file.c

   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * This file is part of UBIFS.
   4 *
   5 * Copyright (C) 2006-2008 Nokia Corporation.
   6 *
   7 * Authors: Artem Bityutskiy (Битюцкий Артём)
   8 *          Adrian Hunter
   9 */
  10
  11/*
  12 * This file implements VFS file and inode operations for regular files, device
  13 * nodes and symlinks as well as address space operations.
  14 *
  15 * UBIFS uses 2 page flags: @PG_private and @PG_checked. @PG_private is set if
  16 * the page is dirty and is used for optimization purposes - dirty pages are
  17 * not budgeted so the flag shows that 'ubifs_write_end()' should not release
  18 * the budget for this page. The @PG_checked flag is set if full budgeting is
  19 * required for the page e.g., when it corresponds to a file hole or it is
  20 * beyond the file size. The budgeting is done in 'ubifs_write_begin()', because
  21 * it is OK to fail in this function, and the budget is released in
  22 * 'ubifs_write_end()'. So the @PG_private and @PG_checked flags carry
  23 * information about how the page was budgeted, to make it possible to release
  24 * the budget properly.
  25 *
  26 * A thing to keep in mind: inode @i_mutex is locked in most VFS operations we
  27 * implement. However, this is not true for 'ubifs_writepage()', which may be
  28 * called with @i_mutex unlocked. For example, when flusher thread is doing
  29 * background write-back, it calls 'ubifs_writepage()' with unlocked @i_mutex.
  30 * At "normal" work-paths the @i_mutex is locked in 'ubifs_writepage()', e.g.
  31 * in the "sys_write -> alloc_pages -> direct reclaim path". So, in
  32 * 'ubifs_writepage()' we are only guaranteed that the page is locked.
  33 *
  34 * Similarly, @i_mutex is not always locked in 'ubifs_read_folio()', e.g., the
  35 * read-ahead path does not lock it ("sys_read -> generic_file_aio_read ->
  36 * ondemand_readahead -> read_folio"). In case of readahead, @I_SYNC flag is not
  37 * set as well. However, UBIFS disables readahead.
  38 */
  39
  40#include "ubifs.h"
  41#include <linux/mount.h>
  42#include <linux/slab.h>
  43#include <linux/migrate.h>
  44
  45static int read_block(struct inode *inode, struct folio *folio, size_t offset,
  46		      unsigned int block, struct ubifs_data_node *dn)
  47{
  48	struct ubifs_info *c = inode->i_sb->s_fs_info;
  49	int err, len, out_len;
  50	union ubifs_key key;
  51	unsigned int dlen;
  52
  53	data_key_init(c, &key, inode->i_ino, block);
  54	err = ubifs_tnc_lookup(c, &key, dn);
  55	if (err) {
  56		if (err == -ENOENT)
  57			/* Not found, so it must be a hole */
  58			folio_zero_range(folio, offset, UBIFS_BLOCK_SIZE);
  59		return err;
  60	}
  61
  62	ubifs_assert(c, le64_to_cpu(dn->ch.sqnum) >
  63		     ubifs_inode(inode)->creat_sqnum);
  64	len = le32_to_cpu(dn->size);
  65	if (len <= 0 || len > UBIFS_BLOCK_SIZE)
  66		goto dump;
  67
  68	dlen = le32_to_cpu(dn->ch.len) - UBIFS_DATA_NODE_SZ;
  69
  70	if (IS_ENCRYPTED(inode)) {
  71		err = ubifs_decrypt(inode, dn, &dlen, block);
  72		if (err)
  73			goto dump;
  74	}
  75
  76	out_len = UBIFS_BLOCK_SIZE;
  77	err = ubifs_decompress_folio(c, &dn->data, dlen, folio, offset,
  78				     &out_len, le16_to_cpu(dn->compr_type));
  79	if (err || len != out_len)
  80		goto dump;
  81
  82	/*
  83	 * Data length can be less than a full block, even for blocks that are
  84	 * not the last in the file (e.g., as a result of making a hole and
  85	 * appending data). Ensure that the remainder is zeroed out.
  86	 */
  87	if (len < UBIFS_BLOCK_SIZE)
  88		folio_zero_range(folio, offset + len, UBIFS_BLOCK_SIZE - len);
  89
  90	return 0;
  91
  92dump:
  93	ubifs_err(c, "bad data node (block %u, inode %lu)",
  94		  block, inode->i_ino);
  95	ubifs_dump_node(c, dn, UBIFS_MAX_DATA_NODE_SZ);
  96	return -EINVAL;
  97}
  98
  99static int do_readpage(struct folio *folio)
 100{
 101	int err = 0, i;
 102	unsigned int block, beyond;
 103	struct ubifs_data_node *dn = NULL;
 104	struct inode *inode = folio->mapping->host;
 105	struct ubifs_info *c = inode->i_sb->s_fs_info;
 106	loff_t i_size = i_size_read(inode);
 107	size_t offset = 0;
 108
 109	dbg_gen("ino %lu, pg %lu, i_size %lld, flags %#lx",
 110		inode->i_ino, folio->index, i_size, folio->flags.f);
 111	ubifs_assert(c, !folio_test_checked(folio));
 112	ubifs_assert(c, !folio->private);
 113
 114	block = folio->index << UBIFS_BLOCKS_PER_PAGE_SHIFT;
 115	beyond = (i_size + UBIFS_BLOCK_SIZE - 1) >> UBIFS_BLOCK_SHIFT;
 116	if (block >= beyond) {
 117		/* Reading beyond inode */
 118		folio_set_checked(folio);
 119		folio_zero_range(folio, 0, folio_size(folio));
 120		goto out;
 121	}
 122
 123	dn = kmalloc(UBIFS_MAX_DATA_NODE_SZ, GFP_NOFS);
 124	if (!dn) {
 125		err = -ENOMEM;
 126		goto out;
 127	}
 128
 129	i = 0;
 130	while (1) {
 131		int ret;
 132
 133		if (block >= beyond) {
 134			/* Reading beyond inode */
 135			err = -ENOENT;
 136			folio_zero_range(folio, offset, UBIFS_BLOCK_SIZE);
 137		} else {
 138			ret = read_block(inode, folio, offset, block, dn);
 139			if (ret) {
 140				err = ret;
 141				if (err != -ENOENT)
 142					break;
 143			} else if (block + 1 == beyond) {
 144				int dlen = le32_to_cpu(dn->size);
 145				int ilen = i_size & (UBIFS_BLOCK_SIZE - 1);
 146
 147				if (ilen && ilen < dlen)
 148					folio_zero_range(folio, offset + ilen, dlen - ilen);
 149			}
 150		}
 151		if (++i >= (UBIFS_BLOCKS_PER_PAGE << folio_order(folio)))
 152			break;
 153		block += 1;
 154		offset += UBIFS_BLOCK_SIZE;
 155	}
 156
 157	if (err) {
 158		struct ubifs_info *c = inode->i_sb->s_fs_info;
 159		if (err == -ENOENT) {
 160			/* Not found, so it must be a hole */
 161			folio_set_checked(folio);
 162			dbg_gen("hole");
 163			err = 0;
 164		} else {
 165			ubifs_err(c, "cannot read page %lu of inode %lu, error %d",
 166				  folio->index, inode->i_ino, err);
 167		}
 168	}
 169
 170out:
 171	kfree(dn);
 172	if (!err)
 173		folio_mark_uptodate(folio);
 174	return err;
 175}
 176
 177/**
 178 * release_new_page_budget - release budget of a new page.
 179 * @c: UBIFS file-system description object
 180 *
 181 * This is a helper function which releases budget corresponding to the budget
 182 * of one new page of data.
 183 */
 184static void release_new_page_budget(struct ubifs_info *c)
 185{
 186	struct ubifs_budget_req req = { .recalculate = 1, .new_page = 1 };
 187
 188	ubifs_release_budget(c, &req);
 189}
 190
 191/**
 192 * release_existing_page_budget - release budget of an existing page.
 193 * @c: UBIFS file-system description object
 194 *
 195 * This is a helper function which releases budget corresponding to the budget
 196 * of changing one page of data which already exists on the flash media.
