fs/iomap.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2010 Red Hat, Inc.
   4  * Copyright (c) 2016-2018 Christoph Hellwig.
   5  */
   6 #include <linux/module.h>
   7 #include <linux/compiler.h>
   8 #include <linux/fs.h>
   9 #include <linux/iomap.h>
  10 #include <linux/uaccess.h>
  11 #include <linux/gfp.h>
  12 #include <linux/migrate.h>
  13 #include <linux/mm.h>
  14 #include <linux/mm_inline.h>
  15 #include <linux/swap.h>
  16 #include <linux/pagemap.h>
  17 #include <linux/pagevec.h>
  18 #include <linux/file.h>
  19 #include <linux/uio.h>
  20 #include <linux/backing-dev.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/task_io_accounting_ops.h>
  23 #include <linux/dax.h>
  24 #include <linux/sched/signal.h>
  25
  26 #include "internal.h"
  27
  28 /*
  29  * Execute a iomap write on a segment of the mapping that spans a
  30  * contiguous range of pages that have identical block mapping state.
  31  *
  32  * This avoids the need to map pages individually, do individual allocations
  33  * for each page and most importantly avoid the need for filesystem specific
  34  * locking per page. Instead, all the operations are amortised over the entire
  35  * range of pages. It is assumed that the filesystems will lock whatever
  36  * resources they require in the iomap_begin call, and release them in the
  37  * iomap_end call.
  38  */
  39 loff_t
  40 iomap_apply(struct inode *inode, loff_t pos, loff_t length, unsigned flags,
  41                 const struct iomap_ops *ops, void *data, iomap_actor_t actor)
  42 {
  43         struct iomap iomap = { 0 };
  44         loff_t written = 0, ret;
  45
  46         /*
  47          * Need to map a range from start position for length bytes. This can
  48          * span multiple pages - it is only guaranteed to return a range of a
  49          * single type of pages (e.g. all into a hole, all mapped or all
  50          * unwritten). Failure at this point has nothing to undo.
  51          *
  52          * If allocation is required for this range, reserve the space now so
  53          * that the allocation is guaranteed to succeed later on. Once we copy
  54          * the data into the page cache pages, then we cannot fail otherwise we
  55          * expose transient stale data. If the reserve fails, we can safely
  56          * back out at this point as there is nothing to undo.
  57          */
  58         ret = ops->iomap_begin(inode, pos, length, flags, &iomap);
  59         if (ret)
  60                 return ret;
  61         if (WARN_ON(iomap.offset > pos))
  62                 return -EIO;
  63         if (WARN_ON(iomap.length == 0))
  64                 return -EIO;
  65
  66         /*
  67          * Cut down the length to the one actually provided by the filesystem,
  68          * as it might not be able to give us the whole size that we requested.
  69          */
  70         if (iomap.offset + iomap.length < pos + length)
  71                 length = iomap.offset + iomap.length - pos;
  72
  73         /*
  74          * Now that we have guaranteed that the space allocation will succeed.
  75          * we can do the copy-in page by page without having to worry about
  76          * failures exposing transient data.
  77          */
  78         written = actor(inode, pos, length, data, &iomap);
  79
  80         /*
  81          * Now the data has been copied, commit the range we've copied.  This
  82          * should not fail unless the filesystem has had a fatal error.
  83          */
  84         if (ops->iomap_end) {
  85                 ret = ops->iomap_end(inode, pos, length,
  86                                      written > 0 ? written : 0,
  87                                      flags, &iomap);
  88         }
  89
  90         return written ? written : ret;
  91 }
  92
  93 static sector_t
  94 iomap_sector(struct iomap *iomap, loff_t pos)
  95 {
  96         return (iomap->addr + pos - iomap->offset) >> SECTOR_SHIFT;
  97 }
  98
  99 static struct iomap_page *
 100 iomap_page_create(struct inode *inode, struct page *page)
 101 {
 102         struct iomap_page *iop = to_iomap_page(page);
 103
 104         if (iop || i_blocksize(inode) == PAGE_SIZE)
 105                 return iop;
 106
 107         iop = kmalloc(sizeof(*iop), GFP_NOFS | __GFP_NOFAIL);
 108         atomic_set(&iop->read_count, 0);
 109         atomic_set(&iop->write_count, 0);
 110         bitmap_zero(iop->uptodate, PAGE_SIZE / SECTOR_SIZE);
 111
 112         /*
 113          * migrate_page_move_mapping() assumes that pages with private data have
 114          * their count elevated by 1.
 115          */
 116         get_page(page);
 117         set_page_private(page, (unsigned long)iop);
 118         SetPagePrivate(page);
 119         return iop;
 120 }
 121
 122 static void
 123 iomap_page_release(struct page *page)
 124 {
 125         struct iomap_page *iop = to_iomap_page(page);
 126
 127         if (!iop)
 128                 return;
 129         WARN_ON_ONCE(atomic_read(&iop->read_count));
 130         WARN_ON_ONCE(atomic_read(&iop->write_count));
 131         ClearPagePrivate(page);
 132         set_page_private(page, 0);
 133         put_page(page);
 134         kfree(iop);
 135 }
 136
 137 /*
 138  * Calculate the range inside the page that we actually need to read.
 139  */
 140 static void
 141 iomap_adjust_read_range(struct inode *inode, struct iomap_page *iop,
 142                 loff_t *pos, loff_t length, unsigned *offp, unsigned *lenp)
 143 {
 144         loff_t orig_pos = *pos;
 145         loff_t isize = i_size_read(inode);
 146         unsigned block_bits = inode->i_blkbits;
 147         unsigned block_size = (1 << block_bits);
 148         unsigned poff = offset_in_page(*pos);
 149         unsigned plen = min_t(loff_t, PAGE_SIZE - poff, length);
 150         unsigned first = poff >> block_bits;
 151         unsigned last = (poff + plen - 1) >> block_bits;
 152
 153         /*
 154          * If the block size is smaller than the page size we need to check the
 155          * per-block uptodate status and adjust the offset and length if needed
 156          * to avoid reading in already uptodate ranges.
 157          */
 158         if (iop) {
 159                 unsigned int i;
 160
 161                 /* move forward for each leading block marked uptodate */
 162                 for (i = first; i <= last; i++) {
 163                         if (!test_bit(i, iop->uptodate))
 164                                 break;
 165                         *pos += block_size;
 166                         poff += block_size;
 167                         plen -= block_size;
 168                         first++;
 169                 }
 170
 171                 /* truncate len if we find any trailing uptodate block(s) */
 172                 for ( ; i <= last; i++) {
 173                         if (test_bit(i, iop->uptodate)) {
 174                                 plen -= (last - i + 1) * block_size;
 175                                 last = i - 1;
 176                                 break;
 177                         }
 178                 }
 179         }
 180
 181         /*
 182          * If the extent spans the block that contains the i_size we need to
 183          * handle both halves separately so that we properly zero data in the
 184          * page cache for blocks that are entirely outside of i_size.
