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
 120         /*
 121          * migrate_page_move_mapping() assumes that pages with private data have
 122          * their count elevated by 1.
 123          */
 124         get_page(page);
 125         set_page_private(page, (unsigned long)iop);
 126         SetPagePrivate(page);
 127         return iop;
 128 }
 129
 130 static void
 131 iomap_page_release(struct page *page)
 132 {
 133         struct iomap_page *iop = to_iomap_page(page);
 134
 135         if (!iop)
 136                 return;
 137         WARN_ON_ONCE(atomic_read(&iop->read_count));
 138         WARN_ON_ONCE(atomic_read(&iop->write_count));
 139         ClearPagePrivate(page);
 140         set_page_private(page, 0);
 141         put_page(page);
 142         kfree(iop);
 143 }
 144
 145 /*
 146  * Calculate the range inside the page that we actually need to read.
 147  */
 148 static void
 149 iomap_adjust_read_range(struct inode *inode, struct iomap_page *iop,
 150                 loff_t *pos, loff_t length, unsigned *offp, unsigned *lenp)
 151 {
 152         loff_t orig_pos = *pos;
 153         loff_t isize = i_size_read(inode);
 154         unsigned block_bits = inode->i_blkbits;
 155         unsigned block_size = (1 << block_bits);
 156         unsigned poff = offset_in_page(*pos);
 157         unsigned plen = min_t(loff_t, PAGE_SIZE - poff, length);
 158         unsigned first = poff >> block_bits;
 159         unsigned last = (poff + plen - 1) >> block_bits;
 160
 161         /*
 162          * If the block size is smaller than the page size we need to check the
 163          * per-block uptodate status and adjust the offset and length if needed
 164          * to avoid reading in already uptodate ranges.
 165          */
 166         if (iop) {
 167                 unsigned int i;
 168
 169                 /* move forward for each leading block marked uptodate */
 170                 for (i = first; i <= last; i++) {
 171                         if (!test_bit(i, iop->uptodate))
 172                                 break;
 173                         *pos += block_size;
 174                         poff += block_size;
 175                         plen -= block_size;
 176                         first++;
 177                 }
 178
 179                 /* truncate len if we find any trailing uptodate block(s) */
 180                 for ( ; i <= last; i++) {
 181                         if (test_bit(i, iop->uptodate)) {
 182                                 plen -= (last - i + 1) * block_size;
 183                                 last = i - 1;
 184                                 break;
 185                         }
 186                 }
 187         }
 188
 189         /*
 190          * If the extent spans the block that contains the i_size we need to
 191          * handle both halves separately so that we properly zero data in the
 192          * page cache for blocks that are entirely outside of i_size.
 193          */
 194         if (orig_pos <= isize && orig_pos + length > isize) {
 195                 unsigned end = offset_in_page(isize - 1) >> block_bits;
 196
 197                 if (first <= end && last > end)
 198                         plen -= (last - end) * block_size;
 199         }
 200
 201         *offp = poff;
 202         *lenp = plen;
 203 }
 204
 205 static void
 206 iomap_set_range_uptodate(struct page *page, unsigned off, unsigned len)
 207 {
 208         struct iomap_page *iop = to_iomap_page(page);
 209         struct inode *inode = page->mapping->host;
 210         unsigned first = off >> inode->i_blkbits;
 211         unsigned last = (off + len - 1) >> inode->i_blkbits;
 212         unsigned int i;
 213         bool uptodate = true;
 214
 215         if (iop) {
 216                 for (i = 0; i < PAGE_SIZE / i_blocksize(inode); i++) {
 217                         if (i >= first && i <= last)
 218                                 set_bit(i, iop->uptodate);
 219                         else if (!test_bit(i, iop->uptodate))
 220                                 uptodate = false;
 221                 }
 222         }
 223
 224         if (uptodate && !PageError(page))
 225                 SetPageUptodate(page);
 226 }
 227
 228 static void
 229 iomap_read_finish(struct iomap_page *iop, struct page *page)
 230 {
 231         if (!iop || atomic_dec_and_test(&iop->read_count))
 232                 unlock_page(page);
 233 }
 234
 235 static void
 236 iomap_read_page_end_io(struct bio_vec *bvec, int error)
 237 {
 238         struct page *page = bvec->bv_page;
 239         struct iomap_page *iop = to_iomap_page(page);
 240
 241         if (unlikely(error)) {
 242                 ClearPageUptodate(page);
 243                 SetPageError(page);
 244         } else {
 245                 iomap_set_range_uptodate(page, bvec->bv_offset, bvec->bv_len);
 246         }
 247
 248         iomap_read_finish(iop, page);
 249 }
 250
 251 static void
 252 iomap_read_inline_data(struct inode *inode, struct page *page,
 253                 struct iomap *iomap)
 254 {
 255         size_t size = i_size_read(inode);
 256         void *addr;
 257
 258         if (PageUptodate(page))
 259                 return;
 260
 261         BUG_ON(page->index);
 262         BUG_ON(size > PAGE_SIZE - offset_in_page(iomap->inline_data));
 263
 264         addr = kmap_atomic(page);
 265         memcpy(addr, iomap->inline_data, size);
 266         memset(addr + size, 0, PAGE_SIZE - size);
 267         kunmap_atomic(addr);
 268         SetPageUptodate(page);
 269 }
 270
 271 static void
 272 iomap_read_end_io(struct bio *bio)
 273 {
 274         int error = blk_status_to_errno(bio->bi_status);
 275         struct bio_vec *bvec;
 276         int i;
 277         struct bvec_iter_all iter_all;
 278
 279         bio_for_each_segment_all(bvec, bio, i, iter_all)
 280                 iomap_read_page_end_io(bvec, error);
 281         bio_put(bio);
 282 }
 283
 284 struct iomap_readpage_ctx {
 285         struct page             *cur_page;
 286         bool                    cur_page_in_bio;
 287         bool                    is_readahead;
 288         struct bio              *bio;
 289         struct list_head        *pages;
 290 };
 291
 292 static loff_t
 293 iomap_readpage_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 294                 struct iomap *iomap)
 295 {
 296         struct iomap_readpage_ctx *ctx = data;
 297         struct page *page = ctx->cur_page;
 298         struct iomap_page *iop = iomap_page_create(inode, page);
 299         bool is_contig = false;
 300         loff_t orig_pos = pos;
 301         unsigned poff, plen;
 302         sector_t sector;
 303
 304         if (iomap->type == IOMAP_INLINE) {
 305                 WARN_ON_ONCE(pos);
 306                 iomap_read_inline_data(inode, page, iomap);
 307                 return PAGE_SIZE;
 308         }
 309
 310         /* zero post-eof blocks as the page may be mapped */
 311         iomap_adjust_read_range(inode, iop, &pos, length, &poff, &plen);
 312         if (plen == 0)
 313                 goto done;
 314
 315         if (iomap->type != IOMAP_MAPPED || pos >= i_size_read(inode)) {
 316                 zero_user(page, poff, plen);
 317                 iomap_set_range_uptodate(page, poff, plen);
 318                 goto done;
 319         }
 320
 321         ctx->cur_page_in_bio = true;
 322
 323         /*
 324          * Try to merge into a previous segment if we can.
 325          */
 326         sector = iomap_sector(iomap, pos);
 327         if (ctx->bio && bio_end_sector(ctx->bio) == sector) {
 328                 if (__bio_try_merge_page(ctx->bio, page, plen, poff, true))
 329                         goto done;
 330                 is_contig = true;
 331         }
 332
 333         /*
 334          * If we start a new segment we need to increase the read count, and we
 335          * need to do so before submitting any previous full bio to make sure
 336          * that we don't prematurely unlock the page.
 337          */
 338         if (iop)
 339                 atomic_inc(&iop->read_count);
 340
 341         if (!ctx->bio || !is_contig || bio_full(ctx->bio)) {
 342                 gfp_t gfp = mapping_gfp_constraint(page->mapping, GFP_KERNEL);
 343                 int nr_vecs = (length + PAGE_SIZE - 1) >> PAGE_SHIFT;
 344
 345                 if (ctx->bio)
 346                         submit_bio(ctx->bio);
 347
 348                 if (ctx->is_readahead) /* same as readahead_gfp_mask */
 349                         gfp |= __GFP_NORETRY | __GFP_NOWARN;
 350                 ctx->bio = bio_alloc(gfp, min(BIO_MAX_PAGES, nr_vecs));
 351                 ctx->bio->bi_opf = REQ_OP_READ;
 352                 if (ctx->is_readahead)
 353                         ctx->bio->bi_opf |= REQ_RAHEAD;
 354                 ctx->bio->bi_iter.bi_sector = sector;
 355                 bio_set_dev(ctx->bio, iomap->bdev);
 356                 ctx->bio->bi_end_io = iomap_read_end_io;
 357         }
 358
 359         bio_add_page(ctx->bio, page, plen, poff);
 360 done:
 361         /*
 362          * Move the caller beyond our range so that it keeps making progress.
 363          * For that we have to include any leading non-uptodate ranges, but
 364          * we can skip trailing ones as they will be handled in the next
 365          * iteration.
