Check that there are sufficient bytes in the buffer before reading the
cipherSize from it. Also, reduce the bufferSize variable by the number
of bytes that make up the cipherSize to avoid reading and writing bytes
beyond the buffer in subsequent steps that do in-place decryption.
This fixes CVE-2023-1017 & CVE-2023-1018.
Signed-off-by: Stefan Berger <stefanb@linux.ibm.com>
Do not set the RSA_FLAG_NO_BLINDING on RSA keys anymore. The BoringSSL
documentation states the following:
"RSA_FLAG_NO_BLINDING disables blinding of private operations, which is
a dangerous thing to do. It is deprecated and should not be used. It will
be ignored whenever possible."
So, follow this documentation. OpenSSL does not seem to say much about it.
Signed-off-by: Stefan Berger <stefanb@linux.ibm.com>
Check the secret size against the size of the buffer, not the size
member that has not been set yet.
Reported by Coverity.
Signed-off-by: Ross Lagerwall <ross.lagerwall@citrix.com>
The build environment used by OSS-Fuzz reports this error.
In file included from /src/libtpms/src/tpm_debug.c:42:
/src/libtpms/src/tpm_debug.h:69:9: error: 'printf' macro redefined [-Werror,-Wmacro-redefined]
#define printf(...) TPMLIB_LogPrintf(__VA_ARGS__);
: ^
/usr/include/x86_64-linux-gnu/bits/stdio2.h:110:11: note: previous definition is here
# define printf(...) \
^
1 error generated.
The simple fix is to #undef printf in case it is #define'd.
Signed-off-by: Stefan Berger <stefanb@linux.ibm.com>
Openssl 3.0 did return an error if EVP_PKEY_CTX_set0_rsa_oaep_label was called
with label size 0. The function should only be called if the size of the label
is greater 0.
With this fix TPM2_RSA_Encrypt/Decrypt did work with OpenSSL 1.1 and 3.0
for encryption without label.
Signed-off-by: Juergen Repp <juergen.repp@sit.fraunhofer.de>
If s_ContextSlotMask was not set since the TPM 2 was not initialized
by a call to TPM_Manufacture() or the state was not resumed, then
initialize the s_ContextSlotMask to 0xffff.
This situation can occur if a VM with an attached swtpm was started
and the VM's firmware either doesn't support TPM or didn't get to
initialize the vTPM.
The following commands recreated the issue with a SeaBIOS-only VM that
had no attached hard disk but an attached TPM 2:
virsh start BIOS-only-VM ; virsh save BIOS-only-VM save.bin ; \
virsh restore save.bin
Error: Failed to restore domain from save.bin
error: internal error: qemu unexpectedly closed the monitor: \
2022-01-04T19:26:18.835851Z qemu-system-x86_64: tpm-emulator: Setting the stateblob (type 2) failed with a TPM error 0x3 a parameter is bad
2022-01-04T19:26:18.835899Z qemu-system-x86_64: error while loading state for instance 0x0 of device 'tpm-emulator'
2022-01-04T19:26:18.835929Z qemu-system-x86_64: load of migration failed: Input/output error
Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=2035731
Signed-off-by: Stefan Berger <stefanb@linux.ibm.com>
exp_array_size is always initialized if `rc == TPM_RC_SUCCESS` and never used
if `rc != TPM_RC_SUCCESS` but some compilers have trouble noticing this.
Signed-off-by: kpcyrd <git@rxv.cc>
To avoid timeouts on short-running commands, such as TPM2_PCR_Extend,
avoid triggering the writing of the permanent state of the TPM 2
if only the clock was updated. So the clock by itself will not cause
the permanent state to be written out anymore but there have to be
other reasons as well.
The state will still be written out upon a TPM2_Shutdown, which is
supposed to be the last command to be sent to the TPM when shutting
down the VM/vTPM. Also, the permanent state will still carry the
latest clock value if it is retrieved via control channel for
VM/VTPM suspend.
The case that may be affected, but is of lesser importance, is the one
where swtpm's volatile state is written to storage using 'swtpm_ioctl -v'
and then swtpm is terminated and restarted (similar to suspend/resume)
and the permanent state file is read from storage but does not contain
the latest clock value. In this case the go.clock will be updated when
the first command after resume is executed.
This fixes the swtpm issue https://github.com/stefanberger/swtpm/issues/597.
Signed-off-by: Stefan Berger <stefanb@linux.ibm.com>
Instead of using -Wno-deprecated-declarations use
-DOPENSSL_SUPPRESS_DEPRECATED to only suppress OpenSSL deprecated
declarations warnings.
Signed-off-by: Stefan Berger <stefanb@linux.ibm.com>
OpenSSL 3.0 has changed the signature of EVP_PKEY_get0_RSA() from
struct rsa_st *EVP_PKEY_get0_RSA(EVP_PKEY *pkey);
to
const struct rsa_st *EVP_PKEY_get0_RSA(const EVP_PKEY *pkey);
We now have to use EVP_PKEY_get1_RSA with this signature so that we can
access the RSA key. The signature of that function hasn't changed between
OpenSSL 1.1.0 and 3.0.0.
struct rsa_st *EVP_PKEY_get1_RSA(EVP_PKEY *pkey);
Free the additional reference held on the RSA key.
Signed-off-by: Stefan Berger <stefanb@linux.ibm.com>
Event sequence objects were never properly marshalled and when their state
was saved and later restored their state may have been corrupted. Fix this
now by also marshalling the state of event sequence objects.
Bump up the version of the HASH_OBJECT's header to '3' so that previously
written state can be resumed if an event sequence object is encountered
and we only unmarshal an event sequence object when the version is at least
'3'.
Fixes issue #259.
Signed-off-by: Stefan Berger <stefanb@linux.ibm.com>
Make the expected array size of compile-time constants dependent on
the version of the header. This way we can add elements to the array
while bumping up the version of the header.
Signed-off-by: Stefan Berger <stefanb@linux.ibm.com>
Extend the test case data generation script with sm4. Since several
distros' openssl do not support sm4, we need to test for whether sm4
is supported by the installed openssl.
Signed-off-by: Stefan Berger <stefanb@linux.ibm.com>
Use the EC_POINT_set/get_affine_coordinates function on OpenSSL >= 1.1.
These function are a 1:1 replacement for the
EC_POINT_set/get_affine_coordinates_GFp functions and are available
since OpenSSL 1.1 and are deprecated in OpenSSL 3.0.
This patch addresses one aspect of the OpenSSL 3.0 issues raised in
issue #215.
Signed-off-by: Stefan Berger <stefanb@linux.ibm.com>
The NVRAM entries in s_indexOrderlyRam array do not need to contain a
0-sized terminating node. Instead, the entries may fill up this 512
byte array so that no NV_RAM_HEADER structure fits anymore. The fact
that no more NV_RAM_HEADER structure fits is also an indicator for the
last entry. We need to account for this in the code marshalling and
unmarshalling the entries so that we stop marshalling the entries
then and similarly stop unmarshalling.
Signed-off-by: Stefan Berger <stefanb@linux.ibm.com>
Initialize a while OBJECT before using it. This is necessary since
an OBJECT may also be used as a HASH_OBJECT via the ANY_OBJECT
union and that HASH_OBJECT can leave bad size inidicators in TPM2B
buffer in the OBJECT. To get rid of this problem we reset the whole
OBJECT to 0 before using it. This is as if the memory for the
OBJECT was just initialized.
Signed-off-by: Stefan Berger <stefanb@linux.ibm.com>
Fix the following compiler warning from gcc 10.3.0 by using memcpy
instead of MemoryCopy (fixes issue #229).
tpm2/NVDynamic.c: In function 'NvRamGetEnd':
tpm2/NVDynamic.c:378:12: warning: function may return address of local variable [-Wreturn-local-addr]
378 | return iter;
| ^
tpm2/NVDynamic.c:339:26: note: declared here
339 | NV_RAM_HEADER header;
| ^
tpm2/NVDynamic.c: In function 'NvRamGetIndex':
tpm2/NVDynamic.c:411:12: warning: function may return address of local variable [-Wreturn-local-addr]
411 | return currentAddr;
| ^
tpm2/NVDynamic.c:339:26: note: declared here
339 | NV_RAM_HEADER header;
| ^
Signed-off-by: Stefan Berger <stefanb@linux.ibm.com>
Restore the original value of the memory location where data from
a stream was unmarshalled and the unmarshalled value was found to
be illegal. The goal is to not keep illegal values in memory.
