Skip to contentSkip to conversation

Andi, you broke my laptop :-)

PATCH Pete Zaitcev 4 participants 6 messages May 9, 2007 – May 10, 2007

lore
Share & tools
lore
#0 lore raw

Hi, Andi:

The attached patch (actually, git show output) makes my Dell 1501 to hang on boot. Sorry, I have no clue why... The culprit is found with git bisect. But yes, it's an AMD MK-36. I use an x86_64 kernel. It is 100% reproducible.

Cheers,
-- Pete

commit c5bcb5635a03da3158f121ae20ccbbf72b4fc62a

Author: Andi Kleen <[email protected]>
Date:   Wed May 2 19:27:21 2007 +0200

[PATCH] x86: Use RDTSCP for synchronous get_cycles if possible

RDTSCP is already synchronous and doesn't need an explicit CPUID. This is a little faster and more importantly avoids VMEXITs on Hypervisors.

Original patch from Joerg Roedel, but reworked by AK Also includes miscompilation fix by Eric Biederman

Cc "Joerg Roedel" <[email protected]>
include/asm-i386/tsc.h +9 -0
index 0181f9d..3f3c1fa 100644
@@ -38,6 +38,15 @@ static __always_inline cycles_t get_cycles_sync(void)
 	unsigned eax;
 
 	/*
+  	 * Use RDTSCP if possible; it is guaranteed to be synchronous
+ 	 * and doesn't cause a VMEXIT on Hypervisors
+	 */
+	alternative_io(ASM_NOP3, ".byte 0x0f,0x01,0xf9", X86_FEATURE_RDTSCP,
+			 	 "=A" (ret), "0" (0ULL) : "ecx", "memory");
+	if (ret)
+		return ret;
+
+	/*
 	 * Don't do an additional sync on CPUs where we know
 	 * RDTSC is already synchronous:
 	 */
#1 lore raw
Quoted 5 lines · Pete Zaitcev

Hi, Andi:

The attached patch (actually, git show output) makes my Dell 1501 to hang on boot. Sorry, I have no clue why... The culprit is found with git bisect. But yes, it's an AMD MK-36. I use an x86_64 kernel. It is 100% reproducible.

MK-36? Does it have SVM?

Anyways we previously had issues with this being miscompiled, but I thought the latest patch should have been ok. What compiler do you use?

Can you send me a disassembly listing of arch/x86_64/kernel/time.o?

-Andi

#2 lore raw
Quoted 7 lines · Andi Kleen
Quoted 5 lines · Pete Zaitcev

Hi, Andi:

The attached patch (actually, git show output) makes my Dell 1501 to hang on boot. Sorry, I have no clue why... The culprit is found with git bisect. But yes, it's an AMD MK-36. I use an x86_64 kernel. It is 100% reproducible.

Quoted 6 lines

MK-36? Does it have SVM?

Anyways we previously had issues with this being miscompiled, but I thought the latest patch should have been ok. What compiler do you use?

Can you send me a disassembly listing of arch/x86_64/kernel/time.o?

I debugged this problem a bit and my compiler[1]interprets the =A constraint as %rax instead of %edx:%eax on x86_64 which causes the problem. The appended patch provides a workaround for this and fixed the hang on my machine.

[1] gcc version 4.1.3 20070429 (prerelease) (Debian 4.1.2-5)

Signature · 6 lines
           |           AMD Saxony Limited Liability Company & Co. KG
 Operating |         Wilschdorfer Landstr. 101, 01109 Dresden, Germany
 System    |                  Register Court Dresden: HRA 4896
 Research  |              General Partner authorized to represent:
 Center    |             AMD Saxony LLC (Wilmington, Delaware, US)
           | General Manager of AMD Saxony LLC: Dr. Hans-R. Deppe, Thomas McCoy
include/asm-i386/alternative.h +6 -0
index 0f70b37..eb7da54 100644
@@ -98,6 +98,12 @@ static inline void alternatives_smp_switch(int smp) {}
 		      ".previous" : output : [feat] "i" (feature), ##input)
 
 /*
+ * use this macro(s) if you need more than one output parameter
+ * in alternative_io
+ */
+#define ASM_OUTPUT2(a, b) a, b
+
+/*
  * Alternative inline assembly for SMP.
  *
  * The LOCK_PREFIX macro defined here replaces the LOCK and
include/asm-i386/tsc.h +4 -2
index 3f3c1fa..62c091f 100644
@@ -35,14 +35,16 @@ static inline cycles_t get_cycles(void)
 static __always_inline cycles_t get_cycles_sync(void)
 {
 	unsigned long long ret;
-	unsigned eax;
+	unsigned eax, edx;
 