 185          */
 186         if (orig_pos <= isize && orig_pos + length > isize) {
 187                 unsigned end = offset_in_page(isize - 1) >> block_bits;
 188
 189                 if (first <= end && last > end)
 190                         plen -= (last - end) * block_size;
 191         }
 192
 193         *offp = poff;
 194         *lenp = plen;
 195 }
 196
 197 static void
 198 iomap_set_range_uptodate(struct page *page, unsigned off, unsigned len)
 199 {
 200         struct iomap_page *iop = to_iomap_page(page);
 201         struct inode *inode = page->mapping->host;
 202         unsigned first = off >> inode->i_blkbits;
 203         unsigned last = (off + len - 1) >> inode->i_blkbits;
 204         unsigned int i;
 205         bool uptodate = true;
 206
 207         if (iop) {
 208                 for (i = 0; i < PAGE_SIZE / i_blocksize(inode); i++) {
 209                         if (i >= first && i <= last)
 210                                 set_bit(i, iop->uptodate);
 211                         else if (!test_bit(i, iop->uptodate))
 212                                 uptodate = false;
 213                 }
 214         }
 215
 216         if (uptodate && !PageError(page))
 217                 SetPageUptodate(page);
 218 }
 219
 220 static void
 221 iomap_read_finish(struct iomap_page *iop, struct page *page)
 222 {
 223         if (!iop || atomic_dec_and_test(&iop->read_count))
 224                 unlock_page(page);
 225 }
 226
 227 static void
 228 iomap_read_page_end_io(struct bio_vec *bvec, int error)
 229 {
 230         struct page *page = bvec->bv_page;
 231         struct iomap_page *iop = to_iomap_page(page);
 232
 233         if (unlikely(error)) {
 234                 ClearPageUptodate(page);
 235                 SetPageError(page);
 236         } else {
 237                 iomap_set_range_uptodate(page, bvec->bv_offset, bvec->bv_len);
 238         }
 239
 240         iomap_read_finish(iop, page);
 241 }
 242
 243 static void
 244 iomap_read_end_io(struct bio *bio)
 245 {
 246         int error = blk_status_to_errno(bio->bi_status);
 247         struct bio_vec *bvec;
 248         struct bvec_iter_all iter_all;
 249
 250         bio_for_each_segment_all(bvec, bio, iter_all)
 251                 iomap_read_page_end_io(bvec, error);
 252         bio_put(bio);
 253 }
 254
 255 struct iomap_readpage_ctx {
 256         struct page             *cur_page;
 257         bool                    cur_page_in_bio;
 258         bool                    is_readahead;
 259         struct bio              *bio;
 260         struct list_head        *pages;
 261 };
 262
 263 static void
 264 iomap_read_inline_data(struct inode *inode, struct page *page,
 265                 struct iomap *iomap)
 266 {
 267         size_t size = i_size_read(inode);
 268         void *addr;
 269
 270         if (PageUptodate(page))
 271                 return;
 272
 273         BUG_ON(page->index);
 274         BUG_ON(size > PAGE_SIZE - offset_in_page(iomap->inline_data));
 275
 276         addr = kmap_atomic(page);
 277         memcpy(addr, iomap->inline_data, size);
 278         memset(addr + size, 0, PAGE_SIZE - size);
 279         kunmap_atomic(addr);
 280         SetPageUptodate(page);
 281 }
 282
 283 static loff_t
 284 iomap_readpage_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 285                 struct iomap *iomap)
 286 {
 287         struct iomap_readpage_ctx *ctx = data;
 288         struct page *page = ctx->cur_page;
 289         struct iomap_page *iop = iomap_page_create(inode, page);
 290         bool same_page = false, is_contig = false;
 291         loff_t orig_pos = pos;
 292         unsigned poff, plen;
 293         sector_t sector;
 294
 295         if (iomap->type == IOMAP_INLINE) {
 296                 WARN_ON_ONCE(pos);
 297                 iomap_read_inline_data(inode, page, iomap);
 298                 return PAGE_SIZE;
 299         }
 300
 301         /* zero post-eof blocks as the page may be mapped */
 302         iomap_adjust_read_range(inode, iop, &pos, length, &poff, &plen);
 303         if (plen == 0)
 304                 goto done;
 305
 306         if (iomap->type != IOMAP_MAPPED || pos >= i_size_read(inode)) {
 307                 zero_user(page, poff, plen);
 308                 iomap_set_range_uptodate(page, poff, plen);
 309                 goto done;
 310         }
 311
 312         ctx->cur_page_in_bio = true;
 313
 314         /*
 315          * Try to merge into a previous segment if we can.
 316          */
 317         sector = iomap_sector(iomap, pos);
 318         if (ctx->bio && bio_end_sector(ctx->bio) == sector)
 319                 is_contig = true;
 320
 321         if (is_contig &&
 322             __bio_try_merge_page(ctx->bio, page, plen, poff, &same_page)) {
 323                 if (!same_page && iop)
 324                         atomic_inc(&iop->read_count);
 325                 goto done;
 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, 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         const struct iomap_page_ops *page_ops = iomap->page_ops;
 664         pgoff_t index = pos >> PAGE_SHIFT;
 665         struct page *page;
 666         int status = 0;
 667
 668         BUG_ON(pos + len > iomap->offset + iomap->length);
 669
 670         if (fatal_signal_pending(current))
 671                 return -EINTR;
 672
 673         if (page_ops && page_ops->page_prepare) {
 674                 status = page_ops->page_prepare(inode, pos, len, iomap);
 675                 if (status)
 676                         return status;
 677         }
 678
 679         page = grab_cache_page_write_begin(inode->i_mapping, index, flags);
 680         if (!page) {
 681                 status = -ENOMEM;
 682                 goto out_no_page;
 683         }
 684
 685         if (iomap->type == IOMAP_INLINE)
 686                 iomap_read_inline_data(inode, page, iomap);
 687         else if (iomap->flags & IOMAP_F_BUFFER_HEAD)
 688                 status = __block_write_begin_int(page, pos, len, NULL, iomap);
 689         else
 690                 status = __iomap_write_begin(inode, pos, len, page, iomap);
 691
 692         if (unlikely(status))
 693                 goto out_unlock;
 694
 695         *pagep = page;
 696         return 0;
 697
 698 out_unlock:
 699         unlock_page(page);
 700         put_page(page);
 701         iomap_write_failed(inode, pos, len);
 702
 703 out_no_page:
 704         if (page_ops && page_ops->page_done)
 705                 page_ops->page_done(inode, pos, 0, NULL, iomap);
 706         return status;
 707 }
 708
 709 int
 710 iomap_set_page_dirty(struct page *page)
 711 {
 712         struct address_space *mapping = page_mapping(page);
 713         int newly_dirty;
 714
 715         if (unlikely(!mapping))
 716                 return !TestSetPageDirty(page);
 717
 718         /*
 719          * Lock out page->mem_cgroup migration to keep PageDirty
 720          * synchronized with per-memcg dirty page counters.
 721          */
 722         lock_page_memcg(page);
 723         newly_dirty = !TestSetPageDirty(page);
 724         if (newly_dirty)
 725                 __set_page_dirty(page, mapping, 0);
 726         unlock_page_memcg(page);
 727
 728         if (newly_dirty)
 729                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 730         return newly_dirty;
 731 }
 732 EXPORT_SYMBOL_GPL(iomap_set_page_dirty);
 733
 734 static int
 735 __iomap_write_end(struct inode *inode, loff_t pos, unsigned len,
 736                 unsigned copied, struct page *page, struct iomap *iomap)
 737 {
 738         flush_dcache_page(page);
 739
 740         /*
 741          * The blocks that were entirely written will now be uptodate, so we
 742          * don't have to worry about a readpage reading them and overwriting a
 743          * partial write.  However if we have encountered a short write and only
 744          * partially written into a block, it will not be marked uptodate, so a
 745          * readpage might come in and destroy our partial write.
 746          *
 747          * Do the simplest thing, and just treat any short write to a non
 748          * uptodate page as a zero-length write, and force the caller to redo
 749          * the whole thing.
 750          */
 751         if (unlikely(copied < len && !PageUptodate(page)))
 752                 return 0;
 753         iomap_set_range_uptodate(page, offset_in_page(pos), len);
 754         iomap_set_page_dirty(page);
 755         return copied;
 756 }
 757
 758 static int
 759 iomap_write_end_inline(struct inode *inode, struct page *page,
 760                 struct iomap *iomap, loff_t pos, unsigned copied)
 761 {
 762         void *addr;
 763
 764         WARN_ON_ONCE(!PageUptodate(page));
 765         BUG_ON(pos + copied > PAGE_SIZE - offset_in_page(iomap->inline_data));
 766
 767         addr = kmap_atomic(page);
 768         memcpy(iomap->inline_data + pos, addr + pos, copied);
 769         kunmap_atomic(addr);
 770
 771         mark_inode_dirty(inode);
 772         return copied;
 773 }
 774
 775 static int
 776 iomap_write_end(struct inode *inode, loff_t pos, unsigned len,
 777                 unsigned copied, struct page *page, struct iomap *iomap)
 778 {
 779         const struct iomap_page_ops *page_ops = iomap->page_ops;
 780         int ret;
 781
 782         if (iomap->type == IOMAP_INLINE) {
 783                 ret = iomap_write_end_inline(inode, page, iomap, pos, copied);
 784         } else if (iomap->flags & IOMAP_F_BUFFER_HEAD) {
 785                 ret = block_write_end(NULL, inode->i_mapping, pos, len, copied,
 786                                 page, NULL);
 787         } else {
 788                 ret = __iomap_write_end(inode, pos, len, copied, page, iomap);
 789         }
 790
 791         __generic_write_end(inode, pos, ret, page);
 792         if (page_ops && page_ops->page_done)
 793                 page_ops->page_done(inode, pos, copied, page, iomap);
 794         put_page(page);
 795
 796         if (ret < len)
 797                 iomap_write_failed(inode, pos, len);
 798         return ret;
 799 }
 800
 801 static loff_t
 802 iomap_write_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 803                 struct iomap *iomap)
 804 {
 805         struct iov_iter *i = data;
 806         long status = 0;
 807         ssize_t written = 0;
 808         unsigned int flags = AOP_FLAG_NOFS;
 809
 810         do {
 811                 struct page *page;
 812                 unsigned long offset;   /* Offset into pagecache page */
 813                 unsigned long bytes;    /* Bytes to write to page */
 814                 size_t copied;          /* Bytes copied from user */
 815
 816                 offset = offset_in_page(pos);
 817                 bytes = min_t(unsigned long, PAGE_SIZE - offset,
 818                                                 iov_iter_count(i));
 819 again:
 820                 if (bytes > length)
 821                         bytes = length;
 822
 823                 /*
 824                  * Bring in the user page that we will copy from _first_.
