]> git.karo-electronics.de Git - linux-beck.git/commitdiff
locking/percpu-rwsem: Optimize readers and reduce global impact
authorPeter Zijlstra <peterz@infradead.org>
Thu, 14 Jul 2016 18:08:46 +0000 (20:08 +0200)
committerIngo Molnar <mingo@kernel.org>
Wed, 10 Aug 2016 12:34:01 +0000 (14:34 +0200)
Currently the percpu-rwsem switches to (global) atomic ops while a
writer is waiting; which could be quite a while and slows down
releasing the readers.

This patch cures this problem by ordering the reader-state vs
reader-count (see the comments in __percpu_down_read() and
percpu_down_write()). This changes a global atomic op into a full
memory barrier, which doesn't have the global cacheline contention.

This also enables using the percpu-rwsem with rcu_sync disabled in order
to bias the implementation differently, reducing the writer latency by
adding some cost to readers.

Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Oleg Nesterov <oleg@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Paul McKenney <paulmck@linux.vnet.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
[ Fixed modular build. ]
Signed-off-by: Ingo Molnar <mingo@kernel.org>
include/linux/percpu-rwsem.h
kernel/locking/percpu-rwsem.c
kernel/rcu/sync.c

index c2fa3ecb0dce57dde9ad4a92a35dc4178d7a159e..146efefde2a157008fce62fcdaa376f13e2c7682 100644 (file)
 
 struct percpu_rw_semaphore {
        struct rcu_sync         rss;
-       unsigned int __percpu   *fast_read_ctr;
+       unsigned int __percpu   *read_count;
        struct rw_semaphore     rw_sem;
-       atomic_t                slow_read_ctr;
-       wait_queue_head_t       write_waitq;
+       wait_queue_head_t       writer;
+       int                     readers_block;
 };
 
-extern void percpu_down_read(struct percpu_rw_semaphore *);
-extern int  percpu_down_read_trylock(struct percpu_rw_semaphore *);
-extern void percpu_up_read(struct percpu_rw_semaphore *);
+extern int __percpu_down_read(struct percpu_rw_semaphore *, int);
+extern void __percpu_up_read(struct percpu_rw_semaphore *);
+
+static inline void percpu_down_read(struct percpu_rw_semaphore *sem)
+{
+       might_sleep();
+
+       rwsem_acquire_read(&sem->rw_sem.dep_map, 0, 0, _RET_IP_);
+
+       preempt_disable();
+       /*
+        * We are in an RCU-sched read-side critical section, so the writer
+        * cannot both change sem->state from readers_fast and start checking
+        * counters while we are here. So if we see !sem->state, we know that
+        * the writer won't be checking until we're past the preempt_enable()
+        * and that one the synchronize_sched() is done, the writer will see
+        * anything we did within this RCU-sched read-size critical section.
+        */
+       __this_cpu_inc(*sem->read_count);
+       if (unlikely(!rcu_sync_is_idle(&sem->rss)))
+               __percpu_down_read(sem, false); /* Unconditional memory barrier */
+       preempt_enable();
+       /*
+        * The barrier() from preempt_enable() prevents the compiler from
+        * bleeding the critical section out.
+        */
+}
+
+static inline int percpu_down_read_trylock(struct percpu_rw_semaphore *sem)
+{
+       int ret = 1;
+
+       preempt_disable();
+       /*
+        * Same as in percpu_down_read().
+        */
+       __this_cpu_inc(*sem->read_count);
+       if (unlikely(!rcu_sync_is_idle(&sem->rss)))
+               ret = __percpu_down_read(sem, true); /* Unconditional memory barrier */
+       preempt_enable();
+       /*
+        * The barrier() from preempt_enable() prevents the compiler from
+        * bleeding the critical section out.
+        */
+
+       if (ret)
+               rwsem_acquire_read(&sem->rw_sem.dep_map, 0, 1, _RET_IP_);
+
+       return ret;
+}
+
+static inline void percpu_up_read(struct percpu_rw_semaphore *sem)
+{
+       /*
+        * The barrier() in preempt_disable() prevents the compiler from
+        * bleeding the critical section out.
+        */
+       preempt_disable();
+       /*
+        * Same as in percpu_down_read().
+        */
+       if (likely(rcu_sync_is_idle(&sem->rss)))
+               __this_cpu_dec(*sem->read_count);
+       else
+               __percpu_up_read(sem); /* Unconditional memory barrier */
+       preempt_enable();
+
+       rwsem_release(&sem->rw_sem.dep_map, 1, _RET_IP_);
+}
 
 extern void percpu_down_write(struct percpu_rw_semaphore *);
 extern void percpu_up_write(struct percpu_rw_semaphore *);
 
 extern int __percpu_init_rwsem(struct percpu_rw_semaphore *,
                                const char *, struct lock_class_key *);
+
 extern void percpu_free_rwsem(struct percpu_rw_semaphore *);
 
