Skip to contentSkip to conversation

Basic perf PMU support for Haswell v11

PATCH In review Andi Kleen 4 participants 20 messages Apr 20, 2013 – Sep 4, 2013

Subsystems (2)
lore
Share & tools
lore

Andi Kleen

author
#0 lore raw

This is based on v7 of the full Haswell PMU support, rebased, and stripped down to the bare bones

Most interesting new features are not in this patchkit
(full version is git://git.kernel.org/pub/scm/linux/kernel/git/ak/linux-misc.git hsw/pmu5)

Contains support for:
- Basic Haswell PMU and PEBS support
- Late unmasking of the PMI
- Basic LBRv4 support

v2: Addressed Stephane's feedback. See individual patches for details.
v3: now even more bite-sized. Qualifier constraints merged earlier.
v4: Rename some variables, add some comments and other minor changes. Add some Reviewed/Tested-bys.
v5: Address some minor review feedback. Port to latest perf/core
v6: Add just some variable names, add comments, edit descriptions, some more testing, rebased to latest perf/core
v7: Expand comment
v8: Rename structure field.
v9: No wide counters, but add basic LBRs. Add some more constraints. Rebase to 3.9rc1
v10: Change some whitespace. Rebase to 3.9rc3
v11: Rebase to perf/core. Fix extra regs. Rename INTX.

-Andi

author
#1 lore raw

PATCH 1/5 perf, x86: Add Haswell PEBS record support v5

From: Andi Kleen <[email protected]>

Add support for the Haswell extended (fmt2) PEBS format.

It has a superset of the nhm (fmt1) PEBS fields, but has a longer record so we need to adjust the code paths.

The main advantage is the new "EventingRip" support which directly gives the instruction, not off-by-one instruction. So with precise == 2 we use that directly and don't try to use LBRs and walking basic blocks. This lowers the overhead of using precise significantly.

Some other features are added in later patches.

v2: Rename various identifiers. Add more comments. Get rid of a cast.
v3: fmt2->hsw rename
v4: ip_of_the_event->real_ip rename
v5: use pr_cont. white space changes.

---
 arch/x86/kernel/cpu/perf_event.c          |    3 +-
 arch/x86/kernel/cpu/perf_event_intel_ds.c |  111 +++++++++++++++++++++++------
 2 files changed, 92 insertions(+), 22 deletions(-)
arch/x86/kernel/cpu/perf_event.c +2 -1
index 5ed7a4c..21cfc52 100644
@@ -397,7 +397,8 @@ int x86_pmu_hw_config(struct perf_event *event)
 		 * check that PEBS LBR correction does not conflict with
 		 * whatever the user is asking with attr->branch_sample_type
 		 */
-		if (event->attr.precise_ip > 1) {
+		if (event->attr.precise_ip > 1 &&
+		    x86_pmu.intel_cap.pebs_format < 2) {
 			u64 *br_type = &event->attr.branch_sample_type;
 
 			if (has_branch_stack(event)) {
arch/x86/kernel/cpu/perf_event_intel_ds.c +90 -21

156 lines not on this page · raw message

Signature · 1 line
1.7.7.6
author
#2 lore raw

PATCH 2/5 perf, x86: Basic Haswell PMU support v8

In reply to Andi Kleen

From: Andi Kleen <[email protected]>

Add basic Haswell PMU support.

Similar to SandyBridge, but has a few new events and two new counter bits.

There are some new counter flags that need to be prevented from being set on fixed counters, and allowed to be set for generic counters.

Also we add support for the counter 2 constraint to handle all raw events.

Contains fixes from Stephane Eranian

v2: Folded TSX bits into standard FIXED_EVENT_CONSTRAINTS
v3: Use SNB LBR init code. Comment fix (Stephane Eranian)
v4: Add the counter2 constraints. Fix comment in the right place.
v5: Expand comment
v6: Add CYCLE_ACTIVITY.* to counter constraints
v7: Follow Linux style, not perf style
v8: Add missing extra regs

---
 arch/x86/include/asm/perf_event.h      |    3 +
 arch/x86/kernel/cpu/perf_event.h       |    5 ++-
 arch/x86/kernel/cpu/perf_event_intel.c |   81 ++++++++++++++++++++++++++++++++
 3 files changed, 88 insertions(+), 1 deletions(-)
arch/x86/include/asm/perf_event.h +3 -0
index 57cb634..b79b6eb 100644
@@ -29,6 +29,9 @@
 #define ARCH_PERFMON_EVENTSEL_INV			(1ULL << 23)
 #define ARCH_PERFMON_EVENTSEL_CMASK			0xFF000000ULL
 
+#define HSW_INTX					(1ULL << 32)
+#define HSW_INTX_CHECKPOINTED				(1ULL << 33)
+
 #define AMD64_EVENTSEL_INT_CORE_ENABLE			(1ULL << 36)
 #define AMD64_EVENTSEL_GUESTONLY			(1ULL << 40)
 #define AMD64_EVENTSEL_HOSTONLY				(1ULL << 41)
arch/x86/kernel/cpu/perf_event.h +4 -1
index ba9aadf..a974fe4 100644
@@ -227,11 +227,14 @@ struct cpu_hw_events {
  *  - inv
  *  - edge
  *  - cnt-mask
+ *  - intx
+ *  - intx_checkpointed
  *  The other filters are supported by fixed counters.
  *  The any-thread option is supported starting with v3.
  */
+#define FIXED_EVENT_FLAGS (X86_RAW_EVENT_MASK|HSW_INTX|HSW_INTX_CHECKPOINTED)
 #define FIXED_EVENT_CONSTRAINT(c, n)	\
-	EVENT_CONSTRAINT(c, (1ULL << (32+n)), X86_RAW_EVENT_MASK)
+	EVENT_CONSTRAINT(c, (1ULL << (32+n)), FIXED_EVENT_FLAGS)
 
