Skip to contentSkip to conversation

perf/mem: AMD IBS and generic tools improvements

PATCH Superseded Ravi Bangoria 3 participants 17 messages Apr 7, 2023 – Jun 30, 2023

Subsystems (2)

v3 of 4 Superseded Latest is v4 →

lore
Share & tools
lore
#0 lore raw

v3 cover perf/mem: AMD IBS and generic tools improvements

Kernel IBS driver wasn't using new PERF_MEM_* APIs due to some of its limitations. Mainly:

1. mem_lvl_num doesn't allow setting multiple sources whereas old API allows it. Setting multiple data sources is useful because IBS on pre-zen4 uarch doesn't provide fine granular DataSrc details (there is only one such DataSrc(2h) though).
2. perf mem sorting logic (sort__lvl_cmp()) ignores mem_lvl_num. perf c2c (c2c_decode_stats()) does not use mem_lvl_num at all. perf mem prints mem_lvl and mem_lvl_num both if both are set, which is ugly.

Set mem_lvl_num, mem_remote and mem_hops for data_src via IBS. Handle first issue using mem_lvl_num = ANY_CACHE | HOPS_0. In addition to setting new API fields, convert all individual field assignments to compile time wrapper macros built using PERF_MEM_S(). Also convert DataSrc conditional code to array lookups.

Interpretation of perf_mem_data_src by perf_mem__lvl_scnprintf() was non-intuitive. Make it sane.

v2: https://lore.kernel.org/r/20230327130851.1565-1-ravi.bangoria%40amd.com
v2->v3:
- IBS: Don't club RmtNode with DataSrc=7 (IO)
- Make perf_mem__lvl_scnprintf() more sane
- Introduce PERF_MEM_LVLNUM_UNC, set it along with PERF_MEM_LVL_UNC and interpreat it in tool.
- Add PERF_MEM_LVLNUM_NA to default data_src value
- Change some of the IBS bit description according to latest PPR

Namhyung Kim (1):
perf/x86/ibs: Set mem_lvl_num, mem_remote and mem_hops for data_src

Ravi Bangoria (8):
perf/mem: Introduce PERF_MEM_LVLNUM_UNC
perf/mem: Add PERF_MEM_LVLNUM_NA to PERF_MEM_NA
perf headers: Sync uapi/linux/perf_event.h
perf mem: Add PERF_MEM_LVLNUM_NA to PERF_MEM_DATA_SRC_NONE perf mem: Add support for printing PERF_MEM_LVLNUM_UNC perf mem: Refactor perf_mem__lvl_scnprintf()
perf mem: Increase HISTC_MEM_LVL column size to 39 chars perf script ibs: Change bit description according to latest PPR

 arch/x86/events/amd/ibs.c             | 156 +++++++++++---------------
 include/linux/perf_event.h            |   3 +-
 include/uapi/linux/perf_event.h       |   3 +-
 tools/include/uapi/linux/perf_event.h |   3 +-
 tools/perf/util/amd-sample-raw.c      |  14 +--
 tools/perf/util/event.h               |   3 +-
 tools/perf/util/hist.c                |   2 +-
 tools/perf/util/mem-events.c          |  90 ++++++++-------
 8 files changed, 132 insertions(+), 142 deletions(-)
Signature · 1 line
2.34.1
#1 lore raw

PATCH v3 1/9 perf/mem: Introduce PERF_MEM_LVLNUM_UNC

Older API PERF_MEM_LVL_UNC can be replaced by PERF_MEM_LVLNUM_UNC.

---
 include/uapi/linux/perf_event.h | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
include/uapi/linux/perf_event.h +2 -1
index 37675437b768..39c6a250dd1b 100644
@@ -1339,7 +1339,8 @@ union perf_mem_data_src {
 #define PERF_MEM_LVLNUM_L2	0x02 /* L2 */
 #define PERF_MEM_LVLNUM_L3	0x03 /* L3 */
 #define PERF_MEM_LVLNUM_L4	0x04 /* L4 */
-/* 5-0x8 available */
+/* 5-0x7 available */
+#define PERF_MEM_LVLNUM_UNC	0x08 /* Uncached */
 #define PERF_MEM_LVLNUM_CXL	0x09 /* CXL */
 #define PERF_MEM_LVLNUM_IO	0x0a /* I/O */
 #define PERF_MEM_LVLNUM_ANY_CACHE 0x0b /* Any cache */
Signature · 1 line
2.34.1
#2 lore raw

PATCH v3 2/9 perf/mem: Add PERF_MEM_LVLNUM_NA to PERF_MEM_NA

In reply to Ravi Bangoria

Add PERF_MEM_LVLNUM_NA wherever PERF_MEM_NA is used to set default values.

---
 include/linux/perf_event.h | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
include/linux/perf_event.h +2 -1
index d5628a7b5eaa..064c5ad7ff11 100644
@@ -1183,7 +1183,8 @@ struct perf_sample_data {
 		    PERF_MEM_S(LVL, NA)   |\
 		    PERF_MEM_S(SNOOP, NA) |\
 		    PERF_MEM_S(LOCK, NA)  |\
-		    PERF_MEM_S(TLB, NA))
+		    PERF_MEM_S(TLB, NA)   |\
+		    PERF_MEM_S(LVLNUM, NA))
 
 static inline void perf_sample_data_init(struct perf_sample_data *data,
 					 u64 addr, u64 period)
Signature · 1 line
2.34.1
#3 lore raw

PATCH v3 3/9 perf/x86/ibs: Set mem_lvl_num, mem_remote and mem_hops for data_src

In reply to Ravi Bangoria

From: Namhyung Kim <[email protected]>

Kernel IBS driver wasn't using new PERF_MEM_* APIs due to some of its limitations. Mainly:

1. mem_lvl_num doesn't allow setting multiple sources whereas old API allows it. Setting multiple data sources is useful because IBS on pre-zen4 uarch doesn't provide fine granular DataSrc details (there is only one such DataSrc(2h) though).
2. perf mem sorting logic (sort__lvl_cmp()) ignores mem_lvl_num. perf c2c (c2c_decode_stats()) does not use mem_lvl_num at all. perf mem prints mem_lvl and mem_lvl_num both if both are set, which is ugly.

