1/*
   2 *  linux/fs/ext4/indirect.c
   3 *
   4 *  from
   5 *
   6 *  linux/fs/ext4/inode.c
   7 *
   8 * Copyright (C) 1992, 1993, 1994, 1995
   9 * Remy Card (card@masi.ibp.fr)
  10 * Laboratoire MASI - Institut Blaise Pascal
  11 * Universite Pierre et Marie Curie (Paris VI)
  12 *
  13 *  from
  14 *
  15 *  linux/fs/minix/inode.c
  16 *
  17 *  Copyright (C) 1991, 1992  Linus Torvalds
  18 *
  19 *  Goal-directed block allocation by Stephen Tweedie
  20 *	(sct@redhat.com), 1993, 1998
  21 */
  22
  23#include "ext4_jbd2.h"
  24#include "truncate.h"
  25#include <linux/dax.h>
  26#include <linux/uio.h>
  27
  28#include <trace/events/ext4.h>
  29
  30typedef struct {
  31	__le32	*p;
  32	__le32	key;
  33	struct buffer_head *bh;
  34} Indirect;
  35
  36static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
  37{
  38	p->key = *(p->p = v);
  39	p->bh = bh;
  40}
  41
  42/**
  43 *	ext4_block_to_path - parse the block number into array of offsets
  44 *	@inode: inode in question (we are only interested in its superblock)
  45 *	@i_block: block number to be parsed
  46 *	@offsets: array to store the offsets in
  47 *	@boundary: set this non-zero if the referred-to block is likely to be
  48 *	       followed (on disk) by an indirect block.
  49 *
  50 *	To store the locations of file's data ext4 uses a data structure common
  51 *	for UNIX filesystems - tree of pointers anchored in the inode, with
  52 *	data blocks at leaves and indirect blocks in intermediate nodes.
  53 *	This function translates the block number into path in that tree -
  54 *	return value is the path length and @offsets[n] is the offset of
  55 *	pointer to (n+1)th node in the nth one. If @block is out of range
  56 *	(negative or too large) warning is printed and zero returned.
  57 *
  58 *	Note: function doesn't find node addresses, so no IO is needed. All
  59 *	we need to know is the capacity of indirect blocks (taken from the
  60 *	inode->i_sb).
  61 */
  62
  63/*
  64 * Portability note: the last comparison (check that we fit into triple
  65 * indirect block) is spelled differently, because otherwise on an
  66 * architecture with 32-bit longs and 8Kb pages we might get into trouble
  67 * if our filesystem had 8Kb blocks. We might use long long, but that would
  68 * kill us on x86. Oh, well, at least the sign propagation does not matter -
  69 * i_block would have to be negative in the very beginning, so we would not
  70 * get there at all.
  71 */
  72
  73static int ext4_block_to_path(struct inode *inode,
  74			      ext4_lblk_t i_block,
  75			      ext4_lblk_t offsets[4], int *boundary)
  76{
  77	int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
  78	int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
  79	const long direct_blocks = EXT4_NDIR_BLOCKS,
  80		indirect_blocks = ptrs,
  81		double_blocks = (1 << (ptrs_bits * 2));
  82	int n = 0;
  83	int final = 0;
  84
  85	if (i_block < direct_blocks) {
  86		offsets[n++] = i_block;
  87		final = direct_blocks;
  88	} else if ((i_block -= direct_blocks) < indirect_blocks) {
  89		offsets[n++] = EXT4_IND_BLOCK;
  90		offsets[n++] = i_block;
  91		final = ptrs;
  92	} else if ((i_block -= indirect_blocks) < double_blocks) {
  93		offsets[n++] = EXT4_DIND_BLOCK;
  94		offsets[n++] = i_block >> ptrs_bits;
  95		offsets[n++] = i_block & (ptrs - 1);
  96		final = ptrs;
  97	} else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
  98		offsets[n++] = EXT4_TIND_BLOCK;
  99		offsets[n++] = i_block >> (ptrs_bits * 2);
 100		offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
 101		offsets[n++] = i_block & (ptrs - 1);
 102		final = ptrs;
 103	} else {
 104		ext4_warning(inode->i_sb, "block %lu > max in inode %lu",
 105			     i_block + direct_blocks +
 106			     indirect_blocks + double_blocks, inode->i_ino);
 107	}
 108	if (boundary)
 109		*boundary = final - 1 - (i_block & (ptrs - 1));
 110	return n;
 111}
 112
 113/**
 114 *	ext4_get_branch - read the chain of indirect blocks leading to data
 115 *	@inode: inode in question
 116 *	@depth: depth of the chain (1 - direct pointer, etc.)
 117 *	@offsets: offsets of pointers in inode/indirect blocks
 118 *	@chain: place to store the result
 119 *	@err: here we store the error value
 120 *
 121 *	Function fills the array of triples <key, p, bh> and returns %NULL
 122 *	if everything went OK or the pointer to the last filled triple
 123 *	(incomplete one) otherwise. Upon the return chain[i].key contains
 124 *	the number of (i+1)-th block in the chain (as it is stored in memory,
 125 *	i.e. little-endian 32-bit), chain[i].p contains the address of that
 126 *	number (it points into struct inode for i==0 and into the bh->b_data
 127 *	for i>0) and chain[i].bh points to the buffer_head of i-th indirect
 128 *	block for i>0 and NULL for i==0. In other words, it holds the block
 129 *	numbers of the chain, addresses they were taken from (and where we can
 130 *	verify that chain did not change) and buffer_heads hosting these
 131 *	numbers.
 132 *
 133 *	Function stops when it stumbles upon zero pointer (absent block)
 134 *		(pointer to last triple returned, *@err == 0)
 135 *	or when it gets an IO error reading an indirect block
 136 *		(ditto, *@err == -EIO)
 137 *	or when it reads all @depth-1 indirect blocks successfully and finds
 138 *	the whole chain, all way to the data (returns %NULL, *err == 0).
 139 *
 140 *      Need to be called with
 141 *      down_read(&EXT4_I(inode)->i_data_sem)
 142 */
 143static Indirect *ext4_get_branch(struct inode *inode, int depth,
 144				 ext4_lblk_t  *offsets,
 145				 Indirect chain[4], int *err)
 146{
 147	struct super_block *sb = inode->i_sb;
 148	Indirect *p = chain;
 149	struct buffer_head *bh;
 150	int ret = -EIO;
 151
 152	*err = 0;
 153	/* i_data is not going away, no lock needed */
 154	add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
 155	if (!p->key)
 156		goto no_block;
 157	while (--depth) {
 158		bh = sb_getblk(sb, le32_to_cpu(p->key));
 159		if (unlikely(!bh)) {
 160			ret = -ENOMEM;
 161			goto failure;
 162		}
 163
 164		if (!bh_uptodate_or_lock(bh)) {
 165			if (bh_submit_read(bh) < 0) {
 166				put_bh(bh);
 167				goto failure;
 168			}
 169			/* validate block references */
 170			if (ext4_check_indirect_blockref(inode, bh)) {
 171				put_bh(bh);
 172				goto failure;
 173			}
 174		}
 175
 176		add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
 177		/* Reader: end */
 178		if (!p->key)
 179			goto no_block;
 180	}
 181	return NULL;
 182
 183failure:
 184	*err = ret;
 185no_block:
 186	return p;
 187}
 188
 189/**
 190 *	ext4_find_near - find a place for allocation with sufficient locality
 191 *	@inode: owner
 192 *	@ind: descriptor of indirect block.
