1/* -*- mode: c; c-basic-offset: 8; -*-
   2 * vim: noexpandtab sw=8 ts=8 sts=0:
   3 *
   4 * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
   5 *
   6 * This program is free software; you can redistribute it and/or
   7 * modify it under the terms of the GNU General Public
   8 * License as published by the Free Software Foundation; either
   9 * version 2 of the License, or (at your option) any later version.
  10 *
  11 * This program is distributed in the hope that it will be useful,
  12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14 * General Public License for more details.
  15 *
  16 * You should have received a copy of the GNU General Public
  17 * License along with this program; if not, write to the
  18 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  19 * Boston, MA 021110-1307, USA.
  20 */
  21
  22#include <linux/fs.h>
  23#include <linux/slab.h>
  24#include <linux/highmem.h>
  25#include <linux/pagemap.h>
  26#include <asm/byteorder.h>
  27#include <linux/swap.h>
  28#include <linux/pipe_fs_i.h>
  29#include <linux/mpage.h>
  30#include <linux/quotaops.h>
  31#include <linux/blkdev.h>
  32#include <linux/uio.h>
 
  33
  34#include <cluster/masklog.h>
  35
  36#include "ocfs2.h"
  37
  38#include "alloc.h"
  39#include "aops.h"
  40#include "dlmglue.h"
  41#include "extent_map.h"
  42#include "file.h"
  43#include "inode.h"
  44#include "journal.h"
  45#include "suballoc.h"
  46#include "super.h"
  47#include "symlink.h"
  48#include "refcounttree.h"
  49#include "ocfs2_trace.h"
  50
  51#include "buffer_head_io.h"
  52#include "dir.h"
  53#include "namei.h"
  54#include "sysfile.h"
  55
  56static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
  57				   struct buffer_head *bh_result, int create)
  58{
  59	int err = -EIO;
  60	int status;
  61	struct ocfs2_dinode *fe = NULL;
  62	struct buffer_head *bh = NULL;
  63	struct buffer_head *buffer_cache_bh = NULL;
  64	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  65	void *kaddr;
  66
  67	trace_ocfs2_symlink_get_block(
  68			(unsigned long long)OCFS2_I(inode)->ip_blkno,
  69			(unsigned long long)iblock, bh_result, create);
  70
  71	BUG_ON(ocfs2_inode_is_fast_symlink(inode));
  72
  73	if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
  74		mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
  75		     (unsigned long long)iblock);
  76		goto bail;
  77	}
  78
  79	status = ocfs2_read_inode_block(inode, &bh);
  80	if (status < 0) {
  81		mlog_errno(status);
  82		goto bail;
  83	}
  84	fe = (struct ocfs2_dinode *) bh->b_data;
  85
  86	if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
  87						    le32_to_cpu(fe->i_clusters))) {
  88		err = -ENOMEM;
  89		mlog(ML_ERROR, "block offset is outside the allocated size: "
  90		     "%llu\n", (unsigned long long)iblock);
  91		goto bail;
  92	}
  93
  94	/* We don't use the page cache to create symlink data, so if
  95	 * need be, copy it over from the buffer cache. */
  96	if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
  97		u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
  98			    iblock;
  99		buffer_cache_bh = sb_getblk(osb->sb, blkno);
 100		if (!buffer_cache_bh) {
 101			err = -ENOMEM;
 102			mlog(ML_ERROR, "couldn't getblock for symlink!\n");
 103			goto bail;
 104		}
 105
 106		/* we haven't locked out transactions, so a commit
 107		 * could've happened. Since we've got a reference on
 108		 * the bh, even if it commits while we're doing the
 109		 * copy, the data is still good. */
 110		if (buffer_jbd(buffer_cache_bh)
 111		    && ocfs2_inode_is_new(inode)) {
 112			kaddr = kmap_atomic(bh_result->b_page);
 113			if (!kaddr) {
 114				mlog(ML_ERROR, "couldn't kmap!\n");
 115				goto bail;
 116			}
 117			memcpy(kaddr + (bh_result->b_size * iblock),
 118			       buffer_cache_bh->b_data,
 119			       bh_result->b_size);
 120			kunmap_atomic(kaddr);
 121			set_buffer_uptodate(bh_result);
 122		}
 123		brelse(buffer_cache_bh);
 124	}
 125
 126	map_bh(bh_result, inode->i_sb,
 127	       le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
 128
 129	err = 0;
 130
 131bail:
 132	brelse(bh);
 133
 134	return err;
 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(&OCFS2_I(inode)->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 = list_entry(pages->prev, struct page, lru);
 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	/* We don't need to lock journal system files, since they aren't
 468	 * accessed concurrently from multiple nodes.
 469	 */
 470	if (!INODE_JOURNAL(inode)) {
 471		err = ocfs2_inode_lock(inode, NULL, 0);
 472		if (err) {
 473			if (err != -ENOENT)
 474				mlog_errno(err);
 475			goto bail;
 476		}
 477		down_read(&OCFS2_I(inode)->ip_alloc_sem);
 478	}
 479
 480	if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
 481		err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
 482						  NULL);
 483
 484	if (!INODE_JOURNAL(inode)) {
 485		up_read(&OCFS2_I(inode)->ip_alloc_sem);
 486		ocfs2_inode_unlock(inode, 0);
 487	}
 488
 489	if (err) {
 490		mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
 491		     (unsigned long long)block);
 492		mlog_errno(err);
 493		goto bail;
 494	}
 495
 496bail:
 497	status = err ? 0 : p_blkno;
 498
 499	return status;
 500}
 501
 502static int ocfs2_releasepage(struct page *page, gfp_t wait)
 503{
 504	if (!page_has_buffers(page))
 505		return 0;
 506	return try_to_free_buffers(page);
 507}
 508
 509static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
 510					    u32 cpos,
 511					    unsigned int *start,
 512					    unsigned int *end)
 513{
 514	unsigned int cluster_start = 0, cluster_end = PAGE_SIZE;
 515
 516	if (unlikely(PAGE_SHIFT > osb->s_clustersize_bits)) {
 517		unsigned int cpp;
 518
 519		cpp = 1 << (PAGE_SHIFT - osb->s_clustersize_bits);
 520
 521		cluster_start = cpos % cpp;
 522		cluster_start = cluster_start << osb->s_clustersize_bits;
 523
 524		cluster_end = cluster_start + osb->s_clustersize;
 525	}
 526
 527	BUG_ON(cluster_start > PAGE_SIZE);
 528	BUG_ON(cluster_end > PAGE_SIZE);
 529
 530	if (start)
 531		*start = cluster_start;
 532	if (end)
 533		*end = cluster_end;
 534}
 535
 536/*
 537 * 'from' and 'to' are the region in the page to avoid zeroing.
 538 *
 539 * If pagesize > clustersize, this function will avoid zeroing outside
 540 * of the cluster boundary.
 541 *
 542 * from == to == 0 is code for "zero the entire cluster region"
 543 */
 544static void ocfs2_clear_page_regions(struct page *page,
 545				     struct ocfs2_super *osb, u32 cpos,
 546				     unsigned from, unsigned to)
 547{
 548	void *kaddr;
 549	unsigned int cluster_start, cluster_end;
 550
 551	ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
 552
 553	kaddr = kmap_atomic(page);
 554
 555	if (from || to) {
 556		if (from > cluster_start)
 557			memset(kaddr + cluster_start, 0, from - cluster_start);
 558		if (to < cluster_end)
 559			memset(kaddr + to, 0, cluster_end - to);
 560	} else {
 561		memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
 562	}
 563
 564	kunmap_atomic(kaddr);
 565}
 566
 567/*
 568 * Nonsparse file systems fully allocate before we get to the write
 569 * code. This prevents ocfs2_write() from tagging the write as an
 570 * allocating one, which means ocfs2_map_page_blocks() might try to
 571 * read-in the blocks at the tail of our file. Avoid reading them by
 572 * testing i_size against each block offset.
 573 */
 574static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
 575				 unsigned int block_start)
 576{
 577	u64 offset = page_offset(page) + block_start;
 578
 579	if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 580		return 1;
 581
 582	if (i_size_read(inode) > offset)
 583		return 1;
 584
 585	return 0;
 586}
 587
 588/*
 589 * Some of this taken from __block_write_begin(). We already have our
 590 * mapping by now though, and the entire write will be allocating or
 591 * it won't, so not much need to use BH_New.
 592 *
 593 * This will also skip zeroing, which is handled externally.
 594 */
 595int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
 596			  struct inode *inode, unsigned int from,
 597			  unsigned int to, int new)
 598{
 599	int ret = 0;
 600	struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
 601	unsigned int block_end, block_start;
 602	unsigned int bsize = 1 << inode->i_blkbits;
 603
 604	if (!page_has_buffers(page))
 605		create_empty_buffers(page, bsize, 0);
 606
 607	head = page_buffers(page);
 608	for (bh = head, block_start = 0; bh != head || !block_start;
 609	     bh = bh->b_this_page, block_start += bsize) {
 610		block_end = block_start + bsize;
 611
 612		clear_buffer_new(bh);
 613
 614		/*
 615		 * Ignore blocks outside of our i/o range -
 616		 * they may belong to unallocated clusters.
 617		 */
 618		if (block_start >= to || block_end <= from) {
 619			if (PageUptodate(page))
 620				set_buffer_uptodate(bh);
 621			continue;
 622		}
 623
 624		/*
 625		 * For an allocating write with cluster size >= page
 626		 * size, we always write the entire page.
 627		 */
 628		if (new)
 629			set_buffer_new(bh);
 630
 631		if (!buffer_mapped(bh)) {
 632			map_bh(bh, inode->i_sb, *p_blkno);
 633			unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
 634		}
 635
 636		if (PageUptodate(page)) {
 637			if (!buffer_uptodate(bh))
 638				set_buffer_uptodate(bh);
 639		} else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
 640			   !buffer_new(bh) &&
 641			   ocfs2_should_read_blk(inode, page, block_start) &&
 642			   (block_start < from || block_end > to)) {
 643			ll_rw_block(READ, 1, &bh);
 644			*wait_bh++=bh;
 645		}
 646
 647		*p_blkno = *p_blkno + 1;
 648	}
 649
 650	/*
 651	 * If we issued read requests - let them complete.
 652	 */
 653	while(wait_bh > wait) {
 654		wait_on_buffer(*--wait_bh);
 655		if (!buffer_uptodate(*wait_bh))
 656			ret = -EIO;
 657	}
 658
 659	if (ret == 0 || !new)
 660		return ret;
 661
 662	/*
 663	 * If we get -EIO above, zero out any newly allocated blocks
 664	 * to avoid exposing stale data.
 665	 */
 666	bh = head;
 667	block_start = 0;
 668	do {
 669		block_end = block_start + bsize;
 670		if (block_end <= from)
 671			goto next_bh;
 672		if (block_start >= to)
 673			break;
 674
 675		zero_user(page, block_start, bh->b_size);
 676		set_buffer_uptodate(bh);
 677		mark_buffer_dirty(bh);
 678
 679next_bh:
 680		block_start = block_end;
 681		bh = bh->b_this_page;
 682	} while (bh != head);
 683
 684	return ret;
 685}
 686
 687#if (PAGE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
 688#define OCFS2_MAX_CTXT_PAGES	1
 689#else
 690#define OCFS2_MAX_CTXT_PAGES	(OCFS2_MAX_CLUSTERSIZE / PAGE_SIZE)
 691#endif
 692
 693#define OCFS2_MAX_CLUSTERS_PER_PAGE	(PAGE_SIZE / OCFS2_MIN_CLUSTERSIZE)
 694
 695struct ocfs2_unwritten_extent {
 696	struct list_head	ue_node;
 697	struct list_head	ue_ip_node;
 698	u32			ue_cpos;
 699	u32			ue_phys;
 700};
 701
 702/*
 703 * Describe the state of a single cluster to be written to.
 704 */
 705struct ocfs2_write_cluster_desc {
 706	u32		c_cpos;
 707	u32		c_phys;
 708	/*
 709	 * Give this a unique field because c_phys eventually gets
 710	 * filled.
 711	 */
 712	unsigned	c_new;
 713	unsigned	c_clear_unwritten;
 714	unsigned	c_needs_zero;
 715};
 716
 717struct ocfs2_write_ctxt {
 718	/* Logical cluster position / len of write */
 719	u32				w_cpos;
 720	u32				w_clen;
 721
 722	/* First cluster allocated in a nonsparse extend */
 723	u32				w_first_new_cpos;
 724
 725	/* Type of caller. Must be one of buffer, mmap, direct.  */
 726	ocfs2_write_type_t		w_type;
 727
 728	struct ocfs2_write_cluster_desc	w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
 729
 730	/*
 731	 * This is true if page_size > cluster_size.
 732	 *
 733	 * It triggers a set of special cases during write which might
 734	 * have to deal with allocating writes to partial pages.
 735	 */
 736	unsigned int			w_large_pages;
 737
 738	/*
 739	 * Pages involved in this write.
 740	 *
 741	 * w_target_page is the page being written to by the user.
 742	 *
 743	 * w_pages is an array of pages which always contains
 744	 * w_target_page, and in the case of an allocating write with
 745	 * page_size < cluster size, it will contain zero'd and mapped
 746	 * pages adjacent to w_target_page which need to be written
 747	 * out in so that future reads from that region will get
 748	 * zero's.
 749	 */
 750	unsigned int			w_num_pages;
 751	struct page			*w_pages[OCFS2_MAX_CTXT_PAGES];
 752	struct page			*w_target_page;
 753
 754	/*
 755	 * w_target_locked is used for page_mkwrite path indicating no unlocking
 756	 * against w_target_page in ocfs2_write_end_nolock.
 757	 */
 758	unsigned int			w_target_locked:1;
 759
 760	/*
 761	 * ocfs2_write_end() uses this to know what the real range to
 762	 * write in the target should be.
 763	 */
 764	unsigned int			w_target_from;
 765	unsigned int			w_target_to;
 766
 767	/*
 768	 * We could use journal_current_handle() but this is cleaner,
 769	 * IMHO -Mark
 770	 */
 771	handle_t			*w_handle;
 772
 773	struct buffer_head		*w_di_bh;
 774
 775	struct ocfs2_cached_dealloc_ctxt w_dealloc;
 776
 777	struct list_head		w_unwritten_list;
 
