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         do {
 162                 struct page *page;
 163                 unsigned long offset;   /* Offset into pagecache page */
 164                 unsigned long bytes;    /* Bytes to write to page */
 165                 size_t copied;          /* Bytes copied from user */
 166
 167                 offset = (pos & (PAGE_SIZE - 1));
 168                 bytes = min_t(unsigned long, PAGE_SIZE - offset,
 169                                                 iov_iter_count(i));
 170 again:
 171                 if (bytes > length)
 172                         bytes = length;
 173
 174                 /*
 175                  * Bring in the user page that we will copy from _first_.
 176                  * Otherwise there's a nasty deadlock on copying from the
 177                  * same page as we're writing to, without it being marked
 178                  * up-to-date.
 179                  *
 180                  * Not only is this an optimisation, but it is also required
 181                  * to check that the address is actually valid, when atomic
 182                  * usercopies are used, below.
 183                  */
 184                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
 185                         status = -EFAULT;
 186                         break;
 187                 }
 188
 189                 status = iomap_write_begin(inode, pos, bytes, flags, &page,
 190                                 iomap);
 191                 if (unlikely(status))
 192                         break;
 193
 194                 if (mapping_writably_mapped(inode->i_mapping))
 195                         flush_dcache_page(page);
 196
 197                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
 198
 199                 flush_dcache_page(page);
 200
 201                 status = iomap_write_end(inode, pos, bytes, copied, page);
 202                 if (unlikely(status < 0))
 203                         break;
 204                 copied = status;
 205
 206                 cond_resched();
 207
 208                 iov_iter_advance(i, copied);
 209                 if (unlikely(copied == 0)) {
 210                         /*
 211                          * If we were unable to copy any data at all, we must
 212                          * fall back to a single segment length write.
 213                          *
 214                          * If we didn't fallback here, we could livelock
 215                          * because not all segments in the iov can be copied at
 216                          * once without a pagefault.
 217                          */
 218                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
 219                                                 iov_iter_single_seg_count(i));
 220                         goto again;
 221                 }
 222                 pos += copied;
 223                 written += copied;
 224                 length -= copied;
 225
 226                 balance_dirty_pages_ratelimited(inode->i_mapping);
 227         } while (iov_iter_count(i) && length);
 228
 229         return written ? written : status;
 230 }
 231
 232 ssize_t
 233 iomap_file_buffered_write(struct kiocb *iocb, struct iov_iter *iter,
 234                 const struct iomap_ops *ops)
 235 {
 236         struct inode *inode = iocb->ki_filp->f_mapping->host;
 237         loff_t pos = iocb->ki_pos, ret = 0, written = 0;
 238
 239         while (iov_iter_count(iter)) {
 240                 ret = iomap_apply(inode, pos, iov_iter_count(iter),
 241                                 IOMAP_WRITE, ops, iter, iomap_write_actor);
 242                 if (ret <= 0)
 243                         break;
 244                 pos += ret;
 245                 written += ret;
 246         }
 247
 248         return written ? written : ret;
 249 }
 250 EXPORT_SYMBOL_GPL(iomap_file_buffered_write);
 251
 252 static struct page *
 253 __iomap_read_page(struct inode *inode, loff_t offset)
 254 {
 255         struct address_space *mapping = inode->i_mapping;
 256         struct page *page;
 257
 258         page = read_mapping_page(mapping, offset >> PAGE_SHIFT, NULL);
 259         if (IS_ERR(page))
 260                 return page;
 261         if (!PageUptodate(page)) {
 262                 put_page(page);
 263                 return ERR_PTR(-EIO);
 264         }
 265         return page;
 266 }
 267
 268 static loff_t
 269 iomap_dirty_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 270                 struct iomap *iomap)
