Skip to contentSkip to conversation

Attempt to cleanup the HSW offcore bits

PATCH In review Peter Zijlstra 6 participants 13 messages Oct 23, 2014 – Oct 29, 2014

Subsystems (2)
lore
Share & tools
lore

Peter Zijlstra

author maintainer
#0 lore raw

cover Attempt to cleanup the HSW offcore bits

So Don asked about offcore and because I forgot I looked at the code and found the terrible mess Andi created with the HSW/BDW bits.

This series attempts to clean some of that up but seeing how it was all magic numbers and no reasons provided for the differences with existing uarchs this might just break things horribly.

That said, if this does break things, fixes will have to explain things, so that's good.

author maintainer
#1 lore raw

PATCH 1/4 perf,x86: De-obfuscate HSW offcore bits

Andi introduced the HSW cache events array, but used magic constants against convention as set by all the other uarchs. Try and deobfuscate these a bit.

This patch should not change the actual values generated; however weird they seems.

In that patch Andi also said there were differences between the SNB/IVB and HSW/BDW offcore tables but failed to specify which and why.

Fixes 86a349a28b24 ("perf/x86/intel: Add Broadwell core support")
Cc Andi Kleen <[email protected]>
Cc Stephane Eranian <[email protected]>
---
 arch/x86/kernel/cpu/perf_event_intel.c |   48 +++++++++++++++++++++++++++------
 1 file changed, 40 insertions(+), 8 deletions(-)
arch/x86/kernel/cpu/perf_event_intel.c +40 -8
@@ -247,6 +247,7 @@ static u64 intel_pmu_event_map(int hw_ev
 #define SNB_BUS_LOCKS		(1ULL << 10)
 #define SNB_STRM_ST		(1ULL << 11)
 #define SNB_OTHER		(1ULL << 15)
+
 #define SNB_RESP_ANY		(1ULL << 16)
 #define SNB_NO_SUPP		(1ULL << 17)
 #define SNB_LLC_HITM		(1ULL << 18)
@@ -255,6 +256,7 @@ static u64 intel_pmu_event_map(int hw_ev
 #define SNB_LLC_HITF		(1ULL << 21)
 #define SNB_LOCAL		(1ULL << 22)
 #define SNB_REMOTE		(0xffULL << 23)
+
 #define SNB_SNP_NONE		(1ULL << 31)
 #define SNB_SNP_NOT_NEEDED	(1ULL << 32)
 #define SNB_SNP_MISS		(1ULL << 33)
@@ -519,6 +521,40 @@ static __initconst const u64 hsw_hw_cach
  },
 };
 
