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 #include <linux/swap.h>
  34
  35 #include "internal.h"
  36
  37 /*
  38  * Execute a iomap write on a segment of the mapping that spans a
  39  * contiguous range of pages that have identical block mapping state.
  40  *
  41  * This avoids the need to map pages individually, do individual allocations
  42  * for each page and most importantly avoid the need for filesystem specific
  43  * locking per page. Instead, all the operations are amortised over the entire
  44  * range of pages. It is assumed that the filesystems will lock whatever
  45  * resources they require in the iomap_begin call, and release them in the
  46  * iomap_end call.
  47  */
  48 loff_t
  49 iomap_apply(struct inode *inode, loff_t pos, loff_t length, unsigned flags,
  50                 const struct iomap_ops *ops, void *data, iomap_actor_t actor)
  51 {
  52         struct iomap iomap = { 0 };
  53         loff_t written = 0, ret;
  54
  55         /*
  56          * Need to map a range from start position for length bytes. This can
  57          * span multiple pages - it is only guaranteed to return a range of a
  58          * single type of pages (e.g. all into a hole, all mapped or all
  59          * unwritten). Failure at this point has nothing to undo.
  60          *
  61          * If allocation is required for this range, reserve the space now so
  62          * that the allocation is guaranteed to succeed later on. Once we copy
  63          * the data into the page cache pages, then we cannot fail otherwise we
  64          * expose transient stale data. If the reserve fails, we can safely
  65          * back out at this point as there is nothing to undo.
  66          */
  67         ret = ops->iomap_begin(inode, pos, length, flags, &iomap);
  68         if (ret)
  69                 return ret;
  70         if (WARN_ON(iomap.offset > pos))
  71                 return -EIO;
  72         if (WARN_ON(iomap.length == 0))
  73                 return -EIO;
  74
  75         /*
  76          * Cut down the length to the one actually provided by the filesystem,
  77          * as it might not be able to give us the whole size that we requested.
  78          */
  79         if (iomap.offset + iomap.length < pos + length)
  80                 length = iomap.offset + iomap.length - pos;
  81
  82         /*
  83          * Now that we have guaranteed that the space allocation will succeed.
  84          * we can do the copy-in page by page without having to worry about
  85          * failures exposing transient data.
  86          */
  87         written = actor(inode, pos, length, data, &iomap);
  88
  89         /*
  90          * Now the data has been copied, commit the range we've copied.  This
  91          * should not fail unless the filesystem has had a fatal error.
  92          */
  93         if (ops->iomap_end) {
  94                 ret = ops->iomap_end(inode, pos, length,
  95                                      written > 0 ? written : 0,
  96                                      flags, &iomap);
  97         }
  98
  99         return written ? written : ret;
 100 }
 101
 102 static sector_t
 103 iomap_sector(struct iomap *iomap, loff_t pos)
 104 {
 105         return (iomap->addr + pos - iomap->offset) >> SECTOR_SHIFT;
 106 }
 107
 108 static struct iomap_page *
 109 iomap_page_create(struct inode *inode, struct page *page)
 110 {
 111         struct iomap_page *iop = to_iomap_page(page);
 112
 113         if (iop || i_blocksize(inode) == PAGE_SIZE)
 114                 return iop;
 115
 116         iop = kmalloc(sizeof(*iop), GFP_NOFS | __GFP_NOFAIL);
 117         atomic_set(&iop->read_count, 0);
 118         atomic_set(&iop->write_count, 0);
 119         bitmap_zero(iop->uptodate, PAGE_SIZE / SECTOR_SIZE);
 120         set_page_private(page, (unsigned long)iop);
 121         SetPagePrivate(page);
 122         return iop;
 123 }
 124
 125 static void
 126 iomap_page_release(struct page *page)
 127 {
 128         struct iomap_page *iop = to_iomap_page(page);
 129
 130         if (!iop)
 131                 return;
 132         WARN_ON_ONCE(atomic_read(&iop->read_count));
 133         WARN_ON_ONCE(atomic_read(&iop->write_count));
 134         ClearPagePrivate(page);
 135         set_page_private(page, 0);
 136         kfree(iop);
 137 }
 138
 139 /*
 140  * Calculate the range inside the page that we actually need to read.
 141  */
 142 static void
 143 iomap_adjust_read_range(struct inode *inode, struct iomap_page *iop,
 144                 loff_t *pos, loff_t length, unsigned *offp, unsigned *lenp)
 145 {
 146         unsigned block_bits = inode->i_blkbits;
 147         unsigned block_size = (1 << block_bits);
 148         unsigned poff = offset_in_page(*pos);
 149         unsigned plen = min_t(loff_t, PAGE_SIZE - poff, length);
 150         unsigned first = poff >> block_bits;
 151         unsigned last = (poff + plen - 1) >> block_bits;
 152         unsigned end = offset_in_page(i_size_read(inode)) >> block_bits;
 153
 154         /*
 155          * If the block size is smaller than the page size we need to check the
 156          * per-block uptodate status and adjust the offset and length if needed
 157          * to avoid reading in already uptodate ranges.
 158          */
 159         if (iop) {
 160                 unsigned int i;
 161
 162                 /* move forward for each leading block marked uptodate */
 163                 for (i = first; i <= last; i++) {
 164                         if (!test_bit(i, iop->uptodate))
 165                                 break;
 166                         *pos += block_size;
 167                         poff += block_size;
 168                         plen -= block_size;
 169                         first++;
 170                 }
 171
 172                 /* truncate len if we find any trailing uptodate block(s) */
 173                 for ( ; i <= last; i++) {
 174                         if (test_bit(i, iop->uptodate)) {
 175                                 plen -= (last - i + 1) * block_size;
 176                                 last = i - 1;
 177                                 break;
 178                         }
 179                 }
 180         }
 181
 182         /*
 183          * If the extent spans the block that contains the i_size we need to
 184          * handle both halves separately so that we properly zero data in the
 185          * page cache for blocks that are entirely outside of i_size.
 186          */
 187         if (first <= end && last > end)
 188                 plen -= (last - end) * block_size;
 189
 190         *offp = poff;
 191         *lenp = plen;
 192 }
 193
 194 static void
 195 iomap_set_range_uptodate(struct page *page, unsigned off, unsigned len)
 196 {
 197         struct iomap_page *iop = to_iomap_page(page);
 198         struct inode *inode = page->mapping->host;
 199         unsigned first = off >> inode->i_blkbits;
 200         unsigned last = (off + len - 1) >> inode->i_blkbits;
 201         unsigned int i;
 202         bool uptodate = true;
 203
 204         if (iop) {
 205                 for (i = 0; i < PAGE_SIZE / i_blocksize(inode); i++) {
 206                         if (i >= first && i <= last)
 207                                 set_bit(i, iop->uptodate);
 208                         else if (!test_bit(i, iop->uptodate))
 209                                 uptodate = false;
 210                 }
 211         }
 212
 213         if (uptodate && !PageError(page))
 214                 SetPageUptodate(page);
 215 }
 216
 217 static void
 218 iomap_read_finish(struct iomap_page *iop, struct page *page)
 219 {
 220         if (!iop || atomic_dec_and_test(&iop->read_count))
 221                 unlock_page(page);
 222 }
 223
 224 static void
 225 iomap_read_page_end_io(struct bio_vec *bvec, int error)
 226 {
 227         struct page *page = bvec->bv_page;
 228         struct iomap_page *iop = to_iomap_page(page);
 229
 230         if (unlikely(error)) {
 231                 ClearPageUptodate(page);
 232                 SetPageError(page);
 233         } else {
 234                 iomap_set_range_uptodate(page, bvec->bv_offset, bvec->bv_len);
 235         }
 236
 237         iomap_read_finish(iop, page);
 238 }
 239
 240 static void
 241 iomap_read_inline_data(struct inode *inode, struct page *page,
 242                 struct iomap *iomap)
 243 {
 244         size_t size = i_size_read(inode);
 245         void *addr;
 246
 247         if (PageUptodate(page))
 248                 return;
 249
 250         BUG_ON(page->index);
 251         BUG_ON(size > PAGE_SIZE - offset_in_page(iomap->inline_data));
 252
 253         addr = kmap_atomic(page);
 254         memcpy(addr, iomap->inline_data, size);
 255         memset(addr + size, 0, PAGE_SIZE - size);
 256         kunmap_atomic(addr);
 257         SetPageUptodate(page);
 258 }
 259
 260 static void
 261 iomap_read_end_io(struct bio *bio)
 262 {
 263         int error = blk_status_to_errno(bio->bi_status);
 264         struct bio_vec *bvec;
 265         int i;
 266
 267         bio_for_each_segment_all(bvec, bio, i)
 268                 iomap_read_page_end_io(bvec, error);
 269         bio_put(bio);
 270 }
 271
 272 struct iomap_readpage_ctx {
 273         struct page             *cur_page;
 274         bool                    cur_page_in_bio;
 275         bool                    is_readahead;
 276         struct bio              *bio;
 277         struct list_head        *pages;
 278 };
 279
 280 static loff_t
 281 iomap_readpage_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 282                 struct iomap *iomap)
 283 {
 284         struct iomap_readpage_ctx *ctx = data;
 285         struct page *page = ctx->cur_page;
 286         struct iomap_page *iop = iomap_page_create(inode, page);
 287         bool is_contig = false;
 288         loff_t orig_pos = pos;
 289         unsigned poff, plen;
 290         sector_t sector;
 291
 292         if (iomap->type == IOMAP_INLINE) {
 293                 WARN_ON_ONCE(pos);
 294                 iomap_read_inline_data(inode, page, iomap);
 295                 return PAGE_SIZE;
 296         }
 297
 298         /* zero post-eof blocks as the page may be mapped */
 299         iomap_adjust_read_range(inode, iop, &pos, length, &poff, &plen);
 300         if (plen == 0)
 301                 goto done;
 302
 303         if (iomap->type != IOMAP_MAPPED || pos >= i_size_read(inode)) {
 304                 zero_user(page, poff, plen);
 305                 iomap_set_range_uptodate(page, poff, plen);
 306                 goto done;
 307         }
 308
 309         ctx->cur_page_in_bio = true;
 310
 311         /*
 312          * Try to merge into a previous segment if we can.