 271 {
 272         long status = 0;
 273         ssize_t written = 0;
 274
 275         do {
 276                 struct page *page, *rpage;
 277                 unsigned long offset;   /* Offset into pagecache page */
 278                 unsigned long bytes;    /* Bytes to write to page */
 279
 280                 offset = (pos & (PAGE_SIZE - 1));
 281                 bytes = min_t(loff_t, PAGE_SIZE - offset, length);
 282
 283                 rpage = __iomap_read_page(inode, pos);
 284                 if (IS_ERR(rpage))
 285                         return PTR_ERR(rpage);
 286
 287                 status = iomap_write_begin(inode, pos, bytes,
 288                                            AOP_FLAG_NOFS, &page, iomap);
 289                 put_page(rpage);
 290                 if (unlikely(status))
 291                         return status;
 292
 293                 WARN_ON_ONCE(!PageUptodate(page));
 294
 295                 status = iomap_write_end(inode, pos, bytes, bytes, page);
 296                 if (unlikely(status <= 0)) {
 297                         if (WARN_ON_ONCE(status == 0))
 298                                 return -EIO;
 299                         return status;
 300                 }
 301
 302                 cond_resched();
 303
 304                 pos += status;
 305                 written += status;
 306                 length -= status;
 307
 308                 balance_dirty_pages_ratelimited(inode->i_mapping);
 309         } while (length);
 310
 311         return written;
 312 }
 313
 314 int
 315 iomap_file_dirty(struct inode *inode, loff_t pos, loff_t len,
 316                 const struct iomap_ops *ops)
 317 {
 318         loff_t ret;
 319
 320         while (len) {
 321                 ret = iomap_apply(inode, pos, len, IOMAP_WRITE, ops, NULL,
 322                                 iomap_dirty_actor);
 323                 if (ret <= 0)
 324                         return ret;
 325                 pos += ret;
 326                 len -= ret;
 327         }
 328
 329         return 0;
 330 }
 331 EXPORT_SYMBOL_GPL(iomap_file_dirty);
 332
 333 static int iomap_zero(struct inode *inode, loff_t pos, unsigned offset,
 334                 unsigned bytes, struct iomap *iomap)
 335 {
 336         struct page *page;
 337         int status;
 338
 339         status = iomap_write_begin(inode, pos, bytes, AOP_FLAG_NOFS, &page,
 340                                    iomap);
 341         if (status)
 342                 return status;
 343
 344         zero_user(page, offset, bytes);
 345         mark_page_accessed(page);
 346
 347         return iomap_write_end(inode, pos, bytes, bytes, page);
 348 }
 349
 350 static int iomap_dax_zero(loff_t pos, unsigned offset, unsigned bytes,
 351                 struct iomap *iomap)
 352 {
 353         sector_t sector = iomap->blkno +
 354                 (((pos & ~(PAGE_SIZE - 1)) - iomap->offset) >> 9);
 355
 356         return __dax_zero_page_range(iomap->bdev, iomap->dax_dev, sector,
 357                         offset, bytes);
 358 }
 359
 360 static loff_t
 361 iomap_zero_range_actor(struct inode *inode, loff_t pos, loff_t count,
 362                 void *data, struct iomap *iomap)
 363 {
 364         bool *did_zero = data;
 365         loff_t written = 0;
 366         int status;
 367
 368         /* already zeroed?  we're done. */
 369         if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
 370                 return count;
 371
 372         do {
 373                 unsigned offset, bytes;
 374
 375                 offset = pos & (PAGE_SIZE - 1); /* Within page */
 376                 bytes = min_t(loff_t, PAGE_SIZE - offset, count);
 377
 378                 if (IS_DAX(inode))
 379                         status = iomap_dax_zero(pos, offset, bytes, iomap);
 380                 else
 381                         status = iomap_zero(inode, pos, offset, bytes, iomap);
 382                 if (status < 0)
 383                         return status;
 384
 385                 pos += bytes;
 386                 count -= bytes;
 387                 written += bytes;
 388                 if (did_zero)
 389                         *did_zero = true;
 390         } while (count > 0);
 391
 392         return written;
 393 }
 394
 395 int
 396 iomap_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
 397                 const struct iomap_ops *ops)
 398 {
 399         loff_t ret;
 400
 401         while (len > 0) {