 193 *
 194 *	This function returns the preferred place for block allocation.
 195 *	It is used when heuristic for sequential allocation fails.
 196 *	Rules are:
 197 *	  + if there is a block to the left of our position - allocate near it.
 198 *	  + if pointer will live in indirect block - allocate near that block.
 199 *	  + if pointer will live in inode - allocate in the same
 200 *	    cylinder group.
 201 *
 202 * In the latter case we colour the starting block by the callers PID to
 203 * prevent it from clashing with concurrent allocations for a different inode
 204 * in the same block group.   The PID is used here so that functionally related
 205 * files will be close-by on-disk.
 206 *
 207 *	Caller must make sure that @ind is valid and will stay that way.
 208 */
 209static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
 210{
 211	struct ext4_inode_info *ei = EXT4_I(inode);
 212	__le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
 213	__le32 *p;
 214
 215	/* Try to find previous block */
 216	for (p = ind->p - 1; p >= start; p--) {
 217		if (*p)
 218			return le32_to_cpu(*p);
 219	}
 220
 221	/* No such thing, so let's try location of indirect block */
 222	if (ind->bh)
 223		return ind->bh->b_blocknr;
 224
 225	/*
 226	 * It is going to be referred to from the inode itself? OK, just put it
 227	 * into the same cylinder group then.
 228	 */
 229	return ext4_inode_to_goal_block(inode);
 230}
 231
 232/**
 233 *	ext4_find_goal - find a preferred place for allocation.
 234 *	@inode: owner
 235 *	@block:  block we want
 236 *	@partial: pointer to the last triple within a chain
 237 *
 238 *	Normally this function find the preferred place for block allocation,
 239 *	returns it.
 240 *	Because this is only used for non-extent files, we limit the block nr
 241 *	to 32 bits.
 242 */
 243static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
 244				   Indirect *partial)
 245{
 246	ext4_fsblk_t goal;
 247
 248	/*
 249	 * XXX need to get goal block from mballoc's data structures
 250	 */
 251
 252	goal = ext4_find_near(inode, partial);
 253	goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
 254	return goal;
 255}
 256
 257/**
 258 *	ext4_blks_to_allocate - Look up the block map and count the number
 259 *	of direct blocks need to be allocated for the given branch.
 260 *
 261 *	@branch: chain of indirect blocks
 262 *	@k: number of blocks need for indirect blocks
 263 *	@blks: number of data blocks to be mapped.
 264 *	@blocks_to_boundary:  the offset in the indirect block
 265 *
 266 *	return the total number of blocks to be allocate, including the
 267 *	direct and indirect blocks.
 268 */
 269static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
 270				 int blocks_to_boundary)
 271{
 272	unsigned int count = 0;
 273
 274	/*
 275	 * Simple case, [t,d]Indirect block(s) has not allocated yet
 276	 * then it's clear blocks on that path have not allocated
 277	 */
 278	if (k > 0) {
 279		/* right now we don't handle cross boundary allocation */
 280		if (blks < blocks_to_boundary + 1)
 281			count += blks;
 282		else
 283			count += blocks_to_boundary + 1;
 284		return count;
 285	}
 286
 287	count++;
 288	while (count < blks && count <= blocks_to_boundary &&
 289		le32_to_cpu(*(branch[0].p + count)) == 0) {
 290		count++;
 291	}
 292	return count;
 293}
 294
 295/**
 296 *	ext4_alloc_branch - allocate and set up a chain of blocks.
 297 *	@handle: handle for this transaction
 298 *	@inode: owner
 299 *	@indirect_blks: number of allocated indirect blocks
 300 *	@blks: number of allocated direct blocks
 301 *	@goal: preferred place for allocation
 302 *	@offsets: offsets (in the blocks) to store the pointers to next.
 303 *	@branch: place to store the chain in.
 304 *
 305 *	This function allocates blocks, zeroes out all but the last one,
 306 *	links them into chain and (if we are synchronous) writes them to disk.
 307 *	In other words, it prepares a branch that can be spliced onto the
 308 *	inode. It stores the information about that chain in the branch[], in
 309 *	the same format as ext4_get_branch() would do. We are calling it after
 310 *	we had read the existing part of chain and partial points to the last
 311 *	triple of that (one with zero ->key). Upon the exit we have the same
 312 *	picture as after the successful ext4_get_block(), except that in one
 313 *	place chain is disconnected - *branch->p is still zero (we did not
 314 *	set the last link), but branch->key contains the number that should
 315 *	be placed into *branch->p to fill that gap.
 316 *
 317 *	If allocation fails we free all blocks we've allocated (and forget
 318 *	their buffer_heads) and return the error value the from failed
 319 *	ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
 320 *	as described above and return 0.
 321 */
 322static int ext4_alloc_branch(handle_t *handle,
 323			     struct ext4_allocation_request *ar,
 324			     int indirect_blks, ext4_lblk_t *offsets,
 325			     Indirect *branch)
 326{
 327	struct buffer_head *		bh;
 328	ext4_fsblk_t			b, new_blocks[4];
 329	__le32				*p;
 330	int				i, j, err, len = 1;
 331
 332	for (i = 0; i <= indirect_blks; i++) {
 333		if (i == indirect_blks) {
 334			new_blocks[i] = ext4_mb_new_blocks(handle, ar, &err);
 335		} else
 336			ar->goal = new_blocks[i] = ext4_new_meta_blocks(handle,
 337					ar->inode, ar->goal,
 338					ar->flags & EXT4_MB_DELALLOC_RESERVED,
 339					NULL, &err);
 340		if (err) {
 341			i--;
 342			goto failed;
 343		}
 344		branch[i].key = cpu_to_le32(new_blocks[i]);
 345		if (i == 0)
 346			continue;
 347
 348		bh = branch[i].bh = sb_getblk(ar->inode->i_sb, new_blocks[i-1]);
 349		if (unlikely(!bh)) {
 350			err = -ENOMEM;
 351			goto failed;
 352		}
 353		lock_buffer(bh);
 354		BUFFER_TRACE(bh, "call get_create_access");
 355		err = ext4_journal_get_create_access(handle, bh);
 356		if (err) {
 357			unlock_buffer(bh);
 358			goto failed;
 359		}
 360
 361		memset(bh->b_data, 0, bh->b_size);
 362		p = branch[i].p = (__le32 *) bh->b_data + offsets[i];
 363		b = new_blocks[i];
 364
 365		if (i == indirect_blks)
 366			len = ar->len;
 367		for (j = 0; j < len; j++)
 368			*p++ = cpu_to_le32(b++);
 369
 370		BUFFER_TRACE(bh, "marking uptodate");
 371		set_buffer_uptodate(bh);
 372		unlock_buffer(bh);
 373
 374		BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
 375		err = ext4_handle_dirty_metadata(handle, ar->inode, bh);
 376		if (err)
 377			goto failed;
 378	}
 379	return 0;
 380failed:
 381	for (; i >= 0; i--) {
 382		/*
 383		 * We want to ext4_forget() only freshly allocated indirect
 384		 * blocks.  Buffer for new_blocks[i-1] is at branch[i].bh and
 385		 * buffer at branch[0].bh is indirect block / inode already
 386		 * existing before ext4_alloc_branch() was called.
