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	/*
 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 = 1 << inode->i_blkbits;
 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};
 788
 789void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
 790{
 791	int i;
 792
 793	for(i = 0; i < num_pages; i++) {
 794		if (pages[i]) {
 795			unlock_page(pages[i]);
 796			mark_page_accessed(pages[i]);
 797			put_page(pages[i]);
 798		}
 799	}
 800}
 801
 802static void ocfs2_unlock_pages(struct ocfs2_write_ctxt *wc)
 803{
 804	int i;
 805
 806	/*
 807	 * w_target_locked is only set to true in the page_mkwrite() case.
 808	 * The intent is to allow us to lock the target page from write_begin()
 809	 * to write_end(). The caller must hold a ref on w_target_page.
 810	 */
 811	if (wc->w_target_locked) {
 812		BUG_ON(!wc->w_target_page);
 813		for (i = 0; i < wc->w_num_pages; i++) {
 814			if (wc->w_target_page == wc->w_pages[i]) {
 815				wc->w_pages[i] = NULL;
 816				break;
 817			}
 818		}
 819		mark_page_accessed(wc->w_target_page);
 820		put_page(wc->w_target_page);
 821	}
 822	ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
 823}
 824
 825static void ocfs2_free_unwritten_list(struct inode *inode,
 826				 struct list_head *head)
 827{
 828	struct ocfs2_inode_info *oi = OCFS2_I(inode);
 829	struct ocfs2_unwritten_extent *ue = NULL, *tmp = NULL;
 830
 831	list_for_each_entry_safe(ue, tmp, head, ue_node) {
 832		list_del(&ue->ue_node);
 833		spin_lock(&oi->ip_lock);
 834		list_del(&ue->ue_ip_node);
 835		spin_unlock(&oi->ip_lock);
 836		kfree(ue);
 837	}
 838}
 839
 840static void ocfs2_free_write_ctxt(struct inode *inode,
 841				  struct ocfs2_write_ctxt *wc)
 842{
 843	ocfs2_free_unwritten_list(inode, &wc->w_unwritten_list);
 844	ocfs2_unlock_pages(wc);
 845	brelse(wc->w_di_bh);
 846	kfree(wc);
 847}
 848
 849static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
 850				  struct ocfs2_super *osb, loff_t pos,
 851				  unsigned len, ocfs2_write_type_t type,
 852				  struct buffer_head *di_bh)
 853{
 854	u32 cend;
 855	struct ocfs2_write_ctxt *wc;
 856
 857	wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
 858	if (!wc)
 859		return -ENOMEM;
 860
 861	wc->w_cpos = pos >> osb->s_clustersize_bits;
 862	wc->w_first_new_cpos = UINT_MAX;
 863	cend = (pos + len - 1) >> osb->s_clustersize_bits;
 864	wc->w_clen = cend - wc->w_cpos + 1;
 865	get_bh(di_bh);
 866	wc->w_di_bh = di_bh;
 867	wc->w_type = type;
 868
 869	if (unlikely(PAGE_SHIFT > osb->s_clustersize_bits))
 870		wc->w_large_pages = 1;
 871	else
 872		wc->w_large_pages = 0;
 873
 874	ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
 875	INIT_LIST_HEAD(&wc->w_unwritten_list);
 876
 877	*wcp = wc;
 878
 879	return 0;
 880}
 881
 882/*
 883 * If a page has any new buffers, zero them out here, and mark them uptodate
 884 * and dirty so they'll be written out (in order to prevent uninitialised
 885 * block data from leaking). And clear the new bit.
 886 */
 887static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
 888{
 889	unsigned int block_start, block_end;
 890	struct buffer_head *head, *bh;
 891
 892	BUG_ON(!PageLocked(page));
 893	if (!page_has_buffers(page))
 894		return;
 895
 896	bh = head = page_buffers(page);
 897	block_start = 0;
 898	do {
 899		block_end = block_start + bh->b_size;
 900
 901		if (buffer_new(bh)) {
 902			if (block_end > from && block_start < to) {
 903				if (!PageUptodate(page)) {
 904					unsigned start, end;
 905
 906					start = max(from, block_start);
 907					end = min(to, block_end);
 908
 909					zero_user_segment(page, start, end);
 910					set_buffer_uptodate(bh);
 911				}
 912
 913				clear_buffer_new(bh);
 914				mark_buffer_dirty(bh);
 915			}
 916		}
 917
 918		block_start = block_end;
 919		bh = bh->b_this_page;
 920	} while (bh != head);
 921}
 922
 923/*
 924 * Only called when we have a failure during allocating write to write
 925 * zero's to the newly allocated region.
 926 */
 927static void ocfs2_write_failure(struct inode *inode,
 928				struct ocfs2_write_ctxt *wc,
 929				loff_t user_pos, unsigned user_len)
 930{
 931	int i;
 932	unsigned from = user_pos & (PAGE_SIZE - 1),
 933		to = user_pos + user_len;
 934	struct page *tmppage;
 935
 936	if (wc->w_target_page)
 937		ocfs2_zero_new_buffers(wc->w_target_page, from, to);
 938
 939	for(i = 0; i < wc->w_num_pages; i++) {
 940		tmppage = wc->w_pages[i];
 941
 942		if (tmppage && page_has_buffers(tmppage)) {
 943			if (ocfs2_should_order_data(inode))
 944				ocfs2_jbd2_file_inode(wc->w_handle, inode);
 
