fs/iomap.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2010 Red Hat, Inc.
   4  * Copyright (c) 2016-2018 Christoph Hellwig.
   5  */
   6 #include <linux/module.h>
   7 #include <linux/compiler.h>
   8 #include <linux/fs.h>
   9 #include <linux/iomap.h>
  10 #include <linux/uaccess.h>
  11 #include <linux/gfp.h>
  12 #include <linux/migrate.h>
  13 #include <linux/mm.h>
  14 #include <linux/mm_inline.h>
  15 #include <linux/swap.h>
  16 #include <linux/pagemap.h>
  17 #include <linux/pagevec.h>
  18 #include <linux/file.h>
  19 #include <linux/uio.h>
  20 #include <linux/backing-dev.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/task_io_accounting_ops.h>
  23 #include <linux/dax.h>
  24 #include <linux/sched/signal.h>
  25
  26 #include "internal.h"
  27
  28 /*
  29  * Execute a iomap write on a segment of the mapping that spans a
  30  * contiguous range of pages that have identical block mapping state.
  31  *
  32  * This avoids the need to map pages individually, do individual allocations
  33  * for each page and most importantly avoid the need for filesystem specific
  34  * locking per page. Instead, all the operations are amortised over the entire
  35  * range of pages. It is assumed that the filesystems will lock whatever
  36  * resources they require in the iomap_begin call, and release them in the
  37  * iomap_end call.
  38  */
  39 loff_t
  40 iomap_apply(struct inode *inode, loff_t pos, loff_t length, unsigned flags,
  41                 const struct iomap_ops *ops, void *data, iomap_actor_t actor)
  42 {
  43         struct iomap iomap = { 0 };
  44         loff_t written = 0, ret;
  45
  46         /*
  47          * Need to map a range from start position for length bytes. This can
  48          * span multiple pages - it is only guaranteed to return a range of a
  49          * single type of pages (e.g. all into a hole, all mapped or all
  50          * unwritten). Failure at this point has nothing to undo.
  51          *
  52          * If allocation is required for this range, reserve the space now so
  53          * that the allocation is guaranteed to succeed later on. Once we copy
  54          * the data into the page cache pages, then we cannot fail otherwise we
  55          * expose transient stale data. If the reserve fails, we can safely
  56          * back out at this point as there is nothing to undo.
  57          */
  58         ret = ops->iomap_begin(inode, pos, length, flags, &iomap);
  59         if (ret)
  60                 return ret;
  61         if (WARN_ON(iomap.offset > pos))
  62                 return -EIO;
  63         if (WARN_ON(iomap.length == 0))
  64                 return -EIO;
  65
  66         /*
  67          * Cut down the length to the one actually provided by the filesystem,
  68          * as it might not be able to give us the whole size that we requested.
  69          */
  70         if (iomap.offset + iomap.length < pos + length)
  71                 length = iomap.offset + iomap.length - pos;
  72
  73         /*
  74          * Now that we have guaranteed that the space allocation will succeed.
  75          * we can do the copy-in page by page without having to worry about
  76          * failures exposing transient data.
  77          */
  78         written = actor(inode, pos, length, data, &iomap);
  79
  80         /*
  81          * Now the data has been copied, commit the range we've copied.  This
  82          * should not fail unless the filesystem has had a fatal error.
  83          */
  84         if (ops->iomap_end) {
  85                 ret = ops->iomap_end(inode, pos, length,
  86                                      written > 0 ? written : 0,
  87                                      flags, &iomap);
  88         }
  89
  90         return written ? written : ret;
  91 }
  92
  93 static sector_t
  94 iomap_sector(struct iomap *iomap, loff_t pos)
  95 {
  96         return (iomap->addr + pos - iomap->offset) >> SECTOR_SHIFT;
  97 }
  98
  99 static struct iomap_page *
 100 iomap_page_create(struct inode *inode, struct page *page)
 101 {
 102         struct iomap_page *iop = to_iomap_page(page);
 103
 104         if (iop || i_blocksize(inode) == PAGE_SIZE)
 105                 return iop;
 106
 107         iop = kmalloc(sizeof(*iop), GFP_NOFS | __GFP_NOFAIL);
 108         atomic_set(&iop->read_count, 0);
 109         atomic_set(&iop->write_count, 0);
 110         bitmap_zero(iop->uptodate, PAGE_SIZE / SECTOR_SIZE);
 111
 112         /*
 113          * migrate_page_move_mapping() assumes that pages with private data have
 114          * their count elevated by 1.
 115          */
 116         get_page(page);
 117         set_page_private(page, (unsigned long)iop);
 118         SetPagePrivate(page);
 119         return iop;
 120 }
 121
 122 static void
 123 iomap_page_release(struct page *page)
 124 {
 125         struct iomap_page *iop = to_iomap_page(page);
 126
 127         if (!iop)
 128                 return;
 129         WARN_ON_ONCE(atomic_read(&iop->read_count));
 130         WARN_ON_ONCE(atomic_read(&iop->write_count));
 131         ClearPagePrivate(page);
 132         set_page_private(page, 0);
 133         put_page(page);
 134         kfree(iop);
 135 }
 136
 137 /*
 138  * Calculate the range inside the page that we actually need to read.
 139  */
 140 static void
 141 iomap_adjust_read_range(struct inode *inode, struct iomap_page *iop,
 142                 loff_t *pos, loff_t length, unsigned *offp, unsigned *lenp)
 143 {
 144         loff_t orig_pos = *pos;
 145         loff_t isize = i_size_read(inode);
 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
 153         /*
 154          * If the block size is smaller than the page size we need to check the
 155          * per-block uptodate status and adjust the offset and length if needed
 156          * to avoid reading in already uptodate ranges.
 157          */
 158         if (iop) {
 159                 unsigned int i;
 160
 161                 /* move forward for each leading block marked uptodate */
 162                 for (i = first; i <= last; i++) {
 163                         if (!test_bit(i, iop->uptodate))
 164                                 break;
 165                         *pos += block_size;
 166                         poff += block_size;
 167                         plen -= block_size;
 168                         first++;
 169                 }
 170
 171                 /* truncate len if we find any trailing uptodate block(s) */
 172                 for ( ; i <= last; i++) {
 173                         if (test_bit(i, iop->uptodate)) {
 174                                 plen -= (last - i + 1) * block_size;
 175                                 last = i - 1;
 176                                 break;
 177                         }
 178                 }
 179         }
 180
 181         /*
 182          * If the extent spans the block that contains the i_size we need to
 183          * handle both halves separately so that we properly zero data in the
 184          * page cache for blocks that are entirely outside of i_size.
 185          */
 186         if (orig_pos <= isize && orig_pos + length > isize) {
 187                 unsigned end = offset_in_page(isize - 1) >> block_bits;
 188
 189                 if (first <= end && last > end)
 190                         plen -= (last - end) * block_size;
 191         }
 192
 193         *offp = poff;
 194         *lenp = plen;
 195 }
 196
 197 static void
 198 iomap_set_range_uptodate(struct page *page, unsigned off, unsigned len)
 199 {
 200         struct iomap_page *iop = to_iomap_page(page);
 201         struct inode *inode = page->mapping->host;
 202         unsigned first = off >> inode->i_blkbits;
 203         unsigned last = (off + len - 1) >> inode->i_blkbits;
 204         unsigned int i;
 205         bool uptodate = true;
 206
 207         if (iop) {
 208                 for (i = 0; i < PAGE_SIZE / i_blocksize(inode); i++) {
 209                         if (i >= first && i <= last)
 210                                 set_bit(i, iop->uptodate);
 211                         else if (!test_bit(i, iop->uptodate))
 212                                 uptodate = false;
 213                 }
 214         }
 215
 216         if (uptodate && !PageError(page))
 217                 SetPageUptodate(page);
 218 }
 219
 220 static void
 221 iomap_read_finish(struct iomap_page *iop, struct page *page)
 222 {
 223         if (!iop || atomic_dec_and_test(&iop->read_count))
 224                 unlock_page(page);
 225 }
 226
 227 static void
 228 iomap_read_page_end_io(struct bio_vec *bvec, int error)
 229 {
 230         struct page *page = bvec->bv_page;
 231         struct iomap_page *iop = to_iomap_page(page);
 232
 233         if (unlikely(error)) {
 234                 ClearPageUptodate(page);
 235                 SetPageError(page);
 236         } else {
 237                 iomap_set_range_uptodate(page, bvec->bv_offset, bvec->bv_len);
 238         }
 239
 240         iomap_read_finish(iop, page);
 241 }
 242
 243 static void
 244 iomap_read_end_io(struct bio *bio)
 245 {
 246         int error = blk_status_to_errno(bio->bi_status);
 247         struct bio_vec *bvec;
 248         struct bvec_iter_all iter_all;
 249
 250         bio_for_each_segment_all(bvec, bio, iter_all)
 251                 iomap_read_page_end_io(bvec, error);
 252         bio_put(bio);
 253 }
 254
 255 struct iomap_readpage_ctx {
 256         struct page             *cur_page;
 257         bool                    cur_page_in_bio;
 258         bool                    is_readahead;
 259         struct bio              *bio;
 260         struct list_head        *pages;
 261 };
 262
 263 static void
 264 iomap_read_inline_data(struct inode *inode, struct page *page,
 265                 struct iomap *iomap)
 266 {
 267         size_t size = i_size_read(inode);
 268         void *addr;
 269
 270         if (PageUptodate(page))
 271                 return;
 272
 273         BUG_ON(page->index);
 274         BUG_ON(size > PAGE_SIZE - offset_in_page(iomap->inline_data));
 275
 276         addr = kmap_atomic(page);
 277         memcpy(addr, iomap->inline_data, size);
 278         memset(addr + size, 0, PAGE_SIZE - size);
 279         kunmap_atomic(addr);
 280         SetPageUptodate(page);
 281 }
 282
 283 static loff_t
 284 iomap_readpage_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 285                 struct iomap *iomap)
 286 {
 287         struct iomap_readpage_ctx *ctx = data;
 288         struct page *page = ctx->cur_page;
 289         struct iomap_page *iop = iomap_page_create(inode, page);
 290         bool same_page = false, is_contig = false;
 291         loff_t orig_pos = pos;
 292         unsigned poff, plen;
 293         sector_t sector;
 294
 295         if (iomap->type == IOMAP_INLINE) {
 296                 WARN_ON_ONCE(pos);
 297                 iomap_read_inline_data(inode, page, iomap);
 298                 return PAGE_SIZE;
 299         }
 300
 301         /* zero post-eof blocks as the page may be mapped */
 302         iomap_adjust_read_range(inode, iop, &pos, length, &poff, &plen);
 303         if (plen == 0)
 304                 goto done;
 305
 306         if (iomap->type != IOMAP_MAPPED || pos >= i_size_read(inode)) {
 307                 zero_user(page, poff, plen);
 308                 iomap_set_range_uptodate(page, poff, plen);
 309                 goto done;
 310         }
 311
 312         ctx->cur_page_in_bio = true;
 313
 314         /*
 315          * Try to merge into a previous segment if we can.