1st one can be handled by using ANY_CACHE with HOPS_0. 2nd is purely perf tool specific issue and should be fixed separately.

---
 arch/x86/events/amd/ibs.c | 156 +++++++++++++++++---------------------
 1 file changed, 68 insertions(+), 88 deletions(-)
arch/x86/events/amd/ibs.c +68 -88
index 64582954b5f6..b6b833f5da50 100644
@@ -703,38 +703,63 @@ static u8 perf_ibs_data_src(union ibs_op_data2 *op_data2)
 	return op_data2->data_src_lo;
 }
 
-static void perf_ibs_get_mem_lvl(union ibs_op_data2 *op_data2,
-				 union ibs_op_data3 *op_data3,
-				 struct perf_sample_data *data)
+#define	L(x)		(PERF_MEM_S(LVL, x) | PERF_MEM_S(LVL, HIT))
+#define	LN(x)		PERF_MEM_S(LVLNUM, x)
+#define	REM		PERF_MEM_S(REMOTE, REMOTE)
+#define	HOPS(x)		PERF_MEM_S(HOPS, x)
+
+static u64 g_data_src[8] = {
+	[IBS_DATA_SRC_LOC_CACHE]	  = L(L3) | L(REM_CCE1) | LN(ANY_CACHE) | HOPS(0),
+	[IBS_DATA_SRC_DRAM]		  = L(LOC_RAM) | LN(RAM),
+	[IBS_DATA_SRC_REM_CACHE]	  = L(REM_CCE2) | LN(ANY_CACHE) | REM | HOPS(1),
+	[IBS_DATA_SRC_IO]		  = L(IO) | LN(IO),
+};
+
+#define RMT_NODE_BITS			(1 << IBS_DATA_SRC_DRAM)
+#define RMT_NODE_APPLICABLE(x)		(RMT_NODE_BITS & (1 << x))
+
+static u64 g_zen4_data_src[32] = {
+	[IBS_DATA_SRC_EXT_LOC_CACHE]	  = L(L3) | LN(L3),
+	[IBS_DATA_SRC_EXT_NEAR_CCX_CACHE] = L(REM_CCE1) | LN(ANY_CACHE) | REM | HOPS(0),
+	[IBS_DATA_SRC_EXT_DRAM]		  = L(LOC_RAM) | LN(RAM),
+	[IBS_DATA_SRC_EXT_FAR_CCX_CACHE]  = L(REM_CCE2) | LN(ANY_CACHE) | REM | HOPS(1),
+	[IBS_DATA_SRC_EXT_PMEM]		  = LN(PMEM),
+	[IBS_DATA_SRC_EXT_IO]		  = L(IO) | LN(IO),
+	[IBS_DATA_SRC_EXT_EXT_MEM]	  = LN(CXL),
+};
+
+#define ZEN4_RMT_NODE_BITS		((1 << IBS_DATA_SRC_EXT_DRAM) | \
+					 (1 << IBS_DATA_SRC_EXT_PMEM) | \
+					 (1 << IBS_DATA_SRC_EXT_EXT_MEM))
+#define ZEN4_RMT_NODE_APPLICABLE(x)	(ZEN4_RMT_NODE_BITS & (1 << x))
+
+static __u64 perf_ibs_get_mem_lvl(union ibs_op_data2 *op_data2,
+				  union ibs_op_data3 *op_data3,
+				  struct perf_sample_data *data)
 {
 	union perf_mem_data_src *data_src = &data->data_src;
 	u8 ibs_data_src = perf_ibs_data_src(op_data2);
 
 	data_src->mem_lvl = 0;
+	data_src->mem_lvl_num = 0;
 
 	/*
 	 * DcMiss, L2Miss, DataSrc, DcMissLat etc. are all invalid for Uncached
 	 * memory accesses. So, check DcUcMemAcc bit early.
 	 */
-	if (op_data3->dc_uc_mem_acc && ibs_data_src != IBS_DATA_SRC_EXT_IO) {
-		data_src->mem_lvl = PERF_MEM_LVL_UNC | PERF_MEM_LVL_HIT;
-		return;
-	}
+	if (op_data3->dc_uc_mem_acc && ibs_data_src != IBS_DATA_SRC_EXT_IO)
+		return L(UNC) | LN(UNC);
 
 	/* L1 Hit */
-	if (op_data3->dc_miss == 0) {
-		data_src->mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_HIT;
-		return;
-	}
+	if (op_data3->dc_miss == 0)
+		return L(L1) | LN(L1);
 
 	/* L2 Hit */
 	if (op_data3->l2_miss == 0) {
 		/* Erratum #1293 */
 		if (boot_cpu_data.x86 != 0x19 || boot_cpu_data.x86_model > 0xF ||
-		    !(op_data3->sw_pf || op_data3->dc_miss_no_mab_alloc)) {
-			data_src->mem_lvl = PERF_MEM_LVL_L2 | PERF_MEM_LVL_HIT;
-			return;
-		}
+		    !(op_data3->sw_pf || op_data3->dc_miss_no_mab_alloc))
+			return L(L2) | LN(L2);
 	}
 
 	/*
@@ -744,82 +769,36 @@ static void perf_ibs_get_mem_lvl(union ibs_op_data2 *op_data2,
 	if (data_src->mem_op != PERF_MEM_OP_LOAD)
 		goto check_mab;
 
-	/* L3 Hit */
 	if (ibs_caps & IBS_CAPS_ZEN4) {
-		if (ibs_data_src == IBS_DATA_SRC_EXT_LOC_CACHE) {
-			data_src->mem_lvl = PERF_MEM_LVL_L3 | PERF_MEM_LVL_HIT;
-			return;
-		}
-	} else {
-		if (ibs_data_src == IBS_DATA_SRC_LOC_CACHE) {
-			data_src->mem_lvl = PERF_MEM_LVL_L3 | PERF_MEM_LVL_REM_CCE1 |
-					    PERF_MEM_LVL_HIT;
-			return;
-		}
-	}
+		u64 val = g_zen4_data_src[ibs_data_src];
 