 387		 */
 388		if (i > 0 && i != indirect_blks && branch[i].bh)
 389			ext4_forget(handle, 1, ar->inode, branch[i].bh,
 390				    branch[i].bh->b_blocknr);
 391		ext4_free_blocks(handle, ar->inode, NULL, new_blocks[i],
 392				 (i == indirect_blks) ? ar->len : 1, 0);
 393	}
 394	return err;
 395}
 396
 397/**
 398 * ext4_splice_branch - splice the allocated branch onto inode.
 399 * @handle: handle for this transaction
 400 * @inode: owner
 401 * @block: (logical) number of block we are adding
 402 * @chain: chain of indirect blocks (with a missing link - see
 403 *	ext4_alloc_branch)
 404 * @where: location of missing link
 405 * @num:   number of indirect blocks we are adding
 406 * @blks:  number of direct blocks we are adding
 407 *
 408 * This function fills the missing link and does all housekeeping needed in
 409 * inode (->i_blocks, etc.). In case of success we end up with the full
 410 * chain to new block and return 0.
 411 */
 412static int ext4_splice_branch(handle_t *handle,
 413			      struct ext4_allocation_request *ar,
 414			      Indirect *where, int num)
 415{
 416	int i;
 417	int err = 0;
 418	ext4_fsblk_t current_block;
 419
 420	/*
 421	 * If we're splicing into a [td]indirect block (as opposed to the
 422	 * inode) then we need to get write access to the [td]indirect block
 423	 * before the splice.
 424	 */
 425	if (where->bh) {
 426		BUFFER_TRACE(where->bh, "get_write_access");
 427		err = ext4_journal_get_write_access(handle, where->bh);
 428		if (err)
 429			goto err_out;
 430	}
 431	/* That's it */
 432
 433	*where->p = where->key;
 434
 435	/*
 436	 * Update the host buffer_head or inode to point to more just allocated
 437	 * direct blocks blocks
 438	 */
 439	if (num == 0 && ar->len > 1) {
 440		current_block = le32_to_cpu(where->key) + 1;
 441		for (i = 1; i < ar->len; i++)
 442			*(where->p + i) = cpu_to_le32(current_block++);
 443	}
 444
 445	/* We are done with atomic stuff, now do the rest of housekeeping */
 446	/* had we spliced it onto indirect block? */
 447	if (where->bh) {
 448		/*
 449		 * If we spliced it onto an indirect block, we haven't
 450		 * altered the inode.  Note however that if it is being spliced
 451		 * onto an indirect block at the very end of the file (the
 452		 * file is growing) then we *will* alter the inode to reflect
 453		 * the new i_size.  But that is not done here - it is done in
 454		 * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
 455		 */
 456		jbd_debug(5, "splicing indirect only\n");
 457		BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
 458		err = ext4_handle_dirty_metadata(handle, ar->inode, where->bh);
 459		if (err)
 460			goto err_out;
 461	} else {
 462		/*
 463		 * OK, we spliced it into the inode itself on a direct block.
 464		 */
 465		ext4_mark_inode_dirty(handle, ar->inode);
 466		jbd_debug(5, "splicing direct\n");
 467	}
 468	return err;
 469
 470err_out:
 471	for (i = 1; i <= num; i++) {
 472		/*
 473		 * branch[i].bh is newly allocated, so there is no
 474		 * need to revoke the block, which is why we don't
 475		 * need to set EXT4_FREE_BLOCKS_METADATA.
 476		 */
 477		ext4_free_blocks(handle, ar->inode, where[i].bh, 0, 1,
 478				 EXT4_FREE_BLOCKS_FORGET);
 479	}
 480	ext4_free_blocks(handle, ar->inode, NULL, le32_to_cpu(where[num].key),
 481			 ar->len, 0);
 482
 483	return err;
 484}
 485
 486/*
 487 * The ext4_ind_map_blocks() function handles non-extents inodes
 488 * (i.e., using the traditional indirect/double-indirect i_blocks
 489 * scheme) for ext4_map_blocks().
 490 *
 491 * Allocation strategy is simple: if we have to allocate something, we will
 492 * have to go the whole way to leaf. So let's do it before attaching anything
 493 * to tree, set linkage between the newborn blocks, write them if sync is
 494 * required, recheck the path, free and repeat if check fails, otherwise
 495 * set the last missing link (that will protect us from any truncate-generated
 496 * removals - all blocks on the path are immune now) and possibly force the
 497 * write on the parent block.
 498 * That has a nice additional property: no special recovery from the failed
 499 * allocations is needed - we simply release blocks and do not touch anything
 500 * reachable from inode.
 501 *
 502 * `handle' can be NULL if create == 0.
 503 *
 504 * return > 0, # of blocks mapped or allocated.
 505 * return = 0, if plain lookup failed.
 506 * return < 0, error case.
 507 *
 508 * The ext4_ind_get_blocks() function should be called with
 509 * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
 510 * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
 511 * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
 512 * blocks.
 513 */
 514int ext4_ind_map_blocks(handle_t *handle, struct inode *inode,
 515			struct ext4_map_blocks *map,
 516			int flags)
 517{
 518	struct ext4_allocation_request ar;
 519	int err = -EIO;
 520	ext4_lblk_t offsets[4];
 521	Indirect chain[4];
 522	Indirect *partial;
 523	int indirect_blks;
 524	int blocks_to_boundary = 0;
 525	int depth;
 526	int count = 0;
 527	ext4_fsblk_t first_block = 0;
 528
 529	trace_ext4_ind_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
 530	J_ASSERT(!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)));
 531	J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
 532	depth = ext4_block_to_path(inode, map->m_lblk, offsets,
 533				   &blocks_to_boundary);
 534
 535	if (depth == 0)
 536		goto out;
 537
 538	partial = ext4_get_branch(inode, depth, offsets, chain, &err);
 539
 540	/* Simplest case - block found, no allocation needed */
 541	if (!partial) {
 542		first_block = le32_to_cpu(chain[depth - 1].key);
 543		count++;
 544		/*map more blocks*/
 545		while (count < map->m_len && count <= blocks_to_boundary) {
 546			ext4_fsblk_t blk;
 547
 548			blk = le32_to_cpu(*(chain[depth-1].p + count));
 549
 550			if (blk == first_block + count)
 551				count++;
 552			else
 553				break;
 554		}
 555		goto got_it;
 556	}
 557
 558	/* Next simple case - plain lookup failed */
 559	if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
 560		unsigned epb = inode->i_sb->s_blocksize / sizeof(u32);
 561		int i;
 562
 563		/* Count number blocks in a subtree under 'partial' */
 564		count = 1;
 565		for (i = 0; partial + i != chain + depth - 1; i++)
 566			count *= epb;
 567		/* Fill in size of a hole we found */
 568		map->m_pblk = 0;
 569		map->m_len = min_t(unsigned int, map->m_len, count);
 570		goto cleanup;
 571	}
 572
 573	/* Failed read of indirect block */
 574	if (err == -EIO)
 575		goto cleanup;
 576
 577	/*
 578	 * Okay, we need to do block allocation.
