Oct. 30th, 2012

It's not difficult to write a UEFI application under Linux. You'll need three things:

  • An installed copy of the gnu-efi library and headers
  • A copy of the UEFI specification from here
  • Something running UEFI (OVMF is a good option if you don't have any UEFI hardware)

So, let's write a trivial UEFI application. Here's one from the gnu-efi source tree:
#include <efi.h>
#include <efilib.h>

EFI_STATUS
efi_main (EFI_HANDLE image, EFI_SYSTEM_TABLE *systab)
{
        SIMPLE_TEXT_OUTPUT_INTERFACE *conout;

        conout = systab->ConOut;
        InitializeLib(image, systab);
        uefi_call_wrapper(conout->OutputString, 2, conout, L"Hello World!\n\r");

        return EFI_SUCCESS;
}
The includes are fairly obvious. efi.h gives you UEFI functions defined by the specification, and efilib.h gives you functions provided by gnu-efi. The runtime will call efi_main() rather than main(), so that's our main function. It returns an EFI_STATUS which will in turn be returned to whatever executed the binary. The EFI_HANDLE image is a pointer to the firmware's context for this application, which is used in certain calls. The EFI_SYSTEM_TABLE *systab is a pointer to the UEFI system table, which in turn points to various other tables.

So far so simple. Now things get interesting. The heart of the UEFI programming model is a set of protocols that provide interfaces. Each of these interfaces is typically a table of function pointers. In this case, we're going to use the simple text protocol to print some text. First, we get a pointer to the simple text output table. This is always referenced from the system table, so we can simply assign it. InitializeLib() just initialises some internal gnu-efi functions - it's not technically required if we're purely calling native UEFI functions as we are here, but it's good style since forgetting it results in weird crashes.

Anyway. We now have conout, a pointer to a SIMPLE_TEXT_OUTPUT_INTERFACE structure. This is described in section 11.4 of the UEFI spec, but looks like this:
typedef struct _EFI_SIMPLE_TEXT_OUTPUT_PROTOCOL {
    EFI_TEXT_RESET Reset;
    EFI_TEXT_STRING OutputString;
    EFI_TEXT_TEST_STRING TestString;
    EFI_TEXT_QUERY_MODE QueryMode;
    EFI_TEXT_SET_MODE SetMode;
    EFI_TEXT_SET_ATTRIBUTE SetAttribute;
    EFI_TEXT_CLEAR_SCREEN ClearScreen;
    EFI_TEXT_SET_CURSOR_POSITION SetCursorPosition;
    EFI_TEXT_ENABLE_CURSOR EnableCursor;
    SIMPLE_TEXT_OUTPUT_MODE *Mode;
} EFI_SIMPLE_TEXT_OUTPUT_PROTOCOL;
The majority of these are functions - the only exception is the SIMPLE_TEXT_OUTPUT_MODE pointer, which points to the current mode. In an ideal world you'd be able to just call these directly, but sadly UEFI and Linux calling conventions are different and we need some thunking. That's where the uefi_call_function() call comes in. This takes a UEFI function as its first argument (in this case, conout->OutputString), the number of arguments (2, in this case) and then the arguments. Checking the spec for OutputString, we see this:
typedef
EFI_STATUS
(EFIAPI *EFI_TEXT_STRING) (
    IN EFI_SIMPLE_TEXT_OUTPUT_PROTOCOL *This,
    IN CHAR16 *String
);
The first argument is a pointer to the specific instance of the simple text output protocol. This permits multiple instances of the same protocol to exist on a single machine, without having to duplicate all of the code. The second argument is simply a UCS-2 string. Our conout pointer is a pointer to the protocol, so we pass that as the first argument. For the second argument, we pass L"Hello World!\n", with the L indicating that this is a wide string rather than a simple 8-bit string. uefi_call_function() then rearranges these arguments into the appropriate calling convention and calls the UEFI function. The firmware then prints "Hello World!" on the screen. Success.