 /*
  * Constraint on the Event code + UMask
arch/x86/kernel/cpu/perf_event_intel.c +81 -0

105 lines not on this page · raw message

Signature · 1 line
1.7.7.6
author
#3 lore raw

PATCH 3/5 perf, x86: Basic Haswell PEBS support v4

In reply to Andi Kleen

From: Andi Kleen <[email protected]>

Add basic PEBS support for Haswell.
The constraints are similar to SandyBridge with a few new events.

v2: Readd missing pebs_aliases
v3: Readd missing hunk. Fix some constraints.
v4: Fix typo in PEBS event table (Stephane Eranian)

---
 arch/x86/kernel/cpu/perf_event.h          |    2 +
 arch/x86/kernel/cpu/perf_event_intel.c    |    6 +++-
 arch/x86/kernel/cpu/perf_event_intel_ds.c |   34 +++++++++++++++++++++++++++++
 3 files changed, 40 insertions(+), 2 deletions(-)
arch/x86/kernel/cpu/perf_event.h +2 -0
index a974fe4..d75d0ff 100644
@@ -636,6 +636,8 @@ extern struct event_constraint intel_snb_pebs_event_constraints[];
 
 extern struct event_constraint intel_ivb_pebs_event_constraints[];
 
+extern struct event_constraint intel_hsw_pebs_event_constraints[];
+
 struct event_constraint *intel_pebs_constraints(struct perf_event *event);
 
 void intel_pmu_pebs_enable(struct perf_event *event);
arch/x86/kernel/cpu/perf_event_intel.c +4 -2
index bee0308..62b6872 100644
@@ -877,7 +877,8 @@ static inline bool intel_pmu_needs_lbr_smpl(struct perf_event *event)
 		return true;
 
 	/* implicit branch sampling to correct PEBS skid */
-	if (x86_pmu.intel_cap.pebs_trap && event->attr.precise_ip > 1)
+	if (x86_pmu.intel_cap.pebs_trap && event->attr.precise_ip > 1 &&
+	    x86_pmu.intel_cap.pebs_format < 2)
 		return true;
 
 	return false;
@@ -2241,8 +2242,9 @@ __init int intel_pmu_init(void)
 		intel_pmu_lbr_init_snb();
 
 		x86_pmu.event_constraints = intel_hsw_event_constraints;
-
+		x86_pmu.pebs_constraints = intel_hsw_pebs_event_constraints;
 		x86_pmu.extra_regs = intel_snb_extra_regs;
+		x86_pmu.pebs_aliases = intel_pebs_aliases_snb;
 		/* all extra regs are per-cpu when HT is on */
 		x86_pmu.er_flags |= ERF_HAS_RSP_1;
 		x86_pmu.er_flags |= ERF_NO_HT_SHARING;
arch/x86/kernel/cpu/perf_event_intel_ds.c +34 -0
index e91d7fa..e0a66f80 100644
@@ -563,6 +563,40 @@ struct event_constraint intel_ivb_pebs_event_constraints[] = {
         EVENT_CONSTRAINT_END
 };
 
+struct event_constraint intel_hsw_pebs_event_constraints[] = {
+	INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
+	INTEL_UEVENT_CONSTRAINT(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
+	INTEL_UEVENT_CONSTRAINT(0x02c2, 0xf), /* UOPS_RETIRED.RETIRE_SLOTS */
+	INTEL_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
+	INTEL_UEVENT_CONSTRAINT(0x01c5, 0xf), /* BR_MISP_RETIRED.CONDITIONAL */
+	INTEL_UEVENT_CONSTRAINT(0x04c5, 0xf), /* BR_MISP_RETIRED.ALL_BRANCHES */
+	INTEL_UEVENT_CONSTRAINT(0x20c5, 0xf), /* BR_MISP_RETIRED.NEAR_TAKEN */
+	INTEL_EVENT_CONSTRAINT(0xcd, 0x8),    /* MEM_TRANS_RETIRED.* */
+ 	/* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
+	INTEL_UEVENT_CONSTRAINT(0x11d0, 0xf),
+       	/* MEM_UOPS_RETIRED.STLB_MISS_STORES */
+	INTEL_UEVENT_CONSTRAINT(0x12d0, 0xf),
+	INTEL_UEVENT_CONSTRAINT(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
+	INTEL_UEVENT_CONSTRAINT(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
+	INTEL_UEVENT_CONSTRAINT(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
+	INTEL_UEVENT_CONSTRAINT(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
+	INTEL_UEVENT_CONSTRAINT(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
+	INTEL_UEVENT_CONSTRAINT(0x01d1, 0xf), /* MEM_LOAD_UOPS_RETIRED.L1_HIT */
+	INTEL_UEVENT_CONSTRAINT(0x02d1, 0xf), /* MEM_LOAD_UOPS_RETIRED.L2_HIT */
+	INTEL_UEVENT_CONSTRAINT(0x04d1, 0xf), /* MEM_LOAD_UOPS_RETIRED.L3_HIT */
+	INTEL_UEVENT_CONSTRAINT(0x40d1, 0xf), /* MEM_LOAD_UOPS_RETIRED.HIT_LFB */
+	/* MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_MISS */
+	INTEL_UEVENT_CONSTRAINT(0x01d2, 0xf),
+	/* MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HIT */
+	INTEL_UEVENT_CONSTRAINT(0x02d2, 0xf),
+       	/* MEM_LOAD_UOPS_LLC_MISS_RETIRED.LOCAL_DRAM */
+	INTEL_UEVENT_CONSTRAINT(0x01d3, 0xf),
+	INTEL_UEVENT_CONSTRAINT(0x04c8, 0xf), /* HLE_RETIRED.Abort */
+	INTEL_UEVENT_CONSTRAINT(0x04c9, 0xf), /* RTM_RETIRED.Abort */
+
+	EVENT_CONSTRAINT_END
+};
+
 struct event_constraint *intel_pebs_constraints(struct perf_event *event)
 {
 	struct event_constraint *c;
Signature · 1 line
1.7.7.6
author
#4 lore raw

