Linux Audio

Check our new training course

Loading...
v5.9
   1// SPDX-License-Identifier: GPL-2.0+
   2/*
   3 * linux/fs/jbd2/commit.c
   4 *
   5 * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
   6 *
   7 * Copyright 1998 Red Hat corp --- All Rights Reserved
   8 *
 
 
 
 
   9 * Journal commit routines for the generic filesystem journaling code;
  10 * part of the ext2fs journaling system.
  11 */
  12
  13#include <linux/time.h>
  14#include <linux/fs.h>
  15#include <linux/jbd2.h>
  16#include <linux/errno.h>
  17#include <linux/slab.h>
  18#include <linux/mm.h>
  19#include <linux/pagemap.h>
  20#include <linux/jiffies.h>
  21#include <linux/crc32.h>
  22#include <linux/writeback.h>
  23#include <linux/backing-dev.h>
  24#include <linux/bio.h>
  25#include <linux/blkdev.h>
  26#include <linux/bitops.h>
  27#include <trace/events/jbd2.h>
  28
  29/*
  30 * IO end handler for temporary buffer_heads handling writes to the journal.
  31 */
  32static void journal_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
  33{
  34	struct buffer_head *orig_bh = bh->b_private;
  35
  36	BUFFER_TRACE(bh, "");
  37	if (uptodate)
  38		set_buffer_uptodate(bh);
  39	else
  40		clear_buffer_uptodate(bh);
  41	if (orig_bh) {
  42		clear_bit_unlock(BH_Shadow, &orig_bh->b_state);
  43		smp_mb__after_atomic();
  44		wake_up_bit(&orig_bh->b_state, BH_Shadow);
  45	}
  46	unlock_buffer(bh);
  47}
  48
  49/*
  50 * When an ext4 file is truncated, it is possible that some pages are not
  51 * successfully freed, because they are attached to a committing transaction.
  52 * After the transaction commits, these pages are left on the LRU, with no
  53 * ->mapping, and with attached buffers.  These pages are trivially reclaimable
  54 * by the VM, but their apparent absence upsets the VM accounting, and it makes
  55 * the numbers in /proc/meminfo look odd.
  56 *
  57 * So here, we have a buffer which has just come off the forget list.  Look to
  58 * see if we can strip all buffers from the backing page.
  59 *
  60 * Called under lock_journal(), and possibly under journal_datalist_lock.  The
  61 * caller provided us with a ref against the buffer, and we drop that here.
  62 */
  63static void release_buffer_page(struct buffer_head *bh)
  64{
  65	struct page *page;
  66
  67	if (buffer_dirty(bh))
  68		goto nope;
  69	if (atomic_read(&bh->b_count) != 1)
  70		goto nope;
  71	page = bh->b_page;
  72	if (!page)
  73		goto nope;
  74	if (page->mapping)
  75		goto nope;
  76
  77	/* OK, it's a truncated page */
  78	if (!trylock_page(page))
  79		goto nope;
  80
  81	get_page(page);
  82	__brelse(bh);
  83	try_to_free_buffers(page);
  84	unlock_page(page);
  85	put_page(page);
  86	return;
  87
  88nope:
  89	__brelse(bh);
  90}
  91
  92static void jbd2_commit_block_csum_set(journal_t *j, struct buffer_head *bh)
  93{
  94	struct commit_header *h;
  95	__u32 csum;
  96
  97	if (!jbd2_journal_has_csum_v2or3(j))
  98		return;
  99
 100	h = (struct commit_header *)(bh->b_data);
 101	h->h_chksum_type = 0;
 102	h->h_chksum_size = 0;
 103	h->h_chksum[0] = 0;
 104	csum = jbd2_chksum(j, j->j_csum_seed, bh->b_data, j->j_blocksize);
 105	h->h_chksum[0] = cpu_to_be32(csum);
 106}
 107
 108/*
 109 * Done it all: now submit the commit record.  We should have
 110 * cleaned up our previous buffers by now, so if we are in abort
 111 * mode we can now just skip the rest of the journal write
 112 * entirely.
 113 *
 114 * Returns 1 if the journal needs to be aborted or 0 on success
 115 */
 116static int journal_submit_commit_record(journal_t *journal,
 117					transaction_t *commit_transaction,
 118					struct buffer_head **cbh,
 119					__u32 crc32_sum)
 120{
 121	struct commit_header *tmp;
 122	struct buffer_head *bh;
 123	int ret;
 124	struct timespec64 now;
 125
 126	*cbh = NULL;
 127
 128	if (is_journal_aborted(journal))
 129		return 0;
 130
 131	bh = jbd2_journal_get_descriptor_buffer(commit_transaction,
 132						JBD2_COMMIT_BLOCK);
 133	if (!bh)
 134		return 1;
 135
 136	tmp = (struct commit_header *)bh->b_data;
 137	ktime_get_coarse_real_ts64(&now);
 
 
 138	tmp->h_commit_sec = cpu_to_be64(now.tv_sec);
 139	tmp->h_commit_nsec = cpu_to_be32(now.tv_nsec);
 140
 141	if (jbd2_has_feature_checksum(journal)) {
 
 142		tmp->h_chksum_type 	= JBD2_CRC32_CHKSUM;
 143		tmp->h_chksum_size 	= JBD2_CRC32_CHKSUM_SIZE;
 144		tmp->h_chksum[0] 	= cpu_to_be32(crc32_sum);
 145	}
 146	jbd2_commit_block_csum_set(journal, bh);
 147
 148	BUFFER_TRACE(bh, "submit commit block");
 149	lock_buffer(bh);
 150	clear_buffer_dirty(bh);
 151	set_buffer_uptodate(bh);
 152	bh->b_end_io = journal_end_buffer_io_sync;
 153
 154	if (journal->j_flags & JBD2_BARRIER &&
 155	    !jbd2_has_feature_async_commit(journal))
 156		ret = submit_bh(REQ_OP_WRITE,
 157			REQ_SYNC | REQ_PREFLUSH | REQ_FUA, bh);
 158	else
 159		ret = submit_bh(REQ_OP_WRITE, REQ_SYNC, bh);
 160
 161	*cbh = bh;
 162	return ret;
 163}
 164
 165/*
 166 * This function along with journal_submit_commit_record
 167 * allows to write the commit record asynchronously.
 168 */
 169static int journal_wait_on_commit_record(journal_t *journal,
 170					 struct buffer_head *bh)
 171{
 172	int ret = 0;
 173
 174	clear_buffer_dirty(bh);
 175	wait_on_buffer(bh);
 176
 177	if (unlikely(!buffer_uptodate(bh)))
 178		ret = -EIO;
 179	put_bh(bh);            /* One for getblk() */
 180
 181	return ret;
 182}
 183
 184/*
 185 * write the filemap data using writepage() address_space_operations.
 186 * We don't do block allocation here even for delalloc. We don't
 187 * use writepages() because with delayed allocation we may be doing
 188 * block allocation in writepages().
 189 */
 190static int journal_submit_inode_data_buffers(struct address_space *mapping,
 191		loff_t dirty_start, loff_t dirty_end)
 192{
 193	int ret;
 194	struct writeback_control wbc = {
 195		.sync_mode =  WB_SYNC_ALL,
 196		.nr_to_write = mapping->nrpages * 2,
 197		.range_start = dirty_start,
 198		.range_end = dirty_end,
 199	};
 200
 201	ret = generic_writepages(mapping, &wbc);
 202	return ret;
 203}
 204
 205/*
 206 * Submit all the data buffers of inode associated with the transaction to
 207 * disk.
 208 *
 209 * We are in a committing transaction. Therefore no new inode can be added to
 210 * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
 211 * operate on from being released while we write out pages.
 212 */
 213static int journal_submit_data_buffers(journal_t *journal,
 214		transaction_t *commit_transaction)
 215{
 216	struct jbd2_inode *jinode;
 217	int err, ret = 0;
 218	struct address_space *mapping;
 219
 220	spin_lock(&journal->j_list_lock);
 221	list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
 222		loff_t dirty_start = jinode->i_dirty_start;
 223		loff_t dirty_end = jinode->i_dirty_end;
 224
 225		if (!(jinode->i_flags & JI_WRITE_DATA))
 226			continue;
 227		mapping = jinode->i_vfs_inode->i_mapping;
 228		jinode->i_flags |= JI_COMMIT_RUNNING;
 229		spin_unlock(&journal->j_list_lock);
 230		/*
 231		 * submit the inode data buffers. We use writepage
 232		 * instead of writepages. Because writepages can do
 233		 * block allocation  with delalloc. We need to write
 234		 * only allocated blocks here.
 235		 */
 236		trace_jbd2_submit_inode_data(jinode->i_vfs_inode);
 237		err = journal_submit_inode_data_buffers(mapping, dirty_start,
 238				dirty_end);
 239		if (!ret)
 240			ret = err;
 241		spin_lock(&journal->j_list_lock);
 242		J_ASSERT(jinode->i_transaction == commit_transaction);
 243		jinode->i_flags &= ~JI_COMMIT_RUNNING;
 244		smp_mb();
 245		wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
 246	}
 247	spin_unlock(&journal->j_list_lock);
 248	return ret;
 249}
 250
 251/*
 252 * Wait for data submitted for writeout, refile inodes to proper
 253 * transaction if needed.
 254 *
 255 */
 256static int journal_finish_inode_data_buffers(journal_t *journal,
 257		transaction_t *commit_transaction)
 258{
 259	struct jbd2_inode *jinode, *next_i;
 260	int err, ret = 0;
 261
 262	/* For locking, see the comment in journal_submit_data_buffers() */
 263	spin_lock(&journal->j_list_lock);
 264	list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
 265		loff_t dirty_start = jinode->i_dirty_start;
 266		loff_t dirty_end = jinode->i_dirty_end;
 267
 268		if (!(jinode->i_flags & JI_WAIT_DATA))
 269			continue;
 270		jinode->i_flags |= JI_COMMIT_RUNNING;
 271		spin_unlock(&journal->j_list_lock);
 272		err = filemap_fdatawait_range_keep_errors(
 273				jinode->i_vfs_inode->i_mapping, dirty_start,
 274				dirty_end);
 275		if (!ret)
 276			ret = err;
 
 
 
 
 
 
 
