Skip to contentSkip to conversation

perf, x86: Enable PEBS mode automatically for mem-{loads,stores} v3

PATCH In review Andi Kleen 4 participants 7 messages Jul 18, 2013 – Jul 23, 2013

Subsystems (2)
lore
Share & tools
lore

Andi Kleen

author
#0 lore raw

PATCH 1/1 perf, x86: Enable PEBS mode automatically for mem-{loads,stores} v3

From: Andi Kleen <[email protected]>

[The patch to enable this in the user tools has been sent separately]

With the earlier patches to automatically try cpu// and add a precise sys attribute, we can now enable PEBS for the mem-loads, mem-stores events everywhere.

This allows to use

perf record -e mem-loads ...

instead of

perf record -e cpu/mem-loads/p ...

Always use precise=2 even though it is costly pre-Haswell

Cc: [email protected]
v2: Different white space
v3: Always use precise=2, as people seem to think overhead doesn't matter.

---
 arch/x86/kernel/cpu/perf_event_intel.c | 14 +++++++++-----
 1 file changed, 9 insertions(+), 5 deletions(-)
arch/x86/kernel/cpu/perf_event_intel.c +9 -5
index fbc9210..ef9236b 100644
@@ -176,9 +176,12 @@ static struct extra_reg intel_snbep_extra_regs[] __read_mostly = {
 	EVENT_EXTRA_END
 };
 
-EVENT_ATTR_STR(mem-loads, mem_ld_nhm, "event=0x0b,umask=0x10,ldlat=3");
-EVENT_ATTR_STR(mem-loads, mem_ld_snb, "event=0xcd,umask=0x1,ldlat=3");
-EVENT_ATTR_STR(mem-stores, mem_st_snb, "event=0xcd,umask=0x2");
+EVENT_ATTR_STR(mem-loads, mem_ld_nhm,
+		"event=0x0b,umask=0x10,ldlat=3,precise=2");
+EVENT_ATTR_STR(mem-loads, mem_ld_snb,
+		"event=0xcd,umask=0x1,ldlat=3,precise=2");
+EVENT_ATTR_STR(mem-stores, mem_st_snb,
+		"event=0xcd,umask=0x2,precise=2");
 
 struct attribute *nhm_events_attrs[] = {
 	EVENT_PTR(mem_ld_nhm),
@@ -2034,8 +2037,9 @@ static __init void intel_nehalem_quirk(void)
 	}
 }
 
-EVENT_ATTR_STR(mem-loads,      mem_ld_hsw,     "event=0xcd,umask=0x1,ldlat=3");
-EVENT_ATTR_STR(mem-stores,     mem_st_hsw,     "event=0xd0,umask=0x82")
+EVENT_ATTR_STR(mem-loads,      mem_ld_hsw,
+		"event=0xcd,umask=0x1,ldlat=3,precise=2");
+EVENT_ATTR_STR(mem-stores,     mem_st_hsw,  "event=0xd0,umask=0x82,precise=2")
 
 static struct attribute *hsw_events_attrs[] = {
 	EVENT_PTR(mem_ld_hsw),
Signature · 1 line
1.8.1.4
#1 lore raw
Quoted 33 lines · Andi Kleen

From: Andi Kleen <[email protected]>

[The patch to enable this in the user tools has been sent separately]

With the earlier patches to automatically try cpu// and add a precise sys attribute, we can now enable PEBS for the mem-loads, mem-stores events everywhere.

This allows to use

perf record -e mem-loads ...

instead of

perf record -e cpu/mem-loads/p ...

Always use precise=2 even though it is costly pre-Haswell

Cc: [email protected]
v2: Different white space
v3: Always use precise=2, as people seem to think overhead doesn't matter.

---
 arch/x86/kernel/cpu/perf_event_intel.c | 14 +++++++++-----
 1 file changed, 9 insertions(+), 5 deletions(-)
arch/x86/kernel/cpu/perf_event_intel.c +0 -0
index fbc9210..ef9236b 100644
@@ -176,9 +176,12 @@ static struct extra_reg intel_snbep_extra_regs[] __read_mostly = {
 	EVENT_EXTRA_END
 };

arch/x86/kernel/cpu/perf_event_intel.c static struct extra_reg intel_snbep_extra_regs[] __read_mostly = { · line 184

Quoted 10 lines
arch/x86/kernel/cpu/perf_event_intel.c +6 -3
 
-EVENT_ATTR_STR(mem-loads, mem_ld_nhm, "event=0x0b,umask=0x10,ldlat=3");
-EVENT_ATTR_STR(mem-loads, mem_ld_snb, "event=0xcd,umask=0x1,ldlat=3");
-EVENT_ATTR_STR(mem-stores, mem_st_snb, "event=0xcd,umask=0x2");
+EVENT_ATTR_STR(mem-loads, mem_ld_nhm,
+		"event=0x0b,umask=0x10,ldlat=3,precise=2");
+EVENT_ATTR_STR(mem-loads, mem_ld_snb,
+		"event=0xcd,umask=0x1,ldlat=3,precise=2");
+EVENT_ATTR_STR(mem-stores, mem_st_snb,
+		"event=0xcd,umask=0x2,precise=2");

Note that here while checkpatch.pl warns about an overlong line, it's pointless to break the line because the result is not improved. Just keep the line overlong in such cases.

checkpatch is a discretionary tool: if it warns then check the place, improve it checkpatch is right and if an improvement is possible - don't make the code harder to read just to placate the checkpatch warning.

arch/x86/kernel/cpu/perf_event_intel.c static __init void intel_nehalem_quirk(void) · line 2041

Quoted 8 lines
arch/x86/kernel/cpu/perf_event_intel.c +2 -2
@@ -2034,8 +2037,9 @@ static __init void intel_nehalem_quirk(void)
 	}
 }
 
-EVENT_ATTR_STR(mem-loads,      mem_ld_hsw,     "event=0xcd,umask=0x1,ldlat=3");
-EVENT_ATTR_STR(mem-stores,     mem_st_hsw,     "event=0xd0,umask=0x82")
+EVENT_ATTR_STR(mem-loads,      mem_ld_hsw,
+		"event=0xcd,umask=0x1,ldlat=3,precise=2");

Ditto.

