fs/iomap/buffered-io.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/pagemap.h>
  11 #include <linux/uio.h>
  12 #include <linux/buffer_head.h>
  13 #include <linux/dax.h>
  14 #include <linux/writeback.h>
  15 #include <linux/swap.h>
  16 #include <linux/bio.h>
  17 #include <linux/sched/signal.h>
  18 #include <linux/migrate.h>
  19
  20 #include "../internal.h"
  21
  22 static struct iomap_page *
  23 iomap_page_create(struct inode *inode, struct page *page)
  24 {
  25         struct iomap_page *iop = to_iomap_page(page);
  26
  27         if (iop || i_blocksize(inode) == PAGE_SIZE)
  28                 return iop;
  29
  30         iop = kmalloc(sizeof(*iop), GFP_NOFS | __GFP_NOFAIL);
  31         atomic_set(&iop->read_count, 0);
  32         atomic_set(&iop->write_count, 0);
  33         bitmap_zero(iop->uptodate, PAGE_SIZE / SECTOR_SIZE);
  34
  35         /*
  36          * migrate_page_move_mapping() assumes that pages with private data have
  37          * their count elevated by 1.
  38          */
  39         get_page(page);
  40         set_page_private(page, (unsigned long)iop);
  41         SetPagePrivate(page);
  42         return iop;
  43 }
  44
  45 static void
  46 iomap_page_release(struct page *page)
  47 {
  48         struct iomap_page *iop = to_iomap_page(page);
  49
  50         if (!iop)
  51                 return;
  52         WARN_ON_ONCE(atomic_read(&iop->read_count));
  53         WARN_ON_ONCE(atomic_read(&iop->write_count));
  54         ClearPagePrivate(page);
  55         set_page_private(page, 0);
  56         put_page(page);
  57         kfree(iop);
  58 }
  59
  60 /*
  61  * Calculate the range inside the page that we actually need to read.
  62  */
  63 static void
  64 iomap_adjust_read_range(struct inode *inode, struct iomap_page *iop,
  65                 loff_t *pos, loff_t length, unsigned *offp, unsigned *lenp)
  66 {
  67         loff_t orig_pos = *pos;
  68         loff_t isize = i_size_read(inode);
  69         unsigned block_bits = inode->i_blkbits;
  70         unsigned block_size = (1 << block_bits);
  71         unsigned poff = offset_in_page(*pos);
  72         unsigned plen = min_t(loff_t, PAGE_SIZE - poff, length);
  73         unsigned first = poff >> block_bits;
  74         unsigned last = (poff + plen - 1) >> block_bits;
  75
  76         /*
  77          * If the block size is smaller than the page size we need to check the
  78          * per-block uptodate status and adjust the offset and length if needed
  79          * to avoid reading in already uptodate ranges.
  80          */
  81         if (iop) {
  82                 unsigned int i;
  83
  84                 /* move forward for each leading block marked uptodate */
  85                 for (i = first; i <= last; i++) {
  86                         if (!test_bit(i, iop->uptodate))
  87                                 break;
  88                         *pos += block_size;
  89                         poff += block_size;
  90                         plen -= block_size;
  91                         first++;
  92                 }
  93
  94                 /* truncate len if we find any trailing uptodate block(s) */
  95                 for ( ; i <= last; i++) {
  96                         if (test_bit(i, iop->uptodate)) {
  97                                 plen -= (last - i + 1) * block_size;
  98                                 last = i - 1;
  99                                 break;
 100                         }
 101                 }
 102         }
 103
 104         /*
 105          * If the extent spans the block that contains the i_size we need to
 106          * handle both halves separately so that we properly zero data in the
 107          * page cache for blocks that are entirely outside of i_size.
