fs/iomap.c

   1 /*
   2  * Copyright (C) 2010 Red Hat, Inc.
   3  * Copyright (c) 2016-2018 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/migrate.h>
  21 #include <linux/mm.h>
  22 #include <linux/mm_inline.h>
  23 #include <linux/swap.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/pagevec.h>
  26 #include <linux/file.h>
  27 #include <linux/uio.h>
  28 #include <linux/backing-dev.h>
  29 #include <linux/buffer_head.h>
  30 #include <linux/task_io_accounting_ops.h>
  31 #include <linux/dax.h>
  32 #include <linux/sched/signal.h>
  33 #include <linux/swap.h>
  34
  35 #include "internal.h"
  36
  37 /*
  38  * Execute a iomap write on a segment of the mapping that spans a
  39  * contiguous range of pages that have identical block mapping state.
  40  *
  41  * This avoids the need to map pages individually, do individual allocations
  42  * for each page and most importantly avoid the need for filesystem specific
  43  * locking per page. Instead, all the operations are amortised over the entire
  44  * range of pages. It is assumed that the filesystems will lock whatever
  45  * resources they require in the iomap_begin call, and release them in the
  46  * iomap_end call.
  47  */
  48 loff_t
  49 iomap_apply(struct inode *inode, loff_t pos, loff_t length, unsigned flags,
  50                 const struct iomap_ops *ops, void *data, iomap_actor_t actor)
  51 {
  52         struct iomap iomap = { 0 };
  53         loff_t written = 0, ret;
  54
  55         /*
  56          * Need to map a range from start position for length bytes. This can
  57          * span multiple pages - it is only guaranteed to return a range of a
  58          * single type of pages (e.g. all into a hole, all mapped or all
  59          * unwritten). Failure at this point has nothing to undo.
  60          *
  61          * If allocation is required for this range, reserve the space now so
  62          * that the allocation is guaranteed to succeed later on. Once we copy
  63          * the data into the page cache pages, then we cannot fail otherwise we
  64          * expose transient stale data. If the reserve fails, we can safely
  65          * back out at this point as there is nothing to undo.
  66          */
  67         ret = ops->iomap_begin(inode, pos, length, flags, &iomap);
  68         if (ret)
  69                 return ret;
  70         if (WARN_ON(iomap.offset > pos))
  71                 return -EIO;
  72         if (WARN_ON(iomap.length == 0))
  73                 return -EIO;
  74
  75         /*
  76          * Cut down the length to the one actually provided by the filesystem,
  77          * as it might not be able to give us the whole size that we requested.
  78          */
  79         if (iomap.offset + iomap.length < pos + length)
  80                 length = iomap.offset + iomap.length - pos;
  81
  82         /*
  83          * Now that we have guaranteed that the space allocation will succeed.
  84          * we can do the copy-in page by page without having to worry about
  85          * failures exposing transient data.
  86          */
  87         written = actor(inode, pos, length, data, &iomap);
  88
  89         /*
  90          * Now the data has been copied, commit the range we've copied.  This
  91          * should not fail unless the filesystem has had a fatal error.
  92          */
  93         if (ops->iomap_end) {
  94                 ret = ops->iomap_end(inode, pos, length,
  95                                      written > 0 ? written : 0,
  96                                      flags, &iomap);
  97         }
  98
  99         return written ? written : ret;
 100 }
 101
 102 static sector_t
 103 iomap_sector(struct iomap *iomap, loff_t pos)
 104 {
 105         return (iomap->addr + pos - iomap->offset) >> SECTOR_SHIFT;
 106 }
 107
 108 static struct iomap_page *
 109 iomap_page_create(struct inode *inode, struct page *page)
 110 {
 111         struct iomap_page *iop = to_iomap_page(page);
 112
 113         if (iop || i_blocksize(inode) == PAGE_SIZE)
 114                 return iop;
 115
 116         iop = kmalloc(sizeof(*iop), GFP_NOFS | __GFP_NOFAIL);
 117         atomic_set(&iop->read_count, 0);
 118         atomic_set(&iop->write_count, 0);
 119         bitmap_zero(iop->uptodate, PAGE_SIZE / SECTOR_SIZE);
 120
 121         /*
 122          * migrate_page_move_mapping() assumes that pages with private data have
 123          * their count elevated by 1.
 124          */
 125         get_page(page);
 126         set_page_private(page, (unsigned long)iop);
 127         SetPagePrivate(page);
 128         return iop;
 129 }
 130
 131 static void
 132 iomap_page_release(struct page *page)
 133 {
 134         struct iomap_page *iop = to_iomap_page(page);
 135
 136         if (!iop)
 137                 return;
 138         WARN_ON_ONCE(atomic_read(&iop->read_count));
 139         WARN_ON_ONCE(atomic_read(&iop->write_count));
 140         ClearPagePrivate(page);
 141         set_page_private(page, 0);
 142         put_page(page);
 143         kfree(iop);
 144 }
 145
 146 /*
 147  * Calculate the range inside the page that we actually need to read.
 148  */
 149 static void
 150 iomap_adjust_read_range(struct inode *inode, struct iomap_page *iop,
 151                 loff_t *pos, loff_t length, unsigned *offp, unsigned *lenp)
 152 {
 153         unsigned block_bits = inode->i_blkbits;
 154         unsigned block_size = (1 << block_bits);
 155         unsigned poff = offset_in_page(*pos);
 156         unsigned plen = min_t(loff_t, PAGE_SIZE - poff, length);
 157         unsigned first = poff >> block_bits;
 158         unsigned last = (poff + plen - 1) >> block_bits;
 159         unsigned end = offset_in_page(i_size_read(inode)) >> block_bits;
 160
 161         /*
 162          * If the block size is smaller than the page size we need to check the
 163          * per-block uptodate status and adjust the offset and length if needed
 164          * to avoid reading in already uptodate ranges.
 165          */
 166         if (iop) {
 167                 unsigned int i;
 168
 169                 /* move forward for each leading block marked uptodate */
 170                 for (i = first; i <= last; i++) {
 171                         if (!test_bit(i, iop->uptodate))
 172                                 break;
 173                         *pos += block_size;
 174                         poff += block_size;
 175                         plen -= block_size;
 176                         first++;
 177                 }
 178
 179                 /* truncate len if we find any trailing uptodate block(s) */
 180                 for ( ; i <= last; i++) {
 181                         if (test_bit(i, iop->uptodate)) {
 182                                 plen -= (last - i + 1) * block_size;
 183                                 last = i - 1;
 184                                 break;
 185                         }
 186                 }
 187         }
 188
 189         /*
 190          * If the extent spans the block that contains the i_size we need to
 191          * handle both halves separately so that we properly zero data in the
 192          * page cache for blocks that are entirely outside of i_size.
 193          */
 194         if (first <= end && last > end)
 195                 plen -= (last - end) * block_size;
 196
 197         *offp = poff;
 198         *lenp = plen;
 199 }
 200
 201 static void
 202 iomap_set_range_uptodate(struct page *page, unsigned off, unsigned len)
 203 {
 204         struct iomap_page *iop = to_iomap_page(page);
 205         struct inode *inode = page->mapping->host;
 206         unsigned first = off >> inode->i_blkbits;
 207         unsigned last = (off + len - 1) >> inode->i_blkbits;
 208         unsigned int i;
 209         bool uptodate = true;
 210
 211         if (iop) {
 212                 for (i = 0; i < PAGE_SIZE / i_blocksize(inode); i++) {
 213                         if (i >= first && i <= last)
 214                                 set_bit(i, iop->uptodate);
 215                         else if (!test_bit(i, iop->uptodate))
 216                                 uptodate = false;
 217                 }
 218         }
 219
 220         if (uptodate && !PageError(page))
 221                 SetPageUptodate(page);
 222 }
 223
 224 static void
 225 iomap_read_finish(struct iomap_page *iop, struct page *page)
 226 {
 227         if (!iop || atomic_dec_and_test(&iop->read_count))
 228                 unlock_page(page);
 229 }
 230
 231 static void
 232 iomap_read_page_end_io(struct bio_vec *bvec, int error)
 233 {
 234         struct page *page = bvec->bv_page;
 235         struct iomap_page *iop = to_iomap_page(page);
 236
 237         if (unlikely(error)) {
 238                 ClearPageUptodate(page);
 239                 SetPageError(page);
 240         } else {
 241                 iomap_set_range_uptodate(page, bvec->bv_offset, bvec->bv_len);
 242         }
 243
 244         iomap_read_finish(iop, page);
 245 }
 246
 247 static void
 248 iomap_read_inline_data(struct inode *inode, struct page *page,
 249                 struct iomap *iomap)
 250 {
 251         size_t size = i_size_read(inode);
 252         void *addr;
 253
 254         if (PageUptodate(page))
 255                 return;
 256
 257         BUG_ON(page->index);
 258         BUG_ON(size > PAGE_SIZE - offset_in_page(iomap->inline_data));
 259
 260         addr = kmap_atomic(page);
 261         memcpy(addr, iomap->inline_data, size);
 262         memset(addr + size, 0, PAGE_SIZE - size);
 263         kunmap_atomic(addr);
 264         SetPageUptodate(page);
 265 }
 266
 267 static void
 268 iomap_read_end_io(struct bio *bio)
 269 {
 270         int error = blk_status_to_errno(bio->bi_status);
 271         struct bio_vec *bvec;
 272         int i;
 273
 274         bio_for_each_segment_all(bvec, bio, i)
 275                 iomap_read_page_end_io(bvec, error);
 276         bio_put(bio);
 277 }
 278
 279 struct iomap_readpage_ctx {
 280         struct page             *cur_page;
 281         bool                    cur_page_in_bio;
 282         bool                    is_readahead;
 283         struct bio              *bio;
 284         struct list_head        *pages;
 285 };
 286
 287 static loff_t
 288 iomap_readpage_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 289                 struct iomap *iomap)
 290 {
 291         struct iomap_readpage_ctx *ctx = data;
 292         struct page *page = ctx->cur_page;
 293         struct iomap_page *iop = iomap_page_create(inode, page);
 294         bool is_contig = false;
 295         loff_t orig_pos = pos;
 296         unsigned poff, plen;
 297         sector_t sector;
 298
 299         if (iomap->type == IOMAP_INLINE) {
 300                 WARN_ON_ONCE(pos);
 301                 iomap_read_inline_data(inode, page, iomap);
 302                 return PAGE_SIZE;
 303         }
 304
 305         /* zero post-eof blocks as the page may be mapped */
 306         iomap_adjust_read_range(inode, iop, &pos, length, &poff, &plen);
 307         if (plen == 0)
 308                 goto done;
 309
 310         if (iomap->type != IOMAP_MAPPED || pos >= i_size_read(inode)) {
 311                 zero_user(page, poff, plen);
 312                 iomap_set_range_uptodate(page, poff, plen);
 313                 goto done;
 314         }
 315
 316         ctx->cur_page_in_bio = true;
 317
 318         /*
 319          * Try to merge into a previous segment if we can.
