From: Thomas Gleixner <tglx@linutronix.de>
Date: Thu, 12 Oct 2017 17:18:06 +0200
Subject: rtmutex: add rwlock implementation based on rtmutex
Git-repo: git://git.kernel.org/pub/scm/linux/kernel/git/rt/linux-rt-devel.git
Git-commit: da77ceac3d20f27310a07a7c346a4ee6b40d6c28
Patch-mainline: Queued in subsystem maintainer repository
References: SLE Realtime Extension
The implementation is bias-based, similar to the rwsem implementation.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Daniel Wagner <dwagner@suse.de>
---
include/linux/rwlock_rt.h | 119 ++++++++++++
include/linux/rwlock_types_rt.h | 55 +++++
kernel/locking/rwlock-rt.c | 368 ++++++++++++++++++++++++++++++++++++++++
3 files changed, 542 insertions(+)
create mode 100644 include/linux/rwlock_rt.h
create mode 100644 include/linux/rwlock_types_rt.h
create mode 100644 kernel/locking/rwlock-rt.c
--- /dev/null
+++ b/include/linux/rwlock_rt.h
@@ -0,0 +1,119 @@
+#ifndef __LINUX_RWLOCK_RT_H
+#define __LINUX_RWLOCK_RT_H
+
+#ifndef __LINUX_SPINLOCK_H
+#error Do not include directly. Use spinlock.h
+#endif
+
+extern void __lockfunc rt_write_lock(rwlock_t *rwlock);
+extern void __lockfunc rt_read_lock(rwlock_t *rwlock);
+extern int __lockfunc rt_write_trylock(rwlock_t *rwlock);
+extern int __lockfunc rt_read_trylock(rwlock_t *rwlock);
+extern void __lockfunc rt_write_unlock(rwlock_t *rwlock);
+extern void __lockfunc rt_read_unlock(rwlock_t *rwlock);
+extern int __lockfunc rt_read_can_lock(rwlock_t *rwlock);
+extern int __lockfunc rt_write_can_lock(rwlock_t *rwlock);
+extern void __rt_rwlock_init(rwlock_t *rwlock, char *name, struct lock_class_key *key);
+
+#define read_can_lock(rwlock) rt_read_can_lock(rwlock)
+#define write_can_lock(rwlock) rt_write_can_lock(rwlock)
+
+#define read_trylock(lock) __cond_lock(lock, rt_read_trylock(lock))
+#define write_trylock(lock) __cond_lock(lock, rt_write_trylock(lock))
+
+static inline int __write_trylock_rt_irqsave(rwlock_t *lock, unsigned long *flags)
+{
+ /* XXX ARCH_IRQ_ENABLED */
+ *flags = 0;
+ return rt_write_trylock(lock);
+}
+
+#define write_trylock_irqsave(lock, flags) \
+ __cond_lock(lock, __write_trylock_rt_irqsave(lock, &(flags)))
+
+#define read_lock_irqsave(lock, flags) \
+ do { \
+ typecheck(unsigned long, flags); \
+ rt_read_lock(lock); \
+ flags = 0; \
+ } while (0)
+
+#define write_lock_irqsave(lock, flags) \
+ do { \
+ typecheck(unsigned long, flags); \
+ rt_write_lock(lock); \
+ flags = 0; \
+ } while (0)
+
+#define read_lock(lock) rt_read_lock(lock)
+
+#define read_lock_bh(lock) \
+ do { \
+ local_bh_disable(); \
+ rt_read_lock(lock); \
+ } while (0)
+
+#define read_lock_irq(lock) read_lock(lock)
+
+#define write_lock(lock) rt_write_lock(lock)
+
+#define write_lock_bh(lock) \
+ do { \
+ local_bh_disable(); \
+ rt_write_lock(lock); \
+ } while (0)
+
+#define write_lock_irq(lock) write_lock(lock)
+
+#define read_unlock(lock) rt_read_unlock(lock)
+
+#define read_unlock_bh(lock) \
+ do { \
+ rt_read_unlock(lock); \
+ local_bh_enable(); \
+ } while (0)
+
+#define read_unlock_irq(lock) read_unlock(lock)
+
+#define write_unlock(lock) rt_write_unlock(lock)
+
+#define write_unlock_bh(lock) \
+ do { \
+ rt_write_unlock(lock); \
+ local_bh_enable(); \
+ } while (0)
+
+#define write_unlock_irq(lock) write_unlock(lock)
+
+#define read_unlock_irqrestore(lock, flags) \
+ do { \
+ typecheck(unsigned long, flags); \
+ (void) flags; \
+ rt_read_unlock(lock); \
+ } while (0)
+
+#define write_unlock_irqrestore(lock, flags) \
+ do { \
+ typecheck(unsigned long, flags); \
+ (void) flags; \
+ rt_write_unlock(lock); \
+ } while (0)
+
+#define rwlock_init(rwl) \
+do { \
+ static struct lock_class_key __key; \
+ \
+ __rt_rwlock_init(rwl, #rwl, &__key); \
+} 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
--- /dev/null
+++ b/include/linux/rwlock_types_rt.h
