Changes in libcfa/src/concurrency/kernel.hfa [e3fea42:dca5802]

File:

• libcfa/src/concurrency/kernel.hfa (modified) (6 diffs)

libcfa/src/concurrency/kernel.hfa

@@ -10 +10 @@
 // Created On       : Tue Jan 17 12:27:26 2017
 // Last Modified By : Peter A. Buhr
-// Last Modified On : Tue Feb  4 12:29:26 2020
-// Update Count     : 22
+// Last Modified On : Wed Dec  4 07:54:51 2019
+// Update Count     : 18
 //
 
@@ -107 +107 @@
         // Cluster from which to get threads
         struct cluster * cltr;
+        unsigned int id;
 
         // Name of the processor
@@ -150 +151 @@
 };
 
-void  ?{}(processor & this, const char name[], struct cluster & cltr);
+void  ?{}(processor & this, const char * name, struct cluster & cltr);
 void ^?{}(processor & this);
 
 static inline void  ?{}(processor & this)                    { this{ "Anonymous Processor", *mainCluster}; }
 static inline void  ?{}(processor & this, struct cluster & cltr)    { this{ "Anonymous Processor", cltr}; }
-static inline void  ?{}(processor & this, const char name[]) { this{name, *mainCluster }; }
+static inline void  ?{}(processor & this, const char * name) { this{name, *mainCluster }; }
 
 static inline [processor *&, processor *& ] __get( processor & this ) {
         return this.node.[next, prev];
 }
+
+
+//-----------------------------------------------------------------------------
+// Cluster Tools
+
+// Cells use by the reader writer lock
+// while not generic it only relies on a opaque pointer
+struct __processor_id;
+
+// Reader-Writer lock protecting the ready-queue
+// while this lock is mostly generic some aspects
+// have been hard-coded to for the ready-queue for
+// simplicity and performance
+struct __clusterRWLock_t {
+        // total cachelines allocated
+        unsigned int max;
+
+        // cachelines currently in use
+        volatile unsigned int alloc;
+
+        // cachelines ready to itereate over
+        // (!= to alloc when thread is in second half of doregister)
+        volatile unsigned int ready;
+
+        // writer lock
+        volatile bool lock;
+
+        // data pointer
+        __processor_id * data;
+};
+
+void  ?{}(__clusterRWLock_t & this);
+void ^?{}(__clusterRWLock_t & this);
+
+// Intrusives lanes which are used by the relaxed ready queue
+struct __attribute__((aligned(128))) __intrusive_lane_t {
+        // spin lock protecting the queue
+        volatile bool lock;
+
+        // anchor for the head and the tail of the queue
+        struct __sentinel_t {
+                // Link lists fields
+                // instrusive link field for threads
+                // must be exactly as in thread_desc
+                __thread_desc_link link;
+        } before, after;
+
+#if defined(__CFA_WITH_VERIFY__)
+        // id of last processor to acquire the lock
+        // needed only to check for mutual exclusion violations
+        unsigned int last_id;
+
+        // number of items on this list
+        // needed only to check for deadlocks
+        unsigned int count;
+#endif
+
+        // Optional statistic counters
+        #if !defined(__CFA_NO_SCHED_STATS__)
+                struct __attribute__((aligned(64))) {
+                        // difference between number of push and pops
+                        ssize_t diff;
+
+                        // total number of pushes and pops
+                        size_t  push;
+                        size_t  pop ;
+                } stat;
+        #endif
+};
+
+void  ?{}(__intrusive_lane_t & this);
+void ^?{}(__intrusive_lane_t & this);
+
+typedef unsigned long long __cfa_readyQ_mask_t;
+
+// enum {
+//      __cfa_ready_queue_mask_size = (64 - sizeof(size_t)) / sizeof(size_t),
+//      __cfa_max_ready_queues = __cfa_ready_queue_mask_size * 8 * sizeof(size_t)
+// };
+
+#define __cfa_lane_mask_size ((64 - sizeof(size_t)) / sizeof(__cfa_readyQ_mask_t))
+#define __cfa_max_lanes (__cfa_lane_mask_size * 8 * sizeof(__cfa_readyQ_mask_t))
+
+//TODO adjust cache size to ARCHITECTURE
+// Structure holding the relaxed ready queue
+struct __attribute__((aligned(128))) __ready_queue_t {
+        // Data tracking how many/which lanes are used
+        // Aligned to 128 for cache locality
+        struct {
+                // number of non-empty lanes
+                volatile size_t count;
+
+                // bit mask, set bits indentify which lanes are non-empty
+                volatile __cfa_readyQ_mask_t mask[ __cfa_lane_mask_size ];
+        } used;
+
+        // Data tracking the actual lanes
+        // On a seperate cacheline from the used struct since
+        // used can change on each push/pop but this data
+        // only changes on shrink/grow
+        struct __attribute__((aligned(64))) {
+                // Arary of lanes
+                __intrusive_lane_t * volatile data;
+
+                // Number of lanes (empty or not)
+                volatile size_t count;
+        } lanes;
+
+        // Statistics
+        #if !defined(__CFA_NO_STATISTICS__)
+                __attribute__((aligned(64))) struct {
+                        struct {
+                                // Push statistic
+                                struct {
+                                        // number of attemps at pushing something
+                                        volatile size_t attempt;
+
+                                        // number of successes at pushing
+                                        volatile size_t success;
+                                } push;
+
+                                // Pop statistic
+                                struct {
+                                        // number of reads of the mask
+                                        // picking an empty __cfa_readyQ_mask_t counts here
+                                        // but not as an attempt
+                                        volatile size_t maskrds;
+
+                                        // number of attemps at poping something
+                                        volatile size_t attempt;
+
+                                        // number of successes at poping
+                                        volatile size_t success;
+                                } pop;
+                        } pick;
+
+                        // stats on the "used" struct of the queue
+                        // tracks average number of queues that are not empty
+                        // when pushing / poping
+                        struct {
+                                volatile size_t value;
+                                volatile size_t count;
+                        } used;
+                } global_stats;
+
+        #endif
+};
+
+void  ?{}(__ready_queue_t & this);
+void ^?{}(__ready_queue_t & this);
 
 //-----------------------------------------------------------------------------
@@ -165 +316 @@
 struct cluster {
         // Ready queue locks
-        __spinlock_t ready_queue_lock;
+        __clusterRWLock_t ready_lock;
 
         // Ready queue for threads
-        __queue_t(thread_desc) ready_queue;
+        __ready_queue_t ready_queue;
 
         // Name of the cluster
@@ -178 +329 @@
         // List of processors
         __spinlock_t proc_list_lock;
-        __dllist_t(struct processor) procs;
         __dllist_t(struct processor) idles;
-        unsigned int nprocessors;
 
         // List of threads
@@ -195 +344 @@
 extern Duration default_preemption();
 
-void ?{} (cluster & this, const char name[], Duration preemption_rate);
+void ?{} (cluster & this, const char * name, Duration preemption_rate);
 void ^?{}(cluster & this);
 
 static inline void ?{} (cluster & this)                           { this{"Anonymous Cluster", default_preemption()}; }
 static inline void ?{} (cluster & this, Duration preemption_rate) { this{"Anonymous Cluster", preemption_rate}; }
-static inline void ?{} (cluster & this, const char name[])        { this{name, default_preemption()}; }
+static inline void ?{} (cluster & this, const char * name)        { this{name, default_preemption()}; }
 
 static inline [cluster *&, cluster *& ] __get( cluster & this ) {