log file parallel write概念介绍

log file parallel write 事件是LGWR进程专属的等待事件,发生在LGWR将日志缓冲区(log_buffer)中的重做日志信息写入联机重做日志文件组的成员文件,LGWR在该事件上等待该写入过程的完成。该事件的等待表示重做日志所处的磁盘设备缓慢或存在争用。下面看看官方一些资料是如何解释log file parallel write等待事件的。

log file parallel write

 

Writing redo records to the redo log files from the log buffer.

Wait Time: Time it takes for the I/Os to complete. Even though redo records are written in parallel, the parallel write is not complete until the last I/O is on disk.

Parameters:

·         P1 = files

·         P2 = blocks

·         P3 = requests

·         files Number of files written to.

If you have more than one log member per group then the files are written to in parallel (if possible). This is the number of redo log members (files) that the writes apply to.

·         blocks Number of blocks to be written.

The number of REDO blocks being written to each log member. eg: A value of 10 indicates 10 redo blocks for each log member.

·         requests Number of I/O requests.

The number of distinct IO requests. The number of blocks to be written is divided up into a number of IO requests.

Wait Time:

The actual wait time is the time taken for all the outstanding I/O requests to complete. Even though the writes may be issued in parallel, LGWR needs to wait for the last I/O to be on disk before the parallel write is considered complete. Hence the wait time depends on the time it takes the OS to complete all requests.

注意:自己看到这个等待事件parallel write,以及网上有“LGWR进程对同一组多个重做日志文件‘同时’写,是通过异步I/O来实现的,因此等待事件log file parallel write应该是在同一组下有多个重做日志文件时才会出现 ”这种说法,其实不管重做日志是一个或多个成员,都会出现log file parallel write等待事件。只要有buffer log写到redo log里面就有这个等待,它不是说并行写多个redo文件,单个文件也是这样的。

log file parallel write出现原因

 

如果数据库出现了这个等待事件,意味着重做日志(redo log)所处的磁盘设备I/O缓慢或存在争用

· 磁盘I/O性能比较差

· REDO文件的分布导致了I/O争用,例如,同一组的REDO成员文件放在相同的磁盘上。

查看log file parallel write等待事件

SQL> select s.event            as event_name

  2        ,s.time_waited/100  as time_waited

  3        ,s.average_wait     as averge_wait

  4  from v$system_event s

  5  where s.event in ('log file parallel write','log file sync');

 

EVENT_NAME                                                       TIME_WAITED AVERGE_WAIT

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

log file parallel write                                             45650.75         .03

log file sync                                                        54045.4         .04

 

SQL> 

上面time_waited、average_wait的单位为百分之一秒(in hundredths of a second), 'log file parallel write'事件的平均等待时间大于 10ms (or 1cs),这通常意味着存在较慢的I/O吞吐量。 LGWR进程写入慢会影响用户事务提交的时间。

If the log file parallel write average wait time is greater than 10ms (or 1cs), this normally indicates slow I/O throughput.The cure is the same as for the db file parallel write waits. Enable asynchronous writes if your redo logs are on raw devices and the operating system supports asynchronous I/O. For redo logs on file systems, use synchronous direct writes.

其它一些查看redo log相关信息的SQL

SELECT b.name, 

       a.value, 

       Round(SYSDATE - c.startup_time) days_old 

FROM   v$sysstat a, 

       v$statname b, 

       v$instance c 

WHERE  a.statistic# = b.statistic# 

       AND b.name IN ( 'redo wastage', 'redo size' ); 

 

 

SELECT sid, 

       value 

FROM   v$sesstat 

WHERE  statistic# = (SELECT statistic# 

                     FROM   v$statname 

                     WHERE  name = 'user commits') 

ORDER  BY value DESC; 

 

如何减少log file parallel write等待时间

要减少log file parallerl write等待时间,可以从下面两个方面入手解决(其它细节参考下面英文资料):

(1)将重做日志文件组放置到高速I/O磁盘上(不要将日志组成员文件放置IO不一样的磁盘上)。不要将重做日志放在RAID5的磁盘上。

(2)尽可能的降低重做数量:

1:尽可能使用nologging选项,例如索引的重建、重组。包括create table...as select...等操作

2:热备可能生成大量的重做信息,所以热备份应该在非高峰时间运行。

3:应用程序的commit的频率小一些也能有一些改善。

Reducing Waits / Wait times:

You might want to reduce "log file parallel write" wait times in order to reduce user waits which depend on LGWR.

