Linux Audio

Check our new training course

Loading...
v4.17
   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 = current_kernel_time64();
 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	tmp->h_commit_sec = cpu_to_be64(now.tv_sec);
 138	tmp->h_commit_nsec = cpu_to_be32(now.tv_nsec);
 139
 140	if (jbd2_has_feature_checksum(journal)) {
 141		tmp->h_chksum_type 	= JBD2_CRC32_CHKSUM;
 142		tmp->h_chksum_size 	= JBD2_CRC32_CHKSUM_SIZE;
 143		tmp->h_chksum[0] 	= cpu_to_be32(crc32_sum);
 144	}
 145	jbd2_commit_block_csum_set(journal, bh);
 146
 147	BUFFER_TRACE(bh, "submit commit block");
 148	lock_buffer(bh);
 149	clear_buffer_dirty(bh);
 150	set_buffer_uptodate(bh);
 151	bh->b_end_io = journal_end_buffer_io_sync;
 152
 153	if (journal->j_flags & JBD2_BARRIER &&
 154	    !jbd2_has_feature_async_commit(journal))
 155		ret = submit_bh(REQ_OP_WRITE,
 156			REQ_SYNC | REQ_PREFLUSH | REQ_FUA, bh);
 157	else
 158		ret = submit_bh(REQ_OP_WRITE, REQ_SYNC, bh);
 159
 160	*cbh = bh;
 161	return ret;
 162}
 163
 164/*
 165 * This function along with journal_submit_commit_record
 166 * allows to write the commit record asynchronously.
 167 */
 168static int journal_wait_on_commit_record(journal_t *journal,
 169					 struct buffer_head *bh)
 170{
 171	int ret = 0;
 172
 173	clear_buffer_dirty(bh);
 174	wait_on_buffer(bh);
 175
 176	if (unlikely(!buffer_uptodate(bh)))
 177		ret = -EIO;
 178	put_bh(bh);            /* One for getblk() */
 179
 180	return ret;
 181}
 182
 183/*
 184 * write the filemap data using writepage() address_space_operations.
 185 * We don't do block allocation here even for delalloc. We don't
 186 * use writepages() because with dealyed allocation we may be doing
 187 * block allocation in writepages().
 188 */
 189static int journal_submit_inode_data_buffers(struct address_space *mapping)
 
 190{
 191	int ret;
 192	struct writeback_control wbc = {
 193		.sync_mode =  WB_SYNC_ALL,
 194		.nr_to_write = mapping->nrpages * 2,
 195		.range_start = 0,
 196		.range_end = i_size_read(mapping->host),
 197	};
 198
 199	ret = generic_writepages(mapping, &wbc);
 200	return ret;
 201}
 202
 203/*
 204 * Submit all the data buffers of inode associated with the transaction to
 205 * disk.
 206 *
 207 * We are in a committing transaction. Therefore no new inode can be added to
 208 * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
 209 * operate on from being released while we write out pages.
 210 */
 211static int journal_submit_data_buffers(journal_t *journal,
 212		transaction_t *commit_transaction)
 213{
 214	struct jbd2_inode *jinode;
 215	int err, ret = 0;
 216	struct address_space *mapping;
 217
 218	spin_lock(&journal->j_list_lock);
 219	list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
 
 
 
 220		if (!(jinode->i_flags & JI_WRITE_DATA))
 221			continue;
 222		mapping = jinode->i_vfs_inode->i_mapping;
 223		jinode->i_flags |= JI_COMMIT_RUNNING;
 224		spin_unlock(&journal->j_list_lock);
 225		/*
 226		 * submit the inode data buffers. We use writepage
 227		 * instead of writepages. Because writepages can do
 228		 * block allocation  with delalloc. We need to write
 229		 * only allocated blocks here.
 230		 */
 231		trace_jbd2_submit_inode_data(jinode->i_vfs_inode);
 232		err = journal_submit_inode_data_buffers(mapping);
 
 233		if (!ret)
 234			ret = err;
 235		spin_lock(&journal->j_list_lock);
 236		J_ASSERT(jinode->i_transaction == commit_transaction);
 237		jinode->i_flags &= ~JI_COMMIT_RUNNING;
 238		smp_mb();
 239		wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
 240	}
 241	spin_unlock(&journal->j_list_lock);
 242	return ret;
 243}
 244
 245/*
 246 * Wait for data submitted for writeout, refile inodes to proper
 247 * transaction if needed.
 248 *
 249 */
 250static int journal_finish_inode_data_buffers(journal_t *journal,
 251		transaction_t *commit_transaction)
 252{
 253	struct jbd2_inode *jinode, *next_i;
 254	int err, ret = 0;
 255
 256	/* For locking, see the comment in journal_submit_data_buffers() */
 257	spin_lock(&journal->j_list_lock);
 258	list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
 
 
 
 259		if (!(jinode->i_flags & JI_WAIT_DATA))
 260			continue;
 261		jinode->i_flags |= JI_COMMIT_RUNNING;
 262		spin_unlock(&journal->j_list_lock);
 263		err = filemap_fdatawait_keep_errors(
 264				jinode->i_vfs_inode->i_mapping);
 
