patches/locking-rt-rwlock--Provide-reader-biased-rwlock-for-RT.patch

Subject: locking/rt/rwlock: Provide reader biased rwlock for RT
From: Thomas Gleixner <tglx@linutronix.de>
Date: Wed, 12 Jul 2017 09:50:11 +0200

Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
---
 include/linux/rwlock_rt.h       |   10 +
 include/linux/rwlock_types_rt.h |   42 ++++++
 kernel/locking/Makefile         |    2 
 kernel/locking/rwlock-rt.c      |  246 ++++++++++++++++++++++++++++++++++++++++
 4 files changed, 299 insertions(+), 1 deletion(-)

--- a/include/linux/rwlock_rt.h
+++ b/include/linux/rwlock_rt.h
@@ -101,4 +101,14 @@ static inline int __write_trylock_rt_irq
 		rt_write_unlock(lock);			\
 	} while (0)
 
+/*
+ * Internal functions made global for CPU pinning
+ */
+void __read_rt_lock(struct rt_rw_lock *lock);
+int __read_rt_trylock(struct rt_rw_lock *lock);
+void __write_rt_lock(struct rt_rw_lock *lock);
+int __write_rt_trylock(struct rt_rw_lock *lock);
+void __read_rt_unlock(struct rt_rw_lock *lock);
+void __write_rt_unlock(struct rt_rw_lock *lock);
+
 #endif
--- a/include/linux/rwlock_types_rt.h
+++ b/include/linux/rwlock_types_rt.h
@@ -29,4 +29,46 @@ typedef struct {
 #define DEFINE_RWLOCK(name) \
 	rwlock_t name = __RW_LOCK_UNLOCKED(name)
 
+#define READER_BIAS	(1U << 31)
+#define WRITER_BIAS	(1U << 30)
+
+/*
+ * A reader biased implementation primarily for CPU pinning.
+ *
+ * Can be selected as general replacement for the single reader RT rwlock
+ * variant
+ */
+struct rt_rw_lock {
+	struct rt_mutex		rtmutex;
+	atomic_t		readers;
+#ifdef CONFIG_DEBUG_LOCK_ALLOC
+	struct lockdep_map	dep_map;
+#endif
+};
+
+#define __RWLOCK_RT_INITIALIZER(name)					\
+{									\
+	.readers = ATOMIC_INIT(READER_BIAS),				\
+	.rtmutex = __RT_MUTEX_INITIALIZER_SAVE_STATE(name.rtmutex),	\
+	RW_DEP_MAP_INIT(name)						\
+}
+
+void __rwlock_biased_rt_init(struct rt_rw_lock *lock, const char *name,
+			     struct lock_class_key *key);
+
+#define rwlock_biased_rt_init(rwlock)					\
+	do {								\
+		static struct lock_class_key __key;			\
+									\
+		__rwlock_biased_rt_init((rwlock), #rwlock, &__key);	\
+	} while (0)
+
+int __read_rt_trylock(struct rt_rw_lock *rwlock);
+void __read_rt_lock(struct rt_rw_lock *rwlock);
+void __read_rt_unlock(struct rt_rw_lock *rwlock);
+
+void __write_rt_lock(struct rt_rw_lock *rwlock);
+int __write_rt_trylock(struct rt_rw_lock *rwlock);
+void __write_rt_unlock(struct rt_rw_lock *rwlock);
+
 #endif
--- a/kernel/locking/Makefile
+++ b/kernel/locking/Makefile
@@ -32,7 +32,7 @@ ifneq ($(CONFIG_PREEMPT_RT_FULL),y)
 obj-$(CONFIG_RWSEM_GENERIC_SPINLOCK) += rwsem-spinlock.o
 obj-$(CONFIG_RWSEM_XCHGADD_ALGORITHM) += rwsem-xadd.o
 endif
-obj-$(CONFIG_PREEMPT_RT_FULL) += rt.o rwsem-rt.o
+obj-$(CONFIG_PREEMPT_RT_FULL) += rt.o rwsem-rt.o rwlock-rt.o
 obj-$(CONFIG_QUEUED_RWLOCKS) += qrwlock.o
 obj-$(CONFIG_LOCK_TORTURE_TEST) += locktorture.o
 obj-$(CONFIG_WW_MUTEX_SELFTEST) += test-ww_mutex.o
--- /dev/null
+++ b/kernel/locking/rwlock-rt.c
@@ -0,0 +1,246 @@
+/*
+ */
+#include <linux/sched/debug.h>
+#include <linux/export.h>
+
+#include "rtmutex_common.h"
+#include <linux/rwlock_types_rt.h>
+
+/*
+ * RT-specific reader/writer locks
+ *
+ * write_lock()
+ *  1) Lock lock->rtmutex
+ *  2) Remove the reader BIAS to force readers into the slow path
+ *  3) Wait until all readers have left the critical region
+ *  4) Mark it write locked
+ *
+ * write_unlock()
+ *  1) Remove the write locked marker
+ *  2) Set the reader BIAS so readers can use the fast path again
+ *  3) Unlock lock->rtmutex to release blocked readers
+ *
+ * read_lock()
+ *  1) Try fast path acquisition (reader BIAS is set)
+ *  2) Take lock->rtmutex.wait_lock which protects the writelocked flag
+ *  3) If !writelocked, acquire it for read
+ *  4) If writelocked, block on lock->rtmutex
+ *  5) unlock lock->rtmutex, goto 1)
+ *
+ * read_unlock()
+ *  1) Try fast path release (reader count != 1)
+ *  2) Wake the writer waiting in write_lock()#3
+ *
+ * read_lock()#3 has the consequence, that rw locks on RT are not writer
+ * fair, but writers, which should be avoided in RT tasks (think tasklist
+ * lock), are subject to the rtmutex priority/DL inheritance mechanism.
+ *
+ * It's possible to make the rw locks writer fair by keeping a list of
+ * active readers. A blocked writer would force all newly incoming readers
+ * to block on the rtmutex, but the rtmutex would have to be proxy locked
+ * for one reader after the other. We can't use multi-reader inheritance
+ * because there is no way to support that with
+ * SCHED_DEADLINE. Implementing the one by one reader boosting/handover
+ * mechanism is a major surgery for a very dubious value.
+ *
+ * The risk of writer starvation is there, but the pathological use cases
+ * which trigger it are not necessarily the typical RT workloads.
+ */
+
+void __rwlock_biased_rt_init(struct rt_rw_lock *lock, const char *name,
+			     struct lock_class_key *key)
+{
+#ifdef CONFIG_DEBUG_LOCK_ALLOC
+	/*
+	 * Make sure we are not reinitializing a held semaphore:
+	 */
+	debug_check_no_locks_freed((void *)lock, sizeof(*lock));
+	lockdep_init_map(&lock->dep_map, name, key, 0);
+#endif
+	atomic_set(&lock->readers, READER_BIAS);
+	rt_mutex_init(&lock->rtmutex);
+	lock->rtmutex.save_state = 1;
+}
+
+int __read_rt_trylock(struct rt_rw_lock *lock)
+{
+	int r, old;
+
+	/*
+	 * Increment reader count, if lock->readers < 0, i.e. READER_BIAS is
+	 * set.
+	 */
+	for (r = atomic_read(&lock->readers); r < 0;) {
+		old = atomic_cmpxchg(&lock->readers, r, r + 1);
+		if (likely(old == r))
+			return 1;
+		r = old;
+	}
+	return 0;
+}
+
