Обсуждение: Dereferenced pointers checked as NULL in btree_utils_var.c

Hi all,

Coverity is pointing out $subject, with the following stuff in gbt_var_same():
        GBT_VARKEY *t1 = (GBT_VARKEY *) DatumGetPointer(d1);
        GBT_VARKEY *t2 = (GBT_VARKEY *) DatumGetPointer(d2);
        GBT_VARKEY_R r1,
                                r2;

        r1 = gbt_var_key_readable(t1); <= t1 dereferenced
        r2 = gbt_var_key_readable(t2); <= t2 dereferenced

        if (t1 && t2)
                result = ((*tinfo->f_cmp) (r1.lower, r2.lower,
collation) == 0 &&
                                  (*tinfo->f_cmp) (r1.upper, r2.upper,
collation) == 0);
        else
                result = (t1 == NULL && t2 == NULL); <= Coverity complains here

        return result;

As Heikki pointed me out on IM, the lack of crash report in this area,
as well as similar coding style in cube/ seem to be sufficient
arguments to simply remove those NULL checks instead of doing more
solid checks on them. Patch is attached.
Regards,
--
Michael

Michael Paquier <michael.paquier@gmail.com> writes:
> Coverity is pointing out $subject, with the following stuff in gbt_var_same():
> ...
> As Heikki pointed me out on IM, the lack of crash report in this area,
> as well as similar coding style in cube/ seem to be sufficient
> arguments to simply remove those NULL checks instead of doing more
> solid checks on them. Patch is attached.

The way to form a convincing argument that these checks are unnecessary
would be to verify that (1) the SQL-accessible functions directly calling
gbt_var_same() are all marked STRICT, and (2) the core GIST code never
passes a NULL to these support functions.  I'm prepared to believe that
(1) and (2) are both true, but it merits checking.
        regards, tom lane

On Tue, Jan 20, 2015 at 11:38 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Michael Paquier <michael.paquier@gmail.com> writes:
>> Coverity is pointing out $subject, with the following stuff in gbt_var_same():
>> ...
>> As Heikki pointed me out on IM, the lack of crash report in this area,
>> as well as similar coding style in cube/ seem to be sufficient
>> arguments to simply remove those NULL checks instead of doing more
>> solid checks on them. Patch is attached.
>
> The way to form a convincing argument that these checks are unnecessary
> would be to verify that (1) the SQL-accessible functions directly calling
> gbt_var_same() are all marked STRICT, and (2) the core GIST code never
> passes a NULL to these support functions.  I'm prepared to believe that
> (1) and (2) are both true, but it merits checking.

Sure. gbt_var_same is called by those functions in btree_gist/:
- gbt_bit_same
- gbt_bytea_same
- gbt_numeric_same
- gbt_text_same
=# select proname, proisstrict from pg_proc
where proname in ('gbt_bit_same', 'gbt_bytea_same',
'gbt_numeric_same', 'gbt_text_same');    proname      | proisstrict
------------------+-------------gbt_text_same    | tgbt_bytea_same   | tgbt_numeric_same | tgbt_bit_same     | t
(4 rows)

For the second point, I have run regression tests with an assertion in
gbt_var_same checking if t1 or t2 are NULL and tests worked. Also,
looking at the code of gist, gbt_var_same is called through
gistKeyIsEQ, which is used for an insertion or for a split. The point
is that when doing an insertion, a call to gistgetadjusted is done and
we look if one of the keys is NULL. If one of them is, code does not
go through gistKeyIsEQ, so the NULL checks do not seem necessary in
gbt_var_same.

Btw, looking at the code of gist, I think that it would be interesting
to add an assertion in gistKeyIsEQ like that:
Assert(DatumGetPointer(a) != NULL && DatumGetPointer(b) != NULL);
Thoughts?
-- 
Michael

On 01/21/2015 07:14 AM, Michael Paquier wrote:
> Also,
> looking at the code of gist, gbt_var_same is called through
> gistKeyIsEQ, which is used for an insertion or for a split. The point
> is that when doing an insertion, a call to gistgetadjusted is done and
> we look if one of the keys is NULL. If one of them is, code does not
> go through gistKeyIsEQ, so the NULL checks do not seem necessary in
> gbt_var_same.

I think you are confusing NULL pointers with an SQL NULL. 
gistgetadjusted checks that it's not an SQL NULL (!oldisnull[i]), but it 
does not check if it's a NULL pointer 
(DatumGetPointer(oldentries[i].key) != NULL). The case we're worried 
about is that the value is not an SQL NULL, i.e. isnull flag is false, 
but the Datum is a NULL pointer.

Nevertheless, looking at the code, I don't that a NULL pointer can ever 
be passed to the same-method, for any of the built-in or contrib 
opclasses, but it's very confusing because some functions check for 
that. My proof goes like this:

1. The key value passed as argument must've been originally formed by 
the compress, union, or picksplit methods.

1.1. Some compress methods do nothing, and just return the passed-in 
key, which comes from the table and cannot be a NULL pointer (the 
compress method is never called for SQL NULLs). Other compress methods 
don't check for a NULL pointer, and would crash if there was one. 
gist_poly_compress() and gist_circle_compress() do check for a NULL, but 
they only return a NULL key if the input key is NULL, which cannot 
happen because the input comes from a table. So I believe the 
NULL-checks in those functions are dead code, and none of the compress 
methods ever return a NULL key.