 945
 946			block_commit_write(tmppage, from, to);
 947		}
 948	}
 949}
 950
 951static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
 952					struct ocfs2_write_ctxt *wc,
 953					struct page *page, u32 cpos,
 954					loff_t user_pos, unsigned user_len,
 955					int new)
 956{
 957	int ret;
 958	unsigned int map_from = 0, map_to = 0;
 959	unsigned int cluster_start, cluster_end;
 960	unsigned int user_data_from = 0, user_data_to = 0;
 961
 962	ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
 963					&cluster_start, &cluster_end);
 964
 965	/* treat the write as new if the a hole/lseek spanned across
 966	 * the page boundary.
 967	 */
 968	new = new | ((i_size_read(inode) <= page_offset(page)) &&
 969			(page_offset(page) <= user_pos));
 970
 971	if (page == wc->w_target_page) {
 972		map_from = user_pos & (PAGE_SIZE - 1);
 973		map_to = map_from + user_len;
 974
 975		if (new)
 976			ret = ocfs2_map_page_blocks(page, p_blkno, inode,
 977						    cluster_start, cluster_end,
 978						    new);
 979		else
 980			ret = ocfs2_map_page_blocks(page, p_blkno, inode,
 981						    map_from, map_to, new);
 982		if (ret) {
 983			mlog_errno(ret);
 984			goto out;
 985		}
 986
 987		user_data_from = map_from;
 988		user_data_to = map_to;
 989		if (new) {
 990			map_from = cluster_start;
 991			map_to = cluster_end;
 992		}
 993	} else {
 994		/*
 995		 * If we haven't allocated the new page yet, we
 996		 * shouldn't be writing it out without copying user
 997		 * data. This is likely a math error from the caller.
 998		 */
 999		BUG_ON(!new);
1000
1001		map_from = cluster_start;
1002		map_to = cluster_end;
1003
1004		ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1005					    cluster_start, cluster_end, new);
1006		if (ret) {
1007			mlog_errno(ret);
1008			goto out;
1009		}
1010	}
1011
1012	/*
1013	 * Parts of newly allocated pages need to be zero'd.
1014	 *
1015	 * Above, we have also rewritten 'to' and 'from' - as far as
1016	 * the rest of the function is concerned, the entire cluster
1017	 * range inside of a page needs to be written.
1018	 *
1019	 * We can skip this if the page is up to date - it's already
1020	 * been zero'd from being read in as a hole.
1021	 */
1022	if (new && !PageUptodate(page))
1023		ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1024					 cpos, user_data_from, user_data_to);
1025
1026	flush_dcache_page(page);
1027
1028out:
1029	return ret;
1030}
1031
1032/*
1033 * This function will only grab one clusters worth of pages.
1034 */
1035static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1036				      struct ocfs2_write_ctxt *wc,
1037				      u32 cpos, loff_t user_pos,
1038				      unsigned user_len, int new,
1039				      struct page *mmap_page)
1040{
1041	int ret = 0, i;
1042	unsigned long start, target_index, end_index, index;
1043	struct inode *inode = mapping->host;
1044	loff_t last_byte;
1045
1046	target_index = user_pos >> PAGE_SHIFT;
1047
1048	/*
1049	 * Figure out how many pages we'll be manipulating here. For
1050	 * non allocating write, we just change the one
1051	 * page. Otherwise, we'll need a whole clusters worth.  If we're
1052	 * writing past i_size, we only need enough pages to cover the
1053	 * last page of the write.
1054	 */
1055	if (new) {
1056		wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1057		start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1058		/*
1059		 * We need the index *past* the last page we could possibly
1060		 * touch.  This is the page past the end of the write or
1061		 * i_size, whichever is greater.
1062		 */
1063		last_byte = max(user_pos + user_len, i_size_read(inode));
1064		BUG_ON(last_byte < 1);
1065		end_index = ((last_byte - 1) >> PAGE_SHIFT) + 1;
1066		if ((start + wc->w_num_pages) > end_index)
1067			wc->w_num_pages = end_index - start;
1068	} else {
1069		wc->w_num_pages = 1;
1070		start = target_index;
1071	}
1072	end_index = (user_pos + user_len - 1) >> PAGE_SHIFT;
1073
1074	for(i = 0; i < wc->w_num_pages; i++) {
1075		index = start + i;
1076
1077		if (index >= target_index && index <= end_index &&
1078		    wc->w_type == OCFS2_WRITE_MMAP) {
1079			/*
1080			 * ocfs2_pagemkwrite() is a little different
1081			 * and wants us to directly use the page
1082			 * passed in.
1083			 */
1084			lock_page(mmap_page);
1085
1086			/* Exit and let the caller retry */
1087			if (mmap_page->mapping != mapping) {
1088				WARN_ON(mmap_page->mapping);
1089				unlock_page(mmap_page);
1090				ret = -EAGAIN;
1091				goto out;
1092			}
1093
1094			get_page(mmap_page);
1095			wc->w_pages[i] = mmap_page;
1096			wc->w_target_locked = true;
1097		} else if (index >= target_index && index <= end_index &&
1098			   wc->w_type == OCFS2_WRITE_DIRECT) {
1099			/* Direct write has no mapping page. */
1100			wc->w_pages[i] = NULL;
1101			continue;
1102		} else {
1103			wc->w_pages[i] = find_or_create_page(mapping, index,
1104							     GFP_NOFS);
1105			if (!wc->w_pages[i]) {
1106				ret = -ENOMEM;
1107				mlog_errno(ret);
1108				goto out;
1109			}
1110		}
1111		wait_for_stable_page(wc->w_pages[i]);
1112
1113		if (index == target_index)
1114			wc->w_target_page = wc->w_pages[i];
1115	}
1116out:
1117	if (ret)
1118		wc->w_target_locked = false;
1119	return ret;
1120}
1121
1122/*
1123 * Prepare a single cluster for write one cluster into the file.
1124 */
1125static int ocfs2_write_cluster(struct address_space *mapping,
1126			       u32 *phys, unsigned int new,
1127			       unsigned int clear_unwritten,
1128			       unsigned int should_zero,
1129			       struct ocfs2_alloc_context *data_ac,
1130			       struct ocfs2_alloc_context *meta_ac,
1131			       struct ocfs2_write_ctxt *wc, u32 cpos,
1132			       loff_t user_pos, unsigned user_len)
1133{
1134	int ret, i;
1135	u64 p_blkno;
1136	struct inode *inode = mapping->host;
1137	struct ocfs2_extent_tree et;
1138	int bpc = ocfs2_clusters_to_blocks(inode->i_sb, 1);
1139
1140	if (new) {
1141		u32 tmp_pos;
1142
1143		/*
1144		 * This is safe to call with the page locks - it won't take
1145		 * any additional semaphores or cluster locks.
1146		 */
1147		tmp_pos = cpos;
1148		ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1149					   &tmp_pos, 1, !clear_unwritten,
1150					   wc->w_di_bh, wc->w_handle,
1151					   data_ac, meta_ac, NULL);
1152		/*
1153		 * This shouldn't happen because we must have already
1154		 * calculated the correct meta data allocation required. The
1155		 * internal tree allocation code should know how to increase