 366          */
 367         return pos - orig_pos + plen;
 368 }
 369
 370 int
 371 iomap_readpage(struct page *page, const struct iomap_ops *ops)
 372 {
 373         struct iomap_readpage_ctx ctx = { .cur_page = page };
 374         struct inode *inode = page->mapping->host;
 375         unsigned poff;
 376         loff_t ret;
 377
 378         for (poff = 0; poff < PAGE_SIZE; poff += ret) {
 379                 ret = iomap_apply(inode, page_offset(page) + poff,
 380                                 PAGE_SIZE - poff, 0, ops, &ctx,
 381                                 iomap_readpage_actor);
 382                 if (ret <= 0) {
 383                         WARN_ON_ONCE(ret == 0);
 384                         SetPageError(page);
 385                         break;
 386                 }
 387         }
 388
 389         if (ctx.bio) {
 390                 submit_bio(ctx.bio);
 391                 WARN_ON_ONCE(!ctx.cur_page_in_bio);
 392         } else {
 393                 WARN_ON_ONCE(ctx.cur_page_in_bio);
 394                 unlock_page(page);
 395         }
 396
 397         /*
 398          * Just like mpage_readpages and block_read_full_page we always
 399          * return 0 and just mark the page as PageError on errors.  This
 400          * should be cleaned up all through the stack eventually.
 401          */
 402         return 0;
 403 }
 404 EXPORT_SYMBOL_GPL(iomap_readpage);
 405
 406 static struct page *
 407 iomap_next_page(struct inode *inode, struct list_head *pages, loff_t pos,
 408                 loff_t length, loff_t *done)
 409 {
 410         while (!list_empty(pages)) {
 411                 struct page *page = lru_to_page(pages);
 412
 413                 if (page_offset(page) >= (u64)pos + length)
 414                         break;
 415
 416                 list_del(&page->lru);
 417                 if (!add_to_page_cache_lru(page, inode->i_mapping, page->index,
 418                                 GFP_NOFS))
 419                         return page;
 420
 421                 /*
 422                  * If we already have a page in the page cache at index we are
 423                  * done.  Upper layers don't care if it is uptodate after the
 424                  * readpages call itself as every page gets checked again once
 425                  * actually needed.
 426                  */
 427                 *done += PAGE_SIZE;
 428                 put_page(page);
 429         }
 430
 431         return NULL;
 432 }
 433
 434 static loff_t
 435 iomap_readpages_actor(struct inode *inode, loff_t pos, loff_t length,
 436                 void *data, struct iomap *iomap)
 437 {
 438         struct iomap_readpage_ctx *ctx = data;
 439         loff_t done, ret;
 440
 441         for (done = 0; done < length; done += ret) {
 442                 if (ctx->cur_page && offset_in_page(pos + done) == 0) {
 443                         if (!ctx->cur_page_in_bio)
 444                                 unlock_page(ctx->cur_page);
 445                         put_page(ctx->cur_page);
 446                         ctx->cur_page = NULL;
 447                 }
 448                 if (!ctx->cur_page) {
 449                         ctx->cur_page = iomap_next_page(inode, ctx->pages,
 450                                         pos, length, &done);
 451                         if (!ctx->cur_page)
 452                                 break;
 453                         ctx->cur_page_in_bio = false;
 454                 }
 455                 ret = iomap_readpage_actor(inode, pos + done, length - done,
 456                                 ctx, iomap);
 457         }
 458
 459         return done;
 460 }
 461
 462 int
 463 iomap_readpages(struct address_space *mapping, struct list_head *pages,
 464                 unsigned nr_pages, const struct iomap_ops *ops)
 465 {
 466         struct iomap_readpage_ctx ctx = {
 467                 .pages          = pages,
 468                 .is_readahead   = true,
 469         };
 470         loff_t pos = page_offset(list_entry(pages->prev, struct page, lru));
 471         loff_t last = page_offset(list_entry(pages->next, struct page, lru));
 472         loff_t length = last - pos + PAGE_SIZE, ret = 0;
 473
 474         while (length > 0) {
 475                 ret = iomap_apply(mapping->host, pos, length, 0, ops,
 476                                 &ctx, iomap_readpages_actor);
 477                 if (ret <= 0) {
 478                         WARN_ON_ONCE(ret == 0);
 479                         goto done;
 480                 }
 481                 pos += ret;
 482                 length -= ret;
 483         }
 484         ret = 0;
 485 done:
 486         if (ctx.bio)
 487                 submit_bio(ctx.bio);
 488         if (ctx.cur_page) {
 489                 if (!ctx.cur_page_in_bio)
 490                         unlock_page(ctx.cur_page);
 491                 put_page(ctx.cur_page);
 492         }
 493
 494         /*
 495          * Check that we didn't lose a page due to the arcance calling
 496          * conventions..
 497          */
 498         WARN_ON_ONCE(!ret && !list_empty(ctx.pages));
 499         return ret;
 500 }
 501 EXPORT_SYMBOL_GPL(iomap_readpages);
 502
 503 /*
 504  * iomap_is_partially_uptodate checks whether blocks within a page are
 505  * uptodate or not.
 506  *
 507  * Returns true if all blocks which correspond to a file portion
 508  * we want to read within the page are uptodate.
 509  */
 510 int
 511 iomap_is_partially_uptodate(struct page *page, unsigned long from,
 512                 unsigned long count)
 513 {
 514         struct iomap_page *iop = to_iomap_page(page);
 515         struct inode *inode = page->mapping->host;
 516         unsigned len, first, last;
 517         unsigned i;
 518
 519         /* Limit range to one page */
 520         len = min_t(unsigned, PAGE_SIZE - from, count);
 521
 522         /* First and last blocks in range within page */
 523         first = from >> inode->i_blkbits;
 524         last = (from + len - 1) >> inode->i_blkbits;
 525
 526         if (iop) {
 527                 for (i = first; i <= last; i++)
 528                         if (!test_bit(i, iop->uptodate))
 529                                 return 0;
 530                 return 1;
 531         }
 532
 533         return 0;
 534 }
 535 EXPORT_SYMBOL_GPL(iomap_is_partially_uptodate);
 536
 537 int
 538 iomap_releasepage(struct page *page, gfp_t gfp_mask)
 539 {
 540         /*
 541          * mm accommodates an old ext3 case where clean pages might not have had
 542          * the dirty bit cleared. Thus, it can send actual dirty pages to
 543          * ->releasepage() via shrink_active_list(), skip those here.
 544          */
 545         if (PageDirty(page) || PageWriteback(page))
 546                 return 0;
 547         iomap_page_release(page);
 548         return 1;
 549 }
 550 EXPORT_SYMBOL_GPL(iomap_releasepage);
 551
 552 void
 553 iomap_invalidatepage(struct page *page, unsigned int offset, unsigned int len)
 554 {
 555         /*
 556          * If we are invalidating the entire page, clear the dirty state from it
 557          * and release it to avoid unnecessary buildup of the LRU.
 558          */
 559         if (offset == 0 && len == PAGE_SIZE) {
 560                 WARN_ON_ONCE(PageWriteback(page));
 561                 cancel_dirty_page(page);
 562                 iomap_page_release(page);
 563         }
 564 }
 565 EXPORT_SYMBOL_GPL(iomap_invalidatepage);
 566
 567 #ifdef CONFIG_MIGRATION
 568 int
 569 iomap_migrate_page(struct address_space *mapping, struct page *newpage,
 570                 struct page *page, enum migrate_mode mode)
 571 {
 572         int ret;
 573
 574         ret = migrate_page_move_mapping(mapping, newpage, page, mode, 0);
 575         if (ret != MIGRATEPAGE_SUCCESS)
 576                 return ret;
 577
 578         if (page_has_private(page)) {
 579                 ClearPagePrivate(page);
 580                 get_page(newpage);
 581                 set_page_private(newpage, page_private(page));
 582                 set_page_private(page, 0);
 583                 put_page(page);
 584                 SetPagePrivate(newpage);
 585         }
 586
 587         if (mode != MIGRATE_SYNC_NO_COPY)
 588                 migrate_page_copy(newpage, page);
 589         else
 590                 migrate_page_states(newpage, page);
 591         return MIGRATEPAGE_SUCCESS;
 592 }
 593 EXPORT_SYMBOL_GPL(iomap_migrate_page);
 594 #endif /* CONFIG_MIGRATION */
 595
 596 static void
 597 iomap_write_failed(struct inode *inode, loff_t pos, unsigned len)
 598 {
 599         loff_t i_size = i_size_read(inode);
 600
 601         /*
 602          * Only truncate newly allocated pages beyoned EOF, even if the
 603          * write started inside the existing inode size.
