Aligned memory access (#68)

* Aligned memory access * Multiple alignment fixes. Main badness comes from non-64 bit alignment and rounded ICE-Controlled attribute which makes the code look really ugly * Pass same call info custom data into Lws callbacks and verify protocol type * Use aligned_mem branch from KVS C/C++ producer SDK * Forward typedef is not allowed by some C compiler * Fix failed unit test
awslabs · Dec 17, 2019 · 3fe9465 · 3fe9465
1 parent 86bd653
commit 3fe9465
Show file tree

Hide file tree

Showing 39 changed files with 145 additions and 293 deletions.
diff --git a/.gitmodules b/.gitmodules
@@ -1,3 +1,4 @@
 [submodule "open-source/amazon-kinesis-video-streams-producer-sdk-cpp"]
 	path = open-source/amazon-kinesis-video-streams-producer-sdk-cpp
 	url = https://github.com/awslabs/amazon-kinesis-video-streams-producer-sdk-cpp.git
+	branch = aligned_mem
diff --git a/samples/Samples.h b/samples/Samples.h
@@ -10,16 +10,12 @@ Shared include file for the samples
 extern "C" {
 #endif
 
-// For tight packing
-#pragma pack(push, samples_i, 1) // for byte alignment
-
 #include <com/amazonaws/kinesis/video/webrtcclient/Include.h>
 
 #define NUMBER_OF_H264_FRAME_FILES                                              403
 #define NUMBER_OF_OPUS_FRAME_FILES                                              618
 #define DEFAULT_FPS_VALUE                                                       25
 #define DEFAULT_MAX_CONCURRENT_STREAMING_SESSION                                10
-#define DEFAULT_TERMINATION_POLLING_INTERVAL                                    5 * HUNDREDS_OF_NANOS_IN_A_SECOND
 
 #define SAMPLE_MASTER_CLIENT_ID                                                 "ProducerMaster"
 #define SAMPLE_VIEWER_CLIENT_ID                                                 "ConsumerViewer"
@@ -112,8 +108,6 @@ VOID sampleFrameHandler(UINT64, PFrame);
 VOID onDataChannel(UINT64, PRtcDataChannel);
 VOID onConnectionStateChange(UINT64, RTC_PEER_CONNECTION_STATE);
 
-#pragma pack(pop, samples_i)
-
 #ifdef  __cplusplus
 }
 #endif

diff --git a/src/include/com/amazonaws/kinesis/video/webrtcclient/Include.h b/src/include/com/amazonaws/kinesis/video/webrtcclient/Include.h
@@ -19,9 +19,6 @@ extern "C" {
 #include <com/amazonaws/kinesis/video/common/Include.h>
 #pragma clang diagnostic pop
 
-// For tight packing
-#pragma pack(push, include, 1) // for byte alignment
-
 ////////////////////////////////////////////////////
 // Status return codes
 ////////////////////////////////////////////////////
@@ -1320,8 +1317,6 @@ PUBLIC_API STATUS signalingClientGetIceConfigInfo(SIGNALING_CLIENT_HANDLE, UINT3
  */
 PUBLIC_API STATUS signalingClientConnectSync(SIGNALING_CLIENT_HANDLE);
 
-#pragma pack(pop, include)
-
 #ifdef  __cplusplus
 }
 #endif

diff --git a/src/source/Dtls/Dtls.h b/src/source/Dtls/Dtls.h
@@ -11,9 +11,6 @@
 extern "C" {
 #endif
 
-// For tight packing
-#pragma pack(push, include_i, 1) // for byte alignment
-
 #define MAX_SRTP_MASTER_KEY_LEN 16
 #define MAX_SRTP_SALT_KEY_LEN 14
 
@@ -88,8 +85,6 @@ STATUS dtlsSessionVerifyRemoteCertificateFingerprint(PDtlsSession, PCHAR);
 STATUS dtlsSessionPutApplicationData(PDtlsSession, PBYTE, INT32);
 STATUS dtlsCheckOutgoingDataBuffer(PDtlsSession);
 
