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