 778};
 779
 780void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
 781{
 782	int i;
 783
 784	for(i = 0; i < num_pages; i++) {
 785		if (pages[i]) {
 786			unlock_page(pages[i]);
 787			mark_page_accessed(pages[i]);
 788			put_page(pages[i]);
 789		}
 790	}
 791}
 792
 793static void ocfs2_unlock_pages(struct ocfs2_write_ctxt *wc)
 794{
 795	int i;
 796
 797	/*
 798	 * w_target_locked is only set to true in the page_mkwrite() case.
 799	 * The intent is to allow us to lock the target page from write_begin()
 800	 * to write_end(). The caller must hold a ref on w_target_page.
 801	 */
 802	if (wc->w_target_locked) {
 803		BUG_ON(!wc->w_target_page);
 804		for (i = 0; i < wc->w_num_pages; i++) {
 805			if (wc->w_target_page == wc->w_pages[i]) {
 806				wc->w_pages[i] = NULL;
 807				break;
 808			}
 809		}
 810		mark_page_accessed(wc->w_target_page);
 811		put_page(wc->w_target_page);
 812	}
 813	ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
 814}
 815
 816static void ocfs2_free_unwritten_list(struct inode *inode,
 817				 struct list_head *head)
 818{
 819	struct ocfs2_inode_info *oi = OCFS2_I(inode);
 820	struct ocfs2_unwritten_extent *ue = NULL, *tmp = NULL;
 821
 822	list_for_each_entry_safe(ue, tmp, head, ue_node) {
 823		list_del(&ue->ue_node);
 824		spin_lock(&oi->ip_lock);
 825		list_del(&ue->ue_ip_node);
 826		spin_unlock(&oi->ip_lock);
 827		kfree(ue);
 828	}
 829}
 830
 831static void ocfs2_free_write_ctxt(struct inode *inode,
 832				  struct ocfs2_write_ctxt *wc)
 833{
 834	ocfs2_free_unwritten_list(inode, &wc->w_unwritten_list);
 835	ocfs2_unlock_pages(wc);
 836	brelse(wc->w_di_bh);
 837	kfree(wc);
 838}
 839
 840static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
 841				  struct ocfs2_super *osb, loff_t pos,
 842				  unsigned len, ocfs2_write_type_t type,
 843				  struct buffer_head *di_bh)
 844{
 845	u32 cend;
 846	struct ocfs2_write_ctxt *wc;
 847
 848	wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
 849	if (!wc)
 850		return -ENOMEM;
 851
 852	wc->w_cpos = pos >> osb->s_clustersize_bits;
 853	wc->w_first_new_cpos = UINT_MAX;
 854	cend = (pos + len - 1) >> osb->s_clustersize_bits;
 855	wc->w_clen = cend - wc->w_cpos + 1;
 856	get_bh(di_bh);
 857	wc->w_di_bh = di_bh;
 858	wc->w_type = type;
 859
 860	if (unlikely(PAGE_SHIFT > osb->s_clustersize_bits))
 861		wc->w_large_pages = 1;
 862	else
 863		wc->w_large_pages = 0;
 864
 865	ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
 866	INIT_LIST_HEAD(&wc->w_unwritten_list);
 867
 868	*wcp = wc;
 869
 870	return 0;
 871}
 872
 873/*
 874 * If a page has any new buffers, zero them out here, and mark them uptodate
 875 * and dirty so they'll be written out (in order to prevent uninitialised
 876 * block data from leaking). And clear the new bit.
 877 */
 878static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
 879{
 880	unsigned int block_start, block_end;
 881	struct buffer_head *head, *bh;
 882
 883	BUG_ON(!PageLocked(page));
 884	if (!page_has_buffers(page))
 885		return;
 886
 887	bh = head = page_buffers(page);
 888	block_start = 0;
 889	do {
 890		block_end = block_start + bh->b_size;
 891
 892		if (buffer_new(bh)) {
 893			if (block_end > from && block_start < to) {
 894				if (!PageUptodate(page)) {
 895					unsigned start, end;
 896
 897					start = max(from, block_start);
 898					end = min(to, block_end);
 899
 900					zero_user_segment(page, start, end);
 901					set_buffer_uptodate(bh);
 902				}
 903
 904				clear_buffer_new(bh);
 905				mark_buffer_dirty(bh);
 906			}
 907		}
 908
 909		block_start = block_end;
 910		bh = bh->b_this_page;
 911	} while (bh != head);
 912}
 913
 914/*
 915 * Only called when we have a failure during allocating write to write
 916 * zero's to the newly allocated region.
 917 */
 918static void ocfs2_write_failure(struct inode *inode,
 919				struct ocfs2_write_ctxt *wc,
 920				loff_t user_pos, unsigned user_len)
 921{
 922	int i;
 923	unsigned from = user_pos & (PAGE_SIZE - 1),
 924		to = user_pos + user_len;
 925	struct page *tmppage;
 926
 927	if (wc->w_target_page)
 928		ocfs2_zero_new_buffers(wc->w_target_page, from, to);
 929
 930	for(i = 0; i < wc->w_num_pages; i++) {
 931		tmppage = wc->w_pages[i];
 932
 933		if (tmppage && page_has_buffers(tmppage)) {
 934			if (ocfs2_should_order_data(inode))
 935				ocfs2_jbd2_file_inode(wc->w_handle, inode);
 