 313          */
 314         sector = iomap_sector(iomap, pos);
 315         if (ctx->bio && bio_end_sector(ctx->bio) == sector) {
 316                 if (__bio_try_merge_page(ctx->bio, page, plen, poff))
 317                         goto done;
 318                 is_contig = true;
 319         }
 320
 321         /*
 322          * If we start a new segment we need to increase the read count, and we
 323          * need to do so before submitting any previous full bio to make sure
 324          * that we don't prematurely unlock the page.
 325          */
 326         if (iop)
 327                 atomic_inc(&iop->read_count);
 328
 329         if (!ctx->bio || !is_contig || bio_full(ctx->bio)) {
 330                 gfp_t gfp = mapping_gfp_constraint(page->mapping, GFP_KERNEL);
 331                 int nr_vecs = (length + PAGE_SIZE - 1) >> PAGE_SHIFT;
 332
 333                 if (ctx->bio)
 334                         submit_bio(ctx->bio);
 335
 336                 if (ctx->is_readahead) /* same as readahead_gfp_mask */
 337                         gfp |= __GFP_NORETRY | __GFP_NOWARN;
 338                 ctx->bio = bio_alloc(gfp, min(BIO_MAX_PAGES, nr_vecs));
 339                 ctx->bio->bi_opf = REQ_OP_READ;
 340                 if (ctx->is_readahead)
 341                         ctx->bio->bi_opf |= REQ_RAHEAD;
 342                 ctx->bio->bi_iter.bi_sector = sector;
 343                 bio_set_dev(ctx->bio, iomap->bdev);
 344                 ctx->bio->bi_end_io = iomap_read_end_io;
 345         }
 346
 347         __bio_add_page(ctx->bio, page, plen, poff);
 348 done:
 349         /*
 350          * Move the caller beyond our range so that it keeps making progress.
 351          * For that we have to include any leading non-uptodate ranges, but
 352          * we can skip trailing ones as they will be handled in the next
 353          * iteration.
 354          */
 355         return pos - orig_pos + plen;
 356 }
 357
 358 int
 359 iomap_readpage(struct page *page, const struct iomap_ops *ops)
 360 {
 361         struct iomap_readpage_ctx ctx = { .cur_page = page };
 362         struct inode *inode = page->mapping->host;
 363         unsigned poff;
 364         loff_t ret;
 365
 366         for (poff = 0; poff < PAGE_SIZE; poff += ret) {
 367                 ret = iomap_apply(inode, page_offset(page) + poff,
 368                                 PAGE_SIZE - poff, 0, ops, &ctx,
 369                                 iomap_readpage_actor);
 370                 if (ret <= 0) {
 371                         WARN_ON_ONCE(ret == 0);
 372                         SetPageError(page);
 373                         break;
 374                 }
 375         }
 376
 377         if (ctx.bio) {
 378                 submit_bio(ctx.bio);
 379                 WARN_ON_ONCE(!ctx.cur_page_in_bio);
 380         } else {
 381                 WARN_ON_ONCE(ctx.cur_page_in_bio);
 382                 unlock_page(page);
 383         }
 384
 385         /*
 386          * Just like mpage_readpages and block_read_full_page we always
 387          * return 0 and just mark the page as PageError on errors.  This
 388          * should be cleaned up all through the stack eventually.
 389          */
 390         return 0;
 391 }
 392 EXPORT_SYMBOL_GPL(iomap_readpage);
 393
 394 static struct page *
 395 iomap_next_page(struct inode *inode, struct list_head *pages, loff_t pos,
 396                 loff_t length, loff_t *done)
 397 {
 398         while (!list_empty(pages)) {
 399                 struct page *page = lru_to_page(pages);
 400
 401                 if (page_offset(page) >= (u64)pos + length)
 402                         break;
 403
 404                 list_del(&page->lru);
 405                 if (!add_to_page_cache_lru(page, inode->i_mapping, page->index,
 406                                 GFP_NOFS))
 407                         return page;
 408
 409                 /*
 410                  * If we already have a page in the page cache at index we are
 411                  * done.  Upper layers don't care if it is uptodate after the
 412                  * readpages call itself as every page gets checked again once
 413                  * actually needed.
 414                  */
 415                 *done += PAGE_SIZE;
 416                 put_page(page);
 417         }
 418
 419         return NULL;
 420 }
 421
 422 static loff_t
 423 iomap_readpages_actor(struct inode *inode, loff_t pos, loff_t length,
 424                 void *data, struct iomap *iomap)
 425 {
 426         struct iomap_readpage_ctx *ctx = data;
 427         loff_t done, ret;
 428
 429         for (done = 0; done < length; done += ret) {
 430                 if (ctx->cur_page && offset_in_page(pos + done) == 0) {
 431                         if (!ctx->cur_page_in_bio)
 432                                 unlock_page(ctx->cur_page);
 433                         put_page(ctx->cur_page);
 434                         ctx->cur_page = NULL;
 435                 }
 436                 if (!ctx->cur_page) {
 437                         ctx->cur_page = iomap_next_page(inode, ctx->pages,
 438                                         pos, length, &done);
 439                         if (!ctx->cur_page)
 440                                 break;
 441                         ctx->cur_page_in_bio = false;
 442                 }
 443                 ret = iomap_readpage_actor(inode, pos + done, length - done,
 444                                 ctx, iomap);
 445         }
 446
 447         return done;
 448 }
 449
 450 int
 451 iomap_readpages(struct address_space *mapping, struct list_head *pages,
 452                 unsigned nr_pages, const struct iomap_ops *ops)
 453 {
 454         struct iomap_readpage_ctx ctx = {
 455                 .pages          = pages,
 456                 .is_readahead   = true,
 457         };
 458         loff_t pos = page_offset(list_entry(pages->prev, struct page, lru));
 459         loff_t last = page_offset(list_entry(pages->next, struct page, lru));
 460         loff_t length = last - pos + PAGE_SIZE, ret = 0;
 461
 462         while (length > 0) {
 463                 ret = iomap_apply(mapping->host, pos, length, 0, ops,
 464                                 &ctx, iomap_readpages_actor);
 465                 if (ret <= 0) {
 466                         WARN_ON_ONCE(ret == 0);
 467                         goto done;
 468                 }
 469                 pos += ret;
 470                 length -= ret;
 471         }
 472         ret = 0;
 473 done:
 474         if (ctx.bio)
 475                 submit_bio(ctx.bio);
 476         if (ctx.cur_page) {
 477                 if (!ctx.cur_page_in_bio)
 478                         unlock_page(ctx.cur_page);
 479                 put_page(ctx.cur_page);
 480         }
 481
 482         /*
 483          * Check that we didn't lose a page due to the arcance calling
 484          * conventions..
 485          */
 486         WARN_ON_ONCE(!ret && !list_empty(ctx.pages));
 487         return ret;
 488 }
 489 EXPORT_SYMBOL_GPL(iomap_readpages);
 490
 491 int
 492 iomap_is_partially_uptodate(struct page *page, unsigned long from,
 493                 unsigned long count)
 494 {
 495         struct iomap_page *iop = to_iomap_page(page);
 496         struct inode *inode = page->mapping->host;
 497         unsigned first = from >> inode->i_blkbits;
 498         unsigned last = (from + count - 1) >> inode->i_blkbits;
 499         unsigned i;
 500
 501         if (iop) {
 502                 for (i = first; i <= last; i++)
 503                         if (!test_bit(i, iop->uptodate))
 504                                 return 0;
 505                 return 1;
 506         }
 507
 508         return 0;
 509 }
 510 EXPORT_SYMBOL_GPL(iomap_is_partially_uptodate);
 511
 512 int
 513 iomap_releasepage(struct page *page, gfp_t gfp_mask)
 514 {
 515         /*
 516          * mm accommodates an old ext3 case where clean pages might not have had
 517          * the dirty bit cleared. Thus, it can send actual dirty pages to
 518          * ->releasepage() via shrink_active_list(), skip those here.
 519          */
 520         if (PageDirty(page) || PageWriteback(page))
 521                 return 0;
 522         iomap_page_release(page);
 523         return 1;
 524 }
 525 EXPORT_SYMBOL_GPL(iomap_releasepage);
 526
 527 void
 528 iomap_invalidatepage(struct page *page, unsigned int offset, unsigned int len)
 529 {
 530         /*
 531          * If we are invalidating the entire page, clear the dirty state from it
 532          * and release it to avoid unnecessary buildup of the LRU.