1.2. None of the union methods return a NULL pointer (nor expect one as 
input).

1.3. None of the picksplit methods return a NULL pointer. They all 
return one of the original values, or one formed with the union method. 
The picksplit method can return a NULL pointer in the spl_ldatum or 
spl_rdatum field, in the degenerate case that it puts all keys on the 
same halve. However, the caller, gistUserPickSplit checks for that and 
immediately overwrites spl_ldatum and spl_rdatum with sane values in 
that case.


I don't understand what this sentence in the documentation on the 
compress method means:

> Depending on your needs, you could also need to care about
> compressing NULL values in there, storing for example (Datum) 0 like
> gist_circle_compress does.

The compress method is never called for NULLs, so the above is nonsense. 
I think it should be removed, as well as the checks in 
gist_circle_compress and gist_poly_compress. According to git history, 
the check in gist_circle_compress been there ever since the module was 
imported into contrib/rtree_gist, in 2001. The documentation was added 
later:

commit a0a3883dd977d6618899ccd14258a0696912a9d2
Author: Tom Lane <tgl@sss.pgh.pa.us>
Date:   Fri Jun 12 19:48:53 2009 +0000
    Improve documentation about GiST opclass support functions.    Dimitri Fontaine

I'm guessing that Tom added that sentence (it was not in the patch that 
Dimitri submitted originally) just because there was that check in the 
existing function, without realizing that the check was in fact dead code.

- Heikki

Heikki Linnakangas <hlinnakangas@vmware.com> writes:
> I think you are confusing NULL pointers with an SQL NULL. 
> gistgetadjusted checks that it's not an SQL NULL (!oldisnull[i]), but it 
> does not check if it's a NULL pointer 
> (DatumGetPointer(oldentries[i].key) != NULL). The case we're worried 
> about is that the value is not an SQL NULL, i.e. isnull flag is false, 
> but the Datum is a NULL pointer.

Actually both of these deserve to be worried about; though it's fairly
clear from looking at the core GIST code that it avoids calling
gistKeyIsEQ on SQL NULLs.

> Nevertheless, looking at the code, I don't that a NULL pointer can ever 
> be passed to the same-method, for any of the built-in or contrib 
> opclasses, but it's very confusing because some functions check for 
> that. My proof goes like this:

> 1. The key value passed as argument must've been originally formed by 
> the compress, union, or picksplit methods.

> 1.1. Some compress methods do nothing, and just return the passed-in 
> key, which comes from the table and cannot be a NULL pointer (the 
> compress method is never called for SQL NULLs). Other compress methods 
> don't check for a NULL pointer, and would crash if there was one. 
> gist_poly_compress() and gist_circle_compress() do check for a NULL, but 
> they only return a NULL key if the input key is NULL, which cannot 
> happen because the input comes from a table. So I believe the 
> NULL-checks in those functions are dead code, and none of the compress 
> methods ever return a NULL key.

> 1.2. None of the union methods return a NULL pointer (nor expect one as 
> input).

> 1.3. None of the picksplit methods return a NULL pointer. They all 
> return one of the original values, or one formed with the union method. 
> The picksplit method can return a NULL pointer in the spl_ldatum or 
> spl_rdatum field, in the degenerate case that it puts all keys on the 
> same halve. However, the caller, gistUserPickSplit checks for that and 
> immediately overwrites spl_ldatum and spl_rdatum with sane values in 
> that case.

Sounds good to me.

> I don't understand what this sentence in the documentation on the 
> compress method means:

>> Depending on your needs, you could also need to care about
>> compressing NULL values in there, storing for example (Datum) 0 like
>> gist_circle_compress does.

I believe you're right that I added this because there were checks for
null pointers in some but not all of the opclass support functions.
It looked to me like some opclasses might be intending to pass around null
pointers as valid (not-SQL-NULL) values.  I think your analysis above
eliminates that idea though.  It's a sufficiently weird concept that
I don't feel a need to document or support it.

So I'm fine with taking out both this documentation text and the dead
null-pointer checks; but let's do that all in one patch not piecemeal.
In any case, this is just cosmetic cleanup; there's no actual hazard
here.
        regards, tom lane

On Wed, Jan 28, 2015 at 3:15 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> So I'm fine with taking out both this documentation text and the dead
> null-pointer checks; but let's do that all in one patch not piecemeal.
> In any case, this is just cosmetic cleanup; there's no actual hazard
> here.
Attached is a patch with all those things done. I added as well an
assertion in gistKeyIsEQ checking if the input datums are NULL. I
believe that this is still useful for developers, feel free to rip it
out from the patch if you think otherwise.
Regards,
--
Michael

On 01/28/2015 02:47 AM, Michael Paquier wrote:
> On Wed, Jan 28, 2015 at 3:15 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> So I'm fine with taking out both this documentation text and the dead
>> null-pointer checks; but let's do that all in one patch not piecemeal.
>> In any case, this is just cosmetic cleanup; there's no actual hazard
>> here.
> Attached is a patch with all those things done.

Thanks, applied.

> I added as well an assertion in gistKeyIsEQ checking if the input
> datums are NULL. I believe that this is still useful for developers,
> feel free to rip it out from the patch if you think otherwise.

I ripped it out because I think was wrong. It assumed that the input 
Datums are pass-by-reference, which is not a given. It looks that's true 
for all the current opclasses, so I wouldn't be surprised if there are 
hidden assumptions on that elsewhere in the code, but it was wrong 
nevertheless.

- Heikki