 604          */
 605         if (pos + len > i_size)
 606                 truncate_pagecache_range(inode, max(pos, i_size), pos + len);
 607 }
 608
 609 static int
 610 iomap_read_page_sync(struct inode *inode, loff_t block_start, struct page *page,
 611                 unsigned poff, unsigned plen, unsigned from, unsigned to,
 612                 struct iomap *iomap)
 613 {
 614         struct bio_vec bvec;
 615         struct bio bio;
 616
 617         if (iomap->type != IOMAP_MAPPED || block_start >= i_size_read(inode)) {
 618                 zero_user_segments(page, poff, from, to, poff + plen);
 619                 iomap_set_range_uptodate(page, poff, plen);
 620                 return 0;
 621         }
 622
 623         bio_init(&bio, &bvec, 1);
 624         bio.bi_opf = REQ_OP_READ;
 625         bio.bi_iter.bi_sector = iomap_sector(iomap, block_start);
 626         bio_set_dev(&bio, iomap->bdev);
 627         __bio_add_page(&bio, page, plen, poff);
 628         return submit_bio_wait(&bio);
 629 }
 630
 631 static int
 632 __iomap_write_begin(struct inode *inode, loff_t pos, unsigned len,
 633                 struct page *page, struct iomap *iomap)
 634 {
 635         struct iomap_page *iop = iomap_page_create(inode, page);
 636         loff_t block_size = i_blocksize(inode);
 637         loff_t block_start = pos & ~(block_size - 1);
 638         loff_t block_end = (pos + len + block_size - 1) & ~(block_size - 1);
 639         unsigned from = offset_in_page(pos), to = from + len, poff, plen;
 640         int status = 0;
 641
 642         if (PageUptodate(page))
 643                 return 0;
 644
 645         do {
 646                 iomap_adjust_read_range(inode, iop, &block_start,
 647                                 block_end - block_start, &poff, &plen);
 648                 if (plen == 0)
 649                         break;
 650
 651                 if ((from > poff && from < poff + plen) ||
 652                     (to > poff && to < poff + plen)) {
 653                         status = iomap_read_page_sync(inode, block_start, page,
 654                                         poff, plen, from, to, iomap);
 655                         if (status)
 656                                 break;
 657                 }
 658
 659         } while ((block_start += plen) < block_end);
 660
 661         return status;
 662 }
 663
 664 static int
 665 iomap_write_begin(struct inode *inode, loff_t pos, unsigned len, unsigned flags,
 666                 struct page **pagep, struct iomap *iomap)
 667 {
 668         pgoff_t index = pos >> PAGE_SHIFT;
 669         struct page *page;
 670         int status = 0;
 671
 672         BUG_ON(pos + len > iomap->offset + iomap->length);
 673
 674         if (fatal_signal_pending(current))
 675                 return -EINTR;
 676
 677         page = grab_cache_page_write_begin(inode->i_mapping, index, flags);
 678         if (!page)
 679                 return -ENOMEM;
 680
 681         if (iomap->type == IOMAP_INLINE)
 682                 iomap_read_inline_data(inode, page, iomap);
 683         else if (iomap->flags & IOMAP_F_BUFFER_HEAD)
 684                 status = __block_write_begin_int(page, pos, len, NULL, iomap);
 685         else
 686                 status = __iomap_write_begin(inode, pos, len, page, iomap);
 687         if (unlikely(status)) {
 688                 unlock_page(page);
 689                 put_page(page);
 690                 page = NULL;
 691
 692                 iomap_write_failed(inode, pos, len);
 693         }
 694
 695         *pagep = page;
 696         return status;
 697 }
 698
 699 int
 700 iomap_set_page_dirty(struct page *page)
 701 {
 702         struct address_space *mapping = page_mapping(page);
 703         int newly_dirty;
 704
 705         if (unlikely(!mapping))
 706                 return !TestSetPageDirty(page);
 707
 708         /*
 709          * Lock out page->mem_cgroup migration to keep PageDirty
 710          * synchronized with per-memcg dirty page counters.
 711          */
 712         lock_page_memcg(page);
 713         newly_dirty = !TestSetPageDirty(page);
 714         if (newly_dirty)
 715                 __set_page_dirty(page, mapping, 0);
 716         unlock_page_memcg(page);
 717
 718         if (newly_dirty)
 719                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 720         return newly_dirty;
 721 }
 722 EXPORT_SYMBOL_GPL(iomap_set_page_dirty);
 723
 724 static int
 725 __iomap_write_end(struct inode *inode, loff_t pos, unsigned len,
 726                 unsigned copied, struct page *page, struct iomap *iomap)
 727 {
 728         flush_dcache_page(page);
 729
 730         /*
 731          * The blocks that were entirely written will now be uptodate, so we
 732          * don't have to worry about a readpage reading them and overwriting a
 733          * partial write.  However if we have encountered a short write and only
 734          * partially written into a block, it will not be marked uptodate, so a
 735          * readpage might come in and destroy our partial write.
 736          *
 737          * Do the simplest thing, and just treat any short write to a non
 738          * uptodate page as a zero-length write, and force the caller to redo
 739          * the whole thing.
 740          */
 741         if (unlikely(copied < len && !PageUptodate(page))) {
 742                 copied = 0;
 743         } else {
 744                 iomap_set_range_uptodate(page, offset_in_page(pos), len);
 745                 iomap_set_page_dirty(page);
 746         }
 747         return __generic_write_end(inode, pos, copied, page);
 748 }
 749
 750 static int
 751 iomap_write_end_inline(struct inode *inode, struct page *page,
 752                 struct iomap *iomap, loff_t pos, unsigned copied)
 753 {
 754         void *addr;
 755
 756         WARN_ON_ONCE(!PageUptodate(page));
 757         BUG_ON(pos + copied > PAGE_SIZE - offset_in_page(iomap->inline_data));
 758
 759         addr = kmap_atomic(page);
 760         memcpy(iomap->inline_data + pos, addr + pos, copied);
 761         kunmap_atomic(addr);
 762
 763         mark_inode_dirty(inode);
 764         __generic_write_end(inode, pos, copied, page);
 765         return copied;
 766 }
 767
 768 static int
 769 iomap_write_end(struct inode *inode, loff_t pos, unsigned len,
 770                 unsigned copied, struct page *page, struct iomap *iomap)
 771 {
 772         int ret;
 773
 774         if (iomap->type == IOMAP_INLINE) {
 775                 ret = iomap_write_end_inline(inode, page, iomap, pos, copied);
 776         } else if (iomap->flags & IOMAP_F_BUFFER_HEAD) {
 777                 ret = generic_write_end(NULL, inode->i_mapping, pos, len,
 778                                 copied, page, NULL);
 779         } else {
 780                 ret = __iomap_write_end(inode, pos, len, copied, page, iomap);
 781         }
 782
 783         if (iomap->page_done)
 784                 iomap->page_done(inode, pos, copied, page, iomap);
 785
 786         if (ret < len)
 787                 iomap_write_failed(inode, pos, len);
 788         return ret;
 789 }
 790
 791 static loff_t
 792 iomap_write_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 793                 struct iomap *iomap)
 794 {
 795         struct iov_iter *i = data;
 796         long status = 0;
 797         ssize_t written = 0;
 798         unsigned int flags = AOP_FLAG_NOFS;
 799
 800         do {
 801                 struct page *page;
 802                 unsigned long offset;   /* Offset into pagecache page */
 803                 unsigned long bytes;    /* Bytes to write to page */
 804                 size_t copied;          /* Bytes copied from user */
 805
 806                 offset = offset_in_page(pos);
 807                 bytes = min_t(unsigned long, PAGE_SIZE - offset,
 808                                                 iov_iter_count(i));
 809 again:
 810                 if (bytes > length)
 811                         bytes = length;
 812
 813                 /*
 814                  * Bring in the user page that we will copy from _first_.
 815                  * Otherwise there's a nasty deadlock on copying from the
 816                  * same page as we're writing to, without it being marked
 817                  * up-to-date.
 818                  *
 819                  * Not only is this an optimisation, but it is also required
 820                  * to check that the address is actually valid, when atomic
 821                  * usercopies are used, below.
 822                  */
 823                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
 824                         status = -EFAULT;
 825                         break;
 826                 }
 827
 828                 status = iomap_write_begin(inode, pos, bytes, flags, &page,
 829                                 iomap);
 830                 if (unlikely(status))
 831                         break;
 832
 833                 if (mapping_writably_mapped(inode->i_mapping))
 834                         flush_dcache_page(page);
 835
 836                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
 837
 838                 flush_dcache_page(page);
 839
 840                 status = iomap_write_end(inode, pos, bytes, copied, page,
 841                                 iomap);
 842                 if (unlikely(status < 0))
 843                         break;
 844                 copied = status;
 845
 846                 cond_resched();
 847
 848                 iov_iter_advance(i, copied);
 849                 if (unlikely(copied == 0)) {
 850                         /*
 851                          * If we were unable to copy any data at all, we must
 852                          * fall back to a single segment length write.
 853                          *
 854                          * If we didn't fallback here, we could livelock
 855                          * because not all segments in the iov can be copied at
 856                          * once without a pagefault.