 265		if (!ret)
 266			ret = err;
 267		spin_lock(&journal->j_list_lock);
 268		jinode->i_flags &= ~JI_COMMIT_RUNNING;
 269		smp_mb();
 270		wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
 271	}
 272
 273	/* Now refile inode to proper lists */
 274	list_for_each_entry_safe(jinode, next_i,
 275				 &commit_transaction->t_inode_list, i_list) {
 276		list_del(&jinode->i_list);
 277		if (jinode->i_next_transaction) {
 278			jinode->i_transaction = jinode->i_next_transaction;
 279			jinode->i_next_transaction = NULL;
 280			list_add(&jinode->i_list,
 281				&jinode->i_transaction->t_inode_list);
 282		} else {
 283			jinode->i_transaction = NULL;
 
 
 284		}
 285	}
 286	spin_unlock(&journal->j_list_lock);
 287
 288	return ret;
 289}
 290
 291static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
 292{
 293	struct page *page = bh->b_page;
 294	char *addr;
 295	__u32 checksum;
 296
 297	addr = kmap_atomic(page);
 298	checksum = crc32_be(crc32_sum,
 299		(void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
 300	kunmap_atomic(addr);
 301
 302	return checksum;
 303}
 304
 305static void write_tag_block(journal_t *j, journal_block_tag_t *tag,
 306				   unsigned long long block)
 307{
 308	tag->t_blocknr = cpu_to_be32(block & (u32)~0);
 309	if (jbd2_has_feature_64bit(j))
 310		tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
 311}
 312
 313static void jbd2_block_tag_csum_set(journal_t *j, journal_block_tag_t *tag,
 314				    struct buffer_head *bh, __u32 sequence)
 315{
 316	journal_block_tag3_t *tag3 = (journal_block_tag3_t *)tag;
 317	struct page *page = bh->b_page;
 318	__u8 *addr;
 319	__u32 csum32;
 320	__be32 seq;
 321
 322	if (!jbd2_journal_has_csum_v2or3(j))
 323		return;
 324
 325	seq = cpu_to_be32(sequence);
 326	addr = kmap_atomic(page);
 327	csum32 = jbd2_chksum(j, j->j_csum_seed, (__u8 *)&seq, sizeof(seq));
 328	csum32 = jbd2_chksum(j, csum32, addr + offset_in_page(bh->b_data),
 329			     bh->b_size);
 330	kunmap_atomic(addr);
 331
 332	if (jbd2_has_feature_csum3(j))
 333		tag3->t_checksum = cpu_to_be32(csum32);
 334	else
 335		tag->t_checksum = cpu_to_be16(csum32);
 336}
 337/*
 338 * jbd2_journal_commit_transaction
 339 *
 340 * The primary function for committing a transaction to the log.  This
 341 * function is called by the journal thread to begin a complete commit.
 342 */
 343void jbd2_journal_commit_transaction(journal_t *journal)
 344{
 345	struct transaction_stats_s stats;
 346	transaction_t *commit_transaction;
 347	struct journal_head *jh;
 348	struct buffer_head *descriptor;
 349	struct buffer_head **wbuf = journal->j_wbuf;
 350	int bufs;
 351	int flags;
 352	int err;
 353	unsigned long long blocknr;
 354	ktime_t start_time;
 355	u64 commit_time;
 356	char *tagp = NULL;
 357	journal_block_tag_t *tag = NULL;
 358	int space_left = 0;
 359	int first_tag = 0;
 360	int tag_flag;
 361	int i;
 362	int tag_bytes = journal_tag_bytes(journal);
 363	struct buffer_head *cbh = NULL; /* For transactional checksums */
 364	__u32 crc32_sum = ~0;
 365	struct blk_plug plug;
 366	/* Tail of the journal */
 367	unsigned long first_block;
 368	tid_t first_tid;
 369	int update_tail;
 370	int csum_size = 0;
 371	LIST_HEAD(io_bufs);
 372	LIST_HEAD(log_bufs);
 373
 374	if (jbd2_journal_has_csum_v2or3(journal))
 375		csum_size = sizeof(struct jbd2_journal_block_tail);
 376
 377	/*
 378	 * First job: lock down the current transaction and wait for
 379	 * all outstanding updates to complete.
 380	 */
 381
 382	/* Do we need to erase the effects of a prior jbd2_journal_flush? */
 383	if (journal->j_flags & JBD2_FLUSHED) {
 384		jbd_debug(3, "super block updated\n");
 385		mutex_lock_io(&journal->j_checkpoint_mutex);
 386		/*
 387		 * We hold j_checkpoint_mutex so tail cannot change under us.
 388		 * We don't need any special data guarantees for writing sb
 389		 * since journal is empty and it is ok for write to be
 390		 * flushed only with transaction commit.
 391		 */
 392		jbd2_journal_update_sb_log_tail(journal,
 393						journal->j_tail_sequence,
 394						journal->j_tail,
 395						REQ_SYNC);
 396		mutex_unlock(&journal->j_checkpoint_mutex);
 397	} else {
 398		jbd_debug(3, "superblock not updated\n");
 399	}
 400
 401	J_ASSERT(journal->j_running_transaction != NULL);
 402	J_ASSERT(journal->j_committing_transaction == NULL);
 403
 404	commit_transaction = journal->j_running_transaction;
 405
 406	trace_jbd2_start_commit(journal, commit_transaction);
 407	jbd_debug(1, "JBD2: starting commit of transaction %d\n",
 408			commit_transaction->t_tid);
 409
 410	write_lock(&journal->j_state_lock);
 411	J_ASSERT(commit_transaction->t_state == T_RUNNING);
 412	commit_transaction->t_state = T_LOCKED;
 413
 414	trace_jbd2_commit_locking(journal, commit_transaction);
 415	stats.run.rs_wait = commit_transaction->t_max_wait;
 416	stats.run.rs_request_delay = 0;
 417	stats.run.rs_locked = jiffies;
 418	if (commit_transaction->t_requested)
 419		stats.run.rs_request_delay =
 420			jbd2_time_diff(commit_transaction->t_requested,
 421				       stats.run.rs_locked);
 422	stats.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
 423					      stats.run.rs_locked);
 424
 425	spin_lock(&commit_transaction->t_handle_lock);
 426	while (atomic_read(&commit_transaction->t_updates)) {
 427		DEFINE_WAIT(wait);
 428
 429		prepare_to_wait(&journal->j_wait_updates, &wait,
 430					TASK_UNINTERRUPTIBLE);
 431		if (atomic_read(&commit_transaction->t_updates)) {
 432			spin_unlock(&commit_transaction->t_handle_lock);
 433			write_unlock(&journal->j_state_lock);
 434			schedule();
 435			write_lock(&journal->j_state_lock);
 436			spin_lock(&commit_transaction->t_handle_lock);
 437		}
 438		finish_wait(&journal->j_wait_updates, &wait);
 439	}
 440	spin_unlock(&commit_transaction->t_handle_lock);
 