 316          */
 317         sector = iomap_sector(iomap, pos);
 318         if (ctx->bio && bio_end_sector(ctx->bio) == sector)
 319                 is_contig = true;
 320
 321         if (is_contig &&
 322             __bio_try_merge_page(ctx->bio, page, plen, poff, &same_page)) {
 323                 if (!same_page && iop)
 324                         atomic_inc(&iop->read_count);
 325                 goto done;
 326         }
 327
 328         /*
 329          * If we start a new segment we need to increase the read count, and we
 330          * need to do so before submitting any previous full bio to make sure
 331          * that we don't prematurely unlock the page.
 332          */
 333         if (iop)
 334                 atomic_inc(&iop->read_count);
 335
 336         if (!ctx->bio || !is_contig || bio_full(ctx->bio, plen)) {
 337                 gfp_t gfp = mapping_gfp_constraint(page->mapping, GFP_KERNEL);
 338                 int nr_vecs = (length + PAGE_SIZE - 1) >> PAGE_SHIFT;
 339
 340                 if (ctx->bio)
 341                         submit_bio(ctx->bio);
 342
 343                 if (ctx->is_readahead) /* same as readahead_gfp_mask */
 344                         gfp |= __GFP_NORETRY | __GFP_NOWARN;
 345                 ctx->bio = bio_alloc(gfp, min(BIO_MAX_PAGES, nr_vecs));
 346                 ctx->bio->bi_opf = REQ_OP_READ;
 347                 if (ctx->is_readahead)
 348                         ctx->bio->bi_opf |= REQ_RAHEAD;
 349                 ctx->bio->bi_iter.bi_sector = sector;
 350                 bio_set_dev(ctx->bio, iomap->bdev);
 351                 ctx->bio->bi_end_io = iomap_read_end_io;
 352         }
 353
 354         bio_add_page(ctx->bio, page, plen, poff);
 355 done:
 356         /*
 357          * Move the caller beyond our range so that it keeps making progress.
 358          * For that we have to include any leading non-uptodate ranges, but
 359          * we can skip trailing ones as they will be handled in the next
 360          * iteration.
 361          */
 362         return pos - orig_pos + plen;
 363 }
 364
 365 int
 366 iomap_readpage(struct page *page, const struct iomap_ops *ops)
 367 {
 368         struct iomap_readpage_ctx ctx = { .cur_page = page };
 369         struct inode *inode = page->mapping->host;
 370         unsigned poff;
 371         loff_t ret;
 372
 373         for (poff = 0; poff < PAGE_SIZE; poff += ret) {
 374                 ret = iomap_apply(inode, page_offset(page) + poff,
 375                                 PAGE_SIZE - poff, 0, ops, &ctx,
 376                                 iomap_readpage_actor);
 377                 if (ret <= 0) {
 378                         WARN_ON_ONCE(ret == 0);
 379                         SetPageError(page);
 380                         break;
 381                 }
 382         }
 383
 384         if (ctx.bio) {
 385                 submit_bio(ctx.bio);
 386                 WARN_ON_ONCE(!ctx.cur_page_in_bio);
 387         } else {
 388                 WARN_ON_ONCE(ctx.cur_page_in_bio);
 389                 unlock_page(page);
 390         }
 391
 392         /*
 393          * Just like mpage_readpages and block_read_full_page we always
 394          * return 0 and just mark the page as PageError on errors.  This
 395          * should be cleaned up all through the stack eventually.
 396          */
 397         return 0;
 398 }
 399 EXPORT_SYMBOL_GPL(iomap_readpage);
 400
 401 static struct page *
 402 iomap_next_page(struct inode *inode, struct list_head *pages, loff_t pos,
 403                 loff_t length, loff_t *done)
 404 {
 405         while (!list_empty(pages)) {
 406                 struct page *page = lru_to_page(pages);
 407
 408                 if (page_offset(page) >= (u64)pos + length)
 409                         break;
 410
 411                 list_del(&page->lru);
 412                 if (!add_to_page_cache_lru(page, inode->i_mapping, page->index,
 413                                 GFP_NOFS))
 414                         return page;
 415
 416                 /*
 417                  * If we already have a page in the page cache at index we are
 418                  * done.  Upper layers don't care if it is uptodate after the
 419                  * readpages call itself as every page gets checked again once
 420                  * actually needed.
 421                  */
 422                 *done += PAGE_SIZE;
 423                 put_page(page);
 424         }
 425
 426         return NULL;
 427 }
 428
 429 static loff_t
 430 iomap_readpages_actor(struct inode *inode, loff_t pos, loff_t length,
 431                 void *data, struct iomap *iomap)
 432 {
 433         struct iomap_readpage_ctx *ctx = data;
 434         loff_t done, ret;
 435
 436         for (done = 0; done < length; done += ret) {
 437                 if (ctx->cur_page && offset_in_page(pos + done) == 0) {
 438                         if (!ctx->cur_page_in_bio)
 439                                 unlock_page(ctx->cur_page);
 440                         put_page(ctx->cur_page);
 441                         ctx->cur_page = NULL;
 442                 }
 443                 if (!ctx->cur_page) {
 444                         ctx->cur_page = iomap_next_page(inode, ctx->pages,
 445                                         pos, length, &done);
 446                         if (!ctx->cur_page)
 447                                 break;
 448                         ctx->cur_page_in_bio = false;
 449                 }
 450                 ret = iomap_readpage_actor(inode, pos + done, length - done,
 451                                 ctx, iomap);
 452         }
 453
 454         return done;
 455 }
 456
 457 int
 458 iomap_readpages(struct address_space *mapping, struct list_head *pages,
 459                 unsigned nr_pages, const struct iomap_ops *ops)
 460 {
 461         struct iomap_readpage_ctx ctx = {
 462                 .pages          = pages,
 463                 .is_readahead   = true,
 464         };
 465         loff_t pos = page_offset(list_entry(pages->prev, struct page, lru));
 466         loff_t last = page_offset(list_entry(pages->next, struct page, lru));
 467         loff_t length = last - pos + PAGE_SIZE, ret = 0;
 468
 469         while (length > 0) {
 470                 ret = iomap_apply(mapping->host, pos, length, 0, ops,
 471                                 &ctx, iomap_readpages_actor);
 472                 if (ret <= 0) {
 473                         WARN_ON_ONCE(ret == 0);
 474                         goto done;
 475                 }
 476                 pos += ret;
 477                 length -= ret;
 478         }
 479         ret = 0;
 480 done:
 481         if (ctx.bio)
 482                 submit_bio(ctx.bio);
 483         if (ctx.cur_page) {
 484                 if (!ctx.cur_page_in_bio)
 485                         unlock_page(ctx.cur_page);
 486                 put_page(ctx.cur_page);
 487         }
 488
 489         /*
 490          * Check that we didn't lose a page due to the arcance calling
 491          * conventions..
 492          */
 493         WARN_ON_ONCE(!ret && !list_empty(ctx.pages));
 494         return ret;
 495 }
 496 EXPORT_SYMBOL_GPL(iomap_readpages);
 497
 498 /*
 499  * iomap_is_partially_uptodate checks whether blocks within a page are
 500  * uptodate or not.
 501  *
 502  * Returns true if all blocks which correspond to a file portion
 503  * we want to read within the page are uptodate.
 504  */
 505 int
 506 iomap_is_partially_uptodate(struct page *page, unsigned long from,
 507                 unsigned long count)
 508 {
 509         struct iomap_page *iop = to_iomap_page(page);
 510         struct inode *inode = page->mapping->host;
 511         unsigned len, first, last;
 512         unsigned i;
 513
 514         /* Limit range to one page */
 515         len = min_t(unsigned, PAGE_SIZE - from, count);
 516
 517         /* First and last blocks in range within page */
 518         first = from >> inode->i_blkbits;
 519         last = (from + len - 1) >> inode->i_blkbits;
 520
 521         if (iop) {
 522                 for (i = first; i <= last; i++)
 523                         if (!test_bit(i, iop->uptodate))
 524                                 return 0;
 525                 return 1;
 526         }
 527
 528         return 0;
 529 }
 530 EXPORT_SYMBOL_GPL(iomap_is_partially_uptodate);
 531
 532 int
 533 iomap_releasepage(struct page *page, gfp_t gfp_mask)
 534 {
 535         /*
 536          * mm accommodates an old ext3 case where clean pages might not have had
 537          * the dirty bit cleared. Thus, it can send actual dirty pages to
 538          * ->releasepage() via shrink_active_list(), skip those here.
 539          */
 540         if (PageDirty(page) || PageWriteback(page))
 541                 return 0;
 542         iomap_page_release(page);
 543         return 1;
 544 }
 545 EXPORT_SYMBOL_GPL(iomap_releasepage);
 546
 547 void
 548 iomap_invalidatepage(struct page *page, unsigned int offset, unsigned int len)
 549 {
 550         /*
 551          * If we are invalidating the entire page, clear the dirty state from it
 552          * and release it to avoid unnecessary buildup of the LRU.
