[PATCH v6 0/5] qrwlock: Introducing a queue read/write lock implementation

[PATCH v6 0/5] qrwlock: Introducing a queue read/write lock implementation

[Waiman Long]

v5->v6:
 - Modify queue_read_can_lock() to avoid false positive result.
 - Move the two slowpath functions' performance tuning change from
   patch 4 to patch 1.
 - Add a new optional patch to use the new smp_store_release() function 
   if that is merged.

v4->v5:
 - Fix wrong definitions for QW_MASK_FAIR & QW_MASK_UNFAIR macros.
 - Add an optional patch 4 which should only be applied after the
   mcs_spinlock.h header file is merged.

v3->v4:
 - Optimize the fast path with better cold cache behavior and
   performance.
 - Removing some testing code.
 - Make x86 use queue rwlock with no user configuration.

v2->v3:
 - Make read lock stealing the default and fair rwlock an option with
   a different initializer.
 - In queue_read_lock_slowpath(), check irq_count() and force spinning
   and lock stealing in interrupt context.
 - Unify the fair and classic read-side code path, and make write-side
   to use cmpxchg with 2 different writer states. This slows down the
   write lock fastpath to make the read side more efficient, but is
   still slightly faster than a spinlock.

v1->v2:
 - Improve lock fastpath performance.
 - Optionally provide classic read/write lock behavior for backward
   compatibility.
 - Use xadd instead of cmpxchg for fair reader code path to make it
   immute to reader contention.
 - Run more performance testing.

As mentioned in the LWN article http://lwn.net/Articles/364583/,
the read/write lock suffer from an unfairness problem that it is
possible for a stream of incoming readers to block a waiting writer
from getting the lock for a long time. Also, a waiting reader/writer
contending a rwlock in local memory will have a higher chance of
acquiring the lock than a reader/writer in remote node.

This patch set introduces a queue-based read/write lock implementation
that can largely solve this unfairness problem if the lock owners
choose to use the fair variant of the lock.

The queue rwlock has two variants selected at initialization time
- unfair (with read lock stealing) and fair (to both readers and
writers). The unfair rwlock is the default.

The read lock slowpath will check if the reader is in an interrupt
context. If so, it will force lock spinning and stealing without
waiting in a queue. This is to ensure the read lock will be granted
as soon as possible.

Even the unfair rwlock is fairer than the current version as there
is a higher chance for writers to get the lock and is fair among
the writers.

The queue write lock can also be used as a replacement for ticket
spinlocks that are highly contended if lock size increase is not
an issue.

This patch set currently provides queue read/write lock support on
x86 architecture only. Support for other architectures can be added
later on once architecture specific support infrastructure is added
and proper testing is done.

The optional patch 4 has a dependency on the the mcs_spinlock.h
header file which has not been merged yet. So this patch should only
be applied after the mcs_spinlock.h header file is merged.

The optional patch 5 use the new smp_store_release() and
smp_load_acquire() functions which are being reviewed & not merged yet.

Waiman Long (5):
  qrwlock: A queue read/write lock implementation
  qrwlock x86: Enable x86 to use queue read/write lock
  qrwlock: Enable fair queue read/write lock
  qrwlock: Use the mcs_spinlock helper functions for MCS queuing
  qrwlock: Use smp_store_release() in write_unlock()

 arch/x86/Kconfig                      |    1 +
 arch/x86/include/asm/spinlock.h       |    2 +
 arch/x86/include/asm/spinlock_types.h |    4 +
 include/asm-generic/qrwlock.h         |  245 +++++++++++++++++++++++++++++++++
 include/linux/rwlock.h                |   15 ++
 include/linux/rwlock_types.h          |   13 ++
 kernel/Kconfig.locks                  |    7 +
 lib/Makefile                          |    1 +
 lib/qrwlock.c                         |  205 +++++++++++++++++++++++++++
 lib/spinlock_debug.c                  |   19 +++
 10 files changed, 512 insertions(+), 0 deletions(-)
 create mode 100644 include/asm-generic/qrwlock.h
 create mode 100644 lib/qrwlock.c

[PATCH v6 1/5] qrwlock: A queue read/write lock implementation

[Waiman Long]

This patch introduces a new read/write lock implementation that put
waiting readers and writers into a queue instead of actively contending
the lock like the current read/write lock implementation. This will
improve performance in highly contended situation by reducing the
cache line bouncing effect.

The queue read/write lock (qrwlock) is mostly fair with respect to
the writers, even though there is still a slight chance of write
lock stealing.

Externally, there are two different types of readers - unfair (the
default) and fair. A unfair reader will try to steal read lock even
if a writer is waiting, whereas a fair reader will be waiting in
the queue under this circumstance.  These variants are chosen at
initialization time by using different initializers. The new *_fair()
initializers are added for selecting the use of fair reader.

Internally, there is a third type of readers which steal lock more
aggressively than the unfair reader. They simply increments the reader
count and wait until the writer releases the lock. The transition to
aggressive reader happens in the read lock slowpath when
1. In an interrupt context.
2. when a classic reader comes to the head of the wait queue.
3. When a fair reader comes to the head of the wait queue and sees
   the release of a write lock.

The fair queue rwlock is more deterministic in the sense that late
comers jumping ahead and stealing the lock is unlikely even though
there is still a very small chance for lock stealing to happen if
the readers or writers come at the right moment.  Other than that,
lock granting is done in a FIFO manner.  As a result, it is possible
to determine a maximum time period after which the waiting is over
and the lock can be acquired.

The queue read lock is safe to use in an interrupt context (softirq
or hardirq) as it will switch to become an aggressive reader in such
environment allowing recursive read lock. However, the fair readers
will not support recursive read lock in a non-interrupt environment
when a writer is waiting.

The only downside of queue rwlock is the size increase in the lock
structure by 4 bytes for 32-bit systems and by 12 bytes for 64-bit
systems.

This patch will replace the architecture specific implementation
of rwlock by this generic version of queue rwlock when the
ARCH_QUEUE_RWLOCK configuration parameter is set.

In term of single-thread performance (no contention), a 256K
lock/unlock loop was run on a 2.4GHz and 2.93Ghz Westmere x86-64
CPUs. The following table shows the average time (in ns) for a single
lock/unlock sequence (including the looping and timing overhead):

Lock Type		    2.4GHz	2.93GHz
---------		    ------	-------
Ticket spinlock		     14.9	 12.3
Read lock		     17.0	 13.5
Write lock		     17.0	 13.5
Queue read lock	     	     16.0	 13.5
Queue fair read lock	     16.0	 13.5
Queue write lock	      9.2	  7.8
Queue fair write lock	     17.5	 14.5

The queue read lock is slightly slower than the spinlock, but is
slightly faster than the read lock. The queue write lock, however,
is the fastest of all. It is almost twice as fast as the write lock
and about 1.5X of the spinlock. The queue fair write lock, on the
other hand, is slightly slower than the write lock.

With lock contention, the speed of each individual lock/unlock function
is less important than the amount of contention-induced delays.

To investigate the performance characteristics of the queue rwlock
compared with the regular rwlock, Ingo's anon_vmas patch that converts
rwsem to rwlock was applied to a 3.12 kernel. This kernel was then
tested under the following 4 conditions:

 1) Plain 3.12
 2) Ingo's patch
 3) Ingo's patch + unfair qrwlock (default)
 4) Ingo's patch + fair qrwlock

Each of the 4 kernels were booted up twice with and without the
"idle=poll" kernel parameter which keeps the CPUs in C0 state while
idling instead of a more energy-saving sleep state.  The jobs per
minutes (JPM) results of the AIM7's high_systime workload at 1500
users on a 8-socket 80-core DL980 (HT off) were:

 Kernel	    JPMs	%Change from (1)
 ------	    ----	----------------
   1	145704/227295		-
   2	229750/236066	    +58%/+3.8%
   3	248144/253521	    +70%/+11.5%
   4	240062/246251	    +65%/+8.3%

The first JPM number is without the "idle=poll" kernel parameter,
the second number is with that parameter. It can be seen that most
of the performance benefit of converting rwsem to rwlock actually
come from the latency improvement of not needing to wake up a CPU
from deep sleep state when work is available.

The use of non-sleeping locks did improve performance by eliminating
the context switching cost. Using queue rwlock gave almost tripling
of performance gain. The performance gain was reduced somewhat with
a fair lock which was to be expected.

Looking at the perf profiles (with idle=poll) below, we can clearly see
that other bottlenecks were constraining the performance improvement.

Perf profile of kernel (2):

 18.65%    reaim  [kernel.kallsyms]  [k] __write_lock_failed
  9.00%    reaim  [kernel.kallsyms]  [k] _raw_spin_lock_irqsave
  5.21%  swapper  [kernel.kallsyms]  [k] cpu_idle_loop
  3.08%    reaim  [kernel.kallsyms]  [k] mspin_lock
  2.50%    reaim  [kernel.kallsyms]  [k] anon_vma_interval_tree_insert
  2.00%       ls  [kernel.kallsyms]  [k] _raw_spin_lock_irqsave
  1.29%    reaim  [kernel.kallsyms]  [k] update_cfs_rq_blocked_load
  1.21%    reaim  [kernel.kallsyms]  [k] __read_lock_failed
  1.12%    reaim  [kernel.kallsyms]  [k] _raw_spin_lock
  1.10%    reaim  [kernel.kallsyms]  [k] perf_event_aux
  1.09%     true  [kernel.kallsyms]  [k] _raw_spin_lock_irqsave

Perf profile of kernel (3):

 20.99%  swapper  [kernel.kallsyms]  [k] cpu_idle_loop
  7.93%    reaim  [kernel.kallsyms]  [k] _raw_spin_lock_irqsave
  5.67%    reaim  [kernel.kallsyms]  [k] mspin_lock
  4.34%    reaim  [kernel.kallsyms]  [k] queue_write_lock_slowpath
  2.13%    reaim  [kernel.kallsyms]  [k] update_cfs_rq_blocked_load
  2.11%       ls  [kernel.kallsyms]  [k] _raw_spin_lock_irqsave
  2.07%    reaim  [kernel.kallsyms]  [k] anon_vma_interval_tree_insert
  1.17%    reaim  [kernel.kallsyms]  [k] queue_write_3step_lock
  1.10%     true  [kernel.kallsyms]  [k] _raw_spin_lock_irqsave
  0.95%    reaim  [kernel.kallsyms]  [k] mutex_spin_on_owner
  0.74%    reaim  [kernel.kallsyms]  [k] queue_read_lock_slowpath

The spinlock bottlenecks were shown below.

