Ext文件系统中的日志处理

下面从ext2_new_inode和ext3_new_inode看看日志和没有日志的区别，在没有涉及到写操作的时候是和原来一样的，下面是在要写inode位图的代码：

// 取得写入权限
err = ext3_journal_get_write_access(handle, bitmap_bh);
// 设置位图中的位
if (!ext3_set_bit_atomic(sb_bgl_lock(sbi, group),ino, bitmap_bh->b_data)) {
	// 标志元数据为脏的
	err = ext3_journal_dirty_metadata(handle,bitmap_bh);
	if (err)
		goto fail;
	goto got;
}
// 放弃写入权限
journal_release_buffer(handle, bitmap_bh);

这里要取的写入权限和平时说的文件权限显然是不一样的。我们来看ext3_journal_get_write_access(handle, bitmap_bh)的代码：

#define ext3_journal_get_write_access(handle, bh) \
	__ext3_journal_get_write_access(__FUNCTION__, (handle), (bh))

__ext3_journal_get_write_access(__FUNCTION__, (handle), (bh))的代码如下：

int __ext3_journal_get_write_access(const char *where, handle_t *handle,struct buffer_head *bh){
	int err = journal_get_write_access(handle, bh);
	if (err)
		ext3_journal_abort_handle(where, __FUNCTION__, bh, handle,err);
	return err;
}

其实到最后就是journal_get_write_access(handle, bh)：

int journal_get_write_access(handle_t *handle, struct buffer_head *bh){
	// 如果有和bh相关的jh就返回它，否则新创建一个并初始化
	struct journal_head *jh = journal_add_journal_head(bh);
	int rc;
	// 真正的执行
	rc = do_get_write_access(handle, jh, 0);
	journal_put_journal_head(jh);
	return rc;
}

可以通过buffer_buffer中b_private字段来取到所在journal_head，真正的操作在do_get_write_access(handle, jh, 0);中完成。这个函数的具体操作如下：如果bh已经是现在的事务的一部分，这时候不需要任何操作。如果bh是前一个事务（正在提交到日志），那么就让它接着写，而是拷贝出来一份，在这个上面写。

if (need_copy) {
	struct page *page;
	int offset;
	char *source;

	J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),"Possible IO failure.\n");
	page = jh2bh(jh)->b_page;
	offset = ((unsigned long) jh2bh(jh)->b_data) & ~PAGE_MASK;
	source = kmap_atomic(page, KM_USER0);
	memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
	kunmap_atomic(source, KM_USER0);
}
jbd_unlock_bh_state(bh);
jbd2_journal_cancel_revoke(handle, jh);

但是可能已经有过一份拷贝了，如果有的话就不需要再取重新拷贝了：

if (jh->b_frozen_data) {
	JBUFFER_TRACE(jh, "has frozen data");
	J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
	jh->b_next_transaction = transaction;
	goto done;
}

到这里也就大概知道write_access其实就是让JBD去准备准备。而int journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)则把一个buffer_head指明是元数据脏缓存。感觉JBD中的麻烦的地方就是处理事务，比如一个buffer_head已经在另一个事务上了，现在的写操作该怎么办？比如判断元数据是不是在正在提交的事务上面，如果是的话就不把它加到现在Metadata对应的链表中，如下：

if (jh->b_transaction != transaction) {
	JBUFFER_TRACE(jh, "already on other transaction");
	J_ASSERT_JH(jh, jh->b_transaction == journal->j_committing_transaction);
	J_ASSERT_JH(jh, jh->b_next_transaction == transaction);
	goto out_unlock_bh;
}

否则就把该bh加入到metadata对应的链表上：

spin_lock(&journal->j_list_lock);
__journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
spin_unlock(&journal->j_list_lock);

各个链表的不同之处在后面给出。通过上面的操作就把包含该元数据的buffer_head标志位脏，因此在下次事务的时候会将该bh写到磁盘上。journal_release_buffer(handle_t *handle, struct buffer_head *bh)函数用来撤销ext3_journal_get_write_access的操作，但是貌似什么都没有做，如下：

journal_release_buffer(handle_t *handle, struct buffer_head *bh){
	BUFFER_TRACE(bh, "entry");
}
#define BUFFER_TRACE(bh, info)  do {} while (0)