 
 277		spin_lock(&journal->j_list_lock);
 278		jinode->i_flags &= ~JI_COMMIT_RUNNING;
 279		smp_mb();
 280		wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
 281	}
 282
 283	/* Now refile inode to proper lists */
 284	list_for_each_entry_safe(jinode, next_i,
 285				 &commit_transaction->t_inode_list, i_list) {
 286		list_del(&jinode->i_list);
 287		if (jinode->i_next_transaction) {
 288			jinode->i_transaction = jinode->i_next_transaction;
 289			jinode->i_next_transaction = NULL;
 290			list_add(&jinode->i_list,
 291				&jinode->i_transaction->t_inode_list);
 292		} else {
 293			jinode->i_transaction = NULL;
 294			jinode->i_dirty_start = 0;
 295			jinode->i_dirty_end = 0;
 296		}
 297	}
 298	spin_unlock(&journal->j_list_lock);
 299
 300	return ret;
 301}
 302
 303static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
 304{
 305	struct page *page = bh->b_page;
 306	char *addr;
 307	__u32 checksum;
 308
 309	addr = kmap_atomic(page);
 310	checksum = crc32_be(crc32_sum,
 311		(void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
 312	kunmap_atomic(addr);
 313
 314	return checksum;
 315}
 316
 317static void write_tag_block(journal_t *j, journal_block_tag_t *tag,
 318				   unsigned long long block)
 319{
 320	tag->t_blocknr = cpu_to_be32(block & (u32)~0);
 321	if (jbd2_has_feature_64bit(j))
 322		tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
 323}
 324
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 325static void jbd2_block_tag_csum_set(journal_t *j, journal_block_tag_t *tag,
 326				    struct buffer_head *bh, __u32 sequence)
 327{
 328	journal_block_tag3_t *tag3 = (journal_block_tag3_t *)tag;
 329	struct page *page = bh->b_page;
 330	__u8 *addr;
 331	__u32 csum32;
 332	__be32 seq;
 333
 334	if (!jbd2_journal_has_csum_v2or3(j))
 335		return;
 336
 337	seq = cpu_to_be32(sequence);
 338	addr = kmap_atomic(page);
 339	csum32 = jbd2_chksum(j, j->j_csum_seed, (__u8 *)&seq, sizeof(seq));
 340	csum32 = jbd2_chksum(j, csum32, addr + offset_in_page(bh->b_data),
 341			     bh->b_size);
 342	kunmap_atomic(addr);
 343
 344	if (jbd2_has_feature_csum3(j))
 345		tag3->t_checksum = cpu_to_be32(csum32);
 346	else
 347		tag->t_checksum = cpu_to_be16(csum32);
 348}
 349/*
 350 * jbd2_journal_commit_transaction
 351 *
 352 * The primary function for committing a transaction to the log.  This
 353 * function is called by the journal thread to begin a complete commit.
 354 */
 355void jbd2_journal_commit_transaction(journal_t *journal)
 356{
 357	struct transaction_stats_s stats;
 358	transaction_t *commit_transaction;
 359	struct journal_head *jh;
 360	struct buffer_head *descriptor;
 361	struct buffer_head **wbuf = journal->j_wbuf;
 362	int bufs;
 363	int flags;
 364	int err;
 365	unsigned long long blocknr;
 366	ktime_t start_time;
 367	u64 commit_time;
 368	char *tagp = NULL;
 
 369	journal_block_tag_t *tag = NULL;
 370	int space_left = 0;
 371	int first_tag = 0;
 372	int tag_flag;
 373	int i;
 374	int tag_bytes = journal_tag_bytes(journal);
 375	struct buffer_head *cbh = NULL; /* For transactional checksums */
 376	__u32 crc32_sum = ~0;
 377	struct blk_plug plug;
 378	/* Tail of the journal */
 379	unsigned long first_block;
 380	tid_t first_tid;
 381	int update_tail;
 382	int csum_size = 0;
 383	LIST_HEAD(io_bufs);
 384	LIST_HEAD(log_bufs);
 385
 386	if (jbd2_journal_has_csum_v2or3(journal))
 387		csum_size = sizeof(struct jbd2_journal_block_tail);
 388
 389	/*
 390	 * First job: lock down the current transaction and wait for
 391	 * all outstanding updates to complete.
 392	 */
 393
 394	/* Do we need to erase the effects of a prior jbd2_journal_flush? */
 395	if (journal->j_flags & JBD2_FLUSHED) {
 396		jbd_debug(3, "super block updated\n");
 397		mutex_lock_io(&journal->j_checkpoint_mutex);
 398		/*
 399		 * We hold j_checkpoint_mutex so tail cannot change under us.
 400		 * We don't need any special data guarantees for writing sb
 401		 * since journal is empty and it is ok for write to be
 402		 * flushed only with transaction commit.
 403		 */
 404		jbd2_journal_update_sb_log_tail(journal,
 405						journal->j_tail_sequence,
 406						journal->j_tail,
 407						REQ_SYNC);
 408		mutex_unlock(&journal->j_checkpoint_mutex);
 409	} else {
 410		jbd_debug(3, "superblock not updated\n");
 411	}
 412
 413	J_ASSERT(journal->j_running_transaction != NULL);
 414	J_ASSERT(journal->j_committing_transaction == NULL);
 415
 416	commit_transaction = journal->j_running_transaction;
 417
 418	trace_jbd2_start_commit(journal, commit_transaction);
 419	jbd_debug(1, "JBD2: starting commit of transaction %d\n",
 420			commit_transaction->t_tid);
 421
 422	write_lock(&journal->j_state_lock);
 423	J_ASSERT(commit_transaction->t_state == T_RUNNING);
 424	commit_transaction->t_state = T_LOCKED;
 425
 426	trace_jbd2_commit_locking(journal, commit_transaction);
 427	stats.run.rs_wait = commit_transaction->t_max_wait;
 428	stats.run.rs_request_delay = 0;
 429	stats.run.rs_locked = jiffies;
 430	if (commit_transaction->t_requested)
 431		stats.run.rs_request_delay =
 432			jbd2_time_diff(commit_transaction->t_requested,
 433				       stats.run.rs_locked);
 434	stats.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
 435					      stats.run.rs_locked);
 436
 437	spin_lock(&commit_transaction->t_handle_lock);
 438	while (atomic_read(&commit_transaction->t_updates)) {
 439		DEFINE_WAIT(wait);
 440
 441		prepare_to_wait(&journal->j_wait_updates, &wait,
 442					TASK_UNINTERRUPTIBLE);
 443		if (atomic_read(&commit_transaction->t_updates)) {
 444			spin_unlock(&commit_transaction->t_handle_lock);
 445			write_unlock(&journal->j_state_lock);
 446			schedule();
 447			write_lock(&journal->j_state_lock);
 448			spin_lock(&commit_transaction->t_handle_lock);
 449		}
 450		finish_wait(&journal->j_wait_updates, &wait);
 451	}
 452	spin_unlock(&commit_transaction->t_handle_lock);
 453	commit_transaction->t_state = T_SWITCH;
 454	write_unlock(&journal->j_state_lock);
 455
 456	J_ASSERT (atomic_read(&commit_transaction->t_outstanding_credits) <=
 457			journal->j_max_transaction_buffers);
 458
 459	/*
 460	 * First thing we are allowed to do is to discard any remaining
 461	 * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
 462	 * that there are no such buffers: if a large filesystem
 463	 * operation like a truncate needs to split itself over multiple
 464	 * transactions, then it may try to do a jbd2_journal_restart() while
 465	 * there are still BJ_Reserved buffers outstanding.  These must
 466	 * be released cleanly from the current transaction.
 467	 *
 468	 * In this case, the filesystem must still reserve write access
 469	 * again before modifying the buffer in the new transaction, but
 470	 * we do not require it to remember exactly which old buffers it
 471	 * has reserved.  This is consistent with the existing behaviour
 472	 * that multiple jbd2_journal_get_write_access() calls to the same
 473	 * buffer are perfectly permissible.
 474	 */
 475	while (commit_transaction->t_reserved_list) {
 476		jh = commit_transaction->t_reserved_list;
 477		JBUFFER_TRACE(jh, "reserved, unused: refile");
 478		/*
 479		 * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
 480		 * leave undo-committed data.
 481		 */
 482		if (jh->b_committed_data) {
 483			struct buffer_head *bh = jh2bh(jh);
 484
 485			spin_lock(&jh->b_state_lock);
 486			jbd2_free(jh->b_committed_data, bh->b_size);
 487			jh->b_committed_data = NULL;
 488			spin_unlock(&jh->b_state_lock);
 489		}
 490		jbd2_journal_refile_buffer(journal, jh);
 491	}
 492
 493	/*
 494	 * Now try to drop any written-back buffers from the journal's
 495	 * checkpoint lists.  We do this *before* commit because it potentially
 496	 * frees some memory
 497	 */
 498	spin_lock(&journal->j_list_lock);
 499	__jbd2_journal_clean_checkpoint_list(journal, false);
 500	spin_unlock(&journal->j_list_lock);
 501
 502	jbd_debug(3, "JBD2: commit phase 1\n");
 503
 504	/*
 505	 * Clear revoked flag to reflect there is no revoked buffers
 506	 * in the next transaction which is going to be started.
 507	 */
 508	jbd2_clear_buffer_revoked_flags(journal);
 509
 510	/*
 511	 * Switch to a new revoke table.
 512	 */
 513	jbd2_journal_switch_revoke_table(journal);
 514
 515	/*
 516	 * Reserved credits cannot be claimed anymore, free them
 517	 */
 518	atomic_sub(atomic_read(&journal->j_reserved_credits),
 519		   &commit_transaction->t_outstanding_credits);
 520
 521	write_lock(&journal->j_state_lock);
 522	trace_jbd2_commit_flushing(journal, commit_transaction);
 523	stats.run.rs_flushing = jiffies;
 524	stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
 525					     stats.run.rs_flushing);
 526
 527	commit_transaction->t_state = T_FLUSH;
 528	journal->j_committing_transaction = commit_transaction;
 529	journal->j_running_transaction = NULL;
 530	start_time = ktime_get();
 531	commit_transaction->t_log_start = journal->j_head;
 532	wake_up(&journal->j_wait_transaction_locked);
 533	write_unlock(&journal->j_state_lock);
 534
 535	jbd_debug(3, "JBD2: commit phase 2a\n");
 536
 537	/*
 538	 * Now start flushing things to disk, in the order they appear
 539	 * on the transaction lists.  Data blocks go first.
 540	 */
 541	err = journal_submit_data_buffers(journal, commit_transaction);
 542	if (err)
 543		jbd2_journal_abort(journal, err);
 544
 545	blk_start_plug(&plug);
 546	jbd2_journal_write_revoke_records(commit_transaction, &log_bufs);
 
 547
 548	jbd_debug(3, "JBD2: commit phase 2b\n");
 549
 550	/*
 551	 * Way to go: we have now written out all of the data for a
 552	 * transaction!  Now comes the tricky part: we need to write out
 553	 * metadata.  Loop over the transaction's entire buffer list:
 554	 */
 555	write_lock(&journal->j_state_lock);
 556	commit_transaction->t_state = T_COMMIT;
 557	write_unlock(&journal->j_state_lock);
 558
 559	trace_jbd2_commit_logging(journal, commit_transaction);
 560	stats.run.rs_logging = jiffies;
 561	stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
 562					       stats.run.rs_logging);
 563	stats.run.rs_blocks = commit_transaction->t_nr_buffers;
 
 564	stats.run.rs_blocks_logged = 0;
 565
 566	J_ASSERT(commit_transaction->t_nr_buffers <=
 567		 atomic_read(&commit_transaction->t_outstanding_credits));
 568
 569	err = 0;
 570	bufs = 0;
 571	descriptor = NULL;
 572	while (commit_transaction->t_buffers) {
 573
 574		/* Find the next buffer to be journaled... */
 575
 576		jh = commit_transaction->t_buffers;
 577
 578		/* If we're in abort mode, we just un-journal the buffer and
 579		   release it. */
 580
 581		if (is_journal_aborted(journal)) {
 582			clear_buffer_jbddirty(jh2bh(jh));
 583			JBUFFER_TRACE(jh, "journal is aborting: refile");
 584			jbd2_buffer_abort_trigger(jh,
 585						  jh->b_frozen_data ?
 586						  jh->b_frozen_triggers :
 587						  jh->b_triggers);
 588			jbd2_journal_refile_buffer(journal, jh);
 589			/* If that was the last one, we need to clean up
 590			 * any descriptor buffers which may have been
 591			 * already allocated, even if we are now
 592			 * aborting. */
 593			if (!commit_transaction->t_buffers)
 594				goto start_journal_io;
 595			continue;
 596		}
 597
 598		/* Make sure we have a descriptor block in which to
 599		   record the metadata buffer. */
 600
 601		if (!descriptor) {
 602			J_ASSERT (bufs == 0);
 603
 604			jbd_debug(4, "JBD2: get descriptor\n");
 605
 606			descriptor = jbd2_journal_get_descriptor_buffer(
 607							commit_transaction,
 608							JBD2_DESCRIPTOR_BLOCK);
 609			if (!descriptor) {
 610				jbd2_journal_abort(journal, -EIO);
 611				continue;
 612			}
 613
 614			jbd_debug(4, "JBD2: got buffer %llu (%p)\n",
 615				(unsigned long long)descriptor->b_blocknr,
 616				descriptor->b_data);
 
