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