 197 */
 198static void release_existing_page_budget(struct ubifs_info *c)
 199{
 200	struct ubifs_budget_req req = { .dd_growth = c->bi.page_budget};
 201
 202	ubifs_release_budget(c, &req);
 203}
 204
 205static int write_begin_slow(struct address_space *mapping,
 206			    loff_t pos, unsigned len, struct folio **foliop)
 207{
 208	struct inode *inode = mapping->host;
 209	struct ubifs_info *c = inode->i_sb->s_fs_info;
 210	pgoff_t index = pos >> PAGE_SHIFT;
 211	struct ubifs_budget_req req = { .new_page = 1 };
 212	int err, appending = !!(pos + len > inode->i_size);
 213	struct folio *folio;
 214
 215	dbg_gen("ino %lu, pos %llu, len %u, i_size %lld",
 216		inode->i_ino, pos, len, inode->i_size);
 217
 218	/*
 219	 * At the slow path we have to budget before locking the folio, because
 220	 * budgeting may force write-back, which would wait on locked folios and
 221	 * deadlock if we had the folio locked. At this point we do not know
 222	 * anything about the folio, so assume that this is a new folio which is
 223	 * written to a hole. This corresponds to largest budget. Later the
 224	 * budget will be amended if this is not true.
 225	 */
 226	if (appending)
 227		/* We are appending data, budget for inode change */
 228		req.dirtied_ino = 1;
 229
 230	err = ubifs_budget_space(c, &req);
 231	if (unlikely(err))
 232		return err;
 233
 234	folio = __filemap_get_folio(mapping, index, FGP_WRITEBEGIN,
 235			mapping_gfp_mask(mapping));
 236	if (IS_ERR(folio)) {
 237		ubifs_release_budget(c, &req);
 238		return PTR_ERR(folio);
 239	}
 240
 241	if (!folio_test_uptodate(folio)) {
 242		if (pos == folio_pos(folio) && len >= folio_size(folio))
 243			folio_set_checked(folio);
 244		else {
 245			err = do_readpage(folio);
 246			if (err) {
 247				folio_unlock(folio);
 248				folio_put(folio);
 249				ubifs_release_budget(c, &req);
 250				return err;
 251			}
 252		}
 253	}
 254
 255	if (folio->private)
 256		/*
 257		 * The folio is dirty, which means it was budgeted twice:
 258		 *   o first time the budget was allocated by the task which
 259		 *     made the folio dirty and set the private field;
 260		 *   o and then we budgeted for it for the second time at the
 261		 *     very beginning of this function.
 262		 *
 263		 * So what we have to do is to release the folio budget we
 264		 * allocated.
 265		 */
 266		release_new_page_budget(c);
 267	else if (!folio_test_checked(folio))
 268		/*
 269		 * We are changing a folio which already exists on the media.
 270		 * This means that changing the folio does not make the amount
 271		 * of indexing information larger, and this part of the budget
 272		 * which we have already acquired may be released.
 273		 */
 274		ubifs_convert_page_budget(c);
 275
 276	if (appending) {
 277		struct ubifs_inode *ui = ubifs_inode(inode);
 278
 279		/*
 280		 * 'ubifs_write_end()' is optimized from the fast-path part of
 281		 * 'ubifs_write_begin()' and expects the @ui_mutex to be locked
 282		 * if data is appended.
 283		 */
 284		mutex_lock(&ui->ui_mutex);
 285		if (ui->dirty)
 286			/*
 287			 * The inode is dirty already, so we may free the
 288			 * budget we allocated.
 289			 */
 290			ubifs_release_dirty_inode_budget(c, ui);
 291	}
 292
 293	*foliop = folio;
 294	return 0;
 295}
 296
 297/**
 298 * allocate_budget - allocate budget for 'ubifs_write_begin()'.
 299 * @c: UBIFS file-system description object
 300 * @folio: folio to allocate budget for
 301 * @ui: UBIFS inode object the page belongs to
 302 * @appending: non-zero if the page is appended
 303 *
 304 * This is a helper function for 'ubifs_write_begin()' which allocates budget
 305 * for the operation. The budget is allocated differently depending on whether
 306 * this is appending, whether the page is dirty or not, and so on. This
 307 * function leaves the @ui->ui_mutex locked in case of appending.
 308 *
 309 * Returns: %0 in case of success and %-ENOSPC in case of failure.
 310 */
 311static int allocate_budget(struct ubifs_info *c, struct folio *folio,
 312			   struct ubifs_inode *ui, int appending)
 313{
 314	struct ubifs_budget_req req = { .fast = 1 };
 315
 316	if (folio->private) {
 317		if (!appending)
 318			/*
 319			 * The folio is dirty and we are not appending, which
 320			 * means no budget is needed at all.
 321			 */
 322			return 0;
 323
 324		mutex_lock(&ui->ui_mutex);
 325		if (ui->dirty)
 326			/*
 327			 * The page is dirty and we are appending, so the inode
 328			 * has to be marked as dirty. However, it is already
 329			 * dirty, so we do not need any budget. We may return,
 330			 * but @ui->ui_mutex hast to be left locked because we
 331			 * should prevent write-back from flushing the inode
 332			 * and freeing the budget. The lock will be released in
 333			 * 'ubifs_write_end()'.
 334			 */
 335			return 0;
 336
 337		/*
 338		 * The page is dirty, we are appending, the inode is clean, so
 339		 * we need to budget the inode change.
 340		 */
 341		req.dirtied_ino = 1;
 342	} else {
 343		if (folio_test_checked(folio))
 344			/*
 345			 * The page corresponds to a hole and does not
 346			 * exist on the media. So changing it makes
 347			 * the amount of indexing information
 348			 * larger, and we have to budget for a new
 349			 * page.
 350			 */
 351			req.new_page = 1;
 352		else
 353			/*
 354			 * Not a hole, the change will not add any new
 355			 * indexing information, budget for page
 356			 * change.
 357			 */
 358			req.dirtied_page = 1;
 359
 360		if (appending) {
 361			mutex_lock(&ui->ui_mutex);
 362			if (!ui->dirty)
 363				/*
 364				 * The inode is clean but we will have to mark
 365				 * it as dirty because we are appending. This
 366				 * needs a budget.
 367				 */
 368				req.dirtied_ino = 1;
 369		}
 370	}
 371
 372	return ubifs_budget_space(c, &req);
 373}
 374
 375/*
 376 * This function is called when a page of data is going to be written. Since
 377 * the page of data will not necessarily go to the flash straight away, UBIFS
 378 * has to reserve space on the media for it, which is done by means of
 379 * budgeting.
 380 *
 381 * This is the hot-path of the file-system and we are trying to optimize it as
 382 * much as possible. For this reasons it is split on 2 parts - slow and fast.
 383 *
 384 * There many budgeting cases:
 385 *     o a new page is appended - we have to budget for a new page and for
 386 *       changing the inode; however, if the inode is already dirty, there is
 387 *       no need to budget for it;
 388 *     o an existing clean page is changed - we have budget for it; if the page
 389 *       does not exist on the media (a hole), we have to budget for a new
 390 *       page; otherwise, we may budget for changing an existing page; the
 391 *       difference between these cases is that changing an existing page does
 392 *       not introduce anything new to the FS indexing information, so it does
 393 *       not grow, and smaller budget is acquired in this case;
 394 *     o an existing dirty page is changed - no need to budget at all, because
 395 *       the page budget has been acquired by earlier, when the page has been
 396 *       marked dirty.
 397 *
 398 * UBIFS budgeting sub-system may force write-back if it thinks there is no
 399 * space to reserve. This imposes some locking restrictions and makes it
 400 * impossible to take into account the above cases, and makes it impossible to
 401 * optimize budgeting.
 402 *
 403 * The solution for this is that the fast path of 'ubifs_write_begin()' assumes
 404 * there is a plenty of flash space and the budget will be acquired quickly,
 405 * without forcing write-back. The slow path does not make this assumption.
