1// SPDX-License-Identifier: GPL-2.0-or-later
   2/*
   3 * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
   4 */
   5
   6#include <linux/fs.h>
   7#include <linux/slab.h>
   8#include <linux/highmem.h>
   9#include <linux/pagemap.h>
  10#include <asm/byteorder.h>
  11#include <linux/swap.h>
  12#include <linux/mpage.h>
  13#include <linux/quotaops.h>
  14#include <linux/blkdev.h>
  15#include <linux/uio.h>
  16#include <linux/mm.h>
  17
  18#include <cluster/masklog.h>
  19
  20#include "ocfs2.h"
  21
  22#include "alloc.h"
  23#include "aops.h"
  24#include "dlmglue.h"
  25#include "extent_map.h"
  26#include "file.h"
  27#include "inode.h"
  28#include "journal.h"
  29#include "suballoc.h"
  30#include "super.h"
  31#include "symlink.h"
  32#include "refcounttree.h"
  33#include "ocfs2_trace.h"
  34
  35#include "buffer_head_io.h"
  36#include "dir.h"
  37#include "namei.h"
  38#include "sysfile.h"
  39
  40static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
  41				   struct buffer_head *bh_result, int create)
  42{
  43	int err = -EIO;
  44	int status;
  45	struct ocfs2_dinode *fe = NULL;
  46	struct buffer_head *bh = NULL;
  47	struct buffer_head *buffer_cache_bh = NULL;
  48	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  49	void *kaddr;
  50
  51	trace_ocfs2_symlink_get_block(
  52			(unsigned long long)OCFS2_I(inode)->ip_blkno,
  53			(unsigned long long)iblock, bh_result, create);
  54
  55	BUG_ON(ocfs2_inode_is_fast_symlink(inode));
  56
  57	if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
  58		mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
  59		     (unsigned long long)iblock);
  60		goto bail;
  61	}
  62
  63	status = ocfs2_read_inode_block(inode, &bh);
  64	if (status < 0) {
  65		mlog_errno(status);
  66		goto bail;
  67	}
  68	fe = (struct ocfs2_dinode *) bh->b_data;
  69
  70	if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
  71						    le32_to_cpu(fe->i_clusters))) {
  72		err = -ENOMEM;
  73		mlog(ML_ERROR, "block offset is outside the allocated size: "
  74		     "%llu\n", (unsigned long long)iblock);
  75		goto bail;
  76	}
  77
  78	/* We don't use the page cache to create symlink data, so if
  79	 * need be, copy it over from the buffer cache. */
  80	if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
  81		u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
  82			    iblock;
  83		buffer_cache_bh = sb_getblk(osb->sb, blkno);
  84		if (!buffer_cache_bh) {
  85			err = -ENOMEM;
  86			mlog(ML_ERROR, "couldn't getblock for symlink!\n");
  87			goto bail;
  88		}
  89
  90		/* we haven't locked out transactions, so a commit
  91		 * could've happened. Since we've got a reference on
  92		 * the bh, even if it commits while we're doing the
  93		 * copy, the data is still good. */
  94		if (buffer_jbd(buffer_cache_bh)
  95		    && ocfs2_inode_is_new(inode)) {
  96			kaddr = kmap_atomic(bh_result->b_page);
  97			if (!kaddr) {
  98				mlog(ML_ERROR, "couldn't kmap!\n");
  99				goto bail;
 100			}
 101			memcpy(kaddr + (bh_result->b_size * iblock),
 102			       buffer_cache_bh->b_data,
 103			       bh_result->b_size);
 104			kunmap_atomic(kaddr);
 105			set_buffer_uptodate(bh_result);
 106		}
 107		brelse(buffer_cache_bh);
 108	}
 109
 110	map_bh(bh_result, inode->i_sb,
 111	       le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
 112
 113	err = 0;
 114
 115bail:
 116	brelse(bh);
 117
 118	return err;
 119}
 120
 121static int ocfs2_lock_get_block(struct inode *inode, sector_t iblock,
 122		    struct buffer_head *bh_result, int create)
 123{
 124	int ret = 0;
 125	struct ocfs2_inode_info *oi = OCFS2_I(inode);
 126
 127	down_read(&oi->ip_alloc_sem);
 128	ret = ocfs2_get_block(inode, iblock, bh_result, create);
 129	up_read(&oi->ip_alloc_sem);
 130
 131	return ret;
 132}
 133
 134int ocfs2_get_block(struct inode *inode, sector_t iblock,
 135		    struct buffer_head *bh_result, int create)
 136{
 137	int err = 0;
 138	unsigned int ext_flags;
 139	u64 max_blocks = bh_result->b_size >> inode->i_blkbits;
 140	u64 p_blkno, count, past_eof;
 141	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 142
 143	trace_ocfs2_get_block((unsigned long long)OCFS2_I(inode)->ip_blkno,
 144			      (unsigned long long)iblock, bh_result, create);
 145
 146	if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
 147		mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
 148		     inode, inode->i_ino);
 149
 150	if (S_ISLNK(inode->i_mode)) {
 151		/* this always does I/O for some reason. */
 152		err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
 153		goto bail;
 154	}
 155
 156	err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, &count,
 157					  &ext_flags);
 158	if (err) {
 159		mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
 160		     "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
 161		     (unsigned long long)p_blkno);
 162		goto bail;
 163	}
 164
 165	if (max_blocks < count)
 166		count = max_blocks;
 167
 168	/*
 169	 * ocfs2 never allocates in this function - the only time we
 170	 * need to use BH_New is when we're extending i_size on a file
 171	 * system which doesn't support holes, in which case BH_New
 172	 * allows __block_write_begin() to zero.
 173	 *
 174	 * If we see this on a sparse file system, then a truncate has
 175	 * raced us and removed the cluster. In this case, we clear
 176	 * the buffers dirty and uptodate bits and let the buffer code
 177	 * ignore it as a hole.
 178	 */
 179	if (create && p_blkno == 0 && ocfs2_sparse_alloc(osb)) {
 180		clear_buffer_dirty(bh_result);
 181		clear_buffer_uptodate(bh_result);
 182		goto bail;
 183	}
 184
 185	/* Treat the unwritten extent as a hole for zeroing purposes. */
 186	if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 187		map_bh(bh_result, inode->i_sb, p_blkno);
 188
 189	bh_result->b_size = count << inode->i_blkbits;
 190
 191	if (!ocfs2_sparse_alloc(osb)) {
 192		if (p_blkno == 0) {
 193			err = -EIO;
 194			mlog(ML_ERROR,
 195			     "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
 196			     (unsigned long long)iblock,
 197			     (unsigned long long)p_blkno,
 198			     (unsigned long long)OCFS2_I(inode)->ip_blkno);
 199			mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
 200			dump_stack();
 201			goto bail;
 202		}
 203	}
 204
 205	past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 206
 207	trace_ocfs2_get_block_end((unsigned long long)OCFS2_I(inode)->ip_blkno,
 208				  (unsigned long long)past_eof);
 209	if (create && (iblock >= past_eof))
 210		set_buffer_new(bh_result);
 211
 212bail:
 213	if (err < 0)
 214		err = -EIO;
 215
 216	return err;
 217}
 218
 219int ocfs2_read_inline_data(struct inode *inode, struct page *page,
 220			   struct buffer_head *di_bh)
 221{
 222	void *kaddr;
 223	loff_t size;
 224	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 225
 226	if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
 227		ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag\n",
 228			    (unsigned long long)OCFS2_I(inode)->ip_blkno);
 229		return -EROFS;
 230	}
 231
 232	size = i_size_read(inode);
 233
 234	if (size > PAGE_SIZE ||
 235	    size > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) {
 236		ocfs2_error(inode->i_sb,
 237			    "Inode %llu has with inline data has bad size: %Lu\n",
 238			    (unsigned long long)OCFS2_I(inode)->ip_blkno,
 239			    (unsigned long long)size);
 240		return -EROFS;
 241	}
 242
 243	kaddr = kmap_atomic(page);
 244	if (size)
 245		memcpy(kaddr, di->id2.i_data.id_data, size);
 246	/* Clear the remaining part of the page */
 247	memset(kaddr + size, 0, PAGE_SIZE - size);
 248	flush_dcache_page(page);
 249	kunmap_atomic(kaddr);
 250
 251	SetPageUptodate(page);
 252
 253	return 0;
 254}
 255
 256static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
 257{
 258	int ret;
 259	struct buffer_head *di_bh = NULL;
 260
 261	BUG_ON(!PageLocked(page));
 262	BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
 263
 264	ret = ocfs2_read_inode_block(inode, &di_bh);
 265	if (ret) {
 266		mlog_errno(ret);
 267		goto out;
 268	}
 269
 270	ret = ocfs2_read_inline_data(inode, page, di_bh);
 271out:
 272	unlock_page(page);
 273
 274	brelse(di_bh);
 275	return ret;
 276}
 277
 278static int ocfs2_read_folio(struct file *file, struct folio *folio)
 279{
 280	struct inode *inode = folio->mapping->host;
 281	struct ocfs2_inode_info *oi = OCFS2_I(inode);
 282	loff_t start = folio_pos(folio);
 283	int ret, unlock = 1;
 284
 285	trace_ocfs2_readpage((unsigned long long)oi->ip_blkno, folio->index);
 286
 287	ret = ocfs2_inode_lock_with_page(inode, NULL, 0, &folio->page);
 288	if (ret != 0) {
 289		if (ret == AOP_TRUNCATED_PAGE)
 290			unlock = 0;
 291		mlog_errno(ret);
 292		goto out;
 293	}
 294
 295	if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
 296		/*
 297		 * Unlock the folio and cycle ip_alloc_sem so that we don't
 298		 * busyloop waiting for ip_alloc_sem to unlock
 299		 */
 300		ret = AOP_TRUNCATED_PAGE;
 301		folio_unlock(folio);
 302		unlock = 0;
 303		down_read(&oi->ip_alloc_sem);
 304		up_read(&oi->ip_alloc_sem);
 305		goto out_inode_unlock;
 306	}
 307
 308	/*
 309	 * i_size might have just been updated as we grabed the meta lock.  We
 310	 * might now be discovering a truncate that hit on another node.
 311	 * block_read_full_folio->get_block freaks out if it is asked to read
 312	 * beyond the end of a file, so we check here.  Callers
 313	 * (generic_file_read, vm_ops->fault) are clever enough to check i_size
 314	 * and notice that the folio they just read isn't needed.