 402                 ret = iomap_apply(inode, pos, len, IOMAP_ZERO,
 403                                 ops, did_zero, iomap_zero_range_actor);
 404                 if (ret <= 0)
 405                         return ret;
 406
 407                 pos += ret;
 408                 len -= ret;
 409         }
 410
 411         return 0;
 412 }
 413 EXPORT_SYMBOL_GPL(iomap_zero_range);
 414
 415 int
 416 iomap_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
 417                 const struct iomap_ops *ops)
 418 {
 419         unsigned int blocksize = i_blocksize(inode);
 420         unsigned int off = pos & (blocksize - 1);
 421
 422         /* Block boundary? Nothing to do */
 423         if (!off)
 424                 return 0;
 425         return iomap_zero_range(inode, pos, blocksize - off, did_zero, ops);
 426 }
 427 EXPORT_SYMBOL_GPL(iomap_truncate_page);
 428
 429 static loff_t
 430 iomap_page_mkwrite_actor(struct inode *inode, loff_t pos, loff_t length,
 431                 void *data, struct iomap *iomap)
 432 {
 433         struct page *page = data;
 434         int ret;
 435
 436         ret = __block_write_begin_int(page, pos, length, NULL, iomap);
 437         if (ret)
 438                 return ret;
 439
 440         block_commit_write(page, 0, length);
 441         return length;
 442 }
 443
 444 int iomap_page_mkwrite(struct vm_fault *vmf, const struct iomap_ops *ops)
 445 {
 446         struct page *page = vmf->page;
 447         struct inode *inode = file_inode(vmf->vma->vm_file);
 448         unsigned long length;
 449         loff_t offset, size;
 450         ssize_t ret;
 451
 452         lock_page(page);
 453         size = i_size_read(inode);
 454         if ((page->mapping != inode->i_mapping) ||
 455             (page_offset(page) > size)) {
 456                 /* We overload EFAULT to mean page got truncated */
 457                 ret = -EFAULT;
 458                 goto out_unlock;
 459         }
 460
 461         /* page is wholly or partially inside EOF */
 462         if (((page->index + 1) << PAGE_SHIFT) > size)
 463                 length = size & ~PAGE_MASK;
 464         else
 465                 length = PAGE_SIZE;
 466
 467         offset = page_offset(page);
 468         while (length > 0) {
 469                 ret = iomap_apply(inode, offset, length,
 470                                 IOMAP_WRITE | IOMAP_FAULT, ops, page,
 471                                 iomap_page_mkwrite_actor);
 472                 if (unlikely(ret <= 0))
 473                         goto out_unlock;
 474                 offset += ret;
 475                 length -= ret;
 476         }
 477
 478         set_page_dirty(page);
 479         wait_for_stable_page(page);
 480         return VM_FAULT_LOCKED;
 481 out_unlock:
 482         unlock_page(page);
 483         return block_page_mkwrite_return(ret);
 484 }
 485 EXPORT_SYMBOL_GPL(iomap_page_mkwrite);
 486
 487 struct fiemap_ctx {
 488         struct fiemap_extent_info *fi;
 489         struct iomap prev;
 490 };
 491
 492 static int iomap_to_fiemap(struct fiemap_extent_info *fi,
 493                 struct iomap *iomap, u32 flags)
 494 {
 495         switch (iomap->type) {
 496         case IOMAP_HOLE:
 497                 /* skip holes */
 498                 return 0;
 499         case IOMAP_DELALLOC:
 500                 flags |= FIEMAP_EXTENT_DELALLOC | FIEMAP_EXTENT_UNKNOWN;
 501                 break;
 502         case IOMAP_UNWRITTEN:
 503                 flags |= FIEMAP_EXTENT_UNWRITTEN;
 504                 break;
 505         case IOMAP_MAPPED:
 506                 break;
 507         }
 508
 509         if (iomap->flags & IOMAP_F_MERGED)
 510                 flags |= FIEMAP_EXTENT_MERGED;
 511         if (iomap->flags & IOMAP_F_SHARED)
 512                 flags |= FIEMAP_EXTENT_SHARED;
 513
 514         return fiemap_fill_next_extent(fi, iomap->offset,
 515                         iomap->blkno != IOMAP_NULL_BLOCK ? iomap->blkno << 9: 0,
 516                         iomap->length, flags);
 517
 518 }
 519
 520 static loff_t
 521 iomap_fiemap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 522                 struct iomap *iomap)
 523 {
 524         struct fiemap_ctx *ctx = data;
 525         loff_t ret = length;
 526
 527         if (iomap->type == IOMAP_HOLE)
 528                 return length;
 529
 530         ret = iomap_to_fiemap(ctx->fi, &ctx->prev, 0);
 531         ctx->prev = *iomap;
 532         switch (ret) {
 533         case 0:         /* success */
 534                 return length;
