6 files changed, 969 insertions, 112 deletions

diff --git a/services/std_svc/lfa/bl31_lfa.c b/services/std_svc/lfa/bl31_lfa.c
index 2a75065e1..cb6f42e0b 100644
--- a/services/std_svc/lfa/bl31_lfa.c
+++ b/services/std_svc/lfa/bl31_lfa.c
@@ -1,28 +1,258 @@
 /*
- * Copyright (c) 2025, Arm Limited. All rights reserved.
+ * Copyright (c) 2025-2026, Arm Limited. All rights reserved.
  *
  * SPDX-License-Identifier: BSD-3-Clause
  */
 
+#include <assert.h>
+#include <errno.h>
+
+#include <common/build_message.h>
+#include <common/debug.h>
+#include <lib/xlat_tables/xlat_tables_v2.h>
+#include <plat/common/platform.h>
 #include <services/bl31_lfa.h>
+#include <services/lfa_holding_pen.h>
 #include <services/lfa_svc.h>
 
-static int lfa_bl31_prime(struct lfa_component_status *activation)
+static const uintptr_t TEXT_START = (uintptr_t)__TEXT_START__;
+
+/* Static function prototypes. */
+static int lfa_bl31_check(uintptr_t base_va, size_t size, uint64_t **entry);
+static int lfa_r_bl31_copy(uintptr_t va, uint64_t *entry, void *src, size_t size, bool reset);
+
+static inline size_t lfa_bl31_errata_size(void)
+{
+	return ERRATA_END - ERRATA_START;
+}
+
+static inline size_t lfa_bl31_cpu_ops_size(void)
+{
+	return CPU_OPS_END - CPU_OPS_START;
+}
+
+static int lfa_bl31_patch_region(struct lfa_component_status *activation,
+				 uintptr_t dst_va, size_t size, bool reset)
+{
+	uint64_t *entry = NULL;
+	size_t src_offset;
+	uintptr_t src_va;
+	uintptr_t patch_start;
+	uintptr_t patch_end;
+	size_t patch_size;
+
+	patch_start = dst_va & ~(PAGE_SIZE - 1U);
+	patch_end = (dst_va + size + PAGE_SIZE - 1U) & ~(PAGE_SIZE - 1U);
+	patch_size = patch_end - patch_start;
+
+	if ((patch_start < TEXT_START) ||
+	    ((patch_size > activation->image_size) &&
+	    (activation->image_size != 0))) {
+		ERROR("BL31 LFA: invalid patch region dst=0x%lx size=0x%zx image_size=0x%lx\n",
+		      dst_va, size, activation->image_size);
+		return LFA_CRITICAL_ERROR;
+	}
+
+	src_offset = patch_start - TEXT_START;
+	if ((activation->image_size != 0) &&
+	    ((src_offset > activation->image_size) ||
+	    (patch_size > (activation->image_size - src_offset)))) {
+		ERROR("BL31 LFA: staged image too small for patch region 0x%lx size 0x%zx "
+		      "aligned_start 0x%lx aligned_size 0x%zx\n",
+		      dst_va, size, patch_start, patch_size);
+		return LFA_CRITICAL_ERROR;
+	}
+
+	src_va = activation->image_address + src_offset;
+
+	if (lfa_bl31_check(patch_start, patch_size, &entry) != 0) {
+		ERROR("BL31 LFA: xlat check failed for patch region dst=0x%lx size=0x%zx "
+		      "aligned_start 0x%lx aligned_size 0x%zx\n",
+		      dst_va, size, patch_start, patch_size);
+		return LFA_CRITICAL_ERROR;
+	}
+
+	lfa_r_bl31_copy(patch_start, entry, (void *)src_va, patch_size, reset);
+	return LFA_SUCCESS;
+}
+
+int32_t lfa_bl31_prime(struct lfa_component_status *activation)
+{
+	// TODO: Check that activation->image_size is not too large.
+	return LFA_SUCCESS;
+}
+
+/* Function that actually performs the activation of the firmware image */
+static int lfa_bl31_image_activation(struct lfa_component_status *activation,
+	uint64_t ep_address, uint64_t context_id)
+{
+	VERBOSE("BL31 LFA: Starting BL31 Image Activation (core %d)\n", plat_my_core_pos());
+	VERBOSE("  Address: 0x%lX\n", activation->image_address);
+	VERBOSE("  Size:    %ld bytes\n", activation->image_size);
+	VERBOSE("  Reset:   %s\n", activation->reset ? "yes" : "no");
+
+#if !ENABLE_RUNTIME_INSTRUMENTATION
+	/*
+	 * Used to avoid console corruption. Not used if measuring performance
+	 * since printouts are removed.
+	 */
+	console_flush();
+#endif
+
+	/* Prepare for warm reset and setup NS entrypoint for primary cores */
+	if (activation->reset) {
+		if (prepare_warm_reset(ep_address, context_id,
+				(uint64_t)&lfa_r_holding_lock_var) != LFA_SUCCESS) {
+			lfa_r_holding_unlock(&lfa_r_holding_lock_var);
+			return LFA_CRITICAL_ERROR;
+		}
