1// SPDX-License-Identifier: GPL-2.0-or-later
   2/* -*- mode: c; c-basic-offset: 8; -*-
   3 * vim: noexpandtab sw=8 ts=8 sts=0:
   4 *
   5 * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
   6 */
   7
   8#include <linux/fs.h>
   9#include <linux/slab.h>
  10#include <linux/highmem.h>
  11#include <linux/pagemap.h>
  12#include <asm/byteorder.h>
  13#include <linux/swap.h>
  14#include <linux/pipe_fs_i.h>
  15#include <linux/mpage.h>
  16#include <linux/quotaops.h>
  17#include <linux/blkdev.h>
  18#include <linux/uio.h>
  19#include <linux/mm.h>
  20
  21#include <cluster/masklog.h>
  22
  23#include "ocfs2.h"
  24
  25#include "alloc.h"
  26#include "aops.h"
  27#include "dlmglue.h"
  28#include "extent_map.h"
  29#include "file.h"
  30#include "inode.h"
  31#include "journal.h"
  32#include "suballoc.h"
  33#include "super.h"
  34#include "symlink.h"
  35#include "refcounttree.h"
  36#include "ocfs2_trace.h"
  37
  38#include "buffer_head_io.h"
  39#include "dir.h"
  40#include "namei.h"
  41#include "sysfile.h"
  42
  43static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
  44				   struct buffer_head *bh_result, int create)
  45{
  46	int err = -EIO;
  47	int status;
  48	struct ocfs2_dinode *fe = NULL;
  49	struct buffer_head *bh = NULL;
  50	struct buffer_head *buffer_cache_bh = NULL;
  51	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  52	void *kaddr;
  53
  54	trace_ocfs2_symlink_get_block(
  55			(unsigned long long)OCFS2_I(inode)->ip_blkno,
  56			(unsigned long long)iblock, bh_result, create);
  57
  58	BUG_ON(ocfs2_inode_is_fast_symlink(inode));
  59
  60	if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
  61		mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
  62		     (unsigned long long)iblock);
  63		goto bail;
  64	}
  65
  66	status = ocfs2_read_inode_block(inode, &bh);
  67	if (status < 0) {
  68		mlog_errno(status);
  69		goto bail;
  70	}
  71	fe = (struct ocfs2_dinode *) bh->b_data;
  72
  73	if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
  74						    le32_to_cpu(fe->i_clusters))) {
  75		err = -ENOMEM;
  76		mlog(ML_ERROR, "block offset is outside the allocated size: "
  77		     "%llu\n", (unsigned long long)iblock);
  78		goto bail;
  79	}
  80
  81	/* We don't use the page cache to create symlink data, so if
  82	 * need be, copy it over from the buffer cache. */
  83	if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
  84		u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
  85			    iblock;
  86		buffer_cache_bh = sb_getblk(osb->sb, blkno);
  87		if (!buffer_cache_bh) {
  88			err = -ENOMEM;
  89			mlog(ML_ERROR, "couldn't getblock for symlink!\n");
  90			goto bail;
  91		}
  92
  93		/* we haven't locked out transactions, so a commit
  94		 * could've happened. Since we've got a reference on
  95		 * the bh, even if it commits while we're doing the
  96		 * copy, the data is still good. */
  97		if (buffer_jbd(buffer_cache_bh)
  98		    && ocfs2_inode_is_new(inode)) {
  99			kaddr = kmap_atomic(bh_result->b_page);
 100			if (!kaddr) {
 101				mlog(ML_ERROR, "couldn't kmap!\n");
 102				goto bail;
 103			}
 104			memcpy(kaddr + (bh_result->b_size * iblock),
 105			       buffer_cache_bh->b_data,
 106			       bh_result->b_size);
 107			kunmap_atomic(kaddr);
 108			set_buffer_uptodate(bh_result);
 109		}
 110		brelse(buffer_cache_bh);
 111	}
 112
 113	map_bh(bh_result, inode->i_sb,
 114	       le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
 115
 116	err = 0;
 117
 118bail:
 119	brelse(bh);
 120
 121	return err;
 122}
 123
 124static int ocfs2_lock_get_block(struct inode *inode, sector_t iblock,
 125		    struct buffer_head *bh_result, int create)
 126{
 127	int ret = 0;
 128	struct ocfs2_inode_info *oi = OCFS2_I(inode);
 129
 130	down_read(&oi->ip_alloc_sem);
 131	ret = ocfs2_get_block(inode, iblock, bh_result, create);
 132	up_read(&oi->ip_alloc_sem);
 133
 134	return ret;
 135}
 136
 137int ocfs2_get_block(struct inode *inode, sector_t iblock,
 138		    struct buffer_head *bh_result, int create)
 139{
 140	int err = 0;
 141	unsigned int ext_flags;
 142	u64 max_blocks = bh_result->b_size >> inode->i_blkbits;
 143	u64 p_blkno, count, past_eof;
 144	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 145
 146	trace_ocfs2_get_block((unsigned long long)OCFS2_I(inode)->ip_blkno,
 147			      (unsigned long long)iblock, bh_result, create);
 148
 149	if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
 150		mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
 151		     inode, inode->i_ino);
 152
 153	if (S_ISLNK(inode->i_mode)) {
 154		/* this always does I/O for some reason. */
 155		err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
 156		goto bail;
 157	}
 158
 159	err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, &count,
 160					  &ext_flags);
 161	if (err) {
 162		mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
 163		     "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
 164		     (unsigned long long)p_blkno);
 165		goto bail;
 166	}
 167
 168	if (max_blocks < count)
 169		count = max_blocks;
 170
 171	/*
 172	 * ocfs2 never allocates in this function - the only time we
 173	 * need to use BH_New is when we're extending i_size on a file
 174	 * system which doesn't support holes, in which case BH_New
 175	 * allows __block_write_begin() to zero.
 176	 *
 177	 * If we see this on a sparse file system, then a truncate has
 178	 * raced us and removed the cluster. In this case, we clear
 179	 * the buffers dirty and uptodate bits and let the buffer code
 180	 * ignore it as a hole.
 181	 */
 182	if (create && p_blkno == 0 && ocfs2_sparse_alloc(osb)) {
 183		clear_buffer_dirty(bh_result);
 184		clear_buffer_uptodate(bh_result);
 185		goto bail;
 186	}
 187
 188	/* Treat the unwritten extent as a hole for zeroing purposes. */
 189	if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 190		map_bh(bh_result, inode->i_sb, p_blkno);
 191
 192	bh_result->b_size = count << inode->i_blkbits;
 193
 194	if (!ocfs2_sparse_alloc(osb)) {
 195		if (p_blkno == 0) {
 196			err = -EIO;
 197			mlog(ML_ERROR,
 198			     "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
 199			     (unsigned long long)iblock,
 200			     (unsigned long long)p_blkno,
 201			     (unsigned long long)OCFS2_I(inode)->ip_blkno);
 202			mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
 203			dump_stack();
 204			goto bail;
 205		}
 206	}
 207
 208	past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 209
 210	trace_ocfs2_get_block_end((unsigned long long)OCFS2_I(inode)->ip_blkno,
 211				  (unsigned long long)past_eof);
 212	if (create && (iblock >= past_eof))
 213		set_buffer_new(bh_result);
 214
 215bail:
 216	if (err < 0)
 217		err = -EIO;
 218
 219	return err;
 220}
 221
 222int ocfs2_read_inline_data(struct inode *inode, struct page *page,
 223			   struct buffer_head *di_bh)
 224{
 225	void *kaddr;
 226	loff_t size;
 227	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 228
 229	if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
 230		ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag\n",
 231			    (unsigned long long)OCFS2_I(inode)->ip_blkno);
 232		return -EROFS;
 233	}
 234
 235	size = i_size_read(inode);
 236
 237	if (size > PAGE_SIZE ||
 238	    size > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) {
 239		ocfs2_error(inode->i_sb,
 240			    "Inode %llu has with inline data has bad size: %Lu\n",
 241			    (unsigned long long)OCFS2_I(inode)->ip_blkno,
 242			    (unsigned long long)size);
 243		return -EROFS;
 244	}
 245
 246	kaddr = kmap_atomic(page);
 247	if (size)
 248		memcpy(kaddr, di->id2.i_data.id_data, size);
 249	/* Clear the remaining part of the page */
 250	memset(kaddr + size, 0, PAGE_SIZE - size);
 251	flush_dcache_page(page);
 252	kunmap_atomic(kaddr);
 253
 254	SetPageUptodate(page);
 255
 256	return 0;
 257}
 258
 259static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
 260{
 261	int ret;
 262	struct buffer_head *di_bh = NULL;
 263
 264	BUG_ON(!PageLocked(page));
 265	BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
 266
 267	ret = ocfs2_read_inode_block(inode, &di_bh);
 268	if (ret) {
 269		mlog_errno(ret);
 270		goto out;
 271	}
 272
 273	ret = ocfs2_read_inline_data(inode, page, di_bh);
 274out:
 275	unlock_page(page);
 276
 277	brelse(di_bh);
 278	return ret;
 279}
 280
 281static int ocfs2_readpage(struct file *file, struct page *page)
 282{
 283	struct inode *inode = page->mapping->host;
 284	struct ocfs2_inode_info *oi = OCFS2_I(inode);
 285	loff_t start = (loff_t)page->index << PAGE_SHIFT;
 286	int ret, unlock = 1;
 287
 288	trace_ocfs2_readpage((unsigned long long)oi->ip_blkno,
 289			     (page ? page->index : 0));
 290
 291	ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page);
 292	if (ret != 0) {
 293		if (ret == AOP_TRUNCATED_PAGE)
 294			unlock = 0;
 295		mlog_errno(ret);
 296		goto out;
 297	}
 298
 299	if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
 300		/*
 301		 * Unlock the page and cycle ip_alloc_sem so that we don't
 302		 * busyloop waiting for ip_alloc_sem to unlock
 303		 */
 304		ret = AOP_TRUNCATED_PAGE;
 305		unlock_page(page);
 306		unlock = 0;
 307		down_read(&oi->ip_alloc_sem);
 308		up_read(&oi->ip_alloc_sem);
 309		goto out_inode_unlock;
 310	}
 311
 312	/*
 313	 * i_size might have just been updated as we grabed the meta lock.  We
 314	 * might now be discovering a truncate that hit on another node.