@@ -0,0 +1,55 @@
+#ifndef __LINUX_RWLOCK_TYPES_RT_H
+#define __LINUX_RWLOCK_TYPES_RT_H
+
+#ifndef __LINUX_SPINLOCK_TYPES_H
+#error "Do not include directly. Include spinlock_types.h instead"
+#endif
+
+#ifdef CONFIG_DEBUG_LOCK_ALLOC
+# define RW_DEP_MAP_INIT(lockname) .dep_map = { .name = #lockname }
+#else
+# define RW_DEP_MAP_INIT(lockname)
+#endif
+
+typedef struct rt_rw_lock rwlock_t;
+
+#define __RW_LOCK_UNLOCKED(name) __RWLOCK_RT_INITIALIZER(name)
+
+#define DEFINE_RWLOCK(name) \
+ rwlock_t name = __RW_LOCK_UNLOCKED(name)
+
+/*
+ * 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 READER_BIAS (1U << 31)
+#define WRITER_BIAS (1U << 30)
+
+#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)
+
+#endif
--- /dev/null
+++ b/kernel/locking/rwlock-rt.c
@@ -0,0 +1,368 @@
+/*
+ */
+#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, true);
+ 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);
+
+ 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);
+
+ for (;;) {
+ /* 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();
+
+ raw_spin_lock_irqsave(&m->wait_lock, flags);
+
+ raw_spin_lock(&self->pi_lock);
+ __set_current_state_no_track(TASK_UNINTERRUPTIBLE);
+ raw_spin_unlock(&self->pi_lock);
+ }
+}
+
+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);
+}
+
+/* Map the reader biased implementation */
+static inline int do_read_rt_trylock(rwlock_t *rwlock)
+{
+ return __read_rt_trylock(rwlock);
+}
+
+static inline int do_write_rt_trylock(rwlock_t *rwlock)
+{
+ return __write_rt_trylock(rwlock);
+}
+
+static inline void do_read_rt_lock(rwlock_t *rwlock)
+{
+ __read_rt_lock(rwlock);
+}
+
+static inline void do_write_rt_lock(rwlock_t *rwlock)
+{
+ __write_rt_lock(rwlock);
+}
+
+static inline void do_read_rt_unlock(rwlock_t *rwlock)
+{
+ __read_rt_unlock(rwlock);
+}
+
+static inline void do_write_rt_unlock(rwlock_t *rwlock)
+{
+ __write_rt_unlock(rwlock);
+}
+
+static inline void do_rwlock_rt_init(rwlock_t *rwlock, const char *name,
+ struct lock_class_key *key)
+{
+ __rwlock_biased_rt_init(rwlock, name, key);
+}
+
+int __lockfunc rt_read_can_lock(rwlock_t *rwlock)
+{
+ return atomic_read(&rwlock->readers) < 0;
+}
+
+int __lockfunc rt_write_can_lock(rwlock_t *rwlock)
+{
+ return atomic_read(&rwlock->readers) == READER_BIAS;
+}
+
+/*
+ * The common functions which get wrapped into the rwlock API.
+ */
+int __lockfunc rt_read_trylock(rwlock_t *rwlock)
+{
+ int ret;
+
+ migrate_disable();
+ ret = do_read_rt_trylock(rwlock);
+ if (ret)
+ rwlock_acquire_read(&rwlock->dep_map, 0, 1, _RET_IP_);
+ else
+ migrate_enable();
+ return ret;
+}
+EXPORT_SYMBOL(rt_read_trylock);
+
+int __lockfunc rt_write_trylock(rwlock_t *rwlock)
+{
+ int ret;
+
+ migrate_disable();
+ ret = do_write_rt_trylock(rwlock);
+ if (ret)
+ rwlock_acquire(&rwlock->dep_map, 0, 1, _RET_IP_);
+ else
+ migrate_enable();
+ return ret;
+}
+EXPORT_SYMBOL(rt_write_trylock);
+
+void __lockfunc rt_read_lock(rwlock_t *rwlock)
+{
+ migrate_disable();
+ rwlock_acquire_read(&rwlock->dep_map, 0, 0, _RET_IP_);
+ do_read_rt_lock(rwlock);
+}
+EXPORT_SYMBOL(rt_read_lock);
+
+void __lockfunc rt_write_lock(rwlock_t *rwlock)
+{
+ migrate_disable();
+ rwlock_acquire(&rwlock->dep_map, 0, 0, _RET_IP_);
+ do_write_rt_lock(rwlock);
+}
+EXPORT_SYMBOL(rt_write_lock);
+
+void __lockfunc rt_read_unlock(rwlock_t *rwlock)
+{
+ rwlock_release(&rwlock->dep_map, _RET_IP_);
+ do_read_rt_unlock(rwlock);
+ migrate_enable();
+}
+EXPORT_SYMBOL(rt_read_unlock);
+
+void __lockfunc rt_write_unlock(rwlock_t *rwlock)
+{
+ rwlock_release(&rwlock->dep_map, _RET_IP_);
+ do_write_rt_unlock(rwlock);
+ migrate_enable();
+}
+EXPORT_SYMBOL(rt_write_unlock);
+
+void __rt_rwlock_init(rwlock_t *rwlock, char *name, struct lock_class_key *key)
+{
+ do_rwlock_rt_init(rwlock, name, key);
+}
+EXPORT_SYMBOL(__rt_rwlock_init);