 315	 *
 316	 * XXX sys_readahead() seems to get that wrong?
 317	 */
 318	if (start >= i_size_read(inode)) {
 319		folio_zero_segment(folio, 0, folio_size(folio));
 320		folio_mark_uptodate(folio);
 321		ret = 0;
 322		goto out_alloc;
 323	}
 324
 325	if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 326		ret = ocfs2_readpage_inline(inode, &folio->page);
 327	else
 328		ret = block_read_full_folio(folio, ocfs2_get_block);
 329	unlock = 0;
 330
 331out_alloc:
 332	up_read(&oi->ip_alloc_sem);
 333out_inode_unlock:
 334	ocfs2_inode_unlock(inode, 0);
 335out:
 336	if (unlock)
 337		folio_unlock(folio);
 338	return ret;
 339}
 340
 341/*
 342 * This is used only for read-ahead. Failures or difficult to handle
 343 * situations are safe to ignore.
 344 *
 345 * Right now, we don't bother with BH_Boundary - in-inode extent lists
 346 * are quite large (243 extents on 4k blocks), so most inodes don't
 347 * grow out to a tree. If need be, detecting boundary extents could
 348 * trivially be added in a future version of ocfs2_get_block().
 349 */
 350static void ocfs2_readahead(struct readahead_control *rac)
 351{
 352	int ret;
 353	struct inode *inode = rac->mapping->host;
 354	struct ocfs2_inode_info *oi = OCFS2_I(inode);
 355
 356	/*
 357	 * Use the nonblocking flag for the dlm code to avoid page
 358	 * lock inversion, but don't bother with retrying.
 359	 */
 360	ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
 361	if (ret)
 362		return;
 363
 364	if (down_read_trylock(&oi->ip_alloc_sem) == 0)
 365		goto out_unlock;
 366
 367	/*
 368	 * Don't bother with inline-data. There isn't anything
 369	 * to read-ahead in that case anyway...
 370	 */
 371	if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 372		goto out_up;
 373
 374	/*
 375	 * Check whether a remote node truncated this file - we just
 376	 * drop out in that case as it's not worth handling here.
 377	 */
 378	if (readahead_pos(rac) >= i_size_read(inode))
 379		goto out_up;
 380
 381	mpage_readahead(rac, ocfs2_get_block);
 382
 383out_up:
 384	up_read(&oi->ip_alloc_sem);
 385out_unlock:
 386	ocfs2_inode_unlock(inode, 0);
 387}
 388
 389/* Note: Because we don't support holes, our allocation has
 390 * already happened (allocation writes zeros to the file data)
 391 * so we don't have to worry about ordered writes in
 392 * ocfs2_writepage.
 393 *
 394 * ->writepage is called during the process of invalidating the page cache
 395 * during blocked lock processing.  It can't block on any cluster locks
 396 * to during block mapping.  It's relying on the fact that the block
 397 * mapping can't have disappeared under the dirty pages that it is
 398 * being asked to write back.
 399 */
 400static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
 
 401{
 402	trace_ocfs2_writepage(
 403		(unsigned long long)OCFS2_I(page->mapping->host)->ip_blkno,
 404		page->index);
 405
 406	return block_write_full_page(page, ocfs2_get_block, wbc);
 407}
 408
 409/* Taken from ext3. We don't necessarily need the full blown
 410 * functionality yet, but IMHO it's better to cut and paste the whole
 411 * thing so we can avoid introducing our own bugs (and easily pick up
 412 * their fixes when they happen) --Mark */
 413int walk_page_buffers(	handle_t *handle,
 414			struct buffer_head *head,
 415			unsigned from,
 416			unsigned to,
 417			int *partial,
 418			int (*fn)(	handle_t *handle,
 419					struct buffer_head *bh))
 420{
 421	struct buffer_head *bh;
 422	unsigned block_start, block_end;
 423	unsigned blocksize = head->b_size;
 424	int err, ret = 0;
 425	struct buffer_head *next;
 426
 427	for (	bh = head, block_start = 0;
 428		ret == 0 && (bh != head || !block_start);
 429	    	block_start = block_end, bh = next)
 430	{
 431		next = bh->b_this_page;
 432		block_end = block_start + blocksize;
 433		if (block_end <= from || block_start >= to) {
 434			if (partial && !buffer_uptodate(bh))
 435				*partial = 1;
 436			continue;
 437		}
 438		err = (*fn)(handle, bh);
 439		if (!ret)
 440			ret = err;
 441	}
 442	return ret;
 443}
 444
 445static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
 446{
 447	sector_t status;
 448	u64 p_blkno = 0;
 449	int err = 0;
 450	struct inode *inode = mapping->host;
 451
 452	trace_ocfs2_bmap((unsigned long long)OCFS2_I(inode)->ip_blkno,
 453			 (unsigned long long)block);
 454
 455	/*
 456	 * The swap code (ab-)uses ->bmap to get a block mapping and then
 457	 * bypasseѕ the file system for actual I/O.  We really can't allow
 458	 * that on refcounted inodes, so we have to skip out here.  And yes,
 459	 * 0 is the magic code for a bmap error..
 460	 */
 461	if (ocfs2_is_refcount_inode(inode))
 462		return 0;
 463
 464	/* We don't need to lock journal system files, since they aren't
 465	 * accessed concurrently from multiple nodes.
 466	 */
 467	if (!INODE_JOURNAL(inode)) {
 468		err = ocfs2_inode_lock(inode, NULL, 0);
 469		if (err) {
 470			if (err != -ENOENT)
 471				mlog_errno(err);
 472			goto bail;
 473		}
 474		down_read(&OCFS2_I(inode)->ip_alloc_sem);
 475	}
 476
 477	if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
 478		err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
 479						  NULL);
 480
 481	if (!INODE_JOURNAL(inode)) {
 482		up_read(&OCFS2_I(inode)->ip_alloc_sem);
 483		ocfs2_inode_unlock(inode, 0);
 484	}
 485
 486	if (err) {
 487		mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
 488		     (unsigned long long)block);
 489		mlog_errno(err);
 490		goto bail;
 491	}
 492
 493bail:
 494	status = err ? 0 : p_blkno;
 495
 496	return status;
 497}
 498
 499static bool ocfs2_release_folio(struct folio *folio, gfp_t wait)
 500{
 501	if (!folio_buffers(folio))
 502		return false;
 503	return try_to_free_buffers(folio);
 504}
 505
 506static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
 507					    u32 cpos,
 508					    unsigned int *start,
 509					    unsigned int *end)
 510{
 511	unsigned int cluster_start = 0, cluster_end = PAGE_SIZE;
 512
 513	if (unlikely(PAGE_SHIFT > osb->s_clustersize_bits)) {
 514		unsigned int cpp;
 515
 516		cpp = 1 << (PAGE_SHIFT - osb->s_clustersize_bits);
 517
 518		cluster_start = cpos % cpp;
 519		cluster_start = cluster_start << osb->s_clustersize_bits;
 520
 521		cluster_end = cluster_start + osb->s_clustersize;
 522	}
 523
 524	BUG_ON(cluster_start > PAGE_SIZE);
 525	BUG_ON(cluster_end > PAGE_SIZE);
 526
 527	if (start)
 528		*start = cluster_start;
 529	if (end)
 530		*end = cluster_end;
 531}
 532
 533/*
 534 * 'from' and 'to' are the region in the page to avoid zeroing.
 535 *
 536 * If pagesize > clustersize, this function will avoid zeroing outside
 537 * of the cluster boundary.
 538 *
 539 * from == to == 0 is code for "zero the entire cluster region"
 540 */
 541static void ocfs2_clear_page_regions(struct page *page,
 542				     struct ocfs2_super *osb, u32 cpos,
 543				     unsigned from, unsigned to)
 544{
 545	void *kaddr;
 546	unsigned int cluster_start, cluster_end;
 547
 548	ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
 549
 550	kaddr = kmap_atomic(page);
 551
 552	if (from || to) {
 553		if (from > cluster_start)
 554			memset(kaddr + cluster_start, 0, from - cluster_start);
 555		if (to < cluster_end)
 556			memset(kaddr + to, 0, cluster_end - to);
 557	} else {
 558		memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
 559	}
 560
 561	kunmap_atomic(kaddr);
 562}
 563
 564/*
 565 * Nonsparse file systems fully allocate before we get to the write
 566 * code. This prevents ocfs2_write() from tagging the write as an
 567 * allocating one, which means ocfs2_map_page_blocks() might try to
 568 * read-in the blocks at the tail of our file. Avoid reading them by
 569 * testing i_size against each block offset.
 570 */
 571static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
 572				 unsigned int block_start)
 573{
 574	u64 offset = page_offset(page) + block_start;
 575
 576	if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 577		return 1;
 578
 579	if (i_size_read(inode) > offset)
 580		return 1;
 581
 582	return 0;
 583}
 584
 585/*
 586 * Some of this taken from __block_write_begin(). We already have our
 587 * mapping by now though, and the entire write will be allocating or
 588 * it won't, so not much need to use BH_New.
 589 *
 590 * This will also skip zeroing, which is handled externally.
 591 */
 592int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
 593			  struct inode *inode, unsigned int from,
 594			  unsigned int to, int new)
 595{
 
 596	int ret = 0;
 597	struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
 598	unsigned int block_end, block_start;
 599	unsigned int bsize = i_blocksize(inode);
 600
 601	if (!page_has_buffers(page))
 602		create_empty_buffers(page, bsize, 0);
 
 603
 604	head = page_buffers(page);
 605	for (bh = head, block_start = 0; bh != head || !block_start;
 606	     bh = bh->b_this_page, block_start += bsize) {
 607		block_end = block_start + bsize;
 608
 609		clear_buffer_new(bh);
 610
 611		/*
 612		 * Ignore blocks outside of our i/o range -
 613		 * they may belong to unallocated clusters.
 614		 */
 615		if (block_start >= to || block_end <= from) {
 616			if (PageUptodate(page))
 617				set_buffer_uptodate(bh);
 618			continue;
 619		}
 620
 621		/*
 622		 * For an allocating write with cluster size >= page
 623		 * size, we always write the entire page.