(Warning: uefi_call_function() does no type checking on its arguments, so if you pass in a struct instead of a pointer it'll build without complaint and then give you a bizarre error at runtime)

Finally, we clean up by simply returning EFI_SUCCESS. Nothing more for us to worry about. That's the simple case. What about a more complicated one? Let's do something with that EFI_HANDLE that got passed into efi_main().
#include <efi.h>
#include <efilib.h>

EFI_STATUS
efi_main (EFI_HANDLE image, EFI_SYSTEM_TABLE *systab)
{
        EFI_LOADED_IMAGE *loaded_image = NULL;
        EFI_GUID loaded_image_protocol = LOADED_IMAGE_PROTOCOL;
        EFI_STATUS status;

        InitializeLib(image, systab);
        status = uefi_call_wrapper(systab->BootServices->HandleProtocol,
                                3,
                                image, 
                                &loaded_image_protocol, 
                                (void **) &loaded_image);
        if (EFI_ERROR(status)) {
                Print(L"handleprotocol: %r\n", status);
        }

        Print(L"Image base        : %lx\n", loaded_image->ImageBase);
        Print(L"Image size        : %lx\n", loaded_image->ImageSize);
        Print(L"Image file        : %s\n", DevicePathToStr(loaded_image->FilePath));
        return EFI_SUCCESS;
}
UEFI handles can have multiple protocols attached to them. A handle may represent a piece of physical hardware, or (as in this case) it can represent a software object. In this case we're going to get a pointer to the loaded image protocol that's associated with the binary we're running. To do this we call the HandleProtocol() function from the Boot Services table. Boot services are documented in section 6 of the UEFI specification and are the interface to the majority of UEFI functionality. HandleProtocol takes an image and a protocol identifier, and hands back a pointer to the protocol. UEFI protocol identifiers are all GUIDs and defined in the specification. If you call HandleProtocol on a handle that doesn't implement the protocol you request, you'll simply get an error back in the status argument. No big deal.

The loaded image protocol is fairly different to the simple text output protocol in that it's almost entirely data rather than function pointers. In this case we're printing the address that our executable was loaded at and the size of our relocated executable. Finally, we print the path of the file. UEFI device paths are documented in section 9.2 of the UEFI specification. They're a binary representation of a file path, which may include the path of the device that the file is on. DevicePathToStr is a gnu-efi helper function that converts the objects it understands into a textual representation. It doesn't cover the full set of specified UEFI device types, so in some corner cases it may print "Unknown()".

That covers how to use protocols attached to the image handle. How about protocols that are attached to other handles? In that case we can do something like this:
#include <efi.h>
#include <efilib.h>

EFI_STATUS
efi_main (EFI_HANDLE image_handle, EFI_SYSTEM_TABLE *systab)
{
        EFI_STATUS status;
        EFI_GRAPHICS_OUTPUT_PROTOCOL *gop;
        EFI_GUID gop_guid = EFI_GRAPHICS_OUTPUT_PROTOCOL_GUID;

        InitializeLib(image_handle, systab);

        status = uefi_call_wrapper(systab->BootServices->LocateProtocol,
                                   3,
                                   &gop_guid,
                                   NULL,
                                   &gop);

        Print(L"Framebuffer base is at %lx\n", gop->Mode->FrameBufferBase);

        return EFI_SUCCESS;
}
This will find the first (where first is an arbitrary platform ordering) instance of a protocol and return a pointer to it. If there may be multiple instances of a protocol (as there often are with the graphics output protocol) you should use LocateHandleBuffer() instead. This returns an array of handles that implement the protocol you asked for - you can then use HandleProtocol() to give you the instance on the specific handle. There's various helpers in gnu-efi like LibLocateHandle() that make these boilerplate tasks easier. You can find the full list in /usr/include/efi/efilib.h.

