Skip to contentSkip to conversation

i386 add Intel PEBS and BTS cpufeature bits and detection

PATCH In review Stephane Eranian 5 participants 15 messages Nov 15, 2006 – Nov 23, 2006

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 i386 add Intel PEBS and BTS cpufeature bits and detection

Hello,

Here is a small patch that adds two cpufeature bits to represent Intel's Precise-Event-Based Sampling (PEBS) and Branch Trace Store (BTS) features. Those features can be found on Intel P4 and Core 2 processors among others and can be used by perfmon.

changelog:
- add CPU_FEATURE_PEBS and CPU_FEATURE_BTS
- add feature detection code

arch/i386/kernel/cpu/intel.c +9 -1
diff --exclude=.git -urNp linux-2.6.orig/arch/i386/kernel/cpu/intel.c linux-2.6.base/arch/i386/kernel/cpu/intel.c
@@ -97,7 +97,7 @@ static int __cpuinit num_cpu_cores(struc
 
 static void __cpuinit init_intel(struct cpuinfo_x86 *c)
 {
-	unsigned int l2 = 0;
+	unsigned int l1, l2 = 0;
 	char *p = NULL;
 
 #ifdef CONFIG_X86_F00F_BUG
@@ -195,6 +195,14 @@ static void __cpuinit init_intel(struct 
 	if ((c->x86 == 0xf && c->x86_model >= 0x03) ||
 		(c->x86 == 0x6 && c->x86_model >= 0x0e))
 		set_bit(X86_FEATURE_CONSTANT_TSC, c->x86_capability);
+
+	if (cpu_has_ds) {
+		rdmsr(MSR_IA32_MISC_ENABLE, l1, l2);
+		if (!(l1 & (1<<11)))
+			set_bit(X86_FEATURE_BTS, c->x86_capability);
+		if (!(l1 & (1<<12)))
+			set_bit(X86_FEATURE_PEBS, c->x86_capability);
+	}
 }
 
include/asm-i386/cpufeature.h +5 -0
diff --exclude=.git -urNp linux-2.6.orig/include/asm-i386/cpufeature.h linux-2.6.base/include/asm-i386/cpufeature.h
@@ -73,6 +73,8 @@
 #define X86_FEATURE_UP		(3*32+ 9) /* smp kernel running on up */
 #define X86_FEATURE_FXSAVE_LEAK (3*32+10) /* FXSAVE leaks FOP/FIP/FOP */
 #define X86_FEATURE_ARCH_PERFMON (3*32+11) /* Intel Architectural PerfMon */
+#define X86_FEATURE_PEBS	(3*32+12)  /* Precise-Event Based Sampling */
+#define X86_FEATURE_BTS		(3*32+13)  /* Branch Trace Store */
 
 /* Intel-defined CPU features, CPUID level 0x00000001 (ecx), word 4 */
 #define X86_FEATURE_XMM3	(4*32+ 0) /* Streaming SIMD Extensions-3 */
@@ -134,6 +136,9 @@
 #define cpu_has_phe_enabled	boot_cpu_has(X86_FEATURE_PHE_EN)
 #define cpu_has_pmm		boot_cpu_has(X86_FEATURE_PMM)
 #define cpu_has_pmm_enabled	boot_cpu_has(X86_FEATURE_PMM_EN)
+#define cpu_has_ds 		boot_cpu_has(X86_FEATURE_DTES)
+#define cpu_has_pebs 		boot_cpu_has(X86_FEATURE_PEBS)
+#define cpu_has_bts		boot_cpu_has(X86_FEATURE_BTS)
 
 #endif /* __ASM_I386_CPUFEATURE_H */
 
#1 lore raw

PATCH 1/1 x86-64 add Intel PEBS and BTS cpufeature bits and detection

Hello,

Here is a small patch that adds two cpufeature bits to represent Intel's Precise-Event-Based Sampling (PEBS) and Branch Trace Store (BTS) features. Those features can be found on Intel P4 and Core 2 processors among others and can be used by perfmon. This patch is for x86-64.

