support SA_RESTART flag and restart syscall

arch/x86/kernel/cpu.c

@@ -604,7 +604,7 @@ void setup_x86_ap(void (*next_func)(void))
 
 void arch_show_interrupt_context(const void *reg);
 void set_signal(int sig, void *regs, struct siginfo *info);
-void check_signal(unsigned long rc, void *regs);
+void check_signal(unsigned long, void *, int);
 extern void tlb_flush_handler(int vector);
 
 void handle_interrupt(int vector, struct x86_user_context *regs)
@@ -678,7 +678,7 @@ void handle_interrupt(int vector, struct x86_user_context *regs)
 		}
 	}
 
-	check_signal(0, regs);
+	check_signal(0, regs, 0);
 	check_need_resched();
 }
 
@@ -691,7 +691,7 @@ void gpe_handler(struct x86_user_context *regs)
 		panic("gpe_handler");
 	}
 	set_signal(SIGSEGV, regs, NULL);
-	check_signal(0, regs);
+	check_signal(0, regs, 0);
 	check_need_resched();
 	// panic("GPF");
 }
@@ -719,7 +719,7 @@ void debug_handler(struct x86_user_context *regs)
 	memset(&info, '\0', sizeof info);
 	info.si_code = si_code;
 	set_signal(SIGTRAP, regs, &info);
-	check_signal(0, regs);
+	check_signal(0, regs, 0);
 	check_need_resched();
 }
 
@@ -736,7 +736,7 @@ void int3_handler(struct x86_user_context *regs)
 	memset(&info, '\0', sizeof info);
 	info.si_code = TRAP_BRKPT;
 	set_signal(SIGTRAP, regs, &info);
-	check_signal(0, regs);
+	check_signal(0, regs, 0);
 	check_need_resched();
 }
 

arch/x86/kernel/include/syscall_list.h

@@ -26,6 +26,7 @@ SYSCALL_DELEGATED(2, open)
 SYSCALL_HANDLED(3, close)
 SYSCALL_DELEGATED(4, stat)
 SYSCALL_DELEGATED(5, fstat)
+SYSCALL_DELEGATED(7, poll)
 SYSCALL_DELEGATED(8, lseek)
 SYSCALL_HANDLED(9, mmap)
 SYSCALL_HANDLED(10, mprotect)
@@ -39,6 +40,7 @@ SYSCALL_DELEGATED(17, pread64)
 SYSCALL_DELEGATED(18, pwrite64)
 SYSCALL_DELEGATED(20, writev)
 SYSCALL_DELEGATED(21, access)
+SYSCALL_DELEGATED(23, select)
 SYSCALL_HANDLED(24, sched_yield)
 SYSCALL_HANDLED(25, mremap)
 SYSCALL_HANDLED(26, msync)
@@ -48,6 +50,7 @@ SYSCALL_HANDLED(29, shmget)
 SYSCALL_HANDLED(30, shmat)
 SYSCALL_HANDLED(31, shmctl)
 SYSCALL_HANDLED(34, pause)
+SYSCALL_DELEGATED(35, nanosleep)
 SYSCALL_HANDLED(39, getpid)
 SYSCALL_HANDLED(56, clone)
 SYSCALL_DELEGATED(57, fork)
@@ -57,7 +60,10 @@ SYSCALL_HANDLED(60, exit)
 SYSCALL_HANDLED(61, wait4)
 SYSCALL_HANDLED(62, kill)
 SYSCALL_DELEGATED(63, uname)
+SYSCALL_DELEGATED(65, semop)
 SYSCALL_HANDLED(67, shmdt)
+SYSCALL_DELEGATED(69, msgsnd)
+SYSCALL_DELEGATED(70, msgrcv)
 SYSCALL_DELEGATED(72, fcntl)
 SYSCALL_DELEGATED(79, getcwd)
 SYSCALL_DELEGATED(89, readlink)
@@ -104,18 +110,25 @@ SYSCALL_DELEGATED(201, time)
 SYSCALL_HANDLED(202, futex)
 SYSCALL_HANDLED(203, sched_setaffinity)
 SYSCALL_HANDLED(204, sched_getaffinity)
+SYSCALL_DELEGATED(208, io_getevents)
 SYSCALL_HANDLED(216, remap_file_pages)
 SYSCALL_DELEGATED(217, getdents64)
 SYSCALL_HANDLED(218, set_tid_address)
+SYSCALL_DELEGATED(220, semtimedop)
+SYSCALL_DELEGATED(230, clock_nanosleep)
 SYSCALL_HANDLED(231, exit_group)
+SYSCALL_DELEGATED(232, epoll_wait)
 SYSCALL_HANDLED(234, tgkill)
 SYSCALL_HANDLED(237, mbind)
 SYSCALL_HANDLED(238, set_mempolicy)
 SYSCALL_HANDLED(239, get_mempolicy)
 SYSCALL_HANDLED(247, waitid)
 SYSCALL_HANDLED(256, migrate_pages)
+SYSCALL_DELEGATED(270, pselect6)
+SYSCALL_DELEGATED(271, ppoll)
 SYSCALL_HANDLED(273, set_robust_list)
 SYSCALL_HANDLED(279, move_pages)
+SYSCALL_DELEGATED(281, epoll_pwait)
 SYSCALL_HANDLED(282, signalfd)
 SYSCALL_HANDLED(289, signalfd4)
 #ifdef DCFA_KMOD

