diff --git a/plat/ti/k3/board/generic/include/board_def.h b/plat/ti/k3/board/generic/include/board_def.h
index c1a5966..0d45116 100644
--- a/plat/ti/k3/board/generic/include/board_def.h
+++ b/plat/ti/k3/board/generic/include/board_def.h
@@ -27,5 +27,6 @@
 
 #define PLAT_PROC_START_ID		32
 #define PLAT_PROC_DEVICE_START_ID	202
+#define PLAT_CLUSTER_DEVICE_START_ID	198
 
 #endif /* BOARD_DEF_H */
diff --git a/plat/ti/k3/common/drivers/ti_sci/ti_sci.c b/plat/ti/k3/common/drivers/ti_sci/ti_sci.c
index ac33278..e390efe 100644
--- a/plat/ti/k3/common/drivers/ti_sci/ti_sci.c
+++ b/plat/ti/k3/common/drivers/ti_sci/ti_sci.c
@@ -20,32 +20,10 @@
 #include "ti_sci_protocol.h"
 #include "ti_sci.h"
 
-/**
- * struct ti_sci_desc - Description of SoC integration
- * @host_id:		Host identifier representing the compute entity
- * @max_msg_size:	Maximum size of data per message that can be handled
- */
-struct ti_sci_desc {
-	uint8_t host_id;
-	int max_msg_size;
-};
-
-/**
- * struct ti_sci_info - Structure representing a TI SCI instance
- * @desc:	SoC description for this instance
- * @seq:	Seq id used for verification for tx and rx message
- */
-struct ti_sci_info {
-	const struct ti_sci_desc desc;
-	uint8_t seq;
-};
-
-static struct ti_sci_info info = {
-	.desc = {
-		.host_id = TI_SCI_HOST_ID,
-		.max_msg_size = TI_SCI_MAX_MESSAGE_SIZE,
-	},
-};
+#if USE_COHERENT_MEM
+__section("tzfw_coherent_mem")
+#endif
+static uint8_t message_sequence;
 
 /**
  * struct ti_sci_xfer - Structure representing a message flow
@@ -82,16 +60,16 @@
 	struct ti_sci_msg_hdr *hdr;
 
 	/* Ensure we have sane transfer sizes */
-	if (rx_message_size > info.desc.max_msg_size ||
-	    tx_message_size > info.desc.max_msg_size ||
+	if (rx_message_size > TI_SCI_MAX_MESSAGE_SIZE ||
+	    tx_message_size > TI_SCI_MAX_MESSAGE_SIZE ||
 	    rx_message_size < sizeof(*hdr) ||
 	    tx_message_size < sizeof(*hdr))
 		return -ERANGE;
 
 	hdr = (struct ti_sci_msg_hdr *)tx_buf;
-	hdr->seq = ++info.seq;
+	hdr->seq = ++message_sequence;
 	hdr->type = msg_type;
-	hdr->host = info.desc.host_id;
+	hdr->host = TI_SCI_HOST_ID;
 	hdr->flags = msg_flags | TI_SCI_FLAG_REQ_ACK_ON_PROCESSED;
 
 	xfer->tx_message.buf = tx_buf;
@@ -131,7 +109,7 @@
 		hdr = (struct ti_sci_msg_hdr *)msg->buf;
 
 		/* Sanity check for message response */
-		if (hdr->seq == info.seq)
+		if (hdr->seq == message_sequence)
 			break;
 		else
 			WARN("Message with sequence ID %u is not expected\n", hdr->seq);
@@ -141,9 +119,9 @@
 		return -EINVAL;
 	}
 
-	if (msg->len > info.desc.max_msg_size) {
+	if (msg->len > TI_SCI_MAX_MESSAGE_SIZE) {
 		ERROR("Unable to handle %lu xfer (max %d)\n",
-		      msg->len, info.desc.max_msg_size);
+		      msg->len, TI_SCI_MAX_MESSAGE_SIZE);
 		return -EINVAL;
 	}
 
@@ -425,13 +403,13 @@
 	int ret;
 
