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         if (WARN_ON(iomap.length == 0))
  69                 return -EIO;
  70
  71         /*
  72          * Cut down the length to the one actually provided by the filesystem,
  73          * as it might not be able to give us the whole size that we requested.
  74          */
  75         if (iomap.offset + iomap.length < pos + length)
  76                 length = iomap.offset + iomap.length - pos;
  77
  78         /*
  79          * Now that we have guaranteed that the space allocation will succeed.
  80          * we can do the copy-in page by page without having to worry about
  81          * failures exposing transient data.
  82          */
  83         written = actor(inode, pos, length, data, &iomap);
  84
  85         /*
  86          * Now the data has been copied, commit the range we've copied.  This
  87          * should not fail unless the filesystem has had a fatal error.
  88          */
  89         if (ops->iomap_end) {
  90                 ret = ops->iomap_end(inode, pos, length,
  91                                      written > 0 ? written : 0,
  92                                      flags, &iomap);
  93         }
  94
  95         return written ? written : ret;
  96 }
  97
  98 static void
  99 iomap_write_failed(struct inode *inode, loff_t pos, unsigned len)
 100 {
 101         loff_t i_size = i_size_read(inode);
 102
 103         /*
 104          * Only truncate newly allocated pages beyoned EOF, even if the
 105          * write started inside the existing inode size.
 106          */
 107         if (pos + len > i_size)
 108                 truncate_pagecache_range(inode, max(pos, i_size), pos + len);
 109 }
 110
 111 static int
 112 iomap_write_begin(struct inode *inode, loff_t pos, unsigned len, unsigned flags,
 113                 struct page **pagep, struct iomap *iomap)
 114 {
 115         pgoff_t index = pos >> PAGE_SHIFT;
 116         struct page *page;
 117         int status = 0;
 118
 119         BUG_ON(pos + len > iomap->offset + iomap->length);
 120
 121         if (fatal_signal_pending(current))
 122                 return -EINTR;
 123
 124         page = grab_cache_page_write_begin(inode->i_mapping, index, flags);
 125         if (!page)
 126                 return -ENOMEM;
 127
 128         status = __block_write_begin_int(page, pos, len, NULL, iomap);
 129         if (unlikely(status)) {
 130                 unlock_page(page);
 131                 put_page(page);
 132                 page = NULL;
 133
 134                 iomap_write_failed(inode, pos, len);
 135         }
 136
 137         *pagep = page;
 138         return status;
 139 }
 140
 141 static int
 142 iomap_write_end(struct inode *inode, loff_t pos, unsigned len,
 143                 unsigned copied, struct page *page)
 144 {
 145         int ret;
 146
 147         ret = generic_write_end(NULL, inode->i_mapping, pos, len,
 148                         copied, page, NULL);
 149         if (ret < len)
 150                 iomap_write_failed(inode, pos, len);
 151         return ret;
 152 }
 153
 154 static loff_t
 155 iomap_write_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 156                 struct iomap *iomap)
 157 {
 158         struct iov_iter *i = data;
 159         long status = 0;
 160         ssize_t written = 0;
 161         unsigned int flags = AOP_FLAG_NOFS;
 162
 163         do {
 164                 struct page *page;
 165                 unsigned long offset;   /* Offset into pagecache page */
 166                 unsigned long bytes;    /* Bytes to write to page */
 167                 size_t copied;          /* Bytes copied from user */
 168
 169                 offset = (pos & (PAGE_SIZE - 1));
 170                 bytes = min_t(unsigned long, PAGE_SIZE - offset,
 171                                                 iov_iter_count(i));
 172 again:
 173                 if (bytes > length)
 174                         bytes = length;
 175
 176                 /*
 177                  * Bring in the user page that we will copy from _first_.
 178                  * Otherwise there's a nasty deadlock on copying from the
 179                  * same page as we're writing to, without it being marked
 180                  * up-to-date.
 181                  *
 182                  * Not only is this an optimisation, but it is also required
 183                  * to check that the address is actually valid, when atomic
 184                  * usercopies are used, below.
 185                  */
 186                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
 187                         status = -EFAULT;
 188                         break;
 189                 }
 190
 191                 status = iomap_write_begin(inode, pos, bytes, flags, &page,
 192                                 iomap);
 193                 if (unlikely(status))
 194                         break;
 195
 196                 if (mapping_writably_mapped(inode->i_mapping))
 197                         flush_dcache_page(page);
 198
 199                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
 200
 201                 flush_dcache_page(page);
 202
 203                 status = iomap_write_end(inode, pos, bytes, copied, page);
 204                 if (unlikely(status < 0))
 205                         break;
 206                 copied = status;
 207
 208                 cond_resched();
 209
 210                 iov_iter_advance(i, copied);
 211                 if (unlikely(copied == 0)) {
 212                         /*
 213                          * If we were unable to copy any data at all, we must
 214                          * fall back to a single segment length write.