 315	 * block_read_full_page->get_block freaks out if it is asked to read
 316	 * beyond the end of a file, so we check here.  Callers
 317	 * (generic_file_read, vm_ops->fault) are clever enough to check i_size
 318	 * and notice that the page they just read isn't needed.
 319	 *
 320	 * XXX sys_readahead() seems to get that wrong?
 321	 */
 322	if (start >= i_size_read(inode)) {
 323		zero_user(page, 0, PAGE_SIZE);
 324		SetPageUptodate(page);
 325		ret = 0;
 326		goto out_alloc;
 327	}
 328
 329	if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 330		ret = ocfs2_readpage_inline(inode, page);
 331	else
 332		ret = block_read_full_page(page, ocfs2_get_block);
 333	unlock = 0;
 334
 335out_alloc:
 336	up_read(&oi->ip_alloc_sem);
 337out_inode_unlock:
 338	ocfs2_inode_unlock(inode, 0);
 339out:
 340	if (unlock)
 341		unlock_page(page);
 342	return ret;
 343}
 344
 345/*
 346 * This is used only for read-ahead. Failures or difficult to handle
 347 * situations are safe to ignore.
 348 *
 349 * Right now, we don't bother with BH_Boundary - in-inode extent lists
 350 * are quite large (243 extents on 4k blocks), so most inodes don't
 351 * grow out to a tree. If need be, detecting boundary extents could
 352 * trivially be added in a future version of ocfs2_get_block().
 353 */
 354static int ocfs2_readpages(struct file *filp, struct address_space *mapping,
 355			   struct list_head *pages, unsigned nr_pages)
 356{
 357	int ret, err = -EIO;
 358	struct inode *inode = mapping->host;
 359	struct ocfs2_inode_info *oi = OCFS2_I(inode);
 360	loff_t start;
 361	struct page *last;
 362
 363	/*
 364	 * Use the nonblocking flag for the dlm code to avoid page
 365	 * lock inversion, but don't bother with retrying.
 366	 */
 367	ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
 368	if (ret)
 369		return err;
 370
 371	if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
 372		ocfs2_inode_unlock(inode, 0);
 373		return err;
 374	}
 375
 376	/*
 377	 * Don't bother with inline-data. There isn't anything
 378	 * to read-ahead in that case anyway...
 379	 */
 380	if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 381		goto out_unlock;
 382
 383	/*
 384	 * Check whether a remote node truncated this file - we just
 385	 * drop out in that case as it's not worth handling here.
 386	 */
 387	last = lru_to_page(pages);
 388	start = (loff_t)last->index << PAGE_SHIFT;
 389	if (start >= i_size_read(inode))
 390		goto out_unlock;
 391
 392	err = mpage_readpages(mapping, pages, nr_pages, ocfs2_get_block);
 393
 394out_unlock:
 395	up_read(&oi->ip_alloc_sem);
 
 396	ocfs2_inode_unlock(inode, 0);
 397
 398	return err;
 399}
 400
 401/* Note: Because we don't support holes, our allocation has
 402 * already happened (allocation writes zeros to the file data)
 403 * so we don't have to worry about ordered writes in
 404 * ocfs2_writepage.
 405 *
 406 * ->writepage is called during the process of invalidating the page cache
 407 * during blocked lock processing.  It can't block on any cluster locks
 408 * to during block mapping.  It's relying on the fact that the block
 409 * mapping can't have disappeared under the dirty pages that it is
 410 * being asked to write back.
 411 */
 412static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
 413{
 414	trace_ocfs2_writepage(
 415		(unsigned long long)OCFS2_I(page->mapping->host)->ip_blkno,
 416		page->index);
 417
 418	return block_write_full_page(page, ocfs2_get_block, wbc);
 419}
 420
 421/* Taken from ext3. We don't necessarily need the full blown
 422 * functionality yet, but IMHO it's better to cut and paste the whole
 423 * thing so we can avoid introducing our own bugs (and easily pick up
 424 * their fixes when they happen) --Mark */
 425int walk_page_buffers(	handle_t *handle,
 426			struct buffer_head *head,
 427			unsigned from,
 428			unsigned to,
 429			int *partial,
 430			int (*fn)(	handle_t *handle,
 431					struct buffer_head *bh))
 432{
 433	struct buffer_head *bh;
 434	unsigned block_start, block_end;
 435	unsigned blocksize = head->b_size;
 436	int err, ret = 0;
 437	struct buffer_head *next;
 438
 439	for (	bh = head, block_start = 0;
 440		ret == 0 && (bh != head || !block_start);
 441	    	block_start = block_end, bh = next)
 442	{
 443		next = bh->b_this_page;
 444		block_end = block_start + blocksize;
 445		if (block_end <= from || block_start >= to) {
 446			if (partial && !buffer_uptodate(bh))
 447				*partial = 1;
 448			continue;
 449		}
 450		err = (*fn)(handle, bh);
 451		if (!ret)
 452			ret = err;
 453	}
 454	return ret;
 455}
 456
 457static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
 458{
 459	sector_t status;
 460	u64 p_blkno = 0;
 461	int err = 0;
 462	struct inode *inode = mapping->host;
 463
 464	trace_ocfs2_bmap((unsigned long long)OCFS2_I(inode)->ip_blkno,
 465			 (unsigned long long)block);
 466
 467	/*
 468	 * The swap code (ab-)uses ->bmap to get a block mapping and then
 469	 * bypasseѕ the file system for actual I/O.  We really can't allow
 470	 * that on refcounted inodes, so we have to skip out here.  And yes,
 471	 * 0 is the magic code for a bmap error..
 472	 */
 473	if (ocfs2_is_refcount_inode(inode))
 474		return 0;
 475
 476	/* We don't need to lock journal system files, since they aren't
 477	 * accessed concurrently from multiple nodes.
 478	 */
 479	if (!INODE_JOURNAL(inode)) {
 480		err = ocfs2_inode_lock(inode, NULL, 0);
 481		if (err) {
 482			if (err != -ENOENT)
 483				mlog_errno(err);
 484			goto bail;
 485		}
 486		down_read(&OCFS2_I(inode)->ip_alloc_sem);
 487	}
 488
 489	if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
 490		err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
 491						  NULL);
 492
 493	if (!INODE_JOURNAL(inode)) {
 494		up_read(&OCFS2_I(inode)->ip_alloc_sem);
 495		ocfs2_inode_unlock(inode, 0);
 496	}
 497
 498	if (err) {
 499		mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
 500		     (unsigned long long)block);
 501		mlog_errno(err);
 502		goto bail;
 503	}
 504
 505bail:
 506	status = err ? 0 : p_blkno;
 507
 508	return status;
 509}
 510
 511static int ocfs2_releasepage(struct page *page, gfp_t wait)
 512{
 513	if (!page_has_buffers(page))
 514		return 0;
 515	return try_to_free_buffers(page);
 516}
 517
 518static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
 519					    u32 cpos,
 520					    unsigned int *start,
 521					    unsigned int *end)
 522{
 523	unsigned int cluster_start = 0, cluster_end = PAGE_SIZE;
 524
 525	if (unlikely(PAGE_SHIFT > osb->s_clustersize_bits)) {
 526		unsigned int cpp;
 527
 528		cpp = 1 << (PAGE_SHIFT - osb->s_clustersize_bits);
 529
 530		cluster_start = cpos % cpp;
 531		cluster_start = cluster_start << osb->s_clustersize_bits;
 532
 533		cluster_end = cluster_start + osb->s_clustersize;
 534	}
 535
 536	BUG_ON(cluster_start > PAGE_SIZE);
 537	BUG_ON(cluster_end > PAGE_SIZE);
 538
 539	if (start)
 540		*start = cluster_start;
 541	if (end)
 542		*end = cluster_end;
 543}
 544
 545/*
 546 * 'from' and 'to' are the region in the page to avoid zeroing.
 547 *
 548 * If pagesize > clustersize, this function will avoid zeroing outside
 549 * of the cluster boundary.
 550 *
 551 * from == to == 0 is code for "zero the entire cluster region"
 552 */
 553static void ocfs2_clear_page_regions(struct page *page,
 554				     struct ocfs2_super *osb, u32 cpos,
 555				     unsigned from, unsigned to)
 556{
 557	void *kaddr;
 558	unsigned int cluster_start, cluster_end;
 559
 560	ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
 561
 562	kaddr = kmap_atomic(page);
 563
 564	if (from || to) {
 565		if (from > cluster_start)
 566			memset(kaddr + cluster_start, 0, from - cluster_start);
 567		if (to < cluster_end)
 568			memset(kaddr + to, 0, cluster_end - to);
 569	} else {
 570		memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
 571	}
 572
 573	kunmap_atomic(kaddr);
 574}
 575
 576/*
 577 * Nonsparse file systems fully allocate before we get to the write
 578 * code. This prevents ocfs2_write() from tagging the write as an
 579 * allocating one, which means ocfs2_map_page_blocks() might try to
 580 * read-in the blocks at the tail of our file. Avoid reading them by
 581 * testing i_size against each block offset.
 582 */
 583static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
 584				 unsigned int block_start)
 585{
 586	u64 offset = page_offset(page) + block_start;
 587
 588	if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 589		return 1;
 590
 591	if (i_size_read(inode) > offset)
 592		return 1;
 593
 594	return 0;
 595}
 596
 597/*
 598 * Some of this taken from __block_write_begin(). We already have our
 599 * mapping by now though, and the entire write will be allocating or
 600 * it won't, so not much need to use BH_New.
 601 *
 602 * This will also skip zeroing, which is handled externally.
 603 */
 604int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
 605			  struct inode *inode, unsigned int from,
 606			  unsigned int to, int new)
 607{
 608	int ret = 0;
 609	struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
 610	unsigned int block_end, block_start;
 611	unsigned int bsize = i_blocksize(inode);
 612
 613	if (!page_has_buffers(page))
 614		create_empty_buffers(page, bsize, 0);
 615
 616	head = page_buffers(page);
 617	for (bh = head, block_start = 0; bh != head || !block_start;
 618	     bh = bh->b_this_page, block_start += bsize) {
 619		block_end = block_start + bsize;
 620
 621		clear_buffer_new(bh);
 622
 623		/*
 624		 * Ignore blocks outside of our i/o range -
 625		 * they may belong to unallocated clusters.
 626		 */
 627		if (block_start >= to || block_end <= from) {
 628			if (PageUptodate(page))
 629				set_buffer_uptodate(bh);
 630			continue;
 631		}
 632
 633		/*
 634		 * For an allocating write with cluster size >= page
 635		 * size, we always write the entire page.
