PAPI support: performance counter's overflow.

and support mckfd fcntl.

arch/x86/kernel/cpu.c

@@ -883,13 +883,36 @@ void handle_interrupt(int vector, struct x86_user_context *regs)
 		dkprintf("timer[%lu]: CPU_FLAG_NEED_RESCHED \n", rdtsc());
 	}
 	else if (vector == LOCAL_PERF_VECTOR) {
+		struct siginfo info;
 		unsigned long value;
+		struct thread *thread = cpu_local_var(current);
+        	struct process *proc = thread->proc;
+		long irqstate;
+		struct mckfd *fdp;
+
+		lapic_write(LAPIC_LVTPC, LOCAL_PERF_VECTOR);
 
 		value = rdmsr(MSR_PERF_GLOBAL_STATUS);
 		wrmsr(MSR_PERF_GLOBAL_OVF_CTRL, value);
 		wrmsr(MSR_PERF_GLOBAL_OVF_CTRL, 0);
-		//TODO: counter overflow signal
+
-		//set_signal(0x1d, regs, NULL); // SIGIO
+		irqstate = ihk_mc_spinlock_lock(&proc->mckfd_lock);
+	        for(fdp = proc->mckfd; fdp; fdp = fdp->next) {
+			if(fdp->sig_no > 0)
+                	        break;
+		}
+	        ihk_mc_spinlock_unlock(&proc->mckfd_lock, irqstate);
+
+		if(fdp) {
+			memset(&info, '\0', sizeof info);
+			info.si_signo = fdp->sig_no;
+			info._sifields._sigfault.si_addr = (void *)regs->gpr.rip;
+			info._sifields._sigpoll.si_fd = fdp->fd;
+			set_signal(fdp->sig_no, regs, &info); 
+		}
+		else {
+			set_signal(SIGIO, regs, NULL);
+		}
 	}
 	else if (vector >= IHK_TLB_FLUSH_IRQ_VECTOR_START && 
 	         vector < IHK_TLB_FLUSH_IRQ_VECTOR_END) {

arch/x86/kernel/include/syscall_list.h

@@ -66,7 +66,7 @@ SYSCALL_DELEGATED(65, semop)
 SYSCALL_HANDLED(67, shmdt)
 SYSCALL_DELEGATED(69, msgsnd)
 SYSCALL_DELEGATED(70, msgrcv)
-SYSCALL_DELEGATED(72, fcntl)
+SYSCALL_HANDLED(72, fcntl)
 SYSCALL_DELEGATED(79, getcwd)
 SYSCALL_DELEGATED(89, readlink)
 SYSCALL_HANDLED(96, gettimeofday)

arch/x86/kernel/perfctr.c

@@ -12,6 +12,7 @@
 #include <errno.h>
 #include <ihk/debug.h>
 #include <registers.h>
+#include <mc_perf_event.h>
 
 extern unsigned int *x86_march_perfmap;
 extern int running_on_kvm(void);
@@ -26,6 +27,7 @@ unsigned long X86_IA32_FIXED_PERF_COUNTERS_MASK = 0;
 
 void x86_init_perfctr(void)
 {
+	int i = 0;
 	unsigned long reg;
 	unsigned long value = 0;
 	uint64_t op;
@@ -61,6 +63,16 @@ void x86_init_perfctr(void)
 				X86_IA32_NUM_PERF_COUNTERS, X86_IA32_NUM_FIXED_PERF_COUNTERS);
 	}
 
+	/* Clear Fixed Counter Control */
+	value = rdmsr(MSR_PERF_FIXED_CTRL);
+	value &= 0xfffffffffffff000L;
+	wrmsr(MSR_PERF_FIXED_CTRL, value);
+
+	/* Clear Generic Counter Control */
+	for(i = 0; i < X86_IA32_NUM_PERF_COUNTERS; i++) {
+		wrmsr(MSR_IA32_PERFEVTSEL0 + i, 0);
+	}
+		
 	/* Enable PMC Control */
 	value = rdmsr(MSR_PERF_GLOBAL_CTRL);
 	value |= X86_IA32_PERF_COUNTERS_MASK;
@@ -98,7 +110,7 @@ static int set_perfctr_x86_direct(int counter, int mode, unsigned int value)
 	return 0;
 }
 
