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         spin_lock_init(&iop->uptodate_lock);
 120         bitmap_zero(iop->uptodate, PAGE_SIZE / SECTOR_SIZE);
 121
 122         /*
 123          * migrate_page_move_mapping() assumes that pages with private data have
 124          * their count elevated by 1.
 125          */
 126         get_page(page);
 127         set_page_private(page, (unsigned long)iop);
 128         SetPagePrivate(page);
 129         return iop;
 130 }
 131
 132 static void
 133 iomap_page_release(struct page *page)
 134 {
 135         struct iomap_page *iop = to_iomap_page(page);
 136
 137         if (!iop)
 138                 return;
 139         WARN_ON_ONCE(atomic_read(&iop->read_count));
 140         WARN_ON_ONCE(atomic_read(&iop->write_count));
 141         ClearPagePrivate(page);
 142         set_page_private(page, 0);
 143         put_page(page);
 144         kfree(iop);
 145 }
 146
 147 /*
 148  * Calculate the range inside the page that we actually need to read.
 149  */
 150 static void
 151 iomap_adjust_read_range(struct inode *inode, struct iomap_page *iop,
 152                 loff_t *pos, loff_t length, unsigned *offp, unsigned *lenp)
 153 {
 154         loff_t orig_pos = *pos;
 155         loff_t isize = i_size_read(inode);
 156         unsigned block_bits = inode->i_blkbits;
 157         unsigned block_size = (1 << block_bits);
 158         unsigned poff = offset_in_page(*pos);
 159         unsigned plen = min_t(loff_t, PAGE_SIZE - poff, length);
 160         unsigned first = poff >> block_bits;
 161         unsigned last = (poff + plen - 1) >> block_bits;
 162
 163         /*
 164          * If the block size is smaller than the page size we need to check the
 165          * per-block uptodate status and adjust the offset and length if needed
 166          * to avoid reading in already uptodate ranges.
 167          */
 168         if (iop) {
 169                 unsigned int i;
 170
 171                 /* move forward for each leading block marked uptodate */
 172                 for (i = first; i <= last; i++) {
 173                         if (!test_bit(i, iop->uptodate))
 174                                 break;
 175                         *pos += block_size;
 176                         poff += block_size;
 177                         plen -= block_size;
 178                         first++;
 179                 }
 180
 181                 /* truncate len if we find any trailing uptodate block(s) */
 182                 for ( ; i <= last; i++) {
 183                         if (test_bit(i, iop->uptodate)) {
 184                                 plen -= (last - i + 1) * block_size;
 185                                 last = i - 1;
 186                                 break;
 187                         }
 188                 }
 189         }
 190
 191         /*
 192          * If the extent spans the block that contains the i_size we need to
 193          * handle both halves separately so that we properly zero data in the
 194          * page cache for blocks that are entirely outside of i_size.
 195          */
 196         if (orig_pos <= isize && orig_pos + length > isize) {
 197                 unsigned end = offset_in_page(isize - 1) >> block_bits;
 198
 199                 if (first <= end && last > end)
 200                         plen -= (last - end) * block_size;
 201         }
 202
 203         *offp = poff;
 204         *lenp = plen;
 205 }
 206
 207 static void
 208 iomap_iop_set_range_uptodate(struct page *page, unsigned off, unsigned len)
 209 {
 210         struct iomap_page *iop = to_iomap_page(page);
 211         struct inode *inode = page->mapping->host;
 212         unsigned first = off >> inode->i_blkbits;
 213         unsigned last = (off + len - 1) >> inode->i_blkbits;
 214         bool uptodate = true;
 215         unsigned long flags;
 216         unsigned int i;
 217
 218         spin_lock_irqsave(&iop->uptodate_lock, flags);
 219         for (i = 0; i < PAGE_SIZE / i_blocksize(inode); i++) {
 220                 if (i >= first && i <= last)
 221                         set_bit(i, iop->uptodate);
 222                 else if (!test_bit(i, iop->uptodate))
 223                         uptodate = false;
 224         }
 225
 226         if (uptodate)
 227                 SetPageUptodate(page);
 228         spin_unlock_irqrestore(&iop->uptodate_lock, flags);
 229 }
 230
 231 static void
 232 iomap_set_range_uptodate(struct page *page, unsigned off, unsigned len)
 233 {
 234         if (PageError(page))
 235                 return;
 236
 237         if (page_has_private(page))
 238                 iomap_iop_set_range_uptodate(page, off, len);
 239         else
 240                 SetPageUptodate(page);
 241 }
 242
 243 static void
 244 iomap_read_finish(struct iomap_page *iop, struct page *page)
 245 {
 246         if (!iop || atomic_dec_and_test(&iop->read_count))
 247                 unlock_page(page);
 248 }
 249
 250 static void
 251 iomap_read_page_end_io(struct bio_vec *bvec, int error)
 252 {
 253         struct page *page = bvec->bv_page;
 254         struct iomap_page *iop = to_iomap_page(page);
 255
 256         if (unlikely(error)) {
 257                 ClearPageUptodate(page);
 258                 SetPageError(page);
 259         } else {
 260                 iomap_set_range_uptodate(page, bvec->bv_offset, bvec->bv_len);
 261         }
 262
 263         iomap_read_finish(iop, page);
 264 }
 265
 266 static void
 267 iomap_read_inline_data(struct inode *inode, struct page *page,
 268                 struct iomap *iomap)
 269 {
 270         size_t size = i_size_read(inode);
 271         void *addr;
 272
 273         if (PageUptodate(page))
 274                 return;
 275
 276         BUG_ON(page->index);
 277         BUG_ON(size > PAGE_SIZE - offset_in_page(iomap->inline_data));
 278
 279         addr = kmap_atomic(page);
 280         memcpy(addr, iomap->inline_data, size);
 281         memset(addr + size, 0, PAGE_SIZE - size);
 282         kunmap_atomic(addr);
 283         SetPageUptodate(page);
 284 }
 285
 286 static void
 287 iomap_read_end_io(struct bio *bio)
 288 {
 289         int error = blk_status_to_errno(bio->bi_status);
 290         struct bio_vec *bvec;
 291         int i;
 292
 293         bio_for_each_segment_all(bvec, bio, i)
 294                 iomap_read_page_end_io(bvec, error);
 295         bio_put(bio);
 296 }
 297
 298 struct iomap_readpage_ctx {
 299         struct page             *cur_page;
 300         bool                    cur_page_in_bio;
 301         bool                    is_readahead;
 302         struct bio              *bio;
 303         struct list_head        *pages;
 304 };
 305
 306 static loff_t
 307 iomap_readpage_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 308                 struct iomap *iomap)
 309 {
 310         struct iomap_readpage_ctx *ctx = data;
 311         struct page *page = ctx->cur_page;
 312         struct iomap_page *iop = iomap_page_create(inode, page);
 313         bool is_contig = false;
 314         loff_t orig_pos = pos;
 315         unsigned poff, plen;
 316         sector_t sector;
 317
 318         if (iomap->type == IOMAP_INLINE) {
 319                 WARN_ON_ONCE(pos);
 320                 iomap_read_inline_data(inode, page, iomap);
 321                 return PAGE_SIZE;
 322         }
 323
 324         /* zero post-eof blocks as the page may be mapped */
 325         iomap_adjust_read_range(inode, iop, &pos, length, &poff, &plen);
 326         if (plen == 0)
 327                 goto done;
 328
 329         if (iomap->type != IOMAP_MAPPED || pos >= i_size_read(inode)) {
 330                 zero_user(page, poff, plen);
 331                 iomap_set_range_uptodate(page, poff, plen);
 332                 goto done;
 333         }
 334
 335         ctx->cur_page_in_bio = true;
 336
 337         /*
 338          * Try to merge into a previous segment if we can.
 339          */
 340         sector = iomap_sector(iomap, pos);
 341         if (ctx->bio && bio_end_sector(ctx->bio) == sector) {
 342                 if (__bio_try_merge_page(ctx->bio, page, plen, poff))
 343                         goto done;
 344                 is_contig = true;
 345         }
 346
 347         /*
 348          * If we start a new segment we need to increase the read count, and we
 349          * need to do so before submitting any previous full bio to make sure
 350          * that we don't prematurely unlock the page.
 351          */
 352         if (iop)
 353                 atomic_inc(&iop->read_count);
 354
 355         if (!ctx->bio || !is_contig || bio_full(ctx->bio)) {
 356                 gfp_t gfp = mapping_gfp_constraint(page->mapping, GFP_KERNEL);
 357                 int nr_vecs = (length + PAGE_SIZE - 1) >> PAGE_SHIFT;
 358
 359                 if (ctx->bio)
 360                         submit_bio(ctx->bio);
 361
 362                 if (ctx->is_readahead) /* same as readahead_gfp_mask */
 363                         gfp |= __GFP_NORETRY | __GFP_NOWARN;
 364                 ctx->bio = bio_alloc(gfp, min(BIO_MAX_PAGES, nr_vecs));
 365                 ctx->bio->bi_opf = REQ_OP_READ;
 366                 if (ctx->is_readahead)
 367                         ctx->bio->bi_opf |= REQ_RAHEAD;
 368                 ctx->bio->bi_iter.bi_sector = sector;
 369                 bio_set_dev(ctx->bio, iomap->bdev);
 370                 ctx->bio->bi_end_io = iomap_read_end_io;
 371         }
 372
 373         __bio_add_page(ctx->bio, page, plen, poff);
 374 done:
 375         /*
 376          * Move the caller beyond our range so that it keeps making progress.