 406 */
 407static int ubifs_write_begin(const struct kiocb *iocb,
 408			     struct address_space *mapping,
 409			     loff_t pos, unsigned len,
 410			     struct folio **foliop, void **fsdata)
 411{
 412	struct inode *inode = mapping->host;
 413	struct ubifs_info *c = inode->i_sb->s_fs_info;
 414	struct ubifs_inode *ui = ubifs_inode(inode);
 415	pgoff_t index = pos >> PAGE_SHIFT;
 416	int err, appending = !!(pos + len > inode->i_size);
 417	int skipped_read = 0;
 418	struct folio *folio;
 419
 420	ubifs_assert(c, ubifs_inode(inode)->ui_size == inode->i_size);
 421	ubifs_assert(c, !c->ro_media && !c->ro_mount);
 422
 423	if (unlikely(c->ro_error))
 424		return -EROFS;
 425
 426	/* Try out the fast-path part first */
 427	folio = __filemap_get_folio(mapping, index, FGP_WRITEBEGIN,
 428			mapping_gfp_mask(mapping));
 429	if (IS_ERR(folio))
 430		return PTR_ERR(folio);
 431
 432	if (!folio_test_uptodate(folio)) {
 433		/* The page is not loaded from the flash */
 434		if (pos == folio_pos(folio) && len >= folio_size(folio)) {
 435			/*
 436			 * We change whole page so no need to load it. But we
 437			 * do not know whether this page exists on the media or
 438			 * not, so we assume the latter because it requires
 439			 * larger budget. The assumption is that it is better
 440			 * to budget a bit more than to read the page from the
 441			 * media. Thus, we are setting the @PG_checked flag
 442			 * here.
 443			 */
 444			folio_set_checked(folio);
 445			skipped_read = 1;
 446		} else {
 447			err = do_readpage(folio);
 448			if (err) {
 449				folio_unlock(folio);
 450				folio_put(folio);
 451				return err;
 452			}
 453		}
 454	}
 455
 456	err = allocate_budget(c, folio, ui, appending);
 457	if (unlikely(err)) {
 458		ubifs_assert(c, err == -ENOSPC);
 459		/*
 460		 * If we skipped reading the page because we were going to
 461		 * write all of it, then it is not up to date.
 462		 */
 463		if (skipped_read)
 464			folio_clear_checked(folio);
 465		/*
 466		 * Budgeting failed which means it would have to force
 467		 * write-back but didn't, because we set the @fast flag in the
 468		 * request. Write-back cannot be done now, while we have the
 469		 * page locked, because it would deadlock. Unlock and free
 470		 * everything and fall-back to slow-path.
 471		 */
 472		if (appending) {
 473			ubifs_assert(c, mutex_is_locked(&ui->ui_mutex));
 474			mutex_unlock(&ui->ui_mutex);
 475		}
 476		folio_unlock(folio);
 477		folio_put(folio);
 478
 479		return write_begin_slow(mapping, pos, len, foliop);
 480	}
 481
 482	/*
 483	 * Whee, we acquired budgeting quickly - without involving
 484	 * garbage-collection, committing or forcing write-back. We return
 485	 * with @ui->ui_mutex locked if we are appending pages, and unlocked
 486	 * otherwise. This is an optimization (slightly hacky though).
 487	 */
 488	*foliop = folio;
 489	return 0;
 490}
 491
 492/**
 493 * cancel_budget - cancel budget.
 494 * @c: UBIFS file-system description object
 495 * @folio: folio to cancel budget for
 496 * @ui: UBIFS inode object the page belongs to
 497 * @appending: non-zero if the page is appended
 498 *
 499 * This is a helper function for a page write operation. It unlocks the
 500 * @ui->ui_mutex in case of appending.
 501 */
 502static void cancel_budget(struct ubifs_info *c, struct folio *folio,
 503			  struct ubifs_inode *ui, int appending)
 504{
 505	if (appending) {
 506		if (!ui->dirty)
 507			ubifs_release_dirty_inode_budget(c, ui);
 508		mutex_unlock(&ui->ui_mutex);
 509	}
 510	if (!folio->private) {
 511		if (folio_test_checked(folio))
 512			release_new_page_budget(c);
 513		else
 514			release_existing_page_budget(c);
 515	}
 516}
 517
 518static int ubifs_write_end(const struct kiocb *iocb,
 519			   struct address_space *mapping, loff_t pos,
 520			   unsigned len, unsigned copied,
 521			   struct folio *folio, void *fsdata)
 522{
 523	struct inode *inode = mapping->host;
 524	struct ubifs_inode *ui = ubifs_inode(inode);
 525	struct ubifs_info *c = inode->i_sb->s_fs_info;
 526	loff_t end_pos = pos + len;
 527	int appending = !!(end_pos > inode->i_size);
 528
 529	dbg_gen("ino %lu, pos %llu, pg %lu, len %u, copied %d, i_size %lld",
 530		inode->i_ino, pos, folio->index, len, copied, inode->i_size);
 531
 532	if (unlikely(copied < len && !folio_test_uptodate(folio))) {
 533		/*
 534		 * VFS copied less data to the folio than it intended and
 535		 * declared in its '->write_begin()' call via the @len
 536		 * argument. If the folio was not up-to-date,
 537		 * the 'ubifs_write_begin()' function did
 538		 * not load it from the media (for optimization reasons). This
 539		 * means that part of the folio contains garbage. So read the
 540		 * folio now.
 541		 */
 542		dbg_gen("copied %d instead of %d, read page and repeat",
 543			copied, len);
 544		cancel_budget(c, folio, ui, appending);
 545		folio_clear_checked(folio);
 546
 547		/*
 548		 * Return 0 to force VFS to repeat the whole operation, or the
 549		 * error code if 'do_readpage()' fails.
 550		 */
 551		copied = do_readpage(folio);
 552		goto out;
 553	}
 554
 555	if (len == folio_size(folio))
 556		folio_mark_uptodate(folio);
 557
 558	if (!folio->private) {
 559		folio_attach_private(folio, (void *)1);
 560		atomic_long_inc(&c->dirty_pg_cnt);
 561		filemap_dirty_folio(mapping, folio);
 562	}
 563
 564	if (appending) {
 565		i_size_write(inode, end_pos);
 566		ui->ui_size = end_pos;
 567		/*
 568		 * We do not set @I_DIRTY_PAGES (which means that
 569		 * the inode has dirty pages), this was done in
 570		 * filemap_dirty_folio().
 571		 */
 572		__mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 573		ubifs_assert(c, mutex_is_locked(&ui->ui_mutex));
 574		mutex_unlock(&ui->ui_mutex);
 575	}
 576
 577out:
 578	folio_unlock(folio);
 579	folio_put(folio);
 580	return copied;
 581}
 582
 583/**
 584 * populate_page - copy data nodes into a page for bulk-read.
 585 * @c: UBIFS file-system description object
 586 * @folio: folio
 587 * @bu: bulk-read information
 588 * @n: next zbranch slot
 589 *
 590 * Returns: %0 on success and a negative error code on failure.