1156		 * transaction credits itself.
1157		 *
1158		 * If need be, we could handle -EAGAIN for a
1159		 * RESTART_TRANS here.
1160		 */
1161		mlog_bug_on_msg(ret == -EAGAIN,
1162				"Inode %llu: EAGAIN return during allocation.\n",
1163				(unsigned long long)OCFS2_I(inode)->ip_blkno);
1164		if (ret < 0) {
1165			mlog_errno(ret);
1166			goto out;
1167		}
1168	} else if (clear_unwritten) {
1169		ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1170					      wc->w_di_bh);
1171		ret = ocfs2_mark_extent_written(inode, &et,
1172						wc->w_handle, cpos, 1, *phys,
1173						meta_ac, &wc->w_dealloc);
1174		if (ret < 0) {
1175			mlog_errno(ret);
1176			goto out;
1177		}
1178	}
1179
1180	/*
1181	 * The only reason this should fail is due to an inability to
1182	 * find the extent added.
1183	 */
1184	ret = ocfs2_get_clusters(inode, cpos, phys, NULL, NULL);
1185	if (ret < 0) {
1186		mlog(ML_ERROR, "Get physical blkno failed for inode %llu, "
1187			    "at logical cluster %u",
1188			    (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
1189		goto out;
1190	}
1191
1192	BUG_ON(*phys == 0);
1193
1194	p_blkno = ocfs2_clusters_to_blocks(inode->i_sb, *phys);
1195	if (!should_zero)
1196		p_blkno += (user_pos >> inode->i_sb->s_blocksize_bits) & (u64)(bpc - 1);
1197
1198	for(i = 0; i < wc->w_num_pages; i++) {
1199		int tmpret;
1200
1201		/* This is the direct io target page. */
1202		if (wc->w_pages[i] == NULL) {
1203			p_blkno++;
1204			continue;
1205		}
1206
1207		tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1208						      wc->w_pages[i], cpos,
1209						      user_pos, user_len,
1210						      should_zero);
1211		if (tmpret) {
1212			mlog_errno(tmpret);
1213			if (ret == 0)
1214				ret = tmpret;
1215		}
1216	}
1217
1218	/*
1219	 * We only have cleanup to do in case of allocating write.
1220	 */
1221	if (ret && new)
1222		ocfs2_write_failure(inode, wc, user_pos, user_len);
1223
1224out:
1225
1226	return ret;
1227}
1228
1229static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1230				       struct ocfs2_alloc_context *data_ac,
1231				       struct ocfs2_alloc_context *meta_ac,
1232				       struct ocfs2_write_ctxt *wc,
1233				       loff_t pos, unsigned len)
1234{
1235	int ret, i;
1236	loff_t cluster_off;
1237	unsigned int local_len = len;
1238	struct ocfs2_write_cluster_desc *desc;
1239	struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1240
1241	for (i = 0; i < wc->w_clen; i++) {
1242		desc = &wc->w_desc[i];
1243
1244		/*
1245		 * We have to make sure that the total write passed in
1246		 * doesn't extend past a single cluster.
1247		 */
1248		local_len = len;
1249		cluster_off = pos & (osb->s_clustersize - 1);
1250		if ((cluster_off + local_len) > osb->s_clustersize)
1251			local_len = osb->s_clustersize - cluster_off;
1252
1253		ret = ocfs2_write_cluster(mapping, &desc->c_phys,
1254					  desc->c_new,
1255					  desc->c_clear_unwritten,
1256					  desc->c_needs_zero,
1257					  data_ac, meta_ac,
1258					  wc, desc->c_cpos, pos, local_len);
1259		if (ret) {
1260			mlog_errno(ret);
1261			goto out;
1262		}
1263
1264		len -= local_len;
1265		pos += local_len;
1266	}
1267
1268	ret = 0;
1269out:
1270	return ret;
1271}
1272
1273/*
1274 * ocfs2_write_end() wants to know which parts of the target page it
1275 * should complete the write on. It's easiest to compute them ahead of
1276 * time when a more complete view of the write is available.
1277 */
1278static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1279					struct ocfs2_write_ctxt *wc,
1280					loff_t pos, unsigned len, int alloc)
1281{
1282	struct ocfs2_write_cluster_desc *desc;
1283
1284	wc->w_target_from = pos & (PAGE_SIZE - 1);
1285	wc->w_target_to = wc->w_target_from + len;
1286
1287	if (alloc == 0)
1288		return;
1289
1290	/*
1291	 * Allocating write - we may have different boundaries based
1292	 * on page size and cluster size.
1293	 *
1294	 * NOTE: We can no longer compute one value from the other as
1295	 * the actual write length and user provided length may be
1296	 * different.
1297	 */
1298
1299	if (wc->w_large_pages) {
1300		/*
1301		 * We only care about the 1st and last cluster within
1302		 * our range and whether they should be zero'd or not. Either
1303		 * value may be extended out to the start/end of a
1304		 * newly allocated cluster.
1305		 */
1306		desc = &wc->w_desc[0];
1307		if (desc->c_needs_zero)
1308			ocfs2_figure_cluster_boundaries(osb,
1309							desc->c_cpos,
1310							&wc->w_target_from,
1311							NULL);
1312
1313		desc = &wc->w_desc[wc->w_clen - 1];
1314		if (desc->c_needs_zero)
1315			ocfs2_figure_cluster_boundaries(osb,
1316							desc->c_cpos,
1317							NULL,
1318							&wc->w_target_to);
1319	} else {
1320		wc->w_target_from = 0;
1321		wc->w_target_to = PAGE_SIZE;
1322	}
1323}
1324
1325/*
1326 * Check if this extent is marked UNWRITTEN by direct io. If so, we need not to
1327 * do the zero work. And should not to clear UNWRITTEN since it will be cleared
1328 * by the direct io procedure.
1329 * If this is a new extent that allocated by direct io, we should mark it in
1330 * the ip_unwritten_list.
1331 */
1332static int ocfs2_unwritten_check(struct inode *inode,
1333				 struct ocfs2_write_ctxt *wc,
1334				 struct ocfs2_write_cluster_desc *desc)
1335{
1336	struct ocfs2_inode_info *oi = OCFS2_I(inode);
1337	struct ocfs2_unwritten_extent *ue = NULL, *new = NULL;
1338	int ret = 0;
1339
1340	if (!desc->c_needs_zero)
1341		return 0;
1342
1343retry:
1344	spin_lock(&oi->ip_lock);
1345	/* Needs not to zero no metter buffer or direct. The one who is zero
1346	 * the cluster is doing zero. And he will clear unwritten after all
1347	 * cluster io finished. */
1348	list_for_each_entry(ue, &oi->ip_unwritten_list, ue_ip_node) {
1349		if (desc->c_cpos == ue->ue_cpos) {
1350			BUG_ON(desc->c_new);
1351			desc->c_needs_zero = 0;
1352			desc->c_clear_unwritten = 0;
1353			goto unlock;
1354		}
1355	}
1356
1357	if (wc->w_type != OCFS2_WRITE_DIRECT)
1358		goto unlock;
1359
1360	if (new == NULL) {
1361		spin_unlock(&oi->ip_lock);
1362		new = kmalloc(sizeof(struct ocfs2_unwritten_extent),
1363			     GFP_NOFS);
1364		if (new == NULL) {
1365			ret = -ENOMEM;
1366			goto out;
1367		}
1368		goto retry;
1369	}
1370	/* This direct write will doing zero. */
1371	new->ue_cpos = desc->c_cpos;
1372	new->ue_phys = desc->c_phys;
1373	desc->c_clear_unwritten = 0;
1374	list_add_tail(&new->ue_ip_node, &oi->ip_unwritten_list);
1375	list_add_tail(&new->ue_node, &wc->w_unwritten_list);
 