 108          */
 109         if (orig_pos <= isize && orig_pos + length > isize) {
 110                 unsigned end = offset_in_page(isize - 1) >> block_bits;
 111
 112                 if (first <= end && last > end)
 113                         plen -= (last - end) * block_size;
 114         }
 115
 116         *offp = poff;
 117         *lenp = plen;
 118 }
 119
 120 static void
 121 iomap_set_range_uptodate(struct page *page, unsigned off, unsigned len)
 122 {
 123         struct iomap_page *iop = to_iomap_page(page);
 124         struct inode *inode = page->mapping->host;
 125         unsigned first = off >> inode->i_blkbits;
 126         unsigned last = (off + len - 1) >> inode->i_blkbits;
 127         unsigned int i;
 128         bool uptodate = true;
 129
 130         if (iop) {
 131                 for (i = 0; i < PAGE_SIZE / i_blocksize(inode); i++) {
 132                         if (i >= first && i <= last)
 133                                 set_bit(i, iop->uptodate);
 134                         else if (!test_bit(i, iop->uptodate))
 135                                 uptodate = false;
 136                 }
 137         }
 138
 139         if (uptodate && !PageError(page))
 140                 SetPageUptodate(page);
 141 }
 142
 143 static void
 144 iomap_read_finish(struct iomap_page *iop, struct page *page)
 145 {
 146         if (!iop || atomic_dec_and_test(&iop->read_count))
 147                 unlock_page(page);
 148 }
 149
 150 static void
 151 iomap_read_page_end_io(struct bio_vec *bvec, int error)
 152 {
 153         struct page *page = bvec->bv_page;
 154         struct iomap_page *iop = to_iomap_page(page);
 155
 156         if (unlikely(error)) {
 157                 ClearPageUptodate(page);
 158                 SetPageError(page);
 159         } else {
 160                 iomap_set_range_uptodate(page, bvec->bv_offset, bvec->bv_len);
 161         }
 162
 163         iomap_read_finish(iop, page);
 164 }
 165
 166 static void
 167 iomap_read_end_io(struct bio *bio)
 168 {
 169         int error = blk_status_to_errno(bio->bi_status);
 170         struct bio_vec *bvec;
 171         struct bvec_iter_all iter_all;
 172
 173         bio_for_each_segment_all(bvec, bio, iter_all)
 174                 iomap_read_page_end_io(bvec, error);
 175         bio_put(bio);
 176 }
 177
 178 struct iomap_readpage_ctx {
 179         struct page             *cur_page;
 180         bool                    cur_page_in_bio;
 181         bool                    is_readahead;
 182         struct bio              *bio;
 183         struct list_head        *pages;
 184 };
 185
 186 static void
 187 iomap_read_inline_data(struct inode *inode, struct page *page,
 188                 struct iomap *iomap)
 189 {
 190         size_t size = i_size_read(inode);
 191         void *addr;
 192
 193         if (PageUptodate(page))
 194                 return;
 195
 196         BUG_ON(page->index);
 197         BUG_ON(size > PAGE_SIZE - offset_in_page(iomap->inline_data));
 198
 199         addr = kmap_atomic(page);
 200         memcpy(addr, iomap->inline_data, size);
 201         memset(addr + size, 0, PAGE_SIZE - size);
 202         kunmap_atomic(addr);
 203         SetPageUptodate(page);
 204 }
 205
 206 static loff_t
 207 iomap_readpage_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 208                 struct iomap *iomap)
 209 {
 210         struct iomap_readpage_ctx *ctx = data;
 211         struct page *page = ctx->cur_page;
 212         struct iomap_page *iop = iomap_page_create(inode, page);
 213         bool same_page = false, is_contig = false;
 214         loff_t orig_pos = pos;
 215         unsigned poff, plen;
 216         sector_t sector;
 217
 218         if (iomap->type == IOMAP_INLINE) {
 219                 WARN_ON_ONCE(pos);
 220                 iomap_read_inline_data(inode, page, iomap);
 221                 return PAGE_SIZE;
 222         }
 223
 224         /* zero post-eof blocks as the page may be mapped */
 225         iomap_adjust_read_range(inode, iop, &pos, length, &poff, &plen);
 226         if (plen == 0)
 227                 goto done;
 228
 229         if (iomap->type != IOMAP_MAPPED || pos >= i_size_read(inode)) {
 230                 zero_user(page, poff, plen);
 231                 iomap_set_range_uptodate(page, poff, plen);
 232                 goto done;
 233         }
 234
 235         ctx->cur_page_in_bio = true;
 236
 237         /*
 238          * Try to merge into a previous segment if we can.
 239          */
 240         sector = iomap_sector(iomap, pos);
 241         if (ctx->bio && bio_end_sector(ctx->bio) == sector)
 242                 is_contig = true;
 243
 244         if (is_contig &&
 245             __bio_try_merge_page(ctx->bio, page, plen, poff, &same_page)) {
 246                 if (!same_page && iop)
 247                         atomic_inc(&iop->read_count);
 248                 goto done;
 249         }
 250
 251         /*
 252          * If we start a new segment we need to increase the read count, and we
 253          * need to do so before submitting any previous full bio to make sure
 254          * that we don't prematurely unlock the page.
 255          */
 256         if (iop)
 257                 atomic_inc(&iop->read_count);
 258
 259         if (!ctx->bio || !is_contig || bio_full(ctx->bio, plen)) {
 260                 gfp_t gfp = mapping_gfp_constraint(page->mapping, GFP_KERNEL);
 261                 int nr_vecs = (length + PAGE_SIZE - 1) >> PAGE_SHIFT;
 262
 263                 if (ctx->bio)
 264                         submit_bio(ctx->bio);
 265
 266                 if (ctx->is_readahead) /* same as readahead_gfp_mask */
 267                         gfp |= __GFP_NORETRY | __GFP_NOWARN;
 268                 ctx->bio = bio_alloc(gfp, min(BIO_MAX_PAGES, nr_vecs));
 269                 ctx->bio->bi_opf = REQ_OP_READ;
 270                 if (ctx->is_readahead)
 271                         ctx->bio->bi_opf |= REQ_RAHEAD;
 272                 ctx->bio->bi_iter.bi_sector = sector;
 273                 bio_set_dev(ctx->bio, iomap->bdev);
 274                 ctx->bio->bi_end_io = iomap_read_end_io;
 275         }
 276
 277         bio_add_page(ctx->bio, page, plen, poff);
 278 done:
 279         /*
 280          * Move the caller beyond our range so that it keeps making progress.