Signed-off-by: Stefan Berger <stefanb@linux.ibm.com>
Add maxSize parameter to TPM2B_Marshal and assert on it checking
the size of the data intended to be marshaled versus the maximum
buffer size.
Signed-off-by: Stefan Berger <stefanb@linux.ibm.com>
Reset the buffer size indicator in a TPM2B type of buffer after it failed
the test for the maximum buffer size it allows. This prevents having bad
buffer sizes in memory that can come to haunt us when writing the volatile
state for example.
Signed-off-by: Stefan Berger <stefanb@linux.ibm.com>
Windows 2019 Server padds the TPM_ContextLoad() command with additional
bytes up to TPM_PT_MAX_OBJECT_CONTEXT for the TPMS_CONTEXT part. Since
libtpms does not use an OBJECT to serialize the keys (anymore) it now
uses less bytes than the MAXimum of TPM_PT_MAX_OBJECT_CONTEXT bytes and
the padding leaves some unconsumed bytes that end up failing the command
since no left-over bytes are allowed in any command.
When unconsumed bytes are left in TPMS_CONTEXT_Unmarshal() we check that
the original passed in size was that of TPM_PT_MAX_OBJECT_CONTEXT and
only then consume the additional padding bytes. Luckily only one command
calls TPMS_CONTEXT_Unmarshal() so that no unwanted side effects should
occur anywhere else, such as no bytes left for unmarshalling the next
structure.
The wisdom behind the padding is not quite clear but it feels like
ill-fixing the code to work around a Windows 2019 server bug...
This patch fixes issed #217
Signed-off-by: Stefan Berger <stefanb@linux.ibm.com>
This patch addresses issue #209.
The context gap for libtpms is currently only 0xff due to the CONTEXT_SLOT
being a UINT8. To extend this to 0xffff, we need to define the CONTEXT_SLOT
as UINT16 and introduce a global variable s_ContextArrayMask that takes on
two valid values, 0xff for simulating the CONTEXT_SLOT when it was UINT8
and 0xffff for usage with the new CONTEXT_SLOT of type UINT16. All
occurrences of casts to CONTEXT_SLOT are replaced with a macro
CONTEXT_SLOT_MASKED that applies this mask to a value instead of using the
cast. We also use it for some calculations to avoid spilling over from
1 byte into 2 bytes for example. The cast with the new code is the same as
applying the mask 0xffff, and using the 0xff mask we can simulate the old
CONTEXT_SLOT (1 byte), which we need for seamlessly resuming old state. We
switch from the 0xff mask to the 0xffff mask when the TPM is reset.
There's one place where the s_ContextArrayMask is initialized to 0xff, and
this is when we resume 'old' STATE_RESET_DATA. The places where it is
intialized to 0xffff are in TPM_Manufacture() and
TPM_SessionStartup(SU_CLEAR), both of which are not called after resuming
state.
Signed-off-by: Stefan Berger <stefanb@linux.ibm.com>
Do not call BLOCK_SKIP_READ once rc has been set to any error value.
Therefore, surround all occurrences of BLOCK_SKIP_READ() with tests
of 'rc'.
Signed-off-by: Stefan Berger <stefanb@linux.ibm.com>
Implement a cache for the private exponent 'D' and prime 'Q' so that we
do not have to recalculate 'Q' and 'D' every time an RSA key is used. For
a cache hit we now use ~34000 cycles and on a cache miss it needs around
130000 cycles. Previously it needed around 100000 cycles to calcuate 'Q'
and 'D'. Assuming that keys will be reused and the cache is big enough
for the number of keys being use (64 entries), it seems well worth it.
This solution is better than extending the OBJECT with 'D' since OBJECT is
kept in the TPM's NVRAM and we would then need more memory to store OBJECTs
there.
Signed-off-by: Stefan Berger <stefanb@linux.ibm.com>