 825                  * Otherwise there's a nasty deadlock on copying from the
 826                  * same page as we're writing to, without it being marked
 827                  * up-to-date.
 828                  *
 829                  * Not only is this an optimisation, but it is also required
 830                  * to check that the address is actually valid, when atomic
 831                  * usercopies are used, below.
 832                  */
 833                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
 834                         status = -EFAULT;
 835                         break;
 836                 }
 837
 838                 status = iomap_write_begin(inode, pos, bytes, flags, &page,
 839                                 iomap);
 840                 if (unlikely(status))
 841                         break;
 842
 843                 if (mapping_writably_mapped(inode->i_mapping))
 844                         flush_dcache_page(page);
 845
 846                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
 847
 848                 flush_dcache_page(page);
 849
 850                 status = iomap_write_end(inode, pos, bytes, copied, page,
 851                                 iomap);
 852                 if (unlikely(status < 0))
 853                         break;
 854                 copied = status;
 855
 856                 cond_resched();
 857
 858                 iov_iter_advance(i, copied);
 859                 if (unlikely(copied == 0)) {
 860                         /*
 861                          * If we were unable to copy any data at all, we must
 862                          * fall back to a single segment length write.
 863                          *
 864                          * If we didn't fallback here, we could livelock
 865                          * because not all segments in the iov can be copied at
 866                          * once without a pagefault.
 867                          */
 868                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
 869                                                 iov_iter_single_seg_count(i));
 870                         goto again;
 871                 }
 872                 pos += copied;
 873                 written += copied;
 874                 length -= copied;
 875
 876                 balance_dirty_pages_ratelimited(inode->i_mapping);
 877         } while (iov_iter_count(i) && length);
 878
 879         return written ? written : status;
 880 }
 881
 882 ssize_t
 883 iomap_file_buffered_write(struct kiocb *iocb, struct iov_iter *iter,
 884                 const struct iomap_ops *ops)
 885 {
 886         struct inode *inode = iocb->ki_filp->f_mapping->host;
 887         loff_t pos = iocb->ki_pos, ret = 0, written = 0;
 888
 889         while (iov_iter_count(iter)) {
 890                 ret = iomap_apply(inode, pos, iov_iter_count(iter),
 891                                 IOMAP_WRITE, ops, iter, iomap_write_actor);
 892                 if (ret <= 0)
 893                         break;
 894                 pos += ret;
 895                 written += ret;
 896         }
 897
 898         return written ? written : ret;
 899 }
 900 EXPORT_SYMBOL_GPL(iomap_file_buffered_write);
 901
 902 static struct page *
 903 __iomap_read_page(struct inode *inode, loff_t offset)
 904 {
 905         struct address_space *mapping = inode->i_mapping;
 906         struct page *page;
 907
 908         page = read_mapping_page(mapping, offset >> PAGE_SHIFT, NULL);
 909         if (IS_ERR(page))
 910                 return page;
 911         if (!PageUptodate(page)) {
 912                 put_page(page);
 913                 return ERR_PTR(-EIO);
 914         }
 915         return page;
 916 }
 917
 918 static loff_t
 919 iomap_dirty_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 920                 struct iomap *iomap)
 921 {
 922         long status = 0;
 923         ssize_t written = 0;
 924
 925         do {
 926                 struct page *page, *rpage;
 927                 unsigned long offset;   /* Offset into pagecache page */
 928                 unsigned long bytes;    /* Bytes to write to page */
 929
 930                 offset = offset_in_page(pos);
 931                 bytes = min_t(loff_t, PAGE_SIZE - offset, length);
 932
 933                 rpage = __iomap_read_page(inode, pos);
 934                 if (IS_ERR(rpage))
 935                         return PTR_ERR(rpage);
 936
 937                 status = iomap_write_begin(inode, pos, bytes,
 938                                            AOP_FLAG_NOFS, &page, iomap);
 939                 put_page(rpage);
 940                 if (unlikely(status))
 941                         return status;
 942
 943                 WARN_ON_ONCE(!PageUptodate(page));
 944
 945                 status = iomap_write_end(inode, pos, bytes, bytes, page, iomap);
 946                 if (unlikely(status <= 0)) {
 947                         if (WARN_ON_ONCE(status == 0))
 948                                 return -EIO;
 949                         return status;
 950                 }
 951
 952                 cond_resched();
 953
 954                 pos += status;
 955                 written += status;
 956                 length -= status;
 957
 958                 balance_dirty_pages_ratelimited(inode->i_mapping);
 959         } while (length);
 960
 961         return written;
 962 }
 963
 964 int
 965 iomap_file_dirty(struct inode *inode, loff_t pos, loff_t len,
 966                 const struct iomap_ops *ops)
 967 {
 968         loff_t ret;
 969
 970         while (len) {
 971                 ret = iomap_apply(inode, pos, len, IOMAP_WRITE, ops, NULL,
 972                                 iomap_dirty_actor);
 973                 if (ret <= 0)
 974                         return ret;
 975                 pos += ret;
 976                 len -= ret;
 977         }
 978
 979         return 0;
 980 }
 981 EXPORT_SYMBOL_GPL(iomap_file_dirty);
 982
 983 static int iomap_zero(struct inode *inode, loff_t pos, unsigned offset,
 984                 unsigned bytes, struct iomap *iomap)
 985 {
 986         struct page *page;
 987         int status;
 988
 989         status = iomap_write_begin(inode, pos, bytes, AOP_FLAG_NOFS, &page,
 990                                    iomap);
 991         if (status)
 992                 return status;
 993
 994         zero_user(page, offset, bytes);
 995         mark_page_accessed(page);
 996
 997         return iomap_write_end(inode, pos, bytes, bytes, page, iomap);
 998 }
 999
1000 static int iomap_dax_zero(loff_t pos, unsigned offset, unsigned bytes,
1001                 struct iomap *iomap)
1002 {
1003         return __dax_zero_page_range(iomap->bdev, iomap->dax_dev,
1004                         iomap_sector(iomap, pos & PAGE_MASK), offset, bytes);
1005 }
1006
1007 static loff_t
1008 iomap_zero_range_actor(struct inode *inode, loff_t pos, loff_t count,
1009                 void *data, struct iomap *iomap)
1010 {
1011         bool *did_zero = data;
1012         loff_t written = 0;
1013         int status;
1014
1015         /* already zeroed?  we're done. */
1016         if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1017                 return count;
1018
1019         do {
1020                 unsigned offset, bytes;
1021
1022                 offset = offset_in_page(pos);
1023                 bytes = min_t(loff_t, PAGE_SIZE - offset, count);
1024
1025                 if (IS_DAX(inode))
1026                         status = iomap_dax_zero(pos, offset, bytes, iomap);
1027                 else
1028                         status = iomap_zero(inode, pos, offset, bytes, iomap);
1029                 if (status < 0)
1030                         return status;
1031
1032                 pos += bytes;
1033                 count -= bytes;
1034                 written += bytes;
1035                 if (did_zero)
1036                         *did_zero = true;
1037         } while (count > 0);
1038
1039         return written;
1040 }
1041
1042 int
1043 iomap_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
1044                 const struct iomap_ops *ops)
1045 {
1046         loff_t ret;
1047
1048         while (len > 0) {
1049                 ret = iomap_apply(inode, pos, len, IOMAP_ZERO,
1050                                 ops, did_zero, iomap_zero_range_actor);
1051                 if (ret <= 0)
1052                         return ret;
1053
1054                 pos += ret;
1055                 len -= ret;
1056         }
1057
