Improper use about DatumGetInt32

Hi

In (/contrib/bloom/blutils.c:277), I found it use DatumGetInt32 to get UInt32 type.
Is it more appropriate to use DatumGetUInt32 here?

See the attachment for the patch

Bes regards,
houzj

On Mon, Sep 21, 2020 at 6:47 AM Hou, Zhijie <[hidden email]> wrote:
>
> In (/contrib/bloom/blutils.c:277), I found it use DatumGetInt32 to get UInt32 type.
> Is it more appropriate to use DatumGetUInt32 here?
>

Makes sense. +1 for the patch. I think with the existing code also we
don't have any problem. If we assume that the hash functions return
uint32, with DatumGetInt32() we are typecasting that uint32 result to
int32, we are assigning it to uint32 i.e. typecasting int32 back to
uint32. Eventually, I think, we will see the proper value in hashVal.
I did a small experiment to prove this [1].

    uint32        hashVal;
    hashVal = DatumGetInt32(FunctionCall1Coll(&state->hashFn[attno],
state->collations[attno], value));

It's good to run a few test cases/test suites(if they exist) that hit
this part of the code, just to ensure we don't break anything.

[1]
int main()
{
    unsigned int u = 3 * 1024 * 1024 * 1024;

    printf("%u\n", u);
    int i = u;
    printf("%d\n", i);
    unsigned int u1 = i;
    printf("%u\n", u1);

    return 0;
}
Output of the above snippet:
3221225472
-1073741824
3221225472

With Regards,
Bharath Rupireddy.
EnterpriseDB: http://www.enterprisedb.com

On Sun, Sep 20, 2020 at 9:17 PM Hou, Zhijie <[hidden email]> wrote:
> In (/contrib/bloom/blutils.c:277), I found it use DatumGetInt32 to get UInt32 type.
> Is it more appropriate to use DatumGetUInt32 here?

Typically, the DatumGetBlah() function that you pick should match the
SQL data type that the function is returning. So if the function
returns pg_catalog.int4, which corresponds to the C data type int32,
you would use DatumGetInt32. There is no SQL type corresponding to the
C data type uint32, so I'm not sure why we even have DatumGetUInt32.
I'm sort of suspicious that there's some fuzzy thinking going on
there.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

Robert Haas <[hidden email]> writes:
> Typically, the DatumGetBlah() function that you pick should match the
> SQL data type that the function is returning. So if the function
> returns pg_catalog.int4, which corresponds to the C data type int32,
> you would use DatumGetInt32. There is no SQL type corresponding to the
> C data type uint32, so I'm not sure why we even have DatumGetUInt32.

xid?

                        regards, tom lane

Hi,

On 2020-09-21 14:08:22 -0400, Robert Haas wrote:
> There is no SQL type corresponding to the C data type uint32, so I'm
> not sure why we even have DatumGetUInt32.  I'm sort of suspicious that
> there's some fuzzy thinking going on there.

I think we mostly use it for the few places where we currently expose
data as a signed integer on the SQL level, but internally actually treat
it as a unsigned data. There's not a lot of those, but there e.g. is
pg_class.relpages.  There also may be places where we use it for
functions that can be created but not called from SQL (using the
INTERNAL type).

- Andres

On Mon, Sep 21, 2020 at 3:53 PM Andres Freund <[hidden email]> wrote:
> On 2020-09-21 14:08:22 -0400, Robert Haas wrote:
> > There is no SQL type corresponding to the C data type uint32, so I'm
> > not sure why we even have DatumGetUInt32.  I'm sort of suspicious that
> > there's some fuzzy thinking going on there.
>
> I think we mostly use it for the few places where we currently expose
> data as a signed integer on the SQL level, but internally actually treat
> it as a unsigned data.

So why is the right solution to that not DatumGetInt32() + a cast to uint32?

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

Robert Haas <[hidden email]> writes:
> On Mon, Sep 21, 2020 at 3:53 PM Andres Freund <[hidden email]> wrote:
>> I think we mostly use it for the few places where we currently expose
>> data as a signed integer on the SQL level, but internally actually treat
>> it as a unsigned data.

> So why is the right solution to that not DatumGetInt32() + a cast to uint32?

You're ignoring the xid use-case, for which DatumGetUInt32 actually is
the right thing.  I tend to agree though that if the SQL argument is
of a signed type, the least API-abusing answer is a signed DatumGetXXX
macro followed by whatever cast you need.

                        regards, tom lane

On Wed, Sep 23, 2020 at 1:41 AM Tom Lane <[hidden email]> wrote: There is DatumGetTransactionId() which should be used instead.
That made me search if there's PG_GETARG_TRANSACTIONID() and yes it's
there but only defined in xid.c. So pg_xact_commit_timestamp(),
pg_xact_commit_timestamp_origin() and pg_get_multixact_members() use
PG_GETARG_UNIT32. IMO those should be changed to use
PG_GETARG_TRANSACTIONID. That would require moving
PG_GETARG_TRANSACTIONID somewhere outside xid.c; may be fmgr.h where
other PG_GETARG_* are.

> I tend to agree though that if the SQL argument is
> of a signed type, the least API-abusing answer is a signed DatumGetXXX
> macro followed by whatever cast you need.
>

I looked for some uses of PG_GETARG_UNIT32() which is the counterpart
of DatumGetUint32(). Found some buggy usages apart from the ones which
can be converted to PG_GETARG_TRANSACTIONID listed above.
normal_rand() for example returns a huge number of rows and takes
forever if we pass a negative first argument to it. Someone could
misuse that for a DOS attack or it could be just an accident that they
pass a negative value to that function and the query takes forever.
explain analyze select count(*) from normal_rand(-1000000, 1.0, 1.0);
                                                               QUERY
PLAN
-----------------------------------------------------------------------------------------------------------------------------------------
 Aggregate  (cost=12.50..12.51 rows=1 width=8) (actual
time=2077574.718..2077574.719 rows=1 loops=1)
   ->  Function Scan on normal_rand  (cost=0.00..10.00 rows=1000
width=0) (actual time=1005176.149..1729994.366 rows=4293967296
loops=1)
 Planning Time: 0.346 ms
 Execution Time: 2079034.835 ms

get_raw_page() also does similar thing but the effect is not as dangerous
SELECT octet_length(get_raw_page('test1', 'main', -1)) AS main_1;
  ERROR:  block number 4294967295 is out of range for relation "test1"
Similarly for bt_page_stats() and bt_page_items()

PFA patches to correct those.

There's Oracle compatible chr() which also uses PG_GETARG_UINT32() but
it's (accidentally?) reporting the negative inputs correctly because
it filters out very large values and reports those using %d. It's
arguable whether we should change that, so I have left it untouched.
But I think we should change that as well and get rid of
PG_GETARG_UNIT32 altogether. This will prevent any future misuse.


--
Best Wishes,
Ashutosh Bapat

On 2020-Sep-23, Ashutosh Bapat wrote:

Hmm, yeah, I think this would be a good idea.

> get_raw_page() also does similar thing but the effect is not as dangerous
> SELECT octet_length(get_raw_page('test1', 'main', -1)) AS main_1;
>   ERROR:  block number 4294967295 is out of range for relation "test1"
> Similarly for bt_page_stats() and bt_page_items()

Hmm, but page numbers above signed INT_MAX are valid.  So this would
prevent reading all legitimate pages past that.