 553          */
 554         if (offset == 0 && len == PAGE_SIZE) {
 555                 WARN_ON_ONCE(PageWriteback(page));
 556                 cancel_dirty_page(page);
 557                 iomap_page_release(page);
 558         }
 559 }
 560 EXPORT_SYMBOL_GPL(iomap_invalidatepage);
 561
 562 #ifdef CONFIG_MIGRATION
 563 int
 564 iomap_migrate_page(struct address_space *mapping, struct page *newpage,
 565                 struct page *page, enum migrate_mode mode)
 566 {
 567         int ret;
 568
 569         ret = migrate_page_move_mapping(mapping, newpage, page, mode, 0);
 570         if (ret != MIGRATEPAGE_SUCCESS)
 571                 return ret;
 572
 573         if (page_has_private(page)) {
 574                 ClearPagePrivate(page);
 575                 get_page(newpage);
 576                 set_page_private(newpage, page_private(page));
 577                 set_page_private(page, 0);
 578                 put_page(page);
 579                 SetPagePrivate(newpage);
 580         }
 581
 582         if (mode != MIGRATE_SYNC_NO_COPY)
 583                 migrate_page_copy(newpage, page);
 584         else
 585                 migrate_page_states(newpage, page);
 586         return MIGRATEPAGE_SUCCESS;
 587 }
 588 EXPORT_SYMBOL_GPL(iomap_migrate_page);
 589 #endif /* CONFIG_MIGRATION */
 590
 591 static void
 592 iomap_write_failed(struct inode *inode, loff_t pos, unsigned len)
 593 {
 594         loff_t i_size = i_size_read(inode);
 595
 596         /*
 597          * Only truncate newly allocated pages beyoned EOF, even if the
 598          * write started inside the existing inode size.
 599          */
 600         if (pos + len > i_size)
 601                 truncate_pagecache_range(inode, max(pos, i_size), pos + len);
 602 }
 603
 604 static int
 605 iomap_read_page_sync(struct inode *inode, loff_t block_start, struct page *page,
 606                 unsigned poff, unsigned plen, unsigned from, unsigned to,
 607                 struct iomap *iomap)
 608 {
 609         struct bio_vec bvec;
 610         struct bio bio;
 611
 612         if (iomap->type != IOMAP_MAPPED || block_start >= i_size_read(inode)) {
 613                 zero_user_segments(page, poff, from, to, poff + plen);
 614                 iomap_set_range_uptodate(page, poff, plen);
 615                 return 0;
 616         }
 617
 618         bio_init(&bio, &bvec, 1);
 619         bio.bi_opf = REQ_OP_READ;
 620         bio.bi_iter.bi_sector = iomap_sector(iomap, block_start);
 621         bio_set_dev(&bio, iomap->bdev);
 622         __bio_add_page(&bio, page, plen, poff);
 623         return submit_bio_wait(&bio);
 624 }
 625
 626 static int
 627 __iomap_write_begin(struct inode *inode, loff_t pos, unsigned len,
 628                 struct page *page, struct iomap *iomap)
 629 {
 630         struct iomap_page *iop = iomap_page_create(inode, page);
 631         loff_t block_size = i_blocksize(inode);
 632         loff_t block_start = pos & ~(block_size - 1);
 633         loff_t block_end = (pos + len + block_size - 1) & ~(block_size - 1);
 634         unsigned from = offset_in_page(pos), to = from + len, poff, plen;
 635         int status = 0;
 636
 637         if (PageUptodate(page))
 638                 return 0;
 639
 640         do {
 641                 iomap_adjust_read_range(inode, iop, &block_start,
 642                                 block_end - block_start, &poff, &plen);
 643                 if (plen == 0)
 644                         break;
 645
 646                 if ((from > poff && from < poff + plen) ||
 647                     (to > poff && to < poff + plen)) {
 648                         status = iomap_read_page_sync(inode, block_start, page,
 649                                         poff, plen, from, to, iomap);
 650                         if (status)
 651                                 break;
 652                 }
 653
 654         } while ((block_start += plen) < block_end);
 655
 656         return status;
 657 }
 658
 659 static int
 660 iomap_write_begin(struct inode *inode, loff_t pos, unsigned len, unsigned flags,
 661                 struct page **pagep, struct iomap *iomap)
 662 {
 663         const struct iomap_page_ops *page_ops = iomap->page_ops;
 664         pgoff_t index = pos >> PAGE_SHIFT;
 665         struct page *page;
 666         int status = 0;
 667
 668         BUG_ON(pos + len > iomap->offset + iomap->length);
 669
 670         if (fatal_signal_pending(current))
 671                 return -EINTR;
 672
 673         if (page_ops && page_ops->page_prepare) {
 674                 status = page_ops->page_prepare(inode, pos, len, iomap);
 675                 if (status)
 676                         return status;
 677         }
 678
 679         page = grab_cache_page_write_begin(inode->i_mapping, index, flags);
 680         if (!page) {
 681                 status = -ENOMEM;
 682                 goto out_no_page;
 683         }
 684
 685         if (iomap->type == IOMAP_INLINE)
 686                 iomap_read_inline_data(inode, page, iomap);
 687         else if (iomap->flags & IOMAP_F_BUFFER_HEAD)
 688                 status = __block_write_begin_int(page, pos, len, NULL, iomap);
 689         else
 690                 status = __iomap_write_begin(inode, pos, len, page, iomap);
 691
 692         if (unlikely(status))
 693                 goto out_unlock;
 694
 695         *pagep = page;
 696         return 0;
 697
 698 out_unlock:
 699         unlock_page(page);
 700         put_page(page);
 701         iomap_write_failed(inode, pos, len);
 702
 703 out_no_page:
 704         if (page_ops && page_ops->page_done)
 705                 page_ops->page_done(inode, pos, 0, NULL, iomap);
 706         return status;
 707 }
 708
 709 int
 710 iomap_set_page_dirty(struct page *page)
 711 {
 712         struct address_space *mapping = page_mapping(page);
 713         int newly_dirty;
 714
 715         if (unlikely(!mapping))
 716                 return !TestSetPageDirty(page);
 717
 718         /*
 719          * Lock out page->mem_cgroup migration to keep PageDirty
 720          * synchronized with per-memcg dirty page counters.
 721          */
 722         lock_page_memcg(page);
 723         newly_dirty = !TestSetPageDirty(page);
 724         if (newly_dirty)
 725                 __set_page_dirty(page, mapping, 0);
 726         unlock_page_memcg(page);
 727
 728         if (newly_dirty)
 729                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 730         return newly_dirty;
 731 }
 732 EXPORT_SYMBOL_GPL(iomap_set_page_dirty);
 733
 734 static int
 735 __iomap_write_end(struct inode *inode, loff_t pos, unsigned len,
 736                 unsigned copied, struct page *page, struct iomap *iomap)
 737 {
 738         flush_dcache_page(page);
 739
 740         /*
 741          * The blocks that were entirely written will now be uptodate, so we
 742          * don't have to worry about a readpage reading them and overwriting a
 743          * partial write.  However if we have encountered a short write and only
 744          * partially written into a block, it will not be marked uptodate, so a
 745          * readpage might come in and destroy our partial write.
 746          *
 747          * Do the simplest thing, and just treat any short write to a non
 748          * uptodate page as a zero-length write, and force the caller to redo
 749          * the whole thing.
 750          */
 751         if (unlikely(copied < len && !PageUptodate(page)))
 752                 return 0;
 753         iomap_set_range_uptodate(page, offset_in_page(pos), len);
 754         iomap_set_page_dirty(page);
 755         return copied;
 756 }
 757
 758 static int
 759 iomap_write_end_inline(struct inode *inode, struct page *page,
 760                 struct iomap *iomap, loff_t pos, unsigned copied)
 761 {
 762         void *addr;
 763
 764         WARN_ON_ONCE(!PageUptodate(page));
 765         BUG_ON(pos + copied > PAGE_SIZE - offset_in_page(iomap->inline_data));
 766
 767         addr = kmap_atomic(page);
 768         memcpy(iomap->inline_data + pos, addr + pos, copied);
 769         kunmap_atomic(addr);
 770
 771         mark_inode_dirty(inode);
 772         return copied;
 773 }
 774
 775 static int
 776 iomap_write_end(struct inode *inode, loff_t pos, unsigned len,
 777                 unsigned copied, struct page *page, struct iomap *iomap)
 778 {
 779         const struct iomap_page_ops *page_ops = iomap->page_ops;
 780         loff_t old_size = inode->i_size;
 781         int ret;
 782
 783         if (iomap->type == IOMAP_INLINE) {
 784                 ret = iomap_write_end_inline(inode, page, iomap, pos, copied);
 785         } else if (iomap->flags & IOMAP_F_BUFFER_HEAD) {
 786                 ret = block_write_end(NULL, inode->i_mapping, pos, len, copied,
 787                                 page, NULL);
 788         } else {
 789                 ret = __iomap_write_end(inode, pos, len, copied, page, iomap);
 790         }
 791
 792         /*
 793          * Update the in-memory inode size after copying the data into the page
 794          * cache.  It's up to the file system to write the updated size to disk,
 795          * preferably after I/O completion so that no stale data is exposed.
 796          */
 797         if (pos + ret > old_size) {
 798                 i_size_write(inode, pos + ret);
 799                 iomap->flags |= IOMAP_F_SIZE_CHANGED;
 800         }
 801         unlock_page(page);
 802
 803         if (old_size < pos)
 804                 pagecache_isize_extended(inode, old_size, pos);
 805         if (page_ops && page_ops->page_done)
 806                 page_ops->page_done(inode, pos, ret, page, iomap);
 807         put_page(page);
 808
 809         if (ret < len)
 810                 iomap_write_failed(inode, pos, len);
 811         return ret;
 812 }
 813
 814 static loff_t
 815 iomap_write_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 816                 struct iomap *iomap)
 817 {
 818         struct iov_iter *i = data;
 819         long status = 0;
 820         ssize_t written = 0;
 821         unsigned int flags = AOP_FLAG_NOFS;
 822
 823         do {
 824                 struct page *page;
 825                 unsigned long offset;   /* Offset into pagecache page */
 826                 unsigned long bytes;    /* Bytes to write to page */
 827                 size_t copied;          /* Bytes copied from user */
 828
 829                 offset = offset_in_page(pos);
 830                 bytes = min_t(unsigned long, PAGE_SIZE - offset,
 831                                                 iov_iter_count(i));
 832 again:
 833                 if (bytes > length)
 834                         bytes = length;
 835
 836                 /*
 837                  * Bring in the user page that we will copy from _first_.