 624		 */
 625		if (new)
 626			set_buffer_new(bh);
 627
 628		if (!buffer_mapped(bh)) {
 629			map_bh(bh, inode->i_sb, *p_blkno);
 630			clean_bdev_bh_alias(bh);
 631		}
 632
 633		if (PageUptodate(page)) {
 634			set_buffer_uptodate(bh);
 635		} else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
 636			   !buffer_new(bh) &&
 637			   ocfs2_should_read_blk(inode, page, block_start) &&
 638			   (block_start < from || block_end > to)) {
 639			bh_read_nowait(bh, 0);
 640			*wait_bh++=bh;
 641		}
 642
 643		*p_blkno = *p_blkno + 1;
 644	}
 645
 646	/*
 647	 * If we issued read requests - let them complete.
 648	 */
 649	while(wait_bh > wait) {
 650		wait_on_buffer(*--wait_bh);
 651		if (!buffer_uptodate(*wait_bh))
 652			ret = -EIO;
 653	}
 654
 655	if (ret == 0 || !new)
 656		return ret;
 657
 658	/*
 659	 * If we get -EIO above, zero out any newly allocated blocks
 660	 * to avoid exposing stale data.
 661	 */
 662	bh = head;
 663	block_start = 0;
 664	do {
 665		block_end = block_start + bsize;
 666		if (block_end <= from)
 667			goto next_bh;
 668		if (block_start >= to)
 669			break;
 670
 671		zero_user(page, block_start, bh->b_size);
 672		set_buffer_uptodate(bh);
 673		mark_buffer_dirty(bh);
 674
 675next_bh:
 676		block_start = block_end;
 677		bh = bh->b_this_page;
 678	} while (bh != head);
 679
 680	return ret;
 681}
 682
 683#if (PAGE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
 684#define OCFS2_MAX_CTXT_PAGES	1
 685#else
 686#define OCFS2_MAX_CTXT_PAGES	(OCFS2_MAX_CLUSTERSIZE / PAGE_SIZE)
 687#endif
 688
 689#define OCFS2_MAX_CLUSTERS_PER_PAGE	(PAGE_SIZE / OCFS2_MIN_CLUSTERSIZE)
 690
 691struct ocfs2_unwritten_extent {
 692	struct list_head	ue_node;
 693	struct list_head	ue_ip_node;
 694	u32			ue_cpos;
 695	u32			ue_phys;
 696};
 697
 698/*
 699 * Describe the state of a single cluster to be written to.
 700 */
 701struct ocfs2_write_cluster_desc {
 702	u32		c_cpos;
 703	u32		c_phys;
 704	/*
 705	 * Give this a unique field because c_phys eventually gets
 706	 * filled.
 707	 */
 708	unsigned	c_new;
 709	unsigned	c_clear_unwritten;
 710	unsigned	c_needs_zero;
 711};
 712
 713struct ocfs2_write_ctxt {
 714	/* Logical cluster position / len of write */
 715	u32				w_cpos;
 716	u32				w_clen;
 717
 718	/* First cluster allocated in a nonsparse extend */
 719	u32				w_first_new_cpos;
 720
 721	/* Type of caller. Must be one of buffer, mmap, direct.  */
 722	ocfs2_write_type_t		w_type;
 723
 724	struct ocfs2_write_cluster_desc	w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
 725
 726	/*
 727	 * This is true if page_size > cluster_size.
 728	 *
 729	 * It triggers a set of special cases during write which might
 730	 * have to deal with allocating writes to partial pages.
 731	 */
 732	unsigned int			w_large_pages;
 733
 734	/*
 735	 * Pages involved in this write.
 736	 *
 737	 * w_target_page is the page being written to by the user.
 738	 *
 739	 * w_pages is an array of pages which always contains
 740	 * w_target_page, and in the case of an allocating write with
 741	 * page_size < cluster size, it will contain zero'd and mapped
 742	 * pages adjacent to w_target_page which need to be written
 743	 * out in so that future reads from that region will get
 744	 * zero's.
 745	 */
 746	unsigned int			w_num_pages;
 747	struct page			*w_pages[OCFS2_MAX_CTXT_PAGES];
 748	struct page			*w_target_page;
 749
 750	/*
 751	 * w_target_locked is used for page_mkwrite path indicating no unlocking
 752	 * against w_target_page in ocfs2_write_end_nolock.
 753	 */
 754	unsigned int			w_target_locked:1;
 755
 756	/*
 757	 * ocfs2_write_end() uses this to know what the real range to
 758	 * write in the target should be.
 759	 */
 760	unsigned int			w_target_from;
 761	unsigned int			w_target_to;
 762
 763	/*
 764	 * We could use journal_current_handle() but this is cleaner,
 765	 * IMHO -Mark
 766	 */
 767	handle_t			*w_handle;
 768
 769	struct buffer_head		*w_di_bh;
 770
 771	struct ocfs2_cached_dealloc_ctxt w_dealloc;
 772
 773	struct list_head		w_unwritten_list;
 774	unsigned int			w_unwritten_count;
 775};
 776
 777void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
 778{
 779	int i;
 780
 781	for(i = 0; i < num_pages; i++) {
 782		if (pages[i]) {
 783			unlock_page(pages[i]);
 784			mark_page_accessed(pages[i]);
 785			put_page(pages[i]);
 786		}
 787	}
 788}
 789
 790static void ocfs2_unlock_pages(struct ocfs2_write_ctxt *wc)
 791{
 792	int i;
 793
 794	/*
 795	 * w_target_locked is only set to true in the page_mkwrite() case.
 796	 * The intent is to allow us to lock the target page from write_begin()
 797	 * to write_end(). The caller must hold a ref on w_target_page.
 798	 */
 799	if (wc->w_target_locked) {
 800		BUG_ON(!wc->w_target_page);
 801		for (i = 0; i < wc->w_num_pages; i++) {
 802			if (wc->w_target_page == wc->w_pages[i]) {
 803				wc->w_pages[i] = NULL;
 804				break;
 805			}
 806		}
 807		mark_page_accessed(wc->w_target_page);
 808		put_page(wc->w_target_page);
 809	}
 810	ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
 811}
 812
 813static void ocfs2_free_unwritten_list(struct inode *inode,
 814				 struct list_head *head)
 815{
 816	struct ocfs2_inode_info *oi = OCFS2_I(inode);
 817	struct ocfs2_unwritten_extent *ue = NULL, *tmp = NULL;
 818
 819	list_for_each_entry_safe(ue, tmp, head, ue_node) {
 820		list_del(&ue->ue_node);
 821		spin_lock(&oi->ip_lock);
 822		list_del(&ue->ue_ip_node);
 823		spin_unlock(&oi->ip_lock);
 824		kfree(ue);
 825	}
 826}
 827
 828static void ocfs2_free_write_ctxt(struct inode *inode,
 829				  struct ocfs2_write_ctxt *wc)
 830{
 831	ocfs2_free_unwritten_list(inode, &wc->w_unwritten_list);
 832	ocfs2_unlock_pages(wc);
 833	brelse(wc->w_di_bh);
 834	kfree(wc);
 835}
 836
 837static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
 838				  struct ocfs2_super *osb, loff_t pos,
 839				  unsigned len, ocfs2_write_type_t type,
 840				  struct buffer_head *di_bh)
 841{
 842	u32 cend;
 843	struct ocfs2_write_ctxt *wc;
 844
 845	wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
 846	if (!wc)
 847		return -ENOMEM;
 848
 849	wc->w_cpos = pos >> osb->s_clustersize_bits;
 850	wc->w_first_new_cpos = UINT_MAX;
 851	cend = (pos + len - 1) >> osb->s_clustersize_bits;
 852	wc->w_clen = cend - wc->w_cpos + 1;
 853	get_bh(di_bh);
 854	wc->w_di_bh = di_bh;
 855	wc->w_type = type;
 856
 857	if (unlikely(PAGE_SHIFT > osb->s_clustersize_bits))
 858		wc->w_large_pages = 1;
 859	else
 860		wc->w_large_pages = 0;
 861
 862	ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
 863	INIT_LIST_HEAD(&wc->w_unwritten_list);
 864
 865	*wcp = wc;
 866
 867	return 0;
 868}
 869
 870/*
 871 * If a page has any new buffers, zero them out here, and mark them uptodate
 872 * and dirty so they'll be written out (in order to prevent uninitialised
 873 * block data from leaking). And clear the new bit.
 874 */
 875static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
 876{
 877	unsigned int block_start, block_end;
 878	struct buffer_head *head, *bh;
 879
 880	BUG_ON(!PageLocked(page));
 881	if (!page_has_buffers(page))
 882		return;
 883
 884	bh = head = page_buffers(page);
 885	block_start = 0;
 886	do {
 887		block_end = block_start + bh->b_size;
 888
 889		if (buffer_new(bh)) {
 890			if (block_end > from && block_start < to) {
 891				if (!PageUptodate(page)) {
 892					unsigned start, end;
 893
 894					start = max(from, block_start);
 895					end = min(to, block_end);
 896
 897					zero_user_segment(page, start, end);
 898					set_buffer_uptodate(bh);
 899				}
 900
 901				clear_buffer_new(bh);
 902				mark_buffer_dirty(bh);
 903			}
 904		}
 905
 906		block_start = block_end;
 907		bh = bh->b_this_page;
 908	} while (bh != head);
 909}
 910
 911/*
 912 * Only called when we have a failure during allocating write to write
 913 * zero's to the newly allocated region.
 914 */
 915static void ocfs2_write_failure(struct inode *inode,
 916				struct ocfs2_write_ctxt *wc,
 917				loff_t user_pos, unsigned user_len)
 918{
 919	int i;
 920	unsigned from = user_pos & (PAGE_SIZE - 1),
 921		to = user_pos + user_len;
 922	struct page *tmppage;
 923
 924	if (wc->w_target_page)
 925		ocfs2_zero_new_buffers(wc->w_target_page, from, to);
 926
 927	for(i = 0; i < wc->w_num_pages; i++) {
 928		tmppage = wc->w_pages[i];
 929
 930		if (tmppage && page_has_buffers(tmppage)) {
 931			if (ocfs2_should_order_data(inode))
 932				ocfs2_jbd2_inode_add_write(wc->w_handle, inode,
 933							   user_pos, user_len);
 934
 935			block_commit_write(tmppage, from, to);
 936		}
 937	}
 938}
 939
 940static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
 941					struct ocfs2_write_ctxt *wc,
 942					struct page *page, u32 cpos,
 943					loff_t user_pos, unsigned user_len,
 944					int new)
 945{
 946	int ret;
 947	unsigned int map_from = 0, map_to = 0;
 948	unsigned int cluster_start, cluster_end;
 949	unsigned int user_data_from = 0, user_data_to = 0;
 950
 951	ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
 952					&cluster_start, &cluster_end);
 953
 954	/* treat the write as new if the a hole/lseek spanned across
 955	 * the page boundary.