+/* HSW Request type */
+#define HSW_DMND_DATA_RD	(1ULL << 0)
+#define HSW_DMND_RFO		(1ULL << 1)
+#define HSW_DMND_IFETCH		(1ULL << 2)
+
+#define HSW_PF_DATA_RD		(1ULL << 4)
+#define HSW_PF_RFO		(1ULL << 5)
+#define HSW_PF_IFETCH		(1ULL << 6)
+
+#define HSW_OTHER		(1ULL << 15)
+
+/* HSW Supplier info := SNB Supplier info */
+/* HSW Snoop Info := SNB Snoop Info */
+
+/*
+ * SNB_LLC_* is specified as 'reserved' in the SDM
+ * but Andi added it anyhow, suggesting HSW/BDW have it just fine
+ *
+ * XXX NHM/WSM/SNB/IVB don't have the PF/IFETCH bits set
+ */
+#define HSW_DMND_READ		(HSW_DMND_DATA_RD|HSW_DMND_IFETCH| \
+				 HSW_PF_DATA_RD|HSW_PF_IFETCH| \
+				 SNB_LLC_DATA_RD|SNB_LLC_IFETCH)
+
+/* XXX NHM/WSM/SNB/IVB dont have the PF bits set */
+#define HSW_DMND_WRITE		(HSW_DMND_RFO|HSW_PF_RFO|SNB_LLC_RFO)
+
+#define HSW_DMND_PREFETCH	(0) /* XXX broken ? */
+
+#define HSW_DRAM_ANY		(SNB_NO_SUPP|SNB_SNP_ANY|(0x78ULL << 23)) /* WTF */
+
+#define HSW_L3_ACCESS		(0) /* XXX no supplier! */
+#define NSW_L3_MISS		(HSW_DRAM_ANY|SNB_NON_DRAM)
+
 static __initconst const u64 hsw_hw_cache_extra_regs
 				[PERF_COUNT_HW_CACHE_MAX]
 				[PERF_COUNT_HW_CACHE_OP_MAX]
@@ -526,16 +562,12 @@ static __initconst const u64 hsw_hw_cach
 {
  [ C(LL  ) ] = {
 	[ C(OP_READ) ] = {
-		/* OFFCORE_RESPONSE:ALL_DATA_RD|ALL_CODE_RD */
-		[ C(RESULT_ACCESS) ] = 0x2d5,
-		/* OFFCORE_RESPONSE:ALL_DATA_RD|ALL_CODE_RD|SUPPLIER_NONE|
-                   L3_MISS|ANY_SNOOP */
-		[ C(RESULT_MISS)   ] = 0x3fbc0202d5ull,
+		[ C(RESULT_ACCESS) ] = HSW_DMND_READ|HSW_L3_ACCESS,
+		[ C(RESULT_MISS)   ] = HSW_DMND_READ|HSW_L3_MISS,
 	},
 	[ C(OP_WRITE) ] = {
-		[ C(RESULT_ACCESS) ] = 0x122, 	/* OFFCORE_RESPONSE:ALL_RFO */
-		/* OFFCORE_RESPONSE:ALL_RFO|SUPPLIER_NONE|L3_MISS|ANY_SNOOP */
-		[ C(RESULT_MISS)   ] = 0x3fbc020122ull,
+		[ C(RESULT_ACCESS) ] = HSW_DMND_WRITE|HSW_L3_ACCESS,
+		[ C(RESULT_MISS)   ] = HSW_DMND_WRITE|HSW_L3_MISS,
 	},
 	[ C(OP_PREFETCH) ] = {
 		[ C(RESULT_ACCESS) ] = 0x0,
maintainer
#2 lore raw

Re: [PATCH 1/4] perf,x86: De-obfuscate HSW offcore bits

Quoted 11 lines · Peter Zijlstra

Andi introduced the HSW cache events array, but used magic constants against convention as set by all the other uarchs. Try and deobfuscate these a bit.

This patch should not change the actual values generated; however weird they seems.

In that patch Andi also said there were differences between the SNB/IVB and HSW/BDW offcore tables but failed to specify which and why.

Quoted 5 lines
Fixes 86a349a28b24 ("perf/x86/intel: Add Broadwell core support")
Cc Andi Kleen <[email protected]>
Cc Stephane Eranian <[email protected]>

This link looks funny.

Quoted 49 lines
---
 arch/x86/kernel/cpu/perf_event_intel.c |   48 +++++++++++++++++++++++++++------
 1 file changed, 40 insertions(+), 8 deletions(-)
arch/x86/kernel/cpu/perf_event_intel.c +25 -0
@@ -247,6 +247,7 @@ static u64 intel_pmu_event_map(int hw_ev
 #define SNB_BUS_LOCKS		(1ULL << 10)
 #define SNB_STRM_ST		(1ULL << 11)
 #define SNB_OTHER		(1ULL << 15)
+
 #define SNB_RESP_ANY		(1ULL << 16)
 #define SNB_NO_SUPP		(1ULL << 17)
 #define SNB_LLC_HITM		(1ULL << 18)
@@ -255,6 +256,7 @@ static u64 intel_pmu_event_map(int hw_ev
 #define SNB_LLC_HITF		(1ULL << 21)
 #define SNB_LOCAL		(1ULL << 22)
 #define SNB_REMOTE		(0xffULL << 23)
+
 #define SNB_SNP_NONE		(1ULL << 31)
 #define SNB_SNP_NOT_NEEDED	(1ULL << 32)
 #define SNB_SNP_MISS		(1ULL << 33)
@@ -519,6 +521,40 @@ static __initconst const u64 hsw_hw_cach
  },
 };
 
+/* HSW Request type */
+#define HSW_DMND_DATA_RD	(1ULL << 0)
+#define HSW_DMND_RFO		(1ULL << 1)
+#define HSW_DMND_IFETCH		(1ULL << 2)
+
+#define HSW_PF_DATA_RD		(1ULL << 4)
+#define HSW_PF_RFO		(1ULL << 5)
+#define HSW_PF_IFETCH		(1ULL << 6)
+
+#define HSW_OTHER		(1ULL << 15)
+
+/* HSW Supplier info := SNB Supplier info */
+/* HSW Snoop Info := SNB Snoop Info */
+
+/*
+ * SNB_LLC_* is specified as 'reserved' in the SDM
+ * but Andi added it anyhow, suggesting HSW/BDW have it just fine
+ *
+ * XXX NHM/WSM/SNB/IVB don't have the PF/IFETCH bits set
+ */
+#define HSW_DMND_READ		(HSW_DMND_DATA_RD|HSW_DMND_IFETCH| \
+				 HSW_PF_DATA_RD|HSW_PF_IFETCH| \
+				 SNB_LLC_DATA_RD|SNB_LLC_IFETCH)

arch/x86/kernel/cpu/perf_event_intel.c static __initconst const u64 hsw_hw_cach · line 556

Quoted 10 lines
arch/x86/kernel/cpu/perf_event_intel.c +10 -0
+
+/* XXX NHM/WSM/SNB/IVB dont have the PF bits set */
+#define HSW_DMND_WRITE		(HSW_DMND_RFO|HSW_PF_RFO|SNB_LLC_RFO)
+
+#define HSW_DMND_PREFETCH	(0) /* XXX broken ? */
+
+#define HSW_DRAM_ANY		(SNB_NO_SUPP|SNB_SNP_ANY|(0x78ULL << 23)) /* WTF */
+
+#define HSW_L3_ACCESS		(0) /* XXX no supplier! */
+#define NSW_L3_MISS		(HSW_DRAM_ANY|SNB_NON_DRAM)

You probably mean HSW and not NSW here...

Signature · 5 lines
Regards/Gruss,
    Boris.

Sent from a fat crate under my desk. Formatting is fine.
--
author maintainer
#3 lore raw

Re: [PATCH 1/4] perf,x86: De-obfuscate HSW offcore bits

Quoted 12 lines · Borislav Petkov
Quoted 11 lines · Peter Zijlstra

Andi introduced the HSW cache events array, but used magic constants against convention as set by all the other uarchs. Try and deobfuscate these a bit.

This patch should not change the actual values generated; however weird they seems.

In that patch Andi also said there were differences between the SNB/IVB and HSW/BDW offcore tables but failed to specify which and why.

Quoted 7 lines
Quoted 5 lines
Fixes 86a349a28b24 ("perf/x86/intel: Add Broadwell core support")
Cc Andi Kleen <[email protected]>
Cc Stephane Eranian <[email protected]>

This link looks funny.

Yes it does, I seem to have not isntalled rndpwd or whatever that script is using. Also I usually grep -v "^Link:" .mbox files before sending, guess what I forgot :-)

author maintainer
#4 lore raw

PATCH 2/4 perf,x86: HSW offcore prefetch events

In reply to Peter Zijlstra

This changes the HSW events to be more inline with the other uarchs and removes the prefetch request bits from the read/write demands and into the prefetch demand.

---
 arch/x86/kernel/cpu/perf_event_intel.c |   14 +++++---------
 1 file changed, 5 insertions(+), 9 deletions(-)
arch/x86/kernel/cpu/perf_event_intel.c +5 -9
@@ -539,16 +539,12 @@ static __initconst const u64 hsw_hw_cach
  * SNB_LLC_* is specified as 'reserved' in the SDM
  * but Andi added it anyhow, suggesting HSW/BDW have it just fine
  *
- * XXX NHM/WSM/SNB/IVB don't have the PF/IFETCH bits set
+ * XXX NHM/WSM/SNB/IVB don't have the IFETCH bits set
  */
 #define HSW_DMND_READ		(HSW_DMND_DATA_RD|HSW_DMND_IFETCH| \
-				 HSW_PF_DATA_RD|HSW_PF_IFETCH| \
 				 SNB_LLC_DATA_RD|SNB_LLC_IFETCH)
-
-/* XXX NHM/WSM/SNB/IVB dont have the PF bits set */
-#define HSW_DMND_WRITE		(HSW_DMND_RFO|HSW_PF_RFO|SNB_LLC_RFO)
-
-#define HSW_DMND_PREFETCH	(0) /* XXX broken ? */
+#define HSW_DMND_WRITE		(HSW_DMND_RFO|SNB_LLC_RFO)
+#define HSW_DMND_PREFETCH	(HSW_PF_DATA_RD|HSW_PF_RFO|HSW_PF_IFETCH)
 
