Windows: speedup PRF autodetection mode by implementing abort mechanism in PBKDF2/Argon2 primitives

2025-11-11 02:58:02 -06:00 · 2025-06-29 21:44:32 +09:00
parent 95659a8563
commit 9dc24ba7d0
21 changed files with 437 additions and 228 deletions
--- a/src/Common/Dlgcode.c
+++ b/src/Common/Dlgcode.c
@@ -6445,32 +6445,32 @@ static BOOL PerformBenchmark(HWND hBenchDlg, HWND hwndDlg)

 				case SHA512:
 					/* PKCS-5 test with HMAC-SHA-512 used as the PRF */
-					derive_key_sha512 ((const unsigned char*) "passphrase-1234567890", 21, (const unsigned char*) tmp_salt, 64, iterations, dk, MASTER_KEYDATA_SIZE);
+					derive_key_sha512 ((const unsigned char*) "passphrase-1234567890", 21, (const unsigned char*) tmp_salt, 64, iterations, dk, MASTER_KEYDATA_SIZE, NULL);
 					break;

 				case SHA256:
 					/* PKCS-5 test with HMAC-SHA-256 used as the PRF */
-					derive_key_sha256 ((const unsigned char*)"passphrase-1234567890", 21, (const unsigned char*)tmp_salt, 64, iterations, dk, MASTER_KEYDATA_SIZE);
+					derive_key_sha256 ((const unsigned char*)"passphrase-1234567890", 21, (const unsigned char*)tmp_salt, 64, iterations, dk, MASTER_KEYDATA_SIZE, NULL);
 					break;
                          #ifndef WOLFCRYPT_BACKEND
 				case BLAKE2S:
 					/* PKCS-5 test with HMAC-BLAKE2s used as the PRF */
-					derive_key_blake2s ((const unsigned char*)"passphrase-1234567890", 21, (const unsigned char*)tmp_salt, 64, iterations, dk, MASTER_KEYDATA_SIZE);
+					derive_key_blake2s ((const unsigned char*)"passphrase-1234567890", 21, (const unsigned char*)tmp_salt, 64, iterations, dk, MASTER_KEYDATA_SIZE, NULL);
 					break;

 				case WHIRLPOOL:
 					/* PKCS-5 test with HMAC-Whirlpool used as the PRF */
-					derive_key_whirlpool ((const unsigned char*)"passphrase-1234567890", 21, (const unsigned char*)tmp_salt, 64, iterations, dk, MASTER_KEYDATA_SIZE);
+					derive_key_whirlpool ((const unsigned char*)"passphrase-1234567890", 21, (const unsigned char*)tmp_salt, 64, iterations, dk, MASTER_KEYDATA_SIZE, NULL);
 					break;

 				case STREEBOG:
 					/* PKCS-5 test with HMAC-STREEBOG used as the PRF */
-					derive_key_streebog((const unsigned char*)"passphrase-1234567890", 21, (const unsigned char*)tmp_salt, 64, iterations, dk, MASTER_KEYDATA_SIZE);
+					derive_key_streebog((const unsigned char*)"passphrase-1234567890", 21, (const unsigned char*)tmp_salt, 64, iterations, dk, MASTER_KEYDATA_SIZE, NULL);
 					break;
 				
 				case ARGON2:
 					/* test with ARGON2 used as the PRF */
-					derive_key_argon2 ((const unsigned char*) "passphrase-1234567890", 21, (const unsigned char*)tmp_salt, 64, iterations, memoryCost, dk, MASTER_KEYDATA_SIZE);
+					derive_key_argon2 ((const unsigned char*) "passphrase-1234567890", 21, (const unsigned char*)tmp_salt, 64, iterations, memoryCost, dk, MASTER_KEYDATA_SIZE, NULL);
 					break;
 				}
 	                   #endif	
--- a/src/Common/EncryptionThreadPool.c
+++ b/src/Common/EncryptionThreadPool.c
@@ -107,7 +107,7 @@ typedef struct EncryptionThreadPoolWorkItemStruct
 			int PasswordLength;
 			int Pkcs5Prf;
 			unsigned char *Salt;
-
+			LONG volatile *pAbortKeyDerivation; 
 		} KeyDerivation;

 		struct
@@ -245,32 +245,32 @@ static TC_THREAD_PROC EncryptionThreadProc (void *threadArg)
 			{
 			case BLAKE2S:
 				derive_key_blake2s (workItem->KeyDerivation.Password, workItem->KeyDerivation.PasswordLength, workItem->KeyDerivation.Salt, PKCS5_SALT_SIZE,
-					workItem->KeyDerivation.IterationCount, workItem->KeyDerivation.DerivedKey, GetMaxPkcs5OutSize());
+					workItem->KeyDerivation.IterationCount, workItem->KeyDerivation.DerivedKey, GetMaxPkcs5OutSize(), workItem->KeyDerivation.pAbortKeyDerivation);
 				break;

 			case SHA512:
 				derive_key_sha512 (workItem->KeyDerivation.Password, workItem->KeyDerivation.PasswordLength, workItem->KeyDerivation.Salt, PKCS5_SALT_SIZE,
-					workItem->KeyDerivation.IterationCount, workItem->KeyDerivation.DerivedKey, GetMaxPkcs5OutSize());
+					workItem->KeyDerivation.IterationCount, workItem->KeyDerivation.DerivedKey, GetMaxPkcs5OutSize(), workItem->KeyDerivation.pAbortKeyDerivation);
 				break;

 			case WHIRLPOOL:
 				derive_key_whirlpool (workItem->KeyDerivation.Password, workItem->KeyDerivation.PasswordLength, workItem->KeyDerivation.Salt, PKCS5_SALT_SIZE,
-					workItem->KeyDerivation.IterationCount, workItem->KeyDerivation.DerivedKey, GetMaxPkcs5OutSize());
+					workItem->KeyDerivation.IterationCount, workItem->KeyDerivation.DerivedKey, GetMaxPkcs5OutSize(), workItem->KeyDerivation.pAbortKeyDerivation);
 				break;

 			case SHA256:
 				derive_key_sha256 (workItem->KeyDerivation.Password, workItem->KeyDerivation.PasswordLength, workItem->KeyDerivation.Salt, PKCS5_SALT_SIZE,
-					workItem->KeyDerivation.IterationCount, workItem->KeyDerivation.DerivedKey, GetMaxPkcs5OutSize());
+					workItem->KeyDerivation.IterationCount, workItem->KeyDerivation.DerivedKey, GetMaxPkcs5OutSize(), workItem->KeyDerivation.pAbortKeyDerivation);
 				break;

 			case STREEBOG:
 				derive_key_streebog(workItem->KeyDerivation.Password, workItem->KeyDerivation.PasswordLength, workItem->KeyDerivation.Salt, PKCS5_SALT_SIZE,
-					workItem->KeyDerivation.IterationCount, workItem->KeyDerivation.DerivedKey, GetMaxPkcs5OutSize());
+					workItem->KeyDerivation.IterationCount, workItem->KeyDerivation.DerivedKey, GetMaxPkcs5OutSize(), workItem->KeyDerivation.pAbortKeyDerivation);
 				break;

 			case ARGON2:
 				derive_key_argon2(workItem->KeyDerivation.Password, workItem->KeyDerivation.PasswordLength, workItem->KeyDerivation.Salt, PKCS5_SALT_SIZE,
-					workItem->KeyDerivation.IterationCount, workItem->KeyDerivation.Memorycost, workItem->KeyDerivation.DerivedKey, GetMaxPkcs5OutSize());
+					workItem->KeyDerivation.IterationCount, workItem->KeyDerivation.Memorycost, workItem->KeyDerivation.DerivedKey, GetMaxPkcs5OutSize(), workItem->KeyDerivation.pAbortKeyDerivation);
 				break;

 			default:
@@ -533,7 +533,7 @@ void EncryptionThreadPoolStop ()
 }


-void EncryptionThreadPoolBeginKeyDerivation (TC_EVENT *completionEvent, TC_EVENT *noOutstandingWorkItemEvent, LONG *completionFlag, LONG *outstandingWorkItemCount, int pkcs5Prf, unsigned char *password, int passwordLength, unsigned char *salt, int iterationCount, int memoryCost, unsigned char *derivedKey)
+void EncryptionThreadPoolBeginKeyDerivation (TC_EVENT *completionEvent, TC_EVENT *noOutstandingWorkItemEvent, LONG *completionFlag, LONG *outstandingWorkItemCount, int pkcs5Prf, unsigned char *password, int passwordLength, unsigned char *salt, int iterationCount, int memoryCost, unsigned char *derivedKey, LONG volatile *pAbortKeyDerivation)
 {
 	EncryptionThreadPoolWorkItem *workItem;

@@ -563,6 +563,7 @@ void EncryptionThreadPoolBeginKeyDerivation (TC_EVENT *completionEvent, TC_EVENT
 	workItem->KeyDerivation.PasswordLength = passwordLength;
 	workItem->KeyDerivation.Pkcs5Prf = pkcs5Prf;
 	workItem->KeyDerivation.Salt = salt;
+	workItem->KeyDerivation.pAbortKeyDerivation = pAbortKeyDerivation;

 	InterlockedIncrement (outstandingWorkItemCount);
 	TC_CLEAR_EVENT (*noOutstandingWorkItemEvent);
--- a/src/Common/EncryptionThreadPool.h
+++ b/src/Common/EncryptionThreadPool.h
@@ -32,7 +32,7 @@ typedef enum
 size_t GetCpuCount (WORD* pGroupCount);
 #endif

-void EncryptionThreadPoolBeginKeyDerivation (TC_EVENT *completionEvent, TC_EVENT *noOutstandingWorkItemEvent, LONG *completionFlag, LONG *outstandingWorkItemCount, int pkcs5Prf, unsigned char *password, int passwordLength, unsigned char *salt, int iterationCount,  int memoryCost, unsigned char *derivedKey);
+void EncryptionThreadPoolBeginKeyDerivation (TC_EVENT *completionEvent, TC_EVENT *noOutstandingWorkItemEvent, LONG *completionFlag, LONG *outstandingWorkItemCount, int pkcs5Prf, unsigned char *password, int passwordLength, unsigned char *salt, int iterationCount,  int memoryCost, unsigned char *derivedKey, LONG volatile *pAbortKeyDerivation);
 void EncryptionThreadPoolBeginReadVolumeHeaderFinalization (TC_EVENT *keyDerivationCompletedEvent, TC_EVENT *noOutstandingWorkItemEvent, LONG* outstandingWorkItemCount, void* keyInfoBuffer, int keyInfoBufferSize, void* keyDerivationWorkItems, int keyDerivationWorkItemsSize);
 void EncryptionThreadPoolDoWork (EncryptionThreadPoolWorkType type, uint8 *data, const UINT64_STRUCT *startUnitNo, uint32 unitCount, PCRYPTO_INFO cryptoInfo);
 BOOL EncryptionThreadPoolStart (size_t encryptionFreeCpuCount);
--- a/src/Common/Pkcs5.c
+++ b/src/Common/Pkcs5.c
@@ -145,7 +145,11 @@ void hmac_sha256
 }
 #endif

