Exploring Qualcomm's TrustZone implementation

In this blog post, we'll be exploring Qualcomm's TrustZone implementation, as present on Snapdragon SoCs. If you haven't already, you might want to read the previous blog post, in which I go into some detail about TrustZone in general.

Where do we start?

First of all, since Qualcomm's TrustZone implementation is closed-source, and as far as I could tell, there are no public documents detailing its architecture or design, we will probably need to reverse-engineer the binary containing the TrustZone code, and analyse it.

Acquiring the TrustZone image

We can attempt to extract the image from two different locations; either from the device itself, or from a factory image of the device.

My personal Nexus 5 device was already rooted, so extracting the image from the device should be pretty straight forward. Since the image is stored on the eMMC chip, and the blocks and partitions of the eMMC chip are available under "/dev/block/platform/msm_sdcc.1", I could simply copy the relevant partition to my desktop (using "dd").

Moreover, the partitions have meaningfully named links to them under "/dev/block/platform/msm_sdcc.1/by-name":

As you can see, there are two partitions here, one named "tz" (short for TrustZone), and one named "tzb", which serves as a backup image to the "tz" image, and is identical to it.

However, having extracted the image this way, I was still rather unsatisfied, for two reasons:
  • Although the TrustZone image is stored on the eMMC chip, it could easily be made inaccessible to the "Normal World" (by requiring the AxPROT bit on the system bus to be set), or several parts of it could be missing.
  • Pulling the entire partition's data doesn't reveal information about the real (logical) boundary of the image, so it will require some extra work to determine where the image actually ends. (Actually, since the "tz" image is an ELF binary, its size is contained within the ELF header, but that's just a fluke on our part).
So, having extracted one image from the device, let's take a look at a factory image.

The Nexus 5's factory images are all available to download from Google. The factory image contains a ZIP with all the default images, and additionally contains the bootloader image.

After downloading the factory image and grepping for strings related to TrustZone, it quickly became apparent that the bootloader image contains the wanted code.

However, there was still a minor problem to solve here - the bootloader image was in an unknown format (although maybe some Google-fu could reveal the answers needed). Regardless, opening the file with a hex-editor and guessing at its structure revealed that the format is actually quite simple:


The bootloader file has the following structure:
  • Magic value ("BOOTLDR!") - 8 bytes
  • The number of images - 4 bytes
  • The offset from the beginning of the file to the beginning of the image's data - 4 bytes
  • The total size of the data contained in the images - 4 bytes
  • An array with a number of entries matching the "number of images" field, above. Each entry in the array has two fields:
    • The image name - 64 bytes (zero padded)
    • The image length - 4 bytes
As you can see in the image above, the bootloader image contains an image called "tz", which is the image we're after. In order to unpack this file, I've written a small python script (available here) which receives a bootloader image and unpacks all of the files contained within it.

After extracting the image, and comparing it to the one extracted previously from the device, I verified that they were indeed identical. So I guess this means we can now move on to examine the TrustZone image.

Fixing up the TrustZone image

First of all, examining the file reveals that it is in fact an ELF file, which is pretty good news! This means that the memory segments and their mapped addresses should be available to us.

After opening the file with IDA Pro and letting the auto-analysis to run for a while, I wanted to start reversing the file. However, surprisingly, there seemed to be a lot of branches to unmapped addresses (or rather, addresses that weren't contained within the "tz" binary).

After taking a closer look, it seemed as though all the absolute branches that pointed to invalid addresses were within the first code segment of the file, and they were pointing into high addresses that weren't mapped. Also, there were no absolute branches to the address of that first code segment.

This seemed a little fishy... So how about we take a look at the ELF file's structure? Executing readelf reveals the following:

There's a NULL segment mapped to a higher address, which actually corresponds with the address range to which the invalid absolute branches were pointing! The guys over at Qualcomm are sneaky pandas :)

Anyway, I made a rather safe guess, which is that the first code segment is in fact mapped to the wrong address, and should actually be mapped to the higher address - 0xFE840000. So naturally, I wanted to rebase the segment using IDA's rebase feature, but lo and behold! This causes IDA to crash spectacularly:

I'm actually not sure if this was intended as an anti-reversing feature by Qualcomm, or if the NULL segment is just a result of their internal build process, but this can be easily bypassed by fixing the ELF file manually. All that's required is to move the NULL segment to an unused address (since it is ignored by IDA anyway), and to move the first code segment from its wrong address (0xFC86000) to the correct address (0xFE840000), like so:

Now, after loading the image in IDA, all the absolute branches are valid! This means we can move on to analyse the image.

