Обсуждение: BUG #17839: Heap-buffer overflow on float8_to_char with invalid template

The following bug has been logged on the website:

Bug reference:      17839
Logged by:          Thiago Nunes
Email address:      thiagotnunes@gmail.com
PostgreSQL version: 15.2
Operating system:   Linux
Description:

Heap-buffer overflow on float8_to_char when format exceeds max double
digits. I noticed this when running tests with memory sanitiser (msan).

The following example triggers the failure (considering max double digits
`DBL_DIG` is 15):

```
float8_to_char(12345678901, "FM9999999999D999990")
```

Explanation below:

1. After parsing the format, `Num.pre` will be 10, `Num.post` will be 6
`Num.zero_end` will be 16
(https://github.com/postgres/postgres/blob/REL_15_2/src/backend/utils/adt/formatting.c#L1196-L1228)
2. The template size is greater than the `DBL_DIG`, `Num.post` will be moved
back here
(https://github.com/postgres/postgres/blob/REL_15_2/src/backend/utils/adt/formatting.c#L6688-L6689).
3. The shortened template with the max `DBL_DIG` will be "stringfied" out
here
(https://github.com/postgres/postgres/blob/REL_15_2/src/backend/utils/adt/formatting.c#L6712-L6717).
The result will be "##########.####" (10 significant digits + '.' + 4
decimal digits).
4. `Np->last_relevant` will be lesser than `Num->zero_end`, so it is updated
to an invalid position in the result above (pointer + 16) here
(https://github.com/postgres/postgres/blob/REL_15_2/src/backend/utils/adt/formatting.c#L5740-L5743).
5. When applying FILLMODE here
(https://github.com/postgres/postgres/blob/REL_15_2/src/backend/utils/adt/formatting.c#L5563),
it will try to get the character at Np->last_relevant, which is out of
bounds.

PG Bug reporting form <noreply@postgresql.org> writes:
> Heap-buffer overflow on float8_to_char when format exceeds max double
> digits. I noticed this when running tests with memory sanitiser (msan).
> The following example triggers the failure (considering max double digits
> `DBL_DIG` is 15):
> float8_to_char(12345678901, "FM9999999999D999990")

Thanks for the report!  For reasons that aren't entirely clear to me,
valgrind didn't whine about this until I added a few more zeroes to the
format string, like

SELECT to_char('12345678901'::float8, 'FM9999999999D9999900000000000000000');

Nonetheless, it did see the issue at that point.  I'm inclined to fix
it by making the code that adjusts Np->last_relevant more paranoid,
as attached.

            regards, tom lane

diff --git a/src/backend/utils/adt/formatting.c b/src/backend/utils/adt/formatting.c
index f3f4db5ef6..e6246dc44b 100644
--- a/src/backend/utils/adt/formatting.c
+++ b/src/backend/utils/adt/formatting.c
@@ -5695,13 +5695,20 @@ NUM_processor(FormatNode *node, NUMDesc *Num, char *inout,

             /*
              * If any '0' specifiers are present, make sure we don't strip
-             * those digits.
+             * those digits.  But don't advance last_relevant beyond the last
+             * character of the Np->number string, which is a hazard if the
+             * number got shortened due to precision limitations.
              */
             if (Np->last_relevant && Np->Num->zero_end > Np->out_pre_spaces)
             {
+                int            last_zero_pos;
                 char       *last_zero;

-                last_zero = Np->number + (Np->Num->zero_end - Np->out_pre_spaces);
+                /* note that Np->number cannot be zero-length here */
+                last_zero_pos = strlen(Np->number) - 1;
+                last_zero_pos = Min(last_zero_pos,
+                                    Np->Num->zero_end - Np->out_pre_spaces);
+                last_zero = Np->number + last_zero_pos;
                 if (Np->last_relevant < last_zero)
                     Np->last_relevant = last_zero;
             }
diff --git a/src/test/regress/expected/numeric.out b/src/test/regress/expected/numeric.out
index 2ac2c99b19..26930f4db4 100644
--- a/src/test/regress/expected/numeric.out
+++ b/src/test/regress/expected/numeric.out
@@ -1929,6 +1929,12 @@ SELECT to_char('100'::numeric, 'FM999');
  100
 (1 row)

+SELECT to_char('12345678901'::float8, 'FM9999999999D9999900000000000000000');
+     to_char
+-----------------
+ ##########.####
+(1 row)
+
 -- Check parsing of literal text in a format string
 SELECT to_char('100'::numeric, 'foo999');
  to_char
diff --git a/src/test/regress/sql/numeric.sql b/src/test/regress/sql/numeric.sql
index 6e9e6145d0..2dddb58625 100644
--- a/src/test/regress/sql/numeric.sql
+++ b/src/test/regress/sql/numeric.sql
@@ -979,6 +979,7 @@ FROM v;
 SELECT to_char('100'::numeric, 'FM999.9');
 SELECT to_char('100'::numeric, 'FM999.');
 SELECT to_char('100'::numeric, 'FM999');
+SELECT to_char('12345678901'::float8, 'FM9999999999D9999900000000000000000');

 -- Check parsing of literal text in a format string
 SELECT to_char('100'::numeric, 'foo999');

Thiago Nunes <thiagotnunes@gmail.com> writes:
> I think your solution deals with all the cases, but I would like to point
> out how I fixed it locally. I recalculated Num.zero_end after this line (
> https://github.com/postgres/postgres/blob/REL_15_2/src/backend/utils/adt/formatting.c#L6716
> ):

> ```
> Num.zero_end = Num.pre + Num.post;
> ```

Hmm ... that seems a bit ad-hoc, because as far as I understand this
code, zero_end is supposed to track where is the last '0' format
character.  That shouldn't change just because we decided that the
data value overflowed.

            regards, tom lane