 215                          *
 216                          * If we didn't fallback here, we could livelock
 217                          * because not all segments in the iov can be copied at
 218                          * once without a pagefault.
 219                          */
 220                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
 221                                                 iov_iter_single_seg_count(i));
 222                         goto again;
 223                 }
 224                 pos += copied;
 225                 written += copied;
 226                 length -= copied;
 227
 228                 balance_dirty_pages_ratelimited(inode->i_mapping);
 229         } while (iov_iter_count(i) && length);
 230
 231         return written ? written : status;
 232 }
 233
 234 ssize_t
 235 iomap_file_buffered_write(struct kiocb *iocb, struct iov_iter *iter,
 236                 const struct iomap_ops *ops)
 237 {
 238         struct inode *inode = iocb->ki_filp->f_mapping->host;
 239         loff_t pos = iocb->ki_pos, ret = 0, written = 0;
 240
 241         while (iov_iter_count(iter)) {
 242                 ret = iomap_apply(inode, pos, iov_iter_count(iter),
 243                                 IOMAP_WRITE, ops, iter, iomap_write_actor);
 244                 if (ret <= 0)
 245                         break;
 246                 pos += ret;
 247                 written += ret;
 248         }
 249
 250         return written ? written : ret;
 251 }
 252 EXPORT_SYMBOL_GPL(iomap_file_buffered_write);
 253
 254 static struct page *
 255 __iomap_read_page(struct inode *inode, loff_t offset)
 256 {
 257         struct address_space *mapping = inode->i_mapping;
 258         struct page *page;
 259
 260         page = read_mapping_page(mapping, offset >> PAGE_SHIFT, NULL);
 261         if (IS_ERR(page))
 262                 return page;
 263         if (!PageUptodate(page)) {
 264                 put_page(page);
 265                 return ERR_PTR(-EIO);
 266         }
 267         return page;
 268 }
 269
 270 static loff_t
 271 iomap_dirty_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 272                 struct iomap *iomap)
 273 {
 274         long status = 0;
 275         ssize_t written = 0;
 276
 277         do {
 278                 struct page *page, *rpage;
 279                 unsigned long offset;   /* Offset into pagecache page */
 280                 unsigned long bytes;    /* Bytes to write to page */
 281
 282                 offset = (pos & (PAGE_SIZE - 1));
 283                 bytes = min_t(loff_t, PAGE_SIZE - offset, length);
 284
 285                 rpage = __iomap_read_page(inode, pos);
 286                 if (IS_ERR(rpage))
 287                         return PTR_ERR(rpage);
 288
 289                 status = iomap_write_begin(inode, pos, bytes,
 290                                            AOP_FLAG_NOFS, &page, iomap);
 291                 put_page(rpage);
 292                 if (unlikely(status))
 293                         return status;
 294
 295                 WARN_ON_ONCE(!PageUptodate(page));
 296
 297                 status = iomap_write_end(inode, pos, bytes, bytes, page);
 298                 if (unlikely(status <= 0)) {
 299                         if (WARN_ON_ONCE(status == 0))
 300                                 return -EIO;
 301                         return status;
 302                 }
 303
 304                 cond_resched();
 305
 306                 pos += status;
 307                 written += status;
 308                 length -= status;
 309
 310                 balance_dirty_pages_ratelimited(inode->i_mapping);
 311         } while (length);
 312
 313         return written;
 314 }
 315
 316 int
 317 iomap_file_dirty(struct inode *inode, loff_t pos, loff_t len,
 318                 const struct iomap_ops *ops)
 319 {
 320         loff_t ret;
 321
 322         while (len) {
 323                 ret = iomap_apply(inode, pos, len, IOMAP_WRITE, ops, NULL,
 324                                 iomap_dirty_actor);
 325                 if (ret <= 0)
 326                         return ret;
 327                 pos += ret;
 328                 len -= ret;
 329         }
 330
 331         return 0;
 332 }
 333 EXPORT_SYMBOL_GPL(iomap_file_dirty);
 334
 335 static int iomap_zero(struct inode *inode, loff_t pos, unsigned offset,
 336                 unsigned bytes, struct iomap *iomap)
 337 {
 338         struct page *page;
 339         int status;
 340
 341         status = iomap_write_begin(inode, pos, bytes, AOP_FLAG_NOFS, &page,
 342                                    iomap);
 343         if (status)
 344                 return status;
 345
 346         zero_user(page, offset, bytes);
 347         mark_page_accessed(page);
 348
 349         return iomap_write_end(inode, pos, bytes, bytes, page);
 350 }
 351
 352 static int iomap_dax_zero(loff_t pos, unsigned offset, unsigned bytes,