 533          */
 534         if (offset == 0 && len == PAGE_SIZE) {
 535                 WARN_ON_ONCE(PageWriteback(page));
 536                 cancel_dirty_page(page);
 537                 iomap_page_release(page);
 538         }
 539 }
 540 EXPORT_SYMBOL_GPL(iomap_invalidatepage);
 541
 542 #ifdef CONFIG_MIGRATION
 543 int
 544 iomap_migrate_page(struct address_space *mapping, struct page *newpage,
 545                 struct page *page, enum migrate_mode mode)
 546 {
 547         int ret;
 548
 549         ret = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
 550         if (ret != MIGRATEPAGE_SUCCESS)
 551                 return ret;
 552
 553         if (page_has_private(page)) {
 554                 ClearPagePrivate(page);
 555                 set_page_private(newpage, page_private(page));
 556                 set_page_private(page, 0);
 557                 SetPagePrivate(newpage);
 558         }
 559
 560         if (mode != MIGRATE_SYNC_NO_COPY)
 561                 migrate_page_copy(newpage, page);
 562         else
 563                 migrate_page_states(newpage, page);
 564         return MIGRATEPAGE_SUCCESS;
 565 }
 566 EXPORT_SYMBOL_GPL(iomap_migrate_page);
 567 #endif /* CONFIG_MIGRATION */
 568
 569 static void
 570 iomap_write_failed(struct inode *inode, loff_t pos, unsigned len)
 571 {
 572         loff_t i_size = i_size_read(inode);
 573
 574         /*
 575          * Only truncate newly allocated pages beyoned EOF, even if the
 576          * write started inside the existing inode size.
 577          */
 578         if (pos + len > i_size)
 579                 truncate_pagecache_range(inode, max(pos, i_size), pos + len);
 580 }
 581
 582 static int
 583 iomap_read_page_sync(struct inode *inode, loff_t block_start, struct page *page,
 584                 unsigned poff, unsigned plen, unsigned from, unsigned to,
 585                 struct iomap *iomap)
 586 {
 587         struct bio_vec bvec;
 588         struct bio bio;
 589
 590         if (iomap->type != IOMAP_MAPPED || block_start >= i_size_read(inode)) {
 591                 zero_user_segments(page, poff, from, to, poff + plen);
 592                 iomap_set_range_uptodate(page, poff, plen);
 593                 return 0;
 594         }
 595
 596         bio_init(&bio, &bvec, 1);
 597         bio.bi_opf = REQ_OP_READ;
 598         bio.bi_iter.bi_sector = iomap_sector(iomap, block_start);
 599         bio_set_dev(&bio, iomap->bdev);
 600         __bio_add_page(&bio, page, plen, poff);
 601         return submit_bio_wait(&bio);
 602 }
 603
 604 static int
 605 __iomap_write_begin(struct inode *inode, loff_t pos, unsigned len,
 606                 struct page *page, struct iomap *iomap)
 607 {
 608         struct iomap_page *iop = iomap_page_create(inode, page);
 609         loff_t block_size = i_blocksize(inode);
 610         loff_t block_start = pos & ~(block_size - 1);
 611         loff_t block_end = (pos + len + block_size - 1) & ~(block_size - 1);
 612         unsigned from = offset_in_page(pos), to = from + len, poff, plen;
 613         int status = 0;
 614
 615         if (PageUptodate(page))
 616                 return 0;
 617
 618         do {
 619                 iomap_adjust_read_range(inode, iop, &block_start,
 620                                 block_end - block_start, &poff, &plen);
 621                 if (plen == 0)
 622                         break;
 623
 624                 if ((from > poff && from < poff + plen) ||
 625                     (to > poff && to < poff + plen)) {
 626                         status = iomap_read_page_sync(inode, block_start, page,
 627                                         poff, plen, from, to, iomap);
 628                         if (status)
 629                                 break;
 630                 }
 631
 632         } while ((block_start += plen) < block_end);
 633
 634         return status;
 635 }
 636
 637 static int
 638 iomap_write_begin(struct inode *inode, loff_t pos, unsigned len, unsigned flags,
 639                 struct page **pagep, struct iomap *iomap)
 640 {
 641         pgoff_t index = pos >> PAGE_SHIFT;
 642         struct page *page;
 643         int status = 0;
 644
 645         BUG_ON(pos + len > iomap->offset + iomap->length);
 646
 647         if (fatal_signal_pending(current))
 648                 return -EINTR;
 649
 650         page = grab_cache_page_write_begin(inode->i_mapping, index, flags);
 651         if (!page)
 652                 return -ENOMEM;
 653
 654         if (iomap->type == IOMAP_INLINE)
 655                 iomap_read_inline_data(inode, page, iomap);
 656         else if (iomap->flags & IOMAP_F_BUFFER_HEAD)
 657                 status = __block_write_begin_int(page, pos, len, NULL, iomap);
 658         else
 659                 status = __iomap_write_begin(inode, pos, len, page, iomap);
 660         if (unlikely(status)) {
 661                 unlock_page(page);
 662                 put_page(page);
 663                 page = NULL;
 664
 665                 iomap_write_failed(inode, pos, len);
 666         }
 667
 668         *pagep = page;
 669         return status;
 670 }
 671
 672 int
 673 iomap_set_page_dirty(struct page *page)
 674 {
 675         struct address_space *mapping = page_mapping(page);
 676         int newly_dirty;
 677
 678         if (unlikely(!mapping))
 679                 return !TestSetPageDirty(page);
 680
 681         /*
 682          * Lock out page->mem_cgroup migration to keep PageDirty
 683          * synchronized with per-memcg dirty page counters.
 684          */
 685         lock_page_memcg(page);
 686         newly_dirty = !TestSetPageDirty(page);
 687         if (newly_dirty)
 688                 __set_page_dirty(page, mapping, 0);
 689         unlock_page_memcg(page);
 690
 691         if (newly_dirty)
 692                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 693         return newly_dirty;
 694 }
 695 EXPORT_SYMBOL_GPL(iomap_set_page_dirty);
 696
 697 static int
 698 __iomap_write_end(struct inode *inode, loff_t pos, unsigned len,
 699                 unsigned copied, struct page *page, struct iomap *iomap)
 700 {
 701         flush_dcache_page(page);
 702
 703         /*
 704          * The blocks that were entirely written will now be uptodate, so we
 705          * don't have to worry about a readpage reading them and overwriting a
 706          * partial write.  However if we have encountered a short write and only
 707          * partially written into a block, it will not be marked uptodate, so a
 708          * readpage might come in and destroy our partial write.
 709          *
 710          * Do the simplest thing, and just treat any short write to a non
 711          * uptodate page as a zero-length write, and force the caller to redo
 712          * the whole thing.
 713          */
 714         if (unlikely(copied < len && !PageUptodate(page))) {
 715                 copied = 0;
 716         } else {
 717                 iomap_set_range_uptodate(page, offset_in_page(pos), len);
 718                 iomap_set_page_dirty(page);
 719         }
 720         return __generic_write_end(inode, pos, copied, page);
 721 }
 722
 723 static int
 724 iomap_write_end_inline(struct inode *inode, struct page *page,
 725                 struct iomap *iomap, loff_t pos, unsigned copied)
 726 {
 727         void *addr;
 728
 729         WARN_ON_ONCE(!PageUptodate(page));
 730         BUG_ON(pos + copied > PAGE_SIZE - offset_in_page(iomap->inline_data));
 731
 732         addr = kmap_atomic(page);
 733         memcpy(iomap->inline_data + pos, addr + pos, copied);
 734         kunmap_atomic(addr);
 735
 736         mark_inode_dirty(inode);
 737         __generic_write_end(inode, pos, copied, page);
 738         return copied;
 739 }
 740
 741 static int
 742 iomap_write_end(struct inode *inode, loff_t pos, unsigned len,
 743                 unsigned copied, struct page *page, struct iomap *iomap)
 744 {
 745         int ret;
 746
 747         if (iomap->type == IOMAP_INLINE) {
 748                 ret = iomap_write_end_inline(inode, page, iomap, pos, copied);
 749         } else if (iomap->flags & IOMAP_F_BUFFER_HEAD) {
 750                 ret = generic_write_end(NULL, inode->i_mapping, pos, len,
 751                                 copied, page, NULL);
 752         } else {
 753                 ret = __iomap_write_end(inode, pos, len, copied, page, iomap);
 754         }
 755
 756         if (iomap->page_done)
 757                 iomap->page_done(inode, pos, copied, page, iomap);
 758
 759         if (ret < len)
 760                 iomap_write_failed(inode, pos, len);
 761         return ret;
 762 }
 763
 764 static loff_t
 765 iomap_write_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 766                 struct iomap *iomap)
 767 {
 768         struct iov_iter *i = data;
 769         long status = 0;
 770         ssize_t written = 0;
 771         unsigned int flags = AOP_FLAG_NOFS;
 772
 773         do {
 774                 struct page *page;
 775                 unsigned long offset;   /* Offset into pagecache page */
 776                 unsigned long bytes;    /* Bytes to write to page */
 777                 size_t copied;          /* Bytes copied from user */
 778
 779                 offset = offset_in_page(pos);
 780                 bytes = min_t(unsigned long, PAGE_SIZE - offset,
 781                                                 iov_iter_count(i));
 782 again:
 783                 if (bytes > length)
 784                         bytes = length;
 785
 786                 /*
 787                  * Bring in the user page that we will copy from _first_.
 788                  * Otherwise there's a nasty deadlock on copying from the
 789                  * same page as we're writing to, without it being marked
 790                  * up-to-date.
 791                  *
 792                  * Not only is this an optimisation, but it is also required
 793                  * to check that the address is actually valid, when atomic
 794                  * usercopies are used, below.