-#pragma pack(pop, include_i)
-
 #ifdef  __cplusplus
 }
 #endif

diff --git a/src/source/Ice/ConnectionListener.h b/src/source/Ice/ConnectionListener.h
@@ -10,9 +10,6 @@ Connection Listener internal include file
 extern "C" {
 #endif
 
-// For tight packing
-#pragma pack(push, include_i, 1) // for byte alignment
-
 #define MAX_UDP_PACKET_SIZE                         65507
 #define SOCKET_WAIT_FOR_DATA_TIMEOUT_SECONDS        2
 
@@ -82,8 +79,6 @@ STATUS connectionListenerStart(PConnectionListener);
 ////////////////////////////////////////////
 PVOID connectionListenerReceiveDataRoutine(PVOID arg);
 
-#pragma pack(pop, include_i)
-
 #ifdef  __cplusplus
 }
 #endif

diff --git a/src/source/Ice/IceAgent.h b/src/source/Ice/IceAgent.h
@@ -10,9 +10,6 @@ IceAgent internal include file
 extern "C" {
 #endif
 
-// For tight packing
-#pragma pack(push, include_i, 1) // for byte alignment
-
 #define KVS_ICE_MAX_CANDIDATE_PAIR_COUNT                                1024
 #define KVS_ICE_MAX_REMOTE_CANDIDATE_COUNT                              100
 #define KVS_ICE_MAX_LOCAL_CANDIDATE_COUNT                               100
@@ -306,8 +303,6 @@ UINT32 computeCandidatePriority(PIceCandidate);
 UINT64 computeCandidatePairPriority(PIceCandidatePair, BOOL);
 PCHAR iceAgentGetCandidateTypeStr(ICE_CANDIDATE_TYPE);
 
-#pragma pack(pop, include_i)
-
 #ifdef  __cplusplus
 }
 #endif

diff --git a/src/source/Ice/IceAgentStateMachine.h b/src/source/Ice/IceAgentStateMachine.h
@@ -10,9 +10,6 @@ Signaling State Machine internal include file
 extern "C" {
 #endif
 
-// For tight packing
-#pragma pack(push, include_i, 1) // for byte alignment
-
 /**
  * Ice states definitions
  *
@@ -72,8 +69,6 @@ STATUS executeDisconnectedIceAgentState(UINT64, UINT64);
 STATUS fromFailedIceAgentState(UINT64, PUINT64);
 STATUS executeFailedIceAgentState(UINT64, UINT64);
 
-#pragma pack(pop, include_i)
-
 #ifdef  __cplusplus
 }
 #endif

diff --git a/src/source/Ice/IceUtils.h b/src/source/Ice/IceUtils.h
@@ -10,9 +10,6 @@ Ice Utils internal include file
 extern "C" {
 #endif
 
-// For tight packing
-#pragma pack(push, include_i, 1) // for byte alignment
-
 #define DEFAULT_MAX_STORED_TRANSACTION_ID_COUNT                         20
 #define MAX_STORED_TRANSACTION_ID_COUNT                                 100
 
@@ -41,8 +38,6 @@ STATUS iceUtilsSendStunPacket(PStunPacket, PBYTE, UINT32, PKvsIpAddress, PSocket
 
 STATUS populateIpFromString(PKvsIpAddress, PCHAR);
 
-#pragma pack(pop, include_i)
-
 #ifdef  __cplusplus
 }
 #endif

diff --git a/src/source/Ice/Network.h b/src/source/Ice/Network.h
@@ -10,9 +10,6 @@ HostInfo internal include file
 extern "C" {
 #endif
 
-// For tight packing
-#pragma pack(push, include_i, 1) // for byte alignment
-
 #define MAX_LOCAL_NETWORK_INTERFACE_COUNT               128
 
 // string buffer size for ipv4 and ipv6. Null terminator included.
@@ -63,8 +60,6 @@ STATUS getIpAddrStr(PKvsIpAddress, PCHAR, UINT32);
 
 BOOL isSameIpAddress(PKvsIpAddress, PKvsIpAddress, BOOL);
 
-#pragma pack(pop, include_i)
-
 #ifdef  __cplusplus
 }
 #endif

diff --git a/src/source/Ice/SocketConnection.h b/src/source/Ice/SocketConnection.h
@@ -10,9 +10,6 @@ Socket Connection internal include file
 extern "C" {
 #endif
 
