Palacios Public Git Repository

To checkout Palacios execute

  git clone http://v3vee.org/palacios/palacios.web/palacios.git
This will give you the master branch. You probably want the devel branch or one of the release branches. To switch to the devel branch, simply execute
  cd palacios
  git checkout --track -b devel origin/devel
The other branches are similar.


Merge branch 'devel'
[palacios.git] / kitten / include / lwk / seqlock.h
diff --git a/kitten/include/lwk/seqlock.h b/kitten/include/lwk/seqlock.h
new file mode 100644 (file)
index 0000000..11ba4ce
--- /dev/null
@@ -0,0 +1,172 @@
+#ifndef _LWK_SEQLOCK_H
+#define _LWK_SEQLOCK_H
+/*
+ * Reader/writer consistent mechanism without starving writers. This type of
+ * lock for data where the reader wants a consitent set of information
+ * and is willing to retry if the information changes.  Readers never
+ * block but they may have to retry if a writer is in
+ * progress. Writers do not wait for readers. 
+ *
+ * This is not as cache friendly as brlock. Also, this will not work
+ * for data that contains pointers, because any writer could
+ * invalidate a pointer that a reader was following.
+ *
+ * Expected reader usage:
+ *     do {
+ *         seq = read_seqbegin(&foo);
+ *     ...
+ *      } while (read_seqretry(&foo, seq));
+ *
+ *
+ * On non-SMP the spin locks disappear but the writer still needs
+ * to increment the sequence variables because an interrupt routine could
+ * change the state of the data.
+ *
+ * Based on x86_64 vsyscall gettimeofday 
+ * by Keith Owens and Andrea Arcangeli
+ */
+
+#include <lwk/spinlock.h>
+
+typedef struct {
+       unsigned sequence;
+       spinlock_t lock;
+} seqlock_t;
+
+/*
+ * These macros triggered gcc-3.x compile-time problems.  We think these are
+ * OK now.  Be cautious.
+ */
+#define SEQLOCK_UNLOCKED { 0, SPIN_LOCK_UNLOCKED }
+#define seqlock_init(x)        do { *(x) = (seqlock_t) SEQLOCK_UNLOCKED; } while (0)
+
+
+/* Lock out other writers and update the count.
+ * Acts like a normal spin_lock/unlock.
+ */
+static inline void write_seqlock(seqlock_t *sl)
+{
+       spin_lock(&sl->lock);
+       ++sl->sequence;
+       smp_wmb();                      
+}      
+
+static inline void write_sequnlock(seqlock_t *sl) 
+{
+       smp_wmb();
+       sl->sequence++;
+       spin_unlock(&sl->lock);
+}
+
+static inline int write_tryseqlock(seqlock_t *sl)
+{
+       int ret = spin_trylock(&sl->lock);
+
+       if (ret) {
+               ++sl->sequence;
+               smp_wmb();                      
+       }
+       return ret;
+}
+
+/* Start of read calculation -- fetch last complete writer token */
+static __always_inline unsigned read_seqbegin(const seqlock_t *sl)
+{
+       unsigned ret = sl->sequence;
+       smp_rmb();
+       return ret;
+}
+
+/* Test if reader processed invalid data.
+ * If initial values is odd, 
+ *     then writer had already started when section was entered
+ * If sequence value changed
+ *     then writer changed data while in section
+ *    
+ * Using xor saves one conditional branch.
+ */
+static __always_inline int read_seqretry(const seqlock_t *sl, unsigned iv)
+{
+       smp_rmb();
+       return (iv & 1) | (sl->sequence ^ iv);
+}
+
+
+/*
+ * Version using sequence counter only.
+ * This can be used when code has its own mutex protecting the
+ * updating starting before the write_seqcountbeqin() and ending
+ * after the write_seqcount_end().
+ */
+
+typedef struct seqcount {
+       unsigned sequence;
+} seqcount_t;
+
+#define SEQCNT_ZERO { 0 }
+#define seqcount_init(x)       do { *(x) = (seqcount_t) SEQCNT_ZERO; } while (0)
+
+/* Start of read using pointer to a sequence counter only.  */
+static inline unsigned read_seqcount_begin(const seqcount_t *s)
+{
+       unsigned ret = s->sequence;
+       smp_rmb();
+       return ret;
+}
+
+/* Test if reader processed invalid data.
+ * Equivalent to: iv is odd or sequence number has changed.
+ *                (iv & 1) || (*s != iv)
+ * Using xor saves one conditional branch.
+ */
+static inline int read_seqcount_retry(const seqcount_t *s, unsigned iv)
+{
+       smp_rmb();
+       return (iv & 1) | (s->sequence ^ iv);
+}
+
+
+/*
+ * Sequence counter only version assumes that callers are using their
+ * own mutexing.
+ */
+static inline void write_seqcount_begin(seqcount_t *s)
+{
+       s->sequence++;
+       smp_wmb();
+}
+
+static inline void write_seqcount_end(seqcount_t *s)
+{
+       smp_wmb();
+       s->sequence++;
+}
+
+/*
+ * Possible sw/hw IRQ protected versions of the interfaces.
+ */
+#define write_seqlock_irqsave(lock, flags)                             \
+       do { local_irq_save(flags); write_seqlock(lock); } while (0)
+#define write_seqlock_irq(lock)                                                \
+       do { local_irq_disable();   write_seqlock(lock); } while (0)
+#define write_seqlock_bh(lock)                                         \
+        do { local_bh_disable();    write_seqlock(lock); } while (0)
+
+#define write_sequnlock_irqrestore(lock, flags)                                \
+       do { write_sequnlock(lock); local_irq_restore(flags); } while(0)
+#define write_sequnlock_irq(lock)                                      \
+       do { write_sequnlock(lock); local_irq_enable(); } while(0)
+#define write_sequnlock_bh(lock)                                       \
+       do { write_sequnlock(lock); local_bh_enable(); } while(0)
+
+#define read_seqbegin_irqsave(lock, flags)                             \
+       ({ local_irq_save(flags);   read_seqbegin(lock); })
+
+#define read_seqretry_irqrestore(lock, iv, flags)                      \
+       ({                                                              \
+               int ret = read_seqretry(lock, iv);                      \
+               local_irq_restore(flags);                               \
+               ret;                                                    \
+       })
+
+#endif /* _LWK_SEQLOCK_H */