 591 */
 592static int populate_page(struct ubifs_info *c, struct folio *folio,
 593			 struct bu_info *bu, int *n)
 594{
 595	int i = 0, nn = *n, offs = bu->zbranch[0].offs, hole = 0, read = 0;
 596	struct inode *inode = folio->mapping->host;
 597	loff_t i_size = i_size_read(inode);
 598	unsigned int page_block;
 599	size_t offset = 0;
 600	pgoff_t end_index;
 601
 602	dbg_gen("ino %lu, pg %lu, i_size %lld, flags %#lx",
 603		inode->i_ino, folio->index, i_size, folio->flags.f);
 604
 605	end_index = (i_size - 1) >> PAGE_SHIFT;
 606	if (!i_size || folio->index > end_index) {
 607		hole = 1;
 608		folio_zero_range(folio, 0, folio_size(folio));
 609		goto out_hole;
 610	}
 611
 612	page_block = folio->index << UBIFS_BLOCKS_PER_PAGE_SHIFT;
 613	while (1) {
 614		int err, len, out_len, dlen;
 615
 616		if (nn >= bu->cnt) {
 617			hole = 1;
 618			folio_zero_range(folio, offset, UBIFS_BLOCK_SIZE);
 619		} else if (key_block(c, &bu->zbranch[nn].key) == page_block) {
 620			struct ubifs_data_node *dn;
 621
 622			dn = bu->buf + (bu->zbranch[nn].offs - offs);
 623
 624			ubifs_assert(c, le64_to_cpu(dn->ch.sqnum) >
 625				     ubifs_inode(inode)->creat_sqnum);
 626
 627			len = le32_to_cpu(dn->size);
 628			if (len <= 0 || len > UBIFS_BLOCK_SIZE)
 629				goto out_err;
 630
 631			dlen = le32_to_cpu(dn->ch.len) - UBIFS_DATA_NODE_SZ;
 632			out_len = UBIFS_BLOCK_SIZE;
 633
 634			if (IS_ENCRYPTED(inode)) {
 635				err = ubifs_decrypt(inode, dn, &dlen, page_block);
 636				if (err)
 637					goto out_err;
 638			}
 639
 640			err = ubifs_decompress_folio(
 641				c, &dn->data, dlen, folio, offset, &out_len,
 642				le16_to_cpu(dn->compr_type));
 643			if (err || len != out_len)
 644				goto out_err;
 645
 646			if (len < UBIFS_BLOCK_SIZE)
 647				folio_zero_range(folio, offset + len,
 648						 UBIFS_BLOCK_SIZE - len);
 649
 650			nn += 1;
 651			read = (i << UBIFS_BLOCK_SHIFT) + len;
 652		} else if (key_block(c, &bu->zbranch[nn].key) < page_block) {
 653			nn += 1;
 654			continue;
 655		} else {
 656			hole = 1;
 657			folio_zero_range(folio, offset, UBIFS_BLOCK_SIZE);
 658		}
 659		if (++i >= UBIFS_BLOCKS_PER_PAGE)
 660			break;
 661		offset += UBIFS_BLOCK_SIZE;
 662		page_block += 1;
 663	}
 664
 665	if (end_index == folio->index) {
 666		int len = i_size & (PAGE_SIZE - 1);
 667
 668		if (len && len < read)
 669			folio_zero_range(folio, len, read - len);
 670	}
 671
 672out_hole:
 673	if (hole) {
 674		folio_set_checked(folio);
 675		dbg_gen("hole");
 676	}
 677
 678	folio_mark_uptodate(folio);
 679	*n = nn;
 680	return 0;
 681
 682out_err:
 683	ubifs_err(c, "bad data node (block %u, inode %lu)",
 684		  page_block, inode->i_ino);
 685	return -EINVAL;
 686}
 687
 688/**
 689 * ubifs_do_bulk_read - do bulk-read.
 690 * @c: UBIFS file-system description object
 691 * @bu: bulk-read information
 692 * @folio1: first folio to read
 693 *
 694 * Returns: %1 if the bulk-read is done, otherwise %0 is returned.
 695 */
 696static int ubifs_do_bulk_read(struct ubifs_info *c, struct bu_info *bu,
 697			      struct folio *folio1)
 698{
 699	pgoff_t offset = folio1->index, end_index;
 700	struct address_space *mapping = folio1->mapping;
 701	struct inode *inode = mapping->host;
 702	struct ubifs_inode *ui = ubifs_inode(inode);
 703	int err, page_idx, page_cnt, ret = 0, n = 0;
 704	int allocate = bu->buf ? 0 : 1;
 705	loff_t isize;
 706	gfp_t ra_gfp_mask = readahead_gfp_mask(mapping) & ~__GFP_FS;
 707
 708	err = ubifs_tnc_get_bu_keys(c, bu);
 709	if (err)
 710		goto out_warn;
 711
 712	if (bu->eof) {
 713		/* Turn off bulk-read at the end of the file */
 714		ui->read_in_a_row = 1;
 715		ui->bulk_read = 0;
 716	}
 717
 718	page_cnt = bu->blk_cnt >> UBIFS_BLOCKS_PER_PAGE_SHIFT;
 719	if (!page_cnt) {
 720		/*
 721		 * This happens when there are multiple blocks per page and the
 722		 * blocks for the first page we are looking for, are not
 723		 * together. If all the pages were like this, bulk-read would
 724		 * reduce performance, so we turn it off for a while.
 725		 */
 726		goto out_bu_off;
 727	}
 728
 729	if (bu->cnt) {
 730		if (allocate) {
 731			/*
 732			 * Allocate bulk-read buffer depending on how many data
 733			 * nodes we are going to read.
 734			 */
 735			bu->buf_len = bu->zbranch[bu->cnt - 1].offs +
 736				      bu->zbranch[bu->cnt - 1].len -
 737				      bu->zbranch[0].offs;
 738			ubifs_assert(c, bu->buf_len > 0);
 739			ubifs_assert(c, bu->buf_len <= c->leb_size);
 740			bu->buf = kmalloc(bu->buf_len, GFP_NOFS | __GFP_NOWARN);
 741			if (!bu->buf)
 742				goto out_bu_off;
 743		}
 744
 745		err = ubifs_tnc_bulk_read(c, bu);
 746		if (err)
 747			goto out_warn;
 748	}
 749
 750	err = populate_page(c, folio1, bu, &n);
 751	if (err)
 752		goto out_warn;
 753
 754	folio_unlock(folio1);
 755	ret = 1;
 756
 757	isize = i_size_read(inode);
 758	if (isize == 0)
 759		goto out_free;
 760	end_index = ((isize - 1) >> PAGE_SHIFT);
 761
 762	for (page_idx = 1; page_idx < page_cnt; page_idx++) {
 763		pgoff_t page_offset = offset + page_idx;
 764		struct folio *folio;
 765
 766		if (page_offset > end_index)
 767			break;
 768		folio = __filemap_get_folio(mapping, page_offset,
 769				 FGP_LOCK|FGP_ACCESSED|FGP_CREAT|FGP_NOWAIT,
 770				 ra_gfp_mask);
 771		if (IS_ERR(folio))
 772			break;
 773		if (!folio_test_uptodate(folio))
 774			err = populate_page(c, folio, bu, &n);
 775		folio_unlock(folio);
 776		folio_put(folio);
 777		if (err)
 778			break;
 779	}
 780
 781	ui->last_page_read = offset + page_idx - 1;
 782
 783out_free:
 784	if (allocate)
 785		kfree(bu->buf);
 786	return ret;
 787
 788out_warn:
 789	ubifs_warn(c, "ignoring error %d and skipping bulk-read", err);
 790	goto out_free;
 791
 792out_bu_off:
 793	ui->read_in_a_row = ui->bulk_read = 0;
 794	goto out_free;
 795}
 796
 797/**
 798 * ubifs_bulk_read - determine whether to bulk-read and, if so, do it.
 799 * @folio: folio from which to start bulk-read.
 800 *
 801 * Some flash media are capable of reading sequentially at faster rates. UBIFS
 802 * bulk-read facility is designed to take advantage of that, by reading in one
 803 * go consecutive data nodes that are also located consecutively in the same
 804 * LEB.
 805 *
 806 * Returns: %1 if a bulk-read is done and %0 otherwise.
 807 */
 808static int ubifs_bulk_read(struct folio *folio)
 809{
 810	struct inode *inode = folio->mapping->host;
 811	struct ubifs_info *c = inode->i_sb->s_fs_info;
 812	struct ubifs_inode *ui = ubifs_inode(inode);
 813	pgoff_t index = folio->index, last_page_read = ui->last_page_read;
 814	struct bu_info *bu;
 815	int err = 0, allocated = 0;
 816
 817	ui->last_page_read = index;
 818	if (!c->bulk_read)
 819		return 0;
 820
 821	/*
 822	 * Bulk-read is protected by @ui->ui_mutex, but it is an optimization,
 823	 * so don't bother if we cannot lock the mutex.
 824	 */
 825	if (!mutex_trylock(&ui->ui_mutex))
 826		return 0;
 827
 828	if (index != last_page_read + 1) {
 829		/* Turn off bulk-read if we stop reading sequentially */
 830		ui->read_in_a_row = 1;
 831		if (ui->bulk_read)
 832			ui->bulk_read = 0;
 833		goto out_unlock;
 834	}
 835
 836	if (!ui->bulk_read) {
 837		ui->read_in_a_row += 1;
 838		if (ui->read_in_a_row < 3)
 839			goto out_unlock;
 840		/* Three reads in a row, so switch on bulk-read */
 841		ui->bulk_read = 1;
 842	}
 843
 844	/*
 845	 * If possible, try to use pre-allocated bulk-read information, which
 846	 * is protected by @c->bu_mutex.
