locking/rt: Add base code for RT rw_semaphore and rwlock
authorThomas Gleixner <tglx@linutronix.de>
Sun, 15 Aug 2021 21:28:03 +0000 (23:28 +0200)
committerIngo Molnar <mingo@kernel.org>
Tue, 17 Aug 2021 15:12:22 +0000 (17:12 +0200)
On PREEMPT_RT, rw_semaphores and rwlocks are substituted with an rtmutex and
a reader count. The implementation is writer unfair, as it is not feasible
to do priority inheritance on multiple readers, but experience has shown
that real-time workloads are not the typical workloads which are sensitive
to writer starvation.

The inner workings of rw_semaphores and rwlocks on RT are almost identical
except for the task state and signal handling. rw_semaphores are not state
preserving over a contention, they are expected to enter and leave with state
== TASK_RUNNING. rwlocks have a mechanism to preserve the state of the task
at entry and restore it after unblocking taking potential non-lock related
wakeups into account. rw_semaphores can also be subject to signal handling
interrupting a blocked state, while rwlocks ignore signals.

To avoid code duplication, provide a shared implementation which takes the
small difference vs. state and signals into account. The code is included
into the relevant rw_semaphore/rwlock base code and compiled for each use
case separately.

Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/20210815211302.957920571@linutronix.de
include/linux/rwbase_rt.h [new file with mode: 0644]
kernel/locking/rwbase_rt.c [new file with mode: 0644]

diff --git a/include/linux/rwbase_rt.h b/include/linux/rwbase_rt.h
new file mode 100644 (file)
index 0000000..1d264dd
--- /dev/null
@@ -0,0 +1,39 @@
+// SPDX-License-Identifier: GPL-2.0-only
+#ifndef _LINUX_RWBASE_RT_H
+#define _LINUX_RWBASE_RT_H
+
+#include <linux/rtmutex.h>
+#include <linux/atomic.h>
+
+#define READER_BIAS            (1U << 31)
+#define WRITER_BIAS            (1U << 30)
+
+struct rwbase_rt {
+       atomic_t                readers;
+       struct rt_mutex_base    rtmutex;
+};
+
+#define __RWBASE_INITIALIZER(name)                             \
+{                                                              \
+       .readers = ATOMIC_INIT(READER_BIAS),                    \
+       .rtmutex = __RT_MUTEX_BASE_INITIALIZER(name.rtmutex),   \
+}
+
+#define init_rwbase_rt(rwbase)                                 \
+       do {                                                    \
+               rt_mutex_base_init(&(rwbase)->rtmutex);         \
+               atomic_set(&(rwbase)->readers, READER_BIAS);    \
+       } while (0)
+
+
+static __always_inline bool rw_base_is_locked(struct rwbase_rt *rwb)
+{
+       return atomic_read(&rwb->readers) != READER_BIAS;
+}
+
+static __always_inline bool rw_base_is_contended(struct rwbase_rt *rwb)
+{
+       return atomic_read(&rwb->readers) > 0;
+}
+
+#endif /* _LINUX_RWBASE_RT_H */
diff --git a/kernel/locking/rwbase_rt.c b/kernel/locking/rwbase_rt.c
new file mode 100644 (file)
index 0000000..4ba1508
--- /dev/null
@@ -0,0 +1,263 @@
+// SPDX-License-Identifier: GPL-2.0-only
+
+/*
+ * RT-specific reader/writer semaphores and reader/writer locks
+ *
+ * down_write/write_lock()
+ *  1) Lock rtmutex
+ *  2) Remove the reader BIAS to force readers into the slow path
+ *  3) Wait until all readers have left the critical section
+ *  4) Mark it write locked
+ *
+ * up_write/write_unlock()
+ *  1) Remove the write locked marker
+ *  2) Set the reader BIAS, so readers can use the fast path again
+ *  3) Unlock rtmutex, to release blocked readers
+ *
+ * down_read/read_lock()
+ *  1) Try fast path acquisition (reader BIAS is set)
+ *  2) Take tmutex::wait_lock, which protects the writelocked flag
+ *  3) If !writelocked, acquire it for read
+ *  4) If writelocked, block on tmutex
+ *  5) unlock rtmutex, goto 1)
+ *
+ * up_read/read_unlock()
+ *  1) Try fast path release (reader count != 1)
+ *  2) Wake the writer waiting in down_write()/write_lock() #3
+ *
+ * down_read/read_lock()#3 has the consequence, that rw semaphores and rw
+ * locks on RT are not writer fair, but writers, which should be avoided in
+ * RT tasks (think mmap_sem), are subject to the rtmutex priority/DL
+ * inheritance mechanism.