其他的写磁盘的操作和这个类似，可以看出，如果没有日志处理，Ext2和Ext3的操作可以说是一模一样的。关键的地方是会有一个后台线程（kjournald）定时启动提交事务，这个时间据说是5秒。下面就来看这个线程做了些什么事情：

if (journal->j_commit_sequence != journal->j_commit_request) {
	jbd_debug(1, "OK, requests differ\n");
	spin_unlock(&journal->j_state_lock);
	del_timer_sync(&journal->j_commit_timer);
	journal_commit_transaction(journal);
	spin_lock(&journal->j_state_lock);
	goto loop;
}

在发现j_commit_request与j_commit_sequence 不相同的时候就需要提交事务了。可见真正提交的工作大部分都在journal_commit_transaction中完成：

commit_transaction->t_state = T_LOCKED;
spin_lock(&commit_transaction->t_handle_lock);
while (commit_transaction->t_updates) {
	DEFINE_WAIT(wait);
	prepare_to_wait(&journal->j_wait_updates, &wait,TASK_UNINTERRUPTIBLE);
	if (commit_transaction->t_updates) {
		spin_unlock(&commit_transaction->t_handle_lock);
		spin_unlock(&journal->j_state_lock);
		schedule();
		spin_lock(&journal->j_state_lock);
		spin_lock(&commit_transaction->t_handle_lock);
	}
	finish_wait(&journal->j_wait_updates, &wait);
}
spin_unlock(&commit_transaction->t_handle_lock);

现在可能任由handle在修改，这样设置T_LOCKED，这样就不会有新的handle被加入该事务，不然就没完没了。t_updates表示有多少handle在处理该事务，下面在while中等待他们完成（journal->j_wait_updates记录了等待者）。

while (commit_transaction->t_reserved_list) {
	jh = commit_transaction->t_reserved_list;
	JBUFFER_TRACE(jh, "reserved, unused: refile");
	if (jh->b_committed_data) {
		struct buffer_head *bh = jh2bh(jh);
		jbd_lock_bh_state(bh);
		jbd_free(jh->b_committed_data, bh->b_size);
		jh->b_committed_data = NULL;
		jbd_unlock_bh_state(bh);
	}
	journal_refile_buffer(journal, jh);
}

在commit_transaction->t_reserved_list上是事务管理但是还没有改变的buffer_head，这些是没有必要提交的，从本事务中删除。

spin_lock(&journal->j_list_lock);
__journal_clean_checkpoint_list(journal);
spin_unlock(&journal->j_list_lock);

处理已经提交的事务的check_point，这样可以从日志文件中空出更多的空间。

journal_switch_revoke_table(journal);

切换revoketable，这样就可以一边提交，一边写了。

commit_transaction->t_state = T_FLUSH;
journal->j_committing_transaction = commit_transaction;
journal->j_running_transaction = NULL;
commit_transaction->t_log_start = journal->j_head;
wake_up(&journal->j_wait_transaction_locked);

这个时候就可以准备开始提交的，设置状态为T_FLUSH，然后可以唤醒等待锁定事务的等待队列。

spin_lock(&journal->j_list_lock);
if (commit_transaction->t_buffers) {
	new_jh = jh = commit_transaction->t_buffers->b_tnext;
	do {
		J_ASSERT_JH(new_jh, new_jh->b_modified == 1 ||new_jh->b_modified == 0);
		new_jh->b_modified = 0;
		new_jh = new_jh->b_tnext;
	} while (new_jh != jh);
}
spin_unlock(&journal->j_list_lock);