  7.93%    reaim  [kernel.kallsyms]  [k] _raw_spin_lock_irqsave
              |--60.57%-- release_pages
              |--36.62%-- pagevec_lru_move_fn
              |--0.80%-- get_page_from_freelist
              |--0.69%-- __page_cache_release
               --1.32%-- [...]

For both release_pages() & pagevec_lru_move_fn() function, the
spinlock contention was on zone->lru_lock. With the queue spinlock
patch, however, the contention went away with a lot more idle time
available and the JPM number went up to 265532 which was an additional
performance improvement.

 28.40%  swapper  [kernel.kallsyms]  [k] cpu_idle_loop
  6.89%    reaim  [kernel.kallsyms]  [k] mspin_lock
  4.17%    reaim  [kernel.kallsyms]  [k] queue_write_lock_slowpath
  2.10%    reaim  [kernel.kallsyms]  [k] anon_vma_interval_tree_insert
  1.82%    reaim  [kernel.kallsyms]  [k] update_cfs_rq_blocked_load
  1.34%    reaim  [kernel.kallsyms]  [k] entity_tick
  1.17%    reaim  [kernel.kallsyms]  [k] queue_write_3step_lock
  1.06%    reaim  [kernel.kallsyms]  [k] mutex_spin_on_owner
  0.86%    reaim  [kernel.kallsyms]  [k] perf_event_aux
  0.83%       ls  [kernel.kallsyms]  [k] mspin_lock
   :
  0.53%    reaim  [kernel.kallsyms]  [k] _raw_spin_lock
  0.14%    reaim  [kernel.kallsyms]  [k] _raw_spin_lock_irqsave

Tim Chen also tested the qrwlock with Ingo's patch on a 4-socket
machine.  It was found the performance improvement of 11% was the
same with regular rwlock or queue rwlock.

Signed-off-by: Waiman Long <Waiman.Long@hp.com>
---
 include/asm-generic/qrwlock.h |  250 +++++++++++++++++++++++++++++++++++++
 kernel/Kconfig.locks          |    7 +
 lib/Makefile                  |    1 +
 lib/qrwlock.c                 |  274 +++++++++++++++++++++++++++++++++++++++++
 4 files changed, 532 insertions(+), 0 deletions(-)
 create mode 100644 include/asm-generic/qrwlock.h
 create mode 100644 lib/qrwlock.c

diff --git a/include/asm-generic/qrwlock.h b/include/asm-generic/qrwlock.h
new file mode 100644
index 0000000..1f68499
--- /dev/null
+++ b/include/asm-generic/qrwlock.h
@@ -0,0 +1,250 @@
+/*
+ * Queue read/write lock
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * (C) Copyright 2013 Hewlett-Packard Development Company, L.P.
+ *
+ * Authors: Waiman Long <waiman.long@hp.com>
+ */
+#ifndef __ASM_GENERIC_QRWLOCK_H
+#define __ASM_GENERIC_QRWLOCK_H
+
+#include <linux/types.h>
+#include <asm/bitops.h>
+#include <asm/cmpxchg.h>
+#include <asm/barrier.h>
+#include <asm/processor.h>
+#include <asm/byteorder.h>
+
+#if !defined(__LITTLE_ENDIAN) && !defined(__BIG_ENDIAN)
+#error "Missing either LITTLE_ENDIAN or BIG_ENDIAN definition."
+#endif
+
+#if (CONFIG_NR_CPUS < 65536)
+typedef u16 __nrcpu_t;
+typedef u32 __nrcpupair_t;
+#define	QRW_READER_BIAS	(1U << 16)
+#else
+typedef u32 __nrcpu_t;
+typedef u64 __nrcpupair_t;
+#define	QRW_READER_BIAS	(1UL << 32)
+#endif
+
+/*
+ * The queue read/write lock data structure
+ *
+ * Read lock stealing can only happen when there is at least one reader
+ * holding the read lock. When the fair flag is not set, it mimics the
+ * behavior of the regular rwlock at the expense that a perpetual stream
+ * of readers could starve a writer for a long period of time. That
+ * behavior, however, may be beneficial to a workload that is reader heavy
+ * with slow writers, and the writers can wait without undesirable consequence.
+ * This fair flag should only be set at initialization time.
+ *
+ * The layout of the structure is endian-sensitive to make sure that adding
+ * QRW_READER_BIAS to the rw field to increment the reader count won't
+ * disturb the writer and the fair fields.
+ */
+struct qrwnode {
+	struct qrwnode *next;
+	bool		wait;	/* Waiting flag */
+};
+
+typedef struct qrwlock {
+	union qrwcnts {
+		struct {
+#ifdef __LITTLE_ENDIAN
+			u8	  writer;	/* Writer state		*/
+			u8	  fair;		/* Fair rwlock flag	*/
+			__nrcpu_t readers;	/* # of active readers	*/
+#else
+			__nrcpu_t readers;	/* # of active readers	*/
+			u8	  fair;		/* Fair rwlock flag	*/
+			u8	  writer;	/* Writer state		*/
+#endif
+		};
+		__nrcpupair_t rw;		/* Reader/writer number pair */
+	} cnts;
+	struct qrwnode *waitq;			/* Tail of waiting queue */
+} arch_rwlock_t;
+
+/*
+ * Writer state values & mask
+ */
+#define	QW_WAITING	1			/* A writer is waiting	   */
+#define	QW_LOCKED	0xff			/* A writer holds the lock */
+#define QW_MASK_FAIR	((u8)~0)		/* Mask for fair reader    */
+#define QW_MASK_UNFAIR	((u8)~QW_WAITING)	/* Mask for unfair reader  */
+#define	QW_GET_WMASK(c) ((c).fair ? QW_MASK_FAIR : QW_MASK_UNFAIR)
+
+/*
+ * External function declarations
+ */
+extern void queue_read_lock_slowpath(struct qrwlock *lock);
+extern void queue_write_lock_slowpath(struct qrwlock *lock);
+
+/**
+ * queue_read_can_lock- would read_trylock() succeed?
+ * @lock: Pointer to queue rwlock structure
+ */
+static inline int queue_read_can_lock(struct qrwlock *lock)
+{
+	union qrwcnts cnts;
+	u8 wmask;
+
+	cnts.rw = ACCESS_ONCE(lock->cnts.rw);
+	wmask   = QW_GET_WMASK(cnts);
+	return !(cnts.writer & wmask);
+}
+
+/**
+ * queue_write_can_lock- would write_trylock() succeed?
+ * @lock: Pointer to queue rwlock structure
+ */
+static inline int queue_write_can_lock(struct qrwlock *lock)
+{
+	union qrwcnts cnts;
+
+	cnts.rw = ACCESS_ONCE(lock->cnts.rw);
+	return !cnts.writer && !cnts.readers;
+}
+
+/**
+ * queue_read_trylock - try to acquire read lock of a queue rwlock
+ * @lock : Pointer to queue rwlock structure
+ * Return: 1 if lock acquired, 0 if failed
+ */
+static inline int queue_read_trylock(struct qrwlock *lock)
+{
+	union qrwcnts cnts;
+	u8 wmask;
+
+	cnts.rw = ACCESS_ONCE(lock->cnts.rw);
+	wmask   = QW_GET_WMASK(cnts);
+	if (likely(!(cnts.writer & wmask))) {
+		cnts.rw = xadd(&lock->cnts.rw, QRW_READER_BIAS);
+		if (likely(!(cnts.writer & wmask)))
+			return 1;
+		add_smp(&lock->cnts.readers, -1);
+	}
+	return 0;
+}
+
+/**
+ * queue_write_trylock - try to acquire write lock of a queue rwlock
+ * @lock : Pointer to queue rwlock structure
+ * Return: 1 if lock acquired, 0 if failed
+ */
+static inline int queue_write_trylock(struct qrwlock *lock)
+{
+	union qrwcnts old, new;
+
+	old.rw = ACCESS_ONCE(lock->cnts.rw);
+	if (likely(!old.writer && !old.readers)) {
+		new.rw = old.rw;
+		new.writer = QW_LOCKED;
+		if (likely(cmpxchg(&lock->cnts.rw, old.rw, new.rw) == old.rw))
+			return 1;
+	}
+	return 0;
+}
+/**
+ * queue_read_lock - acquire read lock of a queue rwlock
+ * @lock: Pointer to queue rwlock structure
+ */
+static inline void queue_read_lock(struct qrwlock *lock)
+{
+	union qrwcnts cnts;
+	u8 wmask;
+
+	cnts.rw = xadd(&lock->cnts.rw, QRW_READER_BIAS);
+	wmask   = QW_GET_WMASK(cnts);
+	if (likely(!(cnts.writer & wmask)))
+		return;
+	/*
+	 * Slowpath will decrement the reader count, if necessary
+	 */
+	queue_read_lock_slowpath(lock);
+}
+
+/**
+ * queue_write_lock - acquire write lock of a queue rwlock
+ * @lock : Pointer to queue rwlock structure
+ */
+static inline void queue_write_lock(struct qrwlock *lock)
+{
+	union qrwcnts old;
+
+	/*
+	 * Optimize for the unfair lock case where the fair flag is 0.
+	 */
+	old.rw = cmpxchg(&lock->cnts.rw, 0, QW_LOCKED);
+	if (likely(old.rw == 0))
+		return;
+	if (likely(!old.writer && !old.readers)) {
+		union qrwcnts new;
+
+		new.rw = old.rw;
+		new.writer = QW_LOCKED;
+		if (likely(cmpxchg(&lock->cnts.rw, old.rw, new.rw) == old.rw))
+			return;
+	}
+	queue_write_lock_slowpath(lock);
+}
+
+/**
+ * queue_read_unlock - release read lock of a queue rwlock
+ * @lock : Pointer to queue rwlock structure
+ */
+static inline void queue_read_unlock(struct qrwlock *lock)
+{
+	/*
+	 * Atomically decrement the reader count
+	 */
+	add_smp(&lock->cnts.readers, -1);
+}
+
+/**
+ * queue_write_unlock - release write lock of a queue rwlock
+ * @lock : Pointer to queue rwlock structure
+ */
+static inline void queue_write_unlock(struct qrwlock *lock)
+{
+	/*
+	 * Make sure that none of the critical section will be leaked out.
+	 */
+	smp_mb__before_clear_bit();
+	ACCESS_ONCE(lock->cnts.writer) = 0;
+	smp_mb__after_clear_bit();
+}
+
+/*
+ * Initializier
+ */
+#define	__ARCH_RW_LOCK_UNLOCKED	{ .cnts = { .rw = 0 }, .waitq = NULL }
+#define	__ARCH_RW_LOCK_UNLOCKED_FAIR	\
+	{ .cnts = { { .writer = 0, .fair = 1, .readers = 0 } }, .waitq = NULL }
+
+/*
+ * Remapping rwlock architecture specific functions to the corresponding
+ * queue rwlock functions.
+ */
+#define arch_read_can_lock(l)	queue_read_can_lock(l)
+#define arch_write_can_lock(l)	queue_write_can_lock(l)
+#define arch_read_lock(l)	queue_read_lock(l)
+#define arch_write_lock(l)	queue_write_lock(l)
+#define arch_read_trylock(l)	queue_read_trylock(l)
+#define arch_write_trylock(l)	queue_write_trylock(l)
+#define arch_read_unlock(l)	queue_read_unlock(l)
+#define arch_write_unlock(l)	queue_write_unlock(l)
+
+#endif /* __ASM_GENERIC_QRWLOCK_H */
diff --git a/kernel/Kconfig.locks b/kernel/Kconfig.locks
index d2b32ac..b665478 100644
--- a/kernel/Kconfig.locks
+++ b/kernel/Kconfig.locks
@@ -223,3 +223,10 @@ endif
 config MUTEX_SPIN_ON_OWNER
 	def_bool y
 	depends on SMP && !DEBUG_MUTEXES
+
+config ARCH_QUEUE_RWLOCK
+	bool
+
+config QUEUE_RWLOCK
+	def_bool y if ARCH_QUEUE_RWLOCK
+	depends on SMP
diff --git a/lib/Makefile b/lib/Makefile
index f3bb2cb..e3175db 100644
--- a/lib/Makefile
+++ b/lib/Makefile
@@ -189,3 +189,4 @@ quiet_cmd_build_OID_registry = GEN     $@
 clean-files	+= oid_registry_data.c
 
