1 files changed, 512 insertions, 87 deletions

diff --git a/tools/perf/util/arm-spe.c b/tools/perf/util/arm-spe.c
index afbd5869f6bf..2a9775649cc2 100644
--- a/tools/perf/util/arm-spe.c
+++ b/tools/perf/util/arm-spe.c
@@ -37,6 +37,8 @@
 #include "../../arch/arm64/include/asm/cputype.h"
 #define MAX_TIMESTAMP (~0ULL)
 
+#define is_ldst_op(op)		(!!((op) & ARM_SPE_OP_LDST))
+
 struct arm_spe {
 	struct auxtrace			auxtrace;
 	struct auxtrace_queues		queues;
@@ -46,7 +48,6 @@ struct arm_spe {
 	struct perf_session		*session;
 	struct machine			*machine;
 	u32				pmu_type;
-	u64				midr;
 
 	struct perf_tsc_conversion	tc;
 
@@ -69,7 +70,7 @@ struct arm_spe {
 	u64				llc_access_id;
 	u64				tlb_miss_id;
 	u64				tlb_access_id;
-	u64				branch_miss_id;
+	u64				branch_id;
 	u64				remote_access_id;
 	u64				memory_id;
 	u64				instructions_id;
@@ -78,6 +79,11 @@ struct arm_spe {
 
 	unsigned long			num_events;
 	u8				use_ctx_pkt_for_pid;
+
+	u64				**metadata;
+	u64				metadata_ver;
+	u64				metadata_nr_cpu;
+	bool				is_homogeneous;
 };
 
 struct arm_spe_queue {
@@ -96,8 +102,22 @@ struct arm_spe_queue {
 	u64				timestamp;
 	struct thread			*thread;
 	u64				period_instructions;
+	u32				flags;
+	struct branch_stack		*last_branch;
+};
+
+struct data_source_handle {
+	const struct midr_range *midr_ranges;
+	void (*ds_synth)(const struct arm_spe_record *record,
+			 union perf_mem_data_src *data_src);
 };
 
+#define DS(range, func)					\
+	{						\
+		.midr_ranges = range,			\
+		.ds_synth = arm_spe__synth_##func,	\
+	}
+
 static void arm_spe_dump(struct arm_spe *spe __maybe_unused,
 			 unsigned char *buf, size_t len)
 {
@@ -118,7 +138,7 @@ static void arm_spe_dump(struct arm_spe *spe __maybe_unused,
 		else
 			pkt_len = 1;
 		printf(".");
-		color_fprintf(stdout, color, "  %08x: ", pos);
+		color_fprintf(stdout, color, "  %08zx: ", pos);
 		for (i = 0; i < pkt_len; i++)
 			color_fprintf(stdout, color, " %02x", buf[i]);
 		for (; i < 16; i++)
@@ -214,6 +234,17 @@ static struct arm_spe_queue *arm_spe__alloc_queue(struct arm_spe *spe,
 	params.get_trace = arm_spe_get_trace;
 	params.data = speq;
 
+	if (spe->synth_opts.last_branch) {
+		size_t sz = sizeof(struct branch_stack);
+
+		/* Allocate up to two entries for PBT + TGT */
+		sz += sizeof(struct branch_entry) *
+			min(spe->synth_opts.last_branch_sz, 2U);
+		speq->last_branch = zalloc(sz);
+		if (!speq->last_branch)
+			goto out_free;
+	}
+
 	/* create new decoder */
 	speq->decoder = arm_spe_decoder_new(&params);
 	if (!speq->decoder)
@@ -223,6 +254,7 @@ static struct arm_spe_queue *arm_spe__alloc_queue(struct arm_spe *spe,
 
 out_free:
 	zfree(&speq->event_buf);
+	zfree(&speq->last_branch);
 	free(speq);
 
 	return NULL;
@@ -273,6 +305,20 @@ static int arm_spe_set_tid(struct arm_spe_queue *speq, pid_t tid)
 	return 0;
 }
 
+static u64 *arm_spe__get_metadata_by_cpu(struct arm_spe *spe, u64 cpu)
+{
+	u64 i;
+
+	if (!spe->metadata)
+		return NULL;
+
+	for (i = 0; i < spe->metadata_nr_cpu; i++)
+		if (spe->metadata[i][ARM_SPE_CPU] == cpu)
+			return spe->metadata[i];
+
+	return NULL;
+}
+
 static struct simd_flags arm_spe__synth_simd_flags(const struct arm_spe_record *record)
 {
 	struct simd_flags simd_flags = {};
@@ -315,6 +361,88 @@ static void arm_spe_prep_sample(struct arm_spe *spe,
 	event->sample.header.size = sizeof(struct perf_event_header);
 }
 