 
 
 
 
 617			tagp = &descriptor->b_data[sizeof(journal_header_t)];
 618			space_left = descriptor->b_size -
 619						sizeof(journal_header_t);
 620			first_tag = 1;
 621			set_buffer_jwrite(descriptor);
 622			set_buffer_dirty(descriptor);
 623			wbuf[bufs++] = descriptor;
 624
 625			/* Record it so that we can wait for IO
 626                           completion later */
 627			BUFFER_TRACE(descriptor, "ph3: file as descriptor");
 628			jbd2_file_log_bh(&log_bufs, descriptor);
 629		}
 630
 631		/* Where is the buffer to be written? */
 632
 633		err = jbd2_journal_next_log_block(journal, &blocknr);
 634		/* If the block mapping failed, just abandon the buffer
 635		   and repeat this loop: we'll fall into the
 636		   refile-on-abort condition above. */
 637		if (err) {
 638			jbd2_journal_abort(journal, err);
 639			continue;
 640		}
 641
 642		/*
 643		 * start_this_handle() uses t_outstanding_credits to determine
 644		 * the free space in the log.
 
 645		 */
 646		atomic_dec(&commit_transaction->t_outstanding_credits);
 647
 648		/* Bump b_count to prevent truncate from stumbling over
 649                   the shadowed buffer!  @@@ This can go if we ever get
 650                   rid of the shadow pairing of buffers. */
 651		atomic_inc(&jh2bh(jh)->b_count);
 652
 653		/*
 654		 * Make a temporary IO buffer with which to write it out
 655		 * (this will requeue the metadata buffer to BJ_Shadow).
 656		 */
 657		set_bit(BH_JWrite, &jh2bh(jh)->b_state);
 658		JBUFFER_TRACE(jh, "ph3: write metadata");
 659		flags = jbd2_journal_write_metadata_buffer(commit_transaction,
 660						jh, &wbuf[bufs], blocknr);
 661		if (flags < 0) {
 662			jbd2_journal_abort(journal, flags);
 663			continue;
 664		}
 665		jbd2_file_log_bh(&io_bufs, wbuf[bufs]);
 666
 667		/* Record the new block's tag in the current descriptor
 668                   buffer */
 669
 670		tag_flag = 0;
 671		if (flags & 1)
 672			tag_flag |= JBD2_FLAG_ESCAPE;
 673		if (!first_tag)
 674			tag_flag |= JBD2_FLAG_SAME_UUID;
 675
 676		tag = (journal_block_tag_t *) tagp;
 677		write_tag_block(journal, tag, jh2bh(jh)->b_blocknr);
 678		tag->t_flags = cpu_to_be16(tag_flag);
 679		jbd2_block_tag_csum_set(journal, tag, wbuf[bufs],
 680					commit_transaction->t_tid);
 681		tagp += tag_bytes;
 682		space_left -= tag_bytes;
 683		bufs++;
 684
 685		if (first_tag) {
 686			memcpy (tagp, journal->j_uuid, 16);
 687			tagp += 16;
 688			space_left -= 16;
 689			first_tag = 0;
 690		}
 691
 692		/* If there's no more to do, or if the descriptor is full,
 693		   let the IO rip! */
 694
 695		if (bufs == journal->j_wbufsize ||
 696		    commit_transaction->t_buffers == NULL ||
 697		    space_left < tag_bytes + 16 + csum_size) {
 698
 699			jbd_debug(4, "JBD2: Submit %d IOs\n", bufs);
 700
 701			/* Write an end-of-descriptor marker before
 702                           submitting the IOs.  "tag" still points to
 703                           the last tag we set up. */
 704
 705			tag->t_flags |= cpu_to_be16(JBD2_FLAG_LAST_TAG);
 706start_journal_io:
 707			if (descriptor)
 708				jbd2_descriptor_block_csum_set(journal,
 709							descriptor);
 710
 
 
 711			for (i = 0; i < bufs; i++) {
 712				struct buffer_head *bh = wbuf[i];
 713				/*
 714				 * Compute checksum.
 715				 */
 716				if (jbd2_has_feature_checksum(journal)) {
 
 717					crc32_sum =
 718					    jbd2_checksum_data(crc32_sum, bh);
 719				}
 720
 721				lock_buffer(bh);
 722				clear_buffer_dirty(bh);
 723				set_buffer_uptodate(bh);
 724				bh->b_end_io = journal_end_buffer_io_sync;
 725				submit_bh(REQ_OP_WRITE, REQ_SYNC, bh);
 726			}
 727			cond_resched();
 
 728
 729			/* Force a new descriptor to be generated next
 730                           time round the loop. */
 731			descriptor = NULL;
 732			bufs = 0;
 733		}
 734	}
 735
 736	err = journal_finish_inode_data_buffers(journal, commit_transaction);
 737	if (err) {
 738		printk(KERN_WARNING
 739			"JBD2: Detected IO errors while flushing file data "
 740		       "on %s\n", journal->j_devname);
 741		if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
 742			jbd2_journal_abort(journal, err);
 743		err = 0;
 744	}
 745
 746	/*
 747	 * Get current oldest transaction in the log before we issue flush
 748	 * to the filesystem device. After the flush we can be sure that
 749	 * blocks of all older transactions are checkpointed to persistent
 750	 * storage and we will be safe to update journal start in the
 751	 * superblock with the numbers we get here.
 752	 */
 753	update_tail =
 754		jbd2_journal_get_log_tail(journal, &first_tid, &first_block);
 755
 756	write_lock(&journal->j_state_lock);
 757	if (update_tail) {
 758		long freed = first_block - journal->j_tail;
 759
 760		if (first_block < journal->j_tail)
 761			freed += journal->j_last - journal->j_first;
 762		/* Update tail only if we free significant amount of space */
 763		if (freed < journal->j_maxlen / 4)
 764			update_tail = 0;
 765	}
 766	J_ASSERT(commit_transaction->t_state == T_COMMIT);
 767	commit_transaction->t_state = T_COMMIT_DFLUSH;
 768	write_unlock(&journal->j_state_lock);
 769
 770	/* 
 771	 * If the journal is not located on the file system device,
 772	 * then we must flush the file system device before we issue
 773	 * the commit record
 774	 */
 775	if (commit_transaction->t_need_data_flush &&
 776	    (journal->j_fs_dev != journal->j_dev) &&
 777	    (journal->j_flags & JBD2_BARRIER))
 778		blkdev_issue_flush(journal->j_fs_dev, GFP_NOFS);
 779
 780	/* Done it all: now write the commit record asynchronously. */
 781	if (jbd2_has_feature_async_commit(journal)) {
 
 782		err = journal_submit_commit_record(journal, commit_transaction,
 783						 &cbh, crc32_sum);
 784		if (err)
 785			jbd2_journal_abort(journal, err);
 786	}
 787
 788	blk_finish_plug(&plug);
 789
 790	/* Lo and behold: we have just managed to send a transaction to
 791           the log.  Before we can commit it, wait for the IO so far to
 792           complete.  Control buffers being written are on the
 793           transaction's t_log_list queue, and metadata buffers are on
 794           the io_bufs list.
 795
 796	   Wait for the buffers in reverse order.  That way we are
 797	   less likely to be woken up until all IOs have completed, and
 798	   so we incur less scheduling load.
 799	*/
 800
 801	jbd_debug(3, "JBD2: commit phase 3\n");
 802
 803	while (!list_empty(&io_bufs)) {
 804		struct buffer_head *bh = list_entry(io_bufs.prev,
 805						    struct buffer_head,
 806						    b_assoc_buffers);
 807
 808		wait_on_buffer(bh);
 809		cond_resched();
 810
 811		if (unlikely(!buffer_uptodate(bh)))
 812			err = -EIO;
 813		jbd2_unfile_log_bh(bh);
 814		stats.run.rs_blocks_logged++;
 815
 816		/*
 817		 * The list contains temporary buffer heads created by
 818		 * jbd2_journal_write_metadata_buffer().
 819		 */
 820		BUFFER_TRACE(bh, "dumping temporary bh");
 821		__brelse(bh);
 822		J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
 823		free_buffer_head(bh);
 824
 825		/* We also have to refile the corresponding shadowed buffer */
 826		jh = commit_transaction->t_shadow_list->b_tprev;
 827		bh = jh2bh(jh);
 828		clear_buffer_jwrite(bh);
 829		J_ASSERT_BH(bh, buffer_jbddirty(bh));
 830		J_ASSERT_BH(bh, !buffer_shadow(bh));
 831
 832		/* The metadata is now released for reuse, but we need
 833                   to remember it against this transaction so that when
 834                   we finally commit, we can do any checkpointing
 835                   required. */
 836		JBUFFER_TRACE(jh, "file as BJ_Forget");
 837		jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
 838		JBUFFER_TRACE(jh, "brelse shadowed buffer");
 839		__brelse(bh);
 840	}
 841
 842	J_ASSERT (commit_transaction->t_shadow_list == NULL);
 843
 844	jbd_debug(3, "JBD2: commit phase 4\n");
 845
 846	/* Here we wait for the revoke record and descriptor record buffers */
 847	while (!list_empty(&log_bufs)) {
 848		struct buffer_head *bh;
 849
 850		bh = list_entry(log_bufs.prev, struct buffer_head, b_assoc_buffers);
 851		wait_on_buffer(bh);
 852		cond_resched();
 853
 854		if (unlikely(!buffer_uptodate(bh)))
 855			err = -EIO;
 856
 857		BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
 858		clear_buffer_jwrite(bh);
 859		jbd2_unfile_log_bh(bh);
 860		stats.run.rs_blocks_logged++;
 861		__brelse(bh);		/* One for getblk */
 862		/* AKPM: bforget here */
 863	}
 864
 865	if (err)
 866		jbd2_journal_abort(journal, err);
 867
 868	jbd_debug(3, "JBD2: commit phase 5\n");
 869	write_lock(&journal->j_state_lock);
 870	J_ASSERT(commit_transaction->t_state == T_COMMIT_DFLUSH);
 871	commit_transaction->t_state = T_COMMIT_JFLUSH;
 872	write_unlock(&journal->j_state_lock);
 873
 874	if (!jbd2_has_feature_async_commit(journal)) {
 
