Advisory: CVE-2014-9707: GoAhead Web Server 3.0.0 - 3.4.1

Affected software: GoAhead Web Server
Affected versions: 3.0.0 - 3.4.1 (3.x.x series before 3.4.2)
CVE ID: CVE-2014-9707

Description: The server incorrectly normalizes HTTP request URIs that
contain path segments that start with a "." but are not entirely equal
to "." or "…" (eg. ".x"). By sending a request with a URI that
contains these incorrectly handled segments, it is possible for remote
attackers to cause a heap overflow with attacker-controlled content or
perform a directory traversal attack.

Fixed version: 3.4.2
Bug entry: https://github.com/embedthis/goahead/issues/106
Fix: https://github.com/embedthis/goahead/commit/eed4a7d177bf94a54c7b06ccce88507fbd76fb77
Reported by: Matthew Daley

Detail:

The vulnerability lies in the websNormalizeUriPath function. This
function correctly handles the normalization of URIs consisting of
normal segments as well as "." and "…" segments, but fails to handle
other segments that start with a '.' character.

A quick runthrough of the important parts of this function:

The function starts by splitting up the URI into segments (at forward
slashes) into an array. At the same time, it calculates the total
length of these segments.

The function then iterates through the resulting array in order to
perform an in-place normalization (both the input and output pointers
point to the same array):

• If a given segment does not start with a '.', it is simply copied from the
  current input pointer to the current output pointer. The for loop's
  increment code will then advance both the input and output pointers.

• Otherwise, if the segment is "." or "…", the input and output pointers are
  adjusted appropriately (taking into account the for loop's increment code)
  but (correctly) no segment is copied.

• Otherwise the segment starts with a '.' but is not "." nor "…"; in this
  case the function incorrectly does nothing and both the input and output
  pointers are simply advanced by the for loop's increment code. This
  effectively skips over a segment in the segment array without any
  modification by the function.

After this iteration has completed, a string buffer for the final
output is allocated. The size used for this allocation comes from the
previously-calculated total segment length, with the addition of space
for forward slashes to join the segments back together again and a
null terminator. The segments in the array up to the final output
pointer are joined together in this buffer with forward slashes
separating them.

There are two ways to exploit this incorrect handling of certain segments:

1) Heap overflow

The heap overflow exploitation lies in the possibility to create a
disconnect between the lengths of the segments left in the segment
array after the iteration has completed and the previously-calculated
total segment length. The previously-calculated length should, in
theory, be the worst-case (longest) final output string buffer size
required (when all segments are left and none are removed by the
normalization iteration). However, since we can force the iteration to
skip over certain segments in the array, it is possible to effectively
duplicate segments in the resulting array; this is done by having the
segment copied from one location to another but then also having the
original copy skipped over, making it appear in the resulting array
twice. When this is done, the previously-calculated length is no
longer long enough for the final output's string buffer, and a heap
overflow occurs while joining together the final result.

As an example, take the following URI as input to the function:
"/./AAAAAAAA/.x".

The URI is first split into the segments "", ".", "AAAAAAAA" and ".",
with the total segment length calculated as 0 + 1 + 8 + 2 = 11 bytes.

The normalization iteration proceeds as follows:

• The "" segment is simply copied from input to output, and hence remains
  unchanged. Both the input and output pointers are then advanced.

• The "." segment causes the output pointer to stay in place while the input
  pointer advances forward.

• The "AAAAAAAA" segment is simply copied from input to output, and hence
  overwrites the previous "." segment. Both the input and output pointers are
  then advanced.

• Finally, the ".x" segment is incorrectly handled: no modification of
  segments is performed but both the input and output pointers are still
  advanced, moving the output pointer over the original "AAAAAAAA" segment.

Hence, the resulting segments in the array that are left up to the
final output pointer are "", "AAAAAAAA" and "AAAAAAAA". Note that the
"AAAAAAAA" segment has been duplicated. These segments, including
space for forward slashes to join them together with and a null
terminator, have a total length of 0 + 8 + 8 + 2 + 1 = 19 bytes.

A string buffer is then allocated for the final output, which uses the
previously-calculated total segment length of 11 bytes plus 3 bytes
for forward slashes and 1 byte for a null terminator, giving a total
size of 11 + 3 + 1 = 15 bytes.

The resulting segments are finally joined together into this final
output string buffer. In doing so in this case, however, the buffer is
overflowed by 19 - 15 = 4 bytes.

So, a remote attacker can make (ie.) a simple HTTP GET request for the
URI in question and cause a heap overflow. ASAN gives the following
output in this case, which shows the exact moment that the heap
overflow occurs:

=================================================================
==2613==ERROR: AddressSanitizer: heap-buffer-overflow on address
0x60200000d47f at pc 0x7ffff6f34020 bp 0x7fffffffd410 sp
0x7fffffffcbd0
WRITE of size 9 at 0x60200000d47f thread T0
#0 0x7ffff6f3401f in __interceptor_strcpy
(/usr/lib/x86_64-linux-gnu/libasan.so.1+0x2f01f)
#1 0x7ffff63a7d6d in websNormalizeUriPath src/http.c:3320
#2 0x7ffff639b4de in parseFirstLine src/http.c:969
#3 0x7ffff639a905 in parseIncoming src/http.c:880
#4 0x7ffff639a4c9 in websPump src/http.c:829
#5 0x7ffff639a19c in readEvent src/http.c:802
#6 0x7ffff6399de7 in socketEvent src/http.c:740
#7 0x7ffff6399cbc in websAccept src/http.c:719
#8 0x7ffff63ac8ed in socketAccept src/socket.c:327
#9 0x7ffff63ade95 in socketDoEvent src/socket.c:638
#10 0x7ffff63add5f in socketProcess src/socket.c:622
#11 0x7ffff639daf8 in websServiceEvents src/http.c:1307
#12 0x401b5c in main src/goahead.c:153
#13 0x7ffff597ab44 in __libc_start_main
(/lib/x86_64-linux-gnu/libc.so.6+0x21b44)
#14 0x4011d8
(/home/matthew/goahead-3.4.1/build/linux-x64-debug/bin/goahead+0x4011d8)