 320          */
 321         sector = iomap_sector(iomap, pos);
 322         if (ctx->bio && bio_end_sector(ctx->bio) == sector) {
 323                 if (__bio_try_merge_page(ctx->bio, page, plen, poff))
 324                         goto done;
 325                 is_contig = true;
 326         }
 327
 328         /*
 329          * If we start a new segment we need to increase the read count, and we
 330          * need to do so before submitting any previous full bio to make sure
 331          * that we don't prematurely unlock the page.
 332          */
 333         if (iop)
 334                 atomic_inc(&iop->read_count);
 335
 336         if (!ctx->bio || !is_contig || bio_full(ctx->bio)) {
 337                 gfp_t gfp = mapping_gfp_constraint(page->mapping, GFP_KERNEL);
 338                 int nr_vecs = (length + PAGE_SIZE - 1) >> PAGE_SHIFT;
 339
 340                 if (ctx->bio)
 341                         submit_bio(ctx->bio);
 342
 343                 if (ctx->is_readahead) /* same as readahead_gfp_mask */
 344                         gfp |= __GFP_NORETRY | __GFP_NOWARN;
 345                 ctx->bio = bio_alloc(gfp, min(BIO_MAX_PAGES, nr_vecs));
 346                 ctx->bio->bi_opf = REQ_OP_READ;
 347                 if (ctx->is_readahead)
 348                         ctx->bio->bi_opf |= REQ_RAHEAD;
 349                 ctx->bio->bi_iter.bi_sector = sector;
 350                 bio_set_dev(ctx->bio, iomap->bdev);
 351                 ctx->bio->bi_end_io = iomap_read_end_io;
 352         }
 353
 354         __bio_add_page(ctx->bio, page, plen, poff);
 355 done:
 356         /*
 357          * Move the caller beyond our range so that it keeps making progress.
 358          * For that we have to include any leading non-uptodate ranges, but
 359          * we can skip trailing ones as they will be handled in the next
 360          * iteration.
 361          */
 362         return pos - orig_pos + plen;
 363 }
 364
 365 int
 366 iomap_readpage(struct page *page, const struct iomap_ops *ops)
 367 {
 368         struct iomap_readpage_ctx ctx = { .cur_page = page };
 369         struct inode *inode = page->mapping->host;
 370         unsigned poff;
 371         loff_t ret;
 372
 373         for (poff = 0; poff < PAGE_SIZE; poff += ret) {
 374                 ret = iomap_apply(inode, page_offset(page) + poff,
 375                                 PAGE_SIZE - poff, 0, ops, &ctx,
 376                                 iomap_readpage_actor);
 377                 if (ret <= 0) {
 378                         WARN_ON_ONCE(ret == 0);
 379                         SetPageError(page);
 380                         break;
 381                 }
 382         }
 383
 384         if (ctx.bio) {
 385                 submit_bio(ctx.bio);
 386                 WARN_ON_ONCE(!ctx.cur_page_in_bio);
 387         } else {
 388                 WARN_ON_ONCE(ctx.cur_page_in_bio);
 389                 unlock_page(page);
 390         }
 391
 392         /*
 393          * Just like mpage_readpages and block_read_full_page we always
 394          * return 0 and just mark the page as PageError on errors.  This
 395          * should be cleaned up all through the stack eventually.
 396          */
 397         return 0;
 398 }
 399 EXPORT_SYMBOL_GPL(iomap_readpage);
 400
 401 static struct page *
 402 iomap_next_page(struct inode *inode, struct list_head *pages, loff_t pos,
 403                 loff_t length, loff_t *done)
 404 {
 405         while (!list_empty(pages)) {
 406                 struct page *page = lru_to_page(pages);
 407
 408                 if (page_offset(page) >= (u64)pos + length)
 409                         break;
 410
 411                 list_del(&page->lru);
 412                 if (!add_to_page_cache_lru(page, inode->i_mapping, page->index,
 413                                 GFP_NOFS))
 414                         return page;
 415
 416                 /*
 417                  * If we already have a page in the page cache at index we are
 418                  * done.  Upper layers don't care if it is uptodate after the
 419                  * readpages call itself as every page gets checked again once
 420                  * actually needed.
 421                  */
 422                 *done += PAGE_SIZE;
 423                 put_page(page);
 424         }
 425
 426         return NULL;
 427 }
 428
 429 static loff_t
 430 iomap_readpages_actor(struct inode *inode, loff_t pos, loff_t length,
 431                 void *data, struct iomap *iomap)
 432 {
 433         struct iomap_readpage_ctx *ctx = data;
 434         loff_t done, ret;
 435
 436         for (done = 0; done < length; done += ret) {
 437                 if (ctx->cur_page && offset_in_page(pos + done) == 0) {
 438                         if (!ctx->cur_page_in_bio)
 439                                 unlock_page(ctx->cur_page);
 440                         put_page(ctx->cur_page);
 441                         ctx->cur_page = NULL;
 442                 }
 443                 if (!ctx->cur_page) {
 444                         ctx->cur_page = iomap_next_page(inode, ctx->pages,
 445                                         pos, length, &done);
 446                         if (!ctx->cur_page)
 447                                 break;
 448                         ctx->cur_page_in_bio = false;
 449                 }
 450                 ret = iomap_readpage_actor(inode, pos + done, length - done,
 451                                 ctx, iomap);
 452         }
 453
 454         return done;
 455 }
 456
 457 int
 458 iomap_readpages(struct address_space *mapping, struct list_head *pages,
 459                 unsigned nr_pages, const struct iomap_ops *ops)
 460 {
 461         struct iomap_readpage_ctx ctx = {
 462                 .pages          = pages,
 463                 .is_readahead   = true,
 464         };
 465         loff_t pos = page_offset(list_entry(pages->prev, struct page, lru));
 466         loff_t last = page_offset(list_entry(pages->next, struct page, lru));
 467         loff_t length = last - pos + PAGE_SIZE, ret = 0;
 468
 469         while (length > 0) {
 470                 ret = iomap_apply(mapping->host, pos, length, 0, ops,
 471                                 &ctx, iomap_readpages_actor);
 472                 if (ret <= 0) {
 473                         WARN_ON_ONCE(ret == 0);
 474                         goto done;
 475                 }
 476                 pos += ret;
 477                 length -= ret;
 478         }
 479         ret = 0;
 480 done:
 481         if (ctx.bio)
 482                 submit_bio(ctx.bio);
 483         if (ctx.cur_page) {
 484                 if (!ctx.cur_page_in_bio)
 485                         unlock_page(ctx.cur_page);
 486                 put_page(ctx.cur_page);
 487         }
 488
 489         /*
 490          * Check that we didn't lose a page due to the arcance calling
 491          * conventions..
 492          */
 493         WARN_ON_ONCE(!ret && !list_empty(ctx.pages));
 494         return ret;
 495 }
 496 EXPORT_SYMBOL_GPL(iomap_readpages);
 497
 498 /*
 499  * iomap_is_partially_uptodate checks whether blocks within a page are
 500  * uptodate or not.
 501  *
 502  * Returns true if all blocks which correspond to a file portion
 503  * we want to read within the page are uptodate.
 504  */
 505 int
 506 iomap_is_partially_uptodate(struct page *page, unsigned long from,
 507                 unsigned long count)
 508 {
 509         struct iomap_page *iop = to_iomap_page(page);
 510         struct inode *inode = page->mapping->host;
 511         unsigned len, first, last;
 512         unsigned i;
 513
 514         /* Limit range to one page */
 515         len = min_t(unsigned, PAGE_SIZE - from, count);
 516
 517         /* First and last blocks in range within page */
 518         first = from >> inode->i_blkbits;
 519         last = (from + len - 1) >> inode->i_blkbits;
 520
 521         if (iop) {
 522                 for (i = first; i <= last; i++)
 523                         if (!test_bit(i, iop->uptodate))
 524                                 return 0;
 525                 return 1;
 526         }
 527
 528         return 0;
 529 }
 530 EXPORT_SYMBOL_GPL(iomap_is_partially_uptodate);
 531
 532 int
 533 iomap_releasepage(struct page *page, gfp_t gfp_mask)
 534 {
 535         /*
 536          * mm accommodates an old ext3 case where clean pages might not have had
 537          * the dirty bit cleared. Thus, it can send actual dirty pages to
 538          * ->releasepage() via shrink_active_list(), skip those here.
