Skip to contentSkip to conversation

powerpc: add BookS wait opcode macro

PATCH In review Nicholas Piggin 2 participants 5 messages Jul 11, 2022 – Jul 20, 2022 PowerPC

lore
Share & tools
lore

Nicholas Piggin

author reviewer
#0 lore raw

PATCH 1/2 powerpc: add BookS wait opcode macro

The wait instruction has a different encoding between BookE and BookS. Add the BookS variant.

---
 arch/powerpc/include/asm/ppc-opcode.h | 3 +++
 1 file changed, 3 insertions(+)
arch/powerpc/include/asm/ppc-opcode.h +3 -0
index 89beabf5325c..46fca27e8101 100644
@@ -331,6 +331,7 @@
 #define __PPC_XSP(s)	((((s) & 0x1e) | (((s) >> 5) & 0x1)) << 21)
 #define __PPC_XTP(s)	__PPC_XSP(s)
 #define __PPC_T_TLB(t)	(((t) & 0x3) << 21)
+#define __PPC_PL(p)	(((p) & 0x3) << 16)
 #define __PPC_WC(w)	(((w) & 0x3) << 21)
 #define __PPC_WS(w)	(((w) & 0x1f) << 11)
 #define __PPC_SH(s)	__PPC_WS(s)
@@ -399,6 +400,7 @@
 #define PPC_RAW_RFMCI			(0x4c00004c)
 #define PPC_RAW_TLBILX(t, a, b)		(0x7c000024 | __PPC_T_TLB(t) | 	__PPC_RA0(a) | __PPC_RB(b))
 #define PPC_RAW_WAIT(w)			(0x7c00007c | __PPC_WC(w))
+#define PPC_RAW_WAIT_BOOKS(w, p)	(0x7c00003c | __PPC_WC(w) | __PPC_PL(p))
 #define PPC_RAW_TLBIE(lp, a)		(0x7c000264 | ___PPC_RB(a) | ___PPC_RS(lp))
 #define PPC_RAW_TLBIE_5(rb, rs, ric, prs, r) \
 	(0x7c000264 | ___PPC_RB(rb) | ___PPC_RS(rs) | ___PPC_RIC(ric) | ___PPC_PRS(prs) | ___PPC_R(r))
@@ -614,6 +616,7 @@
 #define PPC_TLBILX_PID(a, b)	PPC_TLBILX(1, a, b)
 #define PPC_TLBILX_VA(a, b)	PPC_TLBILX(3, a, b)
 #define PPC_WAIT(w)		stringify_in_c(.long PPC_RAW_WAIT(w))
+#define PPC_WAIT_BOOKS(w, p)	stringify_in_c(.long PPC_RAW_WAIT_BOOKS(w, p))
 #define PPC_TLBIE(lp, a) 	stringify_in_c(.long PPC_RAW_TLBIE(lp, a))
 #define	PPC_TLBIE_5(rb, rs, ric, prs, r) \
 				stringify_in_c(.long PPC_RAW_TLBIE_5(rb, rs, ric, prs, r))
Signature · 1 line
2.35.1
author reviewer
#1 lore raw

PATCH 2/2 powerpc/64s: Make POWER10 and later use pause_short in cpu_relax loops

We want to move away from using SMT prioroty updates for cpu_relax, and use a 'wait' instruction which is similar to x86. As well as being a much better fit for what everybody else uses and tests with, priority nops are stateful which is nasty (interrupts have to consider they might be taken at a different priority), and they're expensive to execute, similar to a mtSPR which can effect other threads in the pipe.

Unfortunately qemu TCG does not emulate pause_short properly and will cause hangs. I have a patch for it but not merged yet. But if we tune qspinlock code it would be best to do it with this patch.

---
 arch/powerpc/include/asm/processor.h      | 30 +++++++++++++++++++----
 arch/powerpc/include/asm/vdso/processor.h | 10 +++++++-
 2 files changed, 34 insertions(+), 6 deletions(-)
arch/powerpc/include/asm/processor.h +25 -5
index fdfaae194ddd..61f16515cbe0 100644
@@ -355,11 +355,31 @@ static inline unsigned long __pack_fe01(unsigned int fpmode)
 
 #ifdef CONFIG_PPC64
 
-#define spin_begin()	HMT_low()
-
-#define spin_cpu_relax()	barrier()
-
-#define spin_end()	HMT_medium()
+#define spin_begin()							\
+do {									\
+	asm volatile(ASM_FTR_IFCLR(					\
+		"or 1,1,1", /* HMT_LOW */				\
+		"nop",/* POWER10 onward uses pause_short (wait 2,0) */	\
+				%0) :: "i" (CPU_FTR_ARCH_31) : "memory"); \
+} while (0)
+
+#define spin_cpu_relax()						\
+do {									\
+	asm volatile(ASM_FTR_IFCLR(					\
+		/* Pre-POWER10 uses low ; medium priority nops */	\
+		"nop",							\
+		/* POWER10 onward uses pause_short (wait 2,0) */	\
+		PPC_WAIT_BOOKS(2, 0),					\
+				%0) :: "i" (CPU_FTR_ARCH_31) : "memory"); \
+} while (0)
+
+#define spin_end()							\
+do {									\
+	asm volatile(ASM_FTR_IFCLR(					\
+		"or 2,2,2", /* HMT_MEDIUM */				\
+		"nop",/* POWER10 onward uses pause_short (wait 2,0) */	\
+				%0) :: "i" (CPU_FTR_ARCH_31) : "memory"); \
+} while (0)
 