 377          * For that we have to include any leading non-uptodate ranges, but
 378          * we can skip trailing ones as they will be handled in the next
 379          * iteration.
 380          */
 381         return pos - orig_pos + plen;
 382 }
 383
 384 int
 385 iomap_readpage(struct page *page, const struct iomap_ops *ops)
 386 {
 387         struct iomap_readpage_ctx ctx = { .cur_page = page };
 388         struct inode *inode = page->mapping->host;
 389         unsigned poff;
 390         loff_t ret;
 391
 392         for (poff = 0; poff < PAGE_SIZE; poff += ret) {
 393                 ret = iomap_apply(inode, page_offset(page) + poff,
 394                                 PAGE_SIZE - poff, 0, ops, &ctx,
 395                                 iomap_readpage_actor);
 396                 if (ret <= 0) {
 397                         WARN_ON_ONCE(ret == 0);
 398                         SetPageError(page);
 399                         break;
 400                 }
 401         }
 402
 403         if (ctx.bio) {
 404                 submit_bio(ctx.bio);
 405                 WARN_ON_ONCE(!ctx.cur_page_in_bio);
 406         } else {
 407                 WARN_ON_ONCE(ctx.cur_page_in_bio);
 408                 unlock_page(page);
 409         }
 410
 411         /*
 412          * Just like mpage_readpages and block_read_full_page we always
 413          * return 0 and just mark the page as PageError on errors.  This
 414          * should be cleaned up all through the stack eventually.
 415          */
 416         return 0;
 417 }
 418 EXPORT_SYMBOL_GPL(iomap_readpage);
 419
 420 static struct page *
 421 iomap_next_page(struct inode *inode, struct list_head *pages, loff_t pos,
 422                 loff_t length, loff_t *done)
 423 {
 424         while (!list_empty(pages)) {
 425                 struct page *page = lru_to_page(pages);
 426
 427                 if (page_offset(page) >= (u64)pos + length)
 428                         break;
 429
 430                 list_del(&page->lru);
 431                 if (!add_to_page_cache_lru(page, inode->i_mapping, page->index,
 432                                 GFP_NOFS))
 433                         return page;
 434
 435                 /*
 436                  * If we already have a page in the page cache at index we are
 437                  * done.  Upper layers don't care if it is uptodate after the
 438                  * readpages call itself as every page gets checked again once
 439                  * actually needed.
 440                  */
 441                 *done += PAGE_SIZE;
 442                 put_page(page);
 443         }
 444
 445         return NULL;
 446 }
 447
 448 static loff_t
 449 iomap_readpages_actor(struct inode *inode, loff_t pos, loff_t length,
 450                 void *data, struct iomap *iomap)
 451 {
 452         struct iomap_readpage_ctx *ctx = data;
 453         loff_t done, ret;
 454
 455         for (done = 0; done < length; done += ret) {
 456                 if (ctx->cur_page && offset_in_page(pos + done) == 0) {
 457                         if (!ctx->cur_page_in_bio)
 458                                 unlock_page(ctx->cur_page);
 459                         put_page(ctx->cur_page);
 460                         ctx->cur_page = NULL;
 461                 }
 462                 if (!ctx->cur_page) {
 463                         ctx->cur_page = iomap_next_page(inode, ctx->pages,
 464                                         pos, length, &done);
 465                         if (!ctx->cur_page)
 466                                 break;
 467                         ctx->cur_page_in_bio = false;
 468                 }
 469                 ret = iomap_readpage_actor(inode, pos + done, length - done,
 470                                 ctx, iomap);
 471         }
 472
 473         return done;
 474 }
 475
 476 int
 477 iomap_readpages(struct address_space *mapping, struct list_head *pages,
 478                 unsigned nr_pages, const struct iomap_ops *ops)
 479 {
 480         struct iomap_readpage_ctx ctx = {
 481                 .pages          = pages,
 482                 .is_readahead   = true,
 483         };
 484         loff_t pos = page_offset(list_entry(pages->prev, struct page, lru));
 485         loff_t last = page_offset(list_entry(pages->next, struct page, lru));
 486         loff_t length = last - pos + PAGE_SIZE, ret = 0;
 487
 488         while (length > 0) {
 489                 ret = iomap_apply(mapping->host, pos, length, 0, ops,
 490                                 &ctx, iomap_readpages_actor);
 491                 if (ret <= 0) {
 492                         WARN_ON_ONCE(ret == 0);
 493                         goto done;
 494                 }
 495                 pos += ret;
 496                 length -= ret;
 497         }
 498         ret = 0;
 499 done:
 500         if (ctx.bio)
 501                 submit_bio(ctx.bio);
 502         if (ctx.cur_page) {
 503                 if (!ctx.cur_page_in_bio)
 504                         unlock_page(ctx.cur_page);
 505                 put_page(ctx.cur_page);
 506         }
 507
 508         /*
 509          * Check that we didn't lose a page due to the arcance calling
 510          * conventions..
 511          */
 512         WARN_ON_ONCE(!ret && !list_empty(ctx.pages));
 513         return ret;
 514 }
 515 EXPORT_SYMBOL_GPL(iomap_readpages);
 516
 517 /*
 518  * iomap_is_partially_uptodate checks whether blocks within a page are
 519  * uptodate or not.
 520  *
 521  * Returns true if all blocks which correspond to a file portion
 522  * we want to read within the page are uptodate.
 523  */
 524 int
 525 iomap_is_partially_uptodate(struct page *page, unsigned long from,
 526                 unsigned long count)
 527 {
 528         struct iomap_page *iop = to_iomap_page(page);
 529         struct inode *inode = page->mapping->host;
 530         unsigned len, first, last;
 531         unsigned i;
 532
 533         /* Limit range to one page */
 534         len = min_t(unsigned, PAGE_SIZE - from, count);
 535
 536         /* First and last blocks in range within page */
 537         first = from >> inode->i_blkbits;
 538         last = (from + len - 1) >> inode->i_blkbits;
 539
 540         if (iop) {
 541                 for (i = first; i <= last; i++)
 542                         if (!test_bit(i, iop->uptodate))
 543                                 return 0;
 544                 return 1;
 545         }
 546
 547         return 0;
 548 }
 549 EXPORT_SYMBOL_GPL(iomap_is_partially_uptodate);
 550
 551 int
 552 iomap_releasepage(struct page *page, gfp_t gfp_mask)
 553 {
 554         /*
 555          * mm accommodates an old ext3 case where clean pages might not have had
 556          * the dirty bit cleared. Thus, it can send actual dirty pages to
 557          * ->releasepage() via shrink_active_list(), skip those here.
 558          */
 559         if (PageDirty(page) || PageWriteback(page))
 560                 return 0;
 561         iomap_page_release(page);
 562         return 1;
 563 }
 564 EXPORT_SYMBOL_GPL(iomap_releasepage);
 565
 566 void
 567 iomap_invalidatepage(struct page *page, unsigned int offset, unsigned int len)
 568 {
 569         /*
 570          * If we are invalidating the entire page, clear the dirty state from it
 571          * and release it to avoid unnecessary buildup of the LRU.
 572          */
 573         if (offset == 0 && len == PAGE_SIZE) {
 574                 WARN_ON_ONCE(PageWriteback(page));
 575                 cancel_dirty_page(page);
 576                 iomap_page_release(page);
 577         }
 578 }
 579 EXPORT_SYMBOL_GPL(iomap_invalidatepage);
 580
 581 #ifdef CONFIG_MIGRATION
 582 int
 583 iomap_migrate_page(struct address_space *mapping, struct page *newpage,
 584                 struct page *page, enum migrate_mode mode)
 585 {
 586         int ret;
 587
 588         ret = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
 589         if (ret != MIGRATEPAGE_SUCCESS)
 590                 return ret;
 591
 592         if (page_has_private(page)) {
 593                 ClearPagePrivate(page);
 594                 get_page(newpage);
 595                 set_page_private(newpage, page_private(page));
 596                 set_page_private(page, 0);
 597                 put_page(page);
 598                 SetPagePrivate(newpage);
 599         }
 600
 601         if (mode != MIGRATE_SYNC_NO_COPY)
 602                 migrate_page_copy(newpage, page);
 603         else
 604                 migrate_page_states(newpage, page);
 605         return MIGRATEPAGE_SUCCESS;
 606 }
 607 EXPORT_SYMBOL_GPL(iomap_migrate_page);
 608 #endif /* CONFIG_MIGRATION */
 609
 610 static void
 611 iomap_write_failed(struct inode *inode, loff_t pos, unsigned len)
 612 {
 613         loff_t i_size = i_size_read(inode);
 614
 615         /*
 616          * Only truncate newly allocated pages beyoned EOF, even if the
 617          * write started inside the existing inode size.