-static void derive_u_sha256 (const unsigned char *salt, int salt_len, uint32 iterations, int b, hmac_sha256_ctx* hmac)
+static void derive_u_sha256 (const unsigned char *salt, int salt_len, uint32 iterations, int b, hmac_sha256_ctx* hmac
+#ifndef TC_WINDOWS_BOOT
+	, long volatile *pAbortKeyDerivation
+#endif
+)
 {
 	unsigned char* k = hmac->k;
 	unsigned char* u = hmac->u;
@@ -186,6 +190,11 @@ static void derive_u_sha256 (const unsigned char *salt, int salt_len, uint32 ite
 	/* remaining iterations */
 	while (c > 1)
 	{
+#ifndef TC_WINDOWS_BOOT
+		// CANCELLATION CHECK: Check every 1024 iterations
+		if (pAbortKeyDerivation && (c & 1023) == 0 && *pAbortKeyDerivation == 1)
+			return; // Abort derivation
+#endif
 		hmac_sha256_internal (k, SHA256_DIGESTSIZE, hmac);
 		for (i = 0; i < SHA256_DIGESTSIZE; i++)
 		{
@@ -196,7 +205,11 @@ static void derive_u_sha256 (const unsigned char *salt, int salt_len, uint32 ite
 }


-void derive_key_sha256 (const unsigned char *pwd, int pwd_len, const unsigned char *salt, int salt_len, uint32 iterations, unsigned char *dk, int dklen)
+void derive_key_sha256 (const unsigned char *pwd, int pwd_len, const unsigned char *salt, int salt_len, uint32 iterations, unsigned char *dk, int dklen
+#ifndef TC_WINDOWS_BOOT
+	, long volatile *pAbortKeyDerivation
+#endif
+)
 {	
 	hmac_sha256_ctx hmac;
 	sha256_ctx* ctx;
@@ -264,20 +277,36 @@ void derive_key_sha256 (const unsigned char *pwd, int pwd_len, const unsigned ch
 	/* first l - 1 blocks */
 	for (b = 1; b < l; b++)
 	{
+#ifndef TC_WINDOWS_BOOT
+		derive_u_sha256 (salt, salt_len, iterations, b, &hmac, pAbortKeyDerivation);
+		// Check if the derivation was aborted
+		if (pAbortKeyDerivation && *pAbortKeyDerivation == 1)
+			goto cancelled;
+#else
 		derive_u_sha256 (salt, salt_len, iterations, b, &hmac);
+#endif
 		memcpy (dk, hmac.u, SHA256_DIGESTSIZE);
 		dk += SHA256_DIGESTSIZE;
 	}

 	/* last block */
+#ifndef TC_WINDOWS_BOOT
+	derive_u_sha256 (salt, salt_len, iterations, b, &hmac, pAbortKeyDerivation);
+	// Check if the derivation was aborted (in case of only one block)
+	if (pAbortKeyDerivation && *pAbortKeyDerivation == 1)
+		goto cancelled;
+#else
 	derive_u_sha256 (salt, salt_len, iterations, b, &hmac);
+#endif
 	memcpy (dk, hmac.u, r);

 #if defined (DEVICE_DRIVER) && !defined(_M_ARM64)
 	if (NT_SUCCESS (saveStatus))
 		KeRestoreExtendedProcessorState(&SaveState);
 #endif
-
+#ifndef TC_WINDOWS_BOOT
+cancelled:
+#endif
 	/* Prevent possible leaks. */
 	burn (&hmac, sizeof(hmac));
 #ifndef TC_WINDOWS_BOOT
@@ -395,7 +424,7 @@ void hmac_sha512
 	burn (key, sizeof(key));
 }

-static void derive_u_sha512 (const unsigned char *salt, int salt_len, uint32 iterations, int b, hmac_sha512_ctx* hmac)
+static void derive_u_sha512 (const unsigned char *salt, int salt_len, uint32 iterations, int b, hmac_sha512_ctx* hmac, long volatile *pAbortKeyDerivation)
 {
 	unsigned char* k = hmac->k;
 	unsigned char* u = hmac->u;
@@ -413,6 +442,9 @@ static void derive_u_sha512 (const unsigned char *salt, int salt_len, uint32 ite
 	/* remaining iterations */
 	for (c = 1; c < iterations; c++)
 	{
+		// CANCELLATION CHECK: Check every 1024 iterations
+		if (pAbortKeyDerivation && (c & 1023) == 0 && *pAbortKeyDerivation == 1)
+			return; // Abort derivation
 		hmac_sha512_internal (k, SHA512_DIGESTSIZE, hmac);
 		for (i = 0; i < SHA512_DIGESTSIZE; i++)
 		{
@@ -422,7 +454,7 @@ static void derive_u_sha512 (const unsigned char *salt, int salt_len, uint32 ite
 }


-void derive_key_sha512 (const unsigned char *pwd, int pwd_len, const unsigned char *salt, int salt_len, uint32 iterations, unsigned char *dk, int dklen)
+void derive_key_sha512 (const unsigned char *pwd, int pwd_len, const unsigned char *salt, int salt_len, uint32 iterations, unsigned char *dk, int dklen, long volatile *pAbortKeyDerivation)
 {
 	hmac_sha512_ctx hmac;
 	sha512_ctx* ctx;
@@ -489,20 +521,26 @@ void derive_key_sha512 (const unsigned char *pwd, int pwd_len, const unsigned ch
 	/* first l - 1 blocks */
 	for (b = 1; b < l; b++)
 	{
-		derive_u_sha512 (salt, salt_len, iterations, b, &hmac);
+		derive_u_sha512 (salt, salt_len, iterations, b, &hmac, pAbortKeyDerivation);
+		// Check if the derivation was aborted
+		if (pAbortKeyDerivation && *pAbortKeyDerivation == 1)
+			goto cancelled;
 		memcpy (dk, hmac.u, SHA512_DIGESTSIZE);
 		dk += SHA512_DIGESTSIZE;
 	}

 	/* last block */
-	derive_u_sha512 (salt, salt_len, iterations, b, &hmac);
+	derive_u_sha512 (salt, salt_len, iterations, b, &hmac, pAbortKeyDerivation);
+	// Check if the derivation was aborted (in case of only one block)
+	if (pAbortKeyDerivation && *pAbortKeyDerivation == 1)
+		goto cancelled;
 	memcpy (dk, hmac.u, r);

 #if defined (DEVICE_DRIVER) && !defined(_M_ARM64)
 	if (NT_SUCCESS (saveStatus))
 		KeRestoreExtendedProcessorState(&SaveState);
 #endif
-
+cancelled:
 	/* Prevent possible leaks. */
 	burn (&hmac, sizeof(hmac));
 	burn (key, sizeof(key));
@@ -619,7 +657,11 @@ void hmac_blake2s
 }
 #endif

-static void derive_u_blake2s (const unsigned char *salt, int salt_len, uint32 iterations, int b, hmac_blake2s_ctx* hmac)
+static void derive_u_blake2s (const unsigned char *salt, int salt_len, uint32 iterations, int b, hmac_blake2s_ctx* hmac
+#ifndef TC_WINDOWS_BOOT
+	, volatile long *pAbortKeyDerivation
+#endif
+)
 {
 	unsigned char* k = hmac->k;
 	unsigned char* u = hmac->u;
@@ -660,6 +702,11 @@ static void derive_u_blake2s (const unsigned char *salt, int salt_len, uint32 it
 	/* remaining iterations */
 	while (c > 1)
 	{
+#ifndef TC_WINDOWS_BOOT
+		// CANCELLATION CHECK: Check every 1024 iterations
+		if (pAbortKeyDerivation && (c & 1023) == 0 && *pAbortKeyDerivation)
+			return; // Abort derivation
+#endif
 		hmac_blake2s_internal (k, BLAKE2S_DIGESTSIZE, hmac);
 		for (i = 0; i < BLAKE2S_DIGESTSIZE; i++)
 		{
@@ -670,7 +717,11 @@ static void derive_u_blake2s (const unsigned char *salt, int salt_len, uint32 it
 }


-void derive_key_blake2s (const unsigned char *pwd, int pwd_len, const unsigned char *salt, int salt_len, uint32 iterations, unsigned char *dk, int dklen)
+void derive_key_blake2s (const unsigned char *pwd, int pwd_len, const unsigned char *salt, int salt_len, uint32 iterations, unsigned char *dk, int dklen
+#ifndef TC_WINDOWS_BOOT
+	, volatile long *pAbortKeyDerivation
+#endif
+)
 {	
 	hmac_blake2s_ctx hmac;
 	blake2s_state* ctx;
@@ -738,20 +789,36 @@ void derive_key_blake2s (const unsigned char *pwd, int pwd_len, const unsigned c
 	/* first l - 1 blocks */
 	for (b = 1; b < l; b++)
 	{
+#ifndef TC_WINDOWS_BOOT
+		derive_u_blake2s (salt, salt_len, iterations, b, &hmac, pAbortKeyDerivation);
+		// Check if the derivation was aborted
+		if (pAbortKeyDerivation && *pAbortKeyDerivation)
+			goto cancelled;
+#else
 		derive_u_blake2s (salt, salt_len, iterations, b, &hmac);
+#endif
 		memcpy (dk, hmac.u, BLAKE2S_DIGESTSIZE);
 		dk += BLAKE2S_DIGESTSIZE;
 	}

 	/* last block */
+#ifndef TC_WINDOWS_BOOT
+	derive_u_blake2s (salt, salt_len, iterations, b, &hmac, pAbortKeyDerivation);
+	// Check if the derivation was aborted (in case of only one block)
+	if (pAbortKeyDerivation && *pAbortKeyDerivation)
+		goto cancelled;
+#else
 	derive_u_blake2s (salt, salt_len, iterations, b, &hmac);
+#endif
 	memcpy (dk, hmac.u, r);

 #if defined (DEVICE_DRIVER) && !defined(_M_ARM64)
 	if (NT_SUCCESS (saveStatus))
 		KeRestoreExtendedProcessorState(&SaveState);
 #endif
-
+#ifndef TC_WINDOWS_BOOT
+cancelled:
+#endif
 	/* Prevent possible leaks. */
 	burn (&hmac, sizeof(hmac));
 #ifndef TC_WINDOWS_BOOT
@@ -856,7 +923,7 @@ void hmac_whirlpool
 	burn(&hmac, sizeof(hmac));
 }

-static void derive_u_whirlpool (const unsigned char *salt, int salt_len, uint32 iterations, int b, hmac_whirlpool_ctx* hmac)
+static void derive_u_whirlpool (const unsigned char *salt, int salt_len, uint32 iterations, int b, hmac_whirlpool_ctx* hmac, volatile long *pAbortKeyDerivation)
 {
 	unsigned char* u = hmac->u;
 	unsigned char* k = hmac->k;
@@ -874,6 +941,9 @@ static void derive_u_whirlpool (const unsigned char *salt, int salt_len, uint32
 	/* remaining iterations */
 	for (c = 1; c < iterations; c++)
 	{
+		// CANCELLATION CHECK: Check every 1024 iterations
+		if (pAbortKeyDerivation && (c & 1023) == 0 && *pAbortKeyDerivation)
+			return; // Abort derivation
 		hmac_whirlpool_internal (k, WHIRLPOOL_DIGESTSIZE, hmac);
 		for (i = 0; i < WHIRLPOOL_DIGESTSIZE; i++)
 		{
@@ -882,7 +952,7 @@ static void derive_u_whirlpool (const unsigned char *salt, int salt_len, uint32
 	}
 }

