Skip to contentSkip to conversation

powerpc: queued spinlocks and rwlocks

PATCH Superseded Nicholas Piggin 7 participants 46 messages Jul 6, 2020 – Jul 25, 2020

Subsystems (6)

v3 of 4 Superseded Latest is v4 →

lore
Share & tools
lore

Nicholas Piggin

author reviewer
#0 lore raw

v3 cover powerpc: queued spinlocks and rwlocks

v3 is updated to use __pv_queued_spin_unlock, noticed by Waiman (thank you).

Thanks,
Nick

Nicholas Piggin (6):
powerpc/powernv: must include hvcall.h to get PAPR defines powerpc/pseries: move some PAPR paravirt functions to their own file
powerpc: move spinlock implementation to simple_spinlock powerpc/64s: implement queued spinlocks and rwlocks powerpc/pseries: implement paravirt qspinlocks for SPLPAR powerpc/qspinlock: optimised atomic_try_cmpxchg_lock that adds the
lock hint

 arch/powerpc/Kconfig                          |  13 +
 arch/powerpc/include/asm/Kbuild               |   2 +
 arch/powerpc/include/asm/atomic.h             |  28 ++
 arch/powerpc/include/asm/paravirt.h           |  89 +++++
 arch/powerpc/include/asm/qspinlock.h          |  91 ++++++
 arch/powerpc/include/asm/qspinlock_paravirt.h |   7 +
 arch/powerpc/include/asm/simple_spinlock.h    | 292 +++++++++++++++++
 .../include/asm/simple_spinlock_types.h       |  21 ++
 arch/powerpc/include/asm/spinlock.h           | 308 +-----------------
 arch/powerpc/include/asm/spinlock_types.h     |  17 +-
 arch/powerpc/lib/Makefile                     |   3 +
 arch/powerpc/lib/locks.c                      |  12 +-
 arch/powerpc/platforms/powernv/pci-ioda-tce.c |   1 +
 arch/powerpc/platforms/pseries/Kconfig        |   5 +
 arch/powerpc/platforms/pseries/setup.c        |   6 +-
 include/asm-generic/qspinlock.h               |   4 +
 16 files changed, 577 insertions(+), 322 deletions(-)
 create mode 100644 arch/powerpc/include/asm/paravirt.h
 create mode 100644 arch/powerpc/include/asm/qspinlock.h
 create mode 100644 arch/powerpc/include/asm/qspinlock_paravirt.h
 create mode 100644 arch/powerpc/include/asm/simple_spinlock.h
 create mode 100644 arch/powerpc/include/asm/simple_spinlock_types.h
Signature · 1 line
2.23.0
author reviewer
#1 lore raw

PATCH v3 1/6 powerpc/powernv: must include hvcall.h to get PAPR defines

An include goes away in future patches which breaks compilation without this.

---
 arch/powerpc/platforms/powernv/pci-ioda-tce.c | 1 +
 1 file changed, 1 insertion(+)
arch/powerpc/platforms/powernv/pci-ioda-tce.c +1 -0
index f923359d8afc..8eba6ece7808 100644
@@ -15,6 +15,7 @@
 
 #include <asm/iommu.h>
 #include <asm/tce.h>
+#include <asm/hvcall.h> /* share error returns with PAPR */
 #include "pci.h"
 
 unsigned long pnv_ioda_parse_tce_sizes(struct pnv_phb *phb)
Signature · 1 line
2.23.0
#2 lore raw

Re: [PATCH v3 1/6] powerpc/powernv: must include hvcall.h to get PAPR defines

Quoted 22 lines · Nicholas Piggin

An include goes away in future patches which breaks compilation without this.

---
 arch/powerpc/platforms/powernv/pci-ioda-tce.c | 1 +
 1 file changed, 1 insertion(+)
arch/powerpc/platforms/powernv/pci-ioda-tce.c +1 -0
index f923359d8afc..8eba6ece7808 100644
@@ -15,6 +15,7 @@
 
 #include <asm/iommu.h>
 #include <asm/tce.h>
+#include <asm/hvcall.h> /* share error returns with PAPR */
 #include "pci.h"
 
 unsigned long pnv_ioda_parse_tce_sizes(struct pnv_phb *phb)
Signature · 1 line
2.23.0

This isn't needed anymore AFAICS, since:

5f202c1a1d42 ("powerpc/powernv/ioda: Return correct error if TCE level allocation failed")

cheers

author reviewer
#3 lore raw

PATCH v3 2/6 powerpc/pseries: move some PAPR paravirt functions to their own file

In reply to Nicholas Piggin
---
 arch/powerpc/include/asm/paravirt.h | 61 +++++++++++++++++++++++++++++
 arch/powerpc/include/asm/spinlock.h | 24 +-----------
 arch/powerpc/lib/locks.c            | 12 +++---
 3 files changed, 68 insertions(+), 29 deletions(-)
 create mode 100644 arch/powerpc/include/asm/paravirt.h
arch/powerpc/include/asm/paravirt.h new file +61 -0
new file mode 100644
index 000000000000..7a8546660a63
@@ -0,0 +1,61 @@
+/* SPDX-License-Identifier: GPL-2.0-or-later */
+#ifndef __ASM_PARAVIRT_H
+#define __ASM_PARAVIRT_H
+#ifdef __KERNEL__
+
+#include <linux/jump_label.h>
+#include <asm/smp.h>
+#ifdef CONFIG_PPC64
+#include <asm/paca.h>
+#include <asm/hvcall.h>
+#endif
+
+#ifdef CONFIG_PPC_SPLPAR
+DECLARE_STATIC_KEY_FALSE(shared_processor);
+
+static inline bool is_shared_processor(void)
+{
+	return static_branch_unlikely(&shared_processor);
+}
+
+/* If bit 0 is set, the cpu has been preempted */
+static inline u32 yield_count_of(int cpu)
+{
+	__be32 yield_count = READ_ONCE(lppaca_of(cpu).yield_count);
+	return be32_to_cpu(yield_count);
+}
+
+static inline void yield_to_preempted(int cpu, u32 yield_count)
+{
+	plpar_hcall_norets(H_CONFER, get_hard_smp_processor_id(cpu), yield_count);
+}
+#else
+static inline bool is_shared_processor(void)
+{
+	return false;
+}
+
+static inline u32 yield_count_of(int cpu)
+{
+	return 0;
+}
+
+extern void ___bad_yield_to_preempted(void);
+static inline void yield_to_preempted(int cpu, u32 yield_count)
+{
+	___bad_yield_to_preempted(); /* This would be a bug */
+}
+#endif
+
+#define vcpu_is_preempted vcpu_is_preempted
+static inline bool vcpu_is_preempted(int cpu)
+{
+	if (!is_shared_processor())
+		return false;
+	if (yield_count_of(cpu) & 1)
+		return true;
+	return false;
+}
+
+#endif /* __KERNEL__ */
+#endif /* __ASM_PARAVIRT_H */
arch/powerpc/include/asm/spinlock.h +1 -23
index 2d620896cdae..79be9bb10bbb 100644
@@ -15,11 +15,10 @@
  *
  * (the type definitions are in asm/spinlock_types.h)
  */
-#include <linux/jump_label.h>
 #include <linux/irqflags.h>
+#include <asm/paravirt.h>
 #ifdef CONFIG_PPC64
 #include <asm/paca.h>
-#include <asm/hvcall.h>
 #endif
 #include <asm/synch.h>
 #include <asm/ppc-opcode.h>
@@ -35,18 +34,6 @@
 #define LOCK_TOKEN	1
 #endif
 
-#ifdef CONFIG_PPC_PSERIES
-DECLARE_STATIC_KEY_FALSE(shared_processor);
-
-#define vcpu_is_preempted vcpu_is_preempted
-static inline bool vcpu_is_preempted(int cpu)
-{
-	if (!static_branch_unlikely(&shared_processor))
-		return false;
-	return !!(be32_to_cpu(lppaca_of(cpu).yield_count) & 1);
-}
-#endif
-
 static __always_inline int arch_spin_value_unlocked(arch_spinlock_t lock)
 {
 	return lock.slock = 0;
@@ -110,15 +97,6 @@ static inline void splpar_spin_yield(arch_spinlock_t *lock) {};
 static inline void splpar_rw_yield(arch_rwlock_t *lock) {};
 #endif
 
-static inline bool is_shared_processor(void)
-{
-#ifdef CONFIG_PPC_SPLPAR
-	return static_branch_unlikely(&shared_processor);
-#else
-	return false;
-#endif
-}
-
 static inline void spin_yield(arch_spinlock_t *lock)
 {
 	if (is_shared_processor())
arch/powerpc/lib/locks.c +6 -6
index 6440d5943c00..04165b7a163f 100644
@@ -27,14 +27,14 @@ void splpar_spin_yield(arch_spinlock_t *lock)
 		return;
 	holder_cpu = lock_value & 0xffff;
 	BUG_ON(holder_cpu >= NR_CPUS);
-	yield_count = be32_to_cpu(lppaca_of(holder_cpu).yield_count);
+
+	yield_count = yield_count_of(holder_cpu);
 	if ((yield_count & 1) = 0)
 		return;		/* virtual cpu is currently running */
 	rmb();
 	if (lock->slock != lock_value)
 		return;		/* something has changed */
-	plpar_hcall_norets(H_CONFER,
-		get_hard_smp_processor_id(holder_cpu), yield_count);
+	yield_to_preempted(holder_cpu, yield_count);
 }
 EXPORT_SYMBOL_GPL(splpar_spin_yield);
 
@@ -53,13 +53,13 @@ void splpar_rw_yield(arch_rwlock_t *rw)
 		return;		/* no write lock at present */
 	holder_cpu = lock_value & 0xffff;
 	BUG_ON(holder_cpu >= NR_CPUS);
-	yield_count = be32_to_cpu(lppaca_of(holder_cpu).yield_count);
+
+	yield_count = yield_count_of(holder_cpu);
 	if ((yield_count & 1) = 0)
 		return;		/* virtual cpu is currently running */
 	rmb();
 	if (rw->lock != lock_value)
 		return;		/* something has changed */
-	plpar_hcall_norets(H_CONFER,
-		get_hard_smp_processor_id(holder_cpu), yield_count);
+	yield_to_preempted(holder_cpu, yield_count);
 }
 #endif
Signature · 1 line
2.23.0
#4 lore raw

Re: [PATCH v3 2/6] powerpc/pseries: move some PAPR paravirt functions to their own file

Quoted 1 line · Nicholas Piggin

Little bit of changelog would be nice :D

Quoted 14 lines
---
 arch/powerpc/include/asm/paravirt.h | 61 +++++++++++++++++++++++++++++
 arch/powerpc/include/asm/spinlock.h | 24 +-----------
 arch/powerpc/lib/locks.c            | 12 +++---
 3 files changed, 68 insertions(+), 29 deletions(-)
 create mode 100644 arch/powerpc/include/asm/paravirt.h
arch/powerpc/include/asm/paravirt.h new file +0 -0
new file mode 100644
index 000000000000..7a8546660a63

arch/powerpc/include/asm/paravirt.h line 3

Quoted 3 lines
arch/powerpc/include/asm/paravirt.h new file +3 -0
@@ -0,0 +1,61 @@
+/* SPDX-License-Identifier: GPL-2.0-or-later */
+#ifndef __ASM_PARAVIRT_H
+#define __ASM_PARAVIRT_H

Should be _ASM_POWERPC_PARAVIRT_H

arch/powerpc/include/asm/paravirt.h line 4

Quoted 1 line
arch/powerpc/include/asm/paravirt.h new file +1 -0
+#ifdef __KERNEL__

We shouldn't need __KERNEL__ in here, it's not a uapi header.

cheers

author reviewer
#5 lore raw

PATCH v3 3/6 powerpc: move spinlock implementation to simple_spinlock

In reply to Nicholas Piggin

To prepare for queued spinlocks. This is a simple rename except to update preprocessor guard name and a file reference.

---
 arch/powerpc/include/asm/simple_spinlock.h    | 292 ++++++++++++++++++
 .../include/asm/simple_spinlock_types.h       |  21 ++
 arch/powerpc/include/asm/spinlock.h           | 285 +----------------
 arch/powerpc/include/asm/spinlock_types.h     |  12 +-
 4 files changed, 315 insertions(+), 295 deletions(-)
 create mode 100644 arch/powerpc/include/asm/simple_spinlock.h
 create mode 100644 arch/powerpc/include/asm/simple_spinlock_types.h
arch/powerpc/include/asm/simple_spinlock.h new file +292 -0
new file mode 100644
index 000000000000..e048c041c4a9
@@ -0,0 +1,292 @@
+/* SPDX-License-Identifier: GPL-2.0-or-later */
+#ifndef __ASM_SIMPLE_SPINLOCK_H
+#define __ASM_SIMPLE_SPINLOCK_H
+#ifdef __KERNEL__
+
+/*
+ * Simple spin lock operations.  
+ *
+ * Copyright (C) 2001-2004 Paul Mackerras <[email protected]>, IBM
+ * Copyright (C) 2001 Anton Blanchard <[email protected]>, IBM
+ * Copyright (C) 2002 Dave Engebretsen <[email protected]>, IBM
+ *	Rework to support virtual processors
+ *
+ * Type of int is used as a full 64b word is not necessary.
+ *
+ * (the type definitions are in asm/simple_spinlock_types.h)
+ */
+#include <linux/irqflags.h>
+#include <asm/paravirt.h>
+#ifdef CONFIG_PPC64
+#include <asm/paca.h>
+#endif
+#include <asm/synch.h>
+#include <asm/ppc-opcode.h>
+
+#ifdef CONFIG_PPC64
+/* use 0x800000yy when locked, where yy = CPU number */
+#ifdef __BIG_ENDIAN__
+#define LOCK_TOKEN	(*(u32 *)(&get_paca()->lock_token))
+#else
+#define LOCK_TOKEN	(*(u32 *)(&get_paca()->paca_index))
+#endif
+#else
+#define LOCK_TOKEN	1
+#endif
+
+static __always_inline int arch_spin_value_unlocked(arch_spinlock_t lock)
+{
+	return lock.slock = 0;
+}
+
+static inline int arch_spin_is_locked(arch_spinlock_t *lock)
+{
+	smp_mb();
+	return !arch_spin_value_unlocked(*lock);
+}
+
+/*
+ * This returns the old value in the lock, so we succeeded
+ * in getting the lock if the return value is 0.
+ */
+static inline unsigned long __arch_spin_trylock(arch_spinlock_t *lock)
+{
+	unsigned long tmp, token;
+
+	token = LOCK_TOKEN;
+	__asm__ __volatile__(
+"1:	" PPC_LWARX(%0,0,%2,1) "\n\
+	cmpwi		0,%0,0\n\
+	bne-		2f\n\
+	stwcx.		%1,0,%2\n\
+	bne-		1b\n"
+	PPC_ACQUIRE_BARRIER
+"2:"
+	: "=&r" (tmp)
+	: "r" (token), "r" (&lock->slock)
+	: "cr0", "memory");
+
+	return tmp;
+}
+
+static inline int arch_spin_trylock(arch_spinlock_t *lock)
+{
+	return __arch_spin_trylock(lock) = 0;
+}
+
+/*
+ * On a system with shared processors (that is, where a physical
+ * processor is multiplexed between several virtual processors),
+ * there is no point spinning on a lock if the holder of the lock
+ * isn't currently scheduled on a physical processor.  Instead
+ * we detect this situation and ask the hypervisor to give the
+ * rest of our timeslice to the lock holder.
+ *
+ * So that we can tell which virtual processor is holding a lock,
+ * we put 0x80000000 | smp_processor_id() in the lock when it is
+ * held.  Conveniently, we have a word in the paca that holds this
+ * value.
+ */
+
+#if defined(CONFIG_PPC_SPLPAR)
+/* We only yield to the hypervisor if we are in shared processor mode */
+void splpar_spin_yield(arch_spinlock_t *lock);
+void splpar_rw_yield(arch_rwlock_t *lock);
+#else /* SPLPAR */
+static inline void splpar_spin_yield(arch_spinlock_t *lock) {};
+static inline void splpar_rw_yield(arch_rwlock_t *lock) {};
+#endif
+
+static inline void spin_yield(arch_spinlock_t *lock)
+{
+	if (is_shared_processor())
+		splpar_spin_yield(lock);
+	else
+		barrier();
+}
+
+static inline void rw_yield(arch_rwlock_t *lock)
+{
+	if (is_shared_processor())
+		splpar_rw_yield(lock);
+	else
+		barrier();
+}
+
+static inline void arch_spin_lock(arch_spinlock_t *lock)
+{
+	while (1) {
+		if (likely(__arch_spin_trylock(lock) = 0))
+			break;
+		do {
+			HMT_low();
+			if (is_shared_processor())
+				splpar_spin_yield(lock);
+		} while (unlikely(lock->slock != 0));
+		HMT_medium();
+	}
+}
+
+static inline
+void arch_spin_lock_flags(arch_spinlock_t *lock, unsigned long flags)
+{
+	unsigned long flags_dis;
+
+	while (1) {
+		if (likely(__arch_spin_trylock(lock) = 0))
+			break;
+		local_save_flags(flags_dis);
+		local_irq_restore(flags);
+		do {
+			HMT_low();
+			if (is_shared_processor())
+				splpar_spin_yield(lock);
+		} while (unlikely(lock->slock != 0));
+		HMT_medium();
+		local_irq_restore(flags_dis);
+	}
+}
+#define arch_spin_lock_flags arch_spin_lock_flags
+
+static inline void arch_spin_unlock(arch_spinlock_t *lock)
+{
+	__asm__ __volatile__("# arch_spin_unlock\n\t"
+				PPC_RELEASE_BARRIER: : :"memory");
+	lock->slock = 0;
+}
+
+/*
+ * Read-write spinlocks, allowing multiple readers
+ * but only one writer.
+ *
+ * NOTE! it is quite common to have readers in interrupts
+ * but no interrupt writers. For those circumstances we
+ * can "mix" irq-safe locks - any writer needs to get a
+ * irq-safe write-lock, but readers can get non-irqsafe
+ * read-locks.
+ */
+
+#ifdef CONFIG_PPC64
+#define __DO_SIGN_EXTEND	"extsw	%0,%0\n"
+#define WRLOCK_TOKEN		LOCK_TOKEN	/* it's negative */
+#else
+#define __DO_SIGN_EXTEND
+#define WRLOCK_TOKEN		(-1)
+#endif
+
+/*
+ * This returns the old value in the lock + 1,
+ * so we got a read lock if the return value is > 0.
+ */
+static inline long __arch_read_trylock(arch_rwlock_t *rw)
+{
+	long tmp;
+
+	__asm__ __volatile__(
+"1:	" PPC_LWARX(%0,0,%1,1) "\n"
+	__DO_SIGN_EXTEND
+"	addic.		%0,%0,1\n\
+	ble-		2f\n"
+"	stwcx.		%0,0,%1\n\
+	bne-		1b\n"
+	PPC_ACQUIRE_BARRIER
+"2:"	: "=&r" (tmp)
+	: "r" (&rw->lock)
+	: "cr0", "xer", "memory");
+
+	return tmp;
+}
+
+/*
+ * This returns the old value in the lock,
+ * so we got the write lock if the return value is 0.
+ */
+static inline long __arch_write_trylock(arch_rwlock_t *rw)
+{
+	long tmp, token;
+
+	token = WRLOCK_TOKEN;
+	__asm__ __volatile__(
+"1:	" PPC_LWARX(%0,0,%2,1) "\n\
+	cmpwi		0,%0,0\n\
+	bne-		2f\n"
+"	stwcx.		%1,0,%2\n\
+	bne-		1b\n"
+	PPC_ACQUIRE_BARRIER
+"2:"	: "=&r" (tmp)
+	: "r" (token), "r" (&rw->lock)
+	: "cr0", "memory");
+
+	return tmp;
+}
+
+static inline void arch_read_lock(arch_rwlock_t *rw)
+{
+	while (1) {
+		if (likely(__arch_read_trylock(rw) > 0))
+			break;
+		do {
+			HMT_low();
+			if (is_shared_processor())
+				splpar_rw_yield(rw);
+		} while (unlikely(rw->lock < 0));
+		HMT_medium();
+	}
+}
+
+static inline void arch_write_lock(arch_rwlock_t *rw)
+{
+	while (1) {
+		if (likely(__arch_write_trylock(rw) = 0))
+			break;
+		do {
+			HMT_low();
+			if (is_shared_processor())
+				splpar_rw_yield(rw);
+		} while (unlikely(rw->lock != 0));
+		HMT_medium();
+	}
+}
+
+static inline int arch_read_trylock(arch_rwlock_t *rw)
+{
+	return __arch_read_trylock(rw) > 0;
+}
+
+static inline int arch_write_trylock(arch_rwlock_t *rw)
+{
+	return __arch_write_trylock(rw) = 0;
+}
+
+static inline void arch_read_unlock(arch_rwlock_t *rw)
+{
+	long tmp;
+
+	__asm__ __volatile__(
+	"# read_unlock\n\t"
+	PPC_RELEASE_BARRIER
+"1:	lwarx		%0,0,%1\n\
+	addic		%0,%0,-1\n"
+"	stwcx.		%0,0,%1\n\
+	bne-		1b"
+	: "=&r"(tmp)
+	: "r"(&rw->lock)
+	: "cr0", "xer", "memory");
+}
+
+static inline void arch_write_unlock(arch_rwlock_t *rw)
+{
+	__asm__ __volatile__("# write_unlock\n\t"
+				PPC_RELEASE_BARRIER: : :"memory");
+	rw->lock = 0;
+}
+
+#define arch_spin_relax(lock)	spin_yield(lock)
+#define arch_read_relax(lock)	rw_yield(lock)
+#define arch_write_relax(lock)	rw_yield(lock)
+
+/* See include/linux/spinlock.h */
+#define smp_mb__after_spinlock()   smp_mb()
+
+#endif /* __KERNEL__ */
+#endif /* __ASM_SIMPLE_SPINLOCK_H */
arch/powerpc/include/asm/simple_spinlock_types.h new file +21 -0
new file mode 100644
index 000000000000..7c2b48ce62dc
@@ -0,0 +1,21 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _ASM_POWERPC_SIMPLE_SPINLOCK_TYPES_H
+#define _ASM_POWERPC_SIMPLE_SPINLOCK_TYPES_H
+
+#ifndef __LINUX_SPINLOCK_TYPES_H
+# error "please don't include this file directly"
+#endif
+
+typedef struct {
+	volatile unsigned int slock;
+} arch_spinlock_t;
+
+#define __ARCH_SPIN_LOCK_UNLOCKED	{ 0 }
+
+typedef struct {
+	volatile signed int lock;
+} arch_rwlock_t;
+
+#define __ARCH_RW_LOCK_UNLOCKED		{ 0 }
+
+#endif
arch/powerpc/include/asm/spinlock.h +1 -284