That's a very brief outline of how to use basic UEFI functionality. The spec documents the full set of UEFI functions available to you, including things like direct block access, filesystems and networking. The process is much the same in all cases - you locate the handle that you want to interact with, open the protocol that's installed on it and then make calls using that protocol.

Building these examples

Building these examples is made awkward due to UEFI wanting PE-COFF binaries and Linux toolchains building ELF binaries. You'll want a makefile something like this:
ARCH            = $(shell uname -m | sed s,i[3456789]86,ia32,)
LIB_PATH        = /usr/lib64
EFI_INCLUDE     = /usr/include/efi
EFI_INCLUDES    = -nostdinc -I$(EFI_INCLUDE) -I$(EFI_INCLUDE)/$(ARCH) -I$(EFI_INCLUDE)/protocol

EFI_PATH        = /usr/lib64/gnuefi
EFI_CRT_OBJS    = $(EFI_PATH)/crt0-efi-$(ARCH).o
EFI_LDS         = $(EFI_PATH)/elf_$(ARCH)_efi.lds

CFLAGS          = -fno-stack-protector -fpic -fshort-wchar -mno-red-zone $(EFI_INCLUDES)
ifeq ($(ARCH),x86_64)
        CFLAGS  += -DEFI_FUNCTION_WRAPPER
endif

LDFLAGS         = -nostdlib -znocombreloc -T $(EFI_LDS) -shared -Bsymbolic -L$(EFI_PATH) -L$(LIB_PATH) \
                  $(EFI_CRT_OBJS) -lefi -lgnuefi

TARGET  = test.efi
OBJS    = test.o
SOURCES = test.c

all: $(TARGET)

test.so: $(OBJS)
       $(LD) -o $@ $(LDFLAGS) $^ $(EFI_LIBS)

%.efi: %.so
        objcopy -j .text -j .sdata -j .data \
                -j .dynamic -j .dynsym  -j .rel \
                -j .rela -j .reloc -j .eh_frame \
                --target=efi-app-$(ARCH) $^ $@
There's rather a lot going on here, but the important stuff is the CFLAGS line and everything after that. First, we need to disable the stack protection code - there's nothing in the EFI runtime for it to call into, so we'll build a binary with unresolved symbols otherwise. Second, we need to build position independent code, since UEFI may relocate us anywhere. short-wchar is needed to indicate that strings like L"Hi!" should be 16 bits per character, rather than gcc's default of 32 bits per character. no-red-zone tells the compiler not to assume that there's 256 bytes of stack available to it. Without this, the firmware may modify stack that the binary was depending on.

The linker arguments are less interesting. We simply tell it not to link against the standard libraries, not to merge relocation sections, and to link in the gnu-efi runtime code. Nothing very exciting. What's more interesting is that we build a shared library. The reasoning here is that we want to perform all our linking, but we don't want to build an executable - the Linux executable runtime code would be completely pointless in a UEFI binary. Once we have our .so file, we use objcopy to pull out the various sections and rebuild them into a PE-COFF binary that UEFI will execute.
I've written rather a lot about how we're implementing our Secure Boot support, but there's been a range of subtle changes over time. Here's an overview of how it all fits together.

Secure Boot and signing

Secure Boot is part of the UEFI specification. When enabled, systems will refuse to boot any UEFI executables (such as an OS bootloader) unless they've been signed by a trusted key. The only trusted key effectively guaranteed to be present is Microsoft's, and Microsoft will sign binaries if you have access to the Windows hardware dev center. You gain access to this by purchasing a certificate from Verisign that verifies your identity and then use that to create a developer account at the Sysdev website. Once you've agreed to the legal contracts, you can submit a UEFI executable and Microsoft will then send you back a signed one. The signature is added to the end of the binary, so it's trivial to verify that your executable hasn't otherwise been modified in the process.

Avoiding signing round trips

We didn't want to have to get binaries re-signed every time we updated our bootloader or kernel, so we came up with a compromise approach. We implemented a shim bootloader (cunningly called "Shim") which is signed by Microsoft and includes a copy of a public key that's under our control. Shim will launch any binary signed with either a key installed in the system firmware or the public key built into it. This allows us to build and sign all other binaries ourselves.

