/** Mini-transaction handle and buffer */ struct mtr_t { /** State variables of the mtr */ struct Impl { // mtr 持有锁的栈 /** memo stack for locks etc. */ mtr_buf_t m_memo; // mtr产生的日志 /** mini-transaction log */ mtr_buf_t m_log; // 是否产生buffer pool脏页 /** true if mtr has made at least one buffer pool page dirty */ bool m_made_dirty; // insert buffer 是否修改 /** true if inside ibuf changes */ bool m_inside_ibuf; // 是否修改buffer pool pages /** true if the mini-transaction modified buffer pool pages */ bool m_modifications; /** true if mtr is forced to NO_LOG mode because redo logging is disabled globally. In this case, mtr increments the global counter at ::start and must decrement it back at ::commit. */ bool m_marked_nolog; /** Shard index used for incrementing global counter at ::start. We need to use the same shard while decrementing counter at ::commit. */ size_t m_shard_index; // 一个MTR已产生log记录个数 /** Count of how many page initial log records have been written to the mtr log */ ib_uint32_t m_n_log_recs; /* MTR的日志模式包括 MTR_LOG_ALL：默认模式，记录所有会修改磁盘数据的操作； MTR_LOG_NONE：不记录redo，脏页也不放到flush list上； MTR_LOG_NO_REDO：不记录redo，但脏页放到flush list上，如临时表的修改； MTR_LOG_SHORT_INSERTS：插入记录操作REDO，在将记录从一个page拷贝到 另外一个新建的page时用到，此时忽略写索引信息到redo log中。 */ // 参考函数page_cur_insert_rec_write_log /** specifies which operations should be logged; default value MTR_LOG_ALL */ mtr_log_t m_log_mode; /* mtr状态: MTR_STATE_INIT = 0, TR_STATE_ACTIVE = 12231, MTR_STATE_COMMITTING = 56456, MTR_STATE_COMMITTED = 34676 */ /** State of the transaction */ mtr_state_t m_state; /** Flush Observer */ FlushObserver *m_flush_observer; private: Impl m_impl; /** LSN at commit time */ lsn_t m_commit_lsn; /** true if it is synchronous mini-transaction */ bool m_sync; class Command; friend class Command; };

mtr_memo_slot_t *slot; // 先在动态buffer中申请能容纳锁类型+地址的空间，再对该空间进行初始化 slot = m_impl.m_memo.push<mtr_memo_slot_t *>(sizeof(*slot)); // 锁类型 slot->type = type; // 锁地址或page地址 slot->object = object;

// 预分配待写入的日志空间，若空间不够，则增加新的buffer到动态buffer中。 // 预分配待写入的日志空间，若空间不够，则增加新的buffer到动态buffer中。 byte *mlog_open(mtr_t *mtr, ulint size) do_work(MTR & Btree) --> 对index加锁、修改page // 写入日志类型、space id、page no，且m_n_log_recs加1 mlog_write_initial_log_record_fast() 记录其他相关信息 // 获取mtr buffer（双向链表）的最后一个节点，更新mtr日志大小 mlog_close(mtr, log_ptr) // 将mtr的数据move到mtr buffer(m_log)中 mtr->get_log()->push()

/** Start a mini-transaction. */ #define mtr_start(m) (m)->start()

/** Start a mini-transaction. @param sync true if it is a synchronous mini-transaction @param read_only true if read only mini-transaction */ void mtr_t::start(bool sync, bool read_only) { m_sync = sync; m_commit_lsn = 0; new (&m_impl.m_log) mtr_buf_t(); new (&m_impl.m_memo) mtr_buf_t(); m_impl.m_mtr = this; m_impl.m_log_mode = MTR_LOG_ALL; m_impl.m_inside_ibuf = false; m_impl.m_modifications = false; m_impl.m_made_dirty = false; m_impl.m_n_log_recs = 0; m_impl.m_state = MTR_STATE_ACTIVE; m_impl.m_flush_observer = nullptr; m_impl.m_marked_nolog = false; }