291 lines not on this page · raw message

arch/powerpc/include/asm/spinlock_types.h +1 -11
index 87adaf13b7e8..3906f52dae65 100644
@@ -6,16 +6,6 @@
 # error "please don't include this file directly"
 #endif
 
-typedef struct {
-	volatile unsigned int slock;
-} arch_spinlock_t;
-
-#define __ARCH_SPIN_LOCK_UNLOCKED	{ 0 }
-
-typedef struct {
-	volatile signed int lock;
-} arch_rwlock_t;
-
-#define __ARCH_RW_LOCK_UNLOCKED		{ 0 }
+#include <asm/simple_spinlock_types.h>
 
 #endif
Signature · 1 line
2.23.0
#6 lore raw

Re: [PATCH v3 3/6] powerpc: move spinlock implementation to simple_spinlock

Quoted 19 lines · Nicholas Piggin

To prepare for queued spinlocks. This is a simple rename except to update preprocessor guard name and a file reference.

---
 arch/powerpc/include/asm/simple_spinlock.h    | 292 ++++++++++++++++++
 .../include/asm/simple_spinlock_types.h       |  21 ++
 arch/powerpc/include/asm/spinlock.h           | 285 +----------------
 arch/powerpc/include/asm/spinlock_types.h     |  12 +-
 4 files changed, 315 insertions(+), 295 deletions(-)
 create mode 100644 arch/powerpc/include/asm/simple_spinlock.h
 create mode 100644 arch/powerpc/include/asm/simple_spinlock_types.h
arch/powerpc/include/asm/simple_spinlock.h new file +0 -0
new file mode 100644
index 000000000000..e048c041c4a9

arch/powerpc/include/asm/simple_spinlock.h line 3

Quoted 3 lines
arch/powerpc/include/asm/simple_spinlock.h new file +3 -0
@@ -0,0 +1,292 @@
+/* SPDX-License-Identifier: GPL-2.0-or-later */
+#ifndef __ASM_SIMPLE_SPINLOCK_H
+#define __ASM_SIMPLE_SPINLOCK_H

_ASM_POWERPC_SIMPLE_SPINLOCK_H

arch/powerpc/include/asm/simple_spinlock.h line 4

Quoted 1 line
arch/powerpc/include/asm/simple_spinlock.h new file +1 -0
+#ifdef __KERNEL__

Shouldn't be necessary.

Quoted 7 lines
arch/powerpc/include/asm/simple_spinlock.h new file +7 -0
+/*
+ * Simple spin lock operations.  
+ *
+ * Copyright (C) 2001-2004 Paul Mackerras <[email protected]>, IBM
+ * Copyright (C) 2001 Anton Blanchard <[email protected]>, IBM
+ * Copyright (C) 2002 Dave Engebretsen <[email protected]>, IBM
+ *	Rework to support virtual processors

arch/powerpc/include/asm/simple_spinlock.h line 22

Quoted 10 lines
arch/powerpc/include/asm/simple_spinlock.h new file +10 -0
+ *
+ * Type of int is used as a full 64b word is not necessary.
+ *
+ * (the type definitions are in asm/simple_spinlock_types.h)
+ */
+#include <linux/irqflags.h>
+#include <asm/paravirt.h>
+#ifdef CONFIG_PPC64
+#include <asm/paca.h>
+#endif

I don't think paca.h needs a CONFIG_PPC64 guard, it contains one. I know you're just moving the code, but still nice to cleanup slightly along the way.

cheers

author reviewer
#7 lore raw

PATCH v3 4/6 powerpc/64s: implement queued spinlocks and rwlocks

In reply to Nicholas Piggin

These have shown significantly improved performance and fairness when spinlock contention is moderate to high on very large systems.

[ Numbers hopefully forthcoming after more testing, but initial
results look good ]

Thanks to the fast path, single threaded performance is not noticably hurt.

---
 arch/powerpc/Kconfig                      | 13 ++++++++++++
 arch/powerpc/include/asm/Kbuild           |  2 ++
 arch/powerpc/include/asm/qspinlock.h      | 25 +++++++++++++++++++++++
 arch/powerpc/include/asm/spinlock.h       |  5 +++++
 arch/powerpc/include/asm/spinlock_types.h |  5 +++++
 arch/powerpc/lib/Makefile                 |  3 +++
 include/asm-generic/qspinlock.h           |  2 ++
 7 files changed, 55 insertions(+)
 create mode 100644 arch/powerpc/include/asm/qspinlock.h
arch/powerpc/Kconfig +13 -0
index 24ac85c868db..17663ea57697 100644
@@ -146,6 +146,8 @@ config PPC
 	select ARCH_SUPPORTS_ATOMIC_RMW
 	select ARCH_USE_BUILTIN_BSWAP
 	select ARCH_USE_CMPXCHG_LOCKREF		if PPC64
+	select ARCH_USE_QUEUED_RWLOCKS		if PPC_QUEUED_SPINLOCKS
+	select ARCH_USE_QUEUED_SPINLOCKS	if PPC_QUEUED_SPINLOCKS
 	select ARCH_WANT_IPC_PARSE_VERSION
 	select ARCH_WEAK_RELEASE_ACQUIRE
 	select BINFMT_ELF
@@ -492,6 +494,17 @@ config HOTPLUG_CPU
 
 	  Say N if you are unsure.
 
+config PPC_QUEUED_SPINLOCKS
+	bool "Queued spinlocks"
+	depends on SMP
+	default "y" if PPC_BOOK3S_64
+	help
+	  Say Y here to use to use queued spinlocks which are more complex
+	  but give better salability and fairness on large SMP and NUMA
+	  systems.
+
+	  If unsure, say "Y" if you have lots of cores, otherwise "N".
+
 config ARCH_CPU_PROBE_RELEASE
 	def_bool y
 	depends on HOTPLUG_CPU
arch/powerpc/include/asm/Kbuild +2 -0
index dadbcf3a0b1e..1dd8b6adff5e 100644
@@ -6,5 +6,7 @@ generated-y += syscall_table_spu.h
 generic-y += export.h
 generic-y += local64.h
 generic-y += mcs_spinlock.h
+generic-y += qrwlock.h
+generic-y += qspinlock.h
 generic-y += vtime.h
 generic-y += early_ioremap.h
arch/powerpc/include/asm/qspinlock.h new file +25 -0
new file mode 100644
index 000000000000..c49e33e24edd
@@ -0,0 +1,25 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _ASM_POWERPC_QSPINLOCK_H
+#define _ASM_POWERPC_QSPINLOCK_H
+
+#include <asm-generic/qspinlock_types.h>
+
+#define _Q_PENDING_LOOPS	(1 << 9) /* not tuned */
+
+#define smp_mb__after_spinlock()   smp_mb()
+
+static __always_inline int queued_spin_is_locked(struct qspinlock *lock)
+{
+	/*
+	 * This barrier was added to simple spinlocks by commit 51d7d5205d338,
+	 * but it should now be possible to remove it, asm arm64 has done with
+	 * commit c6f5d02b6a0f.
+	 */
+	smp_mb();
+	return atomic_read(&lock->val);
+}
+#define queued_spin_is_locked queued_spin_is_locked
+
+#include <asm-generic/qspinlock.h>
+
+#endif /* _ASM_POWERPC_QSPINLOCK_H */
arch/powerpc/include/asm/spinlock.h +5 -0
index 21357fe05fe0..434615f1d761 100644
@@ -3,7 +3,12 @@
 #define __ASM_SPINLOCK_H
 #ifdef __KERNEL__
 
+#ifdef CONFIG_PPC_QUEUED_SPINLOCKS
+#include <asm/qspinlock.h>
+#include <asm/qrwlock.h>
+#else
 #include <asm/simple_spinlock.h>
+#endif
 
 #endif /* __KERNEL__ */
 #endif /* __ASM_SPINLOCK_H */
arch/powerpc/include/asm/spinlock_types.h +5 -0
index 3906f52dae65..c5d742f18021 100644
@@ -6,6 +6,11 @@
 # error "please don't include this file directly"
 #endif
 
+#ifdef CONFIG_PPC_QUEUED_SPINLOCKS
+#include <asm-generic/qspinlock_types.h>
+#include <asm-generic/qrwlock_types.h>
+#else
 #include <asm/simple_spinlock_types.h>
+#endif
 
 #endif
arch/powerpc/lib/Makefile +3 -0
index 5e994cda8e40..d66a645503eb 100644
@@ -41,7 +41,10 @@ obj-$(CONFIG_PPC_BOOK3S_64) += copyuser_power7.o copypage_power7.o \
 obj64-y	+= copypage_64.o copyuser_64.o mem_64.o hweight_64.o \
 	   memcpy_64.o memcpy_mcsafe_64.o
 
+ifndef CONFIG_PPC_QUEUED_SPINLOCKS
 obj64-$(CONFIG_SMP)	+= locks.o
+endif
+
 obj64-$(CONFIG_ALTIVEC)	+= vmx-helper.o
 obj64-$(CONFIG_KPROBES_SANITY_TEST)	+= test_emulate_step.o \
 					   test_emulate_step_exec_instr.o
include/asm-generic/qspinlock.h +2 -0
index fde943d180e0..fb0a814d4395 100644
@@ -12,6 +12,7 @@
 
 #include <asm-generic/qspinlock_types.h>
 
+#ifndef queued_spin_is_locked
 /**
  * queued_spin_is_locked - is the spinlock locked?
  * @lock: Pointer to queued spinlock structure
@@ -25,6 +26,7 @@ static __always_inline int queued_spin_is_locked(struct qspinlock *lock)
 	 */
 	return atomic_read(&lock->val);
 }
+#endif
 
 /**
  * queued_spin_value_unlocked - is the spinlock structure unlocked?
Signature · 1 line
2.23.0
#8 lore raw

Re: [PATCH v3 4/6] powerpc/64s: implement queued spinlocks and rwlocks

Quoted 5 lines · Nicholas Piggin

These have shown significantly improved performance and fairness when spinlock contention is moderate to high on very large systems.

[ Numbers hopefully forthcoming after more testing, but initial
results look good ]

Would be good to have something here, even if it's preliminary.

Quoted 12 lines

Thanks to the fast path, single threaded performance is not noticably hurt.

---
 arch/powerpc/Kconfig                      | 13 ++++++++++++
 arch/powerpc/include/asm/Kbuild           |  2 ++
 arch/powerpc/include/asm/qspinlock.h      | 25 +++++++++++++++++++++++
 arch/powerpc/include/asm/spinlock.h       |  5 +++++
 arch/powerpc/include/asm/spinlock_types.h |  5 +++++
 arch/powerpc/lib/Makefile                 |  3 +++
Quoted 1 line
 include/asm-generic/qspinlock.h           |  2 ++

Who's ack do we need for that part?

arch/powerpc/Kconfig config HOTPLUG_CPU · line 500

Quoted 12 lines
arch/powerpc/Kconfig +4 -0
index 24ac85c868db..17663ea57697 100644
@@ -492,6 +494,17 @@ config HOTPLUG_CPU
 
 	  Say N if you are unsure.
 
+config PPC_QUEUED_SPINLOCKS
+	bool "Queued spinlocks"
+	depends on SMP
+	default "y" if PPC_BOOK3S_64

Not sure about default y? At least until we've got a better idea of the perf impact on a range of small/big new/old systems.

arch/powerpc/Kconfig config HOTPLUG_CPU · line 506

Quoted 6 lines
arch/powerpc/Kconfig +6 -0
+	help
+	  Say Y here to use to use queued spinlocks which are more complex
+	  but give better salability and fairness on large SMP and NUMA
+	  systems.
+
+	  If unsure, say "Y" if you have lots of cores, otherwise "N".

Would be nice if we could give a range for "lots".

arch/powerpc/include/asm/Kbuild generated-y += syscall_table_spu.h · line 10

Quoted 10 lines
arch/powerpc/include/asm/Kbuild +2 -0
index dadbcf3a0b1e..1dd8b6adff5e 100644
@@ -6,5 +6,7 @@ generated-y += syscall_table_spu.h
 generic-y += export.h
 generic-y += local64.h
 generic-y += mcs_spinlock.h
+generic-y += qrwlock.h
+generic-y += qspinlock.h

The 2nd line spits a warning about a redundant entry. I think you want to just drop it.

cheers

maintainer Acked-by
#9 lore raw

Re: [PATCH v3 4/6] powerpc/64s: implement queued spinlocks and rwlocks

Quoted 21 lines · Michael Ellerman
Quoted 5 lines · Nicholas Piggin

These have shown significantly improved performance and fairness when spinlock contention is moderate to high on very large systems.

[ Numbers hopefully forthcoming after more testing, but initial
results look good ]

Would be good to have something here, even if it's preliminary.

Quoted 11 lines

Thanks to the fast path, single threaded performance is not noticably hurt.

---
 arch/powerpc/Kconfig                      | 13 ++++++++++++
 arch/powerpc/include/asm/Kbuild           |  2 ++
 arch/powerpc/include/asm/qspinlock.h      | 25 +++++++++++++++++++++++
 arch/powerpc/include/asm/spinlock.h       |  5 +++++
 arch/powerpc/include/asm/spinlock_types.h |  5 +++++
 arch/powerpc/lib/Makefile                 |  3 +++
Quoted 3 lines
Quoted 1 line
 include/asm-generic/qspinlock.h           |  2 ++

Who's ack do we need for that part?

Mine I suppose would do, as discussed earlier, it probably isn't required anymore, but I understand the paranoia of not wanting to change too many things at once :-)

#10 lore raw

Re: [PATCH v3 4/6] powerpc/64s: implement queued spinlocks and rwlocks

In reply to Nicholas Piggin
Quoted 35 lines · Nicholas Piggin

These have shown significantly improved performance and fairness when spinlock contention is moderate to high on very large systems.

[ Numbers hopefully forthcoming after more testing, but initial
results look good ]

Thanks to the fast path, single threaded performance is not noticably hurt.