+static void arm_spe__prep_branch_stack(struct arm_spe_queue *speq)
+{
+	struct arm_spe *spe = speq->spe;
+	struct arm_spe_record *record = &speq->decoder->record;
+	struct branch_stack *bstack = speq->last_branch;
+	struct branch_flags *bs_flags;
+	unsigned int last_branch_sz = spe->synth_opts.last_branch_sz;
+	bool have_tgt = !!(speq->flags & PERF_IP_FLAG_BRANCH);
+	bool have_pbt = last_branch_sz >= (have_tgt + 1U) && record->prev_br_tgt;
+	size_t sz = sizeof(struct branch_stack) +
+		    sizeof(struct branch_entry) * min(last_branch_sz, 2U) /* PBT + TGT */;
+	int i = 0;
+
+	/* Clean up branch stack */
+	memset(bstack, 0x0, sz);
+
+	if (!have_tgt && !have_pbt)
+		return;
+
+	if (have_tgt) {
+		bstack->entries[i].from = record->from_ip;
+		bstack->entries[i].to = record->to_ip;
+
+		bs_flags = &bstack->entries[i].flags;
+		bs_flags->value = 0;
+
+		if (record->op & ARM_SPE_OP_BR_CR_BL) {
+			if (record->op & ARM_SPE_OP_BR_COND)
+				bs_flags->type |= PERF_BR_COND_CALL;
+			else
+				bs_flags->type |= PERF_BR_CALL;
+		/*
+		 * Indirect branch instruction without link (e.g. BR),
+		 * take this case as function return.
+		 */
+		} else if (record->op & ARM_SPE_OP_BR_CR_RET ||
+			   record->op & ARM_SPE_OP_BR_INDIRECT) {
+			if (record->op & ARM_SPE_OP_BR_COND)
+				bs_flags->type |= PERF_BR_COND_RET;
+			else
+				bs_flags->type |= PERF_BR_RET;
+		} else if (record->op & ARM_SPE_OP_BR_CR_NON_BL_RET) {
+			if (record->op & ARM_SPE_OP_BR_COND)
+				bs_flags->type |= PERF_BR_COND;
+			else
+				bs_flags->type |= PERF_BR_UNCOND;
+		} else {
+			if (record->op & ARM_SPE_OP_BR_COND)
+				bs_flags->type |= PERF_BR_COND;
+			else
+				bs_flags->type |= PERF_BR_UNKNOWN;
+		}
+
+		if (record->type & ARM_SPE_BRANCH_MISS) {
+			bs_flags->mispred = 1;
+			bs_flags->predicted = 0;
+		} else {
+			bs_flags->mispred = 0;
+			bs_flags->predicted = 1;
+		}
+
+		if (record->type & ARM_SPE_BRANCH_NOT_TAKEN)
+			bs_flags->not_taken = 1;
+
+		if (record->type & ARM_SPE_IN_TXN)
+			bs_flags->in_tx = 1;
+
+		bs_flags->cycles = min(record->latency, 0xFFFFU);
+		i++;
+	}
+
+	if (have_pbt) {
+		bs_flags = &bstack->entries[i].flags;
+		bs_flags->type |= PERF_BR_UNKNOWN;
+		bstack->entries[i].to = record->prev_br_tgt;
+		i++;
+	}
+
+	bstack->nr = i;
+	bstack->hw_idx = -1ULL;
+}
+
 static int arm_spe__inject_event(union perf_event *event, struct perf_sample *sample, u64 type)
 {
 	event->header.size = perf_event__sample_event_size(sample, type, 0);
@@ -348,8 +476,10 @@ static int arm_spe__synth_mem_sample(struct arm_spe_queue *speq,
 	struct arm_spe *spe = speq->spe;
 	struct arm_spe_record *record = &speq->decoder->record;
 	union perf_event *event = speq->event_buf;
-	struct perf_sample sample = { .ip = 0, };
+	struct perf_sample sample;
+	int ret;
 
+	perf_sample__init(&sample, /*all=*/true);
 	arm_spe_prep_sample(spe, speq, event, &sample);
 
 	sample.id = spe_events_id;
@@ -359,7 +489,9 @@ static int arm_spe__synth_mem_sample(struct arm_spe_queue *speq,
 	sample.data_src = data_src;
 	sample.weight = record->latency;
 
-	return arm_spe_deliver_synth_event(spe, speq, event, &sample);
+	ret = arm_spe_deliver_synth_event(spe, speq, event, &sample);
+	perf_sample__exit(&sample);
+	return ret;
 }
 
 static int arm_spe__synth_branch_sample(struct arm_spe_queue *speq,
@@ -368,16 +500,22 @@ static int arm_spe__synth_branch_sample(struct arm_spe_queue *speq,
 	struct arm_spe *spe = speq->spe;
 	struct arm_spe_record *record = &speq->decoder->record;
 	union perf_event *event = speq->event_buf;
-	struct perf_sample sample = { .ip = 0, };
+	struct perf_sample sample;
+	int ret;
 
+	perf_sample__init(&sample, /*all=*/true);
 	arm_spe_prep_sample(spe, speq, event, &sample);
 
 	sample.id = spe_events_id;
 	sample.stream_id = spe_events_id;
 	sample.addr = record->to_ip;
 	sample.weight = record->latency;
+	sample.flags = speq->flags;
+	sample.branch_stack = speq->last_branch;
 
-	return arm_spe_deliver_synth_event(spe, speq, event, &sample);
+	ret = arm_spe_deliver_synth_event(spe, speq, event, &sample);
+	perf_sample__exit(&sample);
+	return ret;
 }
 
 static int arm_spe__synth_instruction_sample(struct arm_spe_queue *speq,
@@ -386,7 +524,8 @@ static int arm_spe__synth_instruction_sample(struct arm_spe_queue *speq,
 	struct arm_spe *spe = speq->spe;
 	struct arm_spe_record *record = &speq->decoder->record;
 	union perf_event *event = speq->event_buf;
-	struct perf_sample sample = { .ip = 0, };
+	struct perf_sample sample;
+	int ret;
 
 	/*
 	 * Handles perf instruction sampling period.
@@ -396,28 +535,77 @@ static int arm_spe__synth_instruction_sample(struct arm_spe_queue *speq,
 		return 0;
 	speq->period_instructions = 0;
 
