fs/iomap.c

   1 /*
   2  * Copyright (C) 2010 Red Hat, Inc.
   3  * Copyright (c) 2016 Christoph Hellwig.
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms and conditions of the GNU General Public License,
   7  * version 2, as published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12  * more details.
  13  */
  14 #include <linux/module.h>
  15 #include <linux/compiler.h>
  16 #include <linux/fs.h>
  17 #include <linux/iomap.h>
  18 #include <linux/uaccess.h>
  19 #include <linux/gfp.h>
  20 #include <linux/mm.h>
  21 #include <linux/swap.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/file.h>
  24 #include <linux/uio.h>
  25 #include <linux/backing-dev.h>
  26 #include <linux/buffer_head.h>
  27 #include <linux/task_io_accounting_ops.h>
  28 #include <linux/dax.h>
  29 #include <linux/sched/signal.h>
  30
  31 #include "internal.h"
  32
  33 /*
  34  * Execute a iomap write on a segment of the mapping that spans a
  35  * contiguous range of pages that have identical block mapping state.
  36  *
  37  * This avoids the need to map pages individually, do individual allocations
  38  * for each page and most importantly avoid the need for filesystem specific
  39  * locking per page. Instead, all the operations are amortised over the entire
  40  * range of pages. It is assumed that the filesystems will lock whatever
  41  * resources they require in the iomap_begin call, and release them in the
  42  * iomap_end call.
  43  */
  44 loff_t
  45 iomap_apply(struct inode *inode, loff_t pos, loff_t length, unsigned flags,
  46                 const struct iomap_ops *ops, void *data, iomap_actor_t actor)
  47 {
  48         struct iomap iomap = { 0 };
  49         loff_t written = 0, ret;
  50
  51         /*
  52          * Need to map a range from start position for length bytes. This can
  53          * span multiple pages - it is only guaranteed to return a range of a
  54          * single type of pages (e.g. all into a hole, all mapped or all
  55          * unwritten). Failure at this point has nothing to undo.
  56          *
  57          * If allocation is required for this range, reserve the space now so
  58          * that the allocation is guaranteed to succeed later on. Once we copy
  59          * the data into the page cache pages, then we cannot fail otherwise we
  60          * expose transient stale data. If the reserve fails, we can safely
  61          * back out at this point as there is nothing to undo.
  62          */
  63         ret = ops->iomap_begin(inode, pos, length, flags, &iomap);
  64         if (ret)
  65                 return ret;
  66         if (WARN_ON(iomap.offset > pos))
  67                 return -EIO;
  68
  69         /*
  70          * Cut down the length to the one actually provided by the filesystem,
  71          * as it might not be able to give us the whole size that we requested.
  72          */
  73         if (iomap.offset + iomap.length < pos + length)
  74                 length = iomap.offset + iomap.length - pos;
  75
  76         /*
  77          * Now that we have guaranteed that the space allocation will succeed.
  78          * we can do the copy-in page by page without having to worry about
  79          * failures exposing transient data.
  80          */
  81         written = actor(inode, pos, length, data, &iomap);
  82
  83         /*
  84          * Now the data has been copied, commit the range we've copied.  This
  85          * should not fail unless the filesystem has had a fatal error.
  86          */
  87         if (ops->iomap_end) {
  88                 ret = ops->iomap_end(inode, pos, length,
  89                                      written > 0 ? written : 0,
  90                                      flags, &iomap);
  91         }
  92
  93         return written ? written : ret;
  94 }
  95
  96 static void
  97 iomap_write_failed(struct inode *inode, loff_t pos, unsigned len)
  98 {
  99         loff_t i_size = i_size_read(inode);
 100
 101         /*
 102          * Only truncate newly allocated pages beyoned EOF, even if the
 103          * write started inside the existing inode size.