---
 arch/powerpc/Kconfig                      | 13 ++++++++++++
 arch/powerpc/include/asm/Kbuild           |  2 ++
 arch/powerpc/include/asm/qspinlock.h      | 25 +++++++++++++++++++++++
 arch/powerpc/include/asm/spinlock.h       |  5 +++++
 arch/powerpc/include/asm/spinlock_types.h |  5 +++++
 arch/powerpc/lib/Makefile                 |  3 +++
 include/asm-generic/qspinlock.h           |  2 ++
 7 files changed, 55 insertions(+)
 create mode 100644 arch/powerpc/include/asm/qspinlock.h
arch/powerpc/Kconfig +2 -0
index 24ac85c868db..17663ea57697 100644
@@ -146,6 +146,8 @@ config PPC
 	select ARCH_SUPPORTS_ATOMIC_RMW
 	select ARCH_USE_BUILTIN_BSWAP
 	select ARCH_USE_CMPXCHG_LOCKREF		if PPC64
+	select ARCH_USE_QUEUED_RWLOCKS		if PPC_QUEUED_SPINLOCKS
+	select ARCH_USE_QUEUED_SPINLOCKS	if PPC_QUEUED_SPINLOCKS
 	select ARCH_WANT_IPC_PARSE_VERSION
 	select ARCH_WEAK_RELEASE_ACQUIRE
 	select BINFMT_ELF

arch/powerpc/Kconfig config HOTPLUG_CPU · line 503

Quoted 10 lines
arch/powerpc/Kconfig +7 -0
@@ -492,6 +494,17 @@ config HOTPLUG_CPU
 
 	  Say N if you are unsure.
 
+config PPC_QUEUED_SPINLOCKS
+	bool "Queued spinlocks"
+	depends on SMP
+	default "y" if PPC_BOOK3S_64
+	help
+	  Say Y here to use to use queued spinlocks which are more complex
+	  but give better salability and fairness on large SMP and NUMA
                           ^ +c?

Thanks

Michal

Quoted 121 lines
arch/powerpc/Kconfig +4 -0
+	  systems.
+
+	  If unsure, say "Y" if you have lots of cores, otherwise "N".
+
 config ARCH_CPU_PROBE_RELEASE
 	def_bool y
 	depends on HOTPLUG_CPU
arch/powerpc/include/asm/Kbuild +2 -0
index dadbcf3a0b1e..1dd8b6adff5e 100644
@@ -6,5 +6,7 @@ generated-y += syscall_table_spu.h
 generic-y += export.h
 generic-y += local64.h
 generic-y += mcs_spinlock.h
+generic-y += qrwlock.h
+generic-y += qspinlock.h
 generic-y += vtime.h
 generic-y += early_ioremap.h
arch/powerpc/include/asm/qspinlock.h new file +25 -0
new file mode 100644
index 000000000000..c49e33e24edd
@@ -0,0 +1,25 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _ASM_POWERPC_QSPINLOCK_H
+#define _ASM_POWERPC_QSPINLOCK_H
+
+#include <asm-generic/qspinlock_types.h>
+
+#define _Q_PENDING_LOOPS	(1 << 9) /* not tuned */
+
+#define smp_mb__after_spinlock()   smp_mb()
+
+static __always_inline int queued_spin_is_locked(struct qspinlock *lock)
+{
+	/*
+	 * This barrier was added to simple spinlocks by commit 51d7d5205d338,
+	 * but it should now be possible to remove it, asm arm64 has done with
+	 * commit c6f5d02b6a0f.
+	 */
+	smp_mb();
+	return atomic_read(&lock->val);
+}
+#define queued_spin_is_locked queued_spin_is_locked
+
+#include <asm-generic/qspinlock.h>
+
+#endif /* _ASM_POWERPC_QSPINLOCK_H */
arch/powerpc/include/asm/spinlock.h +5 -0
index 21357fe05fe0..434615f1d761 100644
@@ -3,7 +3,12 @@
 #define __ASM_SPINLOCK_H
 #ifdef __KERNEL__
 
+#ifdef CONFIG_PPC_QUEUED_SPINLOCKS
+#include <asm/qspinlock.h>
+#include <asm/qrwlock.h>
+#else
 #include <asm/simple_spinlock.h>
+#endif
 
 #endif /* __KERNEL__ */
 #endif /* __ASM_SPINLOCK_H */
arch/powerpc/include/asm/spinlock_types.h +5 -0
index 3906f52dae65..c5d742f18021 100644
@@ -6,6 +6,11 @@
 # error "please don't include this file directly"
 #endif
 
+#ifdef CONFIG_PPC_QUEUED_SPINLOCKS
+#include <asm-generic/qspinlock_types.h>
+#include <asm-generic/qrwlock_types.h>
+#else
 #include <asm/simple_spinlock_types.h>
+#endif
 
 #endif
arch/powerpc/lib/Makefile +3 -0
index 5e994cda8e40..d66a645503eb 100644
@@ -41,7 +41,10 @@ obj-$(CONFIG_PPC_BOOK3S_64) += copyuser_power7.o copypage_power7.o \
 obj64-y	+= copypage_64.o copyuser_64.o mem_64.o hweight_64.o \
 	   memcpy_64.o memcpy_mcsafe_64.o
 
+ifndef CONFIG_PPC_QUEUED_SPINLOCKS
 obj64-$(CONFIG_SMP)	+= locks.o
+endif
+
 obj64-$(CONFIG_ALTIVEC)	+= vmx-helper.o
 obj64-$(CONFIG_KPROBES_SANITY_TEST)	+= test_emulate_step.o \
 					   test_emulate_step_exec_instr.o
include/asm-generic/qspinlock.h +2 -0
index fde943d180e0..fb0a814d4395 100644
@@ -12,6 +12,7 @@
 
 #include <asm-generic/qspinlock_types.h>
 
+#ifndef queued_spin_is_locked
 /**
  * queued_spin_is_locked - is the spinlock locked?
  * @lock: Pointer to queued spinlock structure
@@ -25,6 +26,7 @@ static __always_inline int queued_spin_is_locked(struct qspinlock *lock)
 	 */
 	return atomic_read(&lock->val);
 }
+#endif
 
 /**
  * queued_spin_value_unlocked - is the spinlock structure unlocked?
Signature · 1 line
2.23.0
author reviewer
#11 lore raw

PATCH v3 5/6 powerpc/pseries: implement paravirt qspinlocks for SPLPAR

In reply to Nicholas Piggin
---
 arch/powerpc/include/asm/paravirt.h           | 28 ++++++++
 arch/powerpc/include/asm/qspinlock.h          | 66 +++++++++++++++++++
 arch/powerpc/include/asm/qspinlock_paravirt.h |  7 ++
 arch/powerpc/platforms/pseries/Kconfig        |  5 ++
 arch/powerpc/platforms/pseries/setup.c        |  6 +-
 include/asm-generic/qspinlock.h               |  2 +
 6 files changed, 113 insertions(+), 1 deletion(-)
 create mode 100644 arch/powerpc/include/asm/qspinlock_paravirt.h
arch/powerpc/include/asm/paravirt.h +28 -0
index 7a8546660a63..f2d51f929cf5 100644
@@ -29,6 +29,16 @@ static inline void yield_to_preempted(int cpu, u32 yield_count)
 {
 	plpar_hcall_norets(H_CONFER, get_hard_smp_processor_id(cpu), yield_count);
 }
+
+static inline void prod_cpu(int cpu)
+{
+	plpar_hcall_norets(H_PROD, get_hard_smp_processor_id(cpu));
+}
+
+static inline void yield_to_any(void)
+{
+	plpar_hcall_norets(H_CONFER, -1, 0);
+}
 #else
 static inline bool is_shared_processor(void)
 {
@@ -45,6 +55,19 @@ static inline void yield_to_preempted(int cpu, u32 yield_count)
 {
 	___bad_yield_to_preempted(); /* This would be a bug */
 }
+
+extern void ___bad_yield_to_any(void);
+static inline void yield_to_any(void)
+{
+	___bad_yield_to_any(); /* This would be a bug */
+}
+
+extern void ___bad_prod_cpu(void);
+static inline void prod_cpu(int cpu)
+{
+	___bad_prod_cpu(); /* This would be a bug */
+}
+
 #endif
 
 #define vcpu_is_preempted vcpu_is_preempted
@@ -57,5 +80,10 @@ static inline bool vcpu_is_preempted(int cpu)
 	return false;
 }
 
+static inline bool pv_is_native_spin_unlock(void)
+{
+     return !is_shared_processor();
+}
+
 #endif /* __KERNEL__ */
 #endif /* __ASM_PARAVIRT_H */
arch/powerpc/include/asm/qspinlock.h +66 -0

81 lines not on this page · raw message

arch/powerpc/include/asm/qspinlock_paravirt.h new file +7 -0
new file mode 100644
index 000000000000..750d1b5e0202
@@ -0,0 +1,7 @@
+/* SPDX-License-Identifier: GPL-2.0-or-later */
+#ifndef __ASM_QSPINLOCK_PARAVIRT_H
+#define __ASM_QSPINLOCK_PARAVIRT_H
+
+EXPORT_SYMBOL(__pv_queued_spin_unlock);
+
+#endif /* __ASM_QSPINLOCK_PARAVIRT_H */
arch/powerpc/platforms/pseries/Kconfig +5 -0
index 24c18362e5ea..756e727b383f 100644
@@ -25,9 +25,14 @@ config PPC_PSERIES
 	select SWIOTLB
 	default y
 
+config PARAVIRT_SPINLOCKS
+	bool
+	default n
+
 config PPC_SPLPAR
 	depends on PPC_PSERIES
 	bool "Support for shared-processor logical partitions"
+	select PARAVIRT_SPINLOCKS if PPC_QUEUED_SPINLOCKS
 	help
 	  Enabling this option will make the kernel run more efficiently
 	  on logically-partitioned pSeries systems which use shared
arch/powerpc/platforms/pseries/setup.c +5 -1
index 2db8469e475f..747a203d9453 100644
@@ -771,8 +771,12 @@ static void __init pSeries_setup_arch(void)
 	if (firmware_has_feature(FW_FEATURE_LPAR)) {
 		vpa_init(boot_cpuid);
 
-		if (lppaca_shared_proc(get_lppaca()))
+		if (lppaca_shared_proc(get_lppaca())) {
 			static_branch_enable(&shared_processor);
+#ifdef CONFIG_PARAVIRT_SPINLOCKS
+			pv_spinlocks_init();
+#endif
+		}
 
 		ppc_md.power_save = pseries_lpar_idle;
 		ppc_md.enable_pmcs = pseries_lpar_enable_pmcs;
include/asm-generic/qspinlock.h +2 -0
index fb0a814d4395..38ca14e79a86 100644
@@ -69,6 +69,7 @@ static __always_inline int queued_spin_trylock(struct qspinlock *lock)
 
 extern void queued_spin_lock_slowpath(struct qspinlock *lock, u32 val);
 
+#ifndef queued_spin_lock
 /**
  * queued_spin_lock - acquire a queued spinlock
  * @lock: Pointer to queued spinlock structure
@@ -82,6 +83,7 @@ static __always_inline void queued_spin_lock(struct qspinlock *lock)
 
 	queued_spin_lock_slowpath(lock, val);
 }
+#endif
 
 #ifndef queued_spin_unlock
 /**
Signature · 1 line
2.23.0
#12 lore raw

Re: [PATCH v3 5/6] powerpc/pseries: implement paravirt qspinlocks for SPLPAR

Quoted 8 lines · Nicholas Piggin
---
 arch/powerpc/include/asm/paravirt.h           | 28 ++++++++
 arch/powerpc/include/asm/qspinlock.h          | 66 +++++++++++++++++++
 arch/powerpc/include/asm/qspinlock_paravirt.h |  7 ++
 arch/powerpc/platforms/pseries/Kconfig        |  5 ++
 arch/powerpc/platforms/pseries/setup.c        |  6 +-
 include/asm-generic/qspinlock.h               |  2 +

Another ack?

Quoted 5 lines
arch/powerpc/include/asm/paravirt.h +0 -0
index 7a8546660a63..f2d51f929cf5 100644

arch/powerpc/include/asm/paravirt.h static inline void yield_to_preempted(int cpu, u32 yield_count) · line 63

Quoted 9 lines
arch/powerpc/include/asm/paravirt.h +6 -0
@@ -45,6 +55,19 @@ static inline void yield_to_preempted(int cpu, u32 yield_count)
 {
 	___bad_yield_to_preempted(); /* This would be a bug */
 }
+
+extern void ___bad_yield_to_any(void);
+static inline void yield_to_any(void)
+{
+	___bad_yield_to_any(); /* This would be a bug */
+}

Why do we do that rather than just not defining yield_to_any() at all and letting the build fail on that?

There's a condition somewhere that we know will false at compile time and drop the call before linking?

arch/powerpc/include/asm/qspinlock_paravirt.h line 3

Quoted 9 lines
arch/powerpc/include/asm/qspinlock_paravirt.h new file +3 -0
new file mode 100644
index 000000000000..750d1b5e0202
@@ -0,0 +1,7 @@
+/* SPDX-License-Identifier: GPL-2.0-or-later */
+#ifndef __ASM_QSPINLOCK_PARAVIRT_H
+#define __ASM_QSPINLOCK_PARAVIRT_H

_ASM_POWERPC_QSPINLOCK_PARAVIRT_H please.

arch/powerpc/include/asm/qspinlock_paravirt.h line 5

Quoted 2 lines
arch/powerpc/include/asm/qspinlock_paravirt.h new file +2 -0
+
+EXPORT_SYMBOL(__pv_queued_spin_unlock);

Why's that in a header? Should that (eventually) go with the generic implementation?

arch/powerpc/platforms/pseries/Kconfig config PPC_PSERIES · line 30

Quoted 11 lines
arch/powerpc/platforms/pseries/Kconfig +3 -0
index 24c18362e5ea..756e727b383f 100644
@@ -25,9 +25,14 @@ config PPC_PSERIES
 	select SWIOTLB
 	default y
 
+config PARAVIRT_SPINLOCKS
+	bool
+	default n

default n is the default.

Quoted 5 lines
arch/powerpc/platforms/pseries/setup.c +0 -0
index 2db8469e475f..747a203d9453 100644

arch/powerpc/platforms/pseries/setup.c static void __init pSeries_setup_arch(void) · line 779

Quoted 10 lines
arch/powerpc/platforms/pseries/setup.c +5 -1
@@ -771,8 +771,12 @@ static void __init pSeries_setup_arch(void)
 	if (firmware_has_feature(FW_FEATURE_LPAR)) {
 		vpa_init(boot_cpuid);
 
-		if (lppaca_shared_proc(get_lppaca()))
+		if (lppaca_shared_proc(get_lppaca())) {
 			static_branch_enable(&shared_processor);
+#ifdef CONFIG_PARAVIRT_SPINLOCKS
+			pv_spinlocks_init();
+#endif
+		}

We could avoid the ifdef with this I think?

arch/powerpc/include/asm/spinlock.h +4 -0
index 434615f1d761..6ec72282888d 100644
@@ -10,5 +10,9 @@
 #include <asm/simple_spinlock.h>
 #endif
 
+#ifndef CONFIG_PARAVIRT_SPINLOCKS
+static inline void pv_spinlocks_init(void) { }
+#endif
+
 #endif /* __KERNEL__ */
 #endif /* __ASM_SPINLOCK_H */

cheers

maintainer Acked-by
#13 lore raw

Re: [PATCH v3 5/6] powerpc/pseries: implement paravirt qspinlocks for SPLPAR

Quoted 11 lines · Michael Ellerman
Quoted 8 lines · Nicholas Piggin
---
 arch/powerpc/include/asm/paravirt.h           | 28 ++++++++
 arch/powerpc/include/asm/qspinlock.h          | 66 +++++++++++++++++++
 arch/powerpc/include/asm/qspinlock_paravirt.h |  7 ++
 arch/powerpc/platforms/pseries/Kconfig        |  5 ++
 arch/powerpc/platforms/pseries/setup.c        |  6 +-
 include/asm-generic/qspinlock.h               |  2 +
Quoted 1 line

Another ack?

reviewer Acked-by
#14 lore raw

Re: [PATCH v3 5/6] powerpc/pseries: implement paravirt qspinlocks for SPLPAR

In reply to Michael Ellerman
Quoted 10 lines · Michael Ellerman
Quoted 8 lines · Nicholas Piggin
---
  arch/powerpc/include/asm/paravirt.h           | 28 ++++++++
  arch/powerpc/include/asm/qspinlock.h          | 66 +++++++++++++++++++
  arch/powerpc/include/asm/qspinlock_paravirt.h |  7 ++
  arch/powerpc/platforms/pseries/Kconfig        |  5 ++
  arch/powerpc/platforms/pseries/setup.c        |  6 +-
  include/asm-generic/qspinlock.h               |  2 +
Quoted 2 lines

Another ack?

I am OK with adding the #ifdef around queued_spin_lock().

Quoted 31 lines
Quoted 5 lines
arch/powerpc/include/asm/paravirt.h +0 -0
index 7a8546660a63..f2d51f929cf5 100644

arch/powerpc/include/asm/paravirt.h static inline void yield_to_preempted(int cpu, u32 yield_count) · line 63

Quoted 9 lines
arch/powerpc/include/asm/paravirt.h +6 -0
@@ -45,6 +55,19 @@ static inline void yield_to_preempted(int cpu, u32 yield_count)
  {
  	___bad_yield_to_preempted(); /* This would be a bug */
  }
+
+extern void ___bad_yield_to_any(void);
+static inline void yield_to_any(void)
+{
+	___bad_yield_to_any(); /* This would be a bug */
+}

Why do we do that rather than just not defining yield_to_any() at all and letting the build fail on that?

There's a condition somewhere that we know will false at compile time and drop the call before linking?

arch/powerpc/include/asm/qspinlock_paravirt.h line 3

Quoted 9 lines
arch/powerpc/include/asm/qspinlock_paravirt.h new file +3 -0
new file mode 100644
index 000000000000..750d1b5e0202
@@ -0,0 +1,7 @@
+/* SPDX-License-Identifier: GPL-2.0-or-later */
+#ifndef __ASM_QSPINLOCK_PARAVIRT_H
+#define __ASM_QSPINLOCK_PARAVIRT_H

_ASM_POWERPC_QSPINLOCK_PARAVIRT_H please.

Quoted 3 lines

arch/powerpc/include/asm/qspinlock_paravirt.h line 5

Quoted 2 lines
arch/powerpc/include/asm/qspinlock_paravirt.h new file +2 -0
+
+EXPORT_SYMBOL(__pv_queued_spin_unlock);

Why's that in a header? Should that (eventually) go with the generic implementation?

The PV qspinlock implementation is not that generic at the moment. Even though native qspinlock is used by a number of archs, PV qspinlock is only currently used in x86. This is certainly an area that needs improvement.

Quoted 46 lines

arch/powerpc/platforms/pseries/Kconfig config PPC_PSERIES · line 30

Quoted 11 lines
arch/powerpc/platforms/pseries/Kconfig +3 -0
index 24c18362e5ea..756e727b383f 100644
@@ -25,9 +25,14 @@ config PPC_PSERIES
  	select SWIOTLB
  	default y
 
+config PARAVIRT_SPINLOCKS
+	bool
+	default n

default n is the default.