 353                 struct iomap *iomap)
 354 {
 355         sector_t sector = (iomap->addr +
 356                            (pos & PAGE_MASK) - iomap->offset) >> 9;
 357
 358         return __dax_zero_page_range(iomap->bdev, iomap->dax_dev, sector,
 359                         offset, bytes);
 360 }
 361
 362 static loff_t
 363 iomap_zero_range_actor(struct inode *inode, loff_t pos, loff_t count,
 364                 void *data, struct iomap *iomap)
 365 {
 366         bool *did_zero = data;
 367         loff_t written = 0;
 368         int status;
 369
 370         /* already zeroed?  we're done. */
 371         if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
 372                 return count;
 373
 374         do {
 375                 unsigned offset, bytes;
 376
 377                 offset = pos & (PAGE_SIZE - 1); /* Within page */
 378                 bytes = min_t(loff_t, PAGE_SIZE - offset, count);
 379
 380                 if (IS_DAX(inode))
 381                         status = iomap_dax_zero(pos, offset, bytes, iomap);
 382                 else
 383                         status = iomap_zero(inode, pos, offset, bytes, iomap);
 384                 if (status < 0)
 385                         return status;
 386
 387                 pos += bytes;
 388                 count -= bytes;
 389                 written += bytes;
 390                 if (did_zero)
 391                         *did_zero = true;
 392         } while (count > 0);
 393
 394         return written;
 395 }
 396
 397 int
 398 iomap_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
 399                 const struct iomap_ops *ops)
 400 {
 401         loff_t ret;
 402
 403         while (len > 0) {
 404                 ret = iomap_apply(inode, pos, len, IOMAP_ZERO,
 405                                 ops, did_zero, iomap_zero_range_actor);
 406                 if (ret <= 0)
 407                         return ret;
 408
 409                 pos += ret;
 410                 len -= ret;
 411         }
 412
 413         return 0;
 414 }
 415 EXPORT_SYMBOL_GPL(iomap_zero_range);
 416
 417 int
 418 iomap_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
 419                 const struct iomap_ops *ops)
 420 {
 421         unsigned int blocksize = i_blocksize(inode);
 422         unsigned int off = pos & (blocksize - 1);
 423
 424         /* Block boundary? Nothing to do */
 425         if (!off)
 426                 return 0;
 427         return iomap_zero_range(inode, pos, blocksize - off, did_zero, ops);
 428 }
 429 EXPORT_SYMBOL_GPL(iomap_truncate_page);
 430
 431 static loff_t
 432 iomap_page_mkwrite_actor(struct inode *inode, loff_t pos, loff_t length,
 433                 void *data, struct iomap *iomap)
 434 {
 435         struct page *page = data;
 436         int ret;
 437
 438         ret = __block_write_begin_int(page, pos, length, NULL, iomap);
 439         if (ret)
 440                 return ret;
 441
 442         block_commit_write(page, 0, length);
 443         return length;
 444 }
 445
 446 int iomap_page_mkwrite(struct vm_fault *vmf, const struct iomap_ops *ops)
 447 {
 448         struct page *page = vmf->page;
 449         struct inode *inode = file_inode(vmf->vma->vm_file);
 450         unsigned long length;
 451         loff_t offset, size;
 452         ssize_t ret;
 453
 454         lock_page(page);
 455         size = i_size_read(inode);
 456         if ((page->mapping != inode->i_mapping) ||
 457             (page_offset(page) > size)) {
 458                 /* We overload EFAULT to mean page got truncated */
 459                 ret = -EFAULT;
 460                 goto out_unlock;
 461         }
 462
 463         /* page is wholly or partially inside EOF */
 464         if (((page->index + 1) << PAGE_SHIFT) > size)
 465                 length = size & ~PAGE_MASK;
 466         else
 467                 length = PAGE_SIZE;
 468
 469         offset = page_offset(page);
 470         while (length > 0) {
 471                 ret = iomap_apply(inode, offset, length,
 472                                 IOMAP_WRITE | IOMAP_FAULT, ops, page,
 473                                 iomap_page_mkwrite_actor);
 474                 if (unlikely(ret <= 0))
 475                         goto out_unlock;
 476                 offset += ret;
 477                 length -= ret;
 478         }
 479
 480         set_page_dirty(page);
 481         wait_for_stable_page(page);
 482         return VM_FAULT_LOCKED;
 483 out_unlock:
 484         unlock_page(page);
 485         return block_page_mkwrite_return(ret);
 486 }
 487 EXPORT_SYMBOL_GPL(iomap_page_mkwrite);
 488
 489 struct fiemap_ctx {
 490         struct fiemap_extent_info *fi;
 491         struct iomap prev;
 492 };
 493
 494 static int iomap_to_fiemap(struct fiemap_extent_info *fi,
 495                 struct iomap *iomap, u32 flags)
 496 {
 497         switch (iomap->type) {
 498         case IOMAP_HOLE:
