On 04/24/2011 01:51 AM, Sergei Golubchik wrote:
Hi, Igor!
On Apr 24, Igor Babaev wrote:
When debugging I discovered the following bad code in sql_select.cc of the current 5.3 tree:
tmp= table->file->read_time(key, 1, min(tmp,s->worst_seeks)-1); <= !!!!! ? tmp*= record_count;
gannotate says that it appeared in the code after your merge with 5.2.
The fact is that 5.2 contains the correct expression here min(tmp,s->worst_seeks) not min(tmp,s->worst_seeks)-1
As a result if the number of records is 1 then min(tmp,s->worst_seeks)-1 is equal to 0 and we pass 0 as the last parameter to table->file->read_time. Naturally it returns 0.
Will you fix this problem or you'd rather want me to do it?
In 5.2 it is
tmp= record_count*min(tmp,s->worst_seeks);
In 5.3, record_count* is moved to a separate line, so it should just be
tmp= min(tmp,s->worst_seeks);
But it uses table->file->read_time(). Which is defined as
virtual double read_time(uint index, uint ranges, ha_rows rows) { return rows2double(ranges+rows); }
In the implementation of innodb/xtradb (ha_innobase::read_time) if the parameter rows = 0 , the function returns 0 :( if (rows <= 2) { return((double) rows); } Regards, Igor.
so, when ranges=1 and we pass rows=min(tmp,s->worst_seeks)-1, we get the same value of min(tmp,s->worst_seeks) as the result.
This was the change of introducing handler::keyread_time(), and at the same time using handler::read_time() where appropriate, to have all values comparable.
Regards, Sergei