 847	 */
 848	if (mutex_trylock(&c->bu_mutex))
 849		bu = &c->bu;
 850	else {
 851		bu = kmalloc(sizeof(struct bu_info), GFP_NOFS | __GFP_NOWARN);
 852		if (!bu)
 853			goto out_unlock;
 854
 855		bu->buf = NULL;
 856		allocated = 1;
 857	}
 858
 859	bu->buf_len = c->max_bu_buf_len;
 860	data_key_init(c, &bu->key, inode->i_ino,
 861		      folio->index << UBIFS_BLOCKS_PER_PAGE_SHIFT);
 862	err = ubifs_do_bulk_read(c, bu, folio);
 863
 864	if (!allocated)
 865		mutex_unlock(&c->bu_mutex);
 866	else
 867		kfree(bu);
 868
 869out_unlock:
 870	mutex_unlock(&ui->ui_mutex);
 871	return err;
 872}
 873
 874static int ubifs_read_folio(struct file *file, struct folio *folio)
 875{
 876	if (ubifs_bulk_read(folio))
 877		return 0;
 878	do_readpage(folio);
 879	folio_unlock(folio);
 880	return 0;
 881}
 882
 883static int do_writepage(struct folio *folio, size_t len)
 884{
 885	int err = 0, blen;
 886	unsigned int block;
 887	size_t offset = 0;
 888	union ubifs_key key;
 889	struct inode *inode = folio->mapping->host;
 890	struct ubifs_info *c = inode->i_sb->s_fs_info;
 891
 892#ifdef UBIFS_DEBUG
 893	struct ubifs_inode *ui = ubifs_inode(inode);
 894	spin_lock(&ui->ui_lock);
 895	ubifs_assert(c, folio->index <= ui->synced_i_size >> PAGE_SHIFT);
 896	spin_unlock(&ui->ui_lock);
 897#endif
 898
 899	folio_start_writeback(folio);
 900
 901	block = folio->index << UBIFS_BLOCKS_PER_PAGE_SHIFT;
 902	for (;;) {
 903		blen = min_t(size_t, len, UBIFS_BLOCK_SIZE);
 904		data_key_init(c, &key, inode->i_ino, block);
 905		err = ubifs_jnl_write_data(c, inode, &key, folio, offset, blen);
 906		if (err)
 907			break;
 908		len -= blen;
 909		if (!len)
 910			break;
 911		block += 1;
 912		offset += blen;
 913	}
 914	if (err) {
 915		mapping_set_error(folio->mapping, err);
 916		ubifs_err(c, "cannot write folio %lu of inode %lu, error %d",
 917			  folio->index, inode->i_ino, err);
 918		ubifs_ro_mode(c, err);
 919	}
 920
 921	ubifs_assert(c, folio->private != NULL);
 922	if (folio_test_checked(folio))
 923		release_new_page_budget(c);
 924	else
 925		release_existing_page_budget(c);
 926
 927	atomic_long_dec(&c->dirty_pg_cnt);
 928	folio_detach_private(folio);
 929	folio_clear_checked(folio);
 930
 931	folio_unlock(folio);
 932	folio_end_writeback(folio);
 933	return err;
 934}
 935
 936/*
 937 * When writing-back dirty inodes, VFS first writes-back pages belonging to the
 938 * inode, then the inode itself. For UBIFS this may cause a problem. Consider a
 939 * situation when a we have an inode with size 0, then a megabyte of data is
 940 * appended to the inode, then write-back starts and flushes some amount of the
 941 * dirty pages, the journal becomes full, commit happens and finishes, and then
 942 * an unclean reboot happens. When the file system is mounted next time, the
 943 * inode size would still be 0, but there would be many pages which are beyond
 944 * the inode size, they would be indexed and consume flash space. Because the
 945 * journal has been committed, the replay would not be able to detect this
 946 * situation and correct the inode size. This means UBIFS would have to scan
 947 * whole index and correct all inode sizes, which is long an unacceptable.
 948 *
 949 * To prevent situations like this, UBIFS writes pages back only if they are
 950 * within the last synchronized inode size, i.e. the size which has been
 951 * written to the flash media last time. Otherwise, UBIFS forces inode
 952 * write-back, thus making sure the on-flash inode contains current inode size,
 953 * and then keeps writing pages back.
 954 *
 955 * Some locking issues explanation. 'ubifs_writepage()' first is called with
 956 * the page locked, and it locks @ui_mutex. However, write-back does take inode
 957 * @i_mutex, which means other VFS operations may be run on this inode at the
 958 * same time. And the problematic one is truncation to smaller size, from where
 959 * we have to call 'truncate_setsize()', which first changes @inode->i_size,
 960 * then drops the truncated pages. And while dropping the pages, it takes the
 961 * page lock. This means that 'do_truncation()' cannot call 'truncate_setsize()'
 962 * with @ui_mutex locked, because it would deadlock with 'ubifs_writepage()'.
 963 * This means that @inode->i_size is changed while @ui_mutex is unlocked.
 964 *
 965 * XXX(truncate): with the new truncate sequence this is not true anymore,
 966 * and the calls to truncate_setsize can be move around freely.  They should
 967 * be moved to the very end of the truncate sequence.
 968 *
 969 * But in 'ubifs_writepage()' we have to guarantee that we do not write beyond
 970 * inode size. How do we do this if @inode->i_size may became smaller while we
 971 * are in the middle of 'ubifs_writepage()'? The UBIFS solution is the
 972 * @ui->ui_isize "shadow" field which UBIFS uses instead of @inode->i_size
 973 * internally and updates it under @ui_mutex.
 974 *
 975 * Q: why we do not worry that if we race with truncation, we may end up with a
 976 * situation when the inode is truncated while we are in the middle of
 977 * 'do_writepage()', so we do write beyond inode size?
 978 * A: If we are in the middle of 'do_writepage()', truncation would be locked
 979 * on the page lock and it would not write the truncated inode node to the
 980 * journal before we have finished.
 981 */
 982static int ubifs_writepage(struct folio *folio, struct writeback_control *wbc)
 983{
 984	struct inode *inode = folio->mapping->host;
 985	struct ubifs_info *c = inode->i_sb->s_fs_info;
 986	struct ubifs_inode *ui = ubifs_inode(inode);
 987	loff_t i_size =  i_size_read(inode), synced_i_size;
 988	int err, len = folio_size(folio);
 989
 990	dbg_gen("ino %lu, pg %lu, pg flags %#lx",
 991		inode->i_ino, folio->index, folio->flags.f);
 992	ubifs_assert(c, folio->private != NULL);
 993
 994	/* Is the folio fully outside @i_size? (truncate in progress) */
 995	if (folio_pos(folio) >= i_size) {
 996		err = 0;
 997		goto out_unlock;
 998	}
 999
1000	spin_lock(&ui->ui_lock);
1001	synced_i_size = ui->synced_i_size;
1002	spin_unlock(&ui->ui_lock);
1003
1004	/* Is the folio fully inside i_size? */
1005	if (folio_pos(folio) + len <= i_size) {
1006		if (folio_pos(folio) + len > synced_i_size) {
1007			err = inode->i_sb->s_op->write_inode(inode, NULL);
1008			if (err)
1009				goto out_redirty;
1010			/*
1011			 * The inode has been written, but the write-buffer has
1012			 * not been synchronized, so in case of an unclean
1013			 * reboot we may end up with some pages beyond inode
1014			 * size, but they would be in the journal (because
1015			 * commit flushes write buffers) and recovery would deal
1016			 * with this.
1017			 */
1018		}
1019		return do_writepage(folio, len);
1020	}
1021
1022	/*
1023	 * The folio straddles @i_size. It must be zeroed out on each and every
1024	 * writepage invocation because it may be mmapped. "A file is mapped
1025	 * in multiples of the page size. For a file that is not a multiple of
1026	 * the page size, the remaining memory is zeroed when mapped, and
1027	 * writes to that region are not written out to the file."