+void __sched __read_rt_lock(struct rt_rw_lock *lock)
+{
+	struct rt_mutex *m = &lock->rtmutex;
+	struct rt_mutex_waiter waiter;
+	unsigned long flags;
+
+	if (__read_rt_trylock(lock))
+		return;
+
+	raw_spin_lock_irqsave(&m->wait_lock, flags);
+	/*
+	 * Allow readers as long as the writer has not completely
+	 * acquired the semaphore for write.
+	 */
+	if (atomic_read(&lock->readers) != WRITER_BIAS) {
+		atomic_inc(&lock->readers);
+		raw_spin_unlock_irqrestore(&m->wait_lock, flags);
+		return;
+	}
+
+	/*
+	 * Call into the slow lock path with the rtmutex->wait_lock
+	 * held, so this can't result in the following race:
+	 *
+	 * Reader1		Reader2		Writer
+	 *			read_lock()
+	 *					write_lock()
+	 *					rtmutex_lock(m)
+	 *					swait()
+	 * read_lock()
+	 * unlock(m->wait_lock)
+	 *			read_unlock()
+	 *			swake()
+	 *					lock(m->wait_lock)
+	 *					lock->writelocked=true
+	 *					unlock(m->wait_lock)
+	 *
+	 *					write_unlock()
+	 *					lock->writelocked=false
+	 *					rtmutex_unlock(m)
+	 *			read_lock()
+	 *					write_lock()
+	 *					rtmutex_lock(m)
+	 *					swait()
+	 * rtmutex_lock(m)
+	 *
+	 * That would put Reader1 behind the writer waiting on
+	 * Reader2 to call read_unlock() which might be unbound.
+	 */
+	rt_mutex_init_waiter(&waiter, false);
+	rt_spin_lock_slowlock_locked(m, &waiter, flags);
+	/*
+	 * The slowlock() above is guaranteed to return with the rtmutex is
+	 * now held, so there can't be a writer active. Increment the reader
+	 * count and immediately drop the rtmutex again.
+	 */
+	atomic_inc(&lock->readers);
+	raw_spin_unlock_irqrestore(&m->wait_lock, flags);
+	rt_spin_lock_slowunlock(m);
+
+	debug_rt_mutex_free_waiter(&waiter);
+}
+
+void __read_rt_unlock(struct rt_rw_lock *lock)
+{
+	struct rt_mutex *m = &lock->rtmutex;
+	struct task_struct *tsk;
+
+	/*
+	 * sem->readers can only hit 0 when a writer is waiting for the
+	 * active readers to leave the critical region.
+	 */
+	if (!atomic_dec_and_test(&lock->readers))
+		return;
+
+	raw_spin_lock_irq(&m->wait_lock);
+	/*
+	 * Wake the writer, i.e. the rtmutex owner. It might release the
+	 * rtmutex concurrently in the fast path, but to clean up the rw
+	 * lock it needs to acquire m->wait_lock. The worst case which can
+	 * happen is a spurious wakeup.
+	 */
+	tsk = rt_mutex_owner(m);
+	if (tsk)
+		wake_up_process(tsk);
+
+	raw_spin_unlock_irq(&m->wait_lock);
+}
+
+static void __write_unlock_common(struct rt_rw_lock *lock, int bias,
+				  unsigned long flags)
+{
+	struct rt_mutex *m = &lock->rtmutex;
+
+	atomic_add(READER_BIAS - bias, &lock->readers);
+	raw_spin_unlock_irqrestore(&m->wait_lock, flags);
+	rt_spin_lock_slowunlock(m);
+}
+
+void __sched __write_rt_lock(struct rt_rw_lock *lock)
+{
+	struct rt_mutex *m = &lock->rtmutex;
+	struct task_struct *self = current;
+	unsigned long flags;
+
+	/* Take the rtmutex as a first step */
+	__rt_spin_lock(m);
+
+	/* Force readers into slow path */
+	atomic_sub(READER_BIAS, &lock->readers);
+
+	for (;;) {
+		raw_spin_lock_irqsave(&m->wait_lock, flags);
+
+		raw_spin_lock(&self->pi_lock);
+		self->saved_state = self->state;
+		__set_current_state_no_track(TASK_UNINTERRUPTIBLE);
+		raw_spin_unlock(&self->pi_lock);
+
+		/* Have all readers left the critical region? */
+		if (!atomic_read(&lock->readers)) {
+			atomic_set(&lock->readers, WRITER_BIAS);
+			raw_spin_lock(&self->pi_lock);
+			__set_current_state_no_track(self->saved_state);
+			self->saved_state = TASK_RUNNING;
+			raw_spin_unlock(&self->pi_lock);
+			raw_spin_unlock_irqrestore(&m->wait_lock, flags);
+			return;
+		}
+
+		raw_spin_unlock_irqrestore(&m->wait_lock, flags);
+
+		if (atomic_read(&lock->readers) != 0)
+			schedule();
+	}
+}
+
+int __write_rt_trylock(struct rt_rw_lock *lock)
+{
+	struct rt_mutex *m = &lock->rtmutex;
+	unsigned long flags;
+
+	if (!rt_mutex_trylock(m))
+		return 0;
+
+	atomic_sub(READER_BIAS, &lock->readers);
+
+	raw_spin_lock_irqsave(&m->wait_lock, flags);
+	if (!atomic_read(&lock->readers)) {
+		atomic_set(&lock->readers, WRITER_BIAS);
+		raw_spin_unlock_irqrestore(&m->wait_lock, flags);
+		return 1;
+	}
+	__write_unlock_common(lock, 0, flags);
+	return 0;
+}
+
+void __write_rt_unlock(struct rt_rw_lock *lock)
+{
+	struct rt_mutex *m = &lock->rtmutex;
+	unsigned long flags;
+
+	raw_spin_lock_irqsave(&m->wait_lock, flags);
+	__write_unlock_common(lock, WRITER_BIAS, flags);
+}