 539          */
 540         if (PageDirty(page) || PageWriteback(page))
 541                 return 0;
 542         iomap_page_release(page);
 543         return 1;
 544 }
 545 EXPORT_SYMBOL_GPL(iomap_releasepage);
 546
 547 void
 548 iomap_invalidatepage(struct page *page, unsigned int offset, unsigned int len)
 549 {
 550         /*
 551          * If we are invalidating the entire page, clear the dirty state from it
 552          * and release it to avoid unnecessary buildup of the LRU.
 553          */
 554         if (offset == 0 && len == PAGE_SIZE) {
 555                 WARN_ON_ONCE(PageWriteback(page));
 556                 cancel_dirty_page(page);
 557                 iomap_page_release(page);
 558         }
 559 }
 560 EXPORT_SYMBOL_GPL(iomap_invalidatepage);
 561
 562 #ifdef CONFIG_MIGRATION
 563 int
 564 iomap_migrate_page(struct address_space *mapping, struct page *newpage,
 565                 struct page *page, enum migrate_mode mode)
 566 {
 567         int ret;
 568
 569         ret = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
 570         if (ret != MIGRATEPAGE_SUCCESS)
 571                 return ret;
 572
 573         if (page_has_private(page)) {
 574                 ClearPagePrivate(page);
 575                 get_page(newpage);
 576                 set_page_private(newpage, page_private(page));
 577                 set_page_private(page, 0);
 578                 put_page(page);
 579                 SetPagePrivate(newpage);
 580         }
 581
 582         if (mode != MIGRATE_SYNC_NO_COPY)
 583                 migrate_page_copy(newpage, page);
 584         else
 585                 migrate_page_states(newpage, page);
 586         return MIGRATEPAGE_SUCCESS;
 587 }
 588 EXPORT_SYMBOL_GPL(iomap_migrate_page);
 589 #endif /* CONFIG_MIGRATION */
 590
 591 static void
 592 iomap_write_failed(struct inode *inode, loff_t pos, unsigned len)
 593 {
 594         loff_t i_size = i_size_read(inode);
 595
 596         /*
 597          * Only truncate newly allocated pages beyoned EOF, even if the
 598          * write started inside the existing inode size.
 599          */
 600         if (pos + len > i_size)
 601                 truncate_pagecache_range(inode, max(pos, i_size), pos + len);
 602 }
 603
 604 static int
 605 iomap_read_page_sync(struct inode *inode, loff_t block_start, struct page *page,
 606                 unsigned poff, unsigned plen, unsigned from, unsigned to,
 607                 struct iomap *iomap)
 608 {
 609         struct bio_vec bvec;
 610         struct bio bio;
 611
 612         if (iomap->type != IOMAP_MAPPED || block_start >= i_size_read(inode)) {
 613                 zero_user_segments(page, poff, from, to, poff + plen);
 614                 iomap_set_range_uptodate(page, poff, plen);
 615                 return 0;
 616         }
 617
 618         bio_init(&bio, &bvec, 1);
 619         bio.bi_opf = REQ_OP_READ;
 620         bio.bi_iter.bi_sector = iomap_sector(iomap, block_start);
 621         bio_set_dev(&bio, iomap->bdev);
 622         __bio_add_page(&bio, page, plen, poff);
 623         return submit_bio_wait(&bio);
 624 }
 625
 626 static int
 627 __iomap_write_begin(struct inode *inode, loff_t pos, unsigned len,
 628                 struct page *page, struct iomap *iomap)
 629 {
 630         struct iomap_page *iop = iomap_page_create(inode, page);
 631         loff_t block_size = i_blocksize(inode);
 632         loff_t block_start = pos & ~(block_size - 1);
 633         loff_t block_end = (pos + len + block_size - 1) & ~(block_size - 1);
 634         unsigned from = offset_in_page(pos), to = from + len, poff, plen;
 635         int status = 0;
 636
 637         if (PageUptodate(page))
 638                 return 0;
 639
 640         do {
 641                 iomap_adjust_read_range(inode, iop, &block_start,
 642                                 block_end - block_start, &poff, &plen);
 643                 if (plen == 0)
 644                         break;
 645
 646                 if ((from > poff && from < poff + plen) ||
 647                     (to > poff && to < poff + plen)) {
 648                         status = iomap_read_page_sync(inode, block_start, page,
 649                                         poff, plen, from, to, iomap);
 650                         if (status)
 651                                 break;
 652                 }
 653
 654         } while ((block_start += plen) < block_end);
 655
 656         return status;
 657 }
 658
 659 static int
 660 iomap_write_begin(struct inode *inode, loff_t pos, unsigned len, unsigned flags,
 661                 struct page **pagep, struct iomap *iomap)
 662 {
 663         pgoff_t index = pos >> PAGE_SHIFT;
 664         struct page *page;
 665         int status = 0;
 666
 667         BUG_ON(pos + len > iomap->offset + iomap->length);
 668
 669         if (fatal_signal_pending(current))
 670                 return -EINTR;
 671
 672         page = grab_cache_page_write_begin(inode->i_mapping, index, flags);
 673         if (!page)
 674                 return -ENOMEM;
 675
 676         if (iomap->type == IOMAP_INLINE)
 677                 iomap_read_inline_data(inode, page, iomap);
 678         else if (iomap->flags & IOMAP_F_BUFFER_HEAD)
 679                 status = __block_write_begin_int(page, pos, len, NULL, iomap);
 680         else
 681                 status = __iomap_write_begin(inode, pos, len, page, iomap);
 682         if (unlikely(status)) {
 683                 unlock_page(page);
 684                 put_page(page);
 685                 page = NULL;
 686
 687                 iomap_write_failed(inode, pos, len);
 688         }
 689
 690         *pagep = page;
 691         return status;
 692 }
 693
 694 int
 695 iomap_set_page_dirty(struct page *page)
 696 {
 697         struct address_space *mapping = page_mapping(page);
 698         int newly_dirty;
 699
 700         if (unlikely(!mapping))
 701                 return !TestSetPageDirty(page);
 702
 703         /*
 704          * Lock out page->mem_cgroup migration to keep PageDirty
 705          * synchronized with per-memcg dirty page counters.
 706          */
 707         lock_page_memcg(page);
 708         newly_dirty = !TestSetPageDirty(page);
 709         if (newly_dirty)
 710                 __set_page_dirty(page, mapping, 0);
 711         unlock_page_memcg(page);
 712
 713         if (newly_dirty)
 714                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 715         return newly_dirty;
 716 }
 717 EXPORT_SYMBOL_GPL(iomap_set_page_dirty);
 718
 719 static int
 720 __iomap_write_end(struct inode *inode, loff_t pos, unsigned len,
 721                 unsigned copied, struct page *page, struct iomap *iomap)
 722 {
 723         flush_dcache_page(page);
 724
 725         /*
 726          * The blocks that were entirely written will now be uptodate, so we
 727          * don't have to worry about a readpage reading them and overwriting a
 728          * partial write.  However if we have encountered a short write and only
 729          * partially written into a block, it will not be marked uptodate, so a
 730          * readpage might come in and destroy our partial write.
 731          *
 732          * Do the simplest thing, and just treat any short write to a non
 733          * uptodate page as a zero-length write, and force the caller to redo
 734          * the whole thing.
 735          */
 736         if (unlikely(copied < len && !PageUptodate(page))) {
 737                 copied = 0;
 738         } else {
 739                 iomap_set_range_uptodate(page, offset_in_page(pos), len);
 740                 iomap_set_page_dirty(page);
 741         }
 742         return __generic_write_end(inode, pos, copied, page);
 743 }
 744
 745 static int
 746 iomap_write_end_inline(struct inode *inode, struct page *page,
 747                 struct iomap *iomap, loff_t pos, unsigned copied)
 748 {
 749         void *addr;
 750
 751         WARN_ON_ONCE(!PageUptodate(page));
 752         BUG_ON(pos + copied > PAGE_SIZE - offset_in_page(iomap->inline_data));
 753
 754         addr = kmap_atomic(page);
 755         memcpy(iomap->inline_data + pos, addr + pos, copied);
 756         kunmap_atomic(addr);
 757
 758         mark_inode_dirty(inode);
 759         __generic_write_end(inode, pos, copied, page);
 760         return copied;
 761 }
 762
 763 static int
 764 iomap_write_end(struct inode *inode, loff_t pos, unsigned len,
 765                 unsigned copied, struct page *page, struct iomap *iomap)
 766 {
 767         int ret;
 768
 769         if (iomap->type == IOMAP_INLINE) {
 770                 ret = iomap_write_end_inline(inode, page, iomap, pos, copied);
 771         } else if (iomap->flags & IOMAP_F_BUFFER_HEAD) {
 772                 ret = generic_write_end(NULL, inode->i_mapping, pos, len,
 773                                 copied, page, NULL);
 774         } else {
 775                 ret = __iomap_write_end(inode, pos, len, copied, page, iomap);
 776         }
 777
 778         if (iomap->page_done)
 779                 iomap->page_done(inode, pos, copied, page, iomap);
 780
 781         if (ret < len)
 782                 iomap_write_failed(inode, pos, len);
 783         return ret;
 784 }
 785
 786 static loff_t
 787 iomap_write_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 788                 struct iomap *iomap)
 789 {
 790         struct iov_iter *i = data;
 791         long status = 0;
 792         ssize_t written = 0;
 793         unsigned int flags = AOP_FLAG_NOFS;
 794
 795         do {
 796                 struct page *page;
 797                 unsigned long offset;   /* Offset into pagecache page */
 798                 unsigned long bytes;    /* Bytes to write to page */
 799                 size_t copied;          /* Bytes copied from user */
 800
 801                 offset = offset_in_page(pos);
 802                 bytes = min_t(unsigned long, PAGE_SIZE - offset,
 803                                                 iov_iter_count(i));
 804 again:
 805                 if (bytes > length)
 806                         bytes = length;
 807
 808                 /*
 809                  * Bring in the user page that we will copy from _first_.