-static int set_pmc_x86_direct(int counter, unsigned long val)
+static int set_pmc_x86_direct(int counter, long val)
 {
 	unsigned long cnt_bit = 0;
 
@@ -106,6 +118,8 @@ static int set_pmc_x86_direct(int counter, unsigned long val)
 		return -EINVAL;
 	}
 
+	val &= 0x000000ffffffffff; // 40bit Mask
+
 	cnt_bit = 1UL << counter;
 	if ( cnt_bit & X86_IA32_PERF_COUNTERS_MASK ) {
 		// set generic pmc
@@ -132,7 +146,7 @@ static int set_perfctr_x86(int counter, int event, int mask, int inv, int count,
 static int set_fixed_counter(int counter, int mode)
 {
 	unsigned long value = 0;
-	unsigned int  ctr_mask = 0x7;
+	unsigned int  ctr_mask = 0xf;
 	int counter_idx = counter - X86_IA32_BASE_FIXED_PERF_COUNTERS ;
 	unsigned int  set_val = 0;
 
@@ -213,6 +227,24 @@ int ihk_mc_perfctr_stop(unsigned long counter_mask)
 	value &= ~counter_mask;
 	wrmsr(MSR_PERF_GLOBAL_CTRL, value);
 
+	if(counter_mask >> 32 & 0x1) {
+		value = rdmsr(MSR_PERF_FIXED_CTRL);
+		value &= ~(0xf);
+		wrmsr(MSR_PERF_FIXED_CTRL, value);
+	}
+		
+	if(counter_mask >> 32 & 0x2) {
+		value = rdmsr(MSR_PERF_FIXED_CTRL);
+		value &= ~(0xf << 4);
+		wrmsr(MSR_PERF_FIXED_CTRL, value);
+	}
+
+	if(counter_mask >> 32 & 0x4) {
+		value = rdmsr(MSR_PERF_FIXED_CTRL);
+		value &= ~(0xf << 8);
+		wrmsr(MSR_PERF_FIXED_CTRL, value);
+	}
+
 	return 0;
 }
 
@@ -220,7 +252,7 @@ int ihk_mc_perfctr_stop(unsigned long counter_mask)
 int ihk_mc_perfctr_fixed_init(int counter, int mode)
 {
 	unsigned long value = 0;
-	unsigned int  ctr_mask = 0x7;
+	unsigned int  ctr_mask = 0xf;
 	int counter_idx = counter - X86_IA32_BASE_FIXED_PERF_COUNTERS ;
 	unsigned int  set_val = 0;
 
@@ -240,6 +272,9 @@ int ihk_mc_perfctr_fixed_init(int counter, int mode)
 		set_val |= 1;
 	}
 
+	// enable PMI on overflow
+	set_val |= 1 << 3;
+
 	set_val <<= counter_idx * 4;
 	value |= set_val;
 
@@ -253,7 +288,7 @@ int ihk_mc_perfctr_reset(int counter)
 	return set_pmc_x86_direct(counter, 0);
 }
 
-int ihk_mc_perfctr_set(int counter, unsigned long val)
+int ihk_mc_perfctr_set(int counter, long val)
 {
 	return set_pmc_x86_direct(counter, val);
 }
@@ -327,23 +362,33 @@ unsigned long ihk_mc_perfctr_read_msr(int counter)
 	return retval;
 }
 
-int ihk_mc_perfctr_alloc_counter(unsigned long pmc_status)
+int ihk_mc_perfctr_alloc_counter(unsigned int *type, unsigned long *config, unsigned long pmc_status)
 {
+	int ret = -1;
 	int i = 0;
-        int ret = -1;
 