Quoted 5 lines
arch/powerpc/platforms/pseries/setup.c +0 -0
index 2db8469e475f..747a203d9453 100644

arch/powerpc/platforms/pseries/setup.c static void __init pSeries_setup_arch(void) · line 779

Quoted 10 lines
arch/powerpc/platforms/pseries/setup.c +5 -1
@@ -771,8 +771,12 @@ static void __init pSeries_setup_arch(void)
  	if (firmware_has_feature(FW_FEATURE_LPAR)) {
  		vpa_init(boot_cpuid);
 
-		if (lppaca_shared_proc(get_lppaca()))
+		if (lppaca_shared_proc(get_lppaca())) {
  			static_branch_enable(&shared_processor);
+#ifdef CONFIG_PARAVIRT_SPINLOCKS
+			pv_spinlocks_init();
+#endif
+		}

We could avoid the ifdef with this I think?

arch/powerpc/include/asm/spinlock.h +4 -0
index 434615f1d761..6ec72282888d 100644
@@ -10,5 +10,9 @@
  #include <asm/simple_spinlock.h>
  #endif
 
+#ifndef CONFIG_PARAVIRT_SPINLOCKS
+static inline void pv_spinlocks_init(void) { }
+#endif
+
  #endif /* __KERNEL__ */
  #endif /* __ASM_SPINLOCK_H */
Quoted 2 lines

cheers

We don't really need to do a pv_spinlocks_init() if pv_kick() isn't supported.

Cheers,
Longman

maintainer
#15 lore raw

Re: [PATCH v3 5/6] powerpc/pseries: implement paravirt qspinlocks for SPLPAR

Quoted 2 lines · Waiman Long

We don't really need to do a pv_spinlocks_init() if pv_kick() isn't supported.

Waiman, if you cannot explain how not having kick is a sane thing, what are you saying here?

reviewer
#16 lore raw

Re: [PATCH v3 5/6] powerpc/pseries: implement paravirt qspinlocks for SPLPAR

Quoted 6 lines · Peter Zijlstra
Quoted 2 lines · Waiman Long

We don't really need to do a pv_spinlocks_init() if pv_kick() isn't supported.

Waiman, if you cannot explain how not having kick is a sane thing, what are you saying here?

The current PPC paravirt spinlock code doesn't do any cpu kick. It does an equivalence of pv_wait by yielding the cpu to the lock holder only. The pv_spinlocks_init() is for setting up the hash table for doing pv_kick. If we don't need to do pv_kick, we don't need the hash table.

I am not saying that pv_kick is not needed for the PPC environment. I was just trying to adapt the pvqspinlock code to such an environment first. Further investigation on how to implement some kind of pv_kick will be something that we may want to do as a follow on.

BTW, do you have any comment on my v2 lock holder cpu info qspinlock patch? I will have to update the patch to fix the reported 0-day test problem, but I want to collect other feedback before sending out v3.

Cheers,
Longman

#17 lore raw

Re: [PATCH v3 5/6] powerpc/pseries: implement paravirt qspinlocks for SPLPAR

Quoted 3 lines · Waiman Long

BTW, do you have any comment on my v2 lock holder cpu info qspinlock patch? I will have to update the patch to fix the reported 0-day test problem, but I want to collect other feedback before sending out v3.

I want to say I hate it all, it adds instructions to a path we spend an aweful lot of time optimizing without really getting anything back for it.

Will, how do you feel about it?

reviewer
#18 lore raw

Re: [PATCH v3 5/6] powerpc/pseries: implement paravirt qspinlocks for SPLPAR

Quoted 7 lines · On 7/23/20 2:47 PM, [email protected] wrote:
Quoted 3 lines · Waiman Long

BTW, do you have any comment on my v2 lock holder cpu info qspinlock patch? I will have to update the patch to fix the reported 0-day test problem, but I want to collect other feedback before sending out v3.

I want to say I hate it all, it adds instructions to a path we spend an aweful lot of time optimizing without really getting anything back for it.

It does add some extra instruction that may slow it down slightly, but I don't agree that it gives nothing back. The cpu lock holder information can be useful in analyzing crash dumps and in some debugging situation. I think it can be useful in RHEL for this readon. How about an x86 config option to allow distros to decide if they want to have it enabled? I will make sure that it will have no performance degradation if the option is not enabled.

Cheers,
Longman

#19 lore raw

Re: [PATCH v3 5/6] powerpc/pseries: implement paravirt qspinlocks for SPLPAR

Quoted 9 lines · Waiman Long
Quoted 7 lines · On 7/23/20 2:47 PM, [email protected] wrote:
Quoted 3 lines · Waiman Long

BTW, do you have any comment on my v2 lock holder cpu info qspinlock patch? I will have to update the patch to fix the reported 0-day test problem, but I want to collect other feedback before sending out v3.

I want to say I hate it all, it adds instructions to a path we spend an aweful lot of time optimizing without really getting anything back for it.

Quoted 6 lines

It does add some extra instruction that may slow it down slightly, but I don't agree that it gives nothing back. The cpu lock holder information can be useful in analyzing crash dumps and in some debugging situation. I think it can be useful in RHEL for this readon. How about an x86 config option to allow distros to decide if they want to have it enabled? I will make sure that it will have no performance degradation if the option is not enabled.

Config knobs suck too; they create a maintenance burden (we get to make sure all the permutations works/build/etc..) and effectively nobody uses them, since world+dog uses what distros pick.

Anyway, instead of adding a second per-cpu variable, can you see how horrible something like this is:

unsigned char adds(unsigned char var, unsigned char val)

{
	unsigned short sat = 0xff, tmp = var;

	asm ("addb	%[val], %b[var];"
	     "cmovc	%[sat], %[var];"
	     : [var] "+r" (tmp)
	     : [val] "ir" (val), [sat] "r" (sat)
	     );

	return tmp;
}

Another thing to try is, instead of threading that lockval throughout the thing, simply:

#define _Q_LOCKED_VAL this_cpu_read_stable(cpu_sat)

or combined with the above

#define _Q_LOCKED_VAL adds(this_cpu_read_stable(cpu_number), 2)

and see if the compiler really makes a mess of things.

#20 lore raw

Re: [PATCH v3 5/6] powerpc/pseries: implement paravirt qspinlocks for SPLPAR

Quoted 5 lines · On Thu, Jul 23, 2020 at 09:58:55PM +0200, [email protected] wrote:
	asm ("addb	%[val], %b[var];"
	     "cmovc	%[sat], %[var];"
	     : [var] "+r" (tmp)
	     : [val] "ir" (val), [sat] "r" (sat)
	     );

"var" (operand 0) needs an earlyclobber ("sat" is read after "var" is written for the first time).

Segher

reviewer
#21 lore raw

Re: [PATCH v3 5/6] powerpc/pseries: implement paravirt qspinlocks for SPLPAR

In reply to [email protected]
Quoted 38 lines · On 7/23/20 3:58 PM, [email protected] wrote:
Quoted 8 lines · Waiman Long
Quoted 7 lines · On 7/23/20 2:47 PM, [email protected] wrote:
Quoted 3 lines · Waiman Long

BTW, do you have any comment on my v2 lock holder cpu info qspinlock patch? I will have to update the patch to fix the reported 0-day test problem, but I want to collect other feedback before sending out v3.

I want to say I hate it all, it adds instructions to a path we spend an aweful lot of time optimizing without really getting anything back for it.

Quoted 6 lines

It does add some extra instruction that may slow it down slightly, but I don't agree that it gives nothing back. The cpu lock holder information can be useful in analyzing crash dumps and in some debugging situation. I think it can be useful in RHEL for this readon. How about an x86 config option to allow distros to decide if they want to have it enabled? I will make sure that it will have no performance degradation if the option is not enabled.

Config knobs suck too; they create a maintenance burden (we get to make sure all the permutations works/build/etc..) and effectively nobody uses them, since world+dog uses what distros pick.

Anyway, instead of adding a second per-cpu variable, can you see how horrible something like this is:

unsigned char adds(unsigned char var, unsigned char val)

{
	unsigned short sat = 0xff, tmp = var;

	asm ("addb	%[val], %b[var];"
	     "cmovc	%[sat], %[var];"
	     : [var] "+r" (tmp)
	     : [val] "ir" (val), [sat] "r" (sat)
	     );

	return tmp;
}

Another thing to try is, instead of threading that lockval throughout the thing, simply:

Quoted 8 lines

#define _Q_LOCKED_VAL this_cpu_read_stable(cpu_sat)

or combined with the above

#define _Q_LOCKED_VAL adds(this_cpu_read_stable(cpu_number), 2)

and see if the compiler really makes a mess of things.

Thanks for the suggestion. I will try that out.

Cheers,
Longman

maintainer
#22 lore raw

Re: [PATCH v3 5/6] powerpc/pseries: implement paravirt qspinlocks for SPLPAR

In reply to [email protected]
Quoted 10 lines · On Thu, Jul 23, 2020 at 08:47:59PM +0200, [email protected] wrote:
Quoted 3 lines · Waiman Long

BTW, do you have any comment on my v2 lock holder cpu info qspinlock patch? I will have to update the patch to fix the reported 0-day test problem, but I want to collect other feedback before sending out v3.

I want to say I hate it all, it adds instructions to a path we spend an aweful lot of time optimizing without really getting anything back for it.

Will, how do you feel about it?

I can see it potentially being useful for debugging, but I hate the limitation to 256 CPUs. Even arm64 is hitting that now.

Also, you're talking ~1% gains here. I think our collective time would be better spent off reviewing the CNA series and trying to make it more deterministic.

Will

reviewer
#23 lore raw

Re: [PATCH v3 5/6] powerpc/pseries: implement paravirt qspinlocks for SPLPAR

Quoted 5 lines · Will Deacon
Quoted 4 lines · On Thu, Jul 23, 2020 at 08:47:59PM +0200, [email protected] wrote:
Quoted 3 lines · Waiman Long

BTW, do you have any comment on my v2 lock holder cpu info qspinlock patch? I will have to update the patch to fix the reported 0-day test problem, but I want to collect other feedback before sending out v3.

Quoted 7 lines
Quoted 5 lines

I want to say I hate it all, it adds instructions to a path we spend an aweful lot of time optimizing without really getting anything back for it.

Will, how do you feel about it?

I can see it potentially being useful for debugging, but I hate the limitation to 256 CPUs. Even arm64 is hitting that now.

After thinking more about that, I think we can use all the remaining bits in the 16-bit locked_pending. Reserving 1 bit for locked and 1 bit for pending, there are 14 bits left. So as long as NR_CPUS < 16k (requirement for 16-bit locked_pending), we can put all possible cpu numbers into the lock. We can also just use smp_processor_id() without additional percpu data.

Quoted 4 lines

Also, you're talking ~1% gains here. I think our collective time would be better spent off reviewing the CNA series and trying to make it more deterministic.

I thought you guys are not interested in CNA. I do want to get CNA merged, if possible. Let review the current version again and see if there are ways we can further improve it.

Cheers,
Longman

reviewer
#24 lore raw

Re: [PATCH v3 5/6] powerpc/pseries: implement paravirt qspinlocks for SPLPAR

Quoted 16 lines · Waiman Long
Quoted 7 lines · Will Deacon
Quoted 6 lines · On Thu, Jul 23, 2020 at 08:47:59PM +0200, [email protected] wrote:
Quoted 5 lines · Waiman Long

BTW, do you have any comment on my v2 lock holder cpu info qspinlock patch?
I will have to update the patch to fix the reported 0-day test problem, but
I want to collect other feedback before sending out v3.

Quoted 7 lines
Quoted 5 lines

I want to say I hate it all, it adds instructions to a path we spend an aweful lot of time optimizing without really getting anything back for it.

Will, how do you feel about it?

I can see it potentially being useful for debugging, but I hate the limitation to 256 CPUs. Even arm64 is hitting that now.

Quoted 6 lines

After thinking more about that, I think we can use all the remaining bits in the 16-bit locked_pending. Reserving 1 bit for locked and 1 bit for pending, there are 14 bits left. So as long as NR_CPUS < 16k (requirement for 16-bit locked_pending), we can put all possible cpu numbers into the lock. We can also just use smp_processor_id() without additional percpu data.

Sorry, that doesn't work. The extra bits in the pending byte won't get cleared on unlock. That will have noticeable performance impact. Clearing the pending byte on unlock will cause other performance problem. So I guess we will have to limit the cpu number in the locked byte.

Regards,
Longman

maintainer
#25 lore raw

Re: [PATCH v3 5/6] powerpc/pseries: implement paravirt qspinlocks for SPLPAR

In reply to Waiman Long
Quoted 11 lines · Waiman Long
Quoted 10 lines · Will Deacon
Quoted 9 lines · On Thu, Jul 23, 2020 at 08:47:59PM +0200, [email protected] wrote:
Quoted 3 lines · Waiman Long

BTW, do you have any comment on my v2 lock holder cpu info qspinlock patch? I will have to update the patch to fix the reported 0-day test problem, but I want to collect other feedback before sending out v3.

I want to say I hate it all, it adds instructions to a path we spend an aweful lot of time optimizing without really getting anything back for it.

Will, how do you feel about it?

Quoted 8 lines
Quoted 2 lines

I can see it potentially being useful for debugging, but I hate the limitation to 256 CPUs. Even arm64 is hitting that now.

After thinking more about that, I think we can use all the remaining bits in the 16-bit locked_pending. Reserving 1 bit for locked and 1 bit for pending, there are 14 bits left. So as long as NR_CPUS < 16k (requirement for 16-bit locked_pending), we can put all possible cpu numbers into the lock. We can also just use smp_processor_id() without additional percpu data.

That sounds horrific, wouldn't that destroy the whole point of using a byte for pending?

Quoted 7 lines
Quoted 3 lines

Also, you're talking ~1% gains here. I think our collective time would be better spent off reviewing the CNA series and trying to make it more deterministic.

I thought you guys are not interested in CNA. I do want to get CNA merged, if possible. Let review the current version again and see if there are ways we can further improve it.

It's not a lack of interrest. We were struggling with the fairness issues and the complexity of the thing. I forgot the current state of matters, but at one point UNLOCK was O(n) in waiters, which is, of course, 'unfortunate'.

I'll have to look up whatever notes remain, but the basic idea of keeping remote nodes on a secondary list is obviously breaking all sorts of fairness. After that they pile on a bunch of hacks to fix the worst of them, but it feels exactly like that, a bunch of hacks.

One of the things I suppose we ought to do is see if some of the ideas of phase-fair locks can be applied to this.

That coupled with a chronic lack of time for anything :-(

reviewer
#26 lore raw

Re: [PATCH v3 5/6] powerpc/pseries: implement paravirt qspinlocks for SPLPAR

Quoted 14 lines · Peter Zijlstra
Quoted 13 lines · Waiman Long
Quoted 12 lines · Will Deacon
Quoted 9 lines · On Thu, Jul 23, 2020 at 08:47:59PM +0200, [email protected] wrote:
Quoted 3 lines · Waiman Long

BTW, do you have any comment on my v2 lock holder cpu info qspinlock patch? I will have to update the patch to fix the reported 0-day test problem, but I want to collect other feedback before sending out v3.

I want to say I hate it all, it adds instructions to a path we spend an aweful lot of time optimizing without really getting anything back for it.

Will, how do you feel about it?

I can see it potentially being useful for debugging, but I hate the limitation to 256 CPUs. Even arm64 is hitting that now.

Quoted 7 lines
Quoted 5 lines

After thinking more about that, I think we can use all the remaining bits in the 16-bit locked_pending. Reserving 1 bit for locked and 1 bit for pending, there are 14 bits left. So as long as NR_CPUS < 16k (requirement for 16-bit locked_pending), we can put all possible cpu numbers into the lock. We can also just use smp_processor_id() without additional percpu data.

That sounds horrific, wouldn't that destroy the whole point of using a byte for pending?

You are right. I realized that later on and had sent a follow-up mail to correct that.

Quoted 11 lines
Quoted 6 lines
Quoted 3 lines

Also, you're talking ~1% gains here. I think our collective time would be better spent off reviewing the CNA series and trying to make it more deterministic.

I thought you guys are not interested in CNA. I do want to get CNA merged, if possible. Let review the current version again and see if there are ways we can further improve it.

It's not a lack of interrest. We were struggling with the fairness issues and the complexity of the thing. I forgot the current state of matters, but at one point UNLOCK was O(n) in waiters, which is, of course, 'unfortunate'.

Quoted 7 lines

I'll have to look up whatever notes remain, but the basic idea of keeping remote nodes on a secondary list is obviously breaking all sorts of fairness. After that they pile on a bunch of hacks to fix the worst of them, but it feels exactly like that, a bunch of hacks.

One of the things I suppose we ought to do is see if some of the ideas of phase-fair locks can be applied to this.

That could be a possible solution to ensure better fairness.

Quoted 3 lines

That coupled with a chronic lack of time for anything :-(

That is always true and I feel this way too:-)

Cheers,
Longman

author reviewer
#27 lore raw

Re: [PATCH v3 5/6] powerpc/pseries: implement paravirt qspinlocks for SPLPAR

In reply to Michael Ellerman

Excerpts from Michael Ellerman's message of July 9, 2020 8:53 pm:

Quoted 31 lines
Quoted 8 lines · Nicholas Piggin
---
 arch/powerpc/include/asm/paravirt.h           | 28 ++++++++
 arch/powerpc/include/asm/qspinlock.h          | 66 +++++++++++++++++++
 arch/powerpc/include/asm/qspinlock_paravirt.h |  7 ++
 arch/powerpc/platforms/pseries/Kconfig        |  5 ++
 arch/powerpc/platforms/pseries/setup.c        |  6 +-
 include/asm-generic/qspinlock.h               |  2 +

Another ack?

Quoted 5 lines
arch/powerpc/include/asm/paravirt.h +0 -0
index 7a8546660a63..f2d51f929cf5 100644

arch/powerpc/include/asm/paravirt.h static inline void yield_to_preempted(int cpu, u32 yield_count) · line 63

Quoted 9 lines
arch/powerpc/include/asm/paravirt.h +6 -0
@@ -45,6 +55,19 @@ static inline void yield_to_preempted(int cpu, u32 yield_count)
 {
 	___bad_yield_to_preempted(); /* This would be a bug */
 }
+
+extern void ___bad_yield_to_any(void);
+static inline void yield_to_any(void)
+{
+	___bad_yield_to_any(); /* This would be a bug */
+}

Why do we do that rather than just not defining yield_to_any() at all and letting the build fail on that?

Quoted 2 lines

There's a condition somewhere that we know will false at compile time and drop the call before linking?

Mainly so you could use it in if (IS_ENABLED()) blocks, but would still catch the (presumably buggy) case where something calls it without the option set.

I think I had it arranged a different way that was using IS_ENABLED earlier and changed it but might as well keep it this way.