 obj-$(CONFIG_UCS2_STRING) += ucs2_string.o
+obj-$(CONFIG_QUEUE_RWLOCK) += qrwlock.o
diff --git a/lib/qrwlock.c b/lib/qrwlock.c
new file mode 100644
index 0000000..4915dc6
--- /dev/null
+++ b/lib/qrwlock.c
@@ -0,0 +1,274 @@
+/*
+ * Queue read/write lock
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * (C) Copyright 2013 Hewlett-Packard Development Company, L.P.
+ *
+ * Authors: Waiman Long <waiman.long@hp.com>
+ */
+#include <linux/smp.h>
+#include <linux/bug.h>
+#include <linux/cpumask.h>
+#include <linux/percpu.h>
+#include <linux/hardirq.h>
+#include <asm-generic/qrwlock.h>
+
+/*
+ * Compared with regular rwlock, the queue rwlock has has the following
+ * advantages:
+ * 1. It is more deterministic for the fair variant. Even though there is
+ *    a slight chance of stealing the lock if come at the right moment, the
+ *    granting of the lock is mostly in FIFO order. Even the default unfair
+ *    variant is fairer at least among the writers.
+ * 2. It is usually faster in high contention situation.
+ *
+ * The only downside is that the lock is 4 bytes larger in 32-bit systems
+ * and 12 bytes larger in 64-bit systems.
+ *
+ * There are two queues for writers. The writer field of the lock is a
+ * one-slot wait queue. The writers that follow will have to wait in the
+ * combined reader/writer queue (waitq).
+ *
+ * Compared with x86 ticket spinlock, the queue rwlock is faster in high
+ * contention situation. The writer lock is also faster in single thread
+ * operations. Therefore, queue rwlock can be considered as a replacement
+ * for those spinlocks that are highly contended as long as an increase
+ * in lock size is not an issue.
+ */
+
+#ifndef arch_mutex_cpu_relax
+# define arch_mutex_cpu_relax() cpu_relax()
+#endif
+
+#ifndef smp_mb__load_acquire
+# ifdef CONFIG_X86
+#   define smp_mb__load_acquire()	barrier()
+# else
+#   define smp_mb__load_acquire()	smp_mb()
+# endif
+#endif
+
+#ifndef smp_mb__store_release
+# ifdef CONFIG_X86
+#   define smp_mb__store_release()	barrier()
+# else
+#   define smp_mb__store_release()	smp_mb()
+# endif
+#endif
+
+/**
+ * wait_in_queue - Add to queue and wait until it is at the head
+ * @lock: Pointer to queue rwlock structure
+ * @node: Node pointer to be added to the queue
+ */
+static __always_inline void
+wait_in_queue(struct qrwlock *lock, struct qrwnode *node)
+{
+	struct qrwnode *prev;
+
+	node->next = NULL;
+	node->wait = true;
+	prev = xchg(&lock->waitq, node);
+	if (prev) {
+		prev->next = node;
+		/*
+		 * Wait until the waiting flag is off
+		 */
+		while (ACCESS_ONCE(node->wait))
+			arch_mutex_cpu_relax();
+		smp_mb__load_acquire();
+	}
+}
+
+/**
+ * signal_next - Signal the next one in queue to be at the head
+ * @lock: Pointer to queue rwlock structure
+ * @node: Node pointer to the current head of queue
+ */
+static __always_inline void
+signal_next(struct qrwlock *lock, struct qrwnode *node)
+{
+	struct qrwnode *next;
+
+	/*
+	 * Try to notify the next node first without disturbing the cacheline
+	 * of the lock. If that fails, check to see if it is the last node
+	 * and so should clear the wait queue.
+	 */
+	next = ACCESS_ONCE(node->next);
+	if (likely(next))
+		goto notify_next;
+
+	/*
+	 * Clear the wait queue if it is the last node
+	 */
+	if ((ACCESS_ONCE(lock->waitq) == node) &&
+	    (cmpxchg(&lock->waitq, node, NULL) == node))
+			return;
+	/*
+	 * Wait until the next one in queue set up the next field
+	 */
+	while (likely(!(next = ACCESS_ONCE(node->next))))
+		arch_mutex_cpu_relax();
+	/*
+	 * The next one in queue is now at the head
+	 */
+notify_next:
+	smp_mb__store_release();
+	ACCESS_ONCE(next->wait) = false;
+}
+
+/**
+ * rspin_until_writer_unlock - inc reader count & spin until writer is gone
+ * @lock: Pointer to queue rwlock structure
+ * @cnts: Current queue rwlock counts structure
+ *
+ * In interrupt context or at the head of the queue, the reader will just
+ * increment the reader count & wait until the writer releases the lock.
+ */
+static __always_inline void
+rspin_until_writer_unlock(struct qrwlock *lock, union qrwcnts cnts)
+{
+	while (cnts.writer == QW_LOCKED) {
+		arch_mutex_cpu_relax();
+		cnts.rw = ACCESS_ONCE(lock->cnts.rw);
+	}
+}
+
+/**
+ * queue_read_lock_slowpath - acquire read lock of a queue rwlock
+ * @lock: Pointer to queue rwlock structure
+ */
+void queue_read_lock_slowpath(struct qrwlock *lock)
+{
+	struct qrwnode node;
+	union qrwcnts cnts;
+
+	/*
+	 * Readers come here when it cannot get the lock without waiting
+	 */
+	if (unlikely(irq_count())) {
+		/*
+		 * Readers in interrupt context will spin until the lock is
+		 * available without waiting in the queue.
+		 */
+		cnts.rw = ACCESS_ONCE(lock->cnts.rw);
+		rspin_until_writer_unlock(lock, cnts);
+		return;
+	}
+	cnts.rw = xadd(&lock->cnts.rw, -QRW_READER_BIAS);
+
+	/*
+	 * Put the reader into the wait queue
+	 */
+	wait_in_queue(lock, &node);
+
+	/*
+	 * At the head of the wait queue now, try to increment the reader
+	 * count and get the lock.
+	 */
+	if (unlikely(cnts.fair)) {
+		/*
+		 * For fair reader, wait until the writer state goes to 0
+		 * before incrementing the reader count.
+		 */
+		while (ACCESS_ONCE(lock->cnts.writer))
+			arch_mutex_cpu_relax();
+	}
+	cnts.rw = xadd(&lock->cnts.rw, QRW_READER_BIAS);
+	rspin_until_writer_unlock(lock, cnts);
+	/*
+	 * Need to have a barrier with read-acquire semantics
+	 */
+	smp_mb__load_acquire();
+	signal_next(lock, &node);
+}
+EXPORT_SYMBOL(queue_read_lock_slowpath);
+
+/**
+ * qwrite_trylock - Try to acquire the write lock
+ * @lock : Pointer to queue rwlock structure
+ * @old  : The current queue rwlock count structure
+ * Return: 1 if lock acquired, 0 otherwise
+ */
+static __always_inline int
+qwrite_trylock(struct qrwlock *lock, union qrwcnts old)
+{
+	register union qrwcnts new;
+
+	new.rw     = old.rw;
+	new.writer = QW_LOCKED;
+	if (likely(cmpxchg(&lock->cnts.rw, old.rw, new.rw) == old.rw))
+		return 1;
+	return 0;
+}
+
+/**
+ * queue_write_3step_lock - acquire write lock in 3 steps
+ * @lock : Pointer to queue rwlock structure
+ * Return: 1 if lock acquired, 0 otherwise
+ *
+ * Step 1 - Try to acquire the lock directly if no reader is present
+ * Step 2 - Set the waiting flag to notify readers that a writer is waiting
+ * Step 3 - When the readers field goes to 0, set the locked flag
+ *
+ * When not in fair mode, the readers actually ignore the second step.
+ * However, this is still necessary to force other writers to fall in line.
+ * In x86, the use of noinline generates a slight better optimized code
+ * with less memory access.
+ */
+static noinline int queue_write_3step_lock(struct qrwlock *lock)
+{
+	register union qrwcnts old;
+
+	old.rw = ACCESS_ONCE(lock->cnts.rw);
+
+	/* Step 1 */
+	if (!old.writer && !old.readers && qwrite_trylock(lock, old))
+		return 1;
+
+	/* Step 2 */
+	if (old.writer || (cmpxchg(&lock->cnts.writer, 0, QW_WAITING) != 0))
+		return 0;
+
+	/* Step 3 */
+	arch_mutex_cpu_relax();
+	old.rw = ACCESS_ONCE(lock->cnts.rw);
+	while (old.readers || !qwrite_trylock(lock, old)) {
+		arch_mutex_cpu_relax();
+		old.rw = ACCESS_ONCE(lock->cnts.rw);
+	}
+	return 1;
+}
+
+/**
+ * queue_write_lock_slowpath - acquire write lock of a queue rwlock
+ * @lock : Pointer to queue rwlock structure
+ */
+void queue_write_lock_slowpath(struct qrwlock *lock)
+{
+	struct qrwnode node;
+
+	/*
+	 * Put the writer into the wait queue
+	 */
+	wait_in_queue(lock, &node);
+
+	/*
+	 * At the head of the wait queue now, call queue_write_3step_lock()
+	 * to acquire the lock until it is done.
+	 */
+	while (!queue_write_3step_lock(lock))
+		arch_mutex_cpu_relax();
+	signal_next(lock, &node);
+}
+EXPORT_SYMBOL(queue_write_lock_slowpath);
-- 
1.7.1

[PATCH v6 1/3] qrwlock: A queue read/write lock implementation

[Waiman Long]

This patch introduces a new read/write lock implementation that put
waiting readers and writers into a queue instead of actively contending
the lock like the current read/write lock implementation. This will
improve performance in highly contended situation by reducing the
cache line bouncing effect.