-void derive_key_whirlpool (const unsigned char *pwd, int pwd_len, const unsigned char *salt, int salt_len, uint32 iterations, unsigned char *dk, int dklen)
+void derive_key_whirlpool (const unsigned char *pwd, int pwd_len, const unsigned char *salt, int salt_len, uint32 iterations, unsigned char *dk, int dklen, volatile long *pAbortKeyDerivation)
 {
 	hmac_whirlpool_ctx hmac;
 	WHIRLPOOL_CTX* ctx;
@@ -942,15 +1012,21 @@ void derive_key_whirlpool (const unsigned char *pwd, int pwd_len, const unsigned
 	/* first l - 1 blocks */
 	for (b = 1; b < l; b++)
 	{
-		derive_u_whirlpool (salt, salt_len, iterations, b, &hmac);
+		derive_u_whirlpool (salt, salt_len, iterations, b, &hmac, pAbortKeyDerivation);
+		// Check if the derivation was aborted
+		if (pAbortKeyDerivation && *pAbortKeyDerivation)
+			goto cancelled;
 		memcpy (dk, hmac.u, WHIRLPOOL_DIGESTSIZE);
 		dk += WHIRLPOOL_DIGESTSIZE;
 	}

 	/* last block */
-	derive_u_whirlpool (salt, salt_len, iterations, b, &hmac);
+	derive_u_whirlpool (salt, salt_len, iterations, b, &hmac, pAbortKeyDerivation);
+	// Check if the derivation was aborted (in case of only one block)
+	if (pAbortKeyDerivation && *pAbortKeyDerivation)
+		goto cancelled;
 	memcpy (dk, hmac.u, r);
-
+cancelled:
 	/* Prevent possible leaks. */
 	burn (&hmac, sizeof(hmac));
 	burn (key, sizeof(key));
@@ -1050,7 +1126,7 @@ void hmac_streebog
 	burn(&hmac, sizeof(hmac));
 }

-static void derive_u_streebog (const unsigned char *salt, int salt_len, uint32 iterations, int b, hmac_streebog_ctx* hmac)
+static void derive_u_streebog (const unsigned char *salt, int salt_len, uint32 iterations, int b, hmac_streebog_ctx* hmac, volatile long *pAbortKeyDerivation)
 {
 	unsigned char* u = hmac->u;
 	unsigned char* k = hmac->k;
@@ -1068,6 +1144,9 @@ static void derive_u_streebog (const unsigned char *salt, int salt_len, uint32 i
 	/* remaining iterations */
 	for (c = 1; c < iterations; c++)
 	{
+		// CANCELLATION CHECK: Check every 1024 iterations
+		if (pAbortKeyDerivation && (c & 1023) == 0 && *pAbortKeyDerivation)
+			return; // Abort derivation
 		hmac_streebog_internal (k, STREEBOG_DIGESTSIZE, hmac);
 		for (i = 0; i < STREEBOG_DIGESTSIZE; i++)
 		{
@@ -1076,7 +1155,7 @@ static void derive_u_streebog (const unsigned char *salt, int salt_len, uint32 i
 	}
 }

-void derive_key_streebog (const unsigned char *pwd, int pwd_len, const unsigned char *salt, int salt_len, uint32 iterations, unsigned char *dk, int dklen)
+void derive_key_streebog (const unsigned char *pwd, int pwd_len, const unsigned char *salt, int salt_len, uint32 iterations, unsigned char *dk, int dklen, volatile long *pAbortKeyDerivation)
 {
 	hmac_streebog_ctx hmac;
 	STREEBOG_CTX* ctx;
@@ -1136,15 +1215,21 @@ void derive_key_streebog (const unsigned char *pwd, int pwd_len, const unsigned
 	/* first l - 1 blocks */
 	for (b = 1; b < l; b++)
 	{
-		derive_u_streebog (salt, salt_len, iterations, b, &hmac);
+		derive_u_streebog (salt, salt_len, iterations, b, &hmac, pAbortKeyDerivation);
+		// Check if the derivation was aborted
+		if (pAbortKeyDerivation && *pAbortKeyDerivation)
+			goto cancelled;
 		memcpy (dk, hmac.u, STREEBOG_DIGESTSIZE);
 		dk += STREEBOG_DIGESTSIZE;
 	}

 	/* last block */
-	derive_u_streebog (salt, salt_len, iterations, b, &hmac);
+	derive_u_streebog (salt, salt_len, iterations, b, &hmac, pAbortKeyDerivation);
+	// Check if the derivation was aborted (in case of only one block)
+	if (pAbortKeyDerivation && *pAbortKeyDerivation)
+		goto cancelled;
 	memcpy (dk, hmac.u, r);
-
+cancelled:
 	/* Prevent possible leaks. */
 	burn (&hmac, sizeof(hmac));
 	burn (key, sizeof(key));
@@ -1245,7 +1330,7 @@ int is_pkcs5_prf_supported (int pkcs5_prf_id, PRF_BOOT_TYPE bootType)

 }

-void derive_key_argon2(const unsigned char *pwd, int pwd_len, const unsigned char *salt, int salt_len, uint32 iterations, uint32 memcost, unsigned char *dk, int dklen)
+void derive_key_argon2(const unsigned char *pwd, int pwd_len, const unsigned char *salt, int salt_len, uint32 iterations, uint32 memcost, unsigned char *dk, int dklen, volatile long *pAbortKeyDerivation)
 {
 #if defined (DEVICE_DRIVER) && !defined(_M_ARM64)
 	NTSTATUS saveStatus = STATUS_INVALID_PARAMETER;
@@ -1259,7 +1344,8 @@ void derive_key_argon2(const unsigned char *pwd, int pwd_len, const unsigned cha
 		1, // parallelism factor (number of threads)
 		pwd, pwd_len, // password and its length
 		salt, salt_len, // salt and its length
-		dk, dklen// derived key and its length
+		dk, dklen,// derived key and its length
+		pAbortKeyDerivation 
 	))
 	{
 		// If the Argon2 derivation fails, we fill the derived key with zeroes
--- a/src/Common/Pkcs5.h
+++ b/src/Common/Pkcs5.h
@@ -20,30 +20,31 @@
 extern "C"
 {
 #endif
+
+#ifndef TC_WINDOWS_BOOT
 /* output written to input_digest which must be at lease 32 bytes long */
 void hmac_blake2s (unsigned char *key, int keylen, unsigned char *input_digest, int len);
-void derive_key_blake2s (const unsigned char *pwd, int pwd_len, const unsigned char *salt, int salt_len, uint32 iterations, unsigned char *dk, int dklen);
+void derive_key_blake2s (const unsigned char *pwd, int pwd_len, const unsigned char *salt, int salt_len, uint32 iterations, unsigned char *dk, int dklen, long volatile *pAbortKeyDerivation);

 /* output written to d which must be at lease 32 bytes long */
 void hmac_sha256 (unsigned char *k, int lk, unsigned char *d, int ld);
-void derive_key_sha256 (const unsigned char *pwd, int pwd_len, const unsigned char *salt, int salt_len, uint32 iterations, unsigned char *dk, int dklen);
+void derive_key_sha256 (const unsigned char *pwd, int pwd_len, const unsigned char *salt, int salt_len, uint32 iterations, unsigned char *dk, int dklen, long volatile *pAbortKeyDerivation);

-#ifndef TC_WINDOWS_BOOT
 /* output written to d which must be at lease 64 bytes long */
 void hmac_sha512 (unsigned char *k, int lk, unsigned char *d, int ld);
-void derive_key_sha512 (const unsigned char *pwd, int pwd_len, const unsigned char *salt, int salt_len, uint32 iterations, unsigned char *dk, int dklen);
+void derive_key_sha512 (const unsigned char *pwd, int pwd_len, const unsigned char *salt, int salt_len, uint32 iterations, unsigned char *dk, int dklen, long volatile *pAbortKeyDerivation);

 /* output written to d which must be at lease 64 bytes long */
 void hmac_whirlpool (unsigned char *k, int lk, unsigned char *d, int ld);
-void derive_key_whirlpool (const unsigned char *pwd, int pwd_len, const unsigned char *salt, int salt_len, uint32 iterations, unsigned char *dk, int dklen);
+void derive_key_whirlpool (const unsigned char *pwd, int pwd_len, const unsigned char *salt, int salt_len, uint32 iterations, unsigned char *dk, int dklen, long volatile *pAbortKeyDerivation);

 void hmac_streebog (unsigned char *k, int lk, unsigned char *d, int ld);
-void derive_key_streebog (const unsigned char *pwd, int pwd_len, const unsigned char *salt, int salt_len, uint32 iterations, unsigned char *dk, int dklen);
+void derive_key_streebog (const unsigned char *pwd, int pwd_len, const unsigned char *salt, int salt_len, uint32 iterations, unsigned char *dk, int dklen, long volatile *pAbortKeyDerivation);

 int get_pkcs5_iteration_count (int pkcs5_prf_id, int pim, BOOL bBoot, int* pMemoryCost);
 wchar_t *get_pkcs5_prf_name (int pkcs5_prf_id);

-void derive_key_argon2(const unsigned char *pwd, int pwd_len, const unsigned char *salt, int salt_len, uint32 iterations, uint32 memcost, unsigned char *dk, int dklen);
+void derive_key_argon2(const unsigned char *pwd, int pwd_len, const unsigned char *salt, int salt_len, uint32 iterations, uint32 memcost, unsigned char *dk, int dklen, long volatile *pAbortKeyDerivation);
 void get_argon2_params(int pim, int* pIterations, int* pMemcost);

 /* check if given PRF supported.*/
@@ -55,6 +56,15 @@ typedef enum
 } PRF_BOOT_TYPE;

 int is_pkcs5_prf_supported (int pkcs5_prf_id, PRF_BOOT_TYPE bootType);
+#else // TC_WINDOWS_BOOT
+/* output written to input_digest which must be at lease 32 bytes long */
+void hmac_blake2s (unsigned char *key, int keylen, unsigned char *input_digest, int len);
+void derive_key_blake2s (const unsigned char *pwd, int pwd_len, const unsigned char *salt, int salt_len, uint32 iterations, unsigned char *dk, int dklen);
+
+/* output written to d which must be at lease 32 bytes long */
+void hmac_sha256 (unsigned char *k, int lk, unsigned char *d, int ld);
+void derive_key_sha256 (const unsigned char *pwd, int pwd_len, const unsigned char *salt, int salt_len, uint32 iterations, unsigned char *dk, int dklen);
+
 #endif