清除b_modified标志，这样以后对这些buffer的访问都会背安排的别的事务上面去。

journal_submit_data_buffers(journal, commit_transaction);
spin_lock(&journal->j_list_lock);
while (commit_transaction->t_locked_list) {
	struct buffer_head *bh;

	jh = commit_transaction->t_locked_list->b_tprev;
	bh = jh2bh(jh);
	get_bh(bh);
	if (buffer_locked(bh)) {
		spin_unlock(&journal->j_list_lock);
		wait_on_buffer(bh);
		if (unlikely(!buffer_uptodate(bh)))
			err = -EIO;
		spin_lock(&journal->j_list_lock);
	}
	if (!inverted_lock(journal, bh)) {
		put_bh(bh);
		spin_lock(&journal->j_list_lock);
		continue;
	}
	if (buffer_jbd(bh) && jh->b_jlist == BJ_Locked) {
		__journal_unfile_buffer(jh);
		jbd_unlock_bh_state(bh);
		journal_remove_journal_head(bh);
		put_bh(bh);
	} else {
		jbd_unlock_bh_state(bh);
	}
	put_bh(bh);
	cond_resched_lock(&journal->j_list_lock);
}
spin_unlock(&journal->j_list_lock);

将BJ_SyncData队列上面的数据写入磁盘，因为日志只是记录对元数据的修改，元数据写入之前要成功写入普通数据，这样其他日志中记录的元数据对应的普通数据也在磁盘上的效果。while等待所有的写入完成。

journal_write_revoke_records(journal, commit_transaction);

写入取消块。接下来就要写入元数据了，这里比较麻烦：

err = journal_next_log_block(journal, &blocknr);

取到往日志的那个位置写入。

if (!descriptor) {
	struct buffer_head *bh;
	J_ASSERT (bufs == 0);
	jbd_debug(4, "JBD: get descriptor\n");
	descriptor = journal_get_descriptor_buffer(journal);
	if (!descriptor) {
		journal_abort(journal, -EIO);
		continue;
	}
	bh = jh2bh(descriptor);
	jbd_debug(4, "JBD: got buffer %llu (%p)\n",(unsigned long long)bh->b_blocknr, bh->b_data);
	header = (journal_header_t *)&bh->b_data[0];
	header->h_magic     = cpu_to_be32(JFS_MAGIC_NUMBER);
	header->h_blocktype = cpu_to_be32(JFS_DESCRIPTOR_BLOCK);
	header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
	tagp = &bh->b_data[sizeof(journal_header_t)];
	space_left = bh->b_size - sizeof(journal_header_t);
	first_tag = 1;
	set_buffer_jwrite(bh);
	set_buffer_dirty(bh);
	wbuf[bufs++] = bh;
	BUFFER_TRACE(bh, "ph3: file as descriptor");
	journal_file_buffer(descriptor, commit_transaction,BJ_LogCtl);
}

上面的这段代码初始化了一个描述符块。首先通过journal_get_descriptor_buffer从日志文件中分配一块存放描述符块，然后设置描述符块的header，接下来计算得到journal_block_tag_t的开始位置，把这个缓冲块加到wbuf（记录所有要写到日志的缓存区）中。将描述符块的缓存区加入到BJ_LogCtl列表中。

flags = journal_write_metadata_buffer(commit_transaction, jh, &new_jh, blocknr);
set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
wbuf[bufs++] = jh2bh(new_jh);

journal_write_metadata_buffer函数并没有执行写入操作，而是根据已存在的元数据缓存区创建一个新的缓存区，在JBD写入日志的时候使用。这时jh会从BJ_Metadata队列移动到BJ_Shadow队列，而new_jh则移动到BJ_IO队列。把new_jh也加入到wbuf中。

tag = (journal_block_tag_t *) tagp;
tag->t_blocknr = cpu_to_be32(jh2bh(jh)->b_blocknr);
tag->t_flags = cpu_to_be32(tag_flag);
tagp += sizeof(journal_block_tag_t);
space_left -= sizeof(journal_block_tag_t);

if (first_tag) {
	memcpy (tagp, journal->j_uuid, 16);
	tagp += 16;
	space_left -= 16;
	first_tag = 0;
}