 
 441
 442	J_ASSERT (atomic_read(&commit_transaction->t_outstanding_credits) <=
 443			journal->j_max_transaction_buffers);
 444
 445	/*
 446	 * First thing we are allowed to do is to discard any remaining
 447	 * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
 448	 * that there are no such buffers: if a large filesystem
 449	 * operation like a truncate needs to split itself over multiple
 450	 * transactions, then it may try to do a jbd2_journal_restart() while
 451	 * there are still BJ_Reserved buffers outstanding.  These must
 452	 * be released cleanly from the current transaction.
 453	 *
 454	 * In this case, the filesystem must still reserve write access
 455	 * again before modifying the buffer in the new transaction, but
 456	 * we do not require it to remember exactly which old buffers it
 457	 * has reserved.  This is consistent with the existing behaviour
 458	 * that multiple jbd2_journal_get_write_access() calls to the same
 459	 * buffer are perfectly permissible.
 460	 */
 461	while (commit_transaction->t_reserved_list) {
 462		jh = commit_transaction->t_reserved_list;
 463		JBUFFER_TRACE(jh, "reserved, unused: refile");
 464		/*
 465		 * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
 466		 * leave undo-committed data.
 467		 */
 468		if (jh->b_committed_data) {
 469			struct buffer_head *bh = jh2bh(jh);
 470
 471			jbd_lock_bh_state(bh);
 472			jbd2_free(jh->b_committed_data, bh->b_size);
 473			jh->b_committed_data = NULL;
 474			jbd_unlock_bh_state(bh);
 475		}
 476		jbd2_journal_refile_buffer(journal, jh);
 477	}
 478
 479	/*
 480	 * Now try to drop any written-back buffers from the journal's
 481	 * checkpoint lists.  We do this *before* commit because it potentially
 482	 * frees some memory
 483	 */
 484	spin_lock(&journal->j_list_lock);
 485	__jbd2_journal_clean_checkpoint_list(journal, false);
 486	spin_unlock(&journal->j_list_lock);
 487
 488	jbd_debug(3, "JBD2: commit phase 1\n");
 489
 490	/*
 491	 * Clear revoked flag to reflect there is no revoked buffers
 492	 * in the next transaction which is going to be started.
 493	 */
 494	jbd2_clear_buffer_revoked_flags(journal);
 495
 496	/*
 497	 * Switch to a new revoke table.
 498	 */
 499	jbd2_journal_switch_revoke_table(journal);
 500
 501	/*
 502	 * Reserved credits cannot be claimed anymore, free them
 503	 */
 504	atomic_sub(atomic_read(&journal->j_reserved_credits),
 505		   &commit_transaction->t_outstanding_credits);
 506
 