1376	new = NULL;
1377unlock:
1378	spin_unlock(&oi->ip_lock);
1379out:
1380	if (new)
1381		kfree(new);
1382	return ret;
1383}
1384
1385/*
1386 * Populate each single-cluster write descriptor in the write context
1387 * with information about the i/o to be done.
1388 *
1389 * Returns the number of clusters that will have to be allocated, as
1390 * well as a worst case estimate of the number of extent records that
1391 * would have to be created during a write to an unwritten region.
1392 */
1393static int ocfs2_populate_write_desc(struct inode *inode,
1394				     struct ocfs2_write_ctxt *wc,
1395				     unsigned int *clusters_to_alloc,
1396				     unsigned int *extents_to_split)
1397{
1398	int ret;
1399	struct ocfs2_write_cluster_desc *desc;
1400	unsigned int num_clusters = 0;
1401	unsigned int ext_flags = 0;
1402	u32 phys = 0;
1403	int i;
1404
1405	*clusters_to_alloc = 0;
1406	*extents_to_split = 0;
1407
1408	for (i = 0; i < wc->w_clen; i++) {
1409		desc = &wc->w_desc[i];
1410		desc->c_cpos = wc->w_cpos + i;
1411
1412		if (num_clusters == 0) {
1413			/*
1414			 * Need to look up the next extent record.
1415			 */
1416			ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1417						 &num_clusters, &ext_flags);
1418			if (ret) {
1419				mlog_errno(ret);
1420				goto out;
1421			}
1422
1423			/* We should already CoW the refcountd extent. */
1424			BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1425
1426			/*
1427			 * Assume worst case - that we're writing in
1428			 * the middle of the extent.
1429			 *
1430			 * We can assume that the write proceeds from
1431			 * left to right, in which case the extent
1432			 * insert code is smart enough to coalesce the
1433			 * next splits into the previous records created.
1434			 */
1435			if (ext_flags & OCFS2_EXT_UNWRITTEN)
1436				*extents_to_split = *extents_to_split + 2;
1437		} else if (phys) {
1438			/*
1439			 * Only increment phys if it doesn't describe
1440			 * a hole.
1441			 */
1442			phys++;
1443		}
1444
1445		/*
1446		 * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1447		 * file that got extended.  w_first_new_cpos tells us
1448		 * where the newly allocated clusters are so we can
1449		 * zero them.
1450		 */
1451		if (desc->c_cpos >= wc->w_first_new_cpos) {
1452			BUG_ON(phys == 0);
1453			desc->c_needs_zero = 1;
1454		}
1455
1456		desc->c_phys = phys;
1457		if (phys == 0) {
1458			desc->c_new = 1;
1459			desc->c_needs_zero = 1;
1460			desc->c_clear_unwritten = 1;
1461			*clusters_to_alloc = *clusters_to_alloc + 1;
1462		}
1463
1464		if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1465			desc->c_clear_unwritten = 1;
1466			desc->c_needs_zero = 1;
1467		}
1468
1469		ret = ocfs2_unwritten_check(inode, wc, desc);
1470		if (ret) {
1471			mlog_errno(ret);
1472			goto out;
1473		}
1474
1475		num_clusters--;
1476	}
1477
1478	ret = 0;
1479out:
1480	return ret;
1481}
1482
1483static int ocfs2_write_begin_inline(struct address_space *mapping,
1484				    struct inode *inode,
1485				    struct ocfs2_write_ctxt *wc)
1486{
1487	int ret;
1488	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1489	struct page *page;
1490	handle_t *handle;
1491	struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1492
1493	handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1494	if (IS_ERR(handle)) {
1495		ret = PTR_ERR(handle);
1496		mlog_errno(ret);
1497		goto out;
1498	}
1499
1500	page = find_or_create_page(mapping, 0, GFP_NOFS);
1501	if (!page) {
1502		ocfs2_commit_trans(osb, handle);
1503		ret = -ENOMEM;
1504		mlog_errno(ret);
1505		goto out;
1506	}
1507	/*
1508	 * If we don't set w_num_pages then this page won't get unlocked
1509	 * and freed on cleanup of the write context.
1510	 */
1511	wc->w_pages[0] = wc->w_target_page = page;
1512	wc->w_num_pages = 1;
1513
1514	ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1515				      OCFS2_JOURNAL_ACCESS_WRITE);
1516	if (ret) {
1517		ocfs2_commit_trans(osb, handle);
1518
1519		mlog_errno(ret);
1520		goto out;
1521	}
1522
1523	if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1524		ocfs2_set_inode_data_inline(inode, di);
1525
1526	if (!PageUptodate(page)) {
1527		ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1528		if (ret) {
1529			ocfs2_commit_trans(osb, handle);
1530
1531			goto out;
1532		}
1533	}
1534
1535	wc->w_handle = handle;
1536out:
1537	return ret;
1538}
1539
1540int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1541{
1542	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1543
1544	if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1545		return 1;
1546	return 0;
1547}
1548
1549static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1550					  struct inode *inode, loff_t pos,
1551					  unsigned len, struct page *mmap_page,
1552					  struct ocfs2_write_ctxt *wc)
1553{
1554	int ret, written = 0;
1555	loff_t end = pos + len;
1556	struct ocfs2_inode_info *oi = OCFS2_I(inode);
1557	struct ocfs2_dinode *di = NULL;
1558
1559	trace_ocfs2_try_to_write_inline_data((unsigned long long)oi->ip_blkno,
1560					     len, (unsigned long long)pos,
1561					     oi->ip_dyn_features);
1562
1563	/*
1564	 * Handle inodes which already have inline data 1st.
1565	 */
1566	if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1567		if (mmap_page == NULL &&
1568		    ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1569			goto do_inline_write;
1570
1571		/*
1572		 * The write won't fit - we have to give this inode an
1573		 * inline extent list now.
1574		 */
1575		ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1576		if (ret)
1577			mlog_errno(ret);
1578		goto out;
1579	}
1580
1581	/*
1582	 * Check whether the inode can accept inline data.
1583	 */
1584	if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1585		return 0;
1586
1587	/*
1588	 * Check whether the write can fit.
1589	 */
1590	di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1591	if (mmap_page ||
1592	    end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1593		return 0;
1594
1595do_inline_write:
1596	ret = ocfs2_write_begin_inline(mapping, inode, wc);
