Changeset 7768b8d for libcfa/src/concurrency/kernel_private.hfa

Timestamp:

Nov 26, 2019, 3:19:20 PM (5 years ago)

Author:

Thierry Delisle <tdelisle@…>

Branches:

ADT, arm-eh, ast-experimental, enum, forall-pointer-decay, jacob/cs343-translation, master, new-ast, new-ast-unique-expr, pthread-emulation, qualifiedEnum

Children:

30763fd

Parents:

21184e3

Message:

First step at adding the new ready queue to Cforall

File:

• libcfa/src/concurrency/kernel_private.hfa (modified) (2 diffs)

libcfa/src/concurrency/kernel_private.hfa

@@ -99 +99 @@
 //-----------------------------------------------------------------------------
 // Utils
-#define KERNEL_STORAGE(T,X) static char storage_##X[sizeof(T)]
+#define KERNEL_STORAGE(T,X) __attribute((aligned(__alignof__(T)))) static char storage_##X[sizeof(T)]
 
 static inline uint32_t tls_rand() {
@@ -115 +115 @@
 void unregister( struct cluster * cltr, struct thread_desc & thrd );
 
-void doregister( struct cluster * cltr, struct processor * proc );
-void unregister( struct cluster * cltr, struct processor * proc );
+//=======================================================================
+// Cluster lock API
+//=======================================================================
+struct __attribute__((aligned(64))) __processor_id {
+        processor * volatile handle;
+        volatile bool lock;
+};
+
+// Lock-Free registering/unregistering of threads
+// Register a processor to a given cluster and get its unique id in return
+unsigned doregister( struct cluster * cltr, struct processor * proc );
+
+// Unregister a processor from a given cluster using its id, getting back the original pointer
+void     unregister( struct cluster * cltr, struct processor * proc );
+
+//=======================================================================
+// Reader-writer lock implementation
+// Concurrent with doregister/unregister,
+//    i.e., threads can be added at any point during or between the entry/exit
+static inline void __atomic_acquire(volatile bool * ll) {
+        while( __builtin_expect(__atomic_exchange_n(ll, (bool)true, __ATOMIC_SEQ_CST), false) ) {
+                while(__atomic_load_n(ll, (int)__ATOMIC_RELAXED))
+                        asm volatile("pause");
+        }
+        /* paranoid */ verify(*ll);
+}
+
+static inline bool __atomic_try_acquire(volatile bool * ll) {
+        return __atomic_exchange_n(ll, (bool)true, __ATOMIC_SEQ_CST);
+}
+
+static inline void __atomic_unlock(volatile bool * ll) {
+        /* paranoid */ verify(*ll);
+        __atomic_store_n(ll, (bool)false, __ATOMIC_RELEASE);
+}
+
+//-----------------------------------------------------------------------
+// Reader side : acquire when using the ready queue to schedule but not
+//  creating/destroying queues
+static inline void ready_schedule_lock( struct cluster & cltr, struct processor * proc) with(cltr.ready_lock) {
+        unsigned iproc = proc->id;
+        /*paranoid*/ verify(data[iproc].handle == proc);
+        /*paranoid*/ verify(iproc < ready);
+
+        // Step 1 : make sure no writer are in the middle of the critical section
+        while(__atomic_load_n(&lock, (int)__ATOMIC_RELAXED))
+                asm volatile("pause");
+
+        // Fence needed because we don't want to start trying to acquire the lock
+        // before we read a false.
+        // Not needed on x86
+        // std::atomic_thread_fence(std::memory_order_seq_cst);
+
+        // Step 2 : acquire our local lock
+        __atomic_acquire( &data[iproc].lock );
+        /*paranoid*/ verify(data[iproc].lock);
+}
+
+static inline void ready_schedule_unlock( struct cluster & cltr, struct processor * proc) with(cltr.ready_lock) {
+        unsigned iproc = proc->id;
+        /*paranoid*/ verify(data[iproc].handle == proc);
+        /*paranoid*/ verify(iproc < ready);
+        /*paranoid*/ verify(data[iproc].lock);
+        __atomic_store_n(&data[iproc].lock, false, __ATOMIC_RELEASE);
+}
+
+//-----------------------------------------------------------------------
+// Writer side : acquire when changing the ready queue, e.g. adding more
+//  queues or removing them.
+uint_fast32_t ready_mutate_lock( struct cluster & cltr );
+
+void ready_mutate_unlock( struct cluster & cltr, uint_fast32_t );
+
+bool push(__intrusive_ready_queue_t & this, thread_desc * node);
+[thread_desc *, bool] pop(__intrusive_ready_queue_t & this);
 
 // Local Variables: //