-// For tight packing
-#pragma pack(push, include_i, 1) // for byte alignment
-
 typedef STATUS (*ConnectionDataAvailableFunc)(UINT64, struct __SocketConnection*, PBYTE, UINT32, PKvsIpAddress, PKvsIpAddress);
 
 typedef struct __SocketConnection SocketConnection;
@@ -116,8 +113,6 @@ STATUS socketConnectionReadData(PSocketConnection, PBYTE, UINT32, PUINT32);
 STATUS createConnectionCertificateAndKey(X509 **, EVP_PKEY **);
 INT32 certificateVerifyCallback(INT32 preverify_ok, X509_STORE_CTX *ctx);
 
-#pragma pack(pop, include_i)
-
 #ifdef  __cplusplus
 }
 #endif

diff --git a/src/source/Ice/TurnConnection.h b/src/source/Ice/TurnConnection.h
@@ -10,36 +10,33 @@ TurnConnection internal include file
 extern "C" {
 #endif
 
-// For tight packing
-#pragma pack(push, include_i, 1) // for byte alignment
-
 // https://en.wikipedia.org/wiki/List_of_IP_protocol_numbers
 #define TURN_REQUEST_TRANSPORT_UDP                                      17
 #define TURN_REQUEST_TRANSPORT_TCP                                      6
 #define DEFAULT_TURN_ALLOCATION_LIFETIME_SECONDS                        600
 // required by rfc5766 to be 300s
-#define TURN_PERMISSION_LIFETIME                                        300 * HUNDREDS_OF_NANOS_IN_A_SECOND
-#define DEFAULT_TURN_TIMER_INTERVAL_BEFORE_READY                        100 * HUNDREDS_OF_NANOS_IN_A_MILLISECOND
-#define DEFAULT_TURN_TIMER_INTERVAL_AFTER_READY                         1 * HUNDREDS_OF_NANOS_IN_A_SECOND
-#define DEFAULT_TURN_SEND_REFRESH_INVERVAL                              1 * HUNDREDS_OF_NANOS_IN_A_SECOND
+#define TURN_PERMISSION_LIFETIME                                        (300 * HUNDREDS_OF_NANOS_IN_A_SECOND)
+#define DEFAULT_TURN_TIMER_INTERVAL_BEFORE_READY                        (100 * HUNDREDS_OF_NANOS_IN_A_MILLISECOND)
+#define DEFAULT_TURN_TIMER_INTERVAL_AFTER_READY                         (1 * HUNDREDS_OF_NANOS_IN_A_SECOND)
+#define DEFAULT_TURN_SEND_REFRESH_INVERVAL                              (1 * HUNDREDS_OF_NANOS_IN_A_SECOND)
 
 // turn state timeouts
-#define DEFAULT_TURN_GET_CREDENTIAL_TIMEOUT                             5 * HUNDREDS_OF_NANOS_IN_A_SECOND
-#define DEFAULT_TURN_ALLOCATION_TIMEOUT                                 5 * HUNDREDS_OF_NANOS_IN_A_SECOND
-#define DEFAULT_TURN_CREATE_PERMISSION_TIMEOUT                          2 * HUNDREDS_OF_NANOS_IN_A_SECOND
-#define DEFAULT_TURN_BIND_CHANNEL_TIMEOUT                               3 * HUNDREDS_OF_NANOS_IN_A_SECOND
-#define DEFAULT_TURN_CLEAN_UP_TIMEOUT                                   10 * HUNDREDS_OF_NANOS_IN_A_SECOND
+#define DEFAULT_TURN_GET_CREDENTIAL_TIMEOUT                             (5 * HUNDREDS_OF_NANOS_IN_A_SECOND)
+#define DEFAULT_TURN_ALLOCATION_TIMEOUT                                 (5 * HUNDREDS_OF_NANOS_IN_A_SECOND)
+#define DEFAULT_TURN_CREATE_PERMISSION_TIMEOUT                          (2 * HUNDREDS_OF_NANOS_IN_A_SECOND)
+#define DEFAULT_TURN_BIND_CHANNEL_TIMEOUT                               (3 * HUNDREDS_OF_NANOS_IN_A_SECOND)
+#define DEFAULT_TURN_CLEAN_UP_TIMEOUT                                   (10 * HUNDREDS_OF_NANOS_IN_A_SECOND)
 