 636		 */
 637		if (new)
 638			set_buffer_new(bh);
 639
 640		if (!buffer_mapped(bh)) {
 641			map_bh(bh, inode->i_sb, *p_blkno);
 642			clean_bdev_bh_alias(bh);
 643		}
 644
 645		if (PageUptodate(page)) {
 646			if (!buffer_uptodate(bh))
 647				set_buffer_uptodate(bh);
 648		} else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
 649			   !buffer_new(bh) &&
 650			   ocfs2_should_read_blk(inode, page, block_start) &&
 651			   (block_start < from || block_end > to)) {
 652			ll_rw_block(REQ_OP_READ, 0, 1, &bh);
 653			*wait_bh++=bh;
 654		}
 655
 656		*p_blkno = *p_blkno + 1;
 657	}
 658
 659	/*
 660	 * If we issued read requests - let them complete.
 661	 */
 662	while(wait_bh > wait) {
 663		wait_on_buffer(*--wait_bh);
 664		if (!buffer_uptodate(*wait_bh))
 665			ret = -EIO;
 666	}
 667
 668	if (ret == 0 || !new)
 669		return ret;
 670
 671	/*
 672	 * If we get -EIO above, zero out any newly allocated blocks
 673	 * to avoid exposing stale data.
 674	 */
 675	bh = head;
 676	block_start = 0;
 677	do {
 678		block_end = block_start + bsize;
 679		if (block_end <= from)
 680			goto next_bh;
 681		if (block_start >= to)
 682			break;
 683
 684		zero_user(page, block_start, bh->b_size);
 685		set_buffer_uptodate(bh);
 686		mark_buffer_dirty(bh);
 687
 688next_bh:
 689		block_start = block_end;
 690		bh = bh->b_this_page;
 691	} while (bh != head);
 692
 693	return ret;
 694}
 695
 696#if (PAGE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
 697#define OCFS2_MAX_CTXT_PAGES	1
 698#else
 699#define OCFS2_MAX_CTXT_PAGES	(OCFS2_MAX_CLUSTERSIZE / PAGE_SIZE)
 700#endif
 701
 702#define OCFS2_MAX_CLUSTERS_PER_PAGE	(PAGE_SIZE / OCFS2_MIN_CLUSTERSIZE)
 703
 704struct ocfs2_unwritten_extent {
 705	struct list_head	ue_node;
 706	struct list_head	ue_ip_node;
 707	u32			ue_cpos;
 708	u32			ue_phys;
 709};
 710
 711/*
 712 * Describe the state of a single cluster to be written to.
 713 */
 714struct ocfs2_write_cluster_desc {
 715	u32		c_cpos;
 716	u32		c_phys;
 717	/*
 718	 * Give this a unique field because c_phys eventually gets
 719	 * filled.
 720	 */
 721	unsigned	c_new;
 722	unsigned	c_clear_unwritten;
 723	unsigned	c_needs_zero;
 724};
 725
 726struct ocfs2_write_ctxt {
 727	/* Logical cluster position / len of write */
 728	u32				w_cpos;
 729	u32				w_clen;
 730
 731	/* First cluster allocated in a nonsparse extend */
 732	u32				w_first_new_cpos;
 733
 734	/* Type of caller. Must be one of buffer, mmap, direct.  */
 735	ocfs2_write_type_t		w_type;
 736
 737	struct ocfs2_write_cluster_desc	w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
 738
 739	/*
 740	 * This is true if page_size > cluster_size.
 741	 *
 742	 * It triggers a set of special cases during write which might
 743	 * have to deal with allocating writes to partial pages.
 744	 */
 745	unsigned int			w_large_pages;
 746
 747	/*
 748	 * Pages involved in this write.
 749	 *
 750	 * w_target_page is the page being written to by the user.
 751	 *
 752	 * w_pages is an array of pages which always contains
 753	 * w_target_page, and in the case of an allocating write with
 754	 * page_size < cluster size, it will contain zero'd and mapped
 755	 * pages adjacent to w_target_page which need to be written
 756	 * out in so that future reads from that region will get
 757	 * zero's.
 758	 */
 759	unsigned int			w_num_pages;
 760	struct page			*w_pages[OCFS2_MAX_CTXT_PAGES];
 761	struct page			*w_target_page;
 762
 763	/*
 764	 * w_target_locked is used for page_mkwrite path indicating no unlocking
 765	 * against w_target_page in ocfs2_write_end_nolock.
 766	 */
 767	unsigned int			w_target_locked:1;
 768
 769	/*
 770	 * ocfs2_write_end() uses this to know what the real range to
 771	 * write in the target should be.
 772	 */
 773	unsigned int			w_target_from;
 774	unsigned int			w_target_to;
 775
 776	/*
 777	 * We could use journal_current_handle() but this is cleaner,
 778	 * IMHO -Mark
 779	 */
 780	handle_t			*w_handle;
 781
 782	struct buffer_head		*w_di_bh;
 783
 784	struct ocfs2_cached_dealloc_ctxt w_dealloc;
 785
 786	struct list_head		w_unwritten_list;
 787	unsigned int			w_unwritten_count;
 788};
 789
 790void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
 791{
 792	int i;
 793
 794	for(i = 0; i < num_pages; i++) {
 795		if (pages[i]) {
 796			unlock_page(pages[i]);
 797			mark_page_accessed(pages[i]);
 798			put_page(pages[i]);
 799		}
 800	}
 801}
 802
 803static void ocfs2_unlock_pages(struct ocfs2_write_ctxt *wc)
 804{
 805	int i;
 806
 807	/*
 808	 * w_target_locked is only set to true in the page_mkwrite() case.
 809	 * The intent is to allow us to lock the target page from write_begin()
 810	 * to write_end(). The caller must hold a ref on w_target_page.
 811	 */
 812	if (wc->w_target_locked) {
 813		BUG_ON(!wc->w_target_page);
 814		for (i = 0; i < wc->w_num_pages; i++) {
 815			if (wc->w_target_page == wc->w_pages[i]) {
 816				wc->w_pages[i] = NULL;
 817				break;
 818			}
 819		}
 820		mark_page_accessed(wc->w_target_page);
 821		put_page(wc->w_target_page);
 822	}
 823	ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
 824}
 825
 826static void ocfs2_free_unwritten_list(struct inode *inode,
 827				 struct list_head *head)
 828{
 829	struct ocfs2_inode_info *oi = OCFS2_I(inode);
 830	struct ocfs2_unwritten_extent *ue = NULL, *tmp = NULL;
 831
 832	list_for_each_entry_safe(ue, tmp, head, ue_node) {
 833		list_del(&ue->ue_node);
 834		spin_lock(&oi->ip_lock);
 835		list_del(&ue->ue_ip_node);
 836		spin_unlock(&oi->ip_lock);
 837		kfree(ue);
 838	}
 839}
 840
 841static void ocfs2_free_write_ctxt(struct inode *inode,
 842				  struct ocfs2_write_ctxt *wc)
 843{
 844	ocfs2_free_unwritten_list(inode, &wc->w_unwritten_list);
 845	ocfs2_unlock_pages(wc);
 846	brelse(wc->w_di_bh);
 847	kfree(wc);
 848}
 849
 850static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
 851				  struct ocfs2_super *osb, loff_t pos,
 852				  unsigned len, ocfs2_write_type_t type,
 853				  struct buffer_head *di_bh)
 854{
 855	u32 cend;
 856	struct ocfs2_write_ctxt *wc;
 857
 858	wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
 859	if (!wc)
 860		return -ENOMEM;
 861
 862	wc->w_cpos = pos >> osb->s_clustersize_bits;
 863	wc->w_first_new_cpos = UINT_MAX;
 864	cend = (pos + len - 1) >> osb->s_clustersize_bits;
 865	wc->w_clen = cend - wc->w_cpos + 1;
 866	get_bh(di_bh);
 867	wc->w_di_bh = di_bh;
 868	wc->w_type = type;
 869
 870	if (unlikely(PAGE_SHIFT > osb->s_clustersize_bits))
 871		wc->w_large_pages = 1;
 872	else
 873		wc->w_large_pages = 0;
 874
 875	ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
 876	INIT_LIST_HEAD(&wc->w_unwritten_list);
 877
 878	*wcp = wc;
 879
 880	return 0;
 881}
 882
 883/*
 884 * If a page has any new buffers, zero them out here, and mark them uptodate
 885 * and dirty so they'll be written out (in order to prevent uninitialised
 886 * block data from leaking). And clear the new bit.
 887 */
 888static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
 889{
 890	unsigned int block_start, block_end;
 891	struct buffer_head *head, *bh;
 892
 893	BUG_ON(!PageLocked(page));
 894	if (!page_has_buffers(page))
 895		return;
 896
 897	bh = head = page_buffers(page);
 898	block_start = 0;
 899	do {
 900		block_end = block_start + bh->b_size;
 901
 902		if (buffer_new(bh)) {
 903			if (block_end > from && block_start < to) {
 904				if (!PageUptodate(page)) {
 905					unsigned start, end;
 906
 907					start = max(from, block_start);
 908					end = min(to, block_end);
 909
 910					zero_user_segment(page, start, end);
 911					set_buffer_uptodate(bh);
 912				}
 913
 914				clear_buffer_new(bh);
 915				mark_buffer_dirty(bh);
 916			}
 917		}
 918
 919		block_start = block_end;
 920		bh = bh->b_this_page;
 921	} while (bh != head);
 922}
 923
 924/*
 925 * Only called when we have a failure during allocating write to write
 926 * zero's to the newly allocated region.
 927 */
 928static void ocfs2_write_failure(struct inode *inode,
 929				struct ocfs2_write_ctxt *wc,
 930				loff_t user_pos, unsigned user_len)
 931{
 932	int i;
 933	unsigned from = user_pos & (PAGE_SIZE - 1),
 934		to = user_pos + user_len;
 935	struct page *tmppage;
 936
 937	if (wc->w_target_page)
 938		ocfs2_zero_new_buffers(wc->w_target_page, from, to);
 939
 940	for(i = 0; i < wc->w_num_pages; i++) {
 941		tmppage = wc->w_pages[i];
 942
 943		if (tmppage && page_has_buffers(tmppage)) {
 944			if (ocfs2_should_order_data(inode))
 945				ocfs2_jbd2_inode_add_write(wc->w_handle, inode,
 946							   user_pos, user_len);
 947
 948			block_commit_write(tmppage, from, to);
 949		}
 950	}
 951}
 952
 953static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
 954					struct ocfs2_write_ctxt *wc,
 955					struct page *page, u32 cpos,
 956					loff_t user_pos, unsigned user_len,
 957					int new)
 958{
 959	int ret;
 960	unsigned int map_from = 0, map_to = 0;
 961	unsigned int cluster_start, cluster_end;
 962	unsigned int user_data_from = 0, user_data_to = 0;
 963
 964	ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
 965					&cluster_start, &cluster_end);
 966
 967	/* treat the write as new if the a hole/lseek spanned across
 968	 * the page boundary.