 #define HSW_DRAM_ANY		(SNB_NO_SUPP|SNB_SNP_ANY|(0x78ULL << 23)) /* WTF */
 
@@ -570,8 +566,8 @@ static __initconst const u64 hsw_hw_cach
 		[ C(RESULT_MISS)   ] = HSW_DMND_WRITE|HSW_L3_MISS,
 	},
 	[ C(OP_PREFETCH) ] = {
-		[ C(RESULT_ACCESS) ] = 0x0,
-		[ C(RESULT_MISS)   ] = 0x0,
+		[ C(RESULT_ACCESS) ] = HSW_DMND_PREFETCH|HSW_L3_ACCESS,
+		[ C(RESULT_MISS)   ] = HSW_DMND_PREFETCH|HSW_L3_MISS,
 	},
  },
 };
author maintainer
#5 lore raw

PATCH 3/4 perf,x86: Attempt to sanitize the HSW supplier info

In reply to Peter Zijlstra

The SDM states the HSW/BDW supplier and snoop info are identical to SNB/IVB, make it so.

Furthermore, it states you have to minimally program a request and supplier type, but the current code does not set a supplier for the ACCESS events.

This significantly alters the actual events and does away with some of the weirdness found by decoding the magic numbers. If this breaks things someone needs to go explain things and augment the SDM.

---
 arch/x86/kernel/cpu/perf_event_intel.c |   17 ++++++-----------
 1 file changed, 6 insertions(+), 11 deletions(-)
arch/x86/kernel/cpu/perf_event_intel.c +6 -11
@@ -544,11 +544,6 @@ static __initconst const u64 hsw_hw_cach
 #define HSW_DMND_WRITE		(HSW_DMND_RFO|SNB_LLC_RFO)
 #define HSW_DMND_PREFETCH	(HSW_PF_DATA_RD|HSW_PF_RFO|HSW_PF_IFETCH)
 
-#define HSW_DRAM_ANY		(SNB_NO_SUPP|SNB_SNP_ANY|(0x78ULL << 23)) /* WTF */
-
-#define HSW_L3_ACCESS		(0) /* XXX no supplier! */
-#define NSW_L3_MISS		(HSW_DRAM_ANY|SNB_NON_DRAM)
-
 static __initconst const u64 hsw_hw_cache_extra_regs
 				[PERF_COUNT_HW_CACHE_MAX]
 				[PERF_COUNT_HW_CACHE_OP_MAX]
@@ -556,16 +551,16 @@ static __initconst const u64 hsw_hw_cach
 {
  [ C(LL  ) ] = {
 	[ C(OP_READ) ] = {
-		[ C(RESULT_ACCESS) ] = HSW_DMND_READ|HSW_L3_ACCESS,
-		[ C(RESULT_MISS)   ] = HSW_DMND_READ|HSW_L3_MISS,
+		[ C(RESULT_ACCESS) ] = HSW_DMND_READ|SNB_L3_ACCESS,
+		[ C(RESULT_MISS)   ] = HSW_DMND_READ|SNB_L3_MISS,
 	},
 	[ C(OP_WRITE) ] = {
-		[ C(RESULT_ACCESS) ] = HSW_DMND_WRITE|HSW_L3_ACCESS,
-		[ C(RESULT_MISS)   ] = HSW_DMND_WRITE|HSW_L3_MISS,
+		[ C(RESULT_ACCESS) ] = HSW_DMND_WRITE|SNB_L3_ACCESS,
+		[ C(RESULT_MISS)   ] = HSW_DMND_WRITE|SNB_L3_MISS,
 	},
 	[ C(OP_PREFETCH) ] = {
-		[ C(RESULT_ACCESS) ] = HSW_DMND_PREFETCH|HSW_L3_ACCESS,
-		[ C(RESULT_MISS)   ] = HSW_DMND_PREFETCH|HSW_L3_MISS,
+		[ C(RESULT_ACCESS) ] = HSW_DMND_PREFETCH|SNB_L3_ACCESS,
+		[ C(RESULT_MISS)   ] = HSW_DMND_PREFETCH|SNB_L3_MISS,
 	},
  },
 };
author maintainer
#6 lore raw

PATCH 4/4 perf,x86: Introduce HSW cache numa events

In reply to Peter Zijlstra

Seeing how HSW-EP is now available and the SDM states the supplier and snoop info is identical to SNB/IVB, provide the cache numa events.

