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