 579	*/
 580	if (ext4_has_feature_bigalloc(inode->i_sb)) {
 581		EXT4_ERROR_INODE(inode, "Can't allocate blocks for "
 582				 "non-extent mapped inodes with bigalloc");
 583		return -EFSCORRUPTED;
 584	}
 585
 586	/* Set up for the direct block allocation */
 587	memset(&ar, 0, sizeof(ar));
 588	ar.inode = inode;
 589	ar.logical = map->m_lblk;
 590	if (S_ISREG(inode->i_mode))
 591		ar.flags = EXT4_MB_HINT_DATA;
 592	if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
 593		ar.flags |= EXT4_MB_DELALLOC_RESERVED;
 594	if (flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
 595		ar.flags |= EXT4_MB_USE_RESERVED;
 596
 597	ar.goal = ext4_find_goal(inode, map->m_lblk, partial);
 598
 599	/* the number of blocks need to allocate for [d,t]indirect blocks */
 600	indirect_blks = (chain + depth) - partial - 1;
 601
 602	/*
 603	 * Next look up the indirect map to count the totoal number of
 604	 * direct blocks to allocate for this branch.
 605	 */
 606	ar.len = ext4_blks_to_allocate(partial, indirect_blks,
 607				       map->m_len, blocks_to_boundary);
 608
 609	/*
 610	 * Block out ext4_truncate while we alter the tree
 611	 */
 612	err = ext4_alloc_branch(handle, &ar, indirect_blks,
 613				offsets + (partial - chain), partial);
 614
 615	/*
 616	 * The ext4_splice_branch call will free and forget any buffers
 617	 * on the new chain if there is a failure, but that risks using
 618	 * up transaction credits, especially for bitmaps where the
 619	 * credits cannot be returned.  Can we handle this somehow?  We
 620	 * may need to return -EAGAIN upwards in the worst case.  --sct
 621	 */
 622	if (!err)
 623		err = ext4_splice_branch(handle, &ar, partial, indirect_blks);
 624	if (err)
 625		goto cleanup;
 626
 627	map->m_flags |= EXT4_MAP_NEW;
 628
 629	ext4_update_inode_fsync_trans(handle, inode, 1);
 630	count = ar.len;
 631got_it:
 632	map->m_flags |= EXT4_MAP_MAPPED;
 633	map->m_pblk = le32_to_cpu(chain[depth-1].key);
 634	map->m_len = count;
 635	if (count > blocks_to_boundary)
 636		map->m_flags |= EXT4_MAP_BOUNDARY;
 637	err = count;
 638	/* Clean up and exit */
 639	partial = chain + depth - 1;	/* the whole chain */
 640cleanup:
 641	while (partial > chain) {
 642		BUFFER_TRACE(partial->bh, "call brelse");
 643		brelse(partial->bh);
 644		partial--;
 645	}
 646out:
 647	trace_ext4_ind_map_blocks_exit(inode, flags, map, err);
 648	return err;
 649}
 650
 651/*
 652 * O_DIRECT for ext3 (or indirect map) based files
 653 *
 654 * If the O_DIRECT write will extend the file then add this inode to the
 655 * orphan list.  So recovery will truncate it back to the original size
 656 * if the machine crashes during the write.
 657 *
 658 * If the O_DIRECT write is intantiating holes inside i_size and the machine
 659 * crashes then stale disk data _may_ be exposed inside the file. But current
 660 * VFS code falls back into buffered path in that case so we are safe.
 661 */
 662ssize_t ext4_ind_direct_IO(struct kiocb *iocb, struct iov_iter *iter,
 663			   loff_t offset)
 664{
 665	struct file *file = iocb->ki_filp;
 666	struct inode *inode = file->f_mapping->host;
 667	struct ext4_inode_info *ei = EXT4_I(inode);
 668	handle_t *handle;
 669	ssize_t ret;
 670	int orphan = 0;
 671	size_t count = iov_iter_count(iter);
 672	int retries = 0;
 673
 674	if (iov_iter_rw(iter) == WRITE) {
 675		loff_t final_size = offset + count;
 676
 677		if (final_size > inode->i_size) {
 678			/* Credits for sb + inode write */
 679			handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 680			if (IS_ERR(handle)) {
 681				ret = PTR_ERR(handle);
 682				goto out;
 683			}
 684			ret = ext4_orphan_add(handle, inode);
 685			if (ret) {
 686				ext4_journal_stop(handle);
 687				goto out;
 688			}
 689			orphan = 1;
 690			ei->i_disksize = inode->i_size;
 691			ext4_journal_stop(handle);
 692		}
 693	}
 694
 695retry:
 696	if (iov_iter_rw(iter) == READ && ext4_should_dioread_nolock(inode)) {
 697		/*
 698		 * Nolock dioread optimization may be dynamically disabled
 699		 * via ext4_inode_block_unlocked_dio(). Check inode's state
 700		 * while holding extra i_dio_count ref.
 701		 */
 702		inode_dio_begin(inode);
 703		smp_mb();
 704		if (unlikely(ext4_test_inode_state(inode,
 705						    EXT4_STATE_DIOREAD_LOCK))) {
 706			inode_dio_end(inode);
 707			goto locked;
 708		}
 709		if (IS_DAX(inode))
 710			ret = dax_do_io(iocb, inode, iter, offset,
 711					ext4_dio_get_block, NULL, 0);
 712		else
 713			ret = __blockdev_direct_IO(iocb, inode,
 714						   inode->i_sb->s_bdev, iter,
 715						   offset, ext4_dio_get_block,
 716						   NULL, NULL, 0);
 717		inode_dio_end(inode);
 718	} else {
 719locked:
 720		if (IS_DAX(inode))
 721			ret = dax_do_io(iocb, inode, iter, offset,
 722					ext4_dio_get_block, NULL, DIO_LOCKING);
 723		else
 724			ret = blockdev_direct_IO(iocb, inode, iter, offset,
 725						 ext4_dio_get_block);
 726
 727		if (unlikely(iov_iter_rw(iter) == WRITE && ret < 0)) {
 728			loff_t isize = i_size_read(inode);
 729			loff_t end = offset + count;
 730
 731			if (end > isize)
 732				ext4_truncate_failed_write(inode);
 733		}
 734	}
 735	if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
 736		goto retry;
 737
 738	if (orphan) {
 739		int err;
 740
 741		/* Credits for sb + inode write */
 742		handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 743		if (IS_ERR(handle)) {
 744			/* This is really bad luck. We've written the data
 745			 * but cannot extend i_size. Bail out and pretend
 746			 * the write failed... */
 747			ret = PTR_ERR(handle);
 748			if (inode->i_nlink)
 749				ext4_orphan_del(NULL, inode);
 750
 751			goto out;
 752		}
 753		if (inode->i_nlink)
 754			ext4_orphan_del(handle, inode);
 755		if (ret > 0) {
 756			loff_t end = offset + ret;
 757			if (end > inode->i_size) {
 758				ei->i_disksize = end;
 759				i_size_write(inode, end);
 760				/*
 761				 * We're going to return a positive `ret'
 762				 * here due to non-zero-length I/O, so there's
 763				 * no way of reporting error returns from
 764				 * ext4_mark_inode_dirty() to userspace.  So
 765				 * ignore it.