+	perf_sample__init(&sample, /*all=*/true);
 	arm_spe_prep_sample(spe, speq, event, &sample);
 
 	sample.id = spe_events_id;
 	sample.stream_id = spe_events_id;
-	sample.addr = record->virt_addr;
+	sample.addr = record->to_ip;
 	sample.phys_addr = record->phys_addr;
 	sample.data_src = data_src;
 	sample.period = spe->instructions_sample_period;
 	sample.weight = record->latency;
+	sample.flags = speq->flags;
+	sample.branch_stack = speq->last_branch;
 
-	return arm_spe_deliver_synth_event(spe, speq, event, &sample);
+	ret = arm_spe_deliver_synth_event(spe, speq, event, &sample);
+	perf_sample__exit(&sample);
+	return ret;
 }
 
-static const struct midr_range neoverse_spe[] = {
+static const struct midr_range common_ds_encoding_cpus[] = {
+	MIDR_ALL_VERSIONS(MIDR_CORTEX_A720),
+	MIDR_ALL_VERSIONS(MIDR_CORTEX_A725),
+	MIDR_ALL_VERSIONS(MIDR_CORTEX_X1C),
+	MIDR_ALL_VERSIONS(MIDR_CORTEX_X3),
+	MIDR_ALL_VERSIONS(MIDR_CORTEX_X925),
 	MIDR_ALL_VERSIONS(MIDR_NEOVERSE_N1),
 	MIDR_ALL_VERSIONS(MIDR_NEOVERSE_N2),
 	MIDR_ALL_VERSIONS(MIDR_NEOVERSE_V1),
+	MIDR_ALL_VERSIONS(MIDR_NEOVERSE_V2),
 	{},
 };
 
-static void arm_spe__synth_data_source_neoverse(const struct arm_spe_record *record,
-						union perf_mem_data_src *data_src)
+static const struct midr_range ampereone_ds_encoding_cpus[] = {
+	MIDR_ALL_VERSIONS(MIDR_AMPERE1A),
+	{},
+};
+
+static void arm_spe__sample_flags(struct arm_spe_queue *speq)
+{
+	const struct arm_spe_record *record = &speq->decoder->record;
+
+	speq->flags = 0;
+	if (record->op & ARM_SPE_OP_BRANCH_ERET) {
+		speq->flags = PERF_IP_FLAG_BRANCH;
+
+		if (record->type & ARM_SPE_BRANCH_MISS)
+			speq->flags |= PERF_IP_FLAG_BRANCH_MISS;
+
+		if (record->type & ARM_SPE_BRANCH_NOT_TAKEN)
+			speq->flags |= PERF_IP_FLAG_NOT_TAKEN;
+
+		if (record->type & ARM_SPE_IN_TXN)
+			speq->flags |= PERF_IP_FLAG_IN_TX;
+
+		if (record->op & ARM_SPE_OP_BR_COND)
+			speq->flags |= PERF_IP_FLAG_CONDITIONAL;
+
+		if (record->op & ARM_SPE_OP_BR_CR_BL)
+			speq->flags |= PERF_IP_FLAG_CALL;
+		else if (record->op & ARM_SPE_OP_BR_CR_RET)
+			speq->flags |= PERF_IP_FLAG_RETURN;
+		/*
+		 * Indirect branch instruction without link (e.g. BR),
+		 * take it as a function return.
+		 */
+		else if (record->op & ARM_SPE_OP_BR_INDIRECT)
+			speq->flags |= PERF_IP_FLAG_RETURN;
+	}
+}
+
+static void arm_spe__synth_data_source_common(const struct arm_spe_record *record,
+					      union perf_mem_data_src *data_src)
 {
 	/*
 	 * Even though four levels of cache hierarchy are possible, no known
@@ -439,17 +627,17 @@ static void arm_spe__synth_data_source_neoverse(const struct arm_spe_record *rec
 	}
 