 969	 */
 970	new = new | ((i_size_read(inode) <= page_offset(page)) &&
 971			(page_offset(page) <= user_pos));
 972
 973	if (page == wc->w_target_page) {
 974		map_from = user_pos & (PAGE_SIZE - 1);
 975		map_to = map_from + user_len;
 976
 977		if (new)
 978			ret = ocfs2_map_page_blocks(page, p_blkno, inode,
 979						    cluster_start, cluster_end,
 980						    new);
 981		else
 982			ret = ocfs2_map_page_blocks(page, p_blkno, inode,
 983						    map_from, map_to, new);
 984		if (ret) {
 985			mlog_errno(ret);
 986			goto out;
 987		}
 988
 989		user_data_from = map_from;
 990		user_data_to = map_to;
 991		if (new) {
 992			map_from = cluster_start;
 993			map_to = cluster_end;
 994		}
 995	} else {
 996		/*
 997		 * If we haven't allocated the new page yet, we
 998		 * shouldn't be writing it out without copying user
 999		 * data. This is likely a math error from the caller.
1000		 */
1001		BUG_ON(!new);
1002
1003		map_from = cluster_start;
1004		map_to = cluster_end;
1005
1006		ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1007					    cluster_start, cluster_end, new);
1008		if (ret) {
1009			mlog_errno(ret);
1010			goto out;
1011		}
1012	}
1013
1014	/*
1015	 * Parts of newly allocated pages need to be zero'd.
1016	 *
1017	 * Above, we have also rewritten 'to' and 'from' - as far as
1018	 * the rest of the function is concerned, the entire cluster
1019	 * range inside of a page needs to be written.
1020	 *
1021	 * We can skip this if the page is up to date - it's already
1022	 * been zero'd from being read in as a hole.
1023	 */
1024	if (new && !PageUptodate(page))
1025		ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1026					 cpos, user_data_from, user_data_to);
1027
1028	flush_dcache_page(page);
1029
1030out:
1031	return ret;
1032}
1033
1034/*
1035 * This function will only grab one clusters worth of pages.
1036 */
1037static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1038				      struct ocfs2_write_ctxt *wc,
1039				      u32 cpos, loff_t user_pos,
1040				      unsigned user_len, int new,
1041				      struct page *mmap_page)
1042{
1043	int ret = 0, i;
1044	unsigned long start, target_index, end_index, index;
1045	struct inode *inode = mapping->host;
1046	loff_t last_byte;
1047
1048	target_index = user_pos >> PAGE_SHIFT;
1049
1050	/*
1051	 * Figure out how many pages we'll be manipulating here. For
1052	 * non allocating write, we just change the one
1053	 * page. Otherwise, we'll need a whole clusters worth.  If we're
1054	 * writing past i_size, we only need enough pages to cover the
1055	 * last page of the write.
1056	 */
1057	if (new) {
1058		wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1059		start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1060		/*
1061		 * We need the index *past* the last page we could possibly
1062		 * touch.  This is the page past the end of the write or
1063		 * i_size, whichever is greater.
1064		 */
1065		last_byte = max(user_pos + user_len, i_size_read(inode));
1066		BUG_ON(last_byte < 1);
1067		end_index = ((last_byte - 1) >> PAGE_SHIFT) + 1;
1068		if ((start + wc->w_num_pages) > end_index)
1069			wc->w_num_pages = end_index - start;
1070	} else {
1071		wc->w_num_pages = 1;
1072		start = target_index;
1073	}
1074	end_index = (user_pos + user_len - 1) >> PAGE_SHIFT;
1075
1076	for(i = 0; i < wc->w_num_pages; i++) {
1077		index = start + i;
1078
1079		if (index >= target_index && index <= end_index &&
1080		    wc->w_type == OCFS2_WRITE_MMAP) {
1081			/*
1082			 * ocfs2_pagemkwrite() is a little different
1083			 * and wants us to directly use the page
1084			 * passed in.
1085			 */
1086			lock_page(mmap_page);
1087
1088			/* Exit and let the caller retry */
1089			if (mmap_page->mapping != mapping) {
1090				WARN_ON(mmap_page->mapping);
1091				unlock_page(mmap_page);
1092				ret = -EAGAIN;
1093				goto out;
1094			}
1095
1096			get_page(mmap_page);
1097			wc->w_pages[i] = mmap_page;
1098			wc->w_target_locked = true;
1099		} else if (index >= target_index && index <= end_index &&
1100			   wc->w_type == OCFS2_WRITE_DIRECT) {
1101			/* Direct write has no mapping page. */
1102			wc->w_pages[i] = NULL;
1103			continue;
1104		} else {
1105			wc->w_pages[i] = find_or_create_page(mapping, index,
1106							     GFP_NOFS);
1107			if (!wc->w_pages[i]) {
1108				ret = -ENOMEM;
1109				mlog_errno(ret);
1110				goto out;
1111			}
1112		}
1113		wait_for_stable_page(wc->w_pages[i]);
1114
1115		if (index == target_index)
1116			wc->w_target_page = wc->w_pages[i];
1117	}
1118out:
1119	if (ret)
1120		wc->w_target_locked = false;
1121	return ret;
1122}
1123
1124/*
1125 * Prepare a single cluster for write one cluster into the file.
1126 */
1127static int ocfs2_write_cluster(struct address_space *mapping,
1128			       u32 *phys, unsigned int new,
1129			       unsigned int clear_unwritten,
1130			       unsigned int should_zero,
1131			       struct ocfs2_alloc_context *data_ac,
1132			       struct ocfs2_alloc_context *meta_ac,
1133			       struct ocfs2_write_ctxt *wc, u32 cpos,
1134			       loff_t user_pos, unsigned user_len)
1135{
1136	int ret, i;
1137	u64 p_blkno;
1138	struct inode *inode = mapping->host;
1139	struct ocfs2_extent_tree et;
1140	int bpc = ocfs2_clusters_to_blocks(inode->i_sb, 1);
1141
1142	if (new) {
1143		u32 tmp_pos;
1144
1145		/*
1146		 * This is safe to call with the page locks - it won't take
1147		 * any additional semaphores or cluster locks.
1148		 */
1149		tmp_pos = cpos;
1150		ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1151					   &tmp_pos, 1, !clear_unwritten,
1152					   wc->w_di_bh, wc->w_handle,
1153					   data_ac, meta_ac, NULL);
1154		/*
1155		 * This shouldn't happen because we must have already
1156		 * calculated the correct meta data allocation required. The
1157		 * internal tree allocation code should know how to increase
1158		 * transaction credits itself.
1159		 *
1160		 * If need be, we could handle -EAGAIN for a
1161		 * RESTART_TRANS here.
1162		 */
1163		mlog_bug_on_msg(ret == -EAGAIN,
1164				"Inode %llu: EAGAIN return during allocation.\n",
1165				(unsigned long long)OCFS2_I(inode)->ip_blkno);
1166		if (ret < 0) {
1167			mlog_errno(ret);
1168			goto out;
1169		}
1170	} else if (clear_unwritten) {
1171		ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1172					      wc->w_di_bh);
1173		ret = ocfs2_mark_extent_written(inode, &et,
1174						wc->w_handle, cpos, 1, *phys,
1175						meta_ac, &wc->w_dealloc);
1176		if (ret < 0) {
1177			mlog_errno(ret);
1178			goto out;
1179		}
1180	}
1181
1182	/*
1183	 * The only reason this should fail is due to an inability to
1184	 * find the extent added.
1185	 */
1186	ret = ocfs2_get_clusters(inode, cpos, phys, NULL, NULL);
1187	if (ret < 0) {
1188		mlog(ML_ERROR, "Get physical blkno failed for inode %llu, "
1189			    "at logical cluster %u",
1190			    (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
1191		goto out;
1192	}
1193
1194	BUG_ON(*phys == 0);
1195
1196	p_blkno = ocfs2_clusters_to_blocks(inode->i_sb, *phys);
1197	if (!should_zero)
1198		p_blkno += (user_pos >> inode->i_sb->s_blocksize_bits) & (u64)(bpc - 1);
1199
1200	for(i = 0; i < wc->w_num_pages; i++) {
1201		int tmpret;
1202
1203		/* This is the direct io target page. */
1204		if (wc->w_pages[i] == NULL) {
1205			p_blkno++;
1206			continue;
1207		}
1208
1209		tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1210						      wc->w_pages[i], cpos,
1211						      user_pos, user_len,
1212						      should_zero);
1213		if (tmpret) {
1214			mlog_errno(tmpret);
1215			if (ret == 0)
1216				ret = tmpret;
1217		}
1218	}
1219
1220	/*
1221	 * We only have cleanup to do in case of allocating write.
1222	 */
1223	if (ret && new)
1224		ocfs2_write_failure(inode, wc, user_pos, user_len);
1225
1226out:
1227
1228	return ret;
1229}
1230
1231static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1232				       struct ocfs2_alloc_context *data_ac,
1233				       struct ocfs2_alloc_context *meta_ac,
1234				       struct ocfs2_write_ctxt *wc,
1235				       loff_t pos, unsigned len)
1236{
1237	int ret, i;
1238	loff_t cluster_off;
1239	unsigned int local_len = len;
1240	struct ocfs2_write_cluster_desc *desc;
1241	struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1242
1243	for (i = 0; i < wc->w_clen; i++) {
1244		desc = &wc->w_desc[i];
1245
1246		/*
1247		 * We have to make sure that the total write passed in
1248		 * doesn't extend past a single cluster.