· Ensure tablespaces are NOT left in HOT BACKUP mode longer than needed.

· Tablespaces in HOT BACKUP mode cause more redo to be generated for each change which can vastly increase the rate of redo generarion.

· NOLOGGING / UNRECOVERABLE operations may be possible for certain operations to reduce the overall rate of redo generation

· Redo log members should ideally be on high speed disks

· eg: Historically RAID5 was not a good candidate for redo log members, although some of the RAID 5 drawbacks have been addressed more recently through larger intelligent disk caches, etc.

· Redo log members should be on disks with little/no IO activity from other sources.

· (including low activity from other sources against the same disk controller)

· With file systems like ASM and ZFS, you may also have to experiment with the disk stripe size to make more friendly with the type of IO LGWR usually performs.

· Check that the system has not become CPU (or memory) bound otherwise it may take LGWR more time to recognize IO has been completed.

· The higher the value for CPU_COUNT, the more strands can be used for writing redo. Having an increased number of strands can lead to an increase the average waits on "log file parallel write" (which in turn can affect "log file sync" waits). Filesystems that do not support Concurrent IO (CIO) are more likely to be affected by this, compared to ASM or raw devices. However, it can also impact filesystems that support CIO. On servers with CPU_COUNT greater than 128 LGWR behaviour may need manually adjusting through underscore parameters. If you think you are encountering long "log file parallel write" due to high CPU counts then contact Oracle Support.

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

Unfortunately, you cannot spawn more than one LGWR process. In this case, it is critical that nothing else is sharing the mount point of the redo log files. Make sure the controller that services the mount point is not overloaded. Moving the redo logs to higher speed disks will also help.

We strongly suggest that you avoid putting redo logs on RAID5 disks, but we also understand that many times you don’t have a choice or a say in it.

Besides improving the I/O throughput, you can also work on lowering the amount of redo entries. This will provide some relief, but not the cure. Whenever possible, use the NOLOGGING option. Indexes should be created or rebuilt with the NOLOGGING option. CTAS operations should also use this option.

Note:

The NOLOGGING option doesn’t apply to normal DML operations such as inserts, updates, and deletes. Objects created with the NOLOGGING option are unrecoverable unless a backup is taken prior to the corruption. If you have to take an additional backup, then the I/Os that you save by not logging will be spent on backup.        Database in FORCE LOGGING mode will log all changes (except for changes in temporary tablespaces), regardless of the tablespace and object settings.

A lower commit frequency at the expense of higher rollback segment usage can also provide some relief.

A high commit frequency causes the LGWR process to be overactive and when coupled with slow I/O throughput will only magnify the log file parallel write waits.

The application may be processing a large set of data in a loop and committing each change, which causes the log buffer to be flushed too frequently. In this case, modify the application to commit at a lower frequency. There could also be many short sessions that log in to the database, perform. a quick DML operation, and log out.

下面提供一个用户做log file sync waits、log file parallel write troubleshoot的一个脚本。官方文档ID为1064487.1

NAME: LFSDIAG.SQL

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

-- AUTHOR: Michael Polaski - Oracle Support Services

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

-- PURPOSE:

-- This script is intended to provide a user friendly guide to troubleshoot

-- log file sync waits. The script will look at important parameters involved

-- in log file sync waits, log file sync wait histogram data, and the script

-- will look at the worst average log file sync times in the active session

-- history data and AWR data and dump information to help determine why those

-- times were the highest. The script will create a file called

-- lfsdiag_<timestamp>.out in your local directory.

 

set echo off

set feedback off

column timecol new_value timestamp

column spool_extension new_value suffix

select to_char(sysdate,'Mondd_hh24mi') timecol,

'.out' spool_extension from sys.dual;

column output new_value dbname

select value || '_' output

from v$parameter where name = 'db_name';

spool lfsdiag_&&dbname&&timestamp&&suffix

set trim on

set trims on

set lines 140

set pages 100

set verify off

alter session set optimizer_features_enable = '10.2.0.4';

 

PROMPT LFSDIAG DATA FOR &&dbname&&timestamp

PROMPT Note: All timings are in milliseconds (1000 milliseconds = 1 second)

 

PROMPT

PROMPT IMPORTANT PARAMETERS RELATING TO LOG FILE SYNC WAITS:

column name format a40 wra

column value format a40 wra

select inst_id, name, value from gv$parameter

where ((value is not null and name like '%log_archive%') or

name like '%commit%' or name like '%event=%' or name like '%lgwr%')

and name not in (select name from gv$parameter where (name like '%log_archive_dest_state%'

and value = 'enable') or name = 'log_archive_format')

order by 1,2,3;

 

PROMPT

PROMPT ASH THRESHOLD...