 795                  */
 796                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
 797                         status = -EFAULT;
 798                         break;
 799                 }
 800
 801                 status = iomap_write_begin(inode, pos, bytes, flags, &page,
 802                                 iomap);
 803                 if (unlikely(status))
 804                         break;
 805
 806                 if (mapping_writably_mapped(inode->i_mapping))
 807                         flush_dcache_page(page);
 808
 809                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
 810
 811                 flush_dcache_page(page);
 812
 813                 status = iomap_write_end(inode, pos, bytes, copied, page,
 814                                 iomap);
 815                 if (unlikely(status < 0))
 816                         break;
 817                 copied = status;
 818
 819                 cond_resched();
 820
 821                 iov_iter_advance(i, copied);
 822                 if (unlikely(copied == 0)) {
 823                         /*
 824                          * If we were unable to copy any data at all, we must
 825                          * fall back to a single segment length write.
 826                          *
 827                          * If we didn't fallback here, we could livelock
 828                          * because not all segments in the iov can be copied at
 829                          * once without a pagefault.
 830                          */
 831                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
 832                                                 iov_iter_single_seg_count(i));
 833                         goto again;
 834                 }
 835                 pos += copied;
 836                 written += copied;
 837                 length -= copied;
 838
 839                 balance_dirty_pages_ratelimited(inode->i_mapping);
 840         } while (iov_iter_count(i) && length);
 841
 842         return written ? written : status;
 843 }
 844
 845 ssize_t
 846 iomap_file_buffered_write(struct kiocb *iocb, struct iov_iter *iter,
 847                 const struct iomap_ops *ops)
 848 {
 849         struct inode *inode = iocb->ki_filp->f_mapping->host;
 850         loff_t pos = iocb->ki_pos, ret = 0, written = 0;
 851
 852         while (iov_iter_count(iter)) {
 853                 ret = iomap_apply(inode, pos, iov_iter_count(iter),
 854                                 IOMAP_WRITE, ops, iter, iomap_write_actor);
 855                 if (ret <= 0)
 856                         break;
 857                 pos += ret;
 858                 written += ret;
 859         }
 860
 861         return written ? written : ret;
 862 }
 863 EXPORT_SYMBOL_GPL(iomap_file_buffered_write);
 864
 865 static struct page *
 866 __iomap_read_page(struct inode *inode, loff_t offset)
 867 {
 868         struct address_space *mapping = inode->i_mapping;
 869         struct page *page;
 870
 871         page = read_mapping_page(mapping, offset >> PAGE_SHIFT, NULL);
 872         if (IS_ERR(page))
 873                 return page;
 874         if (!PageUptodate(page)) {
 875                 put_page(page);
 876                 return ERR_PTR(-EIO);
 877         }
 878         return page;
 879 }
 880
 881 static loff_t
 882 iomap_dirty_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 883                 struct iomap *iomap)
 884 {
 885         long status = 0;
 886         ssize_t written = 0;
 887
 888         do {
 889                 struct page *page, *rpage;
 890                 unsigned long offset;   /* Offset into pagecache page */
 891                 unsigned long bytes;    /* Bytes to write to page */
 892
 893                 offset = offset_in_page(pos);
 894                 bytes = min_t(loff_t, PAGE_SIZE - offset, length);
 895
 896                 rpage = __iomap_read_page(inode, pos);
 897                 if (IS_ERR(rpage))
 898                         return PTR_ERR(rpage);
 899
 900                 status = iomap_write_begin(inode, pos, bytes,
 901                                            AOP_FLAG_NOFS, &page, iomap);
 902                 put_page(rpage);
 903                 if (unlikely(status))
 904                         return status;
 905
 906                 WARN_ON_ONCE(!PageUptodate(page));
 907
 908                 status = iomap_write_end(inode, pos, bytes, bytes, page, iomap);
 909                 if (unlikely(status <= 0)) {
 910                         if (WARN_ON_ONCE(status == 0))
 911                                 return -EIO;
 912                         return status;
 913                 }
 914
 915                 cond_resched();
 916
 917                 pos += status;
 918                 written += status;
 919                 length -= status;
 920
 921                 balance_dirty_pages_ratelimited(inode->i_mapping);
 922         } while (length);
 923
 924         return written;
 925 }
 926
 927 int
 928 iomap_file_dirty(struct inode *inode, loff_t pos, loff_t len,
 929                 const struct iomap_ops *ops)
 930 {
 931         loff_t ret;
 932
 933         while (len) {
 934                 ret = iomap_apply(inode, pos, len, IOMAP_WRITE, ops, NULL,
 935                                 iomap_dirty_actor);
 936                 if (ret <= 0)
 937                         return ret;
 938                 pos += ret;
 939                 len -= ret;
 940         }
 941
 942         return 0;
 943 }
 944 EXPORT_SYMBOL_GPL(iomap_file_dirty);
 945
 946 static int iomap_zero(struct inode *inode, loff_t pos, unsigned offset,
 947                 unsigned bytes, struct iomap *iomap)
 948 {
 949         struct page *page;
 950         int status;
 951
 952         status = iomap_write_begin(inode, pos, bytes, AOP_FLAG_NOFS, &page,
 953                                    iomap);
 954         if (status)
 955                 return status;
 956
 957         zero_user(page, offset, bytes);
 958         mark_page_accessed(page);
 959
 960         return iomap_write_end(inode, pos, bytes, bytes, page, iomap);
 961 }
 962
 963 static int iomap_dax_zero(loff_t pos, unsigned offset, unsigned bytes,
 964                 struct iomap *iomap)
 965 {
 966         return __dax_zero_page_range(iomap->bdev, iomap->dax_dev,
 967                         iomap_sector(iomap, pos & PAGE_MASK), offset, bytes);
 968 }
 969
 970 static loff_t
 971 iomap_zero_range_actor(struct inode *inode, loff_t pos, loff_t count,
 972                 void *data, struct iomap *iomap)
 973 {
 974         bool *did_zero = data;
 975         loff_t written = 0;
 976         int status;
 977
 978         /* already zeroed?  we're done. */
 979         if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
 980                 return count;
 981
 982         do {
 983                 unsigned offset, bytes;
 984
 985                 offset = offset_in_page(pos);
 986                 bytes = min_t(loff_t, PAGE_SIZE - offset, count);
 987
 988                 if (IS_DAX(inode))
 989                         status = iomap_dax_zero(pos, offset, bytes, iomap);
 990                 else
 991                         status = iomap_zero(inode, pos, offset, bytes, iomap);
 992                 if (status < 0)
 993                         return status;
 994
 995                 pos += bytes;
 996                 count -= bytes;
 997                 written += bytes;
 998                 if (did_zero)
 999                         *did_zero = true;
1000         } while (count > 0);
1001
1002         return written;
1003 }
1004
1005 int
1006 iomap_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
1007                 const struct iomap_ops *ops)
1008 {
1009         loff_t ret;
1010
1011         while (len > 0) {
1012                 ret = iomap_apply(inode, pos, len, IOMAP_ZERO,
1013                                 ops, did_zero, iomap_zero_range_actor);
1014                 if (ret <= 0)
1015                         return ret;
1016
1017                 pos += ret;
1018                 len -= ret;
1019         }
1020
1021         return 0;
1022 }
1023 EXPORT_SYMBOL_GPL(iomap_zero_range);
1024
1025 int
1026 iomap_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
1027                 const struct iomap_ops *ops)
1028 {
1029         unsigned int blocksize = i_blocksize(inode);
1030         unsigned int off = pos & (blocksize - 1);
1031
1032         /* Block boundary? Nothing to do */
1033         if (!off)
1034                 return 0;
1035         return iomap_zero_range(inode, pos, blocksize - off, did_zero, ops);
1036 }
1037 EXPORT_SYMBOL_GPL(iomap_truncate_page);
1038
1039 static loff_t
1040 iomap_page_mkwrite_actor(struct inode *inode, loff_t pos, loff_t length,
1041                 void *data, struct iomap *iomap)
1042 {
1043         struct page *page = data;
1044         int ret;
1045
1046         if (iomap->flags & IOMAP_F_BUFFER_HEAD) {
1047                 ret = __block_write_begin_int(page, pos, length, NULL, iomap);
1048                 if (ret)
1049                         return ret;
1050                 block_commit_write(page, 0, length);
1051         } else {
1052                 WARN_ON_ONCE(!PageUptodate(page));
1053                 iomap_page_create(inode, page);
1054         }
1055
1056         return length;
1057 }
1058
1059 int iomap_page_mkwrite(struct vm_fault *vmf, const struct iomap_ops *ops)
1060 {
1061         struct page *page = vmf->page;
1062         struct inode *inode = file_inode(vmf->vma->vm_file);
1063         unsigned long length;
1064         loff_t offset, size;
1065         ssize_t ret;
1066
1067         lock_page(page);
1068         size = i_size_read(inode);
1069         if ((page->mapping != inode->i_mapping) ||
1070             (page_offset(page) > size)) {
1071                 /* We overload EFAULT to mean page got truncated */
1072                 ret = -EFAULT;
1073                 goto out_unlock;