-	/* A peer cache in a near CCX */
-	if (ibs_caps & IBS_CAPS_ZEN4 &&
-	    ibs_data_src == IBS_DATA_SRC_EXT_NEAR_CCX_CACHE) {
-		data_src->mem_lvl = PERF_MEM_LVL_REM_CCE1 | PERF_MEM_LVL_HIT;
-		return;
-	}
+		if (!val)
+			goto check_mab;
 
-	/* A peer cache in a far CCX */
-	if (ibs_caps & IBS_CAPS_ZEN4) {
-		if (ibs_data_src == IBS_DATA_SRC_EXT_FAR_CCX_CACHE) {
-			data_src->mem_lvl = PERF_MEM_LVL_REM_CCE2 | PERF_MEM_LVL_HIT;
-			return;
-		}
-	} else {
-		if (ibs_data_src == IBS_DATA_SRC_REM_CACHE) {
-			data_src->mem_lvl = PERF_MEM_LVL_REM_CCE2 | PERF_MEM_LVL_HIT;
-			return;
+		/* HOPS_1 because IBS doesn't provide remote socket detail */
+		if (op_data2->rmt_node && ZEN4_RMT_NODE_APPLICABLE(ibs_data_src)) {
+			if (ibs_data_src == IBS_DATA_SRC_EXT_DRAM)
+				val = L(REM_RAM1) | LN(RAM) | REM | HOPS(1);
+			else
+				val |= REM | HOPS(1);
 		}
-	}
 
-	/* DRAM */
-	if (ibs_data_src == IBS_DATA_SRC_EXT_DRAM) {
-		if (op_data2->rmt_node == 0)
-			data_src->mem_lvl = PERF_MEM_LVL_LOC_RAM | PERF_MEM_LVL_HIT;
-		else
-			data_src->mem_lvl = PERF_MEM_LVL_REM_RAM1 | PERF_MEM_LVL_HIT;
-		return;
-	}
+		return val;
+	} else {
+		u64 val = g_data_src[ibs_data_src];
 
-	/* PMEM */
-	if (ibs_caps & IBS_CAPS_ZEN4 && ibs_data_src == IBS_DATA_SRC_EXT_PMEM) {
-		data_src->mem_lvl_num = PERF_MEM_LVLNUM_PMEM;
-		if (op_data2->rmt_node) {
-			data_src->mem_remote = PERF_MEM_REMOTE_REMOTE;
-			/* IBS doesn't provide Remote socket detail */
-			data_src->mem_hops = PERF_MEM_HOPS_1;
-		}
-		return;
-	}
+		if (!val)
+			goto check_mab;
 
-	/* Extension Memory */
-	if (ibs_caps & IBS_CAPS_ZEN4 &&
-	    ibs_data_src == IBS_DATA_SRC_EXT_EXT_MEM) {
-		data_src->mem_lvl_num = PERF_MEM_LVLNUM_CXL;
-		if (op_data2->rmt_node) {
-			data_src->mem_remote = PERF_MEM_REMOTE_REMOTE;
-			/* IBS doesn't provide Remote socket detail */
-			data_src->mem_hops = PERF_MEM_HOPS_1;
+		/* HOPS_1 because IBS doesn't provide remote socket detail */
+		if (op_data2->rmt_node && RMT_NODE_APPLICABLE(ibs_data_src)) {
+			if (ibs_data_src == IBS_DATA_SRC_DRAM)
+				val = L(REM_RAM1) | LN(RAM) | REM | HOPS(1);
+			else
+				val |= REM | HOPS(1);
 		}
-		return;
-	}
 
-	/* IO */
-	if (ibs_data_src == IBS_DATA_SRC_EXT_IO) {
-		data_src->mem_lvl = PERF_MEM_LVL_IO;
-		data_src->mem_lvl_num = PERF_MEM_LVLNUM_IO;
-		if (op_data2->rmt_node) {
-			data_src->mem_remote = PERF_MEM_REMOTE_REMOTE;
-			/* IBS doesn't provide Remote socket detail */
-			data_src->mem_hops = PERF_MEM_HOPS_1;
-		}
-		return;
+		return val;
 	}
 
 check_mab:
@@ -830,12 +809,11 @@ static void perf_ibs_get_mem_lvl(union ibs_op_data2 *op_data2,
 	 * DataSrc simultaneously. Prioritize DataSrc over MAB, i.e. set
 	 * MAB only when IBS fails to provide DataSrc.
 	 */
-	if (op_data3->dc_miss_no_mab_alloc) {
-		data_src->mem_lvl = PERF_MEM_LVL_LFB | PERF_MEM_LVL_HIT;
-		return;
-	}
+	if (op_data3->dc_miss_no_mab_alloc)
+		return L(LFB) | LN(LFB);
 
-	data_src->mem_lvl = PERF_MEM_LVL_NA;
+	/* Don't set HIT with NA */
+	return PERF_MEM_S(LVL, NA) | LN(NA);
 }
 
 static bool perf_ibs_cache_hit_st_valid(void)
@@ -925,7 +903,9 @@ static void perf_ibs_get_data_src(struct perf_ibs_data *ibs_data,
 				  union ibs_op_data2 *op_data2,
 				  union ibs_op_data3 *op_data3)
 {
-	perf_ibs_get_mem_lvl(op_data2, op_data3, data);
+	union perf_mem_data_src *data_src = &data->data_src;
+
+	data_src->val |= perf_ibs_get_mem_lvl(op_data2, op_data3, data);
 	perf_ibs_get_mem_snoop(op_data2, data);
 	perf_ibs_get_tlb_lvl(op_data3, data);
 	perf_ibs_get_mem_lock(op_data3, data);
Signature · 1 line
2.34.1
#4 lore raw