 838                  * Otherwise there's a nasty deadlock on copying from the
 839                  * same page as we're writing to, without it being marked
 840                  * up-to-date.
 841                  *
 842                  * Not only is this an optimisation, but it is also required
 843                  * to check that the address is actually valid, when atomic
 844                  * usercopies are used, below.
 845                  */
 846                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
 847                         status = -EFAULT;
 848                         break;
 849                 }
 850
 851                 status = iomap_write_begin(inode, pos, bytes, flags, &page,
 852                                 iomap);
 853                 if (unlikely(status))
 854                         break;
 855
 856                 if (mapping_writably_mapped(inode->i_mapping))
 857                         flush_dcache_page(page);
 858
 859                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
 860
 861                 flush_dcache_page(page);
 862
 863                 status = iomap_write_end(inode, pos, bytes, copied, page,
 864                                 iomap);
 865                 if (unlikely(status < 0))
 866                         break;
 867                 copied = status;
 868
 869                 cond_resched();
 870
 871                 iov_iter_advance(i, copied);
 872                 if (unlikely(copied == 0)) {
 873                         /*
 874                          * If we were unable to copy any data at all, we must
 875                          * fall back to a single segment length write.
 876                          *
 877                          * If we didn't fallback here, we could livelock
 878                          * because not all segments in the iov can be copied at
 879                          * once without a pagefault.
 880                          */
 881                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
 882                                                 iov_iter_single_seg_count(i));
 883                         goto again;
 884                 }
 885                 pos += copied;
 886                 written += copied;
 887                 length -= copied;
 888
 889                 balance_dirty_pages_ratelimited(inode->i_mapping);
 890         } while (iov_iter_count(i) && length);
 891
 892         return written ? written : status;
 893 }
 894
 895 ssize_t
 896 iomap_file_buffered_write(struct kiocb *iocb, struct iov_iter *iter,
 897                 const struct iomap_ops *ops)
 898 {
 899         struct inode *inode = iocb->ki_filp->f_mapping->host;
 900         loff_t pos = iocb->ki_pos, ret = 0, written = 0;
 901
 902         while (iov_iter_count(iter)) {
 903                 ret = iomap_apply(inode, pos, iov_iter_count(iter),
 904                                 IOMAP_WRITE, ops, iter, iomap_write_actor);
 905                 if (ret <= 0)
 906                         break;
 907                 pos += ret;
 908                 written += ret;
 909         }
 910
 911         return written ? written : ret;
 912 }
 913 EXPORT_SYMBOL_GPL(iomap_file_buffered_write);
 914
 915 static struct page *
 916 __iomap_read_page(struct inode *inode, loff_t offset)
 917 {
 918         struct address_space *mapping = inode->i_mapping;
 919         struct page *page;
 920
 921         page = read_mapping_page(mapping, offset >> PAGE_SHIFT, NULL);
 922         if (IS_ERR(page))
 923                 return page;
 924         if (!PageUptodate(page)) {
 925                 put_page(page);
 926                 return ERR_PTR(-EIO);
 927         }
 928         return page;
 929 }
 930
 931 static loff_t
 932 iomap_dirty_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 933                 struct iomap *iomap)
 934 {
 935         long status = 0;
 936         ssize_t written = 0;
 937
 938         do {
 939                 struct page *page, *rpage;
 940                 unsigned long offset;   /* Offset into pagecache page */
 941                 unsigned long bytes;    /* Bytes to write to page */
 942
 943                 offset = offset_in_page(pos);
 944                 bytes = min_t(loff_t, PAGE_SIZE - offset, length);
 945
 946                 rpage = __iomap_read_page(inode, pos);
 947                 if (IS_ERR(rpage))
 948                         return PTR_ERR(rpage);
 949
 950                 status = iomap_write_begin(inode, pos, bytes,
 951                                            AOP_FLAG_NOFS, &page, iomap);
 952                 put_page(rpage);
 953                 if (unlikely(status))
 954                         return status;
 955
 956                 WARN_ON_ONCE(!PageUptodate(page));
 957
 958                 status = iomap_write_end(inode, pos, bytes, bytes, page, iomap);
 959                 if (unlikely(status <= 0)) {
 960                         if (WARN_ON_ONCE(status == 0))
 961                                 return -EIO;
 962                         return status;
 963                 }
 964
 965                 cond_resched();
 966
 967                 pos += status;
 968                 written += status;
 969                 length -= status;
 970
 971                 balance_dirty_pages_ratelimited(inode->i_mapping);
 972         } while (length);
 973
 974         return written;
 975 }
 976
 977 int
 978 iomap_file_dirty(struct inode *inode, loff_t pos, loff_t len,
 979                 const struct iomap_ops *ops)
 980 {
 981         loff_t ret;
 982
 983         while (len) {
 984                 ret = iomap_apply(inode, pos, len, IOMAP_WRITE, ops, NULL,
 985                                 iomap_dirty_actor);
 986                 if (ret <= 0)
 987                         return ret;
 988                 pos += ret;
 989                 len -= ret;
 990         }
 991
 992         return 0;
 993 }
 994 EXPORT_SYMBOL_GPL(iomap_file_dirty);
 995
 996 static int iomap_zero(struct inode *inode, loff_t pos, unsigned offset,
 997                 unsigned bytes, struct iomap *iomap)
 998 {
 999         struct page *page;
1000         int status;
1001
1002         status = iomap_write_begin(inode, pos, bytes, AOP_FLAG_NOFS, &page,
1003                                    iomap);
1004         if (status)
1005                 return status;
1006
1007         zero_user(page, offset, bytes);
1008         mark_page_accessed(page);
1009
1010         return iomap_write_end(inode, pos, bytes, bytes, page, iomap);
1011 }
1012
1013 static int iomap_dax_zero(loff_t pos, unsigned offset, unsigned bytes,
1014                 struct iomap *iomap)
1015 {
1016         return __dax_zero_page_range(iomap->bdev, iomap->dax_dev,
1017                         iomap_sector(iomap, pos & PAGE_MASK), offset, bytes);
1018 }
1019
1020 static loff_t
1021 iomap_zero_range_actor(struct inode *inode, loff_t pos, loff_t count,
1022                 void *data, struct iomap *iomap)
1023 {
1024         bool *did_zero = data;
1025         loff_t written = 0;
1026         int status;
1027
1028         /* already zeroed?  we're done. */
1029         if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1030                 return count;
1031
1032         do {
1033                 unsigned offset, bytes;
1034
1035                 offset = offset_in_page(pos);
1036                 bytes = min_t(loff_t, PAGE_SIZE - offset, count);
1037
1038                 if (IS_DAX(inode))
1039                         status = iomap_dax_zero(pos, offset, bytes, iomap);
1040                 else
1041                         status = iomap_zero(inode, pos, offset, bytes, iomap);
1042                 if (status < 0)
1043                         return status;
1044
1045                 pos += bytes;
1046                 count -= bytes;
1047                 written += bytes;
1048                 if (did_zero)
1049                         *did_zero = true;
1050         } while (count > 0);
1051
1052         return written;
1053 }
1054
1055 int
1056 iomap_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
1057                 const struct iomap_ops *ops)
1058 {
1059         loff_t ret;
1060
1061         while (len > 0) {
1062                 ret = iomap_apply(inode, pos, len, IOMAP_ZERO,
1063                                 ops, did_zero, iomap_zero_range_actor);
1064                 if (ret <= 0)
1065                         return ret;
1066
1067                 pos += ret;
1068                 len -= ret;
1069         }
1070
1071         return 0;
1072 }
1073 EXPORT_SYMBOL_GPL(iomap_zero_range);
1074
1075 int
1076 iomap_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
1077                 const struct iomap_ops *ops)
1078 {
1079         unsigned int blocksize = i_blocksize(inode);
1080         unsigned int off = pos & (blocksize - 1);
1081
1082         /* Block boundary? Nothing to do */
1083         if (!off)
1084                 return 0;
1085         return iomap_zero_range(inode, pos, blocksize - off, did_zero, ops);
1086 }
1087 EXPORT_SYMBOL_GPL(iomap_truncate_page);
1088
1089 static loff_t
1090 iomap_page_mkwrite_actor(struct inode *inode, loff_t pos, loff_t length,
1091                 void *data, struct iomap *iomap)
1092 {
1093         struct page *page = data;
1094         int ret;
1095
1096         if (iomap->flags & IOMAP_F_BUFFER_HEAD) {
1097                 ret = __block_write_begin_int(page, pos, length, NULL, iomap);
1098                 if (ret)
1099                         return ret;
1100                 block_commit_write(page, 0, length);
1101         } else {
1102                 WARN_ON_ONCE(!PageUptodate(page));
1103                 iomap_page_create(inode, page);
1104                 set_page_dirty(page);
1105         }
1106
1107         return length;
1108 }
1109
1110 vm_fault_t iomap_page_mkwrite(struct vm_fault *vmf, const struct iomap_ops *ops)
1111 {
1112         struct page *page = vmf->page;
1113         struct inode *inode = file_inode(vmf->vma->vm_file);
1114         unsigned long length;
1115         loff_t offset, size;
1116         ssize_t ret;
1117
1118         lock_page(page);
1119         size = i_size_read(inode);
1120         if ((page->mapping != inode->i_mapping) ||
1121             (page_offset(page) > size)) {
1122                 /* We overload EFAULT to mean page got truncated */