 956	 */
 957	new = new | ((i_size_read(inode) <= page_offset(page)) &&
 958			(page_offset(page) <= user_pos));
 959
 960	if (page == wc->w_target_page) {
 961		map_from = user_pos & (PAGE_SIZE - 1);
 962		map_to = map_from + user_len;
 963
 964		if (new)
 965			ret = ocfs2_map_page_blocks(page, p_blkno, inode,
 966						    cluster_start, cluster_end,
 967						    new);
 968		else
 969			ret = ocfs2_map_page_blocks(page, p_blkno, inode,
 970						    map_from, map_to, new);
 971		if (ret) {
 972			mlog_errno(ret);
 973			goto out;
 974		}
 975
 976		user_data_from = map_from;
 977		user_data_to = map_to;
 978		if (new) {
 979			map_from = cluster_start;
 980			map_to = cluster_end;
 981		}
 982	} else {
 983		/*
 984		 * If we haven't allocated the new page yet, we
 985		 * shouldn't be writing it out without copying user
 986		 * data. This is likely a math error from the caller.
 987		 */
 988		BUG_ON(!new);
 989
 990		map_from = cluster_start;
 991		map_to = cluster_end;
 992
 993		ret = ocfs2_map_page_blocks(page, p_blkno, inode,
 994					    cluster_start, cluster_end, new);
 995		if (ret) {
 996			mlog_errno(ret);
 997			goto out;
 998		}
 999	}
1000
1001	/*
1002	 * Parts of newly allocated pages need to be zero'd.
1003	 *
1004	 * Above, we have also rewritten 'to' and 'from' - as far as
1005	 * the rest of the function is concerned, the entire cluster
1006	 * range inside of a page needs to be written.
1007	 *
1008	 * We can skip this if the page is up to date - it's already
1009	 * been zero'd from being read in as a hole.
1010	 */
1011	if (new && !PageUptodate(page))
1012		ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1013					 cpos, user_data_from, user_data_to);
1014
1015	flush_dcache_page(page);
1016
1017out:
1018	return ret;
1019}
1020
1021/*
1022 * This function will only grab one clusters worth of pages.
1023 */
1024static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1025				      struct ocfs2_write_ctxt *wc,
1026				      u32 cpos, loff_t user_pos,
1027				      unsigned user_len, int new,
1028				      struct page *mmap_page)
1029{
1030	int ret = 0, i;
1031	unsigned long start, target_index, end_index, index;
1032	struct inode *inode = mapping->host;
1033	loff_t last_byte;
1034
1035	target_index = user_pos >> PAGE_SHIFT;
1036
1037	/*
1038	 * Figure out how many pages we'll be manipulating here. For
1039	 * non allocating write, we just change the one
1040	 * page. Otherwise, we'll need a whole clusters worth.  If we're
1041	 * writing past i_size, we only need enough pages to cover the
1042	 * last page of the write.
1043	 */
1044	if (new) {
1045		wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1046		start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1047		/*
1048		 * We need the index *past* the last page we could possibly
1049		 * touch.  This is the page past the end of the write or
1050		 * i_size, whichever is greater.
1051		 */
1052		last_byte = max(user_pos + user_len, i_size_read(inode));
1053		BUG_ON(last_byte < 1);
1054		end_index = ((last_byte - 1) >> PAGE_SHIFT) + 1;
1055		if ((start + wc->w_num_pages) > end_index)
1056			wc->w_num_pages = end_index - start;
1057	} else {
1058		wc->w_num_pages = 1;
1059		start = target_index;
1060	}
1061	end_index = (user_pos + user_len - 1) >> PAGE_SHIFT;
1062
1063	for(i = 0; i < wc->w_num_pages; i++) {
1064		index = start + i;
1065
1066		if (index >= target_index && index <= end_index &&
1067		    wc->w_type == OCFS2_WRITE_MMAP) {
1068			/*
1069			 * ocfs2_pagemkwrite() is a little different
1070			 * and wants us to directly use the page
1071			 * passed in.
1072			 */
1073			lock_page(mmap_page);
1074
1075			/* Exit and let the caller retry */
1076			if (mmap_page->mapping != mapping) {
1077				WARN_ON(mmap_page->mapping);
1078				unlock_page(mmap_page);
1079				ret = -EAGAIN;
1080				goto out;
1081			}
1082
1083			get_page(mmap_page);
1084			wc->w_pages[i] = mmap_page;
1085			wc->w_target_locked = true;
1086		} else if (index >= target_index && index <= end_index &&
1087			   wc->w_type == OCFS2_WRITE_DIRECT) {
1088			/* Direct write has no mapping page. */
1089			wc->w_pages[i] = NULL;
1090			continue;
1091		} else {
1092			wc->w_pages[i] = find_or_create_page(mapping, index,
1093							     GFP_NOFS);
1094			if (!wc->w_pages[i]) {
1095				ret = -ENOMEM;
1096				mlog_errno(ret);
1097				goto out;
1098			}
1099		}
1100		wait_for_stable_page(wc->w_pages[i]);
1101
1102		if (index == target_index)
1103			wc->w_target_page = wc->w_pages[i];
1104	}
1105out:
1106	if (ret)
1107		wc->w_target_locked = false;
1108	return ret;
1109}
1110
1111/*
1112 * Prepare a single cluster for write one cluster into the file.
1113 */
1114static int ocfs2_write_cluster(struct address_space *mapping,
1115			       u32 *phys, unsigned int new,
1116			       unsigned int clear_unwritten,
1117			       unsigned int should_zero,
1118			       struct ocfs2_alloc_context *data_ac,
1119			       struct ocfs2_alloc_context *meta_ac,
1120			       struct ocfs2_write_ctxt *wc, u32 cpos,
1121			       loff_t user_pos, unsigned user_len)
1122{
1123	int ret, i;
1124	u64 p_blkno;
1125	struct inode *inode = mapping->host;
1126	struct ocfs2_extent_tree et;
1127	int bpc = ocfs2_clusters_to_blocks(inode->i_sb, 1);
1128
1129	if (new) {
1130		u32 tmp_pos;
1131
1132		/*
1133		 * This is safe to call with the page locks - it won't take
1134		 * any additional semaphores or cluster locks.
1135		 */
1136		tmp_pos = cpos;
1137		ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1138					   &tmp_pos, 1, !clear_unwritten,
1139					   wc->w_di_bh, wc->w_handle,
1140					   data_ac, meta_ac, NULL);
1141		/*
1142		 * This shouldn't happen because we must have already
1143		 * calculated the correct meta data allocation required. The
1144		 * internal tree allocation code should know how to increase
1145		 * transaction credits itself.
1146		 *
1147		 * If need be, we could handle -EAGAIN for a
1148		 * RESTART_TRANS here.
1149		 */
1150		mlog_bug_on_msg(ret == -EAGAIN,
1151				"Inode %llu: EAGAIN return during allocation.\n",
1152				(unsigned long long)OCFS2_I(inode)->ip_blkno);
1153		if (ret < 0) {
1154			mlog_errno(ret);
1155			goto out;
1156		}
1157	} else if (clear_unwritten) {
1158		ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1159					      wc->w_di_bh);
1160		ret = ocfs2_mark_extent_written(inode, &et,
1161						wc->w_handle, cpos, 1, *phys,
1162						meta_ac, &wc->w_dealloc);
1163		if (ret < 0) {
1164			mlog_errno(ret);
1165			goto out;
1166		}
1167	}
1168
1169	/*
1170	 * The only reason this should fail is due to an inability to
1171	 * find the extent added.
1172	 */
1173	ret = ocfs2_get_clusters(inode, cpos, phys, NULL, NULL);
1174	if (ret < 0) {
1175		mlog(ML_ERROR, "Get physical blkno failed for inode %llu, "
1176			    "at logical cluster %u",
1177			    (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
1178		goto out;
1179	}
1180
1181	BUG_ON(*phys == 0);
1182
1183	p_blkno = ocfs2_clusters_to_blocks(inode->i_sb, *phys);
1184	if (!should_zero)
1185		p_blkno += (user_pos >> inode->i_sb->s_blocksize_bits) & (u64)(bpc - 1);
1186
1187	for(i = 0; i < wc->w_num_pages; i++) {
1188		int tmpret;
1189
1190		/* This is the direct io target page. */
1191		if (wc->w_pages[i] == NULL) {
1192			p_blkno++;
1193			continue;
1194		}
1195
1196		tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1197						      wc->w_pages[i], cpos,
1198						      user_pos, user_len,
1199						      should_zero);
1200		if (tmpret) {
1201			mlog_errno(tmpret);
1202			if (ret == 0)
1203				ret = tmpret;
1204		}
1205	}
1206
1207	/*
1208	 * We only have cleanup to do in case of allocating write.
1209	 */
1210	if (ret && new)
1211		ocfs2_write_failure(inode, wc, user_pos, user_len);
1212
1213out:
1214
1215	return ret;
1216}
1217
1218static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1219				       struct ocfs2_alloc_context *data_ac,
1220				       struct ocfs2_alloc_context *meta_ac,
1221				       struct ocfs2_write_ctxt *wc,
1222				       loff_t pos, unsigned len)
1223{
1224	int ret, i;
1225	loff_t cluster_off;
1226	unsigned int local_len = len;
1227	struct ocfs2_write_cluster_desc *desc;
1228	struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1229
1230	for (i = 0; i < wc->w_clen; i++) {
1231		desc = &wc->w_desc[i];
1232
1233		/*
1234		 * We have to make sure that the total write passed in
1235		 * doesn't extend past a single cluster.