+ *
+ * It's possible to make the rw primitives 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.
+ *
+ * Common code shared between RT rw_semaphore and rwlock
+ */
+
+static __always_inline int rwbase_read_trylock(struct rwbase_rt *rwb)
+{
+       int r;
+
+       /*
+        * Increment reader count, if sem->readers < 0, i.e. READER_BIAS is
+        * set.
+        */
+       for (r = atomic_read(&rwb->readers); r < 0;) {
+               if (likely(atomic_try_cmpxchg(&rwb->readers, &r, r + 1)))
+                       return 1;
+       }
+       return 0;
+}
+
+static int __sched __rwbase_read_lock(struct rwbase_rt *rwb,
+                                     unsigned int state)
+{
+       struct rt_mutex_base *rtm = &rwb->rtmutex;
+       int ret;
+
+       raw_spin_lock_irq(&rtm->wait_lock);
+       /*
+        * Allow readers, as long as the writer has not completely
+        * acquired the semaphore for write.
+        */
+       if (atomic_read(&rwb->readers) != WRITER_BIAS) {
+               atomic_inc(&rwb->readers);
+               raw_spin_unlock_irq(&rtm->wait_lock);
+               return 0;
+       }
+
+       /*
+        * Call into the slow lock path with the rtmutex->wait_lock
+        * held, so this can't result in the following race:
+        *
+        * Reader1              Reader2         Writer
+        *                      down_read()
+        *                                      down_write()
+        *                                      rtmutex_lock(m)
+        *                                      wait()
+        * down_read()
+        * unlock(m->wait_lock)
+        *                      up_read()
+        *                      wake(Writer)
+        *                                      lock(m->wait_lock)
+        *                                      sem->writelocked=true
+        *                                      unlock(m->wait_lock)
+        *
+        *                                      up_write()
+        *                                      sem->writelocked=false
+        *                                      rtmutex_unlock(m)
+        *                      down_read()
+        *                                      down_write()
+        *                                      rtmutex_lock(m)
+        *                                      wait()
+        * rtmutex_lock(m)
+        *
+        * That would put Reader1 behind the writer waiting on
+        * Reader2 to call up_read(), which might be unbound.
+        */
+
+       /*
+        * For rwlocks this returns 0 unconditionally, so the below
+        * !ret conditionals are optimized out.
+        */
+       ret = rwbase_rtmutex_slowlock_locked(rtm, state);
+
+       /*
+        * On success the rtmutex is held, so there can't be a writer
+        * active. Increment the reader count and immediately drop the
+        * rtmutex again.
+        *
+        * rtmutex->wait_lock has to be unlocked in any case of course.
+        */
+       if (!ret)
+               atomic_inc(&rwb->readers);
+       raw_spin_unlock_irq(&rtm->wait_lock);
+       if (!ret)
+               rwbase_rtmutex_unlock(rtm);
+       return ret;
+}
+
+static __always_inline int rwbase_read_lock(struct rwbase_rt *rwb,
+                                           unsigned int state)
+{
+       if (rwbase_read_trylock(rwb))
+               return 0;
+
+       return __rwbase_read_lock(rwb, state);
+}
+
+static void __sched __rwbase_read_unlock(struct rwbase_rt *rwb,
+                                        unsigned int state)
+{
+       struct rt_mutex_base *rtm = &rwb->rtmutex;
+       struct task_struct *owner;
+
+       raw_spin_lock_irq(&rtm->wait_lock);
+       /*
+        * Wake the writer, i.e. the rtmutex owner. It might release the
+        * rtmutex concurrently in the fast path (due to a signal), but to
+        * clean up rwb->readers it needs to acquire rtm->wait_lock. The
+        * worst case which can happen is a spurious wakeup.