-#define percpu_init_rwsem(brw) \
+#define percpu_init_rwsem(sem)                                 \
 ({                                                             \
        static struct lock_class_key rwsem_key;                 \
-       __percpu_init_rwsem(brw, #brw, &rwsem_key);             \
+       __percpu_init_rwsem(sem, #sem, &rwsem_key);             \
 })
 
-
 #define percpu_rwsem_is_held(sem) lockdep_is_held(&(sem)->rw_sem)
 
 static inline void percpu_rwsem_release(struct percpu_rw_semaphore *sem,
index bec0b647f9cc5291df0211c6532edd5ffe167eb2..ce182599cf2e98b51831adbf5dca6ce545df0d7f 100644 (file)
 #include <linux/sched.h>
 #include <linux/errno.h>
 
-int __percpu_init_rwsem(struct percpu_rw_semaphore *brw,
+int __percpu_init_rwsem(struct percpu_rw_semaphore *sem,
                        const char *name, struct lock_class_key *rwsem_key)
 {
-       brw->fast_read_ctr = alloc_percpu(int);
-       if (unlikely(!brw->fast_read_ctr))
+       sem->read_count = alloc_percpu(int);
+       if (unlikely(!sem->read_count))
                return -ENOMEM;
 
        /* ->rw_sem represents the whole percpu_rw_semaphore for lockdep */
-       __init_rwsem(&brw->rw_sem, name, rwsem_key);
-       rcu_sync_init(&brw->rss, RCU_SCHED_SYNC);
-       atomic_set(&brw->slow_read_ctr, 0);
-       init_waitqueue_head(&brw->write_waitq);
+       rcu_sync_init(&sem->rss, RCU_SCHED_SYNC);
+       __init_rwsem(&sem->rw_sem, name, rwsem_key);
+       init_waitqueue_head(&sem->writer);
+       sem->readers_block = 0;
        return 0;
 }
 EXPORT_SYMBOL_GPL(__percpu_init_rwsem);
 
-void percpu_free_rwsem(struct percpu_rw_semaphore *brw)
+void percpu_free_rwsem(struct percpu_rw_semaphore *sem)
 {
        /*
         * XXX: temporary kludge. The error path in alloc_super()
         * assumes that percpu_free_rwsem() is safe after kzalloc().
         */
-       if (!brw->fast_read_ctr)
+       if (!sem->read_count)
                return;
 
-       rcu_sync_dtor(&brw->rss);
-       free_percpu(brw->fast_read_ctr);
-       brw->fast_read_ctr = NULL; /* catch use after free bugs */
+       rcu_sync_dtor(&sem->rss);
+       free_percpu(sem->read_count);
+       sem->read_count = NULL; /* catch use after free bugs */
 }
 EXPORT_SYMBOL_GPL(percpu_free_rwsem);
 
-/*
- * This is the fast-path for down_read/up_read. If it succeeds we rely
- * on the barriers provided by rcu_sync_enter/exit; see the comments in
- * percpu_down_write() and percpu_up_write().
- *
- * If this helper fails the callers rely on the normal rw_semaphore and
- * atomic_dec_and_test(), so in this case we have the necessary barriers.
- */
-static bool update_fast_ctr(struct percpu_rw_semaphore *brw, unsigned int val)
+int __percpu_down_read(struct percpu_rw_semaphore *sem, int try)
 {
-       bool success;
+       /*
+        * Due to having preemption disabled the decrement happens on
+        * the same CPU as the increment, avoiding the
+        * increment-on-one-CPU-and-decrement-on-another problem.
+        *
+        * If the reader misses the writer's assignment of readers_block, then
+        * the writer is guaranteed to see the reader's increment.
+        *
+        * Conversely, any readers that increment their sem->read_count after
+        * the writer looks are guaranteed to see the readers_block value,
+        * which in turn means that they are guaranteed to immediately
+        * decrement their sem->read_count, so that it doesn't matter that the
+        * writer missed them.
+        */
 