1236		 */
1237		local_len = len;
1238		cluster_off = pos & (osb->s_clustersize - 1);
1239		if ((cluster_off + local_len) > osb->s_clustersize)
1240			local_len = osb->s_clustersize - cluster_off;
1241
1242		ret = ocfs2_write_cluster(mapping, &desc->c_phys,
1243					  desc->c_new,
1244					  desc->c_clear_unwritten,
1245					  desc->c_needs_zero,
1246					  data_ac, meta_ac,
1247					  wc, desc->c_cpos, pos, local_len);
1248		if (ret) {
1249			mlog_errno(ret);
1250			goto out;
1251		}
1252
1253		len -= local_len;
1254		pos += local_len;
1255	}
1256
1257	ret = 0;
1258out:
1259	return ret;
1260}
1261
1262/*
1263 * ocfs2_write_end() wants to know which parts of the target page it
1264 * should complete the write on. It's easiest to compute them ahead of
1265 * time when a more complete view of the write is available.
1266 */
1267static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1268					struct ocfs2_write_ctxt *wc,
1269					loff_t pos, unsigned len, int alloc)
1270{
1271	struct ocfs2_write_cluster_desc *desc;
1272
1273	wc->w_target_from = pos & (PAGE_SIZE - 1);
1274	wc->w_target_to = wc->w_target_from + len;
1275
1276	if (alloc == 0)
1277		return;
1278
1279	/*
1280	 * Allocating write - we may have different boundaries based
1281	 * on page size and cluster size.
1282	 *
1283	 * NOTE: We can no longer compute one value from the other as
1284	 * the actual write length and user provided length may be
1285	 * different.
1286	 */
1287
1288	if (wc->w_large_pages) {
1289		/*
1290		 * We only care about the 1st and last cluster within
1291		 * our range and whether they should be zero'd or not. Either
1292		 * value may be extended out to the start/end of a
1293		 * newly allocated cluster.
1294		 */
1295		desc = &wc->w_desc[0];
1296		if (desc->c_needs_zero)
1297			ocfs2_figure_cluster_boundaries(osb,
1298							desc->c_cpos,
1299							&wc->w_target_from,
1300							NULL);
1301
1302		desc = &wc->w_desc[wc->w_clen - 1];
1303		if (desc->c_needs_zero)
1304			ocfs2_figure_cluster_boundaries(osb,
1305							desc->c_cpos,
1306							NULL,
1307							&wc->w_target_to);
1308	} else {
1309		wc->w_target_from = 0;
1310		wc->w_target_to = PAGE_SIZE;
1311	}
1312}
1313
1314/*
1315 * Check if this extent is marked UNWRITTEN by direct io. If so, we need not to
1316 * do the zero work. And should not to clear UNWRITTEN since it will be cleared
1317 * by the direct io procedure.
1318 * If this is a new extent that allocated by direct io, we should mark it in
1319 * the ip_unwritten_list.
1320 */
1321static int ocfs2_unwritten_check(struct inode *inode,
1322				 struct ocfs2_write_ctxt *wc,
1323				 struct ocfs2_write_cluster_desc *desc)
1324{
1325	struct ocfs2_inode_info *oi = OCFS2_I(inode);
1326	struct ocfs2_unwritten_extent *ue = NULL, *new = NULL;
1327	int ret = 0;
1328
1329	if (!desc->c_needs_zero)
1330		return 0;
1331
1332retry:
1333	spin_lock(&oi->ip_lock);
1334	/* Needs not to zero no metter buffer or direct. The one who is zero
1335	 * the cluster is doing zero. And he will clear unwritten after all
1336	 * cluster io finished. */
1337	list_for_each_entry(ue, &oi->ip_unwritten_list, ue_ip_node) {
1338		if (desc->c_cpos == ue->ue_cpos) {
1339			BUG_ON(desc->c_new);
1340			desc->c_needs_zero = 0;
1341			desc->c_clear_unwritten = 0;
1342			goto unlock;
1343		}
1344	}
1345
1346	if (wc->w_type != OCFS2_WRITE_DIRECT)
1347		goto unlock;
1348
1349	if (new == NULL) {
1350		spin_unlock(&oi->ip_lock);
1351		new = kmalloc(sizeof(struct ocfs2_unwritten_extent),
1352			     GFP_NOFS);
1353		if (new == NULL) {
1354			ret = -ENOMEM;
1355			goto out;
1356		}
1357		goto retry;
1358	}
1359	/* This direct write will doing zero. */
1360	new->ue_cpos = desc->c_cpos;
1361	new->ue_phys = desc->c_phys;
1362	desc->c_clear_unwritten = 0;
1363	list_add_tail(&new->ue_ip_node, &oi->ip_unwritten_list);
1364	list_add_tail(&new->ue_node, &wc->w_unwritten_list);
1365	wc->w_unwritten_count++;
1366	new = NULL;
1367unlock:
1368	spin_unlock(&oi->ip_lock);
1369out:
1370	kfree(new);
1371	return ret;
1372}
1373
1374/*
1375 * Populate each single-cluster write descriptor in the write context
1376 * with information about the i/o to be done.
1377 *
1378 * Returns the number of clusters that will have to be allocated, as
1379 * well as a worst case estimate of the number of extent records that
1380 * would have to be created during a write to an unwritten region.
1381 */
1382static int ocfs2_populate_write_desc(struct inode *inode,
1383				     struct ocfs2_write_ctxt *wc,
1384				     unsigned int *clusters_to_alloc,
1385				     unsigned int *extents_to_split)
1386{
1387	int ret;
1388	struct ocfs2_write_cluster_desc *desc;
1389	unsigned int num_clusters = 0;
1390	unsigned int ext_flags = 0;
1391	u32 phys = 0;
1392	int i;
1393
1394	*clusters_to_alloc = 0;
1395	*extents_to_split = 0;
1396
1397	for (i = 0; i < wc->w_clen; i++) {
1398		desc = &wc->w_desc[i];
1399		desc->c_cpos = wc->w_cpos + i;
1400
1401		if (num_clusters == 0) {
1402			/*
1403			 * Need to look up the next extent record.
1404			 */
1405			ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1406						 &num_clusters, &ext_flags);
1407			if (ret) {
1408				mlog_errno(ret);
1409				goto out;
1410			}
1411
1412			/* We should already CoW the refcountd extent. */
1413			BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1414
1415			/*
1416			 * Assume worst case - that we're writing in
1417			 * the middle of the extent.
1418			 *
1419			 * We can assume that the write proceeds from
1420			 * left to right, in which case the extent
1421			 * insert code is smart enough to coalesce the
1422			 * next splits into the previous records created.
1423			 */
1424			if (ext_flags & OCFS2_EXT_UNWRITTEN)
1425				*extents_to_split = *extents_to_split + 2;
1426		} else if (phys) {
1427			/*
1428			 * Only increment phys if it doesn't describe
1429			 * a hole.
1430			 */
1431			phys++;
1432		}
1433
1434		/*
1435		 * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1436		 * file that got extended.  w_first_new_cpos tells us
1437		 * where the newly allocated clusters are so we can
1438		 * zero them.
1439		 */
1440		if (desc->c_cpos >= wc->w_first_new_cpos) {
1441			BUG_ON(phys == 0);
1442			desc->c_needs_zero = 1;
1443		}
1444
1445		desc->c_phys = phys;
1446		if (phys == 0) {
1447			desc->c_new = 1;
1448			desc->c_needs_zero = 1;
1449			desc->c_clear_unwritten = 1;
1450			*clusters_to_alloc = *clusters_to_alloc + 1;
1451		}
1452
1453		if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1454			desc->c_clear_unwritten = 1;
1455			desc->c_needs_zero = 1;
1456		}
1457
1458		ret = ocfs2_unwritten_check(inode, wc, desc);
1459		if (ret) {
1460			mlog_errno(ret);
1461			goto out;
1462		}
1463
1464		num_clusters--;
1465	}
1466
1467	ret = 0;
1468out:
1469	return ret;
1470}
1471
1472static int ocfs2_write_begin_inline(struct address_space *mapping,
1473				    struct inode *inode,
1474				    struct ocfs2_write_ctxt *wc)
1475{
1476	int ret;
1477	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1478	struct page *page;
1479	handle_t *handle;
1480	struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1481
1482	handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1483	if (IS_ERR(handle)) {
1484		ret = PTR_ERR(handle);
1485		mlog_errno(ret);
1486		goto out;
1487	}
1488
1489	page = find_or_create_page(mapping, 0, GFP_NOFS);
1490	if (!page) {
1491		ocfs2_commit_trans(osb, handle);
1492		ret = -ENOMEM;
1493		mlog_errno(ret);
1494		goto out;
1495	}
1496	/*
1497	 * If we don't set w_num_pages then this page won't get unlocked
1498	 * and freed on cleanup of the write context.
1499	 */
1500	wc->w_pages[0] = wc->w_target_page = page;
1501	wc->w_num_pages = 1;
1502
1503	ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1504				      OCFS2_JOURNAL_ACCESS_WRITE);
1505	if (ret) {
1506		ocfs2_commit_trans(osb, handle);
1507
1508		mlog_errno(ret);
1509		goto out;
1510	}
1511
1512	if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1513		ocfs2_set_inode_data_inline(inode, di);
1514
1515	if (!PageUptodate(page)) {
1516		ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1517		if (ret) {
1518			ocfs2_commit_trans(osb, handle);
1519
1520			goto out;
1521		}
1522	}
1523
1524	wc->w_handle = handle;
1525out:
1526	return ret;
1527}
1528
1529int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1530{
1531	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1532
1533	if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1534		return 1;
1535	return 0;
1536}
1537
1538static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1539					  struct inode *inode, loff_t pos,
1540					  unsigned len, struct page *mmap_page,
1541					  struct ocfs2_write_ctxt *wc)
1542{
1543	int ret, written = 0;
1544	loff_t end = pos + len;
1545	struct ocfs2_inode_info *oi = OCFS2_I(inode);
1546	struct ocfs2_dinode *di = NULL;
1547
1548	trace_ocfs2_try_to_write_inline_data((unsigned long long)oi->ip_blkno,
1549					     len, (unsigned long long)pos,
1550					     oi->ip_dyn_features);
1551
1552	/*
1553	 * Handle inodes which already have inline data 1st.
1554	 */
1555	if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1556		if (mmap_page == NULL &&
1557		    ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1558			goto do_inline_write;
1559
1560		/*
1561		 * The write won't fit - we have to give this inode an
1562		 * inline extent list now.
1563		 */
1564		ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1565		if (ret)
1566			mlog_errno(ret);
1567		goto out;
1568	}
1569
1570	/*
1571	 * Check whether the inode can accept inline data.