 618          */
 619         if (pos + len > i_size)
 620                 truncate_pagecache_range(inode, max(pos, i_size), pos + len);
 621 }
 622
 623 static int
 624 iomap_read_page_sync(struct inode *inode, loff_t block_start, struct page *page,
 625                 unsigned poff, unsigned plen, unsigned from, unsigned to,
 626                 struct iomap *iomap)
 627 {
 628         struct bio_vec bvec;
 629         struct bio bio;
 630
 631         if (iomap->type != IOMAP_MAPPED || block_start >= i_size_read(inode)) {
 632                 zero_user_segments(page, poff, from, to, poff + plen);
 633                 iomap_set_range_uptodate(page, poff, plen);
 634                 return 0;
 635         }
 636
 637         bio_init(&bio, &bvec, 1);
 638         bio.bi_opf = REQ_OP_READ;
 639         bio.bi_iter.bi_sector = iomap_sector(iomap, block_start);
 640         bio_set_dev(&bio, iomap->bdev);
 641         __bio_add_page(&bio, page, plen, poff);
 642         return submit_bio_wait(&bio);
 643 }
 644
 645 static int
 646 __iomap_write_begin(struct inode *inode, loff_t pos, unsigned len,
 647                 struct page *page, struct iomap *iomap)
 648 {
 649         struct iomap_page *iop = iomap_page_create(inode, page);
 650         loff_t block_size = i_blocksize(inode);
 651         loff_t block_start = pos & ~(block_size - 1);
 652         loff_t block_end = (pos + len + block_size - 1) & ~(block_size - 1);
 653         unsigned from = offset_in_page(pos), to = from + len, poff, plen;
 654         int status = 0;
 655
 656         if (PageUptodate(page))
 657                 return 0;
 658
 659         do {
 660                 iomap_adjust_read_range(inode, iop, &block_start,
 661                                 block_end - block_start, &poff, &plen);
 662                 if (plen == 0)
 663                         break;
 664
 665                 if ((from > poff && from < poff + plen) ||
 666                     (to > poff && to < poff + plen)) {
 667                         status = iomap_read_page_sync(inode, block_start, page,
 668                                         poff, plen, from, to, iomap);
 669                         if (status)
 670                                 break;
 671                 }
 672
 673         } while ((block_start += plen) < block_end);
 674
 675         return status;
 676 }
 677
 678 static int
 679 iomap_write_begin(struct inode *inode, loff_t pos, unsigned len, unsigned flags,
 680                 struct page **pagep, struct iomap *iomap)
 681 {
 682         pgoff_t index = pos >> PAGE_SHIFT;
 683         struct page *page;
 684         int status = 0;
 685
 686         BUG_ON(pos + len > iomap->offset + iomap->length);
 687
 688         if (fatal_signal_pending(current))
 689                 return -EINTR;
 690
 691         page = grab_cache_page_write_begin(inode->i_mapping, index, flags);
 692         if (!page)
 693                 return -ENOMEM;
 694
 695         if (iomap->type == IOMAP_INLINE)
 696                 iomap_read_inline_data(inode, page, iomap);
 697         else if (iomap->flags & IOMAP_F_BUFFER_HEAD)
 698                 status = __block_write_begin_int(page, pos, len, NULL, iomap);
 699         else
 700                 status = __iomap_write_begin(inode, pos, len, page, iomap);
 701         if (unlikely(status)) {
 702                 unlock_page(page);
 703                 put_page(page);
 704                 page = NULL;
 705
 706                 iomap_write_failed(inode, pos, len);
 707         }
 708
 709         *pagep = page;
 710         return status;
 711 }
 712
 713 int
 714 iomap_set_page_dirty(struct page *page)
 715 {
 716         struct address_space *mapping = page_mapping(page);
 717         int newly_dirty;
 718
 719         if (unlikely(!mapping))
 720                 return !TestSetPageDirty(page);
 721
 722         /*
 723          * Lock out page->mem_cgroup migration to keep PageDirty
 724          * synchronized with per-memcg dirty page counters.
 725          */
 726         lock_page_memcg(page);
 727         newly_dirty = !TestSetPageDirty(page);
 728         if (newly_dirty)
 729                 __set_page_dirty(page, mapping, 0);
 730         unlock_page_memcg(page);
 731
 732         if (newly_dirty)
 733                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 734         return newly_dirty;
 735 }
 736 EXPORT_SYMBOL_GPL(iomap_set_page_dirty);
 737
 738 static int
 739 __iomap_write_end(struct inode *inode, loff_t pos, unsigned len,
 740                 unsigned copied, struct page *page, struct iomap *iomap)
 741 {
 742         flush_dcache_page(page);
 743
 744         /*
 745          * The blocks that were entirely written will now be uptodate, so we
 746          * don't have to worry about a readpage reading them and overwriting a
 747          * partial write.  However if we have encountered a short write and only
 748          * partially written into a block, it will not be marked uptodate, so a
 749          * readpage might come in and destroy our partial write.
 750          *
 751          * Do the simplest thing, and just treat any short write to a non
 752          * uptodate page as a zero-length write, and force the caller to redo
 753          * the whole thing.
 754          */
 755         if (unlikely(copied < len && !PageUptodate(page))) {
 756                 copied = 0;
 757         } else {
 758                 iomap_set_range_uptodate(page, offset_in_page(pos), len);
 759                 iomap_set_page_dirty(page);
 760         }
 761         return __generic_write_end(inode, pos, copied, page);
 762 }
 763
 764 static int
 765 iomap_write_end_inline(struct inode *inode, struct page *page,
 766                 struct iomap *iomap, loff_t pos, unsigned copied)
 767 {
 768         void *addr;
 769
 770         WARN_ON_ONCE(!PageUptodate(page));
 771         BUG_ON(pos + copied > PAGE_SIZE - offset_in_page(iomap->inline_data));
 772
 773         addr = kmap_atomic(page);
 774         memcpy(iomap->inline_data + pos, addr + pos, copied);
 775         kunmap_atomic(addr);
 776
 777         mark_inode_dirty(inode);
 778         __generic_write_end(inode, pos, copied, page);
 779         return copied;
 780 }
 781
 782 static int
 783 iomap_write_end(struct inode *inode, loff_t pos, unsigned len,
 784                 unsigned copied, struct page *page, struct iomap *iomap)
 785 {
 786         int ret;
 787
 788         if (iomap->type == IOMAP_INLINE) {
 789                 ret = iomap_write_end_inline(inode, page, iomap, pos, copied);
 790         } else if (iomap->flags & IOMAP_F_BUFFER_HEAD) {
 791                 ret = generic_write_end(NULL, inode->i_mapping, pos, len,
 792                                 copied, page, NULL);
 793         } else {
 794                 ret = __iomap_write_end(inode, pos, len, copied, page, iomap);
 795         }
 796
 797         if (iomap->page_done)
 798                 iomap->page_done(inode, pos, copied, page, iomap);
 799
 800         if (ret < len)
 801                 iomap_write_failed(inode, pos, len);
 802         return ret;
 803 }
 804
 805 static loff_t
 806 iomap_write_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 807                 struct iomap *iomap)
 808 {
 809         struct iov_iter *i = data;
 810         long status = 0;
 811         ssize_t written = 0;
 812         unsigned int flags = AOP_FLAG_NOFS;
 813
 814         do {
 815                 struct page *page;
 816                 unsigned long offset;   /* Offset into pagecache page */
 817                 unsigned long bytes;    /* Bytes to write to page */
 818                 size_t copied;          /* Bytes copied from user */
 819
 820                 offset = offset_in_page(pos);
 821                 bytes = min_t(unsigned long, PAGE_SIZE - offset,
 822                                                 iov_iter_count(i));
 823 again:
 824                 if (bytes > length)
 825                         bytes = length;
 826
 827                 /*
 828                  * Bring in the user page that we will copy from _first_.
 829                  * Otherwise there's a nasty deadlock on copying from the
 830                  * same page as we're writing to, without it being marked
 831                  * up-to-date.
 832                  *
 833                  * Not only is this an optimisation, but it is also required
 834                  * to check that the address is actually valid, when atomic
 835                  * usercopies are used, below.
 836                  */
 837                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
 838                         status = -EFAULT;
 839                         break;
 840                 }
 841
 842                 status = iomap_write_begin(inode, pos, bytes, flags, &page,
 843                                 iomap);
 844                 if (unlikely(status))
 845                         break;
 846
 847                 if (mapping_writably_mapped(inode->i_mapping))
 848                         flush_dcache_page(page);
 849
 850                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
 851
 852                 flush_dcache_page(page);
 853
 854                 status = iomap_write_end(inode, pos, bytes, copied, page,
 855                                 iomap);
 856                 if (unlikely(status < 0))
 857                         break;
 858                 copied = status;
 859
 860                 cond_resched();
 861
 862                 iov_iter_advance(i, copied);
 863                 if (unlikely(copied == 0)) {
 864                         /*
 865                          * If we were unable to copy any data at all, we must
 866                          * fall back to a single segment length write.
 867                          *
 868                          * If we didn't fallback here, we could livelock
 869                          * because not all segments in the iov can be copied at
 870                          * once without a pagefault.
