This CL cleans up some bad merges of client ID code, entitlement keys,
and concurrent session access. After this CL, core cdm code on
android should match that on widevine at the commit 2f916720 on branch
master.
CLs merged here are based on:
http://go/wvgerrit/50483 Protect sessions from concurrent access
http://go/wvgerrit/48860 Remove duplicate information from client identification
http://go/wvgerrit/49040 Revert revertion of Client ID Expansion
http://go/wvgerrit/46448 Test Entitlement Licenses
Test: tested as part of http://go/ag/4674759
Change-Id: I45854d6b034c247b16073a96d6ff3ea953ded3ae
Some of http://go/wvgerrit/46251 from Widevine repo.
The rest was merged in the oemcrypto refactor.
When we standardized on BoringSSL, these conditional compilations that
had been added as a stopgap for OpenSSL became unneeded. However, they
were not noticed and removed at the time.
Bug: 72459799
Test: CE CDM Unit Tests
Test: tested as part of http://go/ag/4674759
Change-Id: I693f691ffcb255e03660edaa6743cd0fb9ef12c6
Merge from Widevine repo of http://go/wvgerrit/45940
Since the Level 3 OEMCrypto is being updated to Provisioning 3.0, its
SPOID would be derived from its OEM Certificate, breaking backwards
compatibility. This CL changes how we determine what unique id to use
for SPOIDs by checking to see if OEMCrypto_GetDeviceID is implemented,
and if so, using the id returned from that call. If not and the root
of trust is an OEM Cert, we continue to use that OEM Cert.
This allows Level 3 devices to keep the same SPOID when they undergo a
field update to Provisioning 3.0.
Also, the Level 3 OEMCrypto will share a single OEM certificate across
all devices with the same architecture. Since the OEM Cert is not
unique, it cannot be used to derive a unique id. By using the unique
id returned by OEMCrypto_GetDeviceID, we can generate a unique SPOID.
The id from OEMCrypto_GetDeviceID has always been required to be
unique for devices with keyboxes. The functionality and use of this
function for Provisioning 3.0 devices was introduced in OEMCrypto API
version 14.1.
Test: tested as part of http://go/ag/4674759
Change-Id: I65af8246c9312c75c570a2d518caa3de633007c4
Merge from Widevine repo of http://go/wvgerrit/48841
This test is not providing value to the CDM, as it seems to exist
primarily to validate server behavior. However, it is not doing what
it says it is (the request is rejected because it is using unparseable
garbage data, not because its key ID is unknown) and according to
tinskip@, the behavior it claims to be testing is not valid. (The
licensing service will not fail just because the key ID is unknown.
Indeed, if the test data is fixed to use a valid payload with an
unknown key ID, the test fails because the server does not.)
Bug: 78640287
Test: CE CDM Unit Tests
Test: Android Unit Tests
Test: tested as part of http://go/ag/4674759
Change-Id: Idfcff15ab3d15fdfb6eb111b5dff68aa5a23fb37
Merge from Widevine repo of http://go/wvgerrit/43202
Sync the definition of WidevinePssh data with the latest in support of
entitlement keys.
bug: 73297961 Fix or remove sublicense support.
Test: tested as part of http://go/ag/4674759
Change-Id: Ia9faf82732854a705b4b14430169ce4c8ecbcfcd
Merge from Widevine repo of http://go/wvgerrit/53883
Note: this CL does not modify license_key_status.cpp because the
previous CL already included those changes.
OEMCrypto v14 only supports one entitled key per entitlement key at a
time. Unfortunately, some partners have use cases that require using
old entitlement keys after the new keys have been loaded. Most
notably, when a key rotation occurs, the new PSSH will often be loaded
before the playback position catches up to the PSSH in the stream,
meaning that decryption will need to continue using the old keys for a
bit.
To fix this, EntitlementKeySession now caches the entitled keys when
they are loaded and only loads them under their matching entitlement
key when SelectKey() is called. This ensures that the right entitled
key is loaded for a given entitlement key before decryption.