 766				 */
 767				ext4_mark_inode_dirty(handle, inode);
 768			}
 769		}
 770		err = ext4_journal_stop(handle);
 771		if (ret == 0)
 772			ret = err;
 773	}
 774out:
 775	return ret;
 776}
 777
 778/*
 779 * Calculate the number of metadata blocks need to reserve
 780 * to allocate a new block at @lblocks for non extent file based file
 781 */
 782int ext4_ind_calc_metadata_amount(struct inode *inode, sector_t lblock)
 783{
 784	struct ext4_inode_info *ei = EXT4_I(inode);
 785	sector_t dind_mask = ~((sector_t)EXT4_ADDR_PER_BLOCK(inode->i_sb) - 1);
 786	int blk_bits;
 787
 788	if (lblock < EXT4_NDIR_BLOCKS)
 789		return 0;
 790
 791	lblock -= EXT4_NDIR_BLOCKS;
 792
 793	if (ei->i_da_metadata_calc_len &&
 794	    (lblock & dind_mask) == ei->i_da_metadata_calc_last_lblock) {
 795		ei->i_da_metadata_calc_len++;
 796		return 0;
 797	}
 798	ei->i_da_metadata_calc_last_lblock = lblock & dind_mask;
 799	ei->i_da_metadata_calc_len = 1;
 800	blk_bits = order_base_2(lblock);
 801	return (blk_bits / EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb)) + 1;
 802}
 803
 804/*
 805 * Calculate number of indirect blocks touched by mapping @nrblocks logically
 806 * contiguous blocks
 807 */
 808int ext4_ind_trans_blocks(struct inode *inode, int nrblocks)
 809{
 810	/*
 811	 * With N contiguous data blocks, we need at most
 812	 * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) + 1 indirect blocks,
 813	 * 2 dindirect blocks, and 1 tindirect block
 814	 */
 815	return DIV_ROUND_UP(nrblocks, EXT4_ADDR_PER_BLOCK(inode->i_sb)) + 4;
 816}
 817
 818/*
 819 * Truncate transactions can be complex and absolutely huge.  So we need to
 820 * be able to restart the transaction at a conventient checkpoint to make
 821 * sure we don't overflow the journal.
 822 *
 823 * Try to extend this transaction for the purposes of truncation.  If
 824 * extend fails, we need to propagate the failure up and restart the
 825 * transaction in the top-level truncate loop. --sct
 826 *
 827 * Returns 0 if we managed to create more room.  If we can't create more
 828 * room, and the transaction must be restarted we return 1.
 829 */
 830static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
 831{
 832	if (!ext4_handle_valid(handle))
 833		return 0;
 834	if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
 835		return 0;
 836	if (!ext4_journal_extend(handle, ext4_blocks_for_truncate(inode)))
 837		return 0;
 838	return 1;
 839}
 840
 841/*
 842 * Probably it should be a library function... search for first non-zero word
 843 * or memcmp with zero_page, whatever is better for particular architecture.
 844 * Linus?
 845 */
 846static inline int all_zeroes(__le32 *p, __le32 *q)
 847{
 848	while (p < q)
 849		if (*p++)
 850			return 0;
 851	return 1;
 852}
 853
 854/**
 855 *	ext4_find_shared - find the indirect blocks for partial truncation.
 856 *	@inode:	  inode in question
 857 *	@depth:	  depth of the affected branch
 858 *	@offsets: offsets of pointers in that branch (see ext4_block_to_path)
 859 *	@chain:	  place to store the pointers to partial indirect blocks
 860 *	@top:	  place to the (detached) top of branch
 861 *
 862 *	This is a helper function used by ext4_truncate().
 863 *
 864 *	When we do truncate() we may have to clean the ends of several
 865 *	indirect blocks but leave the blocks themselves alive. Block is
 866 *	partially truncated if some data below the new i_size is referred
 867 *	from it (and it is on the path to the first completely truncated
 868 *	data block, indeed).  We have to free the top of that path along
 869 *	with everything to the right of the path. Since no allocation
 870 *	past the truncation point is possible until ext4_truncate()
 871 *	finishes, we may safely do the latter, but top of branch may
 872 *	require special attention - pageout below the truncation point
 873 *	might try to populate it.
 874 *
 875 *	We atomically detach the top of branch from the tree, store the
 876 *	block number of its root in *@top, pointers to buffer_heads of
 877 *	partially truncated blocks - in @chain[].bh and pointers to
 878 *	their last elements that should not be removed - in
 879 *	@chain[].p. Return value is the pointer to last filled element
 880 *	of @chain.
 881 *
 882 *	The work left to caller to do the actual freeing of subtrees:
 883 *		a) free the subtree starting from *@top
 884 *		b) free the subtrees whose roots are stored in
 885 *			(@chain[i].p+1 .. end of @chain[i].bh->b_data)
 886 *		c) free the subtrees growing from the inode past the @chain[0].
 887 *			(no partially truncated stuff there).  */
 888
 889static Indirect *ext4_find_shared(struct inode *inode, int depth,
 890				  ext4_lblk_t offsets[4], Indirect chain[4],
 891				  __le32 *top)
 892{
 893	Indirect *partial, *p;
 894	int k, err;
 895
 896	*top = 0;
 897	/* Make k index the deepest non-null offset + 1 */
 898	for (k = depth; k > 1 && !offsets[k-1]; k--)
 899		;
 900	partial = ext4_get_branch(inode, k, offsets, chain, &err);
 901	/* Writer: pointers */
 902	if (!partial)
 903		partial = chain + k-1;
 904	/*
 905	 * If the branch acquired continuation since we've looked at it -
 906	 * fine, it should all survive and (new) top doesn't belong to us.
 907	 */
 908	if (!partial->key && *partial->p)
 909		/* Writer: end */
 910		goto no_top;
 911	for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
 912		;
 913	/*
 914	 * OK, we've found the last block that must survive. The rest of our
 915	 * branch should be detached before unlocking. However, if that rest
 916	 * of branch is all ours and does not grow immediately from the inode
 917	 * it's easier to cheat and just decrement partial->p.
 918	 */
 919	if (p == chain + k - 1 && p > chain) {
 920		p->p--;
 921	} else {
 922		*top = *p->p;
 923		/* Nope, don't do this in ext4.  Must leave the tree intact */
 924#if 0
 925		*p->p = 0;
 926#endif
 927	}
 928	/* Writer: end */
 929
 930	while (partial > p) {
 931		brelse(partial->bh);
 932		partial--;
 933	}
 934no_top:
 935	return partial;
 936}
 937
 938/*
 939 * Zero a number of block pointers in either an inode or an indirect block.
 940 * If we restart the transaction we must again get write access to the
 941 * indirect block for further modification.
 942 *
 943 * We release `count' blocks on disk, but (last - first) may be greater
 944 * than `count' because there can be holes in there.
 945 *
 946 * Return 0 on success, 1 on invalid block range
 947 * and < 0 on fatal error.