1597	if (ret) {
1598		mlog_errno(ret);
1599		goto out;
1600	}
1601
1602	/*
1603	 * This signals to the caller that the data can be written
1604	 * inline.
1605	 */
1606	written = 1;
1607out:
1608	return written ? written : ret;
1609}
1610
1611/*
1612 * This function only does anything for file systems which can't
1613 * handle sparse files.
1614 *
1615 * What we want to do here is fill in any hole between the current end
1616 * of allocation and the end of our write. That way the rest of the
1617 * write path can treat it as an non-allocating write, which has no
1618 * special case code for sparse/nonsparse files.
1619 */
1620static int ocfs2_expand_nonsparse_inode(struct inode *inode,
1621					struct buffer_head *di_bh,
1622					loff_t pos, unsigned len,
1623					struct ocfs2_write_ctxt *wc)
1624{
1625	int ret;
1626	loff_t newsize = pos + len;
1627
1628	BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1629
1630	if (newsize <= i_size_read(inode))
1631		return 0;
1632
1633	ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
1634	if (ret)
1635		mlog_errno(ret);
1636
1637	/* There is no wc if this is call from direct. */
1638	if (wc)
1639		wc->w_first_new_cpos =
1640			ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1641
1642	return ret;
1643}
1644
1645static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
1646			   loff_t pos)
1647{
1648	int ret = 0;
1649
1650	BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1651	if (pos > i_size_read(inode))
1652		ret = ocfs2_zero_extend(inode, di_bh, pos);
1653
1654	return ret;
1655}
1656
1657int ocfs2_write_begin_nolock(struct address_space *mapping,
1658			     loff_t pos, unsigned len, ocfs2_write_type_t type,
1659			     struct page **pagep, void **fsdata,
1660			     struct buffer_head *di_bh, struct page *mmap_page)
1661{
1662	int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1663	unsigned int clusters_to_alloc, extents_to_split, clusters_need = 0;
1664	struct ocfs2_write_ctxt *wc;
1665	struct inode *inode = mapping->host;
1666	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1667	struct ocfs2_dinode *di;
1668	struct ocfs2_alloc_context *data_ac = NULL;
1669	struct ocfs2_alloc_context *meta_ac = NULL;
1670	handle_t *handle;
1671	struct ocfs2_extent_tree et;
1672	int try_free = 1, ret1;
1673
1674try_again:
1675	ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, type, di_bh);
1676	if (ret) {
1677		mlog_errno(ret);
1678		return ret;
1679	}
1680
1681	if (ocfs2_supports_inline_data(osb)) {
1682		ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1683						     mmap_page, wc);
1684		if (ret == 1) {
1685			ret = 0;
1686			goto success;
1687		}
1688		if (ret < 0) {
1689			mlog_errno(ret);
1690			goto out;
1691		}
1692	}
1693
1694	/* Direct io change i_size late, should not zero tail here. */
1695	if (type != OCFS2_WRITE_DIRECT) {
1696		if (ocfs2_sparse_alloc(osb))
1697			ret = ocfs2_zero_tail(inode, di_bh, pos);
1698		else
1699			ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos,
1700							   len, wc);
1701		if (ret) {
1702			mlog_errno(ret);
1703			goto out;
1704		}
1705	}
1706
1707	ret = ocfs2_check_range_for_refcount(inode, pos, len);
1708	if (ret < 0) {
1709		mlog_errno(ret);
1710		goto out;
1711	} else if (ret == 1) {
1712		clusters_need = wc->w_clen;
1713		ret = ocfs2_refcount_cow(inode, di_bh,
1714					 wc->w_cpos, wc->w_clen, UINT_MAX);
1715		if (ret) {
1716			mlog_errno(ret);
1717			goto out;
1718		}
1719	}
1720
1721	ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1722					&extents_to_split);
1723	if (ret) {
1724		mlog_errno(ret);
1725		goto out;
1726	}
1727	clusters_need += clusters_to_alloc;
1728
1729	di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1730
1731	trace_ocfs2_write_begin_nolock(
1732			(unsigned long long)OCFS2_I(inode)->ip_blkno,
1733			(long long)i_size_read(inode),
1734			le32_to_cpu(di->i_clusters),
1735			pos, len, type, mmap_page,
1736			clusters_to_alloc, extents_to_split);
1737
1738	/*
1739	 * We set w_target_from, w_target_to here so that
1740	 * ocfs2_write_end() knows which range in the target page to
1741	 * write out. An allocation requires that we write the entire
1742	 * cluster range.
1743	 */
1744	if (clusters_to_alloc || extents_to_split) {
1745		/*
1746		 * XXX: We are stretching the limits of
1747		 * ocfs2_lock_allocators(). It greatly over-estimates
1748		 * the work to be done.
1749		 */
1750		ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1751					      wc->w_di_bh);
1752		ret = ocfs2_lock_allocators(inode, &et,
1753					    clusters_to_alloc, extents_to_split,
1754					    &data_ac, &meta_ac);
1755		if (ret) {
1756			mlog_errno(ret);
1757			goto out;
1758		}
1759
1760		if (data_ac)
1761			data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
1762
1763		credits = ocfs2_calc_extend_credits(inode->i_sb,
1764						    &di->id2.i_list);
1765	} else if (type == OCFS2_WRITE_DIRECT)
1766		/* direct write needs not to start trans if no extents alloc. */
1767		goto success;
1768
1769	/*
1770	 * We have to zero sparse allocated clusters, unwritten extent clusters,
1771	 * and non-sparse clusters we just extended.  For non-sparse writes,
1772	 * we know zeros will only be needed in the first and/or last cluster.
1773	 */
1774	if (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1775			   wc->w_desc[wc->w_clen - 1].c_needs_zero))
1776		cluster_of_pages = 1;
1777	else
1778		cluster_of_pages = 0;
1779
1780	ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1781
1782	handle = ocfs2_start_trans(osb, credits);
1783	if (IS_ERR(handle)) {
1784		ret = PTR_ERR(handle);
1785		mlog_errno(ret);
1786		goto out;
1787	}
1788
1789	wc->w_handle = handle;
1790
1791	if (clusters_to_alloc) {
1792		ret = dquot_alloc_space_nodirty(inode,
1793			ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1794		if (ret)
1795			goto out_commit;
1796	}
1797
1798	ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1799				      OCFS2_JOURNAL_ACCESS_WRITE);
1800	if (ret) {
1801		mlog_errno(ret);
1802		goto out_quota;
1803	}
1804
1805	/*