+	}
+
+	if (lfa_bl31_patch_region(activation, CPU_OPS_START,
+				  lfa_bl31_cpu_ops_size(), false) != LFA_SUCCESS) {
+		return LFA_CRITICAL_ERROR;
+	}
+
+	/*
+	 * Patch the errata section last so reset, if requested, is triggered
+	 * only after cpu_ops has been updated to point at the new errata code.
+	 */
+	return lfa_bl31_patch_region(activation, ERRATA_START,
+				      lfa_bl31_errata_size(),
+				      activation->reset);
+}
+
+int32_t lfa_bl31_activate(struct lfa_component_status *activation,
+	uint64_t ep_address, uint64_t context_id)
 {
-	return LFA_WRONG_STATE;
+	uint32_t core_pos = plat_my_core_pos();
+	int ret = LFA_SUCCESS;
+
+	lfa_r_ep_addresses[core_pos] = ep_address;
+	lfa_r_context_ids[core_pos] = context_id;
+
+	/* Case when Firmware requires CPU Rendezvous for LFA Activation */
+	if (activation->cpu_rendezvous) {
+		/*
+		 * Only one core will return true from this function and it
+		 * controls the activation process.
+		 */
+		if (lfa_holding_start()) {
+			VERBOSE("BL31 LFA: CPU %d is leader core\n", core_pos);
+
+			/*
+			 * Before we release other cores from the holding pen
+			 * try to get the image info. That way if this fails the
+			 * system will still be in a recoverable state.
+			 */
+			ret = plat_lfa_get_image_info(activation->component_id,
+						      &activation->image_address,
+						      &activation->image_size);
+			if (ret) {
+				ERROR("BL31 LFA: Could not get image info!\n");
+				lfa_holding_release(LFA_CRITICAL_ERROR);
+				return LFA_CRITICAL_ERROR;
+			}
+
+			/*
+			 * Only one core will load the relocatable code module and
+			 * acquire the relocatable lock, then release the rest of the
+			 * cores which will then wait for us to release the relocatable
+			 * lock once the activation is complete.
+			 */
+			lfa_load_relocatable();
+			lfa_r_holding_lock(&lfa_r_holding_lock_var);
+			lfa_holding_release(LFA_SUCCESS);
+
+			/*
+			 * Perform image activation. Will not return if reset
+			 * requested.
+			 */
+			ret = lfa_bl31_image_activation(activation, ep_address, context_id);
+
+			/* Unlock holding pen for all other cores */
+			lfa_r_holding_unlock(&lfa_r_holding_lock_var);
+
+			if (ret) {
+				ERROR("BL31 LFA: failed image activation with error: %d\n", ret);
+				return ret;
+			}
+
+			INFO("BL31 LFA: Successful image activation!\n");
+		} else {
+			/* Wait until release from rendezvous holding pen. */
+			ret = lfa_holding_wait();
+			if (ret != LFA_SUCCESS) {
+				return ret;
+			}
+
+			if (activation->reset) {
+				/*
+				 * Prepare for warm reset and setup NS
+				 * entrypoint for secondary cores
+				 */
+				if (prepare_warm_reset(ep_address, context_id,
+						(uint64_t)&lfa_r_holding_lock_var) != LFA_SUCCESS) {
+					return LFA_CRITICAL_ERROR;
+				}
+
+				/*
+				 * Wait until activation of the image is
+				 * complete and warm reset occurs on main core
+				 * then reset.
+				 */
+				lfa_r_holding_wait_warm_reset(&lfa_r_holding_lock_var);
+			} else {
+				/*
+				 * Wait until main core is completed with
+				 * activation
+				 */
+				lfa_r_holding_wait(&lfa_r_holding_lock_var);
+			}
+		}
+	}
+
+	/*
+	 * No CPU rendezvous required. Perform LFA Activate with single core
+	 * NOTE: Assumption that only one core will call activate with
+	 * CPU_RENDEZVOUS=0
+	 */
+	else {
+		/*
+		 * Loading relocatable module normally happens in the holding
+		 * pen. Since there is no holding pen without CPU rendezvous we
+		 * are doing it here instead.
+		 */
+		lfa_load_relocatable();
+
+		/* Perform actual image activation */
+		ret = lfa_bl31_image_activation(activation, ep_address, context_id);
+		if (ret) {
+			ERROR("Failed image activation with error: %d\n", ret);
+			return ret;
+		}
+	}
+
+	NOTICE("BL31 Live Activation Without Reset %u\n", plat_my_core_pos());
+	NOTICE("  Version : %s\n", build_version_string);
+	NOTICE("  %s\n", build_message);
+
+	return ret;
 }
 