 499                 /* skip holes */
 500                 return 0;
 501         case IOMAP_DELALLOC:
 502                 flags |= FIEMAP_EXTENT_DELALLOC | FIEMAP_EXTENT_UNKNOWN;
 503                 break;
 504         case IOMAP_UNWRITTEN:
 505                 flags |= FIEMAP_EXTENT_UNWRITTEN;
 506                 break;
 507         case IOMAP_MAPPED:
 508                 break;
 509         }
 510
 511         if (iomap->flags & IOMAP_F_MERGED)
 512                 flags |= FIEMAP_EXTENT_MERGED;
 513         if (iomap->flags & IOMAP_F_SHARED)
 514                 flags |= FIEMAP_EXTENT_SHARED;
 515         if (iomap->flags & IOMAP_F_DATA_INLINE)
 516                 flags |= FIEMAP_EXTENT_DATA_INLINE;
 517
 518         return fiemap_fill_next_extent(fi, iomap->offset,
 519                         iomap->addr != IOMAP_NULL_ADDR ? iomap->addr : 0,
 520                         iomap->length, flags);
 521 }
 522
 523 static loff_t
 524 iomap_fiemap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 525                 struct iomap *iomap)
 526 {
 527         struct fiemap_ctx *ctx = data;
 528         loff_t ret = length;
 529
 530         if (iomap->type == IOMAP_HOLE)
 531                 return length;
 532
 533         ret = iomap_to_fiemap(ctx->fi, &ctx->prev, 0);
 534         ctx->prev = *iomap;
 535         switch (ret) {
 536         case 0:         /* success */
 537                 return length;
 538         case 1:         /* extent array full */
 539                 return 0;
 540         default:
 541                 return ret;
 542         }
 543 }
 544
 545 int iomap_fiemap(struct inode *inode, struct fiemap_extent_info *fi,
 546                 loff_t start, loff_t len, const struct iomap_ops *ops)
 547 {
 548         struct fiemap_ctx ctx;
 549         loff_t ret;
 550
 551         memset(&ctx, 0, sizeof(ctx));
 552         ctx.fi = fi;
 553         ctx.prev.type = IOMAP_HOLE;
 554
 555         ret = fiemap_check_flags(fi, FIEMAP_FLAG_SYNC);
 556         if (ret)
 557                 return ret;
 558
 559         if (fi->fi_flags & FIEMAP_FLAG_SYNC) {
 560                 ret = filemap_write_and_wait(inode->i_mapping);
 561                 if (ret)
 562                         return ret;
 563         }
 564
 565         while (len > 0) {
 566                 ret = iomap_apply(inode, start, len, IOMAP_REPORT, ops, &ctx,
 567                                 iomap_fiemap_actor);
 568                 /* inode with no (attribute) mapping will give ENOENT */
 569                 if (ret == -ENOENT)
 570                         break;
 571                 if (ret < 0)
 572                         return ret;
 573                 if (ret == 0)
 574                         break;
 575
 576                 start += ret;
 577                 len -= ret;
 578         }
 579
 580         if (ctx.prev.type != IOMAP_HOLE) {
 581                 ret = iomap_to_fiemap(fi, &ctx.prev, FIEMAP_EXTENT_LAST);
 582                 if (ret < 0)
 583                         return ret;
 584         }
 585
 586         return 0;
 587 }
 588 EXPORT_SYMBOL_GPL(iomap_fiemap);
 589
 590 static loff_t
 591 iomap_seek_hole_actor(struct inode *inode, loff_t offset, loff_t length,
 592                       void *data, struct iomap *iomap)
 593 {
 594         switch (iomap->type) {
 595         case IOMAP_UNWRITTEN:
 596                 offset = page_cache_seek_hole_data(inode, offset, length,
 597                                                    SEEK_HOLE);
 598                 if (offset < 0)
 599                         return length;
 600                 /* fall through */
 601         case IOMAP_HOLE:
 602                 *(loff_t *)data = offset;
 603                 return 0;
 604         default:
 605                 return length;
 606         }
 607 }
 608
 609 loff_t
 610 iomap_seek_hole(struct inode *inode, loff_t offset, const struct iomap_ops *ops)
 611 {
 612         loff_t size = i_size_read(inode);
 613         loff_t length = size - offset;
 614         loff_t ret;
 615
 616         /* Nothing to be found before or beyond the end of the file. */
 617         if (offset < 0 || offset >= size)
 618                 return -ENXIO;
 619
 620         while (length > 0) {
 621                 ret = iomap_apply(inode, offset, length, IOMAP_REPORT, ops,
 622                                   &offset, iomap_seek_hole_actor);
 623                 if (ret < 0)
 624                         return ret;
 625                 if (ret == 0)
 626                         break;
 627
 628                 offset += ret;
 629                 length -= ret;
 630         }
 631
 632         return offset;
 633 }
 634 EXPORT_SYMBOL_GPL(iomap_seek_hole);
 635
 636 static loff_t
 637 iomap_seek_data_actor(struct inode *inode, loff_t offset, loff_t length,
 638                       void *data, struct iomap *iomap)
 639 {
 640         switch (iomap->type) {
 641         case IOMAP_HOLE:
 642                 return length;
 643         case IOMAP_UNWRITTEN:
 644                 offset = page_cache_seek_hole_data(inode, offset, length,
 645                                                    SEEK_DATA);
 646                 if (offset < 0)
 647                         return length;
 648                 /*FALLTHRU*/
 649         default:
 650                 *(loff_t *)data = offset;
 651                 return 0;
 652         }
 653 }
 654
 655 loff_t
 656 iomap_seek_data(struct inode *inode, loff_t offset, const struct iomap_ops *ops)
 657 {
 658         loff_t size = i_size_read(inode);
 659         loff_t length = size - offset;
 660         loff_t ret;
 661
 662         /* Nothing to be found before or beyond the end of the file. */
 663         if (offset < 0 || offset >= size)
 664                 return -ENXIO;
 665
 666         while (length > 0) {
 667                 ret = iomap_apply(inode, offset, length, IOMAP_REPORT, ops,
 668                                   &offset, iomap_seek_data_actor);
 669                 if (ret < 0)
 670                         return ret;
 671                 if (ret == 0)
 672                         break;
 673
 674                 offset += ret;
 675                 length -= ret;
 676         }
 677
 678         if (length <= 0)
 679                 return -ENXIO;
 680         return offset;
 681 }
 682 EXPORT_SYMBOL_GPL(iomap_seek_data);
 683
 684 /*
 685  * Private flags for iomap_dio, must not overlap with the public ones in
 686  * iomap.h:
 687  */
 688 #define IOMAP_DIO_WRITE_FUA     (1 << 28)
 689 #define IOMAP_DIO_NEED_SYNC     (1 << 29)
 690 #define IOMAP_DIO_WRITE         (1 << 30)
 691 #define IOMAP_DIO_DIRTY         (1 << 31)
 692
 693 struct iomap_dio {
 694         struct kiocb            *iocb;
 695         iomap_dio_end_io_t      *end_io;
 696         loff_t                  i_size;
 697         loff_t                  size;
 698         atomic_t                ref;
 699         unsigned                flags;
 700         int                     error;
 701
 702         union {
 703                 /* used during submission and for synchronous completion: */
 704                 struct {
 705                         struct iov_iter         *iter;
 706                         struct task_struct      *waiter;
 707                         struct request_queue    *last_queue;
 708                         blk_qc_t                cookie;
 709                 } submit;
 710
 711                 /* used for aio completion: */
 712                 struct {
 713                         struct work_struct      work;
 714                 } aio;
 715         };
 716 };
 717
 718 static ssize_t iomap_dio_complete(struct iomap_dio *dio)
 719 {
 720         struct kiocb *iocb = dio->iocb;
 721         struct inode *inode = file_inode(iocb->ki_filp);
 722         loff_t offset = iocb->ki_pos;
 723         ssize_t ret;
 724
 725         if (dio->end_io) {
 726                 ret = dio->end_io(iocb,
 727                                 dio->error ? dio->error : dio->size,
 728                                 dio->flags);
 729         } else {
 730                 ret = dio->error;
 731         }
 732
 733         if (likely(!ret)) {
 734                 ret = dio->size;
 735                 /* check for short read */
 736                 if (offset + ret > dio->i_size &&
 737                     !(dio->flags & IOMAP_DIO_WRITE))
 738                         ret = dio->i_size - offset;
 739                 iocb->ki_pos += ret;
 740         }
 741
 742         /*
 743          * Try again to invalidate clean pages which might have been cached by
 744          * non-direct readahead, or faulted in by get_user_pages() if the source
 745          * of the write was an mmap'ed region of the file we're writing.  Either
 746          * one is a pretty crazy thing to do, so we don't support it 100%.  If
 747          * this invalidation fails, tough, the write still worked...
 748          *
 749          * And this page cache invalidation has to be after dio->end_io(), as
 750          * some filesystems convert unwritten extents to real allocations in
 751          * end_io() when necessary, otherwise a racing buffer read would cache
 752          * zeros from unwritten extents.
 753          */
 754         if (!dio->error &&
 755             (dio->flags & IOMAP_DIO_WRITE) && inode->i_mapping->nrpages) {
 756                 int err;
 757                 err = invalidate_inode_pages2_range(inode->i_mapping,
 758                                 offset >> PAGE_SHIFT,
 759                                 (offset + dio->size - 1) >> PAGE_SHIFT);
 760                 if (err)
 761                         dio_warn_stale_pagecache(iocb->ki_filp);
 762         }
 763
 764         /*
 765          * If this is a DSYNC write, make sure we push it to stable storage now
 766          * that we've written data.