把jh记录进去，如果是第一块的话还有把UUID拷贝进去，当然后面的就不需要了。

if (bufs == journal->j_wbufsize || commit_transaction->t_buffers == NULL || space_left < sizeof(journal_block_tag_t) + 16) {
	jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
	tag->t_flags |= cpu_to_be32(JFS_FLAG_LAST_TAG);
start_journal_io:
	for (i = 0; i < bufs; i++) {
		struct buffer_head *bh = wbuf[i];
		lock_buffer(bh);
		clear_buffer_dirty(bh);
		set_buffer_uptodate(bh);
		bh->b_end_io = journal_end_buffer_io_sync;
		submit_bh(WRITE, bh);
	}
	cond_resched();
	descriptor = NULL;
	bufs = 0;
}

如果达到日志一次写入缓存区个数，或者BJ_Metadata已经空了，或者描述块已经放满了，现在就可以开始提交了。descriptor = NULL;可以使得下次循环的时候用新的内存。

while (commit_transaction->t_iobuf_list != NULL) {
	struct buffer_head *bh;
	jh = commit_transaction->t_iobuf_list->b_tprev;
	bh = jh2bh(jh);
	if (buffer_locked(bh)) {
		wait_on_buffer(bh);
		goto wait_for_iobuf;
	}
	if (cond_resched())
		goto wait_for_iobuf;
	if (unlikely(!buffer_uptodate(bh)))
		err = -EIO;
	clear_buffer_jwrite(bh);
	JBUFFER_TRACE(jh, "ph4: unfile after journal write");
	journal_unfile_buffer(journal, jh);
	BUFFER_TRACE(bh, "dumping temporary bh");
	journal_put_journal_head(jh);
	__brelse(bh);
	J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
	free_buffer_head(bh);
	jh = commit_transaction->t_shadow_list->b_tprev;
	bh = jh2bh(jh);
	clear_bit(BH_JWrite, &bh->b_state);
	J_ASSERT_BH(bh, buffer_jbddirty(bh));
	JBUFFER_TRACE(jh, "file as BJ_Forget");
	journal_file_buffer(jh, commit_transaction, BJ_Forget);
	wake_up_bit(&bh->b_state, BH_Unshadow);
	JBUFFER_TRACE(jh, "brelse shadowed buffer");
	__brelse(bh);
}

循环等待BJ_IO写入的完成。BJ_IO上的数据和BJ_Shadow是一一对应的，没处理完一个JB_IO上的缓存区，就将其从事务中删除，并将BJ_Shadow上的与其对应的项移动到BJ_Forgot队列上。

while (commit_transaction->t_log_list != NULL) {
	struct buffer_head *bh;
	jh = commit_transaction->t_log_list->b_tprev;
	bh = jh2bh(jh);
	if (buffer_locked(bh)) {
		wait_on_buffer(bh);
		goto wait_for_ctlbuf;
	}
	if (cond_resched())
		goto wait_for_ctlbuf;
	if (unlikely(!buffer_uptodate(bh)))
		err = -EIO;
	BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
	clear_buffer_jwrite(bh);
	journal_unfile_buffer(journal, jh);
	journal_put_journal_head(jh);
	__brelse(bh);
}

等待所有的BJ_LogCtl队列上的项写入日志中，写入成功后将其从事务上删除。

if (journal_write_commit_record(journal, commit_transaction))
	err = -EIO;

写入一个提交块。

J_ASSERT(commit_transaction->t_sync_datalist == NULL);
J_ASSERT(commit_transaction->t_buffers == NULL);
J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
J_ASSERT(commit_transaction->t_iobuf_list == NULL);
J_ASSERT(commit_transaction->t_shadow_list == NULL);
J_ASSERT(commit_transaction->t_log_list == NULL);