---
 arch/x86/kernel/cpu/perf_event_intel.c |   14 ++++++++++++++
 1 file changed, 14 insertions(+)
arch/x86/kernel/cpu/perf_event_intel.c +14 -0
@@ -563,6 +563,20 @@ static __initconst const u64 hsw_hw_cach
 		[ C(RESULT_MISS)   ] = HSW_DMND_PREFETCH|SNB_L3_MISS,
 	},
  },
+ [ C(NODE) ] = {
+	[ C(OP_READ) ] = {
+		[ C(RESULT_ACCESS) ] = HSW_DMND_READ|SNB_DRAM_ANY,
+		[ C(RESULT_MISS)   ] = HSW_DMND_READ|SNB_DRAM_REMOTE,
+	},
+	[ C(OP_WRITE) ] = {
+		[ C(RESULT_ACCESS) ] = HSW_DMND_WRITE|SNB_DRAM_ANY,
+		[ C(RESULT_MISS)   ] = HSW_DMND_WRITE|SNB_DRAM_REMOTE,
+	},
+	[ C(OP_PREFETCH) ] = {
+		[ C(RESULT_ACCESS) ] = HSW_DMND_PREFETCH|SNB_DRAM_ANY,
+		[ C(RESULT_MISS)   ] = HSW_DMND_PREFETCH|SNB_DRAM_REMOTE,
+	},
+ },
 };
 
 static __initconst const u64 westmere_hw_cache_event_ids
#7 lore raw
In reply to Peter Zijlstra
Quoted 5 lines · Peter Zijlstra

So Don asked about offcore and because I forgot I looked at the code and found the terrible mess Andi created with the HSW/BDW bits.

This series attempts to clean some of that up but seeing how it was all magic numbers

All the bits are documented. The actual definitions are available in the JSON offcore definitions at https://download.01.org/perfmon/

Quoted 4 lines

and no reasons provided for the differences with existing uarchs this might just break things horribly.
That said, if this does break things, fixes will have to explain things, so that's good.

I trust you won't submit or merge untested patches.

-Andi

Signature · 1 line
[email protected] -- Speaking for myself only.
author maintainer
#8 lore raw
Quoted 9 lines · Andi Kleen
Quoted 5 lines · Peter Zijlstra

So Don asked about offcore and because I forgot I looked at the code and found the terrible mess Andi created with the HSW/BDW bits.

This series attempts to clean some of that up but seeing how it was all magic numbers

All the bits are documented. The actual definitions are available in the JSON offcore definitions at https://download.01.org/perfmon/

Yeah, no. That's not how we write code. Also, there's no actual JSON offcore file for HSW only some TSV file, and I've no mind to go decode and reverse engineer that stuff.

Also, semi unreadable files on a weird web location is not documentation, the SDM is and the SDM does not explain why you didn't put the PF request bits in the OP_PREFETCH events, nor why you added the IFETCH bits to the OP_READ where all the other uarchs didn't.

Nor does it explain why you set bits 7,9 in OP_READ even though they're listed as 'reserved'.

It also doesn't explain why you don't program a Supplier Info for RESULT_ACCESS.

Nor why you set 0x78 in bits 23:30, which even for SNB are still listed as reserved, even though its official offcore events have that set to 0xFF for all remote events.

And lacking SDM you should have explained all that in your Changelog, but that's also entirely devoid of any usable information.

Quoted 6 lines
Quoted 4 lines

and no reasons provided for the differences with existing uarchs this might just break things horribly.
That said, if this does break things, fixes will have to explain things, so that's good.

I trust you won't submit or merge untested patches.

You trust wrong.. I should never have merged your patches, but seeing they're in it needed cleaning up.

I can revert your patches if you'd rather have that?

#9 lore raw
Quoted 5 lines · Peter Zijlstra
Quoted 4 lines · Andi Kleen
Quoted 3 lines · Peter Zijlstra

So Don asked about offcore and because I forgot I looked at the code and found the terrible mess Andi created with the HSW/BDW bits.

