Heikki Linnakangas <heikki@enterprisedb.com> writes:
> Tom Lane wrote:
>> I can reproduce an out-of-memory condition (basically, replace() is
>> going into an infinite loop because of the invalid input) but I'm
>> not seeing any crash.
> replace_text reads past the end of source string, byte by byte (or
> character by character, not sure), and eventually tries to read from an
> invalid address which causes a segfault. It happens here when start_posn
> == 367368.
Hm, must be memory-layout-dependent. On mine, the output string buffer
is growing fast enough to ensure there's still RAM to read, up till the
kernel says no more.
Anyway the problem is that pg_utf2wchar_with_len silently drops the
trailing incomplete character in its input, causing text_position_next
to think the pattern is empty, causing an infinite loop because
curr_posn never advances. replace_text already tried to guard against
empty pattern, but it doesn't know about this case.
What I intend to do to fix this is to modify the users of
text_position_next to believe the string lengths saved by
text_position_setup, rather than using TEXTLEN() to compute
the lengths. This will effectively make replace_text and
friends consistently act as though the partial character isn't there.
In the long run it might be better to make pg_utf2wchar_with_len throw
an error for bad input, but I'm quite unsure of the consequences of
that, in view of the existing comment "not ours to throw error".
Anyway such a potentially-significant behavioral change doesn't seem
like a good idea to back-patch. (We seem to have this bug in one form
or another clear back to 7.3...)
regards, tom lane