-       preempt_disable();
-       success = rcu_sync_is_idle(&brw->rss);
-       if (likely(success))
-               __this_cpu_add(*brw->fast_read_ctr, val);
-       preempt_enable();
+       smp_mb(); /* A matches D */
 
-       return success;
-}
+       /*
+        * If !readers_block the critical section starts here, matched by the
+        * release in percpu_up_write().
+        */
+       if (likely(!smp_load_acquire(&sem->readers_block)))
+               return 1;
 
-/*
- * Like the normal down_read() this is not recursive, the writer can
- * come after the first percpu_down_read() and create the deadlock.
- *
- * Note: returns with lock_is_held(brw->rw_sem) == T for lockdep,
- * percpu_up_read() does rwsem_release(). This pairs with the usage
- * of ->rw_sem in percpu_down/up_write().
- */
-void percpu_down_read(struct percpu_rw_semaphore *brw)
-{
-       might_sleep();
-       rwsem_acquire_read(&brw->rw_sem.dep_map, 0, 0, _RET_IP_);
+       /*
+        * Per the above comment; we still have preemption disabled and
+        * will thus decrement on the same CPU as we incremented.
+        */
+       __percpu_up_read(sem);
 
-       if (likely(update_fast_ctr(brw, +1)))
-               return;
+       if (try)
+               return 0;
 
-       /* Avoid rwsem_acquire_read() and rwsem_release() */
-       __down_read(&brw->rw_sem);
-       atomic_inc(&brw->slow_read_ctr);
-       __up_read(&brw->rw_sem);
-}
-EXPORT_SYMBOL_GPL(percpu_down_read);
+       /*
+        * We either call schedule() in the wait, or we'll fall through
+        * and reschedule on the preempt_enable() in percpu_down_read().
+        */
+       preempt_enable_no_resched();
 
-int percpu_down_read_trylock(struct percpu_rw_semaphore *brw)
-{
-       if (unlikely(!update_fast_ctr(brw, +1))) {
-               if (!__down_read_trylock(&brw->rw_sem))
-                       return 0;
-               atomic_inc(&brw->slow_read_ctr);
-               __up_read(&brw->rw_sem);
-       }
-
-       rwsem_acquire_read(&brw->rw_sem.dep_map, 0, 1, _RET_IP_);
+       /*
+        * Avoid lockdep for the down/up_read() we already have them.
+        */
+       __down_read(&sem->rw_sem);
+       this_cpu_inc(*sem->read_count);
+       __up_read(&sem->rw_sem);
+
+       preempt_disable();
        return 1;
 }
+EXPORT_SYMBOL_GPL(__percpu_down_read);
 
-void percpu_up_read(struct percpu_rw_semaphore *brw)
+void __percpu_up_read(struct percpu_rw_semaphore *sem)
 {
-       rwsem_release(&brw->rw_sem.dep_map, 1, _RET_IP_);
-
-       if (likely(update_fast_ctr(brw, -1)))
-               return;
+       smp_mb(); /* B matches C */
+       /*
+        * In other words, if they see our decrement (presumably to aggregate
+        * zero, as that is the only time it matters) they will also see our
+        * critical section.
+        */
+       __this_cpu_dec(*sem->read_count);
 
-       /* false-positive is possible but harmless */
-       if (atomic_dec_and_test(&brw->slow_read_ctr))
-               wake_up_all(&brw->write_waitq);
+       /* Prod writer to recheck readers_active */
+       wake_up(&sem->writer);
 }
-EXPORT_SYMBOL_GPL(percpu_up_read);
+EXPORT_SYMBOL_GPL(__percpu_up_read);
+
+#define per_cpu_sum(var)                                               \
+({                                                                     \
+       typeof(var) __sum = 0;                                          \
+       int cpu;                                                        \
+       compiletime_assert_atomic_type(__sum);                          \
+       for_each_possible_cpu(cpu)                                      \
+               __sum += per_cpu(var, cpu);                             \
+       __sum;                                                          \
+})
 