 948 */
 949static int ext4_clear_blocks(handle_t *handle, struct inode *inode,
 950			     struct buffer_head *bh,
 951			     ext4_fsblk_t block_to_free,
 952			     unsigned long count, __le32 *first,
 953			     __le32 *last)
 954{
 955	__le32 *p;
 956	int	flags = EXT4_FREE_BLOCKS_VALIDATED;
 957	int	err;
 958
 959	if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
 960		flags |= EXT4_FREE_BLOCKS_FORGET | EXT4_FREE_BLOCKS_METADATA;
 961	else if (ext4_should_journal_data(inode))
 962		flags |= EXT4_FREE_BLOCKS_FORGET;
 963
 964	if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), block_to_free,
 965				   count)) {
 966		EXT4_ERROR_INODE(inode, "attempt to clear invalid "
 967				 "blocks %llu len %lu",
 968				 (unsigned long long) block_to_free, count);
 969		return 1;
 970	}
 971
 972	if (try_to_extend_transaction(handle, inode)) {
 973		if (bh) {
 974			BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
 975			err = ext4_handle_dirty_metadata(handle, inode, bh);
 976			if (unlikely(err))
 977				goto out_err;
 978		}
 979		err = ext4_mark_inode_dirty(handle, inode);
 980		if (unlikely(err))
 981			goto out_err;
 982		err = ext4_truncate_restart_trans(handle, inode,
 983					ext4_blocks_for_truncate(inode));
 984		if (unlikely(err))
 985			goto out_err;
 986		if (bh) {
 987			BUFFER_TRACE(bh, "retaking write access");
 988			err = ext4_journal_get_write_access(handle, bh);
 989			if (unlikely(err))
 990				goto out_err;
 991		}
 992	}
 993
 994	for (p = first; p < last; p++)
 995		*p = 0;
 996
 997	ext4_free_blocks(handle, inode, NULL, block_to_free, count, flags);
 998	return 0;
 999out_err:
1000	ext4_std_error(inode->i_sb, err);
1001	return err;
1002}
1003
1004/**
1005 * ext4_free_data - free a list of data blocks
1006 * @handle:	handle for this transaction
1007 * @inode:	inode we are dealing with
1008 * @this_bh:	indirect buffer_head which contains *@first and *@last
1009 * @first:	array of block numbers
1010 * @last:	points immediately past the end of array
1011 *
1012 * We are freeing all blocks referred from that array (numbers are stored as
1013 * little-endian 32-bit) and updating @inode->i_blocks appropriately.
1014 *
1015 * We accumulate contiguous runs of blocks to free.  Conveniently, if these
1016 * blocks are contiguous then releasing them at one time will only affect one
1017 * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
1018 * actually use a lot of journal space.
1019 *
1020 * @this_bh will be %NULL if @first and @last point into the inode's direct
1021 * block pointers.
1022 */
1023static void ext4_free_data(handle_t *handle, struct inode *inode,
1024			   struct buffer_head *this_bh,
1025			   __le32 *first, __le32 *last)
1026{
1027	ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
1028	unsigned long count = 0;	    /* Number of blocks in the run */
1029	__le32 *block_to_free_p = NULL;	    /* Pointer into inode/ind
1030					       corresponding to
1031					       block_to_free */
1032	ext4_fsblk_t nr;		    /* Current block # */
1033	__le32 *p;			    /* Pointer into inode/ind
1034					       for current block */
1035	int err = 0;
1036
1037	if (this_bh) {				/* For indirect block */
1038		BUFFER_TRACE(this_bh, "get_write_access");
1039		err = ext4_journal_get_write_access(handle, this_bh);
1040		/* Important: if we can't update the indirect pointers
1041		 * to the blocks, we can't free them. */
1042		if (err)
1043			return;
1044	}
1045
1046	for (p = first; p < last; p++) {
1047		nr = le32_to_cpu(*p);
1048		if (nr) {
1049			/* accumulate blocks to free if they're contiguous */
1050			if (count == 0) {
1051				block_to_free = nr;
1052				block_to_free_p = p;
1053				count = 1;
1054			} else if (nr == block_to_free + count) {
1055				count++;
1056			} else {
1057				err = ext4_clear_blocks(handle, inode, this_bh,
1058						        block_to_free, count,
1059						        block_to_free_p, p);
1060				if (err)
1061					break;
1062				block_to_free = nr;
1063				block_to_free_p = p;
1064				count = 1;
1065			}
1066		}
1067	}
1068
1069	if (!err && count > 0)
1070		err = ext4_clear_blocks(handle, inode, this_bh, block_to_free,
1071					count, block_to_free_p, p);
1072	if (err < 0)
1073		/* fatal error */
1074		return;
1075
1076	if (this_bh) {
1077		BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
1078
1079		/*
1080		 * The buffer head should have an attached journal head at this
1081		 * point. However, if the data is corrupted and an indirect
1082		 * block pointed to itself, it would have been detached when
1083		 * the block was cleared. Check for this instead of OOPSing.
1084		 */
1085		if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
1086			ext4_handle_dirty_metadata(handle, inode, this_bh);
1087		else
1088			EXT4_ERROR_INODE(inode,
1089					 "circular indirect block detected at "
1090					 "block %llu",
1091				(unsigned long long) this_bh->b_blocknr);
1092	}
1093}
1094
1095/**
1096 *	ext4_free_branches - free an array of branches
1097 *	@handle: JBD handle for this transaction
1098 *	@inode:	inode we are dealing with
1099 *	@parent_bh: the buffer_head which contains *@first and *@last
1100 *	@first:	array of block numbers
1101 *	@last:	pointer immediately past the end of array
1102 *	@depth:	depth of the branches to free
1103 *
1104 *	We are freeing all blocks referred from these branches (numbers are
1105 *	stored as little-endian 32-bit) and updating @inode->i_blocks
1106 *	appropriately.
1107 */
1108static void ext4_free_branches(handle_t *handle, struct inode *inode,
1109			       struct buffer_head *parent_bh,
1110			       __le32 *first, __le32 *last, int depth)
1111{
1112	ext4_fsblk_t nr;
1113	__le32 *p;
1114
1115	if (ext4_handle_is_aborted(handle))
1116		return;
1117
1118	if (depth--) {
1119		struct buffer_head *bh;
1120		int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1121		p = last;
1122		while (--p >= first) {
1123			nr = le32_to_cpu(*p);
1124			if (!nr)
1125				continue;		/* A hole */
1126
1127			if (!ext4_data_block_valid(EXT4_SB(inode->i_sb),
1128						   nr, 1)) {
1129				EXT4_ERROR_INODE(inode,
1130						 "invalid indirect mapped "
1131						 "block %lu (level %d)",
1132						 (unsigned long) nr, depth);
1133				break;
1134			}
1135
1136			/* Go read the buffer for the next level down */
1137			bh = sb_bread(inode->i_sb, nr);
1138
1139			/*
1140			 * A read failure? Report error and clear slot
1141			 * (should be rare).
1142			 */
1143			if (!bh) {
1144				EXT4_ERROR_INODE_BLOCK(inode, nr,
1145						       "Read failure");
1146				continue;
1147			}
1148
1149			/* This zaps the entire block.  Bottom up. */
1150			BUFFER_TRACE(bh, "free child branches");
1151			ext4_free_branches(handle, inode, bh,
1152					(__le32 *) bh->b_data,
1153					(__le32 *) bh->b_data + addr_per_block,
1154					depth);
1155			brelse(bh);
1156
1157			/*
1158			 * Everything below this this pointer has been
1159			 * released.  Now let this top-of-subtree go.
1160			 *
1161			 * We want the freeing of this indirect block to be
1162			 * atomic in the journal with the updating of the
1163			 * bitmap block which owns it.  So make some room in
1164			 * the journal.
1165			 *
1166			 * We zero the parent pointer *after* freeing its
1167			 * pointee in the bitmaps, so if extend_transaction()
1168			 * for some reason fails to put the bitmap changes and
1169			 * the release into the same transaction, recovery
1170			 * will merely complain about releasing a free block,
1171			 * rather than leaking blocks.