 871                          */
 872                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
 873                                                 iov_iter_single_seg_count(i));
 874                         goto again;
 875                 }
 876                 pos += copied;
 877                 written += copied;
 878                 length -= copied;
 879
 880                 balance_dirty_pages_ratelimited(inode->i_mapping);
 881         } while (iov_iter_count(i) && length);
 882
 883         return written ? written : status;
 884 }
 885
 886 ssize_t
 887 iomap_file_buffered_write(struct kiocb *iocb, struct iov_iter *iter,
 888                 const struct iomap_ops *ops)
 889 {
 890         struct inode *inode = iocb->ki_filp->f_mapping->host;
 891         loff_t pos = iocb->ki_pos, ret = 0, written = 0;
 892
 893         while (iov_iter_count(iter)) {
 894                 ret = iomap_apply(inode, pos, iov_iter_count(iter),
 895                                 IOMAP_WRITE, ops, iter, iomap_write_actor);
 896                 if (ret <= 0)
 897                         break;
 898                 pos += ret;
 899                 written += ret;
 900         }
 901
 902         return written ? written : ret;
 903 }
 904 EXPORT_SYMBOL_GPL(iomap_file_buffered_write);
 905
 906 static struct page *
 907 __iomap_read_page(struct inode *inode, loff_t offset)
 908 {
 909         struct address_space *mapping = inode->i_mapping;
 910         struct page *page;
 911
 912         page = read_mapping_page(mapping, offset >> PAGE_SHIFT, NULL);
 913         if (IS_ERR(page))
 914                 return page;
 915         if (!PageUptodate(page)) {
 916                 put_page(page);
 917                 return ERR_PTR(-EIO);
 918         }
 919         return page;
 920 }
 921
 922 static loff_t
 923 iomap_dirty_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 924                 struct iomap *iomap)
 925 {
 926         long status = 0;
 927         ssize_t written = 0;
 928
 929         do {
 930                 struct page *page, *rpage;
 931                 unsigned long offset;   /* Offset into pagecache page */
 932                 unsigned long bytes;    /* Bytes to write to page */
 933
 934                 offset = offset_in_page(pos);
 935                 bytes = min_t(loff_t, PAGE_SIZE - offset, length);
 936
 937                 rpage = __iomap_read_page(inode, pos);
 938                 if (IS_ERR(rpage))
 939                         return PTR_ERR(rpage);
 940
 941                 status = iomap_write_begin(inode, pos, bytes,
 942                                            AOP_FLAG_NOFS, &page, iomap);
 943                 put_page(rpage);
 944                 if (unlikely(status))
 945                         return status;
 946
 947                 WARN_ON_ONCE(!PageUptodate(page));
 948
 949                 status = iomap_write_end(inode, pos, bytes, bytes, page, iomap);
 950                 if (unlikely(status <= 0)) {
 951                         if (WARN_ON_ONCE(status == 0))
 952                                 return -EIO;
 953                         return status;
 954                 }
 955
 956                 cond_resched();
 957
 958                 pos += status;
 959                 written += status;
 960                 length -= status;
 961
 962                 balance_dirty_pages_ratelimited(inode->i_mapping);
 963         } while (length);
 964
 965         return written;
 966 }
 967
 968 int
 969 iomap_file_dirty(struct inode *inode, loff_t pos, loff_t len,
 970                 const struct iomap_ops *ops)
 971 {
 972         loff_t ret;
 973
 974         while (len) {
 975                 ret = iomap_apply(inode, pos, len, IOMAP_WRITE, ops, NULL,
 976                                 iomap_dirty_actor);
 977                 if (ret <= 0)
 978                         return ret;
 979                 pos += ret;
 980                 len -= ret;
 981         }
 982
 983         return 0;
 984 }
 985 EXPORT_SYMBOL_GPL(iomap_file_dirty);
 986
 987 static int iomap_zero(struct inode *inode, loff_t pos, unsigned offset,
 988                 unsigned bytes, struct iomap *iomap)
 989 {
 990         struct page *page;
 991         int status;
 992
 993         status = iomap_write_begin(inode, pos, bytes, AOP_FLAG_NOFS, &page,
 994                                    iomap);
 995         if (status)
 996                 return status;
 997
 998         zero_user(page, offset, bytes);
 999         mark_page_accessed(page);
1000
1001         return iomap_write_end(inode, pos, bytes, bytes, page, iomap);
1002 }
1003
1004 static int iomap_dax_zero(loff_t pos, unsigned offset, unsigned bytes,
1005                 struct iomap *iomap)
1006 {
1007         return __dax_zero_page_range(iomap->bdev, iomap->dax_dev,
1008                         iomap_sector(iomap, pos & PAGE_MASK), offset, bytes);
1009 }
1010
1011 static loff_t
1012 iomap_zero_range_actor(struct inode *inode, loff_t pos, loff_t count,
1013                 void *data, struct iomap *iomap)
1014 {
1015         bool *did_zero = data;
1016         loff_t written = 0;
1017         int status;
1018
1019         /* already zeroed?  we're done. */
1020         if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1021                 return count;
1022
1023         do {
1024                 unsigned offset, bytes;
1025
1026                 offset = offset_in_page(pos);
1027                 bytes = min_t(loff_t, PAGE_SIZE - offset, count);
1028
1029                 if (IS_DAX(inode))
1030                         status = iomap_dax_zero(pos, offset, bytes, iomap);
1031                 else
1032                         status = iomap_zero(inode, pos, offset, bytes, iomap);
1033                 if (status < 0)
1034                         return status;
1035
1036                 pos += bytes;
1037                 count -= bytes;
1038                 written += bytes;
1039                 if (did_zero)
1040                         *did_zero = true;
1041         } while (count > 0);
1042
1043         return written;
1044 }
1045
1046 int
1047 iomap_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
1048                 const struct iomap_ops *ops)
1049 {
1050         loff_t ret;
1051
1052         while (len > 0) {
1053                 ret = iomap_apply(inode, pos, len, IOMAP_ZERO,
1054                                 ops, did_zero, iomap_zero_range_actor);
1055                 if (ret <= 0)
1056                         return ret;
1057
1058                 pos += ret;
1059                 len -= ret;
1060         }
1061
1062         return 0;
1063 }
1064 EXPORT_SYMBOL_GPL(iomap_zero_range);
1065
1066 int
1067 iomap_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
1068                 const struct iomap_ops *ops)
1069 {
1070         unsigned int blocksize = i_blocksize(inode);
1071         unsigned int off = pos & (blocksize - 1);
1072
1073         /* Block boundary? Nothing to do */
1074         if (!off)
1075                 return 0;
1076         return iomap_zero_range(inode, pos, blocksize - off, did_zero, ops);
1077 }
1078 EXPORT_SYMBOL_GPL(iomap_truncate_page);
1079
1080 static loff_t
1081 iomap_page_mkwrite_actor(struct inode *inode, loff_t pos, loff_t length,
1082                 void *data, struct iomap *iomap)
1083 {
1084         struct page *page = data;
1085         int ret;
1086
1087         if (iomap->flags & IOMAP_F_BUFFER_HEAD) {
1088                 ret = __block_write_begin_int(page, pos, length, NULL, iomap);
1089                 if (ret)
1090                         return ret;
1091                 block_commit_write(page, 0, length);
1092         } else {
1093                 WARN_ON_ONCE(!PageUptodate(page));
1094                 iomap_page_create(inode, page);
1095                 set_page_dirty(page);
1096         }
1097
1098         return length;
1099 }
1100
1101 int iomap_page_mkwrite(struct vm_fault *vmf, const struct iomap_ops *ops)
1102 {
1103         struct page *page = vmf->page;
1104         struct inode *inode = file_inode(vmf->vma->vm_file);
1105         unsigned long length;
1106         loff_t offset, size;
1107         ssize_t ret;
1108
1109         lock_page(page);
1110         size = i_size_read(inode);
1111         if ((page->mapping != inode->i_mapping) ||
1112             (page_offset(page) > size)) {
1113                 /* We overload EFAULT to mean page got truncated */
1114                 ret = -EFAULT;
1115                 goto out_unlock;
1116         }
1117
1118         /* page is wholly or partially inside EOF */
1119         if (((page->index + 1) << PAGE_SHIFT) > size)
1120                 length = offset_in_page(size);
1121         else
1122                 length = PAGE_SIZE;
1123
1124         offset = page_offset(page);
1125         while (length > 0) {
1126                 ret = iomap_apply(inode, offset, length,
1127                                 IOMAP_WRITE | IOMAP_FAULT, ops, page,
1128                                 iomap_page_mkwrite_actor);
1129                 if (unlikely(ret <= 0))
1130                         goto out_unlock;
1131                 offset += ret;
1132                 length -= ret;
1133         }
1134
1135         wait_for_stable_page(page);
1136         return VM_FAULT_LOCKED;
1137 out_unlock:
1138         unlock_page(page);
1139         return block_page_mkwrite_return(ret);
1140 }
1141 EXPORT_SYMBOL_GPL(iomap_page_mkwrite);
1142
1143 struct fiemap_ctx {
1144         struct fiemap_extent_info *fi;
1145         struct iomap prev;
1146 };
1147
1148 static int iomap_to_fiemap(struct fiemap_extent_info *fi,
1149                 struct iomap *iomap, u32 flags)
1150 {
1151         switch (iomap->type) {
1152         case IOMAP_HOLE:
1153                 /* skip holes */
1154                 return 0;
1155         case IOMAP_DELALLOC:
1156                 flags |= FIEMAP_EXTENT_DELALLOC | FIEMAP_EXTENT_UNKNOWN;
1157                 break;
1158         case IOMAP_MAPPED:
1159                 break;
1160         case IOMAP_UNWRITTEN:
1161                 flags |= FIEMAP_EXTENT_UNWRITTEN;
1162                 break;
1163         case IOMAP_INLINE:
1164                 flags |= FIEMAP_EXTENT_DATA_INLINE;
1165                 break;
1166         }
1167
1168         if (iomap->flags & IOMAP_F_MERGED)
1169                 flags |= FIEMAP_EXTENT_MERGED;
1170         if (iomap->flags & IOMAP_F_SHARED)
1171                 flags |= FIEMAP_EXTENT_SHARED;
1172
1173         return fiemap_fill_next_extent(fi, iomap->offset,
1174                         iomap->addr != IOMAP_NULL_ADDR ? iomap->addr : 0,
1175                         iomap->length, flags);
1176 }
1177
1178 static loff_t
1179 iomap_fiemap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
1180                 struct iomap *iomap)
1181 {
1182         struct fiemap_ctx *ctx = data;
1183         loff_t ret = length;
1184
1185         if (iomap->type == IOMAP_HOLE)
1186                 return length;
1187
1188         ret = iomap_to_fiemap(ctx->fi, &ctx->prev, 0);
1189         ctx->prev = *iomap;
1190         switch (ret) {
1191         case 0:         /* success */
1192                 return length;
1193         case 1:         /* extent array full */
1194                 return 0;
1195         default:
1196                 return ret;
1197         }
1198 }
1199
1200 int iomap_fiemap(struct inode *inode, struct fiemap_extent_info *fi,
1201                 loff_t start, loff_t len, const struct iomap_ops *ops)
1202 {
1203         struct fiemap_ctx ctx;
1204         loff_t ret;
1205
1206         memset(&ctx, 0, sizeof(ctx));
1207         ctx.fi = fi;
1208         ctx.prev.type = IOMAP_HOLE;
1209
1210         ret = fiemap_check_flags(fi, FIEMAP_FLAG_SYNC);
1211         if (ret)
1212                 return ret;
1213
1214         if (fi->fi_flags & FIEMAP_FLAG_SYNC) {
1215                 ret = filemap_write_and_wait(inode->i_mapping);
1216                 if (ret)
1217                         return ret;
1218         }
1219
1220         while (len > 0) {
1221                 ret = iomap_apply(inode, start, len, IOMAP_REPORT, ops, &ctx,
1222                                 iomap_fiemap_actor);
1223                 /* inode with no (attribute) mapping will give ENOENT */
1224                 if (ret == -ENOENT)
1225                         break;
1226                 if (ret < 0)
1227                         return ret;
1228                 if (ret == 0)
1229                         break;
1230
1231                 start += ret;
1232                 len -= ret;
1233         }
1234
1235         if (ctx.prev.type != IOMAP_HOLE) {
1236                 ret = iomap_to_fiemap(fi, &ctx.prev, FIEMAP_EXTENT_LAST);
1237                 if (ret < 0)
1238                         return ret;
1239         }
1240
1241         return 0;
1242 }
1243 EXPORT_SYMBOL_GPL(iomap_fiemap);
1244
1245 /*
1246  * Seek for SEEK_DATA / SEEK_HOLE within @page, starting at @lastoff.
1247  * Returns true if found and updates @lastoff to the offset in file.
1248  */
1249 static bool
1250 page_seek_hole_data(struct inode *inode, struct page *page, loff_t *lastoff,
1251                 int whence)
1252 {
1253         const struct address_space_operations *ops = inode->i_mapping->a_ops;
1254         unsigned int bsize = i_blocksize(inode), off;
1255         bool seek_data = whence == SEEK_DATA;
1256         loff_t poff = page_offset(page);
1257
1258         if (WARN_ON_ONCE(*lastoff >= poff + PAGE_SIZE))
1259                 return false;
1260
1261         if (*lastoff < poff) {
1262                 /*
1263                  * Last offset smaller than the start of the page means we found
1264                  * a hole:
1265                  */
1266                 if (whence == SEEK_HOLE)
1267                         return true;
1268                 *lastoff = poff;
1269         }
1270
1271         /*
1272          * Just check the page unless we can and should check block ranges:
1273          */
1274         if (bsize == PAGE_SIZE || !ops->is_partially_uptodate)
1275                 return PageUptodate(page) == seek_data;
1276
1277         lock_page(page);
1278         if (unlikely(page->mapping != inode->i_mapping))
1279                 goto out_unlock_not_found;
1280
1281         for (off = 0; off < PAGE_SIZE; off += bsize) {
1282                 if (offset_in_page(*lastoff) >= off + bsize)
1283                         continue;
1284                 if (ops->is_partially_uptodate(page, off, bsize) == seek_data) {
1285                         unlock_page(page);
1286                         return true;
1287                 }
1288                 *lastoff = poff + off + bsize;
1289         }
1290
1291 out_unlock_not_found:
1292         unlock_page(page);
1293         return false;
1294 }
1295
1296 /*
1297  * Seek for SEEK_DATA / SEEK_HOLE in the page cache.
1298  *
1299  * Within unwritten extents, the page cache determines which parts are holes
1300  * and which are data: uptodate buffer heads count as data; everything else
1301  * counts as a hole.
1302  *
1303  * Returns the resulting offset on successs, and -ENOENT otherwise.
1304  */
1305 static loff_t
1306 page_cache_seek_hole_data(struct inode *inode, loff_t offset, loff_t length,
1307                 int whence)
1308 {
1309         pgoff_t index = offset >> PAGE_SHIFT;
1310         pgoff_t end = DIV_ROUND_UP(offset + length, PAGE_SIZE);
1311         loff_t lastoff = offset;
1312         struct pagevec pvec;
1313
1314         if (length <= 0)
1315                 return -ENOENT;
1316
1317         pagevec_init(&pvec);
1318
1319         do {
1320                 unsigned nr_pages, i;
1321
1322                 nr_pages = pagevec_lookup_range(&pvec, inode->i_mapping, &index,
1323                                                 end - 1);
1324                 if (nr_pages == 0)
1325                         break;
1326
1327                 for (i = 0; i < nr_pages; i++) {
1328                         struct page *page = pvec.pages[i];
1329
1330                         if (page_seek_hole_data(inode, page, &lastoff, whence))
1331                                 goto check_range;
1332                         lastoff = page_offset(page) + PAGE_SIZE;
1333                 }
1334                 pagevec_release(&pvec);
1335         } while (index < end);
1336
1337         /* When no page at lastoff and we are not done, we found a hole. */
1338         if (whence != SEEK_HOLE)
1339                 goto not_found;
1340
1341 check_range:
1342         if (lastoff < offset + length)
1343                 goto out;
1344 not_found:
1345         lastoff = -ENOENT;
1346 out:
1347         pagevec_release(&pvec);
1348         return lastoff;
1349 }
1350
1351
1352 static loff_t
1353 iomap_seek_hole_actor(struct inode *inode, loff_t offset, loff_t length,
1354                       void *data, struct iomap *iomap)
1355 {
1356         switch (iomap->type) {
1357         case IOMAP_UNWRITTEN:
1358                 offset = page_cache_seek_hole_data(inode, offset, length,
1359                                                    SEEK_HOLE);
1360                 if (offset < 0)
1361                         return length;
1362                 /* fall through */
1363         case IOMAP_HOLE:
1364                 *(loff_t *)data = offset;
1365                 return 0;
1366         default:
1367                 return length;
1368         }
1369 }
1370
1371 loff_t
1372 iomap_seek_hole(struct inode *inode, loff_t offset, const struct iomap_ops *ops)
1373 {
1374         loff_t size = i_size_read(inode);
1375         loff_t length = size - offset;
1376         loff_t ret;
1377
1378         /* Nothing to be found before or beyond the end of the file. */
1379         if (offset < 0 || offset >= size)
1380                 return -ENXIO;
1381
1382         while (length > 0) {
1383                 ret = iomap_apply(inode, offset, length, IOMAP_REPORT, ops,
1384                                   &offset, iomap_seek_hole_actor);
1385                 if (ret < 0)
1386                         return ret;
1387                 if (ret == 0)
1388                         break;
1389
1390                 offset += ret;
1391                 length -= ret;
1392         }
1393
1394         return offset;
1395 }
1396 EXPORT_SYMBOL_GPL(iomap_seek_hole);
1397
1398 static loff_t
1399 iomap_seek_data_actor(struct inode *inode, loff_t offset, loff_t length,
1400                       void *data, struct iomap *iomap)
1401 {
1402         switch (iomap->type) {
1403         case IOMAP_HOLE:
1404                 return length;
1405         case IOMAP_UNWRITTEN:
1406                 offset = page_cache_seek_hole_data(inode, offset, length,
1407                                                    SEEK_DATA);
1408                 if (offset < 0)
1409                         return length;
1410                 /*FALLTHRU*/
1411         default:
1412                 *(loff_t *)data = offset;
1413                 return 0;
1414         }
1415 }
1416
1417 loff_t
1418 iomap_seek_data(struct inode *inode, loff_t offset, const struct iomap_ops *ops)
1419 {
1420         loff_t size = i_size_read(inode);
1421         loff_t length = size - offset;
1422         loff_t ret;
1423
1424         /* Nothing to be found before or beyond the end of the file. */
1425         if (offset < 0 || offset >= size)
1426                 return -ENXIO;
1427
1428         while (length > 0) {
1429                 ret = iomap_apply(inode, offset, length, IOMAP_REPORT, ops,
1430                                   &offset, iomap_seek_data_actor);
1431                 if (ret < 0)
1432                         return ret;
1433                 if (ret == 0)
1434                         break;
1435
1436                 offset += ret;
1437                 length -= ret;
1438         }
1439
1440         if (length <= 0)
1441                 return -ENXIO;
1442         return offset;
1443 }
1444 EXPORT_SYMBOL_GPL(iomap_seek_data);
1445
1446 /*
1447  * Private flags for iomap_dio, must not overlap with the public ones in
1448  * iomap.h:
1449  */
1450 #define IOMAP_DIO_WRITE_FUA     (1 << 28)
1451 #define IOMAP_DIO_NEED_SYNC     (1 << 29)
1452 #define IOMAP_DIO_WRITE         (1 << 30)
1453 #define IOMAP_DIO_DIRTY         (1 << 31)
1454
1455 struct iomap_dio {
1456         struct kiocb            *iocb;
1457         iomap_dio_end_io_t      *end_io;
1458         loff_t                  i_size;
1459         loff_t                  size;
1460         atomic_t                ref;
1461         unsigned                flags;
1462         int                     error;
1463         bool                    wait_for_completion;
1464
1465         union {
1466                 /* used during submission and for synchronous completion: */
1467                 struct {
1468                         struct iov_iter         *iter;
1469                         struct task_struct      *waiter;
1470                         struct request_queue    *last_queue;
1471                         blk_qc_t                cookie;
1472                 } submit;
1473
1474                 /* used for aio completion: */
1475                 struct {
1476                         struct work_struct      work;
1477                 } aio;
1478         };
1479 };
1480
1481 static ssize_t iomap_dio_complete(struct iomap_dio *dio)
1482 {
1483         struct kiocb *iocb = dio->iocb;
1484         struct inode *inode = file_inode(iocb->ki_filp);
1485         loff_t offset = iocb->ki_pos;
1486         ssize_t ret;
1487
1488         if (dio->end_io) {
1489                 ret = dio->end_io(iocb,
1490                                 dio->error ? dio->error : dio->size,
1491                                 dio->flags);
1492         } else {
1493                 ret = dio->error;
1494         }
1495
1496         if (likely(!ret)) {
1497                 ret = dio->size;
1498                 /* check for short read */
1499                 if (offset + ret > dio->i_size &&
1500                     !(dio->flags & IOMAP_DIO_WRITE))
1501                         ret = dio->i_size - offset;
1502                 iocb->ki_pos += ret;
1503         }
1504
1505         /*
1506          * Try again to invalidate clean pages which might have been cached by
1507          * non-direct readahead, or faulted in by get_user_pages() if the source
1508          * of the write was an mmap'ed region of the file we're writing.  Either
1509          * one is a pretty crazy thing to do, so we don't support it 100%.  If
1510          * this invalidation fails, tough, the write still worked...
