Skip to contentSkip to conversation

powerpc/pseries: Kill all prefetch streams on context switch

PATCH In review Michael Neuling 2 participants 3 messages May 30, 2013 – Jun 3, 2013 PowerPC

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 powerpc/pseries: Kill all prefetch streams on context switch

On context switch, we should have no prefetch streams leak from one userspace process to another. This frees up prefetch resources for the next process.

Based on patch from Milton Miller.

arch/powerpc/include/asm/ppc_asm.h +11 -0
index cea8496..2f1b6c5 100644
@@ -523,6 +523,17 @@ END_FTR_SECTION_NESTED(CPU_FTR_HAS_PPR,CPU_FTR_HAS_PPR,946)
 #define PPC440EP_ERR42
 #endif
 
+/* The following stops all load and store data streams associated with stream
+ * ID (ie. streams created explicitly).  The embedded and server mnemonics for
+ * dcbt are different so we use machine "power4" here explicitly.
+ */
+#define DCBT_STOP_ALL_STREAM_IDS(scratch)	\
+.machine push ;					\
+.machine "power4" ;				\
+       lis     scratch,0x60000000@h;		\
+       dcbt    r0,scratch,0b01010;		\
+.machine pop
+
 /*
  * toreal/fromreal/tophys/tovirt macros. 32-bit BookE makes them
  * keep the address intact to be compatible with code shared with
arch/powerpc/kernel/entry_64.S +7 -0
index 0e9095e..246b11c 100644
@@ -501,6 +501,13 @@ BEGIN_FTR_SECTION
 	ldarx	r6,0,r1
 END_FTR_SECTION_IFSET(CPU_FTR_STCX_CHECKS_ADDRESS)
 
+#ifdef CONFIG_PPC_BOOK3S
+/* Cancel all explict user streams as they will have no use after context
+ * switch and will stop the HW from creating streams itself
+ */
+	DCBT_STOP_ALL_STREAM_IDS(r6)
+#endif
+
 	addi	r6,r4,-THREAD	/* Convert THREAD to 'current' */
 	std	r6,PACACURRENT(r13)	/* Set new 'current' */
 
#1 lore raw
Quoted 16 lines · Michael Neuling

On context switch, we should have no prefetch streams leak from one userspace process to another. This frees up prefetch resources for the next process.

Based on patch from Milton Miller.

arch/powerpc/include/asm/ppc_asm.h +0 -0
index cea8496..2f1b6c5 100644
@@ -523,6 +523,17 @@ END_FTR_SECTION_NESTED(CPU_FTR_HAS_PPR,CPU_FTR_HAS_PPR,946)
 #define PPC440EP_ERR42
 #endif
 

arch/powerpc/include/asm/ppc_asm.h END_FTR_SECTION_NESTED(CPU_FTR_HAS_PPR,CPU_FTR_HAS_PPR,946) · line 535

Quoted 10 lines
arch/powerpc/include/asm/ppc_asm.h +10 -0
+/* The following stops all load and store data streams associated with stream
+ * ID (ie. streams created explicitly).  The embedded and server mnemonics for
+ * dcbt are different so we use machine "power4" here explicitly.
+ */
+#define DCBT_STOP_ALL_STREAM_IDS(scratch)	\
+.machine push ;					\
+.machine "power4" ;				\
+       lis     scratch,0x60000000@h;		\
+       dcbt    r0,scratch,0b01010;		\
+.machine pop

I don't see why we need the macro, ie. just stick this code in entry_64.S directly.

cheers

#2 lore raw
Quoted 17 lines · Michael Ellerman
Quoted 16 lines · Michael Neuling

On context switch, we should have no prefetch streams leak from one userspace process to another. This frees up prefetch resources for the next process.

Based on patch from Milton Miller.

arch/powerpc/include/asm/ppc_asm.h +0 -0
index cea8496..2f1b6c5 100644
@@ -523,6 +523,17 @@ END_FTR_SECTION_NESTED(CPU_FTR_HAS_PPR,CPU_FTR_HAS_PPR,946)
 #define PPC440EP_ERR42
 #endif
 
Quoted 13 lines

arch/powerpc/include/asm/ppc_asm.h END_FTR_SECTION_NESTED(CPU_FTR_HAS_PPR,CPU_FTR_HAS_PPR,946) · line 535

Quoted 10 lines
arch/powerpc/include/asm/ppc_asm.h +10 -0
+/* The following stops all load and store data streams associated with stream
+ * ID (ie. streams created explicitly).  The embedded and server mnemonics for
+ * dcbt are different so we use machine "power4" here explicitly.
+ */
+#define DCBT_STOP_ALL_STREAM_IDS(scratch)	\
+.machine push ;					\
+.machine "power4" ;				\
+       lis     scratch,0x60000000@h;		\
+       dcbt    r0,scratch,0b01010;		\
+.machine pop

I don't see why we need the macro, ie. just stick this code in entry_64.S directly.

There's a decent chance we'll want to use this again at some point. We also most stuck it in the error path of the power7 copy user loop but after consulting Anton and heavily reviewing the code, we decoded it shouldn't be needed.

It's in Linus tree now anyway.

Mikey