 104          */
 105         if (pos + len > i_size)
 106                 truncate_pagecache_range(inode, max(pos, i_size), pos + len);
 107 }
 108
 109 static int
 110 iomap_write_begin(struct inode *inode, loff_t pos, unsigned len, unsigned flags,
 111                 struct page **pagep, struct iomap *iomap)
 112 {
 113         pgoff_t index = pos >> PAGE_SHIFT;
 114         struct page *page;
 115         int status = 0;
 116
 117         BUG_ON(pos + len > iomap->offset + iomap->length);
 118
 119         if (fatal_signal_pending(current))
 120                 return -EINTR;
 121
 122         page = grab_cache_page_write_begin(inode->i_mapping, index, flags);
 123         if (!page)
 124                 return -ENOMEM;
 125
 126         status = __block_write_begin_int(page, pos, len, NULL, iomap);
 127         if (unlikely(status)) {
 128                 unlock_page(page);
 129                 put_page(page);
 130                 page = NULL;
 131
 132                 iomap_write_failed(inode, pos, len);
 133         }
 134
 135         *pagep = page;
 136         return status;
 137 }
 138
 139 static int
 140 iomap_write_end(struct inode *inode, loff_t pos, unsigned len,
 141                 unsigned copied, struct page *page)
 142 {
 143         int ret;
 144
 145         ret = generic_write_end(NULL, inode->i_mapping, pos, len,
 146                         copied, page, NULL);
 147         if (ret < len)
 148                 iomap_write_failed(inode, pos, len);
 149         return ret;
 150 }
 151
 152 static loff_t
 153 iomap_write_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 154                 struct iomap *iomap)
 155 {
 156         struct iov_iter *i = data;
 157         long status = 0;
 158         ssize_t written = 0;
 159         unsigned int flags = AOP_FLAG_NOFS;
 160
 161         /*
 162          * Copies from kernel address space cannot fail (NFSD is a big user).
 163          */
 164         if (!iter_is_iovec(i))
 165                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
 166
 167         do {
 168                 struct page *page;
 169                 unsigned long offset;   /* Offset into pagecache page */
 170                 unsigned long bytes;    /* Bytes to write to page */
 171                 size_t copied;          /* Bytes copied from user */
 172
 173                 offset = (pos & (PAGE_SIZE - 1));
 174                 bytes = min_t(unsigned long, PAGE_SIZE - offset,
 175                                                 iov_iter_count(i));
 176 again:
 177                 if (bytes > length)
 178                         bytes = length;
 179
 180                 /*
 181                  * Bring in the user page that we will copy from _first_.
 182                  * Otherwise there's a nasty deadlock on copying from the
 183                  * same page as we're writing to, without it being marked
 184                  * up-to-date.
 185                  *
 186                  * Not only is this an optimisation, but it is also required
 187                  * to check that the address is actually valid, when atomic
 188                  * usercopies are used, below.
 189                  */
 190                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
 191                         status = -EFAULT;
 192                         break;
 193                 }
 194
 195                 status = iomap_write_begin(inode, pos, bytes, flags, &page,
 196                                 iomap);
 197                 if (unlikely(status))
 198                         break;
 199
 200                 if (mapping_writably_mapped(inode->i_mapping))
 201                         flush_dcache_page(page);
 202
 203                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
 204
 205                 flush_dcache_page(page);
 206
 207                 status = iomap_write_end(inode, pos, bytes, copied, page);
 208                 if (unlikely(status < 0))
 209                         break;
 210                 copied = status;
 211
 212                 cond_resched();
 213
 214                 iov_iter_advance(i, copied);
 215                 if (unlikely(copied == 0)) {
 216                         /*
 217                          * If we were unable to copy any data at all, we must
 218                          * fall back to a single segment length write.
 219                          *
 220                          * If we didn't fallback here, we could livelock
 221                          * because not all segments in the iov can be copied at
 222                          * once without a pagefault.