 	/*
   	 * Use RDTSCP if possible; it is guaranteed to be synchronous
  	 * and doesn't cause a VMEXIT on Hypervisors
 	 */
 	alternative_io(ASM_NOP3, ".byte 0x0f,0x01,0xf9", X86_FEATURE_RDTSCP,
-			 	 "=A" (ret), "0" (0ULL) : "ecx", "memory");
+		       ASM_OUTPUT2("=a" (eax), "=d" (edx)),
+		       "a" (0U), "d" (0U) : "ecx", "memory");
+	ret = (((unsigned long long)edx) << 32) | ((unsigned long long)eax);
 	if (ret)
 		return ret;
 
include/asm-x86_64/alternative.h +6 -0
index a09fe85..a094276 100644
@@ -103,6 +103,12 @@ static inline void alternatives_smp_switch(int smp) {}
 		      ".previous" : output : [feat] "i" (feature), ##input)
 
 /*
+ * use this macro(s) if you need more than one output parameter
+ * in alternative_io
+ */
+#define ASM_OUTPUT2(a, b) a, b
+
+/*
  * Alternative inline assembly for SMP.
  *
  * The LOCK_PREFIX macro defined here replaces the LOCK and
#3 lore raw
Quoted 13 lines · Joerg Roedel
Quoted 12 lines · Andi Kleen
Quoted 5 lines · Pete Zaitcev

Hi, Andi:

The attached patch (actually, git show output) makes my Dell 1501 to hang on boot. Sorry, I have no clue why... The culprit is found with git bisect. But yes, it's an AMD MK-36. I use an x86_64 kernel. It is 100% reproducible.

MK-36? Does it have SVM?

Anyways we previously had issues with this being miscompiled, but I thought the latest patch should have been ok. What compiler do you use?

Quoted 6 lines
Quoted 1 line

Can you send me a disassembly listing of arch/x86_64/kernel/time.o?

I debugged this problem a bit and my compiler[1]interprets the =A constraint as %rax instead of %edx:%eax on x86_64 which causes the problem. The appended patch provides a workaround for this and fixed the hang on my machine.

Hmm yes now I can reproduce it too. I didn't see any hangs so i suppose my (and that of most -mm tester's) compiled binary always happened to have a suitable value in edx

Thanks for the patch.

-Andi

#4 lore raw

Joerg's patch works for me too.

Thanks,

Benny

Quoted 31 lines · Andi Kleen
Quoted 9 lines · Joerg Roedel
Quoted 8 lines · Andi Kleen
Quoted 5 lines · Pete Zaitcev

Hi, Andi:

The attached patch (actually, git show output) makes my Dell 1501 to hang on boot. Sorry, I have no clue why... The culprit is found with git bisect. But yes, it's an AMD MK-36. I use an x86_64 kernel. It is 100% reproducible.

MK-36? Does it have SVM?

Quoted 8 lines
Quoted 4 lines

Anyways we previously had issues with this being miscompiled, but I thought the latest patch should have been ok. What compiler do you use?

Can you send me a disassembly listing of arch/x86_64/kernel/time.o?

I debugged this problem a bit and my compiler[1]interprets the =A constraint as %rax instead of %edx:%eax on x86_64 which causes the problem. The appended patch provides a workaround for this and fixed the hang on my machine.

Hmm yes now I can reproduce it too. I didn't see any hangs so i suppose my (and that of most -mm tester's) compiled binary always happened to have a suitable value in edx

Thanks for the patch.

(quoted diff) +0 -2
-Andi
-

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Signature · 8 lines
Benny Halevy
Panasas Inc.
Accelerating Time to Results(TM) with Clustered Storage

www.panasas.com
[email protected]
Tel/Fax: +972-3-647-8340
Mobile: +972-54-802-8340
#5 lore raw
In reply to Joerg Roedel

include/asm-i386/tsc.h static inline cycles_t get_cycles(void) · line 46

Quoted 11 lines · Joerg Roedel

I debugged this problem a bit and my compiler[1]interprets the =A constraint as %rax instead of %edx:%eax on x86_64 which causes the problem. The appended patch provides a workaround for this and fixed the hang on my machine.

[1] gcc version 4.1.3 20070429 (prerelease) (Debian 4.1.2-5)

include/asm-i386/tsc.h +2 -1
 	alternative_io(ASM_NOP3, ".byte 0x0f,0x01,0xf9", X86_FEATURE_RDTSCP,
-			 	 "=A" (ret), "0" (0ULL) : "ecx", "memory");
+		       ASM_OUTPUT2("=a" (eax), "=d" (edx)),
+		       "a" (0U), "d" (0U) : "ecx", "memory");

This works for me. Thanks, Joerg.

gcc version 4.1.2 20070424 (Red Hat 4.1.2-11)

-- Pete