32TB relation size make mdnblocks overflow
Hello
We know that PostgreSQL doesn't support a single relation size over 32TB, limited by the MaxBlockNumber. But if we just 'insert into' one relation over 32TB, it will get an error message 'unexpected data beyond EOF in block 0 of relation' in ReadBuffer_common. The '0 block' is from mdnblocks function where the segment number is over 256 and make segno * RELSEG_SIZE over uint32's max value. So is it necessary to make the error message more readable like 'The relation size is over max value ...' and elog in mdnblocks?
This scene we met is as below, 'shl $0x18, %eax' make $ebx from 256 to 0, which makes segno from 256 to 0.
0x0000000000c2cc51 <+289>: callq 0xc657f0 <PathNameOpenFilePerm>
0x0000000000c2cc56 <+294>: mov -0x8(%r15),%rdi
0x0000000000c2cc5a <+298>: mov %r15,%rsi
0x0000000000c2cc5d <+301>: mov %eax,%r14d
0x0000000000c2cc60 <+304>: mov 0x10(%rdi),%rax
0x0000000000c2cc64 <+308>: callq *0x8(%rax)
0x0000000000c2cc67 <+311>: test %r14d,%r14d
0x0000000000c2cc6a <+314>: jns 0xc2cd68 <mdnblocks+568>
=> 0x0000000000c2cc70 <+320>: add $0x28,%rsp
0x0000000000c2cc74 <+324>: mov %ebx,%eax
0x0000000000c2cc76 <+326>: shl $0x18,%eax
0x0000000000c2cc79 <+329>: pop %rbx
0x0000000000c2cc7a <+330>: pop %r12
0x0000000000c2cc7c <+332>: pop %r13
0x0000000000c2cc7e <+334>: pop %r14
0x0000000000c2cc80 <+336>: pop %r15
0x0000000000c2cc82 <+338>: pop %rbp
0x0000000000c2cc83 <+339>: retq
Hi,
On Tue, Jan 18, 2022 at 02:21:14PM +0800, 陈佳昕(步真) wrote:
We know that PostgreSQL doesn't support a single relation size over 32TB,
limited by the MaxBlockNumber. But if we just 'insert into' one relation over
32TB, it will get an error message 'unexpected data beyond EOF in block 0 of
relation' in ReadBuffer_common. The '0 block' is from mdnblocks function
where the segment number is over 256 and make segno * RELSEG_SIZE over
uint32's max value. So is it necessary to make the error message more
readable like 'The relation size is over max value ...' and elog in
mdnblocks?
I didn't try it but this is supposed to be caught by mdextend():
/*
* If a relation manages to grow to 2^32-1 blocks, refuse to extend it any
* more --- we mustn't create a block whose number actually is
* InvalidBlockNumber. (Note that this failure should be unreachable
* because of upstream checks in bufmgr.c.)
*/
if (blocknum == InvalidBlockNumber)
ereport(ERROR,
(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
errmsg("cannot extend file \"%s\" beyond %u blocks",
relpath(reln->smgr_rnode, forknum),
InvalidBlockNumber)));
Didn't you hit this?
Julien Rouhaud <rjuju123@gmail.com> writes:
On Tue, Jan 18, 2022 at 02:21:14PM +0800, 陈佳昕(步真) wrote:
We know that PostgreSQL doesn't support a single relation size over 32TB,
limited by the MaxBlockNumber. But if we just 'insert into' one relation over
32TB, it will get an error message 'unexpected data beyond EOF in block 0 of
relation' in ReadBuffer_common.
I didn't try it but this is supposed to be caught by mdextend():
...
Didn't you hit this?
Probably not, if the OP was testing something predating 8481f9989,
ie anything older than the latest point releases.
(This report does seem to validate my comment in the commit log
that "I think it might confuse ReadBuffer's logic for data-past-EOF
later on". I'd not bothered to build a non-assert build to check
that, but this looks about like what I guessed would happen.)
regards, tom lane
I think we must meet some corner cases about our storage. The relation has 32TB blocks, so 'mdnblocks' gets the unexpected value, we will check it again.
Thanks a lot.
Import Notes
Reply to msg id not found: a7101d77-17f8-4b01-98b2-01c7d3e309f4.buzhen.cjx@alibaba-inc.com