1249		 */
1250		local_len = len;
1251		cluster_off = pos & (osb->s_clustersize - 1);
1252		if ((cluster_off + local_len) > osb->s_clustersize)
1253			local_len = osb->s_clustersize - cluster_off;
1254
1255		ret = ocfs2_write_cluster(mapping, &desc->c_phys,
1256					  desc->c_new,
1257					  desc->c_clear_unwritten,
1258					  desc->c_needs_zero,
1259					  data_ac, meta_ac,
1260					  wc, desc->c_cpos, pos, local_len);
1261		if (ret) {
1262			mlog_errno(ret);
1263			goto out;
1264		}
1265
1266		len -= local_len;
1267		pos += local_len;
1268	}
1269
1270	ret = 0;
1271out:
1272	return ret;
1273}
1274
1275/*
1276 * ocfs2_write_end() wants to know which parts of the target page it
1277 * should complete the write on. It's easiest to compute them ahead of
1278 * time when a more complete view of the write is available.
1279 */
1280static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1281					struct ocfs2_write_ctxt *wc,
1282					loff_t pos, unsigned len, int alloc)
1283{
1284	struct ocfs2_write_cluster_desc *desc;
1285
1286	wc->w_target_from = pos & (PAGE_SIZE - 1);
1287	wc->w_target_to = wc->w_target_from + len;
1288
1289	if (alloc == 0)
1290		return;
1291
1292	/*
1293	 * Allocating write - we may have different boundaries based
1294	 * on page size and cluster size.
1295	 *
1296	 * NOTE: We can no longer compute one value from the other as
1297	 * the actual write length and user provided length may be
1298	 * different.
1299	 */
1300
1301	if (wc->w_large_pages) {
1302		/*
1303		 * We only care about the 1st and last cluster within
1304		 * our range and whether they should be zero'd or not. Either
1305		 * value may be extended out to the start/end of a
1306		 * newly allocated cluster.
1307		 */
1308		desc = &wc->w_desc[0];
1309		if (desc->c_needs_zero)
1310			ocfs2_figure_cluster_boundaries(osb,
1311							desc->c_cpos,
1312							&wc->w_target_from,
1313							NULL);
1314
1315		desc = &wc->w_desc[wc->w_clen - 1];
1316		if (desc->c_needs_zero)
1317			ocfs2_figure_cluster_boundaries(osb,
1318							desc->c_cpos,
1319							NULL,
1320							&wc->w_target_to);
1321	} else {
1322		wc->w_target_from = 0;
1323		wc->w_target_to = PAGE_SIZE;
1324	}
1325}
1326
1327/*
1328 * Check if this extent is marked UNWRITTEN by direct io. If so, we need not to
1329 * do the zero work. And should not to clear UNWRITTEN since it will be cleared
1330 * by the direct io procedure.
1331 * If this is a new extent that allocated by direct io, we should mark it in
1332 * the ip_unwritten_list.
1333 */
1334static int ocfs2_unwritten_check(struct inode *inode,
1335				 struct ocfs2_write_ctxt *wc,
1336				 struct ocfs2_write_cluster_desc *desc)
1337{
1338	struct ocfs2_inode_info *oi = OCFS2_I(inode);
1339	struct ocfs2_unwritten_extent *ue = NULL, *new = NULL;
1340	int ret = 0;
1341
1342	if (!desc->c_needs_zero)
1343		return 0;
1344
1345retry:
1346	spin_lock(&oi->ip_lock);
1347	/* Needs not to zero no metter buffer or direct. The one who is zero
1348	 * the cluster is doing zero. And he will clear unwritten after all
1349	 * cluster io finished. */
1350	list_for_each_entry(ue, &oi->ip_unwritten_list, ue_ip_node) {
1351		if (desc->c_cpos == ue->ue_cpos) {
1352			BUG_ON(desc->c_new);
1353			desc->c_needs_zero = 0;
1354			desc->c_clear_unwritten = 0;
1355			goto unlock;
1356		}
1357	}
1358
1359	if (wc->w_type != OCFS2_WRITE_DIRECT)
1360		goto unlock;
1361
1362	if (new == NULL) {
1363		spin_unlock(&oi->ip_lock);
1364		new = kmalloc(sizeof(struct ocfs2_unwritten_extent),
1365			     GFP_NOFS);
1366		if (new == NULL) {
1367			ret = -ENOMEM;
1368			goto out;
1369		}
1370		goto retry;
1371	}
1372	/* This direct write will doing zero. */
1373	new->ue_cpos = desc->c_cpos;
1374	new->ue_phys = desc->c_phys;
1375	desc->c_clear_unwritten = 0;
1376	list_add_tail(&new->ue_ip_node, &oi->ip_unwritten_list);
1377	list_add_tail(&new->ue_node, &wc->w_unwritten_list);
1378	wc->w_unwritten_count++;
1379	new = NULL;
1380unlock:
1381	spin_unlock(&oi->ip_lock);
1382out:
1383	kfree(new);
1384	return ret;
1385}
1386
1387/*
1388 * Populate each single-cluster write descriptor in the write context
1389 * with information about the i/o to be done.
1390 *
1391 * Returns the number of clusters that will have to be allocated, as
1392 * well as a worst case estimate of the number of extent records that
1393 * would have to be created during a write to an unwritten region.
1394 */
1395static int ocfs2_populate_write_desc(struct inode *inode,
1396				     struct ocfs2_write_ctxt *wc,
1397				     unsigned int *clusters_to_alloc,
1398				     unsigned int *extents_to_split)
1399{
1400	int ret;
1401	struct ocfs2_write_cluster_desc *desc;
1402	unsigned int num_clusters = 0;
1403	unsigned int ext_flags = 0;
1404	u32 phys = 0;
1405	int i;
1406
1407	*clusters_to_alloc = 0;
1408	*extents_to_split = 0;
1409
1410	for (i = 0; i < wc->w_clen; i++) {
1411		desc = &wc->w_desc[i];
1412		desc->c_cpos = wc->w_cpos + i;
1413
1414		if (num_clusters == 0) {
1415			/*
1416			 * Need to look up the next extent record.
1417			 */
1418			ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1419						 &num_clusters, &ext_flags);
1420			if (ret) {
1421				mlog_errno(ret);
1422				goto out;
1423			}
1424
1425			/* We should already CoW the refcountd extent. */
1426			BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1427
1428			/*
1429			 * Assume worst case - that we're writing in
1430			 * the middle of the extent.
1431			 *
1432			 * We can assume that the write proceeds from
1433			 * left to right, in which case the extent
1434			 * insert code is smart enough to coalesce the
1435			 * next splits into the previous records created.
1436			 */
1437			if (ext_flags & OCFS2_EXT_UNWRITTEN)
1438				*extents_to_split = *extents_to_split + 2;
1439		} else if (phys) {
1440			/*
1441			 * Only increment phys if it doesn't describe
1442			 * a hole.
1443			 */
1444			phys++;
1445		}
1446
1447		/*
1448		 * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1449		 * file that got extended.  w_first_new_cpos tells us
1450		 * where the newly allocated clusters are so we can
1451		 * zero them.
1452		 */
1453		if (desc->c_cpos >= wc->w_first_new_cpos) {
1454			BUG_ON(phys == 0);
1455			desc->c_needs_zero = 1;
1456		}
1457
1458		desc->c_phys = phys;
1459		if (phys == 0) {
1460			desc->c_new = 1;
1461			desc->c_needs_zero = 1;
1462			desc->c_clear_unwritten = 1;
1463			*clusters_to_alloc = *clusters_to_alloc + 1;
1464		}
1465
1466		if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1467			desc->c_clear_unwritten = 1;
1468			desc->c_needs_zero = 1;
1469		}
1470
1471		ret = ocfs2_unwritten_check(inode, wc, desc);
1472		if (ret) {
1473			mlog_errno(ret);
1474			goto out;
1475		}
1476
1477		num_clusters--;
1478	}
1479
1480	ret = 0;
1481out:
1482	return ret;
1483}
1484
1485static int ocfs2_write_begin_inline(struct address_space *mapping,
1486				    struct inode *inode,
1487				    struct ocfs2_write_ctxt *wc)
1488{
1489	int ret;
1490	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1491	struct page *page;
1492	handle_t *handle;
1493	struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1494
1495	handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1496	if (IS_ERR(handle)) {
1497		ret = PTR_ERR(handle);
1498		mlog_errno(ret);
1499		goto out;
1500	}
1501
1502	page = find_or_create_page(mapping, 0, GFP_NOFS);
1503	if (!page) {
1504		ocfs2_commit_trans(osb, handle);
1505		ret = -ENOMEM;
1506		mlog_errno(ret);
1507		goto out;
1508	}
1509	/*
1510	 * If we don't set w_num_pages then this page won't get unlocked
1511	 * and freed on cleanup of the write context.
1512	 */
1513	wc->w_pages[0] = wc->w_target_page = page;
1514	wc->w_num_pages = 1;
1515
1516	ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1517				      OCFS2_JOURNAL_ACCESS_WRITE);
1518	if (ret) {
1519		ocfs2_commit_trans(osb, handle);
1520
1521		mlog_errno(ret);
1522		goto out;
1523	}
1524
1525	if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1526		ocfs2_set_inode_data_inline(inode, di);
1527
1528	if (!PageUptodate(page)) {
1529		ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1530		if (ret) {
1531			ocfs2_commit_trans(osb, handle);
1532
1533			goto out;
1534		}
1535	}
1536
1537	wc->w_handle = handle;
1538out:
1539	return ret;
1540}
1541
1542int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1543{
1544	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1545
1546	if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1547		return 1;
1548	return 0;
1549}
1550
1551static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1552					  struct inode *inode, loff_t pos,
1553					  unsigned len, struct page *mmap_page,
1554					  struct ocfs2_write_ctxt *wc)
1555{
1556	int ret, written = 0;
1557	loff_t end = pos + len;
1558	struct ocfs2_inode_info *oi = OCFS2_I(inode);
1559	struct ocfs2_dinode *di = NULL;
1560
1561	trace_ocfs2_try_to_write_inline_data((unsigned long long)oi->ip_blkno,
1562					     len, (unsigned long long)pos,
1563					     oi->ip_dyn_features);
1564
1565	/*
1566	 * Handle inodes which already have inline data 1st.
1567	 */
1568	if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1569		if (mmap_page == NULL &&
1570		    ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1571			goto do_inline_write;
1572
1573		/*
1574		 * The write won't fit - we have to give this inode an
1575		 * inline extent list now.
1576		 */
1577		ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1578		if (ret)
1579			mlog_errno(ret);
1580		goto out;
1581	}
1582
1583	/*
1584	 * Check whether the inode can accept inline data.