PROMPT

PROMPT This will be the threshold in milliseconds for average log file sync

PROMPT times. This will be used for the next queries to look for the worst

PROMPT 'log file sync' minutes. Any minutes that have an average log file

PROMPT sync time greater than the threshold will be analyzed further.

column threshold_in_ms new_value threshold format 999999999.999

select min(threshold_in_ms) threshold_in_ms

from (select inst_id, to_char(sample_time,'Mondd_hh24mi') minute,

avg(time_waited)/1000 threshold_in_ms

from gv$active_session_history

where event = 'log file sync'

group by inst_id,to_char(sample_time,'Mondd_hh24mi')

order by 3 desc)

where rownum <= 10;

 

PROMPT

PROMPT ASH WORST MINUTES FOR LOG FILE SYNC WAITS:

PROMPT

PROMPT APPROACH: These are the minutes where the avg log file sync time

PROMPT was the highest (in milliseconds).

column minute format a12 tru

column event format a30 tru

column program format a40 tru

column total_wait_time format 999999999999.999

column avg_time_waited format 999999999999.999

select to_char(sample_time,'Mondd_hh24mi') minute, inst_id, event,

sum(time_waited)/1000 TOTAL_WAIT_TIME , count(*) WAITS,

avg(time_waited)/1000 AVG_TIME_WAITED

from gv$active_session_history

where event = 'log file sync'

group by to_char(sample_time,'Mondd_hh24mi'), inst_id, event

having avg(time_waited)/1000 > &&threshold

order by 1,2;

 

PROMPT

PROMPT ASH LFS BACKGROUND PROCESS WAITS DURING WORST MINUTES:

PROMPT

PROMPT APPROACH: What is LGWR doing when 'log file sync' waits

PROMPT are happening? LMS info may be relevent for broadcast

PROMPT on commit and LNS data may be relevant for dataguard.

PROMPT If more details are needed see the ASH DETAILS FOR WORST

PROMPT MINUTES section at the bottom of the report.

column inst format 999

column minute format a12 tru

column event format a30 tru

column program format a40 wra

select to_char(sample_time,'Mondd_hh24mi') minute, inst_id inst, program, event,

sum(time_waited)/1000 TOTAL_WAIT_TIME , count(*) WAITS,

avg(time_waited)/1000 AVG_TIME_WAITED

from gv$active_session_history

where to_char(sample_time,'Mondd_hh24mi') in (select to_char(sample_time,'Mondd_hh24mi')

from gv$active_session_history

where event = 'log file sync'

group by to_char(sample_time,'Mondd_hh24mi'), inst_id

having avg(time_waited)/1000 > &&threshold and sum(time_waited)/1000 > 1)

and (program like '%LGWR%' or program like '%LMS%' or

program like '%LNS%' or event = 'log file sync')

group by to_char(sample_time,'Mondd_hh24mi'), inst_id, program, event

order by 1,2,3,5 desc, 4;

 

PROMPT

PROMPT HISTOGRAM DATA FOR LFS AND OTHER RELATED WAITS:

PROMPT

PROMPT APPROACH: Look at the wait distribution for log file sync waits

PROMPT by looking at "wait_time_milli". Look at the high wait times then

PROMPT see if you can correlate those with other related wait events.

column event format a40 wra

select inst_id, event, wait_time_milli, wait_count

from gv$event_histogram

where event in ('log file sync','gcs log flush sync',

'log file parallel write','wait for scn ack',

'log file switch completion','gc cr grant 2-way',

'gc buffer busy','gc current block 2-way') or

event like '%LGWR%' or event like '%LNS%'

order by 2 desc,1,3;

 

PROMPT

PROMPT ORDERED BY WAIT_TIME_MILLI

select inst_id, event, wait_time_milli, wait_count

from gv$event_histogram

where event in ('log file sync','gcs log flush sync',

'log file parallel write','wait for scn ack',

'log file switch completion','gc cr grant 2-way',

'gc buffer busy','gc current block 2-way')

or event like '%LGWR%' or event like '%LNS%'

order by 3,1,2 desc;

 