1511          *
1512          * And this page cache invalidation has to be after dio->end_io(), as
1513          * some filesystems convert unwritten extents to real allocations in
1514          * end_io() when necessary, otherwise a racing buffer read would cache
1515          * zeros from unwritten extents.
1516          */
1517         if (!dio->error &&
1518             (dio->flags & IOMAP_DIO_WRITE) && inode->i_mapping->nrpages) {
1519                 int err;
1520                 err = invalidate_inode_pages2_range(inode->i_mapping,
1521                                 offset >> PAGE_SHIFT,
1522                                 (offset + dio->size - 1) >> PAGE_SHIFT);
1523                 if (err)
1524                         dio_warn_stale_pagecache(iocb->ki_filp);
1525         }
1526
1527         /*
1528          * If this is a DSYNC write, make sure we push it to stable storage now
1529          * that we've written data.
1530          */
1531         if (ret > 0 && (dio->flags & IOMAP_DIO_NEED_SYNC))
1532                 ret = generic_write_sync(iocb, ret);
1533
1534         inode_dio_end(file_inode(iocb->ki_filp));
1535         kfree(dio);
1536
1537         return ret;
1538 }
1539
1540 static void iomap_dio_complete_work(struct work_struct *work)
1541 {
1542         struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work);
1543         struct kiocb *iocb = dio->iocb;
1544
1545         iocb->ki_complete(iocb, iomap_dio_complete(dio), 0);
1546 }
1547
1548 /*
1549  * Set an error in the dio if none is set yet.  We have to use cmpxchg
1550  * as the submission context and the completion context(s) can race to
1551  * update the error.
1552  */
1553 static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret)
1554 {
1555         cmpxchg(&dio->error, 0, ret);
1556 }
1557
1558 static void iomap_dio_bio_end_io(struct bio *bio)
1559 {
1560         struct iomap_dio *dio = bio->bi_private;
1561         bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
1562
1563         if (bio->bi_status)
1564                 iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
1565
1566         if (atomic_dec_and_test(&dio->ref)) {
1567                 if (dio->wait_for_completion) {
1568                         struct task_struct *waiter = dio->submit.waiter;
1569                         WRITE_ONCE(dio->submit.waiter, NULL);
1570                         wake_up_process(waiter);
1571                 } else if (dio->flags & IOMAP_DIO_WRITE) {
1572                         struct inode *inode = file_inode(dio->iocb->ki_filp);
1573
1574                         INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
1575                         queue_work(inode->i_sb->s_dio_done_wq, &dio->aio.work);
1576                 } else {
1577                         iomap_dio_complete_work(&dio->aio.work);
1578                 }
1579         }
1580
1581         if (should_dirty) {
1582                 bio_check_pages_dirty(bio);
1583         } else {
1584                 struct bio_vec *bvec;
1585                 int i;
1586
1587                 bio_for_each_segment_all(bvec, bio, i)
1588                         put_page(bvec->bv_page);
1589                 bio_put(bio);
1590         }
1591 }
1592
1593 static blk_qc_t
1594 iomap_dio_zero(struct iomap_dio *dio, struct iomap *iomap, loff_t pos,
1595                 unsigned len)
1596 {
1597         struct page *page = ZERO_PAGE(0);
1598         struct bio *bio;
1599
1600         bio = bio_alloc(GFP_KERNEL, 1);
1601         bio_set_dev(bio, iomap->bdev);
1602         bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
1603         bio->bi_private = dio;
1604         bio->bi_end_io = iomap_dio_bio_end_io;
1605
1606         get_page(page);
1607         __bio_add_page(bio, page, len, 0);
1608         bio_set_op_attrs(bio, REQ_OP_WRITE, REQ_SYNC | REQ_IDLE);
1609
1610         atomic_inc(&dio->ref);
1611         return submit_bio(bio);
1612 }
1613
1614 static loff_t
1615 iomap_dio_bio_actor(struct inode *inode, loff_t pos, loff_t length,
1616                 struct iomap_dio *dio, struct iomap *iomap)
1617 {
1618         unsigned int blkbits = blksize_bits(bdev_logical_block_size(iomap->bdev));
1619         unsigned int fs_block_size = i_blocksize(inode), pad;
1620         unsigned int align = iov_iter_alignment(dio->submit.iter);
1621         struct iov_iter iter;
1622         struct bio *bio;
1623         bool need_zeroout = false;
1624         bool use_fua = false;
1625         int nr_pages, ret = 0;
1626         size_t copied = 0;
1627
1628         if ((pos | length | align) & ((1 << blkbits) - 1))
1629                 return -EINVAL;
1630
1631         if (iomap->type == IOMAP_UNWRITTEN) {
1632                 dio->flags |= IOMAP_DIO_UNWRITTEN;
1633                 need_zeroout = true;
1634         }
1635
1636         if (iomap->flags & IOMAP_F_SHARED)
1637                 dio->flags |= IOMAP_DIO_COW;
1638
1639         if (iomap->flags & IOMAP_F_NEW) {
1640                 need_zeroout = true;
1641         } else if (iomap->type == IOMAP_MAPPED) {
1642                 /*
1643                  * Use a FUA write if we need datasync semantics, this is a pure
1644                  * data IO that doesn't require any metadata updates (including
1645                  * after IO completion such as unwritten extent conversion) and
1646                  * the underlying device supports FUA. This allows us to avoid
1647                  * cache flushes on IO completion.
1648                  */
1649                 if (!(iomap->flags & (IOMAP_F_SHARED|IOMAP_F_DIRTY)) &&
1650                     (dio->flags & IOMAP_DIO_WRITE_FUA) &&
1651                     blk_queue_fua(bdev_get_queue(iomap->bdev)))
1652                         use_fua = true;
1653         }
1654
1655         /*
1656          * Operate on a partial iter trimmed to the extent we were called for.
1657          * We'll update the iter in the dio once we're done with this extent.
1658          */
1659         iter = *dio->submit.iter;
1660         iov_iter_truncate(&iter, length);
1661
1662         nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
1663         if (nr_pages <= 0)
1664                 return nr_pages;
1665
1666         if (need_zeroout) {
1667                 /* zero out from the start of the block to the write offset */
1668                 pad = pos & (fs_block_size - 1);
1669                 if (pad)
1670                         iomap_dio_zero(dio, iomap, pos - pad, pad);
1671         }
1672
1673         do {
1674                 size_t n;
1675                 if (dio->error) {
1676                         iov_iter_revert(dio->submit.iter, copied);
1677                         return 0;
1678                 }
1679
1680                 bio = bio_alloc(GFP_KERNEL, nr_pages);
1681                 bio_set_dev(bio, iomap->bdev);
1682                 bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
1683                 bio->bi_write_hint = dio->iocb->ki_hint;
1684                 bio->bi_ioprio = dio->iocb->ki_ioprio;
1685                 bio->bi_private = dio;
1686                 bio->bi_end_io = iomap_dio_bio_end_io;
1687
1688                 ret = bio_iov_iter_get_pages(bio, &iter);
1689                 if (unlikely(ret)) {
1690                         /*
1691                          * We have to stop part way through an IO. We must fall
1692                          * through to the sub-block tail zeroing here, otherwise
1693                          * this short IO may expose stale data in the tail of
1694                          * the block we haven't written data to.