 535         case 1:         /* extent array full */
 536                 return 0;
 537         default:
 538                 return ret;
 539         }
 540 }
 541
 542 int iomap_fiemap(struct inode *inode, struct fiemap_extent_info *fi,
 543                 loff_t start, loff_t len, const struct iomap_ops *ops)
 544 {
 545         struct fiemap_ctx ctx;
 546         loff_t ret;
 547
 548         memset(&ctx, 0, sizeof(ctx));
 549         ctx.fi = fi;
 550         ctx.prev.type = IOMAP_HOLE;
 551
 552         ret = fiemap_check_flags(fi, FIEMAP_FLAG_SYNC);
 553         if (ret)
 554                 return ret;
 555
 556         if (fi->fi_flags & FIEMAP_FLAG_SYNC) {
 557                 ret = filemap_write_and_wait(inode->i_mapping);
 558                 if (ret)
 559                         return ret;
 560         }
 561
 562         while (len > 0) {
 563                 ret = iomap_apply(inode, start, len, IOMAP_REPORT, ops, &ctx,
 564                                 iomap_fiemap_actor);
 565                 /* inode with no (attribute) mapping will give ENOENT */
 566                 if (ret == -ENOENT)
 567                         break;
 568                 if (ret < 0)
 569                         return ret;
 570                 if (ret == 0)
 571                         break;
 572
 573                 start += ret;
 574                 len -= ret;
 575         }
 576
 577         if (ctx.prev.type != IOMAP_HOLE) {
 578                 ret = iomap_to_fiemap(fi, &ctx.prev, FIEMAP_EXTENT_LAST);
 579                 if (ret < 0)
 580                         return ret;
 581         }
 582
 583         return 0;
 584 }
 585 EXPORT_SYMBOL_GPL(iomap_fiemap);
 586
 587 static loff_t
 588 iomap_seek_hole_actor(struct inode *inode, loff_t offset, loff_t length,
 589                       void *data, struct iomap *iomap)
 590 {
 591         switch (iomap->type) {
 592         case IOMAP_UNWRITTEN:
 593                 offset = page_cache_seek_hole_data(inode, offset, length,
 594                                                    SEEK_HOLE);
 595                 if (offset < 0)
 596                         return length;
 597                 /* fall through */
 598         case IOMAP_HOLE:
 599                 *(loff_t *)data = offset;
 600                 return 0;
 601         default:
 602                 return length;
 603         }
 604 }
 605
 606 loff_t
 607 iomap_seek_hole(struct inode *inode, loff_t offset, const struct iomap_ops *ops)
 608 {
 609         loff_t size = i_size_read(inode);
 610         loff_t length = size - offset;
 611         loff_t ret;
 612
 613         /* Nothing to be found before or beyond the end of the file. */
 614         if (offset < 0 || offset >= size)
 615                 return -ENXIO;
 616
 617         while (length > 0) {
 618                 ret = iomap_apply(inode, offset, length, IOMAP_REPORT, ops,
 619                                   &offset, iomap_seek_hole_actor);
 620                 if (ret < 0)
 621                         return ret;
 622                 if (ret == 0)
 623                         break;
 624
 625                 offset += ret;
 626                 length -= ret;
 627         }
 628
 629         return offset;
 630 }
 631 EXPORT_SYMBOL_GPL(iomap_seek_hole);
 632
 633 static loff_t
 634 iomap_seek_data_actor(struct inode *inode, loff_t offset, loff_t length,
 635                       void *data, struct iomap *iomap)
 636 {
 637         switch (iomap->type) {
 638         case IOMAP_HOLE:
 639                 return length;
 640         case IOMAP_UNWRITTEN:
 641                 offset = page_cache_seek_hole_data(inode, offset, length,
 642                                                    SEEK_DATA);
 643                 if (offset < 0)
 644                         return length;
 645                 /*FALLTHRU*/
 646         default:
 647                 *(loff_t *)data = offset;
 648                 return 0;
 649         }
 650 }
 651
 652 loff_t
 653 iomap_seek_data(struct inode *inode, loff_t offset, const struct iomap_ops *ops)
 654 {
 655         loff_t size = i_size_read(inode);
 656         loff_t length = size - offset;
 657         loff_t ret;
 658
 659         /* Nothing to be found before or beyond the end of the file. */
 660         if (offset < 0 || offset >= size)
 661                 return -ENXIO;
 662
 663         while (length > 0) {
 664                 ret = iomap_apply(inode, offset, length, IOMAP_REPORT, ops,
 665                                   &offset, iomap_seek_data_actor);