 281          * For that we have to include any leading non-uptodate ranges, but
 282          * we can skip trailing ones as they will be handled in the next
 283          * iteration.
 284          */
 285         return pos - orig_pos + plen;
 286 }
 287
 288 int
 289 iomap_readpage(struct page *page, const struct iomap_ops *ops)
 290 {
 291         struct iomap_readpage_ctx ctx = { .cur_page = page };
 292         struct inode *inode = page->mapping->host;
 293         unsigned poff;
 294         loff_t ret;
 295
 296         for (poff = 0; poff < PAGE_SIZE; poff += ret) {
 297                 ret = iomap_apply(inode, page_offset(page) + poff,
 298                                 PAGE_SIZE - poff, 0, ops, &ctx,
 299                                 iomap_readpage_actor);
 300                 if (ret <= 0) {
 301                         WARN_ON_ONCE(ret == 0);
 302                         SetPageError(page);
 303                         break;
 304                 }
 305         }
 306
 307         if (ctx.bio) {
 308                 submit_bio(ctx.bio);
 309                 WARN_ON_ONCE(!ctx.cur_page_in_bio);
 310         } else {
 311                 WARN_ON_ONCE(ctx.cur_page_in_bio);
 312                 unlock_page(page);
 313         }
 314
 315         /*
 316          * Just like mpage_readpages and block_read_full_page we always
 317          * return 0 and just mark the page as PageError on errors.  This
 318          * should be cleaned up all through the stack eventually.
 319          */
 320         return 0;
 321 }
 322 EXPORT_SYMBOL_GPL(iomap_readpage);
 323
 324 static struct page *
 325 iomap_next_page(struct inode *inode, struct list_head *pages, loff_t pos,
 326                 loff_t length, loff_t *done)
 327 {
 328         while (!list_empty(pages)) {
 329                 struct page *page = lru_to_page(pages);
 330
 331                 if (page_offset(page) >= (u64)pos + length)
 332                         break;
 333
 334                 list_del(&page->lru);
 335                 if (!add_to_page_cache_lru(page, inode->i_mapping, page->index,
 336                                 GFP_NOFS))
 337                         return page;
 338
 339                 /*
 340                  * If we already have a page in the page cache at index we are
 341                  * done.  Upper layers don't care if it is uptodate after the
 342                  * readpages call itself as every page gets checked again once
 343                  * actually needed.
 344                  */
 345                 *done += PAGE_SIZE;
 346                 put_page(page);
 347         }
 348
 349         return NULL;
 350 }
 351
 352 static loff_t
 353 iomap_readpages_actor(struct inode *inode, loff_t pos, loff_t length,
 354                 void *data, struct iomap *iomap)
 355 {
 356         struct iomap_readpage_ctx *ctx = data;
 357         loff_t done, ret;
 358
 359         for (done = 0; done < length; done += ret) {
 360                 if (ctx->cur_page && offset_in_page(pos + done) == 0) {
 361                         if (!ctx->cur_page_in_bio)
 362                                 unlock_page(ctx->cur_page);
 363                         put_page(ctx->cur_page);
 364                         ctx->cur_page = NULL;
 365                 }
 366                 if (!ctx->cur_page) {
 367                         ctx->cur_page = iomap_next_page(inode, ctx->pages,
 368                                         pos, length, &done);
 369                         if (!ctx->cur_page)
 370                                 break;
 371                         ctx->cur_page_in_bio = false;
 372                 }
 373                 ret = iomap_readpage_actor(inode, pos + done, length - done,
 374                                 ctx, iomap);
 375         }
 376
 377         return done;
 378 }
 379
 380 int
 381 iomap_readpages(struct address_space *mapping, struct list_head *pages,
 382                 unsigned nr_pages, const struct iomap_ops *ops)
 383 {
 384         struct iomap_readpage_ctx ctx = {
 385                 .pages          = pages,
 386                 .is_readahead   = true,
 387         };
 388         loff_t pos = page_offset(list_entry(pages->prev, struct page, lru));
 389         loff_t last = page_offset(list_entry(pages->next, struct page, lru));
 390         loff_t length = last - pos + PAGE_SIZE, ret = 0;
 391
 392         while (length > 0) {
 393                 ret = iomap_apply(mapping->host, pos, length, 0, ops,
 394                                 &ctx, iomap_readpages_actor);
 395                 if (ret <= 0) {
 396                         WARN_ON_ONCE(ret == 0);
 397                         goto done;
 398                 }
 399                 pos += ret;
 400                 length -= ret;
 401         }
 402         ret = 0;
 403 done:
 404         if (ctx.bio)
 405                 submit_bio(ctx.bio);
 406         if (ctx.cur_page) {
 407                 if (!ctx.cur_page_in_bio)
 408                         unlock_page(ctx.cur_page);
 409                 put_page(ctx.cur_page);
 410         }
 411
 412         /*
 413          * Check that we didn't lose a page due to the arcance calling
 414          * conventions..