1028	 */
1029	len = i_size - folio_pos(folio);
1030	folio_zero_segment(folio, len, folio_size(folio));
1031
1032	if (i_size > synced_i_size) {
1033		err = inode->i_sb->s_op->write_inode(inode, NULL);
1034		if (err)
1035			goto out_redirty;
1036	}
1037
1038	return do_writepage(folio, len);
1039out_redirty:
1040	/*
1041	 * folio_redirty_for_writepage() won't call ubifs_dirty_inode() because
1042	 * it passes I_DIRTY_PAGES flag while calling __mark_inode_dirty(), so
1043	 * there is no need to do space budget for dirty inode.
1044	 */
1045	folio_redirty_for_writepage(wbc, folio);
1046out_unlock:
1047	folio_unlock(folio);
1048	return err;
1049}
1050
1051static int ubifs_writepages(struct address_space *mapping,
1052		struct writeback_control *wbc)
1053{
1054	struct folio *folio = NULL;
1055	int error;
1056
1057	while ((folio = writeback_iter(mapping, wbc, folio, &error)))
1058		error = ubifs_writepage(folio, wbc);
1059	return error;
1060}
1061
1062/**
1063 * do_attr_changes - change inode attributes.
1064 * @inode: inode to change attributes for
1065 * @attr: describes attributes to change
1066 */
1067static void do_attr_changes(struct inode *inode, const struct iattr *attr)
1068{
1069	if (attr->ia_valid & ATTR_UID)
1070		inode->i_uid = attr->ia_uid;
1071	if (attr->ia_valid & ATTR_GID)
1072		inode->i_gid = attr->ia_gid;
1073	if (attr->ia_valid & ATTR_ATIME)
1074		inode_set_atime_to_ts(inode, attr->ia_atime);
1075	if (attr->ia_valid & ATTR_MTIME)
1076		inode_set_mtime_to_ts(inode, attr->ia_mtime);
1077	if (attr->ia_valid & ATTR_CTIME)
1078		inode_set_ctime_to_ts(inode, attr->ia_ctime);
1079	if (attr->ia_valid & ATTR_MODE) {
1080		umode_t mode = attr->ia_mode;
1081
1082		if (!in_group_p(inode->i_gid) && !capable(CAP_FSETID))
1083			mode &= ~S_ISGID;
1084		inode->i_mode = mode;
1085	}
1086}
1087
1088/**
1089 * do_truncation - truncate an inode.
1090 * @c: UBIFS file-system description object
1091 * @inode: inode to truncate
1092 * @attr: inode attribute changes description
1093 *
1094 * This function implements VFS '->setattr()' call when the inode is truncated
1095 * to a smaller size.
1096 *
1097 * Returns: %0 in case of success and a negative error code
1098 * in case of failure.
1099 */
1100static int do_truncation(struct ubifs_info *c, struct inode *inode,
1101			 const struct iattr *attr)
1102{
1103	int err;
1104	struct ubifs_budget_req req;
1105	loff_t old_size = inode->i_size, new_size = attr->ia_size;
1106	int offset = new_size & (UBIFS_BLOCK_SIZE - 1), budgeted = 1;
1107	struct ubifs_inode *ui = ubifs_inode(inode);
1108
1109	dbg_gen("ino %lu, size %lld -> %lld", inode->i_ino, old_size, new_size);
1110	memset(&req, 0, sizeof(struct ubifs_budget_req));
1111
1112	/*
1113	 * If this is truncation to a smaller size, and we do not truncate on a
1114	 * block boundary, budget for changing one data block, because the last
1115	 * block will be re-written.
1116	 */
1117	if (new_size & (UBIFS_BLOCK_SIZE - 1))
1118		req.dirtied_page = 1;
1119
1120	req.dirtied_ino = 1;
1121	/* A funny way to budget for truncation node */
1122	req.dirtied_ino_d = UBIFS_TRUN_NODE_SZ;
1123	err = ubifs_budget_space(c, &req);
1124	if (err) {
1125		/*
1126		 * Treat truncations to zero as deletion and always allow them,
1127		 * just like we do for '->unlink()'.
1128		 */
1129		if (new_size || err != -ENOSPC)
1130			return err;
1131		budgeted = 0;
1132	}
1133
1134	truncate_setsize(inode, new_size);
1135
1136	if (offset) {
1137		pgoff_t index = new_size >> PAGE_SHIFT;
1138		struct folio *folio;
1139
1140		folio = filemap_lock_folio(inode->i_mapping, index);
1141		if (!IS_ERR(folio)) {
1142			if (folio_test_dirty(folio)) {
1143				/*
1144				 * 'ubifs_jnl_truncate()' will try to truncate
1145				 * the last data node, but it contains
1146				 * out-of-date data because the page is dirty.
1147				 * Write the page now, so that
1148				 * 'ubifs_jnl_truncate()' will see an already
1149				 * truncated (and up to date) data node.
1150				 */
1151				ubifs_assert(c, folio->private != NULL);
1152
1153				folio_clear_dirty_for_io(folio);
1154				if (UBIFS_BLOCKS_PER_PAGE_SHIFT)
1155					offset = offset_in_folio(folio,
1156							new_size);
1157				err = do_writepage(folio, offset);
1158				folio_put(folio);
1159				if (err)
1160					goto out_budg;
1161				/*
1162				 * We could now tell 'ubifs_jnl_truncate()' not
1163				 * to read the last block.
1164				 */
1165			} else {
1166				/*
1167				 * We could 'kmap()' the page and pass the data
1168				 * to 'ubifs_jnl_truncate()' to save it from
1169				 * having to read it.
1170				 */
1171				folio_unlock(folio);
1172				folio_put(folio);
1173			}
1174		}
1175	}
1176
1177	mutex_lock(&ui->ui_mutex);
1178	ui->ui_size = inode->i_size;
1179	/* Truncation changes inode [mc]time */
1180	inode_set_mtime_to_ts(inode, inode_set_ctime_current(inode));
1181	/* Other attributes may be changed at the same time as well */
1182	do_attr_changes(inode, attr);
1183	err = ubifs_jnl_truncate(c, inode, old_size, new_size);
1184	mutex_unlock(&ui->ui_mutex);
1185
1186out_budg:
1187	if (budgeted)
1188		ubifs_release_budget(c, &req);
1189	else {
1190		c->bi.nospace = c->bi.nospace_rp = 0;
1191		smp_wmb();
1192	}
1193	return err;
1194}
1195
1196/**
1197 * do_setattr - change inode attributes.
1198 * @c: UBIFS file-system description object
1199 * @inode: inode to change attributes for
1200 * @attr: inode attribute changes description
1201 *
1202 * This function implements VFS '->setattr()' call for all cases except
1203 * truncations to smaller size.
1204 *
1205 * Returns: %0 in case of success and a negative
1206 * error code in case of failure.
1207 */
1208static int do_setattr(struct ubifs_info *c, struct inode *inode,
1209		      const struct iattr *attr)
1210{
1211	int err, release;
1212	loff_t new_size = attr->ia_size;
1213	struct ubifs_inode *ui = ubifs_inode(inode);
1214	struct ubifs_budget_req req = { .dirtied_ino = 1,
1215				.dirtied_ino_d = ALIGN(ui->data_len, 8) };
1216
1217	err = ubifs_budget_space(c, &req);
1218	if (err)
1219		return err;
1220
1221	if (attr->ia_valid & ATTR_SIZE) {
1222		dbg_gen("size %lld -> %lld", inode->i_size, new_size);
1223		truncate_setsize(inode, new_size);
1224	}
1225
1226	mutex_lock(&ui->ui_mutex);
1227	if (attr->ia_valid & ATTR_SIZE) {
1228		/* Truncation changes inode [mc]time */
1229		inode_set_mtime_to_ts(inode, inode_set_ctime_current(inode));
1230		/* 'truncate_setsize()' changed @i_size, update @ui_size */
1231		ui->ui_size = inode->i_size;
1232	}
1233
1234	do_attr_changes(inode, attr);
1235
1236	release = ui->dirty;
1237	if (attr->ia_valid & ATTR_SIZE)
1238		/*
1239		 * Inode length changed, so we have to make sure
1240		 * @I_DIRTY_DATASYNC is set.