PATCH v3 4/9 perf headers: Sync uapi/linux/perf_event.h

In reply to Ravi Bangoria

... to bring PERF_MEM_LVLNUM_UNC definition to userspace

---
 tools/include/uapi/linux/perf_event.h | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
tools/include/uapi/linux/perf_event.h +2 -1
index 37675437b768..39c6a250dd1b 100644
@@ -1339,7 +1339,8 @@ union perf_mem_data_src {
 #define PERF_MEM_LVLNUM_L2	0x02 /* L2 */
 #define PERF_MEM_LVLNUM_L3	0x03 /* L3 */
 #define PERF_MEM_LVLNUM_L4	0x04 /* L4 */
-/* 5-0x8 available */
+/* 5-0x7 available */
+#define PERF_MEM_LVLNUM_UNC	0x08 /* Uncached */
 #define PERF_MEM_LVLNUM_CXL	0x09 /* CXL */
 #define PERF_MEM_LVLNUM_IO	0x0a /* I/O */
 #define PERF_MEM_LVLNUM_ANY_CACHE 0x0b /* Any cache */
Signature · 1 line
2.34.1
#5 lore raw

PATCH v3 5/9 perf mem: Add PERF_MEM_LVLNUM_NA to PERF_MEM_DATA_SRC_NONE

In reply to Ravi Bangoria

Add PERF_MEM_LVLNUM_NA wherever PERF_MEM_DATA_SRC_NONE is used to set default values.

---
 tools/perf/util/event.h | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
tools/perf/util/event.h +2 -1
index 6663a676eadc..de20e01c9d72 100644
@@ -89,7 +89,8 @@ enum {
 	 PERF_MEM_S(LVL, NA) |\
 	 PERF_MEM_S(SNOOP, NA) |\
 	 PERF_MEM_S(LOCK, NA) |\
-	 PERF_MEM_S(TLB, NA))
+	 PERF_MEM_S(TLB, NA) |\
+	 PERF_MEM_S(LVLNUM, NA))
 
 /* Attribute type for custom synthesized events */
 #define PERF_TYPE_SYNTH		(INT_MAX + 1U)
Signature · 1 line
2.34.1
#6 lore raw

PATCH v3 6/9 perf mem: Add support for printing PERF_MEM_LVLNUM_UNC

In reply to Ravi Bangoria

Add support for printing PERF_MEM_LVLNUM_UNC in perf mem report.

---
 tools/perf/util/mem-events.c | 1 +
 1 file changed, 1 insertion(+)
tools/perf/util/mem-events.c +1 -0
index b3a91093069a..3a7c72be8326 100644
@@ -295,6 +295,7 @@ static const char * const mem_lvl[] = {
 };
 
 static const char * const mem_lvlnum[] = {
+	[PERF_MEM_LVLNUM_UNC] = "Uncached",
 	[PERF_MEM_LVLNUM_CXL] = "CXL",
 	[PERF_MEM_LVLNUM_IO] = "I/O",
 	[PERF_MEM_LVLNUM_ANY_CACHE] = "Any cache",
Signature · 1 line
2.34.1
#7 lore raw

PATCH v3 7/9 perf mem: Refactor perf_mem__lvl_scnprintf()

In reply to Ravi Bangoria

Interpretation of perf_mem_data_src by perf_mem__lvl_scnprintf() is non-intuitive. For ex, it ignores mem_lvl when mem_hops is set but considers it otherwise. It prints both mem_lvl_num and mem_lvl when mem_hops is not set.

Refactor this function such that it behaves more intuitively: Use new API mem_lvl_num|mem_remote|mem_hops if mem_lvl_num contains value other than PERF_MEM_LVLNUM_NA. Otherwise, fallback to old API mem_lvl. Since new API has no way to indicate MISS, use it from old api, otherwise don't club old and new APIs while parsing as well as printing.

Before:

  $ sudo ./perf mem report -F sample,mem --stdio
  #      Samples  Memory access
  # ............  ........................
  #
          250097  N/A
          188907  L1 hit
            4116  L2 hit
            3496  Remote Cache (1 hop) hit
            3271  Remote Cache (2 hops) hit
             873  L3 hit
             598  Local RAM hit
             438  Remote RAM (1 hop) hit
               1  Uncached hit

After:

  $ sudo ./perf mem report -F sample,mem --stdio
  #      Samples  Memory access
  # ............  .......................................
  #
          255517  N/A
          189989  L1 hit
            4541  L2 hit
            3363  Remote core, same node Any cache hit
            3336  Remote node, same socket Any cache hit
            1275  L3 hit
             743  RAM hit
             545  Remote node, same socket RAM hit
               4  Uncached hit
---
 tools/perf/util/mem-events.c | 89 +++++++++++++++++++-----------------
 1 file changed, 47 insertions(+), 42 deletions(-)
tools/perf/util/mem-events.c +47 -42
index 3a7c72be8326..ed1ee4b05356 100644
@@ -344,66 +344,71 @@ static int perf_mem__op_scnprintf(char *out, size_t sz, struct mem_info *mem_inf
 
 int perf_mem__lvl_scnprintf(char *out, size_t sz, struct mem_info *mem_info)
 {
-	size_t i, l = 0;
-	u64 m =  PERF_MEM_LVL_NA;
-	u64 hit, miss;
+	union perf_mem_data_src data_src;
 	int printed = 0;
-
-	if (mem_info)
-		m  = mem_info->data_src.mem_lvl;
+	size_t l = 0;
+	size_t i;
+	int lvl;
+	char hit_miss[5] = {0};
 
 	sz -= 1; /* -1 for null termination */
 	out[0] = '\0';
 
-	hit = m & PERF_MEM_LVL_HIT;
-	miss = m & PERF_MEM_LVL_MISS;
+	if (!mem_info)
+		goto na;
 
-	/* already taken care of */
-	m &= ~(PERF_MEM_LVL_HIT|PERF_MEM_LVL_MISS);
+	data_src = mem_info->data_src;
 