The queue read/write lock (qrwlock) is mostly fair with respect to
the writers, even though there is still a slight chance of write
lock stealing.

Externally, there are two different types of readers - unfair (the
default) and fair. A unfair reader will try to steal read lock even
if a writer is waiting, whereas a fair reader will be waiting in
the queue under this circumstance.  These variants are chosen at
initialization time by using different initializers. The new *_fair()
initializers are added for selecting the use of fair reader.

Internally, there is a third type of readers which steal lock more
aggressively than the unfair reader. They simply increments the reader
count and wait until the writer releases the lock. The transition to
aggressive reader happens in the read lock slowpath when
1. In an interrupt context.
2. when a classic reader comes to the head of the wait queue.
3. When a fair reader comes to the head of the wait queue and sees
   the release of a write lock.

The fair queue rwlock is more deterministic in the sense that late
comers jumping ahead and stealing the lock is unlikely even though
there is still a very small chance for lock stealing to happen if
the readers or writers come at the right moment.  Other than that,
lock granting is done in a FIFO manner.  As a result, it is possible
to determine a maximum time period after which the waiting is over
and the lock can be acquired.

The queue read lock is safe to use in an interrupt context (softirq
or hardirq) as it will switch to become an aggressive reader in such
environment allowing recursive read lock. However, the fair readers
will not support recursive read lock in a non-interrupt environment
when a writer is waiting.

The only downside of queue rwlock is the size increase in the lock
structure by 4 bytes for 32-bit systems and by 12 bytes for 64-bit
systems.

This patch will replace the architecture specific implementation
of rwlock by this generic version of queue rwlock when the
ARCH_QUEUE_RWLOCK configuration parameter is set.

In term of single-thread performance (no contention), a 256K
lock/unlock loop was run on a 2.4GHz and 2.93Ghz Westmere x86-64
CPUs. The following table shows the average time (in ns) for a single
lock/unlock sequence (including the looping and timing overhead):

Lock Type		    2.4GHz	2.93GHz
---------		    ------	-------
Ticket spinlock		     14.9	 12.3
Read lock		     17.0	 13.5
Write lock		     17.0	 13.5
Queue read lock	     	     16.0	 13.5
Queue fair read lock	     16.0	 13.5
Queue write lock	      9.2	  7.8
Queue fair write lock	     17.5	 14.5

The queue read lock is slightly slower than the spinlock, but is
slightly faster than the read lock. The queue write lock, however,
is the fastest of all. It is almost twice as fast as the write lock
and about 1.5X of the spinlock. The queue fair write lock, on the
other hand, is slightly slower than the write lock.

With lock contention, the speed of each individual lock/unlock function
is less important than the amount of contention-induced delays.

To investigate the performance characteristics of the queue rwlock
compared with the regular rwlock, Ingo's anon_vmas patch that converts
rwsem to rwlock was applied to a 3.12 kernel. This kernel was then
tested under the following 4 conditions:

 1) Plain 3.12
 2) Ingo's patch
 3) Ingo's patch + unfair qrwlock (default)
 4) Ingo's patch + fair qrwlock

Each of the 4 kernels were booted up twice with and without the
"idle=poll" kernel parameter which keeps the CPUs in C0 state while
idling instead of a more energy-saving sleep state.  The jobs per
minutes (JPM) results of the AIM7's high_systime workload at 1500
users on a 8-socket 80-core DL980 (HT off) were:

 Kernel	    JPMs	%Change from (1)
 ------	    ----	----------------
   1	145704/227295		-
   2	229750/236066	    +58%/+3.8%
   3	248144/253521	    +70%/+11.5%
   4	240062/246251	    +65%/+8.3%

The first JPM number is without the "idle=poll" kernel parameter,
the second number is with that parameter. It can be seen that most
of the performance benefit of converting rwsem to rwlock actually
come from the latency improvement of not needing to wake up a CPU
from deep sleep state when work is available.

The use of non-sleeping locks did improve performance by eliminating
the context switching cost. Using queue rwlock gave almost tripling
of performance gain. The performance gain was reduced somewhat with
a fair lock which was to be expected.

Looking at the perf profiles (with idle=poll) below, we can clearly see
that other bottlenecks were constraining the performance improvement.

Perf profile of kernel (2):

 18.65%    reaim  [kernel.kallsyms]  [k] __write_lock_failed
  9.00%    reaim  [kernel.kallsyms]  [k] _raw_spin_lock_irqsave
  5.21%  swapper  [kernel.kallsyms]  [k] cpu_idle_loop
  3.08%    reaim  [kernel.kallsyms]  [k] mspin_lock
  2.50%    reaim  [kernel.kallsyms]  [k] anon_vma_interval_tree_insert
  2.00%       ls  [kernel.kallsyms]  [k] _raw_spin_lock_irqsave
  1.29%    reaim  [kernel.kallsyms]  [k] update_cfs_rq_blocked_load
  1.21%    reaim  [kernel.kallsyms]  [k] __read_lock_failed
  1.12%    reaim  [kernel.kallsyms]  [k] _raw_spin_lock
  1.10%    reaim  [kernel.kallsyms]  [k] perf_event_aux
  1.09%     true  [kernel.kallsyms]  [k] _raw_spin_lock_irqsave

Perf profile of kernel (3):

 20.99%  swapper  [kernel.kallsyms]  [k] cpu_idle_loop
  7.93%    reaim  [kernel.kallsyms]  [k] _raw_spin_lock_irqsave
  5.67%    reaim  [kernel.kallsyms]  [k] mspin_lock
  4.34%    reaim  [kernel.kallsyms]  [k] queue_write_lock_slowpath
  2.13%    reaim  [kernel.kallsyms]  [k] update_cfs_rq_blocked_load
  2.11%       ls  [kernel.kallsyms]  [k] _raw_spin_lock_irqsave
  2.07%    reaim  [kernel.kallsyms]  [k] anon_vma_interval_tree_insert
  1.17%    reaim  [kernel.kallsyms]  [k] queue_write_3step_lock
  1.10%     true  [kernel.kallsyms]  [k] _raw_spin_lock_irqsave
  0.95%    reaim  [kernel.kallsyms]  [k] mutex_spin_on_owner
  0.74%    reaim  [kernel.kallsyms]  [k] queue_read_lock_slowpath

The spinlock bottlenecks were shown below.

  7.93%    reaim  [kernel.kallsyms]  [k] _raw_spin_lock_irqsave
              |--60.57%-- release_pages
              |--36.62%-- pagevec_lru_move_fn
              |--0.80%-- get_page_from_freelist
              |--0.69%-- __page_cache_release
               --1.32%-- [...]

For both release_pages() & pagevec_lru_move_fn() function, the
spinlock contention was on zone->lru_lock. With the queue spinlock
patch, however, the contention went away with a lot more idle time
available and the JPM number went up to 265532 which was an additional
performance improvement.

 28.40%  swapper  [kernel.kallsyms]  [k] cpu_idle_loop
  6.89%    reaim  [kernel.kallsyms]  [k] mspin_lock
  4.17%    reaim  [kernel.kallsyms]  [k] queue_write_lock_slowpath
  2.10%    reaim  [kernel.kallsyms]  [k] anon_vma_interval_tree_insert
  1.82%    reaim  [kernel.kallsyms]  [k] update_cfs_rq_blocked_load
  1.34%    reaim  [kernel.kallsyms]  [k] entity_tick
  1.17%    reaim  [kernel.kallsyms]  [k] queue_write_3step_lock
  1.06%    reaim  [kernel.kallsyms]  [k] mutex_spin_on_owner
  0.86%    reaim  [kernel.kallsyms]  [k] perf_event_aux
  0.83%       ls  [kernel.kallsyms]  [k] mspin_lock
   :
  0.53%    reaim  [kernel.kallsyms]  [k] _raw_spin_lock
  0.14%    reaim  [kernel.kallsyms]  [k] _raw_spin_lock_irqsave

Tim Chen also tested the qrwlock with Ingo's patch on a 4-socket
machine.  It was found the performance improvement of 11% was the
same with regular rwlock or queue rwlock.