changelog:
- add CPU_FEATURE_PEBS and CPU_FEATURE_BTS
- add feature detection code

include/asm-x86_64/cpufeature.h +5 -0
diff --exclude=.git -urNp linux-2.6.orig/include/asm-x86_64/cpufeature.h linux-2.6.base/include/asm-x86_64/cpufeature.h
@@ -68,6 +68,8 @@
 #define X86_FEATURE_FXSAVE_LEAK (3*32+7)  /* FIP/FOP/FDP leaks through FXSAVE */
 #define X86_FEATURE_UP		(3*32+8) /* SMP kernel running on UP */
 #define X86_FEATURE_ARCH_PERFMON (3*32+9) /* Intel Architectural PerfMon */
+#define X86_FEATURE_PEBS	(3*32+10) /* Precise-Event Based Sampling */
+#define X86_FEATURE_BTS		(3*32+11) /* Branch Trace Store */
 
 /* Intel-defined CPU features, CPUID level 0x00000001 (ecx), word 4 */
 #define X86_FEATURE_XMM3	(4*32+ 0) /* Streaming SIMD Extensions-3 */
@@ -112,5 +114,8 @@
 #define cpu_has_cyrix_arr      0
 #define cpu_has_centaur_mcr    0
 #define cpu_has_clflush	       boot_cpu_has(X86_FEATURE_CLFLSH)
+#define cpu_has_ds 	       boot_cpu_has(X86_FEATURE_DTES)
+#define cpu_has_pebs 	       boot_cpu_has(X86_FEATURE_PEBS)
+#define cpu_has_bts 	       boot_cpu_has(X86_FEATURE_BTS)
 
 #endif /* __ASM_X8664_CPUFEATURE_H */
arch/x86_64/kernel/setup.c +9 -0
diff --exclude=.git -urNp linux-2.6.orig/arch/x86_64/kernel/setup.c linux-2.6.base/arch/x86_64/kernel/setup.c
@@ -835,6 +835,15 @@ static void __cpuinit init_intel(struct 
 			set_bit(X86_FEATURE_ARCH_PERFMON, &c->x86_capability);
 	}
 
+	if (cpu_has_ds) {
+		unsigned int l1, l2;
+		rdmsr(MSR_IA32_MISC_ENABLE, l1, l2);
+		if (!(l1 & (1<<11)))
+			set_bit(X86_FEATURE_BTS, c->x86_capability);
+		if (!(l1 & (1<<12)))
+			set_bit(X86_FEATURE_PEBS, c->x86_capability);
+	}
+
 	n = c->extended_cpuid_level;
 	if (n >= 0x80000008) {
 		unsigned eax = cpuid_eax(0x80000008);
#2 lore raw
In reply to Stephane Eranian
Quoted 5 lines · On Wed, 15 Nov 2006 13:32:41 -0800 Stephane Eranian <[email protected]> wrote:

Here is a small patch that adds two cpufeature bits to represent Intel's Precise-Event-Based Sampling (PEBS) and Branch Trace Store (BTS) features. Those features can be found on Intel P4 and Core 2 processors among others and can be used by perfmon.

Andi has already merged a different version of this. If it needs updating, please review his tree (most recent -mm will suit) and send any needed updates.

#3 lore raw

PATCH 1/1 x86-64 add Intel BTS cpufeature bit and detection (take 2)

In reply to Stephane Eranian

Andi,

Here is a small patch for x86-64 which adds a cpufeature flag and detection code for Intel's Branch Trace Store (BTS) feature. This feature can be found on Intel P4 and Core 2 processors among others. It can also be used by perfmon.

