
 662                         */
 663                         if (!(iocb->ki_flags & IOCB_SYNC))
 664                                 dio->flags |= IOMAP_DIO_WRITE_THROUGH;
 665                 }
 666
 667                 /*
 668                  * Try to invalidate cache pages for the range we are writing.
 669                  * If this invalidation fails, let the caller fall back to
 670                  * buffered I/O.
 671                  */
 672                 ret = kiocb_invalidate_pages(iocb, iomi.len);
 673                 if (ret) {
 674                         if (ret != -EAGAIN) {
 675                                 trace_iomap_dio_invalidate_fail(inode, iomi.pos,
 676                                                                 iomi.len);
 677                                 if (iocb->ki_flags & IOCB_ATOMIC) {
 678                                         /*
 679                                          * folio invalidation failed, maybe
 680                                          * this is transient, unlock and see if
 681                                          * the caller tries again.
 682                                          */
 683                                         ret = -EAGAIN;
 684                                 } else {
 685                                         /* fall back to buffered write */
 686                                         ret = -ENOTBLK;
 687                                 }
 688                         }
 689                         goto out_free_dio;
 690                 }
 691
 692                 if (!wait_for_completion && !inode->i_sb->s_dio_done_wq) {
 693                         ret = sb_init_dio_done_wq(inode->i_sb);
 694                         if (ret < 0)
 695                                 goto out_free_dio;
 696                 }
 697         }
 698
 699         inode_dio_begin(inode);
 700
 701         blk_start_plug(&plug);
 702         while ((ret = iomap_iter(&iomi, ops)) > 0) {
 703                 iomi.processed = iomap_dio_iter(&iomi, dio);
 704
 705                 /*
 706                  * We can only poll for single bio I/Os.
 707                  */
 708                 iocb->ki_flags &= ~IOCB_HIPRI;
 709         }
 710
 711         blk_finish_plug(&plug);
 712
 713         /*
 714          * We only report that we've read data up to i_size.
 715          * Revert iter to a state corresponding to that as some callers (such
 716          * as the splice code) rely on it.
 717          */
 718         if (iov_iter_rw(iter) == READ && iomi.pos >= dio->i_size)
 719                 iov_iter_revert(iter, iomi.pos - dio->i_size);
 720
 721         if (ret == -EFAULT && dio->size && (dio_flags & IOMAP_DIO_PARTIAL)) {
 722                 if (!(iocb->ki_flags & IOCB_NOWAIT))
 723                         wait_for_completion = true;
 724                 ret = 0;
 725         }
 726
 727         /* magic error code to fall back to buffered I/O */
 728         if (ret == -ENOTBLK) {
 729                 wait_for_completion = true;
 730                 ret = 0;
 731         }
 732         if (ret < 0)
 733                 iomap_dio_set_error(dio, ret);
 734
 735         /*
 736          * If all the writes we issued were already written through to the
 737          * media, we don't need to flush the cache on IO completion. Clear the
 738          * sync flag for this case.
 739          */
 740         if (dio->flags & IOMAP_DIO_WRITE_THROUGH)
 741                 dio->flags &= ~IOMAP_DIO_NEED_SYNC;
 742
 743         /*
 744          * We are about to drop our additional submission reference, which
 745          * might be the last reference to the dio.  There are three different
 746          * ways we can progress here:
 747          *
 748          *  (a) If this is the last reference we will always complete and free
 749          *      the dio ourselves.
 750          *  (b) If this is not the last reference, and we serve an asynchronous
 751          *      iocb, we must never touch the dio after the decrement, the
 752          *      I/O completion handler will complete and free it.
 753          *  (c) If this is not the last reference, but we serve a synchronous
 754          *      iocb, the I/O completion handler will wake us up on the drop
 755          *      of the final reference, and we will complete and free it here
 756          *      after we got woken by the I/O completion handler.
 757          */
 758         dio->wait_for_completion = wait_for_completion;
 759         if (!atomic_dec_and_test(&dio->ref)) {
 760                 if (!wait_for_completion) {
 761                         trace_iomap_dio_rw_queued(inode, iomi.pos, iomi.len);
 762                         return ERR_PTR(-EIOCBQUEUED);
 763                 }
 764
 765                 for (;;) {
 766                         set_current_state(TASK_UNINTERRUPTIBLE);
 767                         if (!READ_ONCE(dio->submit.waiter))
 768                                 break;
 769
 770                         blk_io_schedule();
 771                 }
 772                 __set_current_state(TASK_RUNNING);
 773         }
 774
 775         return dio;
 776
 777 out_free_dio:
 778         kfree(dio);
 779         if (ret)
 780                 return ERR_PTR(ret);
 781         return NULL;
 782 }
 783 EXPORT_SYMBOL_GPL(__iomap_dio_rw);
 784
 785 ssize_t
 786 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 787                 const struct iomap_ops *ops, const struct iomap_dio_ops *dops,
 788                 unsigned int dio_flags, void *private, size_t done_before)
 789 {
 790         struct iomap_dio *dio;
 791
 792         dio = __iomap_dio_rw(iocb, iter, ops, dops, dio_flags, private,
 793                              done_before);
 794         if (IS_ERR_OR_NULL(dio))
 795                 return PTR_ERR_OR_ZERO(dio);
 796         return iomap_dio_complete(dio);
 797 }
 798 EXPORT_SYMBOL_GPL(iomap_dio_rw);
 799
 800 static int __init iomap_dio_init(void)
 801 {
 802         zero_page = alloc_pages(GFP_KERNEL | __GFP_ZERO,
 803                                 IOMAP_ZERO_PAGE_ORDER);
 804
 805         if (!zero_page)
 806                 return -ENOMEM;
 807
 808         return 0;
 809 }
 810 fs_initcall(iomap_dio_init);