 507	trace_jbd2_commit_flushing(journal, commit_transaction);
 508	stats.run.rs_flushing = jiffies;
 509	stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
 510					     stats.run.rs_flushing);
 511
 512	commit_transaction->t_state = T_FLUSH;
 513	journal->j_committing_transaction = commit_transaction;
 514	journal->j_running_transaction = NULL;
 515	start_time = ktime_get();
 516	commit_transaction->t_log_start = journal->j_head;
 517	wake_up(&journal->j_wait_transaction_locked);
 518	write_unlock(&journal->j_state_lock);
 519
 520	jbd_debug(3, "JBD2: commit phase 2a\n");
 521
 522	/*
 523	 * Now start flushing things to disk, in the order they appear
 524	 * on the transaction lists.  Data blocks go first.
 525	 */
 526	err = journal_submit_data_buffers(journal, commit_transaction);
 527	if (err)
 528		jbd2_journal_abort(journal, err);
 529
 530	blk_start_plug(&plug);
 531	jbd2_journal_write_revoke_records(commit_transaction, &log_bufs);
 532
 533	jbd_debug(3, "JBD2: commit phase 2b\n");
 534
 535	/*
 536	 * Way to go: we have now written out all of the data for a
 537	 * transaction!  Now comes the tricky part: we need to write out
 538	 * metadata.  Loop over the transaction's entire buffer list:
 539	 */
 540	write_lock(&journal->j_state_lock);
 541	commit_transaction->t_state = T_COMMIT;
 542	write_unlock(&journal->j_state_lock);
 543
 544	trace_jbd2_commit_logging(journal, commit_transaction);
 545	stats.run.rs_logging = jiffies;
 546	stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
 547					       stats.run.rs_logging);
 548	stats.run.rs_blocks =
 549		atomic_read(&commit_transaction->t_outstanding_credits);
 550	stats.run.rs_blocks_logged = 0;
 551
 552	J_ASSERT(commit_transaction->t_nr_buffers <=
 553		 atomic_read(&commit_transaction->t_outstanding_credits));
 554
 555	err = 0;
 556	bufs = 0;
 557	descriptor = NULL;
 558	while (commit_transaction->t_buffers) {
 559
 560		/* Find the next buffer to be journaled... */
 561
 562		jh = commit_transaction->t_buffers;
 563
 564		/* If we're in abort mode, we just un-journal the buffer and
 565		   release it. */
 566
 567		if (is_journal_aborted(journal)) {
 568			clear_buffer_jbddirty(jh2bh(jh));
 569			JBUFFER_TRACE(jh, "journal is aborting: refile");
 570			jbd2_buffer_abort_trigger(jh,
 571						  jh->b_frozen_data ?
 572						  jh->b_frozen_triggers :
 573						  jh->b_triggers);
 574			jbd2_journal_refile_buffer(journal, jh);
 575			/* If that was the last one, we need to clean up
 576			 * any descriptor buffers which may have been
 577			 * already allocated, even if we are now
 578			 * aborting. */
 579			if (!commit_transaction->t_buffers)
 580				goto start_journal_io;
 581			continue;
 582		}
 583
 584		/* Make sure we have a descriptor block in which to
 585		   record the metadata buffer. */
 586
 587		if (!descriptor) {
 588			J_ASSERT (bufs == 0);
 589
 590			jbd_debug(4, "JBD2: get descriptor\n");
 591
 592			descriptor = jbd2_journal_get_descriptor_buffer(
 593							commit_transaction,
 594							JBD2_DESCRIPTOR_BLOCK);
 595			if (!descriptor) {
 596				jbd2_journal_abort(journal, -EIO);
 597				continue;
 598			}
 599
 600			jbd_debug(4, "JBD2: got buffer %llu (%p)\n",
 601				(unsigned long long)descriptor->b_blocknr,
 602				descriptor->b_data);
 603			tagp = &descriptor->b_data[sizeof(journal_header_t)];
 604			space_left = descriptor->b_size -
 605						sizeof(journal_header_t);
 606			first_tag = 1;
 607			set_buffer_jwrite(descriptor);
 608			set_buffer_dirty(descriptor);
 609			wbuf[bufs++] = descriptor;
 610
 611			/* Record it so that we can wait for IO
 612                           completion later */
 613			BUFFER_TRACE(descriptor, "ph3: file as descriptor");
 614			jbd2_file_log_bh(&log_bufs, descriptor);
 615		}
 616
 617		/* Where is the buffer to be written? */
 618
 619		err = jbd2_journal_next_log_block(journal, &blocknr);
 620		/* If the block mapping failed, just abandon the buffer
 621		   and repeat this loop: we'll fall into the
 622		   refile-on-abort condition above. */
 623		if (err) {
 624			jbd2_journal_abort(journal, err);
 625			continue;
 626		}
 627
 628		/*
 629		 * start_this_handle() uses t_outstanding_credits to determine
 630		 * the free space in the log, but this counter is changed
 631		 * by jbd2_journal_next_log_block() also.
 632		 */
 633		atomic_dec(&commit_transaction->t_outstanding_credits);
 634
 635		/* Bump b_count to prevent truncate from stumbling over
 636                   the shadowed buffer!  @@@ This can go if we ever get
 637                   rid of the shadow pairing of buffers. */
 638		atomic_inc(&jh2bh(jh)->b_count);
 639
 640		/*
 641		 * Make a temporary IO buffer with which to write it out
 642		 * (this will requeue the metadata buffer to BJ_Shadow).
 643		 */
 644		set_bit(BH_JWrite, &jh2bh(jh)->b_state);
 645		JBUFFER_TRACE(jh, "ph3: write metadata");
 646		flags = jbd2_journal_write_metadata_buffer(commit_transaction,
 647						jh, &wbuf[bufs], blocknr);
 648		if (flags < 0) {
 649			jbd2_journal_abort(journal, flags);
 650			continue;
 651		}
 652		jbd2_file_log_bh(&io_bufs, wbuf[bufs]);
 653
 654		/* Record the new block's tag in the current descriptor
 655                   buffer */
 656
 657		tag_flag = 0;
 658		if (flags & 1)
 659			tag_flag |= JBD2_FLAG_ESCAPE;
 660		if (!first_tag)
 661			tag_flag |= JBD2_FLAG_SAME_UUID;
 662
 663		tag = (journal_block_tag_t *) tagp;
 664		write_tag_block(journal, tag, jh2bh(jh)->b_blocknr);
 665		tag->t_flags = cpu_to_be16(tag_flag);
 666		jbd2_block_tag_csum_set(journal, tag, wbuf[bufs],
 667					commit_transaction->t_tid);
 668		tagp += tag_bytes;
 669		space_left -= tag_bytes;
 670		bufs++;
 671
 672		if (first_tag) {
 673			memcpy (tagp, journal->j_uuid, 16);
 674			tagp += 16;
 675			space_left -= 16;
 676			first_tag = 0;
 677		}
 678
 679		/* If there's no more to do, or if the descriptor is full,
 680		   let the IO rip! */
 681
 682		if (bufs == journal->j_wbufsize ||
 683		    commit_transaction->t_buffers == NULL ||
 684		    space_left < tag_bytes + 16 + csum_size) {
 685
 686			jbd_debug(4, "JBD2: Submit %d IOs\n", bufs);
 687
 688			/* Write an end-of-descriptor marker before
 689                           submitting the IOs.  "tag" still points to
 690                           the last tag we set up. */
 691
 692			tag->t_flags |= cpu_to_be16(JBD2_FLAG_LAST_TAG);
 693
 694			jbd2_descriptor_block_csum_set(journal, descriptor);
 695start_journal_io:
 