 #if defined(__cplusplus)
--- a/src/Common/Tests.c
+++ b/src/Common/Tests.c
@@ -1681,64 +1681,64 @@ BOOL test_pkcs5 ()
 		return FALSE;
 #endif
 	/* PKCS-5 test 1 with HMAC-SHA-256 used as the PRF (https://tools.ietf.org/html/draft-josefsson-scrypt-kdf-00) */
-	derive_key_sha256 ((unsigned char*) "passwd", 6, (unsigned char*) "\x73\x61\x6C\x74", 4, 1, dk, 64);
+	derive_key_sha256 ((unsigned char*) "passwd", 6, (unsigned char*) "\x73\x61\x6C\x74", 4, 1, dk, 64, NULL);
 	if (memcmp (dk, "\x55\xac\x04\x6e\x56\xe3\x08\x9f\xec\x16\x91\xc2\x25\x44\xb6\x05\xf9\x41\x85\x21\x6d\xde\x04\x65\xe6\x8b\x9d\x57\xc2\x0d\xac\xbc\x49\xca\x9c\xcc\xf1\x79\xb6\x45\x99\x16\x64\xb3\x9d\x77\xef\x31\x7c\x71\xb8\x45\xb1\xe3\x0b\xd5\x09\x11\x20\x41\xd3\xa1\x97\x83", 64) != 0)
 		return FALSE;

 	/* PKCS-5 test 2 with HMAC-SHA-256 used as the PRF (https://stackoverflow.com/questions/5130513/pbkdf2-hmac-sha2-test-vectors) */
-	derive_key_sha256 ((unsigned char*) "password", 8, (unsigned char*) "\x73\x61\x6C\x74", 4, 2, dk, 32);
+	derive_key_sha256 ((unsigned char*) "password", 8, (unsigned char*) "\x73\x61\x6C\x74", 4, 2, dk, 32, NULL);
 	if (memcmp (dk, "\xae\x4d\x0c\x95\xaf\x6b\x46\xd3\x2d\x0a\xdf\xf9\x28\xf0\x6d\xd0\x2a\x30\x3f\x8e\xf3\xc2\x51\xdf\xd6\xe2\xd8\x5a\x95\x47\x4c\x43", 32) != 0)
 		return FALSE;

 	/* PKCS-5 test 3 with HMAC-SHA-256 used as the PRF (MS CryptoAPI) */
-	derive_key_sha256 ((unsigned char*)"password", 8, (unsigned char*)"\x12\x34\x56\x78", 4, 5, dk, 4);
+	derive_key_sha256 ((unsigned char*)"password", 8, (unsigned char*)"\x12\x34\x56\x78", 4, 5, dk, 4, NULL);
 	if (memcmp (dk, "\xf2\xa0\x4f\xb2", 4) != 0)
 		return FALSE;

 	/* PKCS-5 test 4 with HMAC-SHA-256 used as the PRF (MS CryptoAPI) */
-	derive_key_sha256 ((unsigned char*)"password", 8, (unsigned char*)"\x12\x34\x56\x78", 4, 5, dk, 144);
+	derive_key_sha256 ((unsigned char*)"password", 8, (unsigned char*)"\x12\x34\x56\x78", 4, 5, dk, 144, NULL);
 	if (memcmp (dk, "\xf2\xa0\x4f\xb2\xd3\xe9\xa5\xd8\x51\x0b\x5c\x06\xdf\x70\x8e\x24\xe9\xc7\xd9\x15\x3d\x22\xcd\xde\xb8\xa6\xdb\xfd\x71\x85\xc6\x99\x32\xc0\xee\x37\x27\xf7\x24\xcf\xea\xa6\xac\x73\xa1\x4c\x4e\x52\x9b\x94\xf3\x54\x06\xfc\x04\x65\xa1\x0a\x24\xfe\xf0\x98\x1d\xa6\x22\x28\xeb\x24\x55\x74\xce\x6a\x3a\x28\xe2\x04\x3a\x59\x13\xec\x3f\xf2\xdb\xcf\x58\xdd\x53\xd9\xf9\x17\xf6\xda\x74\x06\x3c\x0b\x66\xf5\x0f\xf5\x58\xa3\x27\x52\x8c\x5b\x07\x91\xd0\x81\xeb\xb6\xbc\x30\x69\x42\x71\xf2\xd7\x18\x42\xbe\xe8\x02\x93\x70\x66\xad\x35\x65\xbc\xf7\x96\x8e\x64\xf1\xc6\x92\xda\xe0\xdc\x1f\xb5\xf4", 144) != 0)
 		return FALSE;

 	/* PKCS-5 test 1 with HMAC-SHA-512 used as the PRF */
-	derive_key_sha512 ((unsigned char*)"password", 8, (unsigned char*)"\x12\x34\x56\x78", 4, 5, dk, 4);
+	derive_key_sha512 ((unsigned char*)"password", 8, (unsigned char*)"\x12\x34\x56\x78", 4, 5, dk, 4, NULL);
 	if (memcmp (dk, "\x13\x64\xae\xf8", 4) != 0)
 		return FALSE;

 	/* PKCS-5 test 2 with HMAC-SHA-512 used as the PRF (derives a key longer than the underlying
 	hash output size and block size) */
-	derive_key_sha512 ((unsigned char*)"password", 8, (unsigned char*)"\x12\x34\x56\x78", 4, 5, dk, 144);
+	derive_key_sha512 ((unsigned char*)"password", 8, (unsigned char*)"\x12\x34\x56\x78", 4, 5, dk, 144, NULL);
 	if (memcmp (dk, "\x13\x64\xae\xf8\x0d\xf5\x57\x6c\x30\xd5\x71\x4c\xa7\x75\x3f\xfd\x00\xe5\x25\x8b\x39\xc7\x44\x7f\xce\x23\x3d\x08\x75\xe0\x2f\x48\xd6\x30\xd7\x00\xb6\x24\xdb\xe0\x5a\xd7\x47\xef\x52\xca\xa6\x34\x83\x47\xe5\xcb\xe9\x87\xf1\x20\x59\x6a\xe6\xa9\xcf\x51\x78\xc6\xb6\x23\xa6\x74\x0d\xe8\x91\xbe\x1a\xd0\x28\xcc\xce\x16\x98\x9a\xbe\xfb\xdc\x78\xc9\xe1\x7d\x72\x67\xce\xe1\x61\x56\x5f\x96\x68\xe6\xe1\xdd\xf4\xbf\x1b\x80\xe0\x19\x1c\xf4\xc4\xd3\xdd\xd5\xd5\x57\x2d\x83\xc7\xa3\x37\x87\xf4\x4e\xe0\xf6\xd8\x6d\x65\xdc\xa0\x52\xa3\x13\xbe\x81\xfc\x30\xbe\x7d\x69\x58\x34\xb6\xdd\x41\xc6", 144) != 0)
 		return FALSE;

 #ifndef WOLFCRYPT_BACKEND
 	/* PKCS-5 test 1 with HMAC-BLAKE2s used as the PRF */
-	derive_key_blake2s ((unsigned char*)"password", 8, (unsigned char*)"\x12\x34\x56\x78", 4, 5, dk, 4);
+	derive_key_blake2s ((unsigned char*)"password", 8, (unsigned char*)"\x12\x34\x56\x78", 4, 5, dk, 4, NULL);
 	if (memcmp (dk, "\x8d\x51\xfa\x31", 4) != 0)
 		return FALSE;

 	/* PKCS-5 test 2 with HMAC-BLAKE2s used as the PRF (derives a key longer than the underlying hash) */
-	derive_key_blake2s ((unsigned char*)"password", 8, (unsigned char*)"\x12\x34\x56\x78", 4, 5, dk, 48);
+	derive_key_blake2s ((unsigned char*)"password", 8, (unsigned char*)"\x12\x34\x56\x78", 4, 5, dk, 48, NULL);
 	if (memcmp (dk, "\x8d\x51\xfa\x31\x46\x25\x37\x67\xa3\x29\x6b\x3c\x6b\xc1\x5d\xb2\xee\xe1\x6c\x28\x00\x26\xea\x08\x65\x9c\x12\xf1\x07\xde\x0d\xb9\x9b\x4f\x39\xfa\xc6\x80\x26\xb1\x8f\x8e\x48\x89\x85\x2d\x24\x2d", 48) != 0)
 		return FALSE;

 	/* PKCS-5 test 1 with HMAC-Whirlpool used as the PRF */
-	derive_key_whirlpool ((unsigned char*)"password", 8, (unsigned char*)"\x12\x34\x56\x78", 4, 5, dk, 4);
+	derive_key_whirlpool ((unsigned char*)"password", 8, (unsigned char*)"\x12\x34\x56\x78", 4, 5, dk, 4, NULL);
 	if (memcmp (dk, "\x50\x7c\x36\x6f", 4) != 0)
 		return FALSE;

 	/* PKCS-5 test 2 with HMAC-Whirlpool used as the PRF (derives a key longer than the underlying hash) */
-	derive_key_whirlpool ((unsigned char*)"password", 8, (unsigned char*)"\x12\x34\x56\x78", 4, 5, dk, 96);
+	derive_key_whirlpool ((unsigned char*)"password", 8, (unsigned char*)"\x12\x34\x56\x78", 4, 5, dk, 96, NULL);
 	if (memcmp (dk, "\x50\x7c\x36\x6f\xee\x10\x2e\x9a\xe2\x8a\xd5\x82\x72\x7d\x27\x0f\xe8\x4d\x7f\x68\x7a\xcf\xb5\xe7\x43\x67\xaa\x98\x93\x52\x2b\x09\x6e\x42\xdf\x2c\x59\x4a\x91\x6d\x7e\x10\xae\xb2\x1a\x89\x8f\xb9\x8f\xe6\x31\xa9\xd8\x9f\x98\x26\xf4\xda\xcd\x7d\x65\x65\xde\x10\x95\x91\xb4\x84\x26\xae\x43\xa1\x00\x5b\x1e\xb8\x38\x97\xa4\x1e\x4b\xd2\x65\x64\xbc\xfa\x1f\x35\x85\xdb\x4f\x97\x65\x6f\xbd\x24", 96) != 0)
 		return FALSE;

 	/* PKCS-5 test 1 with HMAC-STREEBOG used as the PRF */
-	derive_key_streebog ((unsigned char*)"password", 8, (unsigned char*)"\x12\x34\x56\x78", 4, 5, dk, 4);
+	derive_key_streebog ((unsigned char*)"password", 8, (unsigned char*)"\x12\x34\x56\x78", 4, 5, dk, 4, NULL);
 	if (memcmp (dk, "\xd0\x53\xa2\x30", 4) != 0)
 		return FALSE;

 	/* PKCS-5 test 2 with HMAC-STREEBOG used as the PRF (derives a key longer than the underlying hash) */
-	derive_key_streebog ((unsigned char*)"password", 8, (unsigned char*)"\x12\x34\x56\x78", 4, 5, dk, 96);
+	derive_key_streebog ((unsigned char*)"password", 8, (unsigned char*)"\x12\x34\x56\x78", 4, 5, dk, 96, NULL);
 	if (memcmp (dk, "\xd0\x53\xa2\x30\x6f\x45\x81\xeb\xbc\x06\x81\xc5\xe7\x53\xa8\x5d\xc7\xf1\x23\x33\x1e\xbe\x64\x2c\x3b\x0f\x26\xd7\x00\xe1\x95\xc9\x65\x26\xb1\x85\xbe\x1e\xe2\xf4\x9b\xfc\x6b\x14\x84\xda\x24\x61\xa0\x1b\x9e\x79\x5c\xee\x69\x6e\xf9\x25\xb1\x1d\xca\xa0\x31\xba\x02\x6f\x9e\x99\x0f\xdb\x25\x01\x5b\xf1\xc7\x10\x19\x53\x3b\x29\x3f\x18\x00\xd6\xfc\x85\x03\xdc\xf2\xe5\xe9\x5a\xb1\x1e\x61\xde", 96) != 0)
 		return FALSE;
 #endif
--- a/src/Common/Volumes.c
+++ b/src/Common/Volumes.c
@@ -194,6 +194,7 @@ int ReadVolumeHeader (BOOL bBoot, unsigned char *encryptedHeader, Password *pass
 	int i;
 	int iterationsCount = 0;
 	int memoryCost = 0;
+	LONG volatile abortKeyDerivation = 0;
 #endif
 	size_t queuedWorkItems = 0;