 936
 937			block_commit_write(tmppage, from, to);
 938		}
 939	}
 940}
 941
 942static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
 943					struct ocfs2_write_ctxt *wc,
 944					struct page *page, u32 cpos,
 945					loff_t user_pos, unsigned user_len,
 946					int new)
 947{
 948	int ret;
 949	unsigned int map_from = 0, map_to = 0;
 950	unsigned int cluster_start, cluster_end;
 951	unsigned int user_data_from = 0, user_data_to = 0;
 952
 953	ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
 954					&cluster_start, &cluster_end);
 955
 956	/* treat the write as new if the a hole/lseek spanned across
 957	 * the page boundary.
 958	 */
 959	new = new | ((i_size_read(inode) <= page_offset(page)) &&
 960			(page_offset(page) <= user_pos));
 961
 962	if (page == wc->w_target_page) {
 963		map_from = user_pos & (PAGE_SIZE - 1);
 964		map_to = map_from + user_len;
 965
 966		if (new)
 967			ret = ocfs2_map_page_blocks(page, p_blkno, inode,
 968						    cluster_start, cluster_end,
 969						    new);
 970		else
 971			ret = ocfs2_map_page_blocks(page, p_blkno, inode,
 972						    map_from, map_to, new);
 973		if (ret) {
 974			mlog_errno(ret);
 975			goto out;
 976		}
 977
 978		user_data_from = map_from;
 979		user_data_to = map_to;
 980		if (new) {
 981			map_from = cluster_start;
 982			map_to = cluster_end;
 983		}
 984	} else {
 985		/*
 986		 * If we haven't allocated the new page yet, we
 987		 * shouldn't be writing it out without copying user
 988		 * data. This is likely a math error from the caller.
 989		 */
 990		BUG_ON(!new);
 991
 992		map_from = cluster_start;
 993		map_to = cluster_end;
 994
 995		ret = ocfs2_map_page_blocks(page, p_blkno, inode,
 996					    cluster_start, cluster_end, new);
 997		if (ret) {
 998			mlog_errno(ret);
 999			goto out;
1000		}
1001	}
1002
1003	/*
1004	 * Parts of newly allocated pages need to be zero'd.
1005	 *
1006	 * Above, we have also rewritten 'to' and 'from' - as far as
1007	 * the rest of the function is concerned, the entire cluster
1008	 * range inside of a page needs to be written.
1009	 *
1010	 * We can skip this if the page is up to date - it's already
1011	 * been zero'd from being read in as a hole.
1012	 */
1013	if (new && !PageUptodate(page))
1014		ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1015					 cpos, user_data_from, user_data_to);
1016
1017	flush_dcache_page(page);
1018
1019out:
1020	return ret;
1021}
1022
1023/*
1024 * This function will only grab one clusters worth of pages.
1025 */
1026static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1027				      struct ocfs2_write_ctxt *wc,
1028				      u32 cpos, loff_t user_pos,
1029				      unsigned user_len, int new,
1030				      struct page *mmap_page)
1031{
1032	int ret = 0, i;
1033	unsigned long start, target_index, end_index, index;
1034	struct inode *inode = mapping->host;
1035	loff_t last_byte;
1036
1037	target_index = user_pos >> PAGE_SHIFT;
1038
1039	/*
1040	 * Figure out how many pages we'll be manipulating here. For
1041	 * non allocating write, we just change the one
1042	 * page. Otherwise, we'll need a whole clusters worth.  If we're
1043	 * writing past i_size, we only need enough pages to cover the
1044	 * last page of the write.
1045	 */
1046	if (new) {
1047		wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1048		start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1049		/*
1050		 * We need the index *past* the last page we could possibly
1051		 * touch.  This is the page past the end of the write or
1052		 * i_size, whichever is greater.
1053		 */
1054		last_byte = max(user_pos + user_len, i_size_read(inode));
1055		BUG_ON(last_byte < 1);
1056		end_index = ((last_byte - 1) >> PAGE_SHIFT) + 1;
1057		if ((start + wc->w_num_pages) > end_index)
1058			wc->w_num_pages = end_index - start;
1059	} else {
1060		wc->w_num_pages = 1;
1061		start = target_index;
1062	}
1063	end_index = (user_pos + user_len - 1) >> PAGE_SHIFT;
1064
1065	for(i = 0; i < wc->w_num_pages; i++) {
1066		index = start + i;
1067
1068		if (index >= target_index && index <= end_index &&
1069		    wc->w_type == OCFS2_WRITE_MMAP) {
1070			/*
1071			 * ocfs2_pagemkwrite() is a little different
1072			 * and wants us to directly use the page
1073			 * passed in.
1074			 */
1075			lock_page(mmap_page);
1076
1077			/* Exit and let the caller retry */
1078			if (mmap_page->mapping != mapping) {
1079				WARN_ON(mmap_page->mapping);
1080				unlock_page(mmap_page);
1081				ret = -EAGAIN;
1082				goto out;
1083			}
1084
1085			get_page(mmap_page);
1086			wc->w_pages[i] = mmap_page;
1087			wc->w_target_locked = true;
1088		} else if (index >= target_index && index <= end_index &&
1089			   wc->w_type == OCFS2_WRITE_DIRECT) {
1090			/* Direct write has no mapping page. */
1091			wc->w_pages[i] = NULL;
1092			continue;
1093		} else {
1094			wc->w_pages[i] = find_or_create_page(mapping, index,
1095							     GFP_NOFS);
1096			if (!wc->w_pages[i]) {
1097				ret = -ENOMEM;
1098				mlog_errno(ret);
1099				goto out;
1100			}
1101		}
1102		wait_for_stable_page(wc->w_pages[i]);
1103
1104		if (index == target_index)
1105			wc->w_target_page = wc->w_pages[i];
1106	}
1107out:
1108	if (ret)
1109		wc->w_target_locked = false;
1110	return ret;
1111}
1112
1113/*
1114 * Prepare a single cluster for write one cluster into the file.
1115 */
1116static int ocfs2_write_cluster(struct address_space *mapping,
1117			       u32 *phys, unsigned int new,
1118			       unsigned int clear_unwritten,
1119			       unsigned int should_zero,
1120			       struct ocfs2_alloc_context *data_ac,
1121			       struct ocfs2_alloc_context *meta_ac,
1122			       struct ocfs2_write_ctxt *wc, u32 cpos,
1123			       loff_t user_pos, unsigned user_len)
1124{
1125	int ret, i;
1126	u64 p_blkno;
1127	struct inode *inode = mapping->host;
1128	struct ocfs2_extent_tree et;
1129	int bpc = ocfs2_clusters_to_blocks(inode->i_sb, 1);
1130
1131	if (new) {
1132		u32 tmp_pos;
1133
1134		/*
1135		 * This is safe to call with the page locks - it won't take
1136		 * any additional semaphores or cluster locks.
1137		 */
1138		tmp_pos = cpos;
1139		ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1140					   &tmp_pos, 1, !clear_unwritten,
1141					   wc->w_di_bh, wc->w_handle,
1142					   data_ac, meta_ac, NULL);
1143		/*
1144		 * This shouldn't happen because we must have already
1145		 * calculated the correct meta data allocation required. The
1146		 * internal tree allocation code should know how to increase
1147		 * transaction credits itself.
1148		 *
1149		 * If need be, we could handle -EAGAIN for a
1150		 * RESTART_TRANS here.
1151		 */
1152		mlog_bug_on_msg(ret == -EAGAIN,
1153				"Inode %llu: EAGAIN return during allocation.\n",
1154				(unsigned long long)OCFS2_I(inode)->ip_blkno);
1155		if (ret < 0) {
1156			mlog_errno(ret);
1157			goto out;
1158		}
1159	} else if (clear_unwritten) {
1160		ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1161					      wc->w_di_bh);
1162		ret = ocfs2_mark_extent_written(inode, &et,
1163						wc->w_handle, cpos, 1, *phys,
1164						meta_ac, &wc->w_dealloc);
1165		if (ret < 0) {
1166			mlog_errno(ret);
1167			goto out;
1168		}
1169	}
1170
1171	/*
1172	 * The only reason this should fail is due to an inability to
1173	 * find the extent added.
1174	 */
1175	ret = ocfs2_get_clusters(inode, cpos, phys, NULL, NULL);
1176	if (ret < 0) {
1177		mlog(ML_ERROR, "Get physical blkno failed for inode %llu, "
1178			    "at logical cluster %u",
1179			    (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
1180		goto out;
1181	}
1182
1183	BUG_ON(*phys == 0);
1184
1185	p_blkno = ocfs2_clusters_to_blocks(inode->i_sb, *phys);
1186	if (!should_zero)
1187		p_blkno += (user_pos >> inode->i_sb->s_blocksize_bits) & (u64)(bpc - 1);
1188
1189	for(i = 0; i < wc->w_num_pages; i++) {
1190		int tmpret;
1191
1192		/* This is the direct io target page. */
1193		if (wc->w_pages[i] == NULL) {
1194			p_blkno++;
1195			continue;
1196		}
1197
1198		tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1199						      wc->w_pages[i], cpos,
1200						      user_pos, user_len,
1201						      should_zero);
1202		if (tmpret) {
1203			mlog_errno(tmpret);
1204			if (ret == 0)
1205				ret = tmpret;
1206		}
1207	}
1208
1209	/*
1210	 * We only have cleanup to do in case of allocating write.
1211	 */
1212	if (ret && new)
1213		ocfs2_write_failure(inode, wc, user_pos, user_len);
1214
1215out:
1216
1217	return ret;
1218}
1219
1220static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1221				       struct ocfs2_alloc_context *data_ac,
1222				       struct ocfs2_alloc_context *meta_ac,
1223				       struct ocfs2_write_ctxt *wc,
1224				       loff_t pos, unsigned len)
1225{
1226	int ret, i;
1227	loff_t cluster_off;
1228	unsigned int local_len = len;
1229	struct ocfs2_write_cluster_desc *desc;
1230	struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1231
1232	for (i = 0; i < wc->w_clen; i++) {
1233		desc = &wc->w_desc[i];
1234
1235		/*
1236		 * We have to make sure that the total write passed in
1237		 * doesn't extend past a single cluster.
1238		 */
1239		local_len = len;
1240		cluster_off = pos & (osb->s_clustersize - 1);
1241		if ((cluster_off + local_len) > osb->s_clustersize)
1242			local_len = osb->s_clustersize - cluster_off;
1243
1244		ret = ocfs2_write_cluster(mapping, &desc->c_phys,
1245					  desc->c_new,
1246					  desc->c_clear_unwritten,
1247					  desc->c_needs_zero,
1248					  data_ac, meta_ac,
1249					  wc, desc->c_cpos, pos, local_len);
1250		if (ret) {
1251			mlog_errno(ret);
1252			goto out;
1253		}
1254
1255		len -= local_len;
1256		pos += local_len;
1257	}
1258
1259	ret = 0;
1260out:
1261	return ret;
1262}
1263
1264/*
1265 * ocfs2_write_end() wants to know which parts of the target page it
1266 * should complete the write on. It's easiest to compute them ahead of
1267 * time when a more complete view of the write is available.
1268 */
1269static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1270					struct ocfs2_write_ctxt *wc,
1271					loff_t pos, unsigned len, int alloc)
1272{
1273	struct ocfs2_write_cluster_desc *desc;
1274
1275	wc->w_target_from = pos & (PAGE_SIZE - 1);
1276	wc->w_target_to = wc->w_target_from + len;
1277
1278	if (alloc == 0)
1279		return;
1280
1281	/*
1282	 * Allocating write - we may have different boundaries based
1283	 * on page size and cluster size.
1284	 *
1285	 * NOTE: We can no longer compute one value from the other as
1286	 * the actual write length and user provided length may be
1287	 * different.
1288	 */
1289
1290	if (wc->w_large_pages) {
1291		/*
1292		 * We only care about the 1st and last cluster within
1293		 * our range and whether they should be zero'd or not. Either
1294		 * value may be extended out to the start/end of a
1295		 * newly allocated cluster.
1296		 */
1297		desc = &wc->w_desc[0];
1298		if (desc->c_needs_zero)
1299			ocfs2_figure_cluster_boundaries(osb,
1300							desc->c_cpos,
1301							&wc->w_target_from,
1302							NULL);
1303
1304		desc = &wc->w_desc[wc->w_clen - 1];
1305		if (desc->c_needs_zero)
1306			ocfs2_figure_cluster_boundaries(osb,
1307							desc->c_cpos,
1308							NULL,
1309							&wc->w_target_to);
1310	} else {
1311		wc->w_target_from = 0;
1312		wc->w_target_to = PAGE_SIZE;
1313	}
1314}
1315
1316/*
1317 * Check if this extent is marked UNWRITTEN by direct io. If so, we need not to
1318 * do the zero work. And should not to clear UNWRITTEN since it will be cleared
1319 * by the direct io procedure.
1320 * If this is a new extent that allocated by direct io, we should mark it in
1321 * the ip_unwritten_list.
1322 */
1323static int ocfs2_unwritten_check(struct inode *inode,
1324				 struct ocfs2_write_ctxt *wc,
1325				 struct ocfs2_write_cluster_desc *desc)
1326{
1327	struct ocfs2_inode_info *oi = OCFS2_I(inode);
1328	struct ocfs2_unwritten_extent *ue = NULL, *new = NULL;
1329	int ret = 0;
1330
1331	if (!desc->c_needs_zero)
1332		return 0;
1333
1334retry:
1335	spin_lock(&oi->ip_lock);
1336	/* Needs not to zero no metter buffer or direct. The one who is zero
1337	 * the cluster is doing zero. And he will clear unwritten after all
1338	 * cluster io finished. */
1339	list_for_each_entry(ue, &oi->ip_unwritten_list, ue_ip_node) {
1340		if (desc->c_cpos == ue->ue_cpos) {
1341			BUG_ON(desc->c_new);
1342			desc->c_needs_zero = 0;
1343			desc->c_clear_unwritten = 0;
1344			goto unlock;
1345		}
1346	}
1347
1348	if (wc->w_type != OCFS2_WRITE_DIRECT)
1349		goto unlock;
1350
1351	if (new == NULL) {
1352		spin_unlock(&oi->ip_lock);
1353		new = kmalloc(sizeof(struct ocfs2_unwritten_extent),
1354			     GFP_NOFS);
1355		if (new == NULL) {
1356			ret = -ENOMEM;
1357			goto out;
1358		}
1359		goto retry;
1360	}
1361	/* This direct write will doing zero. */
1362	new->ue_cpos = desc->c_cpos;
1363	new->ue_phys = desc->c_phys;
1364	desc->c_clear_unwritten = 0;
1365	list_add_tail(&new->ue_ip_node, &oi->ip_unwritten_list);
1366	list_add_tail(&new->ue_node, &wc->w_unwritten_list);
 