-static int lfa_bl31_activate(struct lfa_component_status *activation,
-		uint64_t ep_address,
-		uint64_t context_id)
+int32_t lfa_bl31_cancel(struct lfa_component_status *activation)
 {
-	return LFA_WRONG_STATE;
+	/* Let CPUs out of the holding pen if needed. */
+	lfa_holding_release(LFA_ACTIVATION_FAILED);
+
+	return LFA_SUCCESS;
 }
 
-static struct lfa_component_ops bl31_activator = {
+struct lfa_component_ops bl31_activator = {
 	.prime = lfa_bl31_prime,
 	.activate = lfa_bl31_activate,
-	.may_reset_cpu = false,
+	.cancel = lfa_bl31_cancel,
+	.may_reset_cpu = true,
 	.cpu_rendezvous_required = true,
 };
 
@@ -30,3 +260,128 @@ struct lfa_component_ops *get_bl31_activator(void)
 {
 	return &bl31_activator;
 }
+
+uint64_t *lfa_get_xlat_table_entry(uintptr_t virtual_addr, const xlat_ctx_t *ctx)
+{
+	uint64_t *table = ctx->base_table;
+	unsigned int entries = ctx->base_table_entries;
+	unsigned int start_level = GET_XLAT_TABLE_LEVEL_BASE((unsigned long long)ctx->va_max_address + 1ULL);
+
+	for (unsigned int level = start_level; level <= XLAT_TABLE_LEVEL_MAX; level++) {
+		uint64_t idx, desc, desc_type;
+
+		idx = XLAT_TABLE_IDX(virtual_addr, level);
+		if (idx >= entries) {
+			return NULL;
+		}
+
+		desc = table[idx];
+		desc_type = desc & DESC_MASK;
+
+		if (desc_type == INVALID_DESC) {
+			return NULL;
+		}
+
+		if (level == XLAT_TABLE_LEVEL_MAX) {
+			/*
+			 * Only page descriptors allowed at the final lookup
+			 * level.
+			 */
+			assert(desc_type == PAGE_DESC);
+			return &table[idx];
+		}
+
+		if (desc_type == BLOCK_DESC) {
+			return &table[idx];
+		}
+
+		assert(desc_type == TABLE_DESC);
+		table = (uint64_t *)(uintptr_t)(desc & TABLE_ADDR_MASK);
+		entries = XLAT_TABLE_ENTRIES;
+	}
+
+	return NULL;
+}
+
+int lfa_bl31_check(uintptr_t base_va, size_t size, uint64_t **entry)
+{
+	const xlat_ctx_t *ctx = get_xlat_tables();
+
+	assert(ctx != NULL);
+	assert(ctx->initialized);
+
+	if (!IS_PAGE_ALIGNED(base_va)) {
+		WARN("%s: Address 0x%lx is not aligned on a page boundary.\n", __func__, base_va);
+		return -EINVAL;
+	}
+
+	if (size == 0U) {
+		WARN("%s: Size is 0.\n", __func__);
+		return -EINVAL;
+	}
+
+	size_t pages_count = size / PAGE_SIZE;
+
+	VERBOSE("BL31 update will impact %zu pages starting from address 0x%lx...\n",
+		pages_count, base_va);
+
+	*entry = lfa_get_xlat_table_entry(base_va, ctx);
+
+	uint64_t *last = lfa_get_xlat_table_entry(base_va + (pages_count - 1) * PAGE_SIZE, ctx);
+
+	/* Check that all descriptors are in the same page. */
+	if ((((uint64_t)*entry) >> 12) != ((((uint64_t)last) >> 12))) {
+		ERROR("All page descriptors are not part of the same page.\n");
+		return -EINVAL;
+	}
+
+	return 0;
+}
+
+__attribute__((noinline, section(".lfa_relocatable_code")))
+int lfa_r_bl31_copy(uintptr_t base_va, uint64_t *entry, void *src, size_t size, bool reset)
+{
+	do {
+		uint64_t desc = *entry;
+		size_t count = MIN(PAGE_SIZE, size) / sizeof(uint64_t);
+
+		*entry = INVALID_DESC;
+		dsbishst();
+		tlbivae3is(TLBI_ADDR(base_va));
+		dsbish();
+		isb();
+		*entry = (desc & ~LOWER_ATTRS(AP_RO));
+		dsbish();
+
+		for (uint64_t i = 0; i < count; i++) {
+			((uint64_t *) base_va)[i] = ((uint64_t *) src)[i];
+		}
+		if (size <= PAGE_SIZE) {
+			for (uint64_t i = count * sizeof(uint64_t); i < size; i++) {
+				((uint8_t *) base_va)[i] = ((uint8_t *) src)[i];
+			}
+			size = 0;
+		} else {
+			size -= PAGE_SIZE;
+		}
+
+		*entry = INVALID_DESC;
+		dsbishst();
+		tlbivae3is(TLBI_ADDR(base_va));
+		dsbish();
+		isb();
+		*entry = desc;
+
+		base_va += PAGE_SIZE;
+		src += PAGE_SIZE;
+		entry++;
+	} while (size > 0);
+
+	dsbish();
+
+	if (reset) {
+		wfi();
+	}
+
+	return 0;
+}
diff --git a/services/std_svc/lfa/lfa.mk b/services/std_svc/lfa/lfa.mk
index 53af69fd7..d3d5cd864 100644
--- a/services/std_svc/lfa/lfa.mk
+++ b/services/std_svc/lfa/lfa.mk
@@ -6,9 +6,15 @@
 
 LFA_SOURCES	+=	$(addprefix services/std_svc/lfa/, \
 			  lfa_main.c \
-			  bl31_lfa.c \
 			  lfa_holding_pen.c)
 
 ifeq (${ENABLE_RMM}, 1)
 LFA_SOURCES	+=	services/std_svc/rmmd/rmmd_rmm_lfa.c
 endif
+
+ifeq (${ENABLE_LFA_BL31}, 1)
+LFA_SOURCES	+=	$(addprefix services/std_svc/lfa/, \
+			  bl31_lfa.c \
+			  lfa_relocatable.S \
+			  lfa_reset.c)
+endif
diff --git a/services/std_svc/lfa/lfa_holding_pen.c b/services/std_svc/lfa/lfa_holding_pen.c
index 9aa1e8c30..39eac41a8 100644
--- a/services/std_svc/lfa/lfa_holding_pen.c
+++ b/services/std_svc/lfa/lfa_holding_pen.c
@@ -7,10 +7,12 @@
 
 #include <string.h>
 
+#include <common/bl_common.h>
 #include <common/debug.h>
 #include <lib/psci/psci_lib.h>
 #include <lib/spinlock.h>
 #include <lib/utils_def.h>
+#include <lib/xlat_tables/xlat_tables_v2.h>
 #include <plat/common/platform.h>
 #include <services/lfa_holding_pen.h>
 
@@ -18,7 +20,7 @@
 
 static spinlock_t holding_lock;
 static spinlock_t activation_lock;
-static uint32_t activation_count;
+static volatile uint32_t activation_count;
 static enum lfa_retc activation_status;
 
 /**
@@ -44,28 +46,31 @@ static enum lfa_retc activation_status;
  */
 bool lfa_holding_start(void)
 {
-	bool status;
-	unsigned int no_of_cpus;
+	unsigned int no_of_cpus = psci_num_cpus_running_on_safe(plat_my_core_pos());
+	bool last_cpu;
 
+	/*
+	 * This lock ensures only one CPU uses the activation_count
+	 * variable at a time.
+	 */
 	spin_lock(&activation_lock);
 
+	/* First CPU to arrive locks the holding pen. */
 	if (activation_count == 0U) {
-		/* First CPU locks holding lock */
 		spin_lock(&holding_lock);
 	}
 
 	activation_count += 1U;
+	last_cpu = (activation_count == no_of_cpus);
 
-	no_of_cpus = psci_num_cpus_running_on_safe(plat_my_core_pos());
-	status = (activation_count == no_of_cpus);
-	if (!status) {
-		VERBOSE("Hold, %d CPU left\n",
-			 no_of_cpus - activation_count);
+	if (!last_cpu) {
+		VERBOSE("Hold, %d CPU left\n", no_of_cpus - activation_count);
 	}
 
+	/* Release the lock once count is updated. */
 	spin_unlock(&activation_lock);
 