PROMPT

PROMPT REDO WRITE STATS

PROMPT

PROMPT "redo write time" in centiseconds (100 per second)

PROMPT 11.1: "redo write broadcast ack time" in centiseconds (100 per second)

PROMPT 11.2: "redo write broadcast ack time" in microseconds (1000 per millisecond)

column value format 99999999999999999999

column milliseconds format 99999999999999.999

select v.version, ss.inst_id, ss.name, ss.value,

decode(substr(version,1,4),

'11.1',decode (name,'redo write time',value*10,

'redo write broadcast ack time',value*10),

'11.2',decode (name,'redo write time',value*10,

'redo write broadcast ack time',value/1000),

decode (name,'redo write time',value*10)) milliseconds

from gv$sysstat ss, v$instance v

where name like 'redo write%' and value > 0

order by 1,2,3;

 

PROMPT

PROMPT AWR WORST AVG LOG FILE SYNC SNAPS:

PROMPT

PROMPT APPROACH: These are the AWR snaps where the average 'log file sync'

PROMPT times were the highest.

column begin format a12 tru

column end format a12 tru

column name format a13 tru

select dhs.snap_id, dhs.instance_number inst, to_char(dhs.begin_interval_time,'Mondd_hh24mi') BEGIN,

to_char(dhs.end_interval_time,'Mondd_hh24mi') END,

en.name, se.time_waited_micro/1000 total_wait_time, se.total_waits,

se.time_waited_micro/1000 / se.total_waits avg_time_waited

from dba_hist_snapshot dhs, wrh$_system_event se, v$event_name en

where (dhs.snap_id = se.snap_id and dhs.instance_number = se.instance_number)

and se.event_id = en.event_id and en.name = 'log file sync' and

dhs.snap_id in (select snap_id from (

select se.snap_id, se.time_waited_micro/1000 / se.total_waits avg_time_waited

from wrh$_system_event se, v$event_name en

where se.event_id = en.event_id and en.name = 'log file sync'

order by avg_time_waited desc)

where rownum < 4)

order by 1,2;

 

PROMPT

PROMPT AWR REDO WRITE STATS

PROMPT

PROMPT "redo write time" in centiseconds (100 per second)

PROMPT 11.1: "redo write broadcast ack time" in centiseconds (100 per second)

PROMPT 11.2: "redo write broadcast ack time" in microseconds (1000 per millisecond)

column stat_name format a30 tru

select v.version, ss.snap_id, ss.instance_number inst, sn.stat_name, ss.value,

decode(substr(version,1,4),

'11.1',decode (stat_name,'redo write time',value*10,

'redo write broadcast ack time',value*10),

'11.2',decode (stat_name,'redo write time',value*10,

'redo write broadcast ack time',value/1000),

decode (stat_name,'redo write time',value*10)) milliseconds

from wrh$_sysstat ss, wrh$_stat_name sn, v$instance v

where ss.stat_id = sn.stat_id

and sn.stat_name like 'redo write%' and ss.value > 0

and ss.snap_id in (select snap_id from (

select se.snap_id, se.time_waited_micro/1000 / se.total_waits avg_time_waited

from wrh$_system_event se, v$event_name en

where se.event_id = en.event_id and en.name = 'log file sync'

order by avg_time_waited desc)

where rownum < 4)

order by 1,2,3;

 

PROMPT

PROMPT AWR LFS AND OTHER RELATED WAITS FOR WORST LFS AWRs:

PROMPT

PROMPT APPROACH: These are the AWR snaps where the average 'log file sync'

PROMPT times were the highest. Look at related waits at those times.

column name format a40 tru

select se.snap_id, se.instance_number inst, en.name,

se.total_waits, se.time_waited_micro/1000 total_wait_time,

se.time_waited_micro/1000 / se.total_waits avg_time_waited

from wrh$_system_event se, v$event_name en

where se.event_id = en.event_id and (en.name in

('log file sync','gcs log flush sync',

'log file parallel write','wait for scn ack',

'log file switch completion','gc cr grant 2-way',

'gc buffer busy','gc current block 2-way')

or en.name like '%LGWR%' or en.name like '%LNS%')

and se.snap_id in (select snap_id from (

select se.snap_id, se.time_waited_micro/1000 / se.total_waits avg_time_waited

from wrh$_system_event se, v$event_name en

where se.event_id = en.event_id and en.name = 'log file sync'

order by avg_time_waited desc)

where rownum < 4)

order by 1, 6 desc;

 