Preventing Secure Boot circumvention

Secure Boot is intended to prevent scenarios where untrusted code can be run before an OS kernel. Locking down the boot environment avoids the simple case of using a modified bootloader, and requiring that the kernel also be signed avoids simply booting arbitrary attack code that looks roughly like a kernel but does nothing to prevent the more complex case of using a running kernel to launch another kernel. To do that we've had to restrict the behaviour of the kernel in Secure Boot environments. Direct hardware access from userspace has been blocked, and modules must be signed by a key the kernel trusts.

Providing user control over trust

Microsoft requires that it be possible to modify the key database on all Windows-certified x86 systems, but the UI to do so may be inconsistent and awkward. Suse developed a plan to avoid this, involving adding an additional key database to the system. The user enrols a key from the running OS and enters a password. On the next reboot, shim detects that a key is awaiting enrolment and drops to a prompt. If the user does nothing or just hits enter, the system continues to boot. If the user chooses to enrol a new key, the user is asked to re-enter the password in order to confirm that they're the same user that initiated the original request. If successful, the key is then copied into a boot services variable which cannot be directly modified from the OS. Anything signed with the user's key will then be trusted.

Providing user control over signature verification

Some users (such as kernel developers) may be hampered by Secure Boot - forcing them to sign every new kernel they build is time consuming and unhelpful. Windows-certified x86 hardware will be required to provide an option to disable Secure Boot, but again that UI may be inconsistent. We've added an option where a user can generate a request to disable signature validation, again requiring a password. On next boot the user must explicitly choose the "Toggle signature validation" menu item and will then be requested to enter three randomly chosen characters from their password. After confirmation, signature validation in Shim will then be disabled.

Supporting other distributions

Not all distributions will want (or be able) to sign via Microsoft. They'll be able to take any other signed version of Shim and put it on their boot media. If their second stage bootloader is signed by a key that Shim doesn't trust, Shim will pop up a menu permitting the user to choose to enrol a new key off the install media. The user can select this, navigate a file browser and select the key. After confirmation, the key will be added to Shim's trusted key list. The user can then install the distribution.

Third party modules

There's two approaches here. The first is for the third party to provide a module key that the user can install into Shim's database. The kernel will import this at boot time and trust any modules signed with it. The second is for the third party to provide a key signed by Microsoft. David Howells has come up with with the idea of utilising this approach of embedding a public key in a pe binary, with the kernel then having support for importing this key, noticing that it's signed by a trusted key and adding it to the kernel keyring.

How about ARM?

We're not planning on supporting Secure Boot on ARM, for two reasons. The first is that Microsoft require that it be impossible to disable Secure Boot on ARM, and we don't want to support systems unless it's guaranteed that the user will be able to run what they want. The second is that there's no strong expectation that the signing key used for signing third party UEFI binaries will be present on ARM systems.

What are other distributions doing?

As far as I know, Suse and Fedora will be shipping the same code. Ubuntu is shipping an older version of Shim but should pick up the local key management code in the next release. The only significant difference is that Ubuntu doesn't require that kernel modules be signed.
Shim is the first stage bootloader we've been implementing for supporting Secure Boot. It's called Shim because it serves as an object to fit the gap between the Microsoft trust root and our own trust root. As originally envisaged it would do nothing other than load and execute appropriately signed binaries, but it's got a little more complicated than that now. It is, however, basically feature complete at this point - I don't expect it to grow significantly further.

(For those of you following along at home, the Shim source code I'm talking about is here)

First, it checks whether the user has disabled local signature verification in check_mok_sb(). This is done by reading the MokSBState UEFI variable and verifying that it has the correct variable attributes, ensuring that it was created in the trusted boot services environment. If so, and if it's set appropriately, signature verification is disabled for the rest of Shim. Shim then pauses for a couple of seconds while displaying a message indicating that it's running in insecure mode.