1074         }
1075
1076         /* page is wholly or partially inside EOF */
1077         if (((page->index + 1) << PAGE_SHIFT) > size)
1078                 length = offset_in_page(size);
1079         else
1080                 length = PAGE_SIZE;
1081
1082         offset = page_offset(page);
1083         while (length > 0) {
1084                 ret = iomap_apply(inode, offset, length,
1085                                 IOMAP_WRITE | IOMAP_FAULT, ops, page,
1086                                 iomap_page_mkwrite_actor);
1087                 if (unlikely(ret <= 0))
1088                         goto out_unlock;
1089                 offset += ret;
1090                 length -= ret;
1091         }
1092
1093         set_page_dirty(page);
1094         wait_for_stable_page(page);
1095         return VM_FAULT_LOCKED;
1096 out_unlock:
1097         unlock_page(page);
1098         return block_page_mkwrite_return(ret);
1099 }
1100 EXPORT_SYMBOL_GPL(iomap_page_mkwrite);
1101
1102 struct fiemap_ctx {
1103         struct fiemap_extent_info *fi;
1104         struct iomap prev;
1105 };
1106
1107 static int iomap_to_fiemap(struct fiemap_extent_info *fi,
1108                 struct iomap *iomap, u32 flags)
1109 {
1110         switch (iomap->type) {
1111         case IOMAP_HOLE:
1112                 /* skip holes */
1113                 return 0;
1114         case IOMAP_DELALLOC:
1115                 flags |= FIEMAP_EXTENT_DELALLOC | FIEMAP_EXTENT_UNKNOWN;
1116                 break;
1117         case IOMAP_MAPPED:
1118                 break;
1119         case IOMAP_UNWRITTEN:
1120                 flags |= FIEMAP_EXTENT_UNWRITTEN;
1121                 break;
1122         case IOMAP_INLINE:
1123                 flags |= FIEMAP_EXTENT_DATA_INLINE;
1124                 break;
1125         }
1126
1127         if (iomap->flags & IOMAP_F_MERGED)
1128                 flags |= FIEMAP_EXTENT_MERGED;
1129         if (iomap->flags & IOMAP_F_SHARED)
1130                 flags |= FIEMAP_EXTENT_SHARED;
1131
1132         return fiemap_fill_next_extent(fi, iomap->offset,
1133                         iomap->addr != IOMAP_NULL_ADDR ? iomap->addr : 0,
1134                         iomap->length, flags);
1135 }
1136
1137 static loff_t
1138 iomap_fiemap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
1139                 struct iomap *iomap)
1140 {
1141         struct fiemap_ctx *ctx = data;
1142         loff_t ret = length;
1143
1144         if (iomap->type == IOMAP_HOLE)
1145                 return length;
1146
1147         ret = iomap_to_fiemap(ctx->fi, &ctx->prev, 0);
1148         ctx->prev = *iomap;
1149         switch (ret) {
1150         case 0:         /* success */
1151                 return length;
1152         case 1:         /* extent array full */
1153                 return 0;
1154         default:
1155                 return ret;
1156         }
1157 }
1158
1159 int iomap_fiemap(struct inode *inode, struct fiemap_extent_info *fi,
1160                 loff_t start, loff_t len, const struct iomap_ops *ops)
1161 {
1162         struct fiemap_ctx ctx;
1163         loff_t ret;
1164
1165         memset(&ctx, 0, sizeof(ctx));
1166         ctx.fi = fi;
1167         ctx.prev.type = IOMAP_HOLE;
1168
1169         ret = fiemap_check_flags(fi, FIEMAP_FLAG_SYNC);
1170         if (ret)
1171                 return ret;
1172
1173         if (fi->fi_flags & FIEMAP_FLAG_SYNC) {
1174                 ret = filemap_write_and_wait(inode->i_mapping);
1175                 if (ret)
1176                         return ret;
1177         }
1178
1179         while (len > 0) {
1180                 ret = iomap_apply(inode, start, len, IOMAP_REPORT, ops, &ctx,
1181                                 iomap_fiemap_actor);
1182                 /* inode with no (attribute) mapping will give ENOENT */
1183                 if (ret == -ENOENT)
1184                         break;
1185                 if (ret < 0)
1186                         return ret;
1187                 if (ret == 0)
1188                         break;
1189
1190                 start += ret;
1191                 len -= ret;
1192         }
1193
1194         if (ctx.prev.type != IOMAP_HOLE) {
1195                 ret = iomap_to_fiemap(fi, &ctx.prev, FIEMAP_EXTENT_LAST);
1196                 if (ret < 0)
1197                         return ret;
1198         }
1199
1200         return 0;
1201 }
1202 EXPORT_SYMBOL_GPL(iomap_fiemap);
1203
1204 /*
1205  * Seek for SEEK_DATA / SEEK_HOLE within @page, starting at @lastoff.
1206  * Returns true if found and updates @lastoff to the offset in file.
1207  */
1208 static bool
1209 page_seek_hole_data(struct inode *inode, struct page *page, loff_t *lastoff,
1210                 int whence)
1211 {
1212         const struct address_space_operations *ops = inode->i_mapping->a_ops;
1213         unsigned int bsize = i_blocksize(inode), off;
1214         bool seek_data = whence == SEEK_DATA;
1215         loff_t poff = page_offset(page);
1216
1217         if (WARN_ON_ONCE(*lastoff >= poff + PAGE_SIZE))
1218                 return false;
1219
1220         if (*lastoff < poff) {
1221                 /*
1222                  * Last offset smaller than the start of the page means we found
1223                  * a hole:
1224                  */
1225                 if (whence == SEEK_HOLE)
1226                         return true;
1227                 *lastoff = poff;
1228         }
1229
1230         /*
1231          * Just check the page unless we can and should check block ranges:
1232          */
1233         if (bsize == PAGE_SIZE || !ops->is_partially_uptodate)
1234                 return PageUptodate(page) == seek_data;
1235
1236         lock_page(page);
1237         if (unlikely(page->mapping != inode->i_mapping))
1238                 goto out_unlock_not_found;
1239
1240         for (off = 0; off < PAGE_SIZE; off += bsize) {
1241                 if (offset_in_page(*lastoff) >= off + bsize)
1242                         continue;
1243                 if (ops->is_partially_uptodate(page, off, bsize) == seek_data) {
1244                         unlock_page(page);
1245                         return true;
1246                 }
1247                 *lastoff = poff + off + bsize;
1248         }
1249
1250 out_unlock_not_found:
1251         unlock_page(page);
1252         return false;
1253 }
1254
1255 /*
1256  * Seek for SEEK_DATA / SEEK_HOLE in the page cache.
1257  *
1258  * Within unwritten extents, the page cache determines which parts are holes
1259  * and which are data: uptodate buffer heads count as data; everything else
1260  * counts as a hole.
1261  *
1262  * Returns the resulting offset on successs, and -ENOENT otherwise.
1263  */
1264 static loff_t
1265 page_cache_seek_hole_data(struct inode *inode, loff_t offset, loff_t length,
1266                 int whence)
1267 {
1268         pgoff_t index = offset >> PAGE_SHIFT;
1269         pgoff_t end = DIV_ROUND_UP(offset + length, PAGE_SIZE);
1270         loff_t lastoff = offset;
1271         struct pagevec pvec;
1272
1273         if (length <= 0)
1274                 return -ENOENT;
1275
1276         pagevec_init(&pvec);
1277
1278         do {
1279                 unsigned nr_pages, i;
1280
1281                 nr_pages = pagevec_lookup_range(&pvec, inode->i_mapping, &index,
1282                                                 end - 1);
1283                 if (nr_pages == 0)
1284                         break;
1285
1286                 for (i = 0; i < nr_pages; i++) {
1287                         struct page *page = pvec.pages[i];
1288
1289                         if (page_seek_hole_data(inode, page, &lastoff, whence))
1290                                 goto check_range;
1291                         lastoff = page_offset(page) + PAGE_SIZE;
1292                 }
1293                 pagevec_release(&pvec);
1294         } while (index < end);
1295
1296         /* When no page at lastoff and we are not done, we found a hole. */
1297         if (whence != SEEK_HOLE)
1298                 goto not_found;
1299
1300 check_range:
1301         if (lastoff < offset + length)
1302                 goto out;
1303 not_found:
1304         lastoff = -ENOENT;
1305 out:
1306         pagevec_release(&pvec);
1307         return lastoff;
1308 }
1309
1310
1311 static loff_t
1312 iomap_seek_hole_actor(struct inode *inode, loff_t offset, loff_t length,
1313                       void *data, struct iomap *iomap)
1314 {
1315         switch (iomap->type) {
1316         case IOMAP_UNWRITTEN:
1317                 offset = page_cache_seek_hole_data(inode, offset, length,
1318                                                    SEEK_HOLE);
1319                 if (offset < 0)
1320                         return length;
1321                 /* fall through */
1322         case IOMAP_HOLE:
1323                 *(loff_t *)data = offset;
1324                 return 0;
1325         default:
1326                 return length;
1327         }
1328 }
1329
1330 loff_t
1331 iomap_seek_hole(struct inode *inode, loff_t offset, const struct iomap_ops *ops)
1332 {
1333         loff_t size = i_size_read(inode);
1334         loff_t length = size - offset;
1335         loff_t ret;
1336
1337         /* Nothing to be found before or beyond the end of the file. */
1338         if (offset < 0 || offset >= size)
1339                 return -ENXIO;
1340
1341         while (length > 0) {
1342                 ret = iomap_apply(inode, offset, length, IOMAP_REPORT, ops,
1343                                   &offset, iomap_seek_hole_actor);
1344                 if (ret < 0)
1345                         return ret;
1346                 if (ret == 0)
1347                         break;
1348
1349                 offset += ret;
1350                 length -= ret;
1351         }
1352
1353         return offset;
1354 }
1355 EXPORT_SYMBOL_GPL(iomap_seek_hole);
1356
1357 static loff_t
1358 iomap_seek_data_actor(struct inode *inode, loff_t offset, loff_t length,
1359                       void *data, struct iomap *iomap)