-	return status;
+	return last_cpu;
 }
 
 /**
@@ -103,3 +108,42 @@ void lfa_holding_release(enum lfa_retc status)
 	activation_status = status;
 	spin_unlock(&holding_lock);
 }
+
+#if ENABLE_LFA_BL31
+void lfa_load_relocatable(void)
+{
+	int ret;
+
+	/*
+	 * The relocatable code area is outside of the normal BL31 image, so
+	 * before we can use it we have to make it writeable then copy the
+	 * relocatable code into it. Then once the copy is complete we must
+	 * mark it read only and executable again.
+	 */
+	ret = xlat_change_mem_attributes(LFA_RELOCATABLE_CODE_START,
+					 (LFA_RELOCATABLE_DATA_START - LFA_RELOCATABLE_CODE_START),
+					 MT_MEMORY | MT_RW | MT_EXECUTE_NEVER | EL3_PAS);
+	if (ret != 0) {
+		ERROR("LFA: Could not update memory attributes: %d\n", ret);
+		panic();
+	}
+
+	memcpy((void *)LFA_RELOCATABLE_CODE_START, (void *)LFA_RELOCATABLE_LMA,
+	       (uintptr_t)(LFA_RELOCATABLE_DATA_START - LFA_RELOCATABLE_CODE_START));
+
+	ret = xlat_change_mem_attributes(LFA_RELOCATABLE_CODE_START,
+					 (LFA_RELOCATABLE_DATA_START - LFA_RELOCATABLE_CODE_START),
+					 MT_MEMORY | MT_RO | MT_EXECUTE | EL3_PAS);
+	if (ret != 0) {
+		ERROR("LFA: Could not update memory attributes: %d\n", ret);
+		panic();
+	}
+
+	/*
+	 * The relocatable data section contains unknown information at this
+	 * point so zero it out before we use it.
+	 */
+	memset((void *)LFA_RELOCATABLE_DATA_START, 0,
+	       (LFA_RELOCATABLE_DATA_END - LFA_RELOCATABLE_DATA_START));
+}
+#endif /* ENABLE_LFA_BL31 */
diff --git a/services/std_svc/lfa/lfa_main.c b/services/std_svc/lfa/lfa_main.c
index 819b169a7..31c05bf58 100644
--- a/services/std_svc/lfa/lfa_main.c
+++ b/services/std_svc/lfa/lfa_main.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2025, Arm Limited. All rights reserved.
+ * Copyright (c) 2025-2026, Arm Limited. All rights reserved.
  * Copyright (c) 2025, NVIDIA Corporation. All rights reserved.
  *
  * SPDX-License-Identifier: BSD-3-Clause
@@ -15,10 +15,23 @@
 #include <services/rmmd_rmm_lfa.h>
 #include <smccc_helpers.h>
 
-static uint32_t lfa_component_count;
 static plat_lfa_component_info_t *lfa_components;
 static struct lfa_component_status current_activation;
-static bool is_lfa_initialized;
+
+/*
+ * Multiple cores call into LFA service which shares these variables, so ensure
+ * that any accesses to these variables go back to memory. These are sort of
+ * "lazy" state variables where timing isn't really critical so we aren't
+ * explicitly using mutex for access control but we just want to make sure
+ * nothing is optimized away.
+ */
+static volatile uint32_t lfa_component_count;
+static volatile bool is_lfa_initialized;
+static volatile bool activation_in_progress;
+static volatile bool activation_skip_cpu_rendezvous;
+static volatile uint32_t activation_users;
+static volatile bool activation_failed;
+static volatile bool activation_cancel_locked;
 
 /*
  * Spinlock to serialize LFA operations (PRIME, ACTIVATE).
@@ -32,7 +45,12 @@ void lfa_reset_activation(void)
 {
 	current_activation.component_id = LFA_INVALID_COMPONENT;
 	current_activation.prime_status = PRIME_NONE;
-	current_activation.cpu_rendezvous_required = false;
+	current_activation.cpu_rendezvous = false;
+	activation_in_progress = false;
+	activation_skip_cpu_rendezvous = false;
+	activation_users = 0U;
+	activation_failed = false;
+	activation_cancel_locked = false;
 }
 
 static int convert_to_lfa_error(int ret)
@@ -54,8 +72,7 @@ static bool lfa_initialize_components(void)
 	lfa_component_count = plat_lfa_get_components(&lfa_components);
 
 	if (lfa_component_count == 0U || lfa_components == NULL) {
-		/* unlikely to reach here */
-		ERROR("Invalid LFA component setup: count = 0 or components are NULL");
+		ERROR("Could not retrieve LFA components.\n");
 		return false;
 	}
 
@@ -64,8 +81,7 @@ static bool lfa_initialize_components(void)
 
 static uint64_t get_fw_activation_flags(uint32_t fw_seq_id)
 {
-	const plat_lfa_component_info_t *comp =
-				&lfa_components[fw_seq_id];
+	const plat_lfa_component_info_t *comp = &lfa_components[fw_seq_id];
 	uint64_t flags = 0ULL;
 
 	flags |= ((comp->activator == NULL ? 0ULL : 1ULL)
@@ -86,6 +102,7 @@ static uint64_t get_fw_activation_flags(uint32_t fw_seq_id)
 static int lfa_cancel(uint32_t component_id)
 {
 	int ret = LFA_SUCCESS;
+	struct lfa_component_ops *activator;
 
 	if (lfa_component_count == 0U) {
 		return LFA_WRONG_STATE;
@@ -97,76 +114,33 @@ static int lfa_cancel(uint32_t component_id)
 		return LFA_INVALID_PARAMETERS;
 	}
 
-	ret = plat_lfa_cancel(component_id);
-	if (ret != LFA_SUCCESS) {
+	/*
+	 * Deny cancellation once the activation round is no longer cancellable.
+	 */
+	if (activation_in_progress &&
+	    (activation_cancel_locked || activation_skip_cpu_rendezvous)) {
 		return LFA_BUSY;
 	}
 
-	/* TODO: add proper termination prime and activate phases */
-	lfa_reset_activation();
-
-	return ret;
-}
-
-static int lfa_activate(uint32_t component_id, uint64_t flags,
-			uint64_t ep_address, uint64_t context_id)
-{
-	int ret = LFA_ACTIVATION_FAILED;
-	struct lfa_component_ops *activator;
-
-	if ((lfa_component_count == 0U) ||
-	    (!lfa_components[component_id].activation_pending) ||
-	    (current_activation.prime_status != PRIME_COMPLETE)) {
-		return LFA_COMPONENT_WRONG_STATE;
-	}
-
-	/* Check if fw_seq_id is in range. */
-	if ((component_id >= lfa_component_count) ||
-	    (current_activation.component_id != component_id)) {
-		return LFA_INVALID_PARAMETERS;
-	}
-
-	if (lfa_components[component_id].activator == NULL) {
-		return LFA_NOT_SUPPORTED;
-	}
-
-	ret = plat_lfa_notify_activate(component_id);
-	if (ret != 0) {
-		return LFA_ACTIVATION_FAILED;
+	activator = lfa_components[component_id].activator;
+	if (activator->cancel != NULL) {
+		ret = activator->cancel(&current_activation);
+		if (ret != LFA_SUCCESS) {
+			return LFA_BUSY;
+		}
 	}
 