 415          */
 416         WARN_ON_ONCE(!ret && !list_empty(ctx.pages));
 417         return ret;
 418 }
 419 EXPORT_SYMBOL_GPL(iomap_readpages);
 420
 421 /*
 422  * iomap_is_partially_uptodate checks whether blocks within a page are
 423  * uptodate or not.
 424  *
 425  * Returns true if all blocks which correspond to a file portion
 426  * we want to read within the page are uptodate.
 427  */
 428 int
 429 iomap_is_partially_uptodate(struct page *page, unsigned long from,
 430                 unsigned long count)
 431 {
 432         struct iomap_page *iop = to_iomap_page(page);
 433         struct inode *inode = page->mapping->host;
 434         unsigned len, first, last;
 435         unsigned i;
 436
 437         /* Limit range to one page */
 438         len = min_t(unsigned, PAGE_SIZE - from, count);
 439
 440         /* First and last blocks in range within page */
 441         first = from >> inode->i_blkbits;
 442         last = (from + len - 1) >> inode->i_blkbits;
 443
 444         if (iop) {
 445                 for (i = first; i <= last; i++)
 446                         if (!test_bit(i, iop->uptodate))
 447                                 return 0;
 448                 return 1;
 449         }
 450
 451         return 0;
 452 }
 453 EXPORT_SYMBOL_GPL(iomap_is_partially_uptodate);
 454
 455 int
 456 iomap_releasepage(struct page *page, gfp_t gfp_mask)
 457 {
 458         /*
 459          * mm accommodates an old ext3 case where clean pages might not have had
 460          * the dirty bit cleared. Thus, it can send actual dirty pages to
 461          * ->releasepage() via shrink_active_list(), skip those here.
 462          */
 463         if (PageDirty(page) || PageWriteback(page))
 464                 return 0;
 465         iomap_page_release(page);
 466         return 1;
 467 }
 468 EXPORT_SYMBOL_GPL(iomap_releasepage);
 469
 470 void
 471 iomap_invalidatepage(struct page *page, unsigned int offset, unsigned int len)
 472 {
 473         /*
 474          * If we are invalidating the entire page, clear the dirty state from it
 475          * and release it to avoid unnecessary buildup of the LRU.
 476          */
 477         if (offset == 0 && len == PAGE_SIZE) {
 478                 WARN_ON_ONCE(PageWriteback(page));
 479                 cancel_dirty_page(page);
 480                 iomap_page_release(page);
 481         }
 482 }
 483 EXPORT_SYMBOL_GPL(iomap_invalidatepage);
 484
 485 #ifdef CONFIG_MIGRATION
 486 int
 487 iomap_migrate_page(struct address_space *mapping, struct page *newpage,
 488                 struct page *page, enum migrate_mode mode)
 489 {
 490         int ret;
 491
 492         ret = migrate_page_move_mapping(mapping, newpage, page, 0);
 493         if (ret != MIGRATEPAGE_SUCCESS)
 494                 return ret;
 495
 496         if (page_has_private(page)) {
 497                 ClearPagePrivate(page);
 498                 get_page(newpage);
 499                 set_page_private(newpage, page_private(page));
 500                 set_page_private(page, 0);
 501                 put_page(page);
 502                 SetPagePrivate(newpage);
 503         }
 504
 505         if (mode != MIGRATE_SYNC_NO_COPY)
 506                 migrate_page_copy(newpage, page);
 507         else
 508                 migrate_page_states(newpage, page);
 509         return MIGRATEPAGE_SUCCESS;
 510 }
 511 EXPORT_SYMBOL_GPL(iomap_migrate_page);
 512 #endif /* CONFIG_MIGRATION */
 513
 514 static void
 515 iomap_write_failed(struct inode *inode, loff_t pos, unsigned len)
 516 {
 517         loff_t i_size = i_size_read(inode);
 518
 519         /*
 520          * Only truncate newly allocated pages beyoned EOF, even if the
 521          * write started inside the existing inode size.