1360 {
1361         switch (iomap->type) {
1362         case IOMAP_HOLE:
1363                 return length;
1364         case IOMAP_UNWRITTEN:
1365                 offset = page_cache_seek_hole_data(inode, offset, length,
1366                                                    SEEK_DATA);
1367                 if (offset < 0)
1368                         return length;
1369                 /*FALLTHRU*/
1370         default:
1371                 *(loff_t *)data = offset;
1372                 return 0;
1373         }
1374 }
1375
1376 loff_t
1377 iomap_seek_data(struct inode *inode, loff_t offset, const struct iomap_ops *ops)
1378 {
1379         loff_t size = i_size_read(inode);
1380         loff_t length = size - offset;
1381         loff_t ret;
1382
1383         /* Nothing to be found before or beyond the end of the file. */
1384         if (offset < 0 || offset >= size)
1385                 return -ENXIO;
1386
1387         while (length > 0) {
1388                 ret = iomap_apply(inode, offset, length, IOMAP_REPORT, ops,
1389                                   &offset, iomap_seek_data_actor);
1390                 if (ret < 0)
1391                         return ret;
1392                 if (ret == 0)
1393                         break;
1394
1395                 offset += ret;
1396                 length -= ret;
1397         }
1398
1399         if (length <= 0)
1400                 return -ENXIO;
1401         return offset;
1402 }
1403 EXPORT_SYMBOL_GPL(iomap_seek_data);
1404
1405 /*
1406  * Private flags for iomap_dio, must not overlap with the public ones in
1407  * iomap.h:
1408  */
1409 #define IOMAP_DIO_WRITE_FUA     (1 << 28)
1410 #define IOMAP_DIO_NEED_SYNC     (1 << 29)
1411 #define IOMAP_DIO_WRITE         (1 << 30)
1412 #define IOMAP_DIO_DIRTY         (1 << 31)
1413
1414 struct iomap_dio {
1415         struct kiocb            *iocb;
1416         iomap_dio_end_io_t      *end_io;
1417         loff_t                  i_size;
1418         loff_t                  size;
1419         atomic_t                ref;
1420         unsigned                flags;
1421         int                     error;
1422         bool                    wait_for_completion;
1423
1424         union {
1425                 /* used during submission and for synchronous completion: */
1426                 struct {
1427                         struct iov_iter         *iter;
1428                         struct task_struct      *waiter;
1429                         struct request_queue    *last_queue;
1430                         blk_qc_t                cookie;
1431                 } submit;
1432
1433                 /* used for aio completion: */
1434                 struct {
1435                         struct work_struct      work;
1436                 } aio;
1437         };
1438 };
1439
1440 static ssize_t iomap_dio_complete(struct iomap_dio *dio)
1441 {
1442         struct kiocb *iocb = dio->iocb;
1443         struct inode *inode = file_inode(iocb->ki_filp);
1444         loff_t offset = iocb->ki_pos;
1445         ssize_t ret;
1446
1447         if (dio->end_io) {
1448                 ret = dio->end_io(iocb,
1449                                 dio->error ? dio->error : dio->size,
1450                                 dio->flags);
1451         } else {
1452                 ret = dio->error;
1453         }
1454
1455         if (likely(!ret)) {
1456                 ret = dio->size;
1457                 /* check for short read */
1458                 if (offset + ret > dio->i_size &&
1459                     !(dio->flags & IOMAP_DIO_WRITE))
1460                         ret = dio->i_size - offset;
1461                 iocb->ki_pos += ret;
1462         }
1463
1464         /*
1465          * Try again to invalidate clean pages which might have been cached by
1466          * non-direct readahead, or faulted in by get_user_pages() if the source
1467          * of the write was an mmap'ed region of the file we're writing.  Either
1468          * one is a pretty crazy thing to do, so we don't support it 100%.  If
1469          * this invalidation fails, tough, the write still worked...
1470          *
1471          * And this page cache invalidation has to be after dio->end_io(), as
1472          * some filesystems convert unwritten extents to real allocations in
1473          * end_io() when necessary, otherwise a racing buffer read would cache
1474          * zeros from unwritten extents.
1475          */
1476         if (!dio->error &&
1477             (dio->flags & IOMAP_DIO_WRITE) && inode->i_mapping->nrpages) {
1478                 int err;
1479                 err = invalidate_inode_pages2_range(inode->i_mapping,
1480                                 offset >> PAGE_SHIFT,
1481                                 (offset + dio->size - 1) >> PAGE_SHIFT);
1482                 if (err)
1483                         dio_warn_stale_pagecache(iocb->ki_filp);
1484         }
1485
1486         /*
1487          * If this is a DSYNC write, make sure we push it to stable storage now
1488          * that we've written data.
1489          */
1490         if (ret > 0 && (dio->flags & IOMAP_DIO_NEED_SYNC))
1491                 ret = generic_write_sync(iocb, ret);
1492
1493         inode_dio_end(file_inode(iocb->ki_filp));
1494         kfree(dio);
1495
1496         return ret;
1497 }
1498
1499 static void iomap_dio_complete_work(struct work_struct *work)
1500 {
1501         struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work);
1502         struct kiocb *iocb = dio->iocb;
1503
1504         iocb->ki_complete(iocb, iomap_dio_complete(dio), 0);
1505 }
1506
1507 /*
1508  * Set an error in the dio if none is set yet.  We have to use cmpxchg
1509  * as the submission context and the completion context(s) can race to
1510  * update the error.
1511  */
1512 static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret)
1513 {
1514         cmpxchg(&dio->error, 0, ret);
1515 }
1516
1517 static void iomap_dio_bio_end_io(struct bio *bio)
1518 {
1519         struct iomap_dio *dio = bio->bi_private;
1520         bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
1521
1522         if (bio->bi_status)
1523                 iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
1524
1525         if (atomic_dec_and_test(&dio->ref)) {
1526                 if (dio->wait_for_completion) {
1527                         struct task_struct *waiter = dio->submit.waiter;
1528                         WRITE_ONCE(dio->submit.waiter, NULL);
1529                         wake_up_process(waiter);
1530                 } else if (dio->flags & IOMAP_DIO_WRITE) {
1531                         struct inode *inode = file_inode(dio->iocb->ki_filp);
1532
1533                         INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
1534                         queue_work(inode->i_sb->s_dio_done_wq, &dio->aio.work);
1535                 } else {
1536                         iomap_dio_complete_work(&dio->aio.work);
1537                 }
1538         }
1539
1540         if (should_dirty) {
1541                 bio_check_pages_dirty(bio);
1542         } else {
1543                 struct bio_vec *bvec;
1544                 int i;
1545
1546                 bio_for_each_segment_all(bvec, bio, i)
1547                         put_page(bvec->bv_page);
1548                 bio_put(bio);
1549         }
1550 }
1551
1552 static blk_qc_t
1553 iomap_dio_zero(struct iomap_dio *dio, struct iomap *iomap, loff_t pos,
1554                 unsigned len)
1555 {
1556         struct page *page = ZERO_PAGE(0);
1557         struct bio *bio;
1558
1559         bio = bio_alloc(GFP_KERNEL, 1);
1560         bio_set_dev(bio, iomap->bdev);
1561         bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
1562         bio->bi_private = dio;
1563         bio->bi_end_io = iomap_dio_bio_end_io;
1564
1565         get_page(page);
1566         __bio_add_page(bio, page, len, 0);
1567         bio_set_op_attrs(bio, REQ_OP_WRITE, REQ_SYNC | REQ_IDLE);
1568
1569         atomic_inc(&dio->ref);
1570         return submit_bio(bio);
1571 }
1572
1573 static loff_t
1574 iomap_dio_bio_actor(struct inode *inode, loff_t pos, loff_t length,
1575                 struct iomap_dio *dio, struct iomap *iomap)
1576 {
1577         unsigned int blkbits = blksize_bits(bdev_logical_block_size(iomap->bdev));
1578         unsigned int fs_block_size = i_blocksize(inode), pad;
1579         unsigned int align = iov_iter_alignment(dio->submit.iter);
1580         struct iov_iter iter;
1581         struct bio *bio;
1582         bool need_zeroout = false;
1583         bool use_fua = false;
1584         int nr_pages, ret;
1585         size_t copied = 0;
1586
1587         if ((pos | length | align) & ((1 << blkbits) - 1))
1588                 return -EINVAL;
1589
1590         if (iomap->type == IOMAP_UNWRITTEN) {
1591                 dio->flags |= IOMAP_DIO_UNWRITTEN;
1592                 need_zeroout = true;
1593         }
1594
1595         if (iomap->flags & IOMAP_F_SHARED)
1596                 dio->flags |= IOMAP_DIO_COW;
1597
1598         if (iomap->flags & IOMAP_F_NEW) {
1599                 need_zeroout = true;
1600         } else {
1601                 /*
1602                  * Use a FUA write if we need datasync semantics, this
1603                  * is a pure data IO that doesn't require any metadata
1604                  * updates and the underlying device supports FUA. This
1605                  * allows us to avoid cache flushes on IO completion.
1606                  */
1607                 if (!(iomap->flags & (IOMAP_F_SHARED|IOMAP_F_DIRTY)) &&
1608                     (dio->flags & IOMAP_DIO_WRITE_FUA) &&
1609                     blk_queue_fua(bdev_get_queue(iomap->bdev)))
1610                         use_fua = true;
1611         }
1612
1613         /*
1614          * Operate on a partial iter trimmed to the extent we were called for.
1615          * We'll update the iter in the dio once we're done with this extent.