0x60200000d47f is located 0 bytes to the right of 15-byte region
[0x60200000d470,0x60200000d47f)
allocated by thread T0 here:
#0 0x7ffff6f5973f in malloc (/usr/lib/x86_64-linux-gnu/libasan.so.1+0x5473f)
#1 0x7ffff63a7d04 in websNormalizeUriPath src/http.c:3318
#2 0x7ffff639b4de in parseFirstLine src/http.c:969
#3 0x7ffff639a905 in parseIncoming src/http.c:880
#4 0x7ffff639a4c9 in websPump src/http.c:829
#5 0x7ffff639a19c in readEvent src/http.c:802
#6 0x7ffff6399de7 in socketEvent src/http.c:740
#7 0x7ffff6399cbc in websAccept src/http.c:719
#8 0x7ffff63ac8ed in socketAccept src/socket.c:327
#9 0x7ffff63ade95 in socketDoEvent src/socket.c:638
#10 0x7ffff63add5f in socketProcess src/socket.c:622
#11 0x7ffff639daf8 in websServiceEvents src/http.c:1307
#12 0x401b5c in main src/goahead.c:153
#13 0x7ffff597ab44 in __libc_start_main
(/lib/x86_64-linux-gnu/libc.so.6+0x21b44)
(… snip …)

As with all heap overflows, it's likely that this can then go on to be
exploited in order to gain full remote code execution, especially in
embedded systems which are less likely to have heap allocators with
modern hardening techniques.

2) Directory traversal

The directory traversal exploitation lies in the fact that we can
force the normalization iteration to skip over certain segments in the
array; namely, we can force it to skip over a "…" segment. The "…"
segment will pass through unchanged into the final output string
buffer, where it is treated by the rest of the server as an actual
parent-directory relative segment.

As an example, take the following URI as input to the function:
"/…/…/…/…/…/.x/.x/.x/.x/.x/.x/etc/passwd".

The URI is first split into the segments "", "…", "…", "…", "…",
"…", ".x", ".x", ".x", ".x", ".x", ".x", "etc", and "passwd". (The
total segment length that is calculated during this operation is
irrelevant for this mode of exploitation.)

When the normalization iteration reaches the ".x" segments, the
contents of the segment array are still untouched (as all the previous
segments are either empty or are "…") and the output pointer is still
pointing back at the "" segment. The incorrect handling of the ".x"
segments only causes the output (and input) pointers to be advanced
forward over the "" and "…" segments.

When the iteration reaches the "etc" segment, all the "" and "…"
segments have been skipped over; the output pointer is now pointing at
the first ".x" segment. The "etc" is copied over the first ".x"
segment, and the "passwd" segment is copied over the second ".x"
segment.

Hence, the resulting segments in the array that are left up to the
final output pointer are "", "…", "…", "…", "…", "…", "etc" and
"passwd"; note that the "…" segments are still present.

The final output string buffer is created and the resulting segments
are joined together to give a string of "/…/…/…/…/…/etc/passwd".

The rest of the server is expecting that the result from the function
is normalized and that it contains no relative segments. Hence, the
"…" segments go unnoticed when opening the content file while
handling the HTTP request. The end result is that the local filesystem
is traversed up from the administrator-configured web root until
reaching the filesystem's root directory and back down again into the
"/etc/passwd" file. Hence, the file "/etc/passwd" is given in response
to the HTTP request, regardless of the configured web root.

So, a remote attacker can make (ie.) a simple HTTP GET request for the
URI in question and get the contents of the "/etc/passwd" file:

$ echo -ne 'GET /…/…/…/…/…/.x/.x/.x/.x/.x/.x/etc/passwd
HTTP/1.0\r\n\r\n' | nc localhost 4700
HTTP/1.0 200 OK
Server: GoAhead-http
Date: Sun Nov 16 17:21:01 2014
Content-Length: 1346
Connection: close
Last-Modified: Sat Oct 25 17:07:25 2014

root:x:0:0:root:/root:/bin/bash
daemon:x:1:1:daemon:/usr/sbin:/usr/sbin/nologin
bin:x:2:2:bin:/bin:/usr/sbin/nologin
sys:x:3:3:sys:/dev:/usr/sbin/nologin
sync:x:4:65534:sync:/bin:/bin/sync
games:x:5:60:games:/usr/games:/usr/sbin/nologin
man:x:6:12:man:/var/cache/man:/usr/sbin/nologin
lp:x:7:7:lp:/var/spool/lpd:/usr/sbin/nologin
mail:x:8:8:mail:/var/mail:/usr/sbin/nologin
(… snip …)

Of course, 5 "…" segments may not be enough to reach the filesystem's
root directory in all cases and so the crafted URI may have to be
extended with more "…" and ".x" segments.

• Matthew Daley