Analysing the TrustZone image

First, it should be noted that the TrustZone image is a rather large (285.5 KB) binary file, with quite a small amount of strings, and with no public documentation. Moreover, the TrustZone system is comprised of a full kernel with capabilities such as executing applications, and much more. So... it's not clear where we should start, as reversing the whole binary would probably take far too long.

Since we would like to attack the TrustZone kernel from the application processor, the largest attack surface would probably be the secure monitor calls which enable the "Normal World" to interact with the "Secure World".

It should be noted, of course, that there are other vectors with which we can interact with the TrustZone, such as shared memory or maybe even interrupt handling, but since these pose a much smaller attack-surface, it is probably better to start by analysing the SMC calls.

So how do we find where the TrustZone kernel handles the SMC calls? First of all, let's recall that when executing an SMC call, similarly to the handling of SVC calls (that is, regular system calls in the "Normal World"), the "Secure World" must register the address of the vector to which the processor will jump when such an instruction is encountered.

The "Secure World"'s equivalent is the MVBAR (Monitor Vector Base Address Register), which provides the address of the vector containing the handling functions for the different events which are handled by the processor in "Secure World".

Accessing the MVBAR is done using the MRC/MCR opcodes, with the following operands:

So this means we can simply search for an MCR opcode with the following operands in the TrustZone image, and we should be able to find the "Monitor Vector". Indeed, searching for the opcode in IDA returns the following match:

As you can see, the address of the "start" symbol (which is, by the way, the only exported symbol), is loaded into the MVBAR.

According to the ARM documentation, the "Monitor Vector" has the following structure:

Which means that if we look at the "start" symbol mentioned earlier, we can assign the following names to the addresses in that table:

Now, we can analyse the SMC_VECTOR_HANDLER function. Actually, this function is responsible for quite a few tasks; first, it saves all the state registers and the return address in a predefined address (in the "Secure World"), then, it switches over the stack to a preallocated area (also in the "Secure World"). Finally, after performing the necessary preparations, it goes on to analyse the operation requested by the user and operate according to it.

Since the code to issue SMCs is present in the Qualcomm's MSM branch of the Linux kernel, we can take a look at the format of commands which the "Normal World" can issue to the "Secure World".


Confusingly, Qualcomm chose to name the channel through which the "Normal World" interacts with the "Secure World" via SMC opcodes - SCM (Secure Channel Manager).

Anyway, as I've mentioned in the previous blog post, the "qseecom" driver is used to communicate with the "Secure World" using SCMs.

The documentation provided by Qualcomm in the relevant source files is quite extensive, and is enough to get quite a good grip on the format of SCM commands.

Putting it shortly, SCM commands fall into one of two categories:

Regular SCM Call - These calls are used when there is information that needs to be passed from the "Normal World" to the "Secure World", which is needed in order to service the SCM call. The kernel populates the following structure:

And the TrustZone kernel, after servicing the SCM call, writes the response back to the "scm_response" structure:

In order to allocate and fill these structures, the kernel may call the wrapping function "scm_call", which receives pointers to kernel-space buffers containing the data to be sent, the location to which the data should be returned, and most importantly, the service identifier and command identifier.

Each SCM call has a "category", which means which TrustZone kernel subsystem is responsible for handling that call. This is denoted by the service identifier. The command identifier is the code which specifies, within a given service, which command was requested.

After the "scm_call" function allocates and populates the "scm_command" and "scm_response" buffers, it calls an internal "__scm_call" function which flushes all the caches (inner and outer caches), and calls the "smc" function.

This last function actually executes the SMC opcode, transferring control to the TrustZone kernel, like so:

Note that R0 is set to 1, R1 is set to point to a local kernel stack address, which is used as a "context ID" for that call, and R2 is set to point to the physical address of the allocated "scm_command" structure.

This "magic" value set in R0 indicates that this is a regular SCM call, using the "scm_command" structure. However, for certain commands where less data is required, it would be rather wasteful to allocate all these data structures for no reason. In order to address this issue, another form of SCM calls was introduced.

Atomic SCM Call - For calls in which the number of arguments is quite low (up to four arguments), there exists an alternate way to request an SCM call.

There are four wrapper functions, "scm_call_atomic_[1-4]", which correspond to the number of arguments requested. These functions can be called in order to directly issue an SMC for an SCM call with the given service and command IDs, and the given arguments.

Here's the code for the "scm_call_atomic1" function:

Where SCM_ATOMIC is defined as:

Note that both the service ID and the command ID are encoded into R0, along with the number of arguments in the call (in this case, 1). This is instead of the previous "magic" value of 1 used for regular SCM calls.