 810                  * Otherwise there's a nasty deadlock on copying from the
 811                  * same page as we're writing to, without it being marked
 812                  * up-to-date.
 813                  *
 814                  * Not only is this an optimisation, but it is also required
 815                  * to check that the address is actually valid, when atomic
 816                  * usercopies are used, below.
 817                  */
 818                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
 819                         status = -EFAULT;
 820                         break;
 821                 }
 822
 823                 status = iomap_write_begin(inode, pos, bytes, flags, &page,
 824                                 iomap);
 825                 if (unlikely(status))
 826                         break;
 827
 828                 if (mapping_writably_mapped(inode->i_mapping))
 829                         flush_dcache_page(page);
 830
 831                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
 832
 833                 flush_dcache_page(page);
 834
 835                 status = iomap_write_end(inode, pos, bytes, copied, page,
 836                                 iomap);
 837                 if (unlikely(status < 0))
 838                         break;
 839                 copied = status;
 840
 841                 cond_resched();
 842
 843                 iov_iter_advance(i, copied);
 844                 if (unlikely(copied == 0)) {
 845                         /*
 846                          * If we were unable to copy any data at all, we must
 847                          * fall back to a single segment length write.
 848                          *
 849                          * If we didn't fallback here, we could livelock
 850                          * because not all segments in the iov can be copied at
 851                          * once without a pagefault.
 852                          */
 853                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
 854                                                 iov_iter_single_seg_count(i));
 855                         goto again;
 856                 }
 857                 pos += copied;
 858                 written += copied;
 859                 length -= copied;
 860
 861                 balance_dirty_pages_ratelimited(inode->i_mapping);
 862         } while (iov_iter_count(i) && length);
 863
 864         return written ? written : status;
 865 }
 866
 867 ssize_t
 868 iomap_file_buffered_write(struct kiocb *iocb, struct iov_iter *iter,
 869                 const struct iomap_ops *ops)
 870 {
 871         struct inode *inode = iocb->ki_filp->f_mapping->host;
 872         loff_t pos = iocb->ki_pos, ret = 0, written = 0;
 873
 874         while (iov_iter_count(iter)) {
 875                 ret = iomap_apply(inode, pos, iov_iter_count(iter),
 876                                 IOMAP_WRITE, ops, iter, iomap_write_actor);
 877                 if (ret <= 0)
 878                         break;
 879                 pos += ret;
 880                 written += ret;
 881         }
 882
 883         return written ? written : ret;
 884 }
 885 EXPORT_SYMBOL_GPL(iomap_file_buffered_write);
 886
 887 static struct page *
 888 __iomap_read_page(struct inode *inode, loff_t offset)
 889 {
 890         struct address_space *mapping = inode->i_mapping;
 891         struct page *page;
 892
 893         page = read_mapping_page(mapping, offset >> PAGE_SHIFT, NULL);
 894         if (IS_ERR(page))
 895                 return page;
 896         if (!PageUptodate(page)) {
 897                 put_page(page);
 898                 return ERR_PTR(-EIO);
 899         }
 900         return page;
 901 }
 902
 903 static loff_t
 904 iomap_dirty_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 905                 struct iomap *iomap)
 906 {
 907         long status = 0;
 908         ssize_t written = 0;
 909
 910         do {
 911                 struct page *page, *rpage;
 912                 unsigned long offset;   /* Offset into pagecache page */
 913                 unsigned long bytes;    /* Bytes to write to page */
 914
 915                 offset = offset_in_page(pos);
 916                 bytes = min_t(loff_t, PAGE_SIZE - offset, length);
 917
 918                 rpage = __iomap_read_page(inode, pos);
 919                 if (IS_ERR(rpage))
 920                         return PTR_ERR(rpage);
 921
 922                 status = iomap_write_begin(inode, pos, bytes,
 923                                            AOP_FLAG_NOFS, &page, iomap);
 924                 put_page(rpage);
 925                 if (unlikely(status))
 926                         return status;
 927
 928                 WARN_ON_ONCE(!PageUptodate(page));
 929
 930                 status = iomap_write_end(inode, pos, bytes, bytes, page, iomap);
 931                 if (unlikely(status <= 0)) {
 932                         if (WARN_ON_ONCE(status == 0))
 933                                 return -EIO;
 934                         return status;
 935                 }
 936
 937                 cond_resched();
 938
 939                 pos += status;
 940                 written += status;
 941                 length -= status;
 942
 943                 balance_dirty_pages_ratelimited(inode->i_mapping);
 944         } while (length);
 945
 946         return written;
 947 }
 948
 949 int
 950 iomap_file_dirty(struct inode *inode, loff_t pos, loff_t len,
 951                 const struct iomap_ops *ops)
 952 {
 953         loff_t ret;
 954
 955         while (len) {
 956                 ret = iomap_apply(inode, pos, len, IOMAP_WRITE, ops, NULL,
 957                                 iomap_dirty_actor);
 958                 if (ret <= 0)
 959                         return ret;
 960                 pos += ret;
 961                 len -= ret;
 962         }
 963
 964         return 0;
 965 }
 966 EXPORT_SYMBOL_GPL(iomap_file_dirty);
 967
 968 static int iomap_zero(struct inode *inode, loff_t pos, unsigned offset,
 969                 unsigned bytes, struct iomap *iomap)
 970 {
 971         struct page *page;
 972         int status;
 973
 974         status = iomap_write_begin(inode, pos, bytes, AOP_FLAG_NOFS, &page,
 975                                    iomap);
 976         if (status)
 977                 return status;
 978
 979         zero_user(page, offset, bytes);
 980         mark_page_accessed(page);
 981
 982         return iomap_write_end(inode, pos, bytes, bytes, page, iomap);
 983 }
 984
 985 static int iomap_dax_zero(loff_t pos, unsigned offset, unsigned bytes,
 986                 struct iomap *iomap)
 987 {
 988         return __dax_zero_page_range(iomap->bdev, iomap->dax_dev,
 989                         iomap_sector(iomap, pos & PAGE_MASK), offset, bytes);
 990 }
 991
 992 static loff_t
 993 iomap_zero_range_actor(struct inode *inode, loff_t pos, loff_t count,
 994                 void *data, struct iomap *iomap)
 995 {
 996         bool *did_zero = data;
 997         loff_t written = 0;
 998         int status;
 999
1000         /* already zeroed?  we're done. */
1001         if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1002                 return count;
1003
1004         do {
1005                 unsigned offset, bytes;
1006
1007                 offset = offset_in_page(pos);
1008                 bytes = min_t(loff_t, PAGE_SIZE - offset, count);
1009
1010                 if (IS_DAX(inode))
1011                         status = iomap_dax_zero(pos, offset, bytes, iomap);
1012                 else
1013                         status = iomap_zero(inode, pos, offset, bytes, iomap);
1014                 if (status < 0)
1015                         return status;
1016
1017                 pos += bytes;
1018                 count -= bytes;
1019                 written += bytes;
1020                 if (did_zero)
1021                         *did_zero = true;
1022         } while (count > 0);
1023
1024         return written;
1025 }
1026
1027 int
1028 iomap_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
1029                 const struct iomap_ops *ops)
1030 {
1031         loff_t ret;
1032
1033         while (len > 0) {
1034                 ret = iomap_apply(inode, pos, len, IOMAP_ZERO,
1035                                 ops, did_zero, iomap_zero_range_actor);
1036                 if (ret <= 0)
1037                         return ret;
1038
1039                 pos += ret;
1040                 len -= ret;
1041         }
1042
1043         return 0;
1044 }
1045 EXPORT_SYMBOL_GPL(iomap_zero_range);
1046
1047 int
1048 iomap_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
1049                 const struct iomap_ops *ops)
1050 {
1051         unsigned int blocksize = i_blocksize(inode);
1052         unsigned int off = pos & (blocksize - 1);
1053
1054         /* Block boundary? Nothing to do */
1055         if (!off)
1056                 return 0;
1057         return iomap_zero_range(inode, pos, blocksize - off, did_zero, ops);
1058 }
1059 EXPORT_SYMBOL_GPL(iomap_truncate_page);
1060
1061 static loff_t
1062 iomap_page_mkwrite_actor(struct inode *inode, loff_t pos, loff_t length,
1063                 void *data, struct iomap *iomap)
1064 {
1065         struct page *page = data;
1066         int ret;
1067
1068         if (iomap->flags & IOMAP_F_BUFFER_HEAD) {
1069                 ret = __block_write_begin_int(page, pos, length, NULL, iomap);
1070                 if (ret)
1071                         return ret;
1072                 block_commit_write(page, 0, length);
1073         } else {
1074                 WARN_ON_ONCE(!PageUptodate(page));
1075                 iomap_page_create(inode, page);
1076                 set_page_dirty(page);
1077         }
1078
1079         return length;
1080 }
1081
1082 int iomap_page_mkwrite(struct vm_fault *vmf, const struct iomap_ops *ops)
1083 {
1084         struct page *page = vmf->page;
1085         struct inode *inode = file_inode(vmf->vma->vm_file);
1086         unsigned long length;
1087         loff_t offset, size;
1088         ssize_t ret;