 875		err = journal_submit_commit_record(journal, commit_transaction,
 876						&cbh, crc32_sum);
 877		if (err)
 878			jbd2_journal_abort(journal, err);
 879	}
 880	if (cbh)
 881		err = journal_wait_on_commit_record(journal, cbh);
 882	stats.run.rs_blocks_logged++;
 883	if (jbd2_has_feature_async_commit(journal) &&
 884	    journal->j_flags & JBD2_BARRIER) {
 885		blkdev_issue_flush(journal->j_dev, GFP_NOFS);
 886	}
 887
 888	if (err)
 889		jbd2_journal_abort(journal, err);
 890
 891	WARN_ON_ONCE(
 892		atomic_read(&commit_transaction->t_outstanding_credits) < 0);
 893
 894	/*
 895	 * Now disk caches for filesystem device are flushed so we are safe to
 896	 * erase checkpointed transactions from the log by updating journal
 897	 * superblock.
 898	 */
 899	if (update_tail)
 900		jbd2_update_log_tail(journal, first_tid, first_block);
 901
 902	/* End of a transaction!  Finally, we can do checkpoint
 903           processing: any buffers committed as a result of this
 904           transaction can be removed from any checkpoint list it was on
 905           before. */
 906
 907	jbd_debug(3, "JBD2: commit phase 6\n");
 908
 909	J_ASSERT(list_empty(&commit_transaction->t_inode_list));
 910	J_ASSERT(commit_transaction->t_buffers == NULL);
 911	J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
 912	J_ASSERT(commit_transaction->t_shadow_list == NULL);
 913
 914restart_loop:
 915	/*
 916	 * As there are other places (journal_unmap_buffer()) adding buffers
 917	 * to this list we have to be careful and hold the j_list_lock.
 918	 */
 919	spin_lock(&journal->j_list_lock);
 920	while (commit_transaction->t_forget) {
 921		transaction_t *cp_transaction;
 922		struct buffer_head *bh;
 923		int try_to_free = 0;
 924		bool drop_ref;
 925
 926		jh = commit_transaction->t_forget;
 927		spin_unlock(&journal->j_list_lock);
 928		bh = jh2bh(jh);
 929		/*
 930		 * Get a reference so that bh cannot be freed before we are
 931		 * done with it.
 932		 */
 933		get_bh(bh);
 934		spin_lock(&jh->b_state_lock);
 935		J_ASSERT_JH(jh,	jh->b_transaction == commit_transaction);
 936
 937		/*
 938		 * If there is undo-protected committed data against
 939		 * this buffer, then we can remove it now.  If it is a
 940		 * buffer needing such protection, the old frozen_data
 941		 * field now points to a committed version of the
 942		 * buffer, so rotate that field to the new committed
 943		 * data.
 944		 *
 945		 * Otherwise, we can just throw away the frozen data now.
 946		 *
 947		 * We also know that the frozen data has already fired
 948		 * its triggers if they exist, so we can clear that too.
 949		 */
 950		if (jh->b_committed_data) {
 951			jbd2_free(jh->b_committed_data, bh->b_size);
 952			jh->b_committed_data = NULL;
 953			if (jh->b_frozen_data) {
 954				jh->b_committed_data = jh->b_frozen_data;
 955				jh->b_frozen_data = NULL;
 956				jh->b_frozen_triggers = NULL;
 957			}
 958		} else if (jh->b_frozen_data) {
 959			jbd2_free(jh->b_frozen_data, bh->b_size);
 960			jh->b_frozen_data = NULL;
 961			jh->b_frozen_triggers = NULL;
 962		}
 963
 964		spin_lock(&journal->j_list_lock);
 965		cp_transaction = jh->b_cp_transaction;
 966		if (cp_transaction) {
 967			JBUFFER_TRACE(jh, "remove from old cp transaction");
 968			cp_transaction->t_chp_stats.cs_dropped++;
 969			__jbd2_journal_remove_checkpoint(jh);
 970		}
 971
 972		/* Only re-checkpoint the buffer_head if it is marked
 973		 * dirty.  If the buffer was added to the BJ_Forget list
 974		 * by jbd2_journal_forget, it may no longer be dirty and
 975		 * there's no point in keeping a checkpoint record for
 976		 * it. */
 977
 978		/*
 979		 * A buffer which has been freed while still being journaled
 980		 * by a previous transaction, refile the buffer to BJ_Forget of
 981		 * the running transaction. If the just committed transaction
 982		 * contains "add to orphan" operation, we can completely
 983		 * invalidate the buffer now. We are rather through in that
 984		 * since the buffer may be still accessible when blocksize <
 985		 * pagesize and it is attached to the last partial page.
 986		 */
 987		if (buffer_freed(bh) && !jh->b_next_transaction) {
 988			struct address_space *mapping;
 989
 990			clear_buffer_freed(bh);
 991			clear_buffer_jbddirty(bh);
 992
 993			/*
 994			 * Block device buffers need to stay mapped all the
 995			 * time, so it is enough to clear buffer_jbddirty and
 996			 * buffer_freed bits. For the file mapping buffers (i.e.
 997			 * journalled data) we need to unmap buffer and clear
 998			 * more bits. We also need to be careful about the check
 999			 * because the data page mapping can get cleared under
1000			 * our hands. Note that if mapping == NULL, we don't
1001			 * need to make buffer unmapped because the page is
1002			 * already detached from the mapping and buffers cannot
1003			 * get reused.
 
 
 
1004			 */
1005			mapping = READ_ONCE(bh->b_page->mapping);
1006			if (mapping && !sb_is_blkdev_sb(mapping->host->i_sb)) {
 
 
1007				clear_buffer_mapped(bh);
1008				clear_buffer_new(bh);
1009				clear_buffer_req(bh);
1010				bh->b_bdev = NULL;
1011			}
1012		}
1013
1014		if (buffer_jbddirty(bh)) {
1015			JBUFFER_TRACE(jh, "add to new checkpointing trans");
1016			__jbd2_journal_insert_checkpoint(jh, commit_transaction);
1017			if (is_journal_aborted(journal))
1018				clear_buffer_jbddirty(bh);
1019		} else {
1020			J_ASSERT_BH(bh, !buffer_dirty(bh));
1021			/*
1022			 * The buffer on BJ_Forget list and not jbddirty means
1023			 * it has been freed by this transaction and hence it
1024			 * could not have been reallocated until this
1025			 * transaction has committed. *BUT* it could be
1026			 * reallocated once we have written all the data to
1027			 * disk and before we process the buffer on BJ_Forget
1028			 * list.
1029			 */
1030			if (!jh->b_next_transaction)
1031				try_to_free = 1;
1032		}
1033		JBUFFER_TRACE(jh, "refile or unfile buffer");
1034		drop_ref = __jbd2_journal_refile_buffer(jh);
1035		spin_unlock(&jh->b_state_lock);
1036		if (drop_ref)
1037			jbd2_journal_put_journal_head(jh);
1038		if (try_to_free)
1039			release_buffer_page(bh);	/* Drops bh reference */
1040		else
1041			__brelse(bh);
1042		cond_resched_lock(&journal->j_list_lock);
1043	}
1044	spin_unlock(&journal->j_list_lock);
1045	/*
1046	 * This is a bit sleazy.  We use j_list_lock to protect transition
1047	 * of a transaction into T_FINISHED state and calling
1048	 * __jbd2_journal_drop_transaction(). Otherwise we could race with
1049	 * other checkpointing code processing the transaction...
1050	 */
1051	write_lock(&journal->j_state_lock);
1052	spin_lock(&journal->j_list_lock);
1053	/*
1054	 * Now recheck if some buffers did not get attached to the transaction
1055	 * while the lock was dropped...
1056	 */
1057	if (commit_transaction->t_forget) {
1058		spin_unlock(&journal->j_list_lock);
1059		write_unlock(&journal->j_state_lock);
1060		goto restart_loop;
1061	}
1062
1063	/* Add the transaction to the checkpoint list
1064	 * __journal_remove_checkpoint() can not destroy transaction
1065	 * under us because it is not marked as T_FINISHED yet */
1066	if (journal->j_checkpoint_transactions == NULL) {
1067		journal->j_checkpoint_transactions = commit_transaction;
1068		commit_transaction->t_cpnext = commit_transaction;
1069		commit_transaction->t_cpprev = commit_transaction;
1070	} else {
1071		commit_transaction->t_cpnext =
1072			journal->j_checkpoint_transactions;
1073		commit_transaction->t_cpprev =
1074			commit_transaction->t_cpnext->t_cpprev;
1075		commit_transaction->t_cpnext->t_cpprev =
1076			commit_transaction;
1077		commit_transaction->t_cpprev->t_cpnext =
1078				commit_transaction;
1079	}
1080	spin_unlock(&journal->j_list_lock);
1081
1082	/* Done with this transaction! */
1083
1084	jbd_debug(3, "JBD2: commit phase 7\n");
1085
1086	J_ASSERT(commit_transaction->t_state == T_COMMIT_JFLUSH);
1087
1088	commit_transaction->t_start = jiffies;
1089	stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
1090					      commit_transaction->t_start);
1091
1092	/*
1093	 * File the transaction statistics
1094	 */
1095	stats.ts_tid = commit_transaction->t_tid;
1096	stats.run.rs_handle_count =
1097		atomic_read(&commit_transaction->t_handle_count);
1098	trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
1099			     commit_transaction->t_tid, &stats.run);
1100	stats.ts_requested = (commit_transaction->t_requested) ? 1 : 0;
1101
1102	commit_transaction->t_state = T_COMMIT_CALLBACK;
1103	J_ASSERT(commit_transaction == journal->j_committing_transaction);
1104	journal->j_commit_sequence = commit_transaction->t_tid;
1105	journal->j_committing_transaction = NULL;
1106	commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1107
1108	/*
1109	 * weight the commit time higher than the average time so we don't
1110	 * react too strongly to vast changes in the commit time
1111	 */
1112	if (likely(journal->j_average_commit_time))
1113		journal->j_average_commit_time = (commit_time +
1114				journal->j_average_commit_time*3) / 4;
1115	else
1116		journal->j_average_commit_time = commit_time;
1117
1118	write_unlock(&journal->j_state_lock);
1119
1120	if (journal->j_commit_callback)
1121		journal->j_commit_callback(journal, commit_transaction);
1122
1123	trace_jbd2_end_commit(journal, commit_transaction);
1124	jbd_debug(1, "JBD2: commit %d complete, head %d\n",
1125		  journal->j_commit_sequence, journal->j_tail_sequence);
1126
1127	write_lock(&journal->j_state_lock);
1128	spin_lock(&journal->j_list_lock);
1129	commit_transaction->t_state = T_FINISHED;
1130	/* Check if the transaction can be dropped now that we are finished */
1131	if (commit_transaction->t_checkpoint_list == NULL &&
1132	    commit_transaction->t_checkpoint_io_list == NULL) {
1133		__jbd2_journal_drop_transaction(journal, commit_transaction);
1134		jbd2_journal_free_transaction(commit_transaction);
1135	}
1136	spin_unlock(&journal->j_list_lock);
1137	write_unlock(&journal->j_state_lock);
1138	wake_up(&journal->j_wait_done_commit);
1139
1140	/*
1141	 * Calculate overall stats
1142	 */
1143	spin_lock(&journal->j_history_lock);
1144	journal->j_stats.ts_tid++;
1145	journal->j_stats.ts_requested += stats.ts_requested;
1146	journal->j_stats.run.rs_wait += stats.run.rs_wait;
1147	journal->j_stats.run.rs_request_delay += stats.run.rs_request_delay;
1148	journal->j_stats.run.rs_running += stats.run.rs_running;
1149	journal->j_stats.run.rs_locked += stats.run.rs_locked;
1150	journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
1151	journal->j_stats.run.rs_logging += stats.run.rs_logging;
1152	journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
1153	journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
1154	journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
1155	spin_unlock(&journal->j_history_lock);
1156}
v3.15
 