1572	 */
1573	if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1574		return 0;
1575
1576	/*
1577	 * Check whether the write can fit.
1578	 */
1579	di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1580	if (mmap_page ||
1581	    end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1582		return 0;
1583
1584do_inline_write:
1585	ret = ocfs2_write_begin_inline(mapping, inode, wc);
1586	if (ret) {
1587		mlog_errno(ret);
1588		goto out;
1589	}
1590
1591	/*
1592	 * This signals to the caller that the data can be written
1593	 * inline.
1594	 */
1595	written = 1;
1596out:
1597	return written ? written : ret;
1598}
1599
1600/*
1601 * This function only does anything for file systems which can't
1602 * handle sparse files.
1603 *
1604 * What we want to do here is fill in any hole between the current end
1605 * of allocation and the end of our write. That way the rest of the
1606 * write path can treat it as an non-allocating write, which has no
1607 * special case code for sparse/nonsparse files.
1608 */
1609static int ocfs2_expand_nonsparse_inode(struct inode *inode,
1610					struct buffer_head *di_bh,
1611					loff_t pos, unsigned len,
1612					struct ocfs2_write_ctxt *wc)
1613{
1614	int ret;
1615	loff_t newsize = pos + len;
1616
1617	BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1618
1619	if (newsize <= i_size_read(inode))
1620		return 0;
1621
1622	ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
1623	if (ret)
1624		mlog_errno(ret);
1625
1626	/* There is no wc if this is call from direct. */
1627	if (wc)
1628		wc->w_first_new_cpos =
1629			ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1630
1631	return ret;
1632}
1633
1634static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
1635			   loff_t pos)
1636{
1637	int ret = 0;
1638
1639	BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1640	if (pos > i_size_read(inode))
1641		ret = ocfs2_zero_extend(inode, di_bh, pos);
1642
1643	return ret;
1644}
1645
1646int ocfs2_write_begin_nolock(struct address_space *mapping,
1647			     loff_t pos, unsigned len, ocfs2_write_type_t type,
1648			     struct page **pagep, void **fsdata,
1649			     struct buffer_head *di_bh, struct page *mmap_page)
1650{
1651	int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1652	unsigned int clusters_to_alloc, extents_to_split, clusters_need = 0;
1653	struct ocfs2_write_ctxt *wc;
1654	struct inode *inode = mapping->host;
1655	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1656	struct ocfs2_dinode *di;
1657	struct ocfs2_alloc_context *data_ac = NULL;
1658	struct ocfs2_alloc_context *meta_ac = NULL;
1659	handle_t *handle;
1660	struct ocfs2_extent_tree et;
1661	int try_free = 1, ret1;
1662
1663try_again:
1664	ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, type, di_bh);
1665	if (ret) {
1666		mlog_errno(ret);
1667		return ret;
1668	}
1669
1670	if (ocfs2_supports_inline_data(osb)) {
1671		ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1672						     mmap_page, wc);
1673		if (ret == 1) {
1674			ret = 0;
1675			goto success;
1676		}
1677		if (ret < 0) {
1678			mlog_errno(ret);
1679			goto out;
1680		}
1681	}
1682
1683	/* Direct io change i_size late, should not zero tail here. */
1684	if (type != OCFS2_WRITE_DIRECT) {
1685		if (ocfs2_sparse_alloc(osb))
1686			ret = ocfs2_zero_tail(inode, di_bh, pos);
1687		else
1688			ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos,
1689							   len, wc);
1690		if (ret) {
1691			mlog_errno(ret);
1692			goto out;
1693		}
1694	}
1695
1696	ret = ocfs2_check_range_for_refcount(inode, pos, len);
1697	if (ret < 0) {
1698		mlog_errno(ret);
1699		goto out;
1700	} else if (ret == 1) {
1701		clusters_need = wc->w_clen;
1702		ret = ocfs2_refcount_cow(inode, di_bh,
1703					 wc->w_cpos, wc->w_clen, UINT_MAX);
1704		if (ret) {
1705			mlog_errno(ret);
1706			goto out;
1707		}
1708	}
1709
1710	ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1711					&extents_to_split);
1712	if (ret) {
1713		mlog_errno(ret);
1714		goto out;
1715	}
1716	clusters_need += clusters_to_alloc;
1717
1718	di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1719
1720	trace_ocfs2_write_begin_nolock(
1721			(unsigned long long)OCFS2_I(inode)->ip_blkno,
1722			(long long)i_size_read(inode),
1723			le32_to_cpu(di->i_clusters),
1724			pos, len, type, mmap_page,
1725			clusters_to_alloc, extents_to_split);
1726
1727	/*
1728	 * We set w_target_from, w_target_to here so that
1729	 * ocfs2_write_end() knows which range in the target page to
1730	 * write out. An allocation requires that we write the entire
1731	 * cluster range.
1732	 */
1733	if (clusters_to_alloc || extents_to_split) {
1734		/*
1735		 * XXX: We are stretching the limits of
1736		 * ocfs2_lock_allocators(). It greatly over-estimates
1737		 * the work to be done.
1738		 */
1739		ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1740					      wc->w_di_bh);
1741		ret = ocfs2_lock_allocators(inode, &et,
1742					    clusters_to_alloc, extents_to_split,
1743					    &data_ac, &meta_ac);
1744		if (ret) {
1745			mlog_errno(ret);
1746			goto out;
1747		}
1748
1749		if (data_ac)
1750			data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
1751
1752		credits = ocfs2_calc_extend_credits(inode->i_sb,
1753						    &di->id2.i_list);
1754	} else if (type == OCFS2_WRITE_DIRECT)
1755		/* direct write needs not to start trans if no extents alloc. */
1756		goto success;
1757
1758	/*
1759	 * We have to zero sparse allocated clusters, unwritten extent clusters,
1760	 * and non-sparse clusters we just extended.  For non-sparse writes,
1761	 * we know zeros will only be needed in the first and/or last cluster.
1762	 */
1763	if (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1764			   wc->w_desc[wc->w_clen - 1].c_needs_zero))
1765		cluster_of_pages = 1;
1766	else
1767		cluster_of_pages = 0;
1768
1769	ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1770
1771	handle = ocfs2_start_trans(osb, credits);
1772	if (IS_ERR(handle)) {
1773		ret = PTR_ERR(handle);
1774		mlog_errno(ret);
1775		goto out;
1776	}
1777
1778	wc->w_handle = handle;
1779
1780	if (clusters_to_alloc) {
1781		ret = dquot_alloc_space_nodirty(inode,
1782			ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1783		if (ret)
1784			goto out_commit;
1785	}
1786
1787	ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1788				      OCFS2_JOURNAL_ACCESS_WRITE);
1789	if (ret) {
1790		mlog_errno(ret);
1791		goto out_quota;
1792	}
1793
1794	/*
1795	 * Fill our page array first. That way we've grabbed enough so
1796	 * that we can zero and flush if we error after adding the
1797	 * extent.
1798	 */
1799	ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
1800					 cluster_of_pages, mmap_page);
1801	if (ret) {
1802		/*
1803		 * ocfs2_grab_pages_for_write() returns -EAGAIN if it could not lock
1804		 * the target page. In this case, we exit with no error and no target
1805		 * page. This will trigger the caller, page_mkwrite(), to re-try
1806		 * the operation.
1807		 */
1808		if (type == OCFS2_WRITE_MMAP && ret == -EAGAIN) {
1809			BUG_ON(wc->w_target_page);
1810			ret = 0;
1811			goto out_quota;
1812		}
1813
1814		mlog_errno(ret);
1815		goto out_quota;
1816	}
1817
1818	ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1819					  len);
1820	if (ret) {
1821		mlog_errno(ret);
1822		goto out_quota;
1823	}
1824
1825	if (data_ac)
1826		ocfs2_free_alloc_context(data_ac);
1827	if (meta_ac)
1828		ocfs2_free_alloc_context(meta_ac);
1829
1830success:
1831	if (pagep)
1832		*pagep = wc->w_target_page;
1833	*fsdata = wc;
1834	return 0;
1835out_quota:
1836	if (clusters_to_alloc)
1837		dquot_free_space(inode,
1838			  ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1839out_commit:
1840	ocfs2_commit_trans(osb, handle);
1841
1842out:
1843	/*
1844	 * The mmapped page won't be unlocked in ocfs2_free_write_ctxt(),
1845	 * even in case of error here like ENOSPC and ENOMEM. So, we need
1846	 * to unlock the target page manually to prevent deadlocks when
1847	 * retrying again on ENOSPC, or when returning non-VM_FAULT_LOCKED
1848	 * to VM code.
1849	 */
1850	if (wc->w_target_locked)
1851		unlock_page(mmap_page);
1852
1853	ocfs2_free_write_ctxt(inode, wc);
1854
1855	if (data_ac) {
1856		ocfs2_free_alloc_context(data_ac);
1857		data_ac = NULL;
1858	}
1859	if (meta_ac) {
1860		ocfs2_free_alloc_context(meta_ac);
1861		meta_ac = NULL;
1862	}
1863
1864	if (ret == -ENOSPC && try_free) {
1865		/*
1866		 * Try to free some truncate log so that we can have enough
1867		 * clusters to allocate.
1868		 */
1869		try_free = 0;
1870
1871		ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need);
1872		if (ret1 == 1)
1873			goto try_again;
1874
1875		if (ret1 < 0)
1876			mlog_errno(ret1);
1877	}
1878
1879	return ret;
1880}
1881
1882static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1883			     loff_t pos, unsigned len,
1884			     struct page **pagep, void **fsdata)
1885{
1886	int ret;
1887	struct buffer_head *di_bh = NULL;
1888	struct inode *inode = mapping->host;
1889
1890	ret = ocfs2_inode_lock(inode, &di_bh, 1);
1891	if (ret) {
1892		mlog_errno(ret);
1893		return ret;
1894	}
1895
1896	/*
1897	 * Take alloc sem here to prevent concurrent lookups. That way
1898	 * the mapping, zeroing and tree manipulation within
1899	 * ocfs2_write() will be safe against ->read_folio(). This
1900	 * should also serve to lock out allocation from a shared
1901	 * writeable region.