 #endif
 
arch/powerpc/include/asm/vdso/processor.h +9 -1
index 8d79f994b4aa..1116230ebb08 100644
@@ -22,7 +22,15 @@
 #endif
 
 #ifdef CONFIG_PPC64
-#define cpu_relax()	do { HMT_low(); HMT_medium(); barrier(); } while (0)
+#define cpu_relax()							\
+do {									\
+	asm volatile(ASM_FTR_IFCLR(					\
+		/* Pre-POWER10 uses low ; medium priority nops */	\
+		"or 1,1,1 ; or 2,2,2",					\
+		/* POWER10 onward uses pause_short (wait 2,0) */	\
+		PPC_WAIT_BOOKS(2, 0),					\
+				%0) :: "i" (CPU_FTR_ARCH_31) : "memory"); \
+} while (0)
 #else
 #define cpu_relax()	barrier()
 #endif
Signature · 1 line
2.35.1
#2 lore raw
In reply to Nicholas Piggin

Hi!

Quoted 5 lines · Nicholas Piggin

The wait instruction has a different encoding between BookE and BookS. Add the BookS variant.

#define PPC_RAW_WAIT(w) (0x7c00007c | __PPC_WC(w))
+#define PPC_RAW_WAIT_BOOKS(w, p) (0x7c00003c | __PPC_WC(w) | __PPC_PL(p))

The embedded extensions are no longer part of the PowerPC architecture, so wouldn't it be a better way forward to rename the existing one, instead? A bit more work now, but less in the future :-)

Segher

author reviewer
#3 lore raw

Excerpts from Segher Boessenkool's message of July 13, 2022 2:29 am:

Quoted 6 lines

Hi!

Quoted 2 lines · Nicholas Piggin

The wait instruction has a different encoding between BookE and BookS. Add the BookS variant.

Quoted 6 lines
Quoted 2 lines

#define PPC_RAW_WAIT(w) (0x7c00007c | __PPC_WC(w))
+#define PPC_RAW_WAIT_BOOKS(w, p) (0x7c00003c | __PPC_WC(w) | __PPC_PL(p))

The embedded extensions are no longer part of the PowerPC architecture, so wouldn't it be a better way forward to rename the existing one, instead? A bit more work now, but less in the future :-)

And I actually misremembered this too, was off digging and asking about it, but the change isn't strictly BookE vs BookS, but rather the wait opcode was changed in ISA v3.0, which is a bit of an unfortunate landmine.

It seems apparently POWER8 implemented a non-architected instruction 'waitasec' that uses this opcode, then I suppose it was decided to continue with that opcode in v3.0 when BookE was dropped, for reasons. Maybe it was more widely used?

In any case, I will rename it. Precedent is divided. We have PPC_RAW_TLBIEL_v205 for older tlbiel, and PPC_ISA_3_0_INVALIDATE_ERAT for a new ERAT invalidation instruction. I guess making the older instruction the exceptional case ends up being better in the long term.

Thanks,
Nick

#4 lore raw
Quoted 9 lines · Nicholas Piggin

Excerpts from Segher Boessenkool's message of July 13, 2022 2:29 am:

Quoted 3 lines

The embedded extensions are no longer part of the PowerPC architecture, so wouldn't it be a better way forward to rename the existing one, instead? A bit more work now, but less in the future :-)

And I actually misremembered this too, was off digging and asking about it, but the change isn't strictly BookE vs BookS, but rather the wait opcode was changed in ISA v3.0, which is a bit of an unfortunate landmine.

The current wait opcode is *new* in 3.0, but it has the same mnemonic as the old one, more formally.

Quoted 3 lines

It seems apparently POWER8 implemented a non-architected instruction 'waitasec' that uses this opcode, then I suppose it was decided to continue with that opcode in v3.0 when BookE was dropped, for reasons.

The (newer) wait instruction is 0/30 while the old one is 1/30 (secondary opcodes 30 resp. 62), although 0/30 stayed open; maybe there is more history there already? Curious.

Quoted 5 lines

In any case, I will rename it. Precedent is divided. We have PPC_RAW_TLBIEL_v205 for older tlbiel, and PPC_ISA_3_0_INVALIDATE_ERAT for a new ERAT invalidation instruction. I guess making the older instruction the exceptional case ends up being better in the long term.

Yeah exactly, it's better if the simpler name is the one that newer code should use: more readable, more writable, and importantly less room for mistakes :-)

Thanks,

Segher