+struct cfs_cpt_table;
+
+/*
+ * allocate per-cpu-partition data, returned value is an array of pointers,
+ * variable can be indexed by CPU ID.
+ * cptable != NULL: size of array is number of CPU partitions
+ * cptable == NULL: size of array is number of HW cores
+ */
+void *cfs_percpt_alloc(struct cfs_cpt_table *cptab, unsigned int size);
+/*
+ * destory per-cpu-partition variable
+ */
+void cfs_percpt_free(void *vars);
+int cfs_percpt_number(void *vars);
+void *cfs_percpt_current(void *vars);
+void *cfs_percpt_index(void *vars, int idx);
+
+#define cfs_percpt_for_each(var, i, vars) \
+ for (i = 0; i < cfs_percpt_number(vars) && \
+ ((var) = (vars)[i]) != NULL; i++)
+
+/*
+ * allocate a variable array, returned value is an array of pointers.
+ * Caller can specify length of array by count.
+ */
+void *cfs_array_alloc(int count, unsigned int size);
+void cfs_array_free(void *vars);
+
+#define LASSERT_ATOMIC_ENABLED (1)
+
+#if LASSERT_ATOMIC_ENABLED
+
+/** assert value of @a is equal to @v */
+#define LASSERT_ATOMIC_EQ(a, v) \
+do { \
+ LASSERTF(atomic_read(a) == v, \
+ "value: %d\n", atomic_read((a))); \
+} while (0)
+
+/** assert value of @a is unequal to @v */
+#define LASSERT_ATOMIC_NE(a, v) \
+do { \
+ LASSERTF(atomic_read(a) != v, \
+ "value: %d\n", atomic_read((a))); \
+} while (0)
+
+/** assert value of @a is little than @v */
+#define LASSERT_ATOMIC_LT(a, v) \
+do { \
+ LASSERTF(atomic_read(a) < v, \
+ "value: %d\n", atomic_read((a))); \
+} while (0)
+
+/** assert value of @a is little/equal to @v */
+#define LASSERT_ATOMIC_LE(a, v) \
+do { \
+ LASSERTF(atomic_read(a) <= v, \
+ "value: %d\n", atomic_read((a))); \
+} while (0)
+
+/** assert value of @a is great than @v */
+#define LASSERT_ATOMIC_GT(a, v) \
+do { \
+ LASSERTF(atomic_read(a) > v, \
+ "value: %d\n", atomic_read((a))); \
+} while (0)
+
+/** assert value of @a is great/equal to @v */
+#define LASSERT_ATOMIC_GE(a, v) \
+do { \
+ LASSERTF(atomic_read(a) >= v, \
+ "value: %d\n", atomic_read((a))); \
+} while (0)
+
+/** assert value of @a is great than @v1 and little than @v2 */
+#define LASSERT_ATOMIC_GT_LT(a, v1, v2) \
+do { \
+ int __v = atomic_read(a); \
+ LASSERTF(__v > v1 && __v < v2, "value: %d\n", __v); \
+} while (0)
+
+/** assert value of @a is great than @v1 and little/equal to @v2 */
+#define LASSERT_ATOMIC_GT_LE(a, v1, v2) \
+do { \
+ int __v = atomic_read(a); \
+ LASSERTF(__v > v1 && __v <= v2, "value: %d\n", __v); \
+} while (0)
+
+/** assert value of @a is great/equal to @v1 and little than @v2 */
+#define LASSERT_ATOMIC_GE_LT(a, v1, v2) \
+do { \
+ int __v = atomic_read(a); \
+ LASSERTF(__v >= v1 && __v < v2, "value: %d\n", __v); \
+} while (0)
+
+/** assert value of @a is great/equal to @v1 and little/equal to @v2 */
+#define LASSERT_ATOMIC_GE_LE(a, v1, v2) \
+do { \
+ int __v = atomic_read(a); \
+ LASSERTF(__v >= v1 && __v <= v2, "value: %d\n", __v); \
+} while (0)
+
+#else /* !LASSERT_ATOMIC_ENABLED */
+
+#define LASSERT_ATOMIC_EQ(a, v) do {} while (0)
+#define LASSERT_ATOMIC_NE(a, v) do {} while (0)