1123                 ret = -EFAULT;
1124                 goto out_unlock;
1125         }
1126
1127         /* page is wholly or partially inside EOF */
1128         if (((page->index + 1) << PAGE_SHIFT) > size)
1129                 length = offset_in_page(size);
1130         else
1131                 length = PAGE_SIZE;
1132
1133         offset = page_offset(page);
1134         while (length > 0) {
1135                 ret = iomap_apply(inode, offset, length,
1136                                 IOMAP_WRITE | IOMAP_FAULT, ops, page,
1137                                 iomap_page_mkwrite_actor);
1138                 if (unlikely(ret <= 0))
1139                         goto out_unlock;
1140                 offset += ret;
1141                 length -= ret;
1142         }
1143
1144         wait_for_stable_page(page);
1145         return VM_FAULT_LOCKED;
1146 out_unlock:
1147         unlock_page(page);
1148         return block_page_mkwrite_return(ret);
1149 }
1150 EXPORT_SYMBOL_GPL(iomap_page_mkwrite);
1151
1152 struct fiemap_ctx {
1153         struct fiemap_extent_info *fi;
1154         struct iomap prev;
1155 };
1156
1157 static int iomap_to_fiemap(struct fiemap_extent_info *fi,
1158                 struct iomap *iomap, u32 flags)
1159 {
1160         switch (iomap->type) {
1161         case IOMAP_HOLE:
1162                 /* skip holes */
1163                 return 0;
1164         case IOMAP_DELALLOC:
1165                 flags |= FIEMAP_EXTENT_DELALLOC | FIEMAP_EXTENT_UNKNOWN;
1166                 break;
1167         case IOMAP_MAPPED:
1168                 break;
1169         case IOMAP_UNWRITTEN:
1170                 flags |= FIEMAP_EXTENT_UNWRITTEN;
1171                 break;
1172         case IOMAP_INLINE:
1173                 flags |= FIEMAP_EXTENT_DATA_INLINE;
1174                 break;
1175         }
1176
1177         if (iomap->flags & IOMAP_F_MERGED)
1178                 flags |= FIEMAP_EXTENT_MERGED;
1179         if (iomap->flags & IOMAP_F_SHARED)
1180                 flags |= FIEMAP_EXTENT_SHARED;
1181
1182         return fiemap_fill_next_extent(fi, iomap->offset,
1183                         iomap->addr != IOMAP_NULL_ADDR ? iomap->addr : 0,
1184                         iomap->length, flags);
1185 }
1186
1187 static loff_t
1188 iomap_fiemap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
1189                 struct iomap *iomap)
1190 {
1191         struct fiemap_ctx *ctx = data;
1192         loff_t ret = length;
1193
1194         if (iomap->type == IOMAP_HOLE)
1195                 return length;
1196
1197         ret = iomap_to_fiemap(ctx->fi, &ctx->prev, 0);
1198         ctx->prev = *iomap;
1199         switch (ret) {
1200         case 0:         /* success */
1201                 return length;
1202         case 1:         /* extent array full */
1203                 return 0;
1204         default:
1205                 return ret;
1206         }
1207 }
1208
1209 int iomap_fiemap(struct inode *inode, struct fiemap_extent_info *fi,
1210                 loff_t start, loff_t len, const struct iomap_ops *ops)
1211 {
1212         struct fiemap_ctx ctx;
1213         loff_t ret;
1214
1215         memset(&ctx, 0, sizeof(ctx));
1216         ctx.fi = fi;
1217         ctx.prev.type = IOMAP_HOLE;
1218
1219         ret = fiemap_check_flags(fi, FIEMAP_FLAG_SYNC);
1220         if (ret)
1221                 return ret;
1222
1223         if (fi->fi_flags & FIEMAP_FLAG_SYNC) {
1224                 ret = filemap_write_and_wait(inode->i_mapping);
1225                 if (ret)
1226                         return ret;
1227         }
1228
1229         while (len > 0) {
1230                 ret = iomap_apply(inode, start, len, IOMAP_REPORT, ops, &ctx,
1231                                 iomap_fiemap_actor);
1232                 /* inode with no (attribute) mapping will give ENOENT */
1233                 if (ret == -ENOENT)
1234                         break;
1235                 if (ret < 0)
1236                         return ret;
1237                 if (ret == 0)
1238                         break;
1239
1240                 start += ret;
1241                 len -= ret;
1242         }
1243
1244         if (ctx.prev.type != IOMAP_HOLE) {
1245                 ret = iomap_to_fiemap(fi, &ctx.prev, FIEMAP_EXTENT_LAST);
1246                 if (ret < 0)
1247                         return ret;
1248         }
1249
1250         return 0;
1251 }
1252 EXPORT_SYMBOL_GPL(iomap_fiemap);
1253
1254 /*
1255  * Seek for SEEK_DATA / SEEK_HOLE within @page, starting at @lastoff.
1256  * Returns true if found and updates @lastoff to the offset in file.
1257  */
1258 static bool
1259 page_seek_hole_data(struct inode *inode, struct page *page, loff_t *lastoff,
1260                 int whence)
1261 {
1262         const struct address_space_operations *ops = inode->i_mapping->a_ops;
1263         unsigned int bsize = i_blocksize(inode), off;
1264         bool seek_data = whence == SEEK_DATA;
1265         loff_t poff = page_offset(page);
1266
1267         if (WARN_ON_ONCE(*lastoff >= poff + PAGE_SIZE))
1268                 return false;
1269
1270         if (*lastoff < poff) {
1271                 /*
1272                  * Last offset smaller than the start of the page means we found
1273                  * a hole:
1274                  */
1275                 if (whence == SEEK_HOLE)
1276                         return true;
1277                 *lastoff = poff;
1278         }
1279
1280         /*
1281          * Just check the page unless we can and should check block ranges:
1282          */
1283         if (bsize == PAGE_SIZE || !ops->is_partially_uptodate)
1284                 return PageUptodate(page) == seek_data;
1285
1286         lock_page(page);
1287         if (unlikely(page->mapping != inode->i_mapping))
1288                 goto out_unlock_not_found;
1289
1290         for (off = 0; off < PAGE_SIZE; off += bsize) {
1291                 if (offset_in_page(*lastoff) >= off + bsize)
1292                         continue;
1293                 if (ops->is_partially_uptodate(page, off, bsize) == seek_data) {
1294                         unlock_page(page);
1295                         return true;
1296                 }
1297                 *lastoff = poff + off + bsize;
1298         }
1299
1300 out_unlock_not_found:
1301         unlock_page(page);
1302         return false;
1303 }
1304
1305 /*
1306  * Seek for SEEK_DATA / SEEK_HOLE in the page cache.
1307  *
1308  * Within unwritten extents, the page cache determines which parts are holes
1309  * and which are data: uptodate buffer heads count as data; everything else
1310  * counts as a hole.
1311  *
1312  * Returns the resulting offset on successs, and -ENOENT otherwise.
1313  */
1314 static loff_t
1315 page_cache_seek_hole_data(struct inode *inode, loff_t offset, loff_t length,
1316                 int whence)
1317 {
1318         pgoff_t index = offset >> PAGE_SHIFT;
1319         pgoff_t end = DIV_ROUND_UP(offset + length, PAGE_SIZE);
1320         loff_t lastoff = offset;
1321         struct pagevec pvec;
1322
1323         if (length <= 0)
1324                 return -ENOENT;
1325
1326         pagevec_init(&pvec);
1327
1328         do {
1329                 unsigned nr_pages, i;
1330
1331                 nr_pages = pagevec_lookup_range(&pvec, inode->i_mapping, &index,
1332                                                 end - 1);
1333                 if (nr_pages == 0)
1334                         break;
1335
1336                 for (i = 0; i < nr_pages; i++) {
1337                         struct page *page = pvec.pages[i];
1338
1339                         if (page_seek_hole_data(inode, page, &lastoff, whence))
1340                                 goto check_range;
1341                         lastoff = page_offset(page) + PAGE_SIZE;
1342                 }
1343                 pagevec_release(&pvec);
1344         } while (index < end);
1345
1346         /* When no page at lastoff and we are not done, we found a hole. */
1347         if (whence != SEEK_HOLE)
1348                 goto not_found;
1349
1350 check_range:
1351         if (lastoff < offset + length)
1352                 goto out;
1353 not_found:
1354         lastoff = -ENOENT;
1355 out:
1356         pagevec_release(&pvec);
1357         return lastoff;
1358 }
1359
1360
1361 static loff_t
1362 iomap_seek_hole_actor(struct inode *inode, loff_t offset, loff_t length,
1363                       void *data, struct iomap *iomap)
1364 {
1365         switch (iomap->type) {
1366         case IOMAP_UNWRITTEN:
1367                 offset = page_cache_seek_hole_data(inode, offset, length,
1368                                                    SEEK_HOLE);
1369                 if (offset < 0)
1370                         return length;
1371                 /* fall through */
1372         case IOMAP_HOLE:
1373                 *(loff_t *)data = offset;
1374                 return 0;
1375         default:
1376                 return length;
1377         }
1378 }
1379
1380 loff_t
1381 iomap_seek_hole(struct inode *inode, loff_t offset, const struct iomap_ops *ops)
1382 {
1383         loff_t size = i_size_read(inode);
1384         loff_t length = size - offset;
1385         loff_t ret;
1386
1387         /* Nothing to be found before or beyond the end of the file. */
1388         if (offset < 0 || offset >= size)
1389                 return -ENXIO;
1390
1391         while (length > 0) {
1392                 ret = iomap_apply(inode, offset, length, IOMAP_REPORT, ops,
1393                                   &offset, iomap_seek_hole_actor);
1394                 if (ret < 0)
1395                         return ret;
1396                 if (ret == 0)
1397                         break;
1398
1399                 offset += ret;
1400                 length -= ret;
1401         }
1402
1403         return offset;
1404 }
1405 EXPORT_SYMBOL_GPL(iomap_seek_hole);
1406
1407 static loff_t
1408 iomap_seek_data_actor(struct inode *inode, loff_t offset, loff_t length,
1409                       void *data, struct iomap *iomap)
1410 {
1411         switch (iomap->type) {
1412         case IOMAP_HOLE:
1413                 return length;
1414         case IOMAP_UNWRITTEN:
1415                 offset = page_cache_seek_hole_data(inode, offset, length,
1416                                                    SEEK_DATA);