Quoted 8 lines
Quoted 5 lines
Quoted 2 lines

This series attempts to clean some of that up but seeing how it was all magic numbers

All the bits are documented. The actual definitions are available in the JSON offcore definitions at https://download.01.org/perfmon/

Yeah, no. That's not how we write code. Also, there's no actual JSON offcore file for HSW only some TSV file, and I've no mind to go decode

https://download.01.org/perfmon/HSW/Haswell_matrix_V14.json
https://download.01.org/perfmon/HSW/Haswell_matrix_bit_definitions_V14.json

-Andi

author maintainer
#10 lore raw
Quoted 9 lines · Andi Kleen
Quoted 8 lines · Peter Zijlstra
Quoted 7 lines · Andi Kleen
Quoted 5 lines · Peter Zijlstra

So Don asked about offcore and because I forgot I looked at the code and found the terrible mess Andi created with the HSW/BDW bits.

This series attempts to clean some of that up but seeing how it was all magic numbers

Quoted 8 lines
Quoted 5 lines
Quoted 2 lines

All the bits are documented. The actual definitions are available in the JSON offcore definitions at https://download.01.org/perfmon/

Yeah, no. That's not how we write code. Also, there's no actual JSON offcore file for HSW only some TSV file, and I've no mind to go decode

https://download.01.org/perfmon/HSW/Haswell_matrix_V14.json
https://download.01.org/perfmon/HSW/Haswell_matrix_bit_definitions_V14.json

So you expected me to somehow magically know it was not the file with offcore in its name then?

#11 lore raw
In reply to Andi Kleen

Andi,

Quoted 9 lines · Andi Kleen
Quoted 8 lines · Peter Zijlstra
Quoted 7 lines · Andi Kleen
Quoted 5 lines · Peter Zijlstra

So Don asked about offcore and because I forgot I looked at the code and found the terrible mess Andi created with the HSW/BDW bits.

This series attempts to clean some of that up but seeing how it was all magic numbers

Quoted 8 lines
Quoted 5 lines
Quoted 2 lines

All the bits are documented. The actual definitions are available in the JSON offcore definitions at https://download.01.org/perfmon/

Yeah, no. That's not how we write code. Also, there's no actual JSON offcore file for HSW only some TSV file, and I've no mind to go decode

https://download.01.org/perfmon/HSW/Haswell_matrix_V14.json
https://download.01.org/perfmon/HSW/Haswell_matrix_bit_definitions_V14.json

Of course you did not answer any of the other legitimate questions Peter brought up.

Care to answer them w/o pointing to magic json files which lack ANY useful information about the magic bits they provide? And why they are not consistent with the SDM?

Either you come forth with reasonable explanations or I'm going to rip out the mess you created even before Peter can persuade himself to do so.

Thanks,

tglx

#12 lore raw
Quoted 24 lines · Thomas Gleixner

Andi,

Quoted 9 lines · Andi Kleen
Quoted 8 lines · Peter Zijlstra
Quoted 7 lines · Andi Kleen
Quoted 5 lines · Peter Zijlstra

So Don asked about offcore and because I forgot I looked at the code and found the terrible mess Andi created with the HSW/BDW bits.

This series attempts to clean some of that up but seeing how it was all magic numbers

Quoted 8 lines
Quoted 5 lines
Quoted 2 lines

All the bits are documented. The actual definitions are available in the JSON offcore definitions at https://download.01.org/perfmon/

Yeah, no. That's not how we write code. Also, there's no actual JSON offcore file for HSW only some TSV file, and I've no mind to go decode

https://download.01.org/perfmon/HSW/Haswell_matrix_V14.json
https://download.01.org/perfmon/HSW/Haswell_matrix_bit_definitions_V14.json

Of course you did not answer any of the other legitimate questions Peter brought up.

Quoted 7 lines

Care to answer them w/o pointing to magic json files which lack ANY useful information about the magic bits they provide? And why they are not consistent with the SDM?

Either you come forth with reasonable explanations or I'm going to rip out the mess you created even before Peter can persuade himself to do so.

So I waited a week in the hope that all this can be resolved quickly, but the problem is that we are already at -rc3 and we are running out of time and I'm not seeing constructive behavior from Andi.

So I've done a straightforward revert of the messy Broadwell client driver changes, to bring us back to the v3.17 baseline:

1776b10627e4 ("perf/x86/intel: Revert incomplete and undocumented Broadwell client support")

Andi, please rework the series on top of that and also proactively integrate the cleanups, requests and observations from Thomas Gleixner and Peter Zijlstra. Only submit a new series once all feedback given to you in the past has been addressed to the fullest!

If the patches are fixed quickly enough then we might be able to merge support in v3.19. If not then it has to wait until v3.20 or later kernels.

Thanks,

Ingo