| Age | Commit message (Collapse) | Author |
|---|
|
The keymaster1 specification only requires HW modules to implement
SHA256 out of the list of keymaster1 digest modes. That would force
many keys to be software only, and would break legacy scenarios. This
change uses SoftKeymasterDevice to front keymaster modules that don't
implement the full suite of digests, quietly inserting KM_DIGEST_NONE
and KM_PAD_NONE into key generation/import requests when necessary, then
performing the digesting, and sometimes padding, in software, then
delegating crypto operations to the hardware.
This is only done for RSA and EC keys. Software digesting isn't
possible for HMAC or AES-GCM keys.
Note that this is not the complete fix for the bug. Some changes in
keystore are also required, coming in another CL.
Bug: 22529223
Change-Id: I740572eb11341fb0659085309da01d5cbcd3854d
|
|
HMAC and AES-GCM keys must be bound to a mininum MAC/tag length at
creation, and operations may not specify a length smaller than the
minimum, or provide a length smaller than the minimum during
verification.
Bug: 22337277
Change-Id: Id5ae2f4259045ba1418c28e9de8f4a47e67fd433
|
|
Android is built with exceptions disabled, but "operator new" and
"operator new[]" still throw std::bad_alloc on failure rather than
returning new. In general this is a good thing, because it will cause
an immediate crash of the process rather than assigning a null pointer
which is probably not checked. But most memory allocations in Keymaster
are checked, because it's written to run in an environment where new
does *not* throw. This CL updates the code to explicitly use the
non-throwing new.
A handful of throwing news remain, but only in places where a crash on
failure is appropriate.
In addition, this CL also inserts buffer wrap checks in key locations
and changes the development-machine Makefile to build in 32-bit mode, to
make memory problems more apparent.
Bug: 21888473
Change-Id: I8ebc5ec12053e4f5274f6f57ce312abc10611cef
|
|
I should have known better than to make these singletons to begin
with. Globals create problems. This undoes that mistake.
Change-Id: Idf61d5f72e3c34b5c4ddb27cc94b05f506561743
|
|
This reverts commit 13fbe3e93247943c26e7ca2ed27b6d650282b8bf.
Bug: 20912868, 19799085
Change-Id: Iadd6ce5cbe94956c2a2fe277f1bf5b108e4bcf57
|
|
This reverts commit 8ba2a043f0d44ad3f58d4af518f9391c03eca9c3.
I need to update the Volantis non-secure code in sync. Reverting while I get that done.
Change-Id: I0fb9f928e7e624ad678050a04bb873b43b1c9a48
|
|
AndroidKeymaster made a number of assumptions about its context that are
really only valid for TEE-based usage. In addition, KeyFactory made
some similarly TEE-focused assumptions about key blob creation and
parsing.
Both concerns have been moved to a new KeymasterContext class, which is
responsible for building and parsing key blobs in a manner appropriate
for the context in which AndroidKeymaster is running, as well as
providing other context-specific services, such as random number
generation.
In addition, the refactor reduces the need for the KeyBlob and
UnencryptedKeyBlob classes, which encode too many assumptions about blob
formatting and encryption, to the point that they can be removed and
replaced by a handful of utility functions which are much cleaner and
more flexible.
How to review this CL:
I looked hard at breaking this up into smaller CLs, but it's mostly not
feasible. However, it's probably easier to approach it by starting with
the fundamental changes, and then looking at the cascade effects.
1. Look at keymaster_context.h. The core of the change was pulling this
set of features out of AndroidKeymaster. Note that the revised approach
to key blob creation does not involve the KeyBlob and UnencryptedKeyBlob
classes, but instead goes directly from raw key material plus ancillary
data (e.g. auth sets) to a serialized buffer ready to return to
keystore. The same is true in reverse direction for parsing key blobs.
2. Look at key.h. The revised KeyFactory GenerateKey, ImportKey and
LoadKey methods are essential. GenerateKey and ImportKey no longer
produce a Key object, because all that's needed is a returnable blob.
LoadKey produces a Key object, but it starts with raw key material,
rather than an UnencryptedKeyBlob. Also note the change to the Key
class; because Key objects are only created by LoadKey, when there's a
need to use a key, there's only one constructor.
3. Look at asymmetric_key.h, rsa_key.h and rsa_key.cpp. rsa_key.cpp
provides a good example of how the new structure works. GenerateKey and
ImportKey do all of the work necessary to produce an OpenSSL RSA key and
extract the internal representation (using EvpToKeyMaterial; defined in
asymmetric_key.h because it's the same for EC keys). Then, with the raw
key data in hand, they call KeymasterContext::CreateKeyBlob to wrap the
key data in a key blob that can be returned to the caller -- whatever
that wrapping means in the current context. There's a subtlety not
apparent here which is crucial to the rationale for the refactoring:
RsaKeyFactory uses KeymasterContext::get_instance to retrieve the
context, but key factories which depend on operating in a particular
context can use a different way to get their context object, which may
have a larger interface. RsaKeymaster0KeyFactory will do this.
4. Look at soft_keymaster_context. In
particular, SoftKeymasterContext::CreateKeyBlob and ParseKeyBlob.
CreateKeyBlob allocates authorization tags from key_description to
hw_enforced and sw_enforced, then encrypts the key material and
serializes it to a blob. This approach is compatible with the keys
softkeymaster has been producing, but I'm going to change it (post M),
because there's no reason to bother encrypting SW keys with a SW key.
ParseKeyBlob reverses the process to recover the unencrypted key
material and the auth lists. One debatable point was the decision to
implement BuildHiddenAuthorizations and SetAuthorizations here, since
all contexts will need something similar, and they really should all do
it the same. I may refactor later to pull that functionality up to
KeymasterContext; it will depend on what I learn implementing
TrustyKeymasterContext and HybridKeymasterContext (used for the
keymaster0 adapter).
5. Look at ocb_utils and auth_encrypted_key_blob. These contain the key
encryption and key blob serialization code which was formerly split
between AndroidKeymaster::SerializeKeyBlob, UnencryptedKeyBlob and
KeyBlob, now divided into separate encryption and serialization
utilities. Note the refactored key_blob_test.cpp, updated to use the
new utilities rather than UnencryptedKeyBlob.
6. Look at soft_keymaster_device.cpp. Since KeyBlob no longer exists to
provide a nice way to peer into a blob to extract the algorithm, for use
in determining how to parse the keymaster0 signing key params (which
come in as a void*, yuck), we now have to use get_key_characteristics to
recover the params. This was the right way all along; the device layer
should not depend on being able to parse key blobs.
7. The rest.
Bug: 20912868, 19799085
Change-Id: Ieb74b8da39974f674eb8baa959bde75011fdd2e8
|
|
So we can do it differently in the keymaster0 adapter.
Bug: 20912868
Change-Id: If4c602cc0fab3e59cd2e395a97e21dd8cb3a176f
|
|
Change-Id: I7af02342320a9a431cd9845baaf5dbcf61d460c2
|
|
Also modify GoogleKeymaster to query the operation factories to get
lists of supported modes and digests.
Change-Id: Ied30185df5dddaeaeb1106df63237757896d77db
|
|
Change-Id: I6ebab7b44e4a5dbea282397ab8aca437e71bdca0
|
|
Change-Id: I76c73f6e16e5ee4acaf8a78eacd1bfdf3db12b68
|
|
Change-Id: I64c7bdf77388e3cb491b702c52c6746d32f317b0
|
