Re: corrupted tuple (header?), pg_filedump output

Tom Lane wrote:
> Eric Parusel <lists(at)globalrelay(dot)net> writes:
>
>>./pg_filedump -i -f -R 28393 /data1/pgsql/data/base/17760/18004
>>--snip--
>> Item 2 -- Length: 1728 Offset: 6464 (0x1940) Flags: USED
>> XMIN: 12 CMIN: 196608 XMAX: 122552335 CMAX|XVAC: 177664675
>> Block Id: 0 linp Index: 47241 Attributes: 3692 Size: 0
>> infomask: 0x0900 (XMIN_COMMITTED|XMAX_INVALID)
>> Error: Computed header length not equal to header size.
>> Computed <28> Header: <0>
>>--snip--
>
>
> None of those fields seem very sensible. I suspect what's actually
> corrupt is the ItemId in the page header --- ie, it's pointing at
> something that's not a tuple header. You might eyeball the ItemId
> array and see if any pattern of corruption is visible.

Ok, the other items in the dump seemed sensible to me.

Is this what you're speaking of?

Block 28393 ********************************************************
<Header> -----
Block Offset: 0x0ddd2000 Offsets: Lower 56 (0x0038)
Block: Size 8192 Version 2 Upper 1064 (0x0428)
LSN: logid 242 recoff 0x9387bd78 Special 8192 (0x2000)
Items: 9 Free Space: 1008
Length (including item array): 60

0000: f2000000 78bd8793 01000000 38002804 ....x.......8.(.
0010: 00200220 808cb80d 4099800d a485b80d . . (dot)(dot)(dot)(dot)(at)(dot)(dot)(dot)(dot)(dot)(dot)(dot)
0020: ec847001 28848401 9c984801 b8060000 ..p.(.....H.....
0030: 5c93800a 28040000 00000000 \...(.......

Or, this?:

<Data> ------
Item 1 -- Length: 1756 Offset: 3200 (0x0c80) Flags: USED
Item 2 -- Length: 1728 Offset: 6464 (0x1940) Flags: USED
Item 3 -- Length: 1756 Offset: 1444 (0x05a4) Flags: USED
Item 4 -- Length: 184 Offset: 1260 (0x04ec) Flags: USED
Item 5 -- Length: 194 Offset: 1064 (0x0428) Flags: USED
Item 6 -- Length: 164 Offset: 6300 (0x189c) Flags: USED
Item 7 -- Length: 0 Offset: 1720 (0x06b8) Flags: 0x00
Item 8 -- Length: 1344 Offset: 4956 (0x135c) Flags: USED
Item 9 -- Length: 0 Offset: 1064 (0x0428) Flags: 0x00

> (Oh btw: are you certain you have the right version of pg_filedump?
> If it's older than your postmaster it might be giving you wrong
> answers.)

Yup, I do believe so, I've got the PGDG 8.0.1 rpms installed and I just
downloaded the PGDG 8.0.1 contrib srpm, and compiled pg_filedump
(downloaded from
http://sources.redhat.com/rhdb/tools/pg_filedump-4.0.tar)...

> If you can do without the whole page, the easiest way is to zero out the
> entire page with 'dd'. If you just want to kill that one tuple you can
> reset the LP_USED bit in its ItemId. Either way the postmaster needs to
> be stopped to be sure that it's not got the page buffered while you
> change it.

Ok, how do I go about resetting the LP_USED bit? I'll make sure the
postmaster stopped for such an operation...

*******
*******
Oh crap, I just noticed that there is more than just that item that's
inconsistent...

Item 6 -- Length: 164 Offset: 6300 (0x189c) Flags: USED
XMIN: 1852400896 CMIN: 1970348133 XMAX: 543424609 CMAX|XVAC:
1869112179
Block Id: 1819244288 linp Index: 29295 Attributes: 2667 Size: 101
infomask: 0x7270
(HASOID|XMIN_INVALID|MARKED_FOR_UPDATE|UPDATED|MOVED_OFF)
Error: Computed header length not equal to header size.
Computed <32> Header: <101>

Item 8 -- Length: 1344 Offset: 4956 (0x135c) Flags: USED
XMIN: 2100285 CMIN: 1644759593 XMAX: 1634148473 CMAX|XVAC: 2003137536
Block Id: 2036411953 linp Index: 103 Attributes: 28524 Size: 108
infomask: 0x6162 (HASVARWIDTH|XMIN_COMMITTED|UPDATED|MOVED_OFF)
Error: Computed header length not equal to header size.
Computed <28> Header: <108>
*******
*******

Thanks again,
Eric