PATCH 4/5 perf, x86: Move NMI clearing to end of PMI handler after the counter registers are reset

In reply to Andi Kleen

From: Andi Kleen <[email protected]>

This avoids some problems with spurious PMIs on Haswell. Haswell seems to behave more like P4 in this regard. Do the same thing as the P4 perf handler by unmasking the NMI only at the end. Shouldn't make any difference for earlier family 6 cores.

Tested on Haswell, IvyBridge, Westmere, Saltwell (Atom)

---
 arch/x86/kernel/cpu/perf_event_intel.c |   16 ++++++----------
 1 files changed, 6 insertions(+), 10 deletions(-)
arch/x86/kernel/cpu/perf_event_intel.c +6 -10
index 62b6872..4a78745 100644
@@ -1172,16 +1172,6 @@ static int intel_pmu_handle_irq(struct pt_regs *regs)
 
 	cpuc = &__get_cpu_var(cpu_hw_events);
 
-	/*
-	 * Some chipsets need to unmask the LVTPC in a particular spot
-	 * inside the nmi handler.  As a result, the unmasking was pushed
-	 * into all the nmi handlers.
-	 *
-	 * This handler doesn't seem to have any issues with the unmasking
-	 * so it was left at the top.
-	 */
-	apic_write(APIC_LVTPC, APIC_DM_NMI);
-
 	intel_pmu_disable_all();
 	handled = intel_pmu_drain_bts_buffer();
 	status = intel_pmu_get_status();
@@ -1241,6 +1231,12 @@ again:
 
 done:
 	intel_pmu_enable_all(0);
+	/*
+	 * Only unmask the NMI after the overflow counters
+	 * have been reset. This avoids spurious NMIs on
+	 * Haswell CPUs.
+	 */
+	apic_write(APIC_LVTPC, APIC_DM_NMI);
 	return handled;
 }
 
Signature · 1 line
1.7.7.6
author
#5 lore raw

PATCH 5/5 perf, x86: Support Haswell v4 LBR format v2

In reply to Andi Kleen

From: Andi Kleen <[email protected]>

Haswell has two additional LBR from flags for TSX: intx and abort, implemented as a new v4 version of the LBR format.

Handle those in and adjust the sign extension code to still correctly extend. The flags are exported similarly in the LBR record to the existing misprediction flag

v2: Add some _

---
 arch/x86/kernel/cpu/perf_event_intel_lbr.c |   56 +++++++++++++++++++++++++--
 include/linux/perf_event.h                 |    7 +++-
 include/uapi/linux/perf_event.h            |    5 ++-
 3 files changed, 61 insertions(+), 7 deletions(-)
arch/x86/kernel/cpu/perf_event_intel_lbr.c +51 -5
index da02e9c..6f9b794 100644
@@ -12,6 +12,16 @@ enum {
 	LBR_FORMAT_LIP		= 0x01,
 	LBR_FORMAT_EIP		= 0x02,
 	LBR_FORMAT_EIP_FLAGS	= 0x03,
+	LBR_FORMAT_EIP_FLAGS2	= 0x04,
+	LBR_FORMAT_MAX_KNOWN    = LBR_FORMAT_EIP_FLAGS2,
+};
+
+static enum {
+	LBR_EIP_FLAGS		= 1,
+	LBR_TSX			= 2,
+} lbr_desc[LBR_FORMAT_MAX_KNOWN + 1] = {
+	[LBR_FORMAT_EIP_FLAGS]  = LBR_EIP_FLAGS,
+	[LBR_FORMAT_EIP_FLAGS2] = LBR_EIP_FLAGS | LBR_TSX,
 };
 
 /*
@@ -56,6 +66,8 @@ enum {
 	 LBR_FAR)
 
 #define LBR_FROM_FLAG_MISPRED  (1ULL << 63)
+#define LBR_FROM_FLAG_IN_TX    (1ULL << 62)
+#define LBR_FROM_FLAG_ABORT    (1ULL << 61)
 
 #define for_each_branch_sample_type(x) \
 	for ((x) = PERF_SAMPLE_BRANCH_USER; \
@@ -81,9 +93,13 @@ enum {
 	X86_BR_JMP      = 1 << 9, /* jump */
 	X86_BR_IRQ      = 1 << 10,/* hw interrupt or trap or fault */
 	X86_BR_IND_CALL = 1 << 11,/* indirect calls */
+	X86_BR_ABORT    = 1 << 12,/* transaction abort */
+	X86_BR_IN_TX    = 1 << 13,/* in transaction */
+	X86_BR_NO_TX    = 1 << 14,/* not in transaction */
 };
 