1902	 */
1903	down_write(&OCFS2_I(inode)->ip_alloc_sem);
1904
1905	ret = ocfs2_write_begin_nolock(mapping, pos, len, OCFS2_WRITE_BUFFER,
1906				       pagep, fsdata, di_bh, NULL);
1907	if (ret) {
1908		mlog_errno(ret);
1909		goto out_fail;
1910	}
1911
1912	brelse(di_bh);
1913
1914	return 0;
1915
1916out_fail:
1917	up_write(&OCFS2_I(inode)->ip_alloc_sem);
1918
1919	brelse(di_bh);
1920	ocfs2_inode_unlock(inode, 1);
1921
1922	return ret;
1923}
1924
1925static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1926				   unsigned len, unsigned *copied,
1927				   struct ocfs2_dinode *di,
1928				   struct ocfs2_write_ctxt *wc)
1929{
1930	void *kaddr;
1931
1932	if (unlikely(*copied < len)) {
1933		if (!PageUptodate(wc->w_target_page)) {
1934			*copied = 0;
1935			return;
1936		}
1937	}
1938
1939	kaddr = kmap_atomic(wc->w_target_page);
1940	memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1941	kunmap_atomic(kaddr);
1942
1943	trace_ocfs2_write_end_inline(
1944	     (unsigned long long)OCFS2_I(inode)->ip_blkno,
1945	     (unsigned long long)pos, *copied,
1946	     le16_to_cpu(di->id2.i_data.id_count),
1947	     le16_to_cpu(di->i_dyn_features));
1948}
1949
1950int ocfs2_write_end_nolock(struct address_space *mapping,
1951			   loff_t pos, unsigned len, unsigned copied, void *fsdata)
1952{
1953	int i, ret;
1954	unsigned from, to, start = pos & (PAGE_SIZE - 1);
1955	struct inode *inode = mapping->host;
1956	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1957	struct ocfs2_write_ctxt *wc = fsdata;
1958	struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1959	handle_t *handle = wc->w_handle;
1960	struct page *tmppage;
1961
1962	BUG_ON(!list_empty(&wc->w_unwritten_list));
1963
1964	if (handle) {
1965		ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode),
1966				wc->w_di_bh, OCFS2_JOURNAL_ACCESS_WRITE);
1967		if (ret) {
1968			copied = ret;
1969			mlog_errno(ret);
1970			goto out;
1971		}
1972	}
1973
1974	if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1975		ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
1976		goto out_write_size;
1977	}
1978
1979	if (unlikely(copied < len) && wc->w_target_page) {
 
 
1980		if (!PageUptodate(wc->w_target_page))
1981			copied = 0;
1982
1983		ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
1984				       start+len);
 
 
 
 
 
 
 
 
 
 
 
 
 
1985	}
1986	if (wc->w_target_page)
1987		flush_dcache_page(wc->w_target_page);
1988
1989	for(i = 0; i < wc->w_num_pages; i++) {
1990		tmppage = wc->w_pages[i];
1991
1992		/* This is the direct io target page. */
1993		if (tmppage == NULL)
1994			continue;
1995
1996		if (tmppage == wc->w_target_page) {
1997			from = wc->w_target_from;
1998			to = wc->w_target_to;
1999
2000			BUG_ON(from > PAGE_SIZE ||
2001			       to > PAGE_SIZE ||
2002			       to < from);
2003		} else {
2004			/*
2005			 * Pages adjacent to the target (if any) imply
2006			 * a hole-filling write in which case we want
2007			 * to flush their entire range.
2008			 */
2009			from = 0;
2010			to = PAGE_SIZE;
2011		}
2012
2013		if (page_has_buffers(tmppage)) {
2014			if (handle && ocfs2_should_order_data(inode)) {
2015				loff_t start_byte =
2016					((loff_t)tmppage->index << PAGE_SHIFT) +
2017					from;
2018				loff_t length = to - from;
2019				ocfs2_jbd2_inode_add_write(handle, inode,
2020							   start_byte, length);
2021			}
2022			block_commit_write(tmppage, from, to);
2023		}
2024	}
2025
2026out_write_size:
2027	/* Direct io do not update i_size here. */
2028	if (wc->w_type != OCFS2_WRITE_DIRECT) {
2029		pos += copied;
2030		if (pos > i_size_read(inode)) {
2031			i_size_write(inode, pos);
2032			mark_inode_dirty(inode);
2033		}
2034		inode->i_blocks = ocfs2_inode_sector_count(inode);
2035		di->i_size = cpu_to_le64((u64)i_size_read(inode));
2036		inode->i_mtime = inode->i_ctime = current_time(inode);
2037		di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
2038		di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
2039		if (handle)
2040			ocfs2_update_inode_fsync_trans(handle, inode, 1);
2041	}
2042	if (handle)
2043		ocfs2_journal_dirty(handle, wc->w_di_bh);
2044
2045out:
2046	/* unlock pages before dealloc since it needs acquiring j_trans_barrier
2047	 * lock, or it will cause a deadlock since journal commit threads holds
2048	 * this lock and will ask for the page lock when flushing the data.
2049	 * put it here to preserve the unlock order.
2050	 */
2051	ocfs2_unlock_pages(wc);
2052
2053	if (handle)
2054		ocfs2_commit_trans(osb, handle);
2055
2056	ocfs2_run_deallocs(osb, &wc->w_dealloc);
2057
2058	brelse(wc->w_di_bh);
2059	kfree(wc);
2060
2061	return copied;
2062}
2063
2064static int ocfs2_write_end(struct file *file, struct address_space *mapping,
2065			   loff_t pos, unsigned len, unsigned copied,
2066			   struct page *page, void *fsdata)
2067{
2068	int ret;
2069	struct inode *inode = mapping->host;
2070
2071	ret = ocfs2_write_end_nolock(mapping, pos, len, copied, fsdata);
2072
2073	up_write(&OCFS2_I(inode)->ip_alloc_sem);
2074	ocfs2_inode_unlock(inode, 1);
2075
2076	return ret;
2077}
2078
2079struct ocfs2_dio_write_ctxt {
2080	struct list_head	dw_zero_list;
2081	unsigned		dw_zero_count;
2082	int			dw_orphaned;
2083	pid_t			dw_writer_pid;
2084};
2085
2086static struct ocfs2_dio_write_ctxt *
2087ocfs2_dio_alloc_write_ctx(struct buffer_head *bh, int *alloc)
2088{
2089	struct ocfs2_dio_write_ctxt *dwc = NULL;
2090
2091	if (bh->b_private)
2092		return bh->b_private;
2093
2094	dwc = kmalloc(sizeof(struct ocfs2_dio_write_ctxt), GFP_NOFS);
2095	if (dwc == NULL)
2096		return NULL;
2097	INIT_LIST_HEAD(&dwc->dw_zero_list);
2098	dwc->dw_zero_count = 0;
2099	dwc->dw_orphaned = 0;
2100	dwc->dw_writer_pid = task_pid_nr(current);
2101	bh->b_private = dwc;
2102	*alloc = 1;
2103
2104	return dwc;
2105}
2106
2107static void ocfs2_dio_free_write_ctx(struct inode *inode,
2108				     struct ocfs2_dio_write_ctxt *dwc)
2109{
2110	ocfs2_free_unwritten_list(inode, &dwc->dw_zero_list);
2111	kfree(dwc);
2112}
2113
2114/*
2115 * TODO: Make this into a generic get_blocks function.
2116 *
2117 * From do_direct_io in direct-io.c:
2118 *  "So what we do is to permit the ->get_blocks function to populate
2119 *   bh.b_size with the size of IO which is permitted at this offset and
2120 *   this i_blkbits."
2121 *
2122 * This function is called directly from get_more_blocks in direct-io.c.
2123 *
2124 * called like this: dio->get_blocks(dio->inode, fs_startblk,
2125 * 					fs_count, map_bh, dio->rw == WRITE);
2126 */
2127static int ocfs2_dio_wr_get_block(struct inode *inode, sector_t iblock,
2128			       struct buffer_head *bh_result, int create)
2129{
2130	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2131	struct ocfs2_inode_info *oi = OCFS2_I(inode);
2132	struct ocfs2_write_ctxt *wc;
2133	struct ocfs2_write_cluster_desc *desc = NULL;
2134	struct ocfs2_dio_write_ctxt *dwc = NULL;
2135	struct buffer_head *di_bh = NULL;
2136	u64 p_blkno;
2137	unsigned int i_blkbits = inode->i_sb->s_blocksize_bits;
2138	loff_t pos = iblock << i_blkbits;
2139	sector_t endblk = (i_size_read(inode) - 1) >> i_blkbits;
2140	unsigned len, total_len = bh_result->b_size;
2141	int ret = 0, first_get_block = 0;
2142
2143	len = osb->s_clustersize - (pos & (osb->s_clustersize - 1));
2144	len = min(total_len, len);
2145
2146	/*
2147	 * bh_result->b_size is count in get_more_blocks according to write
2148	 * "pos" and "end", we need map twice to return different buffer state:
2149	 * 1. area in file size, not set NEW;
2150	 * 2. area out file size, set  NEW.
2151	 *
2152	 *		   iblock    endblk
2153	 * |--------|---------|---------|---------
2154	 * |<-------area in file------->|
2155	 */
2156
2157	if ((iblock <= endblk) &&
2158	    ((iblock + ((len - 1) >> i_blkbits)) > endblk))
2159		len = (endblk - iblock + 1) << i_blkbits;
2160
2161	mlog(0, "get block of %lu at %llu:%u req %u\n",
2162			inode->i_ino, pos, len, total_len);
2163
2164	/*
2165	 * Because we need to change file size in ocfs2_dio_end_io_write(), or
2166	 * we may need to add it to orphan dir. So can not fall to fast path
2167	 * while file size will be changed.
2168	 */
2169	if (pos + total_len <= i_size_read(inode)) {
2170
2171		/* This is the fast path for re-write. */
2172		ret = ocfs2_lock_get_block(inode, iblock, bh_result, create);
2173		if (buffer_mapped(bh_result) &&
2174		    !buffer_new(bh_result) &&
2175		    ret == 0)
2176			goto out;
2177
2178		/* Clear state set by ocfs2_get_block. */
2179		bh_result->b_state = 0;
2180	}
2181
2182	dwc = ocfs2_dio_alloc_write_ctx(bh_result, &first_get_block);
2183	if (unlikely(dwc == NULL)) {
2184		ret = -ENOMEM;
2185		mlog_errno(ret);
2186		goto out;
2187	}
2188
2189	if (ocfs2_clusters_for_bytes(inode->i_sb, pos + total_len) >
2190	    ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode)) &&
2191	    !dwc->dw_orphaned) {
2192		/*
2193		 * when we are going to alloc extents beyond file size, add the
2194		 * inode to orphan dir, so we can recall those spaces when
2195		 * system crashed during write.