 	switch (record->source) {
-	case ARM_SPE_NV_L1D:
+	case ARM_SPE_COMMON_DS_L1D:
 		data_src->mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_HIT;
 		data_src->mem_lvl_num = PERF_MEM_LVLNUM_L1;
 		data_src->mem_snoop = PERF_MEM_SNOOP_NONE;
 		break;
-	case ARM_SPE_NV_L2:
+	case ARM_SPE_COMMON_DS_L2:
 		data_src->mem_lvl = PERF_MEM_LVL_L2 | PERF_MEM_LVL_HIT;
 		data_src->mem_lvl_num = PERF_MEM_LVLNUM_L2;
 		data_src->mem_snoop = PERF_MEM_SNOOP_NONE;
 		break;
-	case ARM_SPE_NV_PEER_CORE:
+	case ARM_SPE_COMMON_DS_PEER_CORE:
 		data_src->mem_lvl = PERF_MEM_LVL_L2 | PERF_MEM_LVL_HIT;
 		data_src->mem_lvl_num = PERF_MEM_LVLNUM_L2;
 		data_src->mem_snoopx = PERF_MEM_SNOOPX_PEER;
@@ -458,8 +646,8 @@ static void arm_spe__synth_data_source_neoverse(const struct arm_spe_record *rec
 	 * We don't know if this is L1, L2 but we do know it was a cache-2-cache
 	 * transfer, so set SNOOPX_PEER
 	 */
-	case ARM_SPE_NV_LOCAL_CLUSTER:
-	case ARM_SPE_NV_PEER_CLUSTER:
+	case ARM_SPE_COMMON_DS_LOCAL_CLUSTER:
+	case ARM_SPE_COMMON_DS_PEER_CLUSTER:
 		data_src->mem_lvl = PERF_MEM_LVL_L3 | PERF_MEM_LVL_HIT;
 		data_src->mem_lvl_num = PERF_MEM_LVLNUM_L3;
 		data_src->mem_snoopx = PERF_MEM_SNOOPX_PEER;
@@ -467,7 +655,7 @@ static void arm_spe__synth_data_source_neoverse(const struct arm_spe_record *rec
 	/*
 	 * System cache is assumed to be L3
 	 */
-	case ARM_SPE_NV_SYS_CACHE:
+	case ARM_SPE_COMMON_DS_SYS_CACHE:
 		data_src->mem_lvl = PERF_MEM_LVL_L3 | PERF_MEM_LVL_HIT;
 		data_src->mem_lvl_num = PERF_MEM_LVLNUM_L3;
 		data_src->mem_snoop = PERF_MEM_SNOOP_HIT;
@@ -476,13 +664,13 @@ static void arm_spe__synth_data_source_neoverse(const struct arm_spe_record *rec
 	 * We don't know what level it hit in, except it came from the other
 	 * socket
 	 */
-	case ARM_SPE_NV_REMOTE:
+	case ARM_SPE_COMMON_DS_REMOTE:
 		data_src->mem_lvl = PERF_MEM_LVL_REM_CCE1;
 		data_src->mem_lvl_num = PERF_MEM_LVLNUM_ANY_CACHE;
 		data_src->mem_remote = PERF_MEM_REMOTE_REMOTE;
 		data_src->mem_snoopx = PERF_MEM_SNOOPX_PEER;
 		break;
-	case ARM_SPE_NV_DRAM:
+	case ARM_SPE_COMMON_DS_DRAM:
 		data_src->mem_lvl = PERF_MEM_LVL_LOC_RAM | PERF_MEM_LVL_HIT;
 		data_src->mem_lvl_num = PERF_MEM_LVLNUM_RAM;
 		data_src->mem_snoop = PERF_MEM_SNOOP_NONE;
@@ -492,8 +680,51 @@ static void arm_spe__synth_data_source_neoverse(const struct arm_spe_record *rec
 	}
 }
 
-static void arm_spe__synth_data_source_generic(const struct arm_spe_record *record,
-					       union perf_mem_data_src *data_src)
+/*
+ * Source is IMPDEF. Here we convert the source code used on AmpereOne cores
+ * to the common (Neoverse, Cortex) to avoid duplicating the decoding code.
+ */
+static void arm_spe__synth_data_source_ampereone(const struct arm_spe_record *record,
+						 union perf_mem_data_src *data_src)
+{
+	struct arm_spe_record common_record;
+
+	switch (record->source) {
+	case ARM_SPE_AMPEREONE_LOCAL_CHIP_CACHE_OR_DEVICE:
+		common_record.source = ARM_SPE_COMMON_DS_PEER_CORE;
+		break;
+	case ARM_SPE_AMPEREONE_SLC:
+		common_record.source = ARM_SPE_COMMON_DS_SYS_CACHE;
+		break;
+	case ARM_SPE_AMPEREONE_REMOTE_CHIP_CACHE:
+		common_record.source = ARM_SPE_COMMON_DS_REMOTE;
+		break;
+	case ARM_SPE_AMPEREONE_DDR:
+		common_record.source = ARM_SPE_COMMON_DS_DRAM;
+		break;
+	case ARM_SPE_AMPEREONE_L1D:
+		common_record.source = ARM_SPE_COMMON_DS_L1D;
+		break;
+	case ARM_SPE_AMPEREONE_L2D:
+		common_record.source = ARM_SPE_COMMON_DS_L2;
+		break;
+	default:
+		pr_warning_once("AmpereOne: Unknown data source (0x%x)\n",
+				record->source);
+		return;
+	}
+
+	common_record.op = record->op;
+	arm_spe__synth_data_source_common(&common_record, data_src);
+}
+
+static const struct data_source_handle data_source_handles[] = {
+	DS(common_ds_encoding_cpus, data_source_common),
+	DS(ampereone_ds_encoding_cpus, data_source_ampereone),
+};
+
+static void arm_spe__synth_memory_level(const struct arm_spe_record *record,
+					union perf_mem_data_src *data_src)
 {
 	if (record->type & (ARM_SPE_LLC_ACCESS | ARM_SPE_LLC_MISS)) {
 		data_src->mem_lvl = PERF_MEM_LVL_L3;
@@ -515,10 +746,63 @@ static void arm_spe__synth_data_source_generic(const struct arm_spe_record *reco
 		data_src->mem_lvl |= PERF_MEM_LVL_REM_CCE1;
 }
 