arch/x86/kernel/syscall.c

@@ -30,6 +30,7 @@ int copy_from_user(void *dst, const void *src, size_t siz);
 int copy_to_user(void *dst, const void *src, size_t siz);
 int write_process_vm(struct process_vm *vm, void *dst, const void *src, size_t siz);
 long do_sigaction(int sig, struct k_sigaction *act, struct k_sigaction *oact);
+long syscall(int num, ihk_mc_user_context_t *ctx);
 extern void save_fp_regs(struct process *proc);
 
 //#define DEBUG_PRINT_SC
@@ -152,6 +153,8 @@ struct sigsp {
 	unsigned long sigrc;
 	unsigned long sigmask;
 	int ssflags;
+	int num;
+	int restart;
 	siginfo_t info;
 };
 
@@ -160,18 +163,19 @@ SYSCALL_DECLARE(rt_sigreturn)
 	struct process *proc = cpu_local_var(current);
 	struct x86_user_context *regs;
 	struct sigsp *sigsp;
-	long rc = -EFAULT;
 
 	asm("movq %%gs:132, %0" : "=r" (regs));
 	--regs;
 
 	sigsp = (struct sigsp *)regs->gpr.rsp;
+	if(copy_from_user(regs, &sigsp->regs, sizeof(struct x86_user_context)))
+		return -EFAULT;
 	proc->sigmask.__val[0] = sigsp->sigmask;
 	proc->sigstack.ss_flags = sigsp->ssflags;
-	if(copy_from_user(regs, &sigsp->regs, sizeof(struct x86_user_context)))
+	if(sigsp->restart){
-		return rc;
+		return syscall(sigsp->num, (ihk_mc_user_context_t *)regs);
-	copy_from_user(&rc, &sigsp->sigrc, sizeof(long));
+	}
-	return rc;
+	return sigsp->sigrc;
 }
 
 extern struct cpu_local_var *clv;
@@ -495,9 +499,41 @@ void ptrace_report_signal(struct process *proc, int sig)
 	schedule();
 	dkprintf("ptrace_report_signal,wake up\n");
 }
+static int
+isrestart(int num, unsigned long rc, int sig, int restart)
+{
+	if(num == 0 || rc != -EINTR)
+		return 0;
+	switch(num){
+	    case __NR_pause:
+	    case __NR_rt_sigsuspend:
+	    case __NR_rt_sigtimedwait:
+//	    case __NR_rt_sigwaitinfo:
+	    case __NR_epoll_wait:
+	    case __NR_epoll_pwait:
+	    case __NR_poll:
+	    case __NR_ppoll:
+	    case __NR_select:
+	    case __NR_pselect6:
+	    case __NR_msgrcv:
+	    case __NR_msgsnd:
+	    case __NR_semop:
+	    case __NR_semtimedop:
+	    case __NR_clock_nanosleep:
+	    case __NR_nanosleep:
+//	    case __NR_usleep:
+	    case __NR_io_getevents:
+		return 0;
+	}
+	if(sig == SIGCHLD)
+		return 1;
+	if(restart)
+		return 1;
+	return 0;
+}
 