   1/*
   2 * linux/fs/jbd2/commit.c
   3 *
   4 * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
   5 *
   6 * Copyright 1998 Red Hat corp --- All Rights Reserved
   7 *
   8 * This file is part of the Linux kernel and is made available under
   9 * the terms of the GNU General Public License, version 2, or at your
  10 * option, any later version, incorporated herein by reference.
  11 *
  12 * Journal commit routines for the generic filesystem journaling code;
  13 * part of the ext2fs journaling system.
  14 */
  15
  16#include <linux/time.h>
  17#include <linux/fs.h>
  18#include <linux/jbd2.h>
  19#include <linux/errno.h>
  20#include <linux/slab.h>
  21#include <linux/mm.h>
  22#include <linux/pagemap.h>
  23#include <linux/jiffies.h>
  24#include <linux/crc32.h>
  25#include <linux/writeback.h>
  26#include <linux/backing-dev.h>
  27#include <linux/bio.h>
  28#include <linux/blkdev.h>
  29#include <linux/bitops.h>
  30#include <trace/events/jbd2.h>
  31
  32/*
  33 * IO end handler for temporary buffer_heads handling writes to the journal.
  34 */
  35static void journal_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
  36{
  37	struct buffer_head *orig_bh = bh->b_private;
  38
  39	BUFFER_TRACE(bh, "");
  40	if (uptodate)
  41		set_buffer_uptodate(bh);
  42	else
  43		clear_buffer_uptodate(bh);
  44	if (orig_bh) {
  45		clear_bit_unlock(BH_Shadow, &orig_bh->b_state);
  46		smp_mb__after_clear_bit();
  47		wake_up_bit(&orig_bh->b_state, BH_Shadow);
  48	}
  49	unlock_buffer(bh);
  50}
  51
  52/*
  53 * When an ext4 file is truncated, it is possible that some pages are not
  54 * successfully freed, because they are attached to a committing transaction.
  55 * After the transaction commits, these pages are left on the LRU, with no
  56 * ->mapping, and with attached buffers.  These pages are trivially reclaimable
  57 * by the VM, but their apparent absence upsets the VM accounting, and it makes
  58 * the numbers in /proc/meminfo look odd.
  59 *
  60 * So here, we have a buffer which has just come off the forget list.  Look to
  61 * see if we can strip all buffers from the backing page.
  62 *
  63 * Called under lock_journal(), and possibly under journal_datalist_lock.  The
  64 * caller provided us with a ref against the buffer, and we drop that here.
  65 */
  66static void release_buffer_page(struct buffer_head *bh)
  67{
  68	struct page *page;
  69
  70	if (buffer_dirty(bh))
  71		goto nope;
  72	if (atomic_read(&bh->b_count) != 1)
  73		goto nope;
  74	page = bh->b_page;
  75	if (!page)
  76		goto nope;
  77	if (page->mapping)
  78		goto nope;
  79
  80	/* OK, it's a truncated page */
  81	if (!trylock_page(page))
  82		goto nope;
  83
  84	page_cache_get(page);
  85	__brelse(bh);
  86	try_to_free_buffers(page);
  87	unlock_page(page);
  88	page_cache_release(page);
  89	return;
  90
  91nope:
  92	__brelse(bh);
  93}
  94
  95static void jbd2_commit_block_csum_set(journal_t *j, struct buffer_head *bh)
  96{
  97	struct commit_header *h;
  98	__u32 csum;
  99
 100	if (!JBD2_HAS_INCOMPAT_FEATURE(j, JBD2_FEATURE_INCOMPAT_CSUM_V2))
 101		return;
 102
 103	h = (struct commit_header *)(bh->b_data);
 104	h->h_chksum_type = 0;
 105	h->h_chksum_size = 0;
 106	h->h_chksum[0] = 0;
 107	csum = jbd2_chksum(j, j->j_csum_seed, bh->b_data, j->j_blocksize);
 108	h->h_chksum[0] = cpu_to_be32(csum);
 109}
 110
 111/*
 112 * Done it all: now submit the commit record.  We should have
 113 * cleaned up our previous buffers by now, so if we are in abort
 114 * mode we can now just skip the rest of the journal write
 115 * entirely.
 116 *
 117 * Returns 1 if the journal needs to be aborted or 0 on success
 118 */
 119static int journal_submit_commit_record(journal_t *journal,
 120					transaction_t *commit_transaction,
 121					struct buffer_head **cbh,
 122					__u32 crc32_sum)
 123{
 124	struct commit_header *tmp;
 125	struct buffer_head *bh;
 126	int ret;
 127	struct timespec now = current_kernel_time();
 128
 129	*cbh = NULL;
 130
 131	if (is_journal_aborted(journal))
 132		return 0;
 133
 134	bh = jbd2_journal_get_descriptor_buffer(journal);
 
 135	if (!bh)
 136		return 1;
 137
 138	tmp = (struct commit_header *)bh->b_data;
 139	tmp->h_magic = cpu_to_be32(JBD2_MAGIC_NUMBER);
 140	tmp->h_blocktype = cpu_to_be32(JBD2_COMMIT_BLOCK);
 141	tmp->h_sequence = cpu_to_be32(commit_transaction->t_tid);
 142	tmp->h_commit_sec = cpu_to_be64(now.tv_sec);
 143	tmp->h_commit_nsec = cpu_to_be32(now.tv_nsec);
 144
 145	if (JBD2_HAS_COMPAT_FEATURE(journal,
 146				    JBD2_FEATURE_COMPAT_CHECKSUM)) {
 147		tmp->h_chksum_type 	= JBD2_CRC32_CHKSUM;
 148		tmp->h_chksum_size 	= JBD2_CRC32_CHKSUM_SIZE;
 149		tmp->h_chksum[0] 	= cpu_to_be32(crc32_sum);
 150	}
 151	jbd2_commit_block_csum_set(journal, bh);
 152
 153	BUFFER_TRACE(bh, "submit commit block");
 154	lock_buffer(bh);
 155	clear_buffer_dirty(bh);
 156	set_buffer_uptodate(bh);
 157	bh->b_end_io = journal_end_buffer_io_sync;
 158
 159	if (journal->j_flags & JBD2_BARRIER &&
 160	    !JBD2_HAS_INCOMPAT_FEATURE(journal,
 161				       JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT))
 162		ret = submit_bh(WRITE_SYNC | WRITE_FLUSH_FUA, bh);
 163	else
 164		ret = submit_bh(WRITE_SYNC, bh);
 165
 166	*cbh = bh;
 167	return ret;
 168}
 169
 170/*
 171 * This function along with journal_submit_commit_record
 172 * allows to write the commit record asynchronously.
 173 */
 174static int journal_wait_on_commit_record(journal_t *journal,
 175					 struct buffer_head *bh)
 176{
 177	int ret = 0;
 178
 179	clear_buffer_dirty(bh);
 180	wait_on_buffer(bh);
 181
 182	if (unlikely(!buffer_uptodate(bh)))
 183		ret = -EIO;
 184	put_bh(bh);            /* One for getblk() */
 185
 186	return ret;
 187}
 188
 189/*
 190 * write the filemap data using writepage() address_space_operations.
 191 * We don't do block allocation here even for delalloc. We don't
 192 * use writepages() because with dealyed allocation we may be doing
 193 * block allocation in writepages().
 194 */
 195static int journal_submit_inode_data_buffers(struct address_space *mapping)
 
 196{
 197	int ret;
 198	struct writeback_control wbc = {
 199		.sync_mode =  WB_SYNC_ALL,
 200		.nr_to_write = mapping->nrpages * 2,
 201		.range_start = 0,
 202		.range_end = i_size_read(mapping->host),
 203	};
 204
 205	ret = generic_writepages(mapping, &wbc);
 206	return ret;
 207}
 208
 209/*
 210 * Submit all the data buffers of inode associated with the transaction to
 211 * disk.
 212 *
 213 * We are in a committing transaction. Therefore no new inode can be added to
 214 * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
 215 * operate on from being released while we write out pages.
 216 */
 217static int journal_submit_data_buffers(journal_t *journal,
 218		transaction_t *commit_transaction)
 219{
 220	struct jbd2_inode *jinode;
 221	int err, ret = 0;
 222	struct address_space *mapping;
 223
 224	spin_lock(&journal->j_list_lock);
 225	list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
 
 
 
 
 
 226		mapping = jinode->i_vfs_inode->i_mapping;
 227		set_bit(__JI_COMMIT_RUNNING, &jinode->i_flags);
 228		spin_unlock(&journal->j_list_lock);
 229		/*
 230		 * submit the inode data buffers. We use writepage
 231		 * instead of writepages. Because writepages can do
 232		 * block allocation  with delalloc. We need to write
 233		 * only allocated blocks here.
 234		 */
 235		trace_jbd2_submit_inode_data(jinode->i_vfs_inode);
 236		err = journal_submit_inode_data_buffers(mapping);
 
 237		if (!ret)
 238			ret = err;
 239		spin_lock(&journal->j_list_lock);
 240		J_ASSERT(jinode->i_transaction == commit_transaction);
 241		clear_bit(__JI_COMMIT_RUNNING, &jinode->i_flags);
 242		smp_mb__after_clear_bit();
 243		wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
 244	}
 245	spin_unlock(&journal->j_list_lock);
 246	return ret;
 247}
 248
 249/*
 250 * Wait for data submitted for writeout, refile inodes to proper
 251 * transaction if needed.
 252 *
 253 */
 254static int journal_finish_inode_data_buffers(journal_t *journal,
 255		transaction_t *commit_transaction)
 256{
 257	struct jbd2_inode *jinode, *next_i;
 258	int err, ret = 0;
 259
 260	/* For locking, see the comment in journal_submit_data_buffers() */
 261	spin_lock(&journal->j_list_lock);
 262	list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
 263		set_bit(__JI_COMMIT_RUNNING, &jinode->i_flags);
 
 
 
 
 
 264		spin_unlock(&journal->j_list_lock);
 265		err = filemap_fdatawait(jinode->i_vfs_inode->i_mapping);
 266		if (err) {
 267			/*
 268			 * Because AS_EIO is cleared by
 269			 * filemap_fdatawait_range(), set it again so
 270			 * that user process can get -EIO from fsync().
 271			 */
 272			set_bit(AS_EIO,
 273				&jinode->i_vfs_inode->i_mapping->flags);
 274
 275			if (!ret)
 276				ret = err;
 277		}
 278		spin_lock(&journal->j_list_lock);
 279		clear_bit(__JI_COMMIT_RUNNING, &jinode->i_flags);
 280		smp_mb__after_clear_bit();
 281		wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
 282	}
 283
 284	/* Now refile inode to proper lists */
 285	list_for_each_entry_safe(jinode, next_i,
 286				 &commit_transaction->t_inode_list, i_list) {
 287		list_del(&jinode->i_list);
 288		if (jinode->i_next_transaction) {
 289			jinode->i_transaction = jinode->i_next_transaction;
 290			jinode->i_next_transaction = NULL;
 291			list_add(&jinode->i_list,
 292				&jinode->i_transaction->t_inode_list);
 293		} else {
 294			jinode->i_transaction = NULL;
 
