locking/qrwlock: Make use of _{acquire|release|relaxed}() atomics
authorWill Deacon <will.deacon@arm.com>
Thu, 6 Aug 2015 16:54:42 +0000 (17:54 +0100)
committerIngo Molnar <mingo@kernel.org>
Wed, 12 Aug 2015 09:59:06 +0000 (11:59 +0200)
The qrwlock implementation is slightly heavy in its use of memory
barriers, mainly through the use of _cmpxchg() and _return() atomics, which
imply full barrier semantics.

This patch modifies the qrwlock code to use the more relaxed atomic
routines so that we can reduce the unnecessary barrier overhead on
weakly-ordered architectures.

Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Waiman.Long@hp.com
Cc: paulmck@linux.vnet.ibm.com
Link: http://lkml.kernel.org/r/1438880084-18856-7-git-send-email-will.deacon@arm.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
include/asm-generic/qrwlock.h
kernel/locking/qrwlock.c

index eb673dde8879075cd43e95860e2733fe64a1d312..54a8e65e18b622edf10df6d07ca0d84fea9e49ae 100644 (file)
@@ -68,7 +68,7 @@ static inline int queued_read_trylock(struct qrwlock *lock)
 
        cnts = atomic_read(&lock->cnts);
        if (likely(!(cnts & _QW_WMASK))) {
-               cnts = (u32)atomic_add_return(_QR_BIAS, &lock->cnts);
+               cnts = (u32)atomic_add_return_acquire(_QR_BIAS, &lock->cnts);
                if (likely(!(cnts & _QW_WMASK)))
                        return 1;
                atomic_sub(_QR_BIAS, &lock->cnts);
@@ -89,8 +89,8 @@ static inline int queued_write_trylock(struct qrwlock *lock)
        if (unlikely(cnts))
                return 0;
 
-       return likely(atomic_cmpxchg(&lock->cnts,
-                                    cnts, cnts | _QW_LOCKED) == cnts);
+       return likely(atomic_cmpxchg_acquire(&lock->cnts,
+                                            cnts, cnts | _QW_LOCKED) == cnts);
 }
 /**
  * queued_read_lock - acquire read lock of a queue rwlock
@@ -100,7 +100,7 @@ static inline void queued_read_lock(struct qrwlock *lock)
 {
        u32 cnts;
 
-       cnts = atomic_add_return(_QR_BIAS, &lock->cnts);
+       cnts = atomic_add_return_acquire(_QR_BIAS, &lock->cnts);
        if (likely(!(cnts & _QW_WMASK)))
                return;
 
@@ -115,7 +115,7 @@ static inline void queued_read_lock(struct qrwlock *lock)
 static inline void queued_write_lock(struct qrwlock *lock)
 {
        /* Optimize for the unfair lock case where the fair flag is 0. */
-       if (atomic_cmpxchg(&lock->cnts, 0, _QW_LOCKED) == 0)
+       if (atomic_cmpxchg_acquire(&lock->cnts, 0, _QW_LOCKED) == 0)
                return;
 
        queued_write_lock_slowpath(lock);
@@ -130,8 +130,7 @@ static inline void queued_read_unlock(struct qrwlock *lock)
        /*
         * Atomically decrement the reader count
         */
-       smp_mb__before_atomic();
-       atomic_sub(_QR_BIAS, &lock->cnts);
+       (void)atomic_sub_return_release(_QR_BIAS, &lock->cnts);
 }
 
 /**
index 6a7a3b8d5ac9e669b789c6be4542a799542ea40e..f17a3e3b355079b6658b55b2f21811a69e695190 100644 (file)
@@ -55,7 +55,7 @@ rspin_until_writer_unlock(struct qrwlock *lock, u32 cnts)
 {
        while ((cnts & _QW_WMASK) == _QW_LOCKED) {
                cpu_relax_lowlatency();
-               cnts = smp_load_acquire((u32 *)&lock->cnts);
+               cnts = atomic_read_acquire(&lock->cnts);
        }
 }
 
@@ -74,8 +74,9 @@ void queued_read_lock_slowpath(struct qrwlock *lock, u32 cnts)
                 * Readers in interrupt context will get the lock immediately
                 * if the writer is just waiting (not holding the lock yet).
                 * The rspin_until_writer_unlock() function returns immediately
-                * in this case. Otherwise, they will spin until the lock
-                * is available without waiting in the queue.
+                * in this case. Otherwise, they will spin (with ACQUIRE
+                * semantics) until the lock is available without waiting in
+                * the queue.
                 */
                rspin_until_writer_unlock(lock, cnts);
                return;
@@ -88,12 +89,11 @@ void queued_read_lock_slowpath(struct qrwlock *lock, u32 cnts)
        arch_spin_lock(&lock->lock);
 
        /*
-        * At the head of the wait queue now, increment the reader count
-        * and wait until the writer, if it has the lock, has gone away.
-        * At ths stage, it is not possible for a writer to remain in the
-        * waiting state (_QW_WAITING). So there won't be any deadlock.
+        * The ACQUIRE semantics of the following spinning code ensure
+        * that accesses can't leak upwards out of our subsequent critical
+        * section in the case that the lock is currently held for write.
         */
-       cnts = atomic_add_return(_QR_BIAS, &lock->cnts) - _QR_BIAS;
+       cnts = atomic_add_return_acquire(_QR_BIAS, &lock->cnts) - _QR_BIAS;
        rspin_until_writer_unlock(lock, cnts);
 
        /*
@@ -116,7 +116,7 @@ void queued_write_lock_slowpath(struct qrwlock *lock)
 
        /* Try to acquire the lock directly if no reader is present */
        if (!atomic_read(&lock->cnts) &&
-           (atomic_cmpxchg(&lock->cnts, 0, _QW_LOCKED) == 0))
+           (atomic_cmpxchg_acquire(&lock->cnts, 0, _QW_LOCKED) == 0))
                goto unlock;
 
        /*
@@ -127,7 +127,7 @@ void queued_write_lock_slowpath(struct qrwlock *lock)
                struct __qrwlock *l = (struct __qrwlock *)lock;
 
                if (!READ_ONCE(l->wmode) &&
-                  (cmpxchg(&l->wmode, 0, _QW_WAITING) == 0))
+                  (cmpxchg_relaxed(&l->wmode, 0, _QW_WAITING) == 0))
                        break;
 
                cpu_relax_lowlatency();
@@ -137,8 +137,8 @@ void queued_write_lock_slowpath(struct qrwlock *lock)
        for (;;) {
                cnts = atomic_read(&lock->cnts);
                if ((cnts == _QW_WAITING) &&
-                   (atomic_cmpxchg(&lock->cnts, _QW_WAITING,
-                                   _QW_LOCKED) == _QW_WAITING))
+                   (atomic_cmpxchg_acquire(&lock->cnts, _QW_WAITING,
+                                           _QW_LOCKED) == _QW_WAITING))
                        break;
 
                cpu_relax_lowlatency();