-static int clear_fast_ctr(struct percpu_rw_semaphore *brw)
+/*
+ * Return true if the modular sum of the sem->read_count per-CPU variable is
+ * zero.  If this sum is zero, then it is stable due to the fact that if any
+ * newly arriving readers increment a given counter, they will immediately
+ * decrement that same counter.
+ */
+static bool readers_active_check(struct percpu_rw_semaphore *sem)
 {
-       unsigned int sum = 0;
-       int cpu;
+       if (per_cpu_sum(*sem->read_count) != 0)
+               return false;
+
+       /*
+        * If we observed the decrement; ensure we see the entire critical
+        * section.
+        */
 
-       for_each_possible_cpu(cpu) {
-               sum += per_cpu(*brw->fast_read_ctr, cpu);
-               per_cpu(*brw->fast_read_ctr, cpu) = 0;
-       }
+       smp_mb(); /* C matches B */
 
-       return sum;
+       return true;
 }
 
-void percpu_down_write(struct percpu_rw_semaphore *brw)
+void percpu_down_write(struct percpu_rw_semaphore *sem)
 {
+       /* Notify readers to take the slow path. */
+       rcu_sync_enter(&sem->rss);
+
+       down_write(&sem->rw_sem);
+
        /*
-        * Make rcu_sync_is_idle() == F and thus disable the fast-path in
-        * percpu_down_read() and percpu_up_read(), and wait for gp pass.
-        *
-        * The latter synchronises us with the preceding readers which used
-        * the fast-past, so we can not miss the result of __this_cpu_add()
-        * or anything else inside their criticial sections.
+        * Notify new readers to block; up until now, and thus throughout the
+        * longish rcu_sync_enter() above, new readers could still come in.
         */
-       rcu_sync_enter(&brw->rss);
+       WRITE_ONCE(sem->readers_block, 1);
 
-       /* exclude other writers, and block the new readers completely */
-       down_write(&brw->rw_sem);
+       smp_mb(); /* D matches A */
 
-       /* nobody can use fast_read_ctr, move its sum into slow_read_ctr */
-       atomic_add(clear_fast_ctr(brw), &brw->slow_read_ctr);
+       /*
+        * If they don't see our writer of readers_block, then we are
+        * guaranteed to see their sem->read_count increment, and therefore
+        * will wait for them.
+        */
 
-       /* wait for all readers to complete their percpu_up_read() */
-       wait_event(brw->write_waitq, !atomic_read(&brw->slow_read_ctr));
+       /* Wait for all now active readers to complete. */
+       wait_event(sem->writer, readers_active_check(sem));
 }
 EXPORT_SYMBOL_GPL(percpu_down_write);
 
-void percpu_up_write(struct percpu_rw_semaphore *brw)
+void percpu_up_write(struct percpu_rw_semaphore *sem)
 {
-       /* release the lock, but the readers can't use the fast-path */
-       up_write(&brw->rw_sem);
        /*
-        * Enable the fast-path in percpu_down_read() and percpu_up_read()
-        * but only after another gp pass; this adds the necessary barrier
-        * to ensure the reader can't miss the changes done by us.
+        * Signal the writer is done, no fast path yet.
+        *
+        * One reason that we cannot just immediately flip to readers_fast is
+        * that new readers might fail to see the results of this writer's
+        * critical section.
+        *
+        * Therefore we force it through the slow path which guarantees an
+        * acquire and thereby guarantees the critical section's consistency.
+        */
+       smp_store_release(&sem->readers_block, 0);
+
+       /*
+        * Release the write lock, this will allow readers back in the game.
+        */
+       up_write(&sem->rw_sem);
+
+       /*
+        * Once this completes (at least one RCU-sched grace period hence) the
+        * reader fast path will be available again. Safe to use outside the
+        * exclusive write lock because its counting.
         */
-       rcu_sync_exit(&brw->rss);
+       rcu_sync_exit(&sem->rss);
 }
 EXPORT_SYMBOL_GPL(percpu_up_write);
index be922c9f3d37256fc060d5b9ba0aaf6a2d1f085c..198473d90f81d055baecf10f10f5cfe96b308bf9 100644 (file)
@@ -68,6 +68,8 @@ void rcu_sync_lockdep_assert(struct rcu_sync *rsp)
        RCU_LOCKDEP_WARN(!gp_ops[rsp->gp_type].held(),
                         "suspicious rcu_sync_is_idle() usage");
 }
+
+EXPORT_SYMBOL_GPL(rcu_sync_lockdep_assert);
 #endif
 
 /**