Обсуждение: pg_dump 9.1.1 hanging (collectSecLabels gets 0 labels)

We have a cluster running 9.1.1 where pg_dump hangs when we try to dump 
some a database inside of the cluster.  The server is running AIX.

I can see this on clean cluster where we do an initdb, followed by a 
createdb and try running pg_dump.

I've tracked the issue down to collectSecLabels in pg_dump.c

SELECT label, provider, classoid, objoid, objsbid FROM 
pg_catalog.pg_seclabel;

returns 0 rows.

The code in collectSecLabels() is not prepared to deal with a zero row 
result and tries to malloc 0 bytes.

I am not yet sure if the problem is that my pg_seclabel is empty or if 
the issue is in collectSecLabels() or if collectSecLabels shouldn't even 
be called.

Has anyone seen something similar?

Steve Singer <ssinger@ca.afilias.info> writes:
> I've tracked the issue down to collectSecLabels in pg_dump.c

> SELECT label, provider, classoid, objoid, objsbid FROM 
> pg_catalog.pg_seclabel;

> returns 0 rows.

> The code in collectSecLabels() is not prepared to deal with a zero row 
> result and tries to malloc 0 bytes.

pg_seclabel is almost always empty, so I'm not convinced that you've
identified your problem correctly.
        regards, tom lane

On 11-11-09 06:35 PM, Tom Lane wrote:
> Steve Singer<ssinger@ca.afilias.info>  writes:
>> I've tracked the issue down to collectSecLabels in pg_dump.c
>
>> SELECT label, provider, classoid, objoid, objsbid FROM
>> pg_catalog.pg_seclabel;
>
>> returns 0 rows.
>
>> The code in collectSecLabels() is not prepared to deal with a zero row
>> result and tries to malloc 0 bytes.
>
> pg_seclabel is almost always empty, so I'm not convinced that you've
> identified your problem correctly.
>
>             regards, tom lane
>

The attached patch seems to fix the issue.

The man page for malloc on AIX is pretty clear on what happens when you
try to malloc 0 bytes.  It returns NULL.

Steve Singer <ssinger@ca.afilias.info> writes:
> The man page for malloc on AIX is pretty clear on what happens when you 
> try to malloc 0 bytes.  It returns NULL.

Yes, that's a pretty common behavior for malloc(0).  It should not cause
a problem here AFAICS.

... Oh, I see, the problem is that &labels[-1] might not compare to
&labels[0] the way we want.  I think only the first hunk of your
patch is actually necessary.
        regards, tom lane

On 11-11-10 02:00 PM, Tom Lane wrote:
> Steve Singer<ssinger@ca.afilias.info>  writes:
>> The man page for malloc on AIX is pretty clear on what happens when you
>> try to malloc 0 bytes.  It returns NULL.
>
> Yes, that's a pretty common behavior for malloc(0).  It should not cause
> a problem here AFAICS.
>
> ... Oh, I see, the problem is that&labels[-1] might not compare to
> &labels[0] the way we want.  I think only the first hunk of your
> patch is actually necessary.
>
>             regards, tom lane
>

Yes the problem is still fixed if I only apply the first hunk.

Steve Singer <ssinger@ca.afilias.info> writes:
> On 11-11-10 02:00 PM, Tom Lane wrote:
>> ... Oh, I see, the problem is that&labels[-1] might not compare to
>> &labels[0] the way we want.  I think only the first hunk of your
>> patch is actually necessary.

> Yes the problem is still fixed if I only apply the first hunk.

OK, everything seems satisfactorily explained then.  Will commit the
fix, thanks for the report!
        regards, tom lane