+#define LASSERT_ATOMIC_LT(a, v) do {} while (0)
+#define LASSERT_ATOMIC_LE(a, v) do {} while (0)
+#define LASSERT_ATOMIC_GT(a, v) do {} while (0)
+#define LASSERT_ATOMIC_GE(a, v) do {} while (0)
+#define LASSERT_ATOMIC_GT_LT(a, v1, v2) do {} while (0)
+#define LASSERT_ATOMIC_GT_LE(a, v1, v2) do {} while (0)
+#define LASSERT_ATOMIC_GE_LT(a, v1, v2) do {} while (0)
+#define LASSERT_ATOMIC_GE_LE(a, v1, v2) do {} while (0)
+
+#endif /* LASSERT_ATOMIC_ENABLED */
+
+#define LASSERT_ATOMIC_ZERO(a) LASSERT_ATOMIC_EQ(a, 0)
+#define LASSERT_ATOMIC_POS(a) LASSERT_ATOMIC_GT(a, 0)
+
+#define CFS_ALLOC_PTR(ptr) LIBCFS_ALLOC(ptr, sizeof (*(ptr)));
+#define CFS_FREE_PTR(ptr) LIBCFS_FREE(ptr, sizeof (*(ptr)));
+
+/*
+ * percpu partition lock
+ *
+ * There are some use-cases like this in Lustre:
+ * . each CPU partition has it's own private data which is frequently changed,
+ * and mostly by the local CPU partition.
+ * . all CPU partitions share some global data, these data are rarely changed.
+ *
+ * LNet is typical example.
+ * CPU partition lock is designed for this kind of use-cases:
+ * . each CPU partition has it's own private lock
+ * . change on private data just needs to take the private lock
+ * . read on shared data just needs to take _any_ of private locks
+ * . change on shared data needs to take _all_ private locks,
+ * which is slow and should be really rare.
+ */
+
+enum {
+ CFS_PERCPT_LOCK_EX = -1, /* negative */
+};
+
+#ifdef __KERNEL__
+
+struct cfs_percpt_lock {
+ /* cpu-partition-table for this lock */
+ struct cfs_cpt_table *pcl_cptab;
+ /* exclusively locked */
+ unsigned int pcl_locked;
+ /* private lock table */
+ spinlock_t **pcl_locks;
+};
+
+/* return number of private locks */
+#define cfs_percpt_lock_num(pcl) cfs_cpt_number(pcl->pcl_cptab)
+
+#else /* !__KERNEL__ */
+
+# ifdef HAVE_LIBPTHREAD
+
+struct cfs_percpt_lock {
+ pthread_mutex_t pcl_mutex;
+};
+
+# else /* !HAVE_LIBPTHREAD */
+
+struct cfs_percpt_lock {
+ int pcl_lock;
+};
+
+static const struct cfs_percpt_lock CFS_PERCPT_LOCK_MAGIC;
+
+# endif /* HAVE_LIBPTHREAD */
+# define cfs_percpt_lock_num(pcl) 1
+#endif /* __KERNEL__ */
+
+/*
+ * create a cpu-partition lock based on CPU partition table \a cptab,
+ * each private lock has extra \a psize bytes padding data
+ */
+struct cfs_percpt_lock *cfs_percpt_lock_alloc(struct cfs_cpt_table *cptab);
+/* destroy a cpu-partition lock */
+void cfs_percpt_lock_free(struct cfs_percpt_lock *pcl);
+
+/* lock private lock \a index of \a pcl */
+void cfs_percpt_lock(struct cfs_percpt_lock *pcl, int index);
+/* unlock private lock \a index of \a pcl */
+void cfs_percpt_unlock(struct cfs_percpt_lock *pcl, int index);
+/* create percpt (atomic) refcount based on @cptab */
+atomic_t **cfs_percpt_atomic_alloc(struct cfs_cpt_table *cptab, int val);
+/* destroy percpt refcount */
+void cfs_percpt_atomic_free(atomic_t **refs);
+/* return sum of all percpu refs */
+int cfs_percpt_atomic_summary(atomic_t **refs);
+
+
+/** Compile-time assertion.