 
 
 
 696			for (i = 0; i < bufs; i++) {
 697				struct buffer_head *bh = wbuf[i];
 698				/*
 699				 * Compute checksum.
 700				 */
 701				if (jbd2_has_feature_checksum(journal)) {
 702					crc32_sum =
 703					    jbd2_checksum_data(crc32_sum, bh);
 704				}
 705
 706				lock_buffer(bh);
 707				clear_buffer_dirty(bh);
 708				set_buffer_uptodate(bh);
 709				bh->b_end_io = journal_end_buffer_io_sync;
 710				submit_bh(REQ_OP_WRITE, REQ_SYNC, bh);
 711			}
 712			cond_resched();
 713			stats.run.rs_blocks_logged += bufs;
 714
 715			/* Force a new descriptor to be generated next
 716                           time round the loop. */
 717			descriptor = NULL;
 718			bufs = 0;
 719		}
 720	}
 721
 722	err = journal_finish_inode_data_buffers(journal, commit_transaction);
 723	if (err) {
 724		printk(KERN_WARNING
 725			"JBD2: Detected IO errors while flushing file data "
 726		       "on %s\n", journal->j_devname);
 727		if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
 728			jbd2_journal_abort(journal, err);
 729		err = 0;
 730	}
 731
 732	/*
 733	 * Get current oldest transaction in the log before we issue flush
 734	 * to the filesystem device. After the flush we can be sure that
 735	 * blocks of all older transactions are checkpointed to persistent
 736	 * storage and we will be safe to update journal start in the
 737	 * superblock with the numbers we get here.
 738	 */
 739	update_tail =
 740		jbd2_journal_get_log_tail(journal, &first_tid, &first_block);
 741
 742	write_lock(&journal->j_state_lock);
 743	if (update_tail) {
 744		long freed = first_block - journal->j_tail;
 745
 746		if (first_block < journal->j_tail)
 747			freed += journal->j_last - journal->j_first;
 748		/* Update tail only if we free significant amount of space */
 749		if (freed < journal->j_maxlen / 4)
 750			update_tail = 0;
 751	}
 752	J_ASSERT(commit_transaction->t_state == T_COMMIT);
 753	commit_transaction->t_state = T_COMMIT_DFLUSH;
 754	write_unlock(&journal->j_state_lock);
 755
 756	/* 
 757	 * If the journal is not located on the file system device,
 758	 * then we must flush the file system device before we issue
 759	 * the commit record
 760	 */
 761	if (commit_transaction->t_need_data_flush &&
 762	    (journal->j_fs_dev != journal->j_dev) &&
 763	    (journal->j_flags & JBD2_BARRIER))
 764		blkdev_issue_flush(journal->j_fs_dev, GFP_NOFS, NULL);
 765
 766	/* Done it all: now write the commit record asynchronously. */
 767	if (jbd2_has_feature_async_commit(journal)) {
 768		err = journal_submit_commit_record(journal, commit_transaction,
 769						 &cbh, crc32_sum);
 770		if (err)
 771			__jbd2_journal_abort_hard(journal);
 772	}
 773
 774	blk_finish_plug(&plug);
 775
 776	/* Lo and behold: we have just managed to send a transaction to
 777           the log.  Before we can commit it, wait for the IO so far to
 778           complete.  Control buffers being written are on the
 779           transaction's t_log_list queue, and metadata buffers are on
 780           the io_bufs list.
 781
 782	   Wait for the buffers in reverse order.  That way we are
 783	   less likely to be woken up until all IOs have completed, and
 784	   so we incur less scheduling load.
 785	*/
 786
 787	jbd_debug(3, "JBD2: commit phase 3\n");
 788
 789	while (!list_empty(&io_bufs)) {
 790		struct buffer_head *bh = list_entry(io_bufs.prev,
 791						    struct buffer_head,
 792						    b_assoc_buffers);
 793
 794		wait_on_buffer(bh);
 795		cond_resched();
 796
 797		if (unlikely(!buffer_uptodate(bh)))
 798			err = -EIO;
 799		jbd2_unfile_log_bh(bh);
 