 	/* Ensure we have sane transfer size */
-	if (sizeof(req) > info.desc.max_msg_size)
+	if (sizeof(req) > TI_SCI_MAX_MESSAGE_SIZE)
 		return -ERANGE;
 
 	hdr = (struct ti_sci_msg_hdr *)&req;
-	hdr->seq = ++info.seq;
+	hdr->seq = ++message_sequence;
 	hdr->type = TI_SCI_MSG_SET_DEVICE_STATE;
-	hdr->host = info.desc.host_id;
+	hdr->host = TI_SCI_HOST_ID;
 	/* Setup with NORESPONSE flag to keep response queue clean */
 	hdr->flags = TI_SCI_FLAG_REQ_GENERIC_NORESPONSE;
 
@@ -1408,13 +1386,13 @@
 	int ret;
 
 	/* Ensure we have sane transfer size */
-	if (sizeof(req) > info.desc.max_msg_size)
+	if (sizeof(req) > TI_SCI_MAX_MESSAGE_SIZE)
 		return -ERANGE;
 
 	hdr = (struct ti_sci_msg_hdr *)&req;
-	hdr->seq = ++info.seq;
+	hdr->seq = ++message_sequence;
 	hdr->type = TISCI_MSG_SET_PROC_BOOT_CTRL;
-	hdr->host = info.desc.host_id;
+	hdr->host = TI_SCI_HOST_ID;
 	/* Setup with NORESPONSE flag to keep response queue clean */
 	hdr->flags = TI_SCI_FLAG_REQ_GENERIC_NORESPONSE;
 
@@ -1650,13 +1628,13 @@
 	int ret;
 
 	/* Ensure we have sane transfer size */
-	if (sizeof(req) > info.desc.max_msg_size)
+	if (sizeof(req) > TI_SCI_MAX_MESSAGE_SIZE)
 		return -ERANGE;
 
 	hdr = (struct ti_sci_msg_hdr *)&req;
-	hdr->seq = ++info.seq;
+	hdr->seq = ++message_sequence;
 	hdr->type = TISCI_MSG_WAIT_PROC_BOOT_STATUS;
-	hdr->host = info.desc.host_id;
+	hdr->host = TI_SCI_HOST_ID;
 	/* Setup with NORESPONSE flag to keep response queue clean */
 	hdr->flags = TI_SCI_FLAG_REQ_GENERIC_NORESPONSE;
 
diff --git a/plat/ti/k3/common/drivers/ti_sci/ti_sci_protocol.h b/plat/ti/k3/common/drivers/ti_sci/ti_sci_protocol.h
index a921e51..2d23f9a 100644
--- a/plat/ti/k3/common/drivers/ti_sci/ti_sci_protocol.h
+++ b/plat/ti/k3/common/drivers/ti_sci/ti_sci_protocol.h
@@ -563,8 +563,13 @@
 	uint32_t config_flags_clear;
 } __packed;
 
+/* ARMV8 Control Flags */
+#define PROC_BOOT_CTRL_FLAG_ARMV8_ACINACTM      0x00000001
+#define PROC_BOOT_CTRL_FLAG_ARMV8_AINACTS       0x00000002
+#define PROC_BOOT_CTRL_FLAG_ARMV8_L2FLUSHREQ    0x00000100
+
 /* R5 Control Flags */
-#define PROC_BOOT_CTRL_FLAG_R5_CORE_HALT                0x00000001
+#define PROC_BOOT_CTRL_FLAG_R5_CORE_HALT        0x00000001
 
 /**
  * struct ti_sci_msg_req_set_proc_boot_ctrl - Set Processor boot control flags
@@ -618,6 +623,8 @@
 /* ARMv8 Status Flags */
 #define PROC_BOOT_STATUS_FLAG_ARMV8_WFE			0x00000001
 #define PROC_BOOT_STATUS_FLAG_ARMV8_WFI			0x00000002
+#define PROC_BOOT_STATUS_FLAG_ARMV8_L2F_DONE		0x00000010
+#define PROC_BOOT_STATUS_FLAG_ARMV8_STANDBYWFIL2	0x00000020
 