Signed-off-by: Waiman Long <Waiman.Long@hp.com>
---
 include/asm-generic/qrwlock.h |  250 +++++++++++++++++++++++++++++++++++++
 kernel/Kconfig.locks          |    7 +
 lib/Makefile                  |    1 +
 lib/qrwlock.c                 |  273 +++++++++++++++++++++++++++++++++++++++++
 4 files changed, 531 insertions(+), 0 deletions(-)
 create mode 100644 include/asm-generic/qrwlock.h
 create mode 100644 lib/qrwlock.c

diff --git a/include/asm-generic/qrwlock.h b/include/asm-generic/qrwlock.h
new file mode 100644
index 0000000..1f68499
--- /dev/null
+++ b/include/asm-generic/qrwlock.h
@@ -0,0 +1,250 @@
+/*
+ * Queue read/write lock
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * (C) Copyright 2013 Hewlett-Packard Development Company, L.P.
+ *
+ * Authors: Waiman Long <waiman.long@hp.com>
+ */
+#ifndef __ASM_GENERIC_QRWLOCK_H
+#define __ASM_GENERIC_QRWLOCK_H
+
+#include <linux/types.h>
+#include <asm/bitops.h>
+#include <asm/cmpxchg.h>
+#include <asm/barrier.h>
+#include <asm/processor.h>
+#include <asm/byteorder.h>
+
+#if !defined(__LITTLE_ENDIAN) && !defined(__BIG_ENDIAN)
+#error "Missing either LITTLE_ENDIAN or BIG_ENDIAN definition."
+#endif
+
+#if (CONFIG_NR_CPUS < 65536)
+typedef u16 __nrcpu_t;
+typedef u32 __nrcpupair_t;
+#define	QRW_READER_BIAS	(1U << 16)
+#else
+typedef u32 __nrcpu_t;
+typedef u64 __nrcpupair_t;
+#define	QRW_READER_BIAS	(1UL << 32)
+#endif
+
+/*
+ * The queue read/write lock data structure
+ *
+ * Read lock stealing can only happen when there is at least one reader
+ * holding the read lock. When the fair flag is not set, it mimics the
+ * behavior of the regular rwlock at the expense that a perpetual stream
+ * of readers could starve a writer for a long period of time. That
+ * behavior, however, may be beneficial to a workload that is reader heavy
+ * with slow writers, and the writers can wait without undesirable consequence.
+ * This fair flag should only be set at initialization time.
+ *
+ * The layout of the structure is endian-sensitive to make sure that adding
+ * QRW_READER_BIAS to the rw field to increment the reader count won't
+ * disturb the writer and the fair fields.
+ */
+struct qrwnode {
+	struct qrwnode *next;
+	bool		wait;	/* Waiting flag */
+};
+
+typedef struct qrwlock {
+	union qrwcnts {
+		struct {
+#ifdef __LITTLE_ENDIAN
+			u8	  writer;	/* Writer state		*/
+			u8	  fair;		/* Fair rwlock flag	*/
+			__nrcpu_t readers;	/* # of active readers	*/
+#else
+			__nrcpu_t readers;	/* # of active readers	*/
+			u8	  fair;		/* Fair rwlock flag	*/
+			u8	  writer;	/* Writer state		*/
+#endif
+		};
+		__nrcpupair_t rw;		/* Reader/writer number pair */
+	} cnts;
+	struct qrwnode *waitq;			/* Tail of waiting queue */
+} arch_rwlock_t;
+
+/*
+ * Writer state values & mask
+ */
+#define	QW_WAITING	1			/* A writer is waiting	   */
+#define	QW_LOCKED	0xff			/* A writer holds the lock */
+#define QW_MASK_FAIR	((u8)~0)		/* Mask for fair reader    */
+#define QW_MASK_UNFAIR	((u8)~QW_WAITING)	/* Mask for unfair reader  */
+#define	QW_GET_WMASK(c) ((c).fair ? QW_MASK_FAIR : QW_MASK_UNFAIR)
+
+/*
+ * External function declarations
+ */
+extern void queue_read_lock_slowpath(struct qrwlock *lock);
+extern void queue_write_lock_slowpath(struct qrwlock *lock);
+
+/**
+ * queue_read_can_lock- would read_trylock() succeed?
+ * @lock: Pointer to queue rwlock structure
+ */
+static inline int queue_read_can_lock(struct qrwlock *lock)
+{
+	union qrwcnts cnts;
+	u8 wmask;
+
+	cnts.rw = ACCESS_ONCE(lock->cnts.rw);
+	wmask   = QW_GET_WMASK(cnts);
+	return !(cnts.writer & wmask);
+}
+
+/**
+ * queue_write_can_lock- would write_trylock() succeed?
+ * @lock: Pointer to queue rwlock structure
+ */
+static inline int queue_write_can_lock(struct qrwlock *lock)
+{
+	union qrwcnts cnts;
+
+	cnts.rw = ACCESS_ONCE(lock->cnts.rw);
+	return !cnts.writer && !cnts.readers;
+}
+
+/**
+ * queue_read_trylock - try to acquire read lock of a queue rwlock
+ * @lock : Pointer to queue rwlock structure
+ * Return: 1 if lock acquired, 0 if failed
+ */
+static inline int queue_read_trylock(struct qrwlock *lock)
+{
+	union qrwcnts cnts;
+	u8 wmask;
+
+	cnts.rw = ACCESS_ONCE(lock->cnts.rw);
+	wmask   = QW_GET_WMASK(cnts);
+	if (likely(!(cnts.writer & wmask))) {
+		cnts.rw = xadd(&lock->cnts.rw, QRW_READER_BIAS);
+		if (likely(!(cnts.writer & wmask)))
+			return 1;
+		add_smp(&lock->cnts.readers, -1);
+	}
+	return 0;
+}
+
+/**
+ * queue_write_trylock - try to acquire write lock of a queue rwlock
+ * @lock : Pointer to queue rwlock structure
+ * Return: 1 if lock acquired, 0 if failed
+ */
+static inline int queue_write_trylock(struct qrwlock *lock)
+{
+	union qrwcnts old, new;
+
+	old.rw = ACCESS_ONCE(lock->cnts.rw);
+	if (likely(!old.writer && !old.readers)) {
+		new.rw = old.rw;
+		new.writer = QW_LOCKED;
+		if (likely(cmpxchg(&lock->cnts.rw, old.rw, new.rw) == old.rw))
+			return 1;
+	}
+	return 0;
+}
+/**
+ * queue_read_lock - acquire read lock of a queue rwlock
+ * @lock: Pointer to queue rwlock structure
+ */
+static inline void queue_read_lock(struct qrwlock *lock)
+{
+	union qrwcnts cnts;
+	u8 wmask;
+
+	cnts.rw = xadd(&lock->cnts.rw, QRW_READER_BIAS);
+	wmask   = QW_GET_WMASK(cnts);
+	if (likely(!(cnts.writer & wmask)))
+		return;
+	/*
+	 * Slowpath will decrement the reader count, if necessary
+	 */
+	queue_read_lock_slowpath(lock);
+}
+
+/**
+ * queue_write_lock - acquire write lock of a queue rwlock
+ * @lock : Pointer to queue rwlock structure
+ */
+static inline void queue_write_lock(struct qrwlock *lock)
+{
+	union qrwcnts old;
+
+	/*
+	 * Optimize for the unfair lock case where the fair flag is 0.
+	 */
+	old.rw = cmpxchg(&lock->cnts.rw, 0, QW_LOCKED);
+	if (likely(old.rw == 0))
+		return;
+	if (likely(!old.writer && !old.readers)) {
+		union qrwcnts new;
+
+		new.rw = old.rw;
+		new.writer = QW_LOCKED;
+		if (likely(cmpxchg(&lock->cnts.rw, old.rw, new.rw) == old.rw))
+			return;
+	}
+	queue_write_lock_slowpath(lock);
+}
+
+/**
+ * queue_read_unlock - release read lock of a queue rwlock
+ * @lock : Pointer to queue rwlock structure
+ */
+static inline void queue_read_unlock(struct qrwlock *lock)
+{
+	/*
+	 * Atomically decrement the reader count
+	 */
+	add_smp(&lock->cnts.readers, -1);
+}
+
+/**
+ * queue_write_unlock - release write lock of a queue rwlock
+ * @lock : Pointer to queue rwlock structure
+ */
+static inline void queue_write_unlock(struct qrwlock *lock)
+{
+	/*
+	 * Make sure that none of the critical section will be leaked out.
+	 */
+	smp_mb__before_clear_bit();
+	ACCESS_ONCE(lock->cnts.writer) = 0;
+	smp_mb__after_clear_bit();
+}
+
+/*
+ * Initializier
+ */
+#define	__ARCH_RW_LOCK_UNLOCKED	{ .cnts = { .rw = 0 }, .waitq = NULL }
+#define	__ARCH_RW_LOCK_UNLOCKED_FAIR	\
+	{ .cnts = { { .writer = 0, .fair = 1, .readers = 0 } }, .waitq = NULL }
+
+/*
+ * Remapping rwlock architecture specific functions to the corresponding
+ * queue rwlock functions.
+ */
+#define arch_read_can_lock(l)	queue_read_can_lock(l)
+#define arch_write_can_lock(l)	queue_write_can_lock(l)
+#define arch_read_lock(l)	queue_read_lock(l)
+#define arch_write_lock(l)	queue_write_lock(l)
+#define arch_read_trylock(l)	queue_read_trylock(l)
+#define arch_write_trylock(l)	queue_write_trylock(l)
+#define arch_read_unlock(l)	queue_read_unlock(l)
+#define arch_write_unlock(l)	queue_write_unlock(l)
+
+#endif /* __ASM_GENERIC_QRWLOCK_H */
diff --git a/kernel/Kconfig.locks b/kernel/Kconfig.locks
index d2b32ac..b665478 100644
--- a/kernel/Kconfig.locks
+++ b/kernel/Kconfig.locks
@@ -223,3 +223,10 @@ endif
 config MUTEX_SPIN_ON_OWNER
 	def_bool y
 	depends on SMP && !DEBUG_MUTEXES
+
+config ARCH_QUEUE_RWLOCK
+	bool
+
+config QUEUE_RWLOCK
+	def_bool y if ARCH_QUEUE_RWLOCK
+	depends on SMP
diff --git a/lib/Makefile b/lib/Makefile
index f3bb2cb..e3175db 100644
--- a/lib/Makefile
+++ b/lib/Makefile
@@ -189,3 +189,4 @@ quiet_cmd_build_OID_registry = GEN     $@
 clean-files	+= oid_registry_data.c
 
 obj-$(CONFIG_UCS2_STRING) += ucs2_string.o
+obj-$(CONFIG_QUEUE_RWLOCK) += qrwlock.o
diff --git a/lib/qrwlock.c b/lib/qrwlock.c
new file mode 100644
index 0000000..f265c02
--- /dev/null
+++ b/lib/qrwlock.c
@@ -0,0 +1,273 @@
+/*
+ * Queue read/write lock
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * (C) Copyright 2013 Hewlett-Packard Development Company, L.P.
+ *
+ * Authors: Waiman Long <waiman.long@hp.com>
+ */
+#include <linux/smp.h>
+#include <linux/bug.h>
+#include <linux/cpumask.h>
+#include <linux/percpu.h>
+#include <linux/hardirq.h>
+#include <asm-generic/qrwlock.h>
+
+/*
+ * Compared with regular rwlock, the queue rwlock has has the following
+ * advantages:
+ * 1. It is more deterministic for the fair variant. Even though there is
+ *    a slight chance of stealing the lock if come at the right moment, the
+ *    granting of the lock is mostly in FIFO order. Even the default unfair
+ *    variant is fairer at least among the writers.
+ * 2. It is usually faster in high contention situation.
+ *
+ * The only downside is that the lock is 4 bytes larger in 32-bit systems
+ * and 12 bytes larger in 64-bit systems.
+ *
+ * There are two queues for writers. The writer field of the lock is a
+ * one-slot wait queue. The writers that follow will have to wait in the
+ * combined reader/writer queue (waitq).
+ *
+ * Compared with x86 ticket spinlock, the queue rwlock is faster in high
+ * contention situation. The writer lock is also faster in single thread
+ * operations. Therefore, queue rwlock can be considered as a replacement
+ * for those spinlocks that are highly contended as long as an increase
+ * in lock size is not an issue.
+ */
+
+#ifndef arch_mutex_cpu_relax
+# define arch_mutex_cpu_relax() cpu_relax()
+#endif
+
+#ifndef smp_mb__load_acquire
+# ifdef CONFIG_X86
+#   define smp_mb__load_acquire()	barrier()
+# else
+#   define smp_mb__load_acquire()	smp_mb()
+# endif
+#endif
+
+#ifndef smp_mb__store_release
+# ifdef CONFIG_X86
+#   define smp_mb__store_release()	barrier()
+# else
+#   define smp_mb__store_release()	smp_mb()
+# endif
+#endif
+
+/**
+ * wait_in_queue - Add to queue and wait until it is at the head
+ * @lock: Pointer to queue rwlock structure
+ * @node: Node pointer to be added to the queue
+ */
+static __always_inline void
+wait_in_queue(struct qrwlock *lock, struct qrwnode *node)
+{
+	struct qrwnode *prev;
+
+	node->next = NULL;
+	node->wait = true;
+	prev = xchg(&lock->waitq, node);
+	if (prev) {
+		prev->next = node;
+		/*
+		 * Wait until the waiting flag is off
+		 */
+		while (ACCESS_ONCE(node->wait))
+			arch_mutex_cpu_relax();
+		smp_mb__load_acquire();
+	}
+}
+
+/**
+ * signal_next - Signal the next one in queue to be at the head
+ * @lock: Pointer to queue rwlock structure
+ * @node: Node pointer to the current head of queue
+ */
+static __always_inline void
+signal_next(struct qrwlock *lock, struct qrwnode *node)
+{
+	struct qrwnode *next;
+
+	/*
+	 * Try to notify the next node first without disturbing the cacheline
+	 * of the lock. If that fails, check to see if it is the last node
+	 * and so should clear the wait queue.
+	 */
+	next = ACCESS_ONCE(node->next);
+	if (likely(next))
+		goto notify_next;
+
+	/*
+	 * Clear the wait queue if it is the last node
+	 */
+	if ((ACCESS_ONCE(lock->waitq) == node) &&
+	    (cmpxchg(&lock->waitq, node, NULL) == node))
+			return;
+	/*
+	 * Wait until the next one in queue set up the next field
+	 */
+	while (likely(!(next = ACCESS_ONCE(node->next))))
+		arch_mutex_cpu_relax();
+	/*
+	 * The next one in queue is now at the head
+	 */
+notify_next:
+	smp_mb__store_release();
+	ACCESS_ONCE(next->wait) = false;
+}
+
+/**
+ * rspin_until_writer_unlock - inc reader count & spin until writer is gone
+ * @lock: Pointer to queue rwlock structure
+ * @cnts: Current queue rwlock counts structure
+ *
+ * In interrupt context or at the head of the queue, the reader will just
+ * increment the reader count & wait until the writer releases the lock.
+ */
+static __always_inline void
+rspin_until_writer_unlock(struct qrwlock *lock, union qrwcnts cnts)
+{
+	while (cnts.writer == QW_LOCKED) {
+		arch_mutex_cpu_relax();
+		cnts.rw = ACCESS_ONCE(lock->cnts.rw);
+	}
+}
+
+/**
+ * queue_read_lock_slowpath - acquire read lock of a queue rwlock
+ * @lock: Pointer to queue rwlock structure
+ */
+void queue_read_lock_slowpath(struct qrwlock *lock)
+{
+	struct qrwnode node;
+	union qrwcnts cnts;
+
+	/*
+	 * Readers come here when it cannot get the lock without waiting
+	 */
+	if (unlikely(irq_count())) {
+		/*
+		 * Readers in interrupt context will spin until the lock is
+		 * available without waiting in the queue.
+		 */
+		cnts.rw = ACCESS_ONCE(lock->cnts.rw);
+		rspin_until_writer_unlock(lock, cnts);
+		return;
+	}
+	cnts.rw = xadd(&lock->cnts.rw, -QRW_READER_BIAS);
+
+	/*
+	 * Put the reader into the wait queue
+	 */
+	wait_in_queue(lock, &node);
+
+	/*
+	 * At the head of the wait queue now, try to increment the reader
+	 * count and get the lock.
+	 */
+	if (unlikely(cnts.fair)) {
+		/*
+		 * For fair reader, wait until the writer state goes to 0
+		 * before incrementing the reader count.
+		 */
+		while (ACCESS_ONCE(lock->cnts.writer))
+			arch_mutex_cpu_relax();
+	}
+	cnts.rw = xadd(&lock->cnts.rw, QRW_READER_BIAS);
+	rspin_until_writer_unlock(lock, cnts);
+	/*
+	 * Need to have a barrier with read-acquire semantics
+	 */
+	smp_mb__load_acquire();
+	signal_next(lock, &node);
+}
+EXPORT_SYMBOL(queue_read_lock_slowpath);
+
+/**
+ * qwrite_trylock - Try to acquire the write lock
+ * @lock : Pointer to queue rwlock structure
+ * @old  : The current queue rwlock count structure
+ * Return: 1 if lock acquired, 0 otherwise
+ */
+static __always_inline int
+qwrite_trylock(struct qrwlock *lock, union qrwcnts old)
+{
+	register union qrwcnts new;
+
+	new.rw     = old.rw;
+	new.writer = QW_LOCKED;
+	if (likely(cmpxchg(&lock->cnts.rw, old.rw, new.rw) == old.rw))
+		return 1;
+	return 0;
+}
+
+/**
+ * queue_write_3step_lock - acquire write lock in 3 steps
+ * @lock : Pointer to queue rwlock structure
+ * Return: 1 if lock acquired, 0 otherwise
+ *
+ * Step 1 - Try to acquire the lock directly if no reader is present
+ * Step 2 - Set the waiting flag to notify readers that a writer is waiting
+ * Step 3 - When the readers field goes to 0, set the locked flag
+ *
+ * When not in fair mode, the readers actually ignore the second step.
+ * However, this is still necessary to force other writers to fall in line.
+ * In x86, the use of noinline generates a slight better optimized code
+ * with less memory access.
+ */
+static noinline int queue_write_3step_lock(struct qrwlock *lock)
+{
+	register union qrwcnts old;
+
+	old.rw = ACCESS_ONCE(lock->cnts.rw);
+
+	/* Step 1 */
+	if (!old.writer && !old.readers && qwrite_trylock(lock, old))
+		return 1;
+
+	/* Step 2 */
+	if (old.writer || (cmpxchg(&lock->cnts.writer, 0, QW_WAITING) != 0))
+		return 0;
+
+	/* Step 3 */
+	old.rw = ACCESS_ONCE(lock->cnts.rw);
+	while (!qwrite_trylock(lock, old)) {
+		arch_mutex_cpu_relax();
+		old.rw = ACCESS_ONCE(lock->cnts.rw);
+	}
+	return 1;
+}
+
+/**
+ * queue_write_lock_slowpath - acquire write lock of a queue rwlock
+ * @lock : Pointer to queue rwlock structure
+ */
+void queue_write_lock_slowpath(struct qrwlock *lock)
+{
+	struct qrwnode node;
+
+	/*
+	 * Put the writer into the wait queue
+	 */
+	wait_in_queue(lock, &node);
+
+	/*
+	 * At the head of the wait queue now, call queue_write_3step_lock()
+	 * to acquire the lock until it is done.
+	 */
+	while (!queue_write_3step_lock(lock))
+		arch_mutex_cpu_relax();
+	signal_next(lock, &node);
+}
+EXPORT_SYMBOL(queue_write_lock_slowpath);
-- 
1.7.1

[PATCH v6 2/5] qrwlock x86: Enable x86 to use queue read/write lock

[Waiman Long]

This patch makes the necessary changes at the x86 architecture specific
layer to enable the presence of the CONFIG_QUEUE_RWLOCK kernel option
to replace the read/write lock by the queue read/write lock.

It also enables CONFIG_ARCH_QUEUE_RWLOCK which will force the use
of queue read/write lock for x86 which tends to have the largest
NUMA machines compared with the other architectures. This patch will
improve the scalability of those large machines.

Signed-off-by: Waiman Long <Waiman.Long@hp.com>
---
 arch/x86/Kconfig                      |    1 +
 arch/x86/include/asm/spinlock.h       |    2 ++
 arch/x86/include/asm/spinlock_types.h |    4 ++++
 3 files changed, 7 insertions(+), 0 deletions(-)

diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
index f67e839..4f447bb 100644
--- a/arch/x86/Kconfig
+++ b/arch/x86/Kconfig
@@ -123,6 +123,7 @@ config X86
 	select COMPAT_OLD_SIGACTION if IA32_EMULATION
 	select RTC_LIB
 	select HAVE_DEBUG_STACKOVERFLOW
+	select QUEUE_RWLOCK
 
 config INSTRUCTION_DECODER
 	def_bool y
diff --git a/arch/x86/include/asm/spinlock.h b/arch/x86/include/asm/spinlock.h
index bf156de..8fb88c5 100644
--- a/arch/x86/include/asm/spinlock.h
+++ b/arch/x86/include/asm/spinlock.h
@@ -188,6 +188,7 @@ static inline void arch_spin_unlock_wait(arch_spinlock_t *lock)
 		cpu_relax();
 }
 
+#ifndef CONFIG_QUEUE_RWLOCK
 /*
  * Read-write spinlocks, allowing multiple readers
  * but only one writer.
@@ -270,6 +271,7 @@ static inline void arch_write_unlock(arch_rwlock_t *rw)
 	asm volatile(LOCK_PREFIX WRITE_LOCK_ADD(%1) "%0"
 		     : "+m" (rw->write) : "i" (RW_LOCK_BIAS) : "memory");
 }
+#endif /* CONFIG_QUEUE_RWLOCK */
 
 #define arch_read_lock_flags(lock, flags) arch_read_lock(lock)
 #define arch_write_lock_flags(lock, flags) arch_write_lock(lock)
diff --git a/arch/x86/include/asm/spinlock_types.h b/arch/x86/include/asm/spinlock_types.h
index 4f1bea1..a585635 100644
--- a/arch/x86/include/asm/spinlock_types.h
+++ b/arch/x86/include/asm/spinlock_types.h
@@ -34,6 +34,10 @@ typedef struct arch_spinlock {
 
 #define __ARCH_SPIN_LOCK_UNLOCKED	{ { 0 } }
 
+#ifdef CONFIG_QUEUE_RWLOCK
+#include <asm-generic/qrwlock.h>
+#else
 #include <asm/rwlock.h>
+#endif
 
 #endif /* _ASM_X86_SPINLOCK_TYPES_H */
-- 
1.7.1

[PATCH v6 3/5] qrwlock: Enable fair queue read/write lock

[Waiman Long]

By default, queue rwlock is fair among writers and gives preference
to readers allowing them to steal lock even if a writer is
waiting. However, there is a desire to have a fair variant of
rwlock that is more deterministic. To enable this, fair variants
of lock initializers are added by this patch to allow lock owners
to choose to use fair rwlock. These newly added initializers all
have the _fair or _FAIR suffix to indicate the desire to use a fair
rwlock. If the QUEUE_RWLOCK config option is not selected, the fair
rwlock initializers will be the same as the regular ones.

Signed-off-by: Waiman Long <Waiman.Long@hp.com>
---
 include/linux/rwlock.h       |   15 +++++++++++++++
 include/linux/rwlock_types.h |   13 +++++++++++++
 lib/spinlock_debug.c         |   19 +++++++++++++++++++
 3 files changed, 47 insertions(+), 0 deletions(-)

diff --git a/include/linux/rwlock.h b/include/linux/rwlock.h
index bc2994e..5f2628b 100644
--- a/include/linux/rwlock.h
+++ b/include/linux/rwlock.h
@@ -23,9 +23,24 @@ do {								\
 								\
 	__rwlock_init((lock), #lock, &__key);			\
 } while (0)
+
+# ifdef CONFIG_QUEUE_RWLOCK
+extern void __rwlock_init_fair(rwlock_t *lock, const char *name,
+			       struct lock_class_key *key);
+#  define rwlock_init_fair(lock)				\
+do {								\
+	static struct lock_class_key __key;			\
+								\
+	__rwlock_init_fair((lock), #lock, &__key);		\
+} while (0)
+# else
+#  define __rwlock_init_fair(l, n, k)	__rwlock_init(l, n, k)
+# endif /* CONFIG_QUEUE_RWLOCK */
 #else
 # define rwlock_init(lock)					\
 	do { *(lock) = __RW_LOCK_UNLOCKED(lock); } while (0)
+# define rwlock_init_fair(lock)				\
+	do { *(lock) = __RW_LOCK_UNLOCKED_FAIR(lock); } while (0)
 #endif
 
 #ifdef CONFIG_DEBUG_SPINLOCK
diff --git a/include/linux/rwlock_types.h b/include/linux/rwlock_types.h
index cc0072e..d27c812 100644
--- a/include/linux/rwlock_types.h
+++ b/include/linux/rwlock_types.h
@@ -37,12 +37,25 @@ typedef struct {
 				.owner = SPINLOCK_OWNER_INIT,		\
 				.owner_cpu = -1,			\
 				RW_DEP_MAP_INIT(lockname) }
+#define __RW_LOCK_UNLOCKED_FAIR(lockname)				\
+	(rwlock_t)	{	.raw_lock = __ARCH_RW_LOCK_UNLOCKED_FAIR,\
+				.magic = RWLOCK_MAGIC,			\
+				.owner = SPINLOCK_OWNER_INIT,		\
+				.owner_cpu = -1,			\
+				RW_DEP_MAP_INIT(lockname) }
 #else
 #define __RW_LOCK_UNLOCKED(lockname) \
 	(rwlock_t)	{	.raw_lock = __ARCH_RW_LOCK_UNLOCKED,	\
 				RW_DEP_MAP_INIT(lockname) }
+#define __RW_LOCK_UNLOCKED_FAIR(lockname) \
+	(rwlock_t)	{	.raw_lock = __ARCH_RW_LOCK_UNLOCKED_FAIR,\
+				RW_DEP_MAP_INIT(lockname) }
 #endif
 
 #define DEFINE_RWLOCK(x)	rwlock_t x = __RW_LOCK_UNLOCKED(x)
+#define DEFINE_RWLOCK_FAIR(x)	rwlock_t x = __RW_LOCK_UNLOCKED_FAIR(x)
 
+#ifndef	__ARCH_RW_LOCK_UNLOCKED_FAIR
+#define	__ARCH_RW_LOCK_UNLOCKED_FAIR	__ARCH_RW_LOCK_UNLOCKED
+#endif
 #endif /* __LINUX_RWLOCK_TYPES_H */
diff --git a/lib/spinlock_debug.c b/lib/spinlock_debug.c
index 0374a59..d6ef7ce 100644
--- a/lib/spinlock_debug.c
+++ b/lib/spinlock_debug.c
@@ -49,6 +49,25 @@ void __rwlock_init(rwlock_t *lock, const char *name,
 
 EXPORT_SYMBOL(__rwlock_init);
 
+#ifdef CONFIG_QUEUE_RWLOCK
+void __rwlock_init_fair(rwlock_t *lock, const char *name,
+			struct lock_class_key *key)
+{
+#ifdef CONFIG_DEBUG_LOCK_ALLOC
+	/*
+	 * Make sure we are not reinitializing a held lock:
+	 */
+	debug_check_no_locks_freed((void *)lock, sizeof(*lock));
+	lockdep_init_map(&lock->dep_map, name, key, 0);
+#endif
+	lock->raw_lock = (arch_rwlock_t) __ARCH_RW_LOCK_UNLOCKED_FAIR;
+	lock->magic = RWLOCK_MAGIC;
+	lock->owner = SPINLOCK_OWNER_INIT;
+	lock->owner_cpu = -1;
+}
+EXPORT_SYMBOL(__rwlock_init_fair);
+#endif /* CONFIG_QUEUE_RWLOCK */
+
 static void spin_dump(raw_spinlock_t *lock, const char *msg)
 {
 	struct task_struct *owner = NULL;
-- 
1.7.1

[PATCH v6 4/5] qrwlock: Use the mcs_spinlock helper functions for MCS queuing

[Waiman Long]

There is a pending MCS lock patch series that adds a generic MCS
locking helper functions to do MCS-style locking. This patch will
enable the queue rwlock to use that generic MCS lock/unlock primitives
for internal queuing. This patch should only be merged after the
merging of that generic MCS locking patch.

Signed-off-by: Waiman Long <Waiman.Long@hp.com>
---
 include/asm-generic/qrwlock.h |    7 +--
 lib/qrwlock.c                 |   83 +++-------------------------------------
 2 files changed, 9 insertions(+), 81 deletions(-)

diff --git a/include/asm-generic/qrwlock.h b/include/asm-generic/qrwlock.h
index 1f68499..288fb00 100644
--- a/include/asm-generic/qrwlock.h
+++ b/include/asm-generic/qrwlock.h
@@ -54,10 +54,7 @@ typedef u64 __nrcpupair_t;
  * QRW_READER_BIAS to the rw field to increment the reader count won't
  * disturb the writer and the fair fields.
  */
-struct qrwnode {
-	struct qrwnode *next;
-	bool		wait;	/* Waiting flag */
-};
+struct mcs_spinlock;
 
 typedef struct qrwlock {
 	union qrwcnts {
@@ -74,7 +71,7 @@ typedef struct qrwlock {
 		};
 		__nrcpupair_t rw;		/* Reader/writer number pair */
 	} cnts;
-	struct qrwnode *waitq;			/* Tail of waiting queue */
+	struct mcs_spinlock *waitq;		/* Tail of waiting queue */
 } arch_rwlock_t;
 
 /*
diff --git a/lib/qrwlock.c b/lib/qrwlock.c
index 4915dc6..3ae1359 100644
--- a/lib/qrwlock.c
+++ b/lib/qrwlock.c
@@ -20,6 +20,7 @@
 #include <linux/cpumask.h>
 #include <linux/percpu.h>
 #include <linux/hardirq.h>
+#include <linux/mcs_spinlock.h>
 #include <asm-generic/qrwlock.h>
 
 /*
@@ -57,76 +58,6 @@
 # endif
 #endif
 
-#ifndef smp_mb__store_release
-# ifdef CONFIG_X86
-#   define smp_mb__store_release()	barrier()
-# else
-#   define smp_mb__store_release()	smp_mb()
-# endif
-#endif
-
-/**
- * wait_in_queue - Add to queue and wait until it is at the head
- * @lock: Pointer to queue rwlock structure
- * @node: Node pointer to be added to the queue
- */
-static __always_inline void
-wait_in_queue(struct qrwlock *lock, struct qrwnode *node)
-{
-	struct qrwnode *prev;
-
-	node->next = NULL;
-	node->wait = true;
-	prev = xchg(&lock->waitq, node);
-	if (prev) {
-		prev->next = node;
-		/*
-		 * Wait until the waiting flag is off
-		 */
-		while (ACCESS_ONCE(node->wait))
-			arch_mutex_cpu_relax();
-		smp_mb__load_acquire();
-	}
-}
-
-/**
- * signal_next - Signal the next one in queue to be at the head
- * @lock: Pointer to queue rwlock structure
- * @node: Node pointer to the current head of queue
- */
-static __always_inline void
-signal_next(struct qrwlock *lock, struct qrwnode *node)
-{
-	struct qrwnode *next;
-
-	/*
-	 * Try to notify the next node first without disturbing the cacheline
-	 * of the lock. If that fails, check to see if it is the last node
-	 * and so should clear the wait queue.
-	 */
-	next = ACCESS_ONCE(node->next);
-	if (likely(next))
-		goto notify_next;
-
-	/*
-	 * Clear the wait queue if it is the last node
-	 */
-	if ((ACCESS_ONCE(lock->waitq) == node) &&
-	    (cmpxchg(&lock->waitq, node, NULL) == node))
-			return;
-	/*
-	 * Wait until the next one in queue set up the next field
-	 */
-	while (likely(!(next = ACCESS_ONCE(node->next))))
-		arch_mutex_cpu_relax();
-	/*
-	 * The next one in queue is now at the head
-	 */
-notify_next:
-	smp_mb__store_release();
-	ACCESS_ONCE(next->wait) = false;
-}
-
 /**
  * rspin_until_writer_unlock - inc reader count & spin until writer is gone
  * @lock: Pointer to queue rwlock structure
@@ -150,7 +81,7 @@ rspin_until_writer_unlock(struct qrwlock *lock, union qrwcnts cnts)
  */
 void queue_read_lock_slowpath(struct qrwlock *lock)
 {
-	struct qrwnode node;
+	struct mcs_spinlock node;
 	union qrwcnts cnts;
 