 #define X86_BR_PLM (X86_BR_USER | X86_BR_KERNEL)
+#define X86_BR_ANYTX (X86_BR_NO_TX | X86_BR_IN_TX)
 
 #define X86_BR_ANY       \
 	(X86_BR_CALL    |\
@@ -95,6 +111,7 @@ enum {
 	 X86_BR_JCC     |\
 	 X86_BR_JMP	 |\
 	 X86_BR_IRQ	 |\
+	 X86_BR_ABORT	 |\
 	 X86_BR_IND_CALL)
 
 #define X86_BR_ALL (X86_BR_PLM | X86_BR_ANY)
@@ -270,21 +287,31 @@ static void intel_pmu_lbr_read_64(struct cpu_hw_events *cpuc)
 
 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
 		unsigned long lbr_idx = (tos - i) & mask;
-		u64 from, to, mis = 0, pred = 0;
+		u64 from, to, mis = 0, pred = 0, in_tx = 0, abort = 0;
+		int skip = 0;
+		int lbr_flags = lbr_desc[lbr_format];
 
 		rdmsrl(x86_pmu.lbr_from + lbr_idx, from);
 		rdmsrl(x86_pmu.lbr_to   + lbr_idx, to);
 
-		if (lbr_format == LBR_FORMAT_EIP_FLAGS) {
+		if (lbr_flags & LBR_EIP_FLAGS) {
 			mis = !!(from & LBR_FROM_FLAG_MISPRED);
 			pred = !mis;
-			from = (u64)((((s64)from) << 1) >> 1);
+			skip = 1;
+		}
+		if (lbr_flags & LBR_TSX) {
+			in_tx = !!(from & LBR_FROM_FLAG_IN_TX);
+			abort = !!(from & LBR_FROM_FLAG_ABORT);
+			skip = 3;
 		}
+		from = (u64)((((s64)from) << skip) >> skip);
 
 		cpuc->lbr_entries[i].from	= from;
 		cpuc->lbr_entries[i].to		= to;
 		cpuc->lbr_entries[i].mispred	= mis;
 		cpuc->lbr_entries[i].predicted	= pred;
+		cpuc->lbr_entries[i].in_tx	= in_tx;
+		cpuc->lbr_entries[i].abort	= abort;
 		cpuc->lbr_entries[i].reserved	= 0;
 	}
 	cpuc->lbr_stack.nr = i;
@@ -334,6 +361,16 @@ static void intel_pmu_setup_sw_lbr_filter(struct perf_event *event)
 
 	if (br_type & PERF_SAMPLE_BRANCH_IND_CALL)
 		mask |= X86_BR_IND_CALL;
+
+	if (br_type & PERF_SAMPLE_BRANCH_ABORT_TX)
+		mask |= X86_BR_ABORT;
+
+	if (br_type & PERF_SAMPLE_BRANCH_IN_TX)
+		mask |= X86_BR_IN_TX;
+
+	if (br_type & PERF_SAMPLE_BRANCH_NO_TX)
+		mask |= X86_BR_NO_TX;
+
 	/*
 	 * stash actual user request into reg, it may
 	 * be used by fixup code for some CPU
@@ -408,7 +445,7 @@ int intel_pmu_setup_lbr_filter(struct perf_event *event)
  * decoded (e.g., text page not present), then X86_BR_NONE is
  * returned.
  */
-static int branch_type(unsigned long from, unsigned long to)
+static int branch_type(unsigned long from, unsigned long to, int abort)
 {
 	struct insn insn;
 	void *addr;
@@ -428,6 +465,9 @@ static int branch_type(unsigned long from, unsigned long to)
 	if (from == 0 || to == 0)
 		return X86_BR_NONE;
 
+	if (abort)
+		return X86_BR_ABORT | to_plm;
+
 	if (from_plm == X86_BR_USER) {
 		/*
 		 * can happen if measuring at the user level only
@@ -564,7 +604,13 @@ intel_pmu_lbr_filter(struct cpu_hw_events *cpuc)
 		from = cpuc->lbr_entries[i].from;
 		to = cpuc->lbr_entries[i].to;
 
-		type = branch_type(from, to);
+		type = branch_type(from, to, cpuc->lbr_entries[i].abort);
+		if (type != X86_BR_NONE && (br_sel & X86_BR_ANYTX)) {
+			if (cpuc->lbr_entries[i].in_tx)
+				type |= X86_BR_IN_TX;
+			else
+				type |= X86_BR_NO_TX;
+		}
 
 		/* if type does not correspond, then discard */
 		if (type == X86_BR_NONE || (br_sel & type) != type) {
include/linux/perf_event.h +6 -1
index e0373d2..466e378 100644
@@ -73,13 +73,18 @@ struct perf_raw_record {
  *
  * support for mispred, predicted is optional. In case it
  * is not supported mispred = predicted = 0.
+ *
+ *     in_tx: running in a hardware transaction
+ *     abort: aborting a hardware transaction
  */
 struct perf_branch_entry {
 	__u64	from;
 	__u64	to;
 	__u64	mispred:1,  /* target mispredicted */
 		predicted:1,/* target predicted */
-		reserved:62;
+		in_tx:1,    /* in transaction */
+		abort:1,    /* transaction abort */
+		reserved:60;
 };
 