1585	 */
1586	if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1587		return 0;
1588
1589	/*
1590	 * Check whether the write can fit.
1591	 */
1592	di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1593	if (mmap_page ||
1594	    end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1595		return 0;
1596
1597do_inline_write:
1598	ret = ocfs2_write_begin_inline(mapping, inode, wc);
1599	if (ret) {
1600		mlog_errno(ret);
1601		goto out;
1602	}
1603
1604	/*
1605	 * This signals to the caller that the data can be written
1606	 * inline.
1607	 */
1608	written = 1;
1609out:
1610	return written ? written : ret;
1611}
1612
1613/*
1614 * This function only does anything for file systems which can't
1615 * handle sparse files.
1616 *
1617 * What we want to do here is fill in any hole between the current end
1618 * of allocation and the end of our write. That way the rest of the
1619 * write path can treat it as an non-allocating write, which has no
1620 * special case code for sparse/nonsparse files.
1621 */
1622static int ocfs2_expand_nonsparse_inode(struct inode *inode,
1623					struct buffer_head *di_bh,
1624					loff_t pos, unsigned len,
1625					struct ocfs2_write_ctxt *wc)
1626{
1627	int ret;
1628	loff_t newsize = pos + len;
1629
1630	BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1631
1632	if (newsize <= i_size_read(inode))
1633		return 0;
1634
1635	ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
1636	if (ret)
1637		mlog_errno(ret);
1638
1639	/* There is no wc if this is call from direct. */
1640	if (wc)
1641		wc->w_first_new_cpos =
1642			ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1643
1644	return ret;
1645}
1646
1647static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
1648			   loff_t pos)
1649{
1650	int ret = 0;
1651
1652	BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1653	if (pos > i_size_read(inode))
1654		ret = ocfs2_zero_extend(inode, di_bh, pos);
1655
1656	return ret;
1657}
1658
1659int ocfs2_write_begin_nolock(struct address_space *mapping,
1660			     loff_t pos, unsigned len, ocfs2_write_type_t type,
1661			     struct page **pagep, void **fsdata,
1662			     struct buffer_head *di_bh, struct page *mmap_page)
1663{
1664	int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1665	unsigned int clusters_to_alloc, extents_to_split, clusters_need = 0;
1666	struct ocfs2_write_ctxt *wc;
1667	struct inode *inode = mapping->host;
1668	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1669	struct ocfs2_dinode *di;
1670	struct ocfs2_alloc_context *data_ac = NULL;
1671	struct ocfs2_alloc_context *meta_ac = NULL;
1672	handle_t *handle;
1673	struct ocfs2_extent_tree et;
1674	int try_free = 1, ret1;
1675
1676try_again:
1677	ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, type, di_bh);
1678	if (ret) {
1679		mlog_errno(ret);
1680		return ret;
1681	}
1682
1683	if (ocfs2_supports_inline_data(osb)) {
1684		ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1685						     mmap_page, wc);
1686		if (ret == 1) {
1687			ret = 0;
1688			goto success;
1689		}
1690		if (ret < 0) {
1691			mlog_errno(ret);
1692			goto out;
1693		}
1694	}
1695
1696	/* Direct io change i_size late, should not zero tail here. */
1697	if (type != OCFS2_WRITE_DIRECT) {
1698		if (ocfs2_sparse_alloc(osb))
1699			ret = ocfs2_zero_tail(inode, di_bh, pos);
1700		else
1701			ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos,
1702							   len, wc);
1703		if (ret) {
1704			mlog_errno(ret);
1705			goto out;
1706		}
1707	}
1708
1709	ret = ocfs2_check_range_for_refcount(inode, pos, len);
1710	if (ret < 0) {
1711		mlog_errno(ret);
1712		goto out;
1713	} else if (ret == 1) {
1714		clusters_need = wc->w_clen;
1715		ret = ocfs2_refcount_cow(inode, di_bh,
1716					 wc->w_cpos, wc->w_clen, UINT_MAX);
1717		if (ret) {
1718			mlog_errno(ret);
1719			goto out;
1720		}
1721	}
1722
1723	ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1724					&extents_to_split);
1725	if (ret) {
1726		mlog_errno(ret);
1727		goto out;
1728	}
1729	clusters_need += clusters_to_alloc;
1730
1731	di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1732
1733	trace_ocfs2_write_begin_nolock(
1734			(unsigned long long)OCFS2_I(inode)->ip_blkno,
1735			(long long)i_size_read(inode),
1736			le32_to_cpu(di->i_clusters),
1737			pos, len, type, mmap_page,
1738			clusters_to_alloc, extents_to_split);
1739
1740	/*
1741	 * We set w_target_from, w_target_to here so that
1742	 * ocfs2_write_end() knows which range in the target page to
1743	 * write out. An allocation requires that we write the entire
1744	 * cluster range.
1745	 */
1746	if (clusters_to_alloc || extents_to_split) {
1747		/*
1748		 * XXX: We are stretching the limits of
1749		 * ocfs2_lock_allocators(). It greatly over-estimates
1750		 * the work to be done.
1751		 */
1752		ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1753					      wc->w_di_bh);
1754		ret = ocfs2_lock_allocators(inode, &et,
1755					    clusters_to_alloc, extents_to_split,
1756					    &data_ac, &meta_ac);
1757		if (ret) {
1758			mlog_errno(ret);
1759			goto out;
1760		}
1761
1762		if (data_ac)
1763			data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
1764
1765		credits = ocfs2_calc_extend_credits(inode->i_sb,
1766						    &di->id2.i_list);
1767	} else if (type == OCFS2_WRITE_DIRECT)
1768		/* direct write needs not to start trans if no extents alloc. */
1769		goto success;
1770
1771	/*
1772	 * We have to zero sparse allocated clusters, unwritten extent clusters,
1773	 * and non-sparse clusters we just extended.  For non-sparse writes,
1774	 * we know zeros will only be needed in the first and/or last cluster.
1775	 */
1776	if (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1777			   wc->w_desc[wc->w_clen - 1].c_needs_zero))
1778		cluster_of_pages = 1;
1779	else
1780		cluster_of_pages = 0;
1781
1782	ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1783
1784	handle = ocfs2_start_trans(osb, credits);
1785	if (IS_ERR(handle)) {
1786		ret = PTR_ERR(handle);
1787		mlog_errno(ret);
1788		goto out;
1789	}
1790
1791	wc->w_handle = handle;
1792
1793	if (clusters_to_alloc) {
1794		ret = dquot_alloc_space_nodirty(inode,
1795			ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1796		if (ret)
1797			goto out_commit;
1798	}
1799
1800	ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1801				      OCFS2_JOURNAL_ACCESS_WRITE);
1802	if (ret) {
1803		mlog_errno(ret);
1804		goto out_quota;
1805	}
1806
1807	/*
1808	 * Fill our page array first. That way we've grabbed enough so
1809	 * that we can zero and flush if we error after adding the
1810	 * extent.
1811	 */
1812	ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
1813					 cluster_of_pages, mmap_page);
1814	if (ret && ret != -EAGAIN) {
1815		mlog_errno(ret);
1816		goto out_quota;
1817	}
1818
1819	/*
1820	 * ocfs2_grab_pages_for_write() returns -EAGAIN if it could not lock
1821	 * the target page. In this case, we exit with no error and no target
1822	 * page. This will trigger the caller, page_mkwrite(), to re-try
1823	 * the operation.
1824	 */
1825	if (ret == -EAGAIN) {
1826		BUG_ON(wc->w_target_page);
1827		ret = 0;
1828		goto out_quota;
1829	}
1830
1831	ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1832					  len);
1833	if (ret) {
1834		mlog_errno(ret);
1835		goto out_quota;
1836	}
1837
1838	if (data_ac)
1839		ocfs2_free_alloc_context(data_ac);
1840	if (meta_ac)
1841		ocfs2_free_alloc_context(meta_ac);
1842
1843success:
1844	if (pagep)
1845		*pagep = wc->w_target_page;
1846	*fsdata = wc;
1847	return 0;
1848out_quota:
1849	if (clusters_to_alloc)
1850		dquot_free_space(inode,
1851			  ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1852out_commit:
1853	ocfs2_commit_trans(osb, handle);
1854
1855out:
1856	/*
1857	 * The mmapped page won't be unlocked in ocfs2_free_write_ctxt(),
1858	 * even in case of error here like ENOSPC and ENOMEM. So, we need
1859	 * to unlock the target page manually to prevent deadlocks when
1860	 * retrying again on ENOSPC, or when returning non-VM_FAULT_LOCKED
1861	 * to VM code.
1862	 */
1863	if (wc->w_target_locked)
1864		unlock_page(mmap_page);
1865
1866	ocfs2_free_write_ctxt(inode, wc);
1867
1868	if (data_ac) {
1869		ocfs2_free_alloc_context(data_ac);
1870		data_ac = NULL;
1871	}
1872	if (meta_ac) {
1873		ocfs2_free_alloc_context(meta_ac);
1874		meta_ac = NULL;
1875	}
1876
1877	if (ret == -ENOSPC && try_free) {
1878		/*
1879		 * Try to free some truncate log so that we can have enough
1880		 * clusters to allocate.
1881		 */
1882		try_free = 0;
1883
1884		ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need);
1885		if (ret1 == 1)
1886			goto try_again;
1887
1888		if (ret1 < 0)
1889			mlog_errno(ret1);
1890	}
1891
1892	return ret;
1893}
1894
1895static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1896			     loff_t pos, unsigned len, unsigned flags,
1897			     struct page **pagep, void **fsdata)
1898{
1899	int ret;
1900	struct buffer_head *di_bh = NULL;
1901	struct inode *inode = mapping->host;
1902
1903	ret = ocfs2_inode_lock(inode, &di_bh, 1);
1904	if (ret) {
1905		mlog_errno(ret);
1906		return ret;
1907	}
1908
1909	/*
1910	 * Take alloc sem here to prevent concurrent lookups. That way
1911	 * the mapping, zeroing and tree manipulation within
1912	 * ocfs2_write() will be safe against ->readpage(). This
1913	 * should also serve to lock out allocation from a shared
1914	 * writeable region.