PROMPT

PROMPT AWR HISTOGRAM DATA FOR LFS AND OTHER RELATED WAITS FOR WORST LFS AWRs:

PROMPT Note: This query won't work on 10.2 - ORA-942

PROMPT

PROMPT APPROACH: Look at the wait distribution for log file sync waits

PROMPT by looking at "wait_time_milli". Look at the high wait times then

PROMPT see if you can correlate those with other related wait events.

select eh.snap_id, eh.instance_number inst, en.name, eh.wait_time_milli, eh.wait_count

from wrh$_event_histogram eh, v$event_name en

where eh.event_id = en.event_id and

(en.name in ('log file sync','gcs log flush sync',

'log file parallel write','wait for scn ack',

'log file switch completion','gc cr grant 2-way',

'gc buffer busy','gc current block 2-way')

or en.name like '%LGWR%' or en.name like '%LNS%')

and snap_id in (select snap_id from (

select se.snap_id, se.time_waited_micro/1000 / se.total_waits avg_time_waited

from wrh$_system_event se, v$event_name en

where se.event_id = en.event_id and en.name = 'log file sync'

order by avg_time_waited desc)

where rownum < 4)

order by 1,3 desc,2,4;

 

PROMPT

PROMPT ORDERED BY WAIT_TIME_MILLI

PROMPT Note: This query won't work on 10.2 - ORA-942

select eh.snap_id, eh.instance_number inst, en.name, eh.wait_time_milli, eh.wait_count

from wrh$_event_histogram eh, v$event_name en

where eh.event_id = en.event_id and

(en.name in ('log file sync','gcs log flush sync',

'log file parallel write','wait for scn ack',

'log file switch completion','gc cr grant 2-way',

'gc buffer busy','gc current block 2-way')

or en.name like '%LGWR%' or en.name like '%LNS%')

and snap_id in (select snap_id from (

select se.snap_id, se.time_waited_micro/1000 / se.total_waits avg_time_waited

from wrh$_system_event se, v$event_name en

where se.event_id = en.event_id and en.name = 'log file sync'

order by avg_time_waited desc)

where rownum < 4)

order by 1,4,2,3 desc;

 

PROMPT

PROMPT ASH DETAILS FOR WORST MINUTES:

PROMPT

PROMPT APPROACH: If you cannot determine the problem from the data

PROMPT above, you may need to look at the details of what each session

PROMPT is doing during each 'bad' snap. Most likely you will want to

PROMPT note the times of the high log file sync waits, look at what

PROMPT LGWR is doing at those times, and go from there...

column program format a45 wra

column sample_time format a25 tru

column event format a30 tru

column time_waited format 999999.999

column p1 format a40 tru

column p2 format a40 tru

column p3 format a40 tru

select sample_time, inst_id inst, session_id, program, event, time_waited/1000 TIME_WAITED,

p1text||': '||p1 p1,p2text||': '||p2 p2,p3text||': '||p3 p3

from gv$active_session_history

where to_char(sample_time,'Mondd_hh24mi') in (select

to_char(sample_time,'Mondd_hh24mi')

from gv$active_session_history

where event = 'log file sync'

group by to_char(sample_time,'Mondd_hh24mi'), inst_id

having avg(time_waited)/1000 > &&threshold)

and time_waited > 0.5

order by 1,2,3,4,5;

 

select to_char(sysdate,'Mondd hh24:mi:ss') TIME from dual;

 

spool off

 

PROMPT

PROMPT OUTPUT FILE IS: lfsdiag_&&dbname&&timestamp&&suffix

PROMPT

参考资料:

http://www.oraclecommunity.net/profiles/blogs/wait-events-log-file-sync-and-log-file-parallel-write
http://blog.itpub.net/15880878/viewspace-722994/

ORACLE等待事件: log file parallel write的更多相关文章

  1. Oracle等待事件db file parallel read

    SQL> select event#,name,parameter1,parameter2,parameter3 from v$event_name where name = 'db file ...

  2. oracle之 等待事件LOG FILE SYNC (awr)优化

    log file sycn是ORACLE里最普遍的等待事件之一,一般log file sycn的等待时间都非常短 1-5ms,不会有什么问题,但是一旦出问题,往往都比较难解决.什么时候会产生log f ...

  3. Oracle之 等待事件log file sync + log file parallel write (awr优化)

    这是3月份某客户的情况,原因是server硬件故障后进行更换之后,业务翻译偶尔出现提交缓慢的情况.我们先来看下awr的情况. 我们能够看到,该系统的load profile信息事实上并不高,每秒才21 ...