The patch is relative to 2.6.19-rc5-git7 + x86_64-2.6.19-rc5-git7-061116-2

changelog:
- add CPU_FEATURE_BTS
- add Branch Trace Store detection

arch/x86_64/kernel/setup.c +2 -0
diff --exclude=.git -urNp linux-2.6.19-rc5-git7-ak.orig/arch/x86_64/kernel/setup.c linux-2.6.19-rc5-git7-ak/arch/x86_64/kernel/setup.c
@@ -838,6 +838,8 @@ static void __cpuinit init_intel(struct 
 	if (cpu_has_ds) {
 		unsigned int l1, l2;
 		rdmsr(MSR_IA32_MISC_ENABLE, l1, l2);
+		if (!(l1 & (1<<11)))
+			set_bit(X86_FEATURE_BTS, c->x86_capability);
 		if (!(l1 & (1<<12)))
 			set_bit(X86_FEATURE_PEBS, c->x86_capability);
 	}
include/asm-x86_64/cpufeature.h +2 -0
diff --exclude=.git -urNp linux-2.6.19-rc5-git7-ak.orig/include/asm-x86_64/cpufeature.h linux-2.6.19-rc5-git7-ak/include/asm-x86_64/cpufeature.h
@@ -69,6 +69,7 @@
 #define X86_FEATURE_UP		(3*32+8) /* SMP kernel running on UP */
 #define X86_FEATURE_ARCH_PERFMON (3*32+9) /* Intel Architectural PerfMon */
 #define X86_FEATURE_PEBS	(3*32+10) /* Precise-Event Based Sampling */
+#define X86_FEATURE_BTS		(3*32+11) /* Branch Trace Store */
 
 /* Intel-defined CPU features, CPUID level 0x00000001 (ecx), word 4 */
 #define X86_FEATURE_XMM3	(4*32+ 0) /* Streaming SIMD Extensions-3 */
@@ -115,5 +116,6 @@
 #define cpu_has_clflush	       boot_cpu_has(X86_FEATURE_CLFLSH)
 #define cpu_has_ds 	       boot_cpu_has(X86_FEATURE_DS)
 #define cpu_has_pebs 	       boot_cpu_has(X86_FEATURE_PEBS)
+#define cpu_has_bts 	       boot_cpu_has(X86_FEATURE_BTS)
 
 #endif /* __ASM_X8664_CPUFEATURE_H */
#4 lore raw

PATCH 1/1 i386 add Intel BTS cpufeature bit and detection (take 2)

Andi,

Here is a small patch for i386 which adds a cpufeature flag and detection code for Intel's Branch Trace Store (BTS) feature. This feature can be found on Intel P4 and Core 2 processors among others. It can also be used by perfmon.

The patch is relative to 2.6.19-rc5-git7 + x86_64-2.6.19-rc5-git7-061116-2

changelog:
- add CPU_FEATURE_BTS
- add Branch Trace Store detection

arch/i386/kernel/cpu/intel.c +2 -0
diff --exclude=.git -urNp linux-2.6.19-rc5-git7-ak.orig/arch/i386/kernel/cpu/intel.c linux-2.6.19-rc5-git7-ak/arch/i386/kernel/cpu/intel.c
@@ -199,6 +199,8 @@ static void __cpuinit init_intel(struct 
 	if (cpu_has_ds) {
 		unsigned int l1;
 		rdmsr(MSR_IA32_MISC_ENABLE, l1, l2);
+		if (!(l1 & (1<<11)))
+			set_bit(X86_FEATURE_BTS, c->x86_capability);
 		if (!(l1 & (1<<12)))
 			set_bit(X86_FEATURE_PEBS, c->x86_capability);
 	}
include/asm-i386/cpufeature.h +2 -0
diff --exclude=.git -urNp linux-2.6.19-rc5-git7-ak.orig/include/asm-i386/cpufeature.h linux-2.6.19-rc5-git7-ak/include/asm-i386/cpufeature.h
@@ -74,6 +74,7 @@
 #define X86_FEATURE_FXSAVE_LEAK (3*32+10) /* FXSAVE leaks FOP/FIP/FOP */
 #define X86_FEATURE_ARCH_PERFMON (3*32+11) /* Intel Architectural PerfMon */
 #define X86_FEATURE_PEBS	(3*32+12)  /* Precise-Event Based Sampling */
+#define X86_FEATURE_BTS		(3*32+13)  /* Branch Trace Store */
 
 /* Intel-defined CPU features, CPUID level 0x00000001 (ecx), word 4 */
 #define X86_FEATURE_XMM3	(4*32+ 0) /* Streaming SIMD Extensions-3 */
@@ -138,6 +139,7 @@
 #define cpu_has_ds		boot_cpu_has(X86_FEATURE_DS)
 #define cpu_has_pebs 		boot_cpu_has(X86_FEATURE_PEBS)
 #define cpu_has_clflush		boot_cpu_has(X86_FEATURE_CLFLSH)
+#define cpu_has_bts 		boot_cpu_has(X86_FEATURE_BTS)
 