 /*
include/uapi/linux/perf_event.h +4 -1
index fb104e5..0b1df41 100644
@@ -157,8 +157,11 @@ enum perf_branch_sample_type {
 	PERF_SAMPLE_BRANCH_ANY_CALL	= 1U << 4, /* any call branch */
 	PERF_SAMPLE_BRANCH_ANY_RETURN	= 1U << 5, /* any return branch */
 	PERF_SAMPLE_BRANCH_IND_CALL	= 1U << 6, /* indirect calls */
+	PERF_SAMPLE_BRANCH_ABORT_TX	= 1U << 7, /* transaction aborts */
+	PERF_SAMPLE_BRANCH_IN_TX	= 1U << 8, /* in transaction */
+	PERF_SAMPLE_BRANCH_NO_TX	= 1U << 9, /* not in transaction */
 
-	PERF_SAMPLE_BRANCH_MAX		= 1U << 7, /* non-ABI */
+	PERF_SAMPLE_BRANCH_MAX		= 1U << 10, /* non-ABI */
 };
 
 #define PERF_SAMPLE_BRANCH_PLM_ALL \
Signature · 1 line
1.7.7.6
#6 lore raw

Re: [PATCH 5/5] perf, x86: Support Haswell v4 LBR format v2

Quoted 8 lines · Andi Kleen

From: Andi Kleen <[email protected]>

Haswell has two additional LBR from flags for TSX: intx and abort, implemented as a new v4 version of the LBR format.

Handle those in and adjust the sign extension code to still correctly extend. The flags are exported similarly in the LBR record to the existing misprediction flag

I'm trying to update the perf_event_open() manpage for the new changes that were in Linux 3.11 and am having trouble getting info on exactly what these new fields mean.

include/uapi/linux/perf_event.h enum perf_branch_sample_type { · line 162

Quoted 3 lines
include/uapi/linux/perf_event.h +3 -0
+	PERF_SAMPLE_BRANCH_ABORT_TX	= 1U << 7, /* transaction aborts */
+	PERF_SAMPLE_BRANCH_IN_TX	= 1U << 8, /* in transaction */
+	PERF_SAMPLE_BRANCH_NO_TX	= 1U << 9, /* not in transaction */

so if you specify these flags in branch_sample_type, what information appears in the branch record?

If you get an abort, what address appears in the record?

What does it mean in regards to a branch entry to be or not be in a transaction?

If you set "in transaction" does that then only record branches that are in transactions? What happens if you set both in transaction and not in?

Is there some sort of document from intel you can link to that describes all of this?

Thanks,

Vince

#7 lore raw

Re: [PATCH 5/5] perf, x86: Support Haswell v4 LBR format v2

Quoted 6 lines

include/uapi/linux/perf_event.h enum perf_branch_sample_type { · line 162

Quoted 3 lines
include/uapi/linux/perf_event.h +3 -0
+	PERF_SAMPLE_BRANCH_ABORT_TX	= 1U << 7, /* transaction aborts */
+	PERF_SAMPLE_BRANCH_IN_TX	= 1U << 8, /* in transaction */
+	PERF_SAMPLE_BRANCH_NO_TX	= 1U << 9, /* not in transaction */

so if you specify these flags in branch_sample_type, what information appears in the branch record?

This is just a filter, so when set branches that do not satisfy the filter are not reported.

The patches to export the new fields haven't been merged yet.

Quoted 2 lines

If you get an abort, what address appears in the record?

Abort is a jump from the abort point to the abort handler or the HLE XACQUIRE lock instruction.

Quoted 2 lines

What does it mean in regards to a branch entry to be or not be in a transaction?

When the branch is executed the CPU is in the transactional execution state.

Quoted 3 lines

If you set "in transaction" does that then only record branches that are in transactions?

Yes, like all the other branch filter flags.

Quoted 1 line

What happens if you set both in transaction and not in?

Then you get all branches.

Quoted 3 lines

Is there some sort of document from intel you can link to that describes all of this?

http://download.intel.com/products/processor/manual/253669.pdf
Chapter 17.8

-Andi

Signature · 1 line
[email protected] -- Speaking for myself only
#8 lore raw

Re: [PATCH 5/5] perf, x86: Support Haswell v4 LBR format v2

Quoted 9 lines · Andi Kleen
Quoted 6 lines

include/uapi/linux/perf_event.h enum perf_branch_sample_type { · line 162

Quoted 3 lines
include/uapi/linux/perf_event.h +3 -0
+	PERF_SAMPLE_BRANCH_ABORT_TX	= 1U << 7, /* transaction aborts */
+	PERF_SAMPLE_BRANCH_IN_TX	= 1U << 8, /* in transaction */
+	PERF_SAMPLE_BRANCH_NO_TX	= 1U << 9, /* not in transaction */

so if you specify these flags in branch_sample_type, what information appears in the branch record?

This is just a filter, so when set branches that do not satisfy the filter are not reported.

Is the implementation a direct mapping to the LBR documentation or has it been generic so non-Intel architectures can use it?

Quoted 1 line

The patches to export the new fields haven't been merged yet.

What does this mean? The above values are exported as part of
include/uapi/linux/perf_event.h
Do they not work yet?

Quoted 3 lines
Quoted 1 line

What happens if you set both in transaction and not in?

Then you get all branches.

so what happens if you set neither "PERF_SAMPLE_BRANCH_IN_TX" nor "PERF_SAMPLE_BRANCH_NO_TX"? Logically you'd get no branches at all, but that can't be true as all code prior to 3.11 didn't set those values.

