[pgpool-hackers: 6] Re: Fix error on the lsn math

[Tatsuo Ishii]

> On Tue, 2011-11-29 at 09:47 +0900, Tatsuo Ishii wrote:
>> > On Mon, 2011-11-28 at 07:56 +0900, Tatsuo Ishii wrote:
>> >> Guillaume,
>> >> 
>> >> About your commit in the subject.
>> >> 
>> >> Before it used 0xff000000 and changed by me according to the
>> >> suggestion:
>> >> 
>> >> From: Anton Yuzhaninov
>> >> To: pgpool-hackers at pgfoundry.org
>> >> Date: Wed, 04 May 2011 19:56:54 +0400
>> >> 
>> >> > I have two questions/suggestions about replication lag check for
>> >> > straming replication in pgpool-II.
>> >> > 
>> >> > 1. In text_to_lsn() there is formula
>> >> > 
>> >> > lsn = xlogid * 16 * 1024 * 1024 * 255 + xrecoff;
>> >> > 
>> >> > 16 * 1024 * 1024 * 255 = 4278190080 = 0xff000000
>> >> > 
>> >> > where this magic numbers come from?
>> >> > 
>> >> > I think multiplier should be
>> >> > 0xffffffff - XLogSegSize.
>> >> > 
>> >> > In PostgreeSQL source this constant used in
>> >> > src/include/access/xlog_internal.h
>> >> > 
>> >> > #define XLogSegsPerFile (((uint32) 0xffffffff) / XLogSegSize)
>> >> > 
>> >> > But in comment noted, that one segment at the end of each log file is
>> >> > wasted
>> >> > 
>> >> > Each xlogid corresponds to (XLogSegsPerFile - 1) * XLogSegSize =
>> >> > 0xffffffff - XLogSegSize bytes
>> >> 
>> >> I confirmed in the source code and decided to change "16 * 1024 * 1024
>> >> * 255 = 4278190080 = 0xff000000" to 0xffffffff. Now you changed it
>> >> to original constant. Do you have any point against his argument?
>> > 
>> > To be honest, I didn't know about Anton's point of view. I guess he's
>> > wrong. See
>> > http://eulerto.blogspot.com/2011/11/understanding-wal-nomenclature.html
>> > for details. Many agree that we should use 0xff000000.
>> 
>> I think you are plain wrong.
>> 
>> 	lsn = xlogid * ((unsigned long long int)0xff000000 - WALSEGMENTSIZE) + xrecoff;
>> 
>> In the blog he said that correct formula should be:
>> 
>> FF000000 * xlog_master + offset_master
>> 
>> Here xlog_master = our xlogid, and offset_master = our xrecoff.
>> So correct fomula for us is:
>> 
>> 	lsn = xlogid * ((unsigned long long int)0xff000000) + xrecoff;
>> 
>> or 
>> 
>>     lsn = xlogid * ((unsigned long long int)0xffffffff - WALSEGMENTSIZE) + xrecoff;
>> 	(actually that was before you made change. Note that 0xffffffff - WALSEGMENTSIZE = 0xff0000).
>> 
>> In summary you subtracted too much.
> 
> You're right. My apologies to you and Anton, I was too quick on this
> one. Patch reverted.

No problem and thank you for understanding.
--
Tatsuo Ishii
SRA OSS, Inc. Japan
English: http://www.sraoss.co.jp/index_en.php
Japanese: http://www.sraoss.co.jp