@@ -331,7 +332,7 @@ int ReadVolumeHeader (BOOL bBoot, unsigned char *encryptedHeader, Password *pass
 						iterationsCount = get_pkcs5_iteration_count (enqPkcs5Prf, pim, bBoot, &memoryCost);
 						EncryptionThreadPoolBeginKeyDerivation (keyDerivationCompletedEvent, noOutstandingWorkItemEvent,
 							&item->KeyReady, outstandingWorkItemCount, enqPkcs5Prf, keyInfo->userKey,
-							keyInfo->keyLength, keyInfo->salt, iterationsCount, memoryCost, item->DerivedKey);
+							keyInfo->keyLength, keyInfo->salt, iterationsCount, memoryCost, item->DerivedKey, &abortKeyDerivation);

 						++queuedWorkItems;
 						break;
@@ -382,35 +383,35 @@ KeyReady:	;
 			{
 			case SHA512:
 				derive_key_sha512 (keyInfo->userKey, keyInfo->keyLength, keyInfo->salt,
-					PKCS5_SALT_SIZE, keyInfo->noIterations, dk, GetMaxPkcs5OutSize());
+					PKCS5_SALT_SIZE, keyInfo->noIterations, dk, GetMaxPkcs5OutSize(), &abortKeyDerivation);
 				break;

 			case SHA256:
 				derive_key_sha256 (keyInfo->userKey, keyInfo->keyLength, keyInfo->salt,
-					PKCS5_SALT_SIZE, keyInfo->noIterations, dk, GetMaxPkcs5OutSize());
+					PKCS5_SALT_SIZE, keyInfo->noIterations, dk, GetMaxPkcs5OutSize(), &abortKeyDerivation);
 				break;

 #ifndef WOLFCRYPT_BACKEND
 			case BLAKE2S:
 				derive_key_blake2s (keyInfo->userKey, keyInfo->keyLength, keyInfo->salt,
-					PKCS5_SALT_SIZE, keyInfo->noIterations, dk, GetMaxPkcs5OutSize());
+					PKCS5_SALT_SIZE, keyInfo->noIterations, dk, GetMaxPkcs5OutSize(), &abortKeyDerivation);
 				break;

 			case WHIRLPOOL:
 				derive_key_whirlpool (keyInfo->userKey, keyInfo->keyLength, keyInfo->salt,
-					PKCS5_SALT_SIZE, keyInfo->noIterations, dk, GetMaxPkcs5OutSize());
+					PKCS5_SALT_SIZE, keyInfo->noIterations, dk, GetMaxPkcs5OutSize(), &abortKeyDerivation);
 				break;


 			case STREEBOG:
 				derive_key_streebog(keyInfo->userKey, keyInfo->keyLength, keyInfo->salt,
-					PKCS5_SALT_SIZE, keyInfo->noIterations, dk, GetMaxPkcs5OutSize());
+					PKCS5_SALT_SIZE, keyInfo->noIterations, dk, GetMaxPkcs5OutSize(), &abortKeyDerivation);
 				break;


 			case ARGON2:
 				derive_key_argon2(keyInfo->userKey, keyInfo->keyLength, keyInfo->salt,
-					PKCS5_SALT_SIZE, keyInfo->noIterations, keyInfo->memoryCost, dk, GetMaxPkcs5OutSize());
+					PKCS5_SALT_SIZE, keyInfo->noIterations, keyInfo->memoryCost, dk, GetMaxPkcs5OutSize(), &abortKeyDerivation);
 				break;
 #endif	
                        default:
@@ -615,6 +616,12 @@ KeyReady:	;
 				}

 				status = ERR_SUCCESS;
+
+				if ((selected_pkcs5_prf == 0) && (encryptionThreadCount > 1))
+				{
+					// Signal other threads to stop
+					InterlockedExchange(&abortKeyDerivation, 1);
+				}
 				goto ret;
 			}
 		}
@@ -622,6 +629,8 @@ KeyReady:	;
 	status = ERR_PASSWORD_WRONG;

 err:
+	// Signal threads to stop
+	InterlockedExchange(&abortKeyDerivation, 1);
 	if (cryptoInfo != retHeaderCryptoInfo)
 	{
 		crypto_close(cryptoInfo);
@@ -640,19 +649,33 @@ ret:
 #if !defined(_UEFI)
 	if ((selected_pkcs5_prf == 0) && (encryptionThreadCount > 1))
 	{
-		EncryptionThreadPoolBeginReadVolumeHeaderFinalization (keyDerivationCompletedEvent, noOutstandingWorkItemEvent, outstandingWorkItemCount, 
-			keyInfoBuffer, keyInfoBufferSize, 
-			keyDerivationWorkItems, keyDerivationWorkItemsSize);
-	}
-	else
+		// Wait for all outstanding threads to finish or cancel
+		TC_WAIT_EVENT(*noOutstandingWorkItemEvent);
+		// Cleanup is now synchronous because we already waited for all threads to stop.
+		// The asynchronous finalization is no longer needed.
+#if !defined(DEVICE_DRIVER)
+		CloseHandle(*keyDerivationCompletedEvent);
+		CloseHandle(*noOutstandingWorkItemEvent);
 #endif
+		TCfree(keyDerivationCompletedEvent);
+		TCfree(noOutstandingWorkItemEvent);
+		TCfree(outstandingWorkItemCount);
+		if (keyDerivationWorkItems)
 		{
+			burn(keyDerivationWorkItems, keyDerivationWorkItemsSize);
+#if !defined(DEVICE_DRIVER)
+			VirtualUnlock(keyDerivationWorkItems, keyDerivationWorkItemsSize);
+#endif
+			TCfree(keyDerivationWorkItems);
+		}
+	}
+#endif
+
 	burn (keyInfo, sizeof (KEY_INFO));
-#if !defined(DEVICE_DRIVER) && !defined(_UEFI)
+#if !defined(DEVICE_DRIVER)
 	VirtualUnlock (keyInfoBuffer, keyInfoBufferSize);
 #endif
 	TCfree(keyInfoBuffer);
-	}
 	return status;
 }

@@ -1018,33 +1041,33 @@ int CreateVolumeHeaderInMemory (HWND hwndDlg, BOOL bBoot, unsigned char *header,
 		{
 		case SHA512:
 			derive_key_sha512 (keyInfo.userKey, keyInfo.keyLength, keyInfo.salt,
-				PKCS5_SALT_SIZE, keyInfo.noIterations, dk, GetMaxPkcs5OutSize());
+				PKCS5_SALT_SIZE, keyInfo.noIterations, dk, GetMaxPkcs5OutSize(), NULL);
 			break;

 		case SHA256:
 			derive_key_sha256 (keyInfo.userKey, keyInfo.keyLength, keyInfo.salt,
-				PKCS5_SALT_SIZE, keyInfo.noIterations, dk, GetMaxPkcs5OutSize());
+				PKCS5_SALT_SIZE, keyInfo.noIterations, dk, GetMaxPkcs5OutSize(), NULL);
 			break;

        #ifndef WOLFCRYPT_BACKEND
 		case BLAKE2S:
 			derive_key_blake2s (keyInfo.userKey, keyInfo.keyLength, keyInfo.salt,
-				PKCS5_SALT_SIZE, keyInfo.noIterations, dk, GetMaxPkcs5OutSize());
+				PKCS5_SALT_SIZE, keyInfo.noIterations, dk, GetMaxPkcs5OutSize(), NULL);
 			break;

 		case WHIRLPOOL:
 			derive_key_whirlpool (keyInfo.userKey, keyInfo.keyLength, keyInfo.salt,
-				PKCS5_SALT_SIZE, keyInfo.noIterations, dk, GetMaxPkcs5OutSize());
+				PKCS5_SALT_SIZE, keyInfo.noIterations, dk, GetMaxPkcs5OutSize(), NULL);
 			break;

 		case STREEBOG:
 			derive_key_streebog(keyInfo.userKey, keyInfo.keyLength, keyInfo.salt,
-				PKCS5_SALT_SIZE, keyInfo.noIterations, dk, GetMaxPkcs5OutSize());
+				PKCS5_SALT_SIZE, keyInfo.noIterations, dk, GetMaxPkcs5OutSize(), NULL);
 			break;

 		case ARGON2:
 			derive_key_argon2(keyInfo.userKey, keyInfo.keyLength, keyInfo.salt,
-				PKCS5_SALT_SIZE, keyInfo.noIterations, keyInfo.memoryCost, dk, GetMaxPkcs5OutSize());
+				PKCS5_SALT_SIZE, keyInfo.noIterations, keyInfo.memoryCost, dk, GetMaxPkcs5OutSize(), NULL);
 			break;
        #endif
 		default:
--- a/src/Crypto/Argon2/include/argon2.h
+++ b/src/Crypto/Argon2/include/argon2.h
@@ -165,7 +165,9 @@ typedef enum Argon2_ErrorCodes {

    ARGON2_DECODING_LENGTH_FAIL = -34,

-    ARGON2_VERIFY_MISMATCH = -35
+    ARGON2_VERIFY_MISMATCH = -35,
+
+    ARGON2_OPERATION_CANCELLED = -36
 } argon2_error_codes;

 /* Memory allocator types --- for external allocation */
@@ -222,6 +224,9 @@ typedef struct Argon2_Context {

    uint32_t version; /* version number */

+    /* Cancellation token for VeraCrypt */
+    long volatile *pAbortKeyDerivation;
+
    allocate_fptr allocate_cbk; /* pointer to memory allocator */
    deallocate_fptr free_cbk;   /* pointer to memory deallocator */

@@ -275,20 +280,20 @@ ARGON2_PUBLIC int argon2i_hash_raw(const uint32_t t_cost, const uint32_t m_cost,
                                   const uint32_t parallelism, const void *pwd,
                                   const size_t pwdlen, const void *salt,
                                   const size_t saltlen, void *hash,
-                                   const size_t hashlen);
+                                   const size_t hashlen, long volatile* pAbortKeyDerivation);

 ARGON2_PUBLIC int argon2d_hash_raw(const uint32_t t_cost, const uint32_t m_cost,
                                   const uint32_t parallelism, const void *pwd,
                                   const size_t pwdlen, const void *salt,
                                   const size_t saltlen, void *hash,
-                                   const size_t hashlen);
+                                   const size_t hashlen, long volatile* pAbortKeyDerivation);