Vince

author
#9 lore raw

Re: [PATCH 5/5] perf, x86: Support Haswell v4 LBR format v2

Quoted 9 lines · Vince Weaver
Quoted 7 lines · Andi Kleen
Quoted 6 lines

include/uapi/linux/perf_event.h enum perf_branch_sample_type { · line 162

Quoted 3 lines
include/uapi/linux/perf_event.h +3 -0
+	PERF_SAMPLE_BRANCH_ABORT_TX	= 1U << 7, /* transaction aborts */
+	PERF_SAMPLE_BRANCH_IN_TX	= 1U << 8, /* in transaction */
+	PERF_SAMPLE_BRANCH_NO_TX	= 1U << 9, /* not in transaction */

so if you specify these flags in branch_sample_type, what information appears in the branch record?

Quoted 5 lines
Quoted 2 lines

This is just a filter, so when set branches that do not satisfy the filter are not reported.

Is the implementation a direct mapping to the LBR documentation or has it been generic so non-Intel architectures can use it?

It's not a direct mapping (no_tx doesn't exist in the hardware) If other architectures have similar capabilities they can likely use it.

Quoted 6 lines
Quoted 2 lines

The patches to export the new fields haven't been merged yet.

What does this mean? The above values are exported as part of
include/uapi/linux/perf_event.h
Do they not work yet?

You can filter on the fields, but you can't see them outside the kernel driver yet. The patch to see them is still pending.

Quoted 8 lines
Quoted 4 lines
Quoted 2 lines

What happens if you set both in transaction and not in?

Then you get all branches.

so what happens if you set neither "PERF_SAMPLE_BRANCH_IN_TX" nor "PERF_SAMPLE_BRANCH_NO_TX"? Logically you'd get no branches at all, but that can't be true as all code prior to 3.11 didn't set those values.

Then you get all branches too

(that's how all the other filters work too)

-Andi

Signature · 1 line
[email protected] -- Speaking for myself only.
#10 lore raw

Re: [PATCH 5/5] perf, x86: Support Haswell v4 LBR format v2

Quoted 6 lines · Andi Kleen
Quoted 3 lines

What does this mean? The above values are exported as part of
include/uapi/linux/perf_event.h
Do they not work yet?

You can filter on the fields, but you can't see them outside the kernel driver yet. The patch to see them is still pending.

so you can filter for aborts, but they'll never show up in the lbr[] sample buffer?

Quoted 11 lines
Quoted 7 lines
Quoted 3 lines
Quoted 1 line

What happens if you set both in transaction and not in?

Then you get all branches.

so what happens if you set neither "PERF_SAMPLE_BRANCH_IN_TX" nor "PERF_SAMPLE_BRANCH_NO_TX"? Logically you'd get no branches at all, but that can't be true as all code prior to 3.11 didn't set those values.

Then you get all branches too

(that's how all the other filters work too)

This is a really confusing API

so does setting "PERF_SAMPLE_BRANCH_ANY" also enable all of the TX types?

Is leaving branch_sample_type at 0 the same as setting it to all 1s?

Vince

#11 lore raw

Re: [PATCH 5/5] perf, x86: Support Haswell v4 LBR format v2

Quoted 11 lines · Vince Weaver
Quoted 6 lines · Andi Kleen
Quoted 3 lines

What does this mean? The above values are exported as part of
include/uapi/linux/perf_event.h
Do they not work yet?

You can filter on the fields, but you can't see them outside the kernel driver yet. The patch to see them is still pending.

so you can filter for aborts, but they'll never show up in the lbr[] sample buffer?

They will show up, you just don't know that they are aborts because the two new status bits are not exported.

Quoted 11 lines
Quoted 11 lines
Quoted 8 lines
Quoted 4 lines
Quoted 2 lines

What happens if you set both in transaction and not in?

Then you get all branches.

so what happens if you set neither "PERF_SAMPLE_BRANCH_IN_TX" nor "PERF_SAMPLE_BRANCH_NO_TX"? Logically you'd get no branches at all, but that can't be true as all code prior to 3.11 didn't set those values.

Then you get all branches too

Quoted 7 lines
Quoted 1 line

(that's how all the other filters work too)

This is a really confusing API

so does setting "PERF_SAMPLE_BRANCH_ANY" also enable all of the TX types?

Is leaving branch_sample_type at 0 the same as setting it to all 1s?

I believe so.

It may be also that the catch all only works if everything is 0.

-Andi

Signature · 1 line
[email protected] -- Speaking for myself only
#12 lore raw
In reply to Andi Kleen
Quoted 2 lines · Andi Kleen

This is based on v7 of the full Haswell PMU support, rebased, and stripped down to the bare bones

Ok, I found some time to still squeeze this into the v3.10 x86 PMU bits merge window but ran into problems.

You say it's barebones, yet it does not work :-( How well was this patch-set tested on non-Haswell hardware, which makes up 99.99% of our installed base?

In particular, after applying your patches, 'perf top' stopped working on an Intel testbox of mine:

  processor       : 15
  vendor_id       : GenuineIntel
  cpu family      : 6
  model           : 26
  model name      : Intel(R) Xeon(R) CPU           X55600 @ 2.80GHz
  stepping        : 5

'perf top' just does not produce any profiling output - it says 0 events.

Quoted 1 line

Most interesting new features are not in this patchkit

Sigh, we don't want esoteric Haswell-only features that only a very small subset of people will use.