Quoted 13 lines

arch/powerpc/include/asm/qspinlock_paravirt.h line 3

Quoted 10 lines
arch/powerpc/include/asm/qspinlock_paravirt.h new file +3 -0
new file mode 100644
index 000000000000..750d1b5e0202
@@ -0,0 +1,7 @@
+/* SPDX-License-Identifier: GPL-2.0-or-later */
+#ifndef __ASM_QSPINLOCK_PARAVIRT_H
+#define __ASM_QSPINLOCK_PARAVIRT_H

_ASM_POWERPC_QSPINLOCK_PARAVIRT_H please.

Quoted 4 lines

arch/powerpc/include/asm/qspinlock_paravirt.h line 5

Quoted 2 lines
arch/powerpc/include/asm/qspinlock_paravirt.h new file +2 -0
+
+EXPORT_SYMBOL(__pv_queued_spin_unlock);

Why's that in a header? Should that (eventually) go with the generic implementation?

Yeah the qspinlock_paravirt.h header is a bit weird and only gets included into kernel/locking/qspinlock.c

Quoted 22 lines

arch/powerpc/platforms/pseries/Kconfig config PPC_PSERIES · line 30

Quoted 11 lines
arch/powerpc/platforms/pseries/Kconfig +3 -0
index 24c18362e5ea..756e727b383f 100644
@@ -25,9 +25,14 @@ config PPC_PSERIES
 	select SWIOTLB
 	default y
 
+config PARAVIRT_SPINLOCKS
+	bool
+	default n

default n is the default.

Quoted 8 lines
arch/powerpc/platforms/pseries/setup.c +0 -0
index 2db8469e475f..747a203d9453 100644
@@ -771,8 +771,12 @@ static void __init pSeries_setup_arch(void)
 	if (firmware_has_feature(FW_FEATURE_LPAR)) {
 		vpa_init(boot_cpuid);
 
Quoted 9 lines

arch/powerpc/platforms/pseries/setup.c static void __init pSeries_setup_arch(void) · line 779

Quoted 7 lines
arch/powerpc/platforms/pseries/setup.c +5 -1
-		if (lppaca_shared_proc(get_lppaca()))
+		if (lppaca_shared_proc(get_lppaca())) {
 			static_branch_enable(&shared_processor);
+#ifdef CONFIG_PARAVIRT_SPINLOCKS
+			pv_spinlocks_init();
+#endif
+		}

We could avoid the ifdef with this I think?

Yes I think so.

Thanks,
Nick

author reviewer
#28 lore raw

PATCH v3 6/6 powerpc/qspinlock: optimised atomic_try_cmpxchg_lock that adds the lock hint

In reply to Nicholas Piggin

This brings the behaviour of the uncontended fast path back to roughly equivalent to simple spinlocks -- a single atomic op with lock hint.

---
 arch/powerpc/include/asm/atomic.h    | 28 ++++++++++++++++++++++++++++
 arch/powerpc/include/asm/qspinlock.h |  2 +-
 2 files changed, 29 insertions(+), 1 deletion(-)
arch/powerpc/include/asm/atomic.h +28 -0
index 498785ffc25f..f6a3d145ffb7 100644
@@ -193,6 +193,34 @@ static __inline__ int atomic_dec_return_relaxed(atomic_t *v)
 #define atomic_xchg(v, new) (xchg(&((v)->counter), new))
 #define atomic_xchg_relaxed(v, new) xchg_relaxed(&((v)->counter), (new))
 
+/*
+ * Don't want to override the generic atomic_try_cmpxchg_acquire, because
+ * we add a lock hint to the lwarx, which may not be wanted for the
+ * _acquire case (and is not used by the other _acquire variants so it
+ * would be a surprise).
+ */
+static __always_inline bool
+atomic_try_cmpxchg_lock(atomic_t *v, int *old, int new)
+{
+	int r, o = *old;
+
+	__asm__ __volatile__ (
+"1:\t"	PPC_LWARX(%0,0,%2,1) "	# atomic_try_cmpxchg_acquire	\n"
+"	cmpw	0,%0,%3							\n"
+"	bne-	2f							\n"
+"	stwcx.	%4,0,%2							\n"
+"	bne-	1b							\n"
+"\t"	PPC_ACQUIRE_BARRIER "						\n"
+"2:									\n"
+	: "=&r" (r), "+m" (v->counter)
+	: "r" (&v->counter), "r" (o), "r" (new)
+	: "cr0", "memory");
+
+	if (unlikely(r != o))
+		*old = r;
+	return likely(r = o);
+}
+
 /**
  * atomic_fetch_add_unless - add unless the number is a given value
  * @v: pointer of type atomic_t
arch/powerpc/include/asm/qspinlock.h +1 -1
index f5066f00a08c..b752d34517b3 100644
@@ -37,7 +37,7 @@ static __always_inline void queued_spin_lock(struct qspinlock *lock)
 {
 	u32 val = 0;
 
-	if (likely(atomic_try_cmpxchg_acquire(&lock->val, &val, _Q_LOCKED_VAL)))
+	if (likely(atomic_try_cmpxchg_lock(&lock->val, &val, _Q_LOCKED_VAL)))
 		return;
 
 	queued_spin_lock_slowpath(lock, val);
Signature · 1 line
2.23.0
reviewer
#29 lore raw
In reply to Nicholas Piggin
Quoted 31 lines · Nicholas Piggin

v3 is updated to use __pv_queued_spin_unlock, noticed by Waiman (thank you).

Thanks,
Nick

Nicholas Piggin (6):
powerpc/powernv: must include hvcall.h to get PAPR defines powerpc/pseries: move some PAPR paravirt functions to their own file
powerpc: move spinlock implementation to simple_spinlock powerpc/64s: implement queued spinlocks and rwlocks powerpc/pseries: implement paravirt qspinlocks for SPLPAR powerpc/qspinlock: optimised atomic_try_cmpxchg_lock that adds the
lock hint

  arch/powerpc/Kconfig                          |  13 +
  arch/powerpc/include/asm/Kbuild               |   2 +
  arch/powerpc/include/asm/atomic.h             |  28 ++
  arch/powerpc/include/asm/paravirt.h           |  89 +++++
  arch/powerpc/include/asm/qspinlock.h          |  91 ++++++
  arch/powerpc/include/asm/qspinlock_paravirt.h |   7 +
  arch/powerpc/include/asm/simple_spinlock.h    | 292 +++++++++++++++++
  .../include/asm/simple_spinlock_types.h       |  21 ++
  arch/powerpc/include/asm/spinlock.h           | 308 +-----------------
  arch/powerpc/include/asm/spinlock_types.h     |  17 +-
  arch/powerpc/lib/Makefile                     |   3 +
  arch/powerpc/lib/locks.c                      |  12 +-
  arch/powerpc/platforms/powernv/pci-ioda-tce.c |   1 +
  arch/powerpc/platforms/pseries/Kconfig        |   5 +
  arch/powerpc/platforms/pseries/setup.c        |   6 +-
  include/asm-generic/qspinlock.h               |   4 +
  16 files changed, 577 insertions(+), 322 deletions(-)
Quoted 6 lines

create mode 100644 arch/powerpc/include/asm/paravirt.h create mode 100644 arch/powerpc/include/asm/qspinlock.h create mode 100644 arch/powerpc/include/asm/qspinlock_paravirt.h create mode 100644 arch/powerpc/include/asm/simple_spinlock.h create mode 100644 arch/powerpc/include/asm/simple_spinlock_types.h

This patch looks OK to me.

I had run some microbenchmark on powerpc system with or w/o the patch.

On a 2-socket 160-thread SMT4 POWER9 system (not virtualized):

5.8.0-rc4
====
Running locktest with spinlock [runtime = 10s, load = 1] Threads = 160, Min/Mean/Max = 77,665/90,153/106,895 Threads = 160, Total Rate = 1,441,759 op/s; Percpu Rate = 9,011 op/s

Running locktest with rwlock [runtime = 10s, r% = 50%, load = 1] Threads = 160, Min/Mean/Max = 47,879/53,807/63,689 Threads = 160, Total Rate = 860,192 op/s; Percpu Rate = 5,376 op/s

Running locktest with spinlock [runtime = 10s, load = 1] Threads = 80, Min/Mean/Max = 242,907/319,514/463,161 Threads = 80, Total Rate = 2,555 kop/s; Percpu Rate = 32 kop/s

Running locktest with rwlock [runtime = 10s, r% = 50%, load = 1] Threads = 80, Min/Mean/Max = 146,161/187,474/259,270 Threads = 80, Total Rate = 1,498 kop/s; Percpu Rate = 19 kop/s

Running locktest with spinlock [runtime = 10s, load = 1] Threads = 40, Min/Mean/Max = 646,639/1,000,817/1,455,205 Threads = 40, Total Rate = 4,001 kop/s; Percpu Rate = 100 kop/s

Running locktest with rwlock [runtime = 10s, r% = 50%, load = 1] Threads = 40, Min/Mean/Max = 402,165/597,132/814,555 Threads = 40, Total Rate = 2,388 kop/s; Percpu Rate = 60 kop/s

5.8.0-rc4-qlock+
========

Running locktest with spinlock [runtime = 10s, load = 1] Threads = 160, Min/Mean/Max = 123,835/124,580/124,587 Threads = 160, Total Rate = 1,992 kop/s; Percpu Rate = 12 kop/s

Running locktest with rwlock [runtime = 10s, r% = 50%, load = 1] Threads = 160, Min/Mean/Max = 254,210/264,714/276,784 Threads = 160, Total Rate = 4,231 kop/s; Percpu Rate = 26 kop/s

Running locktest with spinlock [runtime = 10s, load = 1] Threads = 80, Min/Mean/Max = 599,715/603,397/603,450 Threads = 80, Total Rate = 4,825 kop/s; Percpu Rate = 60 kop/s

Running locktest with rwlock [runtime = 10s, r% = 50%, load = 1] Threads = 80, Min/Mean/Max = 492,687/525,224/567,456 Threads = 80, Total Rate = 4,199 kop/s; Percpu Rate = 52 kop/s

Running locktest with spinlock [runtime = 10s, load = 1] Threads = 40, Min/Mean/Max = 1,325,623/1,325,628/1,325,636 Threads = 40, Total Rate = 5,299 kop/s; Percpu Rate = 132 kop/s

Running locktest with rwlock [runtime = 10s, r% = 50%, load = 1] Threads = 40, Min/Mean/Max = 1,249,731/1,292,977/1,342,815 Threads = 40, Total Rate = 5,168 kop/s; Percpu Rate = 129 kop/s

On systems on large number of cpus, qspinlock lock is faster and more fair.

With some tuning, we may be able to squeeze out more performance.

Cheers,
Longman

author reviewer
#30 lore raw

Excerpts from Waiman Long's message of July 7, 2020 4:39 am:

Quoted 38 lines · Nicholas Piggin
Quoted 31 lines

v3 is updated to use __pv_queued_spin_unlock, noticed by Waiman (thank you).

Thanks,
Nick

Nicholas Piggin (6):
powerpc/powernv: must include hvcall.h to get PAPR defines powerpc/pseries: move some PAPR paravirt functions to their own file
powerpc: move spinlock implementation to simple_spinlock powerpc/64s: implement queued spinlocks and rwlocks powerpc/pseries: implement paravirt qspinlocks for SPLPAR powerpc/qspinlock: optimised atomic_try_cmpxchg_lock that adds the
lock hint

  arch/powerpc/Kconfig                          |  13 +
  arch/powerpc/include/asm/Kbuild               |   2 +
  arch/powerpc/include/asm/atomic.h             |  28 ++
  arch/powerpc/include/asm/paravirt.h           |  89 +++++
  arch/powerpc/include/asm/qspinlock.h          |  91 ++++++
  arch/powerpc/include/asm/qspinlock_paravirt.h |   7 +
  arch/powerpc/include/asm/simple_spinlock.h    | 292 +++++++++++++++++
  .../include/asm/simple_spinlock_types.h       |  21 ++
  arch/powerpc/include/asm/spinlock.h           | 308 +-----------------
  arch/powerpc/include/asm/spinlock_types.h     |  17 +-
  arch/powerpc/lib/Makefile                     |   3 +
  arch/powerpc/lib/locks.c                      |  12 +-
  arch/powerpc/platforms/powernv/pci-ioda-tce.c |   1 +
  arch/powerpc/platforms/pseries/Kconfig        |   5 +
  arch/powerpc/platforms/pseries/setup.c        |   6 +-
  include/asm-generic/qspinlock.h               |   4 +
  16 files changed, 577 insertions(+), 322 deletions(-)
Quoted 6 lines

create mode 100644 arch/powerpc/include/asm/paravirt.h create mode 100644 arch/powerpc/include/asm/qspinlock.h create mode 100644 arch/powerpc/include/asm/qspinlock_paravirt.h create mode 100644 arch/powerpc/include/asm/simple_spinlock.h create mode 100644 arch/powerpc/include/asm/simple_spinlock_types.h

Quoted 1 line

This patch looks OK to me.

Thanks for reviewing and testing.

Quoted 54 lines

I had run some microbenchmark on powerpc system with or w/o the patch.

On a 2-socket 160-thread SMT4 POWER9 system (not virtualized):

5.8.0-rc4
=========

Running locktest with spinlock [runtime = 10s, load = 1] Threads = 160, Min/Mean/Max = 77,665/90,153/106,895 Threads = 160, Total Rate = 1,441,759 op/s; Percpu Rate = 9,011 op/s

Running locktest with rwlock [runtime = 10s, r% = 50%, load = 1] Threads = 160, Min/Mean/Max = 47,879/53,807/63,689 Threads = 160, Total Rate = 860,192 op/s; Percpu Rate = 5,376 op/s

Running locktest with spinlock [runtime = 10s, load = 1] Threads = 80, Min/Mean/Max = 242,907/319,514/463,161 Threads = 80, Total Rate = 2,555 kop/s; Percpu Rate = 32 kop/s

Running locktest with rwlock [runtime = 10s, r% = 50%, load = 1] Threads = 80, Min/Mean/Max = 146,161/187,474/259,270 Threads = 80, Total Rate = 1,498 kop/s; Percpu Rate = 19 kop/s

Running locktest with spinlock [runtime = 10s, load = 1] Threads = 40, Min/Mean/Max = 646,639/1,000,817/1,455,205 Threads = 40, Total Rate = 4,001 kop/s; Percpu Rate = 100 kop/s

Running locktest with rwlock [runtime = 10s, r% = 50%, load = 1] Threads = 40, Min/Mean/Max = 402,165/597,132/814,555 Threads = 40, Total Rate = 2,388 kop/s; Percpu Rate = 60 kop/s

5.8.0-rc4-qlock+
================

Running locktest with spinlock [runtime = 10s, load = 1] Threads = 160, Min/Mean/Max = 123,835/124,580/124,587 Threads = 160, Total Rate = 1,992 kop/s; Percpu Rate = 12 kop/s

Running locktest with rwlock [runtime = 10s, r% = 50%, load = 1] Threads = 160, Min/Mean/Max = 254,210/264,714/276,784 Threads = 160, Total Rate = 4,231 kop/s; Percpu Rate = 26 kop/s

Running locktest with spinlock [runtime = 10s, load = 1] Threads = 80, Min/Mean/Max = 599,715/603,397/603,450 Threads = 80, Total Rate = 4,825 kop/s; Percpu Rate = 60 kop/s

Running locktest with rwlock [runtime = 10s, r% = 50%, load = 1] Threads = 80, Min/Mean/Max = 492,687/525,224/567,456 Threads = 80, Total Rate = 4,199 kop/s; Percpu Rate = 52 kop/s

Running locktest with spinlock [runtime = 10s, load = 1] Threads = 40, Min/Mean/Max = 1,325,623/1,325,628/1,325,636 Threads = 40, Total Rate = 5,299 kop/s; Percpu Rate = 132 kop/s

Quoted 7 lines

Running locktest with rwlock [runtime = 10s, r% = 50%, load = 1] Threads = 40, Min/Mean/Max = 1,249,731/1,292,977/1,342,815 Threads = 40, Total Rate = 5,168 kop/s; Percpu Rate = 129 kop/s

On systems on large number of cpus, qspinlock lock is faster and more fair.

With some tuning, we may be able to squeeze out more performance.

Yes, powerpc could certainly get more performance out of the slow paths, and then there are a few parameters to tune.

We don't have a good alternate patching for function calls yet, but that would be something to do for native vs pv.

And then there seem to be one or two tunable parameters we could experiment with.

The paravirt locks may need a bit more tuning. Some simple testing under KVM shows we might be a bit slower in some cases. Whether this is fairness or something else I'm not sure. The current simple pv spinlock code can do a directed yield to the lock holder CPU, whereas the pv qspl here just does a general yield. I think we might actually be able to change that to also support directed yield. Though I'm not sure if this is actually the cause of the slowdown yet.

Thanks,
Nick

reviewer
#31 lore raw
Quoted 9 lines · Nicholas Piggin

Yes, powerpc could certainly get more performance out of the slow paths, and then there are a few parameters to tune.

We don't have a good alternate patching for function calls yet, but that would be something to do for native vs pv.

And then there seem to be one or two tunable parameters we could experiment with.

Quoted 7 lines

The paravirt locks may need a bit more tuning. Some simple testing under KVM shows we might be a bit slower in some cases. Whether this is fairness or something else I'm not sure. The current simple pv spinlock code can do a directed yield to the lock holder CPU, whereas the pv qspl here just does a general yield. I think we might actually be able to change that to also support directed yield. Though I'm not sure if this is actually the cause of the slowdown yet.

Regarding the paravirt lock, I have taken a further look into the current PPC spinlock code. There is an equivalent of pv_wait() but no pv_kick(). Maybe PPC doesn't really need that. Attached are two additional qspinlock patches that adds a CONFIG_PARAVIRT_QSPINLOCKS_LITE option to not require pv_kick(). There is also a fixup patch to be applied after your patchset.

I don't have access to a PPC LPAR with shared processor at the moment, so I can't test the performance of the paravirt code. Would you mind adding my patches and do some performance test on your end to see if it gives better result?

Thanks,
Longman

From 161e545523a7eb4c42c145c04e9a5a15903ba3d9 Mon Sep 17 00:00:00 2001
From: Waiman Long <[email protected]>
Date: Tue, 7 Jul 2020 20:46:51 -0400
Subject: [PATCH 1/9] locking/pvqspinlock: Code relocation and extraction

Move pv_kick_node() and the unlock functions up and extract out the hash and lock code from pv_wait_head_or_lock() into pv_hash_lock(). There is no functional change.

---
 kernel/locking/qspinlock_paravirt.h | 302 ++++++++++++++--------------
 1 file changed, 156 insertions(+), 146 deletions(-)
kernel/locking/qspinlock_paravirt.h +156 -146
index e84d21aa0722..8eec58320b85 100644
@@ -55,6 +55,7 @@ struct pv_node {
 
 /*
  * Hybrid PV queued/unfair lock
+ * ----------------------------
  *
  * By replacing the regular queued_spin_trylock() with the function below,
  * it will be called once when a lock waiter enter the PV slowpath before
@@ -259,6 +260,156 @@ static struct pv_node *pv_unhash(struct qspinlock *lock)
 	BUG();
 }
 
+/*
+ * Insert lock into hash and set _Q_SLOW_VAL.
+ * Return true if lock acquired.
+ */
+static inline bool pv_hash_lock(struct qspinlock *lock, struct pv_node *node)
+{
+	struct qspinlock **lp = pv_hash(lock, node);
+
+	/*
+	 * We must hash before setting _Q_SLOW_VAL, such that
+	 * when we observe _Q_SLOW_VAL in __pv_queued_spin_unlock()
+	 * we'll be sure to be able to observe our hash entry.
+	 *
+	 *   [S] <hash>                 [Rmw] l->locked == _Q_SLOW_VAL
+	 *       MB                           RMB
+	 * [RmW] l->locked = _Q_SLOW_VAL  [L] <unhash>
+	 *
+	 * Matches the smp_rmb() in __pv_queued_spin_unlock().
+	 */
+	if (xchg(&lock->locked, _Q_SLOW_VAL) == 0) {
+		/*
+		 * The lock was free and now we own the lock.
+		 * Change the lock value back to _Q_LOCKED_VAL
+		 * and unhash the table.
+		 */
+		WRITE_ONCE(lock->locked, _Q_LOCKED_VAL);
+		WRITE_ONCE(*lp, NULL);
+		return true;
+	}
+	return false;
+}
+
+/*
+ * Called after setting next->locked = 1 when we're the lock owner.
+ *
+ * Instead of waking the waiters stuck in pv_wait_node() advance their state
+ * such that they're waiting in pv_wait_head_or_lock(), this avoids a
+ * wake/sleep cycle.
+ */
+static void pv_kick_node(struct qspinlock *lock, struct mcs_spinlock *node)
+{
+	struct pv_node *pn = (struct pv_node *)node;
+
+	/*
+	 * If the vCPU is indeed halted, advance its state to match that of
+	 * pv_wait_node(). If OTOH this fails, the vCPU was running and will
+	 * observe its next->locked value and advance itself.
+	 *
+	 * Matches with smp_store_mb() and cmpxchg() in pv_wait_node()
+	 *
+	 * The write to next->locked in arch_mcs_spin_unlock_contended()
+	 * must be ordered before the read of pn->state in the cmpxchg()
+	 * below for the code to work correctly. To guarantee full ordering
+	 * irrespective of the success or failure of the cmpxchg(),
+	 * a relaxed version with explicit barrier is used. The control
+	 * dependency will order the reading of pn->state before any
+	 * subsequent writes.
+	 */
+	smp_mb__before_atomic();
+	if (cmpxchg_relaxed(&pn->state, vcpu_halted, vcpu_hashed)
+	    != vcpu_halted)
+		return;
+
+	/*
+	 * Put the lock into the hash table and set the _Q_SLOW_VAL.
+	 *
+	 * As this is the same vCPU that will check the _Q_SLOW_VAL value and
+	 * the hash table later on at unlock time, no atomic instruction is
+	 * needed.
+	 */
+	WRITE_ONCE(lock->locked, _Q_SLOW_VAL);
+	(void)pv_hash(lock, pn);
+}
+
+/*
+ * PV versions of the unlock fastpath and slowpath functions to be used
+ * instead of queued_spin_unlock().
+ */
+__visible void
+__pv_queued_spin_unlock_slowpath(struct qspinlock *lock, u8 locked)
+{
+	struct pv_node *node;
+
+	if (unlikely(locked != _Q_SLOW_VAL)) {
+		WARN(!debug_locks_silent,
+		     "pvqspinlock: lock 0x%lx has corrupted value 0x%x!\n",
+		     (unsigned long)lock, atomic_read(&lock->val));
+		return;
+	}
+
+	/*
+	 * A failed cmpxchg doesn't provide any memory-ordering guarantees,
+	 * so we need a barrier to order the read of the node data in
+	 * pv_unhash *after* we've read the lock being _Q_SLOW_VAL.
+	 *
+	 * Matches the cmpxchg() in pv_wait_head_or_lock() setting _Q_SLOW_VAL.
+	 */
+	smp_rmb();
+
+	/*
+	 * Since the above failed to release, this must be the SLOW path.
+	 * Therefore start by looking up the blocked node and unhashing it.
+	 */
+	node = pv_unhash(lock);
+
+	/*
+	 * Now that we have a reference to the (likely) blocked pv_node,
+	 * release the lock.
+	 */
+	smp_store_release(&lock->locked, 0);
+
+	/*
+	 * At this point the memory pointed at by lock can be freed/reused,
+	 * however we can still use the pv_node to kick the CPU.
+	 * The other vCPU may not really be halted, but kicking an active
+	 * vCPU is harmless other than the additional latency in completing
+	 * the unlock.
+	 */
+	lockevent_inc(pv_kick_unlock);
+	pv_kick(node->cpu);
+}
+
+/*
+ * Include the architecture specific callee-save thunk of the
+ * __pv_queued_spin_unlock(). This thunk is put together with
+ * __pv_queued_spin_unlock() to make the callee-save thunk and the real unlock
+ * function close to each other sharing consecutive instruction cachelines.
+ * Alternatively, architecture specific version of __pv_queued_spin_unlock()
+ * can be defined.
+ */
+#include <asm/qspinlock_paravirt.h>
+
+#ifndef __pv_queued_spin_unlock
+__visible void __pv_queued_spin_unlock(struct qspinlock *lock)
+{
+	u8 locked;
+
+	/*
+	 * We must not unlock if SLOW, because in that case we must first
+	 * unhash. Otherwise it would be possible to have multiple @lock
+	 * entries, which would be BAD.
+	 */
+	locked = cmpxchg_release(&lock->locked, _Q_LOCKED_VAL, 0);
+	if (likely(locked == _Q_LOCKED_VAL))
+		return;
+
+	__pv_queued_spin_unlock_slowpath(lock, locked);
+}
+#endif /* __pv_queued_spin_unlock */
+
 /*
  * Return true if when it is time to check the previous node which is not
  * in a running state.
@@ -350,48 +501,6 @@ static void pv_wait_node(struct mcs_spinlock *node, struct mcs_spinlock *prev)
 	 */
 }
 
-/*
- * Called after setting next->locked = 1 when we're the lock owner.
- *
- * Instead of waking the waiters stuck in pv_wait_node() advance their state
- * such that they're waiting in pv_wait_head_or_lock(), this avoids a
- * wake/sleep cycle.
- */
-static void pv_kick_node(struct qspinlock *lock, struct mcs_spinlock *node)
-{
-	struct pv_node *pn = (struct pv_node *)node;
-
-	/*
-	 * If the vCPU is indeed halted, advance its state to match that of
-	 * pv_wait_node(). If OTOH this fails, the vCPU was running and will
-	 * observe its next->locked value and advance itself.
-	 *
-	 * Matches with smp_store_mb() and cmpxchg() in pv_wait_node()
-	 *
-	 * The write to next->locked in arch_mcs_spin_unlock_contended()
-	 * must be ordered before the read of pn->state in the cmpxchg()
-	 * below for the code to work correctly. To guarantee full ordering
-	 * irrespective of the success or failure of the cmpxchg(),
-	 * a relaxed version with explicit barrier is used. The control
-	 * dependency will order the reading of pn->state before any
-	 * subsequent writes.
-	 */
-	smp_mb__before_atomic();
-	if (cmpxchg_relaxed(&pn->state, vcpu_halted, vcpu_hashed)
-	    != vcpu_halted)
-		return;
-
-	/*
-	 * Put the lock into the hash table and set the _Q_SLOW_VAL.
-	 *
-	 * As this is the same vCPU that will check the _Q_SLOW_VAL value and
-	 * the hash table later on at unlock time, no atomic instruction is
-	 * needed.
-	 */
-	WRITE_ONCE(lock->locked, _Q_SLOW_VAL);
-	(void)pv_hash(lock, pn);
-}
-
 /*
  * Wait for l->locked to become clear and acquire the lock;
  * halt the vcpu after a short spin.
@@ -403,16 +512,13 @@ static u32
 pv_wait_head_or_lock(struct qspinlock *lock, struct mcs_spinlock *node)
 {
 	struct pv_node *pn = (struct pv_node *)node;
-	struct qspinlock **lp = NULL;
 	int waitcnt = 0;
 	int loop;
-
 	/*
-	 * If pv_kick_node() already advanced our state, we don't need to
+	 * If pv_kick_node() had already advanced our state, we don't need to
 	 * insert ourselves into the hash table anymore.
 	 */
-	if (READ_ONCE(pn->state) == vcpu_hashed)
-		lp = (struct qspinlock **)1;
+	bool hashed = (READ_ONCE(pn->state) == vcpu_hashed);
 
 	/*
 	 * Tracking # of slowpath locking operations
@@ -439,30 +545,10 @@ pv_wait_head_or_lock(struct qspinlock *lock, struct mcs_spinlock *node)
 		clear_pending(lock);
 
 
-		if (!lp) { /* ONCE */
-			lp = pv_hash(lock, pn);
-
-			/*
-			 * We must hash before setting _Q_SLOW_VAL, such that
-			 * when we observe _Q_SLOW_VAL in __pv_queued_spin_unlock()
-			 * we'll be sure to be able to observe our hash entry.
-			 *
-			 *   [S] <hash>                 [Rmw] l->locked == _Q_SLOW_VAL
-			 *       MB                           RMB
-			 * [RmW] l->locked = _Q_SLOW_VAL  [L] <unhash>
-			 *
-			 * Matches the smp_rmb() in __pv_queued_spin_unlock().
-			 */
-			if (xchg(&lock->locked, _Q_SLOW_VAL) == 0) {
-				/*
-				 * The lock was free and now we own the lock.
-				 * Change the lock value back to _Q_LOCKED_VAL
-				 * and unhash the table.
-				 */
-				WRITE_ONCE(lock->locked, _Q_LOCKED_VAL);
-				WRITE_ONCE(*lp, NULL);
+		if (!hashed) {	/* ONCE */
+			if (pv_hash_lock(lock, pn))
 				goto gotlock;
-			}
+			hashed = true;
 		}
 		WRITE_ONCE(pn->state, vcpu_hashed);
 		lockevent_inc(pv_wait_head);
@@ -484,79 +570,3 @@ pv_wait_head_or_lock(struct qspinlock *lock, struct mcs_spinlock *node)
 gotlock:
 	return (u32)(atomic_read(&lock->val) | _Q_LOCKED_VAL);
 }
-
-/*
- * PV versions of the unlock fastpath and slowpath functions to be used
- * instead of queued_spin_unlock().
- */
-__visible void
-__pv_queued_spin_unlock_slowpath(struct qspinlock *lock, u8 locked)
-{
-	struct pv_node *node;
-
-	if (unlikely(locked != _Q_SLOW_VAL)) {
-		WARN(!debug_locks_silent,
-		     "pvqspinlock: lock 0x%lx has corrupted value 0x%x!\n",
-		     (unsigned long)lock, atomic_read(&lock->val));
-		return;
-	}
-
-	/*
-	 * A failed cmpxchg doesn't provide any memory-ordering guarantees,
-	 * so we need a barrier to order the read of the node data in
-	 * pv_unhash *after* we've read the lock being _Q_SLOW_VAL.
-	 *
-	 * Matches the cmpxchg() in pv_wait_head_or_lock() setting _Q_SLOW_VAL.
-	 */
-	smp_rmb();
-
-	/*
-	 * Since the above failed to release, this must be the SLOW path.
-	 * Therefore start by looking up the blocked node and unhashing it.
-	 */
-	node = pv_unhash(lock);
-
-	/*
-	 * Now that we have a reference to the (likely) blocked pv_node,
-	 * release the lock.
-	 */
-	smp_store_release(&lock->locked, 0);
-
-	/*
-	 * At this point the memory pointed at by lock can be freed/reused,
-	 * however we can still use the pv_node to kick the CPU.
-	 * The other vCPU may not really be halted, but kicking an active
-	 * vCPU is harmless other than the additional latency in completing
-	 * the unlock.
-	 */
-	lockevent_inc(pv_kick_unlock);
-	pv_kick(node->cpu);
-}
-
-/*
- * Include the architecture specific callee-save thunk of the
- * __pv_queued_spin_unlock(). This thunk is put together with
- * __pv_queued_spin_unlock() to make the callee-save thunk and the real unlock
- * function close to each other sharing consecutive instruction cachelines.
- * Alternatively, architecture specific version of __pv_queued_spin_unlock()
- * can be defined.
- */
-#include <asm/qspinlock_paravirt.h>
-
-#ifndef __pv_queued_spin_unlock
-__visible void __pv_queued_spin_unlock(struct qspinlock *lock)
-{
-	u8 locked;
-
-	/*
-	 * We must not unlock if SLOW, because in that case we must first
-	 * unhash. Otherwise it would be possible to have multiple @lock
-	 * entries, which would be BAD.
-	 */
-	locked = cmpxchg_release(&lock->locked, _Q_LOCKED_VAL, 0);
-	if (likely(locked == _Q_LOCKED_VAL))
-		return;
-
-	__pv_queued_spin_unlock_slowpath(lock, locked);
-}
-#endif /* __pv_queued_spin_unlock */
Signature · 1 line
2.18.1

From 5d7941a498935fb225b2c7a3108cbf590114c3db Mon Sep 17 00:00:00 2001
From: Waiman Long <[email protected]>
Date: Tue, 7 Jul 2020 22:29:16 -0400
Subject: [PATCH 2/9] locking/pvqspinlock: Introduce
CONFIG_PARAVIRT_QSPINLOCKS_LITE

Add a new PARAVIRT_QSPINLOCKS_LITE config option that allows architectures to use the PV qspinlock code without the need to use or implement a pv_kick() function, thus eliminating the atomic unlock overhead. The non-atomic queued_spin_unlock() can be used instead. The pv_wait() function will still be needed, but it can be a dummy function.

With that option set, the hybrid PV queued/unfair locking code should still be able to make it performant enough in a paravirtualized environment.

---
 kernel/Kconfig.locks                |  4 +++
 kernel/locking/lock_events_list.h   |  3 ++
 kernel/locking/qspinlock_paravirt.h | 49 ++++++++++++++++++++++++-----
 kernel/locking/qspinlock_stat.h     |  5 +--
 4 files changed, 52 insertions(+), 9 deletions(-)
kernel/Kconfig.locks +4 -0
index 3de8fd11873b..1824ba8c44a9 100644
@@ -243,6 +243,10 @@ config QUEUED_SPINLOCKS
 	def_bool y if ARCH_USE_QUEUED_SPINLOCKS
 	depends on SMP
 
+config PARAVIRT_QSPINLOCKS_LITE
+	bool
+	depends on QUEUED_SPINLOCKS && PARAVIRT_SPINLOCKS
+
 config BPF_ARCH_SPINLOCK
 	bool
 
kernel/locking/lock_events_list.h +3 -0
index 239039d0ce21..9ae07a7148e8 100644
@@ -22,11 +22,14 @@
 /*
  * Locking events for PV qspinlock.
  */
+#ifndef CONFIG_PARAVIRT_QSPINLOCKS_LITE
 LOCK_EVENT(pv_hash_hops)	/* Average # of hops per hashing operation */
 LOCK_EVENT(pv_kick_unlock)	/* # of vCPU kicks issued at unlock time   */
 LOCK_EVENT(pv_kick_wake)	/* # of vCPU kicks for pv_latency_wake	   */
 LOCK_EVENT(pv_latency_kick)	/* Average latency (ns) of vCPU kick	   */
 LOCK_EVENT(pv_latency_wake)	/* Average latency (ns) of kick-to-wakeup  */
+#endif
+
 LOCK_EVENT(pv_lock_stealing)	/* # of lock stealing operations	   */
 LOCK_EVENT(pv_spurious_wakeup)	/* # of spurious wakeups in non-head vCPUs */
 LOCK_EVENT(pv_wait_again)	/* # of wait's after queue head vCPU kick  */
kernel/locking/qspinlock_paravirt.h +42 -7

73 lines not on this page · raw message

kernel/locking/qspinlock_stat.h +3 -2
index e625bb410aa2..e9f63240785b 100644
@@ -7,7 +7,8 @@
 #include "lock_events.h"
 
 #ifdef CONFIG_LOCK_EVENT_COUNTS
-#ifdef CONFIG_PARAVIRT_SPINLOCKS
+#if defined(CONFIG_PARAVIRT_SPINLOCKS) && \
+    !defined(CONFIG_PARAVIRT_QSPINLOCKS_LITE)
 /*
  * Collect pvqspinlock locking event counts
  */
@@ -133,7 +134,7 @@ static inline void __pv_wait(u8 *ptr, u8 val)
 #define pv_kick(c)	__pv_kick(c)
 #define pv_wait(p, v)	__pv_wait(p, v)
 
-#endif /* CONFIG_PARAVIRT_SPINLOCKS */
+#endif /* CONFIG_PARAVIRT_SPINLOCKS && !CONFIG_PARAVIRT_QSPINLOCKS_LITE */
 
 #else /* CONFIG_LOCK_EVENT_COUNTS */
 
Signature · 1 line
2.18.1

From 655d3a5d48a494a5674cf57454bf3e1f36b6eb83 Mon Sep 17 00:00:00 2001
From: Waiman Long <[email protected]>
Date: Tue, 7 Jul 2020 22:29:20 -0400
Subject: [PATCH 9/9] powerpc/pseries: Fixup

---
 arch/powerpc/include/asm/qspinlock.h   | 22 ----------------------
 arch/powerpc/platforms/pseries/Kconfig |  1 +
 arch/powerpc/platforms/pseries/setup.c |  6 +-----
 3 files changed, 2 insertions(+), 27 deletions(-)
arch/powerpc/include/asm/qspinlock.h +0 -22
index b752d34517b3..1fa724d27a2d 100644
@@ -10,7 +10,6 @@
 #ifdef CONFIG_PARAVIRT_SPINLOCKS
 extern void native_queued_spin_lock_slowpath(struct qspinlock *lock, u32 val);
 extern void __pv_queued_spin_lock_slowpath(struct qspinlock *lock, u32 val);
-extern void __pv_queued_spin_unlock(struct qspinlock *lock);
 
 static __always_inline void queued_spin_lock_slowpath(struct qspinlock *lock, u32 val)
 {
@@ -20,15 +19,6 @@ static __always_inline void queued_spin_lock_slowpath(struct qspinlock *lock, u3
 		__pv_queued_spin_lock_slowpath(lock, val);
 }
 
-#define queued_spin_unlock queued_spin_unlock
-static inline void queued_spin_unlock(struct qspinlock *lock)
-{
-	if (!is_shared_processor())
-		smp_store_release(&lock->locked, 0);
-	else
-		__pv_queued_spin_unlock(lock);
-}
-
 #else
 extern void queued_spin_lock_slowpath(struct qspinlock *lock, u32 val);
 #endif
@@ -72,18 +62,6 @@ static __always_inline void pv_wait(u8 *ptr, u8 val)
 	 */
 }
 
-static __always_inline void pv_kick(int cpu)
-{
-	prod_cpu(cpu);
-}
-
-extern void __pv_init_lock_hash(void);
-
-static inline void pv_spinlocks_init(void)
-{
-	__pv_init_lock_hash();
-}
-
 #endif
 
 #include <asm-generic/qspinlock.h>
arch/powerpc/platforms/pseries/Kconfig +1 -0
index 756e727b383f..1e3bbe27d664 100644
@@ -33,6 +33,7 @@ config PPC_SPLPAR
 	depends on PPC_PSERIES
 	bool "Support for shared-processor logical partitions"
 	select PARAVIRT_SPINLOCKS if PPC_QUEUED_SPINLOCKS
+	select PARAVIRT_QSPINLOCKS_LITE if PPC_QUEUED_SPINLOCKS
 	help
 	  Enabling this option will make the kernel run more efficiently
 	  on logically-partitioned pSeries systems which use shared
arch/powerpc/platforms/pseries/setup.c +1 -5
index 747a203d9453..2db8469e475f 100644
@@ -771,12 +771,8 @@ static void __init pSeries_setup_arch(void)
 	if (firmware_has_feature(FW_FEATURE_LPAR)) {
 		vpa_init(boot_cpuid);
 
-		if (lppaca_shared_proc(get_lppaca())) {
+		if (lppaca_shared_proc(get_lppaca()))
 			static_branch_enable(&shared_processor);
-#ifdef CONFIG_PARAVIRT_SPINLOCKS
-			pv_spinlocks_init();
-#endif
-		}
 
 		ppc_md.power_save = pseries_lpar_idle;
 		ppc_md.enable_pmcs = pseries_lpar_enable_pmcs;
Signature · 1 line
2.18.1
author reviewer
#32 lore raw

Excerpts from Waiman Long's message of July 8, 2020 1:33 pm:

Quoted 18 lines · Nicholas Piggin
Quoted 9 lines

Yes, powerpc could certainly get more performance out of the slow paths, and then there are a few parameters to tune.