The entitlement key integration tests have been updated to verify that
the old entitled keys still work even after loading new entitled keys.
Also, several places in the code that assumed loading new entitled
keys would wipe out the old keys have had to be modified.
Bug: 78652567
Test: CE CDM Unit Tests
Test: tested as part of http://go/ag/4674759
Change-Id: I6fac9dfe2b170ad68fb7cdb5bc8d6a2f35a20c2c
Merge from Widevine repo of http://go/wvgerrit/49003
CopyOldUsageEntry currently needs the old usage table to be loaded via
CreateOldUsageEntry. The CDM uses a workaround by creating a dummy old
entry, but the OEMCrypto code should be responsible for this. However,
since there have been several versions released with the current
OEMCrypto spec, the CDM code would have to still exist to support
implementations of that spec. Therefore, in order to avoid having to
support both a CDM with this workaround removed (as well as updating
the spec) and a CDM with it still in place, this workaround should be
canonicalized.
b/65730828
Test: tested as part of http://go/ag/4674759
Change-Id: I4619c551b79a53746683519d284663bf513ec38d
Merge from Widevine repo of http://go/wvgerrit/53465
It may not be clear that Properties::Init is called multiple
times. This CL adds an explanation in the header file. The function
itself is platform specific.
Test: tested as part of http://go/ag/4674759
Change-Id: I68010c594ec146e36161ae0f4f44d53caefa1896
Merge from Widevine repo of http://go/wvgerrit/53640
While HttpSocket supports both secure and insecure requests, the
error-handling code in HttpSocket::Read() was written assuming that
the insecure code path was previously taken. This resulted in spurious
and misleading error messages being printed when an SSL error
occurred, and it also meant that retryable SSL responses were not
being retried. Also, the code for detecting a closed connection was
technically incorrect, although a quirk of BoringSSL meant that it
happened to work well enough to go unnoticed.
This patch adds separate SSL error handling from the non-secure error
handling. It correctly checks for a closed connection. It will retry
retryable errors after a delay. And it prints the correct BoringSSL
error when an unrecoverable error occurs. There should be no change in
behavior for insecure connections.
Bug: 77338045
Test: CE CDM Unit Tests
Test: tested as part of http://go/ag/4674759
Change-Id: I8c45ca5771f22c11716d2e3649de91ab1acc1954
Merge from Widevine repo of http://go/wvgerrit/50481
Original CL http://go/wvgerrit/47520
The original CL was not completely merged to master on Android, so
this CL only reverts the left-over bits.
The original fix was not sufficient to address all race conditions. A
subsequent CL will address them.
Bug: 73781703
Bug: 79158083
Bug: 79262108
Test: tested as part of http://go/ag/4674759
Change-Id: Ib6c55ab5434e08fe61e0f65623ac8c7b2dc5aaa1
Merge from Widevine repo of http://go/wvgerrit/53625
Passing the third parameter to std::string::insert() as an integer
technically makes for an ambiguous method call, as there are two
overloads that could accept these arguments. While baseline GCC and
Clang discern our intent here correctly, seawardt@ discovered that
XCode is more pedantic and requires that we pass the third parameter
as a char to disambiguate.
Test: WvCdmEnginePreProvTestUat.ProvisioningServiceCertificateInvalidTest
Test: tested as part of http://go/ag/4674759
Change-Id: I65a2506209215cd081c685faac26e08bae486d5e
Merge from Widevine repo of http://go/wvgerrit/53202
and
Merge from Widevine repo of http://go/wvgerrit/53624
This change contains a variety of small tweaks to the
ContentKeySession and EntitlementKeySession classes that were
discovered while fixing b/78652567. There should be no change in
behavior from this patch. The fixes are:
1) Added missing headers and removed unnecessary headers.
2) Removed the unused keys_ member from EntitlementKeySession.