 	/*
@@ -170,7 +101,7 @@ void queue_read_lock_slowpath(struct qrwlock *lock)
 	/*
 	 * Put the reader into the wait queue
 	 */
-	wait_in_queue(lock, &node);
+	mcs_spin_lock(lock, &node);
 
 	/*
 	 * At the head of the wait queue now, try to increment the reader
@@ -190,7 +121,7 @@ void queue_read_lock_slowpath(struct qrwlock *lock)
 	 * Need to have a barrier with read-acquire semantics
 	 */
 	smp_mb__load_acquire();
-	signal_next(lock, &node);
+	mcs_spin_unlock(lock, &node);
 }
 EXPORT_SYMBOL(queue_read_lock_slowpath);
 
@@ -256,12 +187,12 @@ static noinline int queue_write_3step_lock(struct qrwlock *lock)
  */
 void queue_write_lock_slowpath(struct qrwlock *lock)
 {
-	struct qrwnode node;
+	struct mcs_spinlock node;
 
 	/*
 	 * Put the writer into the wait queue
 	 */
-	wait_in_queue(lock, &node);
+	mcs_spin_lock(lock, &node);
 
 	/*
 	 * At the head of the wait queue now, call queue_write_3step_lock()
@@ -269,6 +200,6 @@ void queue_write_lock_slowpath(struct qrwlock *lock)
 	 */
 	while (!queue_write_3step_lock(lock))
 		arch_mutex_cpu_relax();
-	signal_next(lock, &node);
+	mcs_spin_unlock(lock, &node);
 }
 EXPORT_SYMBOL(queue_write_lock_slowpath);
-- 
1.7.1

[PATCH v6 5/5] qrwlock: Use smp_store_release() in write_unlock()

[Waiman Long]

This patch modifies the queue_write_unlock() function to use the
new smp_store_release() function in another pending patch. This patch
should only be merged if the other patch was merged.

Signed-off-by: Waiman Long <Waiman.Long@hp.com>
---
 include/asm-generic/qrwlock.h |    4 +---
 1 files changed, 1 insertions(+), 3 deletions(-)

diff --git a/include/asm-generic/qrwlock.h b/include/asm-generic/qrwlock.h
index 288fb00..479c142 100644
--- a/include/asm-generic/qrwlock.h
+++ b/include/asm-generic/qrwlock.h
@@ -219,9 +219,7 @@ static inline void queue_write_unlock(struct qrwlock *lock)
 	/*
 	 * Make sure that none of the critical section will be leaked out.
 	 */
-	smp_mb__before_clear_bit();
-	ACCESS_ONCE(lock->cnts.writer) = 0;
-	smp_mb__after_clear_bit();
+	smp_store_release(&lock->cnts.writer, 0);
 }
 
 /*
-- 
1.7.1

Re: [PATCH v6 1/3] qrwlock: A queue read/write lock implementation

[Waiman Long]

On 11/12/2013 09:48 AM, Waiman Long wrote:
> This patch introduces a new read/write lock implementation that put
> waiting readers and writers into a queue instead of actively contending
> the lock like the current read/write lock implementation. This will
> improve performance in highly contended situation by reducing the
> cache line bouncing effect.
>
>

Sorry, it is a fat finger. Please ignore this v6 1/3 patch. It is an old 
one.

-Longman

Re: [PATCH v6 3/5] qrwlock: Enable fair queue read/write lock

[Linus Torvalds]

On Tue, Nov 12, 2013 at 6:48 AM, Waiman Long <Waiman.Long@hp.com> wrote:
> By default, queue rwlock is fair among writers and gives preference
> to readers allowing them to steal lock even if a writer is
> waiting. However, there is a desire to have a fair variant of
> rwlock that is more deterministic. To enable this [..]

Is there really any point in having the option for unfair at all?

>From your timings, it looks like the unfair locks are more expensive
for the writer side, but since pretty much the whole point of rwlocks
is when readers are the common case, I don't think we care.

And I'm not at all convinced we want the complexity of two different
kinds of rwlocks with different semantics and extra code for said
semantics..

Your *original* fair rwlocks were unusable, since they didn't allow
for the irq semantics that most users need, but afaik your current
version always makes an irq/bh-context reader work even when the lock
is otherwise trying to be fair, so this whole dual behavior seems to
be largely pointless.

No?

                        Linus

Re: [PATCH v6 3/5] qrwlock: Enable fair queue read/write lock

[Andi Kleen]

On Mon, Nov 18, 2013 at 10:11:48AM -0800, Linus Torvalds wrote:
> On Tue, Nov 12, 2013 at 6:48 AM, Waiman Long <Waiman.Long@hp.com> wrote:
> > By default, queue rwlock is fair among writers and gives preference
> > to readers allowing them to steal lock even if a writer is
> > waiting. However, there is a desire to have a fair variant of
> > rwlock that is more deterministic. To enable this [..]
> 
> Is there really any point in having the option for unfair at all?

FWIW unfair can be faster in some cases. It depends on the workload.

The fair lock doesn't know anything about the topology of the system,
so it will happily spread queuers over all sockets or cores, no
matter how long the latency is.

An unfair lock can do things like "short cut to a very nearby thread".

Some unfairness tends to help with very small critical regions,
when you're not too much contended.

I would assume in cases where we don't expect all CPUs to bang 
on a lock it could be a good idea.

-Andi

Re: [PATCH v6 3/5] qrwlock: Enable fair queue read/write lock

[Linus Torvalds]

On Mon, Nov 18, 2013 at 10:34 AM, Andi Kleen <andi@firstfloor.org> wrote:
>
> FWIW unfair can be faster in some cases. It depends on the workload.

Sure. And we've never cared before.

When we switched over to the ticket spinlocks, we actually had
*numbers* about how the old unfair spinlocks could be faster. And we
still didn't leave a "unfair spinlocks" option. Because it didn't make
sense.

Why would it make sense here?

                Linus

Re: [PATCH v6 3/5] qrwlock: Enable fair queue read/write lock

[Andi Kleen]

> Why would it make sense here?

There may be cases were switching all read locks to unfair may make
concerete workloads slower.

The effect is very visible in (non kernel) lock micro benchmarks,
especially with HyperThreading.

With very high contention or long enough critical sections 
the ordered lock usually wins, but it loses with lower contention.

Unfortunately the "small critical section" case (even though it's
really bad for any contended lock) is reasonably common :-/
[IMHO all of these should be fixed or "batched" somehow, but in some 
cases it is quite hard]

However ordered locks definitely have more consistent performance.
If prioritizing consistency over potential slow down in some cases
is fine only having the ordered option is ok.

-Andi

Re: [PATCH v6 3/5] qrwlock: Enable fair queue read/write lock

[Linus Torvalds]

On Mon, Nov 18, 2013 at 10:46 AM, Andi Kleen <andi@firstfloor.org> wrote:
>> Why would it make sense here?
>
> There may be cases were switching all read locks to unfair may make
> concerete workloads slower.

Sorry, but I don't believe in "may be" as an excuse for complexity.
Especially since the "may be" faster performance is often coupled with
"known latency problems due to unfairness".

I'd want numbers for real loads if we really want the extra complexity.

Right now, the real numbers I can point to is in the size of the
patch, and the extra code complexity. Yes, I see the microbenchmark
numbers, but those are pretty much irrelevant to real loads.

                    Linus

RE: [PATCH v6 3/5] qrwlock: Enable fair queue read/write lock

[Long, Wai Man]

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #1: Type: text/plain; charset="utf-8", Size: 2399 bytes --]

Sorry for the late reply.

The presence of the unfair option as default is largely due to fixing the recursive read lock in interrupt handler problem. You are right that it is no longer a valid reason now. 

An unfair lock may have a better performance characteristics, especially for a highly contended lock. Other than that, I have no other reason to keep an unfair option. I don't mind taking out the unfair option and make the fair qrwlock the default. It does simplify the code. I will send out an updated patchset with only the fair qrwlock.

Thank for suggestion.
-Longman

-----Original Message-----
From: linus971@gmail.com [mailto:linus971@gmail.com] On Behalf Of Linus Torvalds
Sent: Monday, November 18, 2013 1:12 PM
To: Long, Wai Man
Cc: Thomas Gleixner; Ingo Molnar; H. Peter Anvin; Arnd Bergmann; linux-arch@vger.kernel.org; the arch/x86 maintainers; Linux Kernel Mailing List; Peter Zijlstra; Steven Rostedt; Andrew Morton; Michel Lespinasse; Andi Kleen; Rik van Riel; Paul E. McKenney; Raghavendra K T; George Spelvin; Tim Chen; Chandramouleeswaran, Aswin; Norton, Scott J
Subject: Re: [PATCH v6 3/5] qrwlock: Enable fair queue read/write lock

On Tue, Nov 12, 2013 at 6:48 AM, Waiman Long <Waiman.Long@hp.com> wrote:
> By default, queue rwlock is fair among writers and gives preference to 
> readers allowing them to steal lock even if a writer is waiting. 
> However, there is a desire to have a fair variant of rwlock that is 
> more deterministic. To enable this [..]

Is there really any point in having the option for unfair at all?

>From your timings, it looks like the unfair locks are more expensive for the writer side, but since pretty much the whole point of rwlocks is when readers are the common case, I don't think we care.

And I'm not at all convinced we want the complexity of two different kinds of rwlocks with different semantics and extra code for said semantics..

Your *original* fair rwlocks were unusable, since they didn't allow for the irq semantics that most users need, but afaik your current version always makes an irq/bh-context reader work even when the lock is otherwise trying to be fair, so this whole dual behavior seems to be largely pointless.

No?

                        Linus
ÿôèº{.nÇ+‰·Ÿ®‰­†+%ŠËÿ±éݶ\x17¥Šwÿº{.nÇ+‰·¥Š{±þG«éÿŠ{ayº\x1dʇڙë,j\a­¢f£¢·hšïêÿ‘êçz_è®\x03(­éšŽŠÝ¢j"ú\x1a¶^[m§ÿÿ¾\a«þG«éÿ¢¸?™¨è­Ú&£ø§~á¶iO•æ¬z·švØ^\x14\x04\x1a¶^[m§ÿÿÃ\fÿ¶ìÿ¢¸?–I¥