-	if (mem_info && mem_info->data_src.mem_remote) {
-		strcat(out, "Remote ");
-		l += 7;
-	}
+	if (data_src.mem_lvl & PERF_MEM_LVL_HIT)
+		memcpy(hit_miss, "hit", 3);
+	else if (data_src.mem_lvl & PERF_MEM_LVL_MISS)
+		memcpy(hit_miss, "miss", 4);
 
-	/*
-	 * Incase mem_hops field is set, we can skip printing data source via
-	 * PERF_MEM_LVL namespace.
-	 */
-	if (mem_info && mem_info->data_src.mem_hops) {
-		l += scnprintf(out + l, sz - l, "%s ", mem_hops[mem_info->data_src.mem_hops]);
-	} else {
-		for (i = 0; m && i < ARRAY_SIZE(mem_lvl); i++, m >>= 1) {
-			if (!(m & 0x1))
-				continue;
-			if (printed++) {
-				strcat(out, " or ");
-				l += 4;
-			}
-			l += scnprintf(out + l, sz - l, mem_lvl[i]);
+	lvl = data_src.mem_lvl_num;
+	if (lvl && lvl != PERF_MEM_LVLNUM_NA) {
+		if (data_src.mem_remote) {
+			strcat(out, "Remote ");
+			l += 7;
 		}
+
+		if (data_src.mem_hops)
+			l += scnprintf(out + l, sz - l, "%s ", mem_hops[data_src.mem_hops]);
+
+		if (mem_lvlnum[lvl])
+			l += scnprintf(out + l, sz - l, mem_lvlnum[lvl]);
+		else
+			l += scnprintf(out + l, sz - l, "L%d", lvl);
+
+		l += scnprintf(out + l, sz - l, " %s", hit_miss);
+		return l;
 	}
 
-	if (mem_info && mem_info->data_src.mem_lvl_num) {
-		int lvl = mem_info->data_src.mem_lvl_num;
+	lvl = data_src.mem_lvl;
+	if (!lvl)
+		goto na;
+
+	lvl &= ~(PERF_MEM_LVL_NA | PERF_MEM_LVL_HIT | PERF_MEM_LVL_MISS);
+	if (!lvl)
+		goto na;
+
+	for (i = 0; lvl && i < ARRAY_SIZE(mem_lvl); i++, lvl >>= 1) {
+		if (!(lvl & 0x1))
+			continue;
 		if (printed++) {
 			strcat(out, " or ");
 			l += 4;
 		}
-		if (mem_lvlnum[lvl])
-			l += scnprintf(out + l, sz - l, mem_lvlnum[lvl]);
-		else
-			l += scnprintf(out + l, sz - l, "L%d", lvl);
+		l += scnprintf(out + l, sz - l, mem_lvl[i]);
 	}
 
-	if (l == 0)
-		l += scnprintf(out + l, sz - l, "N/A");
-	if (hit)
-		l += scnprintf(out + l, sz - l, " hit");
-	if (miss)
-		l += scnprintf(out + l, sz - l, " miss");
+	if (printed) {
+		l += scnprintf(out + l, sz - l, " %s", hit_miss);
+		return l;
+	}
 
-	return l;
+na:
+	strcat(out, "N/A");
+	return 3;
 }
 
 static const char * const snoop_access[] = {
Signature · 1 line
2.34.1
#8 lore raw

PATCH v3 8/9 perf mem: Increase HISTC_MEM_LVL column size to 39 chars

In reply to Ravi Bangoria

39 is taken from the length of longest printable new API string:
"Remote socket, same board Any cache hit". Although, using old API can result into even longer strings, let's not overkill by making it dynamic length.

---
 tools/perf/util/hist.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
tools/perf/util/hist.c +1 -1
index b6e4b4edde43..46ce4527919a 100644
@@ -207,7 +207,7 @@ void hists__calc_col_len(struct hists *hists, struct hist_entry *h)
 	hists__new_col_len(hists, HISTC_MEM_LOCKED, 6);
 	hists__new_col_len(hists, HISTC_MEM_TLB, 22);
 	hists__new_col_len(hists, HISTC_MEM_SNOOP, 12);
-	hists__new_col_len(hists, HISTC_MEM_LVL, 21 + 3);
+	hists__new_col_len(hists, HISTC_MEM_LVL, 36 + 3);
 	hists__new_col_len(hists, HISTC_LOCAL_WEIGHT, 12);
 	hists__new_col_len(hists, HISTC_GLOBAL_WEIGHT, 12);
 	hists__new_col_len(hists, HISTC_MEM_BLOCKED, 10);
Signature · 1 line
2.34.1
#9 lore raw

PATCH v3 9/9 perf script ibs: Change bit description according to latest PPR

In reply to Ravi Bangoria

Some of the IBS_OP_DATA2 bit descriptions were stale (taken from old version of PPR). Change it according to latest PPR.

---
 tools/perf/util/amd-sample-raw.c | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)
tools/perf/util/amd-sample-raw.c +7 -7
index b0e70ce9d87a..6a6ddba76c75 100644
@@ -105,17 +105,17 @@ static void pr_ibs_op_data2_extended(union ibs_op_data2 reg)
 	static const char * const data_src_str[] = {
 		"",
 		" DataSrc 1=Local L3 or other L1/L2 in CCX",
-		" DataSrc 2=A peer cache in a near CCX",
-		" DataSrc 3=Data returned from DRAM",
+		" DataSrc 2=Another CCX cache in the same NUMA node",
+		" DataSrc 3=DRAM",
 		" DataSrc 4=(reserved)",
-		" DataSrc 5=A peer cache in a far CCX",
-		" DataSrc 6=DRAM address map with \"long latency\" bit set",
-		" DataSrc 7=Data returned from MMIO/Config/PCI/APIC",
-		" DataSrc 8=Extension Memory (S-Link, GenZ, etc)",
+		" DataSrc 5=Another CCX cache in a different NUMA node",
+		" DataSrc 6=Long-latency DIMM",
+		" DataSrc 7=MMIO/Config/PCI/APIC",
+		" DataSrc 8=Extension Memory",
 		" DataSrc 9=(reserved)",
 		" DataSrc 10=(reserved)",
 		" DataSrc 11=(reserved)",
-		" DataSrc 12=Peer Agent Memory",
+		" DataSrc 12=Coherent Memory of a different processor type",
 		/* 13 to 31 are reserved. Avoid printing them. */
 	};
 	int data_src = (reg.data_src_hi << 3) | reg.data_src_lo;
Signature · 1 line
2.34.1
maintainer Acked-by on the v3 series
#10 lore raw
In reply to Ravi Bangoria