 857                          */
 858                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
 859                                                 iov_iter_single_seg_count(i));
 860                         goto again;
 861                 }
 862                 pos += copied;
 863                 written += copied;
 864                 length -= copied;
 865
 866                 balance_dirty_pages_ratelimited(inode->i_mapping);
 867         } while (iov_iter_count(i) && length);
 868
 869         return written ? written : status;
 870 }
 871
 872 ssize_t
 873 iomap_file_buffered_write(struct kiocb *iocb, struct iov_iter *iter,
 874                 const struct iomap_ops *ops)
 875 {
 876         struct inode *inode = iocb->ki_filp->f_mapping->host;
 877         loff_t pos = iocb->ki_pos, ret = 0, written = 0;
 878
 879         while (iov_iter_count(iter)) {
 880                 ret = iomap_apply(inode, pos, iov_iter_count(iter),
 881                                 IOMAP_WRITE, ops, iter, iomap_write_actor);
 882                 if (ret <= 0)
 883                         break;
 884                 pos += ret;
 885                 written += ret;
 886         }
 887
 888         return written ? written : ret;
 889 }
 890 EXPORT_SYMBOL_GPL(iomap_file_buffered_write);
 891
 892 static struct page *
 893 __iomap_read_page(struct inode *inode, loff_t offset)
 894 {
 895         struct address_space *mapping = inode->i_mapping;
 896         struct page *page;
 897
 898         page = read_mapping_page(mapping, offset >> PAGE_SHIFT, NULL);
 899         if (IS_ERR(page))
 900                 return page;
 901         if (!PageUptodate(page)) {
 902                 put_page(page);
 903                 return ERR_PTR(-EIO);
 904         }
 905         return page;
 906 }
 907
 908 static loff_t
 909 iomap_dirty_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 910                 struct iomap *iomap)
 911 {
 912         long status = 0;
 913         ssize_t written = 0;
 914
 915         do {
 916                 struct page *page, *rpage;
 917                 unsigned long offset;   /* Offset into pagecache page */
 918                 unsigned long bytes;    /* Bytes to write to page */
 919
 920                 offset = offset_in_page(pos);
 921                 bytes = min_t(loff_t, PAGE_SIZE - offset, length);
 922
 923                 rpage = __iomap_read_page(inode, pos);
 924                 if (IS_ERR(rpage))
 925                         return PTR_ERR(rpage);
 926
 927                 status = iomap_write_begin(inode, pos, bytes,
 928                                            AOP_FLAG_NOFS, &page, iomap);
 929                 put_page(rpage);
 930                 if (unlikely(status))
 931                         return status;
 932
 933                 WARN_ON_ONCE(!PageUptodate(page));
 934
 935                 status = iomap_write_end(inode, pos, bytes, bytes, page, iomap);
 936                 if (unlikely(status <= 0)) {
 937                         if (WARN_ON_ONCE(status == 0))
 938                                 return -EIO;
 939                         return status;
 940                 }
 941
 942                 cond_resched();
 943
 944                 pos += status;
 945                 written += status;
 946                 length -= status;
 947
 948                 balance_dirty_pages_ratelimited(inode->i_mapping);
 949         } while (length);
 950
 951         return written;
 952 }
 953
 954 int
 955 iomap_file_dirty(struct inode *inode, loff_t pos, loff_t len,
 956                 const struct iomap_ops *ops)
 957 {
 958         loff_t ret;
 959
 960         while (len) {
 961                 ret = iomap_apply(inode, pos, len, IOMAP_WRITE, ops, NULL,
 962                                 iomap_dirty_actor);
 963                 if (ret <= 0)
 964                         return ret;
 965                 pos += ret;
 966                 len -= ret;
 967         }
 968
 969         return 0;
 970 }
 971 EXPORT_SYMBOL_GPL(iomap_file_dirty);
 972
 973 static int iomap_zero(struct inode *inode, loff_t pos, unsigned offset,
 974                 unsigned bytes, struct iomap *iomap)
 975 {
 976         struct page *page;
 977         int status;
 978
 979         status = iomap_write_begin(inode, pos, bytes, AOP_FLAG_NOFS, &page,
 980                                    iomap);
 981         if (status)
 982                 return status;
 983
 984         zero_user(page, offset, bytes);
 985         mark_page_accessed(page);
 986
 987         return iomap_write_end(inode, pos, bytes, bytes, page, iomap);
 988 }
 989
 990 static int iomap_dax_zero(loff_t pos, unsigned offset, unsigned bytes,
 991                 struct iomap *iomap)
 992 {
 993         return __dax_zero_page_range(iomap->bdev, iomap->dax_dev,
 994                         iomap_sector(iomap, pos & PAGE_MASK), offset, bytes);
 995 }
 996
 997 static loff_t
 998 iomap_zero_range_actor(struct inode *inode, loff_t pos, loff_t count,
 999                 void *data, struct iomap *iomap)
1000 {
1001         bool *did_zero = data;
1002         loff_t written = 0;
1003         int status;
1004
1005         /* already zeroed?  we're done. */
1006         if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1007                 return count;
1008
1009         do {
1010                 unsigned offset, bytes;
1011
1012                 offset = offset_in_page(pos);
1013                 bytes = min_t(loff_t, PAGE_SIZE - offset, count);
1014
1015                 if (IS_DAX(inode))
1016                         status = iomap_dax_zero(pos, offset, bytes, iomap);
1017                 else
1018                         status = iomap_zero(inode, pos, offset, bytes, iomap);
1019                 if (status < 0)
1020                         return status;
1021
1022                 pos += bytes;
1023                 count -= bytes;
1024                 written += bytes;
1025                 if (did_zero)
1026                         *did_zero = true;
1027         } while (count > 0);
1028
1029         return written;
1030 }
1031
1032 int
1033 iomap_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
1034                 const struct iomap_ops *ops)
1035 {
1036         loff_t ret;
1037
1038         while (len > 0) {
1039                 ret = iomap_apply(inode, pos, len, IOMAP_ZERO,
1040                                 ops, did_zero, iomap_zero_range_actor);
1041                 if (ret <= 0)
1042                         return ret;
1043
1044                 pos += ret;
1045                 len -= ret;
1046         }
1047
1048         return 0;
1049 }
1050 EXPORT_SYMBOL_GPL(iomap_zero_range);
1051
1052 int
1053 iomap_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
1054                 const struct iomap_ops *ops)
1055 {
1056         unsigned int blocksize = i_blocksize(inode);
1057         unsigned int off = pos & (blocksize - 1);
1058
1059         /* Block boundary? Nothing to do */
1060         if (!off)
1061                 return 0;
1062         return iomap_zero_range(inode, pos, blocksize - off, did_zero, ops);
1063 }
1064 EXPORT_SYMBOL_GPL(iomap_truncate_page);
1065
1066 static loff_t
1067 iomap_page_mkwrite_actor(struct inode *inode, loff_t pos, loff_t length,
1068                 void *data, struct iomap *iomap)
1069 {
1070         struct page *page = data;
1071         int ret;
1072
1073         if (iomap->flags & IOMAP_F_BUFFER_HEAD) {
1074                 ret = __block_write_begin_int(page, pos, length, NULL, iomap);
1075                 if (ret)
1076                         return ret;
1077                 block_commit_write(page, 0, length);
1078         } else {
1079                 WARN_ON_ONCE(!PageUptodate(page));
1080                 iomap_page_create(inode, page);
1081                 set_page_dirty(page);
1082         }
1083
1084         return length;
1085 }
1086
1087 vm_fault_t iomap_page_mkwrite(struct vm_fault *vmf, const struct iomap_ops *ops)
1088 {
1089         struct page *page = vmf->page;
1090         struct inode *inode = file_inode(vmf->vma->vm_file);
1091         unsigned long length;
1092         loff_t offset, size;
1093         ssize_t ret;
1094
1095         lock_page(page);
1096         size = i_size_read(inode);
1097         if ((page->mapping != inode->i_mapping) ||
1098             (page_offset(page) > size)) {
1099                 /* We overload EFAULT to mean page got truncated */
1100                 ret = -EFAULT;
1101                 goto out_unlock;
1102         }
1103
1104         /* page is wholly or partially inside EOF */
1105         if (((page->index + 1) << PAGE_SHIFT) > size)
1106                 length = offset_in_page(size);
1107         else
1108                 length = PAGE_SIZE;
1109
1110         offset = page_offset(page);
1111         while (length > 0) {
1112                 ret = iomap_apply(inode, offset, length,
1113                                 IOMAP_WRITE | IOMAP_FAULT, ops, page,
1114                                 iomap_page_mkwrite_actor);
1115                 if (unlikely(ret <= 0))
1116                         goto out_unlock;
1117                 offset += ret;
1118                 length -= ret;
1119         }
1120
1121         wait_for_stable_page(page);
1122         return VM_FAULT_LOCKED;
1123 out_unlock:
1124         unlock_page(page);
1125         return block_page_mkwrite_return(ret);
1126 }
1127 EXPORT_SYMBOL_GPL(iomap_page_mkwrite);
1128
1129 struct fiemap_ctx {
1130         struct fiemap_extent_info *fi;
1131         struct iomap prev;
1132 };
1133
1134 static int iomap_to_fiemap(struct fiemap_extent_info *fi,
1135                 struct iomap *iomap, u32 flags)
1136 {
1137         switch (iomap->type) {
1138         case IOMAP_HOLE:
1139                 /* skip holes */
1140                 return 0;
1141         case IOMAP_DELALLOC:
1142                 flags |= FIEMAP_EXTENT_DELALLOC | FIEMAP_EXTENT_UNKNOWN;
1143                 break;
1144         case IOMAP_MAPPED:
1145                 break;
1146         case IOMAP_UNWRITTEN:
1147                 flags |= FIEMAP_EXTENT_UNWRITTEN;
1148                 break;
1149         case IOMAP_INLINE:
1150                 flags |= FIEMAP_EXTENT_DATA_INLINE;
1151                 break;
1152         }
1153
1154         if (iomap->flags & IOMAP_F_MERGED)
1155                 flags |= FIEMAP_EXTENT_MERGED;
1156         if (iomap->flags & IOMAP_F_SHARED)
1157                 flags |= FIEMAP_EXTENT_SHARED;
1158
1159         return fiemap_fill_next_extent(fi, iomap->offset,
1160                         iomap->addr != IOMAP_NULL_ADDR ? iomap->addr : 0,
1161                         iomap->length, flags);
1162 }
1163
1164 static loff_t
1165 iomap_fiemap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
1166                 struct iomap *iomap)
1167 {
1168         struct fiemap_ctx *ctx = data;
1169         loff_t ret = length;
1170
1171         if (iomap->type == IOMAP_HOLE)
1172                 return length;
1173
1174         ret = iomap_to_fiemap(ctx->fi, &ctx->prev, 0);
1175         ctx->prev = *iomap;
1176         switch (ret) {
1177         case 0:         /* success */
1178                 return length;
1179         case 1:         /* extent array full */
1180                 return 0;
1181         default:
1182                 return ret;
1183         }
1184 }
1185
1186 int iomap_fiemap(struct inode *inode, struct fiemap_extent_info *fi,
1187                 loff_t start, loff_t len, const struct iomap_ops *ops)
1188 {
1189         struct fiemap_ctx ctx;
1190         loff_t ret;
1191
1192         memset(&ctx, 0, sizeof(ctx));
1193         ctx.fi = fi;
1194         ctx.prev.type = IOMAP_HOLE;
1195
1196         ret = fiemap_check_flags(fi, FIEMAP_FLAG_SYNC);
1197         if (ret)
1198                 return ret;
1199
1200         if (fi->fi_flags & FIEMAP_FLAG_SYNC) {
1201                 ret = filemap_write_and_wait(inode->i_mapping);
1202                 if (ret)
1203                         return ret;
1204         }
1205
1206         while (len > 0) {
1207                 ret = iomap_apply(inode, start, len, IOMAP_REPORT, ops, &ctx,
1208                                 iomap_fiemap_actor);
1209                 /* inode with no (attribute) mapping will give ENOENT */
1210                 if (ret == -ENOENT)
1211                         break;
1212                 if (ret < 0)
1213                         return ret;
1214                 if (ret == 0)
1215                         break;
1216
1217                 start += ret;
1218                 len -= ret;
1219         }
1220
1221         if (ctx.prev.type != IOMAP_HOLE) {
1222                 ret = iomap_to_fiemap(fi, &ctx.prev, FIEMAP_EXTENT_LAST);
1223                 if (ret < 0)
1224                         return ret;
1225         }
1226
1227         return 0;
1228 }
1229 EXPORT_SYMBOL_GPL(iomap_fiemap);
1230
1231 /*
1232  * Seek for SEEK_DATA / SEEK_HOLE within @page, starting at @lastoff.