 223                          */
 224                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
 225                                                 iov_iter_single_seg_count(i));
 226                         goto again;
 227                 }
 228                 pos += copied;
 229                 written += copied;
 230                 length -= copied;
 231
 232                 balance_dirty_pages_ratelimited(inode->i_mapping);
 233         } while (iov_iter_count(i) && length);
 234
 235         return written ? written : status;
 236 }
 237
 238 ssize_t
 239 iomap_file_buffered_write(struct kiocb *iocb, struct iov_iter *iter,
 240                 const struct iomap_ops *ops)
 241 {
 242         struct inode *inode = iocb->ki_filp->f_mapping->host;
 243         loff_t pos = iocb->ki_pos, ret = 0, written = 0;
 244
 245         while (iov_iter_count(iter)) {
 246                 ret = iomap_apply(inode, pos, iov_iter_count(iter),
 247                                 IOMAP_WRITE, ops, iter, iomap_write_actor);
 248                 if (ret <= 0)
 249                         break;
 250                 pos += ret;
 251                 written += ret;
 252         }
 253
 254         return written ? written : ret;
 255 }
 256 EXPORT_SYMBOL_GPL(iomap_file_buffered_write);
 257
 258 static struct page *
 259 __iomap_read_page(struct inode *inode, loff_t offset)
 260 {
 261         struct address_space *mapping = inode->i_mapping;
 262         struct page *page;
 263
 264         page = read_mapping_page(mapping, offset >> PAGE_SHIFT, NULL);
 265         if (IS_ERR(page))
 266                 return page;
 267         if (!PageUptodate(page)) {
 268                 put_page(page);
 269                 return ERR_PTR(-EIO);
 270         }
 271         return page;
 272 }
 273
 274 static loff_t
 275 iomap_dirty_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 276                 struct iomap *iomap)
 277 {
 278         long status = 0;
 279         ssize_t written = 0;
 280
 281         do {
 282                 struct page *page, *rpage;
 283                 unsigned long offset;   /* Offset into pagecache page */
 284                 unsigned long bytes;    /* Bytes to write to page */
 285
 286                 offset = (pos & (PAGE_SIZE - 1));
 287                 bytes = min_t(unsigned long, PAGE_SIZE - offset, length);
 288
 289                 rpage = __iomap_read_page(inode, pos);
 290                 if (IS_ERR(rpage))
 291                         return PTR_ERR(rpage);
 292
 293                 status = iomap_write_begin(inode, pos, bytes,
 294                                 AOP_FLAG_NOFS | AOP_FLAG_UNINTERRUPTIBLE,
 295                                 &page, iomap);
 296                 put_page(rpage);
 297                 if (unlikely(status))
 298                         return status;
 299
 300                 WARN_ON_ONCE(!PageUptodate(page));
 301
 302                 status = iomap_write_end(inode, pos, bytes, bytes, page);
 303                 if (unlikely(status <= 0)) {
 304                         if (WARN_ON_ONCE(status == 0))
 305                                 return -EIO;
 306                         return status;
 307                 }
 308
 309                 cond_resched();
 310
 311                 pos += status;
 312                 written += status;
 313                 length -= status;
 314
 315                 balance_dirty_pages_ratelimited(inode->i_mapping);
 316         } while (length);
 317
 318         return written;
 319 }
 320
 321 int
 322 iomap_file_dirty(struct inode *inode, loff_t pos, loff_t len,
 323                 const struct iomap_ops *ops)
 324 {
 325         loff_t ret;
 326
 327         while (len) {
 328                 ret = iomap_apply(inode, pos, len, IOMAP_WRITE, ops, NULL,
 329                                 iomap_dirty_actor);
 330                 if (ret <= 0)
 331                         return ret;
 332                 pos += ret;
 333                 len -= ret;
 334         }
 335
 336         return 0;
 337 }
 338 EXPORT_SYMBOL_GPL(iomap_file_dirty);
 339
 340 static int iomap_zero(struct inode *inode, loff_t pos, unsigned offset,
 341                 unsigned bytes, struct iomap *iomap)
 342 {
 343         struct page *page;
 344         int status;
 345
 346         status = iomap_write_begin(inode, pos, bytes,
 347                         AOP_FLAG_UNINTERRUPTIBLE | AOP_FLAG_NOFS, &page, iomap);
 348         if (status)
 349                 return status;
 350
 351         zero_user(page, offset, bytes);
 352         mark_page_accessed(page);
 353
 354         return iomap_write_end(inode, pos, bytes, bytes, page);
 355 }
 356
 357 static int iomap_dax_zero(loff_t pos, unsigned offset, unsigned bytes,
 358                 struct iomap *iomap)
 359 {
 360         sector_t sector = iomap->blkno +