1695                          */
1696                         bio_put(bio);
1697                         goto zero_tail;
1698                 }
1699
1700                 n = bio->bi_iter.bi_size;
1701                 if (dio->flags & IOMAP_DIO_WRITE) {
1702                         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
1703                         if (use_fua)
1704                                 bio->bi_opf |= REQ_FUA;
1705                         else
1706                                 dio->flags &= ~IOMAP_DIO_WRITE_FUA;
1707                         task_io_account_write(n);
1708                 } else {
1709                         bio->bi_opf = REQ_OP_READ;
1710                         if (dio->flags & IOMAP_DIO_DIRTY)
1711                                 bio_set_pages_dirty(bio);
1712                 }
1713
1714                 iov_iter_advance(dio->submit.iter, n);
1715
1716                 dio->size += n;
1717                 pos += n;
1718                 copied += n;
1719
1720                 nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
1721
1722                 atomic_inc(&dio->ref);
1723
1724                 dio->submit.last_queue = bdev_get_queue(iomap->bdev);
1725                 dio->submit.cookie = submit_bio(bio);
1726         } while (nr_pages);
1727
1728         /*
1729          * We need to zeroout the tail of a sub-block write if the extent type
1730          * requires zeroing or the write extends beyond EOF. If we don't zero
1731          * the block tail in the latter case, we can expose stale data via mmap
1732          * reads of the EOF block.
1733          */
1734 zero_tail:
1735         if (need_zeroout ||
1736             ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode))) {
1737                 /* zero out from the end of the write to the end of the block */
1738                 pad = pos & (fs_block_size - 1);
1739                 if (pad)
1740                         iomap_dio_zero(dio, iomap, pos, fs_block_size - pad);
1741         }
1742         return copied ? copied : ret;
1743 }
1744
1745 static loff_t
1746 iomap_dio_hole_actor(loff_t length, struct iomap_dio *dio)
1747 {
1748         length = iov_iter_zero(length, dio->submit.iter);
1749         dio->size += length;
1750         return length;
1751 }
1752
1753 static loff_t
1754 iomap_dio_inline_actor(struct inode *inode, loff_t pos, loff_t length,
1755                 struct iomap_dio *dio, struct iomap *iomap)
1756 {
1757         struct iov_iter *iter = dio->submit.iter;
1758         size_t copied;
1759
1760         BUG_ON(pos + length > PAGE_SIZE - offset_in_page(iomap->inline_data));
1761
1762         if (dio->flags & IOMAP_DIO_WRITE) {
1763                 loff_t size = inode->i_size;
1764
1765                 if (pos > size)
1766                         memset(iomap->inline_data + size, 0, pos - size);
1767                 copied = copy_from_iter(iomap->inline_data + pos, length, iter);
1768                 if (copied) {
1769                         if (pos + copied > size)
1770                                 i_size_write(inode, pos + copied);
1771                         mark_inode_dirty(inode);
1772                 }
1773         } else {
1774                 copied = copy_to_iter(iomap->inline_data + pos, length, iter);
1775         }
1776         dio->size += copied;
1777         return copied;
1778 }
1779
1780 static loff_t
1781 iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
1782                 void *data, struct iomap *iomap)
1783 {
1784         struct iomap_dio *dio = data;
1785
1786         switch (iomap->type) {
1787         case IOMAP_HOLE:
1788                 if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
1789                         return -EIO;
1790                 return iomap_dio_hole_actor(length, dio);
1791         case IOMAP_UNWRITTEN:
1792                 if (!(dio->flags & IOMAP_DIO_WRITE))
1793                         return iomap_dio_hole_actor(length, dio);
1794                 return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
1795         case IOMAP_MAPPED:
1796                 return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
1797         case IOMAP_INLINE:
1798                 return iomap_dio_inline_actor(inode, pos, length, dio, iomap);
1799         default:
1800                 WARN_ON_ONCE(1);
1801                 return -EIO;
1802         }
1803 }
1804
1805 /*
1806  * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
1807  * is being issued as AIO or not.  This allows us to optimise pure data writes
1808  * to use REQ_FUA rather than requiring generic_write_sync() to issue a
1809  * REQ_FLUSH post write. This is slightly tricky because a single request here
1810  * can be mapped into multiple disjoint IOs and only a subset of the IOs issued
1811  * may be pure data writes. In that case, we still need to do a full data sync
1812  * completion.
1813  */
1814 ssize_t
1815 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
1816                 const struct iomap_ops *ops, iomap_dio_end_io_t end_io)
1817 {
1818         struct address_space *mapping = iocb->ki_filp->f_mapping;
1819         struct inode *inode = file_inode(iocb->ki_filp);
1820         size_t count = iov_iter_count(iter);
1821         loff_t pos = iocb->ki_pos, start = pos;
1822         loff_t end = iocb->ki_pos + count - 1, ret = 0;
1823         unsigned int flags = IOMAP_DIRECT;
1824         bool wait_for_completion = is_sync_kiocb(iocb);
1825         struct blk_plug plug;
1826         struct iomap_dio *dio;
1827
1828         lockdep_assert_held(&inode->i_rwsem);
1829
1830         if (!count)
1831                 return 0;
1832
1833         dio = kmalloc(sizeof(*dio), GFP_KERNEL);
1834         if (!dio)
1835                 return -ENOMEM;
1836
1837         dio->iocb = iocb;
1838         atomic_set(&dio->ref, 1);
1839         dio->size = 0;
1840         dio->i_size = i_size_read(inode);
1841         dio->end_io = end_io;
1842         dio->error = 0;
1843         dio->flags = 0;
1844
1845         dio->submit.iter = iter;
1846         dio->submit.waiter = current;
1847         dio->submit.cookie = BLK_QC_T_NONE;
1848         dio->submit.last_queue = NULL;
1849
1850         if (iov_iter_rw(iter) == READ) {
1851                 if (pos >= dio->i_size)
1852                         goto out_free_dio;
1853
1854                 if (iter->type == ITER_IOVEC)
1855                         dio->flags |= IOMAP_DIO_DIRTY;
1856         } else {
1857                 flags |= IOMAP_WRITE;
1858                 dio->flags |= IOMAP_DIO_WRITE;
1859
1860                 /* for data sync or sync, we need sync completion processing */
1861                 if (iocb->ki_flags & IOCB_DSYNC)
1862                         dio->flags |= IOMAP_DIO_NEED_SYNC;
1863
1864                 /*
1865                  * For datasync only writes, we optimistically try using FUA for
1866                  * this IO.  Any non-FUA write that occurs will clear this flag,
1867                  * hence we know before completion whether a cache flush is
1868                  * necessary.
1869                  */
1870                 if ((iocb->ki_flags & (IOCB_DSYNC | IOCB_SYNC)) == IOCB_DSYNC)
1871                         dio->flags |= IOMAP_DIO_WRITE_FUA;
1872         }
1873
1874         if (iocb->ki_flags & IOCB_NOWAIT) {
1875                 if (filemap_range_has_page(mapping, start, end)) {
1876                         ret = -EAGAIN;
1877                         goto out_free_dio;
1878                 }
1879                 flags |= IOMAP_NOWAIT;
1880         }
1881
1882         ret = filemap_write_and_wait_range(mapping, start, end);
1883         if (ret)
1884                 goto out_free_dio;
1885
1886         /*
1887          * Try to invalidate cache pages for the range we're direct
1888          * writing.  If this invalidation fails, tough, the write will
1889          * still work, but racing two incompatible write paths is a
1890          * pretty crazy thing to do, so we don't support it 100%.
1891          */
1892         ret = invalidate_inode_pages2_range(mapping,
1893                         start >> PAGE_SHIFT, end >> PAGE_SHIFT);
1894         if (ret)
1895                 dio_warn_stale_pagecache(iocb->ki_filp);
1896         ret = 0;
1897
1898         if (iov_iter_rw(iter) == WRITE && !wait_for_completion &&
1899             !inode->i_sb->s_dio_done_wq) {
1900                 ret = sb_init_dio_done_wq(inode->i_sb);
1901                 if (ret < 0)
1902                         goto out_free_dio;
1903         }
1904
1905         inode_dio_begin(inode);
1906
1907         blk_start_plug(&plug);
1908         do {
1909                 ret = iomap_apply(inode, pos, count, flags, ops, dio,
1910                                 iomap_dio_actor);
1911                 if (ret <= 0) {
1912                         /* magic error code to fall back to buffered I/O */
1913                         if (ret == -ENOTBLK) {
1914                                 wait_for_completion = true;
1915                                 ret = 0;
1916                         }
1917                         break;
1918                 }
1919                 pos += ret;
1920
1921                 if (iov_iter_rw(iter) == READ && pos >= dio->i_size) {
1922                         /*
1923                          * We only report that we've read data up to i_size.
1924                          * Revert iter to a state corresponding to that as
1925                          * some callers (such as splice code) rely on it.
1926                          */
1927                         iov_iter_revert(iter, pos - dio->i_size);
1928                         break;
1929                 }
1930         } while ((count = iov_iter_count(iter)) > 0);
1931         blk_finish_plug(&plug);
1932
1933         if (ret < 0)
1934                 iomap_dio_set_error(dio, ret);
1935
1936         /*
1937          * If all the writes we issued were FUA, we don't need to flush the
1938          * cache on IO completion. Clear the sync flag for this case.
1939          */
1940         if (dio->flags & IOMAP_DIO_WRITE_FUA)
1941                 dio->flags &= ~IOMAP_DIO_NEED_SYNC;
1942
1943         /*
1944          * We are about to drop our additional submission reference, which
1945          * might be the last reference to the dio.  There are three three
1946          * different ways we can progress here:
1947          *
1948          *  (a) If this is the last reference we will always complete and free
1949          *      the dio ourselves.