 522          */
 523         if (pos + len > i_size)
 524                 truncate_pagecache_range(inode, max(pos, i_size), pos + len);
 525 }
 526
 527 static int
 528 iomap_read_page_sync(struct inode *inode, loff_t block_start, struct page *page,
 529                 unsigned poff, unsigned plen, unsigned from, unsigned to,
 530                 struct iomap *iomap)
 531 {
 532         struct bio_vec bvec;
 533         struct bio bio;
 534
 535         if (iomap->type != IOMAP_MAPPED || block_start >= i_size_read(inode)) {
 536                 zero_user_segments(page, poff, from, to, poff + plen);
 537                 iomap_set_range_uptodate(page, poff, plen);
 538                 return 0;
 539         }
 540
 541         bio_init(&bio, &bvec, 1);
 542         bio.bi_opf = REQ_OP_READ;
 543         bio.bi_iter.bi_sector = iomap_sector(iomap, block_start);
 544         bio_set_dev(&bio, iomap->bdev);
 545         __bio_add_page(&bio, page, plen, poff);
 546         return submit_bio_wait(&bio);
 547 }
 548
 549 static int
 550 __iomap_write_begin(struct inode *inode, loff_t pos, unsigned len,
 551                 struct page *page, struct iomap *iomap)
 552 {
 553         struct iomap_page *iop = iomap_page_create(inode, page);
 554         loff_t block_size = i_blocksize(inode);
 555         loff_t block_start = pos & ~(block_size - 1);
 556         loff_t block_end = (pos + len + block_size - 1) & ~(block_size - 1);
 557         unsigned from = offset_in_page(pos), to = from + len, poff, plen;
 558         int status = 0;
 559
 560         if (PageUptodate(page))
 561                 return 0;
 562
 563         do {
 564                 iomap_adjust_read_range(inode, iop, &block_start,
 565                                 block_end - block_start, &poff, &plen);
 566                 if (plen == 0)
 567                         break;
 568
 569                 if ((from > poff && from < poff + plen) ||
 570                     (to > poff && to < poff + plen)) {
 571                         status = iomap_read_page_sync(inode, block_start, page,
 572                                         poff, plen, from, to, iomap);
 573                         if (status)
 574                                 break;
 575                 }
 576
 577         } while ((block_start += plen) < block_end);
 578
 579         return status;
 580 }
 581
 582 static int
 583 iomap_write_begin(struct inode *inode, loff_t pos, unsigned len, unsigned flags,
 584                 struct page **pagep, struct iomap *iomap)
 585 {
 586         const struct iomap_page_ops *page_ops = iomap->page_ops;
 587         pgoff_t index = pos >> PAGE_SHIFT;
 588         struct page *page;
 589         int status = 0;
 590
 591         BUG_ON(pos + len > iomap->offset + iomap->length);
 592
 593         if (fatal_signal_pending(current))
 594                 return -EINTR;
 595
 596         if (page_ops && page_ops->page_prepare) {
 597                 status = page_ops->page_prepare(inode, pos, len, iomap);
 598                 if (status)
 599                         return status;
 600         }
 601
 602         page = grab_cache_page_write_begin(inode->i_mapping, index, flags);
 603         if (!page) {
 604                 status = -ENOMEM;
 605                 goto out_no_page;
 606         }
 607
 608         if (iomap->type == IOMAP_INLINE)
 609                 iomap_read_inline_data(inode, page, iomap);
 610         else if (iomap->flags & IOMAP_F_BUFFER_HEAD)
 611                 status = __block_write_begin_int(page, pos, len, NULL, iomap);
 612         else
 613                 status = __iomap_write_begin(inode, pos, len, page, iomap);
 614
 615         if (unlikely(status))
 616                 goto out_unlock;
 617
 618         *pagep = page;
 619         return 0;
 620
 621 out_unlock:
 622         unlock_page(page);
 623         put_page(page);
 624         iomap_write_failed(inode, pos, len);
 625
 626 out_no_page:
 627         if (page_ops && page_ops->page_done)
 628                 page_ops->page_done(inode, pos, 0, NULL, iomap);
 629         return status;
 630 }
 631
 632 int
 633 iomap_set_page_dirty(struct page *page)
 634 {
 635         struct address_space *mapping = page_mapping(page);
 636         int newly_dirty;
 637
 638         if (unlikely(!mapping))
 639                 return !TestSetPageDirty(page);
 640
 641         /*
 642          * Lock out page->mem_cgroup migration to keep PageDirty
 643          * synchronized with per-memcg dirty page counters.
 644          */
 645         lock_page_memcg(page);
 646         newly_dirty = !TestSetPageDirty(page);
 647         if (newly_dirty)
 648                 __set_page_dirty(page, mapping, 0);
 649         unlock_page_memcg(page);
 650
 651         if (newly_dirty)
 652                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 653         return newly_dirty;
 654 }
 655 EXPORT_SYMBOL_GPL(iomap_set_page_dirty);
 656
 657 static int
 658 __iomap_write_end(struct inode *inode, loff_t pos, unsigned len,
 659                 unsigned copied, struct page *page, struct iomap *iomap)
 660 {
 661         flush_dcache_page(page);
 662
 663         /*
 664          * The blocks that were entirely written will now be uptodate, so we
 665          * don't have to worry about a readpage reading them and overwriting a
 666          * partial write.  However if we have encountered a short write and only
 667          * partially written into a block, it will not be marked uptodate, so a
 668          * readpage might come in and destroy our partial write.