As I mentioned before, we want _EXISTING_ tools to work and we want the patchset to be well-tested. Your Haswell-only extensions might be 'interesting' to you but are completely uninteresting to most users who just want bog standard profiling to work!

Until you don't understand that you will run into problems like this, putting priorities not into making existing stuff work, but working on esoteric extensions and featurities, missing the forest from all the trees...

Your stubborn incompetence is the reason why we are at version 11 of the patch-set (!) which is _still_ trivially unacceptable and there's _still_ no Haswell support in the upstream kernel...

Thanks,

Ingo

#13 lore raw
Quoted 6 lines · Ingo Molnar

You say it's barebones, yet it does not work :-( How well was this patch-set tested on non-Haswell hardware, which makes up 99.99% of our installed base?

In particular, after applying your patches, 'perf top' stopped working on an Intel testbox of mine:

The other problem I noticed was stylistic: when I applied your patches for testing even Git complained about their cleanliness ...

To quote from Documentation/SubmittingPatches:

4) Style check your changes.

Check your patch for basic style violations, details of which can be found in Documentation/CodingStyle. Failure to do so simply wastes the reviewers time and will get your patch rejected, probably without even being read.

At a minimum you should check your patches with the patch style checker prior to submission (scripts/checkpatch.pl). You should be able to justify all violations that remain in your patch.

Please make your patches less sloppy!

Thanks,

Ingo

#14 lore raw
Quoted 22 lines · Ingo Molnar
Quoted 6 lines · Ingo Molnar

You say it's barebones, yet it does not work :-( How well was this patch-set tested on non-Haswell hardware, which makes up 99.99% of our installed base?

In particular, after applying your patches, 'perf top' stopped working on an Intel testbox of mine:

The other problem I noticed was stylistic: when I applied your patches for testing even Git complained about their cleanliness ...

To quote from Documentation/SubmittingPatches:

4) Style check your changes.

Check your patch for basic style violations, details of which can be found in Documentation/CodingStyle. Failure to do so simply wastes the reviewers time and will get your patch rejected, probably without even being read.

Quoted 5 lines

At a minimum you should check your patches with the patch style checker prior to submission (scripts/checkpatch.pl). You should be able to justify all violations that remain in your patch.

Please make your patches less sloppy!

Andi, you have not replied to this mail of mine.

What new measures are you taking to avoid such annoying stylistic problems to creep into your patches?

These problems are regular in your patches and that has been going on for years - causing maintenance overhead for many maintainers, not just me.

Apparently you are not using proper tooling (checkpatch.pl for example) to check your patches. If you refuse to take action I will have to stop dealing with your patches directly altogether - the overhead just does not justify the effort. You'll need to get your patches reviewed by and signed off by a more experienced kernel hacker who knows how to submit patches.

Thanks,

Ingo

author
#15 lore raw
In reply to Ingo Molnar
Quoted 3 lines

How well was this
patch-set tested on non-Haswell hardware, which makes up 99.99% of our installed base?

I tested on a couple systems now and then: usually Haswell, IvyBridge, sometimes also Westmere and Atom. I don't retest every iteration, as you know most of the changes you're requesting don't affect the binary.

My test bed is likely to be smaller than yours though and as usual as you well know some part of the kernel QA is after release.

Quoted 9 lines

In particular, after applying your patches, 'perf top' stopped working on an Intel testbox of mine:

  processor       : 15
  vendor_id       : GenuineIntel
  cpu family      : 6
  model           : 26
  model name      : Intel(R) Xeon(R) CPU           X55600 @ 2.80GHz

I assume the second 0 is a typo?

Quoted 3 lines

stepping : 5

'perf top' just does not produce any profiling output - it says 0 events.

Thanks for testing.

I found a similar system (not same stepping, but same model) and tested perf top works fine here. Also on a couple of other systems.

Since I cannot reproduce I would need your help debugging it.

I assume it worked before my patches. If you don't know please double check. Also I assume there's no general problem between the user land perf you used and the kernel.

The only patch I could think of which may affect other systems is the moving of the APIC ack.

So does it work if you revert

perf, x86: Move NMI clearing to end of PMI handler after ...

If that is it we could white list it for Haswell.

If that's not it I may need a bisect, assuming the problem is stable.

-Andi

author
#16 lore raw

Your action on perf bug report is requested was Re: Basic perf PMU support for Haswell v11

Quoted 4 lines

I found a similar system (not same stepping, but same model) and tested perf top works fine here. Also on a couple of other systems.

Since I cannot reproduce I would need your help debugging it.

Ingo, I haven't heard back from you on this.

You reported an unreproducable bug. I gave you several steps to diagnose the problem, so that we can make progress on this.

You've had several days time now to to this, but I have not heard from you.

If you don't report back from 5/3/2013 I'll assume it was some other mistake on your side.

As a reminder:

I assume it worked before my patches. If you don't know please double check. Also I assume there's no general problem between the user land perf you used and the kernel.

The only patch I could think of which may affect other systems is the moving of the APIC ack.

So does it work if you revert

perf, x86: Move NMI clearing to end of PMI handler after ...

If that is it we could white list it for Haswell.

If that's not it I may need a bisect, assuming the problem is stable.

Thanks for your cooperation.

-Andi

#17 lore raw

Re: Your action on perf bug report is requested was Re: Basic perf PMU support for Haswell v11

Quoted 6 lines · Andi Kleen
Quoted 4 lines

I found a similar system (not same stepping, but same model) and tested perf top works fine here. Also on a couple of other systems.

