Skip to contentSkip to conversation

Do not list or allow unsupported perf events

RFC Tony Jones 3 participants 4 messages Sep 11, 2015 – Sep 13, 2015 3 messages not in the archive

lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]

Tony Jones

author
#1 lore raw

Review

Quoted 11 lines · Tony Jones
(quoted diff) +11 -0
+static int is_event_supported(const struct ctx_opts *ctx)
+{
+	int fd;
+	struct perf_event_attr attr = {
+		.type = ctx->u.perf.type,
+		.config = ctx->u.perf.config,
+		.disabled = 1
+	};
+
+	if (ctx->ctx_type == CONTEXT_PERF_CPU_COUNTER ||
+	    ctx->ctx_type == CONTEXT_PERF_THREAD_COUNTER) {

obviously CONTEXT_PERF_THREAD_COUNTER is UST. Regardless, there are some obvious usability issues (below). I need to look at the code some more.

# lttng add-context -u -t perf:thread:alignment-faults
UST context perf:thread:alignment-faults added to all channels

(invalid, better error needed, also sometimes -k -t perf:thread:* succeeds which seems broken)

# lttng add-context -k -t perf:thread:alignment-faults
Error: perf:thread:alignment-faults: Add kernel context failed
Warning: Some command(s) went wrong
#2 lore raw
In reply to a message not in the archive

CCing the new lttng-tools maintainer, Jérémie. More below,

Quoted 11 lines · ----- On Sep 10, 2015, at 3:21 PM, Tony Jones [email protected] wrote:

Query the perf api to determine the list of events that are actually
supported by the PMU. Reject events that are not supported. On an
AMD Opteron 6128 the following currently listed events are not supported:

perf:bus-cycles, perf:L1-dcache-store-misses, perf:L1-icache-stores, perf:L1-icache-store-misses, perf:L1-icache-prefetch-misses, perf:LLC-store-misses, perf:LLC-prefetches, perf:LLC-prefetch-misses, perf:dTLB-stores, perf:dTLB-store-misses, perf:dTLB-prefetches, perf:dTLB-prefetch-misses

Quoted 8 lines

It's not clear to me (hence the RFC) why parts of the perf ABI definitions had been replicated into add_context.c rather than just including perf_event.h (other than to avoid an autotools dependancy/check on perf_event.h which this patch still needs). It is also possible that querying the available events would be better handled over the lttng user->kernel ABI.

Thoughts?

It might make sense to add a dependency on perf_event.h.

I'm wondering if your detection technique below really try to allocate a PMU counter for a short period of time ? This could be an issue since there is a limited amount of counters available.

Thoughts ?

Thanks,

Mathieu

Quoted 146 lines

---
src/bin/lttng/commands/add_context.c | 92 ++++++++++--------------------------

1 file changed, 25 insertions(+), 67 deletions(-)
src/bin/lttng/commands/add_context.c +25 -67
index cf3f3ef..6f52edb 100644
@@ -26,6 +26,8 @@
 #include <sys/types.h>
 #include <unistd.h>
 
+#include <linux/perf_event.h>
+
 #include <urcu/list.h>
 
 #include <common/mi-lttng.h>
@@ -77,72 +79,6 @@ enum context_type {
 	CONTEXT_PERF_THREAD_COUNTER = 14,
 };
 