1058         return 0;
1059 }
1060 EXPORT_SYMBOL_GPL(iomap_zero_range);
1061
1062 int
1063 iomap_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
1064                 const struct iomap_ops *ops)
1065 {
1066         unsigned int blocksize = i_blocksize(inode);
1067         unsigned int off = pos & (blocksize - 1);
1068
1069         /* Block boundary? Nothing to do */
1070         if (!off)
1071                 return 0;
1072         return iomap_zero_range(inode, pos, blocksize - off, did_zero, ops);
1073 }
1074 EXPORT_SYMBOL_GPL(iomap_truncate_page);
1075
1076 static loff_t
1077 iomap_page_mkwrite_actor(struct inode *inode, loff_t pos, loff_t length,
1078                 void *data, struct iomap *iomap)
1079 {
1080         struct page *page = data;
1081         int ret;
1082
1083         if (iomap->flags & IOMAP_F_BUFFER_HEAD) {
1084                 ret = __block_write_begin_int(page, pos, length, NULL, iomap);
1085                 if (ret)
1086                         return ret;
1087                 block_commit_write(page, 0, length);
1088         } else {
1089                 WARN_ON_ONCE(!PageUptodate(page));
1090                 iomap_page_create(inode, page);
1091                 set_page_dirty(page);
1092         }
1093
1094         return length;
1095 }
1096
1097 vm_fault_t iomap_page_mkwrite(struct vm_fault *vmf, const struct iomap_ops *ops)
1098 {
1099         struct page *page = vmf->page;
1100         struct inode *inode = file_inode(vmf->vma->vm_file);
1101         unsigned long length;
1102         loff_t offset, size;
1103         ssize_t ret;
1104
1105         lock_page(page);
1106         size = i_size_read(inode);
1107         if ((page->mapping != inode->i_mapping) ||
1108             (page_offset(page) > size)) {
1109                 /* We overload EFAULT to mean page got truncated */
1110                 ret = -EFAULT;
1111                 goto out_unlock;
1112         }
1113
1114         /* page is wholly or partially inside EOF */
1115         if (((page->index + 1) << PAGE_SHIFT) > size)
1116                 length = offset_in_page(size);
1117         else
1118                 length = PAGE_SIZE;
1119
1120         offset = page_offset(page);
1121         while (length > 0) {
1122                 ret = iomap_apply(inode, offset, length,
1123                                 IOMAP_WRITE | IOMAP_FAULT, ops, page,
1124                                 iomap_page_mkwrite_actor);
1125                 if (unlikely(ret <= 0))
1126                         goto out_unlock;
1127                 offset += ret;
1128                 length -= ret;
1129         }
1130
1131         wait_for_stable_page(page);
1132         return VM_FAULT_LOCKED;
1133 out_unlock:
1134         unlock_page(page);
1135         return block_page_mkwrite_return(ret);
1136 }
1137 EXPORT_SYMBOL_GPL(iomap_page_mkwrite);
1138
1139 struct fiemap_ctx {
1140         struct fiemap_extent_info *fi;
1141         struct iomap prev;
1142 };
1143
1144 static int iomap_to_fiemap(struct fiemap_extent_info *fi,
1145                 struct iomap *iomap, u32 flags)
1146 {
1147         switch (iomap->type) {
1148         case IOMAP_HOLE:
1149                 /* skip holes */
1150                 return 0;
1151         case IOMAP_DELALLOC:
1152                 flags |= FIEMAP_EXTENT_DELALLOC | FIEMAP_EXTENT_UNKNOWN;
1153                 break;
1154         case IOMAP_MAPPED:
1155                 break;
1156         case IOMAP_UNWRITTEN:
1157                 flags |= FIEMAP_EXTENT_UNWRITTEN;
1158                 break;
1159         case IOMAP_INLINE:
1160                 flags |= FIEMAP_EXTENT_DATA_INLINE;
1161                 break;
1162         }
1163
1164         if (iomap->flags & IOMAP_F_MERGED)
1165                 flags |= FIEMAP_EXTENT_MERGED;
1166         if (iomap->flags & IOMAP_F_SHARED)
1167                 flags |= FIEMAP_EXTENT_SHARED;
1168
1169         return fiemap_fill_next_extent(fi, iomap->offset,
1170                         iomap->addr != IOMAP_NULL_ADDR ? iomap->addr : 0,
1171                         iomap->length, flags);
1172 }
1173
1174 static loff_t
1175 iomap_fiemap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
1176                 struct iomap *iomap)
1177 {
1178         struct fiemap_ctx *ctx = data;
1179         loff_t ret = length;
1180
1181         if (iomap->type == IOMAP_HOLE)
1182                 return length;
1183
1184         ret = iomap_to_fiemap(ctx->fi, &ctx->prev, 0);
1185         ctx->prev = *iomap;
1186         switch (ret) {
1187         case 0:         /* success */
1188                 return length;
1189         case 1:         /* extent array full */
1190                 return 0;
1191         default:
1192                 return ret;
1193         }
1194 }
1195
1196 int iomap_fiemap(struct inode *inode, struct fiemap_extent_info *fi,
1197                 loff_t start, loff_t len, const struct iomap_ops *ops)
1198 {
1199         struct fiemap_ctx ctx;
1200         loff_t ret;
1201
1202         memset(&ctx, 0, sizeof(ctx));
1203         ctx.fi = fi;
1204         ctx.prev.type = IOMAP_HOLE;
1205
1206         ret = fiemap_check_flags(fi, FIEMAP_FLAG_SYNC);
1207         if (ret)
1208                 return ret;
1209
1210         if (fi->fi_flags & FIEMAP_FLAG_SYNC) {
1211                 ret = filemap_write_and_wait(inode->i_mapping);
1212                 if (ret)
1213                         return ret;
1214         }
1215
1216         while (len > 0) {
1217                 ret = iomap_apply(inode, start, len, IOMAP_REPORT, ops, &ctx,
1218                                 iomap_fiemap_actor);
1219                 /* inode with no (attribute) mapping will give ENOENT */
1220                 if (ret == -ENOENT)
1221                         break;
1222                 if (ret < 0)
1223                         return ret;
1224                 if (ret == 0)
1225                         break;
1226
1227                 start += ret;
1228                 len -= ret;
1229         }
1230
1231         if (ctx.prev.type != IOMAP_HOLE) {
1232                 ret = iomap_to_fiemap(fi, &ctx.prev, FIEMAP_EXTENT_LAST);
1233                 if (ret < 0)
1234                         return ret;
1235         }
1236
1237         return 0;
1238 }
1239 EXPORT_SYMBOL_GPL(iomap_fiemap);
1240
1241 /*
1242  * Seek for SEEK_DATA / SEEK_HOLE within @page, starting at @lastoff.
1243  * Returns true if found and updates @lastoff to the offset in file.
1244  */
1245 static bool
1246 page_seek_hole_data(struct inode *inode, struct page *page, loff_t *lastoff,
1247                 int whence)
1248 {
1249         const struct address_space_operations *ops = inode->i_mapping->a_ops;
1250         unsigned int bsize = i_blocksize(inode), off;
1251         bool seek_data = whence == SEEK_DATA;
1252         loff_t poff = page_offset(page);
1253
1254         if (WARN_ON_ONCE(*lastoff >= poff + PAGE_SIZE))
1255                 return false;
1256
1257         if (*lastoff < poff) {
1258                 /*
1259                  * Last offset smaller than the start of the page means we found
1260                  * a hole:
1261                  */
1262                 if (whence == SEEK_HOLE)
1263                         return true;
1264                 *lastoff = poff;
1265         }
1266
1267         /*
1268          * Just check the page unless we can and should check block ranges:
1269          */
1270         if (bsize == PAGE_SIZE || !ops->is_partially_uptodate)
1271                 return PageUptodate(page) == seek_data;
1272
1273         lock_page(page);
1274         if (unlikely(page->mapping != inode->i_mapping))
1275                 goto out_unlock_not_found;
1276
1277         for (off = 0; off < PAGE_SIZE; off += bsize) {
1278                 if (offset_in_page(*lastoff) >= off + bsize)
1279                         continue;
1280                 if (ops->is_partially_uptodate(page, off, bsize) == seek_data) {
1281                         unlock_page(page);
1282                         return true;
1283                 }
1284                 *lastoff = poff + off + bsize;
1285         }
1286
1287 out_unlock_not_found:
1288         unlock_page(page);
1289         return false;
1290 }
1291
1292 /*
1293  * Seek for SEEK_DATA / SEEK_HOLE in the page cache.
1294  *
1295  * Within unwritten extents, the page cache determines which parts are holes
1296  * and which are data: uptodate buffer heads count as data; everything else
1297  * counts as a hole.
1298  *
1299  * Returns the resulting offset on successs, and -ENOENT otherwise.