1233  * Returns true if found and updates @lastoff to the offset in file.
1234  */
1235 static bool
1236 page_seek_hole_data(struct inode *inode, struct page *page, loff_t *lastoff,
1237                 int whence)
1238 {
1239         const struct address_space_operations *ops = inode->i_mapping->a_ops;
1240         unsigned int bsize = i_blocksize(inode), off;
1241         bool seek_data = whence == SEEK_DATA;
1242         loff_t poff = page_offset(page);
1243
1244         if (WARN_ON_ONCE(*lastoff >= poff + PAGE_SIZE))
1245                 return false;
1246
1247         if (*lastoff < poff) {
1248                 /*
1249                  * Last offset smaller than the start of the page means we found
1250                  * a hole:
1251                  */
1252                 if (whence == SEEK_HOLE)
1253                         return true;
1254                 *lastoff = poff;
1255         }
1256
1257         /*
1258          * Just check the page unless we can and should check block ranges:
1259          */
1260         if (bsize == PAGE_SIZE || !ops->is_partially_uptodate)
1261                 return PageUptodate(page) == seek_data;
1262
1263         lock_page(page);
1264         if (unlikely(page->mapping != inode->i_mapping))
1265                 goto out_unlock_not_found;
1266
1267         for (off = 0; off < PAGE_SIZE; off += bsize) {
1268                 if (offset_in_page(*lastoff) >= off + bsize)
1269                         continue;
1270                 if (ops->is_partially_uptodate(page, off, bsize) == seek_data) {
1271                         unlock_page(page);
1272                         return true;
1273                 }
1274                 *lastoff = poff + off + bsize;
1275         }
1276
1277 out_unlock_not_found:
1278         unlock_page(page);
1279         return false;
1280 }
1281
1282 /*
1283  * Seek for SEEK_DATA / SEEK_HOLE in the page cache.
1284  *
1285  * Within unwritten extents, the page cache determines which parts are holes
1286  * and which are data: uptodate buffer heads count as data; everything else
1287  * counts as a hole.
1288  *
1289  * Returns the resulting offset on successs, and -ENOENT otherwise.
1290  */
1291 static loff_t
1292 page_cache_seek_hole_data(struct inode *inode, loff_t offset, loff_t length,
1293                 int whence)
1294 {
1295         pgoff_t index = offset >> PAGE_SHIFT;
1296         pgoff_t end = DIV_ROUND_UP(offset + length, PAGE_SIZE);
1297         loff_t lastoff = offset;
1298         struct pagevec pvec;
1299
1300         if (length <= 0)
1301                 return -ENOENT;
1302
1303         pagevec_init(&pvec);
1304
1305         do {
1306                 unsigned nr_pages, i;
1307
1308                 nr_pages = pagevec_lookup_range(&pvec, inode->i_mapping, &index,
1309                                                 end - 1);
1310                 if (nr_pages == 0)
1311                         break;
1312
1313                 for (i = 0; i < nr_pages; i++) {
1314                         struct page *page = pvec.pages[i];
1315
1316                         if (page_seek_hole_data(inode, page, &lastoff, whence))
1317                                 goto check_range;
1318                         lastoff = page_offset(page) + PAGE_SIZE;
1319                 }
1320                 pagevec_release(&pvec);
1321         } while (index < end);
1322
1323         /* When no page at lastoff and we are not done, we found a hole. */
1324         if (whence != SEEK_HOLE)
1325                 goto not_found;
1326
1327 check_range:
1328         if (lastoff < offset + length)
1329                 goto out;
1330 not_found:
1331         lastoff = -ENOENT;
1332 out:
1333         pagevec_release(&pvec);
1334         return lastoff;
1335 }
1336
1337
1338 static loff_t
1339 iomap_seek_hole_actor(struct inode *inode, loff_t offset, loff_t length,
1340                       void *data, struct iomap *iomap)
1341 {
1342         switch (iomap->type) {
1343         case IOMAP_UNWRITTEN:
1344                 offset = page_cache_seek_hole_data(inode, offset, length,
1345                                                    SEEK_HOLE);
1346                 if (offset < 0)
1347                         return length;
1348                 /* fall through */
1349         case IOMAP_HOLE:
1350                 *(loff_t *)data = offset;
1351                 return 0;
1352         default:
1353                 return length;
1354         }
1355 }
1356
1357 loff_t
1358 iomap_seek_hole(struct inode *inode, loff_t offset, const struct iomap_ops *ops)
1359 {
1360         loff_t size = i_size_read(inode);
1361         loff_t length = size - offset;
1362         loff_t ret;
1363
1364         /* Nothing to be found before or beyond the end of the file. */
1365         if (offset < 0 || offset >= size)
1366                 return -ENXIO;
1367
1368         while (length > 0) {
1369                 ret = iomap_apply(inode, offset, length, IOMAP_REPORT, ops,
1370                                   &offset, iomap_seek_hole_actor);
1371                 if (ret < 0)
1372                         return ret;
1373                 if (ret == 0)
1374                         break;
1375
1376                 offset += ret;
1377                 length -= ret;
1378         }
1379
1380         return offset;
1381 }
1382 EXPORT_SYMBOL_GPL(iomap_seek_hole);
1383
1384 static loff_t
1385 iomap_seek_data_actor(struct inode *inode, loff_t offset, loff_t length,
1386                       void *data, struct iomap *iomap)
1387 {
1388         switch (iomap->type) {
1389         case IOMAP_HOLE:
1390                 return length;
1391         case IOMAP_UNWRITTEN:
1392                 offset = page_cache_seek_hole_data(inode, offset, length,
1393                                                    SEEK_DATA);
1394                 if (offset < 0)
1395                         return length;
1396                 /*FALLTHRU*/
1397         default:
1398                 *(loff_t *)data = offset;
1399                 return 0;
1400         }
1401 }
1402
1403 loff_t
1404 iomap_seek_data(struct inode *inode, loff_t offset, const struct iomap_ops *ops)
1405 {
1406         loff_t size = i_size_read(inode);
1407         loff_t length = size - offset;
1408         loff_t ret;
1409
1410         /* Nothing to be found before or beyond the end of the file. */
1411         if (offset < 0 || offset >= size)
1412                 return -ENXIO;
1413
1414         while (length > 0) {
1415                 ret = iomap_apply(inode, offset, length, IOMAP_REPORT, ops,
1416                                   &offset, iomap_seek_data_actor);
1417                 if (ret < 0)
1418                         return ret;
1419                 if (ret == 0)
1420                         break;
1421
1422                 offset += ret;
1423                 length -= ret;
1424         }
1425
1426         if (length <= 0)
1427                 return -ENXIO;
1428         return offset;
1429 }
1430 EXPORT_SYMBOL_GPL(iomap_seek_data);
1431
1432 /*
1433  * Private flags for iomap_dio, must not overlap with the public ones in
1434  * iomap.h:
1435  */
1436 #define IOMAP_DIO_WRITE_FUA     (1 << 28)
1437 #define IOMAP_DIO_NEED_SYNC     (1 << 29)
1438 #define IOMAP_DIO_WRITE         (1 << 30)
1439 #define IOMAP_DIO_DIRTY         (1 << 31)
1440
1441 struct iomap_dio {
1442         struct kiocb            *iocb;
1443         iomap_dio_end_io_t      *end_io;
1444         loff_t                  i_size;
1445         loff_t                  size;
1446         atomic_t                ref;
1447         unsigned                flags;
1448         int                     error;
1449         bool                    wait_for_completion;
1450
1451         union {
1452                 /* used during submission and for synchronous completion: */
1453                 struct {
1454                         struct iov_iter         *iter;
1455                         struct task_struct      *waiter;
1456                         struct request_queue    *last_queue;
1457                         blk_qc_t                cookie;
1458                 } submit;
1459
1460                 /* used for aio completion: */
1461                 struct {
1462                         struct work_struct      work;
1463                 } aio;
1464         };
1465 };
1466
1467 int iomap_dio_iopoll(struct kiocb *kiocb, bool spin)
1468 {
1469         struct request_queue *q = READ_ONCE(kiocb->private);
1470
1471         if (!q)
1472                 return 0;
1473         return blk_poll(q, READ_ONCE(kiocb->ki_cookie), spin);
1474 }
1475 EXPORT_SYMBOL_GPL(iomap_dio_iopoll);
1476
1477 static void iomap_dio_submit_bio(struct iomap_dio *dio, struct iomap *iomap,
1478                 struct bio *bio)
1479 {
1480         atomic_inc(&dio->ref);
1481
1482         if (dio->iocb->ki_flags & IOCB_HIPRI)
1483                 bio_set_polled(bio, dio->iocb);
1484
1485         dio->submit.last_queue = bdev_get_queue(iomap->bdev);
1486         dio->submit.cookie = submit_bio(bio);
1487 }
1488
1489 static ssize_t iomap_dio_complete(struct iomap_dio *dio)
1490 {
1491         struct kiocb *iocb = dio->iocb;
1492         struct inode *inode = file_inode(iocb->ki_filp);
1493         loff_t offset = iocb->ki_pos;
1494         ssize_t ret;
1495
1496         if (dio->end_io) {
1497                 ret = dio->end_io(iocb,
1498                                 dio->error ? dio->error : dio->size,
1499                                 dio->flags);
1500         } else {
1501                 ret = dio->error;
1502         }
1503
1504         if (likely(!ret)) {
1505                 ret = dio->size;
1506                 /* check for short read */
1507                 if (offset + ret > dio->i_size &&
1508                     !(dio->flags & IOMAP_DIO_WRITE))
1509                         ret = dio->i_size - offset;
1510                 iocb->ki_pos += ret;
1511         }
1512
1513         /*
1514          * Try again to invalidate clean pages which might have been cached by
1515          * non-direct readahead, or faulted in by get_user_pages() if the source
1516          * of the write was an mmap'ed region of the file we're writing.  Either
1517          * one is a pretty crazy thing to do, so we don't support it 100%.  If
1518          * this invalidation fails, tough, the write still worked...
