Linux jbd2_journal_recover日志恢复与superblock标记
Linux jbd2_journal_recover日志恢复与superblock标记
jbd2日志恢复在文件系统mount时由jbd2_journal_recover()触发，其核心任务是从日志磁盘区域中扫描已提交但尚未写入文件系统元数据区域的事务，并重放这些事务的元数据块。
日志恢复的入口函数是jbd2_journal_load()，它在mount流程中调用：
int jbd2_journal_load(journal_t *journal)
{
int err;
journal_superblock_t *sb;
err = jbd2_journal_read_superblock(journal, GFP_KERNEL);
if (err)
return err;
sb = journal->j_superblock;
journal->j_tail = be32_to_cpu(sb->s_start);
journal->j_tail_sequence = be32_to_cpu(sb->s_sequence);
journal->j_first = be32_to_cpu(sb->s_first);
journal->j_last = be32_to_cpu(sb->s_maxlen);
if (journal->j_tail_sequence != be32_to_cpu(sb->s_sequence)) {
err = jbd2_journal_recover(journal);
if (err)
return err;
}
err = jbd2_journal_bmap(journal, 0, &journal->j_blk_offset);
...
return 0;
}
函数首先读取journal superblock，从中获取以下关键字段：s_start（日志起始块的物理块号）、s_sequence（当前日志序列号）、s_first和s_maxlen（日志区域的物理范围）。如果s_sequence不等于j_tail_sequence，表示日志不为空，需要执行恢复。
jbd2_journal_recover()调用do_one_pass()执行实际的扫描和重放操作：
int jbd2_journal_recover(journal_t *journal)
{
struct recovery_info info;
int ret;
memset(&info, 0, sizeof(info));
ret = do_one_pass(journal, &info, PASS_SCAN);
if (ret)
return ret;
jbd2_clear_revoke(journal, &info);
ret = do_one_pass(journal, &info, PASS_REPLAY);
if (ret)
return ret;
jbd2_clear_revoke(journal, &info);
ret = do_one_pass(journal, &info, PASS_REVOKE);
if (ret)
return ret;
jbd2_clear_revoke(journal, &info);
return 0;
}
恢复过程分三个阶段（pass）。PASS_SCAN：从日志尾部开始向前扫描所有日志块，构建事务结构，收集revoke记录。PASS_REPLAY：按事务顺序重放所有已提交的元数据块。PASS_REVOKE：处理revoke记录，跳过被撤销的日志块重放。
do_one_pass()的扫描逻辑如下：
static int do_one_pass(journal_t *journal,
struct recovery_info *info,
enum passtype pass)
{
unsigned int first_commit_ID;
unsigned int next_commit_ID;
unsigned long long next_log_block;
int err, success = 0;
journal_superblock_t *sb;
struct buffer_head *bh;
unsigned int sequence;
int blocktype;
sb = journal->j_superblock;
next_commit_ID = info->end_commit_id;
next_log_block = info->end_transaction;
for (;;) {
bh = __jbd2_journal_read_next_block(journal, &next_log_block,
&next_commit_ID);
if (!bh)
break;
sequence = be32_to_cpu(bh->b_blocknr);
blocktype = be32_to_cpu(((journal_header_t *)bh->b_data)->h_blocktype);
switch (blocktype) {
case JBD2_DESCRIPTOR_BLOCK:
if (pass == PASS_REPLAY) {
err = jbd2_journal_replay_descriptor(journal, bh,
next_log_block);
}
break;
case JBD2_COMMIT_BLOCK:
info->end_commit_id = next_commit_ID;
break;
case JBD2_REVOKE_BLOCK:
if (pass == PASS_REVOKE) {
err = jbd2_journal_scan_revoke_record(journal, bh,
next_log_block);
}
break;
}
brelse(bh);
}
return 0;
}
扫描器从日志中逐个读取块，通过h_blocktype字段判断块类型。JBD2_DESCRIPTOR_BLOCK是描述符块，包含本次事务中所有元数据块的UUID、块号列表和校验信息。JBD2_COMMIT_BLOCK是提交块，包含事务的checksum和flags，标记事务成功提交。JBD2_REVOKE_BLOCK是撤销块，列出了不需要重放的块号列表。
在实际重放阶段，jbd2_journal_replay_descriptor()根据描述符块中的条目逐个处理：
int jbd2_journal_replay_descriptor(journal_t *journal,
struct buffer_head *descriptor,
unsigned long long next_log_block)
{
int i;
int err = 0;
struct buffer_head *bh;
unsigned long blocknr;
journal_block_tag_t *tag;
int flags;
tag = (journal_block_tag_t *)&descriptor->b_data[sizeof(journal_header_t)];
for (i = 0; i < journal->j_blocksize / sizeof(journal_block_tag_t); i++) {
blocknr = be32_to_cpu(tag->t_blocknr);
flags = tag->t_flags;
if (flags & JBD2_FLAG_SAME_UUID) {
bh = __jbd2_journal_get_descriptor_buffer(journal, blocknr);
if (!bh)
return -EIO;
err = jbd2_journal_replay_block(journal, bh, next_log_block);
if (err) {
brelse(bh);
return err;
}
if (flags & JBD2_FLAG_LAST_TAG)
break;
brelse(bh);
}
tag++;
}
return 0;
}
每个描述符块包含多个tag条目，每个tag对应一个日志数据块。t_blocknr是目标块号（文件系统元数据块的物理位置），t_flags包含JBD2_FLAG_SAME_UUID和JBD2_FLAG_LAST_TAG等标志。jbd2_journal_replay_block()将日志块的数据复制到目标元数据块：
int jbd2_journal_replay_block(journal_t *journal,
struct buffer_head *obh,
unsigned long long next_log_block)
{
struct buffer_head *nbh;
int err;
nbh = __getblk(journal->j_fs_dev, next_log_block,
journal->j_blocksize);
lock_buffer(nbh);
memcpy(nbh->b_data, obh->b_data, journal->j_blocksize);
set_buffer_uptodate(nbh);
mark_buffer_dirty(nbh);
unlock_buffer(nbh);
sync_dirty_buffer(nbh);
brelse(nbh);
return 0;
}
恢复完成后，jbd2_journal_recover()更新journal superblock的s_start和s_sequence字段，标记日志为空：
void jbd2_journal_update_sb_log_tail(journal_t *journal,
tid_t tail_tid,
unsigned long long tail_block,
int write_op)
{
journal_superblock_t *sb = journal->j_superblock;
sb->s_sequence = cpu_to_be32(tail_tid);
sb->s_start = cpu_to_be32(tail_block);
jbd2_journal_write_superblock(journal, write_op);
}
s_start被设置为0表示日志已清空，s_sequence更新为最后一个已提交事务的序列号。此时文件系统可以直接挂载，不需要再次恢复。
jbd2的superblock还包含了s_errno字段用于记录上一次mount时的错误码，以及s_feature_compat/s_feature_incompat/s_feature_ro_compat三个特性字段，用于控制日志格式的向后兼容性。当jbd2检测到未知的incompat特性位时，拒绝mount以防止数据损坏。