如果提交成功了，那上面这几个队列都应该是空的。

while (commit_transaction->t_forget) {
	transaction_t *cp_transaction;
	struct buffer_head *bh;
	jh = commit_transaction->t_forget;
	spin_unlock(&journal->j_list_lock);
	bh = jh2bh(jh);
	jbd_lock_bh_state(bh);
	J_ASSERT_JH(jh, jh->b_transaction == commit_transaction || jh->b_transaction == journal->j_running_transaction);
	if (jh->b_committed_data) {
		jbd_free(jh->b_committed_data, bh->b_size);
		jh->b_committed_data = NULL;
		if (jh->b_frozen_data) {
			jh->b_committed_data = jh->b_frozen_data;
			jh->b_frozen_data = NULL;
		}
	} else if (jh->b_frozen_data) {
		jbd_free(jh->b_frozen_data, bh->b_size);
		jh->b_frozen_data = NULL;
	}
	spin_lock(&journal->j_list_lock);
	cp_transaction = jh->b_cp_transaction;
	if (cp_transaction) {
		JBUFFER_TRACE(jh, "remove from old cp transaction");
		__journal_remove_checkpoint(jh);
	}
	if (buffer_freed(bh)) {
		clear_buffer_freed(bh);
		clear_buffer_jbddirty(bh);
	}
	if (buffer_jbddirty(bh)) {
		JBUFFER_TRACE(jh, "add to new checkpointing trans");
		__journal_insert_checkpoint(jh, commit_transaction);
		JBUFFER_TRACE(jh, "refile for checkpoint writeback");
		__journal_refile_buffer(jh);
		jbd_unlock_bh_state(bh);
	} else {
		J_ASSERT_BH(bh, !buffer_dirty(bh));
		JBUFFER_TRACE(jh, "refile or unfile freed buffer");
		__journal_refile_buffer(jh);
		if (!jh->b_transaction) {
			jbd_unlock_bh_state(bh);
			journal_remove_journal_head(bh);
			release_buffer_page(bh);
		} else
			jbd_unlock_bh_state(bh);
	}
	cond_resched_lock(&journal->j_list_lock);
}

上面这段代码实现checkpoint机制。checkpoint机制是针对元数据的，如果数据不仅仅是写到日记中，而是已经写到磁盘上准确的位置了，那么直接从事务中删除就好了。否则将它加入本事务的checkpoint队列。有没有写上去可以通过查看buffer_head是不是脏的来判断。

if (commit_transaction->t_forget) {
	spin_unlock(&journal->j_list_lock);
	spin_unlock(&journal->j_state_lock);
	goto restart_loop;
}

如果在加锁的过程中BJ_Forgot中又有了新的数据，那么跳回上面重新处理checkpoint。

commit_transaction->t_state = T_FINISHED;
J_ASSERT(commit_transaction == journal->j_committing_transaction);
journal->j_commit_sequence = commit_transaction->t_tid;
journal->j_committing_transaction = NULL;
spin_unlock(&journal->j_state_lock);

if (commit_transaction->t_checkpoint_list == NULL &&
	commit_transaction->t_checkpoint_io_list == NULL) {
	__journal_drop_transaction(journal, commit_transaction);
} else {
	if (journal->j_checkpoint_transactions == NULL) {
		journal->j_checkpoint_transactions = commit_transaction;
		commit_transaction->t_cpnext = commit_transaction;
		commit_transaction->t_cpprev = commit_transaction;
	} else {
		commit_transaction->t_cpnext = journal->j_checkpoint_transactions;
		commit_transaction->t_cpprev = commit_transaction->t_cpnext->t_cpprev;
		commit_transaction->t_cpnext->t_cpprev = commit_transaction;
		commit_transaction->t_cpprev->t_cpnext = commit_transaction;
	}
}
spin_unlock(&journal->j_list_lock);
wake_up(&journal->j_wait_done_commit);

到这里，事务的提交已经完成，将该事务加入到journal->j_checkpoint_transaction队列中。在JBD中为什么要将一个个的handle合并成一个事务呢？事务的存在就是为了保证一致性，所有在完成的时候必须将该事务涉及的buffer写入磁盘，但是每个进程修改的文件的元数据的量是非常少的，如果没进行一次修改就写入磁盘这样的效率当然不高了。JBD为保证在开启后台线程的时候是的磁盘的I/O急剧增大也设置了每次事务提交数据的大小。

在事务处理中需要想清楚的地方就是事务与事务之间的关系，比如一个buffer_head同时在两个事务上的时候。还有就是在处理事务之前先处理checkpoint的不仅仅是为了省出空间，也是为了日志系统的争取运行。

-------------------------------

个人理解，欢迎拍砖。