This different value in R0 indicates to the TrustZone kernel that the following SCM call is an atomic call, which means that the arguments will be passed in using R2-R5 (and not using a structure pointed to by R2).

Analysing SCM calls

Now that we understand how SCM calls work, and we've found the handling function in the TrustZone kernel which is used to handle these SCM calls, we can begin disassembling the SCM calls to try and find a vulnerability in one of them.

I'll skip over most of the analysis of the SCM handling function, since most of it is boilerplate handling of user input, etc. However, After switching the stack over to the TrustZone area and saving the original registers with which the call was performed, the handling function goes on to process the service ID and the command ID in order to see which internal handling function should be called.

In order to easily map between the service and command IDs and the relevant handling function, a static list is compiled into the TrustZone image's data segment, and is referenced by the SCM handling function. Here is a short snipped from the list:

As you can see, the list has the following structure:
  • Pointer to the string containing the name of the SCM function
  • "Type" of call
  • Pointer to the handling function
  • Number of arguments
  • Size of each argument (one DWORD for each argument)
  • The Service ID and Command ID, concatenated into a single DWORD - For example, the "tz_blow_sw_fuse" function above, has the type 0x2002 which means it belongs to the service ID 0x20 and its command ID is 0x02.
Now all that's left is to start disassembling each of these functions, and hope to find an exploitable bug.

The Bug!

So after pouring over all of the aforementioned SMC calls (all 69 of them), I finally arrived at the following function:

Normally, when an SCM command is called using the regular SCM call mechanism, R0 will contain the "result address" which points to the "scm_response" buffer which was allocated by the kernel, but which is also validated by the TrustZone kernel to make sure it is actually a physical address within an "allowed" range - that is, a physical address which corresponds to the Linux kernel's memory, and not, for example, a memory location within the TrustZone binary.

This check is performed using an internal function which I will cover in more detail in the next blog post (so keep posted!).

But what happens if we use an atomic SCM call to execute a function? In that case, the "result address" used is the first argument passed by the atomic call.

Now - can you see the bug in the function above?

As opposed to other SCM handling functions, this function fails to validate the value in R0, the "result address", so if we pass in:
  • R1 as a non-zero value (in order to pass the first branch)
  • The fourth argument (which is passed in at var_1C above) is non-zero
  • R0 as any physical address, including an address within the range of the TrustZone address space
The function will reach the left-most branch in the function above, and write a zero DWORD at the address contained in R0.

Responsible Disclosure

I'd like to point out that I've responsibly disclosed this vulnerability to Qualcomm eleven months ago, and the issue has been fixed by them (amazingly fast!). I'll share a detailed timeline and explanation in the next blog post, but I'd like to point out that the people at Qualcomm have been very responsive and a pleasure to work with.

What's next?

In the next blog post I will share a detailed (and quite complex!) exploit for the vulnerability described above, which enables full code execution within the TrustZone kernel. I will also publish the full exploit code, so stay tuned!

Also, since this is only my second blog post, I'm really looking for some (any) input, specifically:
  • What should I write more (or less) about?
  • Blog design issues
  • Research ideas :)


  1. Really interesting, can't wait for more blog posts!

    1. Thank you very much! The next post is already written and will be published within a day :)

  2. impressive work! well done

  3. Just wondering if var_1C should be non-zero (instead of zero) to reach the left most branch at CBZ RO, loc_FE84B372?

    1. Thank you for noticing, you're absolutely right! My bad, I'll fix it right away.

    2. Thanks for your great job

      About tzbsp_es_is_activated, you say "The fourth argument [...] var_1C". But from what I gather tzbsp_es_is_activated does only take 2 arguments (*cmd, len), var_1C would be a local variable that is set by sub_FE814CE6 (passed by address in r3), some kind of boolean.

    3. Also to pass the first branch you need r0 != 0, not r1 != 0. The Z flag is set by MOVS r6, r0. (MOV without Z doesn't set the flag)

  4. Nice blog...waiting for the exploit..Loved the writing style too

  5. Very nice write up! Thanks for posting it.

    In trying to duplicate what you did with my own device, an S4, I dumped the tz image from its partition but it is definitely not an ELF image. I see trustzone type strings in the image so I know it's the right partition. Any thoughts on how to proceed with mapping code and data segments?



    1. Sure!

      I've done something similar with the Moto X's TZ image (which is within a binary called motoboot.img, also not an ELF file) some time ago.

      Generally the bootloader loads the TZ binary, which is why the TZ image can appear in a different format...