1367	new = NULL;
1368unlock:
1369	spin_unlock(&oi->ip_lock);
1370out:
1371	if (new)
1372		kfree(new);
1373	return ret;
1374}
1375
1376/*
1377 * Populate each single-cluster write descriptor in the write context
1378 * with information about the i/o to be done.
1379 *
1380 * Returns the number of clusters that will have to be allocated, as
1381 * well as a worst case estimate of the number of extent records that
1382 * would have to be created during a write to an unwritten region.
1383 */
1384static int ocfs2_populate_write_desc(struct inode *inode,
1385				     struct ocfs2_write_ctxt *wc,
1386				     unsigned int *clusters_to_alloc,
1387				     unsigned int *extents_to_split)
1388{
1389	int ret;
1390	struct ocfs2_write_cluster_desc *desc;
1391	unsigned int num_clusters = 0;
1392	unsigned int ext_flags = 0;
1393	u32 phys = 0;
1394	int i;
1395
1396	*clusters_to_alloc = 0;
1397	*extents_to_split = 0;
1398
1399	for (i = 0; i < wc->w_clen; i++) {
1400		desc = &wc->w_desc[i];
1401		desc->c_cpos = wc->w_cpos + i;
1402
1403		if (num_clusters == 0) {
1404			/*
1405			 * Need to look up the next extent record.
1406			 */
1407			ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1408						 &num_clusters, &ext_flags);
1409			if (ret) {
1410				mlog_errno(ret);
1411				goto out;
1412			}
1413
1414			/* We should already CoW the refcountd extent. */
1415			BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1416
1417			/*
1418			 * Assume worst case - that we're writing in
1419			 * the middle of the extent.
1420			 *
1421			 * We can assume that the write proceeds from
1422			 * left to right, in which case the extent
1423			 * insert code is smart enough to coalesce the
1424			 * next splits into the previous records created.
1425			 */
1426			if (ext_flags & OCFS2_EXT_UNWRITTEN)
1427				*extents_to_split = *extents_to_split + 2;
1428		} else if (phys) {
1429			/*
1430			 * Only increment phys if it doesn't describe
1431			 * a hole.
1432			 */
1433			phys++;
1434		}
1435
1436		/*
1437		 * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1438		 * file that got extended.  w_first_new_cpos tells us
1439		 * where the newly allocated clusters are so we can
1440		 * zero them.
1441		 */
1442		if (desc->c_cpos >= wc->w_first_new_cpos) {
1443			BUG_ON(phys == 0);
1444			desc->c_needs_zero = 1;
1445		}
1446
1447		desc->c_phys = phys;
1448		if (phys == 0) {
1449			desc->c_new = 1;
1450			desc->c_needs_zero = 1;
1451			desc->c_clear_unwritten = 1;
1452			*clusters_to_alloc = *clusters_to_alloc + 1;
1453		}
1454
1455		if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1456			desc->c_clear_unwritten = 1;
1457			desc->c_needs_zero = 1;
1458		}
1459
1460		ret = ocfs2_unwritten_check(inode, wc, desc);
1461		if (ret) {
1462			mlog_errno(ret);
1463			goto out;
1464		}
1465
1466		num_clusters--;
1467	}
1468
1469	ret = 0;
1470out:
1471	return ret;
1472}
1473
1474static int ocfs2_write_begin_inline(struct address_space *mapping,
1475				    struct inode *inode,
1476				    struct ocfs2_write_ctxt *wc)
1477{
1478	int ret;
1479	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1480	struct page *page;
1481	handle_t *handle;
1482	struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1483
1484	handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1485	if (IS_ERR(handle)) {
1486		ret = PTR_ERR(handle);
1487		mlog_errno(ret);
1488		goto out;
1489	}
1490
1491	page = find_or_create_page(mapping, 0, GFP_NOFS);
1492	if (!page) {
1493		ocfs2_commit_trans(osb, handle);
1494		ret = -ENOMEM;
1495		mlog_errno(ret);
1496		goto out;
1497	}
1498	/*
1499	 * If we don't set w_num_pages then this page won't get unlocked
1500	 * and freed on cleanup of the write context.
1501	 */
1502	wc->w_pages[0] = wc->w_target_page = page;
1503	wc->w_num_pages = 1;
1504
1505	ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1506				      OCFS2_JOURNAL_ACCESS_WRITE);
1507	if (ret) {
1508		ocfs2_commit_trans(osb, handle);
1509
1510		mlog_errno(ret);
1511		goto out;
1512	}
1513
1514	if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1515		ocfs2_set_inode_data_inline(inode, di);
1516
1517	if (!PageUptodate(page)) {
1518		ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1519		if (ret) {
1520			ocfs2_commit_trans(osb, handle);
1521
1522			goto out;
1523		}
1524	}
1525
1526	wc->w_handle = handle;
1527out:
1528	return ret;
1529}
1530
1531int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1532{
1533	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1534
1535	if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1536		return 1;
1537	return 0;
1538}
1539
1540static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1541					  struct inode *inode, loff_t pos,
1542					  unsigned len, struct page *mmap_page,
1543					  struct ocfs2_write_ctxt *wc)
1544{
1545	int ret, written = 0;
1546	loff_t end = pos + len;
1547	struct ocfs2_inode_info *oi = OCFS2_I(inode);
1548	struct ocfs2_dinode *di = NULL;
1549
1550	trace_ocfs2_try_to_write_inline_data((unsigned long long)oi->ip_blkno,
1551					     len, (unsigned long long)pos,
1552					     oi->ip_dyn_features);
1553
1554	/*
1555	 * Handle inodes which already have inline data 1st.
1556	 */
1557	if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1558		if (mmap_page == NULL &&
1559		    ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1560			goto do_inline_write;
1561
1562		/*
1563		 * The write won't fit - we have to give this inode an
1564		 * inline extent list now.
1565		 */
1566		ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1567		if (ret)
1568			mlog_errno(ret);
1569		goto out;
1570	}
1571
1572	/*
1573	 * Check whether the inode can accept inline data.
1574	 */
1575	if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1576		return 0;
1577
1578	/*
1579	 * Check whether the write can fit.
1580	 */
1581	di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1582	if (mmap_page ||
1583	    end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1584		return 0;
1585
1586do_inline_write:
1587	ret = ocfs2_write_begin_inline(mapping, inode, wc);
1588	if (ret) {
1589		mlog_errno(ret);
1590		goto out;
1591	}
1592
1593	/*
1594	 * This signals to the caller that the data can be written
1595	 * inline.
1596	 */
1597	written = 1;
1598out:
1599	return written ? written : ret;
1600}
1601
1602/*
1603 * This function only does anything for file systems which can't
1604 * handle sparse files.
1605 *
1606 * What we want to do here is fill in any hole between the current end
1607 * of allocation and the end of our write. That way the rest of the
1608 * write path can treat it as an non-allocating write, which has no
1609 * special case code for sparse/nonsparse files.
1610 */
1611static int ocfs2_expand_nonsparse_inode(struct inode *inode,
1612					struct buffer_head *di_bh,
1613					loff_t pos, unsigned len,
1614					struct ocfs2_write_ctxt *wc)
1615{
1616	int ret;
1617	loff_t newsize = pos + len;
1618
1619	BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1620
1621	if (newsize <= i_size_read(inode))
1622		return 0;
1623
1624	ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
1625	if (ret)
1626		mlog_errno(ret);
1627
1628	/* There is no wc if this is call from direct. */