 669          *
 670          * Do the simplest thing, and just treat any short write to a non
 671          * uptodate page as a zero-length write, and force the caller to redo
 672          * the whole thing.
 673          */
 674         if (unlikely(copied < len && !PageUptodate(page)))
 675                 return 0;
 676         iomap_set_range_uptodate(page, offset_in_page(pos), len);
 677         iomap_set_page_dirty(page);
 678         return copied;
 679 }
 680
 681 static int
 682 iomap_write_end_inline(struct inode *inode, struct page *page,
 683                 struct iomap *iomap, loff_t pos, unsigned copied)
 684 {
 685         void *addr;
 686
 687         WARN_ON_ONCE(!PageUptodate(page));
 688         BUG_ON(pos + copied > PAGE_SIZE - offset_in_page(iomap->inline_data));
 689
 690         addr = kmap_atomic(page);
 691         memcpy(iomap->inline_data + pos, addr + pos, copied);
 692         kunmap_atomic(addr);
 693
 694         mark_inode_dirty(inode);
 695         return copied;
 696 }
 697
 698 static int
 699 iomap_write_end(struct inode *inode, loff_t pos, unsigned len,
 700                 unsigned copied, struct page *page, struct iomap *iomap)
 701 {
 702         const struct iomap_page_ops *page_ops = iomap->page_ops;
 703         loff_t old_size = inode->i_size;
 704         int ret;
 705
 706         if (iomap->type == IOMAP_INLINE) {
 707                 ret = iomap_write_end_inline(inode, page, iomap, pos, copied);
 708         } else if (iomap->flags & IOMAP_F_BUFFER_HEAD) {
 709                 ret = block_write_end(NULL, inode->i_mapping, pos, len, copied,
 710                                 page, NULL);
 711         } else {
 712                 ret = __iomap_write_end(inode, pos, len, copied, page, iomap);
 713         }
 714
 715         /*
 716          * Update the in-memory inode size after copying the data into the page
 717          * cache.  It's up to the file system to write the updated size to disk,
 718          * preferably after I/O completion so that no stale data is exposed.
 719          */
 720         if (pos + ret > old_size) {
 721                 i_size_write(inode, pos + ret);
 722                 iomap->flags |= IOMAP_F_SIZE_CHANGED;
 723         }
 724         unlock_page(page);
 725
 726         if (old_size < pos)
 727                 pagecache_isize_extended(inode, old_size, pos);
 728         if (page_ops && page_ops->page_done)
 729                 page_ops->page_done(inode, pos, ret, page, iomap);
 730         put_page(page);
 731
 732         if (ret < len)
 733                 iomap_write_failed(inode, pos, len);
 734         return ret;
 735 }
 736
 737 static loff_t
 738 iomap_write_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 739                 struct iomap *iomap)
 740 {
 741         struct iov_iter *i = data;
 742         long status = 0;
 743         ssize_t written = 0;
 744         unsigned int flags = AOP_FLAG_NOFS;
 745
 746         do {
 747                 struct page *page;
 748                 unsigned long offset;   /* Offset into pagecache page */
 749                 unsigned long bytes;    /* Bytes to write to page */
 750                 size_t copied;          /* Bytes copied from user */
 751
 752                 offset = offset_in_page(pos);
 753                 bytes = min_t(unsigned long, PAGE_SIZE - offset,
 754                                                 iov_iter_count(i));
 755 again:
 756                 if (bytes > length)
 757                         bytes = length;
 758
 759                 /*
 760                  * Bring in the user page that we will copy from _first_.
 761                  * Otherwise there's a nasty deadlock on copying from the
 762                  * same page as we're writing to, without it being marked
 763                  * up-to-date.
 764                  *
 765                  * Not only is this an optimisation, but it is also required
 766                  * to check that the address is actually valid, when atomic
 767                  * usercopies are used, below.
 768                  */
 769                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
 770                         status = -EFAULT;
 771                         break;
 772                 }
 773
 774                 status = iomap_write_begin(inode, pos, bytes, flags, &page,
 775                                 iomap);
 776                 if (unlikely(status))
 777                         break;
 778
 779                 if (mapping_writably_mapped(inode->i_mapping))
 780                         flush_dcache_page(page);
 781
 782                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
 783
 784                 flush_dcache_page(page);
 785
 786                 status = iomap_write_end(inode, pos, bytes, copied, page,
 787                                 iomap);
 788                 if (unlikely(status < 0))
 789                         break;
 790                 copied = status;
 791
 792                 cond_resched();
 793
 794                 iov_iter_advance(i, copied);
 795                 if (unlikely(copied == 0)) {
 796                         /*
 797                          * If we were unable to copy any data at all, we must
 798                          * fall back to a single segment length write.