/** Redo log - single data structure with state of the redo log system. In future, one could consider splitting this to multiple data structures. */ struct alignas(ut::INNODB_CACHE_LINE_SIZE) log_t { //通常用这个log.sn ,通过log_get_lsn去换算获得当前的lsn atomic_sn_t sn; // log buffer的内存区 aligned_array_pointer<byte, OS_FILE_LOG_BLOCK_SIZE> buf; // 解决并发插入Redo Log Buffer后刷入ib_logfile存在空洞的问题 Link_buf<lsn_t> recent_written; // 解决并发插入flush_list后确认checkpoint_lsn的问题 Link_buf<lsn_t> recent_closed; // write_lsn之前的数据已经写入系统的Cache, 但不保证已经Flush atomic_lsn_t write_lsn; // 已经被flush到磁盘的数据，由log flusher thread 来更新 // log.flushed_to_disk_lsn <= log.write_lsn atomic_lsn_t flushed_to_disk_lsn; // log buffer缓冲区的大小 size_t buf_size; /* 表示可以进行checkpoint的最大的lsn 到这个lsn 为止, 所有的redo log 对应的dirty page 已经flush 到btree 上了， because flush 的时候并不是顺序的flush, 所以有可能存在有空洞的情况, 因此这个lsn 的位置并不是最大的redo log 已经被flush 到btree 的位置. 而是可以作为checkpoint 的最大的位置. 这个值是由log checkpointer thread 来更新 */ lsn_t available_for_checkpoint_lsn; /* 在此lsn之前的所有被添加到buffer pool的flush list的log数据已经被flsuh, 下一次checkpoint可以make在这个lsn. 与last_checkpoint_lsn的区别是 该lsn尚未被真正的checkpoint. */ // 下次需要进行checkpoint的lsn lsn_t requested_checkpoint_lsn; // 目前最新的checkpoint的lsn /* 到这个lsn 为止, 所有的btree dirty page 已经flushed 到disk了, 并且这个lsn 值已经被更新到了ib_logfile0 这个文件去了. 这个lsn 也是下一次recovery 的时候开始的地方, 因为last_checkpoint_lsn 之前的redo log 已经保证都flush 到btree 中去了. 所以比这个lsn 小的redo log 文件已经可以删除了, 因为数据已经都flush 到btree data page 中去了. 这个值是由log checkpointer thread 来更新 log.last_checkpoint_lsn <= log.available_for_checkpoint_lsn <= log.buf_dirty_pages_added_up_to_lsn */ atomic_lsn_t last_checkpoint_lsn; // write ahead 的Buffer大小 uint32_t write_ahead_buf_size; // 当前正在fsync到的LSN,不断增加并且不回环的，它是redo log实际内容在逻辑上的增长 lsn_t current_file_lsn; //代表在所有redo 物理文件的偏移 uint64_t current_file_real_offset; // 当前ib_logfile文件末尾的offset uint64_t current_file_end_offset; // 当前ib_logfile的文件大小 uint64_t file_size; //表示2个文件实际大小总和 file_size * n_files uint64_t files_real_capacity ...... }