-        // find avail generic counter
+	if(*type == PERF_TYPE_HARDWARE) {
-        for(i = 0; i < X86_IA32_NUM_PERF_COUNTERS; i++) {
+		switch(*config){
+		case PERF_COUNT_HW_INSTRUCTIONS :
+			*type = PERF_TYPE_RAW;
+			*config = 0x5300c0;
+			break;
+		default :
+			// Unexpected config
+			return -1;
+		}
+	}
+	else if(*type != PERF_TYPE_RAW) {
+		return -1;
+	}
+
+	// find avail generic counter
+       	for(i = 0; i < X86_IA32_NUM_PERF_COUNTERS; i++) {
 		if(!(pmc_status & (1 << i))) {
 			ret = i;
-			pmc_status |= (1 << i);
 			break;
 		}
 	}
 
-	if(ret < 0){
+	return ret;
-		return ret;
-	}
-
-        return ret;
 }

kernel/include/process.h

@@ -319,12 +319,14 @@ struct process_vm;
 struct mckfd {
 	struct mckfd *next;
 	int fd;
+	int sig_no;
 	long data;
 	void *opt;
 	long (*read_cb)(struct mckfd *, ihk_mc_user_context_t *);
 	int (*ioctl_cb)(struct mckfd *, ihk_mc_user_context_t *);
 	long (*mmap_cb)(struct mckfd *, ihk_mc_user_context_t *);
 	int (*close_cb)(struct mckfd *, ihk_mc_user_context_t *);
+	int (*fcntl_cb)(struct mckfd *, ihk_mc_user_context_t *);
 };
 
 #define SFD_CLOEXEC 02000000

kernel/syscall.c

@@ -2470,6 +2470,31 @@ SYSCALL_DECLARE(close)
 	return rc;
 }
 
+SYSCALL_DECLARE(fcntl)
+{
+	int fd = ihk_mc_syscall_arg0(ctx);
+	// int cmd = ihk_mc_syscall_arg1(ctx);
+	long rc;
+	struct thread *thread = cpu_local_var(current);
+	struct process *proc = thread->proc;
+	struct mckfd *fdp;
+	long irqstate;
+
+	irqstate = ihk_mc_spinlock_lock(&proc->mckfd_lock);
+	for(fdp = proc->mckfd; fdp; fdp = fdp->next)
+		if(fdp->fd == fd)
+			break;
+	ihk_mc_spinlock_unlock(&proc->mckfd_lock, irqstate);
+
+	if(fdp && fdp->fcntl_cb){
+		rc = fdp->fcntl_cb(fdp, ctx);
+	}
+	else{
+		rc = syscall_generic_forwarding(__NR_fcntl, ctx);
+	}
+	return rc;
+}
+
 SYSCALL_DECLARE(rt_sigprocmask)
 {
 	int how = ihk_mc_syscall_arg0(ctx);
@@ -2621,26 +2646,12 @@ perf_counter_alloc(struct mc_perf_event *event)
 	struct perf_event_attr *attr = &event->attr;
 	struct mc_perf_event *leader = event->group_leader;
 
-	if(attr->type == PERF_TYPE_HARDWARE) {
+	ret = ihk_mc_perfctr_alloc_counter(&attr->type, &attr->config, leader->pmc_status); 
-
-		event->counter_id = ihk_mc_perfctr_alloc_counter(leader->pmc_status); 
-	
-	} else if(attr->type == PERF_TYPE_RAW) {
-		// PAPI_REF_CYC counted by fixed counter
-		if((attr->config & 0x0000ffff) == 0x00000300) {
-			event->counter_id = 2 + X86_IA32_BASE_FIXED_PERF_COUNTERS;
-			return ret;
-		}
-
-		event->counter_id = ihk_mc_perfctr_alloc_counter(leader->pmc_status); 
-	} else {
-		// Not supported type.
-		ret = -1;
-	}
 