3) Renamed ContentKeySession's protected member function so that it is
not an overload of the public LoadKeys() function. This makes it
clearer what EntitlementKeySession::LoadKeys() is doing.
4) Added missing "virtual" and "OVERRIDE" keywords.
5) Added missing copyright headers.
6) Ran clang-format with Google style.
7) Correct missing OVERRIDE keywords.
Test: tested as part of http://go/ag/4674759
Change-Id: Icb0af886d7d3eb097b5dffbb716be6ac28f0916d
Merge from Widevine repo of http://go/wvgerrit/48842
In order to work around a limitation of some versions of OEMCrypto,
the packager is going to start generating files with multiple Widevine
PSSH boxes. For backwards-compatibility, the first PSSH will be a
SINGLE-type PSSH while the ENTITLED_KEYS-type PSSH (if any) will come
later. In order to use entitlement licenses, then, the CDM needs to
change how it selects PSSHs from the init data blob.
Previously, the CDM always took the first Widevine PSSH it found. Now,
it must find all the Widevine PSSHs and select the appropriate PSSH
for the OEMCrypto implementation. ENTITLTED_KEYS will be used on OEC
v14 and later, if available, while SINGLE will be preferred on earlier
OEMCrypto versions.
As a side-effect of this, the CDM is now stricter about what PSSH
payloads it will accept. Previously, it would blindly accept the
payload of any PSSH where the wrapper was not malformed. Now, it
sometimes has to actually parse the payload, and therefore PSSHs that
have corrupted payloads will be rejected. This affected a few unit
tests which used PSSHs that were malformed. These tests have been
updated to use PSSHs that do not fail to parse.
Bug: 78142219
Test: CE CDM Unit Tests
Test: Android Unit Tests
Test: Android Google Play & Netflix
Test: tested as part of http://go/ag/4674759
Change-Id: Ia70d627a914299bfbae84b4cb46f100dc5c7a501
Merge from Widevine repo of http://go/wvgerrit/52480
Partners have asked for a way to release offline licenses without
using a release message. This is typically used by cable partners who
are caching licenses ahead of time and do not care about usage
statistics.
As part of implementing this request, CdmSession::DeleteLicense() was
renamed to reflect that it only deletes the *files* associated with a
license, and a new CdmSession::DeleteLicense() has been written that
also cleans up other related data.
Bug: 77955334
Test: CE CDM Unit Tests
Test: tested as part of http://go/ag/4674759
Change-Id: I00d6e20935c5fecb3ac9be6757c0f191d85c6bd6
Merge from Widevine repo of http://go/wvgerrit/49820
Devices with baked-in DRM certs cannot be reprovisioned. As such, we
must protect them against being unprovisioned. Currently, our unit
tests break such devices by attempting to unprovision them. This patch
adds code to block the Unprovision() call on these devices.
Bug: 69264798
Test: CE CDM Unit Tests
Test: tested as part of http://go/ag/4674759
Change-Id: I49322dcb2d3d5c7953e870eb91a9e0b978d4dabe
This CL just updates some comments from the http://go/wvgerrit/51680. Changes
in that CL made to oemcrypto are in the refactor CL.
Test: tested as part of http://go/ag/4674759
Change-Id: Ia619089e146ea635c5a73a53bc81973bb42b42f7
Merge from Widevine repo of http://go/wvgerrit/46204
Refactor utility code - split the mock, step 1
Merge from Widevine repo of http://go/wvgerrit/46205
Move some OEMCrypto types to common header - split the mock, step 2
Merge from Widevine repo of http://go/wvgerrit/46206
Split mock into two -- step 3
Merge from Widevine repo of http://go/wvgerrit/47460
Split the mock into two -- step 3.5
The CL moves several files used by oemcrypto and cdm into a common
subdirectory, so that it may more easily be shared with partners.
The CORE_DISALLOW_COPY_AND_ASSIGN macro was moved to its own header in
the util/include directory.