1417                 if (offset < 0)
1418                         return length;
1419                 /*FALLTHRU*/
1420         default:
1421                 *(loff_t *)data = offset;
1422                 return 0;
1423         }
1424 }
1425
1426 loff_t
1427 iomap_seek_data(struct inode *inode, loff_t offset, const struct iomap_ops *ops)
1428 {
1429         loff_t size = i_size_read(inode);
1430         loff_t length = size - offset;
1431         loff_t ret;
1432
1433         /* Nothing to be found before or beyond the end of the file. */
1434         if (offset < 0 || offset >= size)
1435                 return -ENXIO;
1436
1437         while (length > 0) {
1438                 ret = iomap_apply(inode, offset, length, IOMAP_REPORT, ops,
1439                                   &offset, iomap_seek_data_actor);
1440                 if (ret < 0)
1441                         return ret;
1442                 if (ret == 0)
1443                         break;
1444
1445                 offset += ret;
1446                 length -= ret;
1447         }
1448
1449         if (length <= 0)
1450                 return -ENXIO;
1451         return offset;
1452 }
1453 EXPORT_SYMBOL_GPL(iomap_seek_data);
1454
1455 /*
1456  * Private flags for iomap_dio, must not overlap with the public ones in
1457  * iomap.h:
1458  */
1459 #define IOMAP_DIO_WRITE_FUA     (1 << 28)
1460 #define IOMAP_DIO_NEED_SYNC     (1 << 29)
1461 #define IOMAP_DIO_WRITE         (1 << 30)
1462 #define IOMAP_DIO_DIRTY         (1 << 31)
1463
1464 struct iomap_dio {
1465         struct kiocb            *iocb;
1466         iomap_dio_end_io_t      *end_io;
1467         loff_t                  i_size;
1468         loff_t                  size;
1469         atomic_t                ref;
1470         unsigned                flags;
1471         int                     error;
1472         bool                    wait_for_completion;
1473
1474         union {
1475                 /* used during submission and for synchronous completion: */
1476                 struct {
1477                         struct iov_iter         *iter;
1478                         struct task_struct      *waiter;
1479                         struct request_queue    *last_queue;
1480                         blk_qc_t                cookie;
1481                 } submit;
1482
1483                 /* used for aio completion: */
1484                 struct {
1485                         struct work_struct      work;
1486                 } aio;
1487         };
1488 };
1489
1490 int iomap_dio_iopoll(struct kiocb *kiocb, bool spin)
1491 {
1492         struct request_queue *q = READ_ONCE(kiocb->private);
1493
1494         if (!q)
1495                 return 0;
1496         return blk_poll(q, READ_ONCE(kiocb->ki_cookie), spin);
1497 }
1498 EXPORT_SYMBOL_GPL(iomap_dio_iopoll);
1499
1500 static void iomap_dio_submit_bio(struct iomap_dio *dio, struct iomap *iomap,
1501                 struct bio *bio)
1502 {
1503         atomic_inc(&dio->ref);
1504
1505         if (dio->iocb->ki_flags & IOCB_HIPRI)
1506                 bio_set_polled(bio, dio->iocb);
1507
1508         dio->submit.last_queue = bdev_get_queue(iomap->bdev);
1509         dio->submit.cookie = submit_bio(bio);
1510 }
1511
1512 static ssize_t iomap_dio_complete(struct iomap_dio *dio)
1513 {
1514         struct kiocb *iocb = dio->iocb;
1515         struct inode *inode = file_inode(iocb->ki_filp);
1516         loff_t offset = iocb->ki_pos;
1517         ssize_t ret;
1518
1519         if (dio->end_io) {
1520                 ret = dio->end_io(iocb,
1521                                 dio->error ? dio->error : dio->size,
1522                                 dio->flags);
1523         } else {
1524                 ret = dio->error;
1525         }
1526
1527         if (likely(!ret)) {
1528                 ret = dio->size;
1529                 /* check for short read */
1530                 if (offset + ret > dio->i_size &&
1531                     !(dio->flags & IOMAP_DIO_WRITE))
1532                         ret = dio->i_size - offset;
1533                 iocb->ki_pos += ret;
1534         }
1535
1536         /*
1537          * Try again to invalidate clean pages which might have been cached by
1538          * non-direct readahead, or faulted in by get_user_pages() if the source
1539          * of the write was an mmap'ed region of the file we're writing.  Either
1540          * one is a pretty crazy thing to do, so we don't support it 100%.  If
1541          * this invalidation fails, tough, the write still worked...
1542          *
1543          * And this page cache invalidation has to be after dio->end_io(), as
1544          * some filesystems convert unwritten extents to real allocations in
1545          * end_io() when necessary, otherwise a racing buffer read would cache
1546          * zeros from unwritten extents.
1547          */
1548         if (!dio->error &&
1549             (dio->flags & IOMAP_DIO_WRITE) && inode->i_mapping->nrpages) {
1550                 int err;
1551                 err = invalidate_inode_pages2_range(inode->i_mapping,
1552                                 offset >> PAGE_SHIFT,
1553                                 (offset + dio->size - 1) >> PAGE_SHIFT);
1554                 if (err)
1555                         dio_warn_stale_pagecache(iocb->ki_filp);
1556         }
1557
1558         /*
1559          * If this is a DSYNC write, make sure we push it to stable storage now
1560          * that we've written data.
1561          */
1562         if (ret > 0 && (dio->flags & IOMAP_DIO_NEED_SYNC))
1563                 ret = generic_write_sync(iocb, ret);
1564
1565         inode_dio_end(file_inode(iocb->ki_filp));
1566         kfree(dio);
1567
1568         return ret;
1569 }
1570
1571 static void iomap_dio_complete_work(struct work_struct *work)
1572 {
1573         struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work);
1574         struct kiocb *iocb = dio->iocb;
1575
1576         iocb->ki_complete(iocb, iomap_dio_complete(dio), 0);
1577 }
1578
1579 /*
1580  * Set an error in the dio if none is set yet.  We have to use cmpxchg
1581  * as the submission context and the completion context(s) can race to
1582  * update the error.
1583  */
1584 static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret)
1585 {
1586         cmpxchg(&dio->error, 0, ret);
1587 }
1588
1589 static void iomap_dio_bio_end_io(struct bio *bio)
1590 {
1591         struct iomap_dio *dio = bio->bi_private;
1592         bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
1593
1594         if (bio->bi_status)
1595                 iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
1596
1597         if (atomic_dec_and_test(&dio->ref)) {
1598                 if (dio->wait_for_completion) {
1599                         struct task_struct *waiter = dio->submit.waiter;
1600                         WRITE_ONCE(dio->submit.waiter, NULL);
1601                         blk_wake_io_task(waiter);
1602                 } else if (dio->flags & IOMAP_DIO_WRITE) {
1603                         struct inode *inode = file_inode(dio->iocb->ki_filp);
1604
1605                         INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
1606                         queue_work(inode->i_sb->s_dio_done_wq, &dio->aio.work);
1607                 } else {
1608                         iomap_dio_complete_work(&dio->aio.work);
1609                 }
1610         }
1611
1612         if (should_dirty) {
1613                 bio_check_pages_dirty(bio);
1614         } else {
1615                 bio_release_pages(bio, false);
1616                 bio_put(bio);
1617         }
1618 }
1619
1620 static void
1621 iomap_dio_zero(struct iomap_dio *dio, struct iomap *iomap, loff_t pos,
1622                 unsigned len)
1623 {
1624         struct page *page = ZERO_PAGE(0);
1625         int flags = REQ_SYNC | REQ_IDLE;
1626         struct bio *bio;
1627
1628         bio = bio_alloc(GFP_KERNEL, 1);
1629         bio_set_dev(bio, iomap->bdev);
1630         bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
1631         bio->bi_private = dio;
1632         bio->bi_end_io = iomap_dio_bio_end_io;
1633
1634         get_page(page);
1635         __bio_add_page(bio, page, len, 0);
1636         bio_set_op_attrs(bio, REQ_OP_WRITE, flags);
1637         iomap_dio_submit_bio(dio, iomap, bio);
1638 }
1639
1640 static loff_t
1641 iomap_dio_bio_actor(struct inode *inode, loff_t pos, loff_t length,
1642                 struct iomap_dio *dio, struct iomap *iomap)
1643 {
1644         unsigned int blkbits = blksize_bits(bdev_logical_block_size(iomap->bdev));
1645         unsigned int fs_block_size = i_blocksize(inode), pad;
1646         unsigned int align = iov_iter_alignment(dio->submit.iter);
1647         struct iov_iter iter;
1648         struct bio *bio;
1649         bool need_zeroout = false;
1650         bool use_fua = false;
1651         int nr_pages, ret = 0;
1652         size_t copied = 0;
1653
1654         if ((pos | length | align) & ((1 << blkbits) - 1))
1655                 return -EINVAL;
1656
1657         if (iomap->type == IOMAP_UNWRITTEN) {
1658                 dio->flags |= IOMAP_DIO_UNWRITTEN;
1659                 need_zeroout = true;
1660         }
1661
1662         if (iomap->flags & IOMAP_F_SHARED)
1663                 dio->flags |= IOMAP_DIO_COW;
1664
1665         if (iomap->flags & IOMAP_F_NEW) {
1666                 need_zeroout = true;
1667         } else if (iomap->type == IOMAP_MAPPED) {
1668                 /*
1669                  * Use a FUA write if we need datasync semantics, this is a pure
1670                  * data IO that doesn't require any metadata updates (including
1671                  * after IO completion such as unwritten extent conversion) and
1672                  * the underlying device supports FUA. This allows us to avoid
1673                  * cache flushes on IO completion.
1674                  */
1675                 if (!(iomap->flags & (IOMAP_F_SHARED|IOMAP_F_DIRTY)) &&
1676                     (dio->flags & IOMAP_DIO_WRITE_FUA) &&
1677                     blk_queue_fua(bdev_get_queue(iomap->bdev)))
1678                         use_fua = true;
1679         }
1680
1681         /*
1682          * Operate on a partial iter trimmed to the extent we were called for.
1683          * We'll update the iter in the dio once we're done with this extent.