-	activator = lfa_components[component_id].activator;
-	if (activator->activate != NULL) {
+	ret = plat_lfa_cancel(component_id);
+	if (ret != LFA_SUCCESS) {
 		/*
-		 * Pass skip_cpu_rendezvous (flag[0]) only if flag[0]==1
-		 * & CPU_RENDEZVOUS is not required.
+		 * Cancellation can fail when activation cannot be stopped
+		 * (e.g. activation requested with skip_cpu_rendezvous=1).
 		 */
-		if (flags & LFA_SKIP_CPU_RENDEZVOUS_BIT) {
-			if (!activator->cpu_rendezvous_required) {
-				INFO("Skipping rendezvous requested by caller.\n");
-				current_activation.cpu_rendezvous_required = false;
-			}
-			/*
-			 * Return error if caller tries to skip rendezvous when
-			 * it is required.
-			 */
-			else {
-				ERROR("CPU Rendezvous is required, can't skip.\n");
-				return LFA_INVALID_PARAMETERS;
-			}
-		}
-
-		ret = activator->activate(&current_activation, ep_address,
-					  context_id);
+		return LFA_BUSY;
 	}
 
-	/*
-	 * Update the activation pending flag only if the activation was
-	 * successful.
-	 */
-	if (ret == LFA_SUCCESS) {
-		lfa_components[component_id].activation_pending = false;
-	}
+	/* TODO: add proper termination prime and activate phases */
+	lfa_reset_activation();
 
 	return ret;
 }
@@ -225,7 +199,13 @@ static int lfa_prime(uint32_t component_id, uint64_t *flags)
 			}
 		}
 
+		/*
+		 * Update current activation status fields. CPU rendezvous is enabled
+		 * by default but can be disabled if requested and allowed.
+		 */
+		current_activation.cpu_rendezvous = true;
 		current_activation.prime_status = PRIME_COMPLETE;
+		current_activation.reset = lfa_components[component_id].activator->may_reset_cpu;
 	}
 
 	/*
@@ -264,12 +244,122 @@ int lfa_setup(void)
 	return 0;
 }
 
+static int lfa_activate_prepare(uint32_t component_id, uint64_t flags,
+				struct lfa_component_ops **activator)
+{
+	int ret = LFA_ACTIVATION_FAILED;
+	/* First caller decides activation policy and performs platform notify. */
+	bool first_cpu = !activation_in_progress;
+
+	/* Check if fw_seq_id is in range. */
+	if ((component_id >= lfa_component_count) ||
+	    (current_activation.component_id != component_id)) {
+		return LFA_INVALID_PARAMETERS;
+	}
+
+	if ((lfa_component_count == 0U) ||
+	    (!lfa_components[component_id].activation_pending) ||
+	    (current_activation.prime_status != PRIME_COMPLETE)) {
+		return LFA_COMPONENT_WRONG_STATE;
+	}
+
+	if (lfa_components[component_id].activator == NULL) {
+		return LFA_NOT_SUPPORTED;
+	}
+
+	*activator = lfa_components[component_id].activator;
+
+	if (first_cpu) {
+		activation_failed = false;
+
+		/*
+		 * If rendezvous is optional for this component, default to
+		 * rendezvous unless caller explicitly asks to skip it.
+		 * This choice is latched for the whole activation.
+		 */
+		if (!(*activator)->cpu_rendezvous_required &&
+		    ((flags & LFA_SKIP_CPU_RENDEZVOUS_BIT) == 0U)) {
+			current_activation.cpu_rendezvous = true;
+		} else {
+			current_activation.cpu_rendezvous =
+				(*activator)->cpu_rendezvous_required;
+		}
+		activation_skip_cpu_rendezvous = false;
+		activation_cancel_locked = false;
+	}
+	/*
+	 * Pass skip_cpu_rendezvous (flag[0]) only if flag[0]==1
+	 * & CPU_RENDEZVOUS is not required.
+	 */
+	if (flags & LFA_SKIP_CPU_RENDEZVOUS_BIT) {
+		if (!(*activator)->cpu_rendezvous_required) {
+			/*
+			 * Late skip request is rejected if activation already
+			 * started without skip.
+			 */
+			if (!first_cpu && !activation_skip_cpu_rendezvous) {
+				return LFA_BUSY;
+			}
+			INFO("Skipping rendezvous requested by caller.\n");
+			current_activation.cpu_rendezvous = false;
+			activation_skip_cpu_rendezvous = true;
+		}
+		/*
+		 * Return error if caller tries to skip rendezvous when
+		 * it is required.
+		 */
+		else {
+			ERROR("CPU Rendezvous is required, can't skip.\n");
+			return LFA_INVALID_PARAMETERS;
+		}
+	}
+
+	if (first_cpu) {
+		/* Notify platform once per activation round. */
+		ret = plat_lfa_notify_activate(component_id);
+		if (ret != 0) {
+			return LFA_ACTIVATION_FAILED;
+		}
+		activation_in_progress = true;
+	}
+
+	/* Track how many CPUs have entered LFA_ACTIVATE for this round. */
+	activation_users += 1U;
+	if (activation_users >= psci_num_cpus_running_on_safe(plat_my_core_pos())) {
+	    activation_cancel_locked = true;
+	}
+
+	return LFA_SUCCESS;
+}
+
+
+static void lfa_activate_finish(uint32_t component_id, bool activation_complete)
+{
+	if (activation_users > 0U) {
+		activation_users -= 1U;
+	}
+
+	if (!activation_complete) {
+		return;
+	}
+
+	if (activation_users == 0U) {
+		activation_in_progress = false;
+		activation_skip_cpu_rendezvous = false;
+		activation_cancel_locked = false;
+		if (!activation_failed) {
+			lfa_components[component_id].activation_pending = false;
+		}
+	}
+}
+
 uint64_t lfa_smc_handler(uint32_t smc_fid, u_register_t x1, u_register_t x2,
 			 u_register_t x3, u_register_t x4, void *cookie,
 			 void *handle, u_register_t flags)
 {
-	uint64_t retx1, retx2;
-	uint64_t lfa_flags;
+	uint64_t retx1;
+	uint64_t retx2;
+	uint64_t lfa_flags = 0;
 	uint8_t *uuid_p;
 	uint32_t fw_seq_id = (uint32_t)x1;
 	int ret;
@@ -292,9 +382,9 @@ uint64_t lfa_smc_handler(uint32_t smc_fid, u_register_t x1, u_register_t x2,
 		break;
 
