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