Skip to contentSkip to conversation

perfcounters: fix type/event_id layout on big-endian systems

PATCH In review Paul Mackerras 4 participants 9 messages Mar 21, 2009

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 perfcounters: fix type/event_id layout on big-endian systems

Impact: build fix for powerpc

Commit db3a944aca35ae61 ("perf_counter: revamp syscall input ABI") expanded the hw_event.type field into a union of structs containing bitfields. In particular it introduced a type field and a raw_type field, with the intention that the 1-bit raw_type field should overlay the most-significant bit of the 8-bit type field, and in fact perf_counter_alloc() now assumes that (or at least, assumes that raw_type doesn't overlay any of the bits that are 1 in the values of PERF_TYPE_{HARDWARE,SOFTWARE,TRACEPOINT}).

Unfortunately this is not true on big-endian systems such as PowerPC, where bitfields are laid out from left to right, i.e. from most significant bit to least significant. This means that setting hw_event.type = PERF_TYPE_SOFTWARE will set hw_event.raw_type to 1.

This fixes it by making the layout depend on whether or not __BIG_ENDIAN_BITFIELD is defined. It's a bit ugly, but that's what we get for using bitfields in a user/kernel ABI.

Also, that commit didn't fix up some places in arch/powerpc/kernel/ perf_counter.c where hw_event.raw and hw_event.event_id were used. This fixes them too.

---
This is in the master branch of my perfcounters.git repository at:

git://git.kernel.org/pub/scm/linux/kernel/git/paulus/perfcounters.git master

 arch/powerpc/kernel/perf_counter.c |    9 +++++----
 include/linux/perf_counter.h       |   12 ++++++++++++
 2 files changed, 17 insertions(+), 4 deletions(-)
arch/powerpc/kernel/perf_counter.c +5 -4
index 830ca9c..6413d9c 100644
@@ -602,12 +602,13 @@ hw_perf_counter_init(struct perf_counter *counter)
 		return NULL;
 	if ((s64)counter->hw_event.irq_period < 0)
 		return NULL;
-	ev = counter->hw_event.event_id;
-	if (!counter->hw_event.raw) {
-		if (ev >= ppmu->n_generic ||
-		    ppmu->generic_events[ev] == 0)
+	if (!counter->hw_event.raw_type) {
+		ev = counter->hw_event.event_id;
+		if (ev >= ppmu->n_generic || ppmu->generic_events[ev] == 0)
 			return NULL;
 		ev = ppmu->generic_events[ev];
+	} else {
+		ev = counter->hw_event.raw_event_id;
 	}
 	counter->hw.config_base = ev;
 	counter->hw.idx = 0;
include/linux/perf_counter.h +12 -0
index a4b76c0..98f5990 100644
@@ -15,6 +15,7 @@
 
 #include <linux/types.h>
 #include <linux/ioctl.h>
+#include <asm/byteorder.h>
 
 /*
  * User-space ABI bits:
@@ -86,6 +87,7 @@ enum perf_counter_record_type {
  */
 struct perf_counter_hw_event {
 	union {
+#ifndef __BIG_ENDIAN_BITFIELD
 		struct {
 			__u64			event_id	: 56,
 						type		:  8;
@@ -94,6 +96,16 @@ struct perf_counter_hw_event {
 			__u64			raw_event_id	: 63,
 						raw_type	:  1;
 		};
+#else
+		struct {
+			__u64			type		:  8,
+						event_id	: 56;
+		};
+		struct {
+			__u64			raw_type	:  1,
+						raw_event_id	: 63;
+		};
+#endif /* __BIT_ENDIAN_BITFIELD */
 		__u64		event_config;
 	};
 
Signature · 1 line
1.5.6.3
#1 lore raw
Quoted 10 lines · Paul Mackerras

struct perf_counter_hw_event {

include/linux/perf_counter.h +1 -0
 	union {
+#ifndef __BIG_ENDIAN_BITFIELD
 		struct {
 			__u64			event_id	: 56,
 						type		:  8;
@@ -94,6 +96,16 @@ struct perf_counter_hw_event {
 			__u64			raw_event_id	: 63,
 						raw_type	:  1;
 		};

include/linux/perf_counter.h struct perf_counter_hw_event { · line 108

Quoted 10 lines
include/linux/perf_counter.h +10 -0
+#else
+		struct {
+			__u64			type		:  8,
+						event_id	: 56;
+		};
+		struct {
+			__u64			raw_type	:  1,
+						raw_event_id	: 63;
+		};
+#endif /* __BIT_ENDIAN_BITFIELD */

hm, this ifdef really looks ugly. How about just changing event_id to 64 bits and having a separate u32 type field? The size impact is minimal, the cleanliness win is significant :-)

Ingo

#2 lore raw
Quoted 3 lines · Ingo Molnar

hm, this ifdef really looks ugly. How about just changing event_id to 64 bits and having a separate u32 type field? The size impact is minimal, the cleanliness win is significant :-)

We wanted to get a complete and unique identifier for the event into 64 bits so that we could put it into the ring buffer for PERF_RECORD_GROUP and have it take up only one 8-byte slot and yet identify uniquely which counter's value follows it. I don't know that that is absolutely necessary but it sounds like a nice property.