1089
1090         lock_page(page);
1091         size = i_size_read(inode);
1092         if ((page->mapping != inode->i_mapping) ||
1093             (page_offset(page) > size)) {
1094                 /* We overload EFAULT to mean page got truncated */
1095                 ret = -EFAULT;
1096                 goto out_unlock;
1097         }
1098
1099         /* page is wholly or partially inside EOF */
1100         if (((page->index + 1) << PAGE_SHIFT) > size)
1101                 length = offset_in_page(size);
1102         else
1103                 length = PAGE_SIZE;
1104
1105         offset = page_offset(page);
1106         while (length > 0) {
1107                 ret = iomap_apply(inode, offset, length,
1108                                 IOMAP_WRITE | IOMAP_FAULT, ops, page,
1109                                 iomap_page_mkwrite_actor);
1110                 if (unlikely(ret <= 0))
1111                         goto out_unlock;
1112                 offset += ret;
1113                 length -= ret;
1114         }
1115
1116         wait_for_stable_page(page);
1117         return VM_FAULT_LOCKED;
1118 out_unlock:
1119         unlock_page(page);
1120         return block_page_mkwrite_return(ret);
1121 }
1122 EXPORT_SYMBOL_GPL(iomap_page_mkwrite);
1123
1124 struct fiemap_ctx {
1125         struct fiemap_extent_info *fi;
1126         struct iomap prev;
1127 };
1128
1129 static int iomap_to_fiemap(struct fiemap_extent_info *fi,
1130                 struct iomap *iomap, u32 flags)
1131 {
1132         switch (iomap->type) {
1133         case IOMAP_HOLE:
1134                 /* skip holes */
1135                 return 0;
1136         case IOMAP_DELALLOC:
1137                 flags |= FIEMAP_EXTENT_DELALLOC | FIEMAP_EXTENT_UNKNOWN;
1138                 break;
1139         case IOMAP_MAPPED:
1140                 break;
1141         case IOMAP_UNWRITTEN:
1142                 flags |= FIEMAP_EXTENT_UNWRITTEN;
1143                 break;
1144         case IOMAP_INLINE:
1145                 flags |= FIEMAP_EXTENT_DATA_INLINE;
1146                 break;
1147         }
1148
1149         if (iomap->flags & IOMAP_F_MERGED)
1150                 flags |= FIEMAP_EXTENT_MERGED;
1151         if (iomap->flags & IOMAP_F_SHARED)
1152                 flags |= FIEMAP_EXTENT_SHARED;
1153
1154         return fiemap_fill_next_extent(fi, iomap->offset,
1155                         iomap->addr != IOMAP_NULL_ADDR ? iomap->addr : 0,
1156                         iomap->length, flags);
1157 }
1158
1159 static loff_t
1160 iomap_fiemap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
1161                 struct iomap *iomap)
1162 {
1163         struct fiemap_ctx *ctx = data;
1164         loff_t ret = length;
1165
1166         if (iomap->type == IOMAP_HOLE)
1167                 return length;
1168
1169         ret = iomap_to_fiemap(ctx->fi, &ctx->prev, 0);
1170         ctx->prev = *iomap;
1171         switch (ret) {
1172         case 0:         /* success */
1173                 return length;
1174         case 1:         /* extent array full */
1175                 return 0;
1176         default:
1177                 return ret;
1178         }
1179 }
1180
1181 int iomap_fiemap(struct inode *inode, struct fiemap_extent_info *fi,
1182                 loff_t start, loff_t len, const struct iomap_ops *ops)
1183 {
1184         struct fiemap_ctx ctx;
1185         loff_t ret;
1186
1187         memset(&ctx, 0, sizeof(ctx));
1188         ctx.fi = fi;
1189         ctx.prev.type = IOMAP_HOLE;
1190
1191         ret = fiemap_check_flags(fi, FIEMAP_FLAG_SYNC);
1192         if (ret)
1193                 return ret;
1194
1195         if (fi->fi_flags & FIEMAP_FLAG_SYNC) {
1196                 ret = filemap_write_and_wait(inode->i_mapping);
1197                 if (ret)
1198                         return ret;
1199         }
1200
1201         while (len > 0) {
1202                 ret = iomap_apply(inode, start, len, IOMAP_REPORT, ops, &ctx,
1203                                 iomap_fiemap_actor);
1204                 /* inode with no (attribute) mapping will give ENOENT */
1205                 if (ret == -ENOENT)
1206                         break;
1207                 if (ret < 0)
1208                         return ret;
1209                 if (ret == 0)
1210                         break;
1211
1212                 start += ret;
1213                 len -= ret;
1214         }
1215
1216         if (ctx.prev.type != IOMAP_HOLE) {
1217                 ret = iomap_to_fiemap(fi, &ctx.prev, FIEMAP_EXTENT_LAST);
1218                 if (ret < 0)
1219                         return ret;
1220         }
1221
1222         return 0;
1223 }
1224 EXPORT_SYMBOL_GPL(iomap_fiemap);
1225
1226 /*
1227  * Seek for SEEK_DATA / SEEK_HOLE within @page, starting at @lastoff.
1228  * Returns true if found and updates @lastoff to the offset in file.
1229  */
1230 static bool
1231 page_seek_hole_data(struct inode *inode, struct page *page, loff_t *lastoff,
1232                 int whence)
1233 {
1234         const struct address_space_operations *ops = inode->i_mapping->a_ops;
1235         unsigned int bsize = i_blocksize(inode), off;
1236         bool seek_data = whence == SEEK_DATA;
1237         loff_t poff = page_offset(page);
1238
1239         if (WARN_ON_ONCE(*lastoff >= poff + PAGE_SIZE))
1240                 return false;
1241
1242         if (*lastoff < poff) {
1243                 /*
1244                  * Last offset smaller than the start of the page means we found
1245                  * a hole:
1246                  */
1247                 if (whence == SEEK_HOLE)
1248                         return true;
1249                 *lastoff = poff;
1250         }
1251
1252         /*
1253          * Just check the page unless we can and should check block ranges:
1254          */
1255         if (bsize == PAGE_SIZE || !ops->is_partially_uptodate)
1256                 return PageUptodate(page) == seek_data;
1257
1258         lock_page(page);
1259         if (unlikely(page->mapping != inode->i_mapping))
1260                 goto out_unlock_not_found;
1261
1262         for (off = 0; off < PAGE_SIZE; off += bsize) {
1263                 if (offset_in_page(*lastoff) >= off + bsize)
1264                         continue;
1265                 if (ops->is_partially_uptodate(page, off, bsize) == seek_data) {
1266                         unlock_page(page);
1267                         return true;
1268                 }
1269                 *lastoff = poff + off + bsize;
1270         }
1271
1272 out_unlock_not_found:
1273         unlock_page(page);
1274         return false;
1275 }
1276
1277 /*
1278  * Seek for SEEK_DATA / SEEK_HOLE in the page cache.
1279  *
1280  * Within unwritten extents, the page cache determines which parts are holes
1281  * and which are data: uptodate buffer heads count as data; everything else
1282  * counts as a hole.
1283  *
1284  * Returns the resulting offset on successs, and -ENOENT otherwise.
1285  */
1286 static loff_t
1287 page_cache_seek_hole_data(struct inode *inode, loff_t offset, loff_t length,
1288                 int whence)
1289 {
1290         pgoff_t index = offset >> PAGE_SHIFT;
1291         pgoff_t end = DIV_ROUND_UP(offset + length, PAGE_SIZE);
1292         loff_t lastoff = offset;
1293         struct pagevec pvec;
1294
1295         if (length <= 0)
1296                 return -ENOENT;
1297
1298         pagevec_init(&pvec);
1299
1300         do {
1301                 unsigned nr_pages, i;
1302
1303                 nr_pages = pagevec_lookup_range(&pvec, inode->i_mapping, &index,
1304                                                 end - 1);
1305                 if (nr_pages == 0)
1306                         break;
1307
1308                 for (i = 0; i < nr_pages; i++) {
1309                         struct page *page = pvec.pages[i];
1310
1311                         if (page_seek_hole_data(inode, page, &lastoff, whence))
1312                                 goto check_range;
1313                         lastoff = page_offset(page) + PAGE_SIZE;
1314                 }
1315                 pagevec_release(&pvec);
1316         } while (index < end);
1317
1318         /* When no page at lastoff and we are not done, we found a hole. */
1319         if (whence != SEEK_HOLE)
1320                 goto not_found;
1321
1322 check_range:
1323         if (lastoff < offset + length)
1324                 goto out;
1325 not_found:
1326         lastoff = -ENOENT;
1327 out:
1328         pagevec_release(&pvec);
1329         return lastoff;
1330 }
1331
1332
1333 static loff_t
1334 iomap_seek_hole_actor(struct inode *inode, loff_t offset, loff_t length,
1335                       void *data, struct iomap *iomap)
1336 {
1337         switch (iomap->type) {
1338         case IOMAP_UNWRITTEN:
1339                 offset = page_cache_seek_hole_data(inode, offset, length,
1340                                                    SEEK_HOLE);
1341                 if (offset < 0)
1342                         return length;
1343                 /* fall through */
1344         case IOMAP_HOLE:
1345                 *(loff_t *)data = offset;
1346                 return 0;
1347         default:
1348                 return length;
1349         }
1350 }
1351
1352 loff_t
1353 iomap_seek_hole(struct inode *inode, loff_t offset, const struct iomap_ops *ops)
1354 {
1355         loff_t size = i_size_read(inode);
1356         loff_t length = size - offset;
1357         loff_t ret;
1358
1359         /* Nothing to be found before or beyond the end of the file. */
1360         if (offset < 0 || offset >= size)
1361                 return -ENXIO;
1362
1363         while (length > 0) {
1364                 ret = iomap_apply(inode, offset, length, IOMAP_REPORT, ops,
1365                                   &offset, iomap_seek_hole_actor);
1366                 if (ret < 0)
1367                         return ret;
1368                 if (ret == 0)
1369                         break;
1370
1371                 offset += ret;
1372                 length -= ret;
1373         }
1374
1375         return offset;
1376 }
1377 EXPORT_SYMBOL_GPL(iomap_seek_hole);
1378
1379 static loff_t
1380 iomap_seek_data_actor(struct inode *inode, loff_t offset, loff_t length,
1381                       void *data, struct iomap *iomap)
1382 {
1383         switch (iomap->type) {
1384         case IOMAP_HOLE:
1385                 return length;
1386         case IOMAP_UNWRITTEN:
1387                 offset = page_cache_seek_hole_data(inode, offset, length,
1388                                                    SEEK_DATA);
1389                 if (offset < 0)
1390                         return length;
1391                 /*FALLTHRU*/
1392         default:
1393                 *(loff_t *)data = offset;
1394                 return 0;
1395         }
1396 }
1397
1398 loff_t
1399 iomap_seek_data(struct inode *inode, loff_t offset, const struct iomap_ops *ops)
1400 {
1401         loff_t size = i_size_read(inode);
1402         loff_t length = size - offset;
1403         loff_t ret;
1404
1405         /* Nothing to be found before or beyond the end of the file. */
1406         if (offset < 0 || offset >= size)
1407                 return -ENXIO;
1408
1409         while (length > 0) {
1410                 ret = iomap_apply(inode, offset, length, IOMAP_REPORT, ops,
1411                                   &offset, iomap_seek_data_actor);
1412                 if (ret < 0)
1413                         return ret;
1414                 if (ret == 0)
1415                         break;
1416
1417                 offset += ret;
1418                 length -= ret;
1419         }
1420
1421         if (length <= 0)
1422                 return -ENXIO;
1423         return offset;
1424 }
1425 EXPORT_SYMBOL_GPL(iomap_seek_data);
1426
1427 /*
1428  * Private flags for iomap_dio, must not overlap with the public ones in
1429  * iomap.h:
1430  */
1431 #define IOMAP_DIO_WRITE_FUA     (1 << 28)
1432 #define IOMAP_DIO_NEED_SYNC     (1 << 29)
1433 #define IOMAP_DIO_WRITE         (1 << 30)
1434 #define IOMAP_DIO_DIRTY         (1 << 31)
1435
1436 struct iomap_dio {
1437         struct kiocb            *iocb;
1438         iomap_dio_end_io_t      *end_io;
1439         loff_t                  i_size;
1440         loff_t                  size;
1441         atomic_t                ref;
1442         unsigned                flags;
1443         int                     error;
1444         bool                    wait_for_completion;
1445
1446         union {
1447                 /* used during submission and for synchronous completion: */
1448                 struct {
1449                         struct iov_iter         *iter;
1450                         struct task_struct      *waiter;
1451                         struct request_queue    *last_queue;
1452                         blk_qc_t                cookie;
1453                 } submit;
1454
1455                 /* used for aio completion: */
1456                 struct {
1457                         struct work_struct      work;
1458                 } aio;
1459         };
1460 };
1461
1462 static ssize_t iomap_dio_complete(struct iomap_dio *dio)
1463 {
1464         struct kiocb *iocb = dio->iocb;
1465         struct inode *inode = file_inode(iocb->ki_filp);
1466         loff_t offset = iocb->ki_pos;
1467         ssize_t ret;
1468
1469         if (dio->end_io) {
1470                 ret = dio->end_io(iocb,
1471                                 dio->error ? dio->error : dio->size,
1472                                 dio->flags);
1473         } else {
1474                 ret = dio->error;
1475         }
1476
1477         if (likely(!ret)) {
1478                 ret = dio->size;
1479                 /* check for short read */
1480                 if (offset + ret > dio->i_size &&
1481                     !(dio->flags & IOMAP_DIO_WRITE))
1482                         ret = dio->i_size - offset;
1483                 iocb->ki_pos += ret;
1484         }
1485
1486         /*
1487          * Try again to invalidate clean pages which might have been cached by
1488          * non-direct readahead, or faulted in by get_user_pages() if the source
1489          * of the write was an mmap'ed region of the file we're writing.  Either
1490          * one is a pretty crazy thing to do, so we don't support it 100%.  If
1491          * this invalidation fails, tough, the write still worked...
1492          *
1493          * And this page cache invalidation has to be after dio->end_io(), as
1494          * some filesystems convert unwritten extents to real allocations in
1495          * end_io() when necessary, otherwise a racing buffer read would cache
1496          * zeros from unwritten extents.
1497          */
1498         if (!dio->error &&
1499             (dio->flags & IOMAP_DIO_WRITE) && inode->i_mapping->nrpages) {
1500                 int err;
1501                 err = invalidate_inode_pages2_range(inode->i_mapping,
1502                                 offset >> PAGE_SHIFT,
1503                                 (offset + dio->size - 1) >> PAGE_SHIFT);
1504                 if (err)
1505                         dio_warn_stale_pagecache(iocb->ki_filp);
1506         }
1507
1508         /*
1509          * If this is a DSYNC write, make sure we push it to stable storage now
1510          * that we've written data.
1511          */
1512         if (ret > 0 && (dio->flags & IOMAP_DIO_NEED_SYNC))
1513                 ret = generic_write_sync(iocb, ret);
1514
1515         inode_dio_end(file_inode(iocb->ki_filp));
1516         kfree(dio);
1517
1518         return ret;
1519 }
1520
1521 static void iomap_dio_complete_work(struct work_struct *work)
1522 {
1523         struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work);
1524         struct kiocb *iocb = dio->iocb;
1525
1526         iocb->ki_complete(iocb, iomap_dio_complete(dio), 0);
1527 }
1528
1529 /*
1530  * Set an error in the dio if none is set yet.  We have to use cmpxchg
1531  * as the submission context and the completion context(s) can race to
1532  * update the error.
1533  */
1534 static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret)
1535 {
1536         cmpxchg(&dio->error, 0, ret);
1537 }
1538
1539 static void iomap_dio_bio_end_io(struct bio *bio)
1540 {
1541         struct iomap_dio *dio = bio->bi_private;
1542         bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
1543
1544         if (bio->bi_status)
1545                 iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
1546
1547         if (atomic_dec_and_test(&dio->ref)) {
1548                 if (dio->wait_for_completion) {
1549                         struct task_struct *waiter = dio->submit.waiter;
1550                         WRITE_ONCE(dio->submit.waiter, NULL);
1551                         wake_up_process(waiter);
1552                 } else if (dio->flags & IOMAP_DIO_WRITE) {
1553                         struct inode *inode = file_inode(dio->iocb->ki_filp);
1554
1555                         INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
1556                         queue_work(inode->i_sb->s_dio_done_wq, &dio->aio.work);
1557                 } else {
1558                         iomap_dio_complete_work(&dio->aio.work);
1559                 }
1560         }
1561
1562         if (should_dirty) {
1563                 bio_check_pages_dirty(bio);
1564         } else {
1565                 struct bio_vec *bvec;
1566                 int i;
1567
1568                 bio_for_each_segment_all(bvec, bio, i)
1569                         put_page(bvec->bv_page);
1570                 bio_put(bio);
1571         }
1572 }
1573
1574 static blk_qc_t
1575 iomap_dio_zero(struct iomap_dio *dio, struct iomap *iomap, loff_t pos,
1576                 unsigned len)
1577 {
1578         struct page *page = ZERO_PAGE(0);
1579         struct bio *bio;
1580
1581         bio = bio_alloc(GFP_KERNEL, 1);
1582         bio_set_dev(bio, iomap->bdev);
1583         bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
1584         bio->bi_private = dio;
1585         bio->bi_end_io = iomap_dio_bio_end_io;
1586
1587         get_page(page);
1588         __bio_add_page(bio, page, len, 0);
1589         bio_set_op_attrs(bio, REQ_OP_WRITE, REQ_SYNC | REQ_IDLE);
1590
1591         atomic_inc(&dio->ref);
1592         return submit_bio(bio);
1593 }
1594
1595 static loff_t
1596 iomap_dio_bio_actor(struct inode *inode, loff_t pos, loff_t length,
1597                 struct iomap_dio *dio, struct iomap *iomap)
1598 {
1599         unsigned int blkbits = blksize_bits(bdev_logical_block_size(iomap->bdev));
1600         unsigned int fs_block_size = i_blocksize(inode), pad;
1601         unsigned int align = iov_iter_alignment(dio->submit.iter);
1602         struct iov_iter iter;
1603         struct bio *bio;
1604         bool need_zeroout = false;
1605         bool use_fua = false;
1606         int nr_pages, ret;
1607         size_t copied = 0;
1608
1609         if ((pos | length | align) & ((1 << blkbits) - 1))
1610                 return -EINVAL;
1611
1612         if (iomap->type == IOMAP_UNWRITTEN) {
1613                 dio->flags |= IOMAP_DIO_UNWRITTEN;
1614                 need_zeroout = true;
1615         }
1616
1617         if (iomap->flags & IOMAP_F_SHARED)
1618                 dio->flags |= IOMAP_DIO_COW;
1619
1620         if (iomap->flags & IOMAP_F_NEW) {
1621                 need_zeroout = true;
1622         } else {
1623                 /*
1624                  * Use a FUA write if we need datasync semantics, this
1625                  * is a pure data IO that doesn't require any metadata
1626                  * updates and the underlying device supports FUA. This
1627                  * allows us to avoid cache flushes on IO completion.
1628                  */
1629                 if (!(iomap->flags & (IOMAP_F_SHARED|IOMAP_F_DIRTY)) &&
1630                     (dio->flags & IOMAP_DIO_WRITE_FUA) &&
1631                     blk_queue_fua(bdev_get_queue(iomap->bdev)))
1632                         use_fua = true;
1633         }
1634
1635         /*
1636          * Operate on a partial iter trimmed to the extent we were called for.
1637          * We'll update the iter in the dio once we're done with this extent.