We don't have a good alternate patching for function calls yet, but that would be something to do for native vs pv.

And then there seem to be one or two tunable parameters we could experiment with.

Quoted 7 lines

The paravirt locks may need a bit more tuning. Some simple testing under KVM shows we might be a bit slower in some cases. Whether this is fairness or something else I'm not sure. The current simple pv spinlock code can do a directed yield to the lock holder CPU, whereas the pv qspl here just does a general yield. I think we might actually be able to change that to also support directed yield. Though I'm not sure if this is actually the cause of the slowdown yet.

Quoted 3 lines

Regarding the paravirt lock, I have taken a further look into the current PPC spinlock code. There is an equivalent of pv_wait() but no pv_kick(). Maybe PPC doesn't really need that.

So powerpc has two types of wait, either undirected "all processors" or directed to a specific processor which has been preempted by the hypervisor.

The simple spinlock code does a directed wait, because it knows the CPU which is holding the lock. In this case, there is a sequence that is used to ensure we don't wait if the condition has become true, and the target CPU does not need to kick the waiter it will happen automatically (see splpar_spin_yield). This is preferable because we only wait as needed and don't require the kick operation.

The pv spinlock code I did uses the undirected wait, because we don't know the CPU number which we are waiting on. This is undesirable because it's higher overhead and the wait is not so accurate.

I think perhaps we could change things so we wait on the correct CPU when queued, which might be good enough (we could also put the lock owner CPU in the spinlock word, if we add another format).

Quoted 9 lines

Attached are two
additional qspinlock patches that adds a CONFIG_PARAVIRT_QSPINLOCKS_LITE option to not require pv_kick(). There is also a fixup patch to be applied after your patchset.

I don't have access to a PPC LPAR with shared processor at the moment, so I can't test the performance of the paravirt code. Would you mind adding my patches and do some performance test on your end to see if it gives better result?

Great, I'll do some tests. Any suggestions for what to try?

Thanks,
Nick

reviewer
#33 lore raw
Quoted 25 lines · Nicholas Piggin

Excerpts from Waiman Long's message of July 8, 2020 1:33 pm:

Quoted 17 lines · Nicholas Piggin
Quoted 9 lines

Yes, powerpc could certainly get more performance out of the slow paths, and then there are a few parameters to tune.

We don't have a good alternate patching for function calls yet, but that would be something to do for native vs pv.

And then there seem to be one or two tunable parameters we could experiment with.

Quoted 7 lines

The paravirt locks may need a bit more tuning. Some simple testing under KVM shows we might be a bit slower in some cases. Whether this is fairness or something else I'm not sure. The current simple pv spinlock code can do a directed yield to the lock holder CPU, whereas the pv qspl here just does a general yield. I think we might actually be able to change that to also support directed yield. Though I'm not sure if this is actually the cause of the slowdown yet.

Quoted 3 lines

Regarding the paravirt lock, I have taken a further look into the current PPC spinlock code. There is an equivalent of pv_wait() but no pv_kick(). Maybe PPC doesn't really need that.

So powerpc has two types of wait, either undirected "all processors" or directed to a specific processor which has been preempted by the hypervisor.

Quoted 6 lines

The simple spinlock code does a directed wait, because it knows the CPU which is holding the lock. In this case, there is a sequence that is used to ensure we don't wait if the condition has become true, and the target CPU does not need to kick the waiter it will happen automatically (see splpar_spin_yield). This is preferable because we only wait as needed and don't require the kick operation.

Thanks for the explanation.

Quoted 8 lines

The pv spinlock code I did uses the undirected wait, because we don't know the CPU number which we are waiting on. This is undesirable because it's higher overhead and the wait is not so accurate.

I think perhaps we could change things so we wait on the correct CPU when queued, which might be good enough (we could also put the lock owner CPU in the spinlock word, if we add another format).

The LS byte of the lock word is used to indicate locking status. If we have less than 255 cpus, we can put the (cpu_nr + 1) into the lock byte. The special 0xff value can be used to indicate a cpu number >= 255 for indirect yield. The required change to the qspinlock code will be minimal, I think.

Quoted 10 lines
Quoted 9 lines

Attached are two
additional qspinlock patches that adds a CONFIG_PARAVIRT_QSPINLOCKS_LITE option to not require pv_kick(). There is also a fixup patch to be applied after your patchset.

I don't have access to a PPC LPAR with shared processor at the moment, so I can't test the performance of the paravirt code. Would you mind adding my patches and do some performance test on your end to see if it gives better result?

Great, I'll do some tests. Any suggestions for what to try?

I will just like to see if it will produce some better performance result compared with your current version.

Cheers,
Longman

reviewer
#34 lore raw
Quoted 42 lines · Waiman Long
Quoted 25 lines · Nicholas Piggin

Excerpts from Waiman Long's message of July 8, 2020 1:33 pm:

Quoted 17 lines · Nicholas Piggin
Quoted 9 lines

Yes, powerpc could certainly get more performance out of the slow paths, and then there are a few parameters to tune.

We don't have a good alternate patching for function calls yet, but that would be something to do for native vs pv.

And then there seem to be one or two tunable parameters we could experiment with.

Quoted 7 lines

The paravirt locks may need a bit more tuning. Some simple testing under KVM shows we might be a bit slower in some cases. Whether this is fairness or something else I'm not sure. The current simple pv spinlock code can do a directed yield to the lock holder CPU, whereas the pv qspl here just does a general yield. I think we might actually be able to change that to also support directed yield. Though I'm not sure if this is actually the cause of the slowdown yet.

Quoted 3 lines

Regarding the paravirt lock, I have taken a further look into the current PPC spinlock code. There is an equivalent of pv_wait() but no pv_kick(). Maybe PPC doesn't really need that.

So powerpc has two types of wait, either undirected "all processors" or directed to a specific processor which has been preempted by the hypervisor.

Quoted 6 lines

The simple spinlock code does a directed wait, because it knows the CPU which is holding the lock. In this case, there is a sequence that is used to ensure we don't wait if the condition has become true, and the target CPU does not need to kick the waiter it will happen automatically (see splpar_spin_yield). This is preferable because we only wait as needed and don't require the kick operation.

Thanks for the explanation.

Quoted 8 lines

The pv spinlock code I did uses the undirected wait, because we don't know the CPU number which we are waiting on. This is undesirable because it's higher overhead and the wait is not so accurate.

I think perhaps we could change things so we wait on the correct CPU when queued, which might be good enough (we could also put the lock owner CPU in the spinlock word, if we add another format).

Quoted 5 lines

The LS byte of the lock word is used to indicate locking status. If we have less than 255 cpus, we can put the (cpu_nr + 1) into the lock byte. The special 0xff value can be used to indicate a cpu number >= 255 for indirect yield. The required change to the qspinlock code will be minimal, I think.

BTW, we can also keep track of the previous cpu in the waiting queue. Due to lock stealing, that may not be the cpu that is holding the lock. Maybe we can use this, if available, in case the cpu number is >= 255.

Regards,
Longman

maintainer
#35 lore raw
In reply to Waiman Long
Quoted 13 lines · Waiman Long

From 5d7941a498935fb225b2c7a3108cbf590114c3db Mon Sep 17 00:00:00 2001
From: Waiman Long <[email protected]>
Date: Tue, 7 Jul 2020 22:29:16 -0400
Subject: [PATCH 2/9] locking/pvqspinlock: Introduce
CONFIG_PARAVIRT_QSPINLOCKS_LITE

Add a new PARAVIRT_QSPINLOCKS_LITE config option that allows architectures to use the PV qspinlock code without the need to use or implement a pv_kick() function, thus eliminating the atomic unlock overhead. The non-atomic queued_spin_unlock() can be used instead. The pv_wait() function will still be needed, but it can be a dummy function.

Quoted 2 lines

With that option set, the hybrid PV queued/unfair locking code should still be able to make it performant enough in a paravirtualized

How is this supposed to work? If there is no kick, you have no control over who wakes up and fairness goes out the window entirely.

You don't even begin to explain...

reviewer
#36 lore raw
Quoted 14 lines · Peter Zijlstra
Quoted 13 lines · Waiman Long

From 5d7941a498935fb225b2c7a3108cbf590114c3db Mon Sep 17 00:00:00 2001
From: Waiman Long <[email protected]>
Date: Tue, 7 Jul 2020 22:29:16 -0400
Subject: [PATCH 2/9] locking/pvqspinlock: Introduce
CONFIG_PARAVIRT_QSPINLOCKS_LITE

Add a new PARAVIRT_QSPINLOCKS_LITE config option that allows architectures to use the PV qspinlock code without the need to use or implement a pv_kick() function, thus eliminating the atomic unlock overhead. The non-atomic queued_spin_unlock() can be used instead. The pv_wait() function will still be needed, but it can be a dummy function.

Quoted 7 lines
Quoted 2 lines

With that option set, the hybrid PV queued/unfair locking code should still be able to make it performant enough in a paravirtualized

How is this supposed to work? If there is no kick, you have no control over who wakes up and fairness goes out the window entirely.

You don't even begin to explain...

I don't have a full understanding of how the PPC hypervisor work myself. Apparently, a cpu kick may not be needed.

This is just a test patch to see if it yields better result. It is subjected to further modifcation.

Cheers,
Longman

maintainer
#37 lore raw
In reply to Nicholas Piggin
Quoted 2 lines · Nicholas Piggin

Yes, powerpc could certainly get more performance out of the slow paths, and then there are a few parameters to tune.

Can you clarify? The slow path is already in use on ARM64 which is weak, so I doubt there's superfluous serialization present. And Will spend a fair amount of time on making that thing guarantee forward progressm, so there just isn't too much room to play.

Quoted 2 lines

We don't have a good alternate patching for function calls yet, but that would be something to do for native vs pv.

Going by your jump_label implementation, support for static_call should be fairly straight forward too, no?

https://lkml.kernel.org/r/[email protected]

reviewer
#38 lore raw
Quoted 8 lines · Peter Zijlstra
Quoted 2 lines · Nicholas Piggin

Yes, powerpc could certainly get more performance out of the slow paths, and then there are a few parameters to tune.

Can you clarify? The slow path is already in use on ARM64 which is weak, so I doubt there's superfluous serialization present. And Will spend a fair amount of time on making that thing guarantee forward progressm, so there just isn't too much room to play.

Quoted 7 lines
Quoted 2 lines

We don't have a good alternate patching for function calls yet, but that would be something to do for native vs pv.

Going by your jump_label implementation, support for static_call should be fairly straight forward too, no?

   https://lkml.kernel.org/r/[email protected]

Speaking of static_call, I am also looking forward to it. Do you have an idea when that will be merged?

Cheers,
Longman

maintainer
#39 lore raw
Quoted 11 lines · Waiman Long
Quoted 10 lines · Peter Zijlstra
Quoted 2 lines · Nicholas Piggin

Yes, powerpc could certainly get more performance out of the slow paths, and then there are a few parameters to tune.

Can you clarify? The slow path is already in use on ARM64 which is weak, so I doubt there's superfluous serialization present. And Will spend a fair amount of time on making that thing guarantee forward progressm, so there just isn't too much room to play.

Quoted 2 lines

We don't have a good alternate patching for function calls yet, but that would be something to do for native vs pv.

Quoted 7 lines
Quoted 5 lines

Going by your jump_label implementation, support for static_call should be fairly straight forward too, no?

   https://lkml.kernel.org/r/[email protected]

Speaking of static_call, I am also looking forward to it. Do you have an idea when that will be merged?

0day had one crash on the last round, I think Steve send a fix for that last night and I'll go look at it.

That said, the last posting got 0 feedback, so either everybody is really happy with it, or not interested. So let us know in the thread, with some review feedback.

Once I get through enough of the inbox to actually find the fix and test it, I'll also update the thread, and maybe threaten to merge it if everybody stays silent :-)

author reviewer
#40 lore raw

Excerpts from Peter Zijlstra's message of July 9, 2020 6:31 pm:

Quoted 23 lines · Waiman Long
Quoted 11 lines · Peter Zijlstra
Quoted 10 lines · Nicholas Piggin
Quoted 2 lines

Yes, powerpc could certainly get more performance out of the slow paths, and then there are a few parameters to tune.

Can you clarify? The slow path is already in use on ARM64 which is weak, so I doubt there's superfluous serialization present. And Will spend a fair amount of time on making that thing guarantee forward progressm, so there just isn't too much room to play.

Quoted 2 lines

We don't have a good alternate patching for function calls yet, but that would be something to do for native vs pv.

Quoted 7 lines
Quoted 5 lines

Going by your jump_label implementation, support for static_call should be fairly straight forward too, no?

   https://lkml.kernel.org/r/[email protected]

Speaking of static_call, I am also looking forward to it. Do you have an idea when that will be merged?

0day had one crash on the last round, I think Steve send a fix for that last night and I'll go look at it.

Quoted 7 lines

That said, the last posting got 0 feedback, so either everybody is really happy with it, or not interested. So let us know in the thread, with some review feedback.

Once I get through enough of the inbox to actually find the fix and test it, I'll also update the thread, and maybe threaten to merge it if everybody stays silent :-)

I'd like to use it in powerpc. We have code now for example that patches a branch immediately at the top of memcpy which branches to a different version of the function. pv queued spinlock selection obviously, and there's a bunch of platform ops struct things that get filled in at boot time, etc.

So +1 here if you can get them through. I'm not 100% sure we can do it with existing toolchain and no ugly hacks, but there's no way to structure things that can get around that AFAIKS. We'd eventually use it though, I'd say.

Thanks,
Nick

author reviewer
#41 lore raw
In reply to Peter Zijlstra

Excerpts from Peter Zijlstra's message of July 8, 2020 6:41 pm:

Quoted 4 lines · Nicholas Piggin
Quoted 2 lines

Yes, powerpc could certainly get more performance out of the slow paths, and then there are a few parameters to tune.

Sorry for the delay, got bogged down and distracted by other things :(

Quoted 4 lines

Can you clarify? The slow path is already in use on ARM64 which is weak, so I doubt there's superfluous serialization present. And Will spend a fair amount of time on making that thing guarantee forward progressm, so there just isn't too much room to play.

Sure, the way the pending not-queued slowpath (which I guess is the medium-path) is implemented is just poorly structured for LL/SC. It has one more atomic than necessary (queued_fetch_set_pending_acquire), and a lot of branches in suboptimal order.

Attached patch (completely untested just compiled and looked at asm so far) is a way we can fix this on powerpc I think. It's actually very little generic code change which is good, duplicated medium-path logic unfortunately but that's no worse than something like x86 really.

Quoted 7 lines
Quoted 2 lines

We don't have a good alternate patching for function calls yet, but that would be something to do for native vs pv.

Going by your jump_label implementation, support for static_call should be fairly straight forward too, no?

https://lkml.kernel.org/r/[email protected]

Nice, yeah it should be. I've wanted this for ages!

powerpc is kind of annoying to implement that with limited call range, Hmm, not sure if we'd need a new linker feature to support it. We'd provide call site patch space for indirect branches for those out of range of direct call, so that should work fine. The trick would be patching in the TOC lookup for the function... should be doable somehow.