1629	if (wc)
1630		wc->w_first_new_cpos =
1631			ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1632
1633	return ret;
1634}
1635
1636static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
1637			   loff_t pos)
1638{
1639	int ret = 0;
1640
1641	BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1642	if (pos > i_size_read(inode))
1643		ret = ocfs2_zero_extend(inode, di_bh, pos);
1644
1645	return ret;
1646}
1647
1648/*
1649 * Try to flush truncate logs if we can free enough clusters from it.
1650 * As for return value, "< 0" means error, "0" no space and "1" means
1651 * we have freed enough spaces and let the caller try to allocate again.
1652 */
1653static int ocfs2_try_to_free_truncate_log(struct ocfs2_super *osb,
1654					  unsigned int needed)
1655{
1656	tid_t target;
1657	int ret = 0;
1658	unsigned int truncated_clusters;
1659
1660	inode_lock(osb->osb_tl_inode);
1661	truncated_clusters = osb->truncated_clusters;
1662	inode_unlock(osb->osb_tl_inode);
1663
1664	/*
1665	 * Check whether we can succeed in allocating if we free
1666	 * the truncate log.
1667	 */
1668	if (truncated_clusters < needed)
1669		goto out;
1670
1671	ret = ocfs2_flush_truncate_log(osb);
1672	if (ret) {
1673		mlog_errno(ret);
1674		goto out;
1675	}
1676
1677	if (jbd2_journal_start_commit(osb->journal->j_journal, &target)) {
1678		jbd2_log_wait_commit(osb->journal->j_journal, target);
1679		ret = 1;
1680	}
1681out:
1682	return ret;
1683}
1684
1685int ocfs2_write_begin_nolock(struct address_space *mapping,
1686			     loff_t pos, unsigned len, ocfs2_write_type_t type,
1687			     struct page **pagep, void **fsdata,
1688			     struct buffer_head *di_bh, struct page *mmap_page)
1689{
1690	int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1691	unsigned int clusters_to_alloc, extents_to_split, clusters_need = 0;
1692	struct ocfs2_write_ctxt *wc;
1693	struct inode *inode = mapping->host;
1694	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1695	struct ocfs2_dinode *di;
1696	struct ocfs2_alloc_context *data_ac = NULL;
1697	struct ocfs2_alloc_context *meta_ac = NULL;
1698	handle_t *handle;
1699	struct ocfs2_extent_tree et;
1700	int try_free = 1, ret1;
1701
1702try_again:
1703	ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, type, di_bh);
1704	if (ret) {
1705		mlog_errno(ret);
1706		return ret;
1707	}
1708
1709	if (ocfs2_supports_inline_data(osb)) {
1710		ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1711						     mmap_page, wc);
1712		if (ret == 1) {
1713			ret = 0;
1714			goto success;
1715		}
1716		if (ret < 0) {
1717			mlog_errno(ret);
1718			goto out;
1719		}
1720	}
1721
1722	/* Direct io change i_size late, should not zero tail here. */
1723	if (type != OCFS2_WRITE_DIRECT) {
1724		if (ocfs2_sparse_alloc(osb))
1725			ret = ocfs2_zero_tail(inode, di_bh, pos);
1726		else
1727			ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos,
1728							   len, wc);
1729		if (ret) {
1730			mlog_errno(ret);
1731			goto out;
1732		}
1733	}
1734
1735	ret = ocfs2_check_range_for_refcount(inode, pos, len);
1736	if (ret < 0) {
1737		mlog_errno(ret);
1738		goto out;
1739	} else if (ret == 1) {
1740		clusters_need = wc->w_clen;
1741		ret = ocfs2_refcount_cow(inode, di_bh,
1742					 wc->w_cpos, wc->w_clen, UINT_MAX);
1743		if (ret) {
1744			mlog_errno(ret);
1745			goto out;
1746		}
1747	}
1748
1749	ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1750					&extents_to_split);
1751	if (ret) {
1752		mlog_errno(ret);
1753		goto out;
1754	}
1755	clusters_need += clusters_to_alloc;
1756
1757	di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1758
1759	trace_ocfs2_write_begin_nolock(
1760			(unsigned long long)OCFS2_I(inode)->ip_blkno,
1761			(long long)i_size_read(inode),
1762			le32_to_cpu(di->i_clusters),
1763			pos, len, type, mmap_page,
1764			clusters_to_alloc, extents_to_split);
1765
1766	/*
1767	 * We set w_target_from, w_target_to here so that
1768	 * ocfs2_write_end() knows which range in the target page to
1769	 * write out. An allocation requires that we write the entire
1770	 * cluster range.
1771	 */
1772	if (clusters_to_alloc || extents_to_split) {
1773		/*
1774		 * XXX: We are stretching the limits of
1775		 * ocfs2_lock_allocators(). It greatly over-estimates
1776		 * the work to be done.
1777		 */
1778		ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1779					      wc->w_di_bh);
1780		ret = ocfs2_lock_allocators(inode, &et,
1781					    clusters_to_alloc, extents_to_split,
1782					    &data_ac, &meta_ac);
1783		if (ret) {
1784			mlog_errno(ret);
1785			goto out;
1786		}
1787
1788		if (data_ac)
1789			data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
1790
1791		credits = ocfs2_calc_extend_credits(inode->i_sb,
1792						    &di->id2.i_list);
1793	} else if (type == OCFS2_WRITE_DIRECT)
1794		/* direct write needs not to start trans if no extents alloc. */
1795		goto success;
1796
1797	/*
1798	 * We have to zero sparse allocated clusters, unwritten extent clusters,
1799	 * and non-sparse clusters we just extended.  For non-sparse writes,
1800	 * we know zeros will only be needed in the first and/or last cluster.
1801	 */
1802	if (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1803			   wc->w_desc[wc->w_clen - 1].c_needs_zero))
1804		cluster_of_pages = 1;
1805	else
1806		cluster_of_pages = 0;
1807
1808	ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1809
1810	handle = ocfs2_start_trans(osb, credits);
1811	if (IS_ERR(handle)) {
1812		ret = PTR_ERR(handle);
1813		mlog_errno(ret);
1814		goto out;
1815	}
1816
1817	wc->w_handle = handle;
1818
1819	if (clusters_to_alloc) {
1820		ret = dquot_alloc_space_nodirty(inode,
1821			ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1822		if (ret)
1823			goto out_commit;
1824	}
1825
1826	ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1827				      OCFS2_JOURNAL_ACCESS_WRITE);
1828	if (ret) {
1829		mlog_errno(ret);
1830		goto out_quota;
1831	}
1832
1833	/*
1834	 * Fill our page array first. That way we've grabbed enough so
1835	 * that we can zero and flush if we error after adding the
1836	 * extent.
1837	 */
1838	ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
1839					 cluster_of_pages, mmap_page);
1840	if (ret && ret != -EAGAIN) {
1841		mlog_errno(ret);
1842		goto out_quota;
1843	}
1844
1845	/*
1846	 * ocfs2_grab_pages_for_write() returns -EAGAIN if it could not lock
1847	 * the target page. In this case, we exit with no error and no target
1848	 * page. This will trigger the caller, page_mkwrite(), to re-try
1849	 * the operation.
1850	 */
1851	if (ret == -EAGAIN) {
1852		BUG_ON(wc->w_target_page);
1853		ret = 0;
1854		goto out_quota;
1855	}
1856
1857	ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1858					  len);
1859	if (ret) {
1860		mlog_errno(ret);
1861		goto out_quota;
1862	}
1863
1864	if (data_ac)
1865		ocfs2_free_alloc_context(data_ac);
1866	if (meta_ac)
1867		ocfs2_free_alloc_context(meta_ac);
1868
1869success:
1870	if (pagep)
1871		*pagep = wc->w_target_page;
1872	*fsdata = wc;
1873	return 0;
1874out_quota:
1875	if (clusters_to_alloc)
1876		dquot_free_space(inode,
1877			  ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1878out_commit:
1879	ocfs2_commit_trans(osb, handle);
1880
1881out:
 