Hi Ravi,

Quoted 12 lines · Ravi Bangoria

Kernel IBS driver wasn't using new PERF_MEM_* APIs due to some of its limitations. Mainly:

1. mem_lvl_num doesn't allow setting multiple sources whereas old API allows it. Setting multiple data sources is useful because IBS on pre-zen4 uarch doesn't provide fine granular DataSrc details (there is only one such DataSrc(2h) though).
2. perf mem sorting logic (sort__lvl_cmp()) ignores mem_lvl_num. perf c2c (c2c_decode_stats()) does not use mem_lvl_num at all. perf mem prints mem_lvl and mem_lvl_num both if both are set, which is ugly.

Quoted 8 lines

Set mem_lvl_num, mem_remote and mem_hops for data_src via IBS. Handle first issue using mem_lvl_num = ANY_CACHE | HOPS_0. In addition to setting new API fields, convert all individual field assignments to compile time wrapper macros built using PERF_MEM_S(). Also convert DataSrc conditional code to array lookups.

Interpretation of perf_mem_data_src by perf_mem__lvl_scnprintf() was non-intuitive. Make it sane.

Looks good, but I think you need to split kernel and user patches.

Quoted 13 lines

v2: https://lore.kernel.org/r/20230327130851.1565-1-ravi.bangoria%40amd.com
v2->v3:
- IBS: Don't club RmtNode with DataSrc=7 (IO)
- Make perf_mem__lvl_scnprintf() more sane
- Introduce PERF_MEM_LVLNUM_UNC, set it along with PERF_MEM_LVL_UNC and interpreat it in tool.
- Add PERF_MEM_LVLNUM_NA to default data_src value
- Change some of the IBS bit description according to latest PPR

Namhyung Kim (1):
perf/x86/ibs: Set mem_lvl_num, mem_remote and mem_hops for data_src

Quoted 9 lines

Ravi Bangoria (8):
perf/mem: Introduce PERF_MEM_LVLNUM_UNC
perf/mem: Add PERF_MEM_LVLNUM_NA to PERF_MEM_NA
perf headers: Sync uapi/linux/perf_event.h
perf mem: Add PERF_MEM_LVLNUM_NA to PERF_MEM_DATA_SRC_NONE perf mem: Add support for printing PERF_MEM_LVLNUM_UNC perf mem: Refactor perf_mem__lvl_scnprintf()
perf mem: Increase HISTC_MEM_LVL column size to 39 chars perf script ibs: Change bit description according to latest PPR

Thanks,
Namhyung

Quoted 14 lines
 arch/x86/events/amd/ibs.c             | 156 +++++++++++---------------
 include/linux/perf_event.h            |   3 +-
 include/uapi/linux/perf_event.h       |   3 +-
 tools/include/uapi/linux/perf_event.h |   3 +-
 tools/perf/util/amd-sample-raw.c      |  14 +--
 tools/perf/util/event.h               |   3 +-
 tools/perf/util/hist.c                |   2 +-
 tools/perf/util/mem-events.c          |  90 ++++++++-------
 8 files changed, 132 insertions(+), 142 deletions(-)

--
2.34.1

#11 lore raw
Quoted 21 lines · Namhyung Kim

Hi Ravi,

Quoted 18 lines · Ravi Bangoria

Kernel IBS driver wasn't using new PERF_MEM_* APIs due to some of its limitations. Mainly:

1. mem_lvl_num doesn't allow setting multiple sources whereas old API allows it. Setting multiple data sources is useful because IBS on pre-zen4 uarch doesn't provide fine granular DataSrc details (there is only one such DataSrc(2h) though).
2. perf mem sorting logic (sort__lvl_cmp()) ignores mem_lvl_num. perf c2c (c2c_decode_stats()) does not use mem_lvl_num at all. perf mem prints mem_lvl and mem_lvl_num both if both are set, which is ugly.

Set mem_lvl_num, mem_remote and mem_hops for data_src via IBS. Handle first issue using mem_lvl_num = ANY_CACHE | HOPS_0. In addition to setting new API fields, convert all individual field assignments to compile time wrapper macros built using PERF_MEM_S(). Also convert DataSrc conditional code to array lookups.

Quoted 4 lines
Quoted 2 lines

Interpretation of perf_mem_data_src by perf_mem__lvl_scnprintf() was non-intuitive. Make it sane.

Looks good, but I think you need to split kernel and user patches.

Patch #1 to #3 are kernel changes. Patch #4 to #9 are userspace changes. Arnaldo, Peter, please let me know if you wants to split the series and resend.

Quoted 24 lines
Quoted 14 lines

v2: https://lore.kernel.org/r/20230327130851.1565-1-ravi.bangoria%40amd.com
v2->v3:
- IBS: Don't club RmtNode with DataSrc=7 (IO)
- Make perf_mem__lvl_scnprintf() more sane
- Introduce PERF_MEM_LVLNUM_UNC, set it along with PERF_MEM_LVL_UNC and interpreat it in tool.
- Add PERF_MEM_LVLNUM_NA to default data_src value
- Change some of the IBS bit description according to latest PPR

Namhyung Kim (1):
perf/x86/ibs: Set mem_lvl_num, mem_remote and mem_hops for data_src

Quoted 9 lines

Ravi Bangoria (8):
perf/mem: Introduce PERF_MEM_LVLNUM_UNC
perf/mem: Add PERF_MEM_LVLNUM_NA to PERF_MEM_NA
perf headers: Sync uapi/linux/perf_event.h
perf mem: Add PERF_MEM_LVLNUM_NA to PERF_MEM_DATA_SRC_NONE perf mem: Add support for printing PERF_MEM_LVLNUM_UNC perf mem: Refactor perf_mem__lvl_scnprintf()
perf mem: Increase HISTC_MEM_LVL column size to 39 chars perf script ibs: Change bit description according to latest PPR

Quoted 1 line

Thanks!