 767          */
 768         if (ret > 0 && (dio->flags & IOMAP_DIO_NEED_SYNC))
 769                 ret = generic_write_sync(iocb, ret);
 770
 771         inode_dio_end(file_inode(iocb->ki_filp));
 772         kfree(dio);
 773
 774         return ret;
 775 }
 776
 777 static void iomap_dio_complete_work(struct work_struct *work)
 778 {
 779         struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work);
 780         struct kiocb *iocb = dio->iocb;
 781
 782         iocb->ki_complete(iocb, iomap_dio_complete(dio), 0);
 783 }
 784
 785 /*
 786  * Set an error in the dio if none is set yet.  We have to use cmpxchg
 787  * as the submission context and the completion context(s) can race to
 788  * update the error.
 789  */
 790 static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret)
 791 {
 792         cmpxchg(&dio->error, 0, ret);
 793 }
 794
 795 static void iomap_dio_bio_end_io(struct bio *bio)
 796 {
 797         struct iomap_dio *dio = bio->bi_private;
 798         bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
 799
 800         if (bio->bi_status)
 801                 iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
 802
 803         if (atomic_dec_and_test(&dio->ref)) {
 804                 if (is_sync_kiocb(dio->iocb)) {
 805                         struct task_struct *waiter = dio->submit.waiter;
 806
 807                         WRITE_ONCE(dio->submit.waiter, NULL);
 808                         wake_up_process(waiter);
 809                 } else if (dio->flags & IOMAP_DIO_WRITE) {
 810                         struct inode *inode = file_inode(dio->iocb->ki_filp);
 811
 812                         INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
 813                         queue_work(inode->i_sb->s_dio_done_wq, &dio->aio.work);
 814                 } else {
 815                         iomap_dio_complete_work(&dio->aio.work);
 816                 }
 817         }
 818
 819         if (should_dirty) {
 820                 bio_check_pages_dirty(bio);
 821         } else {
 822                 struct bio_vec *bvec;
 823                 int i;
 824
 825                 bio_for_each_segment_all(bvec, bio, i)
 826                         put_page(bvec->bv_page);
 827                 bio_put(bio);
 828         }
 829 }
 830
 831 static blk_qc_t
 832 iomap_dio_zero(struct iomap_dio *dio, struct iomap *iomap, loff_t pos,
 833                 unsigned len)
 834 {
 835         struct page *page = ZERO_PAGE(0);
 836         struct bio *bio;
 837
 838         bio = bio_alloc(GFP_KERNEL, 1);
 839         bio_set_dev(bio, iomap->bdev);
 840         bio->bi_iter.bi_sector =
 841                 (iomap->addr + pos - iomap->offset) >> 9;
 842         bio->bi_private = dio;
 843         bio->bi_end_io = iomap_dio_bio_end_io;
 844
 845         get_page(page);
 846         if (bio_add_page(bio, page, len, 0) != len)
 847                 BUG();
 848         bio_set_op_attrs(bio, REQ_OP_WRITE, REQ_SYNC | REQ_IDLE);
 849
 850         atomic_inc(&dio->ref);
 851         return submit_bio(bio);
 852 }
 853
 854 static loff_t
 855 iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
 856                 void *data, struct iomap *iomap)
 857 {
 858         struct iomap_dio *dio = data;
 859         unsigned int blkbits = blksize_bits(bdev_logical_block_size(iomap->bdev));
 860         unsigned int fs_block_size = i_blocksize(inode), pad;
 861         unsigned int align = iov_iter_alignment(dio->submit.iter);
 862         struct iov_iter iter;
 863         struct bio *bio;
 864         bool need_zeroout = false;
 865         bool use_fua = false;
 866         int nr_pages, ret;
 867         size_t copied = 0;
 868
 869         if ((pos | length | align) & ((1 << blkbits) - 1))
 870                 return -EINVAL;
 871
 872         switch (iomap->type) {
 873         case IOMAP_HOLE:
 874                 if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
 875                         return -EIO;
 876                 /*FALLTHRU*/
 877         case IOMAP_UNWRITTEN:
 878                 if (!(dio->flags & IOMAP_DIO_WRITE)) {
 879                         length = iov_iter_zero(length, dio->submit.iter);
 880                         dio->size += length;
 881                         return length;
 882                 }
 883                 dio->flags |= IOMAP_DIO_UNWRITTEN;
 884                 need_zeroout = true;
 885                 break;
 886         case IOMAP_MAPPED:
 887                 if (iomap->flags & IOMAP_F_SHARED)
 888                         dio->flags |= IOMAP_DIO_COW;
 889                 if (iomap->flags & IOMAP_F_NEW) {
 890                         need_zeroout = true;
 891                 } else {
 892                         /*
 893                          * Use a FUA write if we need datasync semantics, this
 894                          * is a pure data IO that doesn't require any metadata
 895                          * updates and the underlying device supports FUA. This
 896                          * allows us to avoid cache flushes on IO completion.