2196		 */
2197		ret = ocfs2_add_inode_to_orphan(osb, inode);
2198		if (ret < 0) {
2199			mlog_errno(ret);
2200			goto out;
2201		}
2202		dwc->dw_orphaned = 1;
2203	}
2204
2205	ret = ocfs2_inode_lock(inode, &di_bh, 1);
2206	if (ret) {
2207		mlog_errno(ret);
2208		goto out;
2209	}
2210
2211	down_write(&oi->ip_alloc_sem);
2212
2213	if (first_get_block) {
2214		if (ocfs2_sparse_alloc(osb))
2215			ret = ocfs2_zero_tail(inode, di_bh, pos);
2216		else
2217			ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos,
2218							   total_len, NULL);
2219		if (ret < 0) {
2220			mlog_errno(ret);
2221			goto unlock;
2222		}
2223	}
2224
2225	ret = ocfs2_write_begin_nolock(inode->i_mapping, pos, len,
2226				       OCFS2_WRITE_DIRECT, NULL,
2227				       (void **)&wc, di_bh, NULL);
2228	if (ret) {
2229		mlog_errno(ret);
2230		goto unlock;
2231	}
2232
2233	desc = &wc->w_desc[0];
2234
2235	p_blkno = ocfs2_clusters_to_blocks(inode->i_sb, desc->c_phys);
2236	BUG_ON(p_blkno == 0);
2237	p_blkno += iblock & (u64)(ocfs2_clusters_to_blocks(inode->i_sb, 1) - 1);
2238
2239	map_bh(bh_result, inode->i_sb, p_blkno);
2240	bh_result->b_size = len;
2241	if (desc->c_needs_zero)
2242		set_buffer_new(bh_result);
2243
2244	if (iblock > endblk)
2245		set_buffer_new(bh_result);
2246
2247	/* May sleep in end_io. It should not happen in a irq context. So defer
2248	 * it to dio work queue. */
2249	set_buffer_defer_completion(bh_result);
2250
2251	if (!list_empty(&wc->w_unwritten_list)) {
2252		struct ocfs2_unwritten_extent *ue = NULL;
2253
2254		ue = list_first_entry(&wc->w_unwritten_list,
2255				      struct ocfs2_unwritten_extent,
2256				      ue_node);
2257		BUG_ON(ue->ue_cpos != desc->c_cpos);
2258		/* The physical address may be 0, fill it. */
2259		ue->ue_phys = desc->c_phys;
2260
2261		list_splice_tail_init(&wc->w_unwritten_list, &dwc->dw_zero_list);
2262		dwc->dw_zero_count += wc->w_unwritten_count;
2263	}
2264
2265	ret = ocfs2_write_end_nolock(inode->i_mapping, pos, len, len, wc);
2266	BUG_ON(ret != len);
2267	ret = 0;
2268unlock:
2269	up_write(&oi->ip_alloc_sem);
2270	ocfs2_inode_unlock(inode, 1);
2271	brelse(di_bh);
2272out:
2273	if (ret < 0)
2274		ret = -EIO;
2275	return ret;
2276}
2277
2278static int ocfs2_dio_end_io_write(struct inode *inode,
2279				  struct ocfs2_dio_write_ctxt *dwc,
2280				  loff_t offset,
2281				  ssize_t bytes)
2282{
2283	struct ocfs2_cached_dealloc_ctxt dealloc;
2284	struct ocfs2_extent_tree et;
2285	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2286	struct ocfs2_inode_info *oi = OCFS2_I(inode);
2287	struct ocfs2_unwritten_extent *ue = NULL;
2288	struct buffer_head *di_bh = NULL;
2289	struct ocfs2_dinode *di;
2290	struct ocfs2_alloc_context *data_ac = NULL;
2291	struct ocfs2_alloc_context *meta_ac = NULL;
2292	handle_t *handle = NULL;
2293	loff_t end = offset + bytes;
2294	int ret = 0, credits = 0;
2295
2296	ocfs2_init_dealloc_ctxt(&dealloc);
2297
2298	/* We do clear unwritten, delete orphan, change i_size here. If neither
2299	 * of these happen, we can skip all this. */
2300	if (list_empty(&dwc->dw_zero_list) &&
2301	    end <= i_size_read(inode) &&
2302	    !dwc->dw_orphaned)
2303		goto out;
2304
2305	ret = ocfs2_inode_lock(inode, &di_bh, 1);
2306	if (ret < 0) {
2307		mlog_errno(ret);
2308		goto out;
2309	}
2310
2311	down_write(&oi->ip_alloc_sem);
2312
2313	/* Delete orphan before acquire i_rwsem. */
2314	if (dwc->dw_orphaned) {
2315		BUG_ON(dwc->dw_writer_pid != task_pid_nr(current));
2316
2317		end = end > i_size_read(inode) ? end : 0;
2318
2319		ret = ocfs2_del_inode_from_orphan(osb, inode, di_bh,
2320				!!end, end);
2321		if (ret < 0)
2322			mlog_errno(ret);
2323	}
2324
2325	di = (struct ocfs2_dinode *)di_bh->b_data;
2326
2327	ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
2328
2329	/* Attach dealloc with extent tree in case that we may reuse extents
2330	 * which are already unlinked from current extent tree due to extent
2331	 * rotation and merging.
2332	 */
2333	et.et_dealloc = &dealloc;
2334
2335	ret = ocfs2_lock_allocators(inode, &et, 0, dwc->dw_zero_count*2,
2336				    &data_ac, &meta_ac);
2337	if (ret) {
2338		mlog_errno(ret);
2339		goto unlock;
2340	}
2341
2342	credits = ocfs2_calc_extend_credits(inode->i_sb, &di->id2.i_list);
2343
2344	handle = ocfs2_start_trans(osb, credits);
2345	if (IS_ERR(handle)) {
2346		ret = PTR_ERR(handle);
2347		mlog_errno(ret);
2348		goto unlock;
2349	}
2350	ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
2351				      OCFS2_JOURNAL_ACCESS_WRITE);
2352	if (ret) {
2353		mlog_errno(ret);
2354		goto commit;
2355	}
2356
2357	list_for_each_entry(ue, &dwc->dw_zero_list, ue_node) {
2358		ret = ocfs2_mark_extent_written(inode, &et, handle,
2359						ue->ue_cpos, 1,
2360						ue->ue_phys,
2361						meta_ac, &dealloc);
2362		if (ret < 0) {
2363			mlog_errno(ret);
2364			break;
2365		}
2366	}
2367
2368	if (end > i_size_read(inode)) {
2369		ret = ocfs2_set_inode_size(handle, inode, di_bh, end);
2370		if (ret < 0)
2371			mlog_errno(ret);
2372	}
2373commit:
2374	ocfs2_commit_trans(osb, handle);
2375unlock:
2376	up_write(&oi->ip_alloc_sem);
2377	ocfs2_inode_unlock(inode, 1);
2378	brelse(di_bh);
2379out:
2380	if (data_ac)
2381		ocfs2_free_alloc_context(data_ac);
2382	if (meta_ac)
2383		ocfs2_free_alloc_context(meta_ac);
2384	ocfs2_run_deallocs(osb, &dealloc);
2385	ocfs2_dio_free_write_ctx(inode, dwc);
2386
2387	return ret;
2388}
2389
2390/*
2391 * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
2392 * particularly interested in the aio/dio case.  We use the rw_lock DLM lock
2393 * to protect io on one node from truncation on another.
2394 */
2395static int ocfs2_dio_end_io(struct kiocb *iocb,
2396			    loff_t offset,
2397			    ssize_t bytes,
2398			    void *private)
2399{
2400	struct inode *inode = file_inode(iocb->ki_filp);
2401	int level;
2402	int ret = 0;
2403
2404	/* this io's submitter should not have unlocked this before we could */
2405	BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
2406
2407	if (bytes <= 0)
2408		mlog_ratelimited(ML_ERROR, "Direct IO failed, bytes = %lld",
2409				 (long long)bytes);
2410	if (private) {
2411		if (bytes > 0)
2412			ret = ocfs2_dio_end_io_write(inode, private, offset,
2413						     bytes);
2414		else
2415			ocfs2_dio_free_write_ctx(inode, private);
2416	}
2417
2418	ocfs2_iocb_clear_rw_locked(iocb);
2419
2420	level = ocfs2_iocb_rw_locked_level(iocb);
2421	ocfs2_rw_unlock(inode, level);
2422	return ret;
2423}
2424
2425static ssize_t ocfs2_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2426{
2427	struct file *file = iocb->ki_filp;
2428	struct inode *inode = file->f_mapping->host;
2429	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2430	get_block_t *get_block;
2431
2432	/*
2433	 * Fallback to buffered I/O if we see an inode without
2434	 * extents.
2435	 */
2436	if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2437		return 0;
2438
2439	/* Fallback to buffered I/O if we do not support append dio. */
2440	if (iocb->ki_pos + iter->count > i_size_read(inode) &&
2441	    !ocfs2_supports_append_dio(osb))
2442		return 0;
2443
2444	if (iov_iter_rw(iter) == READ)
2445		get_block = ocfs2_lock_get_block;
2446	else
2447		get_block = ocfs2_dio_wr_get_block;
2448
2449	return __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev,
2450				    iter, get_block,
2451				    ocfs2_dio_end_io, NULL, 0);
2452}
2453
2454const struct address_space_operations ocfs2_aops = {
2455	.dirty_folio		= block_dirty_folio,
2456	.read_folio		= ocfs2_read_folio,
2457	.readahead		= ocfs2_readahead,
2458	.writepage		= ocfs2_writepage,
2459	.write_begin		= ocfs2_write_begin,
2460	.write_end		= ocfs2_write_end,
2461	.bmap			= ocfs2_bmap,
2462	.direct_IO		= ocfs2_direct_IO,
2463	.invalidate_folio	= block_invalidate_folio,
2464	.release_folio		= ocfs2_release_folio,
2465	.migrate_folio		= buffer_migrate_folio,
2466	.is_partially_uptodate	= block_is_partially_uptodate,
2467	.error_remove_page	= generic_error_remove_page,
2468};