[PATCH] PPC: fix LOGMPP instruction opcode and inline asm

Back in 9678cda when we started using the Micro Partition Prefetch Engine
in POWER8 for KVM, there were two mistakes introduced from the original
patch used for investigation and microbenchmarks.

One mistake was that the opcode was constructed incorrectly, putting
the register in the wrong field in the opcode, meaning that we were
asking the chip to read the memory address from some other register than
what we intended - probably r0. For those unfortunate enough to have r0
point somewhere in memory they cared about, the prefetch engine would
gleefully trash all over it leading to some data you cared about being
replaced with a list of physical addresses.

In addition, the logmpp inline function incorrectly used R1 rather than
%0, leading to even if we got the construction of the instruction right,
we'd still generate the wrong thing, looking at the address in r1 rather
than whatever we were asked to look at.

So, this patch fixes the following:
- the inline logmpp function's inline asm to be correct
- puts the register in the right field of the instruction

This bug would overwrite a single 64k page.

https://bugzilla.redhat.com/show_bug.cgi?id=1269653
https://bugzilla.redhat.com/show_bug.cgi?id=1271997

Cc: stable@vger.kernel.org
Fixes: 9678cda ("Use the POWER8 Micro Partition Prefetch Engine in KVM HV")
Reported-by: David Gibson <redacted>
Reported-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Suggested-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Suggested-by: Paul Mackerras <redacted>
Signed-off-by: Stewart Smith <redacted>
---
 arch/powerpc/include/asm/cache.h      | 2 +-
 arch/powerpc/include/asm/ppc-opcode.h | 4 ++--
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/arch/powerpc/include/asm/cache.h b/arch/powerpc/include/asm/cache.h
index 34a05a1a990b..3af1c1e35435 100644
--- a/arch/powerpc/include/asm/cache.h
+++ b/arch/powerpc/include/asm/cache.h

@@ -43,7 +43,7 @@ extern struct ppc64_caches ppc64_caches;
 
 static inline void logmpp(u64 x)
 {
-	asm volatile(PPC_LOGMPP(R1) : : "r" (x));
+	asm volatile(PPC_LOGMPP(%0) : : "r" (x));
 }
 
 #endif /* __powerpc64__ && ! __ASSEMBLY__ */

diff --git a/arch/powerpc/include/asm/ppc-opcode.h b/arch/powerpc/include/asm/ppc-opcode.h
index 65136928a572..0dc2f6f9b445 100644
--- a/arch/powerpc/include/asm/ppc-opcode.h
+++ b/arch/powerpc/include/asm/ppc-opcode.h

@@ -304,8 +304,8 @@
 #define PPC_LDARX(t, a, b, eh)	stringify_in_c(.long PPC_INST_LDARX | \
 					___PPC_RT(t) | ___PPC_RA(a) | \
 					___PPC_RB(b) | __PPC_EH(eh))
-#define PPC_LOGMPP(b)		stringify_in_c(.long PPC_INST_LOGMPP | \
-					__PPC_RB(b))
+#define PPC_LOGMPP(a)		stringify_in_c(.long PPC_INST_LOGMPP | \
+					___PPC_RA(a))
 #define PPC_LWARX(t, a, b, eh)	stringify_in_c(.long PPC_INST_LWARX | \
 					___PPC_RT(t) | ___PPC_RA(a) | \
 					___PPC_RB(b) | __PPC_EH(eh))

-- 
2.1.4

On Fri, 2015-10-16 at 12:20 +1100, Stewart Smith wrote:

Back in 9678cda when we started using the Micro Partition Prefetch Engine
in POWER8 for KVM, there were two mistakes introduced from the original
patch used for investigation and microbenchmarks.

One mistake was that the opcode was constructed incorrectly, putting
the register in the wrong field in the opcode, meaning that we were
asking the chip to read the memory address from some other register than
what we intended - probably r0.

Where is the logmpp instruction documented?

cheers

Michael Ellerman [off-list ref] writes:

On Fri, 2015-10-16 at 12:20 +1100, Stewart Smith wrote:

quoted

Back in 9678cda when we started using the Micro Partition Prefetch Engine
in POWER8 for KVM, there were two mistakes introduced from the original
patch used for investigation and microbenchmarks.

One mistake was that the opcode was constructed incorrectly, putting
the register in the wrong field in the opcode, meaning that we were
asking the chip to read the memory address from some other register than
what we intended - probably r0.

Where is the logmpp instruction documented?

BookIV Section 5.1 - although I think this is now meant to all be in the
"P8 User Manual".

----- Original Message -----

From: "Stewart Smith" <redacted>
To: linuxppc-dev@lists.ozlabs.org
Cc: paulus@samba.org, thuth@redhat.com, jstancek@redhat.com, dgibson@redhat.com, benh@kernel.crashing.org, "Stewart
Smith" [off-list ref], stable@vger.kernel.org
Sent: Friday, 16 October, 2015 3:20:35 AM
Subject: [PATCH] PPC: fix LOGMPP instruction opcode and inline asm

Back in 9678cda when we started using the Micro Partition Prefetch Engine
in POWER8 for KVM, there were two mistakes introduced from the original
patch used for investigation and microbenchmarks.

One mistake was that the opcode was constructed incorrectly, putting
the register in the wrong field in the opcode, meaning that we were
asking the chip to read the memory address from some other register than
what we intended - probably r0. For those unfortunate enough to have r0
point somewhere in memory they cared about, the prefetch engine would
gleefully trash all over it leading to some data you cared about being
replaced with a list of physical addresses.

In addition, the logmpp inline function incorrectly used R1 rather than
%0, leading to even if we got the construction of the instruction right,
we'd still generate the wrong thing, looking at the address in r1 rather
than whatever we were asked to look at.

So, this patch fixes the following:
- the inline logmpp function's inline asm to be correct
- puts the register in the right field of the instruction

This bug would overwrite a single 64k page.

https://bugzilla.redhat.com/show_bug.cgi?id=1269653
https://bugzilla.redhat.com/show_bug.cgi?id=1271997

Cc: stable@vger.kernel.org
Fixes: 9678cda ("Use the POWER8 Micro Partition Prefetch Engine in KVM HV")
Reported-by: David Gibson <redacted>
Reported-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Suggested-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Suggested-by: Paul Mackerras <redacted>
Signed-off-by: Stewart Smith <redacted>

Tested-by: Jan Stancek <jstancek@redhat.com>

After running a kernel with this patch for 36+ hours, I can no longer
reproduce the user-space corruption I have been seeing.

Previously it reproduced fairly quickly (within hour, on IBM Power S814
[8286-41A], 16GB RAM) by starting/stopping couple KVM guests in loop.