 800
 801		/*
 802		 * The list contains temporary buffer heads created by
 803		 * jbd2_journal_write_metadata_buffer().
 804		 */
 805		BUFFER_TRACE(bh, "dumping temporary bh");
 806		__brelse(bh);
 807		J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
 808		free_buffer_head(bh);
 809
 810		/* We also have to refile the corresponding shadowed buffer */
 811		jh = commit_transaction->t_shadow_list->b_tprev;
 812		bh = jh2bh(jh);
 813		clear_buffer_jwrite(bh);
 814		J_ASSERT_BH(bh, buffer_jbddirty(bh));
 815		J_ASSERT_BH(bh, !buffer_shadow(bh));
 816
 817		/* The metadata is now released for reuse, but we need
 818                   to remember it against this transaction so that when
 819                   we finally commit, we can do any checkpointing
 820                   required. */
 821		JBUFFER_TRACE(jh, "file as BJ_Forget");
 822		jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
 823		JBUFFER_TRACE(jh, "brelse shadowed buffer");
 824		__brelse(bh);
 825	}
 826
 827	J_ASSERT (commit_transaction->t_shadow_list == NULL);
 828
 829	jbd_debug(3, "JBD2: commit phase 4\n");
 830
 831	/* Here we wait for the revoke record and descriptor record buffers */
 832	while (!list_empty(&log_bufs)) {
 833		struct buffer_head *bh;
 834
 835		bh = list_entry(log_bufs.prev, struct buffer_head, b_assoc_buffers);
 836		wait_on_buffer(bh);
 837		cond_resched();
 838
 839		if (unlikely(!buffer_uptodate(bh)))
 840			err = -EIO;
 841
 842		BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
 843		clear_buffer_jwrite(bh);
 844		jbd2_unfile_log_bh(bh);
 
 845		__brelse(bh);		/* One for getblk */
 846		/* AKPM: bforget here */
 847	}
 848
 849	if (err)
 850		jbd2_journal_abort(journal, err);
 851
 852	jbd_debug(3, "JBD2: commit phase 5\n");
 853	write_lock(&journal->j_state_lock);
 854	J_ASSERT(commit_transaction->t_state == T_COMMIT_DFLUSH);
 855	commit_transaction->t_state = T_COMMIT_JFLUSH;
 856	write_unlock(&journal->j_state_lock);
 857
 858	if (!jbd2_has_feature_async_commit(journal)) {
 859		err = journal_submit_commit_record(journal, commit_transaction,
 860						&cbh, crc32_sum);
 861		if (err)
 862			__jbd2_journal_abort_hard(journal);
 863	}
 864	if (cbh)
 865		err = journal_wait_on_commit_record(journal, cbh);
 
 866	if (jbd2_has_feature_async_commit(journal) &&
 867	    journal->j_flags & JBD2_BARRIER) {
 868		blkdev_issue_flush(journal->j_dev, GFP_NOFS, NULL);
 869	}
 870
 871	if (err)
 872		jbd2_journal_abort(journal, err);
 873
 
 
 
 874	/*
 875	 * Now disk caches for filesystem device are flushed so we are safe to
 876	 * erase checkpointed transactions from the log by updating journal
 877	 * superblock.
 878	 */
 879	if (update_tail)
 880		jbd2_update_log_tail(journal, first_tid, first_block);
 881
 882	/* End of a transaction!  Finally, we can do checkpoint
 883           processing: any buffers committed as a result of this
 884           transaction can be removed from any checkpoint list it was on
 885           before. */
 886
 887	jbd_debug(3, "JBD2: commit phase 6\n");
 888
 889	J_ASSERT(list_empty(&commit_transaction->t_inode_list));
 890	J_ASSERT(commit_transaction->t_buffers == NULL);
 891	J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
 892	J_ASSERT(commit_transaction->t_shadow_list == NULL);
 893
 894restart_loop:
 895	/*
 896	 * As there are other places (journal_unmap_buffer()) adding buffers
 897	 * to this list we have to be careful and hold the j_list_lock.
 898	 */
 899	spin_lock(&journal->j_list_lock);
 900	while (commit_transaction->t_forget) {
 901		transaction_t *cp_transaction;
 902		struct buffer_head *bh;
 903		int try_to_free = 0;
 