 
 295		}
 296	}
 297	spin_unlock(&journal->j_list_lock);
 298
 299	return ret;
 300}
 301
 302static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
 303{
 304	struct page *page = bh->b_page;
 305	char *addr;
 306	__u32 checksum;
 307
 308	addr = kmap_atomic(page);
 309	checksum = crc32_be(crc32_sum,
 310		(void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
 311	kunmap_atomic(addr);
 312
 313	return checksum;
 314}
 315
 316static void write_tag_block(int tag_bytes, journal_block_tag_t *tag,
 317				   unsigned long long block)
 318{
 319	tag->t_blocknr = cpu_to_be32(block & (u32)~0);
 320	if (tag_bytes > JBD2_TAG_SIZE32)
 321		tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
 322}
 323
 324static void jbd2_descr_block_csum_set(journal_t *j,
 325				      struct buffer_head *bh)
 326{
 327	struct jbd2_journal_block_tail *tail;
 328	__u32 csum;
 329
 330	if (!JBD2_HAS_INCOMPAT_FEATURE(j, JBD2_FEATURE_INCOMPAT_CSUM_V2))
 331		return;
 332
 333	tail = (struct jbd2_journal_block_tail *)(bh->b_data + j->j_blocksize -
 334			sizeof(struct jbd2_journal_block_tail));
 335	tail->t_checksum = 0;
 336	csum = jbd2_chksum(j, j->j_csum_seed, bh->b_data, j->j_blocksize);
 337	tail->t_checksum = cpu_to_be32(csum);
 338}
 339
 340static void jbd2_block_tag_csum_set(journal_t *j, journal_block_tag_t *tag,
 341				    struct buffer_head *bh, __u32 sequence)
 342{
 
 343	struct page *page = bh->b_page;
 344	__u8 *addr;
 345	__u32 csum32;
 346	__be32 seq;
 347
 348	if (!JBD2_HAS_INCOMPAT_FEATURE(j, JBD2_FEATURE_INCOMPAT_CSUM_V2))
 349		return;
 350
 351	seq = cpu_to_be32(sequence);
 352	addr = kmap_atomic(page);
 353	csum32 = jbd2_chksum(j, j->j_csum_seed, (__u8 *)&seq, sizeof(seq));
 354	csum32 = jbd2_chksum(j, csum32, addr + offset_in_page(bh->b_data),
 355			     bh->b_size);
 356	kunmap_atomic(addr);
 357
 358	/* We only have space to store the lower 16 bits of the crc32c. */
 359	tag->t_checksum = cpu_to_be16(csum32);
 
 
 360}
 361/*
 362 * jbd2_journal_commit_transaction
 363 *
 364 * The primary function for committing a transaction to the log.  This
 365 * function is called by the journal thread to begin a complete commit.
 366 */
 367void jbd2_journal_commit_transaction(journal_t *journal)
 368{
 369	struct transaction_stats_s stats;
 370	transaction_t *commit_transaction;
 371	struct journal_head *jh;
 372	struct buffer_head *descriptor;
 373	struct buffer_head **wbuf = journal->j_wbuf;
 374	int bufs;
 375	int flags;
 376	int err;
 377	unsigned long long blocknr;
 378	ktime_t start_time;
 379	u64 commit_time;
 380	char *tagp = NULL;
 381	journal_header_t *header;
 382	journal_block_tag_t *tag = NULL;
 383	int space_left = 0;
 384	int first_tag = 0;
 385	int tag_flag;
 386	int i;
 387	int tag_bytes = journal_tag_bytes(journal);
 388	struct buffer_head *cbh = NULL; /* For transactional checksums */
 389	__u32 crc32_sum = ~0;
 390	struct blk_plug plug;
 391	/* Tail of the journal */
 392	unsigned long first_block;
 393	tid_t first_tid;
 394	int update_tail;
 395	int csum_size = 0;
 396	LIST_HEAD(io_bufs);
 397	LIST_HEAD(log_bufs);
 398
 399	if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_CSUM_V2))
 400		csum_size = sizeof(struct jbd2_journal_block_tail);
 401
 402	/*
 403	 * First job: lock down the current transaction and wait for
 404	 * all outstanding updates to complete.
 405	 */
 406
 407	/* Do we need to erase the effects of a prior jbd2_journal_flush? */
 408	if (journal->j_flags & JBD2_FLUSHED) {
 409		jbd_debug(3, "super block updated\n");
 410		mutex_lock(&journal->j_checkpoint_mutex);
 411		/*
 412		 * We hold j_checkpoint_mutex so tail cannot change under us.
 413		 * We don't need any special data guarantees for writing sb
 414		 * since journal is empty and it is ok for write to be
 415		 * flushed only with transaction commit.
 416		 */
 417		jbd2_journal_update_sb_log_tail(journal,
 418						journal->j_tail_sequence,
 419						journal->j_tail,
 420						WRITE_SYNC);
 421		mutex_unlock(&journal->j_checkpoint_mutex);
 422	} else {
 423		jbd_debug(3, "superblock not updated\n");
 424	}
 425
 426	J_ASSERT(journal->j_running_transaction != NULL);
 427	J_ASSERT(journal->j_committing_transaction == NULL);
 428
 429	commit_transaction = journal->j_running_transaction;
 430
 431	trace_jbd2_start_commit(journal, commit_transaction);
 432	jbd_debug(1, "JBD2: starting commit of transaction %d\n",
 433			commit_transaction->t_tid);
 434
 435	write_lock(&journal->j_state_lock);
 436	J_ASSERT(commit_transaction->t_state == T_RUNNING);
 437	commit_transaction->t_state = T_LOCKED;
 438
 439	trace_jbd2_commit_locking(journal, commit_transaction);
 440	stats.run.rs_wait = commit_transaction->t_max_wait;
 441	stats.run.rs_request_delay = 0;
 442	stats.run.rs_locked = jiffies;
 443	if (commit_transaction->t_requested)
 444		stats.run.rs_request_delay =
 445			jbd2_time_diff(commit_transaction->t_requested,
 446				       stats.run.rs_locked);
 447	stats.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
 448					      stats.run.rs_locked);
 449
 450	spin_lock(&commit_transaction->t_handle_lock);
 451	while (atomic_read(&commit_transaction->t_updates)) {
 452		DEFINE_WAIT(wait);
 453
 454		prepare_to_wait(&journal->j_wait_updates, &wait,
 455					TASK_UNINTERRUPTIBLE);
 456		if (atomic_read(&commit_transaction->t_updates)) {
 457			spin_unlock(&commit_transaction->t_handle_lock);
 458			write_unlock(&journal->j_state_lock);
 459			schedule();
 460			write_lock(&journal->j_state_lock);
 461			spin_lock(&commit_transaction->t_handle_lock);
 462		}
 463		finish_wait(&journal->j_wait_updates, &wait);
 464	}
 465	spin_unlock(&commit_transaction->t_handle_lock);
 
 
 466
 467	J_ASSERT (atomic_read(&commit_transaction->t_outstanding_credits) <=
 468			journal->j_max_transaction_buffers);
 469
 470	/*
 471	 * First thing we are allowed to do is to discard any remaining
 472	 * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
 473	 * that there are no such buffers: if a large filesystem
 474	 * operation like a truncate needs to split itself over multiple
 475	 * transactions, then it may try to do a jbd2_journal_restart() while
 476	 * there are still BJ_Reserved buffers outstanding.  These must
 477	 * be released cleanly from the current transaction.
 478	 *
 479	 * In this case, the filesystem must still reserve write access
 480	 * again before modifying the buffer in the new transaction, but
 481	 * we do not require it to remember exactly which old buffers it
 482	 * has reserved.  This is consistent with the existing behaviour
 483	 * that multiple jbd2_journal_get_write_access() calls to the same
 484	 * buffer are perfectly permissible.
 485	 */
 486	while (commit_transaction->t_reserved_list) {
 487		jh = commit_transaction->t_reserved_list;
 488		JBUFFER_TRACE(jh, "reserved, unused: refile");
 489		/*
 490		 * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
 491		 * leave undo-committed data.
 492		 */
 493		if (jh->b_committed_data) {
 494			struct buffer_head *bh = jh2bh(jh);
 495
 496			jbd_lock_bh_state(bh);
 497			jbd2_free(jh->b_committed_data, bh->b_size);
 498			jh->b_committed_data = NULL;
 499			jbd_unlock_bh_state(bh);
 500		}
 501		jbd2_journal_refile_buffer(journal, jh);
 502	}
 503
 504	/*
 505	 * Now try to drop any written-back buffers from the journal's
 506	 * checkpoint lists.  We do this *before* commit because it potentially
 507	 * frees some memory
 508	 */
 509	spin_lock(&journal->j_list_lock);
 510	__jbd2_journal_clean_checkpoint_list(journal);
 511	spin_unlock(&journal->j_list_lock);
 512
 513	jbd_debug(3, "JBD2: commit phase 1\n");
 514
 515	/*
 516	 * Clear revoked flag to reflect there is no revoked buffers
 517	 * in the next transaction which is going to be started.
 518	 */
 519	jbd2_clear_buffer_revoked_flags(journal);
 520
 521	/*
 522	 * Switch to a new revoke table.
 523	 */
 524	jbd2_journal_switch_revoke_table(journal);
 525
 526	/*
 527	 * Reserved credits cannot be claimed anymore, free them
 528	 */
 529	atomic_sub(atomic_read(&journal->j_reserved_credits),
 530		   &commit_transaction->t_outstanding_credits);
 531
 
 532	trace_jbd2_commit_flushing(journal, commit_transaction);
 533	stats.run.rs_flushing = jiffies;
 534	stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
 535					     stats.run.rs_flushing);
 536
 537	commit_transaction->t_state = T_FLUSH;
 538	journal->j_committing_transaction = commit_transaction;
 539	journal->j_running_transaction = NULL;
 540	start_time = ktime_get();
 541	commit_transaction->t_log_start = journal->j_head;
 542	wake_up(&journal->j_wait_transaction_locked);
 543	write_unlock(&journal->j_state_lock);
 544
 545	jbd_debug(3, "JBD2: commit phase 2a\n");
 546
 547	/*
 548	 * Now start flushing things to disk, in the order they appear
 549	 * on the transaction lists.  Data blocks go first.
 550	 */
 551	err = journal_submit_data_buffers(journal, commit_transaction);
 552	if (err)
 553		jbd2_journal_abort(journal, err);
 554
 555	blk_start_plug(&plug);
 556	jbd2_journal_write_revoke_records(journal, commit_transaction,
 557					  &log_bufs, WRITE_SYNC);
 558
 559	jbd_debug(3, "JBD2: commit phase 2b\n");
 560
 561	/*
 562	 * Way to go: we have now written out all of the data for a
 563	 * transaction!  Now comes the tricky part: we need to write out
 564	 * metadata.  Loop over the transaction's entire buffer list:
 565	 */
 566	write_lock(&journal->j_state_lock);
 567	commit_transaction->t_state = T_COMMIT;
 568	write_unlock(&journal->j_state_lock);
 569
 570	trace_jbd2_commit_logging(journal, commit_transaction);
 571	stats.run.rs_logging = jiffies;
 572	stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
 573					       stats.run.rs_logging);
 574	stats.run.rs_blocks =
 575		atomic_read(&commit_transaction->t_outstanding_credits);
 576	stats.run.rs_blocks_logged = 0;
 577
 578	J_ASSERT(commit_transaction->t_nr_buffers <=
 579		 atomic_read(&commit_transaction->t_outstanding_credits));
 580
 581	err = 0;
 582	bufs = 0;
 583	descriptor = NULL;
 584	while (commit_transaction->t_buffers) {
 585
 586		/* Find the next buffer to be journaled... */
 587
 588		jh = commit_transaction->t_buffers;
 589
 590		/* If we're in abort mode, we just un-journal the buffer and
 591		   release it. */
 592
 593		if (is_journal_aborted(journal)) {
 594			clear_buffer_jbddirty(jh2bh(jh));
 595			JBUFFER_TRACE(jh, "journal is aborting: refile");
 596			jbd2_buffer_abort_trigger(jh,
 597						  jh->b_frozen_data ?
 598						  jh->b_frozen_triggers :
 599						  jh->b_triggers);
 600			jbd2_journal_refile_buffer(journal, jh);
 601			/* If that was the last one, we need to clean up
 602			 * any descriptor buffers which may have been
 603			 * already allocated, even if we are now
 604			 * aborting. */
 605			if (!commit_transaction->t_buffers)
 606				goto start_journal_io;
 607			continue;
 608		}
 609
 610		/* Make sure we have a descriptor block in which to
 611		   record the metadata buffer. */
 612
 613		if (!descriptor) {
 614			J_ASSERT (bufs == 0);
 615
 616			jbd_debug(4, "JBD2: get descriptor\n");
 617
 618			descriptor = jbd2_journal_get_descriptor_buffer(journal);
 