1915	 */
1916	down_write(&OCFS2_I(inode)->ip_alloc_sem);
1917
1918	ret = ocfs2_write_begin_nolock(mapping, pos, len, OCFS2_WRITE_BUFFER,
1919				       pagep, fsdata, di_bh, NULL);
1920	if (ret) {
1921		mlog_errno(ret);
1922		goto out_fail;
1923	}
1924
1925	brelse(di_bh);
1926
1927	return 0;
1928
1929out_fail:
1930	up_write(&OCFS2_I(inode)->ip_alloc_sem);
1931
1932	brelse(di_bh);
1933	ocfs2_inode_unlock(inode, 1);
1934
1935	return ret;
1936}
1937
1938static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1939				   unsigned len, unsigned *copied,
1940				   struct ocfs2_dinode *di,
1941				   struct ocfs2_write_ctxt *wc)
1942{
1943	void *kaddr;
1944
1945	if (unlikely(*copied < len)) {
1946		if (!PageUptodate(wc->w_target_page)) {
1947			*copied = 0;
1948			return;
1949		}
1950	}
1951
1952	kaddr = kmap_atomic(wc->w_target_page);
1953	memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1954	kunmap_atomic(kaddr);
1955
1956	trace_ocfs2_write_end_inline(
1957	     (unsigned long long)OCFS2_I(inode)->ip_blkno,
1958	     (unsigned long long)pos, *copied,
1959	     le16_to_cpu(di->id2.i_data.id_count),
1960	     le16_to_cpu(di->i_dyn_features));
1961}
1962
1963int ocfs2_write_end_nolock(struct address_space *mapping,
1964			   loff_t pos, unsigned len, unsigned copied, void *fsdata)
1965{
1966	int i, ret;
1967	unsigned from, to, start = pos & (PAGE_SIZE - 1);
1968	struct inode *inode = mapping->host;
1969	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1970	struct ocfs2_write_ctxt *wc = fsdata;
1971	struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1972	handle_t *handle = wc->w_handle;
1973	struct page *tmppage;
1974
1975	BUG_ON(!list_empty(&wc->w_unwritten_list));
1976
1977	if (handle) {
1978		ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode),
1979				wc->w_di_bh, OCFS2_JOURNAL_ACCESS_WRITE);
1980		if (ret) {
1981			copied = ret;
1982			mlog_errno(ret);
1983			goto out;
1984		}
1985	}
1986
1987	if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1988		ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
1989		goto out_write_size;
1990	}
1991
1992	if (unlikely(copied < len) && wc->w_target_page) {
1993		if (!PageUptodate(wc->w_target_page))
1994			copied = 0;
1995
1996		ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
1997				       start+len);
1998	}
1999	if (wc->w_target_page)
2000		flush_dcache_page(wc->w_target_page);
2001
2002	for(i = 0; i < wc->w_num_pages; i++) {
2003		tmppage = wc->w_pages[i];
2004
2005		/* This is the direct io target page. */
2006		if (tmppage == NULL)
2007			continue;
2008
2009		if (tmppage == wc->w_target_page) {
2010			from = wc->w_target_from;
2011			to = wc->w_target_to;
2012
2013			BUG_ON(from > PAGE_SIZE ||
2014			       to > PAGE_SIZE ||
2015			       to < from);
2016		} else {
2017			/*
2018			 * Pages adjacent to the target (if any) imply
2019			 * a hole-filling write in which case we want
2020			 * to flush their entire range.
2021			 */
2022			from = 0;
2023			to = PAGE_SIZE;
2024		}
2025
2026		if (page_has_buffers(tmppage)) {
2027			if (handle && ocfs2_should_order_data(inode)) {
2028				loff_t start_byte =
2029					((loff_t)tmppage->index << PAGE_SHIFT) +
2030					from;
2031				loff_t length = to - from;
2032				ocfs2_jbd2_inode_add_write(handle, inode,
2033							   start_byte, length);
2034			}
2035			block_commit_write(tmppage, from, to);
2036		}
2037	}
2038
2039out_write_size:
2040	/* Direct io do not update i_size here. */
2041	if (wc->w_type != OCFS2_WRITE_DIRECT) {
2042		pos += copied;
2043		if (pos > i_size_read(inode)) {
2044			i_size_write(inode, pos);
2045			mark_inode_dirty(inode);
2046		}
2047		inode->i_blocks = ocfs2_inode_sector_count(inode);
2048		di->i_size = cpu_to_le64((u64)i_size_read(inode));
2049		inode->i_mtime = inode->i_ctime = current_time(inode);
2050		di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
2051		di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
2052		if (handle)
2053			ocfs2_update_inode_fsync_trans(handle, inode, 1);
2054	}
2055	if (handle)
2056		ocfs2_journal_dirty(handle, wc->w_di_bh);
2057
2058out:
2059	/* unlock pages before dealloc since it needs acquiring j_trans_barrier
2060	 * lock, or it will cause a deadlock since journal commit threads holds
2061	 * this lock and will ask for the page lock when flushing the data.
2062	 * put it here to preserve the unlock order.
2063	 */
2064	ocfs2_unlock_pages(wc);
2065
2066	if (handle)
2067		ocfs2_commit_trans(osb, handle);
2068
2069	ocfs2_run_deallocs(osb, &wc->w_dealloc);
2070
2071	brelse(wc->w_di_bh);
2072	kfree(wc);
2073
2074	return copied;
2075}
2076
2077static int ocfs2_write_end(struct file *file, struct address_space *mapping,
2078			   loff_t pos, unsigned len, unsigned copied,
2079			   struct page *page, void *fsdata)
2080{
2081	int ret;
2082	struct inode *inode = mapping->host;
2083
2084	ret = ocfs2_write_end_nolock(mapping, pos, len, copied, fsdata);
2085
2086	up_write(&OCFS2_I(inode)->ip_alloc_sem);
2087	ocfs2_inode_unlock(inode, 1);
2088
2089	return ret;
2090}
2091
2092struct ocfs2_dio_write_ctxt {
2093	struct list_head	dw_zero_list;
2094	unsigned		dw_zero_count;
2095	int			dw_orphaned;
2096	pid_t			dw_writer_pid;
2097};
2098
2099static struct ocfs2_dio_write_ctxt *
2100ocfs2_dio_alloc_write_ctx(struct buffer_head *bh, int *alloc)
2101{
2102	struct ocfs2_dio_write_ctxt *dwc = NULL;
2103
2104	if (bh->b_private)
2105		return bh->b_private;
2106
2107	dwc = kmalloc(sizeof(struct ocfs2_dio_write_ctxt), GFP_NOFS);
2108	if (dwc == NULL)
2109		return NULL;
2110	INIT_LIST_HEAD(&dwc->dw_zero_list);
2111	dwc->dw_zero_count = 0;
2112	dwc->dw_orphaned = 0;
2113	dwc->dw_writer_pid = task_pid_nr(current);
2114	bh->b_private = dwc;
2115	*alloc = 1;
2116
2117	return dwc;
2118}
2119
2120static void ocfs2_dio_free_write_ctx(struct inode *inode,
2121				     struct ocfs2_dio_write_ctxt *dwc)
2122{
2123	ocfs2_free_unwritten_list(inode, &dwc->dw_zero_list);
2124	kfree(dwc);
2125}
2126
2127/*
2128 * TODO: Make this into a generic get_blocks function.
2129 *
2130 * From do_direct_io in direct-io.c:
2131 *  "So what we do is to permit the ->get_blocks function to populate
2132 *   bh.b_size with the size of IO which is permitted at this offset and
2133 *   this i_blkbits."
2134 *
2135 * This function is called directly from get_more_blocks in direct-io.c.
2136 *
2137 * called like this: dio->get_blocks(dio->inode, fs_startblk,
2138 * 					fs_count, map_bh, dio->rw == WRITE);
2139 */
2140static int ocfs2_dio_wr_get_block(struct inode *inode, sector_t iblock,
2141			       struct buffer_head *bh_result, int create)
2142{
2143	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2144	struct ocfs2_inode_info *oi = OCFS2_I(inode);
2145	struct ocfs2_write_ctxt *wc;
2146	struct ocfs2_write_cluster_desc *desc = NULL;
2147	struct ocfs2_dio_write_ctxt *dwc = NULL;
2148	struct buffer_head *di_bh = NULL;
2149	u64 p_blkno;
2150	unsigned int i_blkbits = inode->i_sb->s_blocksize_bits;
2151	loff_t pos = iblock << i_blkbits;
2152	sector_t endblk = (i_size_read(inode) - 1) >> i_blkbits;
2153	unsigned len, total_len = bh_result->b_size;
2154	int ret = 0, first_get_block = 0;
2155
2156	len = osb->s_clustersize - (pos & (osb->s_clustersize - 1));
2157	len = min(total_len, len);
2158
2159	/*
2160	 * bh_result->b_size is count in get_more_blocks according to write
2161	 * "pos" and "end", we need map twice to return different buffer state:
2162	 * 1. area in file size, not set NEW;
2163	 * 2. area out file size, set  NEW.
2164	 *
2165	 *		   iblock    endblk
2166	 * |--------|---------|---------|---------
2167	 * |<-------area in file------->|
2168	 */
2169
2170	if ((iblock <= endblk) &&
2171	    ((iblock + ((len - 1) >> i_blkbits)) > endblk))
2172		len = (endblk - iblock + 1) << i_blkbits;
2173
2174	mlog(0, "get block of %lu at %llu:%u req %u\n",
2175			inode->i_ino, pos, len, total_len);
2176
2177	/*
2178	 * Because we need to change file size in ocfs2_dio_end_io_write(), or
2179	 * we may need to add it to orphan dir. So can not fall to fast path
2180	 * while file size will be changed.
2181	 */
2182	if (pos + total_len <= i_size_read(inode)) {
2183
2184		/* This is the fast path for re-write. */
2185		ret = ocfs2_lock_get_block(inode, iblock, bh_result, create);
2186		if (buffer_mapped(bh_result) &&
2187		    !buffer_new(bh_result) &&
2188		    ret == 0)
2189			goto out;
2190
2191		/* Clear state set by ocfs2_get_block. */
2192		bh_result->b_state = 0;
2193	}
2194
2195	dwc = ocfs2_dio_alloc_write_ctx(bh_result, &first_get_block);
2196	if (unlikely(dwc == NULL)) {
2197		ret = -ENOMEM;
2198		mlog_errno(ret);
2199		goto out;
2200	}
2201
2202	if (ocfs2_clusters_for_bytes(inode->i_sb, pos + total_len) >
2203	    ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode)) &&
2204	    !dwc->dw_orphaned) {
2205		/*
2206		 * when we are going to alloc extents beyond file size, add the
2207		 * inode to orphan dir, so we can recall those spaces when
2208		 * system crashed during write.