 666                 if (ret < 0)
 667                         return ret;
 668                 if (ret == 0)
 669                         break;
 670
 671                 offset += ret;
 672                 length -= ret;
 673         }
 674
 675         if (length <= 0)
 676                 return -ENXIO;
 677         return offset;
 678 }
 679 EXPORT_SYMBOL_GPL(iomap_seek_data);
 680
 681 /*
 682  * Private flags for iomap_dio, must not overlap with the public ones in
 683  * iomap.h:
 684  */
 685 #define IOMAP_DIO_WRITE         (1 << 30)
 686 #define IOMAP_DIO_DIRTY         (1 << 31)
 687
 688 struct iomap_dio {
 689         struct kiocb            *iocb;
 690         iomap_dio_end_io_t      *end_io;
 691         loff_t                  i_size;
 692         loff_t                  size;
 693         atomic_t                ref;
 694         unsigned                flags;
 695         int                     error;
 696         bool                    wait_for_completion;
 697
 698         union {
 699                 /* used during submission and for synchronous completion: */
 700                 struct {
 701                         struct iov_iter         *iter;
 702                         struct task_struct      *waiter;
 703                         struct request_queue    *last_queue;
 704                         blk_qc_t                cookie;
 705                 } submit;
 706
 707                 /* used for aio completion: */
 708                 struct {
 709                         struct work_struct      work;
 710                 } aio;
 711         };
 712 };
 713
 714 static ssize_t iomap_dio_complete(struct iomap_dio *dio)
 715 {
 716         struct kiocb *iocb = dio->iocb;
 717         struct inode *inode = file_inode(iocb->ki_filp);
 718         loff_t offset = iocb->ki_pos;
 719         ssize_t ret;
 720
 721         if (dio->end_io) {
 722                 ret = dio->end_io(iocb,
 723                                 dio->error ? dio->error : dio->size,
 724                                 dio->flags);
 725         } else {
 726                 ret = dio->error;
 727         }
 728
 729         if (likely(!ret)) {
 730                 ret = dio->size;
 731                 /* check for short read */
 732                 if (offset + ret > dio->i_size &&
 733                     !(dio->flags & IOMAP_DIO_WRITE))
 734                         ret = dio->i_size - offset;
 735                 iocb->ki_pos += ret;
 736         }
 737
 738         /*
 739          * Try again to invalidate clean pages which might have been cached by
 740          * non-direct readahead, or faulted in by get_user_pages() if the source
 741          * of the write was an mmap'ed region of the file we're writing.  Either
 742          * one is a pretty crazy thing to do, so we don't support it 100%.  If
 743          * this invalidation fails, tough, the write still worked...
 744          *
 745          * And this page cache invalidation has to be after dio->end_io(), as
 746          * some filesystems convert unwritten extents to real allocations in
 747          * end_io() when necessary, otherwise a racing buffer read would cache
 748          * zeros from unwritten extents.
 749          */
 750         if (!dio->error &&
 751             (dio->flags & IOMAP_DIO_WRITE) && inode->i_mapping->nrpages) {
 752                 int err;
 753                 err = invalidate_inode_pages2_range(inode->i_mapping,
 754                                 offset >> PAGE_SHIFT,
 755                                 (offset + dio->size - 1) >> PAGE_SHIFT);
 756                 if (err)
 757                         dio_warn_stale_pagecache(iocb->ki_filp);
 758         }
 759
 760         inode_dio_end(file_inode(iocb->ki_filp));
 761         kfree(dio);
 762
 763         return ret;
 764 }
 765
 766 static void iomap_dio_complete_work(struct work_struct *work)
 767 {
 768         struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work);
 769         struct kiocb *iocb = dio->iocb;
 770         bool is_write = (dio->flags & IOMAP_DIO_WRITE);
 771         ssize_t ret;
 772
 773         ret = iomap_dio_complete(dio);
 774         if (is_write && ret > 0)
 775                 ret = generic_write_sync(iocb, ret);
 776         iocb->ki_complete(iocb, ret, 0);
 777 }
 778
 779 /*
 780  * Set an error in the dio if none is set yet.  We have to use cmpxchg
 781  * as the submission context and the completion context(s) can race to
 782  * update the error.