  4. Oracle 等待事件 db file sequential read

    db file sequential read-数据文件顺序读取 等待事件: "db file sequential read" Reference Note (文档 ID 345 ...

  5. Oracle db file parallel write 和 log file parallel write 等待事件

    一. db file parallel write等待事件 引自如下blog: http://oradbpedia.com/wiki/Wait_Events_-_db_file_parallel_wr ...

  6. Oracle等待事件之log file parallel write

    log file parallel write表示等待 LGWR 向操作系统请求 I/O 开始直到完成 I/O.这种事件发生通常表示日志文件发生了I/O 竞争或者文件所在的驱动器较慢.这说明这种等待与 ...

  7. oracle等待事件以及解决方案

    我们可以通过视图v$session_wait来查看系统当前的等待事件,以及与等待事件相对应的资源的相关信息,从而可确定出产生瓶颈的类型及其对象. v$session_wait的p1.p2.p3告诉我们 ...

  8. Oracle等待事件之等待事件详解

    一. 等待事件的相关知识:1.1 等待事件主要可以分为两类:即空闲(IDLE)等待事件和非空闲(NON-IDLE)等待事件.1). 空闲等待事件指ORACLE正等待某种工作,在诊断和优化数据库的时候, ...

  9. 全面解析Oracle等待事件的分类、发现及优化

    一.等待事件由来 大家可能有些奇怪,为什么说等待事件,先谈到了指标体系.其实,正是因为指标体系的发展,才导致等待事件的引入.总结一下,Oracle的指标体系,大致经历了下面三个阶段: · 以命中率为主 ...

随机推荐

  1. Android 数据库框架OrmLite的使用(二)

    前面说了OrmLite的基本使用,现在说下更规范的做法以及常用的一些功能. 1.DatabaseHelper package com.example.ormlite.db; import java.s ...

  2. Oracle启动报错ORA-27102解决

    环境:RHEL5.5 + Oracle 10.2.0.4 此错误一般是因为数据库的初始化参数文件的内存设置不当导致.本例是因为操作系统参数设置问题导致. 当前现象:Oracle启动报错ORA-2710 ...

  3. 重温Servlet学习笔记--servletContext对象

    一个项目中只有一个ServletContext对象,我们可以在多个servlet中获取这个唯一的对象,使用它可以给多个servlet传递数据,我们通常成servletContext为上下文对象.这个对 ...

  4. ztree + ashx +DataTable +Oracle

    问题描述 好久没有使用ztree了,刚才在使用ztree做导航时遇到了几个小问题: 1.返回数据源是undefined . 2.数据出现后树结构没有出现(pIdKey单词拼写错误). 3.在使用Ora ...

  5. CSS-float详解,深入理解clear:both[转+部分原创]

    首先要知道,div是块级元素,在页面中独占一行,自上而下排列,也就是传说中的流. 可以看出,即使div1的宽度很小,页面中一行可以容下div1和div2,div2也不会排在div1后边,因为div元素 ...

  6. 细说Cookie

    阅读目录 开始 Cookie 概述 Cookie的写.读过程 使用Cookie保存复杂对象 Js中读写Cookie Cookie在Session中的应用 Cookie在身份验证中的应用 Cookie的 ...

  7. linux源码分析(五)-start_kernel

    前置:这里使用的linux版本是4.8,x86体系. local_irq_disable(); 这个函数是做了关闭中断操作.和后面的local_irq_enable相对应.说明启动的下面函数是不允许被 ...

  8. HTML5小游戏之见缝插针

    今天给大家带来的就是一款叫做<见缝插针>的游戏.有空你就往里插,直到你无处可插!看你能过多少关! 简洁大气 黑白搭配游戏画面非常的简洁,米白色的背景中央,放置着一个不断旋转的太阳状的球体, ...

  9. 【转】asp.net(c#)加密解密算法之sha1、md5、des、aes实现源码详解

    原文地址:http://docode.top/Article/Detail/10003 目录: 1..Net(C#)平台下Des加密解密源代码 2..Net(C#)平台下Aes加密解密源代码 3..N ...

  10. .Net 开源项目资源大全

    伯乐在线已在 GitHub 上发起「DotNet 资源大全中文版」的整理.欢迎扩散.欢迎加入. https://github.com/jobbole/awesome-dotnet-cn (注:下面用 ...