1300  */
1301 static loff_t
1302 page_cache_seek_hole_data(struct inode *inode, loff_t offset, loff_t length,
1303                 int whence)
1304 {
1305         pgoff_t index = offset >> PAGE_SHIFT;
1306         pgoff_t end = DIV_ROUND_UP(offset + length, PAGE_SIZE);
1307         loff_t lastoff = offset;
1308         struct pagevec pvec;
1309
1310         if (length <= 0)
1311                 return -ENOENT;
1312
1313         pagevec_init(&pvec);
1314
1315         do {
1316                 unsigned nr_pages, i;
1317
1318                 nr_pages = pagevec_lookup_range(&pvec, inode->i_mapping, &index,
1319                                                 end - 1);
1320                 if (nr_pages == 0)
1321                         break;
1322
1323                 for (i = 0; i < nr_pages; i++) {
1324                         struct page *page = pvec.pages[i];
1325
1326                         if (page_seek_hole_data(inode, page, &lastoff, whence))
1327                                 goto check_range;
1328                         lastoff = page_offset(page) + PAGE_SIZE;
1329                 }
1330                 pagevec_release(&pvec);
1331         } while (index < end);
1332
1333         /* When no page at lastoff and we are not done, we found a hole. */
1334         if (whence != SEEK_HOLE)
1335                 goto not_found;
1336
1337 check_range:
1338         if (lastoff < offset + length)
1339                 goto out;
1340 not_found:
1341         lastoff = -ENOENT;
1342 out:
1343         pagevec_release(&pvec);
1344         return lastoff;
1345 }
1346
1347
1348 static loff_t
1349 iomap_seek_hole_actor(struct inode *inode, loff_t offset, loff_t length,
1350                       void *data, struct iomap *iomap)
1351 {
1352         switch (iomap->type) {
1353         case IOMAP_UNWRITTEN:
1354                 offset = page_cache_seek_hole_data(inode, offset, length,
1355                                                    SEEK_HOLE);
1356                 if (offset < 0)
1357                         return length;
1358                 /* fall through */
1359         case IOMAP_HOLE:
1360                 *(loff_t *)data = offset;
1361                 return 0;
1362         default:
1363                 return length;
1364         }
1365 }
1366
1367 loff_t
1368 iomap_seek_hole(struct inode *inode, loff_t offset, const struct iomap_ops *ops)
1369 {
1370         loff_t size = i_size_read(inode);
1371         loff_t length = size - offset;
1372         loff_t ret;
1373
1374         /* Nothing to be found before or beyond the end of the file. */
1375         if (offset < 0 || offset >= size)
1376                 return -ENXIO;
1377
1378         while (length > 0) {
1379                 ret = iomap_apply(inode, offset, length, IOMAP_REPORT, ops,
1380                                   &offset, iomap_seek_hole_actor);
1381                 if (ret < 0)
1382                         return ret;
1383                 if (ret == 0)
1384                         break;
1385
1386                 offset += ret;
1387                 length -= ret;
1388         }
1389
1390         return offset;
1391 }
1392 EXPORT_SYMBOL_GPL(iomap_seek_hole);
1393
1394 static loff_t
1395 iomap_seek_data_actor(struct inode *inode, loff_t offset, loff_t length,
1396                       void *data, struct iomap *iomap)
1397 {
1398         switch (iomap->type) {
1399         case IOMAP_HOLE:
1400                 return length;
1401         case IOMAP_UNWRITTEN:
1402                 offset = page_cache_seek_hole_data(inode, offset, length,
1403                                                    SEEK_DATA);
1404                 if (offset < 0)
1405                         return length;
1406                 /*FALLTHRU*/
1407         default:
1408                 *(loff_t *)data = offset;
1409                 return 0;
1410         }
1411 }
1412
1413 loff_t
1414 iomap_seek_data(struct inode *inode, loff_t offset, const struct iomap_ops *ops)
1415 {
1416         loff_t size = i_size_read(inode);
1417         loff_t length = size - offset;
1418         loff_t ret;
1419
1420         /* Nothing to be found before or beyond the end of the file. */
1421         if (offset < 0 || offset >= size)
1422                 return -ENXIO;
1423
1424         while (length > 0) {
1425                 ret = iomap_apply(inode, offset, length, IOMAP_REPORT, ops,
1426                                   &offset, iomap_seek_data_actor);
1427                 if (ret < 0)
1428                         return ret;
1429                 if (ret == 0)
1430                         break;
1431
1432                 offset += ret;
1433                 length -= ret;
1434         }
1435
1436         if (length <= 0)
1437                 return -ENXIO;
1438         return offset;
1439 }
1440 EXPORT_SYMBOL_GPL(iomap_seek_data);
1441
1442 /*
1443  * Private flags for iomap_dio, must not overlap with the public ones in
1444  * iomap.h:
1445  */
1446 #define IOMAP_DIO_WRITE_FUA     (1 << 28)
1447 #define IOMAP_DIO_NEED_SYNC     (1 << 29)
1448 #define IOMAP_DIO_WRITE         (1 << 30)
1449 #define IOMAP_DIO_DIRTY         (1 << 31)
1450
1451 struct iomap_dio {
1452         struct kiocb            *iocb;
1453         iomap_dio_end_io_t      *end_io;
1454         loff_t                  i_size;
1455         loff_t                  size;
1456         atomic_t                ref;
1457         unsigned                flags;
1458         int                     error;
1459         bool                    wait_for_completion;
1460
1461         union {
1462                 /* used during submission and for synchronous completion: */
1463                 struct {
1464                         struct iov_iter         *iter;
1465                         struct task_struct      *waiter;
1466                         struct request_queue    *last_queue;
1467                         blk_qc_t                cookie;
1468                 } submit;
1469
1470                 /* used for aio completion: */
1471                 struct {
1472                         struct work_struct      work;
1473                 } aio;
1474         };
1475 };
1476
1477 int iomap_dio_iopoll(struct kiocb *kiocb, bool spin)
1478 {
1479         struct request_queue *q = READ_ONCE(kiocb->private);
1480
1481         if (!q)
1482                 return 0;
1483         return blk_poll(q, READ_ONCE(kiocb->ki_cookie), spin);
1484 }
1485 EXPORT_SYMBOL_GPL(iomap_dio_iopoll);
1486
1487 static void iomap_dio_submit_bio(struct iomap_dio *dio, struct iomap *iomap,
1488                 struct bio *bio)
1489 {
1490         atomic_inc(&dio->ref);
1491
1492         if (dio->iocb->ki_flags & IOCB_HIPRI)
1493                 bio_set_polled(bio, dio->iocb);
1494
1495         dio->submit.last_queue = bdev_get_queue(iomap->bdev);
1496         dio->submit.cookie = submit_bio(bio);
1497 }
1498
1499 static ssize_t iomap_dio_complete(struct iomap_dio *dio)
1500 {
1501         struct kiocb *iocb = dio->iocb;
1502         struct inode *inode = file_inode(iocb->ki_filp);
1503         loff_t offset = iocb->ki_pos;
1504         ssize_t ret;
1505
1506         if (dio->end_io) {
1507                 ret = dio->end_io(iocb,
1508                                 dio->error ? dio->error : dio->size,
1509                                 dio->flags);
1510         } else {
1511                 ret = dio->error;
1512         }
1513
1514         if (likely(!ret)) {
1515                 ret = dio->size;
1516                 /* check for short read */
1517                 if (offset + ret > dio->i_size &&
1518                     !(dio->flags & IOMAP_DIO_WRITE))
1519                         ret = dio->i_size - offset;
1520                 iocb->ki_pos += ret;
1521         }
1522
1523         /*
1524          * Try again to invalidate clean pages which might have been cached by
1525          * non-direct readahead, or faulted in by get_user_pages() if the source
1526          * of the write was an mmap'ed region of the file we're writing.  Either
1527          * one is a pretty crazy thing to do, so we don't support it 100%.  If
1528          * this invalidation fails, tough, the write still worked...
1529          *
1530          * And this page cache invalidation has to be after dio->end_io(), as
1531          * some filesystems convert unwritten extents to real allocations in
1532          * end_io() when necessary, otherwise a racing buffer read would cache
1533          * zeros from unwritten extents.
1534          */
1535         if (!dio->error &&
1536             (dio->flags & IOMAP_DIO_WRITE) && inode->i_mapping->nrpages) {
1537                 int err;
1538                 err = invalidate_inode_pages2_range(inode->i_mapping,
1539                                 offset >> PAGE_SHIFT,
1540                                 (offset + dio->size - 1) >> PAGE_SHIFT);
1541                 if (err)
1542                         dio_warn_stale_pagecache(iocb->ki_filp);
1543         }
1544
1545         /*
1546          * If this is a DSYNC write, make sure we push it to stable storage now
1547          * that we've written data.