 783  */
 784 static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret)
 785 {
 786         cmpxchg(&dio->error, 0, ret);
 787 }
 788
 789 static void iomap_dio_bio_end_io(struct bio *bio)
 790 {
 791         struct iomap_dio *dio = bio->bi_private;
 792         bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
 793
 794         if (bio->bi_status)
 795                 iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
 796
 797         if (atomic_dec_and_test(&dio->ref)) {
 798                 if (dio->wait_for_completion) {
 799                         struct task_struct *waiter = dio->submit.waiter;
 800                         WRITE_ONCE(dio->submit.waiter, NULL);
 801                         wake_up_process(waiter);
 802                 } else if (dio->flags & IOMAP_DIO_WRITE) {
 803                         struct inode *inode = file_inode(dio->iocb->ki_filp);
 804
 805                         INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
 806                         queue_work(inode->i_sb->s_dio_done_wq, &dio->aio.work);
 807                 } else {
 808                         iomap_dio_complete_work(&dio->aio.work);
 809                 }
 810         }
 811
 812         if (should_dirty) {
 813                 bio_check_pages_dirty(bio);
 814         } else {
 815                 struct bio_vec *bvec;
 816                 int i;
 817
 818                 bio_for_each_segment_all(bvec, bio, i)
 819                         put_page(bvec->bv_page);
 820                 bio_put(bio);
 821         }
 822 }
 823
 824 static blk_qc_t
 825 iomap_dio_zero(struct iomap_dio *dio, struct iomap *iomap, loff_t pos,
 826                 unsigned len)
 827 {
 828         struct page *page = ZERO_PAGE(0);
 829         struct bio *bio;
 830
 831         bio = bio_alloc(GFP_KERNEL, 1);
 832         bio_set_dev(bio, iomap->bdev);
 833         bio->bi_iter.bi_sector =
 834                 iomap->blkno + ((pos - iomap->offset) >> 9);
 835         bio->bi_private = dio;
 836         bio->bi_end_io = iomap_dio_bio_end_io;
 837
 838         get_page(page);
 839         if (bio_add_page(bio, page, len, 0) != len)
 840                 BUG();
 841         bio_set_op_attrs(bio, REQ_OP_WRITE, REQ_SYNC | REQ_IDLE);
 842
 843         atomic_inc(&dio->ref);
 844         return submit_bio(bio);
 845 }
 846
 847 static loff_t
 848 iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
 849                 void *data, struct iomap *iomap)
 850 {
 851         struct iomap_dio *dio = data;
 852         unsigned int blkbits = blksize_bits(bdev_logical_block_size(iomap->bdev));
 853         unsigned int fs_block_size = i_blocksize(inode), pad;
 854         unsigned int align = iov_iter_alignment(dio->submit.iter);
 855         struct iov_iter iter;
 856         struct bio *bio;
 857         bool need_zeroout = false;
 858         int nr_pages, ret;
 859
 860         if ((pos | length | align) & ((1 << blkbits) - 1))
 861                 return -EINVAL;
 862
 863         switch (iomap->type) {
 864         case IOMAP_HOLE:
 865                 if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
 866                         return -EIO;
 867                 /*FALLTHRU*/
 868         case IOMAP_UNWRITTEN:
 869                 if (!(dio->flags & IOMAP_DIO_WRITE)) {
 870                         iov_iter_zero(length, dio->submit.iter);
 871                         dio->size += length;
 872                         return length;
 873                 }
 874                 dio->flags |= IOMAP_DIO_UNWRITTEN;
 875                 need_zeroout = true;
 876                 break;
 877         case IOMAP_MAPPED:
 878                 if (iomap->flags & IOMAP_F_SHARED)
 879                         dio->flags |= IOMAP_DIO_COW;
 880                 if (iomap->flags & IOMAP_F_NEW)
 881                         need_zeroout = true;
 882                 break;
 883         default:
 884                 WARN_ON_ONCE(1);
 885                 return -EIO;
 886         }
 887
 888         /*
 889          * Operate on a partial iter trimmed to the extent we were called for.
