spin-wait based kernel timer (for futex() timeout) implementation

kernel/futex.c

@@ -242,11 +242,14 @@ static int futex_wait(uint32_t __user *uaddr, uint32_t val,
 	queue_unlock(queue, irqflags);
 	
 	if (!list_empty(&futex.link)) {
-		// We don't have timers for now, let's sleep forever,
-		// and pretend we were woken up
-		//time_remain = schedule_timeout(timeout);
-		schedule();
-		time_remain = 10;
+		
+		if (timeout) {
+			time_remain = schedule_timeout(timeout);
+		}
+		else {
+			schedule();
+			time_remain = 0;
+		}
 	}
 
 	cpu_local_var(current)->status = PS_RUNNING;

kernel/include/process.h

@@ -22,6 +22,9 @@
 
 #define USE_LARGE_PAGES
 
+#include <waitq.h>
+#include <futex.h>
+
 struct vm_range {
 	struct list_head list;
 	unsigned long start, end;
@@ -52,15 +55,15 @@ struct process {
 	// Runqueue list entry
 	struct list_head sched_list;  
 	
+	aal_spinlock_t spin_sleep_lock;
+	int spin_sleep;
+
 	struct thread {
 		int	*clear_child_tid;
 		unsigned long tlsblock_base, tlsblock_limit;
 	} thread;
 };
 
-#include <waitq.h>
-#include <futex.h>
-
 struct process_vm {
 	aal_atomic_t refcount;
 

kernel/include/timer.h

@@ -18,7 +18,7 @@
 #include <affinity.h>
 #include <time.h>
 
-#define TIMER_CPU_ID	227
+//#define TIMER_CPU_ID	227
 
 struct timer {
 	uint64_t timeout;

kernel/process.c

@@ -58,6 +58,9 @@ struct process *create_process(unsigned long user_pc)
 
 	init_process_vm(proc->vm);
 
+	aal_mc_spinlock_init(&proc->spin_sleep_lock);
+	proc->spin_sleep = 0;
+
 	return proc;
 }
 
@@ -434,10 +437,12 @@ void sched_init(void)
 	cpu_local_var(runq_len) = 0;
 	aal_mc_spinlock_init(&cpu_local_var(runq_lock));
 
+#ifdef TIMER_CPU_ID
 	if (aal_mc_get_processor_id() == TIMER_CPU_ID) {
 		init_timers();
 		wake_timers_loop();
 	}
+#endif
 }
 
 void schedule(void)
@@ -514,8 +519,23 @@ void schedule(void)
 int sched_wakeup_process(struct process *proc, int valid_states)
 {
 	int status;
+	int spin_slept = 0;
 	unsigned long irqstate;
 	struct cpu_local_var *v = get_cpu_local_var(proc->cpu_id);
+	
+	irqstate = aal_mc_spinlock_lock(&(proc->spin_sleep_lock));
+	if (proc->spin_sleep) {
+		dkprintf("sched_wakeup_process() spin wakeup: cpu_id: %d\n", 
+				proc->cpu_id);
+
+		spin_slept = 1;
+		proc->spin_sleep = 0;
+		status = 0;	
+	}
+	aal_mc_spinlock_unlock(&(proc->spin_sleep_lock), irqstate);
+	
+	if (spin_slept)
+		return status;
 
 	irqstate = aal_mc_spinlock_lock(&(v->runq_lock));
 	

kernel/syscall.c

@@ -846,8 +846,6 @@ SYSCALL_DECLARE(madvise)
 
 SYSCALL_DECLARE(futex)
 {
-	// TODO: timespec support!
-	//struct timespec _utime;
 	uint64_t timeout = 0; // No timeout
 	uint32_t val2 = 0;
 
@@ -857,20 +855,47 @@ SYSCALL_DECLARE(futex)
 	struct timespec *utime = (struct timespec*)aal_mc_syscall_arg3(ctx);
 	uint32_t *uaddr2 = (uint32_t *)aal_mc_syscall_arg4(ctx);
 	uint32_t val3 = (uint32_t)aal_mc_syscall_arg5(ctx);
+    
+	dkprintf("futex,uaddr=%lx,op=%x, val=%x, utime=%lx, uaddr2=%lx, val3=%x, []=%x\n", (unsigned long)uaddr, op, val, utime, uaddr2, val3, *uaddr);
 
 	/* Mask off the FUTEX_PRIVATE_FLAG,
 	 * assume all futexes are address space private */
 	op = (op & FUTEX_CMD_MASK);
 