      But IIRC, the TZ image on the S4 is in the MBN file format, right? If so, there's an IDA plugin for MBN files here: https://github.com/daxgr/qcom-mbn-ida-loader, and the format of the file is explained here: http://forum.xda-developers.com/showthread.php?t=2641245

      Let me know if you need more help!

    2. Thanks for the info! Yes the S4 is MBN, That helped a bunch.


  6. Thank you for this nice article. It was very educational. Hope to read more from you in the future.

    Thank you.

  7. I need to know the frame format of Trust zone in snapdragon .how its works can u please give the details how to compile and get the binary images and required apps in userspace

  8. Seems my post a few minutes ago did not upload for some reason; this is my second attempt. First of all thank you for a detailed and useful blog. One suggestion I have is to perhaps try to connect the APIs, and code you refer to above to TZ functionality. That is, you mention SMC call above, and the cmd_id and svc_id. This leaves me wondering how one might leverage these commands in creating credentials for their application. How they can create hashes for instance, and how those hashes can be validated within the TZ. Is there a list of SMC command IDs and the associated numbers?

  9. This comment has been removed by a blog administrator.

  10. Very very nice post!
    I think 'FE810000'(instead of FE810004) is the first address of Monitor Vector. In order to pass the first branch, 'MOVS R6, R0'(instead of MOV R5, R1) can modify the CPSR flag, so the first argument should be non-zero value.

    1. You're right - I wrote this post a year after the original research, and made a few such mistakes in the post. Hopefully I will get around to fixing this when I get a little free time, but if not, my apologies!

  11. Hi, I came across your post while trying to evaluate security of hardware-backed key stores in Android.

    You describe extracting TZ image from eMMC. Is the internal state of the "Secure World" also commonly retained in eMMC? Given sufficient access to the device (e.g. JTAG), should it be possible to extract eMMC, attempt device unlock (e.g. guess PIN), then restore internal state (prior to unlock attempt) and try again?

    1. Hi Nadav,

      I'm not 100% I understand what you mean; the eMMC only contains the code for the TrustZone image, not the actual state - that's stored in RAM. If you had full JTAG access you could halt the application processor in the TrustZone kernel, and try to perform a RAM dump (and restore it later). However - there are many reasons why I suspect this won't work - for one, there are physical XPUs on the SoC which prevent access to different memory locations, based on the processor requesting the access. Depending on which processor you get JTAG access to, you could read certain such parts. Then, even if you were able to restore the state, it is very highly dependant on the state of other subsystems on the SoC (such as the modem, etc.), so simply restoring a huge portion of memory will probably cause a lot of unpredictable behaviour.

    2. Thank you for answering,

      What I meant by internal-state was persistent state, i.e. any state written by the "Secure World" to persistent media during operation. Is all such state commonly persisted to eMMC, or should we assume additional and more elusive storage locations?

      To restate the original question: during incorrect device unlock attempt, what storage locations are generally affected?

    3. Oh I see what you mean -

      Generally, for information persisted by the TZ kernel, this could be a *lot* harder than you would initially assume. First of all, the eMMC is aware of the current world executing (via the AXPROT bus), and therefore disallows access to any location on the eMMC which is used by the secure world to store data. But not only that - there is a secure file-system which is encrypted using a crypto processor on the SoC. I think it uses a hardware key which is unique per-device, and shouldn't be readable in any way using software. This means that even if you de-solder the eMMC and access it fully, you still wouldn't be able to read the data. There's some information in a patent filed by Qualcomm (http://www.google.com/patents/US7921303), but I'm sure if you google SFS you can find out more.

      Ultimately, I would say this: reverse the code responsible for the unlocking attempt and just see where the state is saved. It could be that you're lucky and the data is simple to access, by I would guess otherwise.

    4. What I had in mind is a bit more straightforward:
      1) extract full eMMC, including all partitions — encrypted and otherwise, without attempting decryption
      2) boot the partially disassembled device and attempt a few PIN code guesses (assume device has a 4-digit PIN), stopping short of lockout
      3) restore previously extracted eMMC state via direct write
      4) repeat from step #2 until we guess the PIN code correctly

      Would you expect full read/write operations as stated above to succeed, provided we are able (hypothetically) to repeatedly de-solder and re-solder the eMMC with no damage?

    5. Oh I see what you mean - I assumed earlier you were trying to restore the state on a "live" device.

      I think if the eMMC remains fully intact between attempts, this should work. AFAIK, there is no chip on the SoC which has an internal memory\battery and is used to save the unlocking attempts done so far (like Apple's "Secure Enclave").