1616          */
1617         iter = *dio->submit.iter;
1618         iov_iter_truncate(&iter, length);
1619
1620         nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
1621         if (nr_pages <= 0)
1622                 return nr_pages;
1623
1624         if (need_zeroout) {
1625                 /* zero out from the start of the block to the write offset */
1626                 pad = pos & (fs_block_size - 1);
1627                 if (pad)
1628                         iomap_dio_zero(dio, iomap, pos - pad, pad);
1629         }
1630
1631         do {
1632                 size_t n;
1633                 if (dio->error) {
1634                         iov_iter_revert(dio->submit.iter, copied);
1635                         return 0;
1636                 }
1637
1638                 bio = bio_alloc(GFP_KERNEL, nr_pages);
1639                 bio_set_dev(bio, iomap->bdev);
1640                 bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
1641                 bio->bi_write_hint = dio->iocb->ki_hint;
1642                 bio->bi_ioprio = dio->iocb->ki_ioprio;
1643                 bio->bi_private = dio;
1644                 bio->bi_end_io = iomap_dio_bio_end_io;
1645
1646                 ret = bio_iov_iter_get_pages(bio, &iter);
1647                 if (unlikely(ret)) {
1648                         bio_put(bio);
1649                         return copied ? copied : ret;
1650                 }
1651
1652                 n = bio->bi_iter.bi_size;
1653                 if (dio->flags & IOMAP_DIO_WRITE) {
1654                         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
1655                         if (use_fua)
1656                                 bio->bi_opf |= REQ_FUA;
1657                         else
1658                                 dio->flags &= ~IOMAP_DIO_WRITE_FUA;
1659                         task_io_account_write(n);
1660                 } else {
1661                         bio->bi_opf = REQ_OP_READ;
1662                         if (dio->flags & IOMAP_DIO_DIRTY)
1663                                 bio_set_pages_dirty(bio);
1664                 }
1665
1666                 iov_iter_advance(dio->submit.iter, n);
1667
1668                 dio->size += n;
1669                 pos += n;
1670                 copied += n;
1671
1672                 nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
1673
1674                 atomic_inc(&dio->ref);
1675
1676                 dio->submit.last_queue = bdev_get_queue(iomap->bdev);
1677                 dio->submit.cookie = submit_bio(bio);
1678         } while (nr_pages);
1679
1680         if (need_zeroout) {
1681                 /* zero out from the end of the write to the end of the block */
1682                 pad = pos & (fs_block_size - 1);
1683                 if (pad)
1684                         iomap_dio_zero(dio, iomap, pos, fs_block_size - pad);
1685         }
1686         return copied;
1687 }
1688
1689 static loff_t
1690 iomap_dio_hole_actor(loff_t length, struct iomap_dio *dio)
1691 {
1692         length = iov_iter_zero(length, dio->submit.iter);
1693         dio->size += length;
1694         return length;
1695 }
1696
1697 static loff_t
1698 iomap_dio_inline_actor(struct inode *inode, loff_t pos, loff_t length,
1699                 struct iomap_dio *dio, struct iomap *iomap)
1700 {
1701         struct iov_iter *iter = dio->submit.iter;
1702         size_t copied;
1703
1704         BUG_ON(pos + length > PAGE_SIZE - offset_in_page(iomap->inline_data));
1705
1706         if (dio->flags & IOMAP_DIO_WRITE) {
1707                 loff_t size = inode->i_size;
1708
1709                 if (pos > size)
1710                         memset(iomap->inline_data + size, 0, pos - size);
1711                 copied = copy_from_iter(iomap->inline_data + pos, length, iter);
1712                 if (copied) {
1713                         if (pos + copied > size)
1714                                 i_size_write(inode, pos + copied);
1715                         mark_inode_dirty(inode);
1716                 }
1717         } else {
1718                 copied = copy_to_iter(iomap->inline_data + pos, length, iter);
1719         }
1720         dio->size += copied;
1721         return copied;
1722 }
1723
1724 static loff_t
1725 iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
1726                 void *data, struct iomap *iomap)
1727 {
1728         struct iomap_dio *dio = data;
1729
1730         switch (iomap->type) {
1731         case IOMAP_HOLE:
1732                 if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
1733                         return -EIO;
1734                 return iomap_dio_hole_actor(length, dio);
1735         case IOMAP_UNWRITTEN:
1736                 if (!(dio->flags & IOMAP_DIO_WRITE))
1737                         return iomap_dio_hole_actor(length, dio);
1738                 return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
1739         case IOMAP_MAPPED:
1740                 return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
1741         case IOMAP_INLINE:
1742                 return iomap_dio_inline_actor(inode, pos, length, dio, iomap);
1743         default:
1744                 WARN_ON_ONCE(1);
1745                 return -EIO;
1746         }
1747 }
1748
1749 /*
1750  * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
1751  * is being issued as AIO or not.  This allows us to optimise pure data writes
1752  * to use REQ_FUA rather than requiring generic_write_sync() to issue a
1753  * REQ_FLUSH post write. This is slightly tricky because a single request here
1754  * can be mapped into multiple disjoint IOs and only a subset of the IOs issued
1755  * may be pure data writes. In that case, we still need to do a full data sync
1756  * completion.
1757  */
1758 ssize_t
1759 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
1760                 const struct iomap_ops *ops, iomap_dio_end_io_t end_io)
1761 {
1762         struct address_space *mapping = iocb->ki_filp->f_mapping;
1763         struct inode *inode = file_inode(iocb->ki_filp);
1764         size_t count = iov_iter_count(iter);
1765         loff_t pos = iocb->ki_pos, start = pos;
1766         loff_t end = iocb->ki_pos + count - 1, ret = 0;
1767         unsigned int flags = IOMAP_DIRECT;
1768         struct blk_plug plug;
1769         struct iomap_dio *dio;
1770
1771         lockdep_assert_held(&inode->i_rwsem);
1772
1773         if (!count)
1774                 return 0;
1775
1776         dio = kmalloc(sizeof(*dio), GFP_KERNEL);
1777         if (!dio)
1778                 return -ENOMEM;
1779
1780         dio->iocb = iocb;
1781         atomic_set(&dio->ref, 1);
1782         dio->size = 0;
1783         dio->i_size = i_size_read(inode);
1784         dio->end_io = end_io;
1785         dio->error = 0;
1786         dio->flags = 0;
1787         dio->wait_for_completion = is_sync_kiocb(iocb);
1788
1789         dio->submit.iter = iter;
1790         dio->submit.waiter = current;
1791         dio->submit.cookie = BLK_QC_T_NONE;
1792         dio->submit.last_queue = NULL;
1793
1794         if (iov_iter_rw(iter) == READ) {
1795                 if (pos >= dio->i_size)
1796                         goto out_free_dio;
1797
1798                 if (iter->type == ITER_IOVEC)
1799                         dio->flags |= IOMAP_DIO_DIRTY;
1800         } else {
1801                 flags |= IOMAP_WRITE;
1802                 dio->flags |= IOMAP_DIO_WRITE;
1803
1804                 /* for data sync or sync, we need sync completion processing */
1805                 if (iocb->ki_flags & IOCB_DSYNC)
1806                         dio->flags |= IOMAP_DIO_NEED_SYNC;
1807
1808                 /*
1809                  * For datasync only writes, we optimistically try using FUA for
1810                  * this IO.  Any non-FUA write that occurs will clear this flag,
1811                  * hence we know before completion whether a cache flush is
1812                  * necessary.
1813                  */
1814                 if ((iocb->ki_flags & (IOCB_DSYNC | IOCB_SYNC)) == IOCB_DSYNC)
1815                         dio->flags |= IOMAP_DIO_WRITE_FUA;
1816         }
1817
1818         if (iocb->ki_flags & IOCB_NOWAIT) {
1819                 if (filemap_range_has_page(mapping, start, end)) {
1820                         ret = -EAGAIN;
1821                         goto out_free_dio;
1822                 }
1823                 flags |= IOMAP_NOWAIT;
1824         }
1825
1826         ret = filemap_write_and_wait_range(mapping, start, end);
1827         if (ret)
1828                 goto out_free_dio;
1829
1830         /*
1831          * Try to invalidate cache pages for the range we're direct
1832          * writing.  If this invalidation fails, tough, the write will
1833          * still work, but racing two incompatible write paths is a
1834          * pretty crazy thing to do, so we don't support it 100%.
1835          */
1836         ret = invalidate_inode_pages2_range(mapping,
1837                         start >> PAGE_SHIFT, end >> PAGE_SHIFT);
1838         if (ret)
1839                 dio_warn_stale_pagecache(iocb->ki_filp);
1840         ret = 0;
1841
1842         if (iov_iter_rw(iter) == WRITE && !dio->wait_for_completion &&
1843             !inode->i_sb->s_dio_done_wq) {
1844                 ret = sb_init_dio_done_wq(inode->i_sb);
1845                 if (ret < 0)
1846                         goto out_free_dio;
1847         }
1848
1849         inode_dio_begin(inode);
1850
1851         blk_start_plug(&plug);
1852         do {
1853                 ret = iomap_apply(inode, pos, count, flags, ops, dio,
1854                                 iomap_dio_actor);
1855                 if (ret <= 0) {
1856                         /* magic error code to fall back to buffered I/O */
1857                         if (ret == -ENOTBLK) {
1858                                 dio->wait_for_completion = true;
1859                                 ret = 0;
1860                         }
1861                         break;
1862                 }
1863                 pos += ret;
1864
1865                 if (iov_iter_rw(iter) == READ && pos >= dio->i_size)
1866                         break;
1867         } while ((count = iov_iter_count(iter)) > 0);
1868         blk_finish_plug(&plug);
1869
1870         if (ret < 0)
1871                 iomap_dio_set_error(dio, ret);
1872
1873         /*
1874          * If all the writes we issued were FUA, we don't need to flush the
1875          * cache on IO completion. Clear the sync flag for this case.