We could easily go back to a 1-bit raw field and have the type be either a 64-bit raw value or an 8-bit type plus 32 or 56-bit event_id.

Or we could keep the current layout but use explicit shifts and masks rather than bitfields.

I don't know the current C rules concerning unions very well, but I have the impression that writing to one member of a union and reading another is undefined behaviour, which is another strike against the current code if true...

Anyway, the point is that the current code doesn't compile on powerpc and wouldn't work properly even if it did, so we need to do something.

Paul.

#3 lore raw
Quoted 20 lines · Paul Mackerras
Quoted 3 lines · Ingo Molnar

hm, this ifdef really looks ugly. How about just changing event_id to 64 bits and having a separate u32 type field? The size impact is minimal, the cleanliness win is significant :-)

We wanted to get a complete and unique identifier for the event into 64 bits so that we could put it into the ring buffer for PERF_RECORD_GROUP and have it take up only one 8-byte slot and yet identify uniquely which counter's value follows it. I don't know that that is absolutely necessary but it sounds like a nice property.

We could easily go back to a 1-bit raw field and have the type be either a 64-bit raw value or an 8-bit type plus 32 or 56-bit event_id.

Or we could keep the current layout but use explicit shifts and masks rather than bitfields.

Quoted 8 lines

I don't know the current C rules concerning unions very well, but I have the impression that writing to one member of a union and reading another is undefined behaviour, which is another strike against the current code if true...

Anyway, the point is that the current code doesn't compile on powerpc and wouldn't work properly even if it did, so we need to do something.

yeah - i pulled your fix, thanks Paul.

Maybe the best option is to get rid of the bitfields and use masks
...

Ingo

#4 lore raw
Quoted 1 line · Ingo Molnar

Maybe the best option is to get rid of the bitfields and use masks

Yeah, I think that'll be sanest..

#5 lore raw

PATCH 1/1 perf_counter: remove the event config bitfields

Quoted 5 lines · Peter Zijlstra
Quoted 1 line · Ingo Molnar

Maybe the best option is to get rid of the bitfields and use masks

Yeah, I think that'll be sanest..

Paul, I noticed you assign the raw config without mask, is that ok?

---
Subject: perf_counter: remove the event config bitfields

Since the bitfields turned into a bit of a mess, remove them and rely on good old masks.

---

arch/powerpc/kernel/perf_counter.c +3 -3
index 6413d9c..d056515 100644
@@ -602,13 +602,13 @@ hw_perf_counter_init(struct perf_counter *counter)
 		return NULL;
 	if ((s64)counter->hw_event.irq_period < 0)
 		return NULL;
-	if (!counter->hw_event.raw_type) {
-		ev = counter->hw_event.event_id;
+	if (!perf_event_raw(&counter->hw_event)) {
+		ev = perf_event_id(&counter->hw_event);
 		if (ev >= ppmu->n_generic || ppmu->generic_events[ev] == 0)
 			return NULL;
 		ev = ppmu->generic_events[ev];
 	} else {
-		ev = counter->hw_event.raw_event_id;
+		ev = perf_event_config(&counter->hw_event);
 	}
 	counter->hw.config_base = ev;
 	counter->hw.idx = 0;
arch/x86/kernel/cpu/perf_counter.c +4 -4
index 902282d..3f95b0c 100644
@@ -217,15 +217,15 @@ static int __hw_perf_counter_init(struct perf_counter *counter)
 	/*
 	 * Raw event type provide the config in the event structure
 	 */
-	if (hw_event->raw_type) {
-		hwc->config |= pmc_ops->raw_event(hw_event->raw_event_id);
+	if (perf_event_raw(hw_event)) {
+		hwc->config |= pmc_ops->raw_event(perf_event_config(hw_event));
 	} else {
-		if (hw_event->event_id >= pmc_ops->max_events)
+		if (perf_event_id(hw_event) >= pmc_ops->max_events)
 			return -EINVAL;
 		/*
 		 * The generic map:
 		 */
-		hwc->config |= pmc_ops->event_map(hw_event->event_id);
+		hwc->config |= pmc_ops->event_map(perf_event_id(hw_event));
 	}
 	counter->wakeup_pending = 0;
 
include/linux/perf_counter.h +34 -24

80 lines not on this page · raw message

kernel/perf_counter.c +12 -8
index f054b8c..bbd538a 100644
@@ -1379,7 +1379,7 @@ static void perf_counter_handle_group(struct perf_counter *counter)
 	list_for_each_entry(sub, &leader->sibling_list, list_entry) {
 		if (sub != counter)
 			sub->hw_ops->read(sub);
-		perf_counter_store_irq(counter, sub->hw_event.event_config);
+		perf_counter_store_irq(counter, sub->hw_event.config);
 		perf_counter_store_irq(counter, atomic64_read(&sub->count));
 	}
 }
@@ -1489,13 +1489,13 @@ static int perf_swcounter_match(struct perf_counter *counter,
 	if (counter->state != PERF_COUNTER_STATE_ACTIVE)
 		return 0;
 