 890          * We'll update the iter in the dio once we're done with this extent.
 891          */
 892         iter = *dio->submit.iter;
 893         iov_iter_truncate(&iter, length);
 894
 895         nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
 896         if (nr_pages <= 0)
 897                 return nr_pages;
 898
 899         if (need_zeroout) {
 900                 /* zero out from the start of the block to the write offset */
 901                 pad = pos & (fs_block_size - 1);
 902                 if (pad)
 903                         iomap_dio_zero(dio, iomap, pos - pad, pad);
 904         }
 905
 906         do {
 907                 if (dio->error)
 908                         return 0;
 909
 910                 bio = bio_alloc(GFP_KERNEL, nr_pages);
 911                 bio_set_dev(bio, iomap->bdev);
 912                 bio->bi_iter.bi_sector =
 913                         iomap->blkno + ((pos - iomap->offset) >> 9);
 914                 bio->bi_write_hint = dio->iocb->ki_hint;
 915                 bio->bi_private = dio;
 916                 bio->bi_end_io = iomap_dio_bio_end_io;
 917
 918                 ret = bio_iov_iter_get_pages(bio, &iter);
 919                 if (unlikely(ret)) {
 920                         bio_put(bio);
 921                         return ret;
 922                 }
 923
 924                 if (dio->flags & IOMAP_DIO_WRITE) {
 925                         bio_set_op_attrs(bio, REQ_OP_WRITE, REQ_SYNC | REQ_IDLE);
 926                         task_io_account_write(bio->bi_iter.bi_size);
 927                 } else {
 928                         bio_set_op_attrs(bio, REQ_OP_READ, 0);
 929                         if (dio->flags & IOMAP_DIO_DIRTY)
 930                                 bio_set_pages_dirty(bio);
 931                 }
 932
 933                 dio->size += bio->bi_iter.bi_size;
 934                 pos += bio->bi_iter.bi_size;
 935
 936                 nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
 937
 938                 atomic_inc(&dio->ref);
 939
 940                 dio->submit.last_queue = bdev_get_queue(iomap->bdev);
 941                 dio->submit.cookie = submit_bio(bio);
 942         } while (nr_pages);
 943
 944         /*
 945          * We need to zeroout the tail of a sub-block write if the extent type
 946          * requires zeroing or the write extends beyond EOF. If we don't zero
 947          * the block tail in the latter case, we can expose stale data via mmap
 948          * reads of the EOF block.
 949          */
 950         if (need_zeroout ||
 951             ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode))) {
 952                 /* zero out from the end of the write to the end of the block */
 953                 pad = pos & (fs_block_size - 1);
 954                 if (pad)
 955                         iomap_dio_zero(dio, iomap, pos, fs_block_size - pad);
 956         }
 957
 958         iov_iter_advance(dio->submit.iter, length);
 959         return length;
 960 }
 961
 962 ssize_t
 963 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 964                 const struct iomap_ops *ops, iomap_dio_end_io_t end_io)
 965 {
 966         struct address_space *mapping = iocb->ki_filp->f_mapping;
 967         struct inode *inode = file_inode(iocb->ki_filp);
 968         size_t count = iov_iter_count(iter);
 969         loff_t pos = iocb->ki_pos, start = pos;
 970         loff_t end = iocb->ki_pos + count - 1, ret = 0;
 971         unsigned int flags = IOMAP_DIRECT;
 972         struct blk_plug plug;
 973         struct iomap_dio *dio;
 974
 975         lockdep_assert_held(&inode->i_rwsem);
 976
 977         if (!count)
 978                 return 0;
 979
 980         dio = kmalloc(sizeof(*dio), GFP_KERNEL);
 981         if (!dio)
 982                 return -ENOMEM;
 983
 984         dio->iocb = iocb;
 985         atomic_set(&dio->ref, 1);
 986         dio->size = 0;
 987         dio->i_size = i_size_read(inode);
 988         dio->end_io = end_io;
 989         dio->error = 0;
 990         dio->flags = 0;
 991         dio->wait_for_completion = is_sync_kiocb(iocb);
 992