#12 lore raw

Em Mon, Apr 10, 2023 at 07:53:57AM +0530, Ravi Bangoria escreveu:

Quoted 22 lines · Namhyung Kim
Quoted 21 lines

Hi Ravi,

Quoted 18 lines · Ravi Bangoria

Kernel IBS driver wasn't using new PERF_MEM_* APIs due to some of its limitations. Mainly:

1. mem_lvl_num doesn't allow setting multiple sources whereas old API allows it. Setting multiple data sources is useful because IBS on pre-zen4 uarch doesn't provide fine granular DataSrc details (there is only one such DataSrc(2h) though).
2. perf mem sorting logic (sort__lvl_cmp()) ignores mem_lvl_num. perf c2c (c2c_decode_stats()) does not use mem_lvl_num at all. perf mem prints mem_lvl and mem_lvl_num both if both are set, which is ugly.

Set mem_lvl_num, mem_remote and mem_hops for data_src via IBS. Handle first issue using mem_lvl_num = ANY_CACHE | HOPS_0. In addition to setting new API fields, convert all individual field assignments to compile time wrapper macros built using PERF_MEM_S(). Also convert DataSrc conditional code to array lookups.

Quoted 8 lines
Quoted 4 lines
Quoted 2 lines

Interpretation of perf_mem_data_src by perf_mem__lvl_scnprintf() was non-intuitive. Make it sane.

Looks good, but I think you need to split kernel and user patches.

Patch #1 to #3 are kernel changes. Patch #4 to #9 are userspace changes. Arnaldo, Peter, please let me know if you wants to split the series and resend.

I can always use b4's -P option :-) So no need to resubmit, I can pick the tools bits,

- Arnaldo

Quoted 27 lines
Quoted 24 lines
Quoted 14 lines

v2: https://lore.kernel.org/r/20230327130851.1565-1-ravi.bangoria%40amd.com
v2->v3:
- IBS: Don't club RmtNode with DataSrc=7 (IO)
- Make perf_mem__lvl_scnprintf() more sane
- Introduce PERF_MEM_LVLNUM_UNC, set it along with PERF_MEM_LVL_UNC and interpreat it in tool.
- Add PERF_MEM_LVLNUM_NA to default data_src value
- Change some of the IBS bit description according to latest PPR

Namhyung Kim (1):
perf/x86/ibs: Set mem_lvl_num, mem_remote and mem_hops for data_src

Quoted 9 lines

Ravi Bangoria (8):
perf/mem: Introduce PERF_MEM_LVLNUM_UNC
perf/mem: Add PERF_MEM_LVLNUM_NA to PERF_MEM_NA
perf headers: Sync uapi/linux/perf_event.h
perf mem: Add PERF_MEM_LVLNUM_NA to PERF_MEM_DATA_SRC_NONE perf mem: Add support for printing PERF_MEM_LVLNUM_UNC perf mem: Refactor perf_mem__lvl_scnprintf()
perf mem: Increase HISTC_MEM_LVL column size to 39 chars perf script ibs: Change bit description according to latest PPR

Quoted 1 line

Thanks!

Signature · 1 line
- Arnaldo
#13 lore raw

Em Mon, Apr 10, 2023 at 12:15:48PM -0300, Arnaldo Carvalho de Melo escreveu:

Quoted 26 lines

Em Mon, Apr 10, 2023 at 07:53:57AM +0530, Ravi Bangoria escreveu:

Quoted 25 lines · Namhyung Kim
Quoted 24 lines

Hi Ravi,

Quoted 20 lines · Ravi Bangoria

Kernel IBS driver wasn't using new PERF_MEM_* APIs due to some of its limitations. Mainly:

1. mem_lvl_num doesn't allow setting multiple sources whereas old API allows it. Setting multiple data sources is useful because IBS on pre-zen4 uarch doesn't provide fine granular DataSrc details (there is only one such DataSrc(2h) though).
2. perf mem sorting logic (sort__lvl_cmp()) ignores mem_lvl_num. perf c2c (c2c_decode_stats()) does not use mem_lvl_num at all. perf mem prints mem_lvl and mem_lvl_num both if both are set, which is ugly.

Set mem_lvl_num, mem_remote and mem_hops for data_src via IBS. Handle first issue using mem_lvl_num = ANY_CACHE | HOPS_0. In addition to setting new API fields, convert all individual field assignments to compile time wrapper macros built using PERF_MEM_S(). Also convert DataSrc conditional code to array lookups.

Interpretation of perf_mem_data_src by perf_mem__lvl_scnprintf() was non-intuitive. Make it sane.

Quoted 8 lines
Quoted 5 lines
Quoted 1 line

Looks good, but I think you need to split kernel and user patches.

Patch #1 to #3 are kernel changes. Patch #4 to #9 are userspace changes. Arnaldo, Peter, please let me know if you wants to split the series and resend.

I can always use b4's -P option :-) So no need to resubmit, I can pick

Quoted 1 line

the tools bits,

Done

Quoted 33 lines

- Arnaldo

Quoted 27 lines
Quoted 24 lines
Quoted 14 lines

v2: https://lore.kernel.org/r/20230327130851.1565-1-ravi.bangoria%40amd.com
v2->v3:
- IBS: Don't club RmtNode with DataSrc=7 (IO)
- Make perf_mem__lvl_scnprintf() more sane
- Introduce PERF_MEM_LVLNUM_UNC, set it along with PERF_MEM_LVL_UNC and interpreat it in tool.
- Add PERF_MEM_LVLNUM_NA to default data_src value
- Change some of the IBS bit description according to latest PPR

Namhyung Kim (1):
perf/x86/ibs: Set mem_lvl_num, mem_remote and mem_hops for data_src

Quoted 9 lines

Ravi Bangoria (8):
perf/mem: Introduce PERF_MEM_LVLNUM_UNC
perf/mem: Add PERF_MEM_LVLNUM_NA to PERF_MEM_NA
perf headers: Sync uapi/linux/perf_event.h
perf mem: Add PERF_MEM_LVLNUM_NA to PERF_MEM_DATA_SRC_NONE perf mem: Add support for printing PERF_MEM_LVLNUM_UNC perf mem: Refactor perf_mem__lvl_scnprintf()
perf mem: Increase HISTC_MEM_LVL column size to 39 chars perf script ibs: Change bit description according to latest PPR

Quoted 1 line

Thanks!

