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