-/*
- * Taken from the Perf ABI (all enum perf_*)
- */
-enum perf_type {
-	PERF_TYPE_HARDWARE = 0,
-	PERF_TYPE_SOFTWARE = 1,
-	PERF_TYPE_HW_CACHE = 3,
-};
-
-enum perf_count_hard {
-	PERF_COUNT_HW_CPU_CYCLES		= 0,
-	PERF_COUNT_HW_INSTRUCTIONS		= 1,
-	PERF_COUNT_HW_CACHE_REFERENCES		= 2,
-	PERF_COUNT_HW_CACHE_MISSES		= 3,
-	PERF_COUNT_HW_BRANCH_INSTRUCTIONS	= 4,
-	PERF_COUNT_HW_BRANCH_MISSES		= 5,
-	PERF_COUNT_HW_BUS_CYCLES		= 6,
-	PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	= 7,
-	PERF_COUNT_HW_STALLED_CYCLES_BACKEND	= 8,
-};
-
-enum perf_count_soft {
-	PERF_COUNT_SW_CPU_CLOCK        = 0,
-	PERF_COUNT_SW_TASK_CLOCK       = 1,
-	PERF_COUNT_SW_PAGE_FAULTS      = 2,
-	PERF_COUNT_SW_CONTEXT_SWITCHES = 3,
-	PERF_COUNT_SW_CPU_MIGRATIONS   = 4,
-	PERF_COUNT_SW_PAGE_FAULTS_MIN  = 5,
-	PERF_COUNT_SW_PAGE_FAULTS_MAJ  = 6,
-	PERF_COUNT_SW_ALIGNMENT_FAULTS = 7,
-	PERF_COUNT_SW_EMULATION_FAULTS = 8,
-};
-
-/*
- * Generalized hardware cache events:
- *
- *       { L1-D, L1-I, LLC, ITLB, DTLB, BPU } x
- *       { read, write, prefetch } x
- *       { accesses, misses }
- */
-enum perf_hw_cache_id {
-	PERF_COUNT_HW_CACHE_L1D			= 0,
-	PERF_COUNT_HW_CACHE_L1I			= 1,
-	PERF_COUNT_HW_CACHE_LL			= 2,
-	PERF_COUNT_HW_CACHE_DTLB		= 3,
-	PERF_COUNT_HW_CACHE_ITLB		= 4,
-	PERF_COUNT_HW_CACHE_BPU			= 5,
-
-	PERF_COUNT_HW_CACHE_MAX,		/* non-ABI */
-};
-
-enum perf_hw_cache_op_id {
-	PERF_COUNT_HW_CACHE_OP_READ		= 0,
-	PERF_COUNT_HW_CACHE_OP_WRITE		= 1,
-	PERF_COUNT_HW_CACHE_OP_PREFETCH		= 2,
-
-	PERF_COUNT_HW_CACHE_OP_MAX,		/* non-ABI */
-};
-
-enum perf_hw_cache_op_result_id {
-	PERF_COUNT_HW_CACHE_RESULT_ACCESS	= 0,
-	PERF_COUNT_HW_CACHE_RESULT_MISS		= 1,
-
-	PERF_COUNT_HW_CACHE_RESULT_MAX,		/* non-ABI */
-};
-
 static struct poptOption long_options[] = {
 	/* longName, shortName, argInfo, argPtr, value, descrip, argDesc */
 	{"help",           'h', POPT_ARG_NONE, 0, OPT_HELP, 0, 0},
@@ -461,6 +397,25 @@ struct ctx_type_list {
 	.head = CDS_LIST_HEAD_INIT(ctx_type_list.head),
 };
 
+static int is_event_supported(const struct ctx_opts *ctx)
+{
+	int fd;
+	struct perf_event_attr attr = {
+		.type = ctx->u.perf.type,
+		.config = ctx->u.perf.config,
+		.disabled = 1
+	};
+
+	if (ctx->ctx_type == CONTEXT_PERF_CPU_COUNTER ||
+	    ctx->ctx_type == CONTEXT_PERF_THREAD_COUNTER) {
+		if ((fd = syscall(__NR_perf_event_open, &attr,  0, -1, -1, 0)) < 0)
+			return 0;
+
+		close(fd);
+	}
+	return 1;
+}
+
 /*
 * Pretty print context type.
 */
@@ -473,6 +428,8 @@ static void print_ctx_type(FILE *ofp)
 	fprintf(ofp, "%s", indent);
 	len = indent_len;
 	while (ctx_opts[i].symbol != NULL) {
+		if (!is_event_supported(&ctx_opts[i]))
+			goto next;
 		if (!ctx_opts[i].hide_help) {
 			if (len > indent_len) {
 				if (len + strlen(ctx_opts[i].symbol) + 2
@@ -486,6 +443,7 @@ static void print_ctx_type(FILE *ofp)
 			}
 			len += fprintf(ofp, "%s", ctx_opts[i].symbol);
 		}
+next:
 		i++;
 	}
 }
@@ -538,7 +496,7 @@ static int find_ctx_type_idx(const char *opt)
 	int ret = -1, i = 0;
 