2209		 */
2210		ret = ocfs2_add_inode_to_orphan(osb, inode);
2211		if (ret < 0) {
2212			mlog_errno(ret);
2213			goto out;
2214		}
2215		dwc->dw_orphaned = 1;
2216	}
2217
2218	ret = ocfs2_inode_lock(inode, &di_bh, 1);
2219	if (ret) {
2220		mlog_errno(ret);
2221		goto out;
2222	}
2223
2224	down_write(&oi->ip_alloc_sem);
2225
2226	if (first_get_block) {
2227		if (ocfs2_sparse_alloc(osb))
2228			ret = ocfs2_zero_tail(inode, di_bh, pos);
2229		else
2230			ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos,
2231							   total_len, NULL);
2232		if (ret < 0) {
2233			mlog_errno(ret);
2234			goto unlock;
2235		}
2236	}
2237
2238	ret = ocfs2_write_begin_nolock(inode->i_mapping, pos, len,
2239				       OCFS2_WRITE_DIRECT, NULL,
2240				       (void **)&wc, di_bh, NULL);
2241	if (ret) {
2242		mlog_errno(ret);
2243		goto unlock;
2244	}
2245
2246	desc = &wc->w_desc[0];
2247
2248	p_blkno = ocfs2_clusters_to_blocks(inode->i_sb, desc->c_phys);
2249	BUG_ON(p_blkno == 0);
2250	p_blkno += iblock & (u64)(ocfs2_clusters_to_blocks(inode->i_sb, 1) - 1);
2251
2252	map_bh(bh_result, inode->i_sb, p_blkno);
2253	bh_result->b_size = len;
2254	if (desc->c_needs_zero)
2255		set_buffer_new(bh_result);
2256
2257	if (iblock > endblk)
2258		set_buffer_new(bh_result);
2259
2260	/* May sleep in end_io. It should not happen in a irq context. So defer
2261	 * it to dio work queue. */
2262	set_buffer_defer_completion(bh_result);
2263
2264	if (!list_empty(&wc->w_unwritten_list)) {
2265		struct ocfs2_unwritten_extent *ue = NULL;
2266
2267		ue = list_first_entry(&wc->w_unwritten_list,
2268				      struct ocfs2_unwritten_extent,
2269				      ue_node);
2270		BUG_ON(ue->ue_cpos != desc->c_cpos);
2271		/* The physical address may be 0, fill it. */
2272		ue->ue_phys = desc->c_phys;
2273
2274		list_splice_tail_init(&wc->w_unwritten_list, &dwc->dw_zero_list);
2275		dwc->dw_zero_count += wc->w_unwritten_count;
2276	}
2277
2278	ret = ocfs2_write_end_nolock(inode->i_mapping, pos, len, len, wc);
2279	BUG_ON(ret != len);
2280	ret = 0;
2281unlock:
2282	up_write(&oi->ip_alloc_sem);
2283	ocfs2_inode_unlock(inode, 1);
2284	brelse(di_bh);
2285out:
2286	if (ret < 0)
2287		ret = -EIO;
2288	return ret;
2289}
2290
2291static int ocfs2_dio_end_io_write(struct inode *inode,
2292				  struct ocfs2_dio_write_ctxt *dwc,
2293				  loff_t offset,
2294				  ssize_t bytes)
2295{
2296	struct ocfs2_cached_dealloc_ctxt dealloc;
2297	struct ocfs2_extent_tree et;
2298	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2299	struct ocfs2_inode_info *oi = OCFS2_I(inode);
2300	struct ocfs2_unwritten_extent *ue = NULL;
2301	struct buffer_head *di_bh = NULL;
2302	struct ocfs2_dinode *di;
2303	struct ocfs2_alloc_context *data_ac = NULL;
2304	struct ocfs2_alloc_context *meta_ac = NULL;
2305	handle_t *handle = NULL;
2306	loff_t end = offset + bytes;
2307	int ret = 0, credits = 0, locked = 0;
2308
2309	ocfs2_init_dealloc_ctxt(&dealloc);
2310
2311	/* We do clear unwritten, delete orphan, change i_size here. If neither
2312	 * of these happen, we can skip all this. */
2313	if (list_empty(&dwc->dw_zero_list) &&
2314	    end <= i_size_read(inode) &&
2315	    !dwc->dw_orphaned)
2316		goto out;
2317
2318	/* ocfs2_file_write_iter will get i_mutex, so we need not lock if we
2319	 * are in that context. */
2320	if (dwc->dw_writer_pid != task_pid_nr(current)) {
2321		inode_lock(inode);
2322		locked = 1;
2323	}
2324
2325	ret = ocfs2_inode_lock(inode, &di_bh, 1);
2326	if (ret < 0) {
2327		mlog_errno(ret);
2328		goto out;
2329	}
2330
2331	down_write(&oi->ip_alloc_sem);
2332
2333	/* Delete orphan before acquire i_mutex. */
2334	if (dwc->dw_orphaned) {
2335		BUG_ON(dwc->dw_writer_pid != task_pid_nr(current));
2336
2337		end = end > i_size_read(inode) ? end : 0;
2338
2339		ret = ocfs2_del_inode_from_orphan(osb, inode, di_bh,
2340				!!end, end);
2341		if (ret < 0)
2342			mlog_errno(ret);
2343	}
2344
2345	di = (struct ocfs2_dinode *)di_bh->b_data;
2346
2347	ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
2348
2349	/* Attach dealloc with extent tree in case that we may reuse extents
2350	 * which are already unlinked from current extent tree due to extent
2351	 * rotation and merging.
2352	 */
2353	et.et_dealloc = &dealloc;
2354
2355	ret = ocfs2_lock_allocators(inode, &et, 0, dwc->dw_zero_count*2,
2356				    &data_ac, &meta_ac);
2357	if (ret) {
2358		mlog_errno(ret);
2359		goto unlock;
2360	}
2361
2362	credits = ocfs2_calc_extend_credits(inode->i_sb, &di->id2.i_list);
2363
2364	handle = ocfs2_start_trans(osb, credits);
2365	if (IS_ERR(handle)) {
2366		ret = PTR_ERR(handle);
2367		mlog_errno(ret);
2368		goto unlock;
2369	}
2370	ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
2371				      OCFS2_JOURNAL_ACCESS_WRITE);
2372	if (ret) {
2373		mlog_errno(ret);
2374		goto commit;
2375	}
2376
2377	list_for_each_entry(ue, &dwc->dw_zero_list, ue_node) {
2378		ret = ocfs2_mark_extent_written(inode, &et, handle,
2379						ue->ue_cpos, 1,
2380						ue->ue_phys,
2381						meta_ac, &dealloc);
2382		if (ret < 0) {
2383			mlog_errno(ret);
2384			break;
2385		}
2386	}
2387
2388	if (end > i_size_read(inode)) {
2389		ret = ocfs2_set_inode_size(handle, inode, di_bh, end);
2390		if (ret < 0)
2391			mlog_errno(ret);
2392	}
2393commit:
2394	ocfs2_commit_trans(osb, handle);
2395unlock:
2396	up_write(&oi->ip_alloc_sem);
2397	ocfs2_inode_unlock(inode, 1);
2398	brelse(di_bh);
2399out:
2400	if (data_ac)
2401		ocfs2_free_alloc_context(data_ac);
2402	if (meta_ac)
2403		ocfs2_free_alloc_context(meta_ac);
2404	ocfs2_run_deallocs(osb, &dealloc);
2405	if (locked)
2406		inode_unlock(inode);
2407	ocfs2_dio_free_write_ctx(inode, dwc);
2408
2409	return ret;
2410}
2411
2412/*
2413 * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
2414 * particularly interested in the aio/dio case.  We use the rw_lock DLM lock
2415 * to protect io on one node from truncation on another.
2416 */
2417static int ocfs2_dio_end_io(struct kiocb *iocb,
2418			    loff_t offset,
2419			    ssize_t bytes,
2420			    void *private)
2421{
2422	struct inode *inode = file_inode(iocb->ki_filp);
2423	int level;
2424	int ret = 0;
2425
2426	/* this io's submitter should not have unlocked this before we could */
2427	BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
2428
2429	if (bytes <= 0)
2430		mlog_ratelimited(ML_ERROR, "Direct IO failed, bytes = %lld",
2431				 (long long)bytes);
2432	if (private) {
2433		if (bytes > 0)
2434			ret = ocfs2_dio_end_io_write(inode, private, offset,
2435						     bytes);
2436		else
2437			ocfs2_dio_free_write_ctx(inode, private);
2438	}
2439
2440	ocfs2_iocb_clear_rw_locked(iocb);
2441
2442	level = ocfs2_iocb_rw_locked_level(iocb);
2443	ocfs2_rw_unlock(inode, level);
2444	return ret;
2445}
2446
2447static ssize_t ocfs2_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2448{
2449	struct file *file = iocb->ki_filp;
2450	struct inode *inode = file->f_mapping->host;
2451	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2452	get_block_t *get_block;
2453
2454	/*
2455	 * Fallback to buffered I/O if we see an inode without
2456	 * extents.
2457	 */
2458	if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2459		return 0;
2460
2461	/* Fallback to buffered I/O if we do not support append dio. */
2462	if (iocb->ki_pos + iter->count > i_size_read(inode) &&
2463	    !ocfs2_supports_append_dio(osb))
2464		return 0;
2465
2466	if (iov_iter_rw(iter) == READ)
2467		get_block = ocfs2_lock_get_block;
2468	else
2469		get_block = ocfs2_dio_wr_get_block;
2470
2471	return __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev,
2472				    iter, get_block,
2473				    ocfs2_dio_end_io, NULL, 0);
2474}
2475
2476const struct address_space_operations ocfs2_aops = {
2477	.readpage		= ocfs2_readpage,
2478	.readpages		= ocfs2_readpages,
2479	.writepage		= ocfs2_writepage,
2480	.write_begin		= ocfs2_write_begin,
2481	.write_end		= ocfs2_write_end,
2482	.bmap			= ocfs2_bmap,
2483	.direct_IO		= ocfs2_direct_IO,
2484	.invalidatepage		= block_invalidatepage,
2485	.releasepage		= ocfs2_releasepage,
2486	.migratepage		= buffer_migrate_page,
2487	.is_partially_uptodate	= block_is_partially_uptodate,
2488	.error_remove_page	= generic_error_remove_page,
2489};