 	case LFA_GET_INFO:
-		/**
+		/*
 		 * The current specification limits this input parameter to be zero for
-		 * version 1.0 of LFA
+		 * version 1.0 of LFA.
 		 */
 		if (x1 == 0ULL) {
 			SMC_RET3(handle, LFA_SUCCESS, lfa_component_count, 0);
@@ -309,29 +399,27 @@ uint64_t lfa_smc_handler(uint32_t smc_fid, u_register_t x1, u_register_t x2,
 		}
 
 		/*
-		 * Check if fw_seq_id is in range. LFA_GET_INFO must be called first to scan
-		 * platform firmware and create a valid number of firmware components.
+		 * Check if fw_seq_id is in range. LFA_GET_INFO must be called
+		 * first to scan platform firmware and create a valid number of
+		 * firmware components.
 		 */
 		if (fw_seq_id >= lfa_component_count) {
 			SMC_RET1(handle, LFA_INVALID_PARAMETERS);
 		}
 
-		/*
-		 * grab the UUID of asked fw_seq_id and set the return UUID
-		 * variables
-		 */
+		/* Get the UUID of requested fw_seq_id. */
 		uuid_p = (uint8_t *)&lfa_components[fw_seq_id].uuid;
 		memcpy(&retx1, uuid_p, sizeof(uint64_t));
 		memcpy(&retx2, uuid_p + sizeof(uint64_t), sizeof(uint64_t));
 
 		/*
-		 * check the given fw_seq_id's update available
-		 * and accordingly set the active_pending flag
+		 * Check the given fw_seq_id update available and accordingly
+		 * set the active_pending flag.
 		 */
 		lfa_components[fw_seq_id].activation_pending =
 				is_plat_lfa_activation_pending(fw_seq_id);
 
-		INFO("Component %lu %s live activation:\n", x1,
+		INFO("Component %lu %s live activation.\n", x1,
 		      lfa_components[fw_seq_id].activator ? "supports" :
 		      "does not support");
 
@@ -340,10 +428,7 @@ uint64_t lfa_smc_handler(uint32_t smc_fid, u_register_t x1, u_register_t x2,
 			      lfa_components[fw_seq_id].activation_pending ? "true" : "false");
 		}
 
-		INFO("x1 = 0x%016lx, x2 = 0x%016lx\n", retx1, retx2);
-
 		SMC_RET4(handle, LFA_SUCCESS, retx1, retx2, get_fw_activation_flags(fw_seq_id));
-
 		break;
 
 	case LFA_PRIME:
@@ -369,25 +454,111 @@ uint64_t lfa_smc_handler(uint32_t smc_fid, u_register_t x1, u_register_t x2,
 		break;
 
 	case LFA_ACTIVATE:
-		ret = lfa_activate(fw_seq_id, x2, x3, x4);
-		/* TODO: implement activate again */
-		SMC_RET2(handle, ret, 0ULL);
+		/* LFA_ACTIVATE flow:
+		 * - LFA_SKIP_CPU_RENDEZVOUS_BIT controls skip request on entry.
+		 * - Serialize with lfa_lock, but never block behind a
+		 *   skip-rendezvous activation.
+		 * - lfa_activate_prepare() validates state, selects activator,
+		 *   and latches policy.
+		 * - The first CPU decides rendezvous policy for the whole
+		 *   activation.
+		 * - Skip rendezvous keeps the lock for the full activation;
+		 *   normal path releases it.
+		 * - activator->activate() may return -EAGAIN, mapped to
+		 *   LFA_CALL_AGAIN in lfa_flags.
+		 * - lfa_activate_finish() updates bookkeeping and clears
+		 *   in-progress state when done.
+		 */
+		bool hold_lock = false;
+		bool activation_complete = true;
+		struct lfa_component_ops *activator;
+		/* Caller request: skip CPU rendezvous for this activation. */
+		bool skip_cpu_rendezvous =
+			((x2 & LFA_SKIP_CPU_RENDEZVOUS_BIT) != 0U);
+
+		/*
+		 * When skip_cpu_rendezvous=1 (or when a skip rendezvous
+		 * activation is already in progress), concurrent LFA_ACTIVATE
+		 * calls must not be accepted; return LFA_BUSY instead of
+		 * blocking.
+		 */
+		if (skip_cpu_rendezvous) {
+			/* Non-blocking serialization for skip rendezvous requests. */
+			if (!spin_trylock(&lfa_lock)) {
+				SMC_RET1(handle, LFA_BUSY);
+			}
+		} else {
+			/*
+			 * Normal path: take the lock, but reject if a skip
+			 * activation is in progress.
+			 */
+			spin_lock(&lfa_lock);
+			if (activation_skip_cpu_rendezvous) {
+				spin_unlock(&lfa_lock);
+				SMC_RET1(handle, LFA_BUSY);
+			}
+		}
+
+		ret = lfa_activate_prepare(fw_seq_id, x2, &activator);
+		hold_lock = (ret == LFA_SUCCESS) && activation_skip_cpu_rendezvous;
+
+		/*
+		 * Keep the lock held for the full activation only when skip
+		 * rendezvous is in effect.
+		 */
+		if (!hold_lock) {
+			spin_unlock(&lfa_lock);
+		}
+
+		if (ret != LFA_SUCCESS) {
+			SMC_RET2(handle, ret, 0ULL);
+		}
+
+		if (activator->activate != NULL) {
+			ret = activator->activate(&current_activation, x3, x4);
+		}
+
+		/* Reacquire lock before updating shared activation state. */
+		if (!hold_lock) {
+			spin_lock(&lfa_lock);
+		}
+
+		lfa_flags = 0ULL;
+		if (ret == -EAGAIN) {
+			/* Multi-stage activation: caller must reissue LFA_ACTIVATE. */
+			ret = LFA_SUCCESS;
+			lfa_flags = LFA_CALL_AGAIN;
+			activation_complete = false;
+		} else if (ret != LFA_SUCCESS) {
+			activation_failed = true;
+		}
+
+		/* Update activation bookkeeping and clear in-progress state if complete. */
+		lfa_activate_finish(fw_seq_id, activation_complete);
+
+		spin_unlock(&lfa_lock);
+
+		SMC_RET2(handle, ret, lfa_flags);
 