1806	 * Fill our page array first. That way we've grabbed enough so
1807	 * that we can zero and flush if we error after adding the
1808	 * extent.
1809	 */
1810	ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
1811					 cluster_of_pages, mmap_page);
1812	if (ret && ret != -EAGAIN) {
1813		mlog_errno(ret);
1814		goto out_quota;
1815	}
1816
1817	/*
1818	 * ocfs2_grab_pages_for_write() returns -EAGAIN if it could not lock
1819	 * the target page. In this case, we exit with no error and no target
1820	 * page. This will trigger the caller, page_mkwrite(), to re-try
1821	 * the operation.
1822	 */
1823	if (ret == -EAGAIN) {
1824		BUG_ON(wc->w_target_page);
1825		ret = 0;
1826		goto out_quota;
1827	}
1828
1829	ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1830					  len);
1831	if (ret) {
1832		mlog_errno(ret);
1833		goto out_quota;
1834	}
1835
1836	if (data_ac)
1837		ocfs2_free_alloc_context(data_ac);
1838	if (meta_ac)
1839		ocfs2_free_alloc_context(meta_ac);
1840
1841success:
1842	if (pagep)
1843		*pagep = wc->w_target_page;
1844	*fsdata = wc;
1845	return 0;
1846out_quota:
1847	if (clusters_to_alloc)
1848		dquot_free_space(inode,
1849			  ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1850out_commit:
1851	ocfs2_commit_trans(osb, handle);
1852
1853out:
1854	/*
1855	 * The mmapped page won't be unlocked in ocfs2_free_write_ctxt(),
1856	 * even in case of error here like ENOSPC and ENOMEM. So, we need
1857	 * to unlock the target page manually to prevent deadlocks when
1858	 * retrying again on ENOSPC, or when returning non-VM_FAULT_LOCKED
1859	 * to VM code.
1860	 */
1861	if (wc->w_target_locked)
1862		unlock_page(mmap_page);
1863
1864	ocfs2_free_write_ctxt(inode, wc);
1865
1866	if (data_ac) {
1867		ocfs2_free_alloc_context(data_ac);
1868		data_ac = NULL;
1869	}
1870	if (meta_ac) {
1871		ocfs2_free_alloc_context(meta_ac);
1872		meta_ac = NULL;
1873	}
1874
1875	if (ret == -ENOSPC && try_free) {
1876		/*
1877		 * Try to free some truncate log so that we can have enough
1878		 * clusters to allocate.
1879		 */
1880		try_free = 0;
1881
1882		ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need);
1883		if (ret1 == 1)
1884			goto try_again;
1885
1886		if (ret1 < 0)
1887			mlog_errno(ret1);
1888	}
1889
1890	return ret;
1891}
1892
1893static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1894			     loff_t pos, unsigned len, unsigned flags,
1895			     struct page **pagep, void **fsdata)
1896{
1897	int ret;
1898	struct buffer_head *di_bh = NULL;
1899	struct inode *inode = mapping->host;
1900
1901	ret = ocfs2_inode_lock(inode, &di_bh, 1);
1902	if (ret) {
1903		mlog_errno(ret);
1904		return ret;
1905	}
1906
1907	/*
1908	 * Take alloc sem here to prevent concurrent lookups. That way
1909	 * the mapping, zeroing and tree manipulation within
1910	 * ocfs2_write() will be safe against ->readpage(). This
1911	 * should also serve to lock out allocation from a shared
1912	 * writeable region.
1913	 */
1914	down_write(&OCFS2_I(inode)->ip_alloc_sem);
1915
1916	ret = ocfs2_write_begin_nolock(mapping, pos, len, OCFS2_WRITE_BUFFER,
1917				       pagep, fsdata, di_bh, NULL);
1918	if (ret) {
1919		mlog_errno(ret);
1920		goto out_fail;
1921	}
1922
1923	brelse(di_bh);
1924
1925	return 0;
1926
1927out_fail:
1928	up_write(&OCFS2_I(inode)->ip_alloc_sem);
1929
1930	brelse(di_bh);
1931	ocfs2_inode_unlock(inode, 1);
1932
1933	return ret;
1934}
1935
1936static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1937				   unsigned len, unsigned *copied,
1938				   struct ocfs2_dinode *di,
1939				   struct ocfs2_write_ctxt *wc)
1940{
1941	void *kaddr;
1942
1943	if (unlikely(*copied < len)) {
1944		if (!PageUptodate(wc->w_target_page)) {
1945			*copied = 0;
1946			return;
1947		}
1948	}
1949
1950	kaddr = kmap_atomic(wc->w_target_page);
1951	memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1952	kunmap_atomic(kaddr);
1953
1954	trace_ocfs2_write_end_inline(
1955	     (unsigned long long)OCFS2_I(inode)->ip_blkno,
1956	     (unsigned long long)pos, *copied,
1957	     le16_to_cpu(di->id2.i_data.id_count),
1958	     le16_to_cpu(di->i_dyn_features));
1959}
1960
1961int ocfs2_write_end_nolock(struct address_space *mapping,
1962			   loff_t pos, unsigned len, unsigned copied, void *fsdata)
1963{
1964	int i, ret;
1965	unsigned from, to, start = pos & (PAGE_SIZE - 1);
1966	struct inode *inode = mapping->host;
1967	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1968	struct ocfs2_write_ctxt *wc = fsdata;
1969	struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1970	handle_t *handle = wc->w_handle;
1971	struct page *tmppage;
1972
1973	BUG_ON(!list_empty(&wc->w_unwritten_list));
1974
1975	if (handle) {
1976		ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode),
1977				wc->w_di_bh, OCFS2_JOURNAL_ACCESS_WRITE);
1978		if (ret) {
1979			copied = ret;
1980			mlog_errno(ret);
1981			goto out;
1982		}
1983	}
1984
1985	if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1986		ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
1987		goto out_write_size;
1988	}
1989
1990	if (unlikely(copied < len) && wc->w_target_page) {
1991		if (!PageUptodate(wc->w_target_page))
1992			copied = 0;
1993
1994		ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
1995				       start+len);
1996	}
1997	if (wc->w_target_page)
1998		flush_dcache_page(wc->w_target_page);
1999
2000	for(i = 0; i < wc->w_num_pages; i++) {
2001		tmppage = wc->w_pages[i];
2002
2003		/* This is the direct io target page. */
2004		if (tmppage == NULL)
2005			continue;
2006
2007		if (tmppage == wc->w_target_page) {
2008			from = wc->w_target_from;
2009			to = wc->w_target_to;
2010
2011			BUG_ON(from > PAGE_SIZE ||
2012			       to > PAGE_SIZE ||
2013			       to < from);
2014		} else {
2015			/*
2016			 * Pages adjacent to the target (if any) imply
2017			 * a hole-filling write in which case we want
2018			 * to flush their entire range.
2019			 */
2020			from = 0;
2021			to = PAGE_SIZE;
2022		}
2023
2024		if (page_has_buffers(tmppage)) {
2025			if (handle && ocfs2_should_order_data(inode))
2026				ocfs2_jbd2_file_inode(handle, inode);
 