 ARGON2_PUBLIC int argon2id_hash_raw(const uint32_t t_cost,
                                    const uint32_t m_cost,
                                    const uint32_t parallelism, const void *pwd,
                                    const size_t pwdlen, const void *salt,
                                    const size_t saltlen, void *hash,
-                                    const size_t hashlen);
+                                    const size_t hashlen, long volatile *pAbortKeyDerivation);

 /* generic function underlying the above ones */
 ARGON2_PUBLIC int argon2_hash(const uint32_t t_cost, const uint32_t m_cost,
@@ -296,7 +301,7 @@ ARGON2_PUBLIC int argon2_hash(const uint32_t t_cost, const uint32_t m_cost,
                              const size_t pwdlen, const void *salt,
                              const size_t saltlen, void *hash,
                              const size_t hashlen, argon2_type type,
-                              const uint32_t version);
+                              const uint32_t version, long volatile *pAbortKeyDerivation);

 /**
 * Argon2d: Version of Argon2 that picks memory blocks depending
--- a/src/Crypto/Argon2/src/argon2.c
+++ b/src/Crypto/Argon2/src/argon2.c
@@ -24,6 +24,9 @@
 #include <memory.h>
 #include <stdlib.h>
 #endif
+#include "Crypto/config.h"
+#include "Crypto/cpu.h"
+#include "Crypto/misc.h"

 const char *argon2_type2string(argon2_type type, int uppercase) {
    switch (type) {
@@ -91,6 +94,9 @@ int argon2_ctx(argon2_context *context, argon2_type type) {
    result = fill_memory_blocks(&instance);

    if (ARGON2_OK != result) {
+        // If cancelled, we must still free the allocated memory!
+        free_memory(context, (uint8_t *)instance.memory,
+                    instance.memory_blocks, sizeof(block));
        return result;
    }
    /* 5. Finalization */
@@ -103,7 +109,7 @@ int argon2_hash(const uint32_t t_cost, const uint32_t m_cost,
                const uint32_t parallelism, const void *pwd,
                const size_t pwdlen, const void *salt, const size_t saltlen,
                void *hash, const size_t hashlen, argon2_type type,
-                const uint32_t version){
+                const uint32_t version, long volatile *pAbortKeyDerivation){

    argon2_context context;
    int result;
@@ -148,6 +154,7 @@ int argon2_hash(const uint32_t t_cost, const uint32_t m_cost,
    context.free_cbk = NULL;
    context.flags = ARGON2_DEFAULT_FLAGS;
    context.version = version;
+    context.pAbortKeyDerivation = pAbortKeyDerivation;

    result = argon2_ctx(&context, type);

@@ -171,28 +178,28 @@ int argon2_hash(const uint32_t t_cost, const uint32_t m_cost,
 int argon2i_hash_raw(const uint32_t t_cost, const uint32_t m_cost,
                     const uint32_t parallelism, const void *pwd,
                     const size_t pwdlen, const void *salt,
-                     const size_t saltlen, void *hash, const size_t hashlen) {
+                     const size_t saltlen, void *hash, const size_t hashlen, long volatile* pAbortKeyDerivation) {

    return argon2_hash(t_cost, m_cost, parallelism, pwd, pwdlen, salt, saltlen,
-                       hash, hashlen, Argon2_i, ARGON2_VERSION_NUMBER);
+                       hash, hashlen, Argon2_i, ARGON2_VERSION_NUMBER, pAbortKeyDerivation);
 }

 int argon2d_hash_raw(const uint32_t t_cost, const uint32_t m_cost,
                     const uint32_t parallelism, const void *pwd,
                     const size_t pwdlen, const void *salt,
-                     const size_t saltlen, void *hash, const size_t hashlen) {
+                     const size_t saltlen, void *hash, const size_t hashlen, long volatile* pAbortKeyDerivation) {

    return argon2_hash(t_cost, m_cost, parallelism, pwd, pwdlen, salt, saltlen,
-                       hash, hashlen, Argon2_d, ARGON2_VERSION_NUMBER);
+                       hash, hashlen, Argon2_d, ARGON2_VERSION_NUMBER, pAbortKeyDerivation);
 }

 int argon2id_hash_raw(const uint32_t t_cost, const uint32_t m_cost,
                      const uint32_t parallelism, const void *pwd,
                      const size_t pwdlen, const void *salt,
-                      const size_t saltlen, void *hash, const size_t hashlen) {
+                      const size_t saltlen, void *hash, const size_t hashlen, long volatile *pAbortKeyDerivation) {
    return argon2_hash(t_cost, m_cost, parallelism, pwd, pwdlen, salt, saltlen,
                       hash, hashlen, Argon2_id,
-                       ARGON2_VERSION_NUMBER);
+                       ARGON2_VERSION_NUMBER, pAbortKeyDerivation);
 }

 int argon2d_ctx(argon2_context *context) {
--- a/src/Crypto/Argon2/src/blake2/blake2-impl.h
+++ b/src/Crypto/Argon2/src/blake2/blake2-impl.h
@@ -140,6 +140,8 @@ static BLAKE2_INLINE void store48(void *dst, uint64_t w) {
    *p++ = (uint8_t)w;
 }

+/* removed since they are defined in VeraCrypt headers */
+/*
 static BLAKE2_INLINE uint32_t rotr32(const uint32_t w, const unsigned c) {
    return (w >> c) | (w << (32 - c));
 }
@@ -147,7 +149,7 @@ static BLAKE2_INLINE uint32_t rotr32(const uint32_t w, const unsigned c) {
 static BLAKE2_INLINE uint64_t rotr64(const uint64_t w, const unsigned c) {
    return (w >> c) | (w << (64 - c));
 }
-
+*/
 void clear_internal_memory(void *v, size_t n);

 #endif
--- a/src/Crypto/Argon2/src/blake2/blake2b.c
+++ b/src/Crypto/Argon2/src/blake2/blake2b.c
@@ -15,6 +15,9 @@
 * software. If not, they may be obtained at the above URLs.
 */

+#include "Crypto/config.h"
+#include "Crypto/cpu.h"
+#include "Crypto/misc.h"
 #include "blake2.h"
 #include "blake2-impl.h"

--- a/src/Crypto/Argon2/src/blake2/blamka-round-opt.h
+++ b/src/Crypto/Argon2/src/blake2/blamka-round-opt.h
@@ -20,13 +20,14 @@

 #include "blake2-impl.h"

-#include <emmintrin.h>
+/* remove to solve build errors under Windows Driver since */
+//#include <emmintrin.h>
 #if defined(__SSSE3__)
-#include <tmmintrin.h> /* for _mm_shuffle_epi8 and _mm_alignr_epi8 */
+//#include <tmmintrin.h> /* for _mm_shuffle_epi8 and _mm_alignr_epi8 */
 #endif

 #if defined(__XOP__) && (defined(__GNUC__) || defined(__clang__))
-#include <x86intrin.h>
+//#include <x86intrin.h>
 #endif

 #if !defined(__AVX512F__)
@@ -180,7 +181,7 @@ static BLAKE2_INLINE __m128i fBlaMka(__m128i x, __m128i y) {
    } while ((void)0, 0)
 #else /* __AVX2__ */

-#include <immintrin.h>
+//#include <immintrin.h>

 #define rotr32(x)   _mm256_shuffle_epi32(x, _MM_SHUFFLE(2, 3, 0, 1))
 #define rotr24(x)   _mm256_shuffle_epi8(x, _mm256_setr_epi8(3, 4, 5, 6, 7, 0, 1, 2, 11, 12, 13, 14, 15, 8, 9, 10, 3, 4, 5, 6, 7, 0, 1, 2, 11, 12, 13, 14, 15, 8, 9, 10))
@@ -329,7 +330,7 @@ static BLAKE2_INLINE __m128i fBlaMka(__m128i x, __m128i y) {

 #else /* __AVX512F__ */

-#include <immintrin.h>
+//#include <immintrin.h>

 #define ror64(x, n) _mm512_ror_epi64((x), (n))

--- a/src/Crypto/Argon2/src/core.c
+++ b/src/Crypto/Argon2/src/core.c
@@ -19,6 +19,13 @@


 #include "core.h"
+#include "Crypto/config.h"
+#if !defined(_UEFI)
+#include <memory.h>
+#include <stdlib.h>
+#endif
+#include "Crypto/cpu.h"
+#include "Crypto/misc.h"
 #include "blake2/blake2.h"
 #include "blake2/blake2-impl.h"

@@ -206,12 +213,16 @@ uint32_t index_alpha(const argon2_instance_t *instance,
 /* Single-threaded version for p=1 case */
 static int fill_memory_blocks_st(argon2_instance_t *instance) {
    uint32_t r, s, l;
+    int result = ARGON2_OK;

    for (r = 0; r < instance->passes; ++r) {
        for (s = 0; s < ARGON2_SYNC_POINTS; ++s) {
            for (l = 0; l < instance->lanes; ++l) {
                argon2_position_t position = {r, l, (uint8_t)s, 0};
-                fill_segment(instance, position);
+                result = fill_segment(instance, position);
+                if (result != ARGON2_OK) {
+                    return result;
+                }
            }
        }
 #ifdef GENKAT
--- a/src/Crypto/Argon2/src/core.h
+++ b/src/Crypto/Argon2/src/core.h
@@ -216,7 +216,7 @@ void finalize(const argon2_context *context, argon2_instance_t *instance);
 * @param position Current position
 * @pre all block pointers must be valid
 */
-void fill_segment(const argon2_instance_t *instance,
+int fill_segment(const argon2_instance_t *instance,
                  argon2_position_t position);

 /*
--- a/src/Crypto/Argon2/src/opt_avx2.asm
+++ b/src/Crypto/Argon2/src/opt_avx2.asm
@@ -358,6 +358,8 @@ fill_block:
    jne     .L5
    jmp     .L4

+
+align 16
 next_addresses:
    push    rdi
    push    rbx
@@ -386,6 +388,8 @@ next_addresses:
    pop     rdi
    ret

+
+align 16
 global fill_segment_avx2
 fill_segment_avx2:
    push    r15
@@ -403,7 +407,7 @@ fill_segment_avx2:
    vmovdqu [rsp+48], xmm1
    and     r14, -32
    test    rcx, rcx
-    je      .L44
+    je      .L37
    mov     edx, dword [rcx+36]
    cmp     edx, 1
    je      .L18
@@ -413,47 +417,54 @@ fill_segment_avx2:
    je      .L19
    mov     ebp, dword [rsp+52]
    test    r12d, r12d
-    jne     .L37
-    xor     r12d, r12d
-    test    al, al
-    sete    r12b
+    je      .L51
    xor     r15d, r15d
-    add     r12d, r12d
+    xor     r12d, r12d
 .L20:
-    mov     r8d, dword [rbx+24]
-    mov     r9d, dword [rbx+20]
+    mov     ecx, dword [rbx+24]
+    mov     r8d, dword [rbx+20]
    xor     edx, edx
-    mov     ecx, 128
    mov     rdi, r14
-    imul    ebp, r8d
-    imul    eax, r9d
+    imul    ebp, ecx
+    imul    eax, r8d
    add     ebp, r12d
    add     ebp, eax
    mov     eax, ebp
    lea     r13d, [rbp-1]
-    div     r8d
-    lea     eax, [rbp+r8-1]
+    div     ecx
+    lea     eax, [rbp+rcx-1]
+    mov     ecx, 128
    test    edx, edx
    cmove   r13d, eax
+    lea     rax, [rsp+64]
+    mov     qword [rsp+40], rax
    mov     esi, r13d
    sal     rsi, 10
    add     rsi, qword [rbx]
    rep movsq
-    cmp     r12d, r9d
-    jnb     .L44
-    lea     rax, [rsp+1088]
-    mov     qword [rsp+32], rax
-    lea     rax, [rsp+64]
-    mov     qword [rsp+40], rax
-    jmp     .L35
-    align   16
-    align   8
-.L46:
+    cmp     r12d, r8d
+    jb      .L24
+    jmp     .L36
+
+align 16
+align 8
+.L53:
+    mov     esi, r12d
+    and     esi, 127
+    je      .L52
+.L29:
+    mov     edx, dword [rsp+48]
+    mov     eax, esi
+    mov     ecx, dword [rsp+52]
+    mov     r8, qword [rsp+64+rax*8]
+    test    edx, edx
+    jne     .L31
+.L54:
    cmp     byte [rsp+56], 0
-    jne     .L30
-    mov     rdi, rcx
+    jne     .L31
+    mov     rsi, rcx
    mov     r9d, 1
-.L31:
+.L32:
    lea     rdx, [rsp+48]
    mov     rcx, rbx
    mov     dword [rsp+60], r12d
@@ -461,7 +472,7 @@ fill_segment_avx2:
    mov     edx, dword [rbx+24]
    mov     r8, qword [rbx]
    mov     eax, eax
-    imul    rdx, rdi
+    imul    rdx, rsi
    add     rdx, rax
    mov     eax, ebp
    sal     rdx, 10
@@ -469,55 +480,80 @@ fill_segment_avx2:
    add     rdx, r8
    add     r8, rax
    cmp     dword [rbx+8], 16
-    je      .L32
+    je      .L33
    mov     eax, dword [rsp+48]
    test    eax, eax
-    je      .L32
+    je      .L33
    mov     r9d, 1
    mov     rcx, r14
-    call    fill_block
-.L34:
    add     r12d, 1
-    cmp     r12d, dword [rbx+20]
-    jnb     .L44
-    mov     r8d, dword [rbx+24]
    add     ebp, 1
-.L35:
+    call    fill_block
+    cmp     r12d, dword [rbx+20]
+    jnb     .L36
+.L24:
+    test    r12b, 63
+    jne     .L25
+    mov     rax, qword [rbx+48]
+    mov     rax, qword [rax+96]
+    test    rax, rax
+    je      .L25
+    mov     eax, dword [rax]
+    test    eax, eax
+    jne     .L41
+.L25:
    mov     eax, ebp
    xor     edx, edx
-    div     r8d
+    div     dword [rbx+24]
    cmp     edx, 1
-    je      .L25
+    je      .L26
    mov     eax, r13d
    add     r13d, 1
-.L26:
+.L27:
    test    r15d, r15d
-    je      .L27
-    mov     edi, r12d
-    and     edi, 127
-    je      .L45
-.L28:
-    mov     eax, edi
-    mov     r8, qword [rsp+64+rax*8]
-.L29:
+    jne     .L53
    mov     edx, dword [rsp+48]
+    sal     rax, 10
+    add     rax, qword [rbx]
+    mov     r8, qword [rax]
    mov     ecx, dword [rsp+52]
    test    edx, edx
-    je      .L46
-.L30:
-    mov     edi, dword [rbx+28]
+    je      .L54
+.L31:
+    mov     esi, dword [rbx+28]
    mov     rax, r8
    xor     edx, edx
    xor     r9d, r9d
    shr     rax, 32
-    div     rdi
+    div     rsi
    cmp     rdx, rcx
-    mov     rdi, rdx
+    mov     rsi, rdx
    sete    r9b
-    jmp     .L31
-    align   16
-    align   8
-.L44:
+    jmp     .L32
+
+align 16
+align 8
+.L51:
+    xor     r12d, r12d
+    test    al, al
+    sete    r12b
+    xor     r15d, r15d
+    add     r12d, r12d
+    jmp     .L20
+
+align 16
+align 8
+.L33:
+    xor     r9d, r9d
+    mov     rcx, r14
+    add     r12d, 1
+    add     ebp, 1
+    call    fill_block
+    cmp     r12d, dword [rbx+20]
+    jb      .L24
+.L36:
+    xor     eax, eax
+.L16:
    add     rsp, 3160
    pop     rbx
    pop     rsi
@@ -528,51 +564,35 @@ fill_segment_avx2:
    pop     r14
    pop     r15
    ret
-    align   16
-    align   8
+
+align 16
+align 8
+.L26:
+    lea     eax, [rbp-1]
+    mov     r13d, ebp
+    jmp     .L27
+
+align 16
+align 8
+.L52:
+    mov     rcx, qword [rsp+40]
+    lea     rdx, [rsp+1088]
+    call    next_addresses
+    jmp     .L29
+
+align 16
+align 8
 .L19:
    test    r12d, r12d
-    jne     .L47
+    jne     .L55
    cmp     al, 1
    jbe     .L18
    mov     ebp, dword [rsp+52]
    xor     r15d, r15d
    jmp     .L20
-    align   16
-    align   8
-.L32:
-    xor     r9d, r9d
-    mov     rcx, r14
-    call    fill_block
-    jmp     .L34
-    align   16
-    align   8
-.L27:
-    sal     rax, 10
-    add     rax, qword [rbx]
-    mov     r8, qword [rax]
-    jmp     .L29
-    align   16
-    align   8
-.L25:
-    lea     eax, [rbp-1]
-    mov     r13d, ebp
-    jmp     .L26
-    align   16
-    align   8
-.L45:
-    mov     rdx, qword [rsp+32]
-    mov     rcx, qword [rsp+40]
-    call    next_addresses
-    jmp     .L28
-    align   16
-    align   8
-.L37:
-    xor     r15d, r15d
-    xor     r12d, r12d
-    jmp     .L20
-    align   16
-    align   8
+
+align 16
+align 8
 .L18:
    xor     edx, edx
    lea     rcx, [rsp+1088]
@@ -583,18 +603,18 @@ fill_segment_avx2:
    vpinsrd xmm0, xmm2, dword [rbx+36], 1
    movzx   edx, byte [rsp+56]
    mov     qword [rsp+1088], rax
-    mov     ecx, dword [rbx+16]
+    mov     edi, dword [rbx+16]
    mov     r12, rax
    mov     eax, dword [rsp+52]
    vpmovzxdq xmm0, xmm0
    mov     qword [rsp+1104], rdx
-    mov     qword [rsp+1112], rcx
+    mov     qword [rsp+1112], rdi
    mov     qword [rsp+1096], rax
    mov     rbp, rax
    mov     rax, rdx
    vmovdqu [rsp+1120], xmm0
    test    r12d, r12d
-    jne     .L38
+    jne     .L39
    test    dl, dl
    jne     .L20
    lea     rcx, [rsp+64]
@@ -604,19 +624,28 @@ fill_segment_avx2:
    mov     ebp, dword [rsp+52]
    movzx   eax, byte [rsp+56]
    jmp     .L20
-    align   16
-    align   8
-.L47:
+
+align 16
+align 8
+.L41:
+    mov     eax, -36
+    jmp     .L16
+
+.L55:
    mov     ebp, dword [rsp+52]
    xor     r15d, r15d
    xor     r12d, r12d
    jmp     .L20
-    align   16
-    align   8
-.L38:
+
+.L39:
    xor     r12d, r12d
    jmp     .L20

+.L37:
+    mov     eax, -25
+    jmp     .L16
+
+
 section .rdata align=32
 LC0:
    db 3,4,5,6,7,0,1,2,11,12,13,14,15,8,9,10
@@ -626,8 +655,8 @@ LC1:
    db 2,3,4,5,6,7,0,1,10,11,12,13,14,15,8,9
    db 2,3,4,5,6,7,0,1,10,11,12,13,14,15,8,9

-; External functions
+; External symbols
 extern index_alpha
 extern init_block_value

-; end of file
+; End of file
--- a/src/Crypto/Argon2/src/opt_avx2.c
+++ b/src/Crypto/Argon2/src/opt_avx2.c
@@ -20,6 +20,9 @@

 #include "argon2.h"
 #include "core.h"
+#include "Crypto/config.h"
+#include "Crypto/cpu.h"
+#include "Crypto/misc.h"

 #if defined(__AVX2__)

@@ -88,7 +91,7 @@ static void next_addresses(block *address_block, block *input_block) {
    fill_block(zero2_block, address_block, address_block, 0);
 }

-void fill_segment_avx2(const argon2_instance_t *instance,
+int fill_segment_avx2(const argon2_instance_t *instance,
                  argon2_position_t position) {
    block *ref_block = NULL, *curr_block = NULL;
    block address_block, input_block;
@@ -99,7 +102,7 @@ void fill_segment_avx2(const argon2_instance_t *instance,
    int data_independent_addressing;

    if (instance == NULL) {
-        return;
+        return ARGON2_INCORRECT_PARAMETER;
    }

    data_independent_addressing =
@@ -145,6 +148,12 @@ void fill_segment_avx2(const argon2_instance_t *instance,

    for (i = starting_index; i < instance->segment_length;
         ++i, ++curr_offset, ++prev_offset) {
+        // Check every 64 blocks. This is a good balance for responsiveness.
+        if ((i & 63) == 0 && instance->context_ptr->pAbortKeyDerivation &&
+            *instance->context_ptr->pAbortKeyDerivation)
+        {
+            return ARGON2_OPERATION_CANCELLED; // Return cancellation code
+        }
        /*1.1 Rotating prev_offset if needed */
        if (curr_offset % instance->lane_length == 1) {
            prev_offset = curr_offset - 1;
@@ -191,11 +200,13 @@ void fill_segment_avx2(const argon2_instance_t *instance,
            }
        }
    }
+    return ARGON2_OK;
 }
 #else
-void fill_segment_avx2(const argon2_instance_t* instance,
+int fill_segment_avx2(const argon2_instance_t* instance,
    argon2_position_t position) {
    (void)instance;
    (void)position;
+    return ARGON2_INCORRECT_PARAMETER; /* AVX2 not available */
 }
 #endif
--- a/src/Crypto/Argon2/src/opt_sse2.c
+++ b/src/Crypto/Argon2/src/opt_sse2.c
@@ -19,12 +19,14 @@

 #include "argon2.h"
 #include "core.h"
+#include "Crypto/config.h"
+#include "Crypto/cpu.h"
+#include "Crypto/misc.h"

 #if CRYPTOPP_BOOL_SSE2_INTRINSICS_AVAILABLE

 #include "blake2/blake2.h"
 #include "blake2/blamka-round-opt.h"
-#include "Crypto/config.h"

 /*
 * Function fills a new memory block and optionally XORs the old block over the new one.
@@ -91,7 +93,7 @@ static void next_addresses(block *address_block, block *input_block) {
    fill_block(zero2_block, address_block, address_block, 0);
 }