 897                          */
 898                         if (!(iomap->flags & (IOMAP_F_SHARED|IOMAP_F_DIRTY)) &&
 899                             (dio->flags & IOMAP_DIO_WRITE_FUA) &&
 900                             blk_queue_fua(bdev_get_queue(iomap->bdev)))
 901                                 use_fua = true;
 902                 }
 903                 break;
 904         default:
 905                 WARN_ON_ONCE(1);
 906                 return -EIO;
 907         }
 908
 909         /*
 910          * Operate on a partial iter trimmed to the extent we were called for.
 911          * We'll update the iter in the dio once we're done with this extent.
 912          */
 913         iter = *dio->submit.iter;
 914         iov_iter_truncate(&iter, length);
 915
 916         nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
 917         if (nr_pages <= 0)
 918                 return nr_pages;
 919
 920         if (need_zeroout) {
 921                 /* zero out from the start of the block to the write offset */
 922                 pad = pos & (fs_block_size - 1);
 923                 if (pad)
 924                         iomap_dio_zero(dio, iomap, pos - pad, pad);
 925         }
 926
 927         do {
 928                 size_t n;
 929                 if (dio->error) {
 930                         iov_iter_revert(dio->submit.iter, copied);
 931                         return 0;
 932                 }
 933
 934                 bio = bio_alloc(GFP_KERNEL, nr_pages);
 935                 bio_set_dev(bio, iomap->bdev);
 936                 bio->bi_iter.bi_sector =
 937                         (iomap->addr + pos - iomap->offset) >> 9;
 938                 bio->bi_write_hint = dio->iocb->ki_hint;
 939                 bio->bi_private = dio;
 940                 bio->bi_end_io = iomap_dio_bio_end_io;
 941
 942                 ret = bio_iov_iter_get_pages(bio, &iter);
 943                 if (unlikely(ret)) {
 944                         bio_put(bio);
 945                         return copied ? copied : ret;
 946                 }
 947
 948                 n = bio->bi_iter.bi_size;
 949                 if (dio->flags & IOMAP_DIO_WRITE) {
 950                         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
 951                         if (use_fua)
 952                                 bio->bi_opf |= REQ_FUA;
 953                         else
 954                                 dio->flags &= ~IOMAP_DIO_WRITE_FUA;
 955                         task_io_account_write(n);
 956                 } else {
 957                         bio->bi_opf = REQ_OP_READ;
 958                         if (dio->flags & IOMAP_DIO_DIRTY)
 959                                 bio_set_pages_dirty(bio);
 960                 }
 961
 962                 iov_iter_advance(dio->submit.iter, n);
 963
 964                 dio->size += n;
 965                 pos += n;
 966                 copied += n;
 967
 968                 nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
 969
 970                 atomic_inc(&dio->ref);
 971
 972                 dio->submit.last_queue = bdev_get_queue(iomap->bdev);
 973                 dio->submit.cookie = submit_bio(bio);
 974         } while (nr_pages);
 975
 976         if (need_zeroout) {
 977                 /* zero out from the end of the write to the end of the block */
 978                 pad = pos & (fs_block_size - 1);
 979                 if (pad)
 980                         iomap_dio_zero(dio, iomap, pos, fs_block_size - pad);
 981         }
 982         return copied;
 983 }
 984
 985 /*
 986  * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
 987  * is being issued as AIO or not.  This allows us to optimise pure data writes
 988  * to use REQ_FUA rather than requiring generic_write_sync() to issue a
 989  * REQ_FLUSH post write. This is slightly tricky because a single request here
 990  * can be mapped into multiple disjoint IOs and only a subset of the IOs issued
 991  * may be pure data writes. In that case, we still need to do a full data sync
 992  * completion.
 993  */
 994 ssize_t
 995 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 996                 const struct iomap_ops *ops, iomap_dio_end_io_t end_io)
 997 {
 998         struct address_space *mapping = iocb->ki_filp->f_mapping;
 999         struct inode *inode = file_inode(iocb->ki_filp);
1000         size_t count = iov_iter_count(iter);
1001         loff_t pos = iocb->ki_pos, start = pos;
1002         loff_t end = iocb->ki_pos + count - 1, ret = 0;
1003         unsigned int flags = IOMAP_DIRECT;
1004         struct blk_plug plug;
1005         struct iomap_dio *dio;
1006
1007         lockdep_assert_held(&inode->i_rwsem);
1008
1009         if (!count)
1010                 return 0;
1011
1012         dio = kmalloc(sizeof(*dio), GFP_KERNEL);
1013         if (!dio)
1014                 return -ENOMEM;
1015
1016         dio->iocb = iocb;
1017         atomic_set(&dio->ref, 1);
1018         dio->size = 0;
1019         dio->i_size = i_size_read(inode);
1020         dio->end_io = end_io;
1021         dio->error = 0;
1022         dio->flags = 0;
1023
1024         dio->submit.iter = iter;
1025         if (is_sync_kiocb(iocb)) {
1026                 dio->submit.waiter = current;
1027                 dio->submit.cookie = BLK_QC_T_NONE;
1028                 dio->submit.last_queue = NULL;
1029         }
1030
1031         if (iov_iter_rw(iter) == READ) {
1032                 if (pos >= dio->i_size)
1033                         goto out_free_dio;
1034
1035                 if (iter->type == ITER_IOVEC)
1036                         dio->flags |= IOMAP_DIO_DIRTY;
1037         } else {
1038                 flags |= IOMAP_WRITE;
1039                 dio->flags |= IOMAP_DIO_WRITE;
1040
1041                 /* for data sync or sync, we need sync completion processing */
1042                 if (iocb->ki_flags & IOCB_DSYNC)
1043                         dio->flags |= IOMAP_DIO_NEED_SYNC;
1044
1045                 /*
1046                  * For datasync only writes, we optimistically try using FUA for
1047                  * this IO.  Any non-FUA write that occurs will clear this flag,
1048                  * hence we know before completion whether a cache flush is
1049                  * necessary.
1050                  */
1051                 if ((iocb->ki_flags & (IOCB_DSYNC | IOCB_SYNC)) == IOCB_DSYNC)
1052                         dio->flags |= IOMAP_DIO_WRITE_FUA;
1053         }
1054
1055         if (iocb->ki_flags & IOCB_NOWAIT) {
1056                 if (filemap_range_has_page(mapping, start, end)) {
1057                         ret = -EAGAIN;
1058                         goto out_free_dio;
1059                 }
1060                 flags |= IOMAP_NOWAIT;
1061         }
1062
1063         ret = filemap_write_and_wait_range(mapping, start, end);
1064         if (ret)
1065                 goto out_free_dio;
1066
1067         /*
1068          * Try to invalidate cache pages for the range we're direct
1069          * writing.  If this invalidation fails, tough, the write will
1070          * still work, but racing two incompatible write paths is a
1071          * pretty crazy thing to do, so we don't support it 100%.
1072          */
1073         ret = invalidate_inode_pages2_range(mapping,
1074                         start >> PAGE_SHIFT, end >> PAGE_SHIFT);
1075         if (ret)
1076                 dio_warn_stale_pagecache(iocb->ki_filp);
1077         ret = 0;
1078
1079         if (iov_iter_rw(iter) == WRITE && !is_sync_kiocb(iocb) &&
1080             !inode->i_sb->s_dio_done_wq) {
1081                 ret = sb_init_dio_done_wq(inode->i_sb);
1082                 if (ret < 0)
1083                         goto out_free_dio;
1084         }
1085
1086         inode_dio_begin(inode);
1087
1088         blk_start_plug(&plug);
1089         do {
1090                 ret = iomap_apply(inode, pos, count, flags, ops, dio,
1091                                 iomap_dio_actor);
1092                 if (ret <= 0) {
1093                         /* magic error code to fall back to buffered I/O */
1094                         if (ret == -ENOTBLK)
1095                                 ret = 0;
1096                         break;
1097                 }
1098                 pos += ret;
1099
1100                 if (iov_iter_rw(iter) == READ && pos >= dio->i_size)
1101                         break;
1102         } while ((count = iov_iter_count(iter)) > 0);
1103         blk_finish_plug(&plug);
1104
1105         if (ret < 0)
1106                 iomap_dio_set_error(dio, ret);
1107
1108         /*
1109          * If all the writes we issued were FUA, we don't need to flush the
1110          * cache on IO completion. Clear the sync flag for this case.
1111          */
1112         if (dio->flags & IOMAP_DIO_WRITE_FUA)
1113                 dio->flags &= ~IOMAP_DIO_NEED_SYNC;
1114
1115         if (!atomic_dec_and_test(&dio->ref)) {
1116                 if (!is_sync_kiocb(iocb))
1117                         return -EIOCBQUEUED;
1118
1119                 for (;;) {
1120                         set_current_state(TASK_UNINTERRUPTIBLE);
1121                         if (!READ_ONCE(dio->submit.waiter))
1122                                 break;
1123
1124                         if (!(iocb->ki_flags & IOCB_HIPRI) ||
1125                             !dio->submit.last_queue ||
1126                             !blk_poll(dio->submit.last_queue,
1127                                          dio->submit.cookie))
1128                                 io_schedule();
1129                 }
1130                 __set_current_state(TASK_RUNNING);
1131         }
1132
1133         ret = iomap_dio_complete(dio);
1134
1135         return ret;
1136
1137 out_free_dio:
1138         kfree(dio);
1139         return ret;
1140 }
1141 EXPORT_SYMBOL_GPL(iomap_dio_rw);