Thanks,
Nick

---

arch/powerpc/include/asm/qspinlock.h +46 -5
index b752d34517b3..26d8766a1106 100644
@@ -31,16 +31,57 @@ static inline void queued_spin_unlock(struct qspinlock *lock)
 
 #else
 extern void queued_spin_lock_slowpath(struct qspinlock *lock, u32 val);
+extern void queued_spin_lock_slowpath_queue(struct qspinlock *lock);
 #endif
 
 static __always_inline void queued_spin_lock(struct qspinlock *lock)
 {
-	u32 val = 0;
-
-	if (likely(atomic_try_cmpxchg_lock(&lock->val, &val, _Q_LOCKED_VAL)))
+	atomic_t *a = &lock->val;
+	u32 val;
+
+again:
+	asm volatile(
+"1:\t"	PPC_LWARX(%0,0,%1,1) "	# queued_spin_lock			\n"
+	: "=&r" (val)
+	: "r" (&a->counter)
+	: "memory");
+
+	if (likely(val == 0)) {
+		asm_volatile_goto(
+	"	stwcx.	%0,0,%1							\n"
+	"	bne-	%l[again]						\n"
+	"\t"	PPC_ACQUIRE_BARRIER "						\n"
+		:
+		: "r"(_Q_LOCKED_VAL), "r" (&a->counter)
+		: "cr0", "memory"
+		: again );
 		return;
-
-	queued_spin_lock_slowpath(lock, val);
+	}
+
+	if (likely(val == _Q_LOCKED_VAL)) {
+		asm_volatile_goto(
+	"	stwcx.	%0,0,%1							\n"
+	"	bne-	%l[again]						\n"
+		:
+		: "r"(_Q_LOCKED_VAL | _Q_PENDING_VAL), "r" (&a->counter)
+		: "cr0", "memory"
+		: again );
+
+		atomic_cond_read_acquire(a, !(VAL & _Q_LOCKED_MASK));
+//		clear_pending_set_locked(lock);
+		WRITE_ONCE(lock->locked_pending, _Q_LOCKED_VAL);
+//		lockevent_inc(lock_pending);
+		return;
+	}
+
+	if (val == _Q_PENDING_VAL) {
+		int cnt = _Q_PENDING_LOOPS;
+		val = atomic_cond_read_relaxed(a,
+					       (VAL != _Q_PENDING_VAL) || !cnt--);
+		if (!(val & ~_Q_LOCKED_MASK))
+			goto again;
+        }
+	queued_spin_lock_slowpath_queue(lock);
 }
 #define queued_spin_lock queued_spin_lock
 
kernel/locking/qspinlock.c +24 -7
index b9515fcc9b29..ebcc6f5d99d5 100644
@@ -287,10 +287,14 @@ static __always_inline u32  __pv_wait_head_or_lock(struct qspinlock *lock,
 
 #ifdef CONFIG_PARAVIRT_SPINLOCKS
 #define queued_spin_lock_slowpath	native_queued_spin_lock_slowpath
+#define queued_spin_lock_slowpath_queue	native_queued_spin_lock_slowpath_queue
 #endif
 
 #endif /* _GEN_PV_LOCK_SLOWPATH */
 
+void queued_spin_lock_slowpath_queue(struct qspinlock *lock);
+static void __queued_spin_lock_slowpath_queue(struct qspinlock *lock);
+
 /**
  * queued_spin_lock_slowpath - acquire the queued spinlock
  * @lock: Pointer to queued spinlock structure
@@ -314,12 +318,6 @@ static __always_inline u32  __pv_wait_head_or_lock(struct qspinlock *lock,
  */
 void queued_spin_lock_slowpath(struct qspinlock *lock, u32 val)
 {
-	struct mcs_spinlock *prev, *next, *node;
-	u32 old, tail;
-	int idx;
-
-	BUILD_BUG_ON(CONFIG_NR_CPUS >= (1U << _Q_TAIL_CPU_BITS));
-
 	if (pv_enabled())
 		goto pv_queue;
 
@@ -397,6 +395,26 @@ void queued_spin_lock_slowpath(struct qspinlock *lock, u32 val)
 queue:
 	lockevent_inc(lock_slowpath);
 pv_queue:
+	__queued_spin_lock_slowpath_queue(lock);
+}
+EXPORT_SYMBOL(queued_spin_lock_slowpath);
+
+void queued_spin_lock_slowpath_queue(struct qspinlock *lock)
+{
+	lockevent_inc(lock_slowpath);
+	__queued_spin_lock_slowpath_queue(lock);
+}
+EXPORT_SYMBOL(queued_spin_lock_slowpath_queue);
+
+static void __queued_spin_lock_slowpath_queue(struct qspinlock *lock)
+{
+	struct mcs_spinlock *prev, *next, *node;
+	u32 old, tail;
+	u32 val;
+	int idx;
+
+	BUILD_BUG_ON(CONFIG_NR_CPUS >= (1U << _Q_TAIL_CPU_BITS));
+
 	node = this_cpu_ptr(&qnodes[0].mcs);
 	idx = node->count++;
 	tail = encode_tail(smp_processor_id(), idx);
@@ -559,7 +577,6 @@ void queued_spin_lock_slowpath(struct qspinlock *lock, u32 val)
 	 */
 	__this_cpu_dec(qnodes[0].mcs.count);
 }
-EXPORT_SYMBOL(queued_spin_lock_slowpath);
 
 /*
  * Generate the paravirt code for queued_spin_unlock_slowpath().
reviewer
#42 lore raw
Quoted 57 lines · Nicholas Piggin
arch/powerpc/include/asm/qspinlock.h +39 -5
index b752d34517b3..26d8766a1106 100644
@@ -31,16 +31,57 @@ static inline void queued_spin_unlock(struct qspinlock *lock)
 
  #else
  extern void queued_spin_lock_slowpath(struct qspinlock *lock, u32 val);
+extern void queued_spin_lock_slowpath_queue(struct qspinlock *lock);
  #endif
 
  static __always_inline void queued_spin_lock(struct qspinlock *lock)
  {
-	u32 val = 0;
-
-	if (likely(atomic_try_cmpxchg_lock(&lock->val, &val, _Q_LOCKED_VAL)))
+	atomic_t *a = &lock->val;
+	u32 val;
+
+again:
+	asm volatile(
+"1:\t"	PPC_LWARX(%0,0,%1,1) "	# queued_spin_lock			\n"
+	: "=&r" (val)
+	: "r" (&a->counter)
+	: "memory");
+
+	if (likely(val = 0)) {
+		asm_volatile_goto(
+	"	stwcx.	%0,0,%1							\n"
+	"	bne-	%l[again]						\n"
+	"\t"	PPC_ACQUIRE_BARRIER "						\n"
+		:
+		: "r"(_Q_LOCKED_VAL), "r" (&a->counter)
+		: "cr0", "memory"
+		: again );
  		return;
-
-	queued_spin_lock_slowpath(lock, val);
+	}
+
+	if (likely(val = _Q_LOCKED_VAL)) {
+		asm_volatile_goto(
+	"	stwcx.	%0,0,%1							\n"
+	"	bne-	%l[again]						\n"
+		:
+		: "r"(_Q_LOCKED_VAL | _Q_PENDING_VAL), "r" (&a->counter)
+		: "cr0", "memory"
+		: again );
+
+		atomic_cond_read_acquire(a, !(VAL & _Q_LOCKED_MASK));
+//		clear_pending_set_locked(lock);
+		WRITE_ONCE(lock->locked_pending, _Q_LOCKED_VAL);
+//		lockevent_inc(lock_pending);
+		return;
+	}
+
+	if (val = _Q_PENDING_VAL) {

arch/powerpc/include/asm/qspinlock.h static inline void queued_spin_unlock(struct qspinlock *lock) · line 87

Quoted 10 lines
arch/powerpc/include/asm/qspinlock.h +7 -0
+		int cnt = _Q_PENDING_LOOPS;
+		val = atomic_cond_read_relaxed(a,
+					       (VAL != _Q_PENDING_VAL) || !cnt--);
+		if (!(val & ~_Q_LOCKED_MASK))
+			goto again;
+        }
+	queued_spin_lock_slowpath_queue(lock);
  }
  #define queued_spin_lock queued_spin_lock
 

I am fine with the arch code override some part of the generic code.

Quoted 61 lines
kernel/locking/qspinlock.c +24 -6
index b9515fcc9b29..ebcc6f5d99d5 100644
@@ -287,10 +287,14 @@ static __always_inline u32  __pv_wait_head_or_lock(struct qspinlock *lock,
 
  #ifdef CONFIG_PARAVIRT_SPINLOCKS
  #define queued_spin_lock_slowpath	native_queued_spin_lock_slowpath
+#define queued_spin_lock_slowpath_queue	native_queued_spin_lock_slowpath_queue
  #endif
 
  #endif /* _GEN_PV_LOCK_SLOWPATH */
 
+void queued_spin_lock_slowpath_queue(struct qspinlock *lock);
+static void __queued_spin_lock_slowpath_queue(struct qspinlock *lock);
+
  /**
   * queued_spin_lock_slowpath - acquire the queued spinlock
   * @lock: Pointer to queued spinlock structure
@@ -314,12 +318,6 @@ static __always_inline u32  __pv_wait_head_or_lock(struct qspinlock *lock,
   */
  void queued_spin_lock_slowpath(struct qspinlock *lock, u32 val)
  {
-	struct mcs_spinlock *prev, *next, *node;
-	u32 old, tail;
-	int idx;
-
-	BUILD_BUG_ON(CONFIG_NR_CPUS >= (1U << _Q_TAIL_CPU_BITS));
-
  	if (pv_enabled())
  		goto pv_queue;
 
@@ -397,6 +395,26 @@ void queued_spin_lock_slowpath(struct qspinlock *lock, u32 val)
  queue:
  	lockevent_inc(lock_slowpath);
  pv_queue:
+	__queued_spin_lock_slowpath_queue(lock);
+}
+EXPORT_SYMBOL(queued_spin_lock_slowpath);
+
+void queued_spin_lock_slowpath_queue(struct qspinlock *lock)
+{
+	lockevent_inc(lock_slowpath);
+	__queued_spin_lock_slowpath_queue(lock);
+}
+EXPORT_SYMBOL(queued_spin_lock_slowpath_queue);
+
+static void __queued_spin_lock_slowpath_queue(struct qspinlock *lock)
+{
+	struct mcs_spinlock *prev, *next, *node;
+	u32 old, tail;
+	u32 val;
+	int idx;
+
+	BUILD_BUG_ON(CONFIG_NR_CPUS >= (1U << _Q_TAIL_CPU_BITS));
+
  	node = this_cpu_ptr(&qnodes[0].mcs);
  	idx = node->count++;
  	tail = encode_tail(smp_processor_id(), idx);

kernel/locking/qspinlock.c void queued_spin_lock_slowpath(struct qspinlock *lock, u32 val) · line 582

Quoted 7 lines
kernel/locking/qspinlock.c +0 -1
@@ -559,7 +577,6 @@ void queued_spin_lock_slowpath(struct qspinlock *lock, u32 val)
  	 */
  	__this_cpu_dec(qnodes[0].mcs.count);
  }
-EXPORT_SYMBOL(queued_spin_lock_slowpath);
 
  /*
   * Generate the paravirt code for queued_spin_unlock_slowpath().

I would prefer to extract out the pending bit handling code out into a separate helper function which can be overridden by the arch code instead of breaking the slowpath into 2 pieces.

Cheers,
Longman

author reviewer
#43 lore raw

Excerpts from Waiman Long's message of July 22, 2020 12:36 am:

Quoted 62 lines · Nicholas Piggin
Quoted 61 lines
arch/powerpc/include/asm/qspinlock.h +43 -5
index b752d34517b3..26d8766a1106 100644
@@ -31,16 +31,57 @@ static inline void queued_spin_unlock(struct qspinlock *lock)
 
  #else
  extern void queued_spin_lock_slowpath(struct qspinlock *lock, u32 val);
+extern void queued_spin_lock_slowpath_queue(struct qspinlock *lock);
  #endif
 
  static __always_inline void queued_spin_lock(struct qspinlock *lock)
  {
-	u32 val = 0;
-
-	if (likely(atomic_try_cmpxchg_lock(&lock->val, &val, _Q_LOCKED_VAL)))
+	atomic_t *a = &lock->val;
+	u32 val;
+
+again:
+	asm volatile(
+"1:\t"	PPC_LWARX(%0,0,%1,1) "	# queued_spin_lock			\n"
+	: "=&r" (val)
+	: "r" (&a->counter)
+	: "memory");
+
+	if (likely(val == 0)) {
+		asm_volatile_goto(
+	"	stwcx.	%0,0,%1							\n"
+	"	bne-	%l[again]						\n"
+	"\t"	PPC_ACQUIRE_BARRIER "						\n"
+		:
+		: "r"(_Q_LOCKED_VAL), "r" (&a->counter)
+		: "cr0", "memory"
+		: again );
  		return;
-
-	queued_spin_lock_slowpath(lock, val);
+	}
+
+	if (likely(val == _Q_LOCKED_VAL)) {
+		asm_volatile_goto(
+	"	stwcx.	%0,0,%1							\n"
+	"	bne-	%l[again]						\n"
+		:
+		: "r"(_Q_LOCKED_VAL | _Q_PENDING_VAL), "r" (&a->counter)
+		: "cr0", "memory"
+		: again );
+
+		atomic_cond_read_acquire(a, !(VAL & _Q_LOCKED_MASK));
+//		clear_pending_set_locked(lock);
+		WRITE_ONCE(lock->locked_pending, _Q_LOCKED_VAL);
+//		lockevent_inc(lock_pending);
+		return;
+	}
+
+	if (val == _Q_PENDING_VAL) {
+		int cnt = _Q_PENDING_LOOPS;
+		val = atomic_cond_read_relaxed(a,
+					       (VAL != _Q_PENDING_VAL) || !cnt--);
+		if (!(val & ~_Q_LOCKED_MASK))
Quoted 8 lines

arch/powerpc/include/asm/qspinlock.h static inline void queued_spin_unlock(struct qspinlock *lock) · line 87

Quoted 6 lines
arch/powerpc/include/asm/qspinlock.h +3 -0
+			goto again;
+        }
+	queued_spin_lock_slowpath_queue(lock);
  }
  #define queued_spin_lock queued_spin_lock
 

I am fine with the arch code override some part of the generic code.

Cool.

Quoted 65 lines
Quoted 65 lines
kernel/locking/qspinlock.c +24 -7
index b9515fcc9b29..ebcc6f5d99d5 100644
@@ -287,10 +287,14 @@ static __always_inline u32  __pv_wait_head_or_lock(struct qspinlock *lock,
 
  #ifdef CONFIG_PARAVIRT_SPINLOCKS
  #define queued_spin_lock_slowpath	native_queued_spin_lock_slowpath
+#define queued_spin_lock_slowpath_queue	native_queued_spin_lock_slowpath_queue
  #endif
 
  #endif /* _GEN_PV_LOCK_SLOWPATH */
 
+void queued_spin_lock_slowpath_queue(struct qspinlock *lock);
+static void __queued_spin_lock_slowpath_queue(struct qspinlock *lock);
+
  /**
   * queued_spin_lock_slowpath - acquire the queued spinlock
   * @lock: Pointer to queued spinlock structure
@@ -314,12 +318,6 @@ static __always_inline u32  __pv_wait_head_or_lock(struct qspinlock *lock,
   */
  void queued_spin_lock_slowpath(struct qspinlock *lock, u32 val)
  {
-	struct mcs_spinlock *prev, *next, *node;
-	u32 old, tail;
-	int idx;
-
-	BUILD_BUG_ON(CONFIG_NR_CPUS >= (1U << _Q_TAIL_CPU_BITS));
-
  	if (pv_enabled())
  		goto pv_queue;
 
@@ -397,6 +395,26 @@ void queued_spin_lock_slowpath(struct qspinlock *lock, u32 val)
  queue:
  	lockevent_inc(lock_slowpath);
  pv_queue:
+	__queued_spin_lock_slowpath_queue(lock);
+}
+EXPORT_SYMBOL(queued_spin_lock_slowpath);
+
+void queued_spin_lock_slowpath_queue(struct qspinlock *lock)
+{
+	lockevent_inc(lock_slowpath);
+	__queued_spin_lock_slowpath_queue(lock);
+}
+EXPORT_SYMBOL(queued_spin_lock_slowpath_queue);
+
+static void __queued_spin_lock_slowpath_queue(struct qspinlock *lock)
+{
+	struct mcs_spinlock *prev, *next, *node;
+	u32 old, tail;
+	u32 val;
+	int idx;
+
+	BUILD_BUG_ON(CONFIG_NR_CPUS >= (1U << _Q_TAIL_CPU_BITS));
+
  	node = this_cpu_ptr(&qnodes[0].mcs);
  	idx = node->count++;
  	tail = encode_tail(smp_processor_id(), idx);
@@ -559,7 +577,6 @@ void queued_spin_lock_slowpath(struct qspinlock *lock, u32 val)
  	 */
  	__this_cpu_dec(qnodes[0].mcs.count);
  }
-EXPORT_SYMBOL(queued_spin_lock_slowpath);
 
Quoted 6 lines

kernel/locking/qspinlock.c void queued_spin_lock_slowpath(struct qspinlock *lock, u32 val) · line 582

Quoted 3 lines
kernel/locking/qspinlock.c +0 -0
  /*
   * Generate the paravirt code for queued_spin_unlock_slowpath().

I would prefer to extract out the pending bit handling code out into a separate helper function which can be overridden by the arch code instead of breaking the slowpath into 2 pieces.

You mean have the arch provide a queued_spin_lock_slowpath_pending function that the slow path calls?

I would actually prefer the pending handling can be made inline in the queued_spin_lock function, especially with out-of-line locks it makes sense to put it there.

We could ifdef out queued_spin_lock_slowpath_queue if it's not used, then __queued_spin_lock_slowpath_queue would be inlined into the caller so there would be no split?

Thanks,
Nick

reviewer
#44 lore raw
Quoted 6 lines · Nicholas Piggin
Quoted 3 lines

I would prefer to extract out the pending bit handling code out into a separate helper function which can be overridden by the arch code instead of breaking the slowpath into 2 pieces.

You mean have the arch provide a queued_spin_lock_slowpath_pending function that the slow path calls?

Quoted 7 lines

I would actually prefer the pending handling can be made inline in the queued_spin_lock function, especially with out-of-line locks it makes sense to put it there.

We could ifdef out queued_spin_lock_slowpath_queue if it's not used, then __queued_spin_lock_slowpath_queue would be inlined into the caller so there would be no split?

The pending code is an optimization for lightly contended locks. That is why I think it is appropriate to extract it into a helper function and mark it as such.

You can certainly put the code in the arch's spin_lock code, you just has to override the generic pending code by a null function.

Cheers,
Longman

author reviewer
#45 lore raw

Excerpts from Waiman Long's message of July 24, 2020 12:29 am:

Quoted 15 lines · Nicholas Piggin
Quoted 6 lines
Quoted 3 lines

I would prefer to extract out the pending bit handling code out into a separate helper function which can be overridden by the arch code instead of breaking the slowpath into 2 pieces.

You mean have the arch provide a queued_spin_lock_slowpath_pending function that the slow path calls?

Quoted 7 lines

I would actually prefer the pending handling can be made inline in the queued_spin_lock function, especially with out-of-line locks it makes sense to put it there.

We could ifdef out queued_spin_lock_slowpath_queue if it's not used, then __queued_spin_lock_slowpath_queue would be inlined into the caller so there would be no split?

Quoted 6 lines

The pending code is an optimization for lightly contended locks. That is why I think it is appropriate to extract it into a helper function and mark it as such.

You can certainly put the code in the arch's spin_lock code, you just has to override the generic pending code by a null function.

I see what you mean. I guess that would work fine.

Thanks,
Nick