-void fill_segment_sse2(const argon2_instance_t *instance,
+int fill_segment_sse2(const argon2_instance_t *instance,
                  argon2_position_t position) {
    block *ref_block = NULL, *curr_block = NULL;
    block address_block, input_block;
@@ -102,7 +104,7 @@ void fill_segment_sse2(const argon2_instance_t *instance,
    int data_independent_addressing;

    if (instance == NULL) {
-        return;
+        return ARGON2_INCORRECT_PARAMETER;
    }

    data_independent_addressing =
@@ -148,6 +150,13 @@ void fill_segment_sse2(const argon2_instance_t *instance,

    for (i = starting_index; i < instance->segment_length;
         ++i, ++curr_offset, ++prev_offset) {
+
+        // Check every 64 blocks. This is a good balance for responsiveness.
+        if ((i & 63) == 0 && instance->context_ptr->pAbortKeyDerivation &&
+            *instance->context_ptr->pAbortKeyDerivation)
+        {
+            return ARGON2_OPERATION_CANCELLED; // Return cancellation code
+        }
        /*1.1 Rotating prev_offset if needed */
        if (curr_offset % instance->lane_length == 1) {
            prev_offset = curr_offset - 1;
@@ -194,11 +203,13 @@ void fill_segment_sse2(const argon2_instance_t *instance,
            }
        }
    }
+    return ARGON2_OK;
 }
 #else
-void fill_segment_sse2(const argon2_instance_t* instance,
+int fill_segment_sse2(const argon2_instance_t* instance,
    argon2_position_t position) {
    (void)instance;
    (void)position;
+    return ARGON2_INCORRECT_PARAMETER; // SSE2 not available
 }
 #endif
--- a/src/Crypto/Argon2/src/ref.c
+++ b/src/Crypto/Argon2/src/ref.c
@@ -89,7 +89,7 @@ static void next_addresses(block *address_block, block *input_block,
    fill_block(zero_block, address_block, address_block, 0);
 }

-void fill_segment_ref(const argon2_instance_t *instance,
+int fill_segment_ref(const argon2_instance_t *instance,
                  argon2_position_t position) {
    block *ref_block = NULL, *curr_block = NULL;
    block address_block, input_block, zero_block;
@@ -100,7 +100,7 @@ void fill_segment_ref(const argon2_instance_t *instance,
    int data_independent_addressing;

    if (instance == NULL) {
-        return;
+        return ARGON2_INCORRECT_PARAMETER;
    }

    data_independent_addressing =
@@ -145,6 +145,11 @@ void fill_segment_ref(const argon2_instance_t *instance,

    for (i = starting_index; i < instance->segment_length;
         ++i, ++curr_offset, ++prev_offset) {
+        if ((i & 63) == 0 && instance->context_ptr->pAbortKeyDerivation &&
+            *instance->context_ptr->pAbortKeyDerivation)
+        {
+            return ARGON2_OPERATION_CANCELLED; // Return cancellation code
+        }
        /*1.1 Rotating prev_offset if needed */
        if (curr_offset % instance->lane_length == 1) {
            prev_offset = curr_offset - 1;
@@ -193,29 +198,30 @@ void fill_segment_ref(const argon2_instance_t *instance,
            }
        }
    }
+    return ARGON2_OK;
 }

 #if CRYPTOPP_BOOL_X64 || CRYPTOPP_BOOL_X86 || CRYPTOPP_BOOL_X32
-extern void fill_segment_sse2(const argon2_instance_t* instance,
+extern int fill_segment_sse2(const argon2_instance_t* instance,
 	argon2_position_t position);
-extern void fill_segment_avx2(const argon2_instance_t* instance,
+extern int fill_segment_avx2(const argon2_instance_t* instance,
 	argon2_position_t position);
 #endif

-void fill_segment(const argon2_instance_t* instance,
+int fill_segment(const argon2_instance_t* instance,
    argon2_position_t position) {
 #if CRYPTOPP_BOOL_X64 || CRYPTOPP_BOOL_X86 || CRYPTOPP_BOOL_X32
 	if (HasSAVX2())
 	{
-		fill_segment_avx2(instance, position);
+		return fill_segment_avx2(instance, position);
 	}
 	else if (HasSSE2())
 	{
-		fill_segment_sse2(instance, position);
+		return fill_segment_sse2(instance, position);
 	}
 	else
 #endif
 	{
-		fill_segment_ref(instance, position);
+		return fill_segment_ref(instance, position);
 	}
 }
--- a/src/Crypto/Argon2/src/selftest.c
+++ b/src/Crypto/Argon2/src/selftest.c
@@ -92,6 +92,7 @@ static int argon2_selftest_ctx (void)
   context.free_cbk = NULL;
   context.flags = ARGON2_DEFAULT_FLAGS;
   context.version = ARGON2_VERSION_13;
+   context.pAbortKeyDerivation = NULL; /* No abort function */

   /* Test execution for Argon2d, Argon2i, Argon2id */

@@ -118,7 +119,7 @@ static int hashtest (uint32_t t, uint32_t m, uint32_t p, const char *pwd,
   uint8_t out[32];

   return (argon2_hash(t, 1 << m, p, pwd, strlen (pwd), salt, strlen (salt),
-           out, 32, type, ARGON2_VERSION_NUMBER) == ARGON2_OK &&
+           out, 32, type, ARGON2_VERSION_NUMBER, NULL) == ARGON2_OK &&
           memcmp (out, ref, 32) == 0) ? 0 : 1;
 }

--- a/src/Crypto/cpu.h
+++ b/src/Crypto/cpu.h
@@ -143,6 +143,8 @@ extern __m128 _mm_shuffle_ps(__m128 _A, __m128 _B, unsigned int _Imm8);
 extern __m128i _mm_srli_si128(__m128i _A, int _Imm);
 extern __m128i _mm_slli_si128(__m128i _A, int _Imm);
 extern __m128i _mm_setzero_si128();
+extern __m128i _mm_mul_epu32(__m128i _A, __m128i _B);
+extern __m128i _mm_slli_epi64(__m128i _A, int _Count);
 #define _mm_xor_si64      _m_pxor
 #define _mm_empty         _m_empty
 #define _MM_SHUFFLE(fp3,fp2,fp1,fp0) (((fp3) << 6) | ((fp2) << 4) | \
--- a/src/Volume/Pkcs5Kdf.cpp
+++ b/src/Volume/Pkcs5Kdf.cpp
@@ -74,51 +74,51 @@ namespace VeraCrypt
 	void Pkcs5HmacBlake2s_Boot::DeriveKey (const BufferPtr &key, const VolumePassword &password, const ConstBufferPtr &salt, int iterationCount) const
 	{
 		ValidateParameters (key, password, salt, iterationCount);
-		derive_key_blake2s (password.DataPtr(), (int) password.Size(), salt.Get(), (int) salt.Size(), iterationCount, key.Get(), (int) key.Size());
+		derive_key_blake2s (password.DataPtr(), (int) password.Size(), salt.Get(), (int) salt.Size(), iterationCount, key.Get(), (int) key.Size(), NULL);
 	}

 	void Pkcs5HmacBlake2s::DeriveKey (const BufferPtr &key, const VolumePassword &password, const ConstBufferPtr &salt, int iterationCount) const
 	{
 		ValidateParameters (key, password, salt, iterationCount);
-		derive_key_blake2s (password.DataPtr(), (int) password.Size(), salt.Get(), (int) salt.Size(), iterationCount, key.Get(), (int) key.Size());
+		derive_key_blake2s (password.DataPtr(), (int) password.Size(), salt.Get(), (int) salt.Size(), iterationCount, key.Get(), (int) key.Size(), NULL);
 	}
    #endif

 	void Pkcs5HmacSha256_Boot::DeriveKey (const BufferPtr &key, const VolumePassword &password, const ConstBufferPtr &salt, int iterationCount) const
 	{
 		ValidateParameters (key, password, salt, iterationCount);
-		derive_key_sha256 (password.DataPtr(), (int) password.Size(), salt.Get(), (int) salt.Size(), iterationCount, key.Get(), (int) key.Size());
+		derive_key_sha256 (password.DataPtr(), (int) password.Size(), salt.Get(), (int) salt.Size(), iterationCount, key.Get(), (int) key.Size(), NULL);
 	}

 	void Pkcs5HmacSha256::DeriveKey (const BufferPtr &key, const VolumePassword &password, const ConstBufferPtr &salt, int iterationCount) const
 	{
 		ValidateParameters (key, password, salt, iterationCount);
-		derive_key_sha256 (password.DataPtr(), (int) password.Size(), salt.Get(), (int) salt.Size(), iterationCount, key.Get(), (int) key.Size());
+		derive_key_sha256 (password.DataPtr(), (int) password.Size(), salt.Get(), (int) salt.Size(), iterationCount, key.Get(), (int) key.Size(), NULL);
 	}

 	void Pkcs5HmacSha512::DeriveKey (const BufferPtr &key, const VolumePassword &password, const ConstBufferPtr &salt, int iterationCount) const
 	{
 		ValidateParameters (key, password, salt, iterationCount);
-		derive_key_sha512 (password.DataPtr(), (int) password.Size(), salt.Get(), (int) salt.Size(), iterationCount, key.Get(), (int) key.Size());
+		derive_key_sha512 (password.DataPtr(), (int) password.Size(), salt.Get(), (int) salt.Size(), iterationCount, key.Get(), (int) key.Size(), NULL);
 	}

    #ifndef WOLFCRYPT_BACKEND
 	void Pkcs5HmacWhirlpool::DeriveKey (const BufferPtr &key, const VolumePassword &password, const ConstBufferPtr &salt, int iterationCount) const
 	{
 		ValidateParameters (key, password, salt, iterationCount);
-		derive_key_whirlpool (password.DataPtr(), (int) password.Size(), salt.Get(), (int) salt.Size(), iterationCount, key.Get(), (int) key.Size());
+		derive_key_whirlpool (password.DataPtr(), (int) password.Size(), salt.Get(), (int) salt.Size(), iterationCount, key.Get(), (int) key.Size(), NULL);
 	}
 	
 	void Pkcs5HmacStreebog::DeriveKey (const BufferPtr &key, const VolumePassword &password, const ConstBufferPtr &salt, int iterationCount) const
 	{
 		ValidateParameters (key, password, salt, iterationCount);
-		derive_key_streebog (password.DataPtr(), (int) password.Size(), salt.Get(), (int) salt.Size(), iterationCount, key.Get(), (int) key.Size());
+		derive_key_streebog (password.DataPtr(), (int) password.Size(), salt.Get(), (int) salt.Size(), iterationCount, key.Get(), (int) key.Size(), NULL);
 	}
 	
 	void Pkcs5HmacStreebog_Boot::DeriveKey (const BufferPtr &key, const VolumePassword &password, const ConstBufferPtr &salt, int iterationCount) const
 	{
 		ValidateParameters (key, password, salt, iterationCount);
-		derive_key_streebog (password.DataPtr(), (int) password.Size(), salt.Get(), (int) salt.Size(), iterationCount, key.Get(), (int) key.Size());
+		derive_key_streebog (password.DataPtr(), (int) password.Size(), salt.Get(), (int) salt.Size(), iterationCount, key.Get(), (int) key.Size(), NULL);
 	}
    #endif
 }