1548          */
1549         if (ret > 0 && (dio->flags & IOMAP_DIO_NEED_SYNC))
1550                 ret = generic_write_sync(iocb, ret);
1551
1552         inode_dio_end(file_inode(iocb->ki_filp));
1553         kfree(dio);
1554
1555         return ret;
1556 }
1557
1558 static void iomap_dio_complete_work(struct work_struct *work)
1559 {
1560         struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work);
1561         struct kiocb *iocb = dio->iocb;
1562
1563         iocb->ki_complete(iocb, iomap_dio_complete(dio), 0);
1564 }
1565
1566 /*
1567  * Set an error in the dio if none is set yet.  We have to use cmpxchg
1568  * as the submission context and the completion context(s) can race to
1569  * update the error.
1570  */
1571 static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret)
1572 {
1573         cmpxchg(&dio->error, 0, ret);
1574 }
1575
1576 static void iomap_dio_bio_end_io(struct bio *bio)
1577 {
1578         struct iomap_dio *dio = bio->bi_private;
1579         bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
1580
1581         if (bio->bi_status)
1582                 iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
1583
1584         if (atomic_dec_and_test(&dio->ref)) {
1585                 if (dio->wait_for_completion) {
1586                         struct task_struct *waiter = dio->submit.waiter;
1587                         WRITE_ONCE(dio->submit.waiter, NULL);
1588                         blk_wake_io_task(waiter);
1589                 } else if (dio->flags & IOMAP_DIO_WRITE) {
1590                         struct inode *inode = file_inode(dio->iocb->ki_filp);
1591
1592                         INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
1593                         queue_work(inode->i_sb->s_dio_done_wq, &dio->aio.work);
1594                 } else {
1595                         iomap_dio_complete_work(&dio->aio.work);
1596                 }
1597         }
1598
1599         if (should_dirty) {
1600                 bio_check_pages_dirty(bio);
1601         } else {
1602                 if (!bio_flagged(bio, BIO_NO_PAGE_REF)) {
1603                         struct bvec_iter_all iter_all;
1604                         struct bio_vec *bvec;
1605
1606                         bio_for_each_segment_all(bvec, bio, iter_all)
1607                                 put_page(bvec->bv_page);
1608                 }
1609                 bio_put(bio);
1610         }
1611 }
1612
1613 static void
1614 iomap_dio_zero(struct iomap_dio *dio, struct iomap *iomap, loff_t pos,
1615                 unsigned len)
1616 {
1617         struct page *page = ZERO_PAGE(0);
1618         int flags = REQ_SYNC | REQ_IDLE;
1619         struct bio *bio;
1620
1621         bio = bio_alloc(GFP_KERNEL, 1);
1622         bio_set_dev(bio, iomap->bdev);
1623         bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
1624         bio->bi_private = dio;
1625         bio->bi_end_io = iomap_dio_bio_end_io;
1626
1627         get_page(page);
1628         __bio_add_page(bio, page, len, 0);
1629         bio_set_op_attrs(bio, REQ_OP_WRITE, flags);
1630         iomap_dio_submit_bio(dio, iomap, bio);
1631 }
1632
1633 static loff_t
1634 iomap_dio_bio_actor(struct inode *inode, loff_t pos, loff_t length,
1635                 struct iomap_dio *dio, struct iomap *iomap)
1636 {
1637         unsigned int blkbits = blksize_bits(bdev_logical_block_size(iomap->bdev));
1638         unsigned int fs_block_size = i_blocksize(inode), pad;
1639         unsigned int align = iov_iter_alignment(dio->submit.iter);
1640         struct iov_iter iter;
1641         struct bio *bio;
1642         bool need_zeroout = false;
1643         bool use_fua = false;
1644         int nr_pages, ret = 0;
1645         size_t copied = 0;
1646
1647         if ((pos | length | align) & ((1 << blkbits) - 1))
1648                 return -EINVAL;
1649
1650         if (iomap->type == IOMAP_UNWRITTEN) {
1651                 dio->flags |= IOMAP_DIO_UNWRITTEN;
1652                 need_zeroout = true;
1653         }
1654
1655         if (iomap->flags & IOMAP_F_SHARED)
1656                 dio->flags |= IOMAP_DIO_COW;
1657
1658         if (iomap->flags & IOMAP_F_NEW) {
1659                 need_zeroout = true;
1660         } else if (iomap->type == IOMAP_MAPPED) {
1661                 /*
1662                  * Use a FUA write if we need datasync semantics, this is a pure
1663                  * data IO that doesn't require any metadata updates (including
1664                  * after IO completion such as unwritten extent conversion) and
1665                  * the underlying device supports FUA. This allows us to avoid
1666                  * cache flushes on IO completion.
1667                  */
1668                 if (!(iomap->flags & (IOMAP_F_SHARED|IOMAP_F_DIRTY)) &&
1669                     (dio->flags & IOMAP_DIO_WRITE_FUA) &&
1670                     blk_queue_fua(bdev_get_queue(iomap->bdev)))
1671                         use_fua = true;
1672         }
1673
1674         /*
1675          * Operate on a partial iter trimmed to the extent we were called for.
1676          * We'll update the iter in the dio once we're done with this extent.
1677          */
1678         iter = *dio->submit.iter;
1679         iov_iter_truncate(&iter, length);
1680
1681         nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
1682         if (nr_pages <= 0)
1683                 return nr_pages;
1684
1685         if (need_zeroout) {
1686                 /* zero out from the start of the block to the write offset */
1687                 pad = pos & (fs_block_size - 1);
1688                 if (pad)
1689                         iomap_dio_zero(dio, iomap, pos - pad, pad);
1690         }
1691
1692         do {
1693                 size_t n;
1694                 if (dio->error) {
1695                         iov_iter_revert(dio->submit.iter, copied);
1696                         return 0;
1697                 }
1698
1699                 bio = bio_alloc(GFP_KERNEL, nr_pages);
1700                 bio_set_dev(bio, iomap->bdev);
1701                 bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
1702                 bio->bi_write_hint = dio->iocb->ki_hint;
1703                 bio->bi_ioprio = dio->iocb->ki_ioprio;
1704                 bio->bi_private = dio;
1705                 bio->bi_end_io = iomap_dio_bio_end_io;
1706
1707                 ret = bio_iov_iter_get_pages(bio, &iter);
1708                 if (unlikely(ret)) {
1709                         /*
1710                          * We have to stop part way through an IO. We must fall
1711                          * through to the sub-block tail zeroing here, otherwise
1712                          * this short IO may expose stale data in the tail of
1713                          * the block we haven't written data to.
1714                          */
1715                         bio_put(bio);
1716                         goto zero_tail;
1717                 }
1718
1719                 n = bio->bi_iter.bi_size;
1720                 if (dio->flags & IOMAP_DIO_WRITE) {
1721                         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
1722                         if (use_fua)
1723                                 bio->bi_opf |= REQ_FUA;
1724                         else
1725                                 dio->flags &= ~IOMAP_DIO_WRITE_FUA;
1726                         task_io_account_write(n);
1727                 } else {
1728                         bio->bi_opf = REQ_OP_READ;
1729                         if (dio->flags & IOMAP_DIO_DIRTY)
1730                                 bio_set_pages_dirty(bio);
1731                 }
1732
1733                 iov_iter_advance(dio->submit.iter, n);
1734
1735                 dio->size += n;
1736                 pos += n;
1737                 copied += n;
1738
1739                 nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
1740                 iomap_dio_submit_bio(dio, iomap, bio);
1741         } while (nr_pages);
1742
1743         /*
1744          * We need to zeroout the tail of a sub-block write if the extent type
1745          * requires zeroing or the write extends beyond EOF. If we don't zero
1746          * the block tail in the latter case, we can expose stale data via mmap
1747          * reads of the EOF block.