Since I cannot reproduce I would need your help debugging it.

Ingo, I haven't heard back from you on this.

FYI, the v3.10 merge window has started 3+ days ago.

The merge window is a very busy time period for Linus and maintainers alike, and developers should generally not expect maintainers to deal with new experimental patches near to or especially during the merge window!

As a special exception I tried and tested your patches on Friday and reported back to you the bug, 2 days before the opening of the merge window - but you should not expect out of order treatment of development patches during the merge window

I might have time to look at your patches in a few days. No promises - I still haven't merged all trees to Linus. I think the 11 review cycles of your Haswell patch-set are proof enough of my willingness to deal with your patches.

Quoted 2 lines

You reported an unreproducable bug. I gave you several steps to diagnose the problem, so that we can make progress on this.

It's entirely reproducible here on that testbox, and was caused by your patches - I partially bisected it to your series. (but not to a specific patch in your series - ran out of time.)

Quoted 2 lines

You've had several days time now to to this, but I have not heard from you.

FYI, you are entirely confused about how Linux maintenance works near and during the merge window. Many maintainers don't take any patches but only take fixes for already applied patches. (In -tip we generally try to freeze a week before the merge window, so your patches missed the v3.10 merge window by a wide margin I'm afraid.)

Furthermore, frankly, the nasty, demanding tone of your mail, expecting and demanding a reply to your mail within two work days (!) is ridiculous and unacceptable and does not make it more likely for me to make special exceptions for your patches.

Thanks,

Ingo

#18 lore raw
In reply to Andi Kleen

[ FYI, we are still in the merge window when maintainers are very busy, so
don't expect quick replies to mails that are not about merge window related patches and commits. Those issues are typically handled after -rc1 has been released, once most of the merge fallout in the upstream kernel has been resolved. ]

Quoted 16 lines · Andi Kleen
Quoted 3 lines

How well was this
patch-set tested on non-Haswell hardware, which makes up 99.99% of our installed base?

I tested on a couple systems now and then: usually Haswell, IvyBridge, sometimes also Westmere and Atom. I don't retest every iteration, as you know most of the changes you're requesting don't affect the binary.

My test bed is likely to be smaller than yours though and as usual as you well know some part of the kernel QA is after release.

Quoted 4 lines

In particular, after applying your patches, 'perf top' stopped working on an Intel testbox of mine:

Quoted 7 lines
Quoted 5 lines
  processor       : 15
  vendor_id       : GenuineIntel
  cpu family      : 6
  model           : 26
  model name      : Intel(R) Xeon(R) CPU           X55600 @ 2.80GHz

I assume the second 0 is a typo?

Probably a typo in the BIOS.

Quoted 4 lines
Quoted 3 lines

stepping : 5

'perf top' just does not produce any profiling output - it says 0 events.

Quoted 8 lines

Thanks for testing.

I found a similar system (not same stepping, but same model) and tested perf top works fine here. Also on a couple of other systems.

Since I cannot reproduce I would need your help debugging it.

I assume it worked before my patches.

Yes, obviously.

Here's another easy to test symptom of the bug:

 $ perf record ./hackbench 10
 Time: 0.097
 [ perf record: Woken up 1 times to write data ]
 [ perf record: Captured and wrote 0.043 MB perf.data (~1866 samples) ]
 $ perf report --stdio
 Error:
 The perf.data file has no samples!

Expected result is a profile displayed by 'perf report'.

Quoted 5 lines

[...] If you don't know please double check. Also I assume there's no general problem between the user land perf you used and the kernel.

The only patch I could think of which may affect other systems is the moving of the APIC ack.

Btw., I warned you about the delicate placement of the APIC ACK in my Haswell patches review feedback mail, months ago:

https://lkml.org/lkml/2013/2/13/78

which mail you never replied to and which warning you apparently ignored.

When modifying the PMU ack sequence, please find the relevant Intel SDM that recommends a different ACK sequence from what is implemented currently, and document this in the changelog.

I'm going to ignore your APIC ACK patch until you do it properly.

Quoted 5 lines

So does it work if you revert

perf, x86: Move NMI clearing to end of PMI handler after ...

If that is it we could white list it for Haswell.

No, reverting that patch did not fix the bug.

I have bisected it down to this patch of yours:

"perf/x86: Add Haswell PMU support"

Most of that patch has no effect on non-Haswell machines, so the scope of problematic changes should be pretty small.

My quick guess is that your patch broke fixed counters.

If you find the bug or want me to test anything please send a delta patch, relative to your last series - as I have parts of your patches applied already locally with cleanups, etc.

Thanks,

Ingo

#19 lore raw
In reply to Andi Kleen
Quoted 1 line · Andi Kleen

v11: Rebase to perf/core. Fix extra regs. Rename INTX.

Actually, you did not do what I asked you to do, to rename INTX to IN_TX, you still kept the 'INTX' pattern that is confusingly similar to interrupts related names like 'INT3'.

I still see, even your latest patch, this:

+#define HSW_INTX (1ULL << 32)
+#define HSW_INTX_CHECKPOINTED (1ULL << 33)

Why are you ignoring maintainer requests, repeatedly?

It woulnd't be a big issue and I'd do the rename myself if this wasn't a repeat pattern of passive-aggressive obstruction from you spanning several years, resisting or obstructing maintainer feedback whenever you can ...

Thanks,

Ingo