 799                          *
 800                          * If we didn't fallback here, we could livelock
 801                          * because not all segments in the iov can be copied at
 802                          * once without a pagefault.
 803                          */
 804                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
 805                                                 iov_iter_single_seg_count(i));
 806                         goto again;
 807                 }
 808                 pos += copied;
 809                 written += copied;
 810                 length -= copied;
 811
 812                 balance_dirty_pages_ratelimited(inode->i_mapping);
 813         } while (iov_iter_count(i) && length);
 814
 815         return written ? written : status;
 816 }
 817
 818 ssize_t
 819 iomap_file_buffered_write(struct kiocb *iocb, struct iov_iter *iter,
 820                 const struct iomap_ops *ops)
 821 {
 822         struct inode *inode = iocb->ki_filp->f_mapping->host;
 823         loff_t pos = iocb->ki_pos, ret = 0, written = 0;
 824
 825         while (iov_iter_count(iter)) {
 826                 ret = iomap_apply(inode, pos, iov_iter_count(iter),
 827                                 IOMAP_WRITE, ops, iter, iomap_write_actor);
 828                 if (ret <= 0)
 829                         break;
 830                 pos += ret;
 831                 written += ret;
 832         }
 833
 834         return written ? written : ret;
 835 }
 836 EXPORT_SYMBOL_GPL(iomap_file_buffered_write);
 837
 838 static struct page *
 839 __iomap_read_page(struct inode *inode, loff_t offset)
 840 {
 841         struct address_space *mapping = inode->i_mapping;
 842         struct page *page;
 843
 844         page = read_mapping_page(mapping, offset >> PAGE_SHIFT, NULL);
 845         if (IS_ERR(page))
 846                 return page;
 847         if (!PageUptodate(page)) {
 848                 put_page(page);
 849                 return ERR_PTR(-EIO);
 850         }
 851         return page;
 852 }
 853
 854 static loff_t
 855 iomap_dirty_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
 856                 struct iomap *iomap)
 857 {
 858         long status = 0;
 859         ssize_t written = 0;
 860
 861         do {
 862                 struct page *page, *rpage;
 863                 unsigned long offset;   /* Offset into pagecache page */
 864                 unsigned long bytes;    /* Bytes to write to page */
 865
 866                 offset = offset_in_page(pos);
 867                 bytes = min_t(loff_t, PAGE_SIZE - offset, length);
 868
 869                 rpage = __iomap_read_page(inode, pos);
 870                 if (IS_ERR(rpage))
 871                         return PTR_ERR(rpage);
 872
 873                 status = iomap_write_begin(inode, pos, bytes,
 874                                            AOP_FLAG_NOFS, &page, iomap);
 875                 put_page(rpage);
 876                 if (unlikely(status))
 877                         return status;
 878
 879                 WARN_ON_ONCE(!PageUptodate(page));
 880
 881                 status = iomap_write_end(inode, pos, bytes, bytes, page, iomap);
 882                 if (unlikely(status <= 0)) {
 883                         if (WARN_ON_ONCE(status == 0))
 884                                 return -EIO;
 885                         return status;
 886                 }
 887
 888                 cond_resched();
 889
 890                 pos += status;
 891                 written += status;
 892                 length -= status;
 893
 894                 balance_dirty_pages_ratelimited(inode->i_mapping);
 895         } while (length);
 896
 897         return written;
 898 }
 899
 900 int
 901 iomap_file_dirty(struct inode *inode, loff_t pos, loff_t len,
 902                 const struct iomap_ops *ops)
 903 {
 904         loff_t ret;
 905
 906         while (len) {
 907                 ret = iomap_apply(inode, pos, len, IOMAP_WRITE, ops, NULL,
 908                                 iomap_dirty_actor);
 909                 if (ret <= 0)
 910                         return ret;
 911                 pos += ret;
 912                 len -= ret;
 913         }
 914
 915         return 0;
 916 }
 917 EXPORT_SYMBOL_GPL(iomap_file_dirty);
 918
 919 static int iomap_zero(struct inode *inode, loff_t pos, unsigned offset,
 920                 unsigned bytes, struct iomap *iomap)
 921 {
 922         struct page *page;
 923         int status;
 924
 925         status = iomap_write_begin(inode, pos, bytes, AOP_FLAG_NOFS, &page,
 926                                    iomap);
 927         if (status)
 928                 return status;
 929
 930         zero_user(page, offset, bytes);
 931         mark_page_accessed(page);
 932
 933         return iomap_write_end(inode, pos, bytes, bytes, page, iomap);