 
 619			if (!descriptor) {
 620				jbd2_journal_abort(journal, -EIO);
 621				continue;
 622			}
 623
 624			jbd_debug(4, "JBD2: got buffer %llu (%p)\n",
 625				(unsigned long long)descriptor->b_blocknr,
 626				descriptor->b_data);
 627			header = (journal_header_t *)descriptor->b_data;
 628			header->h_magic     = cpu_to_be32(JBD2_MAGIC_NUMBER);
 629			header->h_blocktype = cpu_to_be32(JBD2_DESCRIPTOR_BLOCK);
 630			header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
 631
 632			tagp = &descriptor->b_data[sizeof(journal_header_t)];
 633			space_left = descriptor->b_size -
 634						sizeof(journal_header_t);
 635			first_tag = 1;
 636			set_buffer_jwrite(descriptor);
 637			set_buffer_dirty(descriptor);
 638			wbuf[bufs++] = descriptor;
 639
 640			/* Record it so that we can wait for IO
 641                           completion later */
 642			BUFFER_TRACE(descriptor, "ph3: file as descriptor");
 643			jbd2_file_log_bh(&log_bufs, descriptor);
 644		}
 645
 646		/* Where is the buffer to be written? */
 647
 648		err = jbd2_journal_next_log_block(journal, &blocknr);
 649		/* If the block mapping failed, just abandon the buffer
 650		   and repeat this loop: we'll fall into the
 651		   refile-on-abort condition above. */
 652		if (err) {
 653			jbd2_journal_abort(journal, err);
 654			continue;
 655		}
 656
 657		/*
 658		 * start_this_handle() uses t_outstanding_credits to determine
 659		 * the free space in the log, but this counter is changed
 660		 * by jbd2_journal_next_log_block() also.
 661		 */
 662		atomic_dec(&commit_transaction->t_outstanding_credits);
 663
 664		/* Bump b_count to prevent truncate from stumbling over
 665                   the shadowed buffer!  @@@ This can go if we ever get
 666                   rid of the shadow pairing of buffers. */
 667		atomic_inc(&jh2bh(jh)->b_count);
 668
 669		/*
 670		 * Make a temporary IO buffer with which to write it out
 671		 * (this will requeue the metadata buffer to BJ_Shadow).
 672		 */
 673		set_bit(BH_JWrite, &jh2bh(jh)->b_state);
 674		JBUFFER_TRACE(jh, "ph3: write metadata");
 675		flags = jbd2_journal_write_metadata_buffer(commit_transaction,
 676						jh, &wbuf[bufs], blocknr);
 677		if (flags < 0) {
 678			jbd2_journal_abort(journal, flags);
 679			continue;
 680		}
 681		jbd2_file_log_bh(&io_bufs, wbuf[bufs]);
 682
 683		/* Record the new block's tag in the current descriptor
 684                   buffer */
 685
 686		tag_flag = 0;
 687		if (flags & 1)
 688			tag_flag |= JBD2_FLAG_ESCAPE;
 689		if (!first_tag)
 690			tag_flag |= JBD2_FLAG_SAME_UUID;
 691
 692		tag = (journal_block_tag_t *) tagp;
 693		write_tag_block(tag_bytes, tag, jh2bh(jh)->b_blocknr);
 694		tag->t_flags = cpu_to_be16(tag_flag);
 695		jbd2_block_tag_csum_set(journal, tag, wbuf[bufs],
 696					commit_transaction->t_tid);
 697		tagp += tag_bytes;
 698		space_left -= tag_bytes;
 699		bufs++;
 700
 701		if (first_tag) {
 702			memcpy (tagp, journal->j_uuid, 16);
 703			tagp += 16;
 704			space_left -= 16;
 705			first_tag = 0;
 706		}
 707
 708		/* If there's no more to do, or if the descriptor is full,
 709		   let the IO rip! */
 710
 711		if (bufs == journal->j_wbufsize ||
 712		    commit_transaction->t_buffers == NULL ||
 713		    space_left < tag_bytes + 16 + csum_size) {
 714
 715			jbd_debug(4, "JBD2: Submit %d IOs\n", bufs);
 716
 717			/* Write an end-of-descriptor marker before
 718                           submitting the IOs.  "tag" still points to
 719                           the last tag we set up. */
 720
 721			tag->t_flags |= cpu_to_be16(JBD2_FLAG_LAST_TAG);
 
 
 
 
 722
 723			jbd2_descr_block_csum_set(journal, descriptor);
 724start_journal_io:
 725			for (i = 0; i < bufs; i++) {
 726				struct buffer_head *bh = wbuf[i];
 727				/*
 728				 * Compute checksum.
 729				 */
 730				if (JBD2_HAS_COMPAT_FEATURE(journal,
 731					JBD2_FEATURE_COMPAT_CHECKSUM)) {
 732					crc32_sum =
 733					    jbd2_checksum_data(crc32_sum, bh);
 734				}
 735
 736				lock_buffer(bh);
 737				clear_buffer_dirty(bh);
 738				set_buffer_uptodate(bh);
 739				bh->b_end_io = journal_end_buffer_io_sync;
 740				submit_bh(WRITE_SYNC, bh);
 741			}
 742			cond_resched();
 743			stats.run.rs_blocks_logged += bufs;
 744
 745			/* Force a new descriptor to be generated next
 746                           time round the loop. */
 747			descriptor = NULL;
 748			bufs = 0;
 749		}
 750	}
 751
 752	err = journal_finish_inode_data_buffers(journal, commit_transaction);
 753	if (err) {
 754		printk(KERN_WARNING
 755			"JBD2: Detected IO errors while flushing file data "
 756		       "on %s\n", journal->j_devname);
 757		if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
 758			jbd2_journal_abort(journal, err);
 759		err = 0;
 760	}
 761
 762	/*
 763	 * Get current oldest transaction in the log before we issue flush
 764	 * to the filesystem device. After the flush we can be sure that
 765	 * blocks of all older transactions are checkpointed to persistent
 766	 * storage and we will be safe to update journal start in the
 767	 * superblock with the numbers we get here.
 768	 */
 769	update_tail =
 770		jbd2_journal_get_log_tail(journal, &first_tid, &first_block);
 771
 772	write_lock(&journal->j_state_lock);
 773	if (update_tail) {
 774		long freed = first_block - journal->j_tail;
 775
 776		if (first_block < journal->j_tail)
 777			freed += journal->j_last - journal->j_first;
 778		/* Update tail only if we free significant amount of space */
 779		if (freed < journal->j_maxlen / 4)
 780			update_tail = 0;
 781	}
 782	J_ASSERT(commit_transaction->t_state == T_COMMIT);
 783	commit_transaction->t_state = T_COMMIT_DFLUSH;
 784	write_unlock(&journal->j_state_lock);
 785
 786	/* 
 787	 * If the journal is not located on the file system device,
 788	 * then we must flush the file system device before we issue
 789	 * the commit record
 790	 */
 791	if (commit_transaction->t_need_data_flush &&
 792	    (journal->j_fs_dev != journal->j_dev) &&
 793	    (journal->j_flags & JBD2_BARRIER))
 794		blkdev_issue_flush(journal->j_fs_dev, GFP_NOFS, NULL);
 795
 796	/* Done it all: now write the commit record asynchronously. */
 797	if (JBD2_HAS_INCOMPAT_FEATURE(journal,
 798				      JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
 799		err = journal_submit_commit_record(journal, commit_transaction,
 800						 &cbh, crc32_sum);
 801		if (err)
 802			__jbd2_journal_abort_hard(journal);
 803	}
 804
 805	blk_finish_plug(&plug);
 806
 807	/* Lo and behold: we have just managed to send a transaction to
 808           the log.  Before we can commit it, wait for the IO so far to
 809           complete.  Control buffers being written are on the
 810           transaction's t_log_list queue, and metadata buffers are on
 811           the io_bufs list.
 812
 813	   Wait for the buffers in reverse order.  That way we are
 814	   less likely to be woken up until all IOs have completed, and
 815	   so we incur less scheduling load.
 816	*/
 817
 818	jbd_debug(3, "JBD2: commit phase 3\n");
 819
 820	while (!list_empty(&io_bufs)) {
 821		struct buffer_head *bh = list_entry(io_bufs.prev,
 822						    struct buffer_head,
 823						    b_assoc_buffers);
 824
 825		wait_on_buffer(bh);
 826		cond_resched();
 827
 828		if (unlikely(!buffer_uptodate(bh)))
 829			err = -EIO;
 830		jbd2_unfile_log_bh(bh);
 
 831
 832		/*
 833		 * The list contains temporary buffer heads created by
 834		 * jbd2_journal_write_metadata_buffer().
 835		 */
 836		BUFFER_TRACE(bh, "dumping temporary bh");
 837		__brelse(bh);
 838		J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
 839		free_buffer_head(bh);
 840
 841		/* We also have to refile the corresponding shadowed buffer */
 842		jh = commit_transaction->t_shadow_list->b_tprev;
 843		bh = jh2bh(jh);
 844		clear_buffer_jwrite(bh);
 845		J_ASSERT_BH(bh, buffer_jbddirty(bh));
 846		J_ASSERT_BH(bh, !buffer_shadow(bh));
 847
 848		/* The metadata is now released for reuse, but we need
 849                   to remember it against this transaction so that when
 850                   we finally commit, we can do any checkpointing
 851                   required. */
 852		JBUFFER_TRACE(jh, "file as BJ_Forget");
 853		jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
 854		JBUFFER_TRACE(jh, "brelse shadowed buffer");
 855		__brelse(bh);
 856	}
 857
 858	J_ASSERT (commit_transaction->t_shadow_list == NULL);
 859
 860	jbd_debug(3, "JBD2: commit phase 4\n");
 861
 862	/* Here we wait for the revoke record and descriptor record buffers */
 863	while (!list_empty(&log_bufs)) {
 864		struct buffer_head *bh;
 865
 866		bh = list_entry(log_bufs.prev, struct buffer_head, b_assoc_buffers);
 867		wait_on_buffer(bh);
 868		cond_resched();
 869
 870		if (unlikely(!buffer_uptodate(bh)))
 871			err = -EIO;
 872
 873		BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
 874		clear_buffer_jwrite(bh);
 875		jbd2_unfile_log_bh(bh);
 