1241		 */
1242		 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1243	else
1244		mark_inode_dirty_sync(inode);
1245	mutex_unlock(&ui->ui_mutex);
1246
1247	if (release)
1248		ubifs_release_budget(c, &req);
1249	if (IS_SYNC(inode))
1250		err = inode->i_sb->s_op->write_inode(inode, NULL);
1251	return err;
1252}
1253
1254int ubifs_setattr(struct mnt_idmap *idmap, struct dentry *dentry,
1255		  struct iattr *attr)
1256{
1257	int err;
1258	struct inode *inode = d_inode(dentry);
1259	struct ubifs_info *c = inode->i_sb->s_fs_info;
1260
1261	dbg_gen("ino %lu, mode %#x, ia_valid %#x",
1262		inode->i_ino, inode->i_mode, attr->ia_valid);
1263	err = setattr_prepare(&nop_mnt_idmap, dentry, attr);
1264	if (err)
1265		return err;
1266
1267	err = dbg_check_synced_i_size(c, inode);
1268	if (err)
1269		return err;
1270
1271	err = fscrypt_prepare_setattr(dentry, attr);
1272	if (err)
1273		return err;
1274
1275	if ((attr->ia_valid & ATTR_SIZE) && attr->ia_size < inode->i_size)
1276		/* Truncation to a smaller size */
1277		err = do_truncation(c, inode, attr);
1278	else
1279		err = do_setattr(c, inode, attr);
1280
1281	return err;
1282}
1283
1284static void ubifs_invalidate_folio(struct folio *folio, size_t offset,
1285				 size_t length)
1286{
1287	struct inode *inode = folio->mapping->host;
1288	struct ubifs_info *c = inode->i_sb->s_fs_info;
1289
1290	ubifs_assert(c, folio_test_private(folio));
1291	if (offset || length < folio_size(folio))
1292		/* Partial folio remains dirty */
1293		return;
1294
1295	if (folio_test_checked(folio))
1296		release_new_page_budget(c);
1297	else
1298		release_existing_page_budget(c);
1299
1300	atomic_long_dec(&c->dirty_pg_cnt);
1301	folio_detach_private(folio);
1302	folio_clear_checked(folio);
1303}
1304
1305int ubifs_fsync(struct file *file, loff_t start, loff_t end, int datasync)
1306{
1307	struct inode *inode = file->f_mapping->host;
1308	struct ubifs_info *c = inode->i_sb->s_fs_info;
1309	int err;
1310
1311	dbg_gen("syncing inode %lu", inode->i_ino);
1312
1313	if (c->ro_mount)
1314		/*
1315		 * For some really strange reasons VFS does not filter out
1316		 * 'fsync()' for R/O mounted file-systems as per 2.6.39.
1317		 */
1318		return 0;
1319
1320	err = file_write_and_wait_range(file, start, end);
1321	if (err)
1322		return err;
1323	inode_lock(inode);
1324
1325	/* Synchronize the inode unless this is a 'datasync()' call. */
1326	if (!datasync || (inode_state_read_once(inode) & I_DIRTY_DATASYNC)) {
1327		err = inode->i_sb->s_op->write_inode(inode, NULL);
1328		if (err)
1329			goto out;
1330	}
1331
1332	/*
1333	 * Nodes related to this inode may still sit in a write-buffer. Flush
1334	 * them.
1335	 */
1336	err = ubifs_sync_wbufs_by_inode(c, inode);
1337out:
1338	inode_unlock(inode);
1339	return err;
1340}
1341
1342/**
1343 * mctime_update_needed - check if mtime or ctime update is needed.
1344 * @inode: the inode to do the check for
1345 * @now: current time
1346 *
1347 * This helper function checks if the inode mtime/ctime should be updated or
1348 * not. If current values of the time-stamps are within the UBIFS inode time
1349 * granularity, they are not updated. This is an optimization.
1350 *
1351 * Returns: %1 if time update is needed, %0 if not
1352 */
1353static inline int mctime_update_needed(const struct inode *inode,
1354				       const struct timespec64 *now)
1355{
1356	struct timespec64 ctime = inode_get_ctime(inode);
1357	struct timespec64 mtime = inode_get_mtime(inode);
1358
1359	if (!timespec64_equal(&mtime, now) || !timespec64_equal(&ctime, now))
1360		return 1;
1361	return 0;
1362}
1363
1364/**
1365 * ubifs_update_time - update time of inode.
1366 * @inode: inode to update
1367 * @flags: time updating control flag determines updating
1368 *	    which time fields of @inode
1369 *
1370 * This function updates time of the inode.
1371 *
1372 * Returns: %0 for success or a negative error code otherwise.
1373 */
1374int ubifs_update_time(struct inode *inode, int flags)
1375{
1376	struct ubifs_inode *ui = ubifs_inode(inode);
1377	struct ubifs_info *c = inode->i_sb->s_fs_info;
1378	struct ubifs_budget_req req = { .dirtied_ino = 1,
1379			.dirtied_ino_d = ALIGN(ui->data_len, 8) };
1380	int err, release;
1381
1382	if (!IS_ENABLED(CONFIG_UBIFS_ATIME_SUPPORT)) {
1383		generic_update_time(inode, flags);
1384		return 0;
1385	}
1386
1387	err = ubifs_budget_space(c, &req);
1388	if (err)
1389		return err;
1390
1391	mutex_lock(&ui->ui_mutex);
1392	inode_update_timestamps(inode, flags);
1393	release = ui->dirty;
1394	__mark_inode_dirty(inode, I_DIRTY_SYNC);
1395	mutex_unlock(&ui->ui_mutex);
1396	if (release)
1397		ubifs_release_budget(c, &req);
1398	return 0;
1399}
1400
1401/**
1402 * update_mctime - update mtime and ctime of an inode.
1403 * @inode: inode to update
1404 *
1405 * This function updates mtime and ctime of the inode if it is not equivalent to
1406 * current time.
1407 *
1408 * Returns: %0 in case of success and a negative error code in
1409 * case of failure.
1410 */
1411static int update_mctime(struct inode *inode)
1412{
1413	struct timespec64 now = current_time(inode);
1414	struct ubifs_inode *ui = ubifs_inode(inode);
1415	struct ubifs_info *c = inode->i_sb->s_fs_info;
1416
1417	if (mctime_update_needed(inode, &now)) {
1418		int err, release;
1419		struct ubifs_budget_req req = { .dirtied_ino = 1,
1420				.dirtied_ino_d = ALIGN(ui->data_len, 8) };
1421
1422		err = ubifs_budget_space(c, &req);
1423		if (err)
1424			return err;
1425
1426		mutex_lock(&ui->ui_mutex);
1427		inode_set_mtime_to_ts(inode, inode_set_ctime_current(inode));
1428		release = ui->dirty;
1429		mark_inode_dirty_sync(inode);
1430		mutex_unlock(&ui->ui_mutex);
1431		if (release)
1432			ubifs_release_budget(c, &req);
1433	}
1434
1435	return 0;
1436}
1437
1438static ssize_t ubifs_write_iter(struct kiocb *iocb, struct iov_iter *from)
1439{
1440	int err = update_mctime(file_inode(iocb->ki_filp));
1441	if (err)
1442		return err;
1443
1444	return generic_file_write_iter(iocb, from);
1445}
1446
1447static bool ubifs_dirty_folio(struct address_space *mapping,
1448		struct folio *folio)
1449{
1450	bool ret;
1451	struct ubifs_info *c = mapping->host->i_sb->s_fs_info;
1452
1453	ret = filemap_dirty_folio(mapping, folio);
1454	/*
1455	 * An attempt to dirty a page without budgeting for it - should not
1456	 * happen.