      However, IIRC the unlocking itself is designed to be computationally intensive, which means each attempt takes ~150 ms for the application processor to perform. This means that you would be limited to the search space * 150 ms. For small PINs (e.g., 4 digits) this is fast enough. For longer ones (passwords or unlock patterns), the search space could be much bigger and it would be infeasible to wait that amount of time.

  12. Cool post! Bud I'm not really understand how to copy image by "dd" command, could you please show me the whole command?

    1. If you don't know how to dd a img to your desired location then you are WAY over your head with this topic were talking about. Not trying to be a dick, just don't want you bricking a device. Use some google-fu and you'll find out how to dd.

  13. Hi,

    Im having an issue after exploiting the tz on my phone (MSM8916). Im not able to read/access the modem ram. It should be mapped somewhere at 0xC0XXXXXX but when i attempt to read the region simply everything hangs. Also the display stops working when exploiting the tz. It still unclear why. Everything else seems to work fine thru, can read tz ram, kernel etc

    1. Hi Raducu,

      You cannot freely read certain regions such as the modem's RAM, since they are protected by a hardware security mechanism (called an XPU). It may be possible to configure XPUs using TrustZone kernel code-execution, though I've never attempted this.

      All the best,

  14. Hi! I'm a vulnerability researcher who recently decided to break into the world of android (for fun, not work). First, I want to say amazing work! What I love most about this field is the clever and ingenious exploitation techniques used to get code exec. You did not disappoint! As I am waiting for my android device to be shipped, I have been reading your blog.

    In this article, you post a picture of what appears to be file-system listings as root. Were these taken from the phone? Is it possible to have a serial terminal with the phone via USB? Or is this some kind of app / software that allows you to access the phone like a linux terminal?

    Thank you and great job!

    1. Hi Mizz,

      Thank you for the kind words! Yes, you can communicate with the phone over USB using the Android Debug Bridge (ADB). In certain configurations, you can also run the debug daemon under root.


  15. Hi! Great post!

    I'm trying to hack the bootloader of my device(which has snapDragon 820), and to make it load the linux kernel at Exception Level 2(that is, hypervisor/virtualization level). Do you have any idea how this can be done?

    I have diseemblied the xbl.elf file where I found lots of *_ELx codes, which is totally a mess. I would really appreciate it if you could give some comments.

    1. Hi Zesen, sorry for the late response -- I haven't been checking the blog lately.

      First of all, you need to consider the fact that Qualcomm's Hypervisor is already running at EL2. However, if you'd like to run your own code in their hypervisor regardless, I suggest taking one of two routes:

      1. Find an exploit a vulnerability in the Hypervisor. For example, you could use the DMA attack that I covered on the Project Zero blog (https://googleprojectzero.blogspot.co.uk/2017/04/over-air-exploiting-broadcoms-wi-fi_11.html) in order to read/write directly into the Hypervisor (allowing you to freely inject code).

      2. You could look for a way to disable secure boot on the device so that you could provide your own unsigned (or self-signed) image for the hypervisor. Some devices don't have secure boot configured in the first place, though that should be quite rare. In other cases, blowing certain fuses could allow loading unsigned images.

      All the best!

  16. Hi Gal.

    Excellent work.

    Just one small question: you imply that for arguments with pointer to addresses passed by atomic SCM calls, the TZ kernel does not verify whether they are in the correct range (e.g., not in the TZ kernel). However, you mention for regular SCM calls, the TZ kernel does perform a verification. Is my understanding correct? if yes why such an implememtation?

    Thank you very much for the education you provide

  17. Hello there,
    Lg h820 at & t phone was dead boot.
    Tot file available, how do we fix this phone with this method?

  18. This is Very very nice article. Everyone should read. Thanks for sharing. Don't miss WORLD'S BEST CarGamesDownload

  19. i am unable to extract keys from my nexus 5 device can you please help me out??

  20. This is such an awesome asset, to the point that you are giving and you give it away for nothing. I adore seeing blog that comprehend the benefit of giving a quality asset to free.jogos online 2019
    play Games friv
    school friv

  21. I visit your web page. It is really useful and easy to understand. Hope everyone get benefit. Thanks for sharing your Knowledge and experience with us.
    Norton setup - Get started with Norton by downloading the setup and installing it on the device. Enter the unique 25-character alphanumeric product key for activation. Check your subscription norton.com/setup | norton.com/setup

  22. Norton Setup-To get protection, access your Norton account and download the Norton product on your device. Installing the Norton software and activate with 25-digit alphanumeric product key at Official platform of Norton Setup.
    norton.com/setup |norton.com/setup

  23. I really happy found this website eventually. Really informative and inoperative, Thanks for the post and effort! Please keep sharing more such


  24. This comment has been removed by the author.