-#if 0 
-	if (utime && (op == FUTEX_WAIT)) {
-		if (copy_from_user(&_utime, utime, sizeof(_utime)) != 0)
+	if (utime && (op == FUTEX_WAIT_BITSET || op == FUTEX_WAIT)) {
+		/* gettimeofday(&tv_now, NULL) from host */
+		struct syscall_request request AAL_DMA_ALIGN; 
+		struct timeval tv_now;
+		request.number = 96;
+		unsigned long __phys;                                          
+
+		dkprintf("futex,utime and FUTEX_WAIT_*, uaddr=%lx, []=%x\n", (unsigned long)uaddr, *uaddr);
+
+		if (aal_mc_pt_virt_to_phys(cpu_local_var(current)->vm->page_table, 
+					(void *)&tv_now, &__phys)) { 
+			return -EFAULT; 
+		}
+
+		request.args[0] = __phys;               
+
+		int r = do_syscall(&request, ctx);
+
+		if (r < 0) {
 			return -EFAULT;
-		if (!timespec_valid(&_utime))
-			return -EINVAL;
-		timeout = timespec_to_ns(_utime);
+		}
+
+		dkprintf("futex, FUTEX_WAIT_*, arg3 != NULL, pc=%lx\n", (unsigned long)aal_mc_syscall_pc(ctx));
+		dkprintf("now->tv_sec=%016ld,tv_nsec=%016ld\n", tv_now.tv_sec, tv_now.tv_usec * 1000);
+		dkprintf("utime->tv_sec=%016ld,tv_nsec=%016ld\n", utime->tv_sec, utime->tv_nsec);
+
+		long nsec_now = ((long)tv_now.tv_sec * 1000000000ULL) + 
+			tv_now.tv_usec * 1000;
+		long nsec_timeout = ((long)utime->tv_sec * 1000000000ULL) + 
+			utime->tv_nsec * 1;
+		long diff_nsec = nsec_timeout - nsec_now;
+
+		timeout = (diff_nsec / 1000) * 1100; // (usec * 1.1GHz)
 	}
-#endif
 
 	/* Requeue parameter in 'utime' if op == FUTEX_CMP_REQUEUE.
 	 * number of waiters to wake in 'utime' if op == FUTEX_WAKE_OP. */

kernel/timer.c

@@ -30,7 +30,7 @@
 #define dkprintf(...)
 #endif
 
-#define LOOP_TIMEOUT 1000 /* default 1000 */
+#define LOOP_TIMEOUT 500
 
 struct list_head timers;
 aal_spinlock_t timers_lock;
@@ -47,6 +47,56 @@ uint64_t schedule_timeout(uint64_t timeout)
 	struct waitq_entry my_wait;
 	struct timer my_timer;
 	unsigned long irqflags;
+	struct process *proc = cpu_local_var(current);
+
+	irqflags = aal_mc_spinlock_lock(&proc->spin_sleep_lock);
+	dkprintf("schedule_timeout() spin sleep timeout: %lu\n", timeout);
+	proc->spin_sleep = 1;
+	aal_mc_spinlock_unlock(&proc->spin_sleep_lock, irqflags);
+
+	/* Spin sleep.. */
+	for (;;) {
+		uint64_t t_s = rdtsc();
+		uint64_t t_e;
+		int spin_over = 0;
+
+		irqflags = aal_mc_spinlock_lock(&proc->spin_sleep_lock);
+		
+		/* Woken up by someone? */
+		if (!proc->spin_sleep) {
+			t_e = rdtsc();
+
+			spin_over = 1;
+			if ((t_e - t_s) < timeout) {
+				timeout -= (t_e - t_s);
+			}
+			else {
+				timeout = 1;
+			}
+		}
+		
+		aal_mc_spinlock_unlock(&proc->spin_sleep_lock, irqflags);
+
+		t_s = rdtsc();
+		
+		while ((rdtsc() - t_s) < LOOP_TIMEOUT) {
+			cpu_pause();
+		}
+
+		if (timeout < LOOP_TIMEOUT) {
+			timeout = 0;
+			spin_over = 1;
+		}
+		else {
+			timeout -= LOOP_TIMEOUT;
+		}
+
+		if (spin_over) {
+			dkprintf("schedule_timeout() spin woken up, timeout: %lu\n", 
+					timeout);
+			return timeout;
+		}
+	}
 
 	/* Init waitq and wait entry for this timer */
 	my_timer.timeout = (timeout < LOOP_TIMEOUT) ? LOOP_TIMEOUT : timeout;
@@ -72,6 +122,7 @@ uint64_t schedule_timeout(uint64_t timeout)
 	if (my_timer.timeout) {
 		list_del(&my_timer.list);
 	}
+	
 	aal_mc_spinlock_unlock(&timers_lock, irqflags);
 
 	dkprintf("schedule_timeout() woken up, timeout: %lu\n",