Signature · 1 line
- Arnaldo
Signature · 1 line
- Arnaldo
#14 lore raw
In reply to Ravi Bangoria
Quoted 22 lines · Ravi Bangoria
Quoted 21 lines · Namhyung Kim

Hi Ravi,

Quoted 18 lines · Ravi Bangoria

Kernel IBS driver wasn't using new PERF_MEM_* APIs due to some of its limitations. Mainly:

1. mem_lvl_num doesn't allow setting multiple sources whereas old API allows it. Setting multiple data sources is useful because IBS on pre-zen4 uarch doesn't provide fine granular DataSrc details (there is only one such DataSrc(2h) though).
2. perf mem sorting logic (sort__lvl_cmp()) ignores mem_lvl_num. perf c2c (c2c_decode_stats()) does not use mem_lvl_num at all. perf mem prints mem_lvl and mem_lvl_num both if both are set, which is ugly.

Set mem_lvl_num, mem_remote and mem_hops for data_src via IBS. Handle first issue using mem_lvl_num = ANY_CACHE | HOPS_0. In addition to setting new API fields, convert all individual field assignments to compile time wrapper macros built using PERF_MEM_S(). Also convert DataSrc conditional code to array lookups.

Quoted 8 lines
Quoted 4 lines
Quoted 2 lines

Interpretation of perf_mem_data_src by perf_mem__lvl_scnprintf() was non-intuitive. Make it sane.

Looks good, but I think you need to split kernel and user patches.

Patch #1 to #3 are kernel changes. Patch #4 to #9 are userspace changes. Arnaldo, Peter, please let me know if you wants to split the series and resend.

Hi Peter, tools/ patches are already upstream. Can you please pick up kernel changes.

Thanks,
Ravi

#15 lore raw
Quoted 26 lines · Ravi Bangoria
Quoted 25 lines · Ravi Bangoria
Quoted 24 lines · Namhyung Kim

Hi Ravi,

Quoted 20 lines · Ravi Bangoria

Kernel IBS driver wasn't using new PERF_MEM_* APIs due to some of its limitations. Mainly:

1. mem_lvl_num doesn't allow setting multiple sources whereas old API allows it. Setting multiple data sources is useful because IBS on pre-zen4 uarch doesn't provide fine granular DataSrc details (there is only one such DataSrc(2h) though).
2. perf mem sorting logic (sort__lvl_cmp()) ignores mem_lvl_num. perf c2c (c2c_decode_stats()) does not use mem_lvl_num at all. perf mem prints mem_lvl and mem_lvl_num both if both are set, which is ugly.

Set mem_lvl_num, mem_remote and mem_hops for data_src via IBS. Handle first issue using mem_lvl_num = ANY_CACHE | HOPS_0. In addition to setting new API fields, convert all individual field assignments to compile time wrapper macros built using PERF_MEM_S(). Also convert DataSrc conditional code to array lookups.

Interpretation of perf_mem_data_src by perf_mem__lvl_scnprintf() was non-intuitive. Make it sane.

Quoted 8 lines
Quoted 5 lines
Quoted 1 line

Looks good, but I think you need to split kernel and user patches.

Patch #1 to #3 are kernel changes. Patch #4 to #9 are userspace changes. Arnaldo, Peter, please let me know if you wants to split the series and resend.

Hi Peter, tools/ patches are already upstream. Can you please pick up kernel changes.

Gentle ping, Peter!

Thanks,
Ravi

#16 lore raw
Quoted 29 lines · Ravi Bangoria
Quoted 28 lines · Ravi Bangoria
Quoted 27 lines · Ravi Bangoria
Quoted 25 lines · Namhyung Kim

Hi Ravi,

Quoted 12 lines · Ravi Bangoria

Kernel IBS driver wasn't using new PERF_MEM_* APIs due to some of its limitations. Mainly:

1. mem_lvl_num doesn't allow setting multiple sources whereas old API allows it. Setting multiple data sources is useful because IBS on pre-zen4 uarch doesn't provide fine granular DataSrc details (there is only one such DataSrc(2h) though).
2. perf mem sorting logic (sort__lvl_cmp()) ignores mem_lvl_num. perf c2c (c2c_decode_stats()) does not use mem_lvl_num at all. perf mem prints mem_lvl and mem_lvl_num both if both are set, which is ugly.

Quoted 8 lines

Set mem_lvl_num, mem_remote and mem_hops for data_src via IBS. Handle first issue using mem_lvl_num = ANY_CACHE | HOPS_0. In addition to setting new API fields, convert all individual field assignments to compile time wrapper macros built using PERF_MEM_S(). Also convert DataSrc conditional code to array lookups.

Interpretation of perf_mem_data_src by perf_mem__lvl_scnprintf() was non-intuitive. Make it sane.

Looks good, but I think you need to split kernel and user patches.

Quoted 8 lines
Quoted 6 lines
Quoted 3 lines

Patch #1 to #3 are kernel changes. Patch #4 to #9 are userspace changes. Arnaldo, Peter, please let me know if you wants to split the series and resend.

Hi Peter, tools/ patches are already upstream. Can you please pick up kernel changes.

Gentle ping, Peter!

Hello Peter, this is pending from long time. Can you please consider taking it. Please let me know if you want me to resend the patches.

Thanks,
Ravi