 #endif /* __ASM_I386_CPUFEATURE_H */
 
#5 lore raw

Re: [PATCH] x86-64 add Intel BTS cpufeature bit and detection (take 2)

In reply to Stephane Eranian
Quoted 6 lines · Stephane Eranian

Andi,

Here is a small patch for x86-64 which adds a cpufeature flag and detection code for Intel's Branch Trace Store (BTS) feature. This feature can be found on Intel P4 and Core 2 processors among others. It can also be used by perfmon.

Added thanks

-Andi

#6 lore raw
In reply to Stephane Eranian
Quoted 5 lines · Stephane Eranian

Here is a small patch that adds two cpufeature bits to represent Intel's Precise-Event-Based Sampling (PEBS) and Branch Trace Store (BTS) features. Those features can be found on Intel P4 and Core 2 processors among others and can be used by perfmon.

I've been thinking it would be useful for kernel debugging if kernel oops messages could use the branch history to show the last few jumps on processors which support it. It would help a lot with the "oh, an oops with eip==esp==0" type crashes, which are otherwise pretty unhelpful.

Do you think that would be easy/possible to support? Would it interfere with other uses of these features?

J

#7 lore raw
Quoted 11 lines · Jeremy Fitzhardinge
Quoted 5 lines · Stephane Eranian

Here is a small patch that adds two cpufeature bits to represent Intel's Precise-Event-Based Sampling (PEBS) and Branch Trace Store (BTS) features. Those features can be found on Intel P4 and Core 2 processors among others and can be used by perfmon.

I've been thinking it would be useful for kernel debugging if kernel oops messages could use the branch history to show the last few jumps on processors which support it. It would help a lot with the "oh, an oops with eip==esp==0" type crashes, which are otherwise pretty unhelpful.

I have had private patches for that myself, using the MSRs on AMD and Intel.

The problem is that you have to insert hooks early into the exception handlers to read the branch history MSRs, and that gets fairly ugly and a little slow and we can't really enable it by default.

But using BTS with a long in memory buffer would be fine. It would just be slower so it couldn't be enabled by default. But as a debugging feature it would be nice.

-Andi

#8 lore raw

Jeremy,

Quoted 13 lines · Andi Kleen
Quoted 11 lines · Jeremy Fitzhardinge
Quoted 5 lines · Stephane Eranian

Here is a small patch that adds two cpufeature bits to represent Intel's Precise-Event-Based Sampling (PEBS) and Branch Trace Store (BTS) features. Those features can be found on Intel P4 and Core 2 processors among others and can be used by perfmon.

I've been thinking it would be useful for kernel debugging if kernel oops messages could use the branch history to show the last few jumps on processors which support it. It would help a lot with the "oh, an oops with eip==esp==0" type crashes, which are otherwise pretty unhelpful.

Quoted 7 lines

I have had private patches for that myself, using the MSRs on AMD and Intel.

The problem is that you have to insert hooks early into the exception handlers to read the branch history MSRs, and that gets fairly ugly and a little slow and we can't really enable it by default.

There are two ways of capturing branches on the Intel processors (I have not looked at AMD): Last Branch Record (LBR) and Branch Trace Store (BTS). The former stores from/to information into MSRs and is very small (4 branches). The later could be as big as you want. On recent processors LBR and BTS can be constrained by priv level. The issue is that they capture ALL taken branches, not just taken function call or return.

Quoted 4 lines

But using BTS with a long in memory buffer would be fine. It would just be slower so it couldn't be enabled by default. But as a debugging feature it would be nice.

Yes, I think it could be used for debugging, you'd need to reserve a few pages and initialize a single MSR (32_DEBUGCTL).

Signature · 1 line
-Stephane
#9 lore raw
Quoted 2 lines

The former
stores from/to information into MSRs and is very small (4 branches).

P4 since Prescott has 16

Quoted 1 line

On recent processors LBR and BTS can be constrained by priv level.

