Fix GH-16878: gmp_fact overflow on memory allocation attempt. #16880
Conversation
ext/gmp/gmp.c
Outdated
| RETURN_THROWS(); | ||
| } | ||
| } else { | ||
| mpz_ptr gmpnum; | ||
| gmp_temp_t temp_a; | ||
|
|
||
| FETCH_GMP_ZVAL(gmpnum, a_arg, temp_a, 1); | ||
| long r = mpz_get_si(gmpnum); |
There was a problem hiding this comment.
Please replace the call to zval_get_long(a_arg) below to use the new variable.
There was a problem hiding this comment.
Don't use mpz_get_si() here in the first place, since it may overflow for large gmpnum. Possibly use mpz_sizeinbase(gmpnum, 2) instead; that might need some fine-tuning of maxbits, though.
ext/gmp/gmp.c
Outdated
| RETURN_THROWS(); | ||
| } | ||
| } else { | ||
| mpz_ptr gmpnum; | ||
| gmp_temp_t temp_a; | ||
|
|
||
| FETCH_GMP_ZVAL(gmpnum, a_arg, temp_a, 1); | ||
| long r = mpz_get_si(gmpnum); |
There was a problem hiding this comment.
Don't use mpz_get_si() here in the first place, since it may overflow for large gmpnum. Possibly use mpz_sizeinbase(gmpnum, 2) instead; that might need some fine-tuning of maxbits, though.
ext/gmp/gmp.c
Outdated
| #if SIZEOF_SIZE_T == 4 | ||
| const zend_long maxbits = ULONG_MAX / GMP_NUMB_BITS; | ||
| #else | ||
| const zend_long maxbits = INT_MAX; | ||
| #endif |
There was a problem hiding this comment.
This should be moved to the top-level, since we're already using it in gmp_random_bits(), and may want to re-use it elsewhere. (Could also be a macro; doesn't really matter, I think.)
| #if SIZEOF_SIZE_T == 4 | ||
| #define GMP_ALLOC_MAXBITS (ULONG_MAX / GMP_NUMB_BITS) | ||
| #else | ||
| #define GMP_ALLOC_MAXBITS INT_MAX | ||
| #endif |
There was a problem hiding this comment.
The logic to avoid overflow in _mpz_realloc() (libgmp 6.3.0) is:
if (sizeof (mp_size_t) == sizeof (int))
{
if (UNLIKELY (new_alloc > ULONG_MAX / GMP_NUMB_BITS))
MPZ_OVERFLOW;
}
else
{
if (UNLIKELY (new_alloc > INT_MAX))
MPZ_OVERFLOW;
}Assuming that mp_size_t is actually size_t, the simplification seems to be correct (we're assuming sizeof(int) == 4 anyway). However, I do not understand how ULONG_MAX fits into the mix. That only works under the assumption that ULONG_MAX == UINT_MAX for 32bit platforms (which is what php-src assumes anyway), but why don't they use UINT_MAX in the first place? I think we should.
| if (mpz_sgn(gmpnum) < 0) { | ||
| zend_argument_value_error(1, "must be greater than or equal to 0"); | ||
| (void)gmpnum; | ||
| val = zval_get_long(a_arg); |
There was a problem hiding this comment.
I don't quite understand how zval_get_long() fits into the mix here. At least, its usage makes FETCH_GMP__ZVAL() above superflous, and we wouldn't even need the special casing for Z_TYPE_P(a_arg) == IS_LONG (since this is already catered to by zval_get_long().
What happens if one calls gmp_fact("18446744073709551617")? Wouldn't that raise a deprecation notice ('Implicit conversion from float-string "18446744073709551617" to int loses precision') instead of throwing an ValueError.
There was a problem hiding this comment.
What happens if one calls
gmp_fact("18446744073709551617")? Wouldn't that raise a deprecation notice ('Implicit conversion from float-string "18446744073709551617" to int loses precision') instead of throwing anValueError.
Nope I get the exception in this case too. I ll see the rest later.
There was a problem hiding this comment.
Ah, indeed! That is because the conversion happens in coercive mode. And that also shows that we can't easily drop the seemingly superfluous FETCH_GMP_ZVAL() since only that will report issues regarding strict typing mode (e.g. passing a float). So zval_get_long() is basically correct here; the saturating behavior is fine since we cannot compute the factorial of PHP_INT_MAX anyway.
Still, there is an issue unrelated to this PR
E.g. for gmp_fact(new GMP("18446744073709551617")) zval_get_long() returns 1 which passes the overflow check, and evaluates to GMP(1). That is because the mpz_get_si() in gmp_cast_object() overflows, while I would expect saturating behavior. This is not related to this PR, though, but rather a general issue with ext/gmp. From the mpz_get_si() documentation:
If op fits into a signed long int return the value of op. Otherwise return the least significant part of op, with the same sign as op.
If op is too big to fit in a signed long int, the returned result is probably not very useful. To find out if the value will fit, use the function mpz_fits_slong_p.
So apparently, there is an easy fix for gmp_cast_object(). However, on LLP64 mpir 3.0.0 mpz_fits_slong_p() returns false for 2147483648. ;)
There was a problem hiding this comment.
I don't quite understand how
zval_get_long()fits into the mix here. At least, its usage makesFETCH_GMP__ZVAL()above superflous, and we wouldn't even need the special casing forZ_TYPE_P(a_arg) == IS_LONG(since this is already catered to byzval_get_long().What happens if one calls
gmp_fact("18446744073709551617")? Wouldn't that raise a deprecation notice ('Implicit conversion from float-string "18446744073709551617" to int loses precision') instead of throwing anValueError.
The zval_get_long() is to convert the GMP object back to an int...
No description provided.