 	while (ctx_opts[i].symbol != NULL) {
-		if (strcmp(opt, ctx_opts[i].symbol) == 0) {
+		if (strcmp(opt, ctx_opts[i].symbol) == 0 && is_event_supported(&ctx_opts[i]))
 {
 			ret = i;
 			goto end;
		}

--
2.4.6

Signature · 8 lines
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com

_______________________________________________
lttng-dev mailing list
[email protected]
http://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev

A message that is not in the archive (sent off-list or not yet received): [email protected]

#4 lore raw

CCing Jeremie.

Quoted 16 lines · ----- On Sep 10, 2015, at 8:15 PM, Tony Jones [email protected] wrote:

Review

Quoted 11 lines · Tony Jones
(quoted diff) +11 -0
+static int is_event_supported(const struct ctx_opts *ctx)
+{
+	int fd;
+	struct perf_event_attr attr = {
+		.type = ctx->u.perf.type,
+		.config = ctx->u.perf.config,
+		.disabled = 1
+	};
+
+	if (ctx->ctx_type == CONTEXT_PERF_CPU_COUNTER ||
+	    ctx->ctx_type == CONTEXT_PERF_THREAD_COUNTER) {

obviously CONTEXT_PERF_THREAD_COUNTER is UST. Regardless, there are some obvious usability issues (below). I need to look at the code some more.

OK, keep us posted. It would be nice if we can list more specifically which counters are supported on the machine.

Thanks,

Mathieu

Quoted 9 lines
# lttng add-context -u -t perf:thread:alignment-faults
UST context perf:thread:alignment-faults added to all channels

(invalid, better error needed, also sometimes -k -t perf:thread:* succeeds which seems broken)

# lttng add-context -k -t perf:thread:alignment-faults
Error: perf:thread:alignment-faults: Add kernel context failed
Warning: Some command(s) went wrong
Signature · 3 lines
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com

A message that is not in the archive (sent off-list or not yet received): [email protected]

#6 lore raw
Quoted 22 lines · Mathieu Desnoyers

CCing the new lttng-tools maintainer, Jérémie. More below,

Quoted 18 lines · ----- On Sep 10, 2015, at 3:21 PM, Tony Jones [email protected] wrote:

Query the perf api to determine the list of events that are actually
supported by the PMU. Reject events that are not supported. On an
AMD Opteron 6128 the following currently listed events are not supported:

perf:bus-cycles, perf:L1-dcache-store-misses, perf:L1-icache-stores, perf:L1-icache-store-misses, perf:L1-icache-prefetch-misses, perf:LLC-store-misses, perf:LLC-prefetches, perf:LLC-prefetch-misses, perf:dTLB-stores, perf:dTLB-store-misses, perf:dTLB-prefetches, perf:dTLB-prefetch-misses

It's not clear to me (hence the RFC) why parts of the perf ABI definitions had been replicated into add_context.c rather than just including perf_event.h (other than to avoid an autotools dependancy/check on perf_event.h which this patch still needs). It is also possible that querying the available events would be better handled over the lttng user->kernel ABI.

Quoted 7 lines
Quoted 1 line

Thoughts?

It might make sense to add a dependency on perf_event.h.

I'm wondering if your detection technique below really try to allocate a PMU counter for a short period of time ? This could be an issue since there is a limited amount of counters available.

It's pretty much the way 'perf list' works but I hear what you are saying. Since it's possible to incrementally add events I expect the detection could fail to allocate any thus reporting no available events. I'd not considered this, I'll have to test the hypothesis.

_______________________________________________
lttng-dev mailing list
[email protected]
http://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev