buffer busy wait--是寫阻塞讀還是讀阻塞寫

如下圖一，假設有個進程A正在讀10號文件的11號數據塊，此時又有一個進程B准備寫該數據塊。首先獲取cbc latch然后復制內存中的10號文件11號塊並新增一個buffer header（bh），並把原來的bh中status改為xcur，復制完成后釋放cbc latch。然后如圖二，獲得cbc latch修改原塊狀態為CR，復制塊狀態為XCUR並把buffer pin修改為X，然后可以釋放latch cbc，然后進程B 可以修改該數據塊，此時如果進程c要讀該數據塊，那么進程c就要進行等待，等待事件即buffer busy wait。附實驗證明

圖一：

圖二：

實驗一：

會話一：

SQL> select distinct sid from v$mystat;

       SID
----------
      853

會話二：

SQL> select distinct sid from v$mystat
  2  ;

       SID
----------
      1713

會話三：

SQL> select sid,event,total_waits,total_timeouts,time_waited_micro/100,time_waited_micro/total_waits/100
  2  from v$session_event where sid in (853,1713) and event in ('buffer busy waits');

no rows selected

會話一進行讀數據塊：

 declare
   j number;
    begin
  for i in 1..3000000 loop
   select id into j from ye where rowid='AAAWt/AAKAAAAHaACV';
   end loop;
   end;
/

會話二進行寫數據塊：

begin
for i in 1..300 loop
update ye set id=id+0 where rowid='AAAWt/AAKAAAAHaACV';
commit;
end loop;
end;
/

會話三查詢等待事件為853會話等待：
       SID EVENT                                    TOTAL_WAITS TOTAL_TIMEOUTS TIME_WAITED_MICRO/100 TIME_WAITED_MICRO/TOTAL_WAITS/100
---------- ---------------------------------------- ----------- -------------- --------------------- ---------------------------------
       853 buffer busy waits                                 81              0                 30.32                        .374320988

上述步驟不變，將會話二中更新數據增到1000次
在會話三中發現也出現了3次buffer busy wait：

SQL> /

       SID EVENT                                    TOTAL_WAITS TOTAL_TIMEOUTS TIME_WAITED_MICRO/100 TIME_WAITED_MICRO/TOTAL_WAITS/100
---------- ---------------------------------------- ----------- -------------- --------------------- ---------------------------------
       853 buffer busy waits                                499              0                238.01                        .476973948
      1713 buffer busy waits                                  3              0                  1.67                        .556666667

這3次等待其實發生在undo block上，當會話853構造cr塊是需要在undo上加上共享模式buffer bin，在會話1713寫block需要向undo寫入前印象，因此需要等待853構造完cr塊。

會話1713分別改為:

begin
for i in 1..60000 loop
update ye set id=id+0 where rowid='AAAWt/AAKAAAAHaACV';
if mod(i,500)=0 then
commit;
end if;
end loop;
commit;
end;
/

begin
for i in 1..60000 loop
update ye set id=id+0 where rowid='AAAWt/AAKAAAAHaACV';
commit;
end loop;
commit;
end;
/

會話三的結果：

SQL> select sid,event,total_waits,total_timeouts,time_waited_micro/100,time_waited_micro/total_waits/100
2 from v$session_event where sid in (853,1713) and event in ('buffer busy waits');

SID EVENT TOTAL_WAITS TOTAL_TIMEOUTS TIME_WAITED_MICRO/100 TIME_WAITED_MICRO/TOTAL_WAITS/100
---------- ---------------------------------------- ----------- -------------- --------------------- ---------------------------------
853 buffer busy waits 12923 0 4291.97 .332118703
1713 buffer busy waits 1448 0 567.66 .392030387

SQL> /

SID EVENT TOTAL_WAITS TOTAL_TIMEOUTS TIME_WAITED_MICRO/100 TIME_WAITED_MICRO/TOTAL_WAITS/100
---------- ---------------------------------------- ----------- -------------- --------------------- ---------------------------------
853 buffer busy waits 32471 0 11886.11 .366053094
1713 buffer busy waits 1568 0 605.15 .3859375

 

說明長事務會等待讀構造cr塊比短事務等待讀構造cr塊更久