This CL removes some references to the mock from other code, and puts
some constants and types, such as the definition of the keybox, into a
header in oemcrypto.
Test: tested as part of http://go/ag/4674759
bug: 76393338
Change-Id: I75b4bde7062ed8ee572c97ebc2f4da018f4be0c9
[ Merge of http://go/wvgerrit/58460 ]
If OEMCrypto runs out of space in the usage table header+entries adding
a new license or loading/using an existing one might fail. This CL makes
two modifications to handle this scenario.
* OEMCrypto_ERROR_INSUFFICIENT_RESOURCES will be returned from
OEMCrypto_CreateNewUsageEntry or OEMCrypto_LoadUsageEntry. An attempt
will be made to release a LRU entry from the usage table and retry
the operation. This may be retried 3 times unless success
occurs earlier.
* On initialization, the usage table header is loaded. If there are more than
the minimum number of usage entries (200), an attempt is made to
add a usage entry. If this fails, we are likely in an unrecoverable
state. We then delete all offline licenses, usage information and
recreate the usage table header. This will allow future playback
attempts to succeed and offline licenses to be able to be downloaded
but will lose all current offline licenses and secure stops.
Bug: 112486006
Test: WV unit/integration tests, GtsMediaDrmTest
Playback tests using Netflix and Play movies.
Change-Id: I41a18d69a329f8a96c7b607d299ce73af3d56177
(This is a merge of http://go/wvgerrit/55265)
Compiling with GCC 7 revealed that a function call in this test was
missing an argument. It meant to be passing the output protection level
to the function, but because of optional arguments, it was instead
passing it as the fourth argument.
Fixing this revealed that the test cases for the test were incorrect in
one case, which has been fixed to expect the correct results.
Thankfully, this part of the code does not appear to have been broken
while the tests had this hole.
Bug: 111648438
Test: build_and_run_all_unit_tests.sh
Change-Id: I6c13d5fecdccc4185ca5e8698fc845929ff16cb1
[ Merge of http://go/wvgerrit/52040 ]
Information stored in files are serialized and protected by an MD5 hash.
When files cannot be read because the MD5 hash computed over it
fails verification, the file is deleted. This allows for recovery.
However if the protobuf deserialization fails we return an error
but do not delete the file. When errors of this sort occur
with usage information files, the CDM cannot recover.
removeAllSecureStops() will fail as well and new licenses
with PSTs cannot be processed. In order to recover the file will
be deleted when a protobuf deserialization error occurs.
Bug: 109765590
Test: WV unit, integration tests. GTS tests. Netflix and Play Movies playback
Change-Id: I408914924e644d5c22b2ba7865d3a7d598788ee6
[ Merge of http://go/wvgerrit/51322 ]
This avoids taking the session_map_lock_ twice.
Bug: 80248149
Test: WV unit/integration tests, GtsMediaTestCases,
2 days of netflix playback.
Change-Id: Iea1c7b7ba08d7d40c227d21c5abfce13c0a8b395
(This is a merge of http://go/wvgerrit/51084)
Nominally, OEMCrypto probably shouldn't modify the buffer descriptor we
pass into OEMCrypto_DecryptCENC(), but in practice, we know some
platforms do this, so we make defensive copies in
CryptoSession::DecryptInChunks() just in case. Turns out, some devices
also behave like this in OEMCrypto_CopyBuffer(), so we should also be
doing defensive copies in CryptoSession::CopyBufferInChunks().
Bug: 79779554
Test: ExoPlayer Demo App, played "Secure Subsample UHD (WebM, VP9)"
Test: build_and_run_all_unit_tests.sh
Change-Id: Ib46043a6cc0aa42d1d1cc85f5adb477c566363e9
Locks in earlier releases controlled access to sessions and the list
of sessions for each CdmEngine instance. This guarded against
concurrent access between session management (OpenSession,
CloseSession, etc), periodic timer calls and calls to Decrypt.
The list of sessions and locking was moved to a separate class
CdmSessionMap. This left open the possibility that a session
might be destructed, while being called to decrypt or invoked through the
timer. An attempt was made to add per-session locks in b/73781703
but this was found insufficient.
Per-session locks will be introduced in a future changelist, but for
now the coarser locks will be reintroduced.
Bug: 73781703
Bug: 79158083
Bug: 79262108
Bug: 79436509
Test: WV unit/integration tests, GTS GtsMediaTestCases tests and
24 hours of continuous Netflix playback.
Change-Id: I30a3ede340192370dfe5c92c01b1c76df16b7123
[ http://go/wvgerrit/50341 ]
The shared_ptr implementation was taken from a google3 implementation.
Updates to the reference counter needed to be atomic and were
platform dependent in the original code. These were not carried
over to this codebase. Race conditions between calls to decrypt and
the periodic timer, led to incorrect reference count values.
CdmSession objects were then destructed while references to
them still existed. Segfaults occurred when they were referenced.
Bug: 79431096
Test: WV unit/integration tests, GTS GtsMediaTestCases tests and
24 hours of continuous Netflix playback.
Change-Id: I6008ddba869efcc58972e5ea8644a204f91410ab
[ Original CL http://ag/3890635,
Merge of http://go/wvgerrit/50340 ]
The original fix was not sufficient to address all race conditions.
A subsequent CL will address them.
Bug: 73781703
Bug: 79158083
Bug: 79262108
Test: WV unit/integration tests, GTS GtsMediaTestCases tests and
24 hours of continuous Netflix playback.
Change-Id: I869c22a250e2467b3d49935815e4157dc012fff5
[ Merge of http://go/wvgerrit/49980 ]
This CL
* corrects some of the test expectations
* switches test content used to test streaming with provider session tokens.
The policy of the earlier test content had changed.
* adds some more information to log messages
Bug: 63819720
Test: WV unit, integration tests, WvCdmExtendedDuraionTest,
GtsMediaDrmTest
Change-Id: I8fdbc9c38d6018cc6e884e1b95b2e9d26e7aa536
[ Merge of http://go/wvgerrit/48640 ]
Usage information is saved periodically, in order to avoid excessive
flash writes. This limits our session usage accuracy to
within a usage save period. Saving usage information when
closing a session is an improvement and addresses some failures
seen with Netflix compliance tests.
Bug: 74015553
Test: WV unit/integration tests
Change-Id: I680aad05922f334df0611ff3933082a512f7c002
[ Merge of http://go/wvgerrit/48720 ]
The device ID does not need to be sent in the client identification
information as it is either present in other fields or ignored
by the license service.
This also allows for build information to be reported during
provisioning for devices with OEM certificates.
Bug: 78578351
Test: WV unit/integration tests. GtsMediaTestCases.
Change-Id: I708c63d34f0e2df7e465154d12096a394a1b23d7
[ Merge of http://go/wvgerrit/48400 ]
Client identification information has recently been enabled in
provisioning messages. For privacy concerns this information
is being encrypted with a default service certificate.
Apps need to be able to override the default one to allow
for provisioning with third party provisioning services.
Bug: 78420508
Test: WV unit, integration tests
New WvCdmRequestLicenseTest.ProvisioningTestWithServiceCertificate test
GTS MediaDrmTestCases
Change-Id: Iee61ad47d33ce011efbea4eb90f7e4b1f032d15f
[ Merge of http://go/wvgerrit/46760 ]
This enables encryption of client ID for provisioning requests for
devices with keyboxes as root of trust. Client ID will not be
provided for those devices with OEM device certificates as root of
trust. That will be addressed in b/78303730.
Bug: 77607585
Test: WV unit/integration tests. Tests with L3 using OEM certs
Change-Id: Id9bd697aa049bd5659ab80714e141dbc50408f6a
Merge from Widevine repo of http://go/wvgerrit/47860
This CL updates the copyright notice to indicate that files
shared with partners are shared under the Widevine Master
License Agreement.
bug: 77926774
test: comment change only
Change-Id: I0423668111578b80fb39a932d763df2827e2dfc3
[ Merge of http://go/wvgerrit/47065 ]
RemoveKeys now resets associated crypto and policy resources,
rather than just closing the crypto session. This results in a
MediaCodec.CryptoException with error code ERROR_NO_KEY
rather than ERROR_SESSION_NOT_OPENED, if decrypt is called
afterwards.
Error SESSION_NOT_FOUND_FOR_DECRYPT is made unique. Error codes
were also synchonized between various branches in the widevine repo.
Bug: 77304819
Test: WV unit/integration tests, VtsHalDrmV1_0Target tests
Change-Id: I6cba2a3e1ce466d58c7727cde2d8f81d9503d655
[ Merge of http://go/wvgerrit/46907 ]
The WV client supports root of trusts as keyboxes or OEM certificates.
Devices with keyboxes use provisioning 2.0 protocol to provision
while those with OEM certificates use 3.0. L3 provisioning failures
occur if the L1 and L3 root of trusts differ.
The provisioning method is now retrieved and cached when the
security level is known, when the session is opened.
Earlier it was retrieved and cached at initialization time and
always set to the value of L1 OEMCrypto (if present). This led
to provisioning failures.
A case of acquiring a lock while one was held in GetProvisioningId()
has also fixed.
Bug: 77606913
Test: WV unit/integration tests
Change-Id: I2d66ee2cf64f846cec4a37fbccb554447c8a0e1d
(This is a merge from http://go/wvgerrit/46447)
The Production Provisioning Service is moving to the "widevine.com"
certificate from the "license.widevine.com" certificate it was using.
This replaces the two places this certificate appears in the source
code. This is expected to be the last such update.
Also, the Staging Provisioning Service was already using this
certificate, but our code had it listed as using the old certificate. It
has also been updated.
Bug: 77244492
Test: CE CDM Unit Tests
Test: Android Unit Tests
Change-Id: I2ce14ea8e672c453ce0f74fbd3345f7e40f2f297
[ Merge of http://go/wvgerrit/46623 ]
If corruption of the usage information file is detected while saving a
streaming license with a PST, usage information file is deleted, so that
a subsequent load keys may succeed.
Also when calling the MediaDrm API releaseAllSecureStops(), an error would
be returned if usage info file was corrupted. Since this file is
deleted successfully, errors have been replaced with warnings.
Bug: 73447733
Test: wv unit/integration tests
Change-Id: Ie4a63ac202fd6009609105f38ffa8a3b23ed334e
[ Merge of http://go/wvgerrit/46622 ]
Secure stop API related changes introduced in b/69674645 caused
segfaults on taimen/walleye but not other devices due to a difference
in OEMCrypto version.
Bug: 77294890
Test: WV unit/integration tests on sailfish and walleye
Change-Id: I8523ef283334d7d32d180e902072fe1dd6e665c1
Previously, we did not have a license request latency metric. This is a
notable limitation in our metrics. This adds a metric that captures the
timing between a GenerateKeyRequest and an AddKey operation.
Bug: 72994956
Test: New unit tests. Google Play
Change-Id: If99c187399c02f9b5d4c355732af7588bbbefb11
A few metrics were missing or not properly collected in the CDM metrics.
This CL addresses them.
Bug: 64570194
Bug: 72866232
Test: Unit tests and Google Play manual test.
Change-Id: I3a3aa4fb3eb8422c9c8c398016f02409307beb33
Merge of http://go/wvgerrit/45520/
Bug: b/70650789
Test: request_license_tests and GTS tests on sailfish and taimen
This is related to b/70650789. An extra call to DeriveKeysFromSessionKey
in the case where there is a provider session token results in the
OEMCrypto mock incorrectly using the derived mac keys to sign the release
message, since a future call to LoadKeys is never called.