Thanks,

Ingo

maintainer
#2 lore raw
In reply to Andi Kleen
Quoted 10 lines · Andi Kleen

From: Andi Kleen <[email protected]>

[The patch to enable this in the user tools has been sent separately]

With the earlier patches to automatically try cpu// and add a precise sys attribute, we can now enable PEBS for the mem-loads, mem-stores events everywhere.

This allows to use

Quoted 7 lines

perf record -e mem-loads ...

instead of

perf record -e cpu/mem-loads/p ...

Always use precise=2 even though it is costly pre-Haswell

This Changelog fails to give a reason _why_ we'd want to do this.

author
#3 lore raw
Quoted 13 lines · Peter Zijlstra
Quoted 12 lines · Andi Kleen

From: Andi Kleen <[email protected]>

[The patch to enable this in the user tools has been sent separately]

With the earlier patches to automatically try cpu// and add a precise sys attribute, we can now enable PEBS for the mem-loads, mem-stores events everywhere.

This allows to use

perf record -e mem-loads ...

Quoted 7 lines
Quoted 5 lines

instead of

perf record -e cpu/mem-loads/p ...

Always use precise=2 even though it is costly pre-Haswell

This Changelog fails to give a reason _why_ we'd want to do this.

The first is much nicer to type and understand? Just in the spirit of making perf easier to use.

-andi

Signature · 1 line
[email protected] -- Speaking for myself only.
maintainer
#4 lore raw
Quoted 16 lines · Andi Kleen
Quoted 15 lines · Peter Zijlstra
Quoted 14 lines · Andi Kleen

From: Andi Kleen <[email protected]>

[The patch to enable this in the user tools has been sent separately]

With the earlier patches to automatically try cpu// and add a precise sys attribute, we can now enable PEBS for the mem-loads, mem-stores events everywhere.

This allows to use

perf record -e mem-loads ...

instead of

Quoted 8 lines
Quoted 5 lines
Quoted 3 lines

perf record -e cpu/mem-loads/p ...

Always use precise=2 even though it is costly pre-Haswell

This Changelog fails to give a reason _why_ we'd want to do this.

The first is much nicer to type and understand? Just in the spirit of making perf easier to use.

And here I was thinking that maybe these events don't make sense without pebs or so. But no, rather than giving an actual useful reason you'd have me look things up myself. *sigh*

#5 lore raw
Quoted 21 lines · Peter Zijlstra
Quoted 20 lines · Andi Kleen
Quoted 19 lines · Peter Zijlstra
Quoted 17 lines · Andi Kleen

From: Andi Kleen <[email protected]>

[The patch to enable this in the user tools has been sent separately]

With the earlier patches to automatically try cpu// and add a precise sys attribute, we can now enable PEBS for the mem-loads, mem-stores events everywhere.

This allows to use

perf record -e mem-loads ...

instead of

perf record -e cpu/mem-loads/p ...

Always use precise=2 even though it is costly pre-Haswell

Quoted 8 lines
Quoted 4 lines
Quoted 1 line

This Changelog fails to give a reason _why_ we'd want to do this.

The first is much nicer to type and understand? Just in the spirit of making perf easier to use.

And here I was thinking that maybe these events don't make sense without pebs or so. But no, rather than giving an actual useful reason you'd have me look things up myself. *sigh*

The loads events using LATENCY_ABOVE_THRESHOLD do not count anything without PEBS (that's for all processors pre-Haswell).

As for forcing precise=2, I think that is what people would expect, i.e., point me
to the load/store instruction. Experts can still force precise=1 because I think the parser uses the value of the last precise= instance.

author
#6 lore raw
In reply to Peter Zijlstra
Quoted 21 lines · Peter Zijlstra
Quoted 20 lines · Andi Kleen
Quoted 19 lines · Peter Zijlstra
Quoted 17 lines · Andi Kleen

From: Andi Kleen <[email protected]>

[The patch to enable this in the user tools has been sent separately]

With the earlier patches to automatically try cpu// and add a precise sys attribute, we can now enable PEBS for the mem-loads, mem-stores events everywhere.

This allows to use

perf record -e mem-loads ...

instead of

perf record -e cpu/mem-loads/p ...

Always use precise=2 even though it is costly pre-Haswell

Quoted 8 lines
Quoted 4 lines
Quoted 1 line

This Changelog fails to give a reason _why_ we'd want to do this.

The first is much nicer to type and understand? Just in the spirit of making perf easier to use.

And here I was thinking that maybe these events don't make sense without pebs or so. But no, rather than giving an actual useful reason you'd have me look things up myself. *sigh*

You can use them without PEBS, but they're much more useful with PEBS (e.g. due to skid or if you want the address or data source, although you can also use perf mem for the later)

I think PEBS on is a better default here.

BTW I should add i have some other events (for HSW TSX) that would really like a PEBS default. That was the original motivation, but then I realized it makes other things nicer too.

-Andi

Signature · 1 line
[email protected] -- Speaking for myself only.