-static u64 arm_spe__synth_data_source(const struct arm_spe_record *record, u64 midr)
+static bool arm_spe__synth_ds(struct arm_spe_queue *speq,
+			      const struct arm_spe_record *record,
+			      union perf_mem_data_src *data_src)
+{
+	struct arm_spe *spe = speq->spe;
+	u64 *metadata = NULL;
+	u64 midr;
+	unsigned int i;
+
+	/* Metadata version 1 assumes all CPUs are the same (old behavior) */
+	if (spe->metadata_ver == 1) {
+		const char *cpuid;
+
+		pr_warning_once("Old SPE metadata, re-record to improve decode accuracy\n");
+		cpuid = perf_env__cpuid(spe->session->evlist->env);
+		midr = strtol(cpuid, NULL, 16);
+	} else {
+		/* CPU ID is -1 for per-thread mode */
+		if (speq->cpu < 0) {
+			/*
+			 * On the heterogeneous system, due to CPU ID is -1,
+			 * cannot confirm the data source packet is supported.
+			 */
+			if (!spe->is_homogeneous)
+				return false;
+
+			/* In homogeneous system, simply use CPU0's metadata */
+			if (spe->metadata)
+				metadata = spe->metadata[0];
+		} else {
+			metadata = arm_spe__get_metadata_by_cpu(spe, speq->cpu);
+		}
+
+		if (!metadata)
+			return false;
+
+		midr = metadata[ARM_SPE_CPU_MIDR];
+	}
+
+	for (i = 0; i < ARRAY_SIZE(data_source_handles); i++) {
+		if (is_midr_in_range_list(midr, data_source_handles[i].midr_ranges)) {
+			data_source_handles[i].ds_synth(record, data_src);
+			return true;
+		}
+	}
+
+	return false;
+}
+
+static u64 arm_spe__synth_data_source(struct arm_spe_queue *speq,
+				      const struct arm_spe_record *record)
 {
 	union perf_mem_data_src	data_src = { .mem_op = PERF_MEM_OP_NA };
-	bool is_neoverse = is_midr_in_range_list(midr, neoverse_spe);
+
+	/* Only synthesize data source for LDST operations */
+	if (!is_ldst_op(record->op))
+		return 0;
 
 	if (record->op & ARM_SPE_OP_LD)
 		data_src.mem_op = PERF_MEM_OP_LOAD;
@@ -527,10 +811,8 @@ static u64 arm_spe__synth_data_source(const struct arm_spe_record *record, u64 m
 	else
 		return 0;
 
-	if (is_neoverse)
-		arm_spe__synth_data_source_neoverse(record, &data_src);
-	else
-		arm_spe__synth_data_source_generic(record, &data_src);
+	if (!arm_spe__synth_ds(speq, record, &data_src))
+		arm_spe__synth_memory_level(record, &data_src);
 
 	if (record->type & (ARM_SPE_TLB_ACCESS | ARM_SPE_TLB_MISS)) {
 		data_src.mem_dtlb = PERF_MEM_TLB_WK;
@@ -551,7 +833,8 @@ static int arm_spe_sample(struct arm_spe_queue *speq)
 	u64 data_src;
 	int err;
 
-	data_src = arm_spe__synth_data_source(record, spe->midr);
+	arm_spe__sample_flags(speq);
+	data_src = arm_spe__synth_data_source(speq, record);
 
 	if (spe->sample_flc) {
 		if (record->type & ARM_SPE_L1D_MISS) {
@@ -601,8 +884,12 @@ static int arm_spe_sample(struct arm_spe_queue *speq)
 		}
 	}
 
-	if (spe->sample_branch && (record->type & ARM_SPE_BRANCH_MISS)) {
-		err = arm_spe__synth_branch_sample(speq, spe->branch_miss_id);
+	if (spe->synth_opts.last_branch &&
+	    (spe->sample_branch || spe->sample_instructions))
+		arm_spe__prep_branch_stack(speq);
+
+	if (spe->sample_branch && (record->op & ARM_SPE_OP_BRANCH_ERET)) {
+		err = arm_spe__synth_branch_sample(speq, spe->branch_id);
 		if (err)
 			return err;
 	}
@@ -619,7 +906,7 @@ static int arm_spe_sample(struct arm_spe_queue *speq)
 	 * When data_src is zero it means the record is not a memory operation,
 	 * skip to synthesize memory sample for this case.
 	 */
-	if (spe->sample_memory && data_src) {
+	if (spe->sample_memory && is_ldst_op(record->op)) {
 		err = arm_spe__synth_mem_sample(speq, spe->memory_id, data_src);
 		if (err)
 			return err;
@@ -899,7 +1186,7 @@ static int arm_spe_context_switch(struct arm_spe *spe, union perf_event *event,
 static int arm_spe_process_event(struct perf_session *session,
 				 union perf_event *event,
 				 struct perf_sample *sample,
-				 struct perf_tool *tool)
+				 const struct perf_tool *tool)
 {
 	int err = 0;
 	u64 timestamp;
@@ -947,7 +1234,7 @@ static int arm_spe_process_event(struct perf_session *session,
 
 static int arm_spe_process_auxtrace_event(struct perf_session *session,
 					  union perf_event *event,
-					  struct perf_tool *tool __maybe_unused)
+					  const struct perf_tool *tool __maybe_unused)
 {
 	struct arm_spe *spe = container_of(session->auxtrace, struct arm_spe,
 					     auxtrace);
@@ -985,7 +1272,7 @@ static int arm_spe_process_auxtrace_event(struct perf_session *session,
 }
 
 static int arm_spe_flush(struct perf_session *session __maybe_unused,
-			 struct perf_tool *tool __maybe_unused)
+			 const struct perf_tool *tool __maybe_unused)
 {
 	struct arm_spe *spe = container_of(session->auxtrace, struct arm_spe,
 			auxtrace);
@@ -1016,6 +1303,73 @@ static int arm_spe_flush(struct perf_session *session __maybe_unused,
 	return 0;
 }
 
+static u64 *arm_spe__alloc_per_cpu_metadata(u64 *buf, int per_cpu_size)
+{
+	u64 *metadata;
+
+	metadata = zalloc(per_cpu_size);
+	if (!metadata)
+		return NULL;
+
+	memcpy(metadata, buf, per_cpu_size);
+	return metadata;
+}
+
+static void arm_spe__free_metadata(u64 **metadata, int nr_cpu)
+{
+	int i;
+
+	for (i = 0; i < nr_cpu; i++)
+		zfree(&metadata[i]);
+	free(metadata);
+}
+
+static u64 **arm_spe__alloc_metadata(struct perf_record_auxtrace_info *info,
+				     u64 *ver, int *nr_cpu)
+{
+	u64 *ptr = (u64 *)info->priv;
+	u64 metadata_size;
+	u64 **metadata = NULL;
+	int hdr_sz, per_cpu_sz, i;
+
+	metadata_size = info->header.size -
+		sizeof(struct perf_record_auxtrace_info);
+
+	/* Metadata version 1 */
+	if (metadata_size == ARM_SPE_AUXTRACE_V1_PRIV_SIZE) {
+		*ver = 1;
+		*nr_cpu = 0;
+		/* No per CPU metadata */
+		return NULL;
+	}
+
+	*ver = ptr[ARM_SPE_HEADER_VERSION];
+	hdr_sz = ptr[ARM_SPE_HEADER_SIZE];
+	*nr_cpu = ptr[ARM_SPE_CPUS_NUM];
+
+	metadata = calloc(*nr_cpu, sizeof(*metadata));
+	if (!metadata)
+		return NULL;
+
+	/* Locate the start address of per CPU metadata */
+	ptr += hdr_sz;
+	per_cpu_sz = (metadata_size - (hdr_sz * sizeof(u64))) / (*nr_cpu);
+
+	for (i = 0; i < *nr_cpu; i++) {
+		metadata[i] = arm_spe__alloc_per_cpu_metadata(ptr, per_cpu_sz);
+		if (!metadata[i])
+			goto err_per_cpu_metadata;
+
+		ptr += per_cpu_sz / sizeof(u64);
+	}
+
+	return metadata;
+
+err_per_cpu_metadata:
+	arm_spe__free_metadata(metadata, *nr_cpu);
+	return NULL;
+}
+
 static void arm_spe_free_queue(void *priv)
 {
 	struct arm_spe_queue *speq = priv;
@@ -1025,6 +1379,7 @@ static void arm_spe_free_queue(void *priv)
 	thread__zput(speq->thread);
 	arm_spe_decoder_free(speq->decoder);
 	zfree(&speq->event_buf);
+	zfree(&speq->last_branch);
 	free(speq);
 }
 
@@ -1050,6 +1405,7 @@ static void arm_spe_free(struct perf_session *session)
 	auxtrace_heap__free(&spe->heap);
 	arm_spe_free_events(session);
 	session->auxtrace = NULL;
+	arm_spe__free_metadata(spe->metadata, spe->metadata_nr_cpu);
 	free(spe);
 }
 
@@ -1061,45 +1417,60 @@ static bool arm_spe_evsel_is_auxtrace(struct perf_session *session,
 	return evsel->core.attr.type == spe->pmu_type;
 }
 
-static const char * const arm_spe_info_fmts[] = {
-	[ARM_SPE_PMU_TYPE]		= "  PMU Type           %"PRId64"\n",
+static const char * const metadata_hdr_v1_fmts[] = {
+	[ARM_SPE_PMU_TYPE]		= "  PMU Type           :%"PRId64"\n",
+	[ARM_SPE_PER_CPU_MMAPS]		= "  Per CPU mmaps      :%"PRId64"\n",
 };
 
-static void arm_spe_print_info(__u64 *arr)
-{
-	if (!dump_trace)
-		return;
-
-	fprintf(stdout, arm_spe_info_fmts[ARM_SPE_PMU_TYPE], arr[ARM_SPE_PMU_TYPE]);
-}
+static const char * const metadata_hdr_fmts[] = {
+	[ARM_SPE_HEADER_VERSION]	= "  Header version     :%"PRId64"\n",
+	[ARM_SPE_HEADER_SIZE]		= "  Header size        :%"PRId64"\n",
+	[ARM_SPE_PMU_TYPE_V2]		= "  PMU type v2        :%"PRId64"\n",
+	[ARM_SPE_CPUS_NUM]		= "  CPU number         :%"PRId64"\n",
+};
 
-struct arm_spe_synth {
-	struct perf_tool dummy_tool;
-	struct perf_session *session;
+static const char * const metadata_per_cpu_fmts[] = {
+	[ARM_SPE_MAGIC]			= "    Magic            :0x%"PRIx64"\n",
+	[ARM_SPE_CPU]			= "    CPU #            :%"PRId64"\n",
+	[ARM_SPE_CPU_NR_PARAMS]		= "    Num of params    :%"PRId64"\n",
+	[ARM_SPE_CPU_MIDR]		= "    MIDR             :0x%"PRIx64"\n",
+	[ARM_SPE_CPU_PMU_TYPE]		= "    PMU Type         :%"PRId64"\n",
+	[ARM_SPE_CAP_MIN_IVAL]		= "    Min Interval     :%"PRId64"\n",
 };
 
-static int arm_spe_event_synth(struct perf_tool *tool,
-			       union perf_event *event,
-			       struct perf_sample *sample __maybe_unused,
-			       struct machine *machine __maybe_unused)
+static void arm_spe_print_info(struct arm_spe *spe, __u64 *arr)
 {
-	struct arm_spe_synth *arm_spe_synth =
-		      container_of(tool, struct arm_spe_synth, dummy_tool);
+	unsigned int i, cpu, hdr_size, cpu_num, cpu_size;
+	const char * const *hdr_fmts;
 
-	return perf_session__deliver_synth_event(arm_spe_synth->session,
-						 event, NULL);
-}
+	if (!dump_trace)
+		return;
 
-static int arm_spe_synth_event(struct perf_session *session,
-			       struct perf_event_attr *attr, u64 id)
-{
-	struct arm_spe_synth arm_spe_synth;
+	if (spe->metadata_ver == 1) {
+		cpu_num = 0;
+		hdr_size = ARM_SPE_AUXTRACE_V1_PRIV_MAX;
+		hdr_fmts = metadata_hdr_v1_fmts;
+	} else {
+		cpu_num = arr[ARM_SPE_CPUS_NUM];
+		hdr_size = arr[ARM_SPE_HEADER_SIZE];
+		hdr_fmts = metadata_hdr_fmts;
+	}
 
-	memset(&arm_spe_synth, 0, sizeof(struct arm_spe_synth));
-	arm_spe_synth.session = session;
+	for (i = 0; i < hdr_size; i++)
+		fprintf(stdout, hdr_fmts[i], arr[i]);
 
-	return perf_event__synthesize_attr(&arm_spe_synth.dummy_tool, attr, 1,
-					   &id, arm_spe_event_synth);
+	arr += hdr_size;
+	for (cpu = 0; cpu < cpu_num; cpu++) {
+		/*
+		 * The parameters from ARM_SPE_MAGIC to ARM_SPE_CPU_NR_PARAMS
+		 * are fixed. The sequential parameter size is decided by the
+		 * field 'ARM_SPE_CPU_NR_PARAMS'.
+		 */
+		cpu_size = (ARM_SPE_CPU_NR_PARAMS + 1) + arr[ARM_SPE_CPU_NR_PARAMS];
+		for (i = 0; i < cpu_size; i++)
+			fprintf(stdout, metadata_per_cpu_fmts[i], arr[i]);
+		arr += cpu_size;
+	}
 }
 
 static void arm_spe_set_event_name(struct evlist *evlist, u64 id,
@@ -1172,7 +1543,7 @@ arm_spe_synth_events(struct arm_spe *spe, struct perf_session *session)
 		spe->sample_flc = true;
 
 		/* Level 1 data cache miss */
-		err = arm_spe_synth_event(session, &attr, id);
+		err = perf_session__deliver_synth_attr_event(session, &attr, id);
 		if (err)
 			return err;
 		spe->l1d_miss_id = id;
@@ -1180,7 +1551,7 @@ arm_spe_synth_events(struct arm_spe *spe, struct perf_session *session)
 		id += 1;
 
 		/* Level 1 data cache access */
-		err = arm_spe_synth_event(session, &attr, id);
+		err = perf_session__deliver_synth_attr_event(session, &attr, id);
 		if (err)
 			return err;
 		spe->l1d_access_id = id;
@@ -1192,7 +1563,7 @@ arm_spe_synth_events(struct arm_spe *spe, struct perf_session *session)
 		spe->sample_llc = true;
 
 		/* Last level cache miss */
-		err = arm_spe_synth_event(session, &attr, id);
+		err = perf_session__deliver_synth_attr_event(session, &attr, id);
 		if (err)
 			return err;
 		spe->llc_miss_id = id;
@@ -1200,7 +1571,7 @@ arm_spe_synth_events(struct arm_spe *spe, struct perf_session *session)
 		id += 1;
 
 		/* Last level cache access */
-		err = arm_spe_synth_event(session, &attr, id);
+		err = perf_session__deliver_synth_attr_event(session, &attr, id);
 		if (err)
 			return err;
 		spe->llc_access_id = id;
@@ -1212,7 +1583,7 @@ arm_spe_synth_events(struct arm_spe *spe, struct perf_session *session)
 		spe->sample_tlb = true;
 
 		/* TLB miss */
-		err = arm_spe_synth_event(session, &attr, id);
+		err = perf_session__deliver_synth_attr_event(session, &attr, id);
 		if (err)
 			return err;
 		spe->tlb_miss_id = id;
@@ -1220,7 +1591,7 @@ arm_spe_synth_events(struct arm_spe *spe, struct perf_session *session)
 		id += 1;
 
 		/* TLB access */
-		err = arm_spe_synth_event(session, &attr, id);
+		err = perf_session__deliver_synth_attr_event(session, &attr, id);
 		if (err)
 			return err;
 		spe->tlb_access_id = id;
@@ -1228,15 +1599,28 @@ arm_spe_synth_events(struct arm_spe *spe, struct perf_session *session)
 		id += 1;
 	}
 
+	if (spe->synth_opts.last_branch) {
+		if (spe->synth_opts.last_branch_sz > 2)
+			pr_debug("Arm SPE supports only two bstack entries (PBT+TGT).\n");
+
+		attr.sample_type |= PERF_SAMPLE_BRANCH_STACK;
+		/*
+		 * We don't use the hardware index, but the sample generation
+		 * code uses the new format branch_stack with this field,
+		 * so the event attributes must indicate that it's present.
+		 */
+		attr.branch_sample_type |= PERF_SAMPLE_BRANCH_HW_INDEX;
+	}
+
 	if (spe->synth_opts.branches) {
 		spe->sample_branch = true;
 
-		/* Branch miss */
-		err = arm_spe_synth_event(session, &attr, id);
+		/* Branch */
+		err = perf_session__deliver_synth_attr_event(session, &attr, id);
 		if (err)
 			return err;
-		spe->branch_miss_id = id;
-		arm_spe_set_event_name(evlist, id, "branch-miss");
+		spe->branch_id = id;
+		arm_spe_set_event_name(evlist, id, "branch");
 		id += 1;
 	}
 
@@ -1244,7 +1628,7 @@ arm_spe_synth_events(struct arm_spe *spe, struct perf_session *session)
 		spe->sample_remote_access = true;
 
 		/* Remote access */
-		err = arm_spe_synth_event(session, &attr, id);
+		err = perf_session__deliver_synth_attr_event(session, &attr, id);
 		if (err)
 			return err;
 		spe->remote_access_id = id;
@@ -1255,7 +1639,7 @@ arm_spe_synth_events(struct arm_spe *spe, struct perf_session *session)
 	if (spe->synth_opts.mem) {
 		spe->sample_memory = true;
 
-		err = arm_spe_synth_event(session, &attr, id);
+		err = perf_session__deliver_synth_attr_event(session, &attr, id);
 		if (err)
 			return err;
 		spe->memory_id = id;
@@ -1276,7 +1660,7 @@ arm_spe_synth_events(struct arm_spe *spe, struct perf_session *session)
 		attr.config = PERF_COUNT_HW_INSTRUCTIONS;
 		attr.sample_period = spe->synth_opts.period;
 		spe->instructions_sample_period = attr.sample_period;
-		err = arm_spe_synth_event(session, &attr, id);
+		err = perf_session__deliver_synth_attr_event(session, &attr, id);
 		if (err)
 			return err;
 		spe->instructions_id = id;
@@ -1287,24 +1671,57 @@ synth_instructions_out:
 	return 0;
 }
 
+static bool arm_spe__is_homogeneous(u64 **metadata, int nr_cpu)
+{
+	u64 midr;
+	int i;
+
+	if (!nr_cpu)
+		return false;
+
+	for (i = 0; i < nr_cpu; i++) {
+		if (!metadata[i])
+			return false;
+
+		if (i == 0) {
+			midr = metadata[i][ARM_SPE_CPU_MIDR];
+			continue;
+		}
+
+		if (midr != metadata[i][ARM_SPE_CPU_MIDR])
+			return false;
+	}
+
+	return true;
+}
+
 int arm_spe_process_auxtrace_info(union perf_event *event,
 				  struct perf_session *session)
 {
 	struct perf_record_auxtrace_info *auxtrace_info = &event->auxtrace_info;
-	size_t min_sz = sizeof(u64) * ARM_SPE_AUXTRACE_PRIV_MAX;
+	size_t min_sz = ARM_SPE_AUXTRACE_V1_PRIV_SIZE;
 	struct perf_record_time_conv *tc = &session->time_conv;
-	const char *cpuid = perf_env__cpuid(session->evlist->env);
-	u64 midr = strtol(cpuid, NULL, 16);
 	struct arm_spe *spe;
-	int err;
+	u64 **metadata = NULL;
+	u64 metadata_ver;
+	int nr_cpu, err;
 
 	if (auxtrace_info->header.size < sizeof(struct perf_record_auxtrace_info) +
 					min_sz)
 		return -EINVAL;
 
+	metadata = arm_spe__alloc_metadata(auxtrace_info, &metadata_ver,
+					   &nr_cpu);
+	if (!metadata && metadata_ver != 1) {
+		pr_err("Failed to parse Arm SPE metadata.\n");
+		return -EINVAL;
+	}
+
 	spe = zalloc(sizeof(struct arm_spe));
-	if (!spe)
-		return -ENOMEM;
+	if (!spe) {
+		err = -ENOMEM;
+		goto err_free_metadata;
+	}
 
 	err = auxtrace_queues__init(&spe->queues);
 	if (err)
@@ -1313,8 +1730,14 @@ int arm_spe_process_auxtrace_info(union perf_event *event,
 	spe->session = session;
 	spe->machine = &session->machines.host; /* No kvm support */
 	spe->auxtrace_type = auxtrace_info->type;
-	spe->pmu_type = auxtrace_info->priv[ARM_SPE_PMU_TYPE];
-	spe->midr = midr;
+	if (metadata_ver == 1)
+		spe->pmu_type = auxtrace_info->priv[ARM_SPE_PMU_TYPE];
+	else
+		spe->pmu_type = auxtrace_info->priv[ARM_SPE_PMU_TYPE_V2];
+	spe->metadata = metadata;
+	spe->metadata_ver = metadata_ver;
+	spe->metadata_nr_cpu = nr_cpu;
+	spe->is_homogeneous = arm_spe__is_homogeneous(metadata, nr_cpu);
 
 	spe->timeless_decoding = arm_spe__is_timeless_decoding(spe);
 
@@ -1347,7 +1770,7 @@ int arm_spe_process_auxtrace_info(union perf_event *event,
 	spe->auxtrace.evsel_is_auxtrace = arm_spe_evsel_is_auxtrace;
 	session->auxtrace = &spe->auxtrace;
 
-	arm_spe_print_info(&auxtrace_info->priv[0]);
+	arm_spe_print_info(spe, &auxtrace_info->priv[0]);
 
 	if (dump_trace)
 		return 0;
@@ -1375,5 +1798,7 @@ err_free_queues:
 	session->auxtrace = NULL;
 err_free:
 	free(spe);
+err_free_metadata:
+	arm_spe__free_metadata(metadata, nr_cpu);
 	return err;
 }