 	if(ret >= 0) {
-		leader->pmc_status |= 1UL << event->counter_id;
+		leader->pmc_status |= 1UL << ret;
 	}
+	event->counter_id = ret;
 
 	return ret;
 }
@@ -2649,7 +2660,6 @@ int
 perf_counter_start(struct mc_perf_event *event)
 {
 	int ret = 0;
-	enum ihk_perfctr_type type;
 	struct perf_event_attr *attr = &event->attr;
 	int mode = 0x00;
 
@@ -2660,37 +2670,16 @@ perf_counter_start(struct mc_perf_event *event)
 		mode |= PERFCTR_USER_MODE;
 	}
 
-	if(attr->type == PERF_TYPE_HARDWARE) {
+	if(event->counter_id >= 0 && event->counter_id < X86_IA32_NUM_PERF_COUNTERS) {
-		switch(attr->config){
-		case PERF_COUNT_HW_CPU_CYCLES :
-			type = APT_TYPE_CYCLE;
-			break;
-		case PERF_COUNT_HW_INSTRUCTIONS :
-			type = APT_TYPE_INSTRUCTIONS;
-			break;
-		default :
-			// Not supported config.
-			type = PERFCTR_MAX_TYPE;
-		}
-
-		ret = ihk_mc_perfctr_init(event->counter_id, type, mode);
-		ihk_mc_perfctr_set(event->counter_id, event->sample_freq * -1);
-		ihk_mc_perfctr_start(1UL << event->counter_id);
-	
-	} else if(attr->type == PERF_TYPE_RAW) {
-		// PAPI_REF_CYC counted by fixed counter
-		if(event->counter_id >= X86_IA32_BASE_FIXED_PERF_COUNTERS) {
-			ret = ihk_mc_perfctr_fixed_init(event->counter_id, mode);
-			ihk_mc_perfctr_set(event->counter_id, event->sample_freq * -1);
-			ihk_mc_perfctr_start(1UL << event->counter_id);
-			return ret;
-		}
-
 		ret = ihk_mc_perfctr_init_raw(event->counter_id, attr->config, mode);
-		ihk_mc_perfctr_set(event->counter_id, event->sample_freq * -1);
 		ihk_mc_perfctr_start(1UL << event->counter_id);
-	} else {
+	}
-		// Not supported type.
+	else if(event->counter_id >= X86_IA32_BASE_FIXED_PERF_COUNTERS &&
+		event->counter_id < X86_IA32_BASE_FIXED_PERF_COUNTERS + X86_IA32_NUM_FIXED_PERF_COUNTERS) {
+		ret = ihk_mc_perfctr_fixed_init(event->counter_id, mode);
+		ihk_mc_perfctr_start(1UL << event->counter_id);
+	}
+	else {
 		ret = -1;
 	}
 		
@@ -2700,12 +2689,15 @@ perf_counter_start(struct mc_perf_event *event)
 unsigned long perf_event_read_value(struct mc_perf_event *event)
 {
 	unsigned long rtn_count = 0;
+	unsigned long pmc_count = 0;
 	int counter_id = event->counter_id;
 
-	if(event->pid == 0)
+	if(event->pid == 0) {
-		event->count = ihk_mc_perfctr_read(counter_id);
+		pmc_count = ihk_mc_perfctr_read(counter_id) + event->attr.sample_freq;
+		pmc_count &= 0x000000ffffffffffL; // 40bit MASK
+	}
 