Doesn't help for kernel debugging.

-Andi

#10 lore raw
Quoted 5 lines · Andi Kleen
Quoted 2 lines

The former
stores from/to information into MSRs and is very small (4 branches).

P4 since Prescott has 16

Yes. I was talking about Core 2

Quoted 4 lines
Quoted 1 line

On recent processors LBR and BTS can be constrained by priv level.

Doesn't help for kernel debugging.

Well, if you set if for kernel level only, you do not capture user level branches. This may happen if you crash soon after you've entered the kernel and you have a small buffer.

Signature · 1 line
-Stephane
#11 lore raw
In reply to Andi Kleen
Quoted 3 lines · Andi Kleen

I have had private patches for that myself, using the MSRs on AMD and Intel.

Would they be something that could be cleaned up into something mergeable? It would be nice to have something that could be left enabled all the time, but an option would at least make the functionality available.

J

#12 lore raw
Quoted 7 lines · Jeremy Fitzhardinge
Quoted 3 lines · Andi Kleen

I have had private patches for that myself, using the MSRs on AMD and Intel.

Would they be something that could be cleaned up into something mergeable?

Hmm maybe.

Quoted 2 lines

It would be nice to have something that could be left enabled all the time,

That would add considerable latency to all exceptions.

And unfortunately we take more than 16 jumps before we figure out an oops, so all the previous data would be gone if it was only done in the error path.

If the CPU had a "disable LBR on exceptions" bit that would work. Unfortunately it hasn't.

Quoted 2 lines

but an option would at least make the
functionality available.

If you have an debugging option you can as well enable the memory based branch tracer, which gives a much larger picture (but has somewhat more overhead too, i don't know how much the difference is) Just someone would need to write a driver for it. I think that would be more useful long term.

Here's the old MSR patch (for 64bit P4) for reference. I had an AMD patch too, but I can't find it right now and on P4 it works much better anyways because it has 16 LBRs instead of 4.

-Andi

Dump last branch information on oopses

arch/x86_64/kernel/entry.S +32 -0
Index: linux/arch/x86_64/kernel/entry.S
===================================================================
@@ -692,8 +692,38 @@ END(spurious_interrupt)
 /*
  * Exception entry points.
  */ 		
+
+	.macro savemsr msr,var
+	movl  $\msr,%ecx
+	rdmsr
+	movl %eax,\var
+	movl %edx,\var+4
+	.endm
+
+	.macro SAVELBR
+	cmpl $0,netburst
+	jz 1f
+	push %rax
+	push %rdx
+	push %rcx
+	savemsr 0x1da,lbr_tos
+	savemsr 0x1d7,ler_from
+	savemsr 0x1d8,ler_to
+	.set cnt,0
+	.rept 16
+	savemsr 0x680+cnt,lbr_from+cnt*8
+	savemsr 0x6c0+cnt,lbr_to+cnt*8
+	.set cnt,cnt+1
+	.endr 
+	pop %rcx
+	pop %rdx
+	pop %rax
+1:
+	.endm
+
 	.macro zeroentry sym
 	INTR_FRAME
+	SAVELBR
 	pushq $0	/* push error code/oldrax */ 
 	CFI_ADJUST_CFA_OFFSET 8
 	pushq %rax	/* push real oldrax to the rdi slot */ 
@@ -705,6 +735,7 @@ END(spurious_interrupt)
 
 	.macro errorentry sym
 	XCPT_FRAME
+	SAVELBR
 	pushq %rax
 	CFI_ADJUST_CFA_OFFSET 8
 	leaq  \sym(%rip),%rax
@@ -715,6 +746,7 @@ END(spurious_interrupt)
 	/* error code is on the stack already */
 	/* handle NMI like exceptions that can happen everywhere */
 	.macro paranoidentry sym, ist=0, irqtrace=1
+	SAVELBR
 	SAVE_ALL
 	cld
 	movl $1,%ebx
arch/x86_64/kernel/setup.c +9 -0
Index: linux/arch/x86_64/kernel/setup.c
===================================================================
@@ -822,10 +822,13 @@ static void srat_detect_node(void)
 #endif
 }
 