 361                 (((pos & ~(PAGE_SIZE - 1)) - iomap->offset) >> 9);
 362
 363         return __dax_zero_page_range(iomap->bdev, iomap->dax_dev, sector,
 364                         offset, bytes);
 365 }
 366
 367 static loff_t
 368 iomap_zero_range_actor(struct inode *inode, loff_t pos, loff_t count,
 369                 void *data, struct iomap *iomap)
 370 {
 371         bool *did_zero = data;
 372         loff_t written = 0;
 373         int status;
 374
 375         /* already zeroed?  we're done. */
 376         if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
 377                 return count;
 378
 379         do {
 380                 unsigned offset, bytes;
 381
 382                 offset = pos & (PAGE_SIZE - 1); /* Within page */
 383                 bytes = min_t(unsigned, PAGE_SIZE - offset, count);
 384
 385                 if (IS_DAX(inode))
 386                         status = iomap_dax_zero(pos, offset, bytes, iomap);
 387                 else
 388                         status = iomap_zero(inode, pos, offset, bytes, iomap);
 389                 if (status < 0)
 390                         return status;
 391
 392                 pos += bytes;
 393                 count -= bytes;
 394                 written += bytes;
 395                 if (did_zero)
 396                         *did_zero = true;
 397         } while (count > 0);
 398
 399         return written;
 400 }
 401
 402 int
 403 iomap_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
 404                 const struct iomap_ops *ops)
 405 {
 406         loff_t ret;
 407
 408         while (len > 0) {
 409                 ret = iomap_apply(inode, pos, len, IOMAP_ZERO,
 410                                 ops, did_zero, iomap_zero_range_actor);
 411                 if (ret <= 0)
 412                         return ret;
 413
 414                 pos += ret;
 415                 len -= ret;
 416         }
 417
 418         return 0;
 419 }
 420 EXPORT_SYMBOL_GPL(iomap_zero_range);
 421
 422 int
 423 iomap_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
 424                 const struct iomap_ops *ops)
 425 {
 426         unsigned int blocksize = i_blocksize(inode);
 427         unsigned int off = pos & (blocksize - 1);
 428
 429         /* Block boundary? Nothing to do */
 430         if (!off)
 431                 return 0;
 432         return iomap_zero_range(inode, pos, blocksize - off, did_zero, ops);
 433 }
 434 EXPORT_SYMBOL_GPL(iomap_truncate_page);
 435
 436 static loff_t
 437 iomap_page_mkwrite_actor(struct inode *inode, loff_t pos, loff_t length,
 438                 void *data, struct iomap *iomap)
 439 {
 440         struct page *page = data;
 441         int ret;
 442
 443         ret = __block_write_begin_int(page, pos, length, NULL, iomap);
 444         if (ret)
 445                 return ret;
 446
 447         block_commit_write(page, 0, length);
 448         return length;
 449 }
 450
 451 int iomap_page_mkwrite(struct vm_fault *vmf, const struct iomap_ops *ops)
 452 {
 453         struct page *page = vmf->page;
 454         struct inode *inode = file_inode(vmf->vma->vm_file);
 455         unsigned long length;
 456         loff_t offset, size;
 457         ssize_t ret;
 458
 459         lock_page(page);
 460         size = i_size_read(inode);
 461         if ((page->mapping != inode->i_mapping) ||
 462             (page_offset(page) > size)) {
 463                 /* We overload EFAULT to mean page got truncated */
 464                 ret = -EFAULT;
 465                 goto out_unlock;
 466         }
 467
 468         /* page is wholly or partially inside EOF */
 469         if (((page->index + 1) << PAGE_SHIFT) > size)
 470                 length = size & ~PAGE_MASK;
 471         else
 472                 length = PAGE_SIZE;
 473
 474         offset = page_offset(page);
 475         while (length > 0) {
 476                 ret = iomap_apply(inode, offset, length,
 477                                 IOMAP_WRITE | IOMAP_FAULT, ops, page,
 478                                 iomap_page_mkwrite_actor);
 479                 if (unlikely(ret <= 0))
 480                         goto out_unlock;
 481                 offset += ret;
 482                 length -= ret;
 483         }
 484
 485         set_page_dirty(page);
 486         wait_for_stable_page(page);
 487         return 0;
 488 out_unlock:
 489         unlock_page(page);
 490         return ret;
 491 }
 492 EXPORT_SYMBOL_GPL(iomap_page_mkwrite);
 493
 494 struct fiemap_ctx {
 495         struct fiemap_extent_info *fi;
 496         struct iomap prev;
 497 };
 498
 499 static int iomap_to_fiemap(struct fiemap_extent_info *fi,
 500                 struct iomap *iomap, u32 flags)
 501 {
 502         switch (iomap->type) {
 503         case IOMAP_HOLE:
 504                 /* skip holes */
 505                 return 0;
 506         case IOMAP_DELALLOC:
 507                 flags |= FIEMAP_EXTENT_DELALLOC | FIEMAP_EXTENT_UNKNOWN;
 508                 break;
 509         case IOMAP_UNWRITTEN:
 510                 flags |= FIEMAP_EXTENT_UNWRITTEN;
 511                 break;
 512         case IOMAP_MAPPED:
 513                 break;
 514         }
 515
 516         if (iomap->flags & IOMAP_F_MERGED)
 517                 flags |= FIEMAP_EXTENT_MERGED;
 518         if (iomap->flags & IOMAP_F_SHARED)
 519                 flags |= FIEMAP_EXTENT_SHARED;
 520
 521         return fiemap_fill_next_extent(fi, iomap->offset,
 522                         iomap->blkno != IOMAP_NULL_BLOCK ? iomap->blkno << 9: 0,
 523                         iomap->length, flags);
 524
 525 }
 526
 527 static loff_t
 528 iomap_fiemap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 529                 struct iomap *iomap)
 530 {
 531         struct fiemap_ctx *ctx = data;
 532         loff_t ret = length;
 533
 534         if (iomap->type == IOMAP_HOLE)
 535                 return length;
 536
 537         ret = iomap_to_fiemap(ctx->fi, &ctx->prev, 0);
 538         ctx->prev = *iomap;
 539         switch (ret) {
 540         case 0:         /* success */
 541                 return length;
 542         case 1:         /* extent array full */
 543                 return 0;
 544         default:
 545                 return ret;
 546         }
 547 }
 548
 549 int iomap_fiemap(struct inode *inode, struct fiemap_extent_info *fi,
 550                 loff_t start, loff_t len, const struct iomap_ops *ops)
 551 {
 552         struct fiemap_ctx ctx;
 553         loff_t ret;
 554
 555         memset(&ctx, 0, sizeof(ctx));
 556         ctx.fi = fi;
 557         ctx.prev.type = IOMAP_HOLE;
 558
 559         ret = fiemap_check_flags(fi, FIEMAP_FLAG_SYNC);
 560         if (ret)
 561                 return ret;
 562
 563         if (fi->fi_flags & FIEMAP_FLAG_SYNC) {
 564                 ret = filemap_write_and_wait(inode->i_mapping);
 565                 if (ret)
 566                         return ret;
 567         }
 568
 569         while (len > 0) {
 570                 ret = iomap_apply(inode, start, len, IOMAP_REPORT, ops, &ctx,
 571                                 iomap_fiemap_actor);
 572                 /* inode with no (attribute) mapping will give ENOENT */
 573                 if (ret == -ENOENT)
 574                         break;
 575                 if (ret < 0)
 576                         return ret;
 577                 if (ret == 0)
 578                         break;
 579
 580                 start += ret;
 581                 len -= ret;
 582         }
 583
 584         if (ctx.prev.type != IOMAP_HOLE) {
 585                 ret = iomap_to_fiemap(fi, &ctx.prev, FIEMAP_EXTENT_LAST);
 586                 if (ret < 0)
 587                         return ret;
 588         }
 589
 590         return 0;
 591 }
 592 EXPORT_SYMBOL_GPL(iomap_fiemap);
 593
 594 /*
 595  * Private flags for iomap_dio, must not overlap with the public ones in
 596  * iomap.h:
 597  */
 598 #define IOMAP_DIO_WRITE         (1 << 30)
 599 #define IOMAP_DIO_DIRTY         (1 << 31)
 600
 601 struct iomap_dio {
 602         struct kiocb            *iocb;
 603         iomap_dio_end_io_t      *end_io;
 604         loff_t                  i_size;
 605         loff_t                  size;
 606         atomic_t                ref;
 607         unsigned                flags;
 608         int                     error;
 609
 610         union {
 611                 /* used during submission and for synchronous completion: */
 612                 struct {
 613                         struct iov_iter         *iter;
 614                         struct task_struct      *waiter;
 615                         struct request_queue    *last_queue;
 616                         blk_qc_t                cookie;
 617                 } submit;
 618
 619                 /* used for aio completion: */
 620                 struct {
 621                         struct work_struct      work;
 622                 } aio;
 623         };
 624 };
 625
 626 static ssize_t iomap_dio_complete(struct iomap_dio *dio)
 627 {
 628         struct kiocb *iocb = dio->iocb;
 629         ssize_t ret;
 630
 631         if (dio->end_io) {
 632                 ret = dio->end_io(iocb,
 633                                 dio->error ? dio->error : dio->size,
 634                                 dio->flags);
 635         } else {
 636                 ret = dio->error;
 637         }
 638
 639         if (likely(!ret)) {
 640                 ret = dio->size;
 641                 /* check for short read */
 642                 if (iocb->ki_pos + ret > dio->i_size &&
 643                     !(dio->flags & IOMAP_DIO_WRITE))
 644                         ret = dio->i_size - iocb->ki_pos;
 645                 iocb->ki_pos += ret;
 646         }
 647
 648         inode_dio_end(file_inode(iocb->ki_filp));
 649         kfree(dio);
 650
 651         return ret;
 652 }
 653
 654 static void iomap_dio_complete_work(struct work_struct *work)
 655 {
 656         struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work);
 657         struct kiocb *iocb = dio->iocb;
 658         bool is_write = (dio->flags & IOMAP_DIO_WRITE);
 659         ssize_t ret;
 660
 661         ret = iomap_dio_complete(dio);
 662         if (is_write && ret > 0)
 663                 ret = generic_write_sync(iocb, ret);
 664         iocb->ki_complete(iocb, ret, 0);
 665 }
 666
 667 /*
 668  * Set an error in the dio if none is set yet.  We have to use cmpxchg
 669  * as the submission context and the completion context(s) can race to
 670  * update the error.
 671  */
 672 static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret)
 673 {
 674         cmpxchg(&dio->error, 0, ret);
 675 }
 676
 677 static void iomap_dio_bio_end_io(struct bio *bio)
 678 {
 679         struct iomap_dio *dio = bio->bi_private;
 680         bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
 681
 682         if (bio->bi_error)
 683                 iomap_dio_set_error(dio, bio->bi_error);
 684
 685         if (atomic_dec_and_test(&dio->ref)) {
 686                 if (is_sync_kiocb(dio->iocb)) {
 687                         struct task_struct *waiter = dio->submit.waiter;
 688
 689                         WRITE_ONCE(dio->submit.waiter, NULL);
 690                         wake_up_process(waiter);
 691                 } else if (dio->flags & IOMAP_DIO_WRITE) {
 692                         struct inode *inode = file_inode(dio->iocb->ki_filp);
 693
 694                         INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
 695                         queue_work(inode->i_sb->s_dio_done_wq, &dio->aio.work);
 696                 } else {
 697                         iomap_dio_complete_work(&dio->aio.work);
 698                 }
 699         }
 700
 701         if (should_dirty) {
 702                 bio_check_pages_dirty(bio);
 703         } else {
 704                 struct bio_vec *bvec;
 705                 int i;
 706
 707                 bio_for_each_segment_all(bvec, bio, i)
 708                         put_page(bvec->bv_page);
 709                 bio_put(bio);
 710         }
 711 }
 712
 713 static blk_qc_t
 714 iomap_dio_zero(struct iomap_dio *dio, struct iomap *iomap, loff_t pos,
 715                 unsigned len)
 716 {
 717         struct page *page = ZERO_PAGE(0);
 718         struct bio *bio;
 719
 720         bio = bio_alloc(GFP_KERNEL, 1);
 721         bio->bi_bdev = iomap->bdev;
 722         bio->bi_iter.bi_sector =
 723                 iomap->blkno + ((pos - iomap->offset) >> 9);
 724         bio->bi_private = dio;
 725         bio->bi_end_io = iomap_dio_bio_end_io;
 726
 727         get_page(page);
 728         if (bio_add_page(bio, page, len, 0) != len)
 729                 BUG();
 730         bio_set_op_attrs(bio, REQ_OP_WRITE, REQ_SYNC | REQ_IDLE);
 731
 732         atomic_inc(&dio->ref);
 733         return submit_bio(bio);
 734 }
 735
 736 static loff_t
 737 iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
 738                 void *data, struct iomap *iomap)
 739 {
 740         struct iomap_dio *dio = data;
 741         unsigned int blkbits = blksize_bits(bdev_logical_block_size(iomap->bdev));
 742         unsigned int fs_block_size = i_blocksize(inode), pad;
 743         unsigned int align = iov_iter_alignment(dio->submit.iter);
 744         struct iov_iter iter;
 745         struct bio *bio;
 746         bool need_zeroout = false;
 747         int nr_pages, ret;
 748
 749         if ((pos | length | align) & ((1 << blkbits) - 1))
 750                 return -EINVAL;
 751
 752         switch (iomap->type) {
 753         case IOMAP_HOLE:
 754                 if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
 755                         return -EIO;
 756                 /*FALLTHRU*/
 757         case IOMAP_UNWRITTEN:
 758                 if (!(dio->flags & IOMAP_DIO_WRITE)) {
 759                         iov_iter_zero(length, dio->submit.iter);
 760                         dio->size += length;
 761                         return length;
 762                 }
 763                 dio->flags |= IOMAP_DIO_UNWRITTEN;
 764                 need_zeroout = true;
 765                 break;
 766         case IOMAP_MAPPED:
 767                 if (iomap->flags & IOMAP_F_SHARED)
 768                         dio->flags |= IOMAP_DIO_COW;
 769                 if (iomap->flags & IOMAP_F_NEW)
 770                         need_zeroout = true;
 771                 break;
 772         default:
 773                 WARN_ON_ONCE(1);
 774                 return -EIO;
 775         }
 776
 777         /*
 778          * Operate on a partial iter trimmed to the extent we were called for.
 779          * We'll update the iter in the dio once we're done with this extent.
 780          */
 781         iter = *dio->submit.iter;
 782         iov_iter_truncate(&iter, length);
 783
 784         nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
 785         if (nr_pages <= 0)
 786                 return nr_pages;
 787
 788         if (need_zeroout) {
 789                 /* zero out from the start of the block to the write offset */
 790                 pad = pos & (fs_block_size - 1);
 791                 if (pad)
 792                         iomap_dio_zero(dio, iomap, pos - pad, pad);
 793         }
 794
 795         do {
 796                 if (dio->error)
 797                         return 0;
 798
 799                 bio = bio_alloc(GFP_KERNEL, nr_pages);
 800                 bio->bi_bdev = iomap->bdev;
 801                 bio->bi_iter.bi_sector =
 802                         iomap->blkno + ((pos - iomap->offset) >> 9);
 803                 bio->bi_private = dio;
 804                 bio->bi_end_io = iomap_dio_bio_end_io;
 805
 806                 ret = bio_iov_iter_get_pages(bio, &iter);
 807                 if (unlikely(ret)) {
 808                         bio_put(bio);
 809                         return ret;
 810                 }
 811
 812                 if (dio->flags & IOMAP_DIO_WRITE) {
 813                         bio_set_op_attrs(bio, REQ_OP_WRITE, REQ_SYNC | REQ_IDLE);
 814                         task_io_account_write(bio->bi_iter.bi_size);
 815                 } else {
 816                         bio_set_op_attrs(bio, REQ_OP_READ, 0);
 817                         if (dio->flags & IOMAP_DIO_DIRTY)
 818                                 bio_set_pages_dirty(bio);
 819                 }
 820
 821                 dio->size += bio->bi_iter.bi_size;
 822                 pos += bio->bi_iter.bi_size;
 823
 824                 nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
 825
 826                 atomic_inc(&dio->ref);
 827
 828                 dio->submit.last_queue = bdev_get_queue(iomap->bdev);
 829                 dio->submit.cookie = submit_bio(bio);
 830         } while (nr_pages);
 831
 832         if (need_zeroout) {
 833                 /* zero out from the end of the write to the end of the block */
 834                 pad = pos & (fs_block_size - 1);
 835                 if (pad)
 836                         iomap_dio_zero(dio, iomap, pos, fs_block_size - pad);
 837         }
 838
 839         iov_iter_advance(dio->submit.iter, length);
 840         return length;
 841 }
 842
 843 ssize_t
 844 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 845                 const struct iomap_ops *ops, iomap_dio_end_io_t end_io)
 846 {
 847         struct address_space *mapping = iocb->ki_filp->f_mapping;
 848         struct inode *inode = file_inode(iocb->ki_filp);
 849         size_t count = iov_iter_count(iter);
 850         loff_t pos = iocb->ki_pos, start = pos;
 851         loff_t end = iocb->ki_pos + count - 1, ret = 0;
 852         unsigned int flags = IOMAP_DIRECT;
 853         struct blk_plug plug;
 854         struct iomap_dio *dio;
 855
 856         lockdep_assert_held(&inode->i_rwsem);
 857
 858         if (!count)
 859                 return 0;
 860
 861         dio = kmalloc(sizeof(*dio), GFP_KERNEL);
 862         if (!dio)
 863                 return -ENOMEM;
 864
 865         dio->iocb = iocb;
 866         atomic_set(&dio->ref, 1);
 867         dio->size = 0;
 868         dio->i_size = i_size_read(inode);
 869         dio->end_io = end_io;
 870         dio->error = 0;
 871         dio->flags = 0;
 872
 873         dio->submit.iter = iter;
 874         if (is_sync_kiocb(iocb)) {
 875                 dio->submit.waiter = current;
 876                 dio->submit.cookie = BLK_QC_T_NONE;
 877                 dio->submit.last_queue = NULL;
 878         }
 879
 880         if (iov_iter_rw(iter) == READ) {
 881                 if (pos >= dio->i_size)
 882                         goto out_free_dio;
 883
 884                 if (iter->type == ITER_IOVEC)
 885                         dio->flags |= IOMAP_DIO_DIRTY;
 886         } else {
 887                 dio->flags |= IOMAP_DIO_WRITE;
 888                 flags |= IOMAP_WRITE;
 889         }
 890
 891         if (mapping->nrpages) {
 892                 ret = filemap_write_and_wait_range(mapping, start, end);
 893                 if (ret)
 894                         goto out_free_dio;
 895
 896                 ret = invalidate_inode_pages2_range(mapping,
 897                                 start >> PAGE_SHIFT, end >> PAGE_SHIFT);
 898                 WARN_ON_ONCE(ret);
 899                 ret = 0;
 900         }
 901
 902         inode_dio_begin(inode);
 903
 904         blk_start_plug(&plug);
 905         do {
 906                 ret = iomap_apply(inode, pos, count, flags, ops, dio,
 907                                 iomap_dio_actor);
 908                 if (ret <= 0) {
 909                         /* magic error code to fall back to buffered I/O */
 910                         if (ret == -ENOTBLK)
 911                                 ret = 0;
 912                         break;
 913                 }
 914                 pos += ret;
 915         } while ((count = iov_iter_count(iter)) > 0);
 916         blk_finish_plug(&plug);
 917
 918         if (ret < 0)
 919                 iomap_dio_set_error(dio, ret);
 920
 921         if (ret >= 0 && iov_iter_rw(iter) == WRITE && !is_sync_kiocb(iocb) &&
 922                         !inode->i_sb->s_dio_done_wq) {
 923                 ret = sb_init_dio_done_wq(inode->i_sb);
 924                 if (ret < 0)
 925                         iomap_dio_set_error(dio, ret);
 926         }
 927
 928         if (!atomic_dec_and_test(&dio->ref)) {
 929                 if (!is_sync_kiocb(iocb))
 930                         return -EIOCBQUEUED;
 931
 932                 for (;;) {
 933                         set_current_state(TASK_UNINTERRUPTIBLE);
 934                         if (!READ_ONCE(dio->submit.waiter))
 935                                 break;
 936
 937                         if (!(iocb->ki_flags & IOCB_HIPRI) ||
 938                             !dio->submit.last_queue ||
 939                             !blk_mq_poll(dio->submit.last_queue,
 940                                          dio->submit.cookie))
 941                                 io_schedule();
 942                 }
 943                 __set_current_state(TASK_RUNNING);
 944         }
 945
 946         ret = iomap_dio_complete(dio);
 947
 948         /*
 949          * Try again to invalidate clean pages which might have been cached by
 950          * non-direct readahead, or faulted in by get_user_pages() if the source
 951          * of the write was an mmap'ed region of the file we're writing.  Either
 952          * one is a pretty crazy thing to do, so we don't support it 100%.  If
 953          * this invalidation fails, tough, the write still worked...
 954          */
 955         if (iov_iter_rw(iter) == WRITE && mapping->nrpages) {
 956                 int err = invalidate_inode_pages2_range(mapping,
 957                                 start >> PAGE_SHIFT, end >> PAGE_SHIFT);
 958                 WARN_ON_ONCE(err);
 959         }
 960
 961         return ret;
 962
 963 out_free_dio:
 964         kfree(dio);
 965         return ret;
 966 }
 967 EXPORT_SYMBOL_GPL(iomap_dio_rw);