 
 
 
 
 
 
 
 
 
1882	ocfs2_free_write_ctxt(inode, wc);
1883
1884	if (data_ac) {
1885		ocfs2_free_alloc_context(data_ac);
1886		data_ac = NULL;
1887	}
1888	if (meta_ac) {
1889		ocfs2_free_alloc_context(meta_ac);
1890		meta_ac = NULL;
1891	}
1892
1893	if (ret == -ENOSPC && try_free) {
1894		/*
1895		 * Try to free some truncate log so that we can have enough
1896		 * clusters to allocate.
1897		 */
1898		try_free = 0;
1899
1900		ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need);
1901		if (ret1 == 1)
1902			goto try_again;
1903
1904		if (ret1 < 0)
1905			mlog_errno(ret1);
1906	}
1907
1908	return ret;
1909}
1910
1911static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1912			     loff_t pos, unsigned len, unsigned flags,
1913			     struct page **pagep, void **fsdata)
1914{
1915	int ret;
1916	struct buffer_head *di_bh = NULL;
1917	struct inode *inode = mapping->host;
1918
1919	ret = ocfs2_inode_lock(inode, &di_bh, 1);
1920	if (ret) {
1921		mlog_errno(ret);
1922		return ret;
1923	}
1924
1925	/*
1926	 * Take alloc sem here to prevent concurrent lookups. That way
1927	 * the mapping, zeroing and tree manipulation within
1928	 * ocfs2_write() will be safe against ->readpage(). This
1929	 * should also serve to lock out allocation from a shared
1930	 * writeable region.
1931	 */
1932	down_write(&OCFS2_I(inode)->ip_alloc_sem);
1933
1934	ret = ocfs2_write_begin_nolock(mapping, pos, len, OCFS2_WRITE_BUFFER,
1935				       pagep, fsdata, di_bh, NULL);
1936	if (ret) {
1937		mlog_errno(ret);
1938		goto out_fail;
1939	}
1940
1941	brelse(di_bh);
1942
1943	return 0;
1944
1945out_fail:
1946	up_write(&OCFS2_I(inode)->ip_alloc_sem);
1947
1948	brelse(di_bh);
1949	ocfs2_inode_unlock(inode, 1);
1950
1951	return ret;
1952}
1953
1954static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1955				   unsigned len, unsigned *copied,
1956				   struct ocfs2_dinode *di,
1957				   struct ocfs2_write_ctxt *wc)
1958{
1959	void *kaddr;
1960
1961	if (unlikely(*copied < len)) {
1962		if (!PageUptodate(wc->w_target_page)) {
1963			*copied = 0;
1964			return;
1965		}
1966	}
1967
1968	kaddr = kmap_atomic(wc->w_target_page);
1969	memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1970	kunmap_atomic(kaddr);
1971
1972	trace_ocfs2_write_end_inline(
1973	     (unsigned long long)OCFS2_I(inode)->ip_blkno,
1974	     (unsigned long long)pos, *copied,
1975	     le16_to_cpu(di->id2.i_data.id_count),
1976	     le16_to_cpu(di->i_dyn_features));
1977}
1978
1979int ocfs2_write_end_nolock(struct address_space *mapping,
1980			   loff_t pos, unsigned len, unsigned copied,
1981			   struct page *page, void *fsdata)
1982{
1983	int i, ret;
1984	unsigned from, to, start = pos & (PAGE_SIZE - 1);
1985	struct inode *inode = mapping->host;
1986	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1987	struct ocfs2_write_ctxt *wc = fsdata;
1988	struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1989	handle_t *handle = wc->w_handle;
1990	struct page *tmppage;
1991
1992	BUG_ON(!list_empty(&wc->w_unwritten_list));
1993
1994	if (handle) {
1995		ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode),
1996				wc->w_di_bh, OCFS2_JOURNAL_ACCESS_WRITE);
1997		if (ret) {
1998			copied = ret;
1999			mlog_errno(ret);
2000			goto out;
2001		}
2002	}
2003
2004	if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
2005		ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
2006		goto out_write_size;
2007	}
2008
2009	if (unlikely(copied < len) && wc->w_target_page) {
2010		if (!PageUptodate(wc->w_target_page))
2011			copied = 0;
2012
2013		ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
2014				       start+len);
2015	}
2016	if (wc->w_target_page)
2017		flush_dcache_page(wc->w_target_page);
2018
2019	for(i = 0; i < wc->w_num_pages; i++) {
2020		tmppage = wc->w_pages[i];
2021
2022		/* This is the direct io target page. */
2023		if (tmppage == NULL)
2024			continue;
2025
2026		if (tmppage == wc->w_target_page) {
2027			from = wc->w_target_from;
2028			to = wc->w_target_to;
2029
2030			BUG_ON(from > PAGE_SIZE ||
2031			       to > PAGE_SIZE ||
2032			       to < from);
2033		} else {
2034			/*
2035			 * Pages adjacent to the target (if any) imply
2036			 * a hole-filling write in which case we want
2037			 * to flush their entire range.
2038			 */
2039			from = 0;
2040			to = PAGE_SIZE;
2041		}
2042
2043		if (page_has_buffers(tmppage)) {
2044			if (handle && ocfs2_should_order_data(inode))
2045				ocfs2_jbd2_file_inode(handle, inode);
 