 		break;
 
 	case LFA_CANCEL:
+		if (!spin_trylock(&lfa_lock)) {
+			SMC_RET1(handle, LFA_BUSY);
+		}
+
 		ret = lfa_cancel(x1);
+
+		spin_unlock(&lfa_lock);
+
 		SMC_RET1(handle, ret);
 		break;
 
 	default:
 		WARN("Unimplemented LFA Service Call: 0x%x\n", smc_fid);
 		SMC_RET1(handle, SMC_UNK);
-		break; /* unreachable */
-
+		break;
 	}
 
-	SMC_RET1(handle, SMC_UNK);
-
 	return 0;
 }
diff --git a/services/std_svc/lfa/lfa_relocatable.S b/services/std_svc/lfa/lfa_relocatable.S
new file mode 100644
index 000000000..7436584bb
--- /dev/null
+++ b/services/std_svc/lfa/lfa_relocatable.S
@@ -0,0 +1,108 @@
+/*
+ * Copyright (c) 2026, Arm Limited and Contributors. All rights reserved.
+ *
+ * SPDX-License-Identifier: BSD-3-Clause
+ */
+
+#include <asm_macros.S>
+#include <platform_def.h>
+
+/* Relocatable code section */
+
+	.globl  lfa_r_holding_lock
+	.globl  lfa_r_holding_wait
+	.globl  lfa_r_holding_unlock
+	.globl  lfa_r_holding_wait_warm_reset
+
+/*
+ * The 'rfunc' macro is a modified version of the 'func' macro in
+ * "include/common/asm_macros_common.S". It differs in that it does not change
+ * the section to ".text.asm". See the original version for comments.
+ */
+	.macro rfunc _name, _align=2
+	.cfi_sections .debug_frame
+	.section .lfa_relocatable_code.\_name, "ax"
+	.type \_name, %function
+	.cfi_startproc
+	.align \_align
+	\_name:
+#if ENABLE_BTI
+	 bti	jc
+#endif
+	.endm
+
+/*
+ * Enable holding lock.
+ *
+ * void lfa_r_holding_lock(spinlock_t *lock);
+ */
+rfunc lfa_r_holding_lock
+	mov	w1, #1
+	stlr	w1, [x0]
+	ret
+endfunc lfa_r_holding_lock
+
+/*
+ * Release hold.
+ *
+ * Use store-release to unconditionally clear the spinlock variable.
+ * Store operation generates an event to all cores waiting in WFE
+ * when address is monitored by the global monitor.
+ *
+ * void lfa_r_holding_unlock(spinlock_t *lock);
+ */
+rfunc lfa_r_holding_unlock
+	mov	w1, #0
+	stlr	w1, [x0]
+	ret
+endfunc lfa_r_holding_unlock
+
+/*
+ * Check lock using load-exclusive instruction.
+ *
+ * void lfa_r_holding_wait(spinlock_t *lock);
+ */
+rfunc lfa_r_holding_wait
+	sevl
+l1:	wfe
+	ldaxr	w1, [x0]
+	cbnz	w1, l1
+	ret
+endfunc lfa_r_holding_wait
+
+/*
+ * Check lock using load-exclusive instruction.
+ * Issue RMR_EL3 for warm reset if lock is unlocked
+ *
+ * void lfa_r_holding_wait(spinlock_t *lock);
+ */
+rfunc lfa_r_holding_wait_warm_reset
+	sevl
+l2:
+	wfe
+	ldaxr	w1, [x0]
+	cbnz	w1, l2
+l3:
+	wfi
+	b l3
+endfunc lfa_r_holding_wait_warm_reset
+
+/* Relocatable data section */
+
+	.globl  lfa_r_holding_lock_var
+	.globl  lfa_r_ep_addresses
+	.globl  lfa_r_context_ids
+
+.align 8
+.section .lfa_relocatable_data.lfa_r_holding_lock_var, "aw"
+lfa_r_holding_lock_var: .quad 0x0
+
+.align 8
+.section .lfa_relocatable_data.lfa_r_ep_addresses, "aw"
+lfa_r_ep_addresses:
+	.fill	PLATFORM_CORE_COUNT, 8, 0x0
+
+.align 8
+.section .lfa_relocatable_data.lfa_r_context_ids, "aw"
+lfa_r_context_ids:
+	.fill	PLATFORM_CORE_COUNT, 8, 0x0
diff --git a/services/std_svc/lfa/lfa_reset.c b/services/std_svc/lfa/lfa_reset.c
new file mode 100644
index 000000000..07db8cc97
--- /dev/null
+++ b/services/std_svc/lfa/lfa_reset.c
@@ -0,0 +1,173 @@
+/*
+ * Copyright (c) 2026, Arm Limited. All rights reserved.
+ *
+ * SPDX-License-Identifier: BSD-3-Clause
+ */
+
+#include <common/build_message.h>
+#include <lib/el3_runtime/context_mgmt.h>
+#include <lib/psci/psci.h>
+#include <lib/spinlock.h>
+#include <plat/arm/common/plat_arm.h>
+#include <plat/common/platform.h>
+#include <services/lfa_svc.h>
+
+/*
+ * These variables are used pre and post live activation so place them in the
+ * relocatable data section to ensure they don't get moved around and persist
+ * across live activation.
+ */
+static lfa_mailbox_message_t lfa_mailbox[PLATFORM_CORE_COUNT] __attribute__((section(".lfa_relocatable_data")));
+static const plat_psci_ops_t *lfa_psci_ops __attribute__((section(".lfa_relocatable_data")));
+static uintptr_t plat_wr_ep __attribute__((section(".lfa_relocatable_data")));
+static uint64_t reset_cpu_count __attribute__((section(".lfa_relocatable_data")));
+static spinlock_t cpucountlock __attribute__((section(".lfa_relocatable_data")));
+
+extern plat_psci_ops_t *psci_plat_pm_ops;