1748          */
1749 zero_tail:
1750         if (need_zeroout ||
1751             ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode))) {
1752                 /* zero out from the end of the write to the end of the block */
1753                 pad = pos & (fs_block_size - 1);
1754                 if (pad)
1755                         iomap_dio_zero(dio, iomap, pos, fs_block_size - pad);
1756         }
1757         return copied ? copied : ret;
1758 }
1759
1760 static loff_t
1761 iomap_dio_hole_actor(loff_t length, struct iomap_dio *dio)
1762 {
1763         length = iov_iter_zero(length, dio->submit.iter);
1764         dio->size += length;
1765         return length;
1766 }
1767
1768 static loff_t
1769 iomap_dio_inline_actor(struct inode *inode, loff_t pos, loff_t length,
1770                 struct iomap_dio *dio, struct iomap *iomap)
1771 {
1772         struct iov_iter *iter = dio->submit.iter;
1773         size_t copied;
1774
1775         BUG_ON(pos + length > PAGE_SIZE - offset_in_page(iomap->inline_data));
1776
1777         if (dio->flags & IOMAP_DIO_WRITE) {
1778                 loff_t size = inode->i_size;
1779
1780                 if (pos > size)
1781                         memset(iomap->inline_data + size, 0, pos - size);
1782                 copied = copy_from_iter(iomap->inline_data + pos, length, iter);
1783                 if (copied) {
1784                         if (pos + copied > size)
1785                                 i_size_write(inode, pos + copied);
1786                         mark_inode_dirty(inode);
1787                 }
1788         } else {
1789                 copied = copy_to_iter(iomap->inline_data + pos, length, iter);
1790         }
1791         dio->size += copied;
1792         return copied;
1793 }
1794
1795 static loff_t
1796 iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
1797                 void *data, struct iomap *iomap)
1798 {
1799         struct iomap_dio *dio = data;
1800
1801         switch (iomap->type) {
1802         case IOMAP_HOLE:
1803                 if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
1804                         return -EIO;
1805                 return iomap_dio_hole_actor(length, dio);
1806         case IOMAP_UNWRITTEN:
1807                 if (!(dio->flags & IOMAP_DIO_WRITE))
1808                         return iomap_dio_hole_actor(length, dio);
1809                 return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
1810         case IOMAP_MAPPED:
1811                 return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
1812         case IOMAP_INLINE:
1813                 return iomap_dio_inline_actor(inode, pos, length, dio, iomap);
1814         default:
1815                 WARN_ON_ONCE(1);
1816                 return -EIO;
1817         }
1818 }
1819
1820 /*
1821  * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
1822  * is being issued as AIO or not.  This allows us to optimise pure data writes
1823  * to use REQ_FUA rather than requiring generic_write_sync() to issue a
1824  * REQ_FLUSH post write. This is slightly tricky because a single request here
1825  * can be mapped into multiple disjoint IOs and only a subset of the IOs issued
1826  * may be pure data writes. In that case, we still need to do a full data sync
1827  * completion.
1828  */
1829 ssize_t
1830 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
1831                 const struct iomap_ops *ops, iomap_dio_end_io_t end_io)
1832 {
1833         struct address_space *mapping = iocb->ki_filp->f_mapping;
1834         struct inode *inode = file_inode(iocb->ki_filp);
1835         size_t count = iov_iter_count(iter);
1836         loff_t pos = iocb->ki_pos, start = pos;
1837         loff_t end = iocb->ki_pos + count - 1, ret = 0;
1838         unsigned int flags = IOMAP_DIRECT;
1839         bool wait_for_completion = is_sync_kiocb(iocb);
1840         struct blk_plug plug;
1841         struct iomap_dio *dio;
1842
1843         lockdep_assert_held(&inode->i_rwsem);
1844
1845         if (!count)
1846                 return 0;
1847
1848         dio = kmalloc(sizeof(*dio), GFP_KERNEL);
1849         if (!dio)
1850                 return -ENOMEM;
1851
1852         dio->iocb = iocb;
1853         atomic_set(&dio->ref, 1);
1854         dio->size = 0;
1855         dio->i_size = i_size_read(inode);
1856         dio->end_io = end_io;
1857         dio->error = 0;
1858         dio->flags = 0;
1859
1860         dio->submit.iter = iter;
1861         dio->submit.waiter = current;
1862         dio->submit.cookie = BLK_QC_T_NONE;
1863         dio->submit.last_queue = NULL;
1864
1865         if (iov_iter_rw(iter) == READ) {
1866                 if (pos >= dio->i_size)
1867                         goto out_free_dio;
1868
1869                 if (iter_is_iovec(iter) && iov_iter_rw(iter) == READ)
1870                         dio->flags |= IOMAP_DIO_DIRTY;
1871         } else {
1872                 flags |= IOMAP_WRITE;
1873                 dio->flags |= IOMAP_DIO_WRITE;
1874
1875                 /* for data sync or sync, we need sync completion processing */
1876                 if (iocb->ki_flags & IOCB_DSYNC)
1877                         dio->flags |= IOMAP_DIO_NEED_SYNC;
1878
1879                 /*
1880                  * For datasync only writes, we optimistically try using FUA for
1881                  * this IO.  Any non-FUA write that occurs will clear this flag,
1882                  * hence we know before completion whether a cache flush is
1883                  * necessary.
1884                  */
1885                 if ((iocb->ki_flags & (IOCB_DSYNC | IOCB_SYNC)) == IOCB_DSYNC)
1886                         dio->flags |= IOMAP_DIO_WRITE_FUA;
1887         }
1888
1889         if (iocb->ki_flags & IOCB_NOWAIT) {
1890                 if (filemap_range_has_page(mapping, start, end)) {
1891                         ret = -EAGAIN;
1892                         goto out_free_dio;
1893                 }
1894                 flags |= IOMAP_NOWAIT;
1895         }
1896
1897         ret = filemap_write_and_wait_range(mapping, start, end);
1898         if (ret)
1899                 goto out_free_dio;
1900
1901         /*
1902          * Try to invalidate cache pages for the range we're direct
1903          * writing.  If this invalidation fails, tough, the write will
1904          * still work, but racing two incompatible write paths is a
1905          * pretty crazy thing to do, so we don't support it 100%.
1906          */
1907         ret = invalidate_inode_pages2_range(mapping,
1908                         start >> PAGE_SHIFT, end >> PAGE_SHIFT);
1909         if (ret)
1910                 dio_warn_stale_pagecache(iocb->ki_filp);
1911         ret = 0;
1912
1913         if (iov_iter_rw(iter) == WRITE && !wait_for_completion &&
1914             !inode->i_sb->s_dio_done_wq) {
1915                 ret = sb_init_dio_done_wq(inode->i_sb);
1916                 if (ret < 0)
1917                         goto out_free_dio;
1918         }
1919
1920         inode_dio_begin(inode);
1921
1922         blk_start_plug(&plug);
1923         do {
1924                 ret = iomap_apply(inode, pos, count, flags, ops, dio,
1925                                 iomap_dio_actor);
1926                 if (ret <= 0) {
1927                         /* magic error code to fall back to buffered I/O */
1928                         if (ret == -ENOTBLK) {
1929                                 wait_for_completion = true;
1930                                 ret = 0;
1931                         }
1932                         break;
1933                 }
1934                 pos += ret;
1935
1936                 if (iov_iter_rw(iter) == READ && pos >= dio->i_size)
1937                         break;
1938         } while ((count = iov_iter_count(iter)) > 0);
1939         blk_finish_plug(&plug);
1940
1941         if (ret < 0)
1942                 iomap_dio_set_error(dio, ret);
1943
1944         /*
1945          * If all the writes we issued were FUA, we don't need to flush the
1946          * cache on IO completion. Clear the sync flag for this case.
1947          */
1948         if (dio->flags & IOMAP_DIO_WRITE_FUA)
1949                 dio->flags &= ~IOMAP_DIO_NEED_SYNC;
1950
1951         WRITE_ONCE(iocb->ki_cookie, dio->submit.cookie);
1952         WRITE_ONCE(iocb->private, dio->submit.last_queue);
1953
1954         /*
1955          * We are about to drop our additional submission reference, which
1956          * might be the last reference to the dio.  There are three three
1957          * different ways we can progress here:
1958          *
1959          *  (a) If this is the last reference we will always complete and free
1960          *      the dio ourselves.
1961          *  (b) If this is not the last reference, and we serve an asynchronous
1962          *      iocb, we must never touch the dio after the decrement, the
1963          *      I/O completion handler will complete and free it.
1964          *  (c) If this is not the last reference, but we serve a synchronous
1965          *      iocb, the I/O completion handler will wake us up on the drop
1966          *      of the final reference, and we will complete and free it here
1967          *      after we got woken by the I/O completion handler.