 
 
 
 
 
2027			block_commit_write(tmppage, from, to);
2028		}
2029	}
2030
2031out_write_size:
2032	/* Direct io do not update i_size here. */
2033	if (wc->w_type != OCFS2_WRITE_DIRECT) {
2034		pos += copied;
2035		if (pos > i_size_read(inode)) {
2036			i_size_write(inode, pos);
2037			mark_inode_dirty(inode);
2038		}
2039		inode->i_blocks = ocfs2_inode_sector_count(inode);
2040		di->i_size = cpu_to_le64((u64)i_size_read(inode));
2041		inode->i_mtime = inode->i_ctime = current_time(inode);
2042		di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
2043		di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
2044		ocfs2_update_inode_fsync_trans(handle, inode, 1);
 
2045	}
2046	if (handle)
2047		ocfs2_journal_dirty(handle, wc->w_di_bh);
2048
2049out:
2050	/* unlock pages before dealloc since it needs acquiring j_trans_barrier
2051	 * lock, or it will cause a deadlock since journal commit threads holds
2052	 * this lock and will ask for the page lock when flushing the data.
2053	 * put it here to preserve the unlock order.
2054	 */
2055	ocfs2_unlock_pages(wc);
2056
2057	if (handle)
2058		ocfs2_commit_trans(osb, handle);
2059
2060	ocfs2_run_deallocs(osb, &wc->w_dealloc);
2061
2062	brelse(wc->w_di_bh);
2063	kfree(wc);
2064
2065	return copied;
2066}
2067
2068static int ocfs2_write_end(struct file *file, struct address_space *mapping,
2069			   loff_t pos, unsigned len, unsigned copied,
2070			   struct page *page, void *fsdata)
2071{
2072	int ret;
2073	struct inode *inode = mapping->host;
2074
2075	ret = ocfs2_write_end_nolock(mapping, pos, len, copied, fsdata);
2076
2077	up_write(&OCFS2_I(inode)->ip_alloc_sem);
2078	ocfs2_inode_unlock(inode, 1);
2079
2080	return ret;
2081}
2082
2083struct ocfs2_dio_write_ctxt {
2084	struct list_head	dw_zero_list;
2085	unsigned		dw_zero_count;
2086	int			dw_orphaned;
2087	pid_t			dw_writer_pid;
2088};
2089
2090static struct ocfs2_dio_write_ctxt *
2091ocfs2_dio_alloc_write_ctx(struct buffer_head *bh, int *alloc)
2092{
2093	struct ocfs2_dio_write_ctxt *dwc = NULL;
2094
2095	if (bh->b_private)
2096		return bh->b_private;
2097
2098	dwc = kmalloc(sizeof(struct ocfs2_dio_write_ctxt), GFP_NOFS);
2099	if (dwc == NULL)
2100		return NULL;
2101	INIT_LIST_HEAD(&dwc->dw_zero_list);
2102	dwc->dw_zero_count = 0;
2103	dwc->dw_orphaned = 0;
2104	dwc->dw_writer_pid = task_pid_nr(current);
2105	bh->b_private = dwc;
2106	*alloc = 1;
2107
2108	return dwc;
2109}
2110
2111static void ocfs2_dio_free_write_ctx(struct inode *inode,
2112				     struct ocfs2_dio_write_ctxt *dwc)
2113{
2114	ocfs2_free_unwritten_list(inode, &dwc->dw_zero_list);
2115	kfree(dwc);
2116}
2117
2118/*
2119 * TODO: Make this into a generic get_blocks function.
2120 *
2121 * From do_direct_io in direct-io.c:
2122 *  "So what we do is to permit the ->get_blocks function to populate
2123 *   bh.b_size with the size of IO which is permitted at this offset and
2124 *   this i_blkbits."
2125 *
2126 * This function is called directly from get_more_blocks in direct-io.c.
2127 *
2128 * called like this: dio->get_blocks(dio->inode, fs_startblk,
2129 * 					fs_count, map_bh, dio->rw == WRITE);
2130 */
2131static int ocfs2_dio_get_block(struct inode *inode, sector_t iblock,
2132			       struct buffer_head *bh_result, int create)
2133{
2134	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2135	struct ocfs2_inode_info *oi = OCFS2_I(inode);
2136	struct ocfs2_write_ctxt *wc;
2137	struct ocfs2_write_cluster_desc *desc = NULL;
2138	struct ocfs2_dio_write_ctxt *dwc = NULL;
2139	struct buffer_head *di_bh = NULL;
2140	u64 p_blkno;
2141	loff_t pos = iblock << inode->i_sb->s_blocksize_bits;
 
 
2142	unsigned len, total_len = bh_result->b_size;
2143	int ret = 0, first_get_block = 0;
2144
2145	len = osb->s_clustersize - (pos & (osb->s_clustersize - 1));
2146	len = min(total_len, len);
2147
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2148	mlog(0, "get block of %lu at %llu:%u req %u\n",
2149			inode->i_ino, pos, len, total_len);
2150
2151	/*
2152	 * Because we need to change file size in ocfs2_dio_end_io_write(), or
2153	 * we may need to add it to orphan dir. So can not fall to fast path
2154	 * while file size will be changed.
2155	 */
2156	if (pos + total_len <= i_size_read(inode)) {
2157		down_read(&oi->ip_alloc_sem);
2158		/* This is the fast path for re-write. */
2159		ret = ocfs2_get_block(inode, iblock, bh_result, create);
2160
2161		up_read(&oi->ip_alloc_sem);
2162
 
 
2163		if (buffer_mapped(bh_result) &&
2164		    !buffer_new(bh_result) &&
2165		    ret == 0)
2166			goto out;
2167
2168		/* Clear state set by ocfs2_get_block. */
2169		bh_result->b_state = 0;
2170	}
2171
2172	dwc = ocfs2_dio_alloc_write_ctx(bh_result, &first_get_block);
2173	if (unlikely(dwc == NULL)) {
2174		ret = -ENOMEM;
2175		mlog_errno(ret);
2176		goto out;
2177	}
2178
2179	if (ocfs2_clusters_for_bytes(inode->i_sb, pos + total_len) >
2180	    ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode)) &&
2181	    !dwc->dw_orphaned) {
2182		/*
2183		 * when we are going to alloc extents beyond file size, add the
2184		 * inode to orphan dir, so we can recall those spaces when
2185		 * system crashed during write.
2186		 */
2187		ret = ocfs2_add_inode_to_orphan(osb, inode);
2188		if (ret < 0) {
2189			mlog_errno(ret);
2190			goto out;
2191		}
2192		dwc->dw_orphaned = 1;
2193	}
2194
2195	ret = ocfs2_inode_lock(inode, &di_bh, 1);
2196	if (ret) {
2197		mlog_errno(ret);
2198		goto out;
2199	}
2200
2201	down_write(&oi->ip_alloc_sem);
2202
2203	if (first_get_block) {
2204		if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
2205			ret = ocfs2_zero_tail(inode, di_bh, pos);
2206		else
2207			ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos,
2208							   total_len, NULL);
2209		if (ret < 0) {
2210			mlog_errno(ret);
2211			goto unlock;
2212		}
2213	}
2214
2215	ret = ocfs2_write_begin_nolock(inode->i_mapping, pos, len,
2216				       OCFS2_WRITE_DIRECT, NULL,
2217				       (void **)&wc, di_bh, NULL);
2218	if (ret) {
2219		mlog_errno(ret);
2220		goto unlock;
2221	}
2222
2223	desc = &wc->w_desc[0];
2224
2225	p_blkno = ocfs2_clusters_to_blocks(inode->i_sb, desc->c_phys);
2226	BUG_ON(p_blkno == 0);
2227	p_blkno += iblock & (u64)(ocfs2_clusters_to_blocks(inode->i_sb, 1) - 1);
2228
2229	map_bh(bh_result, inode->i_sb, p_blkno);
2230	bh_result->b_size = len;
2231	if (desc->c_needs_zero)
2232		set_buffer_new(bh_result);
2233
 
 
 
