Linux jbd2_journal_commit_transaction日志提交与forget链表

Linux jbd2_journal_commit_transaction日志提交与forget链表

jbd2（Journaling Block Device 2）是ext4等文件系统使用的通用日志层。事务提交由jbd2_journal_commit_transaction()实现，这是jbd2中最复杂的函数之一，负责将运行中的事务写入日志磁盘区域并完成原子更新。

每个事务由struct transaction_s表示，包含多个链表关键链表：t_buffers（元数据缓冲区链表）、t_forget（forget链表）、t_shadow_list、t_iobuf_list等。

事务提交的主路径从jbd2_journal_commit_transaction()开始：

int jbd2_journal_commit_transaction(journal_t *journal)
{
transaction_t *commit_transaction;
struct journal_head *jh;
struct buffer_head *bh;
int ret;
int flags;
DEFINE_WAIT(wait);

commit_transaction = journal->j_running_transaction;
commit_transaction->t_state = T_LOCKED;

spin_lock(&commit_transaction->t_handle_lock);
while (atomic_read(&commit_transaction->t_updates)) {
DEFINE_WAIT(wait);
prepare_to_wait(&journal->j_wait_updates, &wait,
TASK_UNINTERRUPTIBLE);
spin_unlock(&commit_transaction->t_handle_lock);
schedule();
spin_lock(&commit_transaction->t_handle_lock);
finish_wait(&journal->j_wait_updates, &wait);
}
spin_unlock(&commit_transaction->t_handle_lock);

jbd2_journal_write_revoke_records(journal, commit_transaction);

jbd2_journal_submit_data_buffers(journal, commit_transaction);

jbd2_journal_write_metadata_buffer(journal, commit_transaction, &wbuf);

commit_transaction->t_state = T_FLUSH;
journal->j_ committing_transaction = commit_transaction;
... // Journal block tag + descriptor block writing

jbd2_journal_update_sb_log_tail(journal,
journal->j_tail,
journal->j_tail_sequence,
GFP_NOFS);
return 0;
}

第一阶段是T_LOCKED状态。函数等待t_updates原子计数器归零，表示所有正在运行的handle都已结束。然后锁定事务，不允许新的handle加入。随后写入revoke记录（用于取消已提交的日志块重放），标记已删除的块不需要在恢复时重做。

第二阶段是元数据缓冲区的flush写入。jbd2_journal_write_metadata_buffer()对每个脏的journal_head执行操作：

int jbd2_journal_write_metadata_buffer(transaction_t *transaction,
struct journal_head *jh_in)
{
struct page *page;
struct page *new_page;
struct buffer_head *bh_in = jh2bh(jh_in);
char *tmp_buffer;
int done;

new_page = jbd2_alloc(journal->j_blocksize, GFP_NOFS);
tmp_buffer = page_address(new_page);

page = bh_in->b_page;
wait_on_buffer(bh_in);

memcpy(tmp_buffer, page_address(page) + bh_in->b_offset,
journal->j_blocksize);

set_bh_page(bh_in, new_page, 0);
jh_in->b_frozen_data = tmp_buffer;
jh_in->b_committed_data = NULL;

J_ASSERT_JH(jh_in, !jh_in->b_committed_data);
jh_in->b_jlist = BJ_Shadow;
set_buffer_shadow(bh_in);
set_buffer_jwrite(bh_in);
return 0;
}

该函数将元数据缓冲区的数据复制到b_frozen_data（冻结快照），保证日志写入时的数据一致性。原始缓冲区被标记为BJ_Shadow加入t_shadow_list，表示正等待日志提交完成后释放。

forget链表的处理是事务提交中的关键环节。当文件系统删除一个块时，jbd2_journal_forget()将该缓冲区的journal_head加入事务的t_forget链表：

int jbd2_journal_forget(journal_t *journal, struct buffer_head *bh)
{
transaction_t *transaction;
struct journal_head *jh;

jh = jbd2_journal_add_journal_head(bh);
transaction = jh->b_transaction;

if (jh->b_cp_transaction) {
JBUFFER_TRACE(jh, "on cp+unfrozen");
__jbd2_journal_unfile_buffer(jh);
jbd2_journal_remove_journal_head(bh);
__brelse(bh);
} else if (transaction == journal->j_committing_transaction) {
set_buffer_freed(bh);
jh->b_frozen_data = NULL;
jh->b_committed_data = NULL;
jh->b_next_transaction = NULL;
jh->b_transaction = NULL;
jbd2_journal_put_journal_head(jh);
} else {
if (jh->b_transaction) {
J_ASSERT_JH(jh, jh->b_jlist == BJ_None ||
jh->b_jlist == BJ_SyncData ||
jh->b_jlist == BJ_Metadata);
if (jh->b_jlist != BJ_None)
__jbd2_journal_refile_buffer(jh);
}
set_buffer_freed(bh);
__jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
}
return 0;
}

当buffer被forget后，它被放入事务的BJ_Forget链表（t_forget_list）。在事务提交时，jbd2_journal_commit_transaction()遍历t_forget链表的journal_head，逐个调用jbd2_journal_invalidatepage()或释放buffer：

commit_transaction->t_state = T_COMMIT;
while (commit_transaction->t_forget_list) {
jh = commit_transaction->t_forget_list;
bh = jh2bh(jh);

if (buffer_freed(bh)) {
clear_buffer_freed(bh);
clear_buffer_jbddirty(bh);
J_ASSERT_JH(jh, !buffer_journaled(jh));
J_ASSERT_JH(jh, !buffer_journal_dirty(jh));
J_ASSERT_JH(jh, !buffer_mapped(bh));
if (jh->b_cp_transaction) {
__jbd2_journal_unfile_buffer(jh);
jbd2_journal_remove_journal_head(bh);
__brelse(bh);
}
} else {
__jbd2_journal_unfile_buffer(jh);
jbd2_journal_remove_journal_head(bh);
__brelse(bh);
}
}

forget链表中的buffer如果是buffer_freed状态，表示对应的block已经在文件系统层面被释放。日志提交时直接清空相关标记并释放journal_head。如果不是freed状态，表示buffer只是从当前事务解绑，但不释放block本身。

事务提交的最后阶段是T_FLUSH和T_COMMIT。在T_FLUSH阶段，jbd2发出屏障IO请求确保日志块先于数据块落盘。在T_COMMIT阶段，写入commit block（包含事务的checksum和时间戳）完成事务提交。

jbd2_journal_commit_transaction()的最终操作是更新journal superblock中的日志尾部位置：调用jbd2_journal_update_sb_log_tail()将j_tail和j_tail_sequence写入journal superblock，标记已经被checkpoint的日志位置。至此，日志空间可被循环复用。