1876          */
1877         if (dio->flags & IOMAP_DIO_WRITE_FUA)
1878                 dio->flags &= ~IOMAP_DIO_NEED_SYNC;
1879
1880         if (!atomic_dec_and_test(&dio->ref)) {
1881                 if (!dio->wait_for_completion)
1882                         return -EIOCBQUEUED;
1883
1884                 for (;;) {
1885                         set_current_state(TASK_UNINTERRUPTIBLE);
1886                         if (!READ_ONCE(dio->submit.waiter))
1887                                 break;
1888
1889                         if (!(iocb->ki_flags & IOCB_HIPRI) ||
1890                             !dio->submit.last_queue ||
1891                             !blk_poll(dio->submit.last_queue,
1892                                          dio->submit.cookie))
1893                                 io_schedule();
1894                 }
1895                 __set_current_state(TASK_RUNNING);
1896         }
1897
1898         ret = iomap_dio_complete(dio);
1899
1900         return ret;
1901
1902 out_free_dio:
1903         kfree(dio);
1904         return ret;
1905 }
1906 EXPORT_SYMBOL_GPL(iomap_dio_rw);
1907
1908 /* Swapfile activation */
1909
1910 #ifdef CONFIG_SWAP
1911 struct iomap_swapfile_info {
1912         struct iomap iomap;             /* accumulated iomap */
1913         struct swap_info_struct *sis;
1914         uint64_t lowest_ppage;          /* lowest physical addr seen (pages) */
1915         uint64_t highest_ppage;         /* highest physical addr seen (pages) */
1916         unsigned long nr_pages;         /* number of pages collected */
1917         int nr_extents;                 /* extent count */
1918 };
1919
1920 /*
1921  * Collect physical extents for this swap file.  Physical extents reported to
1922  * the swap code must be trimmed to align to a page boundary.  The logical
1923  * offset within the file is irrelevant since the swapfile code maps logical
1924  * page numbers of the swap device to the physical page-aligned extents.
1925  */
1926 static int iomap_swapfile_add_extent(struct iomap_swapfile_info *isi)
1927 {
1928         struct iomap *iomap = &isi->iomap;
1929         unsigned long nr_pages;
1930         uint64_t first_ppage;
1931         uint64_t first_ppage_reported;
1932         uint64_t next_ppage;
1933         int error;
1934
1935         /*
1936          * Round the start up and the end down so that the physical
1937          * extent aligns to a page boundary.
1938          */
1939         first_ppage = ALIGN(iomap->addr, PAGE_SIZE) >> PAGE_SHIFT;
1940         next_ppage = ALIGN_DOWN(iomap->addr + iomap->length, PAGE_SIZE) >>
1941                         PAGE_SHIFT;
1942
1943         /* Skip too-short physical extents. */
1944         if (first_ppage >= next_ppage)
1945                 return 0;
1946         nr_pages = next_ppage - first_ppage;
1947
1948         /*
1949          * Calculate how much swap space we're adding; the first page contains
1950          * the swap header and doesn't count.  The mm still wants that first
1951          * page fed to add_swap_extent, however.
1952          */
1953         first_ppage_reported = first_ppage;
1954         if (iomap->offset == 0)
1955                 first_ppage_reported++;
1956         if (isi->lowest_ppage > first_ppage_reported)
1957                 isi->lowest_ppage = first_ppage_reported;
1958         if (isi->highest_ppage < (next_ppage - 1))
1959                 isi->highest_ppage = next_ppage - 1;
1960
1961         /* Add extent, set up for the next call. */
1962         error = add_swap_extent(isi->sis, isi->nr_pages, nr_pages, first_ppage);
1963         if (error < 0)
1964                 return error;
1965         isi->nr_extents += error;
1966         isi->nr_pages += nr_pages;
1967         return 0;
1968 }
1969
1970 /*
1971  * Accumulate iomaps for this swap file.  We have to accumulate iomaps because
1972  * swap only cares about contiguous page-aligned physical extents and makes no
1973  * distinction between written and unwritten extents.
1974  */
1975 static loff_t iomap_swapfile_activate_actor(struct inode *inode, loff_t pos,
1976                 loff_t count, void *data, struct iomap *iomap)
1977 {
1978         struct iomap_swapfile_info *isi = data;
1979         int error;
1980
1981         switch (iomap->type) {
1982         case IOMAP_MAPPED:
1983         case IOMAP_UNWRITTEN:
1984                 /* Only real or unwritten extents. */
1985                 break;
1986         case IOMAP_INLINE:
1987                 /* No inline data. */
1988                 pr_err("swapon: file is inline\n");
1989                 return -EINVAL;
1990         default:
1991                 pr_err("swapon: file has unallocated extents\n");
1992                 return -EINVAL;
1993         }
1994
1995         /* No uncommitted metadata or shared blocks. */
1996         if (iomap->flags & IOMAP_F_DIRTY) {
1997                 pr_err("swapon: file is not committed\n");
1998                 return -EINVAL;
1999         }
2000         if (iomap->flags & IOMAP_F_SHARED) {
2001                 pr_err("swapon: file has shared extents\n");
2002                 return -EINVAL;
2003         }
2004
2005         /* Only one bdev per swap file. */
2006         if (iomap->bdev != isi->sis->bdev) {
2007                 pr_err("swapon: file is on multiple devices\n");
2008                 return -EINVAL;
2009         }
2010
2011         if (isi->iomap.length == 0) {
2012                 /* No accumulated extent, so just store it. */
2013                 memcpy(&isi->iomap, iomap, sizeof(isi->iomap));
2014         } else if (isi->iomap.addr + isi->iomap.length == iomap->addr) {
2015                 /* Append this to the accumulated extent. */
2016                 isi->iomap.length += iomap->length;
2017         } else {
2018                 /* Otherwise, add the retained iomap and store this one. */
2019                 error = iomap_swapfile_add_extent(isi);
2020                 if (error)
2021                         return error;
2022                 memcpy(&isi->iomap, iomap, sizeof(isi->iomap));
2023         }
2024         return count;
2025 }
2026
2027 /*
2028  * Iterate a swap file's iomaps to construct physical extents that can be
2029  * passed to the swapfile subsystem.
2030  */
2031 int iomap_swapfile_activate(struct swap_info_struct *sis,
2032                 struct file *swap_file, sector_t *pagespan,
2033                 const struct iomap_ops *ops)
2034 {
2035         struct iomap_swapfile_info isi = {
2036                 .sis = sis,
2037                 .lowest_ppage = (sector_t)-1ULL,
2038         };
2039         struct address_space *mapping = swap_file->f_mapping;
2040         struct inode *inode = mapping->host;
2041         loff_t pos = 0;
2042         loff_t len = ALIGN_DOWN(i_size_read(inode), PAGE_SIZE);
2043         loff_t ret;
2044
2045         /*
2046          * Persist all file mapping metadata so that we won't have any
2047          * IOMAP_F_DIRTY iomaps.
2048          */
2049         ret = vfs_fsync(swap_file, 1);
2050         if (ret)
2051                 return ret;
2052
2053         while (len > 0) {
2054                 ret = iomap_apply(inode, pos, len, IOMAP_REPORT,
2055                                 ops, &isi, iomap_swapfile_activate_actor);
2056                 if (ret <= 0)
2057                         return ret;
2058
2059                 pos += ret;
2060                 len -= ret;
2061         }
2062
2063         if (isi.iomap.length) {
2064                 ret = iomap_swapfile_add_extent(&isi);
2065                 if (ret)
2066                         return ret;
2067         }
2068
2069         *pagespan = 1 + isi.highest_ppage - isi.lowest_ppage;
2070         sis->max = isi.nr_pages;
2071         sis->pages = isi.nr_pages - 1;
2072         sis->highest_bit = isi.nr_pages - 1;
2073         return isi.nr_extents;
2074 }
2075 EXPORT_SYMBOL_GPL(iomap_swapfile_activate);
2076 #endif /* CONFIG_SWAP */
2077
2078 static loff_t
2079 iomap_bmap_actor(struct inode *inode, loff_t pos, loff_t length,
2080                 void *data, struct iomap *iomap)
2081 {
2082         sector_t *bno = data, addr;
2083
2084         if (iomap->type == IOMAP_MAPPED) {
2085                 addr = (pos - iomap->offset + iomap->addr) >> inode->i_blkbits;
2086                 if (addr > INT_MAX)
2087                         WARN(1, "would truncate bmap result\n");
2088                 else
2089                         *bno = addr;
2090         }
2091         return 0;
2092 }
2093
2094 /* legacy ->bmap interface.  0 is the error return (!) */
2095 sector_t
2096 iomap_bmap(struct address_space *mapping, sector_t bno,
2097                 const struct iomap_ops *ops)
2098 {
2099         struct inode *inode = mapping->host;
2100         loff_t pos = bno << inode->i_blkbits;
2101         unsigned blocksize = i_blocksize(inode);
2102
2103         if (filemap_write_and_wait(mapping))
2104                 return 0;
2105
2106         bno = 0;
2107         iomap_apply(inode, pos, blocksize, 0, ops, &bno, iomap_bmap_actor);
2108         return bno;
2109 }
2110 EXPORT_SYMBOL_GPL(iomap_bmap);