 904
 905		jh = commit_transaction->t_forget;
 906		spin_unlock(&journal->j_list_lock);
 907		bh = jh2bh(jh);
 908		/*
 909		 * Get a reference so that bh cannot be freed before we are
 910		 * done with it.
 911		 */
 912		get_bh(bh);
 913		jbd_lock_bh_state(bh);
 914		J_ASSERT_JH(jh,	jh->b_transaction == commit_transaction);
 915
 916		/*
 917		 * If there is undo-protected committed data against
 918		 * this buffer, then we can remove it now.  If it is a
 919		 * buffer needing such protection, the old frozen_data
 920		 * field now points to a committed version of the
 921		 * buffer, so rotate that field to the new committed
 922		 * data.
 923		 *
 924		 * Otherwise, we can just throw away the frozen data now.
 925		 *
 926		 * We also know that the frozen data has already fired
 927		 * its triggers if they exist, so we can clear that too.
 928		 */
 929		if (jh->b_committed_data) {
 930			jbd2_free(jh->b_committed_data, bh->b_size);
 931			jh->b_committed_data = NULL;
 932			if (jh->b_frozen_data) {
 933				jh->b_committed_data = jh->b_frozen_data;
 934				jh->b_frozen_data = NULL;
 935				jh->b_frozen_triggers = NULL;
 936			}
 937		} else if (jh->b_frozen_data) {
 938			jbd2_free(jh->b_frozen_data, bh->b_size);
 939			jh->b_frozen_data = NULL;
 940			jh->b_frozen_triggers = NULL;
 941		}
 942
 943		spin_lock(&journal->j_list_lock);
 944		cp_transaction = jh->b_cp_transaction;
 945		if (cp_transaction) {
 946			JBUFFER_TRACE(jh, "remove from old cp transaction");
 947			cp_transaction->t_chp_stats.cs_dropped++;
 948			__jbd2_journal_remove_checkpoint(jh);
 949		}
 950
 951		/* Only re-checkpoint the buffer_head if it is marked
 952		 * dirty.  If the buffer was added to the BJ_Forget list
 953		 * by jbd2_journal_forget, it may no longer be dirty and
 954		 * there's no point in keeping a checkpoint record for
 955		 * it. */
 956
 957		/*
 958		* A buffer which has been freed while still being journaled by
 959		* a previous transaction.
 960		*/
 961		if (buffer_freed(bh)) {
 
 
 
 
 
 
 
 
 
 
 962			/*
 963			 * If the running transaction is the one containing
 964			 * "add to orphan" operation (b_next_transaction !=
 965			 * NULL), we have to wait for that transaction to
 966			 * commit before we can really get rid of the buffer.
 967			 * So just clear b_modified to not confuse transaction
 968			 * credit accounting and refile the buffer to
 969			 * BJ_Forget of the running transaction. If the just
 970			 * committed transaction contains "add to orphan"
 971			 * operation, we can completely invalidate the buffer
 972			 * now. We are rather through in that since the
 973			 * buffer may be still accessible when blocksize <
 974			 * pagesize and it is attached to the last partial
 975			 * page.
 976			 */
 977			jh->b_modified = 0;
 978			if (!jh->b_next_transaction) {
 979				clear_buffer_freed(bh);
 980				clear_buffer_jbddirty(bh);
 981				clear_buffer_mapped(bh);
 982				clear_buffer_new(bh);
 983				clear_buffer_req(bh);
 984				bh->b_bdev = NULL;
 985			}
 986		}
 987
 988		if (buffer_jbddirty(bh)) {
 989			JBUFFER_TRACE(jh, "add to new checkpointing trans");
 990			__jbd2_journal_insert_checkpoint(jh, commit_transaction);
 991			if (is_journal_aborted(journal))
 992				clear_buffer_jbddirty(bh);
 993		} else {
 994			J_ASSERT_BH(bh, !buffer_dirty(bh));
 995			/*
 996			 * The buffer on BJ_Forget list and not jbddirty means
 997			 * it has been freed by this transaction and hence it
 998			 * could not have been reallocated until this
 999			 * transaction has committed. *BUT* it could be
1000			 * reallocated once we have written all the data to
1001			 * disk and before we process the buffer on BJ_Forget
1002			 * list.
1003			 */
1004			if (!jh->b_next_transaction)
1005				try_to_free = 1;
1006		}
1007		JBUFFER_TRACE(jh, "refile or unfile buffer");
1008		__jbd2_journal_refile_buffer(jh);
1009		jbd_unlock_bh_state(bh);
 
