Biko's House of Horrors

Shabak Challenge 2021: Paging

This is part of my series of writeups on the Shabak 2021 CTF challenges. See the complete collection here.


The challenge description reads:

After going through the xv6 Memory Management chapter, I decided to challenge myself and implement my own paging mechanism! (In ARM, of course!).

No chance for bugs here, right?

Right, so we're going to be dealing with page tables and ARM code. The referenced xv6 book can be found here1, and should provide the necessary background on paging (I haven't read it, personally 😎). As for the ARM part, I highly recommend the introduction over at Azeria Labs.

With the introductions done, let's get going.

A high-level overview

Looking at the supplied code, there are a lot of moving parts. Before we can get the flag, we'll have to understand how the system works.

The system is an ARMv7-A emulator (as evidenced by the message printed on startup), based on Unicorn. On startup, it reads at most 2048 lines of ARM assembly code, assembles it using Keystone, places it at address 0x4000, then proceeds to execute it. Note that ARMv7-A is a 32-bit system, i.e. pointers are 32-bit in size.

The emulator also provides special services to the emulated code, via the ARM SVC instruction:

  1. Enabling and disabling paging.
  2. Setting the TTBR0 register.
  3. Invoking a "hypercall".
  4. Authentication.
  5. Terminating the emulator.


Looking inside, we can see that we have a two-level paging structure. The physical address of the Page Directory is stored in TTBR0, the PD contains the physical addresses of Page Tables, which in turn point to actual pages with data. Pages are 4kB in size.

Looking at the v_to_p function, and the functions it calls, we can construct a picture of how a virtual address is broken down2:

|          |          |            |
|   PDE    |   PTE    |   Offset   |
|          |          |            |
    10b        10b         12b

That is, the 10 topmost bits are an index into the Page Directory, the next 10 bits are an index into the Page Table, and the final 12 bits are an offset into the page.

What do the entries in a Page Directory/Table look like? The structure is defined in entry.py3:

|                     |        |  |     |        |     |       |
|         PFN         |Reserved|NX|Dirty|Accessed|Write|Present|
|                     |        |  |     |        |     |       |
          20b             7b    1b   1b     1b      1b     1b

The Present bit indicates whether this entry points to a valid table or page. If at any point in the virtual address translation process the emulator encounters an entry with Present == 0, a page fault is generated.

The Write bit indicates whether the target page is writeable. For a page to be writable, all the entries leading up to it must have the Write bit set.

The Accessed bit is set to 1 whenever an entry is accessed during translation.

The Dirty bit is set whenever an entry is accessed during translation, and the target page is being accessed for a write operation.

The NX bit indicates whether the target page is Not eXecutable. For a page to be executable, all the entries leading up to it must have the NX bit set to 0.


The hypercall mechanism provides a lot of functionality, and we'll get back to that later on. For now, we can note the following:

  1. To invoke a hypercall, the code should place the hypercall number in R0, and up to two arguments in R1 and R2 (see hook_intr in
  2. Hypercalls can only be called when paging is enabled (see run_hypercall in
  3. Hypercalls require authentication (see run in, but the authentication service is not implemented (see authenticate in
  4. The hypercall mechanism stores some configuration in the first page of physical memory (see run_hypercall in
    • This configuration is saved to disk when the hypercall mechanism is deactivated.
    • The configuration does not appear to be read from disk on activation, however.

Regarding the hypercall configuration, it has the following format (see _settings_fmt in the Hypercall class in

struct hypercall_config
    uint8_t     groups;
    char        time_activated[19];

        uint8_t group_perm;
    } group_profiles[groups];

Yes, and...

We now understand, in broad strokes, how the system works. However, we are no closer to understanding how we should go about retrieving the flag. The emulator does not read the flag by itself, and it does not expose any functionality for reading files or executing arbitrary commands.

Or does it?

A closer inspection of Hypercall.save_state, the function that saves the hypercall configuration to disk, reveals that it does so in a very peculiar way:

ECHO = 'echo'
STATES_FOLDER = "user_states"

class Hypercall:
    def __init__(self):
        # ...
        self._open_process = os.system
        # ...

    def save_state(self, hypercall_settings, file_name):
        assert ".." not in file_name and "/" not in file_name and "\\" not in file_name
            ECHO +
            f' "{hypercall_settings}"' +
            INTO_FILE +
            ' ' +
            STATES_FOLDER +
            os.path.sep +

In essence, this function invokes:

echo "<hypercall_settings>" > user_states/<file_name>

Can we use this to our advantage? Well, if we could control the filename we could set it to, for instance a;cat flag, which would result in the following command:

echo "<hypercall_settings>" > user_states/a;cat flag

This will save the configuration to a file called user_states/a, and then run cat flag4.