 /* R5 Status Flags */
 #define PROC_BOOT_STATUS_FLAG_R5_WFE			0x00000001
diff --git a/plat/ti/k3/common/k3_psci.c b/plat/ti/k3/common/k3_psci.c
index de9cefe..cf0a21d 100644
--- a/plat/ti/k3/common/k3_psci.c
+++ b/plat/ti/k3/common/k3_psci.c
@@ -17,6 +17,10 @@
 #include <k3_gicv3.h>
 #include <ti_sci.h>
 
+#define CORE_PWR_STATE(state) ((state)->pwr_domain_state[MPIDR_AFFLVL0])
+#define CLUSTER_PWR_STATE(state) ((state)->pwr_domain_state[MPIDR_AFFLVL1])
+#define SYSTEM_PWR_STATE(state) ((state)->pwr_domain_state[PLAT_MAX_PWR_LVL])
+
 uintptr_t k3_sec_entrypoint;
 
 static void k3_cpu_standby(plat_local_state_t cpu_state)
@@ -37,30 +41,40 @@
 
 static int k3_pwr_domain_on(u_register_t mpidr)
 {
-	int core_id, proc, device, ret;
+	int core, proc_id, device_id, ret;
 
-	core_id = plat_core_pos_by_mpidr(mpidr);
-	if (core_id < 0) {
-		ERROR("Could not get target core id: %d\n", core_id);
+	core = plat_core_pos_by_mpidr(mpidr);
+	if (core < 0) {
+		ERROR("Could not get target core id: %d\n", core);
 		return PSCI_E_INTERN_FAIL;
 	}
 
-	proc = PLAT_PROC_START_ID + core_id;
-	device = PLAT_PROC_DEVICE_START_ID + core_id;
+	proc_id = PLAT_PROC_START_ID + core;
+	device_id = PLAT_PROC_DEVICE_START_ID + core;
 
-	ret = ti_sci_proc_request(proc);
+	ret = ti_sci_proc_request(proc_id);
 	if (ret) {
 		ERROR("Request for processor failed: %d\n", ret);
 		return PSCI_E_INTERN_FAIL;
 	}
 
-	ret = ti_sci_proc_set_boot_cfg(proc, k3_sec_entrypoint, 0, 0);
+	ret = ti_sci_proc_set_boot_cfg(proc_id, k3_sec_entrypoint, 0, 0);
 	if (ret) {
 		ERROR("Request to set core boot address failed: %d\n", ret);
 		return PSCI_E_INTERN_FAIL;
 	}
 
+	/* sanity check these are off before starting a core */
+	ret = ti_sci_proc_set_boot_ctrl(proc_id,
+			0, PROC_BOOT_CTRL_FLAG_ARMV8_L2FLUSHREQ |
+			   PROC_BOOT_CTRL_FLAG_ARMV8_AINACTS |
+			   PROC_BOOT_CTRL_FLAG_ARMV8_ACINACTM);
+	if (ret) {
+		ERROR("Request to clear boot configuration failed: %d\n", ret);
+		return PSCI_E_INTERN_FAIL;
+	}
+
-	ret = ti_sci_device_get(device);
+	ret = ti_sci_device_get(device_id);
 	if (ret) {
 		ERROR("Request to start core failed: %d\n", ret);
 		return PSCI_E_INTERN_FAIL;
@@ -71,17 +85,35 @@
 
 void k3_pwr_domain_off(const psci_power_state_t *target_state)
 {
-	int core_id, proc, device, ret;
+	int core, cluster, proc_id, device_id, cluster_id, ret;
+
+	/* At very least the local core should be powering down */
+	assert(CORE_PWR_STATE(target_state) == PLAT_MAX_OFF_STATE);
 
 	/* Prevent interrupts from spuriously waking up this cpu */
 	k3_gic_cpuif_disable();
 