 
1010		if (try_to_free)
1011			release_buffer_page(bh);	/* Drops bh reference */
1012		else
1013			__brelse(bh);
1014		cond_resched_lock(&journal->j_list_lock);
1015	}
1016	spin_unlock(&journal->j_list_lock);
1017	/*
1018	 * This is a bit sleazy.  We use j_list_lock to protect transition
1019	 * of a transaction into T_FINISHED state and calling
1020	 * __jbd2_journal_drop_transaction(). Otherwise we could race with
1021	 * other checkpointing code processing the transaction...
1022	 */
1023	write_lock(&journal->j_state_lock);
1024	spin_lock(&journal->j_list_lock);
1025	/*
1026	 * Now recheck if some buffers did not get attached to the transaction
1027	 * while the lock was dropped...
1028	 */
1029	if (commit_transaction->t_forget) {
1030		spin_unlock(&journal->j_list_lock);
1031		write_unlock(&journal->j_state_lock);
1032		goto restart_loop;
1033	}
1034
1035	/* Add the transaction to the checkpoint list
1036	 * __journal_remove_checkpoint() can not destroy transaction
1037	 * under us because it is not marked as T_FINISHED yet */
1038	if (journal->j_checkpoint_transactions == NULL) {
1039		journal->j_checkpoint_transactions = commit_transaction;
1040		commit_transaction->t_cpnext = commit_transaction;
1041		commit_transaction->t_cpprev = commit_transaction;
1042	} else {
1043		commit_transaction->t_cpnext =
1044			journal->j_checkpoint_transactions;
1045		commit_transaction->t_cpprev =
1046			commit_transaction->t_cpnext->t_cpprev;
1047		commit_transaction->t_cpnext->t_cpprev =
1048			commit_transaction;
1049		commit_transaction->t_cpprev->t_cpnext =
1050				commit_transaction;
1051	}
1052	spin_unlock(&journal->j_list_lock);
1053
1054	/* Done with this transaction! */
1055
1056	jbd_debug(3, "JBD2: commit phase 7\n");
1057
1058	J_ASSERT(commit_transaction->t_state == T_COMMIT_JFLUSH);
1059
1060	commit_transaction->t_start = jiffies;
1061	stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
1062					      commit_transaction->t_start);
1063
1064	/*
1065	 * File the transaction statistics
1066	 */
1067	stats.ts_tid = commit_transaction->t_tid;
1068	stats.run.rs_handle_count =
1069		atomic_read(&commit_transaction->t_handle_count);
1070	trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
1071			     commit_transaction->t_tid, &stats.run);
1072	stats.ts_requested = (commit_transaction->t_requested) ? 1 : 0;
1073
1074	commit_transaction->t_state = T_COMMIT_CALLBACK;
1075	J_ASSERT(commit_transaction == journal->j_committing_transaction);
1076	journal->j_commit_sequence = commit_transaction->t_tid;
1077	journal->j_committing_transaction = NULL;
1078	commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1079
1080	/*
1081	 * weight the commit time higher than the average time so we don't
1082	 * react too strongly to vast changes in the commit time
1083	 */
1084	if (likely(journal->j_average_commit_time))
1085		journal->j_average_commit_time = (commit_time +
1086				journal->j_average_commit_time*3) / 4;
1087	else
1088		journal->j_average_commit_time = commit_time;
1089
1090	write_unlock(&journal->j_state_lock);
1091
1092	if (journal->j_commit_callback)
1093		journal->j_commit_callback(journal, commit_transaction);
1094
1095	trace_jbd2_end_commit(journal, commit_transaction);
1096	jbd_debug(1, "JBD2: commit %d complete, head %d\n",
1097		  journal->j_commit_sequence, journal->j_tail_sequence);
1098
1099	write_lock(&journal->j_state_lock);
1100	spin_lock(&journal->j_list_lock);
1101	commit_transaction->t_state = T_FINISHED;
1102	/* Check if the transaction can be dropped now that we are finished */
1103	if (commit_transaction->t_checkpoint_list == NULL &&
1104	    commit_transaction->t_checkpoint_io_list == NULL) {
1105		__jbd2_journal_drop_transaction(journal, commit_transaction);
1106		jbd2_journal_free_transaction(commit_transaction);
1107	}
1108	spin_unlock(&journal->j_list_lock);
1109	write_unlock(&journal->j_state_lock);
1110	wake_up(&journal->j_wait_done_commit);
1111
1112	/*
1113	 * Calculate overall stats
1114	 */
1115	spin_lock(&journal->j_history_lock);
1116	journal->j_stats.ts_tid++;
1117	journal->j_stats.ts_requested += stats.ts_requested;
1118	journal->j_stats.run.rs_wait += stats.run.rs_wait;
1119	journal->j_stats.run.rs_request_delay += stats.run.rs_request_delay;
1120	journal->j_stats.run.rs_running += stats.run.rs_running;
1121	journal->j_stats.run.rs_locked += stats.run.rs_locked;
1122	journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
1123	journal->j_stats.run.rs_logging += stats.run.rs_logging;
1124	journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
1125	journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
1126	journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
1127	spin_unlock(&journal->j_history_lock);
1128}
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}