Skip to content

Commit

Permalink
CRISv32: don't attempt syscall restart on irq exit
Browse files Browse the repository at this point in the history
r9 is used to determine whether syscall restarting must be performed or
not.  Unfortunately, r9 is never set to zero in the non-syscall path,
and r9 is on top of that a callee-saved register which can be set to
non-zero by the C functions that are called during IRQ handling.

This means that if r10 (used for the syscall return value) is one of the
-ERESTART* values when a hardware interrupt occurs which leads to a
signal being delivered to the process, the kernel will "restart" a
syscall which never occurred.  This will lead to the PC being moved back
by 2 on return to user space.

Fix the problem by setting r9 to zero in the interrupt path.

Test case (should loop forever but ends up executing the break 8 trap
instruction):

  #include <signal.h>
  #include <stdlib.h>
  #include <sys/time.h>

  void f(int n)
  {
  	register int r9 asm ("r9") = 1;
  	register int r10 asm ("r10") = n;

          __asm__ __volatile__(
  		"ba	1f	\n"
  		"nop		\n"
  		"break	8	\n"
  		"1: ba	.	\n"
  		"nop		\n"
  		:
  		: "r" (r9), "r" (r10)
  		: "memory");
  }

  void handler1(int sig) { }

  int main(int argc, char *argv[])
  {
          struct itimerval t1 = { .it_value = {1} };

          signal(SIGALRM, handler1);
          setitimer(ITIMER_REAL, &t1, NULL);

          f(-513); /* -ERESTARTNOINTR */

          return 0;
  }

Signed-off-by: Rabin Vincent <rabin@rab.in>
Signed-off-by: Jesper Nilsson <jespern@axis.com>
  • Loading branch information
Rabin Vincent authored and Jesper Nilsson committed Mar 25, 2015
1 parent 1eb1390 commit db4a35c
Showing 1 changed file with 2 additions and 0 deletions.
2 changes: 2 additions & 0 deletions arch/cris/arch-v32/kernel/entry.S
Original file line number Diff line number Diff line change
Expand Up @@ -99,6 +99,8 @@ ret_from_kernel_thread:

.type ret_from_intr,@function
ret_from_intr:
moveq 0, $r9 ; not a syscall

;; Check for resched if preemptive kernel, or if we're going back to
;; user-mode. This test matches the user_regs(regs) macro. Don't simply
;; test CCS since that doesn't necessarily reflect what mode we'll
Expand Down

0 comments on commit db4a35c

Please sign in to comment.