+extern int psci_validate_entry_point(entry_point_info_t *ep,
+			      uintptr_t entrypoint,
+			      u_register_t context_id);
+extern void bl31_lfa_entrypoint(void);
+extern void bl31_warm_entrypoint(void);
+
+static void add_reset_cpu(void)
+{
+	spin_lock(&cpucountlock);
+	reset_cpu_count++;
+	spin_unlock(&cpucountlock);
+}
+
+static uint64_t remove_reset_cpu(void)
+{
+	uint64_t cpu_count;
+
+	spin_lock(&cpucountlock);
+	if (reset_cpu_count > 0) {
+		reset_cpu_count--;
+	}
+	cpu_count = reset_cpu_count;
+	spin_unlock(&cpucountlock);
+
+	return cpu_count;
+}
+
+/*
+ * Prepare CPU warm reset. This functioned should be called before entering
+ * relocatable code for each CPU.
+ */
+uint64_t prepare_warm_reset(uintptr_t lfa_ns_ep, uint64_t context_id, uint64_t cpu_spinlock)
+{
+	const psci_power_state_t state_info = { {ARM_LOCAL_STATE_OFF} };
+	unsigned int core_pos = plat_my_core_pos();
+	lfa_mailbox_message_t mmesg;
+	int rc;
+	static spinlock_t setuplock;
+
+	spin_lock(&setuplock);
+	if (plat_wr_ep == 0 && plat_lfa_mailbox_base()) {
+		plat_wr_ep = *((uintptr_t *)plat_lfa_mailbox_base());
+	}
+
+	if (lfa_psci_ops == NULL) {
+		plat_setup_psci_ops((uintptr_t)bl31_lfa_entrypoint, &lfa_psci_ops);
+	}
+	spin_unlock(&setuplock);
+
+	rc = psci_validate_entry_point(&mmesg.ep, lfa_ns_ep, (u_register_t)context_id);
+	if (rc != PSCI_E_SUCCESS) {
+		ERROR("BL31 LFA: Invalid NS entrypoint! (error %d)\n", rc);
+		return LFA_CRITICAL_ERROR;
+	}
+
+	/* Stash the holding pen lock so we can release it after reboot. */
+	mmesg.spinlock_p = (uint64_t)cpu_spinlock;
+
+	/* Stash the platform warmboot entry point so we can restore it. */
+	mmesg.plat_wr_mailbox = plat_wr_ep;
+	memcpy((void *)&lfa_mailbox[core_pos], &mmesg, sizeof(lfa_mailbox_message_t));
+
+	/*
+	 * TODO: refactor PSCI suspend API to be usable here. The main issue is that
+	 * PSCI suspend will put the core down immediately, but the use case here is
+	 * slightly different since we have to do the firmware copy operation at the
+	 * very end right before putting the core down for reset.
+	 */
+	if (lfa_psci_ops->pwr_domain_suspend != NULL) {
+		/*
+		 * We can't use the proper PSCI API since it will put the core
+		 * down immediately, we still need to copy BL31.
+		 */
+#if USE_GIC_DRIVER
+		gic_cpuif_disable(core_pos);
+#endif
+		lfa_psci_ops->pwr_domain_suspend(&state_info);
+	} else {
+		ERROR("BL31: PSCI power domain suspend not supported!\n");
+		return LFA_CRITICAL_ERROR;
+	}
+
+	add_reset_cpu();
+
+	cm_el2_sysregs_context_save(NON_SECURE);
+
+	return LFA_SUCCESS;
+}
+
+/*
+ * Entrypoint function for CPU warm reset case.
+ */
+void __no_pauth lfa_warm_reset_entrypoint(void)
+{
+	static spinlock_t resetlock;
+	uint32_t counter_freq;
+	const psci_power_state_t state_info = { {ARM_LOCAL_STATE_OFF} };
+	lfa_mailbox_message_t mmesg;
+	unsigned int core_pos = plat_my_core_pos();
+
+#if ENABLE_FEAT_RME
+	/*
+	 * At warm boot GPT data structures have already been initialized in RAM
+	 * but the sysregs for this CPU need to be initialized. Note that the GPT
+	 * accesses are controlled attributes in GPCCR and do not depend on the
+	 * SCR_EL3.C bit.
+	 */
+	if (gpt_enable() != 0) {
+		panic();
+	}
+#endif
+
+	/* TODO see if we can use psci_warmboot_entrypoint here. */
+	psci_plat_pm_ops->pwr_domain_suspend_finish(&state_info);
+#if USE_GIC_DRIVER
+	gic_cpuif_enable(core_pos);
+#endif
+
+	/* Init registers that never change for the lifetime of TF-A */
+	cm_manage_extensions_el3(core_pos);
+
+	/* Re-init the cntfrq_el0 register */
+	counter_freq = plat_get_syscnt_freq2();
+	write_cntfrq_el0(counter_freq);
+
+	memcpy(&mmesg, (void *)&lfa_mailbox[core_pos], sizeof(lfa_mailbox_message_t));
+
+	spin_lock(&resetlock);
+	NOTICE("BL31 LFA Warm Reset on CPU %u\n", core_pos);
+	NOTICE("  Version : %s\n", build_version_string);
+	NOTICE("  %s\n", build_message);
+	console_flush();
+	spin_unlock(&resetlock);
+
+	/* The last CPU to reach here restores the warm entrypoint. */
+	if (remove_reset_cpu() == 0) {
+		plat_arm_program_trusted_mailbox(mmesg.plat_wr_mailbox);
+		plat_wr_ep = 0;
+		lfa_psci_ops = NULL;
+	}
+
+	spin_unlock((spinlock_t *)mmesg.spinlock_p);
+
+	cm_init_my_context(&mmesg.ep);
+
+	cm_prepare_el3_exit_ns();
+}