Lucky for us, we can set the filename! In theory. The filename used is the time_activated field in the hypercall configuration, and hypercall no. 14 can be used to set this field to an arbitrary value.

And so we have our plan:

  1. Set up paging.
  2. Enable paging.
  3. Authenticate.
  4. Use hypercall 14 to set the timestamp to a;cat flag.
  5. Deactivate paging to execute the command.

Except... authentication doesn't work.

A bit too far

The way authentication is meant to work (according to the code comments), is by setting self._curr_perm to Hypercall.GROUP_PERM_SUPER (by default the value is Hypercall.GROUP_PERM_USER). When a hypercall is invoked, the implementation goes over all group profiles and if any group has a group_perm field equal to self._curr_perm, then access is granted.

There is no code in the emulator that manipulates self._curr_perm, so we can't go that way. But perhaps we can manipulate the hypercall settings to grant access to Hypercall.GROUP_PERM_USER.

Since the settings are stored in the first physical page, can we just write there? We can write to that page while paging is disabled, however as soon as we enable it that page gets replaced with the default hypercall configuration (see activate in When paging is enabled, writes to the first page raise a page fault (see v_to_p in

Okay, so we'll have to write to the first physical page when paging is enabled, but we can't do it directly. What other flows in the emulator result in writes to physical memory? Perhaps we can abuse some of them.

The only writes to physical memory inside occur in _set_physical_mem and in _write_memory. The first function is called only on the activation of paging, and doesn't look very interesting. The second function, however, gets called from several places:

  1. From set_p_value, which is called when a write operation occurs in the emulated code.
  2. From run_hypercall, to store the updated hypercall settings page.
  3. From _validate_entry, to set the Accessed and Dirty bits in a table entry.

The first case is not interesting, since set_p_value is called only after v_to_p translates a virtual address to a physical one, and this will fail if we reference the first physical page.

The second case is not interesting because throws an exception if a hypercall is invoked without prior authentication, and so Paging.run_hypercall exits without storing anything to the first page.

What about the third case? This results in a write of one or two bits to a table entry. Specifically, bit 2 is always set to 1, and bit 3 is set to 1 if we're accessing a page for writing. And note: there is no check here that we aren't writing to the first physical page! Is this useful?

But wait, what do we actually want to achieve here? We want to allow hypercall access for Hypercall.GROUP_PERM_USER. There are two ways to do this:

  1. Set the group_perm field of the default (and only) group to Hypercall.GROUP_PERM_USER == 0.
  2. Add a new group with a group_perm of 0. This requires increasing the groups field and setting the group_perm of the new group to 0.

The first option is immediately out, since we can only write 1 bits. The second option, however... Since the hypercall page is initially zeroed-out (see pack_default_settings in, "simply" increasing groups will give us a new group with Hypercall.GROUP_PERM_USER!

Conveniently for us, the paging structures are little-endian. If we treat the first physical page as a page table, then the first entry in it will look like this:

|                     |                                        |
|time_activated[0...2]|                groups                  |
|                     |                                        |

|                     |        |  |     |        |     |       |
| Rest of table entry |Reserved|NX|Dirty|Accessed|Write|Present|
|                     |        |  |     |        |     |       |
          24b             3b    1b   1b     1b      1b     1b

This means that merely reading from a page that goes through this "table entry" will change groups from 0b1 == 1 to 0b101 == 5, effectively authenticating us!

Note that the page we're accessing need not actually exist in physical memory. In this case, the paging mechanism will throw an exception, which will be suppressed inside hook_read in

Assemble it yourself

Our plan now looks like this:

  1. Set up paging:
    1. Create a Page Directory somewhere in physical memory.
    2. Map physical address 0x4000 to virtual address 0x4000, so that our code will continue executing after paging is enabled.
    3. Make one of the PDEs in the Page Directory point to the first physical page.
    4. Set TTBR0 to the physical address of the Page Directory.
  2. Enable paging.
  3. Authenticate by reading from the magic page.
  4. Use hypercall 14 to set the timestamp to a;cat flag.
  5. Deactivate paging to execute the command.

Note that the emulator disallows the use of . directives in the assembly code, so we can't use .ascii to bring the command with us. Handling this is left as an exercise for the reader 😎.

Some more "features"

While looking into the paging implementation, I noticed it is possible to extend the amount of physical memory by 4 bytes at a time. In fact, one of my early solution attempts made use of this. See if you can find this "feature" 👾.


  1. This appears to be the last edition of the book that targets the x86 architecture. Beginning with the class of 2019, xv6 was ported to RISC-V. ↩︎

  2. All the ASCII diagrams in this post were made using ASCIIFlow Infinity↩︎

  3. PFN is short for Page Frame Number, i.e. the number of the physical page this entry points to. Conveniently, zeroing out all the bits except the PFN yields the physical address of the page. ↩︎

  4. For more such fun shenanigans, see here↩︎