 934 }
 935
 936 static int iomap_dax_zero(loff_t pos, unsigned offset, unsigned bytes,
 937                 struct iomap *iomap)
 938 {
 939         return __dax_zero_page_range(iomap->bdev, iomap->dax_dev,
 940                         iomap_sector(iomap, pos & PAGE_MASK), offset, bytes);
 941 }
 942
 943 static loff_t
 944 iomap_zero_range_actor(struct inode *inode, loff_t pos, loff_t count,
 945                 void *data, struct iomap *iomap)
 946 {
 947         bool *did_zero = data;
 948         loff_t written = 0;
 949         int status;
 950
 951         /* already zeroed?  we're done. */
 952         if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
 953                 return count;
 954
 955         do {
 956                 unsigned offset, bytes;
 957
 958                 offset = offset_in_page(pos);
 959                 bytes = min_t(loff_t, PAGE_SIZE - offset, count);
 960
 961                 if (IS_DAX(inode))
 962                         status = iomap_dax_zero(pos, offset, bytes, iomap);
 963                 else
 964                         status = iomap_zero(inode, pos, offset, bytes, iomap);
 965                 if (status < 0)
 966                         return status;
 967
 968                 pos += bytes;
 969                 count -= bytes;
 970                 written += bytes;
 971                 if (did_zero)
 972                         *did_zero = true;
 973         } while (count > 0);
 974
 975         return written;
 976 }
 977
 978 int
 979 iomap_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
 980                 const struct iomap_ops *ops)
 981 {
 982         loff_t ret;
 983
 984         while (len > 0) {
 985                 ret = iomap_apply(inode, pos, len, IOMAP_ZERO,
 986                                 ops, did_zero, iomap_zero_range_actor);
 987                 if (ret <= 0)
 988                         return ret;
 989
 990                 pos += ret;
 991                 len -= ret;
 992         }
 993
 994         return 0;
 995 }
 996 EXPORT_SYMBOL_GPL(iomap_zero_range);
 997
 998 int
 999 iomap_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
1000                 const struct iomap_ops *ops)
1001 {
1002         unsigned int blocksize = i_blocksize(inode);
1003         unsigned int off = pos & (blocksize - 1);
1004
1005         /* Block boundary? Nothing to do */
1006         if (!off)
1007                 return 0;
1008         return iomap_zero_range(inode, pos, blocksize - off, did_zero, ops);
1009 }
1010 EXPORT_SYMBOL_GPL(iomap_truncate_page);
1011
1012 static loff_t
1013 iomap_page_mkwrite_actor(struct inode *inode, loff_t pos, loff_t length,
1014                 void *data, struct iomap *iomap)
1015 {
1016         struct page *page = data;
1017         int ret;
1018
1019         if (iomap->flags & IOMAP_F_BUFFER_HEAD) {
1020                 ret = __block_write_begin_int(page, pos, length, NULL, iomap);
1021                 if (ret)
1022                         return ret;
1023                 block_commit_write(page, 0, length);
1024         } else {
1025                 WARN_ON_ONCE(!PageUptodate(page));
1026                 iomap_page_create(inode, page);
1027                 set_page_dirty(page);
1028         }
1029
1030         return length;
1031 }
1032
1033 vm_fault_t iomap_page_mkwrite(struct vm_fault *vmf, const struct iomap_ops *ops)
1034 {
1035         struct page *page = vmf->page;
1036         struct inode *inode = file_inode(vmf->vma->vm_file);
1037         unsigned long length;
1038         loff_t offset, size;
1039         ssize_t ret;
1040
1041         lock_page(page);
1042         size = i_size_read(inode);
1043         if ((page->mapping != inode->i_mapping) ||
1044             (page_offset(page) > size)) {
1045                 /* We overload EFAULT to mean page got truncated */
1046                 ret = -EFAULT;
1047                 goto out_unlock;
1048         }
1049
1050         /* page is wholly or partially inside EOF */
1051         if (((page->index + 1) << PAGE_SHIFT) > size)
1052                 length = offset_in_page(size);
1053         else
1054                 length = PAGE_SIZE;
1055
1056         offset = page_offset(page);
1057         while (length > 0) {
1058                 ret = iomap_apply(inode, offset, length,
1059                                 IOMAP_WRITE | IOMAP_FAULT, ops, page,
1060                                 iomap_page_mkwrite_actor);
1061                 if (unlikely(ret <= 0))
1062                         goto out_unlock;
1063                 offset += ret;
1064                 length -= ret;
1065         }
1066
1067         wait_for_stable_page(page);
1068         return VM_FAULT_LOCKED;
1069 out_unlock:
1070         unlock_page(page);
1071         return block_page_mkwrite_return(ret);
1072 }
1073 EXPORT_SYMBOL_GPL(iomap_page_mkwrite);