Next, Shim checks whether the user has set any variables that indicate that a Mok request has been made in check_mok_request(). If so, it launches MokManager. The mechanism by which this is done is described later.

Shim then copies the Mok key database into a runtime variable, which makes it accessible to the kernel (mirror_mok_list()). The kernel can then use the Mok keys when performing module signature verification. Next, as the last step before launching the next stage bootloader, Shim installs a UEFI protocol that permits other applications to call back into Shim. We're now ready to branch off from Shim into the actual bootloader.

To do this, Shim has to do a few things. First, we need to get the bootloader off disk. This involves finding Shim's path, which we can do by opening the loaded image protocol on Shim's image handle, turning that into a string, walking backwards along it until we find a directory separator, stripping the filename and appending the path to the second stage bootloader. This is then turned back into a binary device path, all of this happening in generate_path(). The binary is then read off disk in load_image(), using the simple filesystem protocol. We open the device, open a reference to the file, check how large the file is, allocate a buffer to hold the file and then read it off disk into that buffer. Everything so far has been fairly straightforward, but now things get more complicated.

handle_image() is the real meat of Shim. First it has to examine the header data in read_header(), copying the relevant bits into a context structure that will be used later. Some basic sanity checks on the binary are also performed here. If we're running in secure mode (ie, Secure Boot is enabled and we haven't been toggled into insecure mode) we then need to verify that the binary matches the signature and hasn't been blacklisted.

verify_buffer() has to implement the Microsoft Authenticode specification, which details which sections of the binary have to be hashed and in which order. The actual hashing is done in generate_hash(), which calls into the crypto code to add each binary section to the hashed data. The comments in generate_hash() describe what's going on there clearly enough. The crypto code itself is a copy of the crypto library extracted from the Tiano source tree. It's a relatively thin layer on top of OpenSSL, so the code's been fairly well tested.

Once we have the hash, we need to verify the signature. First, we ensure that the Mok database hasn't been modified from the OS (verify_mok()). Next, we check whether the binary's hash or signature have been blacklisted (check_blacklist()). This is done for each of the SHA1 and SHA256 hashes of the binary (the two hashes implemented in Tiano) and the certificate. If any of these match an entry in the dbx variable or a built-in blacklist, Shim will refuse to run them. Next, we simply do the same for the whitelist. If the binary's hash is either in the db variable or the Mok list, or if it's signed with a certificate that chains back to an entry in either db or the Mok list, we'll run it. Finally, the signature is checked against any built-in keys. Certificate verification is carried out with the AuthenticodeVerify() function, which again comes from Tiano's crypto library and is mostly implemented in OpenSSL.

If the binary passed signature validation we return to handle_image() and now load the binary's sections into their desired addresses. That buffer is handed to relocate_coff() which runs through the binary's relocation entries and fixes them up. There's one last subtlety, which is that because we've been fixing this up by hand, the image handle still refers to the shim binary. We fix the ImageBase and ImageSize values in the loaded image protocol to match the newly relocated binary, which is vital because otherwise grub is unable to find its built in modules. Finally, we jump into the binary's entry point. What the binary does next is up to it - if it's a bootloader, there's a good chance that it'll simply launch an OS and we'll never return. If it does return, we restore ImageBase and ImageSize, uninstall the protocol we installed and exit ourselves. If Shim exits (either because the second stage bootloader wasn't present, didn't verify, or exited) then control will simply pass on to the next boot option in the UEFI priority list. If there's nothing else, the platform will do something platform defined.

And that's how Shim works.

Profile

Matthew Garrett

About Matthew

Power management, mobile and firmware developer on Linux. Security developer at Nebula. Member of the Linux Foundation Technical Advisory Board. Ex-biologist. @mjg59 on Twitter. Content here should not be interpreted as the opinion of my employer.

Expand Cut Tags

No cut tags