1684          */
1685         iter = *dio->submit.iter;
1686         iov_iter_truncate(&iter, length);
1687
1688         nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
1689         if (nr_pages <= 0)
1690                 return nr_pages;
1691
1692         if (need_zeroout) {
1693                 /* zero out from the start of the block to the write offset */
1694                 pad = pos & (fs_block_size - 1);
1695                 if (pad)
1696                         iomap_dio_zero(dio, iomap, pos - pad, pad);
1697         }
1698
1699         do {
1700                 size_t n;
1701                 if (dio->error) {
1702                         iov_iter_revert(dio->submit.iter, copied);
1703                         return 0;
1704                 }
1705
1706                 bio = bio_alloc(GFP_KERNEL, nr_pages);
1707                 bio_set_dev(bio, iomap->bdev);
1708                 bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
1709                 bio->bi_write_hint = dio->iocb->ki_hint;
1710                 bio->bi_ioprio = dio->iocb->ki_ioprio;
1711                 bio->bi_private = dio;
1712                 bio->bi_end_io = iomap_dio_bio_end_io;
1713
1714                 ret = bio_iov_iter_get_pages(bio, &iter);
1715                 if (unlikely(ret)) {
1716                         /*
1717                          * We have to stop part way through an IO. We must fall
1718                          * through to the sub-block tail zeroing here, otherwise
1719                          * this short IO may expose stale data in the tail of
1720                          * the block we haven't written data to.
1721                          */
1722                         bio_put(bio);
1723                         goto zero_tail;
1724                 }
1725
1726                 n = bio->bi_iter.bi_size;
1727                 if (dio->flags & IOMAP_DIO_WRITE) {
1728                         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
1729                         if (use_fua)
1730                                 bio->bi_opf |= REQ_FUA;
1731                         else
1732                                 dio->flags &= ~IOMAP_DIO_WRITE_FUA;
1733                         task_io_account_write(n);
1734                 } else {
1735                         bio->bi_opf = REQ_OP_READ;
1736                         if (dio->flags & IOMAP_DIO_DIRTY)
1737                                 bio_set_pages_dirty(bio);
1738                 }
1739
1740                 iov_iter_advance(dio->submit.iter, n);
1741
1742                 dio->size += n;
1743                 pos += n;
1744                 copied += n;
1745
1746                 nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
1747                 iomap_dio_submit_bio(dio, iomap, bio);
1748         } while (nr_pages);
1749
1750         /*
1751          * We need to zeroout the tail of a sub-block write if the extent type
1752          * requires zeroing or the write extends beyond EOF. If we don't zero
1753          * the block tail in the latter case, we can expose stale data via mmap
1754          * reads of the EOF block.
1755          */
1756 zero_tail:
1757         if (need_zeroout ||
1758             ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode))) {
1759                 /* zero out from the end of the write to the end of the block */
1760                 pad = pos & (fs_block_size - 1);
1761                 if (pad)
1762                         iomap_dio_zero(dio, iomap, pos, fs_block_size - pad);
1763         }
1764         return copied ? copied : ret;
1765 }
1766
1767 static loff_t
1768 iomap_dio_hole_actor(loff_t length, struct iomap_dio *dio)
1769 {
1770         length = iov_iter_zero(length, dio->submit.iter);
1771         dio->size += length;
1772         return length;
1773 }
1774
1775 static loff_t
1776 iomap_dio_inline_actor(struct inode *inode, loff_t pos, loff_t length,
1777                 struct iomap_dio *dio, struct iomap *iomap)
1778 {
1779         struct iov_iter *iter = dio->submit.iter;
1780         size_t copied;
1781
1782         BUG_ON(pos + length > PAGE_SIZE - offset_in_page(iomap->inline_data));
1783
1784         if (dio->flags & IOMAP_DIO_WRITE) {
1785                 loff_t size = inode->i_size;
1786
1787                 if (pos > size)
1788                         memset(iomap->inline_data + size, 0, pos - size);
1789                 copied = copy_from_iter(iomap->inline_data + pos, length, iter);
1790                 if (copied) {
1791                         if (pos + copied > size)
1792                                 i_size_write(inode, pos + copied);
1793                         mark_inode_dirty(inode);
1794                 }
1795         } else {
1796                 copied = copy_to_iter(iomap->inline_data + pos, length, iter);
1797         }
1798         dio->size += copied;
1799         return copied;
1800 }
1801
1802 static loff_t
1803 iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
1804                 void *data, struct iomap *iomap)
1805 {
1806         struct iomap_dio *dio = data;
1807
1808         switch (iomap->type) {
1809         case IOMAP_HOLE:
1810                 if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
1811                         return -EIO;
1812                 return iomap_dio_hole_actor(length, dio);
1813         case IOMAP_UNWRITTEN:
1814                 if (!(dio->flags & IOMAP_DIO_WRITE))
1815                         return iomap_dio_hole_actor(length, dio);
1816                 return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
1817         case IOMAP_MAPPED:
1818                 return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
1819         case IOMAP_INLINE:
1820                 return iomap_dio_inline_actor(inode, pos, length, dio, iomap);
1821         default:
1822                 WARN_ON_ONCE(1);
1823                 return -EIO;
1824         }
1825 }
1826
1827 /*
1828  * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
1829  * is being issued as AIO or not.  This allows us to optimise pure data writes
1830  * to use REQ_FUA rather than requiring generic_write_sync() to issue a
1831  * REQ_FLUSH post write. This is slightly tricky because a single request here
1832  * can be mapped into multiple disjoint IOs and only a subset of the IOs issued
1833  * may be pure data writes. In that case, we still need to do a full data sync
1834  * completion.
1835  */
1836 ssize_t
1837 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
1838                 const struct iomap_ops *ops, iomap_dio_end_io_t end_io)
1839 {
1840         struct address_space *mapping = iocb->ki_filp->f_mapping;
1841         struct inode *inode = file_inode(iocb->ki_filp);
1842         size_t count = iov_iter_count(iter);
1843         loff_t pos = iocb->ki_pos, start = pos;
1844         loff_t end = iocb->ki_pos + count - 1, ret = 0;
1845         unsigned int flags = IOMAP_DIRECT;
1846         bool wait_for_completion = is_sync_kiocb(iocb);
1847         struct blk_plug plug;
1848         struct iomap_dio *dio;
1849
1850         lockdep_assert_held(&inode->i_rwsem);
1851
1852         if (!count)
1853                 return 0;
1854
1855         dio = kmalloc(sizeof(*dio), GFP_KERNEL);
1856         if (!dio)
1857                 return -ENOMEM;
1858
1859         dio->iocb = iocb;
1860         atomic_set(&dio->ref, 1);
1861         dio->size = 0;
1862         dio->i_size = i_size_read(inode);
1863         dio->end_io = end_io;
1864         dio->error = 0;
1865         dio->flags = 0;
1866
1867         dio->submit.iter = iter;
1868         dio->submit.waiter = current;
1869         dio->submit.cookie = BLK_QC_T_NONE;
1870         dio->submit.last_queue = NULL;
1871
1872         if (iov_iter_rw(iter) == READ) {
1873                 if (pos >= dio->i_size)
1874                         goto out_free_dio;
1875
1876                 if (iter_is_iovec(iter) && iov_iter_rw(iter) == READ)
1877                         dio->flags |= IOMAP_DIO_DIRTY;
1878         } else {
1879                 flags |= IOMAP_WRITE;
1880                 dio->flags |= IOMAP_DIO_WRITE;
1881
1882                 /* for data sync or sync, we need sync completion processing */
1883                 if (iocb->ki_flags & IOCB_DSYNC)
1884                         dio->flags |= IOMAP_DIO_NEED_SYNC;
1885
1886                 /*
1887                  * For datasync only writes, we optimistically try using FUA for
1888                  * this IO.  Any non-FUA write that occurs will clear this flag,
1889                  * hence we know before completion whether a cache flush is
1890                  * necessary.
1891                  */
1892                 if ((iocb->ki_flags & (IOCB_DSYNC | IOCB_SYNC)) == IOCB_DSYNC)
1893                         dio->flags |= IOMAP_DIO_WRITE_FUA;
1894         }
1895
1896         if (iocb->ki_flags & IOCB_NOWAIT) {
1897                 if (filemap_range_has_page(mapping, start, end)) {
1898                         ret = -EAGAIN;
1899                         goto out_free_dio;
1900                 }
1901                 flags |= IOMAP_NOWAIT;
1902         }
1903
1904         ret = filemap_write_and_wait_range(mapping, start, end);
1905         if (ret)
1906                 goto out_free_dio;
1907
1908         /*
1909          * Try to invalidate cache pages for the range we're direct
1910          * writing.  If this invalidation fails, tough, the write will
1911          * still work, but racing two incompatible write paths is a
1912          * pretty crazy thing to do, so we don't support it 100%.
1913          */
1914         ret = invalidate_inode_pages2_range(mapping,
1915                         start >> PAGE_SHIFT, end >> PAGE_SHIFT);
1916         if (ret)
1917                 dio_warn_stale_pagecache(iocb->ki_filp);
1918         ret = 0;
1919
1920         if (iov_iter_rw(iter) == WRITE && !wait_for_completion &&
1921             !inode->i_sb->s_dio_done_wq) {
1922                 ret = sb_init_dio_done_wq(inode->i_sb);
1923                 if (ret < 0)
1924                         goto out_free_dio;
1925         }
1926
1927         inode_dio_begin(inode);
1928
1929         blk_start_plug(&plug);
1930         do {
1931                 ret = iomap_apply(inode, pos, count, flags, ops, dio,
1932                                 iomap_dio_actor);
1933                 if (ret <= 0) {
1934                         /* magic error code to fall back to buffered I/O */
1935                         if (ret == -ENOTBLK) {
1936                                 wait_for_completion = true;
1937                                 ret = 0;
1938                         }
1939                         break;
1940                 }
1941                 pos += ret;
1942
1943                 if (iov_iter_rw(iter) == READ && pos >= dio->i_size)
1944                         break;
1945         } while ((count = iov_iter_count(iter)) > 0);
1946         blk_finish_plug(&plug);
1947
1948         if (ret < 0)
1949                 iomap_dio_set_error(dio, ret);
1950
1951         /*
1952          * If all the writes we issued were FUA, we don't need to flush the
1953          * cache on IO completion. Clear the sync flag for this case.