1457	 */
1458	ubifs_assert(c, ret == false);
1459	return ret;
1460}
1461
1462static bool ubifs_release_folio(struct folio *folio, gfp_t unused_gfp_flags)
1463{
1464	struct inode *inode = folio->mapping->host;
1465	struct ubifs_info *c = inode->i_sb->s_fs_info;
1466
1467	if (folio_test_writeback(folio))
1468		return false;
1469
1470	/*
1471	 * Page is private but not dirty, weird? There is one condition
1472	 * making it happened. ubifs_writepage skipped the page because
1473	 * page index beyonds isize (for example. truncated by other
1474	 * process named A), then the page is invalidated by fadvise64
1475	 * syscall before being truncated by process A.
1476	 */
1477	ubifs_assert(c, folio_test_private(folio));
1478	if (folio_test_checked(folio))
1479		release_new_page_budget(c);
1480	else
1481		release_existing_page_budget(c);
1482
1483	atomic_long_dec(&c->dirty_pg_cnt);
1484	folio_detach_private(folio);
1485	folio_clear_checked(folio);
1486	return true;
1487}
1488
1489/*
1490 * mmap()d file has taken write protection fault and is being made writable.
1491 * UBIFS must ensure page is budgeted for.
1492 */
1493static vm_fault_t ubifs_vm_page_mkwrite(struct vm_fault *vmf)
1494{
1495	struct folio *folio = page_folio(vmf->page);
1496	struct inode *inode = file_inode(vmf->vma->vm_file);
1497	struct ubifs_info *c = inode->i_sb->s_fs_info;
1498	struct timespec64 now = current_time(inode);
1499	struct ubifs_budget_req req = { .new_page = 1 };
1500	int err, update_time;
1501
1502	dbg_gen("ino %lu, pg %lu, i_size %lld",	inode->i_ino, folio->index,
1503		i_size_read(inode));
1504	ubifs_assert(c, !c->ro_media && !c->ro_mount);
1505
1506	if (unlikely(c->ro_error))
1507		return VM_FAULT_SIGBUS; /* -EROFS */
1508
1509	/*
1510	 * We have not locked @folio so far so we may budget for changing the
1511	 * folio. Note, we cannot do this after we locked the folio, because
1512	 * budgeting may cause write-back which would cause deadlock.
1513	 *
1514	 * At the moment we do not know whether the folio is dirty or not, so we
1515	 * assume that it is not and budget for a new folio. We could look at
1516	 * the @PG_private flag and figure this out, but we may race with write
1517	 * back and the folio state may change by the time we lock it, so this
1518	 * would need additional care. We do not bother with this at the
1519	 * moment, although it might be good idea to do. Instead, we allocate
1520	 * budget for a new folio and amend it later on if the folio was in fact
1521	 * dirty.
1522	 *
1523	 * The budgeting-related logic of this function is similar to what we
1524	 * do in 'ubifs_write_begin()' and 'ubifs_write_end()'. Glance there
1525	 * for more comments.
1526	 */
1527	update_time = mctime_update_needed(inode, &now);
1528	if (update_time)
1529		/*
1530		 * We have to change inode time stamp which requires extra
1531		 * budgeting.
1532		 */
1533		req.dirtied_ino = 1;
1534
1535	err = ubifs_budget_space(c, &req);
1536	if (unlikely(err)) {
1537		if (err == -ENOSPC)
1538			ubifs_warn(c, "out of space for mmapped file (inode number %lu)",
1539				   inode->i_ino);
1540		return VM_FAULT_SIGBUS;
1541	}
1542
1543	folio_lock(folio);
1544	if (unlikely(folio->mapping != inode->i_mapping ||
1545		     folio_pos(folio) >= i_size_read(inode))) {
1546		/* Folio got truncated out from underneath us */
1547		goto sigbus;
1548	}
1549
1550	if (folio->private)
1551		release_new_page_budget(c);
1552	else {
1553		if (!folio_test_checked(folio))
1554			ubifs_convert_page_budget(c);
1555		folio_attach_private(folio, (void *)1);
1556		atomic_long_inc(&c->dirty_pg_cnt);
1557		filemap_dirty_folio(folio->mapping, folio);
1558	}
1559
1560	if (update_time) {
1561		int release;
1562		struct ubifs_inode *ui = ubifs_inode(inode);
1563
1564		mutex_lock(&ui->ui_mutex);
1565		inode_set_mtime_to_ts(inode, inode_set_ctime_current(inode));
1566		release = ui->dirty;
1567		mark_inode_dirty_sync(inode);
1568		mutex_unlock(&ui->ui_mutex);
1569		if (release)
1570			ubifs_release_dirty_inode_budget(c, ui);
1571	}
1572
1573	folio_wait_stable(folio);
1574	return VM_FAULT_LOCKED;
1575
1576sigbus:
1577	folio_unlock(folio);
1578	ubifs_release_budget(c, &req);
1579	return VM_FAULT_SIGBUS;
1580}
1581
1582static const struct vm_operations_struct ubifs_file_vm_ops = {
1583	.fault        = filemap_fault,
1584	.map_pages = filemap_map_pages,
1585	.page_mkwrite = ubifs_vm_page_mkwrite,
1586};
1587
1588static int ubifs_file_mmap_prepare(struct vm_area_desc *desc)
1589{
1590	int err;
1591
1592	err = generic_file_mmap_prepare(desc);
1593	if (err)
1594		return err;
1595	desc->vm_ops = &ubifs_file_vm_ops;
1596
1597	if (IS_ENABLED(CONFIG_UBIFS_ATIME_SUPPORT))
1598		file_accessed(desc->file);
1599
1600	return 0;
1601}
1602
1603static const char *ubifs_get_link(struct dentry *dentry,
1604					    struct inode *inode,
1605					    struct delayed_call *done)
1606{
1607	struct ubifs_inode *ui = ubifs_inode(inode);
1608
1609	if (!IS_ENCRYPTED(inode))
1610		return ui->data;
1611
1612	if (!dentry)
1613		return ERR_PTR(-ECHILD);
1614
1615	return fscrypt_get_symlink(inode, ui->data, ui->data_len, done);
1616}
1617
1618static int ubifs_symlink_getattr(struct mnt_idmap *idmap,
1619				 const struct path *path, struct kstat *stat,
1620				 u32 request_mask, unsigned int query_flags)
1621{
1622	ubifs_getattr(idmap, path, stat, request_mask, query_flags);
1623
1624	if (IS_ENCRYPTED(d_inode(path->dentry)))
1625		return fscrypt_symlink_getattr(path, stat);
1626	return 0;
1627}
1628
1629const struct address_space_operations ubifs_file_address_operations = {
1630	.read_folio     = ubifs_read_folio,
1631	.writepages     = ubifs_writepages,
1632	.write_begin    = ubifs_write_begin,
1633	.write_end      = ubifs_write_end,
1634	.invalidate_folio = ubifs_invalidate_folio,
1635	.dirty_folio	= ubifs_dirty_folio,
1636	.migrate_folio	= filemap_migrate_folio,
1637	.release_folio	= ubifs_release_folio,
1638};
1639
1640const struct inode_operations ubifs_file_inode_operations = {
1641	.setattr     = ubifs_setattr,
1642	.getattr     = ubifs_getattr,
1643	.listxattr   = ubifs_listxattr,
1644	.update_time = ubifs_update_time,
1645	.fileattr_get = ubifs_fileattr_get,
1646	.fileattr_set = ubifs_fileattr_set,
1647};
1648
1649const struct inode_operations ubifs_symlink_inode_operations = {
1650	.get_link    = ubifs_get_link,
1651	.setattr     = ubifs_setattr,
1652	.getattr     = ubifs_symlink_getattr,
1653	.listxattr   = ubifs_listxattr,
1654	.update_time = ubifs_update_time,
1655};
1656
1657const struct file_operations ubifs_file_operations = {
1658	.llseek         = generic_file_llseek,
1659	.read_iter      = generic_file_read_iter,
1660	.write_iter     = ubifs_write_iter,
1661	.mmap_prepare   = ubifs_file_mmap_prepare,
1662	.fsync          = ubifs_fsync,
1663	.unlocked_ioctl = ubifs_ioctl,
1664	.splice_read	= filemap_splice_read,
1665	.splice_write	= iter_file_splice_write,
1666	.open		= fscrypt_file_open,
1667#ifdef CONFIG_COMPAT
1668	.compat_ioctl   = ubifs_compat_ioctl,
1669#endif
1670};