 993         dio->submit.iter = iter;
 994         dio->submit.waiter = current;
 995         dio->submit.cookie = BLK_QC_T_NONE;
 996         dio->submit.last_queue = NULL;
 997
 998         if (iov_iter_rw(iter) == READ) {
 999                 if (pos >= dio->i_size)
1000                         goto out_free_dio;
1001
1002                 if (iter->type == ITER_IOVEC)
1003                         dio->flags |= IOMAP_DIO_DIRTY;
1004         } else {
1005                 dio->flags |= IOMAP_DIO_WRITE;
1006                 flags |= IOMAP_WRITE;
1007         }
1008
1009         if (iocb->ki_flags & IOCB_NOWAIT) {
1010                 if (filemap_range_has_page(mapping, start, end)) {
1011                         ret = -EAGAIN;
1012                         goto out_free_dio;
1013                 }
1014                 flags |= IOMAP_NOWAIT;
1015         }
1016
1017         ret = filemap_write_and_wait_range(mapping, start, end);
1018         if (ret)
1019                 goto out_free_dio;
1020
1021         /*
1022          * Try to invalidate cache pages for the range we're direct
1023          * writing.  If this invalidation fails, tough, the write will
1024          * still work, but racing two incompatible write paths is a
1025          * pretty crazy thing to do, so we don't support it 100%.
1026          */
1027         ret = invalidate_inode_pages2_range(mapping,
1028                         start >> PAGE_SHIFT, end >> PAGE_SHIFT);
1029         if (ret)
1030                 dio_warn_stale_pagecache(iocb->ki_filp);
1031         ret = 0;
1032
1033         if (iov_iter_rw(iter) == WRITE && !dio->wait_for_completion &&
1034             !inode->i_sb->s_dio_done_wq) {
1035                 ret = sb_init_dio_done_wq(inode->i_sb);
1036                 if (ret < 0)
1037                         goto out_free_dio;
1038         }
1039
1040         inode_dio_begin(inode);
1041
1042         blk_start_plug(&plug);
1043         do {
1044                 ret = iomap_apply(inode, pos, count, flags, ops, dio,
1045                                 iomap_dio_actor);
1046                 if (ret <= 0) {
1047                         /* magic error code to fall back to buffered I/O */
1048                         if (ret == -ENOTBLK) {
1049                                 dio->wait_for_completion = true;
1050                                 ret = 0;
1051                         }
1052                         break;
1053                 }
1054                 pos += ret;
1055
1056                 if (iov_iter_rw(iter) == READ && pos >= dio->i_size) {
1057                         /*
1058                          * We only report that we've read data up to i_size.
1059                          * Revert iter to a state corresponding to that as
1060                          * some callers (such as splice code) rely on it.
1061                          */
1062                         iov_iter_revert(iter, pos - dio->i_size);
1063                         break;
1064                 }
1065         } while ((count = iov_iter_count(iter)) > 0);
1066         blk_finish_plug(&plug);
1067
1068         if (ret < 0)
1069                 iomap_dio_set_error(dio, ret);
1070
1071         if (!atomic_dec_and_test(&dio->ref)) {
1072                 if (!dio->wait_for_completion)
1073                         return -EIOCBQUEUED;
1074
1075                 for (;;) {
1076                         set_current_state(TASK_UNINTERRUPTIBLE);
1077                         if (!READ_ONCE(dio->submit.waiter))
1078                                 break;
1079
1080                         if (!(iocb->ki_flags & IOCB_HIPRI) ||
1081                             !dio->submit.last_queue ||
1082                             !blk_mq_poll(dio->submit.last_queue,
1083                                          dio->submit.cookie))
1084                                 io_schedule();
1085                 }
1086                 __set_current_state(TASK_RUNNING);
1087         }
1088
1089         ret = iomap_dio_complete(dio);
1090
1091         return ret;
1092
1093 out_free_dio:
1094         kfree(dio);
1095         return ret;
1096 }
1097 EXPORT_SYMBOL_GPL(iomap_dio_rw);