1954          */
1955         if (dio->flags & IOMAP_DIO_WRITE_FUA)
1956                 dio->flags &= ~IOMAP_DIO_NEED_SYNC;
1957
1958         WRITE_ONCE(iocb->ki_cookie, dio->submit.cookie);
1959         WRITE_ONCE(iocb->private, dio->submit.last_queue);
1960
1961         /*
1962          * We are about to drop our additional submission reference, which
1963          * might be the last reference to the dio.  There are three three
1964          * different ways we can progress here:
1965          *
1966          *  (a) If this is the last reference we will always complete and free
1967          *      the dio ourselves.
1968          *  (b) If this is not the last reference, and we serve an asynchronous
1969          *      iocb, we must never touch the dio after the decrement, the
1970          *      I/O completion handler will complete and free it.
1971          *  (c) If this is not the last reference, but we serve a synchronous
1972          *      iocb, the I/O completion handler will wake us up on the drop
1973          *      of the final reference, and we will complete and free it here
1974          *      after we got woken by the I/O completion handler.
1975          */
1976         dio->wait_for_completion = wait_for_completion;
1977         if (!atomic_dec_and_test(&dio->ref)) {
1978                 if (!wait_for_completion)
1979                         return -EIOCBQUEUED;
1980
1981                 for (;;) {
1982                         set_current_state(TASK_UNINTERRUPTIBLE);
1983                         if (!READ_ONCE(dio->submit.waiter))
1984                                 break;
1985
1986                         if (!(iocb->ki_flags & IOCB_HIPRI) ||
1987                             !dio->submit.last_queue ||
1988                             !blk_poll(dio->submit.last_queue,
1989                                          dio->submit.cookie, true))
1990                                 io_schedule();
1991                 }
1992                 __set_current_state(TASK_RUNNING);
1993         }
1994
1995         return iomap_dio_complete(dio);
1996
1997 out_free_dio:
1998         kfree(dio);
1999         return ret;
2000 }
2001 EXPORT_SYMBOL_GPL(iomap_dio_rw);
2002
2003 /* Swapfile activation */
2004
2005 #ifdef CONFIG_SWAP
2006 struct iomap_swapfile_info {
2007         struct iomap iomap;             /* accumulated iomap */
2008         struct swap_info_struct *sis;
2009         uint64_t lowest_ppage;          /* lowest physical addr seen (pages) */
2010         uint64_t highest_ppage;         /* highest physical addr seen (pages) */
2011         unsigned long nr_pages;         /* number of pages collected */
2012         int nr_extents;                 /* extent count */
2013 };
2014
2015 /*
2016  * Collect physical extents for this swap file.  Physical extents reported to
2017  * the swap code must be trimmed to align to a page boundary.  The logical
2018  * offset within the file is irrelevant since the swapfile code maps logical
2019  * page numbers of the swap device to the physical page-aligned extents.
2020  */
2021 static int iomap_swapfile_add_extent(struct iomap_swapfile_info *isi)
2022 {
2023         struct iomap *iomap = &isi->iomap;
2024         unsigned long nr_pages;
2025         uint64_t first_ppage;
2026         uint64_t first_ppage_reported;
2027         uint64_t next_ppage;
2028         int error;
2029
2030         /*
2031          * Round the start up and the end down so that the physical
2032          * extent aligns to a page boundary.
2033          */
2034         first_ppage = ALIGN(iomap->addr, PAGE_SIZE) >> PAGE_SHIFT;
2035         next_ppage = ALIGN_DOWN(iomap->addr + iomap->length, PAGE_SIZE) >>
2036                         PAGE_SHIFT;
2037
2038         /* Skip too-short physical extents. */
2039         if (first_ppage >= next_ppage)
2040                 return 0;
2041         nr_pages = next_ppage - first_ppage;
2042
2043         /*
2044          * Calculate how much swap space we're adding; the first page contains
2045          * the swap header and doesn't count.  The mm still wants that first
2046          * page fed to add_swap_extent, however.
2047          */
2048         first_ppage_reported = first_ppage;
2049         if (iomap->offset == 0)
2050                 first_ppage_reported++;
2051         if (isi->lowest_ppage > first_ppage_reported)
2052                 isi->lowest_ppage = first_ppage_reported;
2053         if (isi->highest_ppage < (next_ppage - 1))
2054                 isi->highest_ppage = next_ppage - 1;
2055
2056         /* Add extent, set up for the next call. */
2057         error = add_swap_extent(isi->sis, isi->nr_pages, nr_pages, first_ppage);
2058         if (error < 0)
2059                 return error;
2060         isi->nr_extents += error;
2061         isi->nr_pages += nr_pages;
2062         return 0;
2063 }
2064
2065 /*
2066  * Accumulate iomaps for this swap file.  We have to accumulate iomaps because
2067  * swap only cares about contiguous page-aligned physical extents and makes no
2068  * distinction between written and unwritten extents.
2069  */
2070 static loff_t iomap_swapfile_activate_actor(struct inode *inode, loff_t pos,
2071                 loff_t count, void *data, struct iomap *iomap)
2072 {
2073         struct iomap_swapfile_info *isi = data;
2074         int error;
2075
2076         switch (iomap->type) {
2077         case IOMAP_MAPPED:
2078         case IOMAP_UNWRITTEN:
2079                 /* Only real or unwritten extents. */
2080                 break;
2081         case IOMAP_INLINE:
2082                 /* No inline data. */
2083                 pr_err("swapon: file is inline\n");
2084                 return -EINVAL;
2085         default:
2086                 pr_err("swapon: file has unallocated extents\n");
2087                 return -EINVAL;
2088         }
2089
2090         /* No uncommitted metadata or shared blocks. */
2091         if (iomap->flags & IOMAP_F_DIRTY) {
2092                 pr_err("swapon: file is not committed\n");
2093                 return -EINVAL;
2094         }
2095         if (iomap->flags & IOMAP_F_SHARED) {
2096                 pr_err("swapon: file has shared extents\n");
2097                 return -EINVAL;
2098         }
2099
2100         /* Only one bdev per swap file. */
2101         if (iomap->bdev != isi->sis->bdev) {
2102                 pr_err("swapon: file is on multiple devices\n");
2103                 return -EINVAL;
2104         }
2105
2106         if (isi->iomap.length == 0) {
2107                 /* No accumulated extent, so just store it. */
2108                 memcpy(&isi->iomap, iomap, sizeof(isi->iomap));
2109         } else if (isi->iomap.addr + isi->iomap.length == iomap->addr) {
2110                 /* Append this to the accumulated extent. */
2111                 isi->iomap.length += iomap->length;
2112         } else {
2113                 /* Otherwise, add the retained iomap and store this one. */
2114                 error = iomap_swapfile_add_extent(isi);
2115                 if (error)
2116                         return error;
2117                 memcpy(&isi->iomap, iomap, sizeof(isi->iomap));
2118         }
2119         return count;
2120 }
2121
2122 /*
2123  * Iterate a swap file's iomaps to construct physical extents that can be
2124  * passed to the swapfile subsystem.
2125  */
2126 int iomap_swapfile_activate(struct swap_info_struct *sis,
2127                 struct file *swap_file, sector_t *pagespan,
2128                 const struct iomap_ops *ops)
2129 {
2130         struct iomap_swapfile_info isi = {
2131                 .sis = sis,
2132                 .lowest_ppage = (sector_t)-1ULL,
2133         };
2134         struct address_space *mapping = swap_file->f_mapping;
2135         struct inode *inode = mapping->host;
2136         loff_t pos = 0;
2137         loff_t len = ALIGN_DOWN(i_size_read(inode), PAGE_SIZE);
2138         loff_t ret;
2139
2140         /*
2141          * Persist all file mapping metadata so that we won't have any
2142          * IOMAP_F_DIRTY iomaps.
2143          */
2144         ret = vfs_fsync(swap_file, 1);
2145         if (ret)
2146                 return ret;
2147
2148         while (len > 0) {
2149                 ret = iomap_apply(inode, pos, len, IOMAP_REPORT,
2150                                 ops, &isi, iomap_swapfile_activate_actor);
2151                 if (ret <= 0)
2152                         return ret;
2153
2154                 pos += ret;
2155                 len -= ret;
2156         }
2157
2158         if (isi.iomap.length) {
2159                 ret = iomap_swapfile_add_extent(&isi);
2160                 if (ret)
2161                         return ret;
2162         }
2163
2164         *pagespan = 1 + isi.highest_ppage - isi.lowest_ppage;
2165         sis->max = isi.nr_pages;
2166         sis->pages = isi.nr_pages - 1;
2167         sis->highest_bit = isi.nr_pages - 1;
2168         return isi.nr_extents;
2169 }
2170 EXPORT_SYMBOL_GPL(iomap_swapfile_activate);
2171 #endif /* CONFIG_SWAP */
2172
2173 static loff_t
2174 iomap_bmap_actor(struct inode *inode, loff_t pos, loff_t length,
2175                 void *data, struct iomap *iomap)
2176 {
2177         sector_t *bno = data, addr;
2178
2179         if (iomap->type == IOMAP_MAPPED) {
2180                 addr = (pos - iomap->offset + iomap->addr) >> inode->i_blkbits;
2181                 if (addr > INT_MAX)
2182                         WARN(1, "would truncate bmap result\n");
2183                 else
2184                         *bno = addr;
2185         }
2186         return 0;
2187 }
2188
2189 /* legacy ->bmap interface.  0 is the error return (!) */
2190 sector_t
2191 iomap_bmap(struct address_space *mapping, sector_t bno,
2192                 const struct iomap_ops *ops)
2193 {
2194         struct inode *inode = mapping->host;
2195         loff_t pos = bno << inode->i_blkbits;
2196         unsigned blocksize = i_blocksize(inode);
2197
2198         if (filemap_write_and_wait(mapping))
2199                 return 0;
2200
2201         bno = 0;
2202         iomap_apply(inode, pos, blocksize, 0, ops, &bno, iomap_bmap_actor);
2203         return bno;
2204 }
2205 EXPORT_SYMBOL_GPL(iomap_bmap);