It turns out a7249a65 was masking a second problem - on some binaries,
when we actually don't have any base relocations at all, binutils'
"objcopy --target efi-app-x86_64" is generating a PE header with a base
relocations pointer that happily points into the middle of our text
section. So with shim processing base relocations correctly, it refuses
to load those binaries.
For example, on one binary I just built:
00000130 00 a0 00 00 0a 00 00 00 00 00 00 00 00 00 00 00 |................|
which says there's a Base Relocation Table at 0xa000 that's 0xa bytes long.
That's here:
0000a000 58 00 29 00 00 00 00 00 48 00 44 00 28 00 50 00 |X.).....H.D.(.P.|
0000a010 61 00 72 00 74 00 25 00 64 00 2c 00 53 00 69 00 |a.r.t.%.d.,.S.i.|
0000a020 67 00 25 00 67 00 29 00 00 00 00 00 00 00 00 00 |g.%.g.).........|
0000a030 48 00 44 00 28 00 50 00 61 00 72 00 74 00 25 00 |H.D.(.P.a.r.t.%.|
So the table is:
0000a000 58 00 29 00 00 00 00 00 48 00 |X.).....H. |
That wouldn't be so bad, except those binaries are MokManager.efi,
fallback.efi, and shim.efi, and sometimes they're .reloc, which we're
actually trying to handle correctly now because grub builds with a real
and valid .reloc table. So though I didn't think there was any hair
left on this yak, more shaving ensues.
With this change, instead of letting objcopy do whatever it likes, we
switch to "-O binary" and merely link in a header that's appropriate for
our binaries. This is the same method Ard wrote for aarch64, and it
seems to work fine in either place (modulo some minor changes.)
At some point this should be merged into gnu-efi instead of carrying our
own crt0-efi-x86_64.S, but that's a less immediate problem.
I did not need this problem.
Signed-off-by: Peter Jones <pjones@redhat.com>
When I merged 4bfb13d and fixed the conflicts, I managed to make the
in_protocol test exactly backwards, so that's why we don't currently see
error messages.
Signed-off-by: Peter Jones <pjones@redhat.com>
Actually check the size of our vendor cert quite early, so that there's
no confusion as to what's going on.
This isn't strictly necessary, in that in all cases if vendor_cert_size
is 0, then AuthenticodeVerify -> Pkcs7Verify() -> d2i_X509() will result
in a NULL "Cert", and it will return FALSE, and we'll reject the
signature, but better to avoid all that code in the first place. Belt
and suspenders and whatnot.
Based on a patch from https://github.com/TBOpen .
Signed-off-by: Peter Jones <pjones@redhat.com>
I screwed one of these up when working on 750584c, and it's a real pain
to figure out, so that means we should be validating them.
Signed-off-by: Peter Jones <pjones@redhat.com>
This is mostly based on a patch (https://github.com/mjg59/shim/issues/30)
from https://github.com/TBOpen , which refactors our __LP64__
tests to be tests of the header magic instead. I've simplified things
by using what we've pre-loaded into "context" and making some helper
functions so the conditionals in most of the code say what they do,
instead of how they work.
Note that we're only allowing that from in_protocol's loader - that is,
we'll let 64-bit grub load a 32-bit kernel or 32-bit grub load a 64-bit
kernel, but 32-bit shim isn't loading a 64-bit grub.
Signed-off-by: Peter Jones <pjones@redhat.com>
Currently when we process base relocations, we get the correct Data
Directory pointer from the headers (context->RelocDir), and that header
has been copied into our pristine allocated image when we copied up to
SizeOfHeaders. But the data it points to has not been mirrored in to
the new image, so it is whatever data AllocPool() gave us.
This patch changes relocate_coff() to refer to the base relocation table
from the image we loaded from disk, but apply the fixups to the new
copy.
I have no idea how x86_64 worked without this, but I can't make aarch64
work without it. I also don't know how Ard or Leif have seen aarch64
work. Maybe they haven't? Leif indicated on irc that they may have
only tested shim with simple "hello world" applications from gnu-efi;
they are certainly much less complex than grub.efi, and are generated
through a different linking process.
My only theory is that we're getting recycled data there pretty reliably
that just makes us /not/ process any relocations, but since our
ImageBase is 0, and I don't think we ever load grub with 0 as its base
virtual address, that doesn't follow. I'm open to any other ideas
anybody has.