mtr_t::commit | |-> mtr_t::Command | (m_n_log_recs>0 || m_modifications) | |-> (yes) | v // 将mtr.m_impl->m_log写入公共log buffer，把脏页加入flush list | Command::execute | | | |-> prepare_write | | | | | |-> 若 mtr.m_impl->m_log_mode为 MTR_LOG_NO_REDO或MTR_LOG_NONE， | | | 则直接返回 | | | | | |-> 若 mtr.m_impl->m_n_log_recs==1，则 | | m_log.front()->begin()|=MLOG_SINGLE_REC_FLAG，在日志头Type字段 | | 中标识，否则 m_log->push(MLOG_MULTI_REC_END)，在日志结尾附加1B | | | |-> /* 在公共log buffer中为日志预留空, 不同的mtr会首先调用 | | log_buffer_reserve函数获得起始和终点lsn位置 */ | |-> log_buffer_reserve | | | | | |-> log_buffer_s_lock_enter_reserve | | | | | | | |-> 对 log.pfs_psi加 s-lock | | | | | | | |-> /* 在公共的log buffer中占位, 用自己的REDO长度，原子地对 | | | | 全局偏移log.sn做fetch_add，得到自己在Log Buffer中独享的空间*/ | | | |-> log.sn.fetch_add(mtr.m_impl->m_log.m_size) | | | | | | | |-> /* 若log.sn被SN_LOCKED，则等待log.sn_locked, | | | | 直到log.sn unlock 且超过占位前的log.sn */ | | | |-> log_buffer_s_lock_wait | | | | | |-> /* 将日志内容的偏移量log.sn 转为log block格式的偏移量start_lsn， | | | start_lsn可以唯一表示日志在log block和公共log buffer中的位置*/ | | |-> log_translate_sn_to_lsn | | | | | |-> // 不同mtr并行的将自己的m_log中的数据拷贝到各自独享的空间内。 | | |-> // 若end_sn > log.buf_limit_sn，则等待 | | |-> log_wait_for_space_after_reserving | | | |-> /* 确保这条 redo log 能完整的写入 redo log Buffer， | | | |-> 而且回环后不会覆盖尚未写入磁盘的redo log. */ | | | |-> log_wait_for_space_in_log_buf | | | | |-> log_write_up_to | | | | | |-> log_wait_for_write | | | | | | |-> //log_write_notify会唤醒这个条件变量 | | | | | | |-> os_event_wait_for(log.write_events[slot], | | | | | | | stop_condition); | | | | | |-> log_wait_for_flush | | | |-> // 从mtr的buffer中内容写入到redolog的buffer. | |-> m_impl->m_log.for_each_block(write_log) | | |-> mtr_write_log_t::operator | | | |-> //写入到redo log buffer(log->buf). | | | |-> log_buffer_write | | | |-> //拷贝完成后触发LinkBuf更新,更新recent_written字段 | | | |-> log_buffer_write_completed | | | | |-> //更新本次写入的内容范围对应的LinkBuf内特定的数组项值 | | | | |-> log.recent_written.add_link | |-> // 在加脏页之前需要判断是否link buf已满。 | |-> log_wait_for_space_in_log_recent_closed | | | |-> // 将mtr锁管理中记录的脏页加入flush list | |-> add_dirty_blocks_to_flush_list(mtr.m_impl->m_memo) | | | | | (reverse loop mtr_buf_t::block in m_memo) | | v | | (reverse loop mtr_memo_slot_t in block) | | | | | |-> /* 为了去掉flush_order_mutex，把mtr对应的脏页无序的添加到flush list， | | | 在做checkpoint时, 无法保证flush list 上面最头的page lsn是最小的*/ | | |-> add_to_flush | | | | | | | |-> /* 把修改后的page加入flush list，当mtr_memo_slot_t.type | | | | 为MTR_MEMO_PAGE_X_FIX或MTR_MEMO_PAGE_SX_FIX， | | | | 或为MTR_MEMO_BUF_FIX， | | | | 且mtr_memo_slot_t.object->made_dirty_with_no_latch*/ | | | |-> add_dirty_page_to_flush_list | | | | | | | | | | | | -> buf_flush_note_modification(mtr_memo_slot_t.object) | | | |-> /* 将mtr锁管理中记录的脏页处理完后触发一次Link_buf更新， | | log.recent_closed记录添加到flush list的最大连续日志的lsn, | | 以log.recent_closed.m_tail的lsn来做checkpoint肯定是安全的 */ | |-> log_buffer_close | | | | | | | | |->log_buffer_s_lock_exit_close | | | | | | | |-> 对 log.pfs_psi解锁 s-lock | | | | | | | |-> log.recent_closed.add_link_advance_tail | |-> Command::release_all | | | |-> Release_all(mtr.m_impl->m_memo) 释放mtr持有的锁 | |-> Command::release_resources -> clean mtr.m_impl->m_log & m_memo