+        */
+       owner = rt_mutex_owner(rtm);
+       if (owner)
+               wake_up_state(owner, state);
+
+       raw_spin_unlock_irq(&rtm->wait_lock);
+}
+
+static __always_inline void rwbase_read_unlock(struct rwbase_rt *rwb,
+                                              unsigned int state)
+{
+       /*
+        * rwb->readers can only hit 0 when a writer is waiting for the
+        * active readers to leave the critical section.
+        */
+       if (unlikely(atomic_dec_and_test(&rwb->readers)))
+               __rwbase_read_unlock(rwb, state);
+}
+
+static inline void __rwbase_write_unlock(struct rwbase_rt *rwb, int bias,
+                                        unsigned long flags)
+{
+       struct rt_mutex_base *rtm = &rwb->rtmutex;
+
+       atomic_add(READER_BIAS - bias, &rwb->readers);
+       raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);
+       rwbase_rtmutex_unlock(rtm);
+}
+
+static inline void rwbase_write_unlock(struct rwbase_rt *rwb)
+{
+       struct rt_mutex_base *rtm = &rwb->rtmutex;
+       unsigned long flags;
+
+       raw_spin_lock_irqsave(&rtm->wait_lock, flags);
+       __rwbase_write_unlock(rwb, WRITER_BIAS, flags);
+}
+
+static inline void rwbase_write_downgrade(struct rwbase_rt *rwb)
+{
+       struct rt_mutex_base *rtm = &rwb->rtmutex;
+       unsigned long flags;
+
+       raw_spin_lock_irqsave(&rtm->wait_lock, flags);
+       /* Release it and account current as reader */
+       __rwbase_write_unlock(rwb, WRITER_BIAS - 1, flags);
+}
+
+static int __sched rwbase_write_lock(struct rwbase_rt *rwb,
+                                    unsigned int state)
+{
+       struct rt_mutex_base *rtm = &rwb->rtmutex;
+       unsigned long flags;
+
+       /* Take the rtmutex as a first step */
+       if (rwbase_rtmutex_lock_state(rtm, state))
+               return -EINTR;
+
+       /* Force readers into slow path */
+       atomic_sub(READER_BIAS, &rwb->readers);
+
+       raw_spin_lock_irqsave(&rtm->wait_lock, flags);
+       /*
+        * set_current_state() for rw_semaphore
+        * current_save_and_set_rtlock_wait_state() for rwlock
+        */
+       rwbase_set_and_save_current_state(state);
+
+       /* Block until all readers have left the critical section. */
+       for (; atomic_read(&rwb->readers);) {
+               /* Optimized out for rwlocks */
+               if (rwbase_signal_pending_state(state, current)) {
+                       __set_current_state(TASK_RUNNING);
+                       __rwbase_write_unlock(rwb, 0, flags);
+                       return -EINTR;
+               }
+               raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);
+
+               /*
+                * Schedule and wait for the readers to leave the critical
+                * section. The last reader leaving it wakes the waiter.
+                */
+               if (atomic_read(&rwb->readers) != 0)
+                       rwbase_schedule();
+               set_current_state(state);
+               raw_spin_lock_irqsave(&rtm->wait_lock, flags);
+       }
+
+       atomic_set(&rwb->readers, WRITER_BIAS);
+       rwbase_restore_current_state();
+       raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);
+       return 0;
+}
+
+static inline int rwbase_write_trylock(struct rwbase_rt *rwb)
+{
+       struct rt_mutex_base *rtm = &rwb->rtmutex;
+       unsigned long flags;
+
+       if (!rwbase_rtmutex_trylock(rtm))
+               return 0;
+
+       atomic_sub(READER_BIAS, &rwb->readers);
+
+       raw_spin_lock_irqsave(&rtm->wait_lock, flags);
+       if (!atomic_read(&rwb->readers)) {
+               atomic_set(&rwb->readers, WRITER_BIAS);
+               raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);
+               return 1;
+       }
+       __rwbase_write_unlock(rwb, 0, flags);
+       return 0;
+}