]> git.proxmox.com Git - mirror_spl-debian.git/commitdiff
Simplify rwlock implementation.
authorBrian Behlendorf <behlendorf1@llnl.gov>
Thu, 20 May 2010 21:20:34 +0000 (14:20 -0700)
committerBrian Behlendorf <behlendorf1@llnl.gov>
Thu, 20 May 2010 21:20:34 +0000 (14:20 -0700)
Remove RW_COUNT() from the rwlock implementation.  The idea was that it
could be used as a generic wrapper for getting at the internal state
of a rwlock.  While a good idea it's proven problematic to keep it
correct for multiple archs and internal implementation changes.  In
short it hasn't been worth the trouble.

With that and simplicity in mind things have been updated to use the
rwsem_is_locked() function instead of RW_COUNT for the RW_*_HELD()
functions.  As for rw_upgrade() it remains only implemented for
the generic rwsem implemenation.  It remains to be determined if its
worth the effort of adding a custom implementation for each arch.

include/sys/rwlock.h
module/splat/splat-rwlock.c

index 9c4fb7e0fa33e421e6168ad930884533e5fbb4d9..e2be77c548410d1b4d2125deae3aed55d5ff45dd 100644 (file)
@@ -44,38 +44,8 @@ typedef struct {
         kthread_t *rw_owner;
 } krwlock_t;
 
-/*
- * For the generic implementations of rw-semaphores the following is
- * true.  If your semaphore implementation internally represents the
- * semaphore state differently.  Then special case handling will be
- * required so RW_COUNT() provides these semantics:
- * - if activity/count is 0 then there are no active readers or writers
- * - if activity/count is +ve then that is the number of active readers
- * - if activity/count is -1 then there is one active writer
- */
 #define SEM(rwp)                        ((struct rw_semaphore *)(rwp))
 
-#if defined(CONFIG_RWSEM_GENERIC_SPINLOCK)
-# define RW_COUNT(rwp)                  (SEM(rwp)->activity)
-# define rw_exit_locked(rwp)            __up_read_locked(rwp)
-# define rw_tryenter_locked(rwp)        __down_write_trylock_locked(rwp)
-extern void __up_read_locked(struct rw_semaphore *);
-extern int __down_write_trylock_locked(struct rw_semaphore *);
-#else
-/*
- * 2.6.x  - 2.6.27  use guard macro _I386_RWSEM_H
- * 2.6.28 - 2.6.32+ use guard macro _ASM_X86_RWSEM_H
- */
-# if defined(_I386_RWSEM_H) || defined(_ASM_X86_RWSEM_H)
-#  define RW_COUNT(rwp)                 ((SEM(rwp)->count < 0) ? (-1) : \
-                                        (SEM(rwp)->count & RWSEM_ACTIVE_MASK))
-# else
-#  define RW_COUNT(rwp)                 (SEM(rwp)->count & RWSEM_ACTIVE_MASK)
-# endif
-# define rw_exit_locked(rwp)           up_read(rwp)
-# define rw_tryenter_locked(rwp)       down_write_trylock(rwp)
-#endif
-
 static inline kthread_t *
 spl_rw_get_owner(krwlock_t *rwp)
 {
@@ -122,7 +92,7 @@ RW_READ_HELD(krwlock_t *rwp)
         int rc;
 
         spin_lock_irqsave(&SEM(rwp)->wait_lock, flags);
-        rc = ((RW_COUNT(rwp) > 0) && (spl_rw_get_owner(rwp) == NULL));
+        rc = (rwsem_is_locked(SEM(rwp)) && spl_rw_get_owner(rwp) == NULL);
         spin_unlock_irqrestore(&SEM(rwp)->wait_lock, flags);
 
         return rc;
@@ -135,7 +105,7 @@ RW_WRITE_HELD(krwlock_t *rwp)
         int rc;
 
         spin_lock_irqsave(&SEM(rwp)->wait_lock, flags);
-        rc = ((RW_COUNT(rwp) < 0) && (spl_rw_get_owner(rwp) == current));
+        rc = (rwsem_is_locked(SEM(rwp)) && spl_rw_get_owner(rwp) == current);
         spin_unlock_irqrestore(&SEM(rwp)->wait_lock, flags);
 
         return rc;
@@ -148,7 +118,7 @@ RW_LOCK_HELD(krwlock_t *rwp)
         int rc;
 
         spin_lock_irqsave(&SEM(rwp)->wait_lock, flags);
-        rc = (RW_COUNT(rwp) != 0);
+        rc = rwsem_is_locked(SEM(rwp));
         spin_unlock_irqrestore(&SEM(rwp)->wait_lock, flags);
 
         return rc;
@@ -224,15 +194,28 @@ RW_LOCK_HELD(krwlock_t *rwp)
 })
 