2234	/* May sleep in end_io. It should not happen in a irq context. So defer
2235	 * it to dio work queue. */
2236	set_buffer_defer_completion(bh_result);
2237
2238	if (!list_empty(&wc->w_unwritten_list)) {
2239		struct ocfs2_unwritten_extent *ue = NULL;
2240
2241		ue = list_first_entry(&wc->w_unwritten_list,
2242				      struct ocfs2_unwritten_extent,
2243				      ue_node);
2244		BUG_ON(ue->ue_cpos != desc->c_cpos);
2245		/* The physical address may be 0, fill it. */
2246		ue->ue_phys = desc->c_phys;
2247
2248		list_splice_tail_init(&wc->w_unwritten_list, &dwc->dw_zero_list);
2249		dwc->dw_zero_count++;
2250	}
2251
2252	ret = ocfs2_write_end_nolock(inode->i_mapping, pos, len, len, wc);
2253	BUG_ON(ret != len);
2254	ret = 0;
2255unlock:
2256	up_write(&oi->ip_alloc_sem);
2257	ocfs2_inode_unlock(inode, 1);
2258	brelse(di_bh);
2259out:
2260	if (ret < 0)
2261		ret = -EIO;
2262	return ret;
2263}
2264
2265static int ocfs2_dio_end_io_write(struct inode *inode,
2266				  struct ocfs2_dio_write_ctxt *dwc,
2267				  loff_t offset,
2268				  ssize_t bytes)
2269{
2270	struct ocfs2_cached_dealloc_ctxt dealloc;
2271	struct ocfs2_extent_tree et;
2272	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2273	struct ocfs2_inode_info *oi = OCFS2_I(inode);
2274	struct ocfs2_unwritten_extent *ue = NULL;
2275	struct buffer_head *di_bh = NULL;
2276	struct ocfs2_dinode *di;
2277	struct ocfs2_alloc_context *data_ac = NULL;
2278	struct ocfs2_alloc_context *meta_ac = NULL;
2279	handle_t *handle = NULL;
2280	loff_t end = offset + bytes;
2281	int ret = 0, credits = 0, locked = 0;
2282
2283	ocfs2_init_dealloc_ctxt(&dealloc);
2284
2285	/* We do clear unwritten, delete orphan, change i_size here. If neither
2286	 * of these happen, we can skip all this. */
2287	if (list_empty(&dwc->dw_zero_list) &&
2288	    end <= i_size_read(inode) &&
2289	    !dwc->dw_orphaned)
2290		goto out;
2291
2292	/* ocfs2_file_write_iter will get i_mutex, so we need not lock if we
2293	 * are in that context. */
2294	if (dwc->dw_writer_pid != task_pid_nr(current)) {
2295		inode_lock(inode);
2296		locked = 1;
2297	}
2298
2299	ret = ocfs2_inode_lock(inode, &di_bh, 1);
2300	if (ret < 0) {
2301		mlog_errno(ret);
2302		goto out;
2303	}
2304
2305	down_write(&oi->ip_alloc_sem);
2306
2307	/* Delete orphan before acquire i_mutex. */
2308	if (dwc->dw_orphaned) {
2309		BUG_ON(dwc->dw_writer_pid != task_pid_nr(current));
2310
2311		end = end > i_size_read(inode) ? end : 0;
2312
2313		ret = ocfs2_del_inode_from_orphan(osb, inode, di_bh,
2314				!!end, end);
2315		if (ret < 0)
2316			mlog_errno(ret);
2317	}
2318
2319	di = (struct ocfs2_dinode *)di_bh->b_data;
2320
2321	ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
2322
 
 
 
 
 
 
2323	ret = ocfs2_lock_allocators(inode, &et, 0, dwc->dw_zero_count*2,
2324				    &data_ac, &meta_ac);
2325	if (ret) {
2326		mlog_errno(ret);
2327		goto unlock;
2328	}
2329
2330	credits = ocfs2_calc_extend_credits(inode->i_sb, &di->id2.i_list);
2331
2332	handle = ocfs2_start_trans(osb, credits);
2333	if (IS_ERR(handle)) {
2334		ret = PTR_ERR(handle);
2335		mlog_errno(ret);
2336		goto unlock;
2337	}
2338	ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
2339				      OCFS2_JOURNAL_ACCESS_WRITE);
2340	if (ret) {
2341		mlog_errno(ret);
2342		goto commit;
2343	}
2344
2345	list_for_each_entry(ue, &dwc->dw_zero_list, ue_node) {
2346		ret = ocfs2_mark_extent_written(inode, &et, handle,
2347						ue->ue_cpos, 1,
2348						ue->ue_phys,
2349						meta_ac, &dealloc);
2350		if (ret < 0) {
2351			mlog_errno(ret);
2352			break;
2353		}
2354	}
2355
2356	if (end > i_size_read(inode)) {
2357		ret = ocfs2_set_inode_size(handle, inode, di_bh, end);
2358		if (ret < 0)
2359			mlog_errno(ret);
2360	}
2361commit:
2362	ocfs2_commit_trans(osb, handle);
2363unlock:
2364	up_write(&oi->ip_alloc_sem);
2365	ocfs2_inode_unlock(inode, 1);
2366	brelse(di_bh);
2367out:
2368	if (data_ac)
2369		ocfs2_free_alloc_context(data_ac);
2370	if (meta_ac)
2371		ocfs2_free_alloc_context(meta_ac);
2372	ocfs2_run_deallocs(osb, &dealloc);
2373	if (locked)
2374		inode_unlock(inode);
2375	ocfs2_dio_free_write_ctx(inode, dwc);
2376
2377	return ret;
2378}
2379
2380/*
2381 * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
2382 * particularly interested in the aio/dio case.  We use the rw_lock DLM lock
2383 * to protect io on one node from truncation on another.
2384 */
2385static int ocfs2_dio_end_io(struct kiocb *iocb,
2386			    loff_t offset,
2387			    ssize_t bytes,
2388			    void *private)
2389{
2390	struct inode *inode = file_inode(iocb->ki_filp);
2391	int level;
2392	int ret = 0;
2393
2394	/* this io's submitter should not have unlocked this before we could */
2395	BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
2396
2397	if (bytes > 0 && private)
2398		ret = ocfs2_dio_end_io_write(inode, private, offset, bytes);
 
 
 
 
 
 
 
 
2399
2400	ocfs2_iocb_clear_rw_locked(iocb);
2401
2402	level = ocfs2_iocb_rw_locked_level(iocb);
2403	ocfs2_rw_unlock(inode, level);
2404	return ret;
2405}
2406
2407static ssize_t ocfs2_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2408{
2409	struct file *file = iocb->ki_filp;
2410	struct inode *inode = file->f_mapping->host;
2411	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2412	get_block_t *get_block;
2413
2414	/*
2415	 * Fallback to buffered I/O if we see an inode without
2416	 * extents.
2417	 */
2418	if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2419		return 0;
2420
2421	/* Fallback to buffered I/O if we do not support append dio. */
2422	if (iocb->ki_pos + iter->count > i_size_read(inode) &&
2423	    !ocfs2_supports_append_dio(osb))
2424		return 0;
2425
2426	if (iov_iter_rw(iter) == READ)
2427		get_block = ocfs2_get_block;
2428	else
2429		get_block = ocfs2_dio_get_block;
2430
2431	return __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev,
2432				    iter, get_block,
2433				    ocfs2_dio_end_io, NULL, 0);
2434}
2435
2436const struct address_space_operations ocfs2_aops = {
 
2437	.readpage		= ocfs2_readpage,
2438	.readpages		= ocfs2_readpages,
2439	.writepage		= ocfs2_writepage,
2440	.write_begin		= ocfs2_write_begin,
2441	.write_end		= ocfs2_write_end,
2442	.bmap			= ocfs2_bmap,
2443	.direct_IO		= ocfs2_direct_IO,
2444	.invalidatepage		= block_invalidatepage,
2445	.releasepage		= ocfs2_releasepage,
2446	.migratepage		= buffer_migrate_page,
2447	.is_partially_uptodate	= block_is_partially_uptodate,
2448	.error_remove_page	= generic_error_remove_page,
2449};