1638          */
1639         iter = *dio->submit.iter;
1640         iov_iter_truncate(&iter, length);
1641
1642         nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
1643         if (nr_pages <= 0)
1644                 return nr_pages;
1645
1646         if (need_zeroout) {
1647                 /* zero out from the start of the block to the write offset */
1648                 pad = pos & (fs_block_size - 1);
1649                 if (pad)
1650                         iomap_dio_zero(dio, iomap, pos - pad, pad);
1651         }
1652
1653         do {
1654                 size_t n;
1655                 if (dio->error) {
1656                         iov_iter_revert(dio->submit.iter, copied);
1657                         return 0;
1658                 }
1659
1660                 bio = bio_alloc(GFP_KERNEL, nr_pages);
1661                 bio_set_dev(bio, iomap->bdev);
1662                 bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
1663                 bio->bi_write_hint = dio->iocb->ki_hint;
1664                 bio->bi_ioprio = dio->iocb->ki_ioprio;
1665                 bio->bi_private = dio;
1666                 bio->bi_end_io = iomap_dio_bio_end_io;
1667
1668                 ret = bio_iov_iter_get_pages(bio, &iter);
1669                 if (unlikely(ret)) {
1670                         bio_put(bio);
1671                         return copied ? copied : ret;
1672                 }
1673
1674                 n = bio->bi_iter.bi_size;
1675                 if (dio->flags & IOMAP_DIO_WRITE) {
1676                         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
1677                         if (use_fua)
1678                                 bio->bi_opf |= REQ_FUA;
1679                         else
1680                                 dio->flags &= ~IOMAP_DIO_WRITE_FUA;
1681                         task_io_account_write(n);
1682                 } else {
1683                         bio->bi_opf = REQ_OP_READ;
1684                         if (dio->flags & IOMAP_DIO_DIRTY)
1685                                 bio_set_pages_dirty(bio);
1686                 }
1687
1688                 iov_iter_advance(dio->submit.iter, n);
1689
1690                 dio->size += n;
1691                 pos += n;
1692                 copied += n;
1693
1694                 nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
1695
1696                 atomic_inc(&dio->ref);
1697
1698                 dio->submit.last_queue = bdev_get_queue(iomap->bdev);
1699                 dio->submit.cookie = submit_bio(bio);
1700         } while (nr_pages);
1701
1702         if (need_zeroout) {
1703                 /* zero out from the end of the write to the end of the block */
1704                 pad = pos & (fs_block_size - 1);
1705                 if (pad)
1706                         iomap_dio_zero(dio, iomap, pos, fs_block_size - pad);
1707         }
1708         return copied;
1709 }
1710
1711 static loff_t
1712 iomap_dio_hole_actor(loff_t length, struct iomap_dio *dio)
1713 {
1714         length = iov_iter_zero(length, dio->submit.iter);
1715         dio->size += length;
1716         return length;
1717 }
1718
1719 static loff_t
1720 iomap_dio_inline_actor(struct inode *inode, loff_t pos, loff_t length,
1721                 struct iomap_dio *dio, struct iomap *iomap)
1722 {
1723         struct iov_iter *iter = dio->submit.iter;
1724         size_t copied;
1725
1726         BUG_ON(pos + length > PAGE_SIZE - offset_in_page(iomap->inline_data));
1727
1728         if (dio->flags & IOMAP_DIO_WRITE) {
1729                 loff_t size = inode->i_size;
1730
1731                 if (pos > size)
1732                         memset(iomap->inline_data + size, 0, pos - size);
1733                 copied = copy_from_iter(iomap->inline_data + pos, length, iter);
1734                 if (copied) {
1735                         if (pos + copied > size)
1736                                 i_size_write(inode, pos + copied);
1737                         mark_inode_dirty(inode);
1738                 }
1739         } else {
1740                 copied = copy_to_iter(iomap->inline_data + pos, length, iter);
1741         }
1742         dio->size += copied;
1743         return copied;
1744 }
1745
1746 static loff_t
1747 iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
1748                 void *data, struct iomap *iomap)
1749 {
1750         struct iomap_dio *dio = data;
1751
1752         switch (iomap->type) {
1753         case IOMAP_HOLE:
1754                 if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
1755                         return -EIO;
1756                 return iomap_dio_hole_actor(length, dio);
1757         case IOMAP_UNWRITTEN:
1758                 if (!(dio->flags & IOMAP_DIO_WRITE))
1759                         return iomap_dio_hole_actor(length, dio);
1760                 return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
1761         case IOMAP_MAPPED:
1762                 return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
1763         case IOMAP_INLINE:
1764                 return iomap_dio_inline_actor(inode, pos, length, dio, iomap);
1765         default:
1766                 WARN_ON_ONCE(1);
1767                 return -EIO;
1768         }
1769 }
1770
1771 /*
1772  * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
1773  * is being issued as AIO or not.  This allows us to optimise pure data writes
1774  * to use REQ_FUA rather than requiring generic_write_sync() to issue a
1775  * REQ_FLUSH post write. This is slightly tricky because a single request here
1776  * can be mapped into multiple disjoint IOs and only a subset of the IOs issued
1777  * may be pure data writes. In that case, we still need to do a full data sync
1778  * completion.
1779  */
1780 ssize_t
1781 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
1782                 const struct iomap_ops *ops, iomap_dio_end_io_t end_io)
1783 {
1784         struct address_space *mapping = iocb->ki_filp->f_mapping;
1785         struct inode *inode = file_inode(iocb->ki_filp);
1786         size_t count = iov_iter_count(iter);
1787         loff_t pos = iocb->ki_pos, start = pos;
1788         loff_t end = iocb->ki_pos + count - 1, ret = 0;
1789         unsigned int flags = IOMAP_DIRECT;
1790         bool wait_for_completion = is_sync_kiocb(iocb);
1791         struct blk_plug plug;
1792         struct iomap_dio *dio;
1793
1794         lockdep_assert_held(&inode->i_rwsem);
1795
1796         if (!count)
1797                 return 0;
1798
1799         dio = kmalloc(sizeof(*dio), GFP_KERNEL);
1800         if (!dio)
1801                 return -ENOMEM;
1802
1803         dio->iocb = iocb;
1804         atomic_set(&dio->ref, 1);
1805         dio->size = 0;
1806         dio->i_size = i_size_read(inode);
1807         dio->end_io = end_io;
1808         dio->error = 0;
1809         dio->flags = 0;
1810
1811         dio->submit.iter = iter;
1812         dio->submit.waiter = current;
1813         dio->submit.cookie = BLK_QC_T_NONE;
1814         dio->submit.last_queue = NULL;
1815
1816         if (iov_iter_rw(iter) == READ) {
1817                 if (pos >= dio->i_size)
1818                         goto out_free_dio;
1819
1820                 if (iter->type == ITER_IOVEC)
1821                         dio->flags |= IOMAP_DIO_DIRTY;
1822         } else {
1823                 flags |= IOMAP_WRITE;
1824                 dio->flags |= IOMAP_DIO_WRITE;
1825
1826                 /* for data sync or sync, we need sync completion processing */
1827                 if (iocb->ki_flags & IOCB_DSYNC)
1828                         dio->flags |= IOMAP_DIO_NEED_SYNC;
1829
1830                 /*
1831                  * For datasync only writes, we optimistically try using FUA for
1832                  * this IO.  Any non-FUA write that occurs will clear this flag,
1833                  * hence we know before completion whether a cache flush is
1834                  * necessary.
1835                  */
1836                 if ((iocb->ki_flags & (IOCB_DSYNC | IOCB_SYNC)) == IOCB_DSYNC)
1837                         dio->flags |= IOMAP_DIO_WRITE_FUA;
1838         }
1839
1840         if (iocb->ki_flags & IOCB_NOWAIT) {
1841                 if (filemap_range_has_page(mapping, start, end)) {
1842                         ret = -EAGAIN;
1843                         goto out_free_dio;
1844                 }
1845                 flags |= IOMAP_NOWAIT;
1846         }
1847
1848         ret = filemap_write_and_wait_range(mapping, start, end);
1849         if (ret)
1850                 goto out_free_dio;
1851
1852         /*
1853          * Try to invalidate cache pages for the range we're direct
1854          * writing.  If this invalidation fails, tough, the write will
1855          * still work, but racing two incompatible write paths is a
1856          * pretty crazy thing to do, so we don't support it 100%.
1857          */
1858         ret = invalidate_inode_pages2_range(mapping,
1859                         start >> PAGE_SHIFT, end >> PAGE_SHIFT);
1860         if (ret)
1861                 dio_warn_stale_pagecache(iocb->ki_filp);
1862         ret = 0;
1863
1864         if (iov_iter_rw(iter) == WRITE && !wait_for_completion &&
1865             !inode->i_sb->s_dio_done_wq) {
1866                 ret = sb_init_dio_done_wq(inode->i_sb);
1867                 if (ret < 0)
1868                         goto out_free_dio;
1869         }
1870
1871         inode_dio_begin(inode);
1872
1873         blk_start_plug(&plug);
1874         do {
1875                 ret = iomap_apply(inode, pos, count, flags, ops, dio,
1876                                 iomap_dio_actor);
1877                 if (ret <= 0) {
1878                         /* magic error code to fall back to buffered I/O */
1879                         if (ret == -ENOTBLK) {
1880                                 wait_for_completion = true;
1881                                 ret = 0;
1882                         }
1883                         break;
1884                 }
1885                 pos += ret;
1886
1887                 if (iov_iter_rw(iter) == READ && pos >= dio->i_size)
1888                         break;
1889         } while ((count = iov_iter_count(iter)) > 0);
1890         blk_finish_plug(&plug);
1891
1892         if (ret < 0)
1893                 iomap_dio_set_error(dio, ret);
1894
1895         /*
1896          * If all the writes we issued were FUA, we don't need to flush the
1897          * cache on IO completion. Clear the sync flag for this case.
1898          */
1899         if (dio->flags & IOMAP_DIO_WRITE_FUA)
1900                 dio->flags &= ~IOMAP_DIO_NEED_SYNC;
1901
1902         /*
1903          * We are about to drop our additional submission reference, which
1904          * might be the last reference to the dio.  There are three three
1905          * different ways we can progress here:
1906          *
1907          *  (a) If this is the last reference we will always complete and free
1908          *      the dio ourselves.
1909          *  (b) If this is not the last reference, and we serve an asynchronous
1910          *      iocb, we must never touch the dio after the decrement, the
1911          *      I/O completion handler will complete and free it.
1912          *  (c) If this is not the last reference, but we serve a synchronous
1913          *      iocb, the I/O completion handler will wake us up on the drop
1914          *      of the final reference, and we will complete and free it here
1915          *      after we got woken by the I/O completion handler.
1916          */
1917         dio->wait_for_completion = wait_for_completion;
1918         if (!atomic_dec_and_test(&dio->ref)) {
1919                 if (!wait_for_completion)
1920                         return -EIOCBQUEUED;
1921
1922                 for (;;) {
1923                         set_current_state(TASK_UNINTERRUPTIBLE);
1924                         if (!READ_ONCE(dio->submit.waiter))
1925                                 break;
1926
1927                         if (!(iocb->ki_flags & IOCB_HIPRI) ||
1928                             !dio->submit.last_queue ||
1929                             !blk_poll(dio->submit.last_queue,
1930                                          dio->submit.cookie))
1931                                 io_schedule();
1932                 }
1933                 __set_current_state(TASK_RUNNING);
1934         }
1935
1936         return iomap_dio_complete(dio);
1937
1938 out_free_dio:
1939         kfree(dio);
1940         return ret;
1941 }
1942 EXPORT_SYMBOL_GPL(iomap_dio_rw);
1943
1944 /* Swapfile activation */
1945
1946 #ifdef CONFIG_SWAP
1947 struct iomap_swapfile_info {
1948         struct iomap iomap;             /* accumulated iomap */
1949         struct swap_info_struct *sis;
1950         uint64_t lowest_ppage;          /* lowest physical addr seen (pages) */
1951         uint64_t highest_ppage;         /* highest physical addr seen (pages) */
1952         unsigned long nr_pages;         /* number of pages collected */
1953         int nr_extents;                 /* extent count */
1954 };
1955
1956 /*
1957  * Collect physical extents for this swap file.  Physical extents reported to
1958  * the swap code must be trimmed to align to a page boundary.  The logical
1959  * offset within the file is irrelevant since the swapfile code maps logical
1960  * page numbers of the swap device to the physical page-aligned extents.
1961  */
1962 static int iomap_swapfile_add_extent(struct iomap_swapfile_info *isi)
1963 {
1964         struct iomap *iomap = &isi->iomap;
1965         unsigned long nr_pages;
1966         uint64_t first_ppage;
1967         uint64_t first_ppage_reported;
1968         uint64_t next_ppage;
1969         int error;
1970
1971         /*
1972          * Round the start up and the end down so that the physical
1973          * extent aligns to a page boundary.
1974          */
1975         first_ppage = ALIGN(iomap->addr, PAGE_SIZE) >> PAGE_SHIFT;
1976         next_ppage = ALIGN_DOWN(iomap->addr + iomap->length, PAGE_SIZE) >>
1977                         PAGE_SHIFT;
1978
1979         /* Skip too-short physical extents. */
1980         if (first_ppage >= next_ppage)
1981                 return 0;
1982         nr_pages = next_ppage - first_ppage;
1983
1984         /*
1985          * Calculate how much swap space we're adding; the first page contains
1986          * the swap header and doesn't count.  The mm still wants that first
1987          * page fed to add_swap_extent, however.
1988          */
1989         first_ppage_reported = first_ppage;
1990         if (iomap->offset == 0)
1991                 first_ppage_reported++;
1992         if (isi->lowest_ppage > first_ppage_reported)
1993                 isi->lowest_ppage = first_ppage_reported;
1994         if (isi->highest_ppage < (next_ppage - 1))
1995                 isi->highest_ppage = next_ppage - 1;
1996
1997         /* Add extent, set up for the next call. */
1998         error = add_swap_extent(isi->sis, isi->nr_pages, nr_pages, first_ppage);
1999         if (error < 0)
2000                 return error;
2001         isi->nr_extents += error;
2002         isi->nr_pages += nr_pages;
2003         return 0;
2004 }
2005
2006 /*
2007  * Accumulate iomaps for this swap file.  We have to accumulate iomaps because
2008  * swap only cares about contiguous page-aligned physical extents and makes no
2009  * distinction between written and unwritten extents.
2010  */
2011 static loff_t iomap_swapfile_activate_actor(struct inode *inode, loff_t pos,
2012                 loff_t count, void *data, struct iomap *iomap)
2013 {
2014         struct iomap_swapfile_info *isi = data;
2015         int error;
2016
2017         switch (iomap->type) {
2018         case IOMAP_MAPPED:
2019         case IOMAP_UNWRITTEN:
2020                 /* Only real or unwritten extents. */
2021                 break;
2022         case IOMAP_INLINE:
2023                 /* No inline data. */
2024                 pr_err("swapon: file is inline\n");
2025                 return -EINVAL;
2026         default:
2027                 pr_err("swapon: file has unallocated extents\n");
2028                 return -EINVAL;
2029         }
2030
2031         /* No uncommitted metadata or shared blocks. */
2032         if (iomap->flags & IOMAP_F_DIRTY) {
2033                 pr_err("swapon: file is not committed\n");
2034                 return -EINVAL;
2035         }
2036         if (iomap->flags & IOMAP_F_SHARED) {
2037                 pr_err("swapon: file has shared extents\n");
2038                 return -EINVAL;
2039         }
2040
2041         /* Only one bdev per swap file. */
2042         if (iomap->bdev != isi->sis->bdev) {
2043                 pr_err("swapon: file is on multiple devices\n");
2044                 return -EINVAL;
2045         }
2046
2047         if (isi->iomap.length == 0) {
2048                 /* No accumulated extent, so just store it. */
2049                 memcpy(&isi->iomap, iomap, sizeof(isi->iomap));
2050         } else if (isi->iomap.addr + isi->iomap.length == iomap->addr) {
2051                 /* Append this to the accumulated extent. */
2052                 isi->iomap.length += iomap->length;
2053         } else {
2054                 /* Otherwise, add the retained iomap and store this one. */
2055                 error = iomap_swapfile_add_extent(isi);
2056                 if (error)
2057                         return error;
2058                 memcpy(&isi->iomap, iomap, sizeof(isi->iomap));
2059         }
2060         return count;
2061 }
2062
2063 /*
2064  * Iterate a swap file's iomaps to construct physical extents that can be
2065  * passed to the swapfile subsystem.
2066  */
2067 int iomap_swapfile_activate(struct swap_info_struct *sis,
2068                 struct file *swap_file, sector_t *pagespan,
2069                 const struct iomap_ops *ops)
2070 {
2071         struct iomap_swapfile_info isi = {
2072                 .sis = sis,
2073                 .lowest_ppage = (sector_t)-1ULL,
2074         };
2075         struct address_space *mapping = swap_file->f_mapping;
2076         struct inode *inode = mapping->host;
2077         loff_t pos = 0;
2078         loff_t len = ALIGN_DOWN(i_size_read(inode), PAGE_SIZE);
2079         loff_t ret;
2080
2081         /*
2082          * Persist all file mapping metadata so that we won't have any
2083          * IOMAP_F_DIRTY iomaps.
2084          */
2085         ret = vfs_fsync(swap_file, 1);
2086         if (ret)
2087                 return ret;
2088
2089         while (len > 0) {
2090                 ret = iomap_apply(inode, pos, len, IOMAP_REPORT,
2091                                 ops, &isi, iomap_swapfile_activate_actor);
2092                 if (ret <= 0)
2093                         return ret;
2094
2095                 pos += ret;
2096                 len -= ret;
2097         }
2098
2099         if (isi.iomap.length) {
2100                 ret = iomap_swapfile_add_extent(&isi);
2101                 if (ret)
2102                         return ret;
2103         }
2104
2105         *pagespan = 1 + isi.highest_ppage - isi.lowest_ppage;
2106         sis->max = isi.nr_pages;
2107         sis->pages = isi.nr_pages - 1;
2108         sis->highest_bit = isi.nr_pages - 1;
2109         return isi.nr_extents;
2110 }
2111 EXPORT_SYMBOL_GPL(iomap_swapfile_activate);
2112 #endif /* CONFIG_SWAP */
2113
2114 static loff_t
2115 iomap_bmap_actor(struct inode *inode, loff_t pos, loff_t length,
2116                 void *data, struct iomap *iomap)
2117 {
2118         sector_t *bno = data, addr;
2119
2120         if (iomap->type == IOMAP_MAPPED) {
2121                 addr = (pos - iomap->offset + iomap->addr) >> inode->i_blkbits;
2122                 if (addr > INT_MAX)
2123                         WARN(1, "would truncate bmap result\n");
2124                 else
2125                         *bno = addr;
2126         }
2127         return 0;
2128 }
2129
2130 /* legacy ->bmap interface.  0 is the error return (!) */
2131 sector_t
2132 iomap_bmap(struct address_space *mapping, sector_t bno,
2133                 const struct iomap_ops *ops)
2134 {
2135         struct inode *inode = mapping->host;
2136         loff_t pos = bno << inode->i_blkbits;
2137         unsigned blocksize = i_blocksize(inode);
2138
2139         if (filemap_write_and_wait(mapping))
2140                 return 0;
2141
2142         bno = 0;
2143         iomap_apply(inode, pos, blocksize, 0, ops, &bno, iomap_bmap_actor);
2144         return bno;
2145 }
2146 EXPORT_SYMBOL_GPL(iomap_bmap);