1519          *
1520          * And this page cache invalidation has to be after dio->end_io(), as
1521          * some filesystems convert unwritten extents to real allocations in
1522          * end_io() when necessary, otherwise a racing buffer read would cache
1523          * zeros from unwritten extents.
1524          */
1525         if (!dio->error &&
1526             (dio->flags & IOMAP_DIO_WRITE) && inode->i_mapping->nrpages) {
1527                 int err;
1528                 err = invalidate_inode_pages2_range(inode->i_mapping,
1529                                 offset >> PAGE_SHIFT,
1530                                 (offset + dio->size - 1) >> PAGE_SHIFT);
1531                 if (err)
1532                         dio_warn_stale_pagecache(iocb->ki_filp);
1533         }
1534
1535         /*
1536          * If this is a DSYNC write, make sure we push it to stable storage now
1537          * that we've written data.
1538          */
1539         if (ret > 0 && (dio->flags & IOMAP_DIO_NEED_SYNC))
1540                 ret = generic_write_sync(iocb, ret);
1541
1542         inode_dio_end(file_inode(iocb->ki_filp));
1543         kfree(dio);
1544
1545         return ret;
1546 }
1547
1548 static void iomap_dio_complete_work(struct work_struct *work)
1549 {
1550         struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work);
1551         struct kiocb *iocb = dio->iocb;
1552
1553         iocb->ki_complete(iocb, iomap_dio_complete(dio), 0);
1554 }
1555
1556 /*
1557  * Set an error in the dio if none is set yet.  We have to use cmpxchg
1558  * as the submission context and the completion context(s) can race to
1559  * update the error.
1560  */
1561 static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret)
1562 {
1563         cmpxchg(&dio->error, 0, ret);
1564 }
1565
1566 static void iomap_dio_bio_end_io(struct bio *bio)
1567 {
1568         struct iomap_dio *dio = bio->bi_private;
1569         bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
1570
1571         if (bio->bi_status)
1572                 iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
1573
1574         if (atomic_dec_and_test(&dio->ref)) {
1575                 if (dio->wait_for_completion) {
1576                         struct task_struct *waiter = dio->submit.waiter;
1577                         WRITE_ONCE(dio->submit.waiter, NULL);
1578                         blk_wake_io_task(waiter);
1579                 } else if (dio->flags & IOMAP_DIO_WRITE) {
1580                         struct inode *inode = file_inode(dio->iocb->ki_filp);
1581
1582                         INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
1583                         queue_work(inode->i_sb->s_dio_done_wq, &dio->aio.work);
1584                 } else {
1585                         iomap_dio_complete_work(&dio->aio.work);
1586                 }
1587         }
1588
1589         if (should_dirty) {
1590                 bio_check_pages_dirty(bio);
1591         } else {
1592                 if (!bio_flagged(bio, BIO_NO_PAGE_REF)) {
1593                         struct bvec_iter_all iter_all;
1594                         struct bio_vec *bvec;
1595                         int i;
1596
1597                         bio_for_each_segment_all(bvec, bio, i, iter_all)
1598                                 put_page(bvec->bv_page);
1599                 }
1600                 bio_put(bio);
1601         }
1602 }
1603
1604 static void
1605 iomap_dio_zero(struct iomap_dio *dio, struct iomap *iomap, loff_t pos,
1606                 unsigned len)
1607 {
1608         struct page *page = ZERO_PAGE(0);
1609         int flags = REQ_SYNC | REQ_IDLE;
1610         struct bio *bio;
1611
1612         bio = bio_alloc(GFP_KERNEL, 1);
1613         bio_set_dev(bio, iomap->bdev);
1614         bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
1615         bio->bi_private = dio;
1616         bio->bi_end_io = iomap_dio_bio_end_io;
1617
1618         get_page(page);
1619         __bio_add_page(bio, page, len, 0);
1620         bio_set_op_attrs(bio, REQ_OP_WRITE, flags);
1621         iomap_dio_submit_bio(dio, iomap, bio);
1622 }
1623
1624 static loff_t
1625 iomap_dio_bio_actor(struct inode *inode, loff_t pos, loff_t length,
1626                 struct iomap_dio *dio, struct iomap *iomap)
1627 {
1628         unsigned int blkbits = blksize_bits(bdev_logical_block_size(iomap->bdev));
1629         unsigned int fs_block_size = i_blocksize(inode), pad;
1630         unsigned int align = iov_iter_alignment(dio->submit.iter);
1631         struct iov_iter iter;
1632         struct bio *bio;
1633         bool need_zeroout = false;
1634         bool use_fua = false;
1635         int nr_pages, ret = 0;
1636         size_t copied = 0;
1637
1638         if ((pos | length | align) & ((1 << blkbits) - 1))
1639                 return -EINVAL;
1640
1641         if (iomap->type == IOMAP_UNWRITTEN) {
1642                 dio->flags |= IOMAP_DIO_UNWRITTEN;
1643                 need_zeroout = true;
1644         }
1645
1646         if (iomap->flags & IOMAP_F_SHARED)
1647                 dio->flags |= IOMAP_DIO_COW;
1648
1649         if (iomap->flags & IOMAP_F_NEW) {
1650                 need_zeroout = true;
1651         } else if (iomap->type == IOMAP_MAPPED) {
1652                 /*
1653                  * Use a FUA write if we need datasync semantics, this is a pure
1654                  * data IO that doesn't require any metadata updates (including
1655                  * after IO completion such as unwritten extent conversion) and
1656                  * the underlying device supports FUA. This allows us to avoid
1657                  * cache flushes on IO completion.
1658                  */
1659                 if (!(iomap->flags & (IOMAP_F_SHARED|IOMAP_F_DIRTY)) &&
1660                     (dio->flags & IOMAP_DIO_WRITE_FUA) &&
1661                     blk_queue_fua(bdev_get_queue(iomap->bdev)))
1662                         use_fua = true;
1663         }
1664
1665         /*
1666          * Operate on a partial iter trimmed to the extent we were called for.
1667          * We'll update the iter in the dio once we're done with this extent.
1668          */
1669         iter = *dio->submit.iter;
1670         iov_iter_truncate(&iter, length);
1671
1672         nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
1673         if (nr_pages <= 0)
1674                 return nr_pages;
1675
1676         if (need_zeroout) {
1677                 /* zero out from the start of the block to the write offset */
1678                 pad = pos & (fs_block_size - 1);
1679                 if (pad)
1680                         iomap_dio_zero(dio, iomap, pos - pad, pad);
1681         }
1682
1683         do {
1684                 size_t n;
1685                 if (dio->error) {
1686                         iov_iter_revert(dio->submit.iter, copied);
1687                         return 0;
1688                 }
1689
1690                 bio = bio_alloc(GFP_KERNEL, nr_pages);
1691                 bio_set_dev(bio, iomap->bdev);
1692                 bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
1693                 bio->bi_write_hint = dio->iocb->ki_hint;
1694                 bio->bi_ioprio = dio->iocb->ki_ioprio;
1695                 bio->bi_private = dio;
1696                 bio->bi_end_io = iomap_dio_bio_end_io;
1697
1698                 ret = bio_iov_iter_get_pages(bio, &iter);
1699                 if (unlikely(ret)) {
1700                         /*
1701                          * We have to stop part way through an IO. We must fall
1702                          * through to the sub-block tail zeroing here, otherwise
1703                          * this short IO may expose stale data in the tail of
1704                          * the block we haven't written data to.