-	if (counter->hw_event.raw_type)
+	if (perf_event_raw(&counter->hw_event))
 		return 0;
 
-	if (counter->hw_event.type != type)
+	if (perf_event_type(&counter->hw_event) != type)
 		return 0;
 
-	if (counter->hw_event.event_id != event)
+	if (perf_event_id(&counter->hw_event) != event)
 		return 0;
 
 	if (counter->hw_event.exclude_user && user_mode(regs))
@@ -1763,7 +1763,7 @@ static void tp_perf_counter_destroy(struct perf_counter *counter)
 static const struct hw_perf_counter_ops *
 tp_perf_counter_init(struct perf_counter *counter)
 {
-	int event_id = counter->hw_event.event_id;
+	int event_id = perf_event_id(&counter->hw_event);
 	int ret;
 
 	ret = ftrace_profile_enable(event_id);
@@ -1797,7 +1797,7 @@ sw_perf_counter_init(struct perf_counter *counter)
 	 * to be kernel events, and page faults are never hypervisor
 	 * events.
 	 */
-	switch (counter->hw_event.event_id) {
+	switch (perf_event_id(&counter->hw_event)) {
 	case PERF_COUNT_CPU_CLOCK:
 		hw_ops = &perf_ops_cpu_clock;
 
@@ -1882,9 +1882,12 @@ perf_counter_alloc(struct perf_counter_hw_event *hw_event,
 
 	hw_ops = NULL;
 
-	if (hw_event->raw_type)
+	if (perf_event_raw(hw_event)) {
 		hw_ops = hw_perf_counter_init(counter);
-	else switch (hw_event->type) {
+		goto done;
+	}
+
+	switch (perf_event_type(hw_event)) {
 	case PERF_TYPE_HARDWARE:
 		hw_ops = hw_perf_counter_init(counter);
 		break;
@@ -1902,6 +1905,7 @@ perf_counter_alloc(struct perf_counter_hw_event *hw_event,
 		kfree(counter);
 		return NULL;
 	}
+done:
 	counter->hw_ops = hw_ops;
 
 	return counter;
#6 lore raw

Re: [PATCH] perf_counter: remove the event config bitfields

Quoted 1 line · Peter Zijlstra

Paul, I noticed you assign the raw config without mask, is that ok?

Yep. That doesn't get assigned directly to a hardware register; it undergoes further processing first, which just pulls out the bits it needs. (It's not used directly because the hardware has a couple of registers that configure all the counters at once, rather than a configuration register per counter.)

Paul.

#7 lore raw
In reply to Paul Mackerras
Quoted 11 lines · Paul Mackerras

Impact: build fix for powerpc

Commit db3a944aca35ae61 ("perf_counter: revamp syscall input ABI") expanded the hw_event.type field into a union of structs containing bitfields. In particular it introduced a type field and a raw_type field, with the intention that the 1-bit raw_type field should overlay the most-significant bit of the 8-bit type field, and in fact perf_counter_alloc() now assumes that (or at least, assumes that raw_type doesn't overlay any of the bits that are 1 in the values of PERF_TYPE_{HARDWARE,SOFTWARE,TRACEPOINT}).

Quoted 8 lines

Unfortunately this is not true on big-endian systems such as PowerPC, where bitfields are laid out from left to right, i.e. from most significant bit to least significant. This means that setting hw_event.type = PERF_TYPE_SOFTWARE will set hw_event.raw_type to 1.

This fixes it by making the layout depend on whether or not __BIG_ENDIAN_BITFIELD is defined. It's a bit ugly, but that's what we get for using bitfields in a user/kernel ABI.

Hmm, does userspace know about __BIG_ENDIAN_BITFIELD too? If not, we've got ourselves a problem here.

Sorry about missing those powerpc bits, I guess I should dust-off the cross-compiler.

Quoted 37 lines
include/linux/perf_counter.h +12 -0
index a4b76c0..98f5990 100644
@@ -15,6 +15,7 @@
 
 #include <linux/types.h>
 #include <linux/ioctl.h>
+#include <asm/byteorder.h>
 
 /*
  * User-space ABI bits:
@@ -86,6 +87,7 @@ enum perf_counter_record_type {
  */
 struct perf_counter_hw_event {
 	union {
+#ifndef __BIG_ENDIAN_BITFIELD
 		struct {
 			__u64			event_id	: 56,
 						type		:  8;
@@ -94,6 +96,16 @@ struct perf_counter_hw_event {
 			__u64			raw_event_id	: 63,
 						raw_type	:  1;
 		};
+#else
+		struct {
+			__u64			type		:  8,
+						event_id	: 56;
+		};
+		struct {
+			__u64			raw_type	:  1,
+						raw_event_id	: 63;
+		};
+#endif /* __BIT_ENDIAN_BITFIELD */
 		__u64		event_config;
 	};
 
#8 lore raw
Quoted 2 lines · Peter Zijlstra

Hmm, does userspace know about __BIG_ENDIAN_BITFIELD too? If not, we've got ourselves a problem here.

Yes, it does (I did test the patch), but I think going to explicit shifting and masking will be better.

Paul.