 /*
  * if no application data is sent through turn for this much time then we assume that a better connection is found
  * and initiate turn clean up.
  */
-#define DEFAULT_TURN_START_CLEAN_UP_TIMEOUT                             10 * HUNDREDS_OF_NANOS_IN_A_SECOND
-#define DEFAULT_TURN_ALLOCATION_REFRESH_GRACE_PERIOD                    30 * HUNDREDS_OF_NANOS_IN_A_SECOND
-#define DEFAULT_TURN_PERMISSION_REFRESH_GRACE_PERIOD                    30 * HUNDREDS_OF_NANOS_IN_A_SECOND
+#define DEFAULT_TURN_START_CLEAN_UP_TIMEOUT                             (10 * HUNDREDS_OF_NANOS_IN_A_SECOND)
+#define DEFAULT_TURN_ALLOCATION_REFRESH_GRACE_PERIOD                    (30 * HUNDREDS_OF_NANOS_IN_A_SECOND)
+#define DEFAULT_TURN_PERMISSION_REFRESH_GRACE_PERIOD                    (30 * HUNDREDS_OF_NANOS_IN_A_SECOND)
 
-#define DEFAULT_TURN_MESSAGE_SEND_CHANNEL_DATA_BUFFER_LEN               10 * 1024
-#define DEFAULT_TURN_MESSAGE_RECV_CHANNEL_DATA_BUFFER_LEN               10 * 1024
+#define DEFAULT_TURN_MESSAGE_SEND_CHANNEL_DATA_BUFFER_LEN               (10 * 1024)
+#define DEFAULT_TURN_MESSAGE_RECV_CHANNEL_DATA_BUFFER_LEN               (10 * 1024)
 
 // all turn channel numbers must be greater than 0x4000 and less than 0x7FFF
 #define TURN_CHANNEL_BIND_CHANNEL_NUMBER_BASE                           (UINT16) 0x4000
@@ -185,8 +182,6 @@ STATUS turnConnectionHandleStun(PTurnConnection, PSocketConnection, PBYTE, UINT3
 STATUS turnConnectionHandleStunError(PTurnConnection, PSocketConnection, PBYTE, UINT32);
 STATUS turnConnectionHandleChannelDataTcpMode(PTurnConnection, PSocketConnection, PBYTE, UINT32);
 
-#pragma pack(pop, include_i)
-
 #ifdef  __cplusplus
 }
 #endif

diff --git a/src/source/Include_i.h b/src/source/Include_i.h
@@ -50,9 +50,6 @@ struct __SocketConnection;
 #include <arpa/inet.h>
 #endif
 
-// For tight packing
-#pragma pack(push, include_i, 1) // for byte alignment
-
 // Max uFrag and uPwd length as documented in https://tools.ietf.org/html/rfc5245#section-15.4
 #define ICE_MAX_UFRAG_LEN               256
 #define ICE_MAX_UPWD_LEN                256
@@ -134,8 +131,6 @@ typedef struct {
 // Project internal functions
 ////////////////////////////////////////////////////
 
-#pragma pack(pop, include_i)
-
 #ifdef  __cplusplus
 }
 #endif

diff --git a/src/source/PeerConnection/DataChannel.h b/src/source/PeerConnection/DataChannel.h
@@ -10,9 +10,6 @@ DataChannel internal include file
 extern "C" {
 #endif
 
-// For tight packing
-#pragma pack(push, include_i, 1) // for byte alignment
-
 typedef struct {
     RtcDataChannel dataChannel;
 
@@ -23,9 +20,6 @@ typedef struct {
     RtcOnMessage onMessage;
 } KvsDataChannel, *PKvsDataChannel;
 
-
-#pragma pack(pop, include_i)
-
 #ifdef  __cplusplus
 }
 #endif

diff --git a/src/source/PeerConnection/JitterBuffer.h b/src/source/PeerConnection/JitterBuffer.h
@@ -10,9 +10,6 @@ PeerConnection internal include file
 extern "C" {
 #endif
 