1705                          */
1706                         bio_put(bio);
1707                         goto zero_tail;
1708                 }
1709
1710                 n = bio->bi_iter.bi_size;
1711                 if (dio->flags & IOMAP_DIO_WRITE) {
1712                         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
1713                         if (use_fua)
1714                                 bio->bi_opf |= REQ_FUA;
1715                         else
1716                                 dio->flags &= ~IOMAP_DIO_WRITE_FUA;
1717                         task_io_account_write(n);
1718                 } else {
1719                         bio->bi_opf = REQ_OP_READ;
1720                         if (dio->flags & IOMAP_DIO_DIRTY)
1721                                 bio_set_pages_dirty(bio);
1722                 }
1723
1724                 iov_iter_advance(dio->submit.iter, n);
1725
1726                 dio->size += n;
1727                 pos += n;
1728                 copied += n;
1729
1730                 nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
1731                 iomap_dio_submit_bio(dio, iomap, bio);
1732         } while (nr_pages);
1733
1734         /*
1735          * We need to zeroout the tail of a sub-block write if the extent type
1736          * requires zeroing or the write extends beyond EOF. If we don't zero
1737          * the block tail in the latter case, we can expose stale data via mmap
1738          * reads of the EOF block.
1739          */
1740 zero_tail:
1741         if (need_zeroout ||
1742             ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode))) {
1743                 /* zero out from the end of the write to the end of the block */
1744                 pad = pos & (fs_block_size - 1);
1745                 if (pad)
1746                         iomap_dio_zero(dio, iomap, pos, fs_block_size - pad);
1747         }
1748         return copied ? copied : ret;
1749 }
1750
1751 static loff_t
1752 iomap_dio_hole_actor(loff_t length, struct iomap_dio *dio)
1753 {
1754         length = iov_iter_zero(length, dio->submit.iter);
1755         dio->size += length;
1756         return length;
1757 }
1758
1759 static loff_t
1760 iomap_dio_inline_actor(struct inode *inode, loff_t pos, loff_t length,
1761                 struct iomap_dio *dio, struct iomap *iomap)
1762 {
1763         struct iov_iter *iter = dio->submit.iter;
1764         size_t copied;
1765
1766         BUG_ON(pos + length > PAGE_SIZE - offset_in_page(iomap->inline_data));
1767
1768         if (dio->flags & IOMAP_DIO_WRITE) {
1769                 loff_t size = inode->i_size;
1770
1771                 if (pos > size)
1772                         memset(iomap->inline_data + size, 0, pos - size);
1773                 copied = copy_from_iter(iomap->inline_data + pos, length, iter);
1774                 if (copied) {
1775                         if (pos + copied > size)
1776                                 i_size_write(inode, pos + copied);
1777                         mark_inode_dirty(inode);
1778                 }
1779         } else {
1780                 copied = copy_to_iter(iomap->inline_data + pos, length, iter);
1781         }
1782         dio->size += copied;
1783         return copied;
1784 }
1785
1786 static loff_t
1787 iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
1788                 void *data, struct iomap *iomap)
1789 {
1790         struct iomap_dio *dio = data;
1791
1792         switch (iomap->type) {
1793         case IOMAP_HOLE:
1794                 if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
1795                         return -EIO;
1796                 return iomap_dio_hole_actor(length, dio);
1797         case IOMAP_UNWRITTEN:
1798                 if (!(dio->flags & IOMAP_DIO_WRITE))
1799                         return iomap_dio_hole_actor(length, dio);
1800                 return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
1801         case IOMAP_MAPPED:
1802                 return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
1803         case IOMAP_INLINE:
1804                 return iomap_dio_inline_actor(inode, pos, length, dio, iomap);
1805         default:
1806                 WARN_ON_ONCE(1);
1807                 return -EIO;
1808         }
1809 }
1810
1811 /*
1812  * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
1813  * is being issued as AIO or not.  This allows us to optimise pure data writes
1814  * to use REQ_FUA rather than requiring generic_write_sync() to issue a
1815  * REQ_FLUSH post write. This is slightly tricky because a single request here
1816  * can be mapped into multiple disjoint IOs and only a subset of the IOs issued
1817  * may be pure data writes. In that case, we still need to do a full data sync
1818  * completion.
1819  */
1820 ssize_t
1821 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
1822                 const struct iomap_ops *ops, iomap_dio_end_io_t end_io)
1823 {
1824         struct address_space *mapping = iocb->ki_filp->f_mapping;
1825         struct inode *inode = file_inode(iocb->ki_filp);
1826         size_t count = iov_iter_count(iter);
1827         loff_t pos = iocb->ki_pos, start = pos;
1828         loff_t end = iocb->ki_pos + count - 1, ret = 0;
1829         unsigned int flags = IOMAP_DIRECT;
1830         bool wait_for_completion = is_sync_kiocb(iocb);
1831         struct blk_plug plug;
1832         struct iomap_dio *dio;
1833
1834         lockdep_assert_held(&inode->i_rwsem);
1835
1836         if (!count)
1837                 return 0;
1838
1839         dio = kmalloc(sizeof(*dio), GFP_KERNEL);
1840         if (!dio)
1841                 return -ENOMEM;
1842
1843         dio->iocb = iocb;
1844         atomic_set(&dio->ref, 1);
1845         dio->size = 0;
1846         dio->i_size = i_size_read(inode);
1847         dio->end_io = end_io;
1848         dio->error = 0;
1849         dio->flags = 0;
1850
1851         dio->submit.iter = iter;
1852         dio->submit.waiter = current;
1853         dio->submit.cookie = BLK_QC_T_NONE;
1854         dio->submit.last_queue = NULL;
1855
1856         if (iov_iter_rw(iter) == READ) {
1857                 if (pos >= dio->i_size)
1858                         goto out_free_dio;
1859
1860                 if (iter_is_iovec(iter) && iov_iter_rw(iter) == READ)
1861                         dio->flags |= IOMAP_DIO_DIRTY;
1862         } else {
1863                 flags |= IOMAP_WRITE;
1864                 dio->flags |= IOMAP_DIO_WRITE;
1865
1866                 /* for data sync or sync, we need sync completion processing */
1867                 if (iocb->ki_flags & IOCB_DSYNC)
1868                         dio->flags |= IOMAP_DIO_NEED_SYNC;
1869
1870                 /*
1871                  * For datasync only writes, we optimistically try using FUA for
1872                  * this IO.  Any non-FUA write that occurs will clear this flag,
1873                  * hence we know before completion whether a cache flush is
1874                  * necessary.
1875                  */
1876                 if ((iocb->ki_flags & (IOCB_DSYNC | IOCB_SYNC)) == IOCB_DSYNC)
1877                         dio->flags |= IOMAP_DIO_WRITE_FUA;
1878         }
1879
1880         if (iocb->ki_flags & IOCB_NOWAIT) {
1881                 if (filemap_range_has_page(mapping, start, end)) {
1882                         ret = -EAGAIN;
1883                         goto out_free_dio;
1884                 }
1885                 flags |= IOMAP_NOWAIT;
1886         }
1887
1888         ret = filemap_write_and_wait_range(mapping, start, end);
1889         if (ret)
1890                 goto out_free_dio;
1891
1892         /*
1893          * Try to invalidate cache pages for the range we're direct
1894          * writing.  If this invalidation fails, tough, the write will
1895          * still work, but racing two incompatible write paths is a
1896          * pretty crazy thing to do, so we don't support it 100%.
1897          */
1898         ret = invalidate_inode_pages2_range(mapping,
1899                         start >> PAGE_SHIFT, end >> PAGE_SHIFT);
1900         if (ret)
1901                 dio_warn_stale_pagecache(iocb->ki_filp);
1902         ret = 0;
1903
1904         if (iov_iter_rw(iter) == WRITE && !wait_for_completion &&
1905             !inode->i_sb->s_dio_done_wq) {
1906                 ret = sb_init_dio_done_wq(inode->i_sb);
1907                 if (ret < 0)
1908                         goto out_free_dio;
1909         }
1910
1911         inode_dio_begin(inode);
1912
1913         blk_start_plug(&plug);
1914         do {
1915                 ret = iomap_apply(inode, pos, count, flags, ops, dio,
1916                                 iomap_dio_actor);
1917                 if (ret <= 0) {
1918                         /* magic error code to fall back to buffered I/O */
1919                         if (ret == -ENOTBLK) {
1920                                 wait_for_completion = true;
1921                                 ret = 0;
1922                         }
1923                         break;
1924                 }
1925                 pos += ret;
1926
1927                 if (iov_iter_rw(iter) == READ && pos >= dio->i_size)
1928                         break;
1929         } while ((count = iov_iter_count(iter)) > 0);
1930         blk_finish_plug(&plug);
1931
1932         if (ret < 0)
1933                 iomap_dio_set_error(dio, ret);
1934
1935         /*
1936          * If all the writes we issued were FUA, we don't need to flush the
1937          * cache on IO completion. Clear the sync flag for this case.
1938          */
1939         if (dio->flags & IOMAP_DIO_WRITE_FUA)
1940                 dio->flags &= ~IOMAP_DIO_NEED_SYNC;
1941
1942         WRITE_ONCE(iocb->ki_cookie, dio->submit.cookie);
1943         WRITE_ONCE(iocb->private, dio->submit.last_queue);
1944
1945         /*
1946          * We are about to drop our additional submission reference, which
1947          * might be the last reference to the dio.  There are three three
1948          * different ways we can progress here:
1949          *
1950          *  (a) If this is the last reference we will always complete and free
1951          *      the dio ourselves.