/** The recent written buffer. Protected by: locking sn not to add. */ Link_buf<lsn_t> recent_written;

|--> log_writer | |--> /*endless loop*/ | |--> //Innodb将redo log buffer内容写入日志文件时需要保证不能存在空洞， | |--> //即在写入前需要获得当前最大的无空洞lsn。这依赖LinkBuf。 | |--> //在后台写日志线程log_writer的log_advance_ready_for_write_lsn函数中完成。 | |--> log_advance_ready_for_write_lsn | | |--> //获取之前的tail值，这里仅为验证作用 | | |--> log_buffer_ready_for_write_lsn | | |--> //推进Link_buf::m_tail，同时回收之前空间 | | |--> log.recent_written.advance_tail_until(stop_condition) | |--> log_buffer_ready_for_write_lsn | |--> //Do the actual work | |--> log_writer_write_buffer | | |--> /* Do the write to the log files */ | | |--> log_files_write_buffer | | | |--> notify_about_advanced_write_lsn | | | | |--> //去唤醒write_notifier_event | | | | |--> os_event_set(log.write_notifier_event); | | | |--> log_update_buf_limit

log_write_notifier { /*endless loop*/ while(1) { sleep(); if (log.write_lsn.load() >= lsn) { while (lsn <= notified_up_to_lsn) { const auto slot = log_compute_write_event_slot(log, lsn); lsn += OS_FILE_LOG_BLOCK_SIZE; //唤醒在等待的用户线程 os_event_set(log.write_events[slot]); } lsn = write_lsn + 1; } } }

os_event_wait_time_low(log.flusher_event, flush_every_us - time_elapsed_us, 0);

|--> log_flusher | |--> /*endless loop*/ | |--> if (srv_flush_log_at_trx_commit != 1) | |--> os_event_wait_time_low(log.flusher_event) | |--> endif | |--> if (last_flush_lsn < log.write_lsn.load()) | |--> log_flush_low | |--> endif

log_flush_notifier { /*endless loop*/ while(1) { if (log.flushed_to_disk_lsn.load() >= lsn) { while (lsn <= notified_up_to_lsn) { const auto slot = log_compute_flush_event_slot(log, lsn); lsn += OS_FILE_LOG_BLOCK_SIZE; //唤醒在等待的log_flusher线程 os_event_set(log.flush_events[slot]); } lsn = flush_lsn + 1; } } }

/* LWM lsn for unflushed dirty pages in Buffer Pool / lsn_t lwm_lsn = buf_pool_get_oldest_modification_lwm(); / Note lsn up to which all dirty pages have already been added into Buffer Pool */ const lsn_t dpa_lsn = log_buffer_dirty_pages_added_up_to_lsn(log); lsn_t checkpoint_lsn = std::min(lwm_lsn, dpa_lsn);

/* LWM lsn for unflushed dirty pages in Buffer Pool / const lsn_t lsn = buf_pool_get_oldest_modification_approx(); const lsn_t lag = log.recent_closed.capacity(); lsn_t lwm_lsn = lsn - lag; / Note lsn up to which all dirty pages have already been added into Buffer Pool */ const lsn_t dpa_lsn = log_buffer_dirty_pages_added_up_to_lsn(log); lsn_t checkpoint_lsn = std::min(lwm_lsn, dpa_lsn);

|--> log_checkpointer | |--> /*endless loop*/ | |--> log_update_available_for_checkpoint_lsn(log) | | |--> log_compute_available_for_checkpoint_lsn(log) | | | |--> //当前连续加入Buffer Pool脏页的最大LSN: dpa_lsn | | | |--> dpa_lsn = log_buffer_dirty_pages_added_up_to_lsn(log) | | | |--> //获得当前所有脏页对应的最小REDO LSN：lwm_lsn | | | |--> lwm_lsn = buf_pool_get_oldest_modification_lwm(); | | | |--> flushed_lsn = log.flushed_to_disk_lsn.load(); | | | |--> lsn = std::min(dpa_lsn, lwm_lsn, flushed_lsn) | |--> /* Consider flushing some dirty pages. */ | |--> sync_flushed = log_consider_sync_flush(log); | | |--> log_request_sync_flush | | | |--> buf_flush_request_force | | | | |--> //buf_flush_sync_lsn作为page cleaner线程同步刷脏的依据 | | | | |--> buf_flush_sync_lsn = lsn_target; | | | | |--> os_event_set(buf_flush_event); | | |--> log_update_available_for_checkpoint_lsn | |--> /* Consider writing checkpoint. | |--> 把checkpoint 信息写入到ib_logfile0 文件中*/ | |--> checkpointed = log_consider_checkpoint(log) | | |--> log_checkpoint | | | |--> log_files_write_checkpoint(log, checkpoint_lsn); | | | | |--> fil_redo_io