Re: [HACKERS] Current 6.3 issues

From: "Vadim B(dot) Mikheev" <vadim(at)sable(dot)krasnoyarsk(dot)su>
To: PostgreSQL-development <hackers(at)postgreSQL(dot)org>
Cc: tih(at)Hamartun(dot)Priv(dot)NO
Subject: Re: [HACKERS] Current 6.3 issues
Date: 1998-02-23 14:04:55
Message-ID: 34F18207.9EFAB2C2@sable.krasnoyarsk.su
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-hackers

> Check select_views regression test

Second query is replaced -> there shouldn't be differences any more...

> 'Can not write block blind' error on createdb and regression collision(Vadim)

New bug was introduced in mdblindwrt(): backend tried to use current
database path when writing data for another data base => path was invalid
=> elog(ERROR). I fixed this: now backend uses GetRawDatabaseInfo()
to get real path for given data base, BUT I don't like this way:
GetRawDatabaseInfo() reads pg_database to get path and this is
really bad for performance!

Keeping in mind implementation of TABLESPACEs in _near_ future
we have to either use symlinks to give backend simple and fast way
to constract full path to a relation or store full path in
buffer header. Shared memory is critical resource but nevertheless
I prefer the second way.

Now about hanging another backends after 'blind' error. This is
very old bug in buffer manager: before doing blind write backend locks
buffer for IO using local spinlock of this buffer, but elog(ERROR)
doesn't releases spinlocks of this type! I tried to fix this too
but decided that there is not enough time to do this now: there are
another "bad places" in bufmgr. Re-visiting of bufmgr was in my
plans after 6.0, but ... hope to do this for 6.4.

Vadim

In response to

Browse pgsql-hackers by date

  From Date Subject
Next Message orion.SAPserv.Hamburg.dsh.de!wieck 1998-02-23 14:12:41 Re: [HACKERS] Here it is - view permissions]
Previous Message Vadim B. Mikheev 1998-02-23 13:53:32 ExecReScan implemented for Sort & Unique (DISTINCT) nodes (empty).