 876		__brelse(bh);		/* One for getblk */
 877		/* AKPM: bforget here */
 878	}
 879
 880	if (err)
 881		jbd2_journal_abort(journal, err);
 882
 883	jbd_debug(3, "JBD2: commit phase 5\n");
 884	write_lock(&journal->j_state_lock);
 885	J_ASSERT(commit_transaction->t_state == T_COMMIT_DFLUSH);
 886	commit_transaction->t_state = T_COMMIT_JFLUSH;
 887	write_unlock(&journal->j_state_lock);
 888
 889	if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
 890				       JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
 891		err = journal_submit_commit_record(journal, commit_transaction,
 892						&cbh, crc32_sum);
 893		if (err)
 894			__jbd2_journal_abort_hard(journal);
 895	}
 896	if (cbh)
 897		err = journal_wait_on_commit_record(journal, cbh);
 898	if (JBD2_HAS_INCOMPAT_FEATURE(journal,
 899				      JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT) &&
 900	    journal->j_flags & JBD2_BARRIER) {
 901		blkdev_issue_flush(journal->j_dev, GFP_NOFS, NULL);
 902	}
 903
 904	if (err)
 905		jbd2_journal_abort(journal, err);
 906
 
 
 
 907	/*
 908	 * Now disk caches for filesystem device are flushed so we are safe to
 909	 * erase checkpointed transactions from the log by updating journal
 910	 * superblock.
 911	 */
 912	if (update_tail)
 913		jbd2_update_log_tail(journal, first_tid, first_block);
 914
 915	/* End of a transaction!  Finally, we can do checkpoint
 916           processing: any buffers committed as a result of this
 917           transaction can be removed from any checkpoint list it was on
 918           before. */
 919
 920	jbd_debug(3, "JBD2: commit phase 6\n");
 921
 922	J_ASSERT(list_empty(&commit_transaction->t_inode_list));
 923	J_ASSERT(commit_transaction->t_buffers == NULL);
 924	J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
 925	J_ASSERT(commit_transaction->t_shadow_list == NULL);
 926
 927restart_loop:
 928	/*
 929	 * As there are other places (journal_unmap_buffer()) adding buffers
 930	 * to this list we have to be careful and hold the j_list_lock.
 931	 */
 932	spin_lock(&journal->j_list_lock);
 933	while (commit_transaction->t_forget) {
 934		transaction_t *cp_transaction;
 935		struct buffer_head *bh;
 936		int try_to_free = 0;
 
 937
 938		jh = commit_transaction->t_forget;
 939		spin_unlock(&journal->j_list_lock);
 940		bh = jh2bh(jh);
 941		/*
 942		 * Get a reference so that bh cannot be freed before we are
 943		 * done with it.
 944		 */
 945		get_bh(bh);
 946		jbd_lock_bh_state(bh);
 947		J_ASSERT_JH(jh,	jh->b_transaction == commit_transaction);
 948
 949		/*
 950		 * If there is undo-protected committed data against
 951		 * this buffer, then we can remove it now.  If it is a
 952		 * buffer needing such protection, the old frozen_data
 953		 * field now points to a committed version of the
 954		 * buffer, so rotate that field to the new committed
 955		 * data.
 956		 *
 957		 * Otherwise, we can just throw away the frozen data now.
 958		 *
 959		 * We also know that the frozen data has already fired
 960		 * its triggers if they exist, so we can clear that too.
 961		 */
 962		if (jh->b_committed_data) {
 963			jbd2_free(jh->b_committed_data, bh->b_size);
 964			jh->b_committed_data = NULL;
 965			if (jh->b_frozen_data) {
 966				jh->b_committed_data = jh->b_frozen_data;
 967				jh->b_frozen_data = NULL;
 968				jh->b_frozen_triggers = NULL;
 969			}
 970		} else if (jh->b_frozen_data) {
 971			jbd2_free(jh->b_frozen_data, bh->b_size);
 972			jh->b_frozen_data = NULL;
 973			jh->b_frozen_triggers = NULL;
 974		}
 975
 976		spin_lock(&journal->j_list_lock);
 977		cp_transaction = jh->b_cp_transaction;
 978		if (cp_transaction) {
 979			JBUFFER_TRACE(jh, "remove from old cp transaction");
 980			cp_transaction->t_chp_stats.cs_dropped++;
 981			__jbd2_journal_remove_checkpoint(jh);
 982		}
 983
 984		/* Only re-checkpoint the buffer_head if it is marked
 985		 * dirty.  If the buffer was added to the BJ_Forget list
 986		 * by jbd2_journal_forget, it may no longer be dirty and
 987		 * there's no point in keeping a checkpoint record for
 988		 * it. */
 989
 990		/*
 991		* A buffer which has been freed while still being journaled by
 992		* a previous transaction.
 993		*/
 994		if (buffer_freed(bh)) {
 
 
 
 
 
 
 
 
 
 
 995			/*
 996			 * If the running transaction is the one containing
 997			 * "add to orphan" operation (b_next_transaction !=
 998			 * NULL), we have to wait for that transaction to
 999			 * commit before we can really get rid of the buffer.
1000			 * So just clear b_modified to not confuse transaction
1001			 * credit accounting and refile the buffer to
1002			 * BJ_Forget of the running transaction. If the just
1003			 * committed transaction contains "add to orphan"
1004			 * operation, we can completely invalidate the buffer
1005			 * now. We are rather through in that since the
1006			 * buffer may be still accessible when blocksize <
1007			 * pagesize and it is attached to the last partial
1008			 * page.
1009			 */
1010			jh->b_modified = 0;
1011			if (!jh->b_next_transaction) {
1012				clear_buffer_freed(bh);
1013				clear_buffer_jbddirty(bh);
1014				clear_buffer_mapped(bh);
1015				clear_buffer_new(bh);
1016				clear_buffer_req(bh);
1017				bh->b_bdev = NULL;
1018			}
1019		}
1020
1021		if (buffer_jbddirty(bh)) {
1022			JBUFFER_TRACE(jh, "add to new checkpointing trans");
1023			__jbd2_journal_insert_checkpoint(jh, commit_transaction);
1024			if (is_journal_aborted(journal))
1025				clear_buffer_jbddirty(bh);
1026		} else {
1027			J_ASSERT_BH(bh, !buffer_dirty(bh));
1028			/*
1029			 * The buffer on BJ_Forget list and not jbddirty means
1030			 * it has been freed by this transaction and hence it
1031			 * could not have been reallocated until this
1032			 * transaction has committed. *BUT* it could be
1033			 * reallocated once we have written all the data to
1034			 * disk and before we process the buffer on BJ_Forget
1035			 * list.
1036			 */
1037			if (!jh->b_next_transaction)
1038				try_to_free = 1;
1039		}
1040		JBUFFER_TRACE(jh, "refile or unfile buffer");
1041		__jbd2_journal_refile_buffer(jh);
1042		jbd_unlock_bh_state(bh);
 
 
1043		if (try_to_free)
1044			release_buffer_page(bh);	/* Drops bh reference */
1045		else
1046			__brelse(bh);
1047		cond_resched_lock(&journal->j_list_lock);
1048	}
1049	spin_unlock(&journal->j_list_lock);
1050	/*
1051	 * This is a bit sleazy.  We use j_list_lock to protect transition
1052	 * of a transaction into T_FINISHED state and calling
1053	 * __jbd2_journal_drop_transaction(). Otherwise we could race with
1054	 * other checkpointing code processing the transaction...
1055	 */
1056	write_lock(&journal->j_state_lock);
1057	spin_lock(&journal->j_list_lock);
1058	/*
1059	 * Now recheck if some buffers did not get attached to the transaction
1060	 * while the lock was dropped...
1061	 */
1062	if (commit_transaction->t_forget) {
1063		spin_unlock(&journal->j_list_lock);
1064		write_unlock(&journal->j_state_lock);
1065		goto restart_loop;
1066	}
1067
1068	/* Add the transaction to the checkpoint list
1069	 * __journal_remove_checkpoint() can not destroy transaction
1070	 * under us because it is not marked as T_FINISHED yet */
1071	if (journal->j_checkpoint_transactions == NULL) {
1072		journal->j_checkpoint_transactions = commit_transaction;
1073		commit_transaction->t_cpnext = commit_transaction;
1074		commit_transaction->t_cpprev = commit_transaction;
1075	} else {
1076		commit_transaction->t_cpnext =
1077			journal->j_checkpoint_transactions;
1078		commit_transaction->t_cpprev =
1079			commit_transaction->t_cpnext->t_cpprev;
1080		commit_transaction->t_cpnext->t_cpprev =
1081			commit_transaction;
1082		commit_transaction->t_cpprev->t_cpnext =
1083				commit_transaction;
1084	}
1085	spin_unlock(&journal->j_list_lock);
1086
1087	/* Done with this transaction! */
1088
1089	jbd_debug(3, "JBD2: commit phase 7\n");
1090
1091	J_ASSERT(commit_transaction->t_state == T_COMMIT_JFLUSH);
1092
1093	commit_transaction->t_start = jiffies;
1094	stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
1095					      commit_transaction->t_start);
1096
1097	/*
1098	 * File the transaction statistics
1099	 */
1100	stats.ts_tid = commit_transaction->t_tid;
1101	stats.run.rs_handle_count =
1102		atomic_read(&commit_transaction->t_handle_count);
1103	trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
1104			     commit_transaction->t_tid, &stats.run);
1105	stats.ts_requested = (commit_transaction->t_requested) ? 1 : 0;
1106
1107	commit_transaction->t_state = T_COMMIT_CALLBACK;
1108	J_ASSERT(commit_transaction == journal->j_committing_transaction);
1109	journal->j_commit_sequence = commit_transaction->t_tid;
1110	journal->j_committing_transaction = NULL;
1111	commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1112
1113	/*
1114	 * weight the commit time higher than the average time so we don't
1115	 * react too strongly to vast changes in the commit time
1116	 */
1117	if (likely(journal->j_average_commit_time))
1118		journal->j_average_commit_time = (commit_time +
1119				journal->j_average_commit_time*3) / 4;
1120	else
1121		journal->j_average_commit_time = commit_time;
1122
1123	write_unlock(&journal->j_state_lock);
1124
1125	if (journal->j_commit_callback)
1126		journal->j_commit_callback(journal, commit_transaction);
1127
1128	trace_jbd2_end_commit(journal, commit_transaction);
1129	jbd_debug(1, "JBD2: commit %d complete, head %d\n",
1130		  journal->j_commit_sequence, journal->j_tail_sequence);
1131
1132	write_lock(&journal->j_state_lock);
1133	spin_lock(&journal->j_list_lock);
1134	commit_transaction->t_state = T_FINISHED;
1135	/* Check if the transaction can be dropped now that we are finished */
1136	if (commit_transaction->t_checkpoint_list == NULL &&
1137	    commit_transaction->t_checkpoint_io_list == NULL) {
1138		__jbd2_journal_drop_transaction(journal, commit_transaction);
1139		jbd2_journal_free_transaction(commit_transaction);
1140	}
1141	spin_unlock(&journal->j_list_lock);
1142	write_unlock(&journal->j_state_lock);
1143	wake_up(&journal->j_wait_done_commit);
1144
1145	/*
1146	 * Calculate overall stats
1147	 */
1148	spin_lock(&journal->j_history_lock);
1149	journal->j_stats.ts_tid++;
1150	journal->j_stats.ts_requested += stats.ts_requested;
1151	journal->j_stats.run.rs_wait += stats.run.rs_wait;
1152	journal->j_stats.run.rs_request_delay += stats.run.rs_request_delay;
1153	journal->j_stats.run.rs_running += stats.run.rs_running;
1154	journal->j_stats.run.rs_locked += stats.run.rs_locked;
1155	journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
1156	journal->j_stats.run.rs_logging += stats.run.rs_logging;
1157	journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
1158	journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
1159	journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
1160	spin_unlock(&journal->j_history_lock);
1161}