 void
-do_signal(unsigned long rc, void *regs0, struct process *proc, struct sig_pending *pending)
+do_signal(unsigned long rc, void *regs0, struct process *proc, struct sig_pending *pending, int num)
 {
 	struct x86_user_context *regs = regs0;
 	struct k_sigaction *k;
@@ -564,6 +600,10 @@ do_signal(unsigned long rc, void *regs0, struct process *proc, struct sig_pendin
 		}
 		sigsp->sigmask = mask;
 		sigsp->ssflags = ssflags;
+		sigsp->num = num;
+		sigsp->restart = isrestart(num, rc, sig, k->sa.sa_flags & SA_RESTART);
+		if(num != 0 && rc == -EINTR && sig == SIGCHLD)
+			sigsp->restart = 1;
 		memcpy(&sigsp->info, &pending->info, sizeof(siginfo_t));
 
 		usp = (unsigned long *)sigsp;
@@ -697,6 +737,9 @@ getsigpending(struct process *proc, int delflag){
 	struct sig_pending *pending;
 	__sigset_t w;
 	int	irqstate;
+	__sigset_t x;
+	int sig;
+	struct k_sigaction *k;
 
 	w = proc->sigmask.__val[0];
 
@@ -705,11 +748,18 @@ getsigpending(struct process *proc, int delflag){
 	for(;;){
 		irqstate = ihk_mc_spinlock_lock(lock);
 		list_for_each_entry_safe(pending, next, head, list){
-			if(!(pending->sigmask.__val[0] & w)){
+			for(x = pending->sigmask.__val[0], sig = 0; x; sig++, x >>= 1);
-				if(delflag)
+			k = proc->sighandler->action + sig - 1;
-					list_del(&pending->list);
+			if(delflag ||
-				ihk_mc_spinlock_unlock(lock, irqstate);
+			   (sig != SIGCHLD && sig != SIGURG) ||
-				return pending;
+			   (k->sa.sa_handler != (void *)1 &&
+			    k->sa.sa_handler != NULL)){
+				if(!(pending->sigmask.__val[0] & w)){
+					if(delflag)
+						list_del(&pending->list);
+					ihk_mc_spinlock_unlock(lock, irqstate);
+					return pending;
+				}
 			}
 		}
 		ihk_mc_spinlock_unlock(lock, irqstate);
@@ -730,7 +780,7 @@ hassigpending(struct process *proc)
 }
 
 void
-check_signal(unsigned long rc, void *regs0)
+check_signal(unsigned long rc, void *regs0, int num)
 {
 	struct x86_user_context *regs = regs0;
 	struct process *proc;
@@ -770,7 +820,7 @@ check_signal(unsigned long rc, void *regs0)
 			return;
 		}
 
-		do_signal(rc, regs, proc, pending);
+		do_signal(rc, regs, proc, pending, num);
 	}
 }
 

kernel/mem.c

@@ -171,7 +171,7 @@ static struct ihk_mc_interrupt_handler query_free_mem_handler = {
 };
 
 void set_signal(int sig, void *regs, struct siginfo *info);
-void check_signal(unsigned long rc, void *regs);
+void check_signal(unsigned long, void *, int);
 int gencore(struct process *, void *, struct coretable **, int *);
 void freecore(struct coretable **);
 
@@ -416,7 +416,7 @@ static void page_fault_handler(void *fault_addr, uint64_t reason, void *regs)
 			info._sifields._sigfault.si_addr = fault_addr;
 			set_signal(SIGSEGV, regs, &info);
 		}
-		check_signal(0, regs);
+		check_signal(0, regs, 0);
 		goto out;
 	}
 

kernel/syscall.c

@@ -94,7 +94,7 @@ static char *syscall_name[] MCKERNEL_UNUSED = {
 #undef	SYSCALL_DELEGATED
 };
 
-void check_signal(unsigned long rc, void *regs);
+void check_signal(unsigned long, void *, int);
 void do_signal(long rc, void *regs, struct process *proc, struct sig_pending *pending);
 extern unsigned long do_kill(int pid, int tid, int sig, struct siginfo *info, int ptracecont);
 extern struct sigpending *hassigpending(struct process *proc);
@@ -6059,7 +6059,7 @@ long syscall(int num, ihk_mc_user_context_t *ctx)
 		l = syscall_generic_forwarding(num, ctx);
 	}
 
-	check_signal(l, NULL);
+	check_signal(l, NULL, num);
 	check_need_resched();
 
 	if (cpu_local_var(current)->ftn->ptrace) {