 #if defined(CONFIG_RWSEM_GENERIC_SPINLOCK)
+/*
+ * For the generic implementations of rw-semaphores the following is
+ * true.  If your semaphore implementation internally represents the
+ * semaphore state differently then special case handling is required.
+ * - if activity/count is 0 then there are no active readers or writers
+ * - if activity/count is +ve then that is the number of active readers
+ * - if activity/count is -1 then there is one active writer
+ */
+
+extern void __up_read_locked(struct rw_semaphore *);
+extern int __down_write_trylock_locked(struct rw_semaphore *);
+
 #define rw_tryupgrade(rwp)                                              \
 ({                                                                      \
         unsigned long _flags_;                                          \
         int _rc_ = 0;                                                   \
                                                                         \
         spin_lock_irqsave(&SEM(rwp)->wait_lock, _flags_);               \
-        if (list_empty(&SEM(rwp)->wait_list) && (RW_COUNT(rwp) == 1)) { \
-                rw_exit_locked(SEM(rwp));                               \
-                VERIFY(_rc_ = rw_tryenter_locked(SEM(rwp)));            \
+        if ((list_empty(&SEM(rwp)->wait_list)) &&                       \
+            (SEM(rwp)->activity == 1)) {                                \
+                __up_read_locked(SEM(rwp));                             \
+                VERIFY(_rc_ = __down_write_trylock_locked(SEM(rwp)));   \
                 (rwp)->rw_owner = current;                              \
         }                                                               \
         spin_unlock_irqrestore(&SEM(rwp)->wait_lock, _flags_);          \
@@ -240,9 +223,11 @@ RW_LOCK_HELD(krwlock_t *rwp)
 })
 #else
 /*
- * This can be done correctly but for each supported arch we will need
- * a custom cmpxchg() to atomically check and promote the rwsem.  That's
- * not worth the trouble for now so rw_tryupgrade() will always fail.
+ * rw_tryupgrade() can be implemented correctly but for each supported
+ * arch we will need a custom implementation.  For the x86 implementation
+ * it looks like a custom cmpxchg() to atomically check and promote the
+ * rwsem would be safe.  For now that's not worth the trouble so in this
+ * case rw_tryupgrade() has just been disabled.
  */
 #define rw_tryupgrade(rwp)      ({ 0; })
 #endif
index 6b2ecbe8298eaf645f29e847fa03a640a10e26fb..dc560c2b9ccdbad58124fa54cba59b633b17cd06 100644 (file)
@@ -571,7 +571,7 @@ static int
 splat_rwlock_test6(struct file *file, void *arg)
 {
        rw_priv_t *rwp;
-       int rc = -EINVAL;
+       int rc;
 
        rwp = (rw_priv_t *)kmalloc(sizeof(*rwp), GFP_KERNEL);
        if (rwp == NULL)
@@ -584,15 +584,18 @@ splat_rwlock_test6(struct file *file, void *arg)
                splat_vprint(file, SPLAT_RWLOCK_TEST6_NAME,
                             "rwlock should be read lock: %d\n",
                             RW_READ_HELD(&rwp->rw_rwlock));
+               rc = -ENOLCK;
                goto out;
        }
 
-       /* With one reader upgrade should never fail */
+#if defined(CONFIG_RWSEM_GENERIC_SPINLOCK)
+       /* With one reader upgrade should never fail. */
        rc = rw_tryupgrade(&rwp->rw_rwlock);
        if (!rc) {
                splat_vprint(file, SPLAT_RWLOCK_TEST6_NAME,
-                            "rwlock contended preventing upgrade: %ld\n",
-                            (long int)RW_COUNT(&rwp->rw_rwlock));
+                            "rwlock failed upgrade from reader: %d\n",
+                            RW_READ_HELD(&rwp->rw_rwlock));
+               rc = -ENOLCK;
                goto out;
        }
 
@@ -607,6 +610,11 @@ splat_rwlock_test6(struct file *file, void *arg)
        rc = 0;
        splat_vprint(file, SPLAT_RWLOCK_TEST6_NAME, "%s",
                     "rwlock properly upgraded\n");
+#else
+       rc = 0;
+       splat_vprint(file, SPLAT_RWLOCK_TEST6_NAME, "%s",
+                    "rw_tryupgrade() is disabled for this arch\n");
+#endif
 out:
        rw_exit(&rwp->rw_rwlock);
        rw_destroy(&rwp->rw_rwlock);