 
 
 
 
 
2046			block_commit_write(tmppage, from, to);
2047		}
2048	}
2049
2050out_write_size:
2051	/* Direct io do not update i_size here. */
2052	if (wc->w_type != OCFS2_WRITE_DIRECT) {
2053		pos += copied;
2054		if (pos > i_size_read(inode)) {
2055			i_size_write(inode, pos);
2056			mark_inode_dirty(inode);
2057		}
2058		inode->i_blocks = ocfs2_inode_sector_count(inode);
2059		di->i_size = cpu_to_le64((u64)i_size_read(inode));
2060		inode->i_mtime = inode->i_ctime = CURRENT_TIME;
2061		di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
2062		di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
2063		ocfs2_update_inode_fsync_trans(handle, inode, 1);
 
2064	}
2065	if (handle)
2066		ocfs2_journal_dirty(handle, wc->w_di_bh);
2067
2068out:
2069	/* unlock pages before dealloc since it needs acquiring j_trans_barrier
2070	 * lock, or it will cause a deadlock since journal commit threads holds
2071	 * this lock and will ask for the page lock when flushing the data.
2072	 * put it here to preserve the unlock order.
2073	 */
2074	ocfs2_unlock_pages(wc);
2075
2076	if (handle)
2077		ocfs2_commit_trans(osb, handle);
2078
2079	ocfs2_run_deallocs(osb, &wc->w_dealloc);
2080
2081	brelse(wc->w_di_bh);
2082	kfree(wc);
2083
2084	return copied;
2085}
2086
2087static int ocfs2_write_end(struct file *file, struct address_space *mapping,
2088			   loff_t pos, unsigned len, unsigned copied,
2089			   struct page *page, void *fsdata)
2090{
2091	int ret;
2092	struct inode *inode = mapping->host;
2093
2094	ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
2095
2096	up_write(&OCFS2_I(inode)->ip_alloc_sem);
2097	ocfs2_inode_unlock(inode, 1);
2098
2099	return ret;
2100}
2101
2102struct ocfs2_dio_write_ctxt {
2103	struct list_head	dw_zero_list;
2104	unsigned		dw_zero_count;
2105	int			dw_orphaned;
2106	pid_t			dw_writer_pid;
2107};
2108
2109static struct ocfs2_dio_write_ctxt *
2110ocfs2_dio_alloc_write_ctx(struct buffer_head *bh, int *alloc)
2111{
2112	struct ocfs2_dio_write_ctxt *dwc = NULL;
2113
2114	if (bh->b_private)
2115		return bh->b_private;
2116
2117	dwc = kmalloc(sizeof(struct ocfs2_dio_write_ctxt), GFP_NOFS);
2118	if (dwc == NULL)
2119		return NULL;
2120	INIT_LIST_HEAD(&dwc->dw_zero_list);
2121	dwc->dw_zero_count = 0;
2122	dwc->dw_orphaned = 0;
2123	dwc->dw_writer_pid = task_pid_nr(current);
2124	bh->b_private = dwc;
2125	*alloc = 1;
2126
2127	return dwc;
2128}
2129
2130static void ocfs2_dio_free_write_ctx(struct inode *inode,
2131				     struct ocfs2_dio_write_ctxt *dwc)
2132{
2133	ocfs2_free_unwritten_list(inode, &dwc->dw_zero_list);
2134	kfree(dwc);
2135}
2136
2137/*
2138 * TODO: Make this into a generic get_blocks function.
2139 *
2140 * From do_direct_io in direct-io.c:
2141 *  "So what we do is to permit the ->get_blocks function to populate
2142 *   bh.b_size with the size of IO which is permitted at this offset and
2143 *   this i_blkbits."
2144 *
2145 * This function is called directly from get_more_blocks in direct-io.c.
2146 *
2147 * called like this: dio->get_blocks(dio->inode, fs_startblk,
2148 * 					fs_count, map_bh, dio->rw == WRITE);
2149 */
2150static int ocfs2_dio_get_block(struct inode *inode, sector_t iblock,
2151			       struct buffer_head *bh_result, int create)
2152{
2153	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2154	struct ocfs2_inode_info *oi = OCFS2_I(inode);
2155	struct ocfs2_write_ctxt *wc;
2156	struct ocfs2_write_cluster_desc *desc = NULL;
2157	struct ocfs2_dio_write_ctxt *dwc = NULL;
2158	struct buffer_head *di_bh = NULL;
2159	u64 p_blkno;
2160	loff_t pos = iblock << inode->i_sb->s_blocksize_bits;
 
 
2161	unsigned len, total_len = bh_result->b_size;
2162	int ret = 0, first_get_block = 0;
2163
2164	len = osb->s_clustersize - (pos & (osb->s_clustersize - 1));
2165	len = min(total_len, len);
2166
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2167	mlog(0, "get block of %lu at %llu:%u req %u\n",
2168			inode->i_ino, pos, len, total_len);
2169
2170	/*
2171	 * Because we need to change file size in ocfs2_dio_end_io_write(), or
2172	 * we may need to add it to orphan dir. So can not fall to fast path
2173	 * while file size will be changed.
2174	 */
2175	if (pos + total_len <= i_size_read(inode)) {
2176		down_read(&oi->ip_alloc_sem);
2177		/* This is the fast path for re-write. */
2178		ret = ocfs2_get_block(inode, iblock, bh_result, create);
2179
2180		up_read(&oi->ip_alloc_sem);
2181
 
 
2182		if (buffer_mapped(bh_result) &&
2183		    !buffer_new(bh_result) &&
2184		    ret == 0)
2185			goto out;
2186
2187		/* Clear state set by ocfs2_get_block. */
2188		bh_result->b_state = 0;
2189	}
2190
2191	dwc = ocfs2_dio_alloc_write_ctx(bh_result, &first_get_block);
2192	if (unlikely(dwc == NULL)) {
2193		ret = -ENOMEM;
2194		mlog_errno(ret);
2195		goto out;
2196	}
2197
2198	if (ocfs2_clusters_for_bytes(inode->i_sb, pos + total_len) >
2199	    ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode)) &&
2200	    !dwc->dw_orphaned) {
2201		/*
2202		 * when we are going to alloc extents beyond file size, add the
2203		 * inode to orphan dir, so we can recall those spaces when
2204		 * system crashed during write.
2205		 */
2206		ret = ocfs2_add_inode_to_orphan(osb, inode);
2207		if (ret < 0) {
2208			mlog_errno(ret);
2209			goto out;
2210		}
2211		dwc->dw_orphaned = 1;
2212	}
2213
2214	ret = ocfs2_inode_lock(inode, &di_bh, 1);
2215	if (ret) {
2216		mlog_errno(ret);
2217		goto out;
2218	}
2219
2220	down_write(&oi->ip_alloc_sem);
2221
2222	if (first_get_block) {
2223		if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
2224			ret = ocfs2_zero_tail(inode, di_bh, pos);
2225		else
2226			ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos,
2227							   total_len, NULL);
2228		if (ret < 0) {
2229			mlog_errno(ret);
2230			goto unlock;
2231		}
2232	}
2233
2234	ret = ocfs2_write_begin_nolock(inode->i_mapping, pos, len,
2235				       OCFS2_WRITE_DIRECT, NULL,
2236				       (void **)&wc, di_bh, NULL);
2237	if (ret) {
2238		mlog_errno(ret);
2239		goto unlock;
2240	}
2241
2242	desc = &wc->w_desc[0];
2243
2244	p_blkno = ocfs2_clusters_to_blocks(inode->i_sb, desc->c_phys);
2245	BUG_ON(p_blkno == 0);
2246	p_blkno += iblock & (u64)(ocfs2_clusters_to_blocks(inode->i_sb, 1) - 1);
2247
2248	map_bh(bh_result, inode->i_sb, p_blkno);
2249	bh_result->b_size = len;
2250	if (desc->c_needs_zero)
2251		set_buffer_new(bh_result);
2252
 