1952          *  (b) If this is not the last reference, and we serve an asynchronous
1953          *      iocb, we must never touch the dio after the decrement, the
1954          *      I/O completion handler will complete and free it.
1955          *  (c) If this is not the last reference, but we serve a synchronous
1956          *      iocb, the I/O completion handler will wake us up on the drop
1957          *      of the final reference, and we will complete and free it here
1958          *      after we got woken by the I/O completion handler.
1959          */
1960         dio->wait_for_completion = wait_for_completion;
1961         if (!atomic_dec_and_test(&dio->ref)) {
1962                 if (!wait_for_completion)
1963                         return -EIOCBQUEUED;
1964
1965                 for (;;) {
1966                         set_current_state(TASK_UNINTERRUPTIBLE);
1967                         if (!READ_ONCE(dio->submit.waiter))
1968                                 break;
1969
1970                         if (!(iocb->ki_flags & IOCB_HIPRI) ||
1971                             !dio->submit.last_queue ||
1972                             !blk_poll(dio->submit.last_queue,
1973                                          dio->submit.cookie, true))
1974                                 io_schedule();
1975                 }
1976                 __set_current_state(TASK_RUNNING);
1977         }
1978
1979         return iomap_dio_complete(dio);
1980
1981 out_free_dio:
1982         kfree(dio);
1983         return ret;
1984 }
1985 EXPORT_SYMBOL_GPL(iomap_dio_rw);
1986
1987 /* Swapfile activation */
1988
1989 #ifdef CONFIG_SWAP
1990 struct iomap_swapfile_info {
1991         struct iomap iomap;             /* accumulated iomap */
1992         struct swap_info_struct *sis;
1993         uint64_t lowest_ppage;          /* lowest physical addr seen (pages) */
1994         uint64_t highest_ppage;         /* highest physical addr seen (pages) */
1995         unsigned long nr_pages;         /* number of pages collected */
1996         int nr_extents;                 /* extent count */
1997 };
1998
1999 /*
2000  * Collect physical extents for this swap file.  Physical extents reported to
2001  * the swap code must be trimmed to align to a page boundary.  The logical
2002  * offset within the file is irrelevant since the swapfile code maps logical
2003  * page numbers of the swap device to the physical page-aligned extents.
2004  */
2005 static int iomap_swapfile_add_extent(struct iomap_swapfile_info *isi)
2006 {
2007         struct iomap *iomap = &isi->iomap;
2008         unsigned long nr_pages;
2009         uint64_t first_ppage;
2010         uint64_t first_ppage_reported;
2011         uint64_t next_ppage;
2012         int error;
2013
2014         /*
2015          * Round the start up and the end down so that the physical
2016          * extent aligns to a page boundary.
2017          */
2018         first_ppage = ALIGN(iomap->addr, PAGE_SIZE) >> PAGE_SHIFT;
2019         next_ppage = ALIGN_DOWN(iomap->addr + iomap->length, PAGE_SIZE) >>
2020                         PAGE_SHIFT;
2021
2022         /* Skip too-short physical extents. */
2023         if (first_ppage >= next_ppage)
2024                 return 0;
2025         nr_pages = next_ppage - first_ppage;
2026
2027         /*
2028          * Calculate how much swap space we're adding; the first page contains
2029          * the swap header and doesn't count.  The mm still wants that first
2030          * page fed to add_swap_extent, however.
2031          */
2032         first_ppage_reported = first_ppage;
2033         if (iomap->offset == 0)
2034                 first_ppage_reported++;
2035         if (isi->lowest_ppage > first_ppage_reported)
2036                 isi->lowest_ppage = first_ppage_reported;
2037         if (isi->highest_ppage < (next_ppage - 1))
2038                 isi->highest_ppage = next_ppage - 1;
2039
2040         /* Add extent, set up for the next call. */
2041         error = add_swap_extent(isi->sis, isi->nr_pages, nr_pages, first_ppage);
2042         if (error < 0)
2043                 return error;
2044         isi->nr_extents += error;
2045         isi->nr_pages += nr_pages;
2046         return 0;
2047 }
2048
2049 /*
2050  * Accumulate iomaps for this swap file.  We have to accumulate iomaps because
2051  * swap only cares about contiguous page-aligned physical extents and makes no
2052  * distinction between written and unwritten extents.
2053  */
2054 static loff_t iomap_swapfile_activate_actor(struct inode *inode, loff_t pos,
2055                 loff_t count, void *data, struct iomap *iomap)
2056 {
2057         struct iomap_swapfile_info *isi = data;
2058         int error;
2059
2060         switch (iomap->type) {
2061         case IOMAP_MAPPED:
2062         case IOMAP_UNWRITTEN:
2063                 /* Only real or unwritten extents. */
2064                 break;
2065         case IOMAP_INLINE:
2066                 /* No inline data. */
2067                 pr_err("swapon: file is inline\n");
2068                 return -EINVAL;
2069         default:
2070                 pr_err("swapon: file has unallocated extents\n");
2071                 return -EINVAL;
2072         }
2073
2074         /* No uncommitted metadata or shared blocks. */
2075         if (iomap->flags & IOMAP_F_DIRTY) {
2076                 pr_err("swapon: file is not committed\n");
2077                 return -EINVAL;
2078         }
2079         if (iomap->flags & IOMAP_F_SHARED) {
2080                 pr_err("swapon: file has shared extents\n");
2081                 return -EINVAL;
2082         }
2083
2084         /* Only one bdev per swap file. */
2085         if (iomap->bdev != isi->sis->bdev) {
2086                 pr_err("swapon: file is on multiple devices\n");
2087                 return -EINVAL;
2088         }
2089
2090         if (isi->iomap.length == 0) {
2091                 /* No accumulated extent, so just store it. */
2092                 memcpy(&isi->iomap, iomap, sizeof(isi->iomap));
2093         } else if (isi->iomap.addr + isi->iomap.length == iomap->addr) {
2094                 /* Append this to the accumulated extent. */
2095                 isi->iomap.length += iomap->length;
2096         } else {
2097                 /* Otherwise, add the retained iomap and store this one. */
2098                 error = iomap_swapfile_add_extent(isi);
2099                 if (error)
2100                         return error;
2101                 memcpy(&isi->iomap, iomap, sizeof(isi->iomap));
2102         }
2103         return count;
2104 }
2105
2106 /*
2107  * Iterate a swap file's iomaps to construct physical extents that can be
2108  * passed to the swapfile subsystem.
2109  */
2110 int iomap_swapfile_activate(struct swap_info_struct *sis,
2111                 struct file *swap_file, sector_t *pagespan,
2112                 const struct iomap_ops *ops)
2113 {
2114         struct iomap_swapfile_info isi = {
2115                 .sis = sis,
2116                 .lowest_ppage = (sector_t)-1ULL,
2117         };
2118         struct address_space *mapping = swap_file->f_mapping;
2119         struct inode *inode = mapping->host;
2120         loff_t pos = 0;
2121         loff_t len = ALIGN_DOWN(i_size_read(inode), PAGE_SIZE);
2122         loff_t ret;
2123
2124         /*
2125          * Persist all file mapping metadata so that we won't have any
2126          * IOMAP_F_DIRTY iomaps.
2127          */
2128         ret = vfs_fsync(swap_file, 1);
2129         if (ret)
2130                 return ret;
2131
2132         while (len > 0) {
2133                 ret = iomap_apply(inode, pos, len, IOMAP_REPORT,
2134                                 ops, &isi, iomap_swapfile_activate_actor);
2135                 if (ret <= 0)
2136                         return ret;
2137
2138                 pos += ret;
2139                 len -= ret;
2140         }
2141
2142         if (isi.iomap.length) {
2143                 ret = iomap_swapfile_add_extent(&isi);
2144                 if (ret)
2145                         return ret;
2146         }
2147
2148         *pagespan = 1 + isi.highest_ppage - isi.lowest_ppage;
2149         sis->max = isi.nr_pages;
2150         sis->pages = isi.nr_pages - 1;
2151         sis->highest_bit = isi.nr_pages - 1;
2152         return isi.nr_extents;
2153 }
2154 EXPORT_SYMBOL_GPL(iomap_swapfile_activate);
2155 #endif /* CONFIG_SWAP */
2156
2157 static loff_t
2158 iomap_bmap_actor(struct inode *inode, loff_t pos, loff_t length,
2159                 void *data, struct iomap *iomap)
2160 {
2161         sector_t *bno = data, addr;
2162
2163         if (iomap->type == IOMAP_MAPPED) {
2164                 addr = (pos - iomap->offset + iomap->addr) >> inode->i_blkbits;
2165                 if (addr > INT_MAX)
2166                         WARN(1, "would truncate bmap result\n");
2167                 else
2168                         *bno = addr;
2169         }
2170         return 0;
2171 }
2172
2173 /* legacy ->bmap interface.  0 is the error return (!) */
2174 sector_t
2175 iomap_bmap(struct address_space *mapping, sector_t bno,
2176                 const struct iomap_ops *ops)
2177 {
2178         struct inode *inode = mapping->host;
2179         loff_t pos = bno << inode->i_blkbits;
2180         unsigned blocksize = i_blocksize(inode);
2181
2182         if (filemap_write_and_wait(mapping))
2183                 return 0;
2184
2185         bno = 0;
2186         iomap_apply(inode, pos, blocksize, 0, ops, &bno, iomap_bmap_actor);
2187         return bno;
2188 }
2189 EXPORT_SYMBOL_GPL(iomap_bmap);