-	rtn_count += event->count;
+	rtn_count += event->count + pmc_count;
 
 	if(event->attr.inherit)
 		rtn_count += event->child_count_total;
@@ -2922,11 +2914,21 @@ perf_ioctl(struct mckfd *sfd, ihk_mc_user_context_t *ctx)
                 break;
         case PERF_EVENT_IOC_RESET:
 		// TODO: reset other process
-		ihk_mc_perfctr_reset(counter_id);
+		ihk_mc_perfctr_set(counter_id, event->attr.sample_freq * -1);
+		event->count = 0L;
                 break;
         case PERF_EVENT_IOC_REFRESH:
 		// TODO: refresh other process
-		ihk_mc_perfctr_set(counter_id, event->sample_freq * -1);
+		
+		// not supported on inherited events
+		if(event->attr.inherit)
+			return -EINVAL;
+
+		event->count += event->attr.sample_freq;
+		ihk_mc_perfctr_set(counter_id, event->attr.sample_freq * -1);
+
+		perf_start(event);
+
 		break;
 	default :
 		return -1;
@@ -2945,6 +2947,28 @@ perf_close(struct mckfd *sfd, ihk_mc_user_context_t *ctx)
 	return 0;
 }
 
+static int
+perf_fcntl(struct mckfd *sfd, ihk_mc_user_context_t *ctx)
+{
+	int cmd = ihk_mc_syscall_arg1(ctx);
+	long arg = ihk_mc_syscall_arg2(ctx);
+	int rc = 0;
+
+	switch(cmd) {
+	case 10: // F_SETSIG
+		sfd->sig_no = arg;
+		break;
+	case 0xf: // F_SETOWN_EX
+		break;
+	default : 
+		break;
+	}
+
+	rc = syscall_generic_forwarding(__NR_fcntl, ctx);
+
+	return rc;
+}
+
 static long
 perf_mmap(struct mckfd *sfd, ihk_mc_user_context_t *ctx)
 {
@@ -2963,6 +2987,7 @@ perf_mmap(struct mckfd *sfd, ihk_mc_user_context_t *ctx)
 
 	// setup perf_event_mmap_page
 	page = (struct perf_event_mmap_page *)rc;
+	page->data_head = 16;
 	page->cap_user_rdpmc = 1;
 
 	return rc;
@@ -3014,7 +3039,7 @@ SYSCALL_DECLARE(perf_event_open)
 
 	event->sample_freq = attr->sample_freq;
 	event->nr_siblings = 0;
-	event->count = 0;
+	event->count = 0L;
 	event->child_count_total = 0;
 	event->parent = NULL;
 	event->pid = pid;
@@ -3050,10 +3075,12 @@ SYSCALL_DECLARE(perf_event_open)
 	if(!sfd)
 		return -ENOMEM;
 	sfd->fd = fd;
+	sfd->sig_no = -1;
 	sfd->read_cb = perf_read;
 	sfd->ioctl_cb = perf_ioctl;
 	sfd->close_cb = perf_close;
 	sfd->mmap_cb = perf_mmap;
+	sfd->fcntl_cb = perf_fcntl;
 	sfd->data = (long)event;
 	irqstate = ihk_mc_spinlock_lock(&proc->mckfd_lock);
 

lib/include/ihk/perfctr.h

@@ -54,11 +54,11 @@ int ihk_mc_perfctr_start(unsigned long counter_mask);
 int ihk_mc_perfctr_stop(unsigned long counter_mask);
 int ihk_mc_perfctr_fixed_init(int counter, int mode);
 int ihk_mc_perfctr_reset(int counter);
-int ihk_mc_perfctr_set(int counter, unsigned long value);
+int ihk_mc_perfctr_set(int counter, long value);
 int ihk_mc_perfctr_read_mask(unsigned long counter_mask, unsigned long *value);
 unsigned long ihk_mc_perfctr_read(int counter);
 unsigned long ihk_mc_perfctr_read_msr(int counter);
-int ihk_mc_perfctr_alloc_counter(unsigned long pmc_status);
+int ihk_mc_perfctr_alloc_counter(unsigned int *type, unsigned long *config, unsigned long pmc_status);
 
 #endif