-	core_id = plat_my_core_pos();
-	proc = PLAT_PROC_START_ID + core_id;
-	device = PLAT_PROC_DEVICE_START_ID + core_id;
+	core = plat_my_core_pos();
+	cluster = MPIDR_AFFLVL1_VAL(read_mpidr_el1());
+	proc_id = PLAT_PROC_START_ID + core;
+	device_id = PLAT_PROC_DEVICE_START_ID + core;
+	cluster_id = PLAT_CLUSTER_DEVICE_START_ID + (cluster * 2);
+
+	/*
+	 * If we are the last core in the cluster then we take a reference to
+	 * the cluster device so that it does not get shutdown before we
+	 * execute the entire cluster L2 cleaning sequence below.
+	 */
+	if (CLUSTER_PWR_STATE(target_state) == PLAT_MAX_OFF_STATE) {
+		ret = ti_sci_device_get(cluster_id);
+		if (ret) {
+			ERROR("Request to get cluster failed: %d\n", ret);
+			return;
+		}
+	}
 
 	/* Start by sending wait for WFI command */
-	ret = ti_sci_proc_wait_boot_status_no_wait(proc,
+	ret = ti_sci_proc_wait_boot_status_no_wait(proc_id,
 			/*
 			 * Wait maximum time to give us the best chance to get
 			 * to WFI before this command timeouts
@@ -95,11 +127,72 @@
 	}
 
 	/* Now queue up the core shutdown request */
-	ret = ti_sci_device_put_no_wait(device);
+	ret = ti_sci_device_put_no_wait(device_id);
 	if (ret) {
 		ERROR("Sending core shutdown message failed (%d)\n", ret);
 		return;
 	}
+
+	/* If our cluster is not going down we stop here */
+	if (CLUSTER_PWR_STATE(target_state) != PLAT_MAX_OFF_STATE)
+		return;
+
+	/* set AINACTS */
+	ret = ti_sci_proc_set_boot_ctrl_no_wait(proc_id,
+			PROC_BOOT_CTRL_FLAG_ARMV8_AINACTS, 0);
+	if (ret) {
+		ERROR("Sending set control message failed (%d)\n", ret);
+		return;
+	}
+
+	/* set L2FLUSHREQ */
+	ret = ti_sci_proc_set_boot_ctrl_no_wait(proc_id,
+			PROC_BOOT_CTRL_FLAG_ARMV8_L2FLUSHREQ, 0);
+	if (ret) {
+		ERROR("Sending set control message failed (%d)\n", ret);
+		return;
+	}
+
+	/* wait for L2FLUSHDONE*/
+	ret = ti_sci_proc_wait_boot_status_no_wait(proc_id,
+			UINT8_MAX, 2, UINT8_MAX, UINT8_MAX,
+			PROC_BOOT_STATUS_FLAG_ARMV8_L2F_DONE, 0, 0, 0);
+	if (ret) {
+		ERROR("Sending wait message failed (%d)\n", ret);
+		return;
+	}
+
+	/* clear L2FLUSHREQ */
+	ret = ti_sci_proc_set_boot_ctrl_no_wait(proc_id,
+			0, PROC_BOOT_CTRL_FLAG_ARMV8_L2FLUSHREQ);
+	if (ret) {
+		ERROR("Sending set control message failed (%d)\n", ret);
+		return;
+	}
+
+	/* set ACINACTM */
+	ret = ti_sci_proc_set_boot_ctrl_no_wait(proc_id,
+			PROC_BOOT_CTRL_FLAG_ARMV8_ACINACTM, 0);
+	if (ret) {
+		ERROR("Sending set control message failed (%d)\n", ret);
+		return;
+	}
+
+	/* wait for STANDBYWFIL2 */
+	ret = ti_sci_proc_wait_boot_status_no_wait(proc_id,
+			UINT8_MAX, 2, UINT8_MAX, UINT8_MAX,
+			PROC_BOOT_STATUS_FLAG_ARMV8_STANDBYWFIL2, 0, 0, 0);
+	if (ret) {
+		ERROR("Sending wait message failed (%d)\n", ret);
+		return;
+	}
+
+	/* Now queue up the cluster shutdown request */
+	ret = ti_sci_device_put_no_wait(cluster_id);
+	if (ret) {
+		ERROR("Sending cluster shutdown message failed (%d)\n", ret);
+		return;
+	}
 }
 
 void k3_pwr_domain_on_finish(const psci_power_state_t *target_state)