1172			 */
1173			if (ext4_handle_is_aborted(handle))
1174				return;
1175			if (try_to_extend_transaction(handle, inode)) {
1176				ext4_mark_inode_dirty(handle, inode);
1177				ext4_truncate_restart_trans(handle, inode,
1178					    ext4_blocks_for_truncate(inode));
1179			}
1180
1181			/*
1182			 * The forget flag here is critical because if
1183			 * we are journaling (and not doing data
1184			 * journaling), we have to make sure a revoke
1185			 * record is written to prevent the journal
1186			 * replay from overwriting the (former)
1187			 * indirect block if it gets reallocated as a
1188			 * data block.  This must happen in the same
1189			 * transaction where the data blocks are
1190			 * actually freed.
1191			 */
1192			ext4_free_blocks(handle, inode, NULL, nr, 1,
1193					 EXT4_FREE_BLOCKS_METADATA|
1194					 EXT4_FREE_BLOCKS_FORGET);
1195
1196			if (parent_bh) {
1197				/*
1198				 * The block which we have just freed is
1199				 * pointed to by an indirect block: journal it
1200				 */
1201				BUFFER_TRACE(parent_bh, "get_write_access");
1202				if (!ext4_journal_get_write_access(handle,
1203								   parent_bh)){
1204					*p = 0;
1205					BUFFER_TRACE(parent_bh,
1206					"call ext4_handle_dirty_metadata");
1207					ext4_handle_dirty_metadata(handle,
1208								   inode,
1209								   parent_bh);
1210				}
1211			}
1212		}
1213	} else {
1214		/* We have reached the bottom of the tree. */
1215		BUFFER_TRACE(parent_bh, "free data blocks");
1216		ext4_free_data(handle, inode, parent_bh, first, last);
1217	}
1218}
1219
1220void ext4_ind_truncate(handle_t *handle, struct inode *inode)
1221{
1222	struct ext4_inode_info *ei = EXT4_I(inode);
1223	__le32 *i_data = ei->i_data;
1224	int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1225	ext4_lblk_t offsets[4];
1226	Indirect chain[4];
1227	Indirect *partial;
1228	__le32 nr = 0;
1229	int n = 0;
1230	ext4_lblk_t last_block, max_block;
1231	unsigned blocksize = inode->i_sb->s_blocksize;
1232
1233	last_block = (inode->i_size + blocksize-1)
1234					>> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1235	max_block = (EXT4_SB(inode->i_sb)->s_bitmap_maxbytes + blocksize-1)
1236					>> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1237
1238	if (last_block != max_block) {
1239		n = ext4_block_to_path(inode, last_block, offsets, NULL);
1240		if (n == 0)
1241			return;
1242	}
1243
1244	ext4_es_remove_extent(inode, last_block, EXT_MAX_BLOCKS - last_block);
1245
1246	/*
1247	 * The orphan list entry will now protect us from any crash which
1248	 * occurs before the truncate completes, so it is now safe to propagate
1249	 * the new, shorter inode size (held for now in i_size) into the
1250	 * on-disk inode. We do this via i_disksize, which is the value which
1251	 * ext4 *really* writes onto the disk inode.
1252	 */
1253	ei->i_disksize = inode->i_size;
1254
1255	if (last_block == max_block) {
1256		/*
1257		 * It is unnecessary to free any data blocks if last_block is
1258		 * equal to the indirect block limit.
1259		 */
1260		return;
1261	} else if (n == 1) {		/* direct blocks */
1262		ext4_free_data(handle, inode, NULL, i_data+offsets[0],
1263			       i_data + EXT4_NDIR_BLOCKS);
1264		goto do_indirects;
1265	}
1266
1267	partial = ext4_find_shared(inode, n, offsets, chain, &nr);
1268	/* Kill the top of shared branch (not detached) */
1269	if (nr) {
1270		if (partial == chain) {
1271			/* Shared branch grows from the inode */
1272			ext4_free_branches(handle, inode, NULL,
1273					   &nr, &nr+1, (chain+n-1) - partial);
1274			*partial->p = 0;
1275			/*
1276			 * We mark the inode dirty prior to restart,
1277			 * and prior to stop.  No need for it here.
1278			 */
1279		} else {
1280			/* Shared branch grows from an indirect block */
1281			BUFFER_TRACE(partial->bh, "get_write_access");
1282			ext4_free_branches(handle, inode, partial->bh,
1283					partial->p,
1284					partial->p+1, (chain+n-1) - partial);
1285		}
1286	}
1287	/* Clear the ends of indirect blocks on the shared branch */
1288	while (partial > chain) {
1289		ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
1290				   (__le32*)partial->bh->b_data+addr_per_block,
1291				   (chain+n-1) - partial);
1292		BUFFER_TRACE(partial->bh, "call brelse");
1293		brelse(partial->bh);
1294		partial--;
1295	}
1296do_indirects:
1297	/* Kill the remaining (whole) subtrees */
1298	switch (offsets[0]) {
1299	default:
1300		nr = i_data[EXT4_IND_BLOCK];
1301		if (nr) {
1302			ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
1303			i_data[EXT4_IND_BLOCK] = 0;
1304		}
1305	case EXT4_IND_BLOCK:
1306		nr = i_data[EXT4_DIND_BLOCK];
1307		if (nr) {
1308			ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
1309			i_data[EXT4_DIND_BLOCK] = 0;
1310		}
1311	case EXT4_DIND_BLOCK:
1312		nr = i_data[EXT4_TIND_BLOCK];
1313		if (nr) {
1314			ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
1315			i_data[EXT4_TIND_BLOCK] = 0;
1316		}
1317	case EXT4_TIND_BLOCK:
1318		;
1319	}
1320}
1321
1322/**
1323 *	ext4_ind_remove_space - remove space from the range
1324 *	@handle: JBD handle for this transaction
1325 *	@inode:	inode we are dealing with
1326 *	@start:	First block to remove
1327 *	@end:	One block after the last block to remove (exclusive)
1328 *
1329 *	Free the blocks in the defined range (end is exclusive endpoint of
1330 *	range). This is used by ext4_punch_hole().
1331 */
1332int ext4_ind_remove_space(handle_t *handle, struct inode *inode,
1333			  ext4_lblk_t start, ext4_lblk_t end)
1334{
1335	struct ext4_inode_info *ei = EXT4_I(inode);
1336	__le32 *i_data = ei->i_data;
1337	int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1338	ext4_lblk_t offsets[4], offsets2[4];
1339	Indirect chain[4], chain2[4];
1340	Indirect *partial, *partial2;
1341	ext4_lblk_t max_block;
1342	__le32 nr = 0, nr2 = 0;
1343	int n = 0, n2 = 0;
1344	unsigned blocksize = inode->i_sb->s_blocksize;
1345
1346	max_block = (EXT4_SB(inode->i_sb)->s_bitmap_maxbytes + blocksize-1)
1347					>> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1348	if (end >= max_block)
1349		end = max_block;
1350	if ((start >= end) || (start > max_block))
1351		return 0;
1352
1353	n = ext4_block_to_path(inode, start, offsets, NULL);
1354	n2 = ext4_block_to_path(inode, end, offsets2, NULL);
1355
1356	BUG_ON(n > n2);
1357
1358	if ((n == 1) && (n == n2)) {
1359		/* We're punching only within direct block range */
1360		ext4_free_data(handle, inode, NULL, i_data + offsets[0],
1361			       i_data + offsets2[0]);
1362		return 0;
1363	} else if (n2 > n) {
1364		/*
1365		 * Start and end are on a different levels so we're going to
1366		 * free partial block at start, and partial block at end of
1367		 * the range. If there are some levels in between then
1368		 * do_indirects label will take care of that.