1968          */
1969         dio->wait_for_completion = wait_for_completion;
1970         if (!atomic_dec_and_test(&dio->ref)) {
1971                 if (!wait_for_completion)
1972                         return -EIOCBQUEUED;
1973
1974                 for (;;) {
1975                         set_current_state(TASK_UNINTERRUPTIBLE);
1976                         if (!READ_ONCE(dio->submit.waiter))
1977                                 break;
1978
1979                         if (!(iocb->ki_flags & IOCB_HIPRI) ||
1980                             !dio->submit.last_queue ||
1981                             !blk_poll(dio->submit.last_queue,
1982                                          dio->submit.cookie, true))
1983                                 io_schedule();
1984                 }
1985                 __set_current_state(TASK_RUNNING);
1986         }
1987
1988         return iomap_dio_complete(dio);
1989
1990 out_free_dio:
1991         kfree(dio);
1992         return ret;
1993 }
1994 EXPORT_SYMBOL_GPL(iomap_dio_rw);
1995
1996 /* Swapfile activation */
1997
1998 #ifdef CONFIG_SWAP
1999 struct iomap_swapfile_info {
2000         struct iomap iomap;             /* accumulated iomap */
2001         struct swap_info_struct *sis;
2002         uint64_t lowest_ppage;          /* lowest physical addr seen (pages) */
2003         uint64_t highest_ppage;         /* highest physical addr seen (pages) */
2004         unsigned long nr_pages;         /* number of pages collected */
2005         int nr_extents;                 /* extent count */
2006 };
2007
2008 /*
2009  * Collect physical extents for this swap file.  Physical extents reported to
2010  * the swap code must be trimmed to align to a page boundary.  The logical
2011  * offset within the file is irrelevant since the swapfile code maps logical
2012  * page numbers of the swap device to the physical page-aligned extents.
2013  */
2014 static int iomap_swapfile_add_extent(struct iomap_swapfile_info *isi)
2015 {
2016         struct iomap *iomap = &isi->iomap;
2017         unsigned long nr_pages;
2018         uint64_t first_ppage;
2019         uint64_t first_ppage_reported;
2020         uint64_t next_ppage;
2021         int error;
2022
2023         /*
2024          * Round the start up and the end down so that the physical
2025          * extent aligns to a page boundary.
2026          */
2027         first_ppage = ALIGN(iomap->addr, PAGE_SIZE) >> PAGE_SHIFT;
2028         next_ppage = ALIGN_DOWN(iomap->addr + iomap->length, PAGE_SIZE) >>
2029                         PAGE_SHIFT;
2030
2031         /* Skip too-short physical extents. */
2032         if (first_ppage >= next_ppage)
2033                 return 0;
2034         nr_pages = next_ppage - first_ppage;
2035
2036         /*
2037          * Calculate how much swap space we're adding; the first page contains
2038          * the swap header and doesn't count.  The mm still wants that first
2039          * page fed to add_swap_extent, however.
2040          */
2041         first_ppage_reported = first_ppage;
2042         if (iomap->offset == 0)
2043                 first_ppage_reported++;
2044         if (isi->lowest_ppage > first_ppage_reported)
2045                 isi->lowest_ppage = first_ppage_reported;
2046         if (isi->highest_ppage < (next_ppage - 1))
2047                 isi->highest_ppage = next_ppage - 1;
2048
2049         /* Add extent, set up for the next call. */
2050         error = add_swap_extent(isi->sis, isi->nr_pages, nr_pages, first_ppage);
2051         if (error < 0)
2052                 return error;
2053         isi->nr_extents += error;
2054         isi->nr_pages += nr_pages;
2055         return 0;
2056 }
2057
2058 /*
2059  * Accumulate iomaps for this swap file.  We have to accumulate iomaps because
2060  * swap only cares about contiguous page-aligned physical extents and makes no
2061  * distinction between written and unwritten extents.
2062  */
2063 static loff_t iomap_swapfile_activate_actor(struct inode *inode, loff_t pos,
2064                 loff_t count, void *data, struct iomap *iomap)
2065 {
2066         struct iomap_swapfile_info *isi = data;
2067         int error;
2068
2069         switch (iomap->type) {
2070         case IOMAP_MAPPED:
2071         case IOMAP_UNWRITTEN:
2072                 /* Only real or unwritten extents. */
2073                 break;
2074         case IOMAP_INLINE:
2075                 /* No inline data. */
2076                 pr_err("swapon: file is inline\n");
2077                 return -EINVAL;
2078         default:
2079                 pr_err("swapon: file has unallocated extents\n");
2080                 return -EINVAL;
2081         }
2082
2083         /* No uncommitted metadata or shared blocks. */
2084         if (iomap->flags & IOMAP_F_DIRTY) {
2085                 pr_err("swapon: file is not committed\n");
2086                 return -EINVAL;
2087         }
2088         if (iomap->flags & IOMAP_F_SHARED) {
2089                 pr_err("swapon: file has shared extents\n");
2090                 return -EINVAL;
2091         }
2092
2093         /* Only one bdev per swap file. */
2094         if (iomap->bdev != isi->sis->bdev) {
2095                 pr_err("swapon: file is on multiple devices\n");
2096                 return -EINVAL;
2097         }
2098
2099         if (isi->iomap.length == 0) {
2100                 /* No accumulated extent, so just store it. */
2101                 memcpy(&isi->iomap, iomap, sizeof(isi->iomap));
2102         } else if (isi->iomap.addr + isi->iomap.length == iomap->addr) {
2103                 /* Append this to the accumulated extent. */
2104                 isi->iomap.length += iomap->length;
2105         } else {
2106                 /* Otherwise, add the retained iomap and store this one. */
2107                 error = iomap_swapfile_add_extent(isi);
2108                 if (error)
2109                         return error;
2110                 memcpy(&isi->iomap, iomap, sizeof(isi->iomap));
2111         }
2112         return count;
2113 }
2114
2115 /*
2116  * Iterate a swap file's iomaps to construct physical extents that can be
2117  * passed to the swapfile subsystem.
2118  */
2119 int iomap_swapfile_activate(struct swap_info_struct *sis,
2120                 struct file *swap_file, sector_t *pagespan,
2121                 const struct iomap_ops *ops)
2122 {
2123         struct iomap_swapfile_info isi = {
2124                 .sis = sis,
2125                 .lowest_ppage = (sector_t)-1ULL,
2126         };
2127         struct address_space *mapping = swap_file->f_mapping;
2128         struct inode *inode = mapping->host;
2129         loff_t pos = 0;
2130         loff_t len = ALIGN_DOWN(i_size_read(inode), PAGE_SIZE);
2131         loff_t ret;
2132
2133         /*
2134          * Persist all file mapping metadata so that we won't have any
2135          * IOMAP_F_DIRTY iomaps.
2136          */
2137         ret = vfs_fsync(swap_file, 1);
2138         if (ret)
2139                 return ret;
2140
2141         while (len > 0) {
2142                 ret = iomap_apply(inode, pos, len, IOMAP_REPORT,
2143                                 ops, &isi, iomap_swapfile_activate_actor);
2144                 if (ret <= 0)
2145                         return ret;
2146
2147                 pos += ret;
2148                 len -= ret;
2149         }
2150
2151         if (isi.iomap.length) {
2152                 ret = iomap_swapfile_add_extent(&isi);
2153                 if (ret)
2154                         return ret;
2155         }
2156
2157         *pagespan = 1 + isi.highest_ppage - isi.lowest_ppage;
2158         sis->max = isi.nr_pages;
2159         sis->pages = isi.nr_pages - 1;
2160         sis->highest_bit = isi.nr_pages - 1;
2161         return isi.nr_extents;
2162 }
2163 EXPORT_SYMBOL_GPL(iomap_swapfile_activate);
2164 #endif /* CONFIG_SWAP */
2165
2166 static loff_t
2167 iomap_bmap_actor(struct inode *inode, loff_t pos, loff_t length,
2168                 void *data, struct iomap *iomap)
2169 {
2170         sector_t *bno = data, addr;
2171
2172         if (iomap->type == IOMAP_MAPPED) {
2173                 addr = (pos - iomap->offset + iomap->addr) >> inode->i_blkbits;
2174                 if (addr > INT_MAX)
2175                         WARN(1, "would truncate bmap result\n");
2176                 else
2177                         *bno = addr;
2178         }
2179         return 0;
2180 }
2181
2182 /* legacy ->bmap interface.  0 is the error return (!) */
2183 sector_t
2184 iomap_bmap(struct address_space *mapping, sector_t bno,
2185                 const struct iomap_ops *ops)
2186 {
2187         struct inode *inode = mapping->host;
2188         loff_t pos = bno << inode->i_blkbits;
2189         unsigned blocksize = i_blocksize(inode);
2190
2191         if (filemap_write_and_wait(mapping))
2192                 return 0;
2193
2194         bno = 0;
2195         iomap_apply(inode, pos, blocksize, 0, ops, &bno, iomap_bmap_actor);
2196         return bno;
2197 }
2198 EXPORT_SYMBOL_GPL(iomap_bmap);