+int netburst;
+
 static void __cpuinit init_intel(struct cpuinfo_x86 *c)
 {
 	/* Cache sizes */
 	unsigned n;
+	unsigned long val;
 
 	init_intel_cacheinfo(c);
 	if (c->cpuid_level > 9 ) {
@@ -867,6 +870,12 @@ static void __cpuinit init_intel(struct 
  	c->x86_max_cores = intel_num_cpu_cores(c);
 
 	srat_detect_node();
+
+	if (c->x86 == 15) { 
+		rdmsrl(MSR_IA32_DEBUGCTLMSR, val);
+		wrmsrl(MSR_IA32_DEBUGCTLMSR, val | 1);
+		netburst = 1;
+	}
 }
 
 static void __cpuinit get_cpu_vendor(struct cpuinfo_x86 *c)
arch/x86_64/kernel/process.c +16 -0
Index: linux/arch/x86_64/kernel/process.c
===================================================================
@@ -36,6 +36,7 @@
 #include <linux/random.h>
 #include <linux/notifier.h>
 #include <linux/kprobes.h>
+#include <linux/kallsyms.h>
 
 #include <asm/uaccess.h>
 #include <asm/pgtable.h>
@@ -278,6 +279,9 @@ static int __init idle_setup (char *str)
 
 __setup("idle=", idle_setup);
 
+unsigned long lbr_tos, lbr_from[16], lbr_to[16], ler_from, ler_to;
+extern int netburst;
+
 /* Prints also some state that isn't saved in the pt_regs */ 
 void __show_regs(struct pt_regs * regs)
 {
@@ -326,6 +330,18 @@ void __show_regs(struct pt_regs * regs)
 	       fs,fsindex,gs,gsindex,shadowgs); 
 	printk("CS:  %04x DS: %04x ES: %04x CR0: %016lx\n", cs, ds, es, cr0); 
 	printk("CR2: %016lx CR3: %016lx CR4: %016lx\n", cr2, cr3, cr4);
+
+	if (netburst) { 
+		unsigned i;
+		printk("LBR: TOS %lx", lbr_tos);
+		print_symbol(" LER_FROM %s", ler_from);
+		print_symbol(" LER_TO %s\n", ler_to);
+		for (i = 0; i < 16; i++) { 
+			printk("    [%d]", i);
+			print_symbol(" FROM %s", lbr_from[i]);
+			print_symbol(" TO %s\n", lbr_from[i]);
+		}
+	}
 }
 
 void show_regs(struct pt_regs *regs)
#13 lore raw
Quoted 14 lines · Andi Kleen (on Sat, 18 Nov 2006 09:24:01 +0100) wrote:
Quoted 7 lines · Jeremy Fitzhardinge
Quoted 3 lines · Andi Kleen

I have had private patches for that myself, using the MSRs on AMD and Intel.

Would they be something that could be cleaned up into something mergeable?

Hmm maybe.

Quoted 2 lines

It would be nice to have something that could be left enabled all the time,

Quoted 8 lines

That would add considerable latency to all exceptions.

And unfortunately we take more than 16 jumps before we figure out an oops, so all the previous data would be gone if it was only done in the error path.

If the CPU had a "disable LBR on exceptions" bit that would work. Unfortunately it hasn't.

LBR is mainly useful on wild branches to random addresses. As such, you really only need to disable LBR in the page fault handler. For a long time, KDB had LBR support with this replacement for the i386 page fault handler.

#if defined(CONFIG_KDB)
ENTRY(page_fault_mca)

       pushl %ecx
       pushl %edx
       pushl %eax
       movl  $473,%ecx
       rdmsr
       andl  $0xfffffffe,%eax          /* Disable last branch recording */
       wrmsr
       popl  %eax
       popl  %edx
       popl  %ecx
       pushl $do_page_fault
       jmp error_code

#endif

Nobody was using the LBR feature of KDB so I removed it in 2.6.17.

#14 lore raw
Quoted 1 line

LBR is mainly useful on wild branches to random addresses. As such,

The page fault handler is one of the most performance critical exceptions. And then on x86-64 a lot of stray pointers actually result in a #GP, not a page fault.

-Andi