-// For tight packing
-#pragma pack(push, include_i, 1) // for byte alignment
-
 typedef STATUS (*FrameReadyFunc)(UINT64, UINT16, UINT16, UINT32);
 typedef STATUS (*FrameDroppedFunc)(UINT64, UINT32);
 #define UINT16_DEC(a) ((UINT16) ((a) - 1))
@@ -24,8 +21,8 @@ typedef struct {
     DepayRtpPayloadFunc depayPayloadFn;
 
     UINT32 lastPushTimestamp;
-    UINT16 lastRemovedSequenceNumber;
     UINT32 lastRemovedTimestamp;
+    UINT16 lastRemovedSequenceNumber;
     UINT16 lastPopSequenceNumber;
     UINT32 lastPopTimestamp;
     UINT64 maxLatency;
@@ -41,8 +38,6 @@ STATUS jitterBufferPop(PJitterBuffer, BOOL);
 STATUS jitterBufferDropBufferData(PJitterBuffer, UINT16, UINT16, UINT32);
 STATUS jitterBufferFillFrameData(PJitterBuffer, PBYTE, UINT32, PUINT32, UINT16, UINT16);
 
-#pragma pack(pop, include_i)
-
 #ifdef  __cplusplus
 }
 #endif

diff --git a/src/source/PeerConnection/PeerConnection.h b/src/source/PeerConnection/PeerConnection.h
@@ -10,9 +10,6 @@ PeerConnection internal include file
 extern "C" {
 #endif
 
-// For tight packing
-#pragma pack(push, include_i, 1) // for byte alignment
-
 #define LOCAL_ICE_UFRAG_LEN             4
 #define LOCAL_ICE_PWD_LEN               24
 #define LOCAL_CNAME_LEN                 16
@@ -30,8 +27,8 @@ extern "C" {
 
 #define CODEC_HASH_TABLE_BUCKET_COUNT                   50
 #define CODEC_HASH_TABLE_BUCKET_LENGTH                  2
-#define RTX_HASH_TABLE_BUCKET_COUNT                   50
-#define RTX_HASH_TABLE_BUCKET_LENGTH                  2
+#define RTX_HASH_TABLE_BUCKET_COUNT                     50
+#define RTX_HASH_TABLE_BUCKET_LENGTH                    2
 
 #define DATA_CHANNEL_HASH_TABLE_BUCKET_COUNT            200
 #define DATA_CHANNEL_HASH_TABLE_BUCKET_LENGTH           2
@@ -103,8 +100,6 @@ STATUS sendPacketToRtpReceiver(PKvsPeerConnection, PBYTE, UINT32);
 
 STATUS generateJSONSafeString(PCHAR, UINT32);
 
-#pragma pack(pop, include_i)
-
 #ifdef  __cplusplus
 }
 #endif

diff --git a/src/source/PeerConnection/Retransmitter.h b/src/source/PeerConnection/Retransmitter.h
@@ -10,22 +10,17 @@ Retransmitter internal include file
 extern "C" {
 #endif
 
-// For tight packing
-#pragma pack(push, include_i, 1) // for byte alignment
-
 typedef struct {
     PUINT16 sequenceNumberList;
     UINT32 seqNumListLen;
-    PUINT64 validIndexList;
     UINT32 validIndexListLen;
+    PUINT64 validIndexList;
 } Retransmitter, *PRetransmitter;
 
 STATUS createRetransmitter(UINT32, UINT32, PRetransmitter*);
 STATUS freeRetransmitter(PRetransmitter*);
 STATUS resendPacketOnNack(PRtcpPacket, PKvsPeerConnection);
 
-#pragma pack(pop, include_i)
-
 #ifdef  __cplusplus
 }
 #endif

diff --git a/src/source/PeerConnection/Rtcp.h b/src/source/PeerConnection/Rtcp.h
@@ -7,13 +7,8 @@
 extern "C" {
 #endif
 
-// For tight packing
-#pragma pack(push, include_i, 1) // for byte alignment
-
 STATUS onRtcpPacket(PKvsPeerConnection, PBYTE, INT32);
 
-#pragma pack(pop, include_i)
-
 #ifdef  __cplusplus
 }
 #endif