 
 
2253	/* May sleep in end_io. It should not happen in a irq context. So defer
2254	 * it to dio work queue. */
2255	set_buffer_defer_completion(bh_result);
2256
2257	if (!list_empty(&wc->w_unwritten_list)) {
2258		struct ocfs2_unwritten_extent *ue = NULL;
2259
2260		ue = list_first_entry(&wc->w_unwritten_list,
2261				      struct ocfs2_unwritten_extent,
2262				      ue_node);
2263		BUG_ON(ue->ue_cpos != desc->c_cpos);
2264		/* The physical address may be 0, fill it. */
2265		ue->ue_phys = desc->c_phys;
2266
2267		list_splice_tail_init(&wc->w_unwritten_list, &dwc->dw_zero_list);
2268		dwc->dw_zero_count++;
2269	}
2270
2271	ret = ocfs2_write_end_nolock(inode->i_mapping, pos, len, len, NULL, wc);
2272	BUG_ON(ret != len);
2273	ret = 0;
2274unlock:
2275	up_write(&oi->ip_alloc_sem);
2276	ocfs2_inode_unlock(inode, 1);
2277	brelse(di_bh);
2278out:
2279	if (ret < 0)
2280		ret = -EIO;
2281	return ret;
2282}
2283
2284static void ocfs2_dio_end_io_write(struct inode *inode,
2285				   struct ocfs2_dio_write_ctxt *dwc,
2286				   loff_t offset,
2287				   ssize_t bytes)
2288{
2289	struct ocfs2_cached_dealloc_ctxt dealloc;
2290	struct ocfs2_extent_tree et;
2291	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2292	struct ocfs2_inode_info *oi = OCFS2_I(inode);
2293	struct ocfs2_unwritten_extent *ue = NULL;
2294	struct buffer_head *di_bh = NULL;
2295	struct ocfs2_dinode *di;
2296	struct ocfs2_alloc_context *data_ac = NULL;
2297	struct ocfs2_alloc_context *meta_ac = NULL;
2298	handle_t *handle = NULL;
2299	loff_t end = offset + bytes;
2300	int ret = 0, credits = 0, locked = 0;
2301
2302	ocfs2_init_dealloc_ctxt(&dealloc);
2303
2304	/* We do clear unwritten, delete orphan, change i_size here. If neither
2305	 * of these happen, we can skip all this. */
2306	if (list_empty(&dwc->dw_zero_list) &&
2307	    end <= i_size_read(inode) &&
2308	    !dwc->dw_orphaned)
2309		goto out;
2310
2311	/* ocfs2_file_write_iter will get i_mutex, so we need not lock if we
2312	 * are in that context. */
2313	if (dwc->dw_writer_pid != task_pid_nr(current)) {
2314		mutex_lock(&inode->i_mutex);
2315		locked = 1;
2316	}
2317
2318	ret = ocfs2_inode_lock(inode, &di_bh, 1);
2319	if (ret < 0) {
2320		mlog_errno(ret);
2321		goto out;
2322	}
2323
2324	down_write(&oi->ip_alloc_sem);
2325
2326	/* Delete orphan before acquire i_mutex. */
2327	if (dwc->dw_orphaned) {
2328		BUG_ON(dwc->dw_writer_pid != task_pid_nr(current));
2329
2330		end = end > i_size_read(inode) ? end : 0;
2331
2332		ret = ocfs2_del_inode_from_orphan(osb, inode, di_bh,
2333				!!end, end);
2334		if (ret < 0)
2335			mlog_errno(ret);
2336	}
2337
2338	di = (struct ocfs2_dinode *)di_bh;
2339
2340	ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
2341
 
 
 
 
 
 
2342	ret = ocfs2_lock_allocators(inode, &et, 0, dwc->dw_zero_count*2,
2343				    &data_ac, &meta_ac);
2344	if (ret) {
2345		mlog_errno(ret);
2346		goto unlock;
2347	}
2348
2349	credits = ocfs2_calc_extend_credits(inode->i_sb, &di->id2.i_list);
2350
2351	handle = ocfs2_start_trans(osb, credits);
2352	if (IS_ERR(handle)) {
2353		ret = PTR_ERR(handle);
2354		mlog_errno(ret);
2355		goto unlock;
2356	}
2357	ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
2358				      OCFS2_JOURNAL_ACCESS_WRITE);
2359	if (ret) {
2360		mlog_errno(ret);
2361		goto commit;
2362	}
2363
2364	list_for_each_entry(ue, &dwc->dw_zero_list, ue_node) {
2365		ret = ocfs2_mark_extent_written(inode, &et, handle,
2366						ue->ue_cpos, 1,
2367						ue->ue_phys,
2368						meta_ac, &dealloc);
2369		if (ret < 0) {
2370			mlog_errno(ret);
2371			break;
2372		}
2373	}
2374
2375	if (end > i_size_read(inode)) {
2376		ret = ocfs2_set_inode_size(handle, inode, di_bh, end);
2377		if (ret < 0)
2378			mlog_errno(ret);
2379	}
2380commit:
2381	ocfs2_commit_trans(osb, handle);
2382unlock:
2383	up_write(&oi->ip_alloc_sem);
2384	ocfs2_inode_unlock(inode, 1);
2385	brelse(di_bh);
2386out:
2387	if (data_ac)
2388		ocfs2_free_alloc_context(data_ac);
2389	if (meta_ac)
2390		ocfs2_free_alloc_context(meta_ac);
2391	ocfs2_run_deallocs(osb, &dealloc);
2392	if (locked)
2393		mutex_unlock(&inode->i_mutex);
2394	ocfs2_dio_free_write_ctx(inode, dwc);
 
 
2395}
2396
2397/*
2398 * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
2399 * particularly interested in the aio/dio case.  We use the rw_lock DLM lock
2400 * to protect io on one node from truncation on another.
2401 */
2402static int ocfs2_dio_end_io(struct kiocb *iocb,
2403			    loff_t offset,
2404			    ssize_t bytes,
2405			    void *private)
2406{
2407	struct inode *inode = file_inode(iocb->ki_filp);
2408	int level;
2409
2410	if (bytes <= 0)
2411		return 0;
2412
2413	/* this io's submitter should not have unlocked this before we could */
2414	BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
2415
2416	if (private)
2417		ocfs2_dio_end_io_write(inode, private, offset, bytes);
 
 
 
 
 
 
 
 
2418
2419	ocfs2_iocb_clear_rw_locked(iocb);
2420
2421	level = ocfs2_iocb_rw_locked_level(iocb);
2422	ocfs2_rw_unlock(inode, level);
2423	return 0;
2424}
2425
2426static ssize_t ocfs2_direct_IO(struct kiocb *iocb, struct iov_iter *iter,
2427			       loff_t offset)
2428{
2429	struct file *file = iocb->ki_filp;
2430	struct inode *inode = file_inode(file)->i_mapping->host;
2431	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2432	loff_t end = offset + iter->count;
2433	get_block_t *get_block;
2434
2435	/*
2436	 * Fallback to buffered I/O if we see an inode without
2437	 * extents.
2438	 */
2439	if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2440		return 0;
2441
2442	/* Fallback to buffered I/O if we do not support append dio. */
2443	if (end > i_size_read(inode) && !ocfs2_supports_append_dio(osb))
 
2444		return 0;
2445
2446	if (iov_iter_rw(iter) == READ)
2447		get_block = ocfs2_get_block;
2448	else
2449		get_block = ocfs2_dio_get_block;
2450
2451	return __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev,
2452				    iter, offset, get_block,
2453				    ocfs2_dio_end_io, NULL, 0);
2454}
2455
2456const struct address_space_operations ocfs2_aops = {
 
2457	.readpage		= ocfs2_readpage,
2458	.readpages		= ocfs2_readpages,
2459	.writepage		= ocfs2_writepage,
2460	.write_begin		= ocfs2_write_begin,
2461	.write_end		= ocfs2_write_end,
2462	.bmap			= ocfs2_bmap,
2463	.direct_IO		= ocfs2_direct_IO,
2464	.invalidatepage		= block_invalidatepage,
2465	.releasepage		= ocfs2_releasepage,
2466	.migratepage		= buffer_migrate_page,
2467	.is_partially_uptodate	= block_is_partially_uptodate,
2468	.error_remove_page	= generic_error_remove_page,
2469};