1369		 */
1370
1371		if (n == 1) {
1372			/*
1373			 * Start is at the direct block level, free
1374			 * everything to the end of the level.
1375			 */
1376			ext4_free_data(handle, inode, NULL, i_data + offsets[0],
1377				       i_data + EXT4_NDIR_BLOCKS);
1378			goto end_range;
1379		}
1380
1381
1382		partial = ext4_find_shared(inode, n, offsets, chain, &nr);
1383		if (nr) {
1384			if (partial == chain) {
1385				/* Shared branch grows from the inode */
1386				ext4_free_branches(handle, inode, NULL,
1387					   &nr, &nr+1, (chain+n-1) - partial);
1388				*partial->p = 0;
1389			} else {
1390				/* Shared branch grows from an indirect block */
1391				BUFFER_TRACE(partial->bh, "get_write_access");
1392				ext4_free_branches(handle, inode, partial->bh,
1393					partial->p,
1394					partial->p+1, (chain+n-1) - partial);
1395			}
1396		}
1397
1398		/*
1399		 * Clear the ends of indirect blocks on the shared branch
1400		 * at the start of the range
1401		 */
1402		while (partial > chain) {
1403			ext4_free_branches(handle, inode, partial->bh,
1404				partial->p + 1,
1405				(__le32 *)partial->bh->b_data+addr_per_block,
1406				(chain+n-1) - partial);
1407			BUFFER_TRACE(partial->bh, "call brelse");
1408			brelse(partial->bh);
1409			partial--;
1410		}
1411
1412end_range:
1413		partial2 = ext4_find_shared(inode, n2, offsets2, chain2, &nr2);
1414		if (nr2) {
1415			if (partial2 == chain2) {
1416				/*
1417				 * Remember, end is exclusive so here we're at
1418				 * the start of the next level we're not going
1419				 * to free. Everything was covered by the start
1420				 * of the range.
1421				 */
1422				goto do_indirects;
1423			}
1424		} else {
1425			/*
1426			 * ext4_find_shared returns Indirect structure which
1427			 * points to the last element which should not be
1428			 * removed by truncate. But this is end of the range
1429			 * in punch_hole so we need to point to the next element
1430			 */
1431			partial2->p++;
1432		}
1433
1434		/*
1435		 * Clear the ends of indirect blocks on the shared branch
1436		 * at the end of the range
1437		 */
1438		while (partial2 > chain2) {
1439			ext4_free_branches(handle, inode, partial2->bh,
1440					   (__le32 *)partial2->bh->b_data,
1441					   partial2->p,
1442					   (chain2+n2-1) - partial2);
1443			BUFFER_TRACE(partial2->bh, "call brelse");
1444			brelse(partial2->bh);
1445			partial2--;
1446		}
1447		goto do_indirects;
1448	}
1449
1450	/* Punch happened within the same level (n == n2) */
1451	partial = ext4_find_shared(inode, n, offsets, chain, &nr);
1452	partial2 = ext4_find_shared(inode, n2, offsets2, chain2, &nr2);
1453
1454	/* Free top, but only if partial2 isn't its subtree. */
1455	if (nr) {
1456		int level = min(partial - chain, partial2 - chain2);
1457		int i;
1458		int subtree = 1;
1459
1460		for (i = 0; i <= level; i++) {
1461			if (offsets[i] != offsets2[i]) {
1462				subtree = 0;
1463				break;
1464			}
1465		}
1466
1467		if (!subtree) {
1468			if (partial == chain) {
1469				/* Shared branch grows from the inode */
1470				ext4_free_branches(handle, inode, NULL,
1471						   &nr, &nr+1,
1472						   (chain+n-1) - partial);
1473				*partial->p = 0;
1474			} else {
1475				/* Shared branch grows from an indirect block */
1476				BUFFER_TRACE(partial->bh, "get_write_access");
1477				ext4_free_branches(handle, inode, partial->bh,
1478						   partial->p,
1479						   partial->p+1,
1480						   (chain+n-1) - partial);
1481			}
1482		}
1483	}
1484
1485	if (!nr2) {
1486		/*
1487		 * ext4_find_shared returns Indirect structure which
1488		 * points to the last element which should not be
1489		 * removed by truncate. But this is end of the range
1490		 * in punch_hole so we need to point to the next element
1491		 */
1492		partial2->p++;
1493	}
1494
1495	while (partial > chain || partial2 > chain2) {
1496		int depth = (chain+n-1) - partial;
1497		int depth2 = (chain2+n2-1) - partial2;
1498
1499		if (partial > chain && partial2 > chain2 &&
1500		    partial->bh->b_blocknr == partial2->bh->b_blocknr) {
1501			/*
1502			 * We've converged on the same block. Clear the range,
1503			 * then we're done.
1504			 */
1505			ext4_free_branches(handle, inode, partial->bh,
1506					   partial->p + 1,
1507					   partial2->p,
1508					   (chain+n-1) - partial);
1509			BUFFER_TRACE(partial->bh, "call brelse");
1510			brelse(partial->bh);
1511			BUFFER_TRACE(partial2->bh, "call brelse");
1512			brelse(partial2->bh);
1513			return 0;
1514		}
1515
1516		/*
1517		 * The start and end partial branches may not be at the same
1518		 * level even though the punch happened within one level. So, we
1519		 * give them a chance to arrive at the same level, then walk
1520		 * them in step with each other until we converge on the same
1521		 * block.
1522		 */
1523		if (partial > chain && depth <= depth2) {
1524			ext4_free_branches(handle, inode, partial->bh,
1525					   partial->p + 1,
1526					   (__le32 *)partial->bh->b_data+addr_per_block,
1527					   (chain+n-1) - partial);
1528			BUFFER_TRACE(partial->bh, "call brelse");
1529			brelse(partial->bh);
1530			partial--;
1531		}
1532		if (partial2 > chain2 && depth2 <= depth) {
1533			ext4_free_branches(handle, inode, partial2->bh,
1534					   (__le32 *)partial2->bh->b_data,
1535					   partial2->p,
1536					   (chain2+n2-1) - partial2);
1537			BUFFER_TRACE(partial2->bh, "call brelse");
1538			brelse(partial2->bh);
1539			partial2--;
1540		}
1541	}
1542	return 0;
1543
1544do_indirects:
1545	/* Kill the remaining (whole) subtrees */
1546	switch (offsets[0]) {
1547	default:
1548		if (++n >= n2)
1549			return 0;
1550		nr = i_data[EXT4_IND_BLOCK];
1551		if (nr) {
1552			ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
1553			i_data[EXT4_IND_BLOCK] = 0;
1554		}
1555	case EXT4_IND_BLOCK:
1556		if (++n >= n2)
1557			return 0;
1558		nr = i_data[EXT4_DIND_BLOCK];
1559		if (nr) {
1560			ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
1561			i_data[EXT4_DIND_BLOCK] = 0;
1562		}
1563	case EXT4_DIND_BLOCK:
1564		if (++n >= n2)
1565			return 0;
1566		nr = i_data[EXT4_TIND_BLOCK];
1567		if (nr) {
1568			ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
1569			i_data[EXT4_TIND_BLOCK] = 0;
1570		}
1571	case EXT4_TIND_BLOCK:
1572		;
1573	}
1574	return 0;
1575}