I do know that on x86_64 (and presumably aarch64 as well), we don't
actually start seeing *symptoms* of this bug until the first chunk[0] of
94c9a77f is applied[1]. Once that is applied, relocate_coff() starts
seeing zero[2] for both RelocBase->VirtualAddress and
RelocBase->SizeOfBlock, because RelocBase is a (generated, relative)
pointer that only makes sense in the context of the original binary, not
our partial copy. Since RelocBase->SizeOfBlock is tested first,
relocate_base() gives us "Reloc block size is invalid"[3] and returns
EFI_UNSUPPORTED. At that point shim exits with an error.
[0] The second chunk of 94c9a77f patch makes no difference on this
issue.
[1] I don't see why at all.
[2] Which could really be any value since it's AllocatePool() and not
AllocateZeroPool() results, but 0 is all I've observed; I think
AllocatePool() has simply never recycled any memory in my test
cases.
[3] which is silent because perror() tries to avoid talking because that
has caused much crashing in the past; work needs to go in to 0.9 for
this.
Signed-off-by: Peter Jones <pjones@redhat.com>
Since in theory you could, for example, get an x86_64 binary signed that
also behaves as an ARM executable, we should be checking this before
people build on other architectures.
Signed-off-by: Peter Jones <pjones@redhat.com>
On aarch64 due to some terrifying include chain we wind up with
Cryptlib's definition of exit here. I'm not a glutton for punishment,
so I'm just changing the name so it's not coliding.
Signed-off-by: Peter Jones <pjones@redhat.com>
On archs where no EFI aware objcopy is available, the generated PE/COFF
header contains a .reloc section which is completely empty. Handle this by
- returning early from relocate_coff() with EFI_SUCCESS,
- ignoring discardable sections in the section loader.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
We don't need to .data entries; the second one should be .data*. He's
since fixed this in his tree, but I'd already pulled it and pushed to
master.
Signed-off-by: Peter Jones <pjones@redhat.com>
This patch cleans up and refactors the Makefiles to better allow new
architectures to be added:
- remove unused Makefile definitions
- import Makefile definitions from top level rather than redefining
- move x86 specific CFLAGS to inside ifeq() blocks
- remove x86 inline asm
- allow $(FORMAT) to be overridden: this is necessary as there exists no
EFI or PE/COFF aware objcopy for ARM
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Prevent unhook_system_services() from dereferencing a NULL systab, which
may occur if hook_system_services() has never been called.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Upstream GNU-EFI contains changes to efistdarg.h resulting in the va_start,
va_arg and va_end macros to be #defined unconditionally. Make sure we #undef
them before overriding the definitions.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
MokSBState and MokDBState are just 1 byte variables, so a UINT8
local variable is sufficient to include the content.
Signed-off-by: Gary Ching-Pang Lin <glin@suse.com>
Conflicts:
shim.c
There are functions defined in lib to check the secure variables.
Use the functions to shun the duplicate code.
Signed-off-by: Gary Ching-Pang Lin <glin@suse.com>
Conflicts:
shim.c
When grub2 invokes the functions of shim protocol in gfx mode,
OutputString in shim could distort the screen.
Signed-off-by: Gary Ching-Pang Lin <glin@suse.com>
Conflicts:
shim.c
(modified by pjones to include some newer Prints that weren't there when
Gary did the initial work here.)
If ca.crt was added into the certificate database, ca.crt would be the first
certificate in the signature. Because shim couldn't verify ca.crt with the
embedded shim.cer, it failed to load MokManager.efi.signed and
fallback.efi.signed.
Signed-off-by: Gary Ching-Pang Lin <glin@suse.com>
On some machines, even though the key event was signaled, ReadKeyStroke
still got EFI_NOT_READY. This commit handles the error status to avoid
console_get_keystroke from returning unexpected keys.
Signed-off-by: Gary Ching-Pang Lin <glin@suse.com>
Conflicts:
MokManager.c
A non-DER encoding x509 certificate may be mistakenly enrolled into
db or MokList. This commit checks the first 4 bytes of the certificate
to ensure that it's DER encoding.
This commit also removes the iteration of the x509 signature list.
Per UEFI SPEC, each x509 signature list contains only one x509 certificate.
Besides, the size of certificate is incorrect. The size of the header must
be substracted from the signature size.
Signed-off-by: Gary Ching-Pang Lin <glin@suse.com>
The previous strategy is to locate the first available PXE_BASE_CODE
protocol and to fetch the second stage image from it, and this may
cause shim to fetch the wrong second stage image, i.e. grub.efi.
Consider the machine with the following boot order:
1. PXE Boot
2. Hard Drive
Assume that the EFI image, e.g. bootx64.efi, in the PXE server is
broken, then "PXE Boot" will fail and fallback to "Hard Drive". While
shim.efi in "Hard Drive" is loaded, it will find the PXE protocol is
available and fetch grub.efi from the PXE server, not grub.efi in the
disk.
This commit checks the DeviceHandle from Loaded Image. If the device
supports PXE, then shim fetches grub.efi with the PXE protocol. Otherwise,
shim loads grub.efi from the disk.
Signed-off-by: Gary Ching-Pang Lin <glin@suse.com>
Some UEFI implementations never care the boot options, so the
restored boot options could be just ignored and this results in
endless reboot. To avoid this situation, this commit makes
fallback.efi to load the first matched boot option even if there
is no boot option to be restored. It may not be perfect, but at
least the bootloader is loaded...
Signed-off-by: Gary Ching-Pang Lin <glin@suse.com>
set_boot_order() already copies the old BootOrder to the variable,
bootorder. Besides, we can adjust BootOrder when adding the newly
generated boot option. So, we don't have to copy the old one again
in update_boot_order(). This avoid the duplicate entries in BootOrder.
Signed-off-by: Gary Ching-Pang Lin <glin@suse.com>
In read_header, we adjust context->PEHdr's address by doshdr->e_lfanew.
If we're going to recompute that address, we have to adjust it here
too.
Signed-off-by: Peter Jones <pjones@redhat.com>
This adds additional bounds-checking on the section sizes. Also adds
-Wsign-compare to the Makefile and replaces some signed variables with
unsigned counteparts for robustness.
Signed-off-by: Kees Cook <kees@ubuntu.com>
Track use of the system's LoadImage(), and when the next StartImage()
call is for an image the system verified, allow that to count as
participating, since it has been verified by the system's db.
Signed-off-by: Peter Jones <pjones@redhat.com>
Because you know you wanted a test plan. You feel it deeply inside.
Note that none of the /negative/ cases are tested yet.
Signed-off-by: Peter Jones <pjones@redhat.com>
Some firmwares seem to ignore our boot entries and put their fallback
entries back on top. Right now that results in a lot of boot entries
for our stuff, a la https://bugzilla.redhat.com/show_bug.cgi?id=995834 .
Instead of that happening, if we simply find existing entries that match
the entry we would create and move them to the top of the boot order,
the machine will continue to operate in failure mode (which we can't
avoid), but at least we won't create thousands of extra entries.
Signed-off-by: Peter Jones <pjones@redhat.com>
UEFI 2.x section 3.1.2 provides for "short-form device path", where the
first element specified is a "hard drive media device path", so that you
can move a disk around on different buses without invalidating your
device path. Fallback has not been using this option, though in most
cases efibootmgr has.
Note that we still keep the full device path, because LoadImage()
isn't necessarily the layer where HD() works - one some systems BDS is
responsible for resolving the full path and passes that to LoadImage()
instead. So we have to do LoadImage() with the full path.
The things we do for our tools. In this case, make the AllocatePool()
happen outside of a conditional, even though that conditional will
always bee satisfied. This way coverity won't think we're setting fi
to NULL and passing it to StrCaseCmp.
Signed-off-by: Peter Jones <pjones@redhat.com>
Right now we always look for e.g. "\grubx64.efi", which is completely
wrong. This makes it look for the path shim was loaded from and modify
that to end in a sanitized version of our default loader name.
Resolves: rhbz#1032583
Signed-off-by: Peter Jones <pjones@redhat.com>
Shim should only need to enforce its security policy when its launching
binaries signed with its built-in key. Binaries signed by keys in db or
Mokdb should be able to rely on their own security policy.
Signed-off-by: Matthew Garrett <matthew.garrett@nebula.com>