1950          *  (b) If this is not the last reference, and we serve an asynchronous
1951          *      iocb, we must never touch the dio after the decrement, the
1952          *      I/O completion handler will complete and free it.
1953          *  (c) If this is not the last reference, but we serve a synchronous
1954          *      iocb, the I/O completion handler will wake us up on the drop
1955          *      of the final reference, and we will complete and free it here
1956          *      after we got woken by the I/O completion handler.
1957          */
1958         dio->wait_for_completion = wait_for_completion;
1959         if (!atomic_dec_and_test(&dio->ref)) {
1960                 if (!wait_for_completion)
1961                         return -EIOCBQUEUED;
1962
1963                 for (;;) {
1964                         set_current_state(TASK_UNINTERRUPTIBLE);
1965                         if (!READ_ONCE(dio->submit.waiter))
1966                                 break;
1967
1968                         if (!(iocb->ki_flags & IOCB_HIPRI) ||
1969                             !dio->submit.last_queue ||
1970                             !blk_poll(dio->submit.last_queue,
1971                                          dio->submit.cookie))
1972                                 io_schedule();
1973                 }
1974                 __set_current_state(TASK_RUNNING);
1975         }
1976
1977         return iomap_dio_complete(dio);
1978
1979 out_free_dio:
1980         kfree(dio);
1981         return ret;
1982 }
1983 EXPORT_SYMBOL_GPL(iomap_dio_rw);
1984
1985 /* Swapfile activation */
1986
1987 #ifdef CONFIG_SWAP
1988 struct iomap_swapfile_info {
1989         struct iomap iomap;             /* accumulated iomap */
1990         struct swap_info_struct *sis;
1991         uint64_t lowest_ppage;          /* lowest physical addr seen (pages) */
1992         uint64_t highest_ppage;         /* highest physical addr seen (pages) */
1993         unsigned long nr_pages;         /* number of pages collected */
1994         int nr_extents;                 /* extent count */
1995 };
1996
1997 /*
1998  * Collect physical extents for this swap file.  Physical extents reported to
1999  * the swap code must be trimmed to align to a page boundary.  The logical
2000  * offset within the file is irrelevant since the swapfile code maps logical
2001  * page numbers of the swap device to the physical page-aligned extents.
2002  */
2003 static int iomap_swapfile_add_extent(struct iomap_swapfile_info *isi)
2004 {
2005         struct iomap *iomap = &isi->iomap;
2006         unsigned long nr_pages;
2007         uint64_t first_ppage;
2008         uint64_t first_ppage_reported;
2009         uint64_t next_ppage;
2010         int error;
2011
2012         /*
2013          * Round the start up and the end down so that the physical
2014          * extent aligns to a page boundary.
2015          */
2016         first_ppage = ALIGN(iomap->addr, PAGE_SIZE) >> PAGE_SHIFT;
2017         next_ppage = ALIGN_DOWN(iomap->addr + iomap->length, PAGE_SIZE) >>
2018                         PAGE_SHIFT;
2019
2020         /* Skip too-short physical extents. */
2021         if (first_ppage >= next_ppage)
2022                 return 0;
2023         nr_pages = next_ppage - first_ppage;
2024
2025         /*
2026          * Calculate how much swap space we're adding; the first page contains
2027          * the swap header and doesn't count.  The mm still wants that first
2028          * page fed to add_swap_extent, however.
2029          */
2030         first_ppage_reported = first_ppage;
2031         if (iomap->offset == 0)
2032                 first_ppage_reported++;
2033         if (isi->lowest_ppage > first_ppage_reported)
2034                 isi->lowest_ppage = first_ppage_reported;
2035         if (isi->highest_ppage < (next_ppage - 1))
2036                 isi->highest_ppage = next_ppage - 1;
2037
2038         /* Add extent, set up for the next call. */
2039         error = add_swap_extent(isi->sis, isi->nr_pages, nr_pages, first_ppage);
2040         if (error < 0)
2041                 return error;
2042         isi->nr_extents += error;
2043         isi->nr_pages += nr_pages;
2044         return 0;
2045 }
2046
2047 /*
2048  * Accumulate iomaps for this swap file.  We have to accumulate iomaps because
2049  * swap only cares about contiguous page-aligned physical extents and makes no
2050  * distinction between written and unwritten extents.
2051  */
2052 static loff_t iomap_swapfile_activate_actor(struct inode *inode, loff_t pos,
2053                 loff_t count, void *data, struct iomap *iomap)
2054 {
2055         struct iomap_swapfile_info *isi = data;
2056         int error;
2057
2058         switch (iomap->type) {
2059         case IOMAP_MAPPED:
2060         case IOMAP_UNWRITTEN:
2061                 /* Only real or unwritten extents. */
2062                 break;
2063         case IOMAP_INLINE:
2064                 /* No inline data. */
2065                 pr_err("swapon: file is inline\n");
2066                 return -EINVAL;
2067         default:
2068                 pr_err("swapon: file has unallocated extents\n");
2069                 return -EINVAL;
2070         }
2071
2072         /* No uncommitted metadata or shared blocks. */
2073         if (iomap->flags & IOMAP_F_DIRTY) {
2074                 pr_err("swapon: file is not committed\n");
2075                 return -EINVAL;
2076         }
2077         if (iomap->flags & IOMAP_F_SHARED) {
2078                 pr_err("swapon: file has shared extents\n");
2079                 return -EINVAL;
2080         }
2081
2082         /* Only one bdev per swap file. */
2083         if (iomap->bdev != isi->sis->bdev) {
2084                 pr_err("swapon: file is on multiple devices\n");
2085                 return -EINVAL;
2086         }
2087
2088         if (isi->iomap.length == 0) {
2089                 /* No accumulated extent, so just store it. */
2090                 memcpy(&isi->iomap, iomap, sizeof(isi->iomap));
2091         } else if (isi->iomap.addr + isi->iomap.length == iomap->addr) {
2092                 /* Append this to the accumulated extent. */
2093                 isi->iomap.length += iomap->length;
2094         } else {
2095                 /* Otherwise, add the retained iomap and store this one. */
2096                 error = iomap_swapfile_add_extent(isi);
2097                 if (error)
2098                         return error;
2099                 memcpy(&isi->iomap, iomap, sizeof(isi->iomap));
2100         }
2101         return count;
2102 }
2103
2104 /*
2105  * Iterate a swap file's iomaps to construct physical extents that can be
2106  * passed to the swapfile subsystem.
2107  */
2108 int iomap_swapfile_activate(struct swap_info_struct *sis,
2109                 struct file *swap_file, sector_t *pagespan,
2110                 const struct iomap_ops *ops)
2111 {
2112         struct iomap_swapfile_info isi = {
2113                 .sis = sis,
2114                 .lowest_ppage = (sector_t)-1ULL,
2115         };
2116         struct address_space *mapping = swap_file->f_mapping;
2117         struct inode *inode = mapping->host;
2118         loff_t pos = 0;
2119         loff_t len = ALIGN_DOWN(i_size_read(inode), PAGE_SIZE);
2120         loff_t ret;
2121
2122         /*
2123          * Persist all file mapping metadata so that we won't have any
2124          * IOMAP_F_DIRTY iomaps.
2125          */
2126         ret = vfs_fsync(swap_file, 1);
2127         if (ret)
2128                 return ret;
2129
2130         while (len > 0) {
2131                 ret = iomap_apply(inode, pos, len, IOMAP_REPORT,
2132                                 ops, &isi, iomap_swapfile_activate_actor);
2133                 if (ret <= 0)
2134                         return ret;
2135
2136                 pos += ret;
2137                 len -= ret;
2138         }
2139
2140         if (isi.iomap.length) {
2141                 ret = iomap_swapfile_add_extent(&isi);
2142                 if (ret)
2143                         return ret;
2144         }
2145
2146         *pagespan = 1 + isi.highest_ppage - isi.lowest_ppage;
2147         sis->max = isi.nr_pages;
2148         sis->pages = isi.nr_pages - 1;
2149         sis->highest_bit = isi.nr_pages - 1;
2150         return isi.nr_extents;
2151 }
2152 EXPORT_SYMBOL_GPL(iomap_swapfile_activate);
2153 #endif /* CONFIG_SWAP */
2154
2155 static loff_t
2156 iomap_bmap_actor(struct inode *inode, loff_t pos, loff_t length,
2157                 void *data, struct iomap *iomap)
2158 {
2159         sector_t *bno = data, addr;
2160
2161         if (iomap->type == IOMAP_MAPPED) {
2162                 addr = (pos - iomap->offset + iomap->addr) >> inode->i_blkbits;
2163                 if (addr > INT_MAX)
2164                         WARN(1, "would truncate bmap result\n");
2165                 else
2166                         *bno = addr;
2167         }
2168         return 0;
2169 }
2170
2171 /* legacy ->bmap interface.  0 is the error return (!) */
2172 sector_t
2173 iomap_bmap(struct address_space *mapping, sector_t bno,
2174                 const struct iomap_ops *ops)
2175 {
2176         struct inode *inode = mapping->host;
2177         loff_t pos = bno << inode->i_blkbits;
2178         unsigned blocksize = i_blocksize(inode);
2179
2180         if (filemap_write_and_wait(mapping))
2181                 return 0;
2182
2183         bno = 0;
2184         iomap_apply(inode, pos, blocksize